content
stringlengths 1
15.9M
|
|---|
\section{}
\label{}
\section{}
\label{}
\section{}
\label{}
\section{Introduction}
Suppose a number of mobile agents are moving on a row. Some of the agents need to travel left, and the other agents need to travel right to arrive at their destination. The agents are not allowed to collide, but have access to another adjacent, initially empty row that they can use to manoeuvre past each other. What is the most efficient way for the agents to achieve their goal and arrive at their desired left-side or right-side destinations?
In this work we study a discrete formalization of this problem. Given a collection of red and blue mobile agents located on parallel grid rows of equal length, we seek a centralized algorithm to move all the blue agents to the far left side and all the red agents to the far right side columns, thus \textit{sorting} the agents according to color (see \cref{fig:typical.configs}). We assume all agents are initially located on the bottom row. Agents can move simultaneously at discrete time steps $T=0, 1, \ldots$, and must not collide with each other (two agents may never occupy or attempt to move to the same location). Our goal is to design a centralized algorithm that controls the agents so as to sort them in the least number of time steps.
This problem first arose as part of the authors' ongoing research into traffic management algorithms for self-driving vehicles on a freeway. In the context of traffic management, the two rows represent a moving subsection of an upwards-facing freeway. The subsection tracks a set of vehicles all driving at the same forward velocity. We assume that freeway exits might be placed to the left or right of the road. Thus, it is necessary to shift all vehicles that need to exit the freeway leftwards or rightwards ahead of time depending on their desired exit direction, i.e., to ``sort'' them (see \cref{fig:car_frame}). In this work we focus on the special case where the top row is initially empty, which we believe to be independently interesting.
\begin{figure}[!ht]
\begin{subfigure}{.48\textwidth}
\centering
\scalebox{1.1}{\import{./images/}{road.frame.cars.only}}
\caption{}
\label{sub:unsorted.cars}
\end{subfigure}
\hfill
\begin{subfigure}{.48\textwidth}
\centering
\scalebox{1.1}{\import{./images/}{road.frame.cars.sorted}}
\caption{}
\label{sub:sorted.cars}
\end{subfigure}
\caption{Vehicles driving on a freeway. The blue vehicles want to exit the freeway via an upcoming left exit (not illustrated), and the red vehicles want to exit the freeway via an upcoming right exit. Vehicles need to shift their position to the left or right ahead of time to prepare for exiting the freeway. (\subref{sub:unsorted.cars}) shows an unsorted configuration, and (\subref{sub:sorted.cars}) shows the sorted configuration, after the vehicles have adjusted their positions.}
\label{fig:car_frame}
\end{figure}
Similar problems in vehicular control, warehouse management, and combinatorial puzzles have been investigated in the literature and might broadly be referred to as ``physical sorting'' problems \cite{zhou2016asymmetric, krupke2015parallel, litus2010fall, shang2016approach, shang2014swarm,ratner1990n2,johnson1879notes}. In physical sorting problems, a collection of mobile agents that occupy physical space must bypass each other without colliding in order to arrive at some predefined sorted configuration.
A trivial algorithm that accomplishes our agents' sorting task is the following: move all the agents of one color (say, red) to the top row; then let the red agents move right and the blue agents move left in their respective row (\cref{fig:2Delta images}, a-e). This algorithm does not require complex computation nor even centralized decision-making--it is a straightforward, decentralized, local strategy that can be executed by simple autonomous agents without requiring any global knowledge on the agent configuration.
\begin{figure}
\centering
\begin{subfigure}{.48\textwidth}
\resizebox{\textwidth}{!}{\import{./images}{Configuration.Typical.Initial}}
\caption{}
\label{sub:initial.config}
\end{subfigure}
\hfill
\begin{subfigure}{.48\textwidth}
\resizebox{\textwidth}{!}{\import{./images}{Configuration.Typical.Final}}
\caption{}
\label{sub:final.config}
\end{subfigure}
\caption{(\subref{sub:initial.config}) illustrates an initial configuration of agents. (\subref{sub:final.config}) illustrates a sorted configuration (there are many possible such configurations).}
\label{fig:typical.configs}
\end{figure}
How does the above simple algorithm fare compared to an optimal centralized sorting algorithm? A priori, since the number of possible strategies for the agents is enormous, one would expect far better sorting strategies are available. However, in this work, we shall prove the surprising result that this ``trivial'' distributed strategy is at most one time step short of optimal for a very large class of initial agent configurations called \textit{normal configurations} (configurations where the leftmost agent wants to go right and the rightmost agent wants to go left), and is in fact optimal over such configurations assuming we choose the correct color to move to the spare row.
In the general case where we also consider non-normal configurations, we show that the optimal makespan is determined by the maximal normal subconfiguration, and find a provably optimal sorting algorithm for the agents (\cref{NonNormalConfigurationSection}). Furthermore, we derive an \textbf{exact} lower bound for the amount of time it takes to sort the agents given any starting configuration (\cref{theorem:makespanresult}). Our proposed optimal algorithm matches this lower bound, thus it is instance optimal in the sense of \cite{afshani2017instance}, attaining the best possible sorting time for any initial configuration.
The algorithm (\cref{algo:general.movement}) can be understood as blending two strategies: inside the maximal normal subconfiguration, we split the agents into rows based on their desired direction of motion, according to the aforementioned ``trivial'' strategy. Outside this subconfiguration, agents split between both rows to move faster regardless of their color (\cref{figure:subnormalalgoexample}). The idea behind this algorithm is simple to understand, but the implementation requires several delicate caveats which are discussed throughout this work.
Physical sorting can sometimes be unintuitive: the addition of a single agent can completely change the optimal strategy and double the time to completion of the sorting (\cref{fig:2Delta images}). Furthermore, when attempting to derive lower bounds on the makespan, the infinite set of strategies available to the agents makes it difficult to keep track of each agent, which greatly complicates the mathematical proofs. We overcome these complexities by identifying a small set of \textit{critical agents} (see \cref{definition:critical-agents}) and showing, roughly, that only the movements of the critical agents at a given time step can affect any algorithm's makespan.
\begin{figure}[!ht]
\centering
\begin{minipage}{0.48\textwidth}
\foreach \tick in {0, ..., 4}
{
\begin{subfigure}{\textwidth}
\resizebox{\textwidth}{!}{\import{./images}{Example.2Delta}}
\caption{$t=\tick$}
\label{sub:Example.2Delta.\tick}
\end{subfigure}
}
\end{minipage}
\hfill
\begin{minipage}{0.48\textwidth}
\foreach \tick in {0, ..., 4}
{
\begin{subfigure}{\textwidth}
\resizebox{\textwidth}{!}{\import{./images}{Example.Delta+1}}
\caption{$t=\tick$}
\label{sub:Example.Delta+1.\tick}
\end{subfigure}
}
\end{minipage}
\caption{(\subref{sub:Example.2Delta.0})-(\subref{sub:Example.2Delta.4}) illustrate the first five time steps of an optimal strategy for sorting a normal initial configuration with a single blue agent placed in front of a red agent group. No sorting is possible in faster than $2\Delta$ ticks. Subfigures
(\subref{sub:Example.Delta+1.0})-(\subref{sub:Example.Delta+1.4}) illustrate non-normal initial configuration that can be sorted in $\Delta + 1$ ticks via splitting the agents between the rows in an alternating fashion (\cref{algo:general.movement}). The first five time steps of this splitting strategy are shown.
Both initial configurations share a very long sequence of red agents. The only difference is a lone blue agent. However, the optimal sorting times are strikingly different.}
\label{fig:2Delta images}
\end{figure}
We believe that the expression for the lower bound obtained in \cref{theorem:makespanresult} is independently interesting. As an example application, \cref{corollary:singlerowtime}, which is a small special case of the expression, computes the amount of time it takes for a ``traffic jam'' of agents on a single row to get to the right (left) side of the row, assuming each agent moves right (left) at every time step where there is no agent in front of them (see \cref{fig:One row TASEP - like}). This corollary relates to the study of the \textit{totally asymmetric simple exclusion process} in statistical mechanics, which has been used to study transport phenomena such as traffic flow and biological transport \cite{chou2011biologicaltasep,kriecherbauer2010pedestrian,chowdhury2000statistical}. The lower bound of \cref{theorem:makespanresult} can be applied to compute the time it takes an arbitrary TASEP particle configuration with synchronized waking times to get from one end of the row to the other, assuming $p_{right} = 1$ (probability of moving right) and $p_{left} = 0$. Although this corollary is fairly straightforward, we could not find a similar result in the TASEP literature.
\begin{figure}
\centering
\foreach \tick in {0, ..., 4}
{
\begin{subfigure}{0.7\textwidth}
\resizebox{\textwidth}{!}{\import{./images}{Example.OneRow}}
\caption{$t=\tick$}
\label{sub:Example.OneRow.\tick}
\end{subfigure}
}
\caption{Subfigures (\subref{sub:Example.OneRow.0})-(\subref{sub:Example.OneRow.4}) illustrate a single row model with agents moving in one direction. The model could be viewed as a TASEP with $p_{\text{right}} = 1$ and $p_{\text{left}}=0$. In the provided example, all agents reach the far-right side on the $9$th time step, which is equal to what we define as the $f_{max}$ value of the configuration (see \cref{def:f} and \cref{corollary:singlerowtime})}
\label{fig:One row TASEP - like}
\end{figure}
\section{Related Work}
Various ``physical sorting'' problems appear in the literature \cite{zhou2016asymmetric, krupke2015parallel, litus2010fall, shang2016approach, shang2014swarm} with applications to warehouse logistics, where loads must be efficiently moved to different ends of a warehouse (in particular, our assumptions resemble those of puzzle based storage systems \cite{gue2007puzzlebasedstorage}); servicing, where mobile robots self-sort according to some priority ordering; assembling, where tasks should be carried in a predefined order which respects physical space; and transportation. In such problems, mobile agents (robots, vehicles, etc.) that take up physical space are required to attain some kind of sorted configuration while avoiding collisions and maneuvering around each other, making their decisions in either a distributed \cite{zhou2016asymmetric} or centralized \cite{petig2018changing} fashion. For example, similar to this work, in \cite{petig2018changing}, the authors consider a task of sorting mobile vehicles in two rows, developing approximation algorithms and proving computational hardness results. Their goal, however, is to \textit{split} the vehicles between the rows according to their color as fast as possible, as opposed to our goal of bringing robots to the left-hand or right-hand side depending on their color.
The famous ``15 puzzle", where numbered tiles must be slid across a grid until they are ordered, can also be seen as a physical sorting task for which algorithms and computational hardness results are available \cite{johnson1879notes,ratner1990n2}.
Lane changing has been intensively studied in the past in many traffic models \cite{hatipoglu1995optimal,chee1994vehicle,naranjo2008lane,atagoziyev2016lane}, with recent results in both centralized and distributed control protocols \cite{visintainer2016towards, ekenstedt2019membership,schubert2010situation}.
As previously mentioned, a side application of our results (\cref{corollary:singlerowtime}) relates to \textit{totally simple exclusion processes} (TASEPs) in statistical mechanics \cite{chou2011biologicaltasep,kriecherbauer2010pedestrian,chowdhury2000statistical}. TASEPs have been used, in particular, as an idealized model that captures many of the phenomena of real-world vehicular traffic. The general model we present in this work is, however, not directly related to any TASEP model, since we study deterministic mobile agents that act according to an intelligent, centralized algorithm, rather than a predetermined stochastic process.
Broadly speaking, our mathematical modelling of the problem (\cref{section:model}) also relates to various discrete grid-like settings in multi-robot and multi-agent systems, wherein a large number of robots whose spatial locations are represented as coordinates on a grid-like region all move synchronously according to a local or centralized algorithm while avoiding collisions. The goal of agents in these settings can be, for example, fast deployment, gathering, or formation on an a priori unknown grid environment \cite{arxivminimizingtravel,uniformdispersion,barraswarm1,altshuler2018introduction,siamcomputing_dieudonne2015meetagentsasynchronously,siamcomputing_fujinaga2015patternmobilerobot,hideg2020asynchronous}.
\section{Model}
\label{section:model}
We are given $n$ mobile agents on a $2 \times m$ grid environment, such that $n_1$ agents are red and $n_2$ agents are blue ($n_1 + n_2 = n$). Every agent begins in an $(x,y)$ coordinate of the form $(\cdot, 1)$ (see \cref{sub:initial.config}), starting at $x=1$. A column of the environment is called \textit{red-occupied} if it contains at least one red agent and no blue agents, \textit{blue-occupied} if it contains at least one blue agent and no red agents, \textit{mixed} if it contains both colors of agents, and \textit{empty} otherwise. The goal of the agents is to move to a configuration such that the red-occupied columns are the rightmost columns, and the blue-occupied columns are the leftmost columns. Formally, the following conditions must be fulfilled:
\begin{enumerate}
\item There are no mixed columns.
\item Every blue-occupied column is to the left of every red-occupied column and every empty column.
\item Every red-occupied column is to the right of every blue-occupied column and every empty column.
\end{enumerate}
When the agents achieve such a configuration we say the system is \textit{sorted} (see \cref{sub:final.config}), as it separates all agents in the configuration to the far left or far right columns of the environment based on color. It might be desirable to require, in addition to the above three conditions, that all agents end up on the bottom row (just as they began): in all algorithms we present here, this can be achieved at the cost of at most one additional time step that moves all the robots downwards, and analogous optimality results hold.
We define a model of agent motion based on common assumptions in synchronous and semi-synchronous models of traffic flow \cite{chou2011biologicaltasep}, biological systems \cite{amir2022locust,amir2020discretelocusts}, and the theory of mobile multi-robot systems \cite{arxivminimizingtravel,uniformdispersion,barraswarm1,altshuler2018introduction,siamcomputing_dieudonne2015meetagentsasynchronously,siamcomputing_fujinaga2015patternmobilerobot}.
Time is discretized into steps $t = 0, 1, \ldots$. At every time step, agents may move to adjacent empty locations (up, down, left, or right, non-diagonally). Agents cannot move to a location that already contains an agent, nor can two agents move to the same empty location at the same time.
Note that, according to the above assumptions, a robot can only move into a location that is unoccupied, but two or more adjacent agents followed by an empty location cannot both move right (or left) in the same time step, meaning that such situations result in a traffic jam or ``shock'': the agents at the leftmost (rightmost) end of the line must wait several time steps for empty space to be created between them and the rest of the agents (see the first column of \cref{fig:2Delta images}). This is a common assumption in the literature on traffic flow, reflecting the fact that adjacent agents (even those receiving commands transmitted by a central algorithm) do not have perfectly coordinated clocks nor perfect motion detection systems and thus cannot begin moving at the exact same time and at the exact same speed without risking collisions.
The \textit{beginning} of a time step refers to the configuration before any agent movements, and the \textit{end} of the time step refers to the configuration after agent movements (so the configuration at the beginning of time $t+1$ is the same as at the end of time $t$). Unless explicitly stated otherwise, anywhere in this work, when we refer to the agents' configuration ``at time $t$", we mean the configuration at the beginning of that time step.
The agents' actions are controlled by a sorting algorithm. The number of time steps it takes an algorithm to move the agents from the initial configuration to a sorted configuration is called the \textit{makespan} of that algorithm with respect to the initial configuration (i.e., the makespan is the first $t$ such that the configuration is sorted at the beginning of time step $t$).
\begin{definition}\label{def:normal-configuration}
A \textbf{normal} initial configuration is an initial configuration of agents such that the rightmost agent is blue and the leftmost agent is red.
\end{definition}
Unless stated otherwise, we assume in this work that the initial agent configuration is normal. We will specifically handle non-normal initial configurations in \cref{NonNormalConfigurationSection}. Certain strategies, such as an alternating split between the rows that enables each agent to move horizontally faster, are effective in non-normal configurations but ineffective in normal configurations, thus the optimal strategy can be different. An example is shown in \cref{fig:2Delta images}, which compares a normal configuration to a non-normal configuration and shows an optimal strategy for each case. In \cref{fig:2Delta images}, the addition of a single blue agent completely alters the optimal strategy, and doubles the optimal makespan.
\section{A lower bound on makespan}
In this section, we will prove a lower bound on the makespan of any (centralized or distributed) sorting algorithm for normal configurations. In the next section, we will show this lower bound is tight for normal configurations: there is a centralized algorithm that matches it exactly.
The lower bound is defined as a function of the initial configuration of agents and empty columns. For every agent $A$, we denote $A$'s location at time $t$ as $(A_x(t), A_y(t))$. Let $\mathcal{A} = \mathcal{R} \cup \mathcal{B}$ be the set of all agents, where $\mathcal{R}$ is the set of red agents and $\mathcal{B}$ is the set of blue agents.
\begin{definition}
\label{def:frontback}
For any agent $A$, we define $front(A)$ and $back(A)$.
\begin{enumerate}
\item If $A \in \mathcal{R}$, then $front(A)$ is the set of all \textit{not}-red-occupied columns (i.e., empty or blue-occupied) to the right right of $A$ (i.e., with strictly greater $x$ coordinate) and $back(A)$ is the set of all red-occupied columns to the left of $A$ (i.e., with strictly smaller $x$ coordinate) at time $0$.
\item If $A \in \mathcal{B}$, then $front(A)$ is the set of all \textit{not}-blue-occupied columns to the left of $A$ and $back(A)$ is the set of all blue-occupied columns to the right of $A$ at time $0$.
\end{enumerate}
\end{definition}
\cref{def:frontback} is illustrated in \cref{fig:example: front and back}.
\begin{figure}[!ht]
\centering
\resizebox{.9\textwidth}{!}{\import{./images}{Example.Explanations.Front.n.Back}}
\caption{The $front$ and the $back$ of the red agent $A$ (highlighted in yellow). We see that $f(A) = 8$ (\cref{def:f}).}
\label{fig:example: front and back}
\end{figure}
For any agent $A$ we define the value $f(A)$ which is dependent on the initial configuration. We will show that every value $f(A)$ in an initial configuration is a lower bound on the makespan of \textit{any} sorting algorithm executed over that initial configuration. Thus the maximum $f_{max}$ of all these values is a lower bound as well.
\begin{definition}\label{def:f}
For every agent $A$, we define $f(A)$ as follows (see \cref{fig:example: front and back}):
$$f(A) = |front(A)| + |back(A)|,$$
We further define $f_{max} = \max_{A \in \mathcal{A}}f(A)$.
\end{definition}
\begin{definition} \label{definition:critical-agents}
An agent $A$ for which $f(A) = f_{max}$ is called a \textbf{critical agent}.
\end{definition}
We will show that in many initial configurations, $f_{max}$ is the exact lower bound for the makespan of any sorting algorithm. However, in some situations, such as when the agents are too densely packed together and must waste time without being able to move, it is possible that the algorithm requires one extra time step to sort the agents. To derive an exact lower bound, we will have to take into account this ``plus 1,'' whose preconditions are captured by the following definition:
\begin{definition}
\label{Vmaxdef}
We define $\mathcal{V}$ to equal $1$ if:
\begin{enumerate}
\item There is both a red and a blue critical agent, or
\item There is a critical agent $A$ such that, in the initial configuration, another agent is located immediately in front of it (i.e., at $(A_x(0)+1,1)$ if $A$ is red and $(A_x(0)-1,1)$ if $A$ is blue).
\end{enumerate}
Otherwise, $\mathcal{V}=0$.
\end{definition}
The goal of this section and the next one is to prove the following result:
\begin{theorem}
\label{theorem:makespanresult}
The makespan of any physical sorting algorithm for a given normal initial configuration is at least $f_{max} + \mathcal{V}$. This lower bound is precise; there is an algorithm that achieves sorting in exactly $f_{max}+\mathcal{V}$ time steps.
\end{theorem}
For the sake of the proof, it is helpful to track the relative ordering of red and blue agents based on their $x$ coordinate. To this end, we assign each red agent an integer called a \textit{label}. At time $t=0$, the labels depend on the relative $x$-coordinate position of the agents, such that the \textit{leftmost} red agent has the label $n_1$, the second-leftmost red agent has the label $n_1 - 1$, and so on, with the rightmost red agent having label $1$. Two agents $R$ and $R'$ \textit{exchange} their labels at time $t$ if $R_x(t-1) \leq R_x'(t-1)$ but $R_x(t) > R_x'(t)$ (in other words, they pass each other horizontally). When this happens, $R'$ receives the label $R$ had at time $t-1$, and vice-versa.
We similarly define labels for blue agents, but in the \textit{opposite} order of $x$ coordinates, such that the \textit{rightmost} blue agent receives label $n_2$, and the \textit{leftmost} blue agent receives label $1$. Label exchanging is defined as before.
Note that labels can only be exchanged by agents of the same color. When a pair of red and blue agents pass each other, labels are unchanged.
We define $ord(A, t)$ to equal the agent $A$'s label at time $t$.
Let $\mathcal{P}^B_i(t)$ denote the blue agent $A$ with $ord(A,t) = i$ at time $t$. When $t$ is implicitly obvious, we will simply write $\mathcal{P}^{B}_{i}$. In the text, we will intuitively think of $\mathcal{P}^{B}_i$ \textbf{as having a ``persistent'' identity}, like an agent, and track its position. We will likewise define $P^R_{i}(t)$ for red agents.
\begin{observation}\label{observation:TagSlowMovement}
Let $ord(A, t) = ord(A', t + 1)$, where $A$ and $A'$ are agents of the same color. Then $|A_x(t) - A_x'(t+1)| \leq 1$, i.e. a label can not travel faster than one step horizontally in a single time tick.
\end{observation}
\cref{observation:TagSlowMovement} follows from the fact that agents can only ``exchange" labels by bypassing each other.
We will now prove several claims about blue agents which will establish lower bounds. Symmetrically, all such claims will hold for red agents, which will allow us to wrap up the proof at the end of the section.
We will henceforth assume the agents move according to some makespan-optimal algorithm $\textit{ALG}$. Per \cref{theorem:makespanresult}, our end-goal is to show that $\textit{ALG}$'s makespan is at least $f_{max} + \mathcal{V}$.
\begin{definition}
\label{definition:meeting}
A red agent $R$ and a blue agent $B$ are said to \textbf{meet} at time $t$ if $$\sign{(R_{x}(t-1) - B_{x}(t-1))} \in \{1, -1\}$$ and $$\sign{(R_{x}(t-1) - B_{x}(t-1))} \neq \sign{(R_{x}(t) - B_{x}(t))},$$ where $\sign{(\cdot)}$ denotes the sign ($-1$, $0$ or $1$) of an integer. (See \cref{fig:example.configs}.)
Two labels $P^B$ and $P^R$ are said to meet at time $t$ if their respective agents meet at time $t$.
\end{definition}
\begin{figure}[!ht]
\centering
\resizebox{.45\textwidth}{!}{\import{./images}{Meeting.Last.SameColumn}}
\caption{The possible outcomes of a ``meeting'' between blue and red agents. On the left are two possible states of the agents at time $t-1$ before meeting, and on the right are the possible states they can transition into after meeting.}
\label{fig:example.configs}
\end{figure}
\begin{definition}
\label{definition:agentVvalues}
Define $\mathcal{V}_i^{\textit{B}} = 0$ if during the execution of \textit{ALG}, $\mathcal{P}^{B}_i$ decreases its $x$ coordinate at \textit{every} time step before it arrives at column $i$ for the first time \textbf{and} before it meets $\mathcal{P}^{R}_{n_1}$ . Otherwise let $\mathcal{V}_i^{\textit{B}} = 1$.
Define $\mathcal{V}_i^{\textit{R}} = 0$ if during the execution of \textit{ALG}, $\mathcal{P}^{R}_i$ increases its $x$ coordinate at \textit{every} time step before it arrives at column $n-i+1$ for the first time \textbf{and} before it meets $\mathcal{P}^{B}_{n_2}$. Otherwise let $\mathcal{V}_i^{\textit{R}} = 1$.
\end{definition}
Informally speaking, the values $\mathcal{V}_i^{\textit{B}}$ and $\mathcal{V}_i^{\textit{R}}$ are a measure of whether a given label ever performs a movement that will propagate backwards and slow down all the labels behind it. We will see how this occurs in \cref{lemma:lower.bound}, but the intuitive idea is this: if there is any time step where, say, the label $\mathcal{P}_i^B$ doesn't move on the $x$ axis towards its final destination (which must be the column $i$ or to the left of it), then every label $P_j^{B}$ with $j > i$ might get obstructed by this movement due to a back-propagating slowdown. For technical reasons apparent in the proof of \cref{lemma:lower.bound}, we want to only factor such movements into our bound if they occur \textit{before} $\mathcal{P}_i^B$ meets the leftmost red label $\mathcal{P}^{R}_{n_1}$.
The general plan is to eventually tie \cref{definition:agentVvalues} to the value $\mathcal{V}$ in \cref{Vmaxdef}.
\begin{lemma}\label{lemma:t_icrudelowerboundlemma}
Let $t_i$ be the first time when $\mathcal{P}^B_i$ reaches column $i$, i.e. an agent with label $i$ arrives at column $i$ for the first time. Then $t_i \geq |front(\mathcal{P}^B_i)| + \mathcal{V}_i^{\textit{B}}$.
\end{lemma}
\begin{proof}
Since in the initial configuration $\mathcal{P}^B_i$ sees $i-1$ blue agents in front of it, we have that
\begin{equation*}
\left|(\mathcal{P}^B_i)_x - i\right| = \left|front(\mathcal{P}^B_i)\right|
\end{equation*}
(e.g., see \cref{sub:Lemma.5.12.Initial}). Hence $\mathcal{P}^B_i$ requires at least $|front(\mathcal{P}^B_i)|$ horizontal moves to get to column $i$, and hence requires this many time steps (\cref{observation:TagSlowMovement}).
If $\mathcal{V}_i^{\textit{B}} = 1$ then ${\mathcal{P}^B_i}$ performs a non-leftward movement before arriving at column $i$, and this adds one time step.
\end{proof}
\begin{observation}
\label{observation:labelsmeetoneatatime}
A given label $P^B$ cannot meet more than one label at any time step.
\end{observation}
\begin{proof}
A given agent $A$ cannot meet more than one agent at any time step, since there are only two rows in the grid (see \cref{fig:example.configs}). Due to \cref{observation:TagSlowMovement}, the same is true of labels.
\end{proof}
We now proceed to the proof of the main technical lemma of this section.
\begin{figure}
\centering
\begin{minipage}{0.88\textwidth}
\begin{subfigure}{\textwidth}
\resizebox{\textwidth}{!}{\import{./images}{Lemma.5.12.Initial}}
\caption{}
\label{sub:Lemma.5.12.Initial}
\end{subfigure}
\par\bigskip
\begin{subfigure}{\textwidth}
\resizebox{\textwidth}{!}{\import{./images}{Lemma.5.12.Case.1}}
\caption{}
\label{sub:Lemma.5.12.Case.1}
\end{subfigure}
\par\bigskip
\begin{subfigure}{\textwidth}
\resizebox{\textwidth}{!}{\import{./images}{Lemma.5.12.Case.2}}
\caption{}
\label{sub:Lemma.5.12.Case.2}
\end{subfigure}
\end{minipage}
\caption{\cref{lemma:lower.bound}. (\subref{sub:Lemma.5.12.Initial}) At time $t=0$, $\mathcal{P}^B_i$ is at distance $|front(\mathcal{P})|$ from column $i$. (\subref{sub:Lemma.5.12.Case.1}) Case 1: $C_i \leq i$. $\mathcal{P}^R_{n_1}$ must meet at least $|back(\mathcal{P}^i_B)|$ blue agents after meeting $\mathcal{P}^i_B$. (\subref{sub:Lemma.5.12.Case.2}) Case 2: $C_i > i$. $\mathcal{P}^B_{i+1}$ needs at least two time steps to enter column $C_i$, and at least $b_i+1$ time steps to arrive at column $i+1$.}
\end{figure}
\begin{lemma}\label{lemma:lower.bound}
Let $t_{final}$ be the makespan of \textit{ALG} on the given initial configuration. Then for any $1 \leq i \leq n_2$,
\begin{align*}
t_{final} \geq \left|front\left(\mathcal{P}^B_i\left(0\right)\right)\right| + \left|back\left(\mathcal{P}^B_i\left(0\right)\right)\right| + \mathcal{V}_i^{\textit{B}}
\end{align*}
\end{lemma}
\begin{proof}
At time $t_{final}$, $P^B_i$ must be located at or to the left of column $i$ (see \cref{sub:Lemma.5.12.Initial}). Let $t_i$ denote the first time $P^B_i$ reaches column $i$, and let $$b_i = |front(\mathcal{P}^B_i)| + \mathcal{V}_i^{\textit{B}}.$$ In \cref{lemma:t_icrudelowerboundlemma} we show that $t_i \geq b_i$. Note that $t_i = b_i$ if and only if $\mathcal{P}^B_i$ moves left at all times $t < t_i$ except at most $\mathcal{V}_i^{\textit{B}}$ time steps.
We will prove the lemma by tracking $\mathcal{P}^R_{n_1}$. At time $t=0$, since the initial configuration is normal, $\mathcal{P}^R_{n_1}$ is the leftmost agent. Hence, between time $0$ and $t_{final}$ it ought meet $\mathcal{P}^B_i$. Let $M_i$ be the time of their first meeting, and $C_i$ be the $x$-coordinate of $\mathcal{P}^B_i$ at time $M_i$.
We separate the proof into cases.
\textit{Case 1.} Assume $C_i \leq i$. If $C_i \leq i$ then $M_i \geq t_i \geq b_i$. Since in a sorted configuration $\mathcal{P}^R_{n_1}$ must be located to the right of all blue-occupied and empty columns, $\mathcal{P}^R_{n_1}$ must meet the blue labels $\mathcal{P}^B_{i+1}, \mathcal{P}^B_{i+2}, \ldots, \mathcal{P}^B_{n_2}$ before time $t_{final}$. By \cref{observation:labelsmeetoneatatime}, this must take
$n_2 - i = |back(\mathcal{P}^B_i)|$ time ticks. Hence in total \textit{ALG} requires at least $|front(\mathcal{P}^B_i)| + |back(\mathcal{P}^B_i)|+ \mathcal{V}_i^{\textit{B}} $ time steps to complete, as desired. (See \cref{sub:Lemma.5.12.Case.1})
\textit{Case 2.} Assume $C_i > i$. In this case, we claim that $\mathcal{P}^B_{i+1}$ cannot reach column $C_i$ before time $M_i + 2$. If column $C_i$ is a mixed column at time $M_i$ then no agent can move into $C_i$ at time $M_i+1$ (see \cref{sub:Lemma.5.12.Case.2}). Otherwise, $C_i$ is occupied by $\mathcal{P}^B_i$, while $\mathcal{P}^R_{n_1}$ occupies $C_i+1$ in a different row than $\mathcal{P}^B_i$. Since both rows are obstructed by an agent, there is no empty location that will enable $\mathcal{P}^B_{i+1}$ to move into column $C_i$ before time $M_i+2$.
The label $\mathcal{P}^B_{i+1}$ must reach column $i+1$ before time $t_{final}$. By the above, $\mathcal{P}^B_{i+1}$ can reach column $C_i$ for the first time only $2$ time steps after $\mathcal{P}^B_{i}$. The fastest $\mathcal{P}^B_{i}$ can reach column $i+1$ is within $b_i - 1$ time steps, and this can only occur when it is able to move left at every time step after it reaches column $C_i$. Hence the fastest time $\mathcal{P}^B_{i+1}$ can reach column $i+1$ is $b_i - 1 + 2$.
We can now repeat the argument above for $\mathcal{P}^B_{i+1}$. Note that by \cref{observation:labelsmeetoneatatime}, $\mathcal{P}^B_{i+1}$ cannot meet $\mathcal{P}^R_{n_1}$ before time $M_i+1$. We split the argument into the same two cases:
In Case 1, we assume $\mathcal{P}^B_{i+1}$ reaches column $i+1$ before it meets $\mathcal{P}^R_{n_1}$ or at the same time step. Moreover, $\mathcal{P}^R_{n_1}$ must spend $n_2 - i - 1 = |back(\mathcal{P}^B_i)| - 1$ time steps meeting the blue labels $\mathcal{P}^B_{i+2}, \mathcal{P}^B_{i+3}, \ldots, \mathcal{P}^B_{n_2}$. Since $\mathcal{P}^B_{i+1}$ needs at least $b_i - 1 + 2$ time steps to reach column $i+1$, we have that $t_{final} \geq b_i + |back(\mathcal{P}^B_i)|$ just like before, and we are done.
In Case 2, $\mathcal{P}^B_{i+1}$ meets $\mathcal{P}^R_{n_1}$ before reaching column $i+1$. Denote the $x$-coordinate of $\mathcal{P}^B_{i+1}$ at the time of the meeting by $C_{i+1}$. By the same argument as before, $\mathcal{P}^B_{i+2}$ can arrive at column $C_{i+1}$ only $2$ time steps after $\mathcal{P}^B_{i+1}$. $\mathcal{P}^B_{i+1}$ can arrive at column $i+2$ within $b_i+1-1$ time steps only if it moves left at every time step after meeting $\mathcal{P}^R_{n_1}$. Hence like in the previous argument, the fastest time $\mathcal{P}^B_{i+2}$ can reach column $i+2$ is $b_i + 2$.
We can extend this exact argument by induction to all the blue agents $\mathcal{P}^B_{i+3}$, $\ldots$, $\mathcal{P}^B_{n_2}$. At every stage of the induction we separate into two cases. In the first case, $\mathcal{P}^B_{j}$ arrives at column $j$ before it meets $\mathcal{P}^R_{n_1}$ or at the same time step, in which case we can show that $t_{final} \geq b_i + |back(\mathcal{P}^B_i)|$. In the second case, $\mathcal{P}^B_{j}$ meets $\mathcal{P}^R_{n_1}$ before reaching column $j$, which implies it needs at least $b_i + (j - i)$ time steps to arrive at column $j$. By continuing the induction up to $\mathcal{P}^B_{n_2}$ we deduce that the makespan must be at least $$b_i + (n_2 - i) = b_i + |back(\mathcal{P}^B_i)| = |front(\mathcal{P}^B_i)| + |back(\mathcal{P}^B_i)| + \mathcal{V}_i^{B} $$ in all cases, so we are done.
\end{proof}
\cref{lemma:lower.bound} provides a lower bound on the makespan of any sorting algorithm $ALG$ with respect to a given initial configuration. The lower bound is based on the $front$ and $back$s of the blue agents and on the $\mathcal{V}_i^{B}$ values induced by $ALG$. By symmetry, a similar argument holds for the red agents, and consequently we may establish:
\begin{corollary}\label{lemma:optimal-lower-bound}
Let $t_{final}$ be the makespan of \textit{ALG} on the given initial configuration. Then
\begin{equation}\label{equation_optimal_lower_bound}
t_{final} \geq \max\bigg(\max\limits_{1 \leq i \leq n_1} \big(f(P_i^R(0)) + \mathcal{V}_i^{R}\big),\max\limits_{1 \leq i \leq n_2} \big(f(P_i^B(0)) + \mathcal{V}_i^{B}\big)\bigg)
\end{equation}
\end{corollary}
The somewhat unwieldy expression in \cref{lemma:optimal-lower-bound} just says that $t_{final}$ is bounded below by $f_{max}$ plus at most $1$, depending on the value of the $\mathcal{V}_i$s. Recalling \cref{Vmaxdef}, we would like to show that the right-hand side of \eqref{equation_optimal_lower_bound} is at least as large as $f_{max} + \mathcal{V}$. To show this will require a bit more work.
\begin{lemma}\label{lemma:critical-agents-face-each-other}
If a normal initial configuration of agents has both red and blue critical agents, then there must be some blue critical agent to the right of some red critical agent.
\end{lemma}
\begin{figure}[!ht]
\centering
\resizebox{\textwidth}{!}{\import{./images}{Configuration.General.Critical.Agents}}
\caption{A fictitious configuration with critical agents $A_2$ (blue) and $A_3$ (red) not facing each other, per the construction given in \cref{lemma:critical-agents-face-each-other}. Column ranges between agents are marked ($I-V$).}
\label{fig:general.config.faceless.critical.agents}
\end{figure}
\begin{proof}
To prove the Lemma we will assume that the claim is wrong and deduce a contradiction.
Suppose there exists an initial configuration with red and blue critical agents, where all the red critical agents are to the right of all the blue critical agents. Let $A_2$ be some blue critical agent and $A_3$ be some red critical agent. Let $A_1$ be the right-most red agent to the left of $A_2$, and $A_4$ the left-most blue agent to the right of $A_3$. We know these agents exist since the initial configuration is normal. Denote the areas between the agents in the following manner: $I$ denotes the set of columns to the left of $A_1$, $II$ denotes the columns between $A_1$ and $A_2$, and so forth until $V$, which denotes the columns to the right of $A_4$. This construction is illustrated in \cref{fig:general.config.faceless.critical.agents}.
Denote by $\mathbf{E}_x, \mathbf{R}_x$ and $\mathbf{B}_x$, respectively, the number of empty, red-occupied, and blue-occupied columns in the set $x \in \{I,II,\ldots, V\}$. We will write down the value of $f(\cdot)$ for all four agents $A_i$. By definition,
\begin{equation*}
f(A_2) = \mathbf{R}_{I} + \mathbf{E}_{I} + 1 + \mathbf{E}_{II} + \mathbf{B}_{III} + 1 + \mathbf{B}_{V},
\end{equation*}
where the first $+1$ counts $A_1$, and the second $+1$ counts $A_4$.
Similarly,
\begin{equation*}
f(A_3) = \mathbf{R}_{I} + 1 + \mathbf{R}_{III} + \mathbf{E}_{IV} + 1 + \mathbf{B}_{V}+ \mathbf{E}_{V}
\end{equation*}
\begin{equation*}
f(A_1) = \mathbf{R}_{I} + \mathbf{B}_{II} + \mathbf{E}_{II} + 1 + \mathbf{B}_{III} + \mathbf{E}_{III} + \mathbf{E}_{IV} + 1 + \mathbf{B}_{V} + \mathbf{E}_{V}
\end{equation*}
\begin{equation*}
f(A_4) = \mathbf{R}_{I} + \mathbf{E}_{I} + 1 + \mathbf{E}_{II} + \mathbf{R}_{III} + \mathbf{E}_{III} + 1 + \mathbf{R}_{IV} + \mathbf{E}_{IV} + \mathbf{B}_{V}
\end{equation*}
By algebra, we have that if $\mathbf{B}_{III} \geq \mathbf{R}_{III}$ then $f(A_1) \geq f(A_3)$, and otherwise $f(A_4) > f(A_2)$. In the case $f(A_1) \geq f(A_3)$, we have by definition that $A_1$ is a red critical agent to the left of the blue critical agent $A_2$, in contradiction to our initial assumption. In the case $f(A_4) > f(A_2)$, we arrive at a contradiction, since $f(A_4)$ cannot exceed $f_{max} = f(A_2)$. This completes the proof.
\end{proof}
Let us now show that \cref{lemma:optimal-lower-bound} bounds from above our desired lower bound of $f_{max} + \mathcal{V}$, by proving the following:
\begin{lemma}
\label{lemma:inequality_v_final}
\begin{equation}
\label{equation:inequality_v_final}
\max\bigg(\max\limits_{1 \leq i \leq n_1} \big(f(P_i^R(0)) + \mathcal{V}_i^{R}\big),\max\limits_{1 \leq i \leq n_2} \big(f(P_i^B(0)) + \mathcal{V}_i^{B}\big)\bigg) \geq f_{max} + \mathcal{V}
\end{equation}
\end{lemma}
\begin{proof}
The case where $\mathcal{V} = 0$ is trivial, since the maximum of $f$ taken over all agents is by definition $f_{max}$ and our $\mathcal{V}_i^R, \mathcal{V}_i^B$s are non-negative. So let us assume $\mathcal{V} = 1$. By definition, this means either (a) there is a critical agent $A$ that sees another agent immediately in front of it in the initial configuration, or (b) there are two critical agents of different colors in the initial configuration.
In case (a), let us suppose without loss of generality that $A$ is red, and let $ord(A) = i$. Then necessarily $f(P_i^R(0)) = f_{max}$ and, since $A$ cannot move horizontally nor change labels in the first time step of $ALG$, $\mathcal{V}_i^R = 1$. Hence $f(P_i^R(0)) + \mathcal{V}_i^R = f_{max} + \mathcal{V}$ which proves Inequality \eqref{equation:inequality_v_final}.
In case (b), \cref{lemma:critical-agents-face-each-other} tells us that there exists a pair of \textit{red and blue critical agents} $A^R$ and $A^B$ facing each other. Let $ord(A^R) = i$ and $ord(A^B) = j$. If $A^R$ and $A^B$ do not change labels before meeting (i.e., for all times $t$ prior to the meeting, $ord(A^R,t) = i$ and $ord(A^B,t) = j$), then necessarily one of the labels $\mathcal{P}_i^R$ or $\mathcal{P}_j^B$ must move to the upper row so that they can bypass each other. Hence $\mathcal{V}_i^R = 1$ or $\mathcal{V}_j^B = 1$, and similar to case (a), this establishes Inequality \eqref{equation:inequality_v_final}.
Otherwise, either $A^R$ or $A^B$ changes their label before meeting. Let us suppose wlog that $A^R$ changes its label at time $t_0$, i.e. $ord(A^R, t_0-1) = i \land ord(A^R, t_0) \neq i$. At $t_0$, $A^R$'s label can be either $i+1$ or $i-1$. It is straightforward to verify by hand that in either case, $\mathcal{V}^R_i = 1$, since in order for two adjacent agents to swap labels, one of them must wait or move toward the other, and this causes $\mathcal{V}^R$ to become $1$ for both agents that swapped labels. Hence, similar to case (a) and (b), this establishes Inequality \eqref{equation:inequality_v_final}.
\end{proof}
\cref{lemma:optimal-lower-bound} and \cref{lemma:inequality_v_final} together establish the lower bound $f_{max}+\mathcal{V}$ of \cref{theorem:makespanresult}. It remains to show an algorithm that achieves this bound exactly.
\section{Optimal algorithm for normal configurations}
\label{sec:TheAlgorithm}
In this section we present a simple agent sorting algorithm. We show that despite its simplicity, the algorithm is optimal for normal initial configurations, in the sense that its makespan \textbf{always} meets the lower bound established in \cref{theorem:makespanresult}, thus completing the proof of \cref{theorem:makespanresult}.
\begin{algorithm
\SetAlgoLined
\KwResult{The configuration is sorted. }
\kwInit{$t \gets 0$}
\ForEach{agent $A \in \mathcal{A}$}
{
calculate $f(A)$ to determine the critical agents\;
}
\uIf{there are critical agents of both colors}{
$c \gets \text{red}$\;
}
\Else{
$c \gets$ the color of critical agents\;
}
\ForEach{agent $A \in \mathcal{A}$}{
\uIf{$A$'s color is $c$}{
move one step in the desired direction if the adjacent location in that direction is not occupied at the beginning of time $t=0$
}
\Else{
move to the second row\;
}
}
\While{configuration is not sorted} {
$t \gets t + 1$\;
\ForEach{red agent $\in \mathcal{A}$}{
\If{right-adjacent location is empty at the beginning of $t$} {
move one step right\;
}
}
\ForEach{blue agent $\in \mathcal{A}$}{
\If{left-adjacent location is empty at the beginning of $t$} {
move one step left\;
}
}
}
\caption{Optimal sorting algorithm for unsorted normal initial configurations.}
\label{algo:trivial.centralized}
\end{algorithm}
\cref{algo:trivial.centralized} is simple to describe: \textit{aim to move in your desired direction},
left for blue agents and right for red agents. The only twist is the first tick, where either all red or all blue agents move to the second row. Which color should we move vertically? The answer comes from \cref{theorem:makespanresult}; we want to avoid paying more than $f_{max} + \mathcal{V}$ time steps by moving vertically only the color without critical agents. This enables critical agents to move horizontally without being delayed by a vertical movement. If there are red critical agents - we move blue agents to the second row, otherwise red agents are moved.
Denote the above algorithm ${ALG}_1$. To establish its makespan, let us first study the simple scenario where all blue agents are in the one row and all red agents are in the other, and at every time step they simply move horizontally in the desired direction:
\begin{lemma}\label{theorem:strange-model-algo-runtime}
Consider a normal configuration where all red agents are at the top row and all blue agents are at the bottom row, and every column contains at most one agent.
Let $ALG^{\bot}_1$ denote the algorithm that says: at every time step, red agents move right unless the adjacent location to their right is occupied, and blue agents move left unless the adjacent location to their left is occupied. Then the makespan of $ALG^{\bot}_1$ on this configuration is at most $f_{max}$.
\end{lemma}
\begin{proof}
For any red agent $R$, we define $front^{\bot}(R, t)$ as the set of all empty locations on $R$'s row which are to the right of $R$ at time $t$. Furthermore, define $back^{\bot}(R,t)$ to be the set of all red agents to the left of $R$.
Similarly, for any blue agent $B$, we define $front^{\bot}(B, t)$ as the set of all empty locations on $B$'s row which are to the left of $B$ at time $t$, and define $back^{\bot}(B,t)$ to be the set of all blue agents to the right of $B$.
For every agent $A$ define $f^{\bot}(A,t)$ at time $t$ in the following manner:
\begin{itemize}
\item if $front^{\bot}(A,t)$ contains no empty locations, $f^{\bot}(A,t)=0$
\item otherwise, $f^{\bot}(A,t) = |front^{\bot}(A,t)|+|back^{\bot}(A,t)|$
\end{itemize}
In the Lemma's assumed agent configuration, $|front^{\bot}(A,0)| > 0$ for any agent $A$. Hence for all agents $f^{\bot}(A,0) \neq 0$. Recalling \cref{def:f}, this means $f^{\bot}(A,0)=f(A)$ for any agent $A$. Hence, there is a critical agent $A$ for which $f^{\bot}(A,0)=f_{max}$.
Let us define $f_{max}^{\bot}(t) = \max_{A \in \mathcal{A}}f^{\bot}(A,t)$. By the above, $f_{max}^{\bot}(0)=f_{max}$. When $f_{max}^{\bot}(t) = 0$, the agent configuration is sorted, since every agent has moved as far as it could in its desired direction. We will show that $f_{max}^{\bot}(t)$ decreases every time step as long as it is not $0$, which completes the proof.
At time $t$, let $A^*$ be an agent for which $f^{\bot}(A^*,t)=f_{max}^{\bot}(t)$. Let us suppose w.l.o.g. that $A^*$ is a red agent. Note that if a red agent $R$ is an adjacent rightward neighbor of $A^*$ at time $t$ then
\begin{enumerate}[label=(\alph*)]
\item if $f^{\bot}(R, t) = 0$ then $f^{\bot}(A^*, t) = 0$, and
\item \label{enum:option-b} $f^{\bot}(A^*,t) = f^{\bot}(R, t) - 1$ otherwise.
\end{enumerate}
\ref{enum:option-b} is of course impossible, since $A^*$ maximizes $f^{\bot}(\cdot,t)$. Hence, if $A^*$ cannot move right at time $t$, we have that $f^{\bot}(A^*,t) = f_{max}(t) = 0$, so the configuration is sorted. Otherwise, $A^*$ sees an empty location immediately to its right, and it moves to it at time $t$. This causes $f^{\bot}(A^*,t)$ to decrease by $1$.
Since the above argument is true for \textit{any} agent that maximizes $f^{\bot}(\cdot, t)$, we see that as long as $f_{max}^{\bot}(t) \neq 0$, $f_{max}$ decreases at every time step.
\end{proof}
An immediate corollary of \cref{theorem:strange-model-algo-runtime}, which (as mentioned in the introduction) relates to asymmetric simple exclusion processes and may be of independent interest, is:
\begin{corollary}
\label{corollary:singlerowtime}
Assume $j$ agents located on a single row (without access to a spare row) move right at every time step where they are unobstructed by another agent. Then after exactly $f_{max}$ time steps,
the $j$ right-most columns are occupied by an agent.
\end{corollary}
\begin{theorem}
\label{theorem:optimal-sorting-alg1}
For a given initial normal configuration, ${ALG}_1$'s makespan is $f_{max} + \mathcal{V}$.
\end{theorem}
\begin{proof}
By \cref{theorem:makespanresult}, we know the makespan is at least $f_{max} + \mathcal{V}$. Let us show it is at most this.
Let us define $f_{max}'$ to be calculated like $f_{max}$ in \cref{def:f}, but over the agent configuration at time $t=1$. Note that $f_{max} \geq f_{max}'$, since $ALG_1$ never moves agents in a way that can increase this value.
Note that from time $t=1$ onwards, all agents are on two separate rows and execute $ALG_1^{\bot}$ (of \cref{theorem:strange-model-algo-runtime}) in their respective row. Hence, by \cref{theorem:strange-model-algo-runtime}, our algorithm's makespan is at most $f_{max}' + 1 \leq f_{max} + 1$. Hence, if $\mathcal{V} = 1$, we are done.
Let us deal with the case $\mathcal{V} = 0$. We assume, w.l.o.g., that there are red critical agents. Since $\mathcal{V} = 0$, this means there are no blue critical agents, and also that every red critical agent takes a step to the right in the first time step of the algorithm. Since the maximum value of $f$ is obtained over a red agent, we know that
\begin{equation*}
\max\limits_{B \in \mathcal{B}} f(B) \leq f_{max} - 1.
\end{equation*}
Furthermore, since every red critical agent moves right at time $t=0$, when we re-calculate $f$ over the configuration at time $t=1$, $f_{max}$ will have decreased by $1$ (since every critical agent $A$ has shifted an empty location from $front(A)$ to $back(A)$). Hence $f_{max}' = f_{max} - 1$, and so our makespan is at most $f_{max}$, as claimed.
\end{proof}
\cref{theorem:optimal-sorting-alg1} and the lower bound result of the previous section establish \cref{theorem:makespanresult}. We see that ${ALG}_1$ is an optimal algorithm for physically sorting any normal initial configuration. In the next section, we will extend this result to non-normal configurations.
\section{Non-normal initial configurations}
\label{NonNormalConfigurationSection}
In the previous sections we have handled a subset of all the possible initial configurations - the normal configurations (\cref{def:normal-configuration}). In this section, we extend our previously derived lower bound to all possible initial configurations, and extend our optimal algorithm to such configurations.
\begin{definition}
Consider the configuration of agents at time $t$. Let $R$ be the leftmost red agent and let $B$ be the rightmost blue agent. The \textbf{maximal normal sub-configuration at time $t$} is the set of all columns in the interval $[R_x(t), B_x(t)]$. Note that this set is empty when $R_x(t) > B_x(t)$.
Furthermore, let $\mathcal{S}$ be the (possibly empty) maximal normal sub-configuration at time $t=0$. Let $\mathcal{S}^c = \mathcal{A} \setminus \mathcal{S}$ be the set of columns outside $\mathcal{S}$.
\end{definition}
We define a function $f^*$ to equal $f$ for all agents in (the columns of) $\mathcal{S}$, and for any agent $A$ in $\mathcal{S}^c$ we set $f^*(A) = |front(A)|$. We further define $f^*_{max} = \max\limits_{A \in \mathcal{A}} f^*(A)$.
\begin{definition}\label{def:f*-critical agents}
An agent $A \in \mathcal{A}$ for which $f^*(A) = f^*_{max}$ is called an \textbf{$f^*$-critical agent}.
\end{definition}
We will define $\mathcal{V}^*$ to closely resemble $\mathcal{V}$:
\begin{definition}\label{def:V*}
$\mathcal{V}^*$ will equal $1$ if:
\begin{enumerate}[label={\arabic*.}]
\item There is both a red and a blue $f^*$-critical agent \textbf{and} $\mathcal{S} \neq \varnothing$, or
\item There is an $f^*$-critical agent such that, in the initial configuration, another agent is located immediately in front of it \textbf{and} $f^*_{max} > 0$.
\end{enumerate}
Otherwise, $\mathcal{V}^* = 0$.
\end{definition}
\cref{lemma:lower.bound} can now be extended to non-normal initial configurations as follows:
\begin{lemma}
\label{lemma:boundextended}
The makespan of any algorithm that brings the initial configuration to a sorted configuration is at least $f_{max}^* + \mathcal{V}^*$.
\end{lemma}
\begin{proof}
The argument is a rather straightforward generalization of the previous sections.
We assume first that $\mathcal{S} \neq \varnothing$. We note that in this case, agents outside $\mathcal{S}$ are not $f^*$-critical. Indeed, for any same-colored agents $A_1 \in \mathcal{S}$ and $A_2 \notin \mathcal{S}$, $front(A_2) \subsetneq front(A_1)$. Therefore, we trivially have that $f^*(A_2) < f^*(A_1)$. We note that the argument in \cref{lemma:lower.bound} generalizes genuinely to agents inside $\mathcal{S}$. Consequently, \cref{lemma:lower.bound} provides a lower bound on makespan, since $f \equiv f^*$ on $\mathcal{S}$ and when restricted to this set, $\mathcal{V}^*$ is equivalent to the definition of $\mathcal{V}$.
Now let us assume $\mathcal{S} = \varnothing$, i.e. in the initial configuration no pair of red and blue agents face each other. Each label $\mathcal{P}^B_i$ must traverse at least $|front(\mathcal{P}^B_i(0))|$ columns before a sorted configuration can be reached and the same claim is true for $\mathcal{P}^R_i$. In particular, by definition, $f^*$-critical agents will need to traverse $f^*_{max}$ columns to reach their position. If $\mathcal{V}^* = 0$, this establishes our desired lower bound.
If $\mathcal{V}^* = 1$, we have that there is an $f^*$-critical agent that is blocked by an adjacent agent, hence cannot change columns in the first time step. Furthermore, since $f^*_{max} > 0$, the configuration is not sorted, so this agent \textit{must} traverse at least one column. Hence, at least one $f^*$-critical agent must spend $\mathcal{V}^*$ ticks before starting to move horizontally, establishing a lower bound of $f^*_{max} + \mathcal{V}^*$ time steps before a sorted configuration is reached.
\end{proof}
We are now interested in designing an optimal physical sorting algorithm for non-normal configurations - \cref{algo:general.movement}. The algorithm we present will work as follows: the agents that are inside $S$ at time $t=0$ will continue to execute \cref{algo:trivial.centralized} as before, whereas all agents initialized outside $\mathcal{S}$ will execute an ``alternating row split'' strategy. These strategies are executed independently by the two sets of agents: agents initialized in $\mathcal{S}$ will not reach the $\mathcal{S}^c$ columns fast enough to interact with agents initialized outside $\mathcal{S}^c$.
\begin{definition}
Let $r_0$ be the label of the leftmost red agent in $\mathcal{S}^c$ and $b_0$ be the label of the rightmost blue agent in $\mathcal{S}^c$ at time $t=0$. $\mathcal{P}^R_{r_0}(0)$ and $\mathcal{P}^B_{b_0}(0)$ will be called red and blue \textbf{anchor} agents respectively.
\label{definition:anchoragents}
\end{definition}
\begin{algorithm}
\footnotesize
\SetAlgoLined
\KwResult{The configuration is sorted.}
\kwInit{$t \gets 0$, $\mathcal{Q} \gets$ all agents initialized in the columns of $\mathcal{S}$, $\mathcal{P} \gets$ all agents initialized in the columns of $\mathcal{S}^c$}
\ForEach{agent $A \in \mathcal{A}$}
{
calculate $f^*(A)$ to determine the critical agents\;
}
\uIf{there are $f^*$-critical agents of both colors}{
$c \gets \text{red}$\;
}
\Else{
$c \gets$ the color of the $f^*$-critical agents\;
}
move all agents in $\mathcal{Q}$ whose color is not $c$ to the second row\;
\ForEach(\tcp*[h]{agents initialized in $S^c$ split themselves between the upper and lower rows}){agent $A \in \mathcal{P}$}{
\uIf{$A$ is red \And ($r_0 - ord(A, 0)$) is odd}{
move $A$ to the second row\;
}
\ElseIf{$A$ is blue \And ($b_0 - ord(A, 0)$) is odd}{
move $A$ to the second row\;
}
}
\ForEach{agent $A \in \mathcal{A}$}{
\If{$A$ has not moved due to a previous step of the algorithm}{
move one step in the desired direction (left for blue agents, right for red agents) if the adjacent location in that direction is not occupied at the beginning of time $t=0$\;
}
}
\While(\tcp*[h]{at times $t\geq1$ all agents move horizontally toward their destination}){the configuration is not sorted} {
$t \gets t + 1$\;
\ForEach{red agent $A$}{
\If{there is a mixed, empty, or blue-occupied column to the right of $A$ at the beginning of time $t$} {
move one step right\;
}
}
\ForEach{blue agent $A$}{
\If{there is a mixed, empty, or red-occupied column to the left of $B$ at the beginning of time $t$} {
move one step left\;
}
}
}
\caption{Optimal sorting algorithm for any unsorted initial configuration.}
\label{algo:general.movement}
\end{algorithm}
The ``alternating row split'' strategy is based on the following simple idea: at time $t=0$, split agents outside $\mathcal{S}$ in an alternating fashion between the two rows, so that there is at least one empty space between each pair of same-colored agents in the same row (see \cref{figure:subnormalalgoexample}). Once this is done, each agent in $\mathcal{S}^c$ can move horizontally in its desired direction at every time step without any further delays - see \cref{figure:subnormalalgoexample}.
\begin{figure}[!ht]
\centering
\foreach \tick in {0, ..., 4}
{
\begin{subfigure}{\textwidth}
\resizebox{\textwidth}{!}{\import{./images}{Example.Complex.NonNormal}}
\caption{$t=\tick$}
\label{sub:Example.Complex.\tick}
\end{subfigure}
}
\caption{We illustrate $5$ time steps of a run of \cref{algo:general.movement} on a non-normal initial configuration. The maximal normal sub-configuration $\mathcal{S}$ is highlighted in green \protect\tikz{\protect\draw[dashed, ultra thin, fill = green!20] (0,0) rectangle (0.28, 0.28);}. The left-hand and right-hand sides of $\mathcal{S}^c$ are highlighted in blue \protect\tikz{\protect\draw[dashed, ultra thin, fill = blue!20] (0,0) rectangle (0.28, 0.28);} and red \protect\tikz{\protect\draw[dashed, ultra thin, fill = red!20] (0,0) rectangle (0.28, 0.28);} respectively. The blue $\mathcal{P}^B_{b_0}$ and red $\mathcal{P}^R_{r_0}$ \textbf{anchor agents} are circled in blue and red respectively.}
\label{figure:subnormalalgoexample}
\end{figure}
\FloatBarrier
We will show that \cref{algo:general.movement} is an optimal sorting algorithm.
\begin{lemma}\label{lemma:Alternating-Shift-Algo-runtime}
If $\mathcal{S} = \varnothing$, the makespan of \cref{algo:general.movement} is exactly
\begin{equation*}
f_{max}^* + \mathcal{V}^*
\end{equation*}
\end{lemma}
\begin{proof}
\cref{lemma:boundextended} shows $f_{max}^* + \mathcal{V}^*$ is a lower bound on the makespan, so we just need to prove that it is also an upper bound.
If $\mathcal{S} = \varnothing$, then at time $t=0$ all the agents split between the two rows in an alternating fashion (see \cref{figure:subnormalalgoexample}). In every subsequent time step any agent that has not reached its final sorted position moves horizontally along its row. Thus, each agent $A \in \mathcal{S}^c$ takes at most $front(A)+1 \leq f^*(A)+1$ time steps to reach its final position in the sorted configuration.
We split the proof into cases.
Case 1: There are only red $f^*$-critical agents. Recalling \cref{definition:anchoragents}, we infer that the red anchor agent is necessarily $f^*$-critical. Denote this agent $W$.
Suppose $W$ moves horizontally (i.e., rightwards) at time $t=0$. It will continue to do so at every subsequent time-step, thus reaching its desired column in $front(W) = f^*(W)$ time steps. Note that if $W$ moves horizontally at time $t=0$ according to \cref{algo:general.movement}, it is necessarily the \textbf{only} red $f^*$-critical agent, because it has an empty space in front of it that no other agent counts towards its $f^*$ value. Thus in this case, $\mathcal{V}^*=0$.
Suppose on the other hand that $W$ doesn't move horizontally at time $t=0$. Then, since $W$ is a critical agent, $\mathcal{V}^*=1$, and $W$ will reach its desired column in $front(W) + 1 = f^*(W)+\mathcal{V}^*$ time steps.
Every other agent $A$ reaches its desired column in at most $f^*(A) + 1 \leq f^*(W) + \mathcal{V}^*$ time steps. Thus \cref{algo:general.movement}'s makespan is at most $f_{max}^* + \mathcal{V}^*$.
Case 2: There are only blue $f^*$-critical agents. This is the same as case 1.
Case 3: There are both red and blue $f^*$-critical agents. Since when $\mathcal{S} = \varnothing$, red and blue agents both move independently on disjoint sets of columns, thus we may combine the arguments of Case 1 and Case 2 to infer that the makespan is $f_{max}^* + \mathcal{V}^*$.
\end{proof}
\begin{observation}\label{observation:fcriticalalwaysinS}
If $\mathcal{S} \neq \varnothing$ then any $f^*$-critical agent is necessarily in a column of $S$ at time $t=0$.
\end{observation}
\begin{proof}
Suppose there is an $f^*$-critical agent $W$ in a column of $\mathcal{S}^c$. Assume without loss of generality that $W$ is red. Since $\mathcal{S} \neq \varnothing$, there is necessarily a red agent $W'$ in a column of $\mathcal{S}$ with smaller $x$-coordinate than $W$. We have that $|front(W')| > |front(W)|$, since everything that is in front of $W$ is also in front of $W'$, but $front(W')$ contains a blue-occupied column between $W$ and $W'$ which is not in $front(W')$. By definition $f^*(W') \geq |front(W')| > |front(W)| = f^*(W)$, a contradiction to the assumption that $W$ is a critical agent.
\end{proof}
\begin{proposition}
\label{proposition:optimalgeneralmovement}
In every (normal or non-normal) configuration, \cref{algo:general.movement} completes in exactly
\begin{equation*}
f_{max}^* + \mathcal{V}^*
\end{equation*}
time steps.
\end{proposition}
\begin{proof}
If $\mathcal{S} = \varnothing$, the claim follows from \cref{lemma:Alternating-Shift-Algo-runtime}. We assume therefore that $\mathcal{S} \neq \varnothing$.
Note that at every time step starting at $t = 1$ all agents execute the same logic: red agents move right and blue agents move left whenever they are unobstructed.
In \cref{algo:general.movement}, after time $t=0$, any blue or red agent initialized in the columns of $\mathcal{S}^c$ can move in its desired direction until it settles in its final column in the physical sorting. Furthermore, agents initialized in the columns of $\mathcal{S}$ need at least three time ticks to reach any column of $\mathcal{S}^c$ (see \cref{figure:subnormalalgoexample}). Therefore, agents initialized in $\mathcal{S}$'s columns can never catch up with agents initialized in $\mathcal{S}^c$ before the $\mathcal{S}^c$ agents reach their final sorted position. In other words, agents in $\mathcal{S}^c$ never obstruct agents in $\mathcal{S}$. Moreover, the agents initialized in $\mathcal{S}$ arrive at their final sorted position strictly later than agents in $\mathcal{S}^c$.
Let $r$ be the number of red agents in the columns of $\mathcal{S}^c$. Note that no agent initialized in $\mathcal{S}$ will ever enter the rightmost $r$ columns, since these will be occupied by red $\mathcal{S}^c$ agents and \cref{algo:general.movement} does not let two red agents occupy the same column at any time step. Analogously, let $b$ be the number of blue agents in the columns of $\mathcal{S}^c$. Then no agent initialized in $\mathcal{S}$ will ever enter the leftmost $b$ columns.
Let $\mathcal{C}^*$ be our initial agent configuration, and let $\mathcal{C}$ be the initial configuration where we delete all agents in the $\mathcal{S}^c$ columns and delete the rightmost $r$ columns and the leftmost $b$ columns. It is not difficult to show from the previous two paragraphs that the makespan of \cref{algo:general.movement} on $\mathcal{C}^*$ is the same as the makespan of \cref{algo:trivial.centralized} on $\mathcal{C}$.
According to \cref{observation:fcriticalalwaysinS}, all critical agents in the configuration $\mathcal{C}^*$ are initialized in $\mathcal{S}$. Let $A$ be any red agent initialized in $\mathcal{S}$. Every red agent that we remove from $\mathcal{C}^*$ to create $\mathcal{C}$ (i.e., red agents in $\mathcal{C}^* \setminus \mathcal{C}$) increases $|front(A)|$ by 1. Every column we remove from the right-hand side decreases $|front(A)|$ by $1$ and offsets this. Thus the value of $|front(A)|$ does not change between $\mathcal{C}$ and $\mathcal{C}^*$. The size of $front$ similarly remains unchanged for every blue agent initialized in $\mathcal{S}$. $|back(A)|$ remains unchanged for red and blue agents by definition.
Consequently, $f_{max}^* + \mathcal{V}^*$ over $\mathcal{C}^*$ is equal to $f_{max} + \mathcal{V}$ over $\mathcal{C}$. Thus the makespan of \cref{algo:general.movement} over $\mathcal{C}^*$ is $f_{max}^* + \mathcal{V}^*$.
\end{proof}
In conclusion, the makespan of any sorting algorithm over any initial configuration is at least $f_{max}^* + \mathcal{V}^*$. This lower bound is precise; \cref{algo:general.movement} meets it exactly. This generalizes \cref{theorem:makespanresult} to non-normal configurations.
\begin{theorem}
The makespan of any sorting algorithm over a given (normal or non-normal) initial configuration is at least $f_{max}^* + \mathcal{V}^*$. This lower bound is precise; \cref{algo:general.movement} meets it exactly.
\end{theorem}
\section{Discussion}
We studied the problem of sorting vehicles or ``agents'' on a horizontal two-row highway, sending red vehicles to the right and blue vehicles to the left as quickly as possible. We derived an exact lower bound for the amount of time it takes an arbitrary configuration of such vehicles starting at the bottom row to arrive at a sorted configuration, and presented an optimal sorting algorithm that attains this lower bound.
The instance-optimal algorithm we presented for sorting normal configurations (\cref{algo:trivial.centralized}) requires global knowledge of the initial agent configuration to compute $f_{max}$, as the value of $f_{max}$ determines which color of agent it raises to the upper row. Consider the algorithm that, instead of computing $f_{max}$, simply raises all the \textit{blue} agents to the upper row, and otherwise proceeds the same as \cref{algo:trivial.centralized}. This is a straightforward algorithm that requires no global knowledge; in fact, it can be implemented by decentralized agents with local sensing and no communication (see Appendix A, \cref{algo:trivial.distributed}). \cref{theorem:makespanresult} shows that this algorithm is at most $1$ time step slower than \cref{algo:trivial.centralized}. \cref{fig:comparison run} in Appendix A shows an example run of both algorithms. The authors find it significant that the instance-optimal strategy for this problem can be approximated by a very simple, decentralized and local sensing-based algorithm.
It is not clear whether a simple decentralized strategy exists for non-normal configurations. In our general sorting algorithm (\cref{algo:general.movement}), agents must know whether they are in the maximal subnormal configuration or outside of it to determine their movements, and so a local decentralized algorithm with nearly identical performance is more difficult to conceive of.
In future work, it will be interesting to consider the following extensions of our problem:
\begin{enumerate}
\item Settings where there are more than two rows which the agents may use, and where agents are initialized on arbitrary rows and columns.
\item General permutations. Suppose that instead of just $2$ colors, there are $k$ different colors. What is the optimal way to sort the agents, such that all agents of color $i$ are between those of color $i-1$ and $i+1$?
\item Agents moving at different velocities, e.g., a grid square every $2$ or $4$ time steps instead of $1$.
\end{enumerate}
Physical sorting problems such as the one described in this paper force algorithm designers to take into account a number of factors that are not present in more traditional settings, such as physical motion and collision avoidance. When compared to traditional combinatorial algorithms, we believe that there is a lot of room for further theoretical developments in this domain, as even results that at a glance might seem straightforward currently require specialized analysis.
\section*{Acknowledgments}
The research presented in this paper was partially funded by the The Israeli Smart Transportation Research Center (ISTRC).
\clearpage
\begin{appendix}
\section{An almost optimal distributed solution}
The proposed centralized solution can be easily parallelized to produce an almost trivial decentralized algorithm, that, nonetheless, sorts agents in only a single tick slower than the instance-optimal \cref{algo:trivial.centralized}. Every agent $A$ has a memory bit $d$ which is used to track whether the current time step is $0$ or not. At time $0$, all the blue agents move to the upper row. In \cref{algo:trivial.distributed}, we describe what each (decentralized) mobile agent does at every time tick $t$:
\begin{algorithm
\SetAlgoLined
\KwResult{A sorted agent configuration.}
\kwInit{memory bit $d \gets 0$}
\uIf{I am a blue agent}
{
\uIf{$d = 0$}{
move to the second row\;
}
\Else{
\If{adjacent location to the left is empty} {
move one step left\;
}
}
}
\Else
{
\If{adjacent location to the right is empty} {
move one step right\;
}
}
$d \gets 1$\;
\caption{Almost optimal decentralized sorting algorithm for an unsorted initial normal configuration. }
\label{algo:trivial.distributed}
\end{algorithm}
Based on the analysis of \cref{theorem:makespanresult}, this algorithm is at most $1$ time step slower than the optimal time for any given initial configuration. The additional time tick is incurred because in an instance-optimal solution, sometimes we must raise the red agents to the second row rather than the blue agents. In \cref{fig:comparison run}, we provide an example run comparing \cref{algo:trivial.centralized} and \cref{algo:trivial.distributed}.
\FloatBarrier
\begin{figure}
\centering
\resizebox{\textwidth}{!}{\import{./images}{Comparison.Algos}}
\caption{A side-by-side algorithm execution on a given initial grid configuration. (Left) centralized \cref{algo:trivial.centralized}, (right) distributed \cref{algo:trivial.distributed}. The centralized algorithm finishes $1$ time step before the decentralized algorithm ($t=7$ and $t=8$ respectively).}
\label{fig:comparison run}
\end{figure}
\end{appendix}
\bibliographystyle{elsarticle-num}
|
\section{Introduction} \label{sec:intro}
Detecting planets in the early stages of formation is key to reconstructing the history and diversity of fully developed planetary systems, including our own. However, the dense and opaque environment where planets are assembled --protoplanetary discs-- make the direct observation of these bodies a challenging task. To date, PDS\,70 is the only system in which forming planets have been convincingly detected by direct imaging \citep{keppler+2018, haffert+2019}. Nevertheless, our growing understanding of how young planets interact with the disc material has stimulated the development of novel, albeit less direct, detection techniques.
Embedded planets are expected to produce velocity disturbances observable in molecular line emission through the gaseous component of their hosting disc \citep{perez+2015, perez+2018}. In fact, there has been an increasing number of ALMA observations reporting localised \citep{pinte+2018b, casassus+2019, pinte+2019} and extended deviations from Keplerian rotation \citep{teague+2018a, teague+2019nat} attributed to the presence of planets, which have naturally inspired theoretical efforts on the characterisation and detection of planet-driven perturbations \citep{diskdynamics+2020, bollati+2021, rabago+2021, izquierdo+2021}. In particular, the disc around HD\,163296{}, a Herbig Ae star at 101.5 pc from Earth \citep{bailer+2018}, has become one of the most interesting laboratories for the study of planet-disc interactions.
It displays strong indications of embedded planets such as gaps and rings in the dust, and non-Keplerian deviations to the velocity field in the gas.
As reported by \citet{isella+2016, isella+2018} using Band 6 continuum data, the dust gaps of the HD\,163296{} disc are located at a radial distance of 10, 45, 86, and 141 au from the star (referred to as D10, D45, D86, D141), and the dust emission rings at 14, 67, 100, and 159 au (referred to as B14, B67, B100, B159). An additional dust gap appears at 270 au\footnote{Rescaled to the latest distance to the source.} (or D270) according to former observations in optical scattered light by \citet{grady+2000}. One way to explain these gaps is by invoking multiple embedded planets with masses between $\sim\!0.1-4\,\rm M_{Jup}$ depending on the physical properties of the disc \citep{zhang+2018}. However, multiple gaps driven by spiral waves from a single planet \citep{bae+2017}, or induced by non-planetary mechanisms should not be discarded \citep[see e.g.][for a review]{andrews+2020}.
Luckily, further constraints on the presence of planets have been possible thanks to recent kinematical analyses of the molecular gas in the disc. For instance, \citet{pinte+2018b} observed a localised velocity perturbation in $^{12}$CO{} channel maps, also known as a `kink', consistent with the presence of a 2\,$\rm M_{Jup}$ planet at an orbital distance of 260\,au according to hydrodynamic simulations. From here on we refer to this kink as K260.
In the same disc, \citet{teague+2018a} detected azimuthally extended deviations from Keplerian rotation driven by radially localised pressure gradients typical of gas gaps \citep{kanagawa+2015}. By modelling the rotation curve of the system, the authors found a plausible scenario consisting of gas gaps carved by two Jupiter-mass planets orbiting at 83 and 137\,au. Moreover, at the radial location of these gaps, \citet{teague+2019nat} would later report the discovery of meridional circulation of gas flowing from the disc surface towards the midplane, providing further evidence for the presence of strong depletions in the surface density of the disc.
However, these large-scale fluctuations associated with gas gaps should ideally be accompanied by the study of azimuthally localised perturbations in the velocity field in order to reduce the ambiguity that exists between planetary and non-planetary mechanisms to explain the origin of such substructures \citep[see e.g.][]{rabago+2021, izquierdo+2021}.
In this article, we apply the \textsc{discminer}{} channel-map fitting analysis and statistical framework presented in \citet{izquierdo+2021} to search for embedded planets in the disc around HD\,163296{} using DSHARP $^{12}$CO\,$J=2-1${} archival data.
The technique detects two azimuthally localised velocity perturbations possibly driven by two giant planets; one at $R=94$\,au, $\phi=50^\circ$, within the D86 dust gap, and another at $R=261$\,au, $\phi=57^\circ$, potentially linked to the K260 velocity kink. Additionally, we use a best-fit model of the channel maps from the data to study the vertical structure of the disc, as well as the radial gradient of temperatures and line widths observed on the upper and lower emitting surfaces of $^{12}$CO{}.
While this paper was under review, further data on the source was published by the MAPS collaboration \citep{oberg+2021}. We have repeated the analysis of this work on the new data without finding differences affecting our conclusions. We will show a detailed analysis of the new data in a future publication.
\section{Line intensity model of the HD\,163296 disc}
\subsection{Dataset}
In this work we use $^{12}$CO\,$J=2-1${} line observations of the disc around HD\,163296{} obtained by the DSHARP ALMA Large Program \citep{andrews+2018, isella+2016, isella+2018}. The synthesized beam of the data is $0\farcs{104}\times0\farcs{095}$, and the velocity channels are spaced by $0.32$\,km\,s$^{-1}$. The rms noise per channel is 0.84\,mJy\,beam$^{-1}$. The data cube is available at \url{https://almascience.eso.org/almadata/lp/DSHARP}. Details on the calibration of the data and imaging process can be found in \citet{andrews+2018}.
\subsection{Discminer model setup} \label{sec:discminer}
To model the line intensity and kinematics of the disc we use the \textsc{discminer}{} package introduced in \citet[][hereafter Paper I]{izquierdo+2021}. The \textsc{discminer}{} assumes parametric prescriptions for the peak intensity, line width, rotation velocity, and height of the emitting surfaces of the disc to produce intensity channel maps. It then invokes the \textsc{emcee}{}
sampler \citep{foreman+2013} to find the model parameters that best reproduce the intensity of the input data cube. This approach guarantees that the physical and morphological attributes of the disc are modelled simultaneously, providing a comprehensive picture of the gas disc structure and kinematics.
The \textsc{discminer}{} pieces the disc attributes together in a predefined kernel to generate model line profiles and channel maps. As in \citetalias{izquierdo+2021}, we adopt a generalised bell kernel to shape the model intensity, $I_m$, as a function of the disc cylindrical coordinates $(R,z)$ as follows,
\begin{equation} \label{eq:kernel}
I_m(R, z; \upsilon_{\rm ch}) = I_p \left(1+\left|\frac{\upsilon_{\rm ch} - \upsilon_{\rm k^{l.o.s}}}{L_w} \right|^{2L_s} \right)^{-1},
\end{equation}
where $I_p$ is the peak intensity, $L_w$ is half the line width at half power, or just line width from now on, $L_s$ is the line slope. For simplicity, we parameterise these attributes as power laws of the disc cylindrical coordinates ($R$, $z$). On the other hand, $\upsilon_{\rm ch}$ is the channel velocity where the intensity is to be computed, and $\upsilon_{\rm k^{l.o.s}}$ is the Keplerian line-of-sight velocity.
The vertical coordinate $z$ is determined by the height of the upper and lower emission surfaces, which implies that each attribute (but the line slope) has two different representations. The height of each surface is parameterised independently using a combination of two radial power laws. The exact functional form of each attribute and the free parameters of the model are summarised in Table \ref{table:attributes_parameters}.
To merge the contribution of the upper and lower emitting surfaces into a single line profile, on each velocity channel and pixel we select the highest intensity between bell profiles computed for both surfaces independently. In terms of radiative transfer this type of masking is more precise than adding up both intensity profiles directly. The reason is that in a real scenario, the emission from the lower surface can only be distinguished when the upper surface becomes optically thin enough, which for $^{12}$CO{} is mainly limited to the wings of the profile\footnote{Assuming that the sensitivity, and the angular and spectral resolution of the observation are sufficient to resolve both emitting surfaces.}.
It should be noted that the model attributes introduced here are merely descriptive and are not the result of detailed radiative transfer. Also, these are constrained to the upper and lower emitting surfaces of $^{12}$CO\,$J=2-1${}, and therefore any extrapolation to other scale heights in the disc should be done with caution.
\subsection{Parameter search with \textsc{emcee}{}}
\label{subsec:emcee}
\begin{figure*}
\centering
\includegraphics[width=1.0\textwidth]{channel_maps_25to30_hd163296.png}
\caption{ Selected channel maps from DSHARP $^{12}$CO\,$J=2-1${} data of the HD\,163296{} disc (top row), compared to those from the best-fit model found by the \textsc{discminer}{} (middle row). Line-of-sight velocity contours from the model upper and lower surfaces of the disc are overlaid on the model channel maps as solid and dashed white lines, respectively. Also shown are residual brightness temperatures for each velocity channel (bottom row). For reference, the best-fit systemic velocity is $\upsilon_{\rm sys}=5.77$\,km\,s$^{-1}$. The synthesized beam of the observation is shown in the bottom left of the top left panel as a cyan ellipse. Velocity channels are spaced by 0.32\,km\,s$^{-1}$. Residuals with magnitudes lower than the rms noise of the data were masked. Green crosses highlight the position of the localised velocity perturbations, P94 and P261, detected by our clustering algorithm in Sect. \ref{subsec:planets}. Dashed black lines mark the projected location of the D86 and D141 dust gaps \citep{isella+2018}, while the dotted purple line shows the radial location of the K260 kink \citep{pinte+2018b}. From visual inspection, a localised kink-like feature near the P94 perturbation is observed in the $\upsilon_{\rm ch}=5.96,\,6.28$\,km\,s$^{-1}$ channels. The K260 kink near the P261 perturbation is more extended, spanning from $\upsilon_{\rm ch}=5.96$ to at least $\upsilon_{\rm ch}=6.92$\,km\,s$^{-1}$ (see also Fig. \ref{fig:kink_channels}).
}
\label{fig:channels}
\end{figure*}
We initialise the \textsc{emcee}{} sampler with a first guess of parameters according to previous measurements of the inclination, position angle, and stellar mass \citep{isella+2018,teague+2019nat}. The other initial parameters are guessed using the prototyping tool of the \textsc{discminer}{} which allows for a quick comparison of the morphology of model channel maps with respect to the data. The MCMC search is performed with 256 walkers which evolve for 3000 steps for an initial burn-in stage. Next, to sample the posterior distributions and assess convergence of parameters we run the same number of walkers for 10000, 20000, and 50000 steps. We note that the variance and the median of parameter walkers remain almost unchanged after $\sim\!10000$ steps.
The best-fit parameters summarised in Table \ref{table:attributes_parameters} are the median of the posterior distributions in the last 5000 steps of the 20000 step run, thinned by half the auto-correlation times of the parameter chains in order to minimise the impact of non-independent samples on the posterior statistics. In Figure \ref{fig:channels}, we compare selected channel maps from the data and the best-fit model obtained with these parameters.
On a side note, we observe that the best-fit stellar mass retrieved by our model is affected by the choice of the disc outer radius. For a disc radius of $R_d=380$\,au we find a stellar mass of $M_\star=2.02$\,M$_\odot$,
whereas for $R_d=450$\,au the stellar mass decreases to $M_\star=1.97$\,M$_\odot$. This behaviour is expected as the model tries to compensate for the sub-Keplerian rotation supported by steep pressure gradients at large disc radii \citep[][]{dullemond+2020}. However, none of the results of this paper is affected by such a small variation in the stellar mass.
The noise of the data is taken into consideration for the parameter search as follows. At each pixel of the data, we compute the standard deviation of the residual intensities in line-free channels and take it as the weighting factor of the likelihood function to be maximised by the sampler \citepalias[see Eq. 1 of][]{izquierdo+2021}. To ensure that the noise of individual pixels is approximately independent from neighbouring pixels, we down-sample the data and the model grid so that the pixels are separated by $\sim$1.5 beams between each other. Also, this enables us to safely consider the variance of the posterior distribution of model parameters and their (anti--)correlations to quantify uncertainties in observables derived from the model. Details on the analytic propagation of errors taking into account parameter correlations are presented in Appendix \ref{sec:appendix_errors}.
\section{Results} \label{sec:results}
\subsection{Physical attributes of the $^{12}$CO{} disc}
\begin{figure*}
\centering
\includegraphics[width=1.0\textwidth]{emission_surface_attributes.pdf}
\caption{Best-fit attributes derived by the \textsc{discminer}{} for the upper (solid orange) and lower (solid blue) emitting surfaces of $^{12}$CO\,$J=2-1${} in the disc of HD\,163296. The functional forms of the model attributes and best-fit parameters are presented in Table \ref{table:attributes_parameters}. \textit{Left}: Height of the model emission surfaces. The background dots illustrate the height of the upper (orange) and lower (blue) surfaces derived independently with the method of \citet{pinte+2018a}. See Sect. \ref{subsec:height} for details on how to use the method on the lower surface alone. The connected circles mark the mean value of dots within 10\,au ($\sim$beam size) intervals. Also shown is the height reported by \citet{teague+2019nat}. \textit{Top right}: Model peak brightness temperature, computed with the full Planck law. The background dots are the primary and secondary peaks, associated with emission from the upper and lower surface, respectively, as retrieved from double-Bell profiles fitted along the velocity axis of the data cube (see Sec \ref{subsec:height}). \textit{Bottom right}: Model half line widths at half maximum. The dashed lines indicate half the thermal broadening at half maximum for $^{12}$CO{} at the model temperatures of the top right panel. The difference between the thermal and the observed supra-thermal line width is given by the contribution of density and turbulence to the optical depth of the line and by velocity mixing.
}
\label{fig:emission_surfaces}
\end{figure*}
In this Section, we discuss the form of the main physical attributes retrieved by the \textsc{discminer}{} model of the $^{12}$CO\,$J=2-1${} emission from the disc around HD\,163296{}. In Figure \ref{fig:emission_surfaces}, peak brightness temperature, line width and height of the upper and lower emitting surfaces are shown as a function of the radial location in the disc.
\subsubsection{Emission height}
\label{subsec:height}
Our best-fit model indicates that the scale height of the upper surface of the disc is around $z/R\approx0.26$, similar to the findings of \citet{rosenfeld+2013}. This scale height is also in good agreement with the kinematical model of the upper surface reported by \citet{teague+2019nat}, although they diverge substantially on the outskirts of the disc ($R>300$\,au). For further comparison, we also determined the height of the upper surface using the \textsc{disksurf} code \citep{teague+2021disksurf}, which is an open source implementation of the geometrical method introduced by \citet{pinte+2018a} for measuring the altitude of molecular line emission in discs. This independent experiment is better reproduced by the upper surface of the \textsc{discminer}{} up to $R=400$\, au, although beyond that radius the three methods seem to differ. One of the reasons for this discrepancy may be the fact that the boundary to distinguish intensities from the upper and lower surfaces is diffuse towards the edge of the disc, and hence the extraction of the upper surface might be biased by the contribution of the lower counterpart. Nevertheless, our analysis of gas substructures and detection of planets takes place within $R<300$\,au, where the three measurements agree.
Unlike previous methods, the \textsc{discminer}{} allows us to infer the height of the lower emitting surface too. Our model finds that the lower surface stands around $z/R\approx0.2$ scale heights above the midplane of the disc. To validate this part of the modelling, we performed an empirical reconstruction of the lower surface intensity from the data and then estimated the altitude of its emission using once again the geometrical method of \citet{pinte+2018a}. To do this, we first fit double-bell profiles along the velocity axis of the pixels of the data cube in an attempt to separate the upper and lower surface emission. Next, on each pixel we extract the bell profile with the smaller peak intensity of the two, which is normally associated with the lower surface contribution to the line intensity profile \citep[see e.g.][]{dullemond+2020}. Finally, we combine these secondary bell profiles from all pixels to generate channel maps of the lower surface alone. These channel maps can then be an input to the method of \citet{pinte+2018a} to determine the altitude of the lower surface emission, which is illustrated by the blue dots in Fig. \ref{fig:emission_surfaces}, left panel. The \textsc{discminer}{} height of the lower surface is consistent with this independent estimate, at least for $R<300$\,au where our following analyses take place. Centroid velocities of the lower and upper surfaces obtained separately from this empirical reconstruction are presented in Fig. \ref{fig:centroids_low_up}.
\subsubsection{Line width and Brightness temperature}
\label{subsec:linewidth}
Because of the high densities present in protoplanetary discs, the emission from abundant molecules such as $^{12}$CO{} is optically thick almost everywhere. For a related reason, the level populations of these molecules can be safely considered in local thermodynamic equilibrium \citep{weaver+2018}. Both these facts imply that the peak intensity at the central channel of the line emission converges to the kinetic temperature of the gas, and it saturates over neighbouring channels until the optical depth becomes low at the line wings. Thus, the extent of the plateau at the top of the line is highly influenced by the density of the species, and consequently so is the observed line width. Conversely, if the transition was optically thin, the line broadening would be primarily dominated by thermal and turbulent motions \citep{hacar+2016}.
While the thermal broadening at half maximum for $^{12}$CO{} at 50\,K is 0.21\,km\,s$^{-1}$, the best-fit model line width at the same temperature (i.e. at $R=140$\,au, on the upper surface) can be as high as 0.43\,km\,s$^{-1}$, and the contrast becomes even larger on the lower surface of the disc (0.15 vs 0.60\,km\,s$^{-1}$). This noticeable excess in line widths across the entire disc could be explained by strong turbulent motions in the gas.
However, using observations of the HD\,163296{} disc with different molecules, \citet{flaherty+2017} obtained upper limits for turbulent broadening of only $\delta_{\rm turb}<0.06c_s$, where $c_s$ is the sound speed of the medium ($c_s=0.50$\,km\,s$^{-1}$ at 50\,K, assuming a unit mass for the medium of $\mu=2.37$\,u), which means that the supra-thermal line widths should instead be dominated by the density of the species in most of the disc\footnote{At small radii, due to the finite angular and spectral resolutions, velocity mixing becomes the dominant source of line broadening.}. Such an opacity effect could explain why the retrieved line widths on the lower surface are generally higher than those on the upper surface despite the fact that the lower surface is considerably cooler. The width of the $^{12}$CO{} line profile is thus an indirect window to the gas density, which we exploit in Sect. \ref{subsec:gaps} to determine the location of gas gaps and analyse asymmetries in the gas substructure.
The peak brightness temperature of the model upper surface is consistent with the peak intensities measured directly from the data, which span a wide range of temperatures between $\sim20\!-\!70$\,K \citep[see also][]{isella+2018}. The peak brightness temperatures of the model lower surface oscillate between $\sim20\!-\!30$\,K, which is compatible with temperatures where CO molecules are expected to freeze-out onto dust grains \citep[see e.g.,][]{miotello+2014, woitke+2016}. They also agree with previous radiative transfer models \citep{flaherty+2015}, and direct estimates \citep{dullemond+2020}, as well as with the secondary peak intensities retrieved independently by the double-bell fits introduced in Sect. \ref{subsec:height}.
\subsection{Residual maps} \label{subsec:residuals}
\begin{figure*}
\centering
\includegraphics[width=1.0\textwidth]{residuals_2D_upper_upper_lower.pdf}
\caption{Deprojected residuals retrieved by the \textsc{discminer}{} (a, b, c) and \textsc{bettermoments} (d) by comparing $^{12}$CO\,$J=2-1${} observations of the disc around HD\,163296{} and best-fit line profiles found by the \textsc{discminer}{} as described in Sect. \ref{subsec:residuals}. The cartoons next to each panel illustrate how the corresponding residuals are computed. The synthesized beam of the observation is shown in the bottom left of all panels as a cyan ellipse. The beam is displayed as it is projected on the sky plane. For reference, the north axis of the sky plane and the rotation direction of the disc are marked in panel d. \textit{Top row}: (a) Line width and (b) peak intensity residuals. The shadows indicate regions excluded from the analysis due to systematic residuals possibly caused by velocity averaging in channels around the projected minor axis of the disc. \textit{Bottom row}: Velocity residuals obtained by subtracting the absolute value of data and model line centroids, which are (c) computed from full line profiles to account for the contribution of the upper and lower surfaces of the disc, or (d) measured around the peak of line profiles to which the upper surface contributes the most. Blue and red lines overlaid on velocity residuals are the spines of sub-- and super-Keplerian filamentary structures found by \textsc{filfinder}. The dashed lines indicate the location of the D45, D86 and D141 dust continuum gaps registered by \citet{isella+2018}. The outer dotted line marks the radial distance of the K260 kink reported by \citet{pinte+2018b}.
}
\label{fig:residuals_2d}
\end{figure*}
To extract observables and quantify line profile differences between the data cube and the smooth Keplerian model of the disc, we fit a Gaussian profile to each pixel of the data and the model cube\footnote{We fit Gaussian and not double-bell profiles because the data line centroids from the latter have high pixel-to-pixel variations, of the order of the channel width. This stage is not to be confused with the generation of model line profiles for the MCMC minimisation of intensity differences discussed in Sect. \ref{sec:discminer}.}.
The standard deviation, amplitude, and mean value of each Gaussian profile represent, respectively, the line width, peak intensity, and centroid velocity of the corresponding pixel. The Gaussian properties of the model line profiles are then subtracted from those of the data to produce line width, peak intensity and centroid velocity residuals, as illustrated in Figure \ref{fig:residuals_2d}. Analogous to \citetalias{izquierdo+2021}, the velocity residuals reported in this work are defined as the difference between the absolute value of data line centroids and the absolute value of model line centroids. Considering the absolute value before subtraction of velocities can be convenient for visualisation because it makes residuals on the blueshifted side of the disc to switch sign with respect to residuals computed from direct subtraction. This occurs in such a way that sub(super) Keplerian perturbations in the azimuthal component of the velocity, as those expected around gas gaps, appear blue(red) in the residuals map. Differences between velocity residuals computed by direct subtraction and by subtraction of absolute values of line centroids are illustrated in Appendix \ref{sec:supporting_figures}, Fig. \ref{fig:kind_residuals}.
It should be noted that our forward-modelling of channel maps allows us to account for the effect of intensity variations on the retrieval of gas velocities from both model and data. For this reason, fitting Gaussians to full line profiles can be safely done. Note that this approach implies that the retrieved centroid velocities and line widths are the result of the combined contribution of the upper and lower surfaces to the intensity of the disc. For comparison, we explore another possibility using the quadratic fit method supported by the \textsc{bettermoments} package \citep{teague+2018_bettermoments} which operates with intensity channels around line profile peaks to determine line-of-sight velocities, primarily representative of the upper surface of the disc (see Fig. \ref{fig:residuals_2d}d).
Additionally, we use both kinds of velocity residuals, `upper+lower' and `upper--only', to find coherent filamentary structures with the \textsc{filfinder} package \citep{koch+2015}. To start the search, we assume a smoothing size of 10\,au, similar to the extent of the synthesized beam of the data, and a minimum size of 1500 pixels for a filament to be considered in the analysis. The red and blue lines overplotted in the bottom panels of Fig. \ref{fig:residuals_2d} are the medial axes of the filamentary structures found by the algorithm in our velocity residuals. In contrast, intensity and line width residuals do not exhibit elongated signatures as clearly. Nevertheless, in Sect. \ref{subsec:gaps} we show that all three types of residuals provide remarkable clues about the gas substructure in the disc.
On the other hand, velocity residuals can also be used to hunt for candidate embedded planets. As demonstrated in \citetalias{izquierdo+2021}, the presence of a planet is closely related to spatially localised velocity perturbations in the gas disc, whose magnitude and location should be retrievable as long as the resolution and signal-to-noise ratio of the data allow it.
In what remains, we focus our analysis on these three types of residuals to track gas substructure, and to search for localised velocity perturbations possibly associated with the presence of young planets in the disc of HD\,163296{}.
\subsection{Gas gaps} \label{subsec:gaps}
\begin{figure*}
\centering
\includegraphics[width=0.95\textwidth]{averaged_lw_peakint_gasgaps_stderror.pdf}
\caption{Azimuthally averaged line width, peak intensity and centroid velocity residuals extracted from the blue-- and redshifted halves of the disc, computed within $\pm70^\circ$ azimuths to avoid systematic residuals near the disc minor axis (see Fig. \ref{fig:residuals_2d}, top row). The radial location of dust gaps, dust rings, and the K260 kink is marked by the dashed, solid and dotted lines, respectively. In the top panel is also shown the radial location of the two planet detections reported in Sect. \ref{subsec:planets}. The shaded regions represent one standard deviation from the mean value, divided by the square root of the number of independent beams along each projected annulus. Gas gaps on the redshifted side of the disc are indicated by arrows and brackets; arrows mark local line width minima near dust gaps, while brackets enclose positive velocity gradients indicative of localised variations in the gas pressure.
}
\label{fig:azimuthal_average}
\end{figure*}
As discussed in Sect. \ref{subsec:linewidth}, the line broadening of highly optically thick tracers is dominated by the species density, which in other words means that only a fraction of the line width originates from (non--)thermal motions \citep{hacar+2016}. Therefore, $^{12}$CO{}-depleted regions in discs are expected to drive prominent line width minima as illustrated in \citetalias{izquierdo+2021} for a planet-carved gap.
Line centroids are also sensitive to the presence of gaps due to the fact that any gas substructure triggers local pressure forces which induce deviations from Keplerian rotation that follow the geometry of the pressure gradient. In the gas gap scenario, one can expect axisymmetric velocity perturbations with a positive radial gradient enclosed within the edges of the gap \citep[][]{kanagawa+2015, teague+2018a}. To exploit this background knowledge and search for gas gaps in the disc of HD\,163296{}, we compute azimuthally averaged profiles of line width, peak intensity, and velocity residuals as displayed in Figure \ref{fig:azimuthal_average}. This is done separately for the red-- and blueshifted halves of the disc to highlight large-scale azimuthal asymmetries.
\subsubsection{Azimuthally averaged residuals}
From the average line width and velocity residual profiles there are clear indications of gas gaps near the D45, D86 and D141 dust gaps. Line width residuals exhibit local minima at $R=38$\,au, $R=88$\,au and $R=136$\,au, closely coexisting with positive velocity gradients, and in good agreement with the location of gas gaps proposed by the radiative transfer modelling of \citet{isella+2016} and \citet{liu+2018}. Also, the positive velocity gradients observed in our average velocity residuals are consistent with rotation curves of the same disc reported in recent studies \citep{teague+2018a, rosotti+2020}. We note additional evidence of gas substructure near the K260 kink and the D270 dust gap, although this time the line width trough is not as clear as the closest positive velocity gradient, centred around $R=245$\,au on the redshifted half of the disc for both types of centroid residuals.
On the other hand, the average peak brightness temperature displays several local minima that do not overlap with the line width minima, nor with dust gaps or rings. Instead, some local maxima seem to co-locate with line width and dust gaps (D45\footnote{However, the temperature peak at D45 is likely due to continuum subtraction and beam dilution in the inner 30\,au of the disc \citep{teague+2019nat}. This behaviour is captured by the model at the cost of underestimating temperatures around 50\,au (see also Fig. \ref{fig:emission_surfaces}, top right).}, D141), but there are multiple exceptions. Thus, looking at line width fluctuations in optically thick tracers appears to be a more reliable alternative to probing gas substructure. The reason is presumably that the line width is measured over the full line profile, so it can trace the systematic effect that the varying optical depth spawns over different velocity channels, while the peak intensity is measured on a single channel, making it more subject to local thermodynamic fluctuations and noise. Moreover, this means that the way in which gas gap attributes are retrieved could strongly impact the interpretation of hydrodynamic properties of discs and, more specifically, the inference of planet masses from radiative transfer models of gaps.
\subsubsection{Non-axisymmetric gas substructure}
Our separate analysis of both halves of the disc allows us to comment on asymmetries present in the gas distribution and kinematics. Asymmetries in the gas velocities are subtle throughout most of the disc, but near the location of the K260 kink there is a significant radial shift of about $\sim\!50$\,au between the peak of positive velocity gradients on each half of the disc. This is possibly an effect of non-axisymmetric spiral-like perturbations (see Fig. \ref{fig:residuals_2d}, bottom row), potentially linked to the hydrodynamic interaction of the disc and the massive planet detected by \citet{pinte+2018b} at $R=260$\,au, on the redshifted side. Another substantial asymmetry in average gas velocities is observed between 140\,au and 200\,au. However, this feature appears on the `upper+lower' centroid residuals only, suggesting that instead of an actual asymmetry in the velocity field it could be related to contamination of the lower surface emission by dust absorption in the midplane of the disc \citep[see][]{isella+2018}.
Unsurprisingly, there are asymmetries in the azimuthally averaged line width and intensity profiles too. Line widths are systematically lower on the redshifted part of the disc, between about B67 and B159, while peak intensities are slightly higher than those on the blueshifted side, in the same radial section. This finding reveals azimuthal fluctuations in the density and temperature of the disc: the blueshifted side of the disc is denser and cooler than the redshifted half. A plausible origin of these asymmetries may be the presence of massive planets, which are capable of producing vertical and turbulent motions that induce azimuthal gradients of velocity dispersion, density and temperature around their orbit \citep{dong+2019}.
The spatial coincidence of these features with the D86 and D141 gaps is in agreement with such a scenario. In particular, the blueshifted side of the disc has a prominent line width excess of $\sim\!50$\,m\,s$^{-1}$ in the D86 gap, which is similar to the expected excess that a massive planet would trigger in and around its gap \citep[see Fig. 8 of][]{dong+2019}. In Sect. \ref{subsec:planets}, we detect an azimuthally localised velocity perturbation near D86 which strengthens the idea of an embedded planet at this location. Future studies of optically thin isotopologues could offer additional clues about candidate massive planets by characterising line width and temperature asymmetries in the disc.
\subsection{Detection of planets} \label{subsec:planets}
\begin{figure*}
\centering
\includegraphics[width=0.83\textwidth]{clusters_2D_folded_residuals_upperonly.pdf}\\ \vspace{0cm}
\includegraphics[width=0.83\textwidth]{clusters_2D_folded_residuals_upper_lower.pdf}
\caption{Folded velocity residuals (left column) and detected clusters of peak velocities (right) using seven radial and eight azimuthal clusters. The centroid velocities are measured either around the peak of the line profiles, namely on the upper surface of the disc (top row), or using full line profiles, that is, considering the combined contribution of the upper and lower surfaces (bottom row). The green crosses mark the inferred position of the localised velocity perturbations detected by our clustering algorithm, P94 and P261, possibly driven by embedded planets. The grey regions in the left panels correspond to masked portions due to high velocity uncertainties (top; see Appendix \ref{sec:appendix_errors}), or possible contamination of the lower surface by dust absorption in the midplane (bottom).
}
\label{fig:planet_detection}
\end{figure*}
We employ the statistical analysis developed in \citetalias{izquierdo+2021} to search for localised planet-driven perturbations in the gas disc kinematics of HD\,163296{}. More specifically, we use the so-called Variance Peak method of that framework to exploit the fact that peak deviations from Keplerian rotation are expected to be spatially clustered near planets.
The first step of the Variance Peak method consists of folding centroid velocity residuals (Fig. \ref{fig:residuals_2d}, bottom row) along the projected minor axis of the disc to get rid of any axisymmetric feature driven by symmetric substructures such as gas gaps.
In other words, we subtract line centroids on the blueshifted side of the disc from their mirror location on the redshifted half. Next, a radial scan is performed over the map of folded centroids to search for peak velocity residuals, and to record their magnitude, their azimuth, and radial location, as displayed in Figure \ref{fig:folded_azimuthal}. With this information we run a K-means clustering algorithm \citep{macqueen+1967, scikit-learn+2011} along the radial and azimuthal coordinates, independently, to look for localised velocity perturbations.
The K-means algorithm subdivides the input residuals into a predefined number of clusters in such a way that the centre of each cluster is the closest centre to all the residuals in the cluster. Said differently, the input data are iteratively partitioned into Voronoi cells until convergence is reached, which in this case means until the sum of squared distances from the peak residuals to the centre of their clusters is minimised. We refer the reader to \citetalias{izquierdo+2021}, Sect. 4.4, for further details on the K-means algorithm applied to the analysis of velocity residuals.
If a velocity perturbation is strong and coherent, the variance of its corresponding cluster of peak velocity residuals should be high and exceed the variance of the background clusters, as predicted in \citetalias{izquierdo+2021}. If such an excess is higher than three times the dispersion of the background variances (i.e. $>\!3\sigma$) we claim for the presence of a localised perturbation possibly driven by a planet.
This is illustrated in Figure \ref{fig:planet_detection}, where we present folded centroid velocity maps for one half of the deprojected disc, and the localised perturbations detected by the Variance Peak method. The detection algorithm is applied on centroid velocities derived from the upper surface alone (top row), and from the combined contribution of the upper and lower surfaces (bottom row). For the latter case, we masked the inner part of the disc where there is obscuration of the lower surface emission due to dust absorption in the midplane \citep{isella+2018}. Non-uniform absorption features would trigger spurious velocity fluctuations that may bias the detection process.
\subsubsection{P94 perturbation}
\label{sec:p94}
On the upper surface, our method finds a strong localised perturbation with an amplitude of $\delta\upsilon=0.41$\,km\,s$^{-1}$, centred at $R=94\pm6$\,au, $\phi=50\pm3^\circ$ (hereafter P94)\footnote{Because of the folding procedure applied to the velocity residuals, the azimuth of the detected perturbation is degenerate with respect to the other half of the disc. That is, the same perturbation is found, with opposite sign, at $\phi=180^\circ-50^\circ=123^\circ$, near a candidate kink reported by \citet{pinte+2020} in the D86 gap. Nevertheless, since the localised perturbation has a `Doppler flip' morphology, we favour the detection at $\phi=50^\circ$ where the super-Keplerian part of it is exterior to the orbit of the planet, and the sub-Keplerian part is interior. This is expected so long the observed perturbation is not strongly dominated by radial velocity fluctuations, which is true for massive planets at azimuths between $\phi=0^\circ$ and $\phi=70^\circ$, as illustrated in Fig. 5 of \citetalias{izquierdo+2021}.} in the disc frame of reference. This perturbation is in good agreement with the presence of a planet previously hypothesized as the main driver of the D86 dust gap observed in continuum \citep{isella+2018}, and of the corresponding gas gap inferred from pressure bumps in the rotation curve of the disc \citep{teague+2018a}. The prominent line width asymmetries around the D86 gap further support the presence of this planet as discussed in Sect. \ref{subsec:gaps}. Also, P94 features a Doppler flip typical of spiral wakes generated by an embedded planet \citep{perez+2018, casassus+2019}. We provide a rough estimate of the mass of this planet by rescaling the hydrodynamic simulations presented in \citetalias{izquierdo+2021}, for a stellar mass of 2\,M$_{\odot}$ and a planet at $R=94$\,au (see Fig. \ref{fig:simulations}). Omitting radiative transfer effects, our simulations would predict that a 3\,$\rm M_{Jup}$ planet can produce perturbations with an amplitude similar to that of P94.
However, in \citetalias{izquierdo+2021} we demonstrated that peak velocity fluctuations observed in folded velocity residuals can be amplified due to radiative transfer and projection effects. In particular, in Figs. 5 and 10 of that work we showed that at intermediate azimuths, between $\phi\approx30^\circ$ and $\phi\approx60^\circ$, an intrinsic perturbation of $\sim0.25$\,km\,s$^{-1}$ can become as large as $\sim0.4$\,km\,s$^{-1}$ when both effects are considered. Extending that result to this scenario, and by inspection of Fig. \ref{fig:simulations}, left panel, a 1\,$\rm M_{Jup}$ planet would be sufficient to explain the P94 perturbation, which is the same planet mass suggested by \citet{teague+2018a} near the D86 gap.
From \citetalias{izquierdo+2021}, we also note that line width residuals as low as the observed $\Delta L_w\approx-0.1$\,km\,s$^{-1}$ at this radius are compatible with a deep gas gap carved by a $1\,\rm M_{Jup}$ planet too.
Nevertheless, our mass estimate conflicts with the intermediate planet masses ($0.1\!-0.3\,\rm M_{Jup}$) proposed by \citet{zhang+2018} at D86 based on hydrodynamic simulations of the dust gaps in the disc. This tension is expected because the dynamical properties of dust, which dictate how the dust grains interact with the gas in the disc, are still poorly constrained by observations. On the contrary, the use of kinematic measurements of planetary masses can help model the local properties of the dust with greater precision \citep{pinte+2019}.
Nevertheless, we note that the accuracy of planet mass estimates from forward modelling of hydro simulations can still be systematically affected by a simplistic treatment of the gas thermodynamics \citep[see e.g.][]{bae+2021}.
\subsubsection{P261 perturbation}
\label{sec:p261}
From the upper surface velocity residuals we do not find azimuthally localised perturbations linked to the K260 kink registered by \citet{pinte+2018b}. Instead, we report that such a kink is actually driven by a long filamentary structure in the gas kinematics, spanning from $\sim\!90^\circ$ to $20^\circ$ azimuths, and centred around $R\approx270$\,au. This is illustrated in Figure \ref{fig:kink_channels}, where we compare isovelocity contours of the data against those of the \textsc{discminer}{} model, and highlight that the kink is present over the channels that cross the $\delta\upsilon\approx0.15$\,km\,s$^{-1}$ filament on the top right corner of the deprojected disc. Conversely, there are a few other elongated deviations from Keplerian rotation that do not exhibit clear kinks to the eye because they are either weaker or stand at azimuths close to the main axes of the disc.
Nevertheless, when the combined contribution of the upper and lower surfaces is considered, a localised fluctuation of magnitude $\delta\upsilon=0.40$\,km\,s$^{-1}$ appears at $R=261\pm4$\,au, $\phi=57\pm1^\circ$, in the region associated with the K260 kink. In \citetalias{izquierdo+2021}, we have shown that planet-driven perturbations are sometimes best observed on the lower emitting surface due to projection effects, especially for planets at intermediate azimuths in the near side of the disc, as in this case. Furthermore, it is also known that the magnitude of the three-dimensional velocity perturbations around a planet change with scale-height \citep[see e.g.,][]{rabago+2021}, which could explain why the morphology and magnitude of the perturbation vary when the lower surface of the disc is taken into account.
Both effects strengthen the idea that this localised perturbation (hereafter P261) should be caused by an embedded planet, which at the same time is likely to be the main driver of the long kinematic filament associated with the K260 kink.
\subsubsection{Detection significance of the localised perturbations}
The significance of the P94 and P261 perturbations is above an acceptance threshold of $3\sigma$ with respect to the background velocity residuals. We note that both localised signatures are robustly detected regardless of the number of clusters considered for the K-Means algorithm, which we tested using six to ten clusters. The P94 detection yields an average significance of $19.4\sigma$ in radius and $7.5\sigma$ in azimuth, while the P261 detection has an average significance of $5.2\sigma$ in radius and $4.6\sigma$ in azimuth.
Typical mean values of background cluster variances are between (0.3--0.4)$\times 10^{-3}$ and (0.5--0.6)$\times 10^{-3}$\,km$^2$\,s$^{-2}$ for the P94 and P261 analyses, respectively, while 1$\sigma$ values are within (0.1--0.2)$\times 10^{-3}$ and (0.3--0.4)$\times 10^{-3}$\,km$^2$\,s$^{-2}$, where $\sigma$ represents the standard deviation of background cluster variances. The reported orbital radius and azimuth of P94 and P261 is the mean value of the detected location weighted by the statistical significance of the measurement, while the reported uncertainty is the weighted standard deviation of the detected locations in all realisations.
\begin{figure*}
\centering
\includegraphics[width=1.0\textwidth]{summary_2D_upperonly.pdf}
\caption{Summary of the main results of this work. The green crosses mark the inferred location of the P94 and P261 planets, and the blue--red contours, ranging from $\pm$0.15 to $\pm$0.30 in steps of 0.05\,km\,s$^{-1}$, are nearby velocity residuals associated with the detections. The background colours are azimuthally averaged line width residuals to highlight the location of gas gaps. In light red and blue are overlaid the kinematic filamentary structures extracted from centroid velocity residuals on the upper surface of the disc. The quarter circle on the right is a zoom-in around the P94 perturbation, whose nearby velocity contours are this time coloured according to their magnitude. The green ellipse centered on the P94 cross indicates the spatial error of the detection.
}
\label{fig:summary}
\end{figure*}
\subsubsection{Non-detections}
Our algorithm does not detect any localised perturbation around the D141 gap. Assuming that this gap originates from a planet, our rescaled simulations suggest that it should be a low-mass planet of the order of 0.5\,$\rm M_{Jup}$ or less, so that the planet-driven perturbations are blurred with the background $\delta\upsilon\approx0.1$\,km\,s$^{-1}$ velocities (see Fig. \ref{fig:simulations}, right panel). Furthermore, as demonstrated in Fig. 6 of \citetalias{izquierdo+2021}, average line width residuals of only $\Delta L_w\approx-0.05$\,km\,s$^{-1}$ as those around D141 would be compatible with a gas gap opened by a low-mass planet ($<\!1\,\rm M_{Jup}$). On the other hand, as displayed in Fig. \ref{fig:planet_detection}, velocity fluctuations in the D45 gap can be as high as $\delta\upsilon\approx0.45$\,km\,s$^{-1}$, consistent with an embedded giant planet. However, the large azimuthal extent of these perturbations, spanning from $-30^\circ$ to $+60^\circ$ in the disc, prevents the method from detecting any localised signal there, and affects the detection of features beyond. Moreover, D45 is inside the region where the errors in the observed velocities exceed $0.1$\,km\,s$^{-1}$. For these reasons we exclude the D45 gap from the detection analysis.
\section{Conclusions} \label{sec:conclusions}
We employed the \textsc{discminer}{} channel-map modelling framework and statistical analysis introduced in \citet{izquierdo+2021} to search for localised velocity perturbations in the disc of HD\,163296{} using $^{12}$CO\,$J=2-1${} DSHARP data. Our study aims at retrieving not only radial distance but also azimuth of the localised perturbations, which is a natural step forward in the field. We report the robust detection of two coherent, localised fluctuations possibly driven by two giant planets at $R=94\pm6$\,au, $\phi=50\pm3^\circ$, and $R=261\pm4$\,au, $\phi=57\pm1^\circ$, labeled here as P94 and P261, respectively. The P261 perturbation is in the region of a kink-like feature previously observed by \citet{pinte+2018b} in intensity channel maps, and attributed to an unseen 2\,$\rm M_{Jup}$ planet. The P94 perturbation is consistent with the presence of a 1\,$\rm M_{Jup}$ planet near the centre of the D86 dust gap, which is in turn potentially linked to the radially localised pressure bump reported by \citet{teague+2018a} at $R=83$\,au. The presence of this massive planet could also explain the non-axisymmetric line widths retrieved by our analysis around the D86 gap.
Additionally, we use line profile properties to infer the location of gas gaps and non-axisymmetric substructures in the disc. Based on line width residuals, we detect three gaps centred at $R=38$\,au, $R=88$\,au and $R=136$\,au, compatible with prior radiative transfer models and kinematical measurements of radially localised pressure gradients. On the other hand, the height of the upper emitting surface retrieved by our model at $z/R\approx0.26$ is in good agreement with previous estimates from geometrical and kinematical models. Simultaneously, we provide a model for the lower emitting surface of the disc, which stands at an altitude of $z/R\approx0.2$ above the midplane, and displays brightness temperatures near the CO freeze-out temperature. An illustrative diagram summarising the main findings of this article is presented in Figure \ref{fig:summary}.
\acknowledgments
The authors would like to thank the anonymous referee for their constructive remarks and suggestions which allowed to improve the robustness of the results and the overall quality of the manuscript. This work was partly supported by the Italian Ministero dell Istruzione, Universit\`a e Ricerca through the grant Progetti Premiali 2012 – iALMA (CUP C$52$I$13000140001$),
by the Deutsche Forschungs-gemeinschaft (DFG, German Research Foundation) - Ref no. FOR $2634$/$1$ TE $1024$/$1$-$1$,
and by the DFG cluster of excellence Origins (www.origins-cluster.de).
This project has received funding from the European Union's Horizon 2020 research and innovation programme under the Marie Sklodowska-Curie grant agreement No 823823 (DUSTBUSTERS) and from the European Research Council (ERC) via the ERC Synergy Grant {\em ECOGAL} (grant 855130). SF acknowledges an ESO fellowship. GR acknowledges support from the Netherlands Organisation for Scientific Research (NWO, program number 016.Veni.192.233) and from an STFC Ernest Rutherford Fellowship (grant number ST/T003855/1).
\software{
\textsc{bettermoments} \citep{teague+2018_bettermoments},
\textsc{discminer} \citep{izquierdo+2021},
\textsc{disksurf}
\citep{teague+2021disksurf},
\textsc{emcee} \citep{foreman+2013},
\textsc{filfinder} \citep{koch+2015},
\textsc{matplotlib} \citep{hunter+2007},
\textsc{scikit-learn} \citep{scikit-learn+2011}.
}
\setlength{\tabcolsep}{8.0pt}
\begin{table*}
\centering
{\renewcommand{\arraystretch}{1.0
\caption{List of attributes considered for the \textsc{discminer}{} model of $^{12}$CO\,$J=2-1${} intensity channel maps from the disc around HD\,163296{}, and the corresponding best-fit parameters. $D_0=100$\,au is a normalisation constant, $z$ is the height above the disc midplane, $R$ is the cylindrical radius, and $r$ is the spherical radius. The (down-sampled) pixel size of the model is 15.83\,au. PA is the position angle of the semi-major axis of the disc on the redshifted side.}
\label{table:attributes_parameters}
\begin{tabular}{ llllll }
\toprule
\toprule
Attribute & Prescription & \multicolumn{4}{c}{Best-fit parameters for $^{12}$CO\,$J=2-1${}} \\
\midrule
Inclination & $i$ & $i=45.71^\circ$ & -- & -- & -- \vspace{0.15cm} \\
Position angle & PA & ${\rm PA}=312.35^\circ$ & -- & -- & -- \vspace{0.15cm} \\
\midrule
Systemic velocity & $\upsilon_{\rm sys}$ & $\upsilon_{\rm sys}=5.77$\,km\,s$^{-1}$ & -- & -- & -- \vspace{0.15cm} \\
Rotation velocity & $\upsilon_k = \sqrt{\frac{GM_\star}{r^3}}R$ & $M_\star =1.97$\,M$_\odot$ & -- & -- & -- \\ \midrule
Upper surface & $z_U = z_0 (R/D_0)^p - z_1 (R/D_0)^q$ & $z_0 = 29.78$\,au & $p=1.21$ & $z_1 = 4.36$\,au & $q=1.98$ \vspace{0.15cm} \\
Lower surface & $z_L = z_0 (R/D_0)^p - z_1 (R/D_0)^q$ & $z_0 = 19.91$\,au & $p=1.09$ & $z_1 = 0.03$\,au & $q=4.18$ \\
\midrule
Peak intensity & $I_p = I_0 (R/D_0)^p (z/D_0)^q$ & $I_0 = 8.23$\,Jy\,pix$^{-1}$ & $p=-4.16$ & $q=3.68$ & -- \vspace{0.15cm} \\
Line width & $L_w = L_{w0} (R/D_0)^p (z/D_0)^q$ & $L_{w0} = 0.08$\,km\,s$^{-1}$ & $p=0.86$ & $q=-1.38$ & -- \vspace{0.15cm} \\
Line slope & $L_s = L_{s0} (R/D_0)^p$ & $L_{s0} = 1.85$ & $p=0.21$ & -- & -- \\
\bottomrule
\end{tabular}
}
\end{table*}
\begin{figure*}
\centering
\includegraphics[width=0.7\textwidth]{centroids_low_up_hd163296.pdf}
\caption{Empirical reconstruction of centroid velocities from the lower and upper emitting surfaces of the HD\,163296{} disc as observed in $^{12}$CO\,$J=2-1${}, shifted to $\upsilon_{\rm sys}=5.77$\,km\,s$^{-1}$. As explained in Sect. \ref{subsec:height}, this model-independent reconstruction consists of fitting double-Bell profiles to the datacube pixels along the velocity axis. For comparison, black contours are line-of-sight velocities from the \textsc{discminer}{} best-fit model of the lower and upper surfaces of the disc, ranging from $-3.5$ to $3.5$\,km\,s$^{-1}$ in steps of 1.0\,km\,s$^{-1}$.
}
\label{fig:centroids_low_up}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=0.7\textwidth]{skeleton_chans_2D_folded_upperonly.pdf}
\caption{Illustrating isovelocity contours and deviations from Keplerian rotation related to the presence of coherent structures in the kinematics of HD\,163296{} as seen in $^{12}$CO\,$J=2-1${}. The kink-like feature reported by \citet{pinte+2018b} is driven by the long red filament around $R=260$\,au (see also Fig. \ref{fig:residuals_2d}d), spanning from $90^\circ$ to $20^\circ$ azimuths, which corresponds to a range of $\Delta\upsilon\approx\!1.2$\,km\,s$^{-1}$ in velocity channels. Some of these channels are labelled at the top of the plot within the `Kink zone', and displayed in the right column in sky coordinates. This long kinematic substructure could be closely linked to the localised P261 perturbation and the associated planet candidate reported in Sect. \ref{sec:p261}.
}
\label{fig:kink_channels}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=0.47\textwidth]{sim_peak_residuals_3jup_0p065_final_P86_hd163296.png} \includegraphics[width=0.47\textwidth]{sim_peak_residuals_0p3jup_0p065_final_P141_hd163296.png}
\caption{Simulated intrinsic peak velocity fluctuations as a function of planet azimuth, projected along a line-of-sight parallel to the minor axis of the disc, with an inclination of $45^\circ$. No radiative transfer effects are considered. These simulations are rescaled versions of those presented in \citetalias{izquierdo+2021}, assuming a stellar mass of 2\,M$_\odot$, and three different planet masses at a radius of $R=94$\,au (left) and $R=141$\,au (right). The P94 perturbation detected in this work at $R=94$\,au, $\phi=50^\circ$ on the near side of the disc, is consistent with a $\sim\!1\,\rm M_{Jup}$ planet, whose perturbation is expected to be amplified in folded residual maps due to radiative transfer and projection effects as discussed in Sect. \ref{sec:p94}. The non-detection reported around the D141 gap suggests that an embedded planet at such an orbital radius, if any, should be less massive than half a Jupiter so that its kinematical perturbations go unnoticed with respect to the background fluctuations.
}
\label{fig:simulations}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=0.8\textwidth]{spotplanet_peak_residuals_hd163296_independentpixels_noavg_in0au_mc_upperonly.png} \includegraphics[width=0.8\textwidth]{spotplanet_peak_residuals_hd163296_independentpixels_noavg_in0au_mc_upper+lower.png} \\
\caption{ Location of peak velocity residuals in azimuth (left column), and radial distance in the disc (right), obtained from analyses of the upper surface alone (top row) and of the combined contribution of upper and lower surfaces (bottom). Using this kinematical information, the clustering algorithm introduced in Sect. \ref{subsec:planets} detects two localised perturbations, P94 (top) and P261 (bottom), possibly associated with the presence of two giant planets in the disc of HD\,163296{}. Blue circles are residuals extracted from 60\,au wide radial sectors, centred at 86 (top) and 260\,au (bottom). Error bars are computed according to the analytical treatment of uncertainties presented in Appendix \ref{sec:appendix_errors}. The global peaks shown as black lines mark the median location of peak velocity residuals above 3$\sigma$ significance with respect to the background velocities, which nearly follow a normal distribution as displayed in the rightmost subpanels. Note that outliers in the normal distributions are related to the localised perturbations. The mean value of all peak velocity residuals is 0.16 and 0.18\,km\,s$^{-1}$ for the P94 and P261 analyses, respectively, while 1$\sigma$ equals 0.05 and 0.06\,km\,s$^{-1}$, which represents the standard deviation of velocity residuals weighted by their uncertainties.
}
\label{fig:folded_azimuthal}
\end{figure*}
|
\section{Introduction}
With the huge widespread of social media and audio networks, it is mandatory to develop machine learning tools to identify and recognize joined connected speakers in joint conversations. Such tools are able to detect fake audio recordings and recognize falsehood human speeches. Communication systems over various social media need to first authenticate communicating users before actual conversations are being congregated.
The sounds other than the speech of humans (objects) are defined as background noise. Background noise can affect some attributes of speeches: intelligibility, clearness, and quality~\cite{Michelsanti2019}.
The problem of speech identification based on a given set of speeches has been investigated recently by many researchers~\cite{Bai2012},~\cite{Shahin2021a},~ \cite{Shahin2021b},~\cite{Asda2016}. The goal is to produce a speech-to-identify and recognize machine learning system. Speaker identification can be identified based on either text-independent or text-dependant.
Recent work on speech recognition focuses on the way speakers are stressed, emotional, and disguised in their speeches~\cite{aly2021a}. In this work, we aim to develop a deep learning model for voice identification in Arabic speech. We use a Quranic dataset developed by ~\cite{Nahar2019}.
Unlike Google speech recognition API~\cite{GoogleSTT}, Wav2vec2.0, with its remarkable series wav2vec, wav2vecU Unsupervised~\cite{Baevski2021wav2vecU}, is a well-recognized API library recently developed by Facebook folks~\cite{wav2vec2}. Wav2Vec2.0 is self-supervised learning of representations from raw audio data, that assured the framework can enable automatic speech recognition with just only 10 minutes of transcribed audio data. Wav2Vec2.0 model is pre-trained on 16 kHz frequency. The mentioned model wav2vec2.0 enables speech recognition frameworks at a word error rate (WER) of $8.6$ percent on noisy speech and $5.2$ percent on clean speech on the standard LibriSpeech benchmark~\cite{wav2vec2}. A cross-lingual approach model, called XLSR, can learn speech units common to several languages. Wav2vec2.0 was trained by predicting speech units for masked parts of the audio, in the same manner as Bidirectional Encoder Representations from Transformers (BERT).
Recently, wav2vecU Unsupervised model is proposed to train speech recognition models without any labeled data~~\cite{Baevski2021wav2vecU}. Wav2vec-U decreases the phoneme error rate from 26.1 to 11.3 on the TIMIT standard in comprising to unsupervised wav2vec.
The paper structure is described as follows. In Section~\ref{sec:relatedwork} we present related work. In Sections~\ref{sec:dataset} we describe the AR-DAD dataset that is used in our methods, which is developed in Section~\ref{sec:methods}. In Section~\ref{sec:analysis} we show some analysis for the deep learning algorithms, and simulation studies for the proposed algorithms are demonstrated in Section~\ref{sec:simulation}. Finally, the paper is concluded in Section~\ref{sec:conclusion}.
\begin{figure*}[t]
\centering
\includegraphics[width=\textwidth,height=7cm]{fig/model1-4.png}
\caption{Architecture of the system for identifying a set of speeches}\label{fig:model1}
\end{figure*}
\section{Related Work}\label{sec:relatedwork}
There have been several research efforts in speaker identifications. Several search attempts for the Quran reciter’s identification and verification are proposed, using recurrent neural networks and deep learning, see~\cite{Ahmed2017,Sajjad2017,Lataifeha2020,Nahar2019,Elnagar2018,Lataifeha2020}.
Recently, Lataifeha etc.~\cite{Lataifeha2020} introduced a speaker identification model of 30 known reciters of the Holy Quran. In their work, an audio assembled dataset is presented, which consists of 15810 audio clips that are used for training, and another 397 audio clips that are used for top imitators. In addition, for classification, six top-achieving classical and two deep learning classifiers are used.
Noteworthy, Elnagar etc.~\cite{Elnagar2020} described a comparative analysis for a supervised classification system of Quranic audio clips of several reciters and evaluated different classifiers' performance to assess the model. Elnagar etc.~\cite{Elnagar2018} described a supervised classification system of the holy Quran audio clips of various reciters and constructed a representative dataset of audio clips for seven popular reciters, particularly from Saudi Arabia.
Sajjad etc.~\cite{Sajjad2017} introduced a model for Speaker Identification \& Verification Using MFCC \& SVM.
Khelifa etc.~\cite{Khelifa2017StrategiesFI} demonstrated a practical model for an optimal ASR system development for Quranic audio
recognition, in which a statistical approach of Hidden Markov Models (HMMs) and the Cambridge HTK tools are used.
Mohamed etc.~\cite{Mohamed2014VirtualLS} developed a virtual learning recitations system for Sighted and Blind Students and developed an efficient speech recognition engine that is a speaker and accent independent.
The authors in~\cite{Shahin2021a} proposed a classifier that is based on exploiting supervised Convolutional Neural Network (CNN) using three distinct speech databases: local Arabic Emirati-accented, global English SUSAS, and global English RAVDESS. The most recent used classifiers are the Gaussian Mixture Model, Vector Quantization, and Hidden Markov Model. The author in this work developed a new classifier based on CNN model in stressful talking environments using English and Arabic Emirati-accented databases~\cite{Shahin2021a}.
\section{AR-DAD Dataset}\label{sec:dataset}
This work uses Arabic Diversified Audio Dataset~\cite{ARDADdataset}, denoted AR-DAD, that is delivered online by several reciters from several countries. AR-DAD ~\cite{Lataifeh2020ArDADAD,ARDADdataset} includes approximately 15810 Arabic-based audio clips, which are taken from 37 chapters from the Holy Quran for 30 well-known reciters. In addition, there are additional 397 audio clips for 12 competent imitators of those distinguished reciters that have the most downloaded statistics and are authorized by many agencies.
The AR-DAD dataset is prearranged into three main directories: reciters directory, Textual directory, and Imitators’ directory. In this work, we only used 1000 records of the AR-DAD dataset with only used 10 reciters.
\section{Proposed Methods}\label{sec:methods}
In this section, we present the feature extraction and classification. In our model, we acquire raw audio clips from 10 reciters of the Holy Quran.
\subsection{Feature Extraction}
In our work, we use two models for Feature extractions: wav2vec2.0 and HuBERT. The reason we use these models is that they can obtain high-level contextual representation and learn basic units for less labeled data. It can also capture aspects without a clear segmentation for words.
\textbf{Wav2vec2.0~\cite{wav2vec2}:}
wav2vec2.0 process the raw audio data with a multilayer convolutional neural network (CNN) to obtain latent audio representations of 25ms each. The representations are encapsulated into a quantizer and a transformer for feature extraction and selection as shown in Fig.~\ref{fig:model1}.
The quantizer and transformer processes are described in detail in~\cite{wav2vec2}. Quantization is achieved via Gumbel and K-means. Wav2vec2.0 properties include fully convolutional, binary cross-entropy loss, and its representations used to improve ASR tasks.
Codebook diversity penalty to encourage more codes to be used. The strength of wav2vec2.0 comes from the idea that it performs joint learning between quantization of the latent representation. This feature was not existing in previous versions of Wav2vec.
\textbf{HuBERT:}
Wei-Ning Hsu etc. proposed Hidden-Unit BERT (HuBERT)~\cite{hsu2021hubert}; Self-Supervised Speech Representation Learning by Masked Prediction of Hidden Units in June 2021 as a remarkable tool for Self-supervised approaches for speech representation training and learning. As stated in the seminal paper, HuBERT employs the prediction loss above the masked regions only that influences the model to learn a combined acoustic and contextual model over the continuous inputs. It relies mainly on the consistency of the unsupervised clustering step rather than the intrinsic quality of the allocated cluster labels.
In a performance study, the HuBERT either matches or improves over wav2vec 2.0 performance on the Librispeech (960h) and Libri-light (60,000h) benchmarks with 10min, 1h, 10h, 100h, and 960h fine-tuning subsets for a simple k-means teacher of 100 clusters, and using only two iterations of clustering, see~\cite{hsu2021hubert}. To produce a better representation, HuBERT replies on the consistency of unsupervised learning using k-means teacher.
\subsection{Classifiers}
Connectionist Temporal Classification (CTC) is a sequence modeling algorithm that is plugged in wav2vec2.0 overcomes the problem of accurate alignment and variation lengths of inputs and outputs. CTC is suitable for speech and handwriting recognition.
We have 10 classes and our target is to identify the reciters' voices. We perform the classification process by using three different algorithms: MLP, RNN, and CNN. We perform end-to-end training in which we added a classification head in the wave2vec and HuBERT.
\begin{table}
\centering
\caption{Average accuracy of various classifiers: Speaker identification accuracy evaluation using based on CNN, NB, RBF, SVM, KNN, andMLP}\label{classify_table}
\begin{tabular}{|c|c|}
\hline
classifier & Accuracy\\
\hline
MLP & 0000\\
CNN & 0000 \\
Bi-lstm& 0000 \\
\hline
\end{tabular}
\end{table}
As stated in Table~\ref{classify_table}, different classifiers achieve the greatest results under neutral giving conditions only without any stressful conditions. Moreover, the average of each classifier has been calculated and the result shows that CNN is superior to each of MLP, KNN, NB, RBF, and SVM in all talking conditions.
\section{Training Procedure}\label{sec:analysis}
We train the HuBERT and wav2vec2.0 models for 1000 records with 52 min. In detail, HuBERT-base and HuBERT-large took 240 steps and a long time to converge. This is due to the fact that HuBERT is trained in multiple languages, and did not fine-tune previously in Arabic speeches. In contrast, there is a version of wav2vec2.0 large, which is fine-tuned in Arabic speeches, and it makes the convergence fast.
The algorithms steps can be stated as follows.
Sample a mini-batch of feature vector lab pairs from the training set.
Compute $h_o$ as the initialization of the attention-network layer (A.N.L.).
Compute loss between predicted label $y$ and target $Y$
Let $TL$ and $VL$ be the training loss and validation loss, respectively.
\begin{algorithm}
\caption{Training Procedure of AR-Wav2Wav speaker identification Algorithm}\label{alg:cap}
\hspace*{\algorithmicindent} \textbf{Input:} {Raw audio sequence A} \\
\hspace*{\algorithmicindent} \textbf{Output: }{ Acoustic Speaker Identification}
\algrule
\begin{algorithmic}[1]
\Procedure{AR-wave2wav}{$A$}\Comment{Audio Identification}
\State ${A}\gets []$
\State $V \gets []$
\vskip 5mm
\For{i =1 to \text{max-iter}}
\State Sample a mini-batch of pairs from ${A}$
\State \texttt{ $V \gets wav2vec~FeatureExtractor({A_i})$}
\State Compute $h_o$ as the init. of A.N.L.
\State Compute loss for predicted $y$ and target $Y$
\State \textbf{Print } $TL$
\State \textbf{Print } $VL$
\EndFor
\State Print ${a_j}$ speaker identification
\EndProcedure
\end{algorithmic}
\end{algorithm}
\begin{figure*}[t]
\centering
\includegraphics[width=5.9cm,height=6cm]{fig/train_val_loss_hubert_large.png}
\includegraphics[width=6cm,height=6cm]{fig/train_val_loss_hubert_base.png}
\includegraphics[width=5.9cm,height=6cm]{fig/train_val_loss_hubert_large.png}
\caption{Results and accuracy of the proposed models}\label{fig:f1score}
\end{figure*}
We conducted weighted cross-entropy to overcome the problem of the unbalanced dataset.
\section{Results and Performance Evaluations}\label{sec:simulation}
We performed several experimental tests to measure the success of our proposed model. The measurement accuracy is described in Table~\ref{table:result1}.
\begin{table}[h]
\centering
\caption{Measured accuracy of the three models}\label{table:result1}
\begin{tabular}{|c|c|c|c|c|c|}
\hline
Data size & Length & Model & Clip &reciters & F1-\\
Clips &&&Len.&&Score\\
\hline
1000 & 53-Min & Wav2vec.2.0 Large & $<=$ 4 & 10& 97\% \\
1000 & 53-Min & HuBERT Base & $<= 4$ & 10& 98\% \\
1000 & 53-Min & HuBERT Large & $<= 4$ & 10& 99\% \\
\hline
\end{tabular}
\end{table}
\subsection{F-1 Score}
To measure the precision and recall, we used F1-score as described in~\ref{fig:f1score}.
\begin{equation}\label{}
F_1= 2 \frac{precison. recall}{precision + recall}
\end{equation}
The reason we use the F-1 score is that it gives better measurement for unbalanced data.
Figure~\ref{fig:f1score} shows the validation accuracy across the three models: wav2vec2.0, HuBERT large, and HuBERT base. At the beginning of the training, Wav2vec2.0 converges much faster than other models because it was previously fine-tuned on Arabic corpus. However, after completion of the training process, HuBERT large and HuBERT achieved better validation accuracy.
\subsection{Validation Loss and Training Loss}
Figure~~\ref{fig:accuracy_all_models} presents a comparison between the accuracy of wav2vec2.0, HuBERT large, and HuBERT base models.
Figure~~\ref{fig:validation_loss_all_models} presents the comparison between training and validation loss of wav2vec2.0, HuBERT large, and HuBERT base models. After certain steps, the training loss converges to validation loss.
\begin{figure}[h]
\centering
\includegraphics[width=8.5cm,height=7cm]{fig/accuracy_all_models.png}
\caption{Results and accuracy of the proposed models}\label{fig:accuracy_all_models}
\end{figure}
\begin{figure}[h]
\centering
\includegraphics[width=8.5cm,height=7cm]{fig/validation_loss_all_models.png}
\caption{Results and accuracy of the validation loss all models}\label{fig:validation_loss_all_models}
\end{figure}
\begin{figure*}[t]
\centering
\includegraphics[width=5.8cm,height=6cm]{fig/wav2vec_conf_matrix.png}
\includegraphics[width=6cm,height=6cm]{fig/hubert_base_conf_matrix.png}
\includegraphics[width=5.8cm,height=6cm]{fig/hubert_large_conf_matrix.png}
\caption{(a) Results and accuracy of the wav2vec2.0 conf matrix, (b) Results and accuracy of the HuBERT base conf matrix, (c) Results and accuracy of the HuBERT large conf matrix}\label{fig:all_confusion_matrix}
\end{figure*}
\subsection{Confusion Matrix}
Figure~\ref{fig:all_confusion_matrix} displays that the prediction matrix of the wa2vec confusion matrix model among all ten speakers. The figure shows that speaker number five has the highest prediction value. Figure~\ref{fig:all_confusion_matrix} (a) shows that R01 has an intersection with R03, and the wav2vec2.0 model was not able to differentiate between them. Also, in Figure 6 (b) the HuBERT base model was not able to differentiate between R06 and R10. These wav2vec2, HuBERT base, HuBERT large models miss only $4$, $3$, and $1$ miss-classified samples, respectively. It was noticed that the three models did not miss in the same sample.
\section{Discussions and Future Research Directions}
The titanic success of neural networks in the machine and deep learning are transforming the interaction between humans and objects around in the globe. The famous types of neural networks including CNN, RNN, ANN are at the core of the machine and deep learning revolution with various real applications specifically in speech recognition, image, and video transformations.
In this work, we used only $1000$ records. Previous conducted by~\cite{Lataifeha2020} used $15000$ records and obtained $98\%$.
Speaker speech identification has various applications. Some recent trusted systems depend on voice recognition in order to access the system.
Other speaker identification tools used classical neural networks such as RNN and CNN to be able to recognize the target speaker. New tools for Automated Speech Recognition (ASR) did achieve high accuracy for two reasons: \begin{itemize} \item recently proposed speech recognitions representation tools such as wav2vec2.0 and HuBERT present high accuracy and low validation loss. \item applying innovative approaches that use deep learning concepts. \item several datasets have been available online that made is easy for researchers to download, then train and test their developed models.
\end{itemize}
As a research direction, HuBERT can be fine-turned for a large Arabic corpus. This may improve the performance of the model and obtain a better representation of the Arabic speeches.
A large dataset can also be tested using this model to classify more audio acoustics.
\section{Conclusion}\label{sec:conclusion}
We developed wav2wav2.0 and HuBERT deep learning models for Arabic speakers' identification. We implemented these learning models and demonstrated their results on the Arabic audio BASED dataset. Several experiments are performed using wav2vec2 and HuBERT different validation techniques. The model is successfully performed by using wav2vec2 and yielded an accuracy of $96.23\%$ with data augmentation. In future work, we plan to extend the proposed method to incorporate more feature sets and increase the size of the dataset for words, sentences, paragraph recognition.
The proposed model can be deployed in various practical systems in the cases of biometric gate entrance, bank account checking, critical security access control, etc.
\bigskip
\section*{Acknowledgement}
This research is partially funded by a grant from the academy of scientific research and technology (ASRT), 2020-2021, research grant number 6547.
\bigskip
\bibliographystyle{plain}
\input{audioreferences.bbl}
\end{document}
|
\section{Introduction}
In this paper, which consists of three closely related parts, we investigate the relationship between the filtration by powers of the maximal ideal in the Lubin--Tate ring and chromatic homotopy theory from several perspectives. In the first part, we focus on Morava $K$-theories and their associated Hopf algebroids. In particular, we give an invariant description of their cohomology and prove its finiteness.
We then establish convergence of the $K$-based Adams spectral sequence and relate it to the $K$-local Adams--Novikov and filtration by powers spectral sequences through a Miller square, and describe it completely at large enough primes. In the last part, we construct Hopkins' spectral sequence relating homology of the $K$-local sphere with derived functors of completion, and identify the latter with cohomology of the Morava stabilizer group. As an application, we compute the zeroth limit at all primes and heights.
\subsection*{Main results} A classical approach to the study of an arithmetic problem is to first consider its reduction to residue fields and to then reassemble the local solutions. Over the $p$-local integers, there are just two reside fields, namely the field $\mathbb{Q}$ of rational numbers and the finite field $\mathbb{F}_p$ corresponding to the closed point. Informally, these detect, respectively, the torsion-free and torsion phenomena.
The approach of reducing to residue fields also works very well in stable homotopy theory, where one is interested in classifying stable homotopy classes of maps between finite complexes. Identifying a classical field $k$ with the corresponding Eilenberg--MacLane spectrum, we have a descent spectral sequence of signature
\begin{equation}\label{eq:cass}
E_2^{s,t}\cong \Ext_{\pi_*(k \otimes_{S^0} k)}^{s,t}(k, k) \implies \pi_{t-s}(S^0_{k}),
\end{equation}
where $S^{0}_{k}$ is the appropriate Bousfield localization. When $k = \mathbb{Q}$, this collapses immediately, recovering Serre's calculation that $\pi_{*} S^{0} \otimes _{\mathbb{Z}} \mathbb{Q} \simeq \mathbb{Q}$.
The situation is much more complicated when $k = \mathbb{F}_{p}$, in which case $\cA_* = \pi_*(\mathbb{F}_p\otimes_{S^0}\mathbb{F}_p)$ is the dual Steenrod algebra and the above recovers the classical Adams spectral sequence. In this case, the spectral sequence converges completely, but it does not collapse at any finite page and there is no known algorithmic way of understanding its structure\footnote{A common saying about the classical Adams spectral sequence is ``\emph{Every differential is a theorem.}''}. It is arguably the most important tool for computing the stable homotopy groups of spheres.
A new feature in homotopy theory which is not visible in the algebraic contexts is that, even $p$-locally, $\mathbb{Q}$ and $\mathbb{F}_p$ are not the only residue fields of spectra. Instead, we have an infinite family of prime fields of ``intermediate characteristic'', given by the Morava $K$-theories $K(n)$ for $0<n<\infty$. These are homotopy ring spectra with the properties that
\begin{enumerate}
\item $\pi_{*} K(n) \simeq k [u^{\pm 1}]$, where $k$ is a perfect field of characteristic $p$ and $u$ is an invertible variable of degree $|u| = 2$ and
\item the Quillen formal group $\textbf{G}_{0} := \textnormal{Spf}(K(n)^{0}(\mathbf{CP}^{\infty}))$ over $k$ is of finite height $n$.
\end{enumerate}
One can show that such a homotopy ring spectrum exists for any $\textbf{G}_{0}$ of finite height over a perfect field. By a celebrated result of Devinatz, Hopkins, and Smith, together with $\mathbb{Q}$ and $\mathbb{F}_p$ these are the only residue fields of $p$-local spectra.
Morava $K$-theories are very calculable for at least two distinct reasons:
\begin{enumerate}
\item the coefficients $K(n)_{*}$ form a graded field, so that we have an unrestricted K\"{u}nneth isomorphism $K(n)_{*}(X \otimes Y) \simeq K(n)_{*}(X) \otimes_{K(n)_{*}} K(n)_{*}(Y)$ for any $X, Y$;
\item there's a canonical Poincare duality isomorphism $K(n)_{*}(X) \simeq K(n)^{*}(X)$ whenever $X$ is a $\pi$-finite space, in particular a classifying space of a finite group.
\end{enumerate}
The second property, known as ambidexterity, is one way in which $K(n)$ behave as if they were of characteristic zero, despite $K(n)_{*}$ being of positive characteristic. This surprising feature makes Morava $K$-theories a powerful tool also in more geometric contexts, such as in the breakthrough work of Abouzaid and Blumberg on the Arnold conjecture in symplectic topology \cite{abouzaid2021arnold}.
In the present paper, we set up and analyse the descent spectral sequence associated to the Morava $K$-theories. As a first result, we verify that we indeed do have convergence.
\begin{thma}[\ref{corollary:k_local_adams_conditionally_convergent}, \ref{proposition:complete_convergence_of_k_local_adams_for_dualizable_spectra}]
The $K(n)$-based Adams spectral sequence of signature
\[
\Ext^{s, t}_{K(n)_{*}K(n)}(K(n)_*,K(n)_*(X)) \implies \pi_{t-s}L_{K(n)}X
\]
is conditionally convergent for any spectrum $X$ and converges completely whenever $X$ is $K(n)$-locally dualizable.
\end{thma}
At first glance, this result is quite surprising, as each group on the second page of this spectral sequence for the sphere is torsion, while the abutment $\pi_*L_{K(n)}S^0$ is known to contain torsion-free summands. In particular, this implies that there cannot exist a horizontal vanishing line on any finite page of the spectral sequence. In order to establish conditional convergence, we instead reduce to the case of a finite $n$ complex, where such vanishing lines exist. Complete convergence is the consequence of the degreewise finiteness of the $E_{2}$-term, which we prove for dualizable spectra using the Cartan--Eilenberg spectral sequence
Going back to the case of non-localized spectra, the utility of the classical Adams spectral sequence based on $\mathbb{F}_{p}$ is further amplified when coupled with the Adams--Novikov spectral sequence, which is the descent spectral sequence based on the Brown--Peterson spectrum $BP$. As one spectacular application, Miller constructs a square relating the two spectral sequences and compares the differentials to compute the $v_1$-periodic homotopy of Moore spaces, thereby proving the telescope conjecture at height 1 and for all odd primes \cite{miller_adams}.
The situation is similar in the case of the $K(n)$-based Adams, which has a cousin which might be more familiar to the working chromatic homotopy theorist, namely the $K(n)$-local $E_n$-based Adams spectral sequence. Here, $E_n$ denotes the Lubin--Tate spectrum associated to the formal group $\mathbf{G}_{0}$, parametrizing derived deformations of the latter. We show that the relation between the two spectral sequences is again governed by a square of spectral sequences of the following explicit form.
\begin{thmb}[\ref{corollary:millers_may_sseq_is_filtration_by_powers}, \ref{proposition:mahowald_sseq_is_ce_sseq}]
\label{thmb:introduction}
The $K(n)$-local Miller square associated to $(E,K)=(E_n, K(n))$ is of the form
\[
\begin{tikzcd}
& {\Ext_{E_{*}^{\vee}E}(E_{*}, \bigoplus_k \mfrak^{k}/\mfrak^{k+1} E_{*}^{\vee}X)} \\
{\Ext_{E_{*}^{\vee}E}(E_{*}, E^{\vee}_{*}X)} && {\Ext_{K_{*}K}(K_{*}, K_{*}X)} \\
& {\pi_{*}L_{K}X}
\arrow["{E-\textnormal{Adams}}"', Rightarrow, from=2-1, to=3-2]
\arrow["{K-\textnormal{Adams}}", Rightarrow, from=2-3, to=3-2]
\arrow["{\textnormal{May}}"', Rightarrow, from=1-2, to=2-1]
\arrow["{\textnormal{Mahowald}}", Rightarrow, from=1-2, to=2-3],
\end{tikzcd}
\]
where the May spectral sequence is obtained from the $\mfrak$-adic filtration of $E_{*}^{\vee}X$.
\end{thmb}
We also identify the Mahowald spectral sequence in the square above with a Cartan--Eilenberg spectral sequence of a certain extension of Hopf algebroids canonically attached to each Morava $K$-theory, see \cref{sec:finiteheightmillersquare} for details.
A theorem of Miller relates the $d_{2}$-differentials in the $K$-Adams and May spectral sequences, but at large primes we can say much more. If $2p - 2 > n^{2}+n$, then it is well-known that the $K$-local $E$-based Adams spectral sequence for the sphere collapses at the second page. Using a variation on the classical argument of Milnor, we equip $K_{*}K$ at odd primes with an additional grading which also forces the collapse of the Mahowald spectral sequence.
If follows that at sufficiently large primes, two of the spectral sequences in Miller's square collapse, and it is thus natural to expect that the other two can be identified. This is indeed the case, as we show the following.
\begin{thmc}[\ref{theorem:k_based_adams_at_large_primes_iso_to_filtration_by_powers}]
\label{thmc:introductionthmc}
If $2p-2 > n^{2}+n+1$, then the $K$-based Adams spectral sequence for $S^{0}_{K}$ can be given an additional grading so that it becomes isomorphic to the May spectral sequence
\begin{equation}
\label{equation:may_spectral_seq_in_intro}
\Ext_{E_{*}^{\vee}E}(E_{*}, \bigoplus_k \mfrak^{k}/\mfrak^{k+1} E_{*}) \implies \Ext_{E_{*}^{\vee}E}(E_{*}, E_{*})
\end{equation}
induced by the filtration of $E_{*}$ by the powers of the maximal ideal.
\end{thmc}
Note that one new major phenomenon in intermediate characteristic is that Morava $K$-theories do depend on choices; in particular, on the choice of a formal group. By a result of Lazard, any two such formal groups of the same height are isomorphic over the separable closure; it follows that for our purposes all $K(n)$ of the same height are essentially interchangeable, in particular, they determine the same Adams spectral sequence.
To be more precise, while the Hopf algebroid $K(n)_{*}K(n)$ depends on the choice of a Morava $K$-theory, its category of comodules does not, by a result of Hovey and Strickland \cite{hovey2005comodules}. In the body of the text, we describe a different derivation of this result, identifying the category of $K(n)_{*}K(n)$-comodules with the category $\Mil_{abs}$ of \emph{absolute Milnor modules}, which are certain sheaves on finite spectra, see \cref{sec:milnormodules}. This gives an approach to the Adams spectral sequence based on Morava $K$-theory which only depends on the prime and the height, but no other choices.
On a different note, the completion tower of comodules of the form
\[
\ldots \rightarrow E_{*}/\mfrak^{3} \rightarrow E_{*}/\mfrak^{2} \rightarrow E_{*}/ \mfrak
\]
appearing in \hyperref[thmc:introductionthmc]{Theorem C} also leads to a \emph{another} spectral sequence, first studied by Hopkins and Sadofsky\footnote{The construction of Hopkins and Sadofsky is unpublished.}. The latter spectral sequence approaches the transchromatic information contained in $L_{K(n)}S^0$ through its \emph{uncompleted} $E$-homology $E_*(L_{K(n)}S^0)$, as opposed to the homotopy groups, as we now describe in more detail.
To provide some context, recall that it is known that at large primes there exists an equivalence $h \spectra_{E} \simeq h \dcat^{per}(E_{*}E)$ between the homotopy categories of $E$-local spectra and differential comodules. Under this equivalence, the limit $\varprojlim E_{*} / \mfrak^{k}$ in the derived $\infty$-category corresponds to the $K$-local sphere $S^{0}_{K}$ \cite{pstrkagowski2018chromatic}; thus, the derived completion of $E_{*}$ can be thought of as the algebraic analogue of the $K$-local sphere.
At smaller primes, we do not have algebraic models for the homotopy category, but we instead have a spectral sequence relating the homology groups of the derived completion to $E_{*} (S^{0}_{K})$. This spectral sequence, originally due to Hopkins and Sadofsky, was subsequently worked out by Peterson \cite{peterson2020coalgebraic}, adapting the classical construction of the Adams spectral sequence. We follow a slightly different approach, based on the notion of an adapted homology theory and Devinatz--Hopkins' modified Adams spectral sequence \cite{dev_morava}.
Our construction is closely related to work of Hovey on derived products of comodules \cite{hovey_product}, and in fact works in much greater generality, see \cref{section:homology_of_inverse_limits}. In the case of $E$-homology, we are able to express these derived limits as cohomology of the Morava stabilizer group, leading to the following statement.
\begin{thmd}[\ref{thm:limss}, \ref{prop:limssconvergence}, \ref{prop:inverselimascontcohom}]
\label{thmd:introduction}
Let $E_{*}E := \pi_{*}(E \otimes E)$ be the uncompleted $E$-homology of itself, given the unique topology which restricts to the $\mfrak$-adic topology on each finitely generated $E_{*}$-submodule. Then, there exists a canonical isomorphism
\[
\rmH^{s}_{\cts}(\mathbb{G}_{n}, E_{t}E) \cong (\textstyle\varprojlim_{E_{*}E}^{s} E_{*}/\mfrak^{k})_{t}
\]
between the continuous cohomology of the Morava stabilizer group and the derived functors of the limit in $E_{*}E$-comodules. Moreover, there exists a spectral sequence of $E_*E$-comodules of the form
\begin{equation}
E_2^{s,t} \cong \rmH^{s}_{\cts}(\mathbb{G}_{n}, E_{t}E) \implies E_{t-s}(L_{K}S^0)
\end{equation}
with differentials $d_r^{s,t}\colon E_r^{s,t} \to E_r^{s+r,t+r-1}$. This spectral sequences converges completely and collapses at a finite page with a horizontal vanishing line.
\end{thmd}
The subtle point in this result is the convergence of the spectral sequence, which relies crucially on the finite virtual cohomological dimension of the Morava stabilizer group. Note that here it is important to consider the derived sequential limits in the category of comodules as opposed to ordinary modules over $E_*$; indeed, in positive degrees the latter all vanish in this case and there would be no hope of having a convergent spectral sequence.
The above spectral sequence underlies one approach towards Hopkins' chromatic splitting conjecture \cite{report_on_e_theory_conjectures}, \cite{hovey_csc}. In a nutshell, the latter provides a precise description of the transchromatic behaviour of the $K(n)$-local sphere; that is, its behaviour under applying chromatic localizations $L_h$ for intermediate heights $0 \le h <n$. As such it has been verified for all heights $n \le 2$ and all primes \cite{shimomura_yabe, goerss2012rational, beaudry2017chromatic}, with a minor modification to the original prediction at $n=p=2$ (\cite{bgh_chromaticsplitting}). Beyond height 2, the conjecture remains wide open.
The standard way of approaching the splitting conjecture is through an explicit computation of the homotopy groups of $L_{K(n)}S^0$. However, as mentioned, its transchromatic information is also encoded in the \emph{uncompleted} $E$-homology, and this is what the spectral sequence \eqref{eq:generalss} abuts to.
Using the identification of derived functors of the limit with continuous cohomology of the Morava stabilizer group, we are able to compute the zeroth limit at all primes and heights.
\begin{thme}[{\ref{thm:zeroth_derived_limit_of_completion}}]
There is a canonical isomorphism
\[
\textstyle\varprojlim_{E_{*}E} E_{*} / \mfrak^{k} E_{*} \simeq E_{*} \otimes_{\mathbb{Z}} \mathbb{Z}_{p}
\]
where the limit on the left is taken in the category of $E_{*}E$-comodules.
\end{thme}
Note that $E_{*} \otimes_{\mathbb{Z}} \mathbb{Z}_{p} \simeq E_{*}(S^{0}_{p})$, the homology of the $p$-complete sphere. Thus, the above result can be interpreted as detecting the conjectured copy of $L_{n-1}S^{0}_{p}$ inside $L_{n-1}S^{0}_{K}$ through $E$-homology.
At height one, the Morava stabilizer group is particularly simple, and we are able to compute all of the derived limits.
\begin{thmf}[\ref{theorem:cohomology_of_ee_at_height_one}]
At height $n=1$ and any prime, we have
\[
\rmH_{\cts}^{s}(\mathbb{G}_{1}, E_{*}E) \simeq \begin{cases}
E_{*} \otimes_{\mathbb{Z}} \mathbb{Z}_{p} & \mbox{when } $s=0$,\\
E_{*} \otimes_{\mathbb{Z}} \mathbb{Q}_{p}& \mbox{when } $s=1$,\\
0 & \mbox{otherwise}.
\end{cases}
\]
\end{thmf}
Note that the above in particular completely recovers the height one case of the chromatic splitting conjecture, which states that $L_{0} L_{K(1)} S^{0} \simeq L_{0} S^{0}_{p} \oplus L_{0} S^{-1}_{p}$. The latter can be equivalently obtained by explicitly computing $\pi_{*} L_{K}S^0$, which if we ignore torsion contains exactly two $\mathbb{Z}_{p}$ summands, one for each sphere. However, the computation of homology gives more precise information, as the two copies of $\mathbb{Z}_{p}$ are isomorphic as abelian groups, but the comodules $E_{*} \otimes_{\mathbb{Z}} \mathbb{Z}_{p}$ and $E_{*} \otimes_{\mathbb{Z}} \mathbb{Q}$ are not.
\subsection*{Outline of document}
We hope that the structure of the document becomes apparent from the choice of section titles. In brief, the first part of the paper, comprising \cref{sec:moravaktheories} through \cref{sec:milnorfiltration} deals with Morava $K$-theories, their Hopf algebroids, categories of comodules, and their cohomology, through the perspective of Milnor modules. We pay special attention to the (in)dependence of our constructions on the choice of Morava $K$-theories, and revisit the relation to continuous cohomology.
The $K$-based Adams spectral sequence and the finite height Miller square are the topic of Part 2. We establish convergence properties in \cref{sec:knbasedass}, then construct the finite height Miller square, and use it to relate it to the filtration by powers spectral sequence. In \cref{sec:somecomputations}, the final section of this part, we illustrate our results by going through explicit computations at height 1.
The third part of the paper consists of \cref{section:homology_of_inverse_limits} to \cref{sec:heightone}. Here, the focus lies on the derived category of $E_*E$-comodules and the construction and study of the inverse limit spectral sequence. We then give our application to the algebraic chromatic splitting conjecture based on the connection to the continuous cohomology of the Morava stabilizer group. Each of the two parts ends with some explicit computations at height one, which we hope elucidate our methods and results.
\subsection*{Acknowledgements}
We would like to thank Agn\`es Beaudry, Robert Burklund, Mike Hopkins, Eric Peterson, and Hal Sadofsky for useful conversations related to this work. We acknowledge the hospitality of the Max Planck Institute for Mathematics in Bonn.
\part{Morava $K$-theories and Milnor modules} In this part of the paper, we study Morava $K$-theories, which play the role of residue fields of Lubin--Tate spectra. We describe their Hopf algebroids, categories of comodules, and cohomology.
\section{Morava $K$-theories and their Hopf algebroids}
\label{sec:moravaktheories}
Let $\kappa$ be a perfect field equipped with a choice of a height $n$ formal group $\gzero$. Associated to this data is the Lubin--Tate ring $E_{0}$ classifying deformations \cite{lubin1966formal}. This is a complete local $W(\kappa)$-algebra such that there exists a non-canonical choice of regular generators inducing an isomorphism of rings $E_{0} \simeq W(\kappa)\llbracket u_{1}, \ldots, u_{n-1}\rrbracket$. We write $\mfrak = (p, u_{1} \ldots, u_{n-1})$ for the maximal ideal.
The ring $E_{0}$ equipped with the universal deformation is Landweber exact and so can be lifted to a $2$-periodic ring spectrum $E$ with $E_{*} \simeq E_{0}[u^{\pm 1}]$ for some unit $u \in E_{2}$, and the property that the formal group $\g := \textnormal{Spf}(E^{0}(BS^{1}))$ is the universal deformation of $\gzero$. It is well-known that $E$ admits a unique $\mathbf{E}_{\infty}$-ring structure, functorial in the choice of $\kappa$ and the formal group, see \cite{goerss2005moduli}. More recently, it was proven by Lurie that $E$ arises as a solution to a moduli problem involving formal groups over $\mathbf{E}_{\infty}$-rings, see \cite{lurie2018elliptic}.
The canonical nature of the Lubin--Tate spectrum makes it a good starting point for the study of chromatic phenomena. Since $E_{0}$ is local, one would like to construct a spectrum which plays the role of the residue field of $E$. Following Hopkins and Lurie \cite{hopkins2017brauer}, we make the following definition.
\begin{definition}
\label{definition:morava_k_theory}
An $\mathbf{E}_{1}$-$E$-algebra $K$ is a \emph{Morava $K$-theory} if the unit map $E \rightarrow K$ induces an isomorphism $K_{*} \simeq E_{*} / \mfrak$.
\end{definition}
One can show that an algebra satisfying the above conditions always exists, for any Lubin--Tate spectrum $E$; in fact:
\begin{enumerate}
\item there are uncountably many Morava $K$-theories which are not equivalent as $E$-algebras,
\item none of which is preferred and
\item none of which can be promoted to an $\mathbf{E}_{2}$-$E$-algebra,
\end{enumerate}
see \cite{hopkins2017brauer} for more information.
\begin{warning}
Large parts of the literature are written in terms of the $(2p^{n}-2)$-periodic spectrum $K(n)$ with $\pi_{*}K(n) \simeq \mathbb{F}_{p}[v_{n}^{\pm 1}]$; this cannot be a made into a Morava $K$-theory according to the above definition as it not $2$-periodic. We will not use $K(n)$ in the current work, but in the interest of completness we collect some results about it in \S\ref{section:digression_minimal_morava_k_theory} below.
\end{warning}
Throughout the rest of this section, the letter $K$ will denote a choice of a Morava $K$-theory in the sense of \cref{definition:morava_k_theory}.
\begin{remark}
While two different Morava $K$-theories need not be equivalent as $E$-algebras, one can show that they are always equivalent as $E$-modules \cite[Corollary 3.6.6]{hopkins2017brauer}. In particular, they are always equivalent as spectra.
\end{remark}
Observe that we have $K_{*} \simeq E_{*} / \mfrak E_{*}$ by definition, so $K_{*} \simeq \kappa[u^{\pm 1}]$, which is a graded field. Note that the first isomorphism is completely canonical, while the second is not, as it depends on the choice of a unit $u \in K_{2}$. In any case, it follows that the ring spectrum $K$ is even periodic, and so complex-orientable. The reduction map $E^{0}(BS^{1}) \rightarrow K^{0}(BS^{1})$ induces a canonical isomorphism $\textnormal{Spf}(K^{0}(BS^{1})) \simeq \gzero$.
\begin{remark}
It is not strictly necessary to start relative to a chosen Lubin--Tate spectrum. Instead, one could say that an $\mathbf{E}_{1}$-algebra $K$ in spectra is a \emph{Morava $K$-theory} if there exists an isomorphism $K_{*} \simeq \kappa[u^{\pm 1}]$ with $\kappa$ a perfect field and such that $\gzero := \textnormal{Spf}(K^{0}(BS^{1}))$ is of finite height. One can show that in this case $K$ can be canonically lifted to an algebra over the corresponding Lubin--Tate ring spectrum. Thus, we lose no generality by working relative to $E$.
\end{remark}
In this paper, we will be interested in the $K$-based Adams spectral sequence. By standard arguments, the $E_{2}$-page of this spectral sequence has a description in terms of homological algebra of comodules over the Hopf algebroid $K_{*}K$. Thus, we begin by giving a partial description of the latter.
Our strategy is to exploit the $E$-algebra structure on $K$ to divide $K_{*}K$ into two parts, the first of which admits a convenient interpretation in terms of formal groups and the second of which is more mysterious, but manageably small. We start with the latter, for which we will need to work relative to $E$.
\begin{notation}
If $M, N$ are $E$-modules, we will write
\[
M_{*}^{E}N := \pi_{*}(M \otimes_{E} N)
\]
for their homology relative to $E$ and
\[
M^{*}_{E}N := [N, M]^{*}_{E} \simeq \pi_{-*} F_{E}(N, M)
\]
for their relative cohomology. Here, $F_{E}$ is the internal mapping $E$-module; that is, the right adjoint to the tensor product of $E$-modules.
\end{notation}
By usual arguments involving flatness of $K_{*}^{E}K$ over $K_{*}$, the map
\[
K \otimes_{E} K \rightarrow K \otimes_{E} K \otimes_{E} K
\]
induces a $K_{*}$-coalgebra structure on $K_{*}^{E}K$. Note that this is an honest coalgebra structure; that is, the left and right units coincide, as they are necessarily maps of $E_{*}$-algebras, of which $K_{*}$ is a quotient. Together with multiplication, this makes $K_{*}^{E}K$ into a Hopf algebra, which a priori need not be either commutative nor cocommutative.
\begin{lemma}
\label{lemma:internal_k_homology_dual_to_exterior_algebra}
The coalgebra structure on $K_{*}^{E}K$ is dual to an exterior algebra on an $n$-dimensional vector space $V$ in degree $-1$. In particular, $K_{*}^{E}K$ is cocommutative.
\end{lemma}
\begin{proof}
Since the coefficients of $K$ form a field, we have an isomorphism of algebras
\[
\Hom_{K_{*}}(K_{*}^{E}K, K_{*}) \simeq [K, K]_{E},
\]
where on the right hand side we have maps of $E$-modules. One can show that the latter is always isomorphic to an exterior algebra over a vector space of the needed dimension, see \cite[Proposition 6.5.1]{hopkins2017brauer}.
\end{proof}
\begin{remark}
\label{remark:relative_homology_always_commutative}
Hopkins and Lurie show that $K_{*}^{E}K$ is actually isomorphic to an exterior algebra \emph{as a Hopf algebra}, though there is in general no canonical such isomorphism \cite[Proposition 5.2.4]{hopkins2017brauer}. In particular, it is also always commutative.
\end{remark}
\begin{remark}
\label{remark:kunneth_spectral_sequence_collapses_for_kek}
As both sides are $K_{*}$-vector spaces of dimension $2^{n}$, the latter by \cref{lemma:internal_k_homology_dual_to_exterior_algebra}, the K\"{u}nneth spectral sequence
\[
\Tor^{t, s}_{E_{*}}(K_{*}, K_{*}) \implies K_{s+t}^{E}K
\]
collapses. Indeed, since $E_{0}$ is regular ring of dimension $n$, the relevang $\Tor$-groups are canonically isomorphic to an exterior algebra on $n$ generators.
\end{remark}
The second, ``understandable'' part of $K_{*}K$ is the image of the canonical map from $E_{*}E$, which can be identified with $K_{*}E$. The analogue of this Hopf algebroid for the minimal Morava $K$-theory of \S\ref{section:digression_minimal_morava_k_theory} is denoted in Ravenel's book by $\Sigma(n)$, see \cite[6.2]{ravenel2003complex}.
Notice that since $E_{*}E$ is flat over $E_{*}$, we have canonical isomorphisms
\[
K_{*}E \simeq K_{*} \otimes_{E_{*}} E_{*}E \simeq E_{*} E / \mfrak,
\]
which in turn has the following consequence.
\begin{proposition}
\label{proposition:ke_comodules_independnent_of_k}
The category of $K_{*}E$-comodules is independent of $K$; more precisely, for any two choices of Morava $K$-theories, possibly over different Lubin--Tate spectra, these categories of comodules are canonically equivalent as symmetric monoidal categories.
\end{proposition}
\begin{proof}
By a result of Hovey--Strickland, see \cite[Theorem C]{hovey2005comodules}, the categories of $E_{*}E$-comodules for varying Lubin--Tate spectra $E$ are canonically equivalent as locally graded symmetric monoidal categories.
The category of $K_{*}E$-comodules can be identified with $K_{*}$-modules in the symmetric moonidal category $\Comod_{E_*E}$, and $K_{*} \simeq E_{*} / \mfrak$ itself is uniquely determined as the minimal quotient of $E_{*}$ as a comodule by \cite[Theorem D]{hovey2005comodules}. Thus, the equivalence of Hovey and Strickland induces one between categories of $K_{*}E$-comodules.
\end{proof}
In fact, we can give a direct algebro-geometric description of this category of comodules. By standard results about Landweber exact homology theories we have
\[
E_{*}E \simeq E_{*} \otimes_{BP_{*}} BP_{*}BP \otimes_{BP_{*}} E_{*}.
\]
and by again invoking flatness we see that
\[
K_{*}E \simeq K_{*} \otimes_{BP_{*}} BP_{*}BP \otimes_{BP_{*}} E_{*}
\]
Since the ideal $I_{n}$ is invariant, we can instead write
\[
K_{*}E \simeq K_{*} \otimes_{BP_{*}} BP_{*}BP \otimes_{BP_{*}} K_{*}.
\]
This is a familiar Hopf algebroid, commutative ring homomorphisms out of which classify pairs of homomorphisms $f_{1}, f_{2}: K_{*} \rightarrow A$ together with a strict isomorphism of the resulting formal groups $(f_{1})^{*} \gzero = (f_{2})^{*} \gzero$. In other words, we have a pullback of algebraic stacks
\[\begin{tikzcd}
{\mathrm{Spec}(K_{*}E)} & {\mathrm{Spec}(K_{*})} \\
{\mathrm{Spec}(K_{*})} & {\mathcal{M}_{fg}^{\omega=\mathrm{triv}}}
\arrow[from=1-2, to=2-2]
\arrow[from=2-1, to=2-2]
\arrow[from=1-1, to=2-1]
\arrow[from=1-1, to=1-2],
\end{tikzcd}\]
where $\mathcal{M}_{fg}^{\omega=\mathrm{1}}$ is the moduli of formal groups with a trivialized Lie algebra. The maps from $\Spec(K_{*})$ are faithfully flat surjections onto the height $n$ point in this moduli stack, and we deduce that the groupoid $(\Spec(K_{*}), \Spec(K_{*}E))$ is its presentation.
\begin{proposition}
\label{proposition:ke_comodules_same_as_qcoh_sheaves}
The category of even graded $K_{*}E$-comodules is equivalent to the category of quasi-coherent sheaves over the moduli stack $\mathcal{M}_{fg}^{n}$ of formal groups of height exactly $n$.
\end{proposition}
\begin{proof}
By the discussion above, the category of ungraded $K_{*}E$-comodules can be identified with the category of quasi-coherent sheaves over the moduli stack $\mathcal{M}_{fg}^{n, \omega=1}$ of formal groups of height $n$ with trivialized Lie algebra.
The even grading of $K_{*}E$ corresponds under this equivalence to the $\mathbf{G}_{m}$-action on the chosen trivialization of the Lie algebra, so that even graded comodules can be identified with the quasi-coherent sheaves over the quotient stack
\[
\faktor{\mathcal{M}_{fg}^{n, \omega=1}}{\mathbf{G}_{m}} \simeq \mathcal{M}_{fg}^{n},
\]
which is exactly the claim.
\end{proof}
The above discussion identifies $K_{*}E$ with familiar objects from the theory of formal groups, which is why we referred to it above as the ``understandable'' part of $K_{*}K$. The following gives some control over how these two parts are related.
\begin{lemma}
\label{lemma:eotimesk_is_a_direct_sum_of_ks_as_an_emodule}
Consider $E \otimes K$ as an $E$-module with the module structure inherited only from the left factor. Then, $E \otimes K$ is equivalent as an $E$-module to a direct sum of $K$.
\end{lemma}
\begin{proof}
Since $E_{*}E$ is flat over $E_{*}$, it follows from \cref{corollary:flat_emodule_is_klocally_profree} below that $L_{K}(E \otimes E)$ is equivalent as an $E$-modules to a $K$-local direct sum of $E$. This is a corollary to the work of Hovey and it is independent of the arguments given here.
Thus, it follows that there is a map $\bigoplus E \rightarrow L_{K}(E \otimes E)$ which is an equivalence after applying $- \otimes K$. As the relative tensor product $- \otimes_{E} K$ can be computed using the bar construction, it also takes $K$-local equivalences of $E$-modules to equivalences. We deduce that
\[
\bigoplus K \simeq (\bigoplus E) \otimes_{E} K \simeq L_{K}(E \otimes E) \otimes_{E} K \simeq E \otimes E \otimes_{E} K \simeq E \otimes K,
\]
which is what we wanted to show.
\end{proof}
\begin{remark}
Note that one can alternatively consider $E \otimes K$ as an $E$-module with the module structure inherited form the right factor. In this case, it is clear that it is a direct sum of $K$ as an $E$-module, as it is even a module over the $E$-algebra $K$.
\end{remark}
\begin{corollary}
\label{corollary:kunneth_for_kk_collapses}
Consider $K \otimes E$ as an $E$-module with the module structure inherited only from the right. Then, the K\"{u}nneth spectral sequence
\[
\Tor_{E_{*}}(K_{*}E, K_{*}) \implies \pi_{*}((K \otimes E) \otimes_{E} K) \simeq K_{*}K
\]
in $E$-modules collapses.
\end{corollary}
\begin{proof}
This is immediate from \cref{lemma:eotimesk_is_a_direct_sum_of_ks_as_an_emodule} and \cref{remark:kunneth_spectral_sequence_collapses_for_kek}, as the K\"{u}nneth spectral sequence depends only on the $E$-module structure.
\end{proof}
\begin{proposition}
\label{proposition:map_from_ke_to_kk_injective}
The map $K_{*}E \rightarrow K_{*}K$ is injective, central, and presents the target as a free $K_{*}E$-module of rank $2^{n}$.
\end{proposition}
\begin{proof}
To see that the map is central, notice that $K \otimes K$ is canonically an $E \otimes E$-algebra. It follows that the map $E_{*}E \rightarrow K_{*}K$ is central, and thus so must be its image $K_{*}E \simeq K_{*} \otimes_{E_{*}} E_{*}E$.
As a consequence of \cref{corollary:kunneth_for_kk_collapses}, the K\"{u}nneth spectral sequence
\[
K_{*}E \otimes \Tor_{E_{*}}(K_{*}, K_{*}) \simeq \Tor_{E_{*}}(K_{*}E, K_{*}) \implies K_{*}K
\]
in $E$-modules collapses. Since the $\Tor$-groups on the left form a $K_{*}$-vector space of dimension $2^{n}$, it follows from the above collapse that $K_{*}K$ has a finite filtration as a $K_{*}E$-module such that the associated graded is free of rank $2^{n}$. Thus, $K_{*}K$ itself must be free of this rank as well.
\end{proof}
\begin{proposition}
\label{proposition:k_homology_of_k_tensored_down}
Passing to homotopy groups in the diagram
\[\begin{tikzcd}
{K \otimes E} & {K \otimes K} \\
{K \otimes_{E} E } & {K \otimes_{E} K}
\arrow[from=1-1, to=2-1]
\arrow[from=1-1, to=1-2]
\arrow[from=2-1, to=2-2]
\arrow[from=1-2, to=2-2]
\end{tikzcd}\]
induces an an isomorphism
\[
K _{*}K \otimes_{K_{*}E} K_{*} \simeq K_{*}^{E}K.
\]
\end{proposition}
\begin{proof}
Both of the K\"{u}nneth spectral sequences for $K \otimes K \simeq (K \otimes E) \otimes _{E} K$ and $K \otimes_{E} K$ collapse, as a consequence of \cref{corollary:kunneth_for_kk_collapses} and \cref{remark:kunneth_spectral_sequence_collapses_for_kek}. Thus, the map $K_{*}K \rightarrow K_{*}^{E}K$ is surjective after passing to associated graded, and so surjective by an inductive argument. The conclusion follows by observing that $K_{*} \otimes_{K_{*}E} K _{*}K \rightarrow K_{*}^{E}K$ is a surjective map of $K_{*}$-vector spaces of the same dimension.
\end{proof}
We are now ready to assemble the above information. By standard arguments, the pair $(K_{*}, K_{*}K)$ acquires the structure of a Hopf algebroid. More precisely, the two maps $K \rightarrow K \otimes K$ induce left and right units, both of which are central as a consequence of \cref{proposition:map_from_ke_to_kk_injective}, as they factor through $K_{*}E$. These make $K_{*}K$ into a $K_{*}$-bimodule, and similarly we get a suitable comultiplication and antipode.
Note that it is common in algebraic topology literature to assume in the definition of a Hopf algebroid that multiplication is commutative. However, this is not strictly necessary; in our case, as the units are central, $K_{*}K$ is a Hopf algebroid in the more general sense of Maltsiniotis \cite{maltsiniotis1992groupoides}. The category of comodules can be defined in the usual way, it will be monoidal with the tensor product lifting that of $K_{*}$-modules. It need not in general be symmetric monoidal.
\begin{remark}
\label{remark:k_k_always_commutative_question}
We believe it is plausible that $K_{*}K$ is in fact always commutative; it is a finite algebra over the commutative $K_{*}E$ with the relative tensor product $K_{*} \otimes _{K_{*}E} K_{*}K \simeq K_{*}^{E} K$ also commutative, as a consequence of \cref{remark:relative_homology_always_commutative}. We were, however, not able to resolve this question. Note that $K_{*}K$ is automatically commutative if we choose a homotopy commutative Morava $K$-theory, as can always be done when $p > 2$ \cite[Proposition 3.5.2]{hopkins2017brauer}.
\end{remark}
\begin{theorem}
\label{theorem:ke_kk_kek_is_an_extension}
The natural maps
\begin{align}
\label{align:short_exact_sequence_of_hopf_algebroids}
(K_{*}, K_{*}E) \rightarrow (K_{*}, K_{*}K) \rightarrow (K_{*}, K_{*}^{E}K)
\end{align}
form an extension of Hopf algebroids in the sense of Ravenel.
\end{theorem}
\begin{proof}
Since $K_{*}^{E}K$ is a cocommutative, consulting Ravenel's definition of an extension \cite[\S A.1.1]{ravenel2003complex}, we see that we have to verify that
\[
\Ext^{0}_{K_{*}^{E}K}(K_{*}, K_{*}) \simeq K_{*}
\]
and
\[
\Ext^{0}_{K_{*}^{E}K}(K_{*}, K_{*}K) \simeq K_{*}E.
\]
The first part is clear, as $K_{*}^{E}K$ is a Hopf algebra, while the second part requires a little bit of work. If we write
\[
K_{*}K \simeq \pi_{*}(K \otimes K) \simeq \pi_{*}(K \otimes_{E} E \otimes K) \simeq K_{*}^{E}(E \otimes K),
\]
then the left $K_{*}^{E}K$-comodule structure on $K_{*}K$ given by the quotient map $K_{*}K \rightarrow K_{*}^{E}K$ corresponds to the standard comodule structure
\[
K_{*}^{E}(E \otimes K) \rightarrow K_{*}^{E}K \otimes_{K_{*}} K_{*}^{E}(E \otimes K)
\]
on the homology $K_{*}^{E}(E \otimes K)$ relative to the $\infty$-category of $E$-modules.
The comodule structure map is determined by the structure of $E \otimes K$ as a left $E$-module with module structure coming only from the left factor. By \cref{lemma:eotimesk_is_a_direct_sum_of_ks_as_an_emodule}, $E \otimes K$ is equivalent as an $E$-module to a direct sum of $K$, and so admits a compatible structure of a $K$-module. Choosing such a $K$-module structure determines an isomorphism
\[
K_{*}^{E}(E \otimes K) \simeq K_{*}^{E}K \otimes \pi_{*}(E \otimes K) \simeq K_{*}^{E}K \otimes E_{*}K
\]
so that
\[
\Ext^{0}_{K_{*}^{E}K}(K_{*}, K_{*}K) \simeq \Ext^{0}_{K_{*}^{E}K}(K_{*}, K_{*}^{E}(E \otimes K)) \simeq \Ext^{0}_{K_{*}^{E}K}(K_{*}, K_{*}^{E}K \otimes E_{*}K) \simeq
E_{*}K
\]
as needed.
\end{proof}
\section{Digression: Minimal Morava $K$-theory}
\label{section:digression_minimal_morava_k_theory}
Large parts of the literature on Morava $K$-theories are written in terms of the \emph{minimal Morava $K$-theory spectrum} $K(n)$, the unique up to equivalence $BP$-module with $K(n)_{*} \simeq \mathbb{F}_{p}[v_{n}^{\pm 1}]$. Note that it is not a Morava $K$-theory in the sense of \cref{definition:morava_k_theory} as it is not $2$-periodic, but as a consequence of nilpotence theorem any Morava $K$-theory is equivalent as a spectrum to a direct sum of $K(n)$ \cite[Lemma 1.8]{nilpotence2}.
Our focus on $2$-periodic Morava $K$-theories stems from our preference for Lubin--Tate spectra as they admit canonical $\mathbf{E}_{\infty}$-structures, unlike the Brown--Peterson spectrum. In this short section, we collect some results about $K(n)$ to highlight similarities and differences with the $E$-algebra version. We do this for completness; these results will not be used elsewhere in the current work.
It is a result of Robinson that $K(n)$ can be made into an $\mathbf{E}_{1}$-ring spectrum, but there is no canonical way to do so \cite{robinson1989obstruction}. A surprising result of Angelveit shows that all choices of $\mathbf{E}_{1}$-multiplication yield equivalent $\mathbf{E}_{1}$-algebras in spectra \cite{angeltveit2011uniqueness}.
One can also ask about less structured multiplication. By a theorem of W\"{u}rgler \cite{wurgler1991morava}, in the homotopy category $h \spectra$ of spectra, the spectrum $K(n)$ admits
\begin{enumerate}
\item if $p > 2$, a unique product $K(n) \otimes K(n) \rightarrow K(n)$;
\item if $p = 2$, exactly two products $K(n) \otimes K(n) \rightarrow K(n)$ differing by the symmetry on $K(n) \otimes K(n)$,
\end{enumerate}
which make it into an associative $BP$-algebra in a way compatible with its $BP$-module structure \cite[Theorem 1.5]{wurgler1991morava}.
The ring of cooperations of $K(n)$ has been computed completely \cite[Theorem 2.4]{wurgler1991morava}. The map $K(n)_{*}BP \rightarrow K(n)_{*}K(n)$ is injective and we have an isomorphism
\[
K(n)_{*}BP \simeq K(n)_{*}[t_{i}] / (v_{n} t_{i}^{p^{n}} - v_{n}^{p^{i}} t_{i}),
\]
where $|t_{i}| = 2p^{i}-2$ for $i \geq 1$. At odd primes, this extends to an isomorphism
\[
K(n)_{*}K(n) \simeq K(n)_{*}BP \otimes _{K(n)_{*}} \Lambda_{K(n)_{*}}(\tau_{0}, \ldots, \tau_{n-1}),
\]
where the latter factor is an exterior algebra on $|\tau_{i}| = 2p^{i}-1$. When $p = 2$, we instead have that
\[
K(n)_{*}K(n) \simeq K(n)_{*}BP \otimes _{K(n)_{*}} K(n)_{*}[\tau_{0}, \ldots, \tau_{n-1}] / (\tau_{i}^{2} = t_{i+1}).
\]
This is analogous to the variation in the dual Steenrod algebra depending on whether $p$ is even or odd. Note that the latter computation implies that $K(n)_{*}K(n)$ is commutative also at $p=2$, despite $K(n)$ not being homotopy commutative. The corresponding question in the $2$-periodic case is open, see \cref{remark:k_k_always_commutative_question}.
\section{Milnor modules}\label{sec:milnormodules}
As we have seen in the previous section, there are many different Morava $K$-theory spectra, which makes any discussion of this subject dependent on these choices. However, as in the case of \cref{proposition:ke_comodules_independnent_of_k}, one show that many cohomological invariants are canonical and do not depend on any choices. In this section, we will make this canonical nature transparent by introducing an absolute variant of Hopkins and Lurie's categories of \emph{Milnor modules}, see \cite{hopkins2017brauer}.
Before we work in the absolute case, let us first describe the case relative to a fixed Lubin--Tate spectrum $E$. Note that by the uniqueness part of Goerss--Hopkins--Miller, this is the same as choosing a perfect field equipped with a finite height formal group.
\begin{definition}
We say that an $E$-module is \emph{molecular} if it is equivalent to a finite sum of shifts of Morava $K$-theories of $E$, and denote the full subcategory of $E$-modules on the molecular objects by $\Mod^{mol}_{E}$.
\end{definition}
\begin{remark}
Note that since any two Morava $K$-theories of $E$ are equivalent as $E$-modules, the notion of being molecular does not depend on any choices.
\end{remark}
\begin{definition}
The category of $E$-based \emph{Milnor modules} is the category
\[
\Mil_{E} := P_{\Sigma}(\Mod^{mol}_{E}, \Ab)
\]
of additive presheaves of abelian groups on molecular $E$-modules.
\end{definition}
By construction, the category of Milnor modules is a compactly generated Grothendieck abelian category. Since molecular $E$-modules are stable under the $E$-tensor product, it acquires a non-unital symmetric monoidal structure via left Kan extension.
One can show this extends uniquely to a unital symmetric monoidal structure \cite[Propositions 4.4.1 and 4.4.10]{hopkins2017brauer}.
The main property of Milnor modules is that it canonically captures cohomology of \emph{any} Morava $K$-theory relative to $E$, in the following sense.
\begin{proposition}
\label{proposition:milnor_modules_same_as_kek_comodules}
Let $K$ be a Morava $K$-theory of $E$. Then, the left Kan extension of the functor
\[
K_{*}^{E}(-): \Mod_{E}^{mol} \rightarrow \Comod_{K_{*}^{E}K}
\]
which associates to any molecular $E$-module its relative $K$-homology induces a monoidal equivalence
\[
\Mil_{E} \simeq \Comod_{K_{*}^{E}K}
\]
between Milnor modules and $K_{*}^{E}K$-comodules.
\end{proposition}
\begin{proof}
This is \cite[Corollary 6.4.13]{hopkins2017brauer}, where we replaced the category of modules over $K^{*}_{E}K$ with comodules over its linear dual.
\end{proof}
\begin{corollary}
\label{corollary:kek_comod_doesnt_depend_on_k}
As a monoidal category, $\Comod_{K_{*}^{E}K}$ does not depend on the choice of $K$, but only on its corresponding formal group.
\end{corollary}
\begin{remark}
Note that in the context of \cref{corollary:kek_comod_doesnt_depend_on_k}, the situation is somewhat better than just saying that the categories $\Comod_{K_{*}^{E}K}$ for fixed $E$ and varying $K$ are all equivalent to each other. In fact, they are all canonically equivalent to a single category which doesn't depend on any choices, namely the category of Milnor modules of $E$.
\end{remark}
\begin{remark}
\label{remark:milnor_modules_and_homotopy_commutativity}
The category $\Mil_{E}$ is not only canonical, but in a certain sense gives the right answer even if $K_{*}^{E}K$-comodules do not. As an example, the restricted Yoneda embedding
\[
y\colon \Mod_{E} \rightarrow \Mil_{E},
\]
which can be thought of as a way of taking relative $K$-homology without choosing $K$, is always a symmetric monoidal functor \cite[Variant 4.4.11]{hopkins2017brauer}.
On the other hand, even if $K$ is not homotopy commutative, $K_{*}^{E}K$ is both commutative and cocommutative so that the category $\Comod_{K_{*}^{E}K}$ acquires a symmetric monoidal structure from $K_{*}$-vector spaces. However, in this case the relative $K$-homology functor
\[
K_{*}^{E}(-): \Mod_{E} \rightarrow \Comod_{K_{*}^{E}K}
\]
need not be symmetric monoidal. In fact, one can show that if $p = 2$ and $n>0$, this functor is \emph{never} symmetric monoidal, even when restricted to molecular modules. Indeed, any such functor would induce a symmetric monoidal equivalence $\Comod_{K_{*}^{E}K} \simeq \Mil_{E}$, which one can show is not possible at the even prime\footnote{Personal communication with Jacob Lurie.}.
If $K$ is homotopy commutative, then the above homology theory is canonically symmetric monoidal so that we have a symmetric monoidal equivalence $\Comod_{K_{*}^{E}K} \simeq \Mil_{E}$. Informally, any algebraic structure which one can put on $\Comod_{K_{*}^{E}K}$ in a way compatible with the homology functor from $E$-modules must already be present in the category of Milnor modules.
\end{remark}
Let us now describe the global analogue of this situation, which describes the Adams spectral sequence in spectra rather than $E$-modules; alternatively, which describes comodules over $K_{*}K$ rather than the relative variant $K_{*}^{E}K$.
In Miller's approach to the Adams spectral sequence, the latter is determined by the class of maps which are $K$-split; that is, which are split epimorphisms after applying $K \otimes -$. This only depends on the structure of $K$ as a spectrum and does not require any coherent multiplication.
As a consequence of the nilpotence theorem, all Morava $K$-theories at a fixed prime and height are, as spectra, direct sums of the minimal Morava $K$-theory spectrum of \cref{section:digression_minimal_morava_k_theory}. It follows that they all determine the same class of epimorphisms, and hence isomorphic Adams spectral sequences. In fact, the class of epimorphisms is enough to describe canonically the abelian category whose $\Ext$-groups form the $E_{2}$-page. To do so, we will consider sheaves with respect to a certain topology, as in the construction of synthetic spectra \cite{pstrkagowski2018synthetic}.
\begin{definition}
The \emph{$K_{*}$-epimorphism} topology on the $\infty$-category $\spectra^{fin}$ of finite spectra is the Grothendieck pretopology in which a covering families $\{ X \rightarrow Y \}$ consist of a single map such that $K_{*} X \rightarrow K_{*} Y$ is surjective.
\end{definition}
\begin{remark}
A reader familiar with \cite{pstrkagowski2018synthetic} might recall that in the construction of synthetic spectra, one works with the site of those finite spectra $X$ such that $K_{*}X$ is projective as a $K_{*}$-module \cite[Definition 3.12]{pstrkagowski2018synthetic}. Since Morava $K$-theories are fields, this yields the class of all finite spectra, as above.
\end{remark}
\begin{definition}
The category of \emph{absolute Milnor modules} is the category
\[
\Mil_{abs} := \Shv_{\Sigma}(\Sp^{fin}, \Ab)
\]
of additive sheaves of abelian groups on finite spectra with respect to the $K_{*}$-epimorphism topology.
\end{definition}
The category of absolute Milnor modules is a compactly-generated Grothendieck abelian category with a symmetric monoidal structure induced by that of the tensor product of finite spectra.
\begin{remark}
A reader familiar with synthethic spectra will immediately recognize that the category of Milnor modules can be canonically identified as $\Mil_{abs} \simeq \euscr{S}yn_K^{\heartsuit}$, the heart of the stable $\infty$-category of $K$-based synthetic spectra of \cite{pstrkagowski2018synthetic}.
\end{remark}
The following is the absolute analogue of \cref{proposition:milnor_modules_same_as_kek_comodules}.
\begin{proposition}
\label{proposition:absolute_milnor_modules_same_as_comodules}
If $K$ is a Morava $K$-theory, then the left Kan extension of the functor
\[
K_{*}: \Sp^{fin} \rightarrow \Comod_{K_{*}K}
\]
induces a monoidal equivalence
\[
\Mil_{abs} \simeq \Comod_{K_{*}K}
\]
between absolute Milnor modules and $K_{*}K$-comodules.
\end{proposition}
\begin{proof}
This is an instance of \cite[Remark 3.26]{pstrkagowski2018synthetic}.
\end{proof}
As a consequence, we deduce the following result proven earlier by Hovey and Strickland using slightly different methods \cite[\S 8]{hovey2005comodules}.
\begin{corollary}
\label{corollary:kk_comodules_independent_of_k}
The monoidal category $\Comod_{K_{*}K}$ of $K_{*}K$-comodules does not depend on the choice of Morava $K$-theory, but only on the prime and height.
\end{corollary}
\begin{remark}
By the same argument as in the relative case of \cref{remark:milnor_modules_and_homotopy_commutativity}, if $K$ is homotopy commutative, then we have a canonical symmetric monoidal equivalence $\Mil_{abs} \simeq \Comod_{K_{*}K}$.
\end{remark}
\begin{remark}[Homological residue fields]
In the terminology of Balmer, the category of Milnor modules is the \emph{homological residue field} of the tensor-triangulated category of finite spectra at the prime ideal of spectra of type at least $n$, see \cite{balmer2017nilpotence}, \cite{balmer2020computing}.
\end{remark}
\begin{remark}[Milnor modules as sheaves]
\label{remark:relation_between_milnor_modules_plus_independence_of_ce_sseq}
The reader can observe that there is a certain asymmetry in our definitions of Milnor modules and absolute Milnor modules; namely, in the former case we use as our indexing category only molecular $E$-modules and in the latter case we use \emph{all} finite spectra. Moreover, in the latter case we use sheaves while in the former presheaves are already enough.
In fact, this difference is but a trick of light, as there is an alternative description of Milnor modules which is more in line with the absolute case. Namely, one can show that $\Mil_E$ is equivalent to the category of sheaves on the $\infty$-category $\Mod^{perf}_{E}$ of perfect $E$-modules with respect to the $K_{*}^{E}(-)$-epimorphism topology. Under this topology, the functor
\[
E \otimes -\colon \spectra^{fin} \rightarrow \Mod_{E}^{perf}
\]
becomes a morphism of sites, so that left Kan extension yields a cocontinuous, symmetric monoidal functor
\[
\Mil_{abs} \rightarrow \Mil_{E}.
\]
In this sense absolute Milnor modules determine a relative Milnor module, for any choice of $E$, justifying the terminology.
Under the equivalences of \cref{proposition:milnor_modules_same_as_kek_comodules} and \cref{proposition:absolute_milnor_modules_same_as_comodules} the above left Kan extension corresponds to the forgetful functor
\[
\Comod_{K_{*}K} \rightarrow \Comod_{K_{*}^{E}K}.
\]
The image of the monoidal unit under the corresponding right adjoint can be identified with the Hopf algebroid $K_{*}E$ studied previously. As the Adams spectral sequence based on $K_{*}E$ in the appropriate derived $\infty$-categories can be identified with the Cartan--Eilenberg spectral sequence associated to this extension of Hopf algebroids \cite[1.1]{belmont2020cartan}, we deduce that the Cartan--Eilenberg spectral sequence itself can be constructed using only the categories of Milnor modules, so it only depends on $E$.
\end{remark}
\section{Cohomology}
\label{sec:cohomology}
In this section, we establish finiteness properties of the cohomology of $K_{*}K$-comodules using the Cartan--Eilenberg spectral sequence.
We have seen in \cref{theorem:ke_kk_kek_is_an_extension} that we have an extension of Hopf algebroids
\[
(K_{*}, K_{*}E) \rightarrow (K_{*}, K_{*}K) \rightarrow (K_{*}, K_{*}^{E}K)
\]
which induces a Cartan--Eilenberg spectral sequence of signature
\[
\Ext^{s}_{K_{*}E}(K_{*}, \Ext^{s^{\prime}}_{K_{*}^{E}K}(K_{*}, K_{*})) \implies \Ext^{s + s^{\prime}}_{K_{*}K}(K_{*}, K_{*}),
\]
see \cite[A.1.3.14]{ravenel2003complex}. This can be thought as approximating the cohomology of $K_{*}K$ using simplier pieces, namely the cohomology of $K_{*}E$ and $K_{*}^{E}K$. To make best use of this, we need to begin by understanding the latter two.
We have seen in \cref{sec:moravaktheories} that the Hopf algebroid $(K_{*}, K_{*}E)$ is an even graded presentation of the moduli of formal groups of height exactly $n$, see \cref{proposition:ke_comodules_same_as_qcoh_sheaves}. To identify its cohomology, it is convenient to choose a particularly small presentation of the latter moduli stack, which is what we will do.
\begin{notation}
\label{notation:small_choice_of_morava_e_theory}
Let $\gzero$ be the Honda formal group law over $\mathbb{F}_{q}$, where $q = p^{n}$, and let $E_{n}$ be the associated Lubin--Tate spectrum and $K_{n}$ a Morava $K$-theory associated to $E_{n}$. Note that we have a non-canonical isomorphism
\[
\pi_{0} E_{n} \simeq W(\mathbb{F}_{q})[[u_{1}, \ldots, u_{n-1}]].
\]
We write $\mathbb{G}_{n} \simeq \mathbb{S}_{n} \rtimes \textnormal{Gal}(\mathbb{F}_{q} / \mathbb{F}_{p})$ for the \emph{extended Morava stabilizer group}, the automorphism group of the pair $\mathbb{F}_{q}$ together with $\gzero$. By the uniquenses part of Goerss--Hopkins--Miller, the action of $\mathbb{G}_{n}$ extends to an action on the Lubin--Tate spectrum $E_{n}$.
\end{notation}
\begin{lemma}
\label{lemma:iso_of_hopf_algebroid_of_kn_and_the_gn_one}
The association $\mathbb{G}_{n} \times (K_{n})_{*}E_{n} \rightarrow (K_{n})_{*}$ sending an element $g \in \mathbb{G}_{n}$ and a homotopy class $S^{k} \rightarrow K_{n} \otimes E_{n}$ to the homotopy class of the composite
\[
\begin{tikzcd}
{S^{k}} & {K_{n} \otimes E_{n}} & {K_{n} \otimes E_{n} } & K_{n}
\arrow["", from=1-1, to=1-2]
\arrow["{id \otimes g}", from=1-2, to=1-3]
\arrow["m", from=1-3, to=1-4],
\end{tikzcd}
\]
induces an isomorphism $\map_{\cts}(\mathbb{G}_{n}, (K_{n})_{*}) \simeq (K_{n})_{*}E_{n}$ of algebras between the $K_{n}$-homology of $E_{n}$ and continuous functions on the Morava stabilizer group.
\end{lemma}
\begin{proof}
This is a result of Strickland, reduced modulo $\mathfrak{m}$ \cite[Theorem 12]{strickland2000gross}.
\end{proof}
One can check that the above isomorphism is compatible with the Hopf algebroid structure, which is induced from the group structure on $\mathbb{G}_{n}$ and its action on $(K_{n})_{*} \simeq \mathbb{F}_{q}[u^{\pm 1}]$, which yields the following.
\begin{corollary}
\label{corollary:ke_comodules_same_as_vector_spaces_with_gn_action}
For any Morava $K$-theory $K$, there is a canonical equivalence of categories between $K_{*}E$-comodules and graded $\mathbb{F}_{q}[u^{\pm 1}]$-vector spaces equipped with a continuous action of the extended Morava stabilizer group $\mathbb{G}_{n}$.
\end{corollary}
\begin{proof}
We have proven previously in \cref{proposition:ke_comodules_independnent_of_k} that the category of $K_{*}E$-comodules is indendent of the choice of $K$. For the particular choice of $K = K_{n}$ as above, this isomorphism is induced by \cref{lemma:iso_of_hopf_algebroid_of_kn_and_the_gn_one}.
\end{proof}
\begin{remark}
In terms of the equivalence $\Comod_{K_{*}E}^{ev} \simeq \QCoh(\mathcal{M}_{fg}^{n})$ of \cref{proposition:ke_comodules_same_as_qcoh_sheaves}, \cref{corollary:ke_comodules_same_as_vector_spaces_with_gn_action} is a consequence of the fact that the map $\mathbb{F}_{q} \rightarrow \mathcal{M}_{fg}^{n}$ classifying the Honda formal group law is a Galois covering with automorphism group $\mathbb{G}_{n}$; we refer to \cite{goerss2008stacks} for further details.
\end{remark}
\begin{corollary}
\label{corollary:k_e_cohomology_same_as_morava_stabilizer}
For any Morava $K$-theory, there is a canonical isomorphism
\[
\Ext_{K_{*}E}^{s, t}(K_{*}, K_{*}) \simeq \mathrm{H}^{*}_{\cts}(\mathbb{G}_{n}, \mathbb{F}_{q}[u^{\pm 1}])
\]
between the cohomology of the Hopf algebraid $(K_{*}, K_{*}E)$ and the continuous cohomology of the Morava stabilizer group.
\end{corollary}
The group $\mathbb{G}_{n} \simeq \mathbb{S}_{n} \rtimes \textnormal{Gal}(\mathbb{F}_{q} / \mathbb{F}_{p})$ is a semidirect product of the automorphism group $\mathbb{S}_{n}$ of $\gzero$ and the Galois group. The former is a $p$-adic Lie group, in fact a group of units in integers in a central division algebra over $\mathbb{Q}_{p}$ of Hasse invariant $\frac{1}{n}$, which gives its excellent cohomological properties.
The action of the Morava stabilizer group on $\mathbb{F}_{q}[u^{\pm 1}]$ is easy to describe. The Galois group acts by its standard action on $\mathbb{F}_{q}$. The action of $\mathbb{S}_{n}$ factors through the leading coefficient homomorphism $\rho\colon \mathbb{S}_{n} \rightarrow \mathbb{F}_{q}^{\times}$ and is determined by the formula $s \cdot u = p(s)u$ for $s \in \mathbb{S}_{n}$. The kernel of $\rho$ is the strict Morava stabilizer group $S_{n}$; it is a pro-$p$-group acting trivially. Morava \cite{morava1985noetherian} essentially proved:
\begin{theorem}[Morava]
\label{thm:cohomfiniteness}
We have isomorphisms of continuous cohomology groups
\[
\rmH_{\cts}^*(\mathbb{G}_n,\mathbb{F}_{q}[u^{\pm 1}]) \simeq \rmH^{*}_{\cts}(\mathbb{S}_{n}, \mathbb{F}_{q}[u^{\pm 1}])^{\mathrm{Gal}} \simeq \rmH^{*}_{\cts}(S_{n}, \mathbb{F}_{q}[u^{\pm 1}])^{\mathbb{F}_{q}^{\times} \rtimes \mathrm{Gal}}.
\]
Moreover, these groups are finite in each bidegree.
If $(p-1) \nmid n$, this is a Poincar\'{e} duality algebra over $\rmH_{\cts}^{0} \simeq \mathbb{F}_{p}[v_{n}^{\pm 1}]$ of dimension $n^{2}$, where $v_{n} := u^{p^{n}-1}$, in particular their total dimension is finite.
\end{theorem}
\begin{proof}
These isomorphism are induced by the collapse of the respective Lyndon--Hochschild--Serre spectral sequences, as neither $\mathrm{Gal}$ nor $\mathbb{F}_{q}^{\times}$ have any higher cohomology with these coefficients, the first one by Galois theory and the second one as it is of order coprime to $p$.
The finiteness statements are due to Lazard, as $S_{n}$ is a $p$-adic analytic Lie group of dimension $n^{2}$ and so has a finite index subgroup whose cohomology is isomorphic to an exterior algebra \cite[Theorem 5.1.5]{symonds2000cohomology} of dimension $n^{2}$. If $(p-1) \nmid n$, then $S_{n}$ has no $p$-torsion, see \cite[Theorem 3.2.1]{henn1998centralizers}, and so it has the same cohomological dimension as all of its finite index subgroups \cite{serre1965dimension}.
\end{proof}
\begin{corollary}\label{cor:cohomfiniteness}
\label{corollary:coh_finiteness_of_ke_comodules}
For any finite-dimensional $K_{*}E$-comodules $M, N$, the groups $\Ext^{t, s}_{K_{*}E}(M, N)$ are finite in each bidegree. If $(p-1) \nmid n$, then they vanish for $s > n^{2}$.
\end{corollary}
\begin{proof}
We first claim that any finite-dimensional $K_{*}E$-comodule $M$ has an invariant element; that is, there exists a non-zero map $K_{*} \rightarrow M$ of comodules. Observe that $M$ can be considered as an $E_{*}E$-comodule using restriction along the maps $E_{*}E \rightarrow K_{*}E$ and $E_{*} \rightarrow K_{*}$. As $M$ is finitely generated over $E_{*}$, it admits a non-zero map $E_{*} \rightarrow M$ of comodules by a result of Hovey and Strickland \cite[Theorem D]{hovey2005comodules}. As the target is $\mfrak$-torsion, the map will factor through a map $K_{*} \rightarrow M$ of $K_{*}E$-comodules, which is what we were looking for.
Observe that the result holds when $M = N = K_{*}$ as a combination of \cref{corollary:k_e_cohomology_same_as_morava_stabilizer} and \cref{thm:cohomfiniteness}. Since $K_{*}$ is a field, it follows that from the above that any finite-dimensional $K_{*}E$ comodule can be obtained as an iterated extension of $K_{*}$. Thus, the general case is proven by presenting $M, N$ as such iterated extensions and using the corresponding long exact sequences of $\Ext$-groups.
\end{proof}
To get hold of the Cartan--Eilenberg spectral sequence, we also need to have some control over the cohomology of $K_{*}^{E}K$. By a result of Hopkins and Lurie, see \cref{remark:relative_homology_always_commutative}, the latter is isomorphic to an exterior algebras as a Hopf algebra and so its cohomology is necessarily a polynomial ring. In fact, the same authors identify the generators of this polynomial ring, as we now recall.
Suppose that $x \in \mfrak$ is an element of the maximal ideal. We then have a fibre sequence
\[
\Sigma^{-1} E \rightarrow E/x \rightarrow E
\]
of $E$-modules which induces a short exact sequence
\[
0 \rightarrow K_{*}[-1] \rightarrow K_{*}^{E}(E/x) \rightarrow K_{*} \rightarrow 0
\]
of $K_{*}^{E}K$-comodules. Let us denote by $\psi(x)$ the correspoding element of $\Ext^{1, -1}_{K_{*}^{E}K}(K_{*}, K_{*})$.
\begin{proposition}[Hopkins--Lurie]
\label{proposition:hl_cohomology_of_kek}
The association $x \mapsto \psi(x)$ induces a map
\[
\mfrak / \mfrak^{2} \rightarrow \Ext_{K_{*}^{E}K}^{1, -1}(K_{*}, K_{*})
\]
which extends to an isomorphism of bigraded rings
\[
K_{*} \otimes_{\kappa} \Sym_{\kappa}^{*} (\mfrak / \mfrak^{2}) \simeq \Ext_{K_{*}^{E}K}(K_{*}, K_{*}).
\]
\end{proposition}
\begin{proof}
As we show in \cref{proposition:milnor_modules_same_as_kek_comodules}, the category of $K_{*}^{E}K$-comodules can be identified, as a monoidal category, with the category of Milnor modules associated to $E$. It follows that we have
\[
\Ext^{*}_{K_{*}^{E}K}(K_{*}, K_{*}) \simeq \Ext^{*}_{\Mil_{E}}(y(E), y(E)),
\]
the cohomology in Milnor modules, where $y(E)$ is the Milnor module associated to $E$. For the right hand side, this isomorphism is \cite[Proposition 7.2.7]{hopkins2017brauer}.
\end{proof}
\begin{remark}
If we again let $K_{n}$ be any Morava $K$-theory associated to the Honda formal group over $\mathbb{F}_{q}$, it follows that the Cartan--Eilenberg spectral sequence for cohomology of the unit has signature
\[
\rmH_{\cts}^{s_{0}}(\mathbb{G}_{n}, \mathbb{F}_{q}[u^{\pm 1}] \otimes_{\mathbb{F}_{q}} \Sym^{s_{1}}_{\mathbb{F}_{q}}(\mfrak / \mfrak^{2})) \implies \Ext^{s_{0} + s_{1}}_{K_{*}K}(K_{*}, K_{*})
\]
To determine the action on the symmetric algebra, it is convenient to work in the context of Milnor modules, as in the proof of \cref{proposition:hl_cohomology_of_kek}. The category of Milnor modules $\Mil_{E}$ is by construction functorial in automorphisms of $E$ and it it is clear that the association $x \mapsto \psi(x)$ is natural with respect to the action of the Morava stabilizer group. It follows that the action of $\mathbb{G}_{n}$ on this symmetric algebra is the one induced by its action on the cotangent space $\mfrak / \mfrak^{2}$ of the Lubin--Tate ring.
\end{remark}
Using the Cartan--Eilenberg spectral sequence, we can prove the following finiteness statement.
\begin{theorem}
\label{theorem:finiteness_of_ext_kk}
Let $M, N$ be finite-dimensional $K_{*}K$-comodules. Then, the group $\Ext^{s, t}_{K_{*}K}(M, N)$ is finite in each bidegree.
\end{theorem}
\begin{proof}
By replacing $N$ by $N \otimes M^{\vee}$, where $M^{\vee} := \Hom_{K_{*}}(M, K_{*})$ is the linear dual, we can assume that $M = K_{*}$. We have a Cartan--Eilenberg spectral sequence of signature
\[
\Ext^{s_{0}}_{K_{*}E}(K_{*}, \Ext^{s_{1}}_{K_{*}^{E}K}(K_{*}, N)) \implies \Ext^{s_{0}+s_{1}}_{K_{*}K}(K_{*}, N)
\]
The groups contributing to $\Ext^{s}_{K_{*}K}(K_{*}, N)$ are those indexed by $s_{0}, s_{1}$ with $s_{0} + s_{1} = s$, of which there are finitely many. For a fixed $s_{1}$, the group $\Ext^{s_{1}}_{K_{*}^{E}K}(K_{*}, N)$ is finite-dimensional over $K_{*}$ since both $M$ and $K_{*}^{E}K$ are, and we can compute this $\Ext$-group using the cobar complex. The finiteness of $\Ext^{s_{0}}_{K_{*}E}(K_{*}, -)$ applied to these groups follows from \cref{cor:cohomfiniteness}.
\end{proof}
\section{Milnor filtration}\label{sec:milnorfiltration}
In this section we describe a filtration on $K_{*}K$ which arises from its module structure over $E$, which we call the \emph{Milnor filtration}. We show that at odd primes this filtration can be canonically refined to a grading which forces the Cartan--Eilenberg spectral sequence associated to the extension $K_{*}E \rightarrow K_{*}K \rightarrow K_{*}^{E}K$ to collapse.
\begin{remark}
This existence of an additional grading at $p > 2$ is a finite height analogue of the classical argument involving the Cartan--Eilenberg spectral sequence associated to $H \mathbb{F}_{p}$, see \cite[Theorem 4.3.3]{ravenel2003complex}.
\end{remark}
The Milnor filtration will be induced by a suitable filtration of the spectrum $K \otimes K$. Before we work with such objects, we first recall some standard definitions and establish terminology.
A non-negatively \emph{filtered spectrum} is a functor $P_{\bullet}\colon \mathbb{N} \rightarrow \spectra$, where we consider the natural numbers as a partially ordered set. If $X$ is a spectrum, then a \emph{filtration} on $X$ is a filtered spectrum $P_{\bullet}$ together with an equivalence $\varinjlim P_{\bullet} \simeq X$.
\begin{remark}
Note that according to our terminology, any filtered spectrum is canonically a filtration of its colimit.
\end{remark}
To any filtration $P_{\bullet}$ we can attach an associated non-negatively graded spectrum
\[
\gr(P_{\bullet}) = \bigoplus_s\mathrm{gr}_{s}(P_{\bullet})
\]
with the degree $s$ part defined by
\[
\mathrm{gr}_{s}(P_{\bullet}) := \mathrm{cofib}(P_{s-1} \rightarrow P_{s}),
\]
where $P_{-1} = 0$ by convention. The boundary maps
\[
\Sigma^{-1} \mathrm{gr}_{s}(P_{\bullet}) \rightarrow P_{s-1} \rightarrow \mathrm{gr}_{s-1}(P_{\bullet})
\]
make $\Sigma^{-s} \gr_{s}(P_{\bullet})$ into a chain complex in the homotopy category of spectra.
Any filtration has a corresponding spectral sequence of signature
\[
E^{1}_{s, t} := \pi_{t} \mathrm{gr}_{s}(\Sigma^{-s} P_{\bullet}) \implies \pi_{s+t} X,
\]
where the first differential is induced by the chain complex structure on the associated graded \cite[\S 1.2.2]{lurie2016higher}. The increasing filtration on $\pi_{*}X$ associated to this spectral sequence is given by the images of maps $\pi_{*} P_{s} \rightarrow \pi_{*}X$.
There is a natural tensor product on the $\infty$-category of filtered spectra, given by Day convolution. This has the properties that
\begin{enumerate}
\item if $P_{\bullet}$ and $S_{\bullet}$ are filtrations of $X, Y$, then $P_{\bullet} \otimes S_{\bullet}$ is a filtration of $X \otimes Y$ and
\item $\mathrm{gr}(P_{\bullet} \otimes S_{\bullet}) \simeq \gr(P_{\bullet}) \otimes \gr(S_{\bullet})$, where the latter is the tensor product of graded objects.
\end{enumerate}
In fact, the latter equivalence can be upgraded to an isomorphism in the category of chain complexes in the homotopy category of spectra, where we consider the latter as symmetric monoidal with the differential on the tensor product given by the usual Leibniz rule \cite[3.2.15]{raksit2020hochschild}.
\begin{example}
We will need certain special filtrations of $E$-modules, the main reference here being Lurie \cite[\S 7.2.1]{lurie2016higher}, who works with simplicial objects. By the stable Dold--Kan equivalence, this is equivalent to working with filtered objects as we do.
If $P_{\bullet}$ is an $E$-module filtration of some $M$, then we say $P_{\bullet}$ is \emph{$E$-free} if the associated simplicial object is $S$-free with respect to the set $S = \{ \ \Sigma^{n} E \ | \ n \in \mathbb{Z} \ \}$. We say it is an $E$-\emph{resolution} if it is $E$-free and an $S$-hypercover. For precise definitions of these terms, see \cite[Definition 7.2.1.2.]{lurie2016higher}. The two main facts we will need are that
\begin{enumerate}
\item if $P_{\bullet}$ is $E$-free, then the associated graded $E$-module is levelwise free and
\item if it is a resolution, then $\pi_{*} \Sigma^{-n} \gr_{n}(P_{\bullet})$ is a projective resolution of $M_{*}$ in $E$-modules.
\end{enumerate}
Lurie shows that any $E$-module $M$ admits a resolution, which is unique (as an augmented simplicial object) up to simplicial homotopy.
If $P_{\bullet}$ is an $E$-module resolution of $M$ and $X$ is an arbitrary spectrum, then $P_{\bullet} \otimes X$ is a filtration of $M \otimes X$ and so we get a spectral sequence whose signature is readily shown to be
\[
E^{2}_{t, s} := \Tor^{s, t}_{E_{*}}(M_{*}, E_{*}X) \implies M_{*}X
\]
This is the K\"{u}nneth spectral sequence which we previously mentioned in \cref{corollary:kunneth_for_kk_collapses}.
\end{example}
We are now ready to make the key definition of this section.
\begin{definition}
The \emph{Milnor filtration} on $K_{*}K \simeq \pi_{*}(K \otimes K)$ is the filtration induced by the filtration $P_{\bullet} \otimes P_{\bullet}$ of $K \otimes K$, where $P_{\bullet}$ is an $E$-free resolution of $K$ as an $E$-module.
\end{definition}
In fact, the above filtration is readily identified with a familiar one.
\begin{lemma}
\label{lemma:milnor_filtration_coincides_with_the_kunneth_filtration}
The Milnor filtration coincides with the filtration induced by the K\"{u}nneth spectral sequence of signature
\[
\Tor_{E_{*}}^{s, t}(K_{*}, E_{*}K) \implies K_{t+s}K.
\]
\end{lemma}
\begin{proof}
The K\"{u}nneth spectral sequence is induced by the filtration $P_{\bullet} \otimes K$, and the obvious map $P_{\bullet} \rightarrow K$ into the constant filtration of the latter induces a comparison map $P_{\bullet} \otimes P_{\bullet} \rightarrow P_{\bullet} \otimes K$ of filtrations of $K \otimes K$. We claim this induces an isomorphism of spectral sequences starting from the second page onwards, and hence an isomorphism of filtrations of $K_{*}K$.
On the first pages of the corresponding spectral sequences, we are looking at the map
\[
\pi_{*} (\gr(P_{\bullet}) \otimes \gr(P_{\bullet})) \rightarrow \pi_{*}(\gr(P_{\bullet} \otimes K)),
\]
where we have suppressed the suspensions needed to make the above into chain complexes of graded abelian groups.
Note that if $M$ is an arbitrary $E$-module, then $E_{*}M \simeq E_{*}E \otimes_{E_{*}} M$ as $E_{*}E$ is flat over $E_{*}$ so that the corresponding K\"{u}nneth spectral sequence collapses. It follows that since each $\gr(P_{\bullet})$ is levelwise free as an $E$-module, we can rewrite the above map on first pages of the corresponding spectral sequences as
\[
\pi_{*} \gr(P_{\bullet}) \otimes_{E_{*}} E_{*}E \otimes_{E_{*}} \pi_{*} \gr(P_{\bullet}) \rightarrow \pi_{*} \gr(P_{\bullet}) \otimes_{E_{*}} E_{*}E \otimes_{E_{*}} K_{*}.
\]
The tensor product of the first two factors is a non-negatively graded complex of flat $E_{*}$-modules, so that tensoring with it preserves quasi-isomorphism. As $\gr_{*}(P_{\bullet}) \rightarrow K_{*}$ is a quasi-isomorphism, we deduce that so is the map on first pages, ending the argument.
\end{proof}
\begin{corollary}
The Milnor filtration does not depend on the choice of the $E$-free resolution.
\end{corollary}
\begin{proof}
This is immediate from \cref{lemma:milnor_filtration_coincides_with_the_kunneth_filtration}, as any two $E$-free filtrations $P_{\bullet}^{1}$, $P_{\bullet}^{2}$ induce the same K\"{u}nneth filtration, as $P_{\bullet}^{1} \otimes K$ and $P_{\bullet}^{2} \otimes K$ are simplicially homotopic.
\end{proof}
\begin{remark}
\label{remark:milnor_filtration_also_coincides_with_other_kunneth_filtration}
By symmetry, \cref{lemma:milnor_filtration_coincides_with_the_kunneth_filtration} implies that the Milnor filtration also coincides with the filtration induced by the other K\"{u}nneth spectral sequence, namely that of signature
\[
\Tor_{E_{*}}^{s, t}(K_{*}E, K_{*}) \implies K_{t+s}K.
\]
The main advantage of using the Milnor filtration is that its definition is manifestly symmetric, so that it is easier to work with.
\end{remark}
Note that since the Milnor filtration coincides with the K\"{u}nneth one, we immediately deduce that the associated graded object is given by
\[
\Tor_{E_{*}}^{*, *}(K_{*}, E_{*}K) \simeq \Tor_{E_{*}}^{*, *}(K_{*}, K_{*}) \otimes E_{*}K,
\]
where the additional grading comes from the external $\Tor$-grading. In particular, the elements of filtration zero are given by
\[
F_{0}^{Mil}(K_{*}K) = E_{*}K,
\]
the image of the canonical map $E_{*}E \rightarrow K_{*}K$. As the latter is a sub-Hopf algebroid, this suggests that the Milnor filtration is compatible with the Hopf algebroid structure. This is indeed the case, as we will now show.
\begin{lemma}
\label{lemma:milnor_filtration_is_multiplicative}
The Milnor filtration on $K_{*}K$ is multiplicative.
\end{lemma}
\begin{proof}
We first claim that the multiplication map $K \otimes K \rightarrow K$ lifts to a morphism $P_{\bullet} \otimes P_{\bullet} \rightarrow P_{\bullet}$ of filtrations.
Observe that since $K$ is an $E$-algebra, the multiplication factors canonically through a map $K \otimes_{E} K \rightarrow K$. Observe that $P_{\bullet} \otimes_{E} P_{\bullet}$ is an $E$-free filtration of $K \otimes_{E} K$ (although not a resolution, but this is not needed). Then, since $P_{\bullet}$ is an $E$-module resolution of $K$, it follows from \cite[Proposition 7.2.1.5]{lurie2016higher} (where $C = K$) that the map $K \otimes_{E} K \rightarrow K$ of $E$-modules lifts to a map $P_{\bullet} \otimes_{E} P_{\bullet} \rightarrow P_{\bullet}$ of filtrations. The needed lift of $K \otimes K \rightarrow K$ is then given as composition
\[
P_{\bullet} \otimes P_{\bullet} \rightarrow P_{\bullet} \otimes_{E} P_{\bullet} \rightarrow P_{\bullet},
\]
where the first map is the canonical one to the relative tensor product.
We now claim that the multiplication on $K \otimes K$ also lifts to a map of corresponding filtrations of spectra, showing that it respects filtrations on homotopy groups. This product map is given up to homotopy by the composite
\[
\begin{tikzcd}
{K \otimes K \otimes K \otimes K} & {} & {K \otimes K \otimes K \otimes K} && {K \otimes K}
\arrow["{K \otimes \sigma \otimes K}", from=1-1, to=1-3]
\arrow["{m \otimes m}", from=1-3, to=1-5]
\end{tikzcd},
\]
where $\sigma$ is the twist in the symmetric monoidal structure. This is lifted to a map of filtrations by
\[\begin{tikzcd}
{(P_{\bullet} \otimes P_{\bullet}) \otimes (P_{\bullet} \otimes P_{\bullet})} && {P_{\bullet} \otimes P_{\bullet} \otimes P_{\bullet} \otimes P_{\bullet}} && {P_{\bullet} \otimes P_{\bullet} }
\arrow["{P_{\bullet} \otimes \sigma \otimes P_{\bullet}}", from=1-1, to=1-3]
\arrow["{\widetilde{m} \otimes \widetilde{m}}", from=1-3, to=1-5],
\end{tikzcd}\]
where $\widetilde{m}$ is the lift of the multiplication map constructed above. This ends the argument.
\end{proof}
\begin{theorem}
\label{theorem:milnor_filtration_makes_kk_into_a_filtered_hopf_algebroid}
The Milnor filtration makes $K_{*}K$ into a filtered Hopf algebroid over $K_{*}$.
\end{theorem}
\begin{proof}
We have already checked that the Milnor filtration is preserved by multiplication on $K_{*}K$ in \cref{lemma:milnor_filtration_is_multiplicative}. Thus, we are left with comultiplication and antipode. The latter is induced by the twist map $K \otimes K$, which can be lifted to a map of filtration by the twist on $P_{\bullet} \otimes P_{\bullet}$, hence it preserves the filtration on homotopy groups.
We move on to comultiplication $\Delta\colon K_{*}K \rightarrow K_{*}K \otimes_{K_{*}} K_{*}K$. This is induced by the map
\[
\begin{tikzcd}
{K \otimes K \simeq K \otimes S^{0} \otimes K} & {} & {K \otimes K \otimes K \simeq (K \otimes K) \otimes_{K} (K \otimes K)}
\arrow["{K \otimes u \otimes K}", from=1-1, to=1-3],
\end{tikzcd}
\]
where $u\colon S^{0} \rightarrow K$ is the unit and we implicitly use the K\"{u}nneth isomorphism
\[
\pi_{*}((K \otimes K) \otimes_{K} (K \otimes K)) \simeq K_{*}K \otimes_{K_{*}} K_{*}K,
\]
which holds in full generality since $K_{*}$ is a field.
As $P_{\bullet}$ is an $E$-free filtration of $K_{*}$, it induces a trivial filtration on $K_{*}$, so that the unit factors through some map $S^{0} \rightarrow P_{0}$. The latter induces a map $\widetilde{u}\colon S^{0} \rightarrow P_{\bullet}$ of filtrations, where we equip the sphere with the trivial filtration, which makes it into the monoidal unit of filtered spectra.
Consider the composite
\[
\begin{tikzcd}
{P_{\bullet} \otimes P_{\bullet} } && {P_{\bullet} \otimes P_{\bullet} \otimes P_{\bullet} } & {P_{\bullet} \otimes K \otimes P_{\bullet} \simeq (P_{\bullet} \otimes K) \otimes_{K} (K \otimes P_{\bullet}) }
\arrow["{P_{\bullet} \otimes \widetilde{u} \otimes P_{\bullet}}", from=1-1, to=1-3]
\arrow[from=1-3, to=1-4]
\end{tikzcd}
\]
which by construction lifts the map inducing comultiplication. As the filtrations of $K_{*}K$ induced by $P_{\bullet} \otimes K$ and $K \otimes P_{\bullet}$ coincide with the Milnor filtration by \cref{lemma:milnor_filtration_coincides_with_the_kunneth_filtration} and
\cref{remark:milnor_filtration_also_coincides_with_other_kunneth_filtration}, we deduce that comultiplication preserves the Milnor filtration as needed.
\end{proof}
We will show that if $p > 2$ and $K_{*}K$ is commutative---for example, if $K$ is homotopy commutative---then the Milnor filtration can be canonically promoted to a grading. We will prove that the existence of this grading implies that the Cartan--Eilenberg spectral sequence associated to the extension $K_{*}E \rightarrow K_{*}K \rightarrow K_{*}^{E}K$ collapses.
Assume that $p > 2$ and $K_{*}K$ is commutative, we will first use the Milnor filtration to describe $K_{*}K$ completely as an algebra. By \cref{lemma:milnor_filtration_coincides_with_the_kunneth_filtration}, the associated graded to the Milnor filtration is given by
\[
\Tor^{*, *}_{E_{*}}(K_{*}, E_{*}K) \simeq \Tor^{*, *}_{E_{*}}(K_{*}, K_{*}) \otimes E_{*}K.
\]
Observe that since $F_{0}^{Mil}(K_{*}K) \simeq E_{*}K$ is in even degree, the external degree one part in the associated graded object, which is
\[
\Tor^{1}_{E_{*}}(K_{*}, K_{*}) \otimes E_{*}K,
\]
can be identified with elements of Milnor filtration at most one in $K_{*}K$ of odd internal degree. Thus, the inclusion
\[
\Tor^{1}_{E_{*}}(K_{*}, K_{*}) \hookrightarrow \Tor^{1}_{E_{*}}(K_{*}, K_{*}) \otimes E_{*}K,
\]
can be canonically lifted to a unique map \[
\Tor^{1}_{E_{*}}(K_{*}, K_{*}) \rightarrow K_{*}K.
\]
which reduces to the canonical inclusion relative to elements of Milnor filtration zero.
Since $\Tor^{*, *}_{E_{*}}(K_{*}, K_{*})$ is an exterior algebra on elements of external degree one, when $p > 2$ this $\Tor$-algebra is a free commutative $K_{*}$-algebra on the vector space $\Tor^{1, *}$. If $K_{*}K$ is commutative, it follows that the above lift uniquely extends to a map of $K_{*}$-algebras
\[
\Tor^{*}_{E_{*}}(K_{*}, K_{*}) \rightarrow K_{*}K.
\]
with the following property.
\begin{proposition}
\label{proposition:kk_at_odd_primes_a_tensor_of_ke_and_tor}
Let $p > 2$ and assume that $K_{*}K$ is commutative. Then, the induced map of $E_{*}K$-algebras
\begin{align}
\label{align:ravenel_iso_of_algebras_involving_kk}
\Tor_{E_{*}}^{*}(K_{*}, K_{*}) \otimes_{K_{*}} E_{*}K \rightarrow K_{*}K
\end{align}
is an isomorphism.
\end{proposition}
\begin{proof}
By construction, the above map is surjective on the associated graded of $K_{*}K$ with respect to the Milnor filtration. Thus, it is surjective by an inductive argument, and as a surjective map of free $E_{*}K$-modules of the same rank it must be an isomorphism.
\end{proof}
As a consequence of \cref{proposition:kk_at_odd_primes_a_tensor_of_ke_and_tor}, $K_{*}K$ inherits an additional external grading from the grading of $\Tor$-groups. This grading can be made explicit, as in the following definition.
\begin{definition}
\label{definition:milnor_grading}
Let $p > 2$ and assume that $K_{*}K$ is commutative. Then, the \emph{Milnor grading} on $K_{*}K$ is the unique algebra grading such that
\begin{enumerate}
\item elements of Milnor filtration zero are of Milnor degree zero and
\item odd internal degree elements of Milnor filtration at most one are of Milnor degree one.
\end{enumerate}
\end{definition}
\begin{lemma}
The Milnor grading is well-defined.
\end{lemma}
\begin{proof}
This is immediate from \cref{proposition:kk_at_odd_primes_a_tensor_of_ke_and_tor}, as the external grading on the algebra tensor product is $\Tor^{*}_{E_{*}}(K_{*}, K_{*}) \otimes_{K_{*}} E_{*}K$ is uniquely determined by which elements are of degrees zero and one.
\end{proof}
\begin{remark}
The increasing filtration associated to the Milnor grading is exactly the Milnor filtration. Thus, this grading can be considered as a refinement of the latter canonical filtration, which exists at all primes and with no assumptions on commutativity.
\end{remark}
\begin{theorem}
\label{theorem:existence_and_independence_of_milnor_degree}
Let $p > 2$ and assume that $K_{*}K$ is commutative. Then, the Milnor grading on $K_{*}K$ is compatible with comultiplication and antipode, making it into a bigraded Hopf algebroid.
\end{theorem}
\begin{proof}
As $K_{*}K$ is generated as an algebra by elements in Milnor degrees zero and one, it is enough to check that these two degrees are preserved. The subspace of elements of Milnor degree most one coincides with elements of filtration at most one, which get preserved by comultiplication and antipode by \cref{theorem:milnor_filtration_makes_kk_into_a_filtered_hopf_algebroid}.
Since the ones in Milnor degree zero are precisely those of even internal degree, and the ones in Milnor degree one are exactly the ones of odd internal degree, we deduce that each degree gets preserved separately by both comultiplication and antipode.
\end{proof}
\begin{remark}
To see what goes wrong if $p = 2$, observe that even if $K_{*}K$ were commutative, odd degree elements do not have to square to zero and so there need not be a multiplicative extension of the canonical map $\Tor^{1}_{E_{*}}(K_{*}, K_{*}) \rightarrow K_{*}K$ to a map on the whole $\Tor$-algebra.
One can try to define the Milnor grading at $p = 2$ directly similarly to \cref{definition:milnor_grading}, by declaring that elements in Milnor degree $s$ are the $s$-fold products of those in Milnor degree one. This is bound to fail; the calculations of W\"{u}rgler in the case of the minimal Morava $K$-theory given in \cref{section:digression_minimal_morava_k_theory} show that the elements $\tau_{i}$ of Milnor degree one square to non-zero elements of Milnor degree zero.
\end{remark}
\begin{theorem}
\label{theorem:collapse_of_ce_sseq_at_odd_p}
Suppose that $p > 2$ and that $K_{*}K$ is commutative. Then, the Cartan--Eilenberg spectral sequence associated to the extension
\[
K_{*}E \rightarrow K_{*}K \rightarrow K_{*}^{E}K
\]
collapses on the second page, inducing a canonical isomorphism
\[
\Ext_{K_{*}K}(K_{*}, K_{*}) \simeq \Ext_{K_{*}E}(K_*,\Ext_{K_{*}^{E}K}(K_{*}, K_{*})).
\]
\end{theorem}
\begin{proof}
Under the given assumptions, $K_{*}K$ is bigraded using its internal grading and the Milnor grading of \cref{definition:milnor_grading}. This grading is compatible with the above extension, with all of $K_{*}E$ of Milnor degree zero.
As $K_{*}^{E}K \simeq \Lambda_{K_{*}}(V)$ as Hopf algebras, where $V$ is the $K_{*}$-vector subspace of primitive elements, we have
\[
\Ext^{*, *, *}_{K_{*}^{E}K}(K_{*}, K_{*}) \simeq \Sym_{K_{*}}(V^{\vee}),
\]
an isomorphism with the symmetric algebra. Here, $V^{\vee}$ is of homological degree one and Milnor degree one, and so the whole above $\Ext$-algebra is concentrated on the plane of elements for which Milnor degree is equal to the homological one.
Since the Cartan--Eilenberg differentials lower the homological degree and preserve the Milnor degree, we deduce that they are all zero. This gives an isomorphism as above up to passing to associated graded to the homological filtration, as the extension problems all get trivialized by the additional grading.
\end{proof}
\part{The $K$-based Adams spectral sequence}
In this part, we study the Adams spectral sequence based on a Morava $K$-theory $K$ at arbitrary height $n$ and prime $p$, constructed as part of a finite height analogue of Miller's square of spectral sequences. In particular, we establish its (perhaps surprisingly) good convergence properties and identify it with the filtration by powers spectral sequence, at least for sufficiently large primes $p$. We then conclude this part with an illustration of the spectral sequence at heights 1 and odd primes, already exhibiting the existence of arbitrary long differentials.
\begin{remark}[Adams spectral sequence based on Milnor modules]
\label{remark:k_based_adams_spectral_sequence_in_terms_of_milnor_modules}
We will not proceed in this way, but one can phrase the $K$-based Adams spectral sequence entirely in terms of absolute Milnor modules, following Miller's observation that the Adams spectral sequence is determined by the class of $K_{*}$-injective maps.
To see this, notice that the restricted Yoneda embedding, followed by a sheafification, gives a symmetric monoidal functor
\[
y\colon \Sp \rightarrow \Mil_{abs}
\]
which can be considered a Milnor modules-valued homology theory. That is, the functor $y$ takes cofibre sequences to exact sequences, preserves arbitrary direct sums, and takes the suspension in spectra to the distinguished auto-equivalence of Milnor modules induced by suspension functor on $\Sp^{fin}$.
By construction, any injective object in Milnor modules can be lifted to a representing object in spectra, and this allows one to construct a ``$y$-based'' Adams spectral sequence based on injectives, where the $E_{2}$-terms will be canonically given by $\Ext$-groups in absolute Milnor modules \cite{patchkoria2021adams}. Through the equivalence of \cref{proposition:absolute_milnor_modules_same_as_comodules}, this spectral sequence will be canonically isomorphic to the $K$-based Adams spectral sequence for any choice of Morava $K$-theory.
\end{remark}
\section{Convergence of the $K$-based Adams spectral sequence}\label{sec:knbasedass}
In this short section, we prove that the $K$-based Adams spectral sequence is conditionally convergent for any $K$-local spectrum, and it converges completely in the dualizable case.
\begin{remark}
As observed in \cref{remark:k_based_adams_spectral_sequence_in_terms_of_milnor_modules}, the $K$-based Adams spectral sequence depends only on the choice of the prime and the height. However, this independence from choices will not be needed in this section, as the arguments we use apply equally well to any Morava $K$-theory.
\end{remark}
By standard arguments, the $K$-based Adams spectral sequence is isomorphic to the totalization spectral sequence obtained by mapping into the \emph{canonical $K$-Adams resolution}
\[
X \rightarrow K \otimes X \rightrightarrows K \otimes K \otimes X \Rrightarrow \ldots
\]
Thus, it follows that the $K$-based Adams spectral sequence converges to the homotopy of the \emph{$K$-nilpotent completion of $X$}:
\[
\Tot(K^{\otimes \bullet +1} \otimes X) := \textstyle\varprojlim_{s \in \Delta} K^{\otimes \bullet +1} \otimes X.
\]
Since the totalization is $K$-local as a limit of $K$-local spectra, there is a canonical comparison map from the $K$-localization of $X$. Our first task is to determine when this map is an equivalence.
\begin{proposition}
\label{prop:knnilpotentcompletion}
For any spectrum $X$, the $K$-nilpotent completion of $X$ is equivalent to the $K$-localization of $X$; that is, the canonical map
\begin{equation}
\label{equation:k_localization_into_k_completion_map_in_thm}
L_{K}X \rightarrow \Tot(K^{\otimes \bullet+1} \otimes X)
\end{equation}
is an equivalence.
\end{proposition}
\begin{proof}
Since both sides are $K$-local, it is enough to show that this map is a $K_{*}$-isomorphism. The latter can be checked after tensoring with any non-$K$-acyclic spectrum as $K_{*}$ is a field, so that we have the K\"{u}nneth isomorphism, and tensoring with a non-zero $K_{*}$-vector space reflects isomorphisms.
After tensoring with a finite spectrum $F(n)$ of type $n$, the above map becomes
\[
L_{K}(X \otimes F(n)) \rightarrow \Tot(K^{\otimes \bullet+1} \otimes X \otimes F(n)),
\]
as both sides of (\ref{equation:k_localization_into_k_completion_map_in_thm}) define exact functors in $X$ and so commute with tensoring with a finite spectrum. Thus, it is enough to check the result holds for $X \otimes F(n)$.
The proof of the smash product theorem in \cite[Chapter 8]{orangebook} shows that $S^0$ is $E$-prenilpotent; that is, $L_ES^0 \in \Thick^{\otimes}(E)$, where $\Thick^{\otimes}(Y)$ denotes the smallest full subcategory of $\spectra$ which contains $Y$, is thick, and is closed under smashing with arbitrary objects in $\spectra$. It follows that
\[
L_{K}F(n) \simeq L_E F(n) \in \Thick^{\otimes}(E\otimes F(n)) \subseteq \Thick^{\otimes}(K).
\]
In other words, $F(n)$ is $K$-prenilpotent. By definition, $X \otimes F(n)$ is thus also $K$-prenilpotent for all $X \in \spectra$. Therefore, Bousfield's theorem \cite[Theorem 6.10]{bousfield_locspectra} applies to provide an equivalence $L_{K}(X \otimes F(n)) \simeq \Tot (K^{\otimes \bullet+1} \otimes X \otimes F(n))$, which is what we wanted.
\end{proof}
If $X, Y$ are spectra, then by applying $[Y, -]$ to the canonical $K$-Adams resolution of $X$ we obtain the Adams spectral sequence as the totalization spectral sequence. Since $K_{*}$ is a field, $K_{*}Y$ is projective over the base ring, so that by standard arguments the resulting spectral sequence has signature
\[
E_2^{s,t}(Y, X) \cong \Ext_{K_*K}^{s,t}(K_*Y,K_*X) \implies [Y, L_{K}X]^{t-s}.
\]
Indeed, if we use the canonical Adams resolution, then the first page of the Adams spectral sequence will be given by the cobar complex, which in the projective case computes the $\Ext$-groups in comodules \cite[Corollary A.1.2.12]{ravenel2003complex}.
\begin{corollary}
\label{corollary:k_local_adams_conditionally_convergent}
For any $K$-local spectrum $X$ and any $Y$, the $K$-based Adams spectral sequence converges conditionally in the sense of Boardman to $[Y, L_{K}X]$.
\end{corollary}
\begin{proof}
The totalization spectral sequence is the spectral sequence induced by the tower of partial totalizations. Conditional convergence to $[-, L_{K}X]$ in the sense of Boardman is equivalent to the natural map from $L_{K}X$ into the limit of this tower being an equivalence, so that the statement follows from \cref{prop:knnilpotentcompletion}.
\end{proof}
Let us write $E_{r}^{*, *}(Y, X)$ for the $r$-th page of the $K$-based Adams spectral sequence converging conditionally to $[Y, L_{K}X]$, so that we have $E_{2}^{s, t}(Y, X) \simeq \Ext_{K_{*}K}^{s, t}(K_{*}Y, K_{*}X)$. The differentials in this spectral sequence are of the form
\[
d_r\colon E_r^{s,t}(Y, X) \to E_r^{s+r,t+r-1}(Y, X).
\]
In particular, for $r>s$ there are inclusions $E_{r+1}^{s,t}(Y, X) \subseteq E_{r}^{s,t}(Y, X)$, and we write
\[
E_{\infty}^{s,t}(Y, X) = \lim_rE_{r}^{s,t}(Y, X).
\]
Let $F^*[Y,L_{K}X]_*$ denote the filtration on $[Y,L_{K}X]_*$ induced by the spectral sequence. Following Bousfield \cite[\S6]{bousfield_locspectra}, we say that the $K$-based Adams spectral sequence for $X$ \emph{converges completely} if the canonical maps
\begin{equation*}
[Y,L_{K}X]_* \twoheadrightarrow \lim_s [Y,L_{K}X]_* / F^s[Y,L_{K}X]_*
\end{equation*}
and
\begin{equation*}
F^s[Y,L_{K}X]_* / F^{s+1}[Y,L_{K}X]_* \hookrightarrow E_{\infty}^{s,s+*}(Y, X)
\end{equation*}
are isomorphisms. By the work of Bousfield, see \cite[Proposition 6.3]{bousfield_locspectra}, this is equivalent to the condition that $\lim_{r>s}^1E_{r}^{s,t}(Y, X)$ vanishes.
\begin{proposition}
\label{proposition:complete_convergence_of_k_local_adams_for_dualizable_spectra}
The $K$-based Adams spectral sequence of signature
\[
\Ext_{K_*K}^{s,t}(K_*Y,K_*X) \implies [Y, L_{K}X]^{t-s}.
\]
converges completely for any spectra $X$ and $Y$ with $K_{*}X$ and $K_{*}Y$ finite-dimensional.
\end{proposition}
\begin{proof}
By \cref{theorem:finiteness_of_ext_kk}, the groups in the $E_{2}$-term are all finite. Therefore, the system
\[
(E_r^{s,t}(Y, X))_{r>s}
\]
is Mittag-Leffler, so that Bousfield's convergence criterion \cite[Proposition 6.3]{bousfield_locspectra} applies.
\end{proof}
\begin{remark}
For a spectrum $X$, the homology $K_{*}X$ being finite-dimensional is equivalent to $X$ being $K$-locally dualizable, see \cite[Theorem 8.6]{hovey_morava_1999}.
\end{remark}
\begin{remark}
In general, the $K$-based Adams spectral sequence may fail to converge completely in the sense of Bousfield, see \cref{ex:noncompleteconvergence}.
\end{remark}
\section{The finite height Miller square}\label{sec:finiteheightmillersquare}
In his seminal paper \cite{miller_adams} on relations between Adams spectral sequences, Miller uses the relationship between the Adams--Novikov and classical Adams spectral sequences to prove the telescope conjecture at height one and odd primes. In this section, we apply his axiomatic framework to the case of Lubin--Tate and Morava $K$-theory spectra.
In Miller's framework, one starts with a map $A \rightarrow B$ of homotopy ring spectra; that is, an arrow in the category $\Alg(\ho(\Sp))$. In this context, it is not hard to show that a map $X \rightarrow Y$ of spectra which is \emph{$A$-monic}, in the sense that $A \otimes X \rightarrow A \otimes Y$ is a split inclusion, is also $B$-monic. Thus, any $B$-injective spectrum is also $A$-injective, see \cite[Lemma 2.1]{miller_adams}.
\begin{definition}[Miller]
\label{definition:ab_primary_spectrum}
A spectrum $X$ is \emph{$(A, B)$-primary} if there exists an $A$-Adams resolution $X \rightarrow I^{\bullet}$ such that for each ${m}$, the $B$-Adams spectral sequence computing $\pi_{*} I^{m}$ converges and collapses on the second page.
\end{definition}
When $X$ is $(A, B)$-primary, Miller constructs a square of spectral sequences
\[\begin{tikzcd}
& {E^{2}_{A, B}(X)} \\
{E^{2}_{A}(X)} && {E^{2}_{B}(X)} \\
& {\pi_{*}X}
\arrow["{A-\textnormal{Adams}}"', Rightarrow, from=2-1, to=3-2]
\arrow["{B-\textnormal{Adams}}", Rightarrow, from=2-3, to=3-2]
\arrow["{\textnormal{May}}"', Rightarrow, from=1-2, to=2-1]
\arrow["{\textnormal{Mahowald}}", Rightarrow, from=1-2, to=2-3]
\end{tikzcd}\]
which we will now explain. The terms $E_{A}^{2}(X), E^{2}_{B}(X)$ form the second pages of the relevant Adams spectral sequences, and the bottom two arrows are the corresponding Adams spectral sequences.
The construction of the other two spectral sequences is more involved, but let us briefly recall the arguments of Miller. Observe that $E^{2}_{A}(X)$ arises as homology of the cochain complex $\pi_{*} I^{\bullet}$, where $I^{\bullet}$ is an $A$-Adams resolution as in \cref{definition:ab_primary_spectrum}. As this cochain complex is given by homotopy groups, it has an additional filtration given by the $B$-Adams filtration. By definition, the term $E^{2}_{A, B}(X)$ is the cohomology of the associated graded of $\pi_{*} I^{\bullet}$ with respect to this filtration, and the top left arrow is the associated cohomology of filtered complex spectral sequence, which Miller calls the \emph{May spectral sequence}.
If $X$ is primary in the sense of \cref{definition:ab_primary_spectrum}, the cohomology $E^{2}_{A, B}(X)$ of the associated graded of $\pi_{*} I^{\bullet}$ is the same as the cohomology of the complex $E^{2}_{B}(I^{\bullet})$ consisting of the $B$-Adams $E_{2}$-pages of the spectra $I^{\bullet}$. As all $A$-split maps are $B$-split, these $E_{2}$-pages are related by long exact sequences and can be collected into an $E_{1}$ of an exact couple using that $I^{\bullet}$ is a resolution of $X$, yielding a spectral sequence computing $E^{2}_{B}(X)$. This is the spectral sequence in the top right corner, which Miller calls the \emph{Mahowald spectral sequence}.
\begin{remark}
Miller proves that the above diagram of spectral sequences is commutative in the sense that, after making appropriate choices, the $d_{2}$-differentials in the May spectral sequence can be identified with $d_{2}$-differentials in the $B$-Adams spectral sequence \cite[Theorem 4.2]{miller_adams}.
In the case of Morava $K$-theory and under the assumption that the prime is large enough compared to the height, we will generalize this result to higher differentials in \cref{proposition:k_adams_in_algebraic_model}.
\end{remark}
Let us give an explicit description of these spectral sequences for Lubin--Tate spectra and Morava $K$-theories. Note that in this case, the $E$-based Adams spectral sequence computes the $E$-local homotopy groups, which are quite different from the $K$-local ones. Thus, to get a commutative square, we should work internally to $K$-local spectra.
Our first goal is to show that the $K$-based Adams spectral sequence has a particularly simple form for $E$-modules with flat homotopy groups. To do so, we will reduce to the pro-free case using the work of Hovey.
Before we proceed, let us recall that the completion functor $M_{*} \mapsto (M_{*})_{\mfrak}^{\vee} := \varprojlim M_{*} / \mfrak^{n} M_{*}$ is neither left or right exact, but it has left derived functors $L_{s}\colon \Mod_{E_{*}} \rightarrow \Mod_{E_{*}}$. For more properties of these functors, see \cite[Appendix B]{hovey_morava_1999} or \cite[Appendix A]{barthelfrankland2015}.
\begin{lemma}
\label{lemma:flat_modules_complete_to_profree_ones}
Let $M_*$ be an $E_{*}$-module such that $\Tor^{s}_{E_{*}}(M_{*}, K_{*}) = 0$ for $s > 0$; for example, $M_*$ can be flat. Then, the completion $(M_{*})^{\vee}_{\mfrak}$ is pro-free.
\end{lemma}
\begin{proof}
This is proven by Hovey in \cite[1.2]{hovey2004some}.
\end{proof}
\begin{lemma}
\label{lemma:k_localization_of_flat_e_module_is_completion}
Let $M$ be an $E$-module such that $M_{*}$ is flat over $E_{*}$. Then, the $K$-localization map $M \rightarrow L_{K} M$ induces an isomorphism $\pi_{*} L_{K} M \simeq (M_{*})^{\vee}_{\mfrak}$ with the completion.
\end{lemma}
\begin{proof}
Hovey constructs for any $E$-module $M$ a strongly convergent spectral sequence of signature
\[
L_{s} M_{t} \rightarrow \pi_{t+s} L_{K} M.
\]
In the present case, the spectral is concentrated on the zero line by \cref{lemma:flat_modules_complete_to_profree_ones}, hence the spectral sequence collapses inducing the needed isomorphism.
\end{proof}
\begin{corollary}
\label{corollary:flat_emodule_is_klocally_profree}
Let $M$ be an $E$-module such that $M_{*}$ is flat. Then, $L_{K}M$ is equivalent as an $E$-module to the $K$-localization of a direct sum of copies of $E$.
\end{corollary}
\begin{proof}
We have $\pi_{*} L_{K} M \simeq (M_{*})^{\vee}_{\mfrak}$ by \cref{lemma:k_localization_of_flat_e_module_is_completion} and $(M_{*})^{\vee}_{\mfrak}$ is pro-free by \cref{lemma:flat_modules_complete_to_profree_ones}. Choosing pro-generators of the latter induces a map $L_{K}(\bigoplus E) \rightarrow M$ from a direct sum of shifts of $E$ which is an equivalence by inspection of homotopy groups.
\end{proof}
\begin{proposition}
\label{proposition:collapse_of_k_based_ass_for_flat_e_modules}
If $M$ is an $E$-module such that $M_{*}$ is $E_{*}$-flat, then the $K$-Adams filtration on $M_{*}$ is the $\mfrak$-adic one and the $K$-based Adams spectral sequence for $M$ converges completely to $\pi_{*} L_{K} M := (M_{*})^{\vee}_{\mfrak}$ and collapses on the second page.
\end{proposition}
\begin{proof}
By flatness we have $\pi_{*} L_{K}M \simeq (M_{*})^{\vee}_{\mfrak}$. Since the $\mfrak$-adic filtration on $M_{*}$ is induced by that of completion, we can replace $M$ by its $K$-localization. In this case, $M$ is $K$-locally equivalent to a direct sum of $E$ by \cref{corollary:flat_emodule_is_klocally_profree}. As the Adams spectral sequence of a direct sum is a direct sum of spectral sequences, we can assume that $M = E$.
We have a base-change isomorphism
\[
\Ext_{K_{*}K}(K_{*}E, K_{*}) \simeq \Ext_{K_{*}^{E}K}(K_{*}, K_{*})
\]
and by \cref{proposition:hl_cohomology_of_kek}
a further isomorphism
\[
\Ext_{K_{*}^{E}K}(K_{*}, K_{*}) \simeq K_{*} \otimes_{\kappa} \Sym_{\kappa}^{*} (\mfrak / \mfrak^{2}).
\]
Since $E_{0}$ is regular, $\Sym_{\kappa}^{*} (\mfrak / \mfrak^{2})$ can be identified with the associated graded to the $\mfrak$-adic filtration on $E_{0}$.
We deduce from the above that the $E_{2}$-term of the $K$-based Adams for $E$ is abstractly isomorphic to the associated graded of the $\mfrak$-adic filtration on $E_{*}$. Moreover, since the $E_{2}$-term is concentrated in even Adams degree, we deduce that the Adams spectral sequence collapses.
We are left with showing that the $K$-based Adams filtration on $E_{*}$ coincides with the $\mfrak$-adic one. Observe that since $E \rightarrow K$ is a $K_{*}$-monomorphism by \cref{proposition:map_from_ke_to_kk_injective}, the elements of positive Adams filtration are precisely given by $\mfrak E_{*} = \mathrm{ker}(E_{*} \rightarrow K_{*})$.
Using compatibility of the Adams filtration with products, we deduce that any element of $\mfrak^{n} E_{*}$ is of Adams filtration at least $n$. Thus, one filtration is contained in the other. As their associated gradeds have the same dimension over $K_{*}$ by what was said above, we deduce that the two filtrations must coincide, as needed.
\end{proof}
\begin{proposition}
\label{proposition:primarity_from_flat_homology}
Let $X$ be a spectrum such that $E_{*}^{\vee}(X) := \pi_{*} L_{K} (E \otimes X)$ is pro-free; for example, such that $E_{*}X$ is flat. Then, $X$ is $(E, K)$-primary in the $K$-local category.
\end{proposition}
\begin{proof}
The implication that $E_{*}X$ flat implies $E_{*}^{\vee}X$ pro-free is a combination of \cref{lemma:flat_modules_complete_to_profree_ones} and \cref{proposition:collapse_of_k_based_ass_for_flat_e_modules}.
Since $E_{*}E$ is flat over $E_{*}$, this in particular implies that $E_{*}^{\vee} E$ is pro-free, which can be also proven directly using evenness, see \cite[THeorem 8.6]{hovey_morava_1999}. It follows (as in \cite[Corollary 1.24]{bh_enass}, for example) that
\[
\pi_{*} L_{K}(E^{\bullet} \otimes X) \simeq (E_{*}^{\vee}E)^{\otimes^{\vee}_{E_{*}} \bullet-1} \otimes^{\vee}_{E_{*}} E_{*}^{\vee} X,
\]
where on the right hand side we have the completed tensor products, which are again pro-free. This together with \cref{proposition:collapse_of_k_based_ass_for_flat_e_modules} implies that the standard $K$-local $E$-based Adams resolution
\[\begin{tikzcd}
L_{K}X & {L_{K}(E \otimes X)} & {L_{K}(E \otimes \overline{E} \otimes X)} & \ldots \\
& {L_{K}(\overline{E} \otimes M)} & {L_{K}(\overline{E} \otimes \overline{E} \otimes X) }
\arrow[from=1-1, to=1-2]
\arrow[from=1-2, to=2-2]
\arrow[from=1-2, to=1-3]
\arrow[from=1-3, to=2-3]
\arrow[from=1-3, to=1-4]
\arrow[dashed, from=2-2, to=1-1]
\arrow[dashed, from=2-3, to=1-2]
\end{tikzcd}\]
satisfies the primarity condition.
\end{proof}
If $E_{*}^{\vee}X$ is pro-free, then the $E_{1}$-page of the $K$-local $E$-based Adams spectral sequence is given by the cobar complex
\[
\pi_{*} L_{K}(E^{\bullet} \otimes X) \simeq (E_{*}^{\vee}E)^{\otimes^{\vee}_{E_{*}} \bullet-1} \otimes_{E_{*}} E_{*}^{\vee} X.
\]
It follows from \cref{proposition:collapse_of_k_based_ass_for_flat_e_modules} that the $K$-Adams filtration on these homotopy groups coincides with the filtration by powers of the maximal ideal $\mfrak$; this is the ``filtration by powers'' appearing in the title. This has the following consequence.
\begin{corollary}
\label{corollary:millers_may_sseq_is_filtration_by_powers}
If $E_{*}^{\vee}X$ is pro-free, then Miller's May spectral sequence associated to $(E, K)$ coincides with the spectral sequence of signature
\[
\Ext_{E_{*}^{\vee}E}^{s,t}(E_{*}, \bigoplus \mfrak^{n}/\mfrak^{n+1} E_{*}^{\vee}X) \implies \Ext_{E_{*}^{\vee}E}^{t-s}(E_{*}, E^{\vee}_{*}X)
\]
obtained by filtering $E_{*}^{\vee}X$ by powers of the maximal ideal; that is, by $\mfrak^{n} E_{*}^{\vee}X$.
\end{corollary}
\begin{proof}
Since Miller's May spectral sequence is induced by the $K$-Adams filtration on the homotopy groups of the $K$-local $E$-Adams resolution, this is a consequence of \cref{proposition:collapse_of_k_based_ass_for_flat_e_modules}, as under the above assumption all of the spectra in the resolution are pro-free.
\end{proof}
\begin{remark}
The identification of the $E_{2}$-page of the $K$-local $E$-based Adams spectral sequence is somewhat subtle. However, in the case where $E_{*}^{\vee}X$ is pro-free, it follows from \cite{bh_enass} that it acquires a structure of a suitably complete comodule over $E_{*}^{\vee}E$ and that the $E_{2}$-page is given by the $\Ext$-groups in comodules as above.
\end{remark}
We deduce that the $K$-local Miller square associated to $(E, K)$ is of the form
\begin{equation}
\label{equation:miller_square}
\begin{tikzcd}
& {\Ext_{E_{*}^{\vee}E}(E_{*}, \bigoplus \mfrak^{n}/\mfrak^{n+1} E_{*}^{\vee}X)} \\
{\Ext_{E_{*}^{\vee}E}(E_{*}, E^{\vee}_{*}X)} && {\Ext_{K_{*}K}(K_{*}, K_{*}X)} \\
& {\pi_{*}L_{K}X}
\arrow["{E-\textnormal{Adams}}"', Rightarrow, from=2-1, to=3-2]
\arrow["{K-\textnormal{Adams}}", Rightarrow, from=2-3, to=3-2]
\arrow["{\textnormal{May}}"', Rightarrow, from=1-2, to=2-1]
\arrow["{\textnormal{Mahowald}}", Rightarrow, from=1-2, to=2-3],
\end{tikzcd}
\end{equation}
whenever $E^{\vee}_{*}X$ is profree. This leaves the question of identifying the Mahowald spectral sequence, which is not too difficult.
\begin{proposition}
\label{proposition:mahowald_sseq_is_ce_sseq}
If $E_{*}^{\vee}X$ is pro-free, then Miller's Mahowald spectral sequence based on $(E, K)$ can be identified with the Cartan--Eilenberg spectral sequence associated to the extension of Hopf algebroids
\[
K_{*}E \rightarrow K_{*}K \rightarrow K_{*}^{E}K.
\]
\end{proposition}
\begin{proof}
Taking the standard $K$-local $E$-Adams resolution
\[
X \rightarrow L_{K}(E \otimes X) \rightarrow L_{K}(E \otimes \overline{E} \otimes X) \rightarrow \ldots
\]
and applying $K$-homology gives the long exact sequence of comodules of the form
\[
K_{*}X \rightarrow K_{*}E \otimes K_{*}X \rightarrow K_{*}E \otimes_{K_{*}} \overline{K_{*}E} \otimes_{K_{*}} K_{*}X \rightarrow \ldots.
\]
Applying $\Ext_{K_{*}K}(K_{*}, -)$ to the above leads to an exact couple which yields the Mahowald spectral sequence.
Note that from the description of the $E_{1}$-term we see that this is just the $K_{*}E$-based Adams spectral sequence in the derived category of $K_{*}K$-comodules, sometimes called the Margolis--Palmieri spectral sequence. This is known to be isomorphic to the Cartan--Eilenberg spectral sequence associated to the extension, see \cite[Theorem 4.2]{belmont2020cartan}.
\end{proof}
\begin{remark}
\label{remark:collapses_of_mahowald_sseq}
We have shown in \cref{theorem:collapse_of_ce_sseq_at_odd_p} that if $p > 2$, then the Cartan--Eilenberg spectral sequence computing $\Ext_{K_{*}K}(K_{*}, K_{*})$ collapses on the second page. It follows from the isomorphism of \cref{proposition:mahowald_sseq_is_ce_sseq} that the same is true for Miller's $(E,K)$-Mahowald spectral sequence associated to $X = S^{0}$.
\end{remark}
\begin{remark}
Note that applying \cref{proposition:mahowald_sseq_is_ce_sseq} to the starting terms of the respective spectral sequences implies in particular that
\[
\Ext_{E_{*}^{\vee}E}(E_{*}, \bigoplus \mfrak^{n}/\mfrak^{n+1} E_{*}) \simeq \Ext_{E_{*}K}(K_{*}, \Ext_{K_{*}^{E}K}(K_{*}, K_{*})).
\]
This can be proven directly, without resorting to Miller's work, using Hopkins and Lurie's calculation of the cohomology of $K_{*}^{E}K$, which we stated as \cref{proposition:hl_cohomology_of_kek}. The key is that since $E_{0}$ is regular, we have $\Sym_{\kappa}(\mfrak / \mfrak^{2}) \simeq \bigoplus \mfrak^{n} / \mfrak^{n+1}$.
\end{remark}
\begin{remark}
If we again pick $E = E_{n}$, the Lubin--Tate spectrum associated to the Honda formal group law of height $n$ over $\mathbb{F}_{p^{n}}$, then the relevant $\Ext$-groups appearing in the Miller square can be described in terms of the cohomology of the Morava stabilizer group, see \cite[Theorem 4.3]{bh_enass}. Keeping this choice of $E$ in mind, we can rewrite the Miller square as
\[\begin{tikzcd}
& {\rmH^{*}(\mathbb{G}_{n}, \bigoplus \mfrak^{n}/\mfrak^{n+1} E_{*}^{\vee}X)} \\
{\rmH^{*}(\mathbb{G}_{n},E^{\vee}_{*}X)} && {\Ext_{K_{*}K}(K_{*}, K_{*}X)} \\
& {\pi_{*}L_{K}X}
\arrow["{E-\textnormal{Adams}}"', Rightarrow, from=2-1, to=3-2]
\arrow["{K-\textnormal{Adams}}", Rightarrow, from=2-3, to=3-2]
\arrow["{\textnormal{May}}"', Rightarrow, from=1-2, to=2-1]
\arrow["{\textnormal{Mahowald}}", Rightarrow, from=1-2, to=2-3].
\end{tikzcd}\]
\end{remark}
\begin{remark}
In the important case of $(MU, H \mathbb{F}_{p})$, the Miller square can also be described explicitly, see for example \cite{gheorghe2018special}. In this context, the Miller square foreshadows the later development of the $\mathrm{C}\tau$-philosophy of Gheorghe, Isaksen, Wang and Xu which applied motivic methods to breakthrough results in stable homotopy groups of spheres \cite{isaksen2020stable}.
\end{remark}
\section{The $K$-based Adams spectral sequence at large primes}
\label{sec:algebraicmodelkbasedass}
It is well-known that if $p$ is large compared to the height, then the $E$-based Adams--Novikov spectral sequence for the $K$-local sphere collapses. As the same is true for the Mahowald spectral sequence, as we have shown in \cref{remark:collapses_of_mahowald_sseq}, we deduce that under these assumptions, two of the sides of the Miller square of (\ref{equation:miller_square}) represent collapsing spectral sequences.
It is then only natural to expect that the other two sides can be identified up to regrading; in this section, we show that this is indeed the case. More precisely, we will show that if $2p-2 > n^{2}+n+1$, then the $K$-based Adams spectral sequence for the sphere is isomorphic to the filtration by powers spectral sequence
\[
\rmH^{*}(\mathbb{G}_{n}, \bigoplus \mfrak^{n}/\mfrak^{n+1}) \implies \rmH^{*}(\mathbb{G}_{n}, E_{*}),
\]
which we can identify with Miller's May spectral sequence by \cref{corollary:millers_may_sseq_is_filtration_by_powers}.
\begin{remark}
Note that a result of Miller \cite[Theorem 4.2]{miller_adams} relates the $d_{2}$-s in the May and $K$-based Adams spectral sequence, at all primes and heights. Our first approach to this problem was to try to push Miller's methods further into higher differentials, hoping that the collapses of the other two spectral sequences in the square would simplify the situation. This approach was unsuccessful, and below we use quite different methods.
\end{remark}
\begin{remark}
While we will show in this section that at large primes the $K$-based Adams spectral sequence admits an algebraic description, it \emph{never} collapses for $n>0$, unlike the Adams--Novikov spectral sequence. Indeed, the element $v_{n} = u^{p^{n}-1} \in \Ext^{0, 2p^{n}-2}_{K_{*}K}(K_{*}, K_{*})$ is never a permanent cycle, as the $K$-local sphere is of type $0$ and so cannot admit a $v_{n}$-periodic map.
\end{remark}
To prove our comparison result, we will use an algebraicity result for homotopy categories of $E$-local spectra at large primes due to the second author, strengthened in recent work with Patchkoria, which compares it the following algebraic construction.
\begin{notation}
Let us write $\eD(E_{*}E)$ for the derived $\infty$-category of the abelian category of comodules. This is a symmetric monoidal stable $\infty$-category, with monoidal unit given by $E_{*}$ itself, considered as a chain complex concentrated in degree zero. The unit defines for any $X \in \eD(E_{*}E)$ its \emph{homotopy groups}
\[
\Ext_{\eD(E_{*}E)}^{s, t}(X) := [\Sigma^{s} E_{*}[t], X]_{\eD(E_{*}E)}.
\]
Notice that the homotopy groups are bigraded, since the abelian category of $E_{*}E$-comodules has an internal grading. This endows the derived $\infty$-category with \emph{two} gradings,
\begin{enumerate}
\item the \emph{homological} one given by the suspension and corresponding above to $s$ and
\item the \emph{internal one} induced from the grading shift functor on $\Comod_{E_*E}$ and corresponding above to $t$.
\end{enumerate}
In the special case of $X$ being a comodule, considered as an object of the heart, these homotopy groups encode the $E_{2}$-term of the $E$-based Adams spectral sequence in the $\infty$-category of spectra.
\end{notation}
\begin{definition}
\label{definition:periodicity_algebra}
The \emph{periodicity algebra} $P(E_{*})$ is the $\mathbf{E}_{\infty}$-algebra in $\dcat(E_{*}E)$ determined by the commutative algebra
\[
E_{*}[\tau^{\pm 1}], \textnormal{where } | \tau | = (1, -1)
\]
in chain complexes of comodules, equipped with the zero differential.
\end{definition}
\begin{notation}
For brevity, let us write
\[
\Mod_{P(E_{*})} := \Mod_{P(E_{*})}(\dcat(E_{*}E)).
\]
\end{notation}
The following result shows that modules over the periodicity algebra give an algebraic model for the homotopy category of $E$-local spectra.
\begin{theorem}
\label{theorem:chromatic_algebraicity}
If $2p-2 > n^{2}+n$, there exists an equivalence
\[
\phi\colon h \spectra_{E} \simeq h \Mod_{P(E_{*})}\dcat(E_{*}E)
\]
of homotopy categories of $E$-local spectra and $P(E_{*})$-modules which is moreover compatible with homology functors in the sense that the following diagram commutes
\[
\begin{tikzcd}
{h \spectra_{E}} & {} & {h \Mod_{P(E_{*})}\dcat(E_{*}E)} \\
& \Comod_{E_*E}.
\arrow["{E_{*}(-)}"', from=1-1, to=2-2]
\arrow["{\rmH_{0}}", from=1-3, to=2-2]
\arrow["\phi", from=1-1, to=1-3]
\end{tikzcd}
\]
\end{theorem}
\begin{proof}
This is proven under the assumption $2p-2 > 2(n^{2}+n)$ in the second author's thesis \cite{pstrkagowski2018chromatic}. The bound was improved to the one given above in joint work with Patchkoria \cite{patchkoria2021adams}.
\end{proof}
\begin{remark}
Note that as an $\mathbf{E}_{1}$-algebra in $\dcat(E_{*}E)$, the connective cover of the periodicity algebra $P(E_{*})$ can be identified with the free $\mathbf{E}_{1}$-algebra on the object $\Sigma E_{*}[-1]$, and so is uniquely determined by its homology groups.
The category of modules over $P(E_{*})$ (considered only as a stable $\infty$-category) can be identified with the subcategory of modules over the connective cover spanned by those modules on which $\tau$ acts invertibly. Thus, to define the target of the equivalence of \cref{theorem:chromatic_algebraicity}, it is not strictly necessary to work with algebras in chain complexes as in \cref{definition:periodicity_algebra}, but we do so out of convenience.
\end{remark}
\begin{remark}
\label{remark:the_many_models_for_e_local_category}
For any $M, N \in \Mod_{P(E_{*})}$ we have an Adams spectral sequence of signature
\[
\Ext^{s, t}(\rmH_{0}(M), \rmH_{0}(N)) \implies [M, N]_{t-s},
\]
where on the right hand side we have homotopy classes of maps of $P(E_{*})$-modules.
This observation, due to Franke \cite{franke_exotic}, is the starting point for the comparison of $\Mod_{P(E_{*})}$ to the $E$-local category. The general form of Franke's conjecture implies that all stable $\infty$-categories admitting a suitably convergent Adams-type spectral sequence of this signature must have equivalent homotopy categories \cite{patchkoria2021adams}.
\end{remark}
One of the advantage of $P(E_{*})$-modules over $E$-local spectra is the existence of the free module functor
\[
P(E_{*}) \otimes -\colon \eD(E_{*}E) \rightarrow \Mod_{P(E_{*})}
\]
(also called the \emph{periodicization}) which is cocontinuous and symmetric monoidal. The homotopy groups of free modules can be understood explicitly, as the following shows.
\begin{remark}
\label{remark:periodicization_simple_homotopy_groups}
Since the periodicity algebra is a sum of shifts of a unit as an object of the derived $\infty$-category, the homotopy groups of the free module are simple to understand. More precisely, for any $M \in \eD(E_{*}E)$ we have
\[
[P(E_{*}), P(E_{*}) \otimes_{E_{*}} M]_{P(E_{*})}^{t} \simeq [E_{*}, P(E_{*}) \otimes_{E_{*}} M)]_{\dcat(E_{*}E)}^{t} \simeq \bigoplus_{s} \Ext^{s, t-s}_{E_{*}E}(E_{*}, M).
\]
As a consequence, whenever a $P(E_{*})$-module can be written as a periodicization of some $M$ as above, its homotopy groups acquire an additional grading. Note that any given module can be written as a periodicization in many different ways and this additional grading depends on that choice.
\end{remark}
Our plan is to use \cref{theorem:chromatic_algebraicity} to compare the $K$-based Adams spectral sequence in $E$-local spectra to the one in its algebraic model, and subsequently give an explicit description of the latter in terms of filtration by powers.
\begin{lemma}
\label{lemma:exotic_equivalence_preserves_sphere_and_k}
Let $X$ be either the $E$-local sphere or a Morava $K$-theory; that is, $X = S^{0}_{E}$ or $X = K$. Then, there exists a canonical isomorphism
\[
\phi(X) \simeq P(E_{*}) \otimes_{E_{*}} E_{*}X,
\]
in $h \Mod_{P(E_{*})}$, where $\phi$ is the algebraicity equivalence of \cref{theorem:chromatic_algebraicity}.
\end{lemma}
\begin{proof}
It will be convenient to work here with the Johnson-Wilson homology, the $p$-local Landweber exact homology theory with
\[
\pi_{*}E(n) \simeq \mathbb{Z}_{(p)}[v_{1}, \ldots, v_{n-1}, v_{n}^{\pm 1}]
\]
Both this ring and $E(n)_{*}E(n)$ are concentrated in degrees divisible by $2p-2$, which is a key ingredient in the construction of the equivalence $\phi$ \cite[\S 2]{pstrkagowski2018chromatic}.
Choosing coordinates for the Quillen formal group over $E_{*}$ such that $v_{i} = 0$ for $i > n$ we obtain a classifying map $E(n)_{*} \rightarrow E_{*}$ which is faithfully flat and hence
\[
E_{*}X \simeq E_{*} \otimes_{E(n)_{*}} E(n)_{*}X
\]
The induced functor $\Comod_{E(n)_{*}E(n)} \rightarrow \Comod_{E_{*}E}$ is an equivalence of categories by a result of Hovey and Strickland \cite{hovey2005comodules}.
Now suppose that $X$ is a spectrum with $E(n)_{*}X$ concentrated in degrees divisible by $2p-2$. Since $\phi$ is compatible with taking homology, we have
\[
\rmH_{0}(\phi(X)) \simeq E_{*}X \simeq E_{*} \otimes_{E(n)_{*}} E(n)_{*}X
\]
as comodules. By \cref{remark:the_many_models_for_e_local_category}, we have an Adams spectral sequence
\[
\Ext^{s, t}(E_{*}X, E_{*}X) \implies [\phi(X), P(E_{*}) \otimes_{E_{*}} E_{*}X]^{t-s}
\]
Using the equivalence of categories of comodules, we can rewrite the $E_{2}$-page as
\[
\Ext^{s, t}(E(n)_{*}X, E(n)_{*}X) \simeq \Ext^{s, t}(E_{*}X, E_{*}X)
\]
This vanishes above $n^{2}+n$ by \cite[Remark 2.5]{pstrkagowski2018chromatic} and is concentrated in degrees $t$ divisible by $2p-2$. Since $2p-2 > n^{2}+n$ by assumption, we deduce that the spectral sequence collapses on the second page, so that the identity of $E_{*}X$ descends to an equivalence $\phi(X) \simeq P(E_{*}) \otimes_{E_{*}} E_{*}X$.
This establishes the result for $X = S^{0}_{E}$, which has $E(n)_{*}X$ homology concentrated in degrees divisible by $2p-2$. If $K$ is a Morava $K$-theory, then as a consequence of the nilpotence theorem it is a module over the minimal Morava $K$-theory of \S\ref{section:digression_minimal_morava_k_theory} and so equivalent as a spectrum to a direct sum of $K(n)$. As $\phi$ preserves direct sums and
\[
E(n)_{*}K(n) \simeq K(n)_{*} \otimes_{E(n)_{*}} E(n)_{*}E(n)
\]
is concentrated in degrees divisible by $2p-2$, the result follows also for $X = K$.
\end{proof}
\begin{remark}
The detour into Johnson-Wilson homology \cref{lemma:exotic_equivalence_preserves_sphere_and_k} can be avoided, by directly defining a splitting of order $(2p-2)$ on the category $\Comod_{E_{*}E}$ by using the equivalence with quasi-coherent sheaves on the moduli of formal groups of height at most $n$. We decided against it, as the degree divisibility argument involving $E(n)$ is classical and well-known.
\end{remark}
Let us write $K_{alg} := P(E_{*}) \otimes _{E_{*}} E_{*}K$ for the algebraic model for the Morava $K$-theory spectrum. Note that its homotopy groups satisfy
\[
\pi_{s} K_{alg} := [P(E_{*}), K_{alg}]^{P(E_{*})}_{s} \simeq K_{s},
\]
so that $K_{alg}$ is a field object in $P(E_{*})$. Consequently, for a map $M \rightarrow N$ of $P(E_{*})$-modules, the following two conditions are equivalent:
\begin{enumerate}
\item $K_{alg} \otimes_{P(E_{*})} M \rightarrow K_{alg} \otimes_{P(E_{*})} N$ is a split monomorphism of $P(E_{*})$-modules or
\item $[N, K_{alg}]_{*} \rightarrow [M, K_{alg}]_{*}$ is an epimorphism of graded abelian groups.
\end{enumerate}
This forms a class of monomorphisms and so determines an Adams spectral sequence \cite[\S3.1]{patchkoria2021adams}.
\begin{definition}
We call the Adams spectral sequence in $P(E_{*})$-modules determined by the above class of monomorphisms the \emph{$K_{alg}$-based Adams spectral sequence}.
\end{definition}
We are now ready to verify the correspondence between the topological and algebraic Adams spectral sequences; we learned the following elegant argument from Robert Burklund.
\begin{proposition}
\label{proposition:k_adams_in_algebraic_model}
If $2p-2 > n^{2}+n+1$, then the $K$-based Adams spectral sequence in $\spectra_{E}$ for $S^{0}_{E}$ is isomorphic to the $K_{alg}$-based Adams spectral sequence for $P(E_{*})$ in $P(E_{*})$-modules.
\end{proposition}
\begin{proof}
We claim that the equivalence $\phi$ of \cref{theorem:chromatic_algebraicity} induces an isomorphism of exact couples leading to these spectral sequences. Both of these are obtained by mapping into suitable towers, so we have to verify that $\phi$ takes a $K$-Adams resolution of $S^{0}_{E}$ to a $K_{alg}$-Adams resolution of $\phi(S^{0}_{E)}) \simeq P(E_{*})$.
If $2p-2 > n^{2}+n+1$, the equivalence $\phi$ is compatible with triangulated structures \cite[B.8]{pstrkagowski2018chromatic}, so it remains to show that it identifies $K$-injectives with $K_{alg}$-injectives, as the corresponding class of monomorphism is then uniquely determined. This follows immediately from \cref{lemma:exotic_equivalence_preserves_sphere_and_k}, as on both sides each injective is a direct sum of shifts of, respectively, $K$ and $K_{alg}$.
\end{proof}
In our case, in \cref{proposition:k_adams_in_algebraic_model} we considered a $K_{alg} = P(E_{*}) \otimes E_{*}K$-based Adams spectral sequence for the monoidal unit $P(E_{*})$. Both objects arise through periodicization, giving us additional information, which we will now make explicit.
\begin{proposition}
\label{proposition:k_based_adams_has_an_additional_grading_for_which_it_is_ek_adams}
If $2p-2 > n^{2}+n+1$, then the $K$-based Adams spectral sequence for $S^{0}_{E}$ can be given an additional grading so that it becomes isomorphic to the $E_{*}K$-based Adams spectral sequence in $\eD(E_{*}E)$.
\end{proposition}
\begin{proof}
By \cref{proposition:k_adams_in_algebraic_model}, the $K$-Adams spectral sequence in $\spectra_{E}$ is isomorphic to the $K_{alg}$-based Adams spectral sequence in $P(E_{*})$-modules. The latter is isomorphic to the spectral sequence induced by the Amitsur resolution
\[
P(E_{*}) \rightarrow K_{alg} \rightrightarrows K_{alg} \otimes_{P(E_{*})} K_{alg} \Rrightarrow \ldots
\]
which is image of the $E_{*}K$-Adams resolution of $E_{*}$ in $\eD(E_{*}E)$ under the periodicization functor. The statement follows from the isomorphism of homotopy groups of \cref{remark:periodicization_simple_homotopy_groups} applied to the exact couple giving rise to the spectral sequence.
\end{proof}
The above result reduces the study of the $K$-based Adams spectral sequence at large primes to the study of the algebrac $E_{*}K$-based in the derived $\infty$-category of comodules. The latter can be described explicitly in terms of the cohomology of the Morava stabilizer group, with no assumptions on the prime, which is our next step.
We will need to recall some results on d\'{e}calage, following Deligne and Levine \cite{levine2015adams}.
\begin{construction}
\label{construction:decalage}
If $X^{\bullet}$ is a cosimplicial spectrum, we have an associated conditionally convergent spectral sequence
\[
\prescript{\Tot}{}{E}^{2}_{s, t} := \rmH^{s} \pi_{t} X^{\bullet} \implies \pi_{t-s} \Tot(X^{\bullet}).
\]
The \textit{d\'{e}calage} of this cosimplicial object is the tower
\[
\Dec_{n} := \Tot(\tau_{\leq n} X^{\bullet})
\]
obtained by totalizing the Postnikov truncations of $X$. Since limits commute with limits we have
\[
\varprojlim \Dec_{n} \simeq \varprojlim \Tot(\tau_{\leq n} X^{\bullet}) \simeq \Tot(\varprojlim \tau_{\leq n} X^{\bullet}) \simeq \Tot(X^{\bullet}).
\]
The above equivalence induces a spectral sequence associated to the tower on the left, which is a conditionally convergent and of signature
\[
\prescript{\Dec}{}{E}^{1}_{s, t} := H^{s} \pi_{t}X^{\bullet} \implies \pi_{t-s} \Tot(X^{\bullet}).
\]
Levine proves that the evident isomorphisms between the $E^{2}$-term of the $\Tot$-spectral sequence and the $E_{1}$-term of the spectral sequence associated to the d\'{e}calage extends to an isomorphism of spectral sequences \cite{levine2015adams}.
\end{construction}
In our case, we will be working with cosimplicial objects in $\eD(E_{*}E)$, rather than in spectra. Thus, the role of the Postnikov towers will be played by the truncations in the standard $t$-structure on the derived category.
\begin{definition}
Let $X^{\bullet}\colon \Delta \rightarrow \eD(E_{*}E)$ be a cosimplicial object in the derived $\infty$-category of comodules. The \emph{homological d\'{e}calage} of $X^{\bullet}$ is the tower
\[
h\Dec_{n} := \Tot(\tau_{\leq n} X^{\bullet}),
\]
where $\tau_{\leq n}$ denotes the truncation in the standard $t$-structure on the derived $\infty$-category.
\end{definition}
\begin{warning}
Note that if $X^{\bullet}$ is a cosimplicial object in $\eD(E_{*}E)$, then the homological d\'{e}calage tower induces a spectral sequence obtained by applying $\Ext(E_{*}, -)$. Since Postnikov towers converge in $\dcat(E_{*}E)$, this will be conditionally convergent to $\Ext(E_{*}, \Tot(X^{\bullet}))$.
The layers of the tower are given, up to a shift, by $\Tot(H_{k}(X^{\bullet}))$ so that this spectral sequence has signature
\[
\Ext(E_{*}, \Tot(H_{k}(X^{\bullet})) \implies \Ext(E_{*}, \Tot(X^{\bullet})).
\]
This spectral sequence is usually not the same as the one obtained by applying $\Ext(E_{*}, -)$ to $X^{\bullet}$ itself, because the homological $t$-structure does not interact in an easy way with $E_{*}$-homotopy groups, unlike in the case of the standard $t$-structure on the $\infty$-category of spectra.
\end{warning}
Despite the above warning, for a certain restricted class of cosimplicial objets, we \emph{do} have an isomorphism between the spectral sequence of a cosimplicial object and its homological d\'{e}calage.
\begin{lemma}
\label{lemma:homological_decalage_and_spectral_decalage}
Let $X^{\bullet}\colon \Delta \rightarrow \eD(E_{*}E)$ be a cosimplicial object and assume that each $X^{m}$ is a direct sum of shifts of objects in the heart, each of which is a cofree $E_{*}E$-comodule. Then
\[
F(E_{*}[t], h\Dec_{n} X^{\bullet}) \simeq \Dec_{n} F(E_{*}, X^{\bullet}),
\]
where $F(E_{*}[t], -)$ is the internal mapping spectrum in $\eD(E_{*}E)$. That is, for such cosimplicial objects, mapping out of $E_{*}[t]$ takes homological d\'{e}calage to the spectral one.
\end{lemma}
\begin{proof}
Since $F(E_{*}, -)$ preserves totalizations, it is enough to check that $F(E_{*}, -)$ takes homological truncations to Postnikov truncations. Since $X^{\bullet}$ is levelwise a direct sum of shifts of cofree comodules, it is enough to check this in the latter case, where it is clear since $\Ext^{s}(E_{*}[t], C) = 0$ for $s > 0$ whenever $C$ is cofree.
\end{proof}
\begin{corollary}
\label{corollary:adams_spectral_sequence_for_cofree_things_same_as_homological_decalage}
The $E_{*}K$-based Adams spectral sequence in $\eD(E_{*}E)$ has a second page isomorphic to the first page of the spectral sequence associated to the homological d\'{e}calage $h\Dec_{n} (E_{*}K)^{\otimes \bullet}$ of the Amitsur resolution.
\end{corollary}
\begin{proof}
Both the Adams spectral sequence and the spectral sequence of the tower are obtained by applying $F(E_{*}[t], -)$ (for all $t$ at once) and using the relevant spectral sequence in spectra. Thus, the result then follows from Levine's work and \cref{lemma:homological_decalage_and_spectral_decalage}, since
\[
E_{*}K^{\otimes n} \simeq (E_{*}E)^{\otimes n} \otimes_{E_{*}} K_{*} ^{\otimes_{E_*} n}
\]
and
\[
K_{*}^{\otimes_{E_*} n} \simeq \Tor_{E_{*}}(K_{*}, K_{*}) \otimes_{K_{*}} \ldots \otimes_{K_{*}} \Tor_{E_{*}}(K_{*}, K_{*}),
\]
where the right hand side has $(n-1)$ factors and the unadorned tensor products are the derived tensor products over $E_{*}$. This is a direct sum of shifts of $K_{*}$, as needed.
\end{proof}
The idea is now to relate the homological d\'{e}calage of $E_{*}K$ to that of $K_{*}$.
\begin{lemma}
\label{lemma:k_mapping_into_ek_a_decalage_isomorphism}
The morphism $K_{*} \rightarrow E_{*}E \otimes K_{*} \simeq E_{*}K$ of algebras in the derived category of $E_*E$-comodules induces an isomorphism of spectral sequence associated to the homological d\'{e}calage of Amitsur resolutions.
\end{lemma}
\begin{proof}
Since $E_{*}E$ is flat, for any $X \in \eD(E_{*}E)$ we have $\rmH_{n}(E_{*}E \otimes X) \simeq E_{*}E \otimes \rmH_{n}(X)$. Thus, the map between the layers of the homological d\'{e}calage towers is given by applying totalization to the map
\[
\rmH_{n}(K_{*}^{\otimes \bullet}) \rightarrow E_{*}E^{\otimes \bullet} \otimes \rmH_{n}(K_{*}^{\otimes \bullet}).
\]
The map is obtained by tensoring the map
\[
E_{*}^{\otimes \bullet} \rightarrow E_{*}E^{\otimes \bullet},
\]
which is a quasi-isomorphism of levelwise flat cosimplicial comodules with $\rmH_{n}(K_{*}^{\otimes \bullet})$, so it will be a quasi-isomorphism again. It follows that it is an equivalence after taking totalizations.
\end{proof}
\begin{proposition}
\label{proposition:adams_sseq_assocaited_to_e_k}
From the second page on, the Adams spectral sequence associated to $E_{*}K$ is isomorphic to the spectral sequence obtained from the tower
\[
\ldots \rightarrow E_{*} / \mfrak^{2} \rightarrow E_{*} / \mfrak,
\]
that is, it is the filtration by powers spectral sequence.
\end{proposition}
\begin{proof}
We know from \cref{corollary:adams_spectral_sequence_for_cofree_things_same_as_homological_decalage} that this Adams spectral sequence has second page isomorphic to the spectral sequence of the homological d\'{e}calage tower, which is in turn equivalent to the homological d\'{e}calage tower associated to the $K_{*}$-Adams resolution by \cref{lemma:k_mapping_into_ek_a_decalage_isomorphism}. Thus, it is enough to identify the latter.
We're interested in the cosimplicial object
\[
K_{*} \rightrightarrows K_{*} \otimes _{E_{*}} K_{*} \Rrightarrow \ldots,
\]
all tensor products being implicitly derived. Everything here is $2$-periodic and concentrated in even degrees in the internal grading, so that we can instead focus on the cosimplicial object
\[
K_{0} \rightrightarrows K_{0} \otimes_{E_0} K_{0} \Rrightarrow \ldots,
\]
working in the derived category of $E_{0}$-modules. As an object of the latter, $K_{0} \otimes_{E_0} K_{0}$ is a direct sum of its homology groups, and we have
\[
H_{*}(K_{0} \otimes_{E_{0}} K_{0}) \simeq \Tor^{E_{0}}_{*}(K_{0}, K_{0})
\]
and more generally
\begin{equation}
\label{equation:homology_of_cosimplicial_kzero_resolution_and_tor}
H_{*}(K_{0}^{\otimes_{E_0} n}) \simeq \Tor_{*}^{E_{0}}(K_{0}, K_{0})^{\otimes_{K_{0}} n-1},
\end{equation}
where on the left we have the derived tensor product of $E_{0}$-modules and on the right the ordinary tensor product of graded $K_{0}$-modules.
Using that $E_{0}$ is a regular ring with residue field $K_{0}$, the
canonical isomorphism
\[
\Tor^{1}_{E_{0}}(K_{0}, K_{0}) \simeq (\mfrak / \mfrak^{2})^{\vee}
\]
with the tangent space extends to a grading-preserving isomorphism
\[
\Tor_{*}^{E_{0}}(K_{0}, K_{0}) \simeq \Lambda_{K_{0}}((\mfrak / \mfrak^{2})^{\vee})
\]
with the exterior algebra. This is flat over $K_{0}$, which is a field, so by standard arguments the cosimplicial object $K_{0}^{\otimes \bullet}$ on the right of (\ref{equation:homology_of_cosimplicial_kzero_resolution_and_tor}) encodes the cobar complex computing
\[
\textnormal{Cotor}_{\Lambda_{K_{0}}((\mfrak / \mfrak^{2})^{\vee})}(K_{0}, K_{0}) \simeq \Sym_{K_{0}}(\mfrak/\mfrak^{2}) \simeq \bigoplus_{n \geq 0} \mfrak^{n}/\mfrak^{n+1}.
\]
Note that this is bigraded using the internal grading of the exterior algebra and the $\textnormal{Cotor}$-grading, with the summand $\mfrak^{n} / \mfrak^{n+1}$ being concentrated in degrees $(n, n)$.
Using the isomorphism in \cref{equation:homology_of_cosimplicial_kzero_resolution_and_tor} and passing to cohomology, we deduce that
\[
H^{s}(H_{n}(K_{0}^{\otimes_{E_{0}} \bullet})) \simeq \begin{cases}
\mfrak^{n}/\mfrak^{n+1} & \mbox{when } $s=n$,\\
0 & \mbox{otherwise,}
\end{cases}
\]
where $H^{s}$ is the cohomology of the given cosimplicial abelian group. Thus, the totalization spectral sequence for $H_{n}(K_{0}^{\otimes_{E_{0}} \bullet})$, which we consider as a cosimplicial object of the heart of the derived $\infty$-category of $E_{0}$-modules, collapses and we deduce that
\[
\Tot(H_{n}(K_{0}^{\otimes \bullet})) \simeq \Sigma^{-n} \mfrak^{n} / \mfrak^{n+1},
\]
It follows that the graded pieces of the homological d\'{e}calage filtration of $E_{*}$ are given by
\[
\mathrm{fib}(\Dec_{n} \rightarrow \Dec_{n-1}) \simeq \Tot(\Sigma^{n} H_{n}(K_{*}^{\otimes \bullet})) \simeq \mfrak^{n} / \mfrak^{n+1} \otimes_{E_{0}} E_{*}
\]
and so by induction the tower itself must be of the claimed form.
\end{proof}
\begin{theorem}
\label{theorem:k_based_adams_at_large_primes_iso_to_filtration_by_powers}
If $2p-2 > n^{2}+n+1$, the $K$-based Adams spectral sequence can be given an additional grading so that it becomes isomorphic to the spectral sequence induced by applying $\Ext_{E_{*}E}(E_{*}, -)$ to the filtration of $E_{*}$ given by powers $\mfrak^{n} E_{*}$ of the maximal ideal.
\end{theorem}
\begin{proof}
This is a combination of \cref{proposition:k_based_adams_has_an_additional_grading_for_which_it_is_ek_adams} and \cref{proposition:adams_sseq_assocaited_to_e_k}.
\end{proof}
\begin{remark}
\label{remark:k_based_adams_at_large_primes_same_as_filtration_by_powers_in_group_cohomology}
If we pick $E = E_{n}$ to be the Morava $E$-theory of the Honda formal group law over $\mathbb{F}_{p^{n}}$, then since $\mfrak^{n} / \mfrak^{n+1} E_{*}$ is $\mfrak$-torsion, we have
\[
\Ext_{E_{*}E}^*(E_{*}, \mfrak^{n} / \mfrak^{n+1} E_{*}) \simeq \rmH_{\cts}^{*}(\mathbb{G}_{n}, \mfrak^{n} / \mfrak^{n+1} E_{*}).
\]
Thus, the spectral sequence of \cref{theorem:k_based_adams_at_large_primes_iso_to_filtration_by_powers} is isomorphic to the conditionally convergent spectral sequence
\[
\rmH_{\cts}^{*}(\mathbb{G}_{n}, \mfrak^{n}/\mfrak^{n+1} E_{*}) \implies \rmH_{\cts}^{*}(\mathbb{G}_{n}, E_{*})
\]
obtained by the same filtration of $E_{*}$, but in continuous $\mathbb{G}_{n}$-modules.
\end{remark}
\section{Height one $K$-based Adams spectral sequence at an odd prime}
\label{sec:somecomputations}
Let us describe how using our methods one obtains an explicit description of the $K$-based Adams spectral sequence at $n=1$ and $p$ odd. Since in this case
\[
2p-2 \geq 4 > 3 = n^{2}+n+1,
\]
\cref{theorem:k_based_adams_at_large_primes_iso_to_filtration_by_powers} and \cref{remark:k_based_adams_at_large_primes_same_as_filtration_by_powers_in_group_cohomology} imply that the relevant spectral sequence is isomorphic to the one computing cohomology $\rmH_{\cts}^{*}(\mathbb{G}_{n}, E_{*})$ using the filtration by powers of the maximal ideal, where $E_{0}$ is the Lubin--Tate ring of the Honda formal group law.
\begin{notation}
We have $E_{*} \simeq \mathbb{Z}_{p}[u^{\pm 1}]$, $\mathfrak{m} = (p)$ and $\mathbb{G}_{1} \simeq \mathbb{Z}_{p}^{\times}$ with the action determined by $\lambda_{*}(u) = \lambda u$ for $\lambda \in \mathbb{Z}_{p}$. We will denote the associated graded of $E_{*}$ by
\[
A_{*} \simeq \mathbb{F}_{p}[b][u^{\pm1}],
\]
with $|b| = (0, 1)$ the equivalence class of $p \in E_{0}$, and $|u| = (2, 0)$, where the first degree is internal and the second one is coming from the filtration.
\end{notation}
The $p$-adic group $\mathbb{Z}_{p}^{\times}$ is topologically cyclic at odd primes, generated by $\psi = \sigma (1 + p)$ for $\sigma$ a primitive $(p-1)$-th root of unity. It follows that for a profinite $\mathbb{Z}_{p}^{\times}$-module $M$, its cohomology can be computed using the $2$-term complex
\begin{equation}
\label{equation:2_term_complex_computing_cohomology}
\partial\colon M \rightarrow M,
\end{equation}
where $\partial = id_{M} - \psi_{*}$.
The complex of (\ref{equation:2_term_complex_computing_cohomology}) is a quasi-isomorphic quotient of the standard group cohomology cochain complex
\begin{equation}
\label{equation:standard_cochain_complex_computing_zptimes_cohomology}
M \rightarrow \Map_{\cts}(\mathbb{Z}_{p}^{\times}, M) \rightarrow \Map_{\cts}(\mathbb{Z}^{\times}_{p} \times \mathbb{Z}_{p}^{\times}, M) \rightarrow \ldots,
\end{equation}
the quotient map being given by the identity in degree zero and evaluation at the generator $\psi \in \mathbb{Z}_{p}^{\times}$ in degree one.
From this explicit description we see that if we equip $M$ with a filtration, the above quotient map is a quasi-isomorphism of filtered complexes, where we consider both (\ref{equation:2_term_complex_computing_cohomology}) and (\ref{equation:standard_cochain_complex_computing_zptimes_cohomology}) with the filtration induced from that of $M$. In particular, both filtrations induced isomorphic spectral sequences.
\begin{remark}
The 2-term complex of (\ref{equation:2_term_complex_computing_cohomology}) can be used to directly compute the cohomology groups $\rmH_{\cts}^{*}(\mathbb{Z}_{p}^{\times}, \mathbb{Z}_{p}[u^{\pm 1}])$, recovering the classical image of $J$ pattern \cite[Lecture 35]{lurie2010chromatic}. Our goal is not to compute the above groups, which are well-known, but rather analyze the structure of the spectral sequence associated to filtration by powers of $p$.
\end{remark}
We first determine the cohomology groups of the associated graded $A_*$. We have $\psi_{*} b = b$, because the latter can be represented by $p \in E_{0}$ which is necessarily acted on trivially, and since $\psi_{*} u \equiv \sigma u \pmod p$, we have $\psi_{*} u = [\sigma] u$ in the associated graded, where $[\sigma] \in \mathbb{F}_{p}$ is the image of $\sigma$. Thus, in the associated graded we have that $\partial(b) = 0$ and $\partial(u^{k})$ is a unit multiple of $u^{k}$ for $(p-1) \nmid k$ and zero otherwise. We deduce the following.
\begin{proposition}
\label{proposition:e2_term_of_kass_at_n1_and_p_odd}
The $E_{2}$-term of the $K$-based Adams for $S^{0}$ at $n=1$ and $p > 2$ is given by
\[
\Ext_{K_{*}K}^*(K_{*}, K_{*}) \cong \rmH_{\cts}^{*}(\mathbb{G}_{1}, A_{*}) \cong \rmH_{\cts}^{*}(\mathbb{Z}_{p}^{\times}, \mathbb{F}_{p}[b][u^{\pm 1}]) \cong \mathbb{F}_{p}[b][v_{1}^{\pm 1}] \otimes \Lambda(\zeta),
\]
where $v_{1} = u^{p-1}$ is of degree $(2p-2, 0, 0)$ and $\zeta$ is the class of $b$ of degree $(0, 1, 1)$, with the last degree the cohomological one.
\end{proposition}
By \cref{remark:k_based_adams_at_large_primes_same_as_filtration_by_powers_in_group_cohomology}, if the $K$-based Adams is isomorphic to the spectral sequence computing $\rmH_{\cts}^{*}(\mathbb{Z}_{p}^{\times}, \mathbb{Z}_{p}[u^{\pm 1}])$ induced by the filtration of (\ref{equation:2_term_complex_computing_cohomology}) by powers of $p$, with $E_{2}$-term as above and with differentials $d_{r}\colon E_{r} \rightarrow E_{r}$ of degree $(0, r, 1)$. We will now analyze this spectral sequence.
We know that $\partial(p) = 0$ in $E_{0}$, so that $b$ is a permanent cycle. The same is true for $\zeta$, for degree reasons, and we deduce that all of the differentials are both $b$ and $\zeta$-linear, and so the structure of the spectral sequence is completely determined by what it does on $v_{1}^{k}$. One computes with no difficulty that if we write $k = p^{n}m$, where $p \nmid m$, then
\[
\partial(u^{(p-1)k}) \equiv \text{unit} \cdot p^{n+1} u^{(p-1)k} \pmod {p^{n+2}}
\]
from which we deduce that
\begin{equation}
\label{equation:differentials_in_k_based_ass_for_sphere_at_height_one_odd_prime}
d_{i}(v_{1}^{k}) =
\begin{cases}
0 & \text{for } i < n+1 \\
\zeta b^{n+1} v_{1}^{k} & \text{for } i = {n+1}.
\end{cases}
\end{equation}
Note that the first computation is also implied by the Leibniz rule for $p$-th powers. The extension problems are resolved using multiplication by $b$, recovering the classical answer that
\[
\rmH_{\cts}^{s, t}(\mathbb{Z}_{p}^{\times}, \mathbb{Z}_{p}[u^{\pm 1}]) = \begin{cases} \mathbb{Z}_{p} & \mbox{if } (s, t) = (0, 0), \\
\mathbb{Z}_{p} \{\zeta\} & \mbox{if } (s, t) = (1, 0), \\
\mathbb{Z}/p^{n+1} \{\zeta u^{(p-1)p^{n}m} \} & \mbox{if } (s, t) = (1, (2p-2)p^{n}m) \mbox{, where } p \nmid m, \\
0 & \mbox{otherwise,} \end{cases}
\]
where $\{-\}$ denotes a generator of the given group. Thus, the spectral sequence is as in the following familiar ``image of J'' pattern
\[
\begin{tikzcd}[column sep=tiny]
&& {} \\
&& {} & \bullet & \ldots && \bullet & \ldots && \bullet & \ldots && \bullet & \ldots \\
&&& \bullet & \bullet && \bullet & \bullet && \bullet & \bullet && \bullet & \bullet \\
&&& {\bullet_{\zeta b}} & {\bullet_{b^{2}}} && \bullet & \bullet && \bullet & \bullet && \bullet & \bullet \\
{} &&& {\bullet_{\zeta}} & {\bullet_{b}} && {\bullet_{v_{1} \zeta}} & {\bullet_{v_{1}b}} && \bullet & \bullet && \bullet & \bullet \\
{} &&& {} & {\bullet_{1}} &&& {\bullet_{v_{1}}} &&& {\bullet_{v_{1}^{2}}} &&& {\bullet_{v_{1}^{p}}} \\
&& {} & {-1} & 0 &&& {2p-2} & \ldots &&& \ldots &&& {} \\
&& {} &&&&&&&& {} && {} & {} & {} & {}
\arrow[color={rgb,255:red,214;green,92;blue,92}, from=6-8, to=4-7]
\arrow[color={rgb,255:red,214;green,92;blue,92}, from=5-8, to=3-7]
\arrow[color={rgb,255:red,214;green,92;blue,92}, from=6-11, to=4-10]
\arrow[color={rgb,255:red,214;green,92;blue,92}, from=5-11, to=3-10]
\arrow[color={rgb,255:red,214;green,153;blue,92}, from=6-14, to=3-13]
\arrow["{t-s}", shift left=5, from=7-3, to=7-15]
\arrow[color={rgb,255:red,214;green,92;blue,92}, from=4-8, to=2-7]
\arrow[color={rgb,255:red,214;green,92;blue,92}, from=4-11, to=2-10]
\arrow[color={rgb,255:red,214;green,153;blue,92}, from=5-14, to=2-13]
\arrow["s", shorten >=12pt, from=7-3, to=1-3]
\end{tikzcd}
\]
with even longer differentials supported on $v_{1}^{p^{k}}$ for $k \geq 1$.
It is interesting to observe that this shows that already at height one, the $K$-based Adams spectral sequence has non-zero differentials of arbitrary length.
\begin{example}[Non-completely convergent $K$-Adams]
\label{ex:noncompleteconvergence}
Using the above calculation, we can give an example of a $K$-local spectrum for which its $K$-based Adams spectral sequence is not completely convergent. Note that we know that it is always conditionally convergent by \cref{prop:knnilpotentcompletion}.
To this end, we will construct a spectrum $X$ for which $\lim_r^1E_r^{0,0}(X)$ does not vanish at $n =1$ and $p > 2$. Set $\lambda_k = 2(p-1)p^k$ and consider the spectrum
\[
X = L_{K(1)}\bigoplus_{k\ge 0}S^{-\lambda_k}.
\]
In bidegree $(s,t) = (0,0)$, the $E_2$-page is given by
\[
E_2^{0,0}(X) \cong \bigoplus_{k\ge 0} \mathbb{F}_p\{v_1^{p^k}\},
\]
where the generators correspond to the maps $v_1^{p^k}\colon K(n)_* \to K(n)_*S^{\lambda_k}$. The formulas for the differentials of \cref{equation:differentials_in_k_based_ass_for_sphere_at_height_one_odd_prime} show that (up to reindexing) the filtration on $E_2^{0,0}(X)$ is given by
\[
\ldots \subset \bigoplus_{k \ge 2}\mathbb{F}_p\{v_1^{p^k}\} \subset \bigoplus_{k \ge 1}\mathbb{F}_p\{v_1^{p^k}\} \subset \bigoplus_{k \ge 0}\mathbb{F}_p\{v_1^{p^k}\} = E_2^{0,0}(X).
\]
Therefore, we obtain an exact sequence
\[
0 \to \lim_rE_r^{s,t}(X) \to E_2^{0,0}(X) = \bigoplus_{k\ge 0} \mathbb{F}_p\{v_1^{p^k}\} \to \prod_{k \ge 0}\mathbb{F}_p\{v_1^{p^k}\} \to \lim_r^1E_r^{s,t}(X) \to 0.
\]
In particular, we see that $\lim_rE_r^{s,t}(X) = 0$, while $\lim_r^1E_r^{s,t}(X) \neq 0$.
\end{example}
\part{Homology of inverse limits and the algebraic chromatic splitting conjecture}
In the third part of the work, we largely move away from Morava $K$-theory, and we focus on homology of $K$-local spectra. From the perspective of $E$-local category, $K$-localization is a form of completion, and so we first focus on homology of inverse limits, construction a spectral sequence computing these in the generality of an arbitrary adapted homology theory.
In the case of $K$-localization, the $E_{2}$-page of this spectral sequence is given by derived functors of limits in $E_{*}E$-comodules. We describe the latter in terms of cohomology of the Morava stabilizer group, and use it to compute the zeroth one at all heights and primes. At height one, we compute these derived functors completely.
\section{The homology of inverse limits}
\label{section:homology_of_inverse_limits}
We begin with a general review of the context for the construction of the modified Adams spectral sequence, following the approach of Devinatz and Hopkins \cite{dev_morava}. This approach will then be employed to construct a spectral sequence that computes the homology of inverse limits in a presentable stable $\infty$-category.
In particular, this gives rise to a spectral sequence for computing the $E$-homology of the inverse limit of a tower of spectra $(X_{\alpha})$ from the derived functors of inverse limits of the tower of $E_*E$-comodules $(E_*X_{\alpha})$, for suitable ring spectra $E$. The question of when this spectral sequence converges is subtle, and will be studied in detail in the case of Morava $E$-theory. The material in this section is based on unpublished work of Mike Hopkins and Hal Sadofsky.
\subsection{Adapted homology theories}\label{ssec:adaptedhomology}
In this short section we will recall basic facts about adapted homology theories. Informally, these are exactly those homology theories which admit an Adams spectral sequence based on injectives. Everything here is classical, although our presentation will be most close to \cite[\S 2]{patchkoria2021adams}.
A locally graded $\infty$-category is an $\infty$-category equipped with a distinguished autoequivalence $[1]_{\euscr{C}}\colon \euscr{C} \rightarrow \euscr{C}$. A locally graded functor $f\colon \euscr{C} \rightarrow \dcat$ is functor equipped with a natural isomorphism $f \circ [1]_{\euscr{C}} \simeq [1]_{\dcat} \circ f$.
\begin{definition}
Let $\euscr{C}$ be a stable $\infty$-category, considered as a locally graded $\infty$-category using the suspension functor, and $\euscr{A}$ an arbitrary locally graded abelian category. We say a locally graded functor $H\colon \euscr{C} \rightarrow \euscr{A}$ is a \emph{homology theory} if for any cofibre sequence
\[
X \rightarrow Y \rightarrow Z
\]
in $\euscr{C}$, the induced diagram
\[
H(X) \rightarrow H(Y) \rightarrow H(Z)
\]
in $\euscr{A}$ is exact in the middle.
\end{definition}
\begin{remark}
The requirement that $H$ is locally graded amounts to specifying an isomorphism
\[
H(\Sigma X) \simeq H(X)[1]
\]
natural in $X \in \euscr{C}$.
\end{remark}
In cases of interest to us, $H$ will be \emph{Grothendieck} in the sense that
\begin{enumerate}
\item $\euscr{C}$ is presentable,
\item $\euscr{A}$ is Grothendieck abelian and
\item $H$ preserves arbitrary direct sums.
\end{enumerate}
In particular, this means that $\euscr{A}$ has enough injectives. Any injective object $I$ of $\euscr{A}$ gives rise to a cohomological functor
\[
\Hom_{\euscr{A}}(H(-),I)\colon \euscr{C}^{\op} \to \euscr{A}.
\]
Since $\euscr{C}$ is presentable and stable, Brown's representability theorem holds~\cite[Proposition 2.15]{patchkoria2021adams}, so that there exists some $D(I) \in \Ho\euscr{C}$ representing $\Hom_{\euscr{A}}(H(-),I)$. It follows that, for any object $X \in \euscr{C}$, there is a natural equivalence
\[
\Hom_{\euscr{A}}(H(X),I) \simeq \pi_{0}\Hom_{\euscr{C}}(X,D(I)).
\]
Note that the object $D(I)$ is $H$-local, in the sense that for any object $Y \in \euscr{C}$ with $H(Y) = 0$ we have $\Hom(Y,D(I)) =0$. Moreover, the identity map on $D(I)$ corresponds under this isomorphism to a natural counit map $H(D(I)) \to I$.
\begin{definition}\label{def:adapted}
We say $H\colon \euscr{C} \to \euscr{A}$ is \emph{adapted} if the counit map $H(D(I)) \to I$ is an equivalence for any injective $I \in \euscr{A}$.
\end{definition}
\begin{example}
Rational homology, viewed as a functor from spectra to graded rational vector spaces, is an adapted homology theory. In this case, the lift $D(V)$ of a (graded) $\mathbb{Q}$-vector space $V$ is the (generalized) Eilenberg--MacLane spectrum $HV$.
\end{example}
\begin{example}
\label{example:mod_p_homology_adapted_when_valued_in_comodules}
In contrast to the previous example, mod $p$ homology is not adapted for any $p$; indeed, the lift of $\mathbb{F}_p$ is $H\mathbb{F}_p$, but its mod $p$-cohomology is the mod $p$ Steenrod algebra $\cA_p$ which is larger than $\mathbb{F}_p$.
However, $H_*(-,\mathbb{F}_p)$ has more structure: it takes values in the category $\Comod_{\cA_p}$ of graded $\cA_p$-comodules. It turns out that the homology theory
\[
H_*(-;\mathbb{F}_p)\colon \Sp \to \Comod_{\cA_p}
\]
is adapted.
\end{example}
\begin{remark}
The phenomena visible in \cref{example:mod_p_homology_adapted_when_valued_in_comodules} is typical in the sense that asking for $H$ to be adapted is to ask for $\euscr{A}$ to encode all available homological information. This generalizes considerably, one can show that any homology theory factors uniquely through an adapted one followed by an exact comonadic functor of abelian categories \cite[\S 3.3]{patchkoria2021adams}
\end{remark}
In fact, the previous example generalizes considerably. Let $R$ be a homotopy commutative ring spectrum which satisfies the Adams condition, that is, it can be written as a filtered colimit of finite spectra $X_{\alpha}$ with the properties that
\begin{enumerate}
\item $R_*X_{\alpha}$ is a finitely generated projective $R_*$-module and
\item the K\"{u}nneth map $E^{*}X_{\alpha} \rightarrow \Hom_{E_{*}}(E_{*}X_{\alpha}, E_{*})$ is an isomorphism.
\end{enumerate}
Note that the second condition follows from the first whenever $R$ can be made $\mathbf{E}_{1}$.
For example, $H \mathbb{F}_{p}$ or any Landweber exact homology theory satisfy the Adams condition \cite[Theorems 1.4.7-9]{hovey_htptheory}, but $H\mathbb{Z}$ does not~\cite[Page 17]{goersshopkins_moduli}. This condition guarantees that the associated Hopf algebroid $(R_*,R_*R)$ is Adams and thus has a well-behaved category of comodules $\Comod_{R_*R}$; see \cite[Section 1.4]{hovey_htptheory} for details.
One way in which the Adams-type condition implies that the category $\Comod_{R_{*}R}$ is well-behaved also in the topological sense is the following result of Devinatz.
\begin{lemma}[{\cite[Theorem 1.5 and Section 2]{dev_morava}}]\label{lem:dlift}
Let $R$ be a topologically flat commutative ring spectrum. For any injective $R_*R$-comodule $I$, the identity map on $D(I)$ induces an isomorphism $R_*D(I) \cong I$. In other words, the functor
\[
R_0\colon \spectra \to \Comod_{R_*R}
\]
is an adapted homology theory.
\qed
\end{lemma}
\subsection{Construction of the spectral sequence}
\label{subsection:construction_of_the_spectral_sequence}
The starting point of our construction is the existence of modified Adams towers for any homology theory $H\colon \euscr{C} \to \euscr{A}$. We collect its properties in the next result:
\begin{lemma}\label{lem:modadamstower}
Consider an adapted homology theory $H\colon \euscr{C} \to \euscr{A}$. Let $X \in \euscr{C}$ and suppose
\[
\xymatrix{0 \ar[r] & H(X)=C_0 \ar[r]^-{\eta_0} & I_0 \ar[r]^-{\tau_0} & I_1 \ar[r]^-{\tau_1} & \ldots \\
& & C_0 \ar[u]^-{\eta_0} & C_1 \ar[u]^-{\eta_1} & \ldots}
\]
is an injective resolution of $H(X)$ in $\euscr{A}$, where $C_i = \ker(\tau_i)$ for all $i$. There exists a tower of objects in $\euscr{C}$ over $X$ of the form
\[
\xymatrix{X=X_0 \ar[d]_{f_0} & X_1 \ar[l]_-{g_0} \ar[d]_{f_1} & X_2 \ar[l]_-{g_1} \ar[d]_{f_2} & \ldots \ar[l]_-{g_2} \\
\Sigma^{0}D(I_0) \ar@{-->}[ru]_{\delta_0} & \Sigma^{-1}D(I_1) \ar@{-->}[ru]_{\delta_1} & \Sigma^{-2}D(I_2) \ar@{-->}[ru]_{\delta_2} & \ldots,}
\]
where the dotted maps are the boundary maps shifting degree by $1$, and such that the following properties are satisfied for all $i$:
\begin{enumerate}
\item $X_{i+1} \simeq \fib(f_i)$;
\item $H(X_i) \cong C_i[{-i}]$;
\item $H(f_i) = \eta_i[{-i}]$;
\item $H(g_i)=0$;
\item $H(\delta_i)$ is the surjection $I_i[{-i}] \to C_{i+1}[{-i}]$.
\end{enumerate}
If $X$ is $H$-local, then $X_i$ and $D(I_i)$ are $H$-local for all $i$. Moreover, this tower is weakly functorial; that is, a map $X \to Y$ can be extended to a (non-canonical) map of towers $(X_i) \to (Y_i)$.
\end{lemma}
\begin{proof}
This lemma is an axiomatization of the results proven in \cite[Section 1]{dev_morava}. The tower so constructed consists of $H$-local spectra as noticed in \cref{ssec:adaptedhomology}, while naturality follows from the naturality of injective resolutions.
\end{proof}
\begin{notation}
We refer to the tower $(X_i)$ constructed in \cref{lem:modadamstower} as the modified Adams tower of $X$ with respect to the adapted homology theory $H\colon \euscr{C} \to \euscr{A}$.
\end{notation}
Since the limit $\varprojlim I_{\alpha}$ of an injective tower of injectives in $\euscr{A}$ is injective, the construction of the previous subsection yields an object $D(\varprojlim I_{\alpha})$ of $\euscr{C}$. In order to proceed, we need an auxiliary characterization of injective towers of objects in a Grothendieck abelian category $\euscr{A}$.
\begin{lemma}\label{lem:charinjtower}
Let $\euscr{A}$ be a Grothendieck abelian category. An object $I = (I_{\alpha}) \in \euscr{A}^{\mathbb{N}}$ is injective if and only if $I_{\alpha}$ is injective for all $\alpha \in J$ and all structure maps in $I$ are split epimorphisms.
\end{lemma}
\begin{proof}
This is proven in \cite[Proposition 1.1]{jannsen_contetcohom}. We sketch the argument of the implication we need. Let $j\colon \mathbb{N}^{\delta} \to \mathbb{N}$ be the inclusion of the discrete set of natural numbers into the poset of natural numbers. This induces an adjunction $(j^*,j_*)$ between diagram categories. If $I$ is injective, then the natural monomorphism $I \to j_*j^*I$ is split. By construction, the structure maps of $j_*j^*I$ are projections, so $I$ also has split surjective structure maps, and the claim follows.
\end{proof}
\begin{proposition}\label{prop:dlim}
If $I=(I_{\alpha}) \in \Inj(\cA^{\mathbb{N}})$ is an injective tower in $\euscr{A}$, then there is a preferred equivalence
\[
\xymatrix{D(\varprojlim_{\euscr{A}} I_{\alpha}) \ar[r]^-{\sim} & \varprojlim_{\euscr{C}} D(I_{\alpha}).}
\]
of objects in $\euscr{C}$, well-defined up to homotopy.
\end{proposition}
\begin{proof}
For brevity, let us write $I_{\infty} := \varprojlim_{\euscr{A}} I_{\alpha}$; note that this is an injective object of $\euscr{A}$. The canonical structure maps $\varprojlim_{\euscr{A})} I_{\alpha} \to I_{\alpha}$ give arrows
\[
D(I_{\infty}) \rightarrow D(I_{\alpha}),
\]
well-defined and compatible up to homotopy. By the Milnor exact sequence, these can be lifted to a homotopy class of maps
\[
\xymatrix{\phi\colon D(I_{\infty}) \ar[r] & \varprojlim D(I_\alpha),}
\]
in $\euscr{C}$. We claim this is an equivalence, which we will check by verifying that both sides represent the same functor in the homotopy category.
Indeed, for any $X \in \euscr{C}$, there are isomorphisms
\begin{align*}
\textstyle
[X,D(I_{\infty})]_* & \cong \Hom_{\euscr{A}}^*(H(X), I_{\infty}) \\
& \cong \varprojlim \ \Hom_{\euscr{A}}^*(H(X), I_\alpha) \\
& \cong \varprojlim \ [X,D(I_\alpha)]_*.
\end{align*}
The Milnor sequence associated to the limit $\varprojlim D(I_{n})$ takes the form
\begin{equation}\label{eq:milnorseq}
\xymatrix{0 \ar[r] & \varprojlim^1[X,D(I_{n})]_{*+1} \ar[r] & [X,\varprojlim D(I_{\alpha})]_* \ar[r] & \varprojlim[X,D(I_{\alpha})]_{*} \ar[r] & 0.}
\end{equation}
It follows from \cref{lem:charinjtower} that the structure maps in the tower
\[
([X,D(I_{\alpha})]_{*+1}) \cong (\Hom_{\euscr{A}}^{*+1}(H(X),I_{\alpha}))
\]
are split epimorphisms, so the $\varprojlim^1$-term in \eqref{eq:milnorseq} vanishes. Thus, the Yoneda lemma implies that $\phi$ is an equivalence, as claimed.
\end{proof}
The spectral sequence we construct will involve derived functors of the limit, and so we first establish a consistent notation for these.
\begin{notation}
If $\euscr{A}$ is a Grothendieck abelian category and $F\colon J \rightarrow \euscr{A}$ is a diagram, then we write $\textstyle \varprojlim_{\euscr{A}} F$ for the limit taken in $\euscr{A}$ itself. Formation of limits defines a functor
\[
\textstyle\varprojlim_{\euscr{A}}\colon \Fun(J, \euscr{A}) \rightarrow \euscr{A}
\]
and we denote its right derived functors by
\[
\textstyle\varprojlim^{s}_{\euscr{A}} F.
\]
Alternatively, $\euscr{A}$ has an associated derived $\infty$-category, and we write
\[
\textstyle\varprojlim_{\dcat(\euscr{A})} F
\]
for the limit of the composite of $F$ with the inclusion with the heart. These two different notions of limits, one in $\euscr{A}$ and the other in the derived $\infty$-category, are related by the formula
\[
\textstyle\varprojlim^{s}_{\euscr{A}} F \simeq H_{-s}(\textstyle\varprojlim_{\dcat(\euscr{A})} F),
\]
where on the right hand side we have the homology of an object of the derived $\infty$-category. Thus, $\textstyle\varprojlim_{\dcat(\euscr{A})} F$ can be thought of as the total derived functor of the limit.
\end{notation}
\begin{theorem}
\label{thm:limss}
Let $H\colon \euscr{C} \to \euscr{A}$ be an adapted homology theory. If $(X^{\alpha}) \in \euscr{C}^{\mathbb{N}}$ is a tower in $\euscr{C}$, then there exists a natural spectral sequence in $\euscr{A}$ of the form
\begin{equation}\label{eq:generalss}
\textstyle E_2^{s,t} \cong \varprojlim_{\euscr{A}}^s H(X^{\alpha})[-t]_{\euscr{A}} \implies H(\varprojlim X^{\alpha})[s-t]_{\euscr{A}}
\end{equation}
with differentials $d_r^{s,t}\colon E_r^{s,t} \to E_r^{s+r,t+r-1}$.
\end{theorem}
\begin{proof}
If $H$ is adapted, then so is the induced homology theory
\[
H\colon \euscr{C}^{\mathbb{N}} \rightarrow \euscr{A}^{\mathbb{N}}
\]
between the $\infty$-categories of towers \cite[Example 8.24]{patchkoria2021adams}. Thus, we have Adams resolutions of towers and we let $(X_i^{\alpha})_{i \geq 0}$
be a modified Adams tower of $(X^{\alpha})$ as in \Cref{lem:modadamstower}.
Taking limits in the Adams resolution of $X^{\alpha}$ yields a tower in $\euscr{C}$ of the form
\[
\xymatrix{\varprojlim X_0^{\alpha} \ar[d]_{f_0} & \varprojlim X_1^{\alpha} \ar[l]_-{g_0} \ar[d]_{f_1} & \varprojlim X_2^{\alpha} \ar[l]_-{g_1} \ar[d]_{f_2} & \ldots \ar[l]_-{g_2} \\
\varprojlim\Sigma^{0}D(I_0^{\alpha}) \ar@{-->}[ru]_{\delta_0} & \varprojlim\Sigma^{-1}D(I_1^{\alpha}) \ar@{-->}[ru]_{\delta_1} & \varprojlim\Sigma^{-2}D(I_2^{\alpha}) \ar@{-->}[ru]_{\delta_2} & \ldots.}
\]
For brevity, let us write $[k] := [k]_{\euscr{A}}: \euscr{A} \rightarrow \euscr{A}$ for the internal degree shift. Setting $D_1^{s,t} = H(\varprojlim X_{s}^{\alpha})[s-t]$ and $E_1^{s,t}=H(\varprojlim \Sigma^{-s}D(I_s^{\alpha}))[s-t]$, this induces an exact couple of bigraded objects in $\euscr{A}$ of the form
\[
\xymatrix{D \ar[rr]^-{i_1=H(g_i)[s-t]} && D \ar[ld]^{j_1=H(f_i)[s-t]} \\
& E, \ar[lu]^{k_1=H(\delta_i)[s-t]} }
\]
where the maps have bidegree $|i_1|=(-1,-1)$, $|j_1|=(0,0)$, and $|k_1|=(1,0)$. The $E_1$-term can be identified using \Cref{prop:dlim} and \Cref{lem:dlift} as
\[
\textstyle
E_1^{s,t} = \varprojlim \Sigma^{-s}D(I_s^{\alpha})[s-t] \cong D(\varprojlim_{\euscr{A}} I_s^{\alpha})[-t] = (\varprojlim_{\euscr{A}} I_s^{\alpha})[-t]
\]
and $d_1 = \varprojlim_{\euscr{A}} \tau_s^{\alpha}$. This gives the $E_2$-term
\[
\textstyle E_2^{s,t} \cong \varprojlim_{\euscr{A}}^s(H_*X^{\alpha})[-t],
\]
as needed. It is clear from this construction and \Cref{lem:modadamstower} that the resulting spectral sequence is natural in the tower $X^{\alpha}$ and that all pages and differentials are in $\euscr{A}$.
\end{proof}
\begin{warning}
Note that since the local grading $[1]_{\euscr{A}}\colon \euscr{A} \rightarrow \euscr{A}$ is an equivalence, it commutes with derived functors of the limit, and in the spectral sequence of \cref{thm:limss} we have
\[
E_{2}^{s, t} := \textstyle\varprojlim_{\euscr{A}}^{s} H(X^{\alpha})[-t]_{\euscr{A}} \simeq (\textstyle\varprojlim_{\euscr{A}}^{s} H(x^{\alpha}))[-t]_{\euscr{A}} \simeq E_{2}^{s}[-t]_{\euscr{A}}
\]
Thus, the $t = 0$ line in the spectral sequence already determines the whole $E_{2}$-page. This is similar to the case of the Bockstein spectral sequence, which also has a periodic $E_{2}$-page.
\end{warning}
\begin{remark}
Suppose that $(X^{\alpha})$ is a tower under $X$. Then, the image
\[
\mathrm{im}(H(X)) \subseteq \textstyle\varprojlim^{0}_{\euscr{A}} H(X^{\alpha})
\]
consists of permanent cycles; that is, it is in the kernel of all the differentials in the spectral sequence of \cref{thm:limss}.
To see this, note that the assumption gives a map from the constant tower on $X$ to $(X^{\alpha})$, inducing a map of spectral sequences
\[
\xymatrix{\varprojlim_{\euscr{A}}^sH(X) \ar@{=>}[r] \ar[d] & H(X) \ar[d] \\
\varprojlim^{s}_{\euscr{A}} H(X^{\alpha}) \ar@{=>}[r] & H(\varprojlim X^{\alpha})}
\]
The $E_2$-page of the top spectral sequence is concentrated in
\[
E_2^{0,t} \cong \textstyle\varprojlim_{\euscr{A}}^{0}H(X)[-t]_{\euscr{A}} \cong H(X)[-t]_{\euscr{A}},
\]
so that the spectral sequence collapses. The comparison map is readily identified with the induced morphism $H(X) \to \varprojlim_{\euscr{A}} H(X^{\alpha})$, so the claim follows.
\end{remark}
\subsection{Derived limits of comodules}
We now specialize the above to the case of a homology theory on spectra corresponding to a ring spectrum of Adams type. In order to study the $E_2$-term of the spectral sequence of \cref{thm:limss}, we begin with a general discussion of inverse limits of comodules over suitable flat Hopf algebroids.
Let $(A, \Psi)$ be an Adams Hopf algebroid in the sense of \cite{hovey_htptheory}[1.4.3], so that $\Psi$ is a filtered colimit of dualizable comodules. We will describe an approach to computing limits in the derived $\infty$-category $\dcat(\Psi) = \dcat(\Comod_{\Psi})$ of comodules by reducing to the case of modules. The forgetful functor $\epsilon_*\colon \Comod_{\Psi} \to \Mod(A)$ is \emph{left} adjoint to the extended (or cofree) comodule functor $\epsilon^*$ which sends an $A$-module $M$ to $\Psi \otimes_A M$. These functors are exact and thus give rise to an adjunction:
\[
\epsilon_{*} \dashv \epsilon^{*}\colon \dcat(\Psi) \leftrightarrows \dcat(A).
\]
In terms of chain complexes, both can be computed levelwise.
The corresponding monad $\epsilon^*\epsilon_*$ can be used to construct a resolution of any comodule. Concretely, if $M$ is a $\Psi$-comodule, then we have the associated cobar complex
\[
0 \rightarrow M \rightarrow \Psi \otimes _{A} M \rightarrow \Psi \otimes _{A} \Psi \otimes _{A} M \rightarrow \ldots,
\]
which is a resolution of $M$ by extended comodules, that is, those of the form $\epsilon^{*} N = \Psi \otimes _{A} N$ for an $A$-module $N$.
One way to phrase that the cobar complex is exact is to say that the augmented cosimplicial object from which it arises, namely
\begin{equation}\label{eq:amitsur}
M \rightarrow \Psi \otimes _{A} M \rightrightarrows \Psi \otimes _{A} \Psi \otimes_{A} M \Rrightarrow \ldots,
\end{equation}
is a limit diagram in the derived $\infty$-category $\dcat(\Psi)$. Now, since limit diagrams are stable under levelwise limits, it follows that if $(M_{i})$ is a diagram of comodules, then the diagram
\[
\textstyle\varprojlim _{\dcat(\Psi)} M_{i} \rightarrow \varprojlim _{\dcat(\Psi)} \Psi \otimes _{A} M_{i} \rightrightarrows \varprojlim _{\dcat(\Psi)} \Psi \otimes _{A} \Psi \otimes_{A} M_i \Rrightarrow \ldots,
\]
is also limit, giving an approach to computing $\varprojlim _{\dcat(\Psi)} M_{i}$. To see this, notice that since $\epsilon^{*}$ is a right adjoint and thus preserves limits, we have
\[
\textstyle\varprojlim _{\dcat(\Psi)} \Psi \otimes _{A} N_{i} \simeq \varprojlim _{\dcat(\Psi)} \epsilon^{*} N_{i} \simeq \epsilon^{*} (\varprojlim _{\dcat(A)} N_{i}) \simeq \Psi \otimes _{A} (\varprojlim _{\dcat(A)} N_{i}).
\]
Applying this to the cosimplicial diagram \eqref{eq:amitsur}, we see that we get a limit diagram of the form
\[
\textstyle\varprojlim _{\dcat(\Psi)} M_{i} \rightarrow \Psi \otimes _{A} (\varprojlim _{\dcat(A)} M_{i}) \rightrightarrows \Psi \otimes _{A} (\varprojlim _{\dcat(A)} \Psi \otimes_A M_{i}) \Rrightarrow \ldots
\]
which expresses the limit in the derived category of comodules using only limits taken in the category of modules. Passing to homology, which encodes the derived functors of the limit in the categories of $A$-modules and $\Psi$-comodules, we deduce the following.
\begin{proposition}
\label{prop:spectral_sequence_computing_limits_of_comodules}
If $M_{i}$ is a diagram of comodules, then there exists a spectral sequence of comodules of signature
\[
E_{1}^{s, t} := \textstyle\Psi \otimes_{A} \varprojlim_{A}^s (\Psi^{\otimes_{A} t} \otimes_{A} M_{i}) \implies \varprojlim_{\Psi}^{s+t} M_{i}.
\]
computing the derived functors of the limit in comodules.
\end{proposition}
\begin{remark}
Note that in the particular case of sequential inverse limits, that is, diagrams of the form
\[
\ldots \rightarrow M_{2} \rightarrow M_{1} \rightarrow M_{0},
\]
this spectral sequence will collapse after at most two pages, since in this case we have $\textstyle\varprojlim^{s} _{A} N_{i} = 0$ for $s \neq 0, -1$ and any diagram $N_{i}$ of $A$-modules.
\end{remark}
\begin{remark}
\label{rem:limit_of_comodules_as_a_chain_complex_under_mittag_leffler}
In an even more specific case, let us assume that we work with sequential inverse limits and that the maps $M_{n+1} \rightarrow M_{n}$ are all epimorphisms. If that is the case, the same is true for $\Psi \otimes _{A } \ldots \otimes _{A} M_{n+1} \rightarrow \Psi \otimes _{A} \ldots \otimes _{A} M_{n}$ and we deduce using the Mittag-Leffler condition that all of the $\textstyle\varprojlim^{1} _{A}$-terms vanish. In this case the spectral sequence of \cref{prop:spectral_sequence_computing_limits_of_comodules} has only one potentially non-zero differential.
Phrasing it without the use of spectral sequences, we see that in this the diagram of comodules
\[
\textstyle\Psi \otimes _{A} (\varprojlim _{A} M_{i}) \rightarrow \Psi \otimes _{A} (\varprojlim _{A} \Psi \otimes _{A} M_{i}) \rightarrow \ldots,
\]
is an explicit presentation of the derived limit $\varprojlim _{\dcat(\Psi)} M_{i}$ as a chain complex.
\end{remark}
\begin{example}\label{ex:limss}
Suppose that $R$ is a homotopy commutative ring spectrum of Adams-type. As $R_{*}: \spectra \rightarrow \Comod_{R_{*}R}$ is an adapted homology theory by \cref{lem:dlift}, \cref{thm:limss} specializes to give a spectral sequence
\[
\textstyle E_2^{s,t} \cong \varprojlim_{R_*R}^s(R_{*+t}X^{\alpha}) \implies R_{*+s+t}(\lim_{\alpha} X^{\alpha})
\]
for a tower of spectra $(X^{\alpha})$, where the derived functors of the limit are taken in $R_{*}R$-comodules.
This is a spectral sequence in $R_{*}R$-comodules, so that in particular the $E_{2}$-page is \emph{trigraded}. Passing to internal degree zero, we obtain a more-pleasant looking bigraded spectral sequence
\[
\textstyle E_2^{s,t} \cong (\varprojlim_{R_*R}^s R_{*}X^{\alpha})_{t} \implies R_{s+t}(\lim_{\alpha} X^{\alpha})
\]
Note that in general we cannot rewrite the $E_{2}$-page as ``$\textstyle\varprojlim^{s}_{R_{*}R}(R_{t}X^{\alpha})$'', as the derived functors of the limit are computed in $R_{*}R$-comodules, a structure which cannot be restricted to a single degree.
\end{example}
\begin{warning}
Even for quite reasonable $R$, such as $R = H\mathbb{Q}$, convergence of the spectral sequence of \cref{ex:limss} is a subtle problem. For example, let $(M_{p^i})$ be the tower of mod $p^i$ Moore spectra with the canonical structure maps, and suppose $R=H\mathbb{Q}$. Since
\[
\lim(\ldots \longrightarrow M_{p^2} \longrightarrow M_{p}) \simeq S_p^0,
\]
the $p$-complete sphere, the abutment of the spectral sequence is non-trivial, while the $E_2$-term is zero. We will study this question in more detail in the next section in the case of Morava $E$-theory.
\end{warning}
\begin{remark}
Same methods as those leading to \cref{ex:limss} have been employed by Hovey to set up a spectral sequence computing the $R$-homology of a product of spectra, see \cite{hovey_product}. In \cite[Appendix A]{peterson_coalg}, a different construction of the spectral sequence of \cref{ex:limss} is given, based on Adams resolutions rather than modified Adams resolutions.
\end{remark}
\section{The Morava $E$-homology of inverse limits}
\label{section:morava_e_homology_of_inverse_limits}
The goal of this section is to study the spectral sequence constructed in \cref{thm:limss} in the special case of (uncompleted) Morava $E$-homology. In particular, we will identify the $E_2$-page for the sphere spectrum in terms of the continuous cohomology of the Morava stablizer group and exhibit conditions on the tower that ensure convergence.
We will work with a particular form of Morava $E$-theory, which we used previously in \S\ref{sec:cohomology}. For convenience of the reader, let us fix our notation for the remainder of the current work.
We fix a prime $p$ and a height $n > 0$ and we let $\mathbf{G}_{0}$ be the Honda formal group law over $\mathbb{F}_{q} := \mathbb{F}_{p^{n}}$. We write $E$ for the associated Lubin-Tate spectrum, so that we have a non-canonical isomorphism
\[
E_{*} \simeq W(\mathbb{F}_{p^{n}})\llbracket u_{1}, \ldots, u_{n-1}\rrbracket[u^{\pm 1}].
\]
We write $\mfrak = (p,u_1,\ldots, u_{n-1})$ for the maximal ideal of $E_0$. By Goerss-Hopkins-Miller, the spectrum $E$ is acted on by the extended Morava stabilizer group $\mathbb{G}_{n} := \mathrm{Aut}(\mathbf{G}_{0} / \mathbb{F}_{p}) \rtimes \mathrm{Gal}(\mathbb{F}_{q} / \mathbb{F}_{p})$.
\begin{warning}
We remind the reader that, unless otherwise noted, we work with uncompleted $E$-homology, the uncompleted homology cooperations $E_{*}E := \pi_{*}(E \otimes E)$, and uncompleted $E_*E$-comodules.
\end{warning}
\subsection{Continuous cohomology of filtered colimits}
In this subsection, we prove that continuous cohomology groups of $\mathbb{G}_n$ commute with certain filtered colimits. To this end, it will be useful for us to consider the following variant on the $\mfrak$-adic topology on an $E_{0}$-module.
\begin{definition}
Let $M$ be an $E_{0}$-module. The \emph{local $\mfrak$-adic} topology on $M$ is the linear topology in which a submodule $U \subseteq M$ is open if for every finitely generated $E_{0}$-submodule $M^{\prime} \subseteq M$,
\[
U \cap M^{\prime} \subseteq M^{\prime}
\]
is open in the usual $\mfrak$-adic topology on $M^{\prime}$; that is, we have $\mfrak^{n} M^{\prime} \subseteq U \cap M^{\prime}$ for some $n$ depending on $M^{\prime}$.
\end{definition}
\begin{example}
Suppose that $M$ is finitely generated over $E_{0}$. Then, the local $\mfrak$-adic topology on $M$ coincides with the usual $\mfrak$-adic topology.
\end{example}
\begin{example}
\label{example:local_m_adic_topology_a_filtered_colimit}
Suppose that $M$ is an arbitrary $E_{0}$-module. Then, we can write $M \simeq \varinjlim M_{\alpha}$ as a filtered colimit of finitely generated $E_0$-modules. The local $\mfrak$-adic topology coincides with the colimit topology if we equip each $M_{\alpha}$ with its usual $\mfrak$-adic topology.
\end{example}
The importance of the local $\mfrak$-adic topology in our context comes down to the following observation.
\begin{remark}
Let $X$ be a spectrum. Then, $E_{0}X$ is a continuous $\mathbb{G}_{n}$-module with respect to its local $\mfrak$-adic topology. To see this, observe that we can write $X \simeq \varinjlim X_{\alpha}$ as a filtered colimit of finite spectra, so that
\begin{equation}
\label{equation:homology_a_colimit_of_homologies_of_finite_spectra}
E_{0}X \simeq \varinjlim E_{0} X_{\alpha}.
\end{equation}
Each of $E_{0} X_{\alpha}$ is a continuous $\mathbb{G}_{n}$-module when considered with its $\mfrak$-adic topology, and $E_{0}X$ becomes a topological $\mathbb{G}_{n}$-module since it has the colimit topology with respect to (\ref{equation:homology_a_colimit_of_homologies_of_finite_spectra})
\end{remark}
We will be interested in the continuous cohomology of $\mathbb{G}_{n}$ with coefficients in an $E_{0}$-module equipped with its local $\mfrak$-adic topology. As we will see, this has an elegant description in terms of cohomologies of its finitely generated submodules. The key is the following lemma.
\begin{lemma}
\label{lemma:mapping_into_local_madic_module_lands_in_a_fg_submodule}
Let $K$ be a compact Hausdorff topological space and $M$ be an $E_{0}$-module equipped with its local $\mfrak$-adic topology. Then, any continuous map $f\colon K \rightarrow M$ factors through a finitely generated submodule. In other words, we have
\[
\map_{\cts}(K, M) \simeq \varinjlim \map_{\cts}(K, M_{\alpha}),
\]
where $M_{\alpha}$ is the poset of finitely generated submodules of $M$.
\end{lemma}
\begin{proof}
Suppose by contradiction that for every finitely generated submodule $M^{\prime}$, there exists a $k \in K$ such that $f(k) \not\in M^{\prime}$. Proceeding inductively, we produce an increasing sequence
\[
M_{0} \subseteq M_{1} \subseteq M_{2} \subseteq \ldots
\]
of finitely generated submodules and points $k_{n} \in K$ such that $f(k_{n}) \in M_{n+1}$, but $f(k_{n}) \notin M_{n}$.
We claim that an arbitrary finitely generated submodule $N$ contains at most finitely many of the $f(k_{n})$. To see this, let us write $M_{\infty} = \cup M_{n}$ and consider $N \cap M_{\infty}$; the latter is again finitely generated, as $E_{0}$ is noetherian, and its intersection with $\{ f(k_{n}) \}$ is the same as that of $N$. By finite generation, we have $N \cap M_{\infty} = N \cap M_{n}$ for sufficiently large $n$, proving the claim, as the latter contains at most $n$ of the $f(k_{n})$.
Since each finitely generated submodule is Hausdorff, its finite subspaces are closed and discrete and we deduce that $T = \{ f(k_{n}) \}$ is closed and discrete as a subspace of $M$, as this is true for its intersection with each finitely generated submodule. This is a contradiction, as $T = f(f^{-1}(T))$ must be compact Hausdorff as an image of a closed subspace of $K$, so it cannot be both infinite and discrete.
\end{proof}
\begin{proposition}
\label{proposition:mapping_out_of_ch_top_space_commutes_with_filtered_colimits_of_modules_with_local_madic_top}
Let $M \simeq \varinjlim M_{\alpha}$ be a filtered colimit diagram of $E_{0}$-modules equipped with their local $\mfrak$-adic topologies and let $K$ be compact Hausdorff. Then, the induced map
\[
\theta\colon \varinjlim \map_{\cts}(K, M_{\alpha}) \rightarrow \map_{\cts}(K, M)
\]
is an isomorphism of abelian groups.
\end{proposition}
\begin{proof}
We first show surjectivity of $\theta$. Suppose that $f\colon K \rightarrow M$ is a continuous map, by \cref{lemma:mapping_into_local_madic_module_lands_in_a_fg_submodule} it factors through a finitely generated submodule $N \subseteq M$. If we write $N_{\alpha} = N \times_{M} M_{\alpha}$, then since filtered colimits are exact we have
\[
N \simeq \varinjlim N_{\alpha}
\]
As $N$ is finitely generated and $E_{0}$ is noetherian, it is finitely presented and we deduce that there exists a section $N \rightarrow N_{\alpha}$ for some $\alpha$. Then, the composite
\[
K \rightarrow N \rightarrow N_{\alpha} \rightarrow M_{\alpha}
\]
determines the needed element of the filtered colimit.
For injectivity, suppose that we have a continuous map $f\colon K \rightarrow M_{\alpha}$ such that the composite $s_{\alpha} \circ f\colon K \rightarrow M$ where $s_{\alpha}\colon M_{\alpha} \rightarrow M$ is the canonical map, is zero. By another application of \cref{lemma:mapping_into_local_madic_module_lands_in_a_fg_submodule}, $K$ factors through a finitely generated submodule $N_{\alpha} \subseteq M_{\alpha}$ and we necessarily have $N_{\alpha} \subseteq \mathrm{ker}(s_{\alpha})$. As $N_{\alpha}$ is finitely generated, the second conidtion implies that we can find a larger index $\beta$ such that the composite
\[
N_{\alpha} \rightarrow M_{\alpha} \rightarrow M_{\beta}
\]
is zero. It follows that $f$ determines the zero element of the filtered colimit, as needed.
\end{proof}
\begin{corollary}
\label{corollary:continuous_cohomology_commutes_with_filtered_colimits}
Let $M \simeq \varinjlim M_{\alpha}$ be a filtered colimit of $E_{0}$-modules equipped with their local $\mfrak$-adic topology and compatible continuous $\mathbb{G}_{n}$-actions. Then,
\[
\rmH^{*}_{\cts}(\mathbb{G}_{n}, M) \simeq \varinjlim \rmH^{*}_{\cts}(\mathbb{G}_{n}, M_{\alpha}).
\]
In particular, for any $E_{0}$-module $N$, we have
\[
\rmH^{*}_{\cts}(\mathbb{G}_{n}, N) \simeq \varinjlim \rmH^{*}_{\cts}(\mathbb{G}_{n}, N_{\alpha}).
\]
where the colimit is taken over the poset $N_{\alpha}$ of finitely generated submodules.
\end{corollary}
\begin{proof}
The continuous cohomology is computed by the standard cochain complex
\[
M \rightarrow \map_{\cts}(\mathbb{G}_{n}, M) \rightarrow \map_{\cts}(\mathbb{G}_{n} \times \mathbb{G}_{n}, M) \rightarrow \ldots
\]
of continuous cochains. The statement follows from an application of \cref{proposition:mapping_out_of_ch_top_space_commutes_with_filtered_colimits_of_modules_with_local_madic_top} to each term separately, as $\mathbb{G}_{n} \times \ldots \times \mathbb{G}_{n}$ is compact Hausdorff and taking cohomology of cochain complexes commutes with filtered colimits.
\end{proof}
\subsection{Inverse limits of $E_*E$-comodules and continuous cohomology}\label{ssec:inverselimitsofcomodules}
In this section, we will give a description of derived functors of the limit
\[
\textstyle\varprojlim_{E_{*}E}^{s} M / \mfrak^{k} M \simeq \rmH_{-s}(\ \textstyle\varprojlim_{\dcat(E_{*}E)} M / \mfrak^{k} M)
\]
in $E_{*}E$-comodules as cohomology of the Morava stabilizer group. The importance of these derived functors stems from the fact that they form the $E_{2}$-page of the spectral sequence constructed in \S\ref{subsection:construction_of_the_spectral_sequence}.
\begin{lemma}
\label{lem:inverselimascontcohom}
If $M$ is a dualizable $E_{*}E$-comodule, then we have a canonical isomorphism
\[
\Ext_{\dcat(E_{*}E)}^{s}(E_{*}, \textstyle\varprojlim_{\dcat(E_{*}E)} M / \mfrak^{k} M) := \pi_{-s}\Hom_{\dcat(E_{*}E)}(E_{*}, \textstyle\varprojlim_{\dcat(E_{*}E)} M / \mfrak^{k} M) \simeq \rmH_{\cts}^{s}(\mathbb{G}_{n}, M)
\]
for any $s\ge 0$, where $E_{*} / \mfrak^{k}$ is the $\mfrak$-adic tower of $E_{*}$.
\end{lemma}
\begin{proof}
For brevity, if $X \in \dcat(E_{*}E)$, let us write
\[
\Ext^{s}(X) := \Ext^{s}_{\dcat(E_{*}E)}(E_{*}, X) \simeq \pi_{-s} \Hom_{\dcat(E_{*}E)}(E_{*}, X)
\]
Since we have
\[
\Hom_{\dcat(E_{*}E)}(E, \textstyle\varprojlim _{\dcat(E_{*}E)} M / \mfrak^{k} M) \simeq \varprojlim\Hom_{\dcat(E_{*}E)}(E_{*}, M / \mfrak^{k} M),
\]
there is a Milnor sequence
\[
0 \to \textstyle\varprojlim^1 \Ext^{s-1}(M / \mfrak^{k} M) \to \Ext^{s}(\textstyle\varprojlim_{\dcat(E_{*}E)} M / \mfrak^{k} M) \to \varprojlim \Ext^{s}(M / \mfrak^{k} M) \to 0,
\]
where we have omitted subscripts for simplicity and the derived functors of the limit are taken in abelian groups.
We first claim that the $\varprojlim^1$-term above vanishes. Note that since $M / \mfrak^{k}M$ is finitely generated and $\mfrak$-torsion for any $k \geq 1$, we have a canonical isomorphism
\[
\Ext(M / \mfrak^{k}M) \simeq \rmH_{\cts}^{s}(\mathbb{G}_{n}, M/\mfrak^{k}M)
\]
with the cohomology of the Morava stabilizer group. For $k = 1$, we have a further isomorphism
\[
\Ext(M / \mfrak M) \simeq \Ext_{E_{*}E}(E_{*}, K_{*} \otimes_{E_{*}} M) \simeq \Ext_{E_{*}K}(K_{*}, K_{*} \otimes_{E_{*}} M)
\]
and the group on the right is degreewise finite by \cref{corollary:coh_finiteness_of_ke_comodules}. It follows by induction that $\Ext(M_{*} / \mfrak^k M)$ is degreewise finite for each $k\ge 1$. Therefore, the groups $\Ext(M / \mfrak^k M)$ satisfy the Mittag-Leffler condition as $k$ varies, so that the corresponding $\varprojlim^1$-term vanishes as claimed.
Consequently, we get a string of isomorphisms
\begin{align*}
\Ext^{s}(\textstyle\varprojlim_{\dcat(E_{*}E)} M/ \mfrak^{k} M) & \cong \varprojlim \Ext^{s}(E_{*}, M / \mfrak^{k} M) \\
& \cong \textstyle\varprojlim \rmH_{\cts}^s(\mathbb{G}_n, M / \mfrak^k M) \\
& \cong \rmH_{\cts}^s(\mathbb{G}_n, M).
\end{align*}
Here, the last isomorphism uses that $M$ is $\mfrak$-adically complete as a finitely generated $E_{*}$-module, and that taking continuous cohomology of a finitely generated profinite group commutes with taking countable inverse limits of finite modules \cite[\S 7]{neukirch2013cohomology}.
\end{proof}
In order to pass from \cref{lem:inverselimascontcohom} to a description of the derived functors $\textstyle\varprojlim^{s}_{E_{*}E} E_{*} / \mfrak^{k}$, we have to understand the homology groups of $\varprojlim_{\dcat(E_{*}E)} M / \mfrak^{k} M$, rather than its homotopy. To do so, we will make use of the Adams condition, which involves certain filtered colimits. To commute past the latter, we will need a technical lemma stating that homotopy groups
\[
\Ext^{s} (E_{*}, -) := \pi_{-s} \Hom_{\dcat(E_{*}E)}(E_*,-)
\]
preserve certain filtered colimits of bounded complexes. If $p > n+1$, then the monoidal unit of $\dcat(E_{*}E)$ is compact, so no boundedness hypotheses are required, but this can fail for general $n$. In general, we still have the following result.
\begin{lemma}\label{lem:boundedcompactness}
Suppose $(L_{\alpha})_{\alpha}$ is a filtered system of complexes of $E_*E$-comodules which are uniformly bounded above in the standard $t$-structure on $\dcat(E_{*}E)$; that is, there exists an $N$ such that for all $s>N$ and all $\alpha$ we have $H_s(L_{\alpha}) = 0$. Then the canonical comparison map
\[
\phi\colon \varinjlim_{\alpha} \Ext^{s}(L_{\alpha}) \longrightarrow \Ext^{s}(\varinjlim_{\alpha}L_{\alpha})
\]
is an isomorphism for any $s \in \mathbb{Z}$.
\end{lemma}
\begin{proof}
By shifting the system $(L_{\alpha})_{\alpha}$ if necessary, we may reduce to the case of $s = 0$. In other words, we have to show that that the canonical map
\[
\varinjlim_{\alpha} \ [E_*,L_{\alpha}] \longrightarrow [E_*,\varinjlim_{\alpha}L_{\alpha}]
\]
between homotopy classes of maps in the derived $\infty$-category, is an isomorphism.
Since $E_*$ is connective in the standard $t$-structure, we can then replace $L_{\alpha}$ by its connective cover $\tau_{\ge 0}L_{\alpha}$; that is, reduce to the case in which there exists a non-negative integer $N$ such that $L_{\alpha}$ has homology concentrated in degrees $[0,N]$ for all $\alpha$. By induction and the five-lemma, we can reduce further to the case that $L_{\alpha}$ is concentrated in a single non-negative degree $[s,s]$ for all $\alpha$.
Summarizing these reduction steps, it remains to prove that the canonical map
\[
\varinjlim_{\alpha}\Ext_{E_{*}E}^s(E_*,L_{\alpha}) \longrightarrow \Ext_{E_{*}E}^s(E_*,\varinjlim_{\alpha}L_{\alpha})
\]
is an isomorphism for any filtered system of comodules $L_{\alpha} \in \Comod_{E_*E}$. This is clear, these $\Ext$-groups can be computed by the cobar complex, formation of which commutes with filtered colimits.
\end{proof}
\begin{proposition}\label{prop:inverselimascontcohom}
For any dualizable $E_{*}E$-comodule $M$, there exists a canonical isomorphism
\[
\textstyle\varprojlim_{E_{*}E}^s M / \mfrak^{k} M \cong \rmH^s_{\cts}(\mathbb{G}_n,E_*E \otimes_{E_{*}} M).
\]
between the derived functors of the limit in comodules and the continuous cohomology of the Morava stabilizer group with coefficients in $E_{*}E \otimes_{E_{*}} M$ equipped with its local $\mfrak$-adic topology.
\end{proposition}
\begin{proof}
We have
\[
\textstyle{\varprojlim_{E_{*}E}^s} M_{*} / \mfrak^{k} M \simeq H_{-s}(\textstyle\varprojlim_{\dcat(E_{*}E)} M_{*} / \mfrak^{k} M)
\]
which we can further rewrite as
\[
H_{-s}(\textstyle\varprojlim_{\dcat(E_{*}E)} M / \mfrak^{k} M) \simeq \Ext^{s}(E_{*}E \otimes_{E_{*}} \varprojlim_{\dcat(E_{*}E)} M_{*} / \mfrak^{k} M),
\]
where the tensor product is the derived one. Writing $E_{*}E \simeq \varinjlim N_{\alpha}$ as a filtered colimit of dualizable comodules, we have
\begin{align*}
\Ext^{s}(E_{*}E \otimes_{E_{*}} \textstyle\varprojlim_{\dcat(E_{*}E)} M / \mfrak^{k} M) & \cong \Ext^{s}(\varinjlim \ (N \otimes_{E_{*}} \textstyle\varprojlim_{\dcat(E_{*}E)} M / \mfrak^{k} M)) \\
& \cong \varinjlim \Ext^{s}(N \otimes_{E_{*}} \textstyle\varprojlim_{\dcat(E_{*}E)} M / \mfrak^{k} M) \\
& \cong \varinjlim \Ext^{s}(\textstyle\varprojlim_{\dcat(E_{*}E)} N \otimes_{E_{*}} M / \mfrak^{k}(N \otimes_{E_{*}} M)) \\
& \cong \varinjlim \ \rmH_{\cts}^{s}(\mathbb{G}_{n}, N \otimes_{E_{*}} M) \\
& \cong \rmH_{\cts}^{s}(\mathbb{G}_{n}, \varinjlim \ N \otimes_{E_{*}} M) \\
& \cong \rmH^s_{\cts}(\mathbb{G}_n,E_{*}E \otimes_{E_{*}} M).
\end{align*}
Here, the first isomorphism uses that the derived tensor product commutes with filtered colimits and the second one is \cref{lem:boundedcompactness}. The third isomorphism uses that tensoring with a dualizable object commutes with limit, the fourth one is \cref{lem:inverselimascontcohom} while the fifth one is \cref{corollary:continuous_cohomology_commutes_with_filtered_colimits}.
\end{proof}
\begin{remark}
The relationship between derived functors of the limit and derived completion in the context of comodules over a flat Hopf algebroid is studied further in \cite{bhv4}.
\end{remark}
\subsection{The inverse limit spectral sequence and its convergence}
In this subsection, we study the convergence of the inverse limit spectral sequence based on Morava $E$-theory and combine the previous results in a spectral sequence computing the (uncompleted) $E$-homology of the $K$-local sphere spectrum from continuous group cohomology with coefficients in the (uncompleted) cooperations $E_*E$.
\begin{proposition}
\label{prop:limssconvergence}
If $R=E$ is Morava $E$-theory and $(X^{\alpha}) \in \spectra_E^{\mathbb{N}}$ is a tower of $E$-local spectra, then the spectral sequence of \cref{ex:limss} converges conditionally and strongly, and has a horizontal vanishing line at the $E_r$-page for some $r\ge 2$.
\end{proposition}
\begin{proof}
By \Cref{lem:modadamstower} and assumption on $X^{\alpha}$, the modified Adams tower $(X_i^{\alpha})$ is $E$-local for all $i$. Therefore, the proof of the smash product theorem due to Hopkins and Ravenel~\cite[Theorem 7.5.6]{orangebook} yields a constant $N$ independent of $\alpha$ such that
\[
\xymatrix{g_s^{\alpha}\circ \ldots \circ g_{s+N-1}^{\alpha}\colon X_{s+N}^{\alpha} \ar[r] & X_s^{\alpha}}
\]
is null for all $\alpha$; see \cite[Theorem 5.10]{hs_localcohom}. Consequently,
\begin{equation}\label{eq:zerocomposite}
\xymatrix{g_s\circ \ldots \circ g_{s+N-1}\colon \varprojlim X_{s+N}^{\alpha} \ar[r] & \varprojlim X_s^{\alpha}}
\end{equation}
is null as well, so \cite{boardman_convergence} shows that the spectral sequence is conditionally convergent. Similarly, we obtain the horizontal vanishing line from \eqref{eq:zerocomposite}, which in turn gives strong convergence.
\end{proof}
\begin{theorem}
If $(X^{\alpha}) \in \spectra_E^{\mathbb{N}}$ is a tower of $E$-local spectra, then there is a strongly convergent spectral sequence
\[
\xymatrix{E_2^{s,t} \cong (\varprojlim_{E_*E}^s E_*X^{\alpha})_t \implies E_{t-s}(\varprojlim X^{\alpha}),}
\]
with a horizontal vanishing line at the $E_r$-page for some $r\ge 2$. In particular, this spectral sequence specializes to:
\begin{equation}\label{eq:elimssk(n)localsphere}
\xymatrix{\rmH_{\cts}^s(\mathbb{G}_n,E_{t}E) \cong (\varprojlim_{E_*E}^s E_*/\mfrak^k)_t \implies E_{t-s}(L_KS^0).}
\end{equation}
\end{theorem}
\begin{proof}
The spectral sequence was constructed in \cref{thm:limss}, see also \cref{ex:limss}, while the convergence properties were established in \cref{prop:limssconvergence}. The final statement follows by applying this spectral sequence to a cofinal tower of $E$-local generalized type $n$ Moore spectra as in \cite[Section 4]{hovey_morava_1999}. The identification of the $E_2$-page in terms of continuous cohomology was proven in \cref{prop:inverselimascontcohom}.
\end{proof}
\begin{remark}
The spectral sequence \eqref{eq:elimssk(n)localsphere} may be thought of as an $E_*$-homology version of the $K$-local $E$-based Adams spectral sequence for the sphere, which has signature
\[
\rmH_{\cts}^s(\mathbb{G}_n,E_*)_t \implies \pi_*L_{K}S^0
\]
and is also strongly convergent. In contrast, working with completed $E_*$-homology would result in an isomorphism $E_* \cong E_*^{\vee}L_{K}S^0 \cong \rmH_{\cts}^*(\mathbb{G}_n,E_*^{\vee}E_*)$. This amply highlights the extra information contained in the terms $\rmH_{\cts}^*(\mathbb{G}_n,E_*E)$.
\end{remark}
\section{Digression: Cohomology of $\mathbb{G}_{n}$ as $\Ext$ in comodules}\label{sec:cohomofgn}
In \cref{lem:inverselimascontcohom}, we have expressed continuous cohomology of a Morava stabilizer group with coefficients in a dualizable $E_{*}E$-comodule $M$ as homotopy groups of a the derived completion. In this short section, we will use a form of local duality in the context of comodules to express it as an $\Ext$-groups between actual comodules, rather than objects of the derived $\infty$-category. Using this calculation we will define a comparison map between $\Ext$-groups and continuous cohomology.
Let $I_k = (p,v_1,\ldots,v_{k-1})$ be the $k$-th invariant chromatic ideal in $E_*$; in particular, $I_n = \mathfrak{m}$. We define $E_* = E_*/I_0^{\infty}$ and then iteratively construct $E_*E$-comodules via the cofiber sequences
\begin{equation}\label{eq:cousin}
E_{*} / I_k^{\infty} \rightarrow v_{k}^{-1} E_{*} / I_k^{\infty} \rightarrow E_{*} / I_{k+1}^{\infty}
\end{equation}
in $\dcat(E_*E)$. Note that $E_{*} / I_k^{\infty} \simeq \varinjlim E_{*} / v_{0}^{l_{0}}, \ldots, v_{k-1}^{l_{k-1}}$, where the colimit is taken over a set of indices with $l_{i} \rightarrow \infty$. For the constituent pieces, we have the following self-duality result:
\begin{lemma}
\label{lemma:quotients_of_estar_are_selfdual_in_the_derived_category}
In the derived category of $E_{*}E$-comodules, the comodule $E_{*} / v_{0}^{i_{0}}, v_{1}^{i_{1}}, \ldots, v_{k}^{i_{k}}$ is self-dual with a shift in the sense that there exists an equivalence
\[
D(E_{*} / v_{0}^{i_{0}}, v_{1}^{i_{1}}, \ldots, v_{k}^{i_{k}}) \simeq \Sigma^{-k} E_{*} / v_{0}^{i_{0}}, v_{1}^{i_{1}}, \ldots, v_{k}^{i_{k}},
\]
where $DX = F(X, E_{*})$ denotes the monoidal dual.
\end{lemma}
\begin{proof}
We prove this by induction on $k$, where $k = -1$ is clear, since $E_{*}$ is the monoidal unit.
Now assume that the statement is known for $E_{*} / v_{0}^{i_{0}}, v_{1}^{i_{1}}, \ldots, v_{k-1}^{i_{k-1}}$. Multiplication by $v_{k}^{i_{k}}$ induces a cofiber sequence
\[
E_{*} / v_{0}^{i_{0}}, v_{1}^{i_{1}}, \ldots, v_{k-1}^{i_{k-1}} \rightarrow E_{*} / v_{0}^{i_{0}}, v_{1}^{i_{1}}, \ldots, v_{k-1}^{i_{k-1}} \rightarrow E_{*} / v_{0}^{i_{0}}, v_{1}^{i_{1}}, \ldots, v_{k}^{i_{k}}
\]
which after taking duals yields a cofiber sequence
\[
D(E_{*} / v_{0}^{i_{0}}, v_{1}^{i_{1}}, \ldots, v_{k}^{i_{k}}) \rightarrow D(E_{*} / v_{0}^{i_{0}}, v_{1}^{i_{1}}, \ldots, v_{k-1}^{i_{k-1}}) \rightarrow D(E_{*} / v_{0}^{i_{0}}, v_{1}^{i_{1}}, \ldots, v_{k-1}^{i_{k-1}}).
\]
By the inductive assumption, this sequence can be rewritten as
\[
D(E_{*} / v_{0}^{i_{0}}, v_{1}^{i_{1}}, \ldots, v_{k}^{i_{k}}) \rightarrow \Sigma^{k-1} E_{*} / v_{0}^{i_{0}}, v_{1}^{i_{1}}, \ldots, v_{k-1}^{i_{k-1}} \rightarrow \Sigma^{k-1} E_{*} / v_{0}^{i_{0}}, v_{1}^{i_{1}}, \ldots, v_{k-1}^{i_{k-1}}.
\]
The self-map of $\Sigma^{k-1} E_{*} / v_{0}^{i_{0}}, v_{1}^{i_{1}}, \ldots, v_{k-1}^{i_{k-1}} \rightarrow \Sigma^{k-1} E_{*} / v_{0}^{i_{0}}, v_{1}^{i_{1}}, \ldots, v_{k-1}^{i_{k-1}}$ appearing on the right cannot be non-zero, since it is a dual to the non-zero class $v_{k}^{i_{k}}$. It follows that it must be itself a unit times $v_{k}^{i_{k}}$ since there is nothing else in this degree, see for example \cite[Thm.~4.3.2(b)]{greenbook}. The assertion then follows.
\end{proof}
\begin{lemma}\label{lem:comodulecompletion}
If $M$ is a finitely generated $E_{*}E$-comodule, then its derived comodule completion of $M$ is given by
\[
\textstyle\varprojlim_{\dcat(E_*E)}(M\otimes_{E_{*}} E_*/v_{0}^{k_{0}}, \ldots, v_{n-1}^{k_{n-1}}) \simeq \textstyle\varprojlim_{\dcat(E_*E)}M/v_{0}^{k_{0}}, \ldots, v_{n-1}^{k_{n-1}},
\]
where on the left hand side we have used the derived tensor product of comodules.
\end{lemma}
\begin{proof}
This is proven in \cite[Proposition 2.25]{bhv4}
\end{proof}
The next result is an incarnation of local duality for comodules \cite[\S5]{bhv1}, relating local cohomology (derived torsion) and local homology (derived completion), combined with \cref{lem:inverselimascontcohom}.
\begin{theorem}
\label{thm:cohomology_of_gn_as_ext_in_comodules}
Let $M$ be a finitely generated $E_{*}E$-comodule. Then, there exists an isomorphism
\[
\Ext^{n+s}_{E_{*}E}(E_{*} / I_n^{\infty}, M) \simeq \rmH_{\cts}^{s}(\mathbb{G}_{n}, M)
\]
between $\Ext$-groups in comodules and the cohomology of $\mathbb{G}_{n}$, for any $s\ge 0$.
\end{theorem}
\begin{proof}
As noted before, we have $E_{*} / I_n^{\infty} \simeq \varinjlim E_{*} / v_{0}^{k_{0}}, \ldots, v_{n-1}^{k_{n-1}}$, where the colimit is taken over a set of indices with $k_{i} \rightarrow \infty$. By by the duality equivalence of \cref{lemma:quotients_of_estar_are_selfdual_in_the_derived_category}
we have
\[
\Hom_{\dcat(E_{*}E)}(\varinjlim\Sigma^{-n}E_{*} / v_{0}^{k_{0}}, \ldots, v_{n-1}^{k_{n-1}}, M) \simeq \Hom_{\dcat(E_{*}E)}(E_{*}, \textstyle\varprojlim_{\dcat(E_{*}E)} M \otimes E_*/v_{0}^{k_{0}}, \ldots, v_{n-1}^{k_{n-1}}).
\]
In light of the \cref{lem:comodulecompletion}, the latter term is equivalent to $\Hom_{\dcat(E_{*}E)}(E_{*}, \varprojlim M/v_{0}^{k_{0}}, \ldots, v_{n-1}^{k_{n-1}})$. It follows that
\[
\Ext^{n+s}_{E_{*}E}(E_{*} / I_n^{\infty}, M) \simeq \varprojlim \Ext^{s}_{E_{*}E}(E_{*}, M/v_{0}^{k_{0}}, \ldots, v_{n-1}^{k_{n-1}}),
\]
up to possible $\varprojlim^{1}$-terms. However, all of the Ext groups on the right are finite, since $M$ is finitely generated, so these $\varprojlim^{1}$-terms vanish. We then deduce as in \cref{lem:inverselimascontcohom} that
\[
\Ext^{n+s}_{E_{*}E}(E_{*} / I_n^{\infty}, M) \simeq \varprojlim \Ext_{E_{*}E}^{s}(E_{*}, M / v_{0}^{k_{0}}, \ldots, v_{n-1}^{k_{n-1}}) \simeq \rmH_{\cts}^{s}(\mathbb{G}_{n}, M),
\]
which is what we wanted to show.
\end{proof}
The boundary maps $\delta_k\colon E_*/I_{k+1}^{\infty} \to \Sigma E_*/I_{k}^{\infty}$ associated with the cofiber sequences \eqref{eq:cousin} compose to give a `residue map' in the derived category:
\[
\delta := \delta_{n-1} \circ \ldots \circ \delta_{0} \colon E_*/I_{n}^{\infty} \to \Sigma^n E_*.
\]
For any finitely generated $E_*E$-comodule $N$, the map $\delta$ induces a natural map on Ext-groups that fits, for any $s$, into a commutative square
\begin{equation}\label{eq:comparisonmap}
\vcenter{
\xymatrix{\Ext^{s}_{E_{*}E}(E_{*}, N) \ar[r]^-{\Ext(\delta,N)} \ar[d]_{\cong} & \Ext^{s+n}_{E_{*}E}(E_*/I_{n}^{\infty}, N) \ar[d]^{\cong} \\
[N^{\dual},E_*]^{-s} \ar[r] & [N^{\dual}, \varprojlim _{\dcat(E_{*}E)} E_{*} / I_n^{k}]^{-s}.}
}
\end{equation}
Here, the left vertical isomorphism is by duality, while the right vertical isomorphism is established as in the proof of the previous theorem. Unwinding the construction, the bottom horizontal map is then induced by the canonical completion map $E_* \to \varprojlim _{\dcat(E_{*}E)} E_{*} / I_n^{k}$.
\begin{definition}\label{def:comparisonmap}
For a finitely generated $E_*E$-comodule $N$, we define the comparison map $\alpha_N$ as the composite
\[
\alpha_N\colon \Ext^{s}_{E_{*}E}(E_{*}, N) \to \Ext^{n+s}_{E_{*}E}(E_{*} / I_n^{\infty}, N) \simeq \rmH_{\cts}^{s}(\mathbb{G}_{n}, N)
\]
of the map in \eqref{eq:comparisonmap} and the isomorphism of \cref{thm:cohomology_of_gn_as_ext_in_comodules}.
\end{definition}
Note that, by construction, the maps $\alpha_{N}$ are natural and compatible with the long exact sequences of $\Ext$-groups and cohomology.
\begin{remark}
Alternatively, it is possible to define the comparison map $\alpha_M$ directly using the cobar complex. More precisely, one can show that if $M$ is a finitely generated $E_{*}E$-comodule, then the differentials
in the cobar complex
\[
M \rightarrow E_{*}E \otimes_{E_{*}} M \rightarrow E_{*}E \otimes_{E_{*}} E_{*}E \otimes_{E_{*}} M \rightarrow \ldots
\]
respect the $\mfrak$-adic filtration (even though they are not $E_{*}$-linear). The completion of the cobar complex is exactly the cochain complex computing continuous $\mathbb{G}_{n}$-cohomology, and the induced map on homology can be shown to be exactly $\alpha_{M}$.
\end{remark}
\section{Algebraic chromatic splitting conjecture}\label{sec:specialcaseofacss}
In this penultimate section we use the comparison map of the previous section to compute the zeroth derived limit as $\varprojlim^{0}_{E_{*}E} E_{*} / \mfrak^{k} \simeq E_{*} \otimes _{\mathbb{Z}} \mathbb{Z}_{p}$. This result should be viewed in light of Hopkins' algebraic chromatic splitting conjecture, see \cite[\S14]{report_on_e_theory_conjectures}, which we briefly review for context.
\begin{notation}\label{not:abbreviations}
To simplify notation, if $N$ is a finitely generated $E_{*}E$-comodule, we will sometimes use abbreviations $\Ext^{s}(N) := \Ext^{s}_{E_{*}E}(E_{*}, N)$ and $\rmH^{s}(N) := \rmH_{\cts}^{s}(\mathbb{G}_{n}, N)$.
\end{notation}
\subsection{The algebraic chromatic splitting conjecture}
Motivated by computations in height 2, this conjecture was formulated by Hopkins and recorded later in \cite[Conjecture 134]{report_on_e_theory_conjectures}\footnote{We note that the statement given there contains a typo: the left hand side of the equation should have $v_{n-1}$ inverted.} as well as \cite[Conjecture 6.10]{barthel2020chromatic}, and studied further in \cite{bh_acht}. To state it, we first recall the construction of the class $\zeta \in \pi_{-1}L_{K(n)}S^0$ due to Devinatz and Hopkins~\cite[Section 8]{devinatzhopkins_hfp}.
If we write $\mathbb{G}_n^1$ for the kernel of the determinant map $\det\colon \mathbb{G}_n \to \mathbb{Z}_p$, there is a residual action of $\mathbb{Z}_p$ on $E_n^{h\mathbb{G}_n^1}$. For $t$ is a topological generator of $\mathbb{Z}_p$, we then obtain a fiber sequence of spectra
\[
\xymatrix{L_{K(n)}S^0 \ar[r] & E_n^{h\mathbb{G}_n^1} \ar[r]^-{\id-t} & E_n^{h\mathbb{G}_n^1} \ar[r]^-{\delta} & L_{K(n)}S^1.}
\]
The composition $\delta\circ \eta \colon S^0 \to L_{K(n)}S^0$ of the unit map with the boundary map gives rise to an element $\zeta \in \pi_{-1}L_{K(n)}S^0$. In \cite[Proposition 8.2]{devinatzhopkins_hfp}, Devinatz and Hopkins prove that $\zeta$ is non-trivial as long as the height $n>0$.
\begin{conjecture}[Algebraic chromatic splitting conjecture]\label{conj:acsc}
For all heights $n \ge 2$ and sufficiently large $p$, there are isomorphisms of $E_*E$-comodules
\[
\textstyle\varprojlim_{E_{*}E}^sE_*/(p,\ldots,v_{n-2},v_{n-1}^i) \cong
\begin{cases}
E_*/(p,\ldots,v_{n-2}) & s=0 \\
v_{n-1}^{-1}E_*/(p,\ldots,v_{n-2}) & s=1 \\
0 & \mathrm{otherwise.}
\end{cases}
\]
Moreover, these isomorphisms are topologically realized by the unit map $\iota\colon S^0 \to L_{K(n)}S^0$ and the map $\zeta\colon S^{-1} \to L_{K(n)}S^0$, respectively.
\end{conjecture}
\begin{remark}
As the name suggests, the algebraic chromatic splitting conjecture provides an algebraic counterpart to Hopkins' chromatic splitting conjecture \cite[Conjecture 4.2]{hovey_csc} in the edge case: If the algebraic chromatic splitting conjecture holds at height $n$ and prime $p$, then the maps $\iota$ and $\zeta$ induce a splitting
\[
(\iota,\zeta)\colon L_{n-1}F \oplus \Sigma^{-1}L_{n-1}F \simeq L_{n-1}L_{K(n)}F
\]
for any finite spectrum $F$ of type $n-1$. Indeed, the conclusion of the algebraic chromatic splitting conjecture forces the inverse limit sequence (\ref{eq:generalss}) for $L_{K(n)}F$ to collapse at the $E_2$-page, thereby showing $(\iota,\zeta)$ is an $E_*$-isomorphism and hence an $E$-local equivalence.
\end{remark}
\begin{example}
In \cref{sec:heightone}, we will verify a height 1 variant of the algebraic chromatic splitting conjecture at any prime, see in particular \cref{theorem:cohomology_of_ee_at_height_one} and also \cite[\S14]{report_on_e_theory_conjectures}. The appearance of the additional tensor factors is the only reason we excluded the case of height 1 from the statement of the algebraic chromatic splitting conjecture above.
\end{example}
\subsection{The comparison map in degree 0}
We work with \cref{not:abbreviations}. Our goal is to study the comparison map $\Ext^{s}(M) \rightarrow \rmH^{s}(M)$, constructed in \cref{def:comparisonmap}. We will focus on the particular case when $s = 0$; note that in this case this map is always injective as both the source and target are abelian subgroups of $M$.
Thus, the real question is how far this comparison map is from being surjective. When $M$ is the monoidal unit, we have the following folklore result of Hopkins.
\begin{lemma}[Hopkins]
\label{lemma:invariants_of_the_lubin_tate_ring}
We have $\rmH^{0}(E_{*}) \simeq \mathbb{Z}_{p}$ and the map $\Ext^{0}(E_{*}) \rightarrow \rmH^{0}(E_{*})$ is a $p$-completion; that is, it induces an isomorphism $\mathbb{Z}_{p} \otimes_{\mathbb{Z}} \Ext^{0}(E_{*}) \simeq \rmH^{0}(E_{*})$.
\end{lemma}
\begin{proof}
The first part is \cite[Lemma 1.33]{bobkova2018topological}. The second part follows from the fact that $\Ext^{0}(E_{*}) \simeq \mathbb{Z}_{(p)}$, as in both cases the relevant invariants are generated by the unit of the Lubin--Tate ring.
\end{proof}
\begin{lemma}
\label{lemma:ext_h_comparison_mono_on_h1_for_e*}
The map $\Ext^{1}(E_{*}) \rightarrow \rmH^{1}(E_{*})$ is a monomorphism.
\end{lemma}
\begin{proof}
Since it vanishes rationally, the source of this map is $p$-torsion. For a given $x \in \Ext^{1}(E_{*})$ we can thus find a $k$ and $y \in \Ext^{0}(E_{*}/p^{k})$ so that $x$ can be written as $x = \delta(y)$, where $\delta$ denotes the boundary homomorphism. We have a commutative diagram
\[\begin{tikzcd}
{\Ext^{0}(E_{*})} & {\Ext^{0}(E_{*})} & {\Ext^{0}(E_{*}/p^{k})} & {\Ext^{1}(E_{*})} \\
{\rmH^{0}(E_{*})} & {\rmH^{0}(E_{*})} & {\rmH^{0}(E_{*}/p^{k})} & {\rmH^{1}(E_{*})}
\arrow["{p^{k}}", from=2-1, to=2-2]
\arrow[from=2-2, to=2-3]
\arrow["\delta", from=2-3, to=2-4]
\arrow[from=1-4, to=2-4]
\arrow[from=1-3, to=2-3]
\arrow["\delta", from=1-3, to=1-4]
\arrow[from=1-2, to=2-2]
\arrow[from=1-2, to=1-3]
\arrow[from=1-1, to=2-1]
\arrow[from=1-1, to=1-2]
\end{tikzcd}\]
with exact rows. Thus, it is enough to show that if the image $\widetilde{y} \in \rmH^{0}(E_{*}/p^{k})$ of $y$ can be lifted to $\rmH^{0}(E_{*})$, then $y$ can be lifted to $\Ext^{0}(E_{*})$. By \cref{lemma:invariants_of_the_lubin_tate_ring}, the latter two groups are isomorphic to respectively $\mathbb{Z}_{p}$ and $\mathbb{Z}_{(p)}$.
By multiplying by an appropriate unit $\sigma$ in the $p$-adics congruent to $1$ modulo $p^{k}$, and replacing the chosen lift $\widetilde{y} \in \rmH^{0}(E_{*}) \simeq \mathbb{Z}_{(p)}$ of $\widetilde{x}$ by $\sigma \widetilde{y}$, we can assume that $\widetilde{y} \in \mathbb{Z}_{(p)}$. It follows that $x$ can be lifted to $\Ext^{0}(E_{*})$, which is what we wanted.
\end{proof}
Our goal is to use an inductive argument, building on the calculation of invariants of $E_{*}$ due to Hopkins and extending it to a larger class of comodules. The starting point will be the following class.
\begin{definition}
\label{definition:pure_comodule}
We say a finitely generated $E_{*}E$-comodule $M$ is \emph{pure} if it belongs to the smallest class of comodules closed under extensions and containing all shifts of $E_{*}$.
\end{definition}
\begin{remark}
\label{remark:pure_comodules_closed_under_taking_duals}
Any pure $M$ is finitely generated and projective over $E_{*}$, and so dualizable. It follows that if
\[
0 \rightarrow K \rightarrow M \rightarrow N \rightarrow 0
\]
is a short exact sequence of pure comodules, then it is split over $E_{*}$, so that the induced sequence
\[
0 \rightarrow N^{\dual} \rightarrow M^{\dual} \rightarrow K^{\dual} \rightarrow 0
\]
of linear duals is again exact. We deduce that the class of pure comodules is closed under taking duals.
\end{remark}
\begin{remark}
It is plausible that it follows from the arguments of Hovey and Strickland, who establish an analogue of the Landweber filtration for $E_{*}E$-comodules \cite{hovey2005comodules}, that all comodules finitely generated and projective over $E_{*}$ are pure in the sense of \cref{definition:pure_comodule}. We did not check whether this is true, as we do not need this fact in our arguments.
\end{remark}
\begin{lemma}
\label{lemma:comodules_generated_by_extensions_of_e*}
The category of $E_{*}E$-comodules is generated under colimits by pure comodules.
\end{lemma}
\begin{proof}
Since $E \simeq \varinjlim E_{\alpha}$ is a filtered colimit of finite spectra, it follows formally that the category of $E_{*}E$-comodules is generated under colimits by shifts of comodules of the form $E_{*} E_{\alpha}$ \cite[Proposition 1.4.4]{hovey_htptheory}.
Furthermore, we claim that since the $E_{\alpha}$ only have even cells, the resulting comodules $E_{*} E_{\alpha}$ are pure. Using a cell decomposition of $E_{\alpha}$ and the associated long exact sequences in homology, which will be short exact here, we deduce that each $E_{*}E_{\alpha}$ is a comodule which can be obtained using iterated extensions of shifts of $E_{*}$ and so is pure.
\end{proof}
Recall that a $E_{*}E$-comodule is dualizable if and only if the underlying $E_*$-module is finitely generated and projective.
\begin{theorem}
\label{theorem:ext_0_coincides_with_cohomology_after_p_completion_for_fp_comodules}
If $N$ is a dualizable $E_{*}E$-comodule, then
\begin{enumerate}
\item $\Ext^{0}(N)$ is a finitely generated $\mathbb{Z}_{(p)}$-module and
\item the map $\alpha_{N}\colon \Ext^{0}(N) \rightarrow \rmH^{0}(N)$ is a $p$-completion, that is, it induces an isomorphism $\Ext^{0}(N) \otimes _{\mathbb{Z}} \mathbb{Z}_{p} \simeq \rmH^{0}(N)$.
\end{enumerate}
\end{theorem}
\begin{proof}
For $N = E_{*}$ or its shifts, this is \cref{lemma:invariants_of_the_lubin_tate_ring}.
We first claim that the result is true for comodules $C$ which are pure in the sense of \cref{definition:pure_comodule}. By induction on rank, we can assume that we have a short exact sequence
\[
0 \rightarrow E_{*} \rightarrow C \rightarrow C^{\prime} \rightarrow 0
\]
and that the result already holds for $C^{\prime}$. The long exact sequence of $\Ext$-groups implies that $\Ext^{0}(C)$ is a finitely generated module over $\mathbb{Z}_{(p)}$. Thus, we have a commutative diagram
\[\begin{tikzcd}
0 & {\mathbb{Z}_{p} \otimes \Ext^{0}(E_{*})} & {\mathbb{Z}_{p} \otimes \Ext^{0}(C)} & {\mathbb{Z}_{p} \otimes \Ext^{0}(C^{\prime})} & {\mathbb{Z}_{p} \otimes \Ext^{1}(E_{*})} \\
0 & {\rmH^{0}(E_{*})} & {\rmH^{0}(C)} & {\rmH^{0}(C^{\prime})} & {\rmH^{1}(E_{*}),}
\arrow[from=1-5, to=2-5]
\arrow[from=1-4, to=2-4]
\arrow[from=1-3, to=2-3]
\arrow[from=1-2, to=2-2]
\arrow[from=1-1, to=1-2]
\arrow[from=2-1, to=2-2]
\arrow[from=1-2, to=1-3]
\arrow[from=2-2, to=2-3]
\arrow[from=1-3, to=1-4]
\arrow[from=2-3, to=2-4]
\arrow[from=1-4, to=1-5]
\arrow[from=2-4, to=2-5]
\end{tikzcd}
\]
where the top row is again exact, since $\mathbb{Z}_{p}$ is flat. By induction on rank, counting from the left, the first and third vertical maps are isomorphisms. By \cref{lemma:ext_h_comparison_mono_on_h1_for_e*}, the fourth one is a monomorphism, as $\mathbb{Z}_{p} \otimes \Ext^{1}(E_{*}) \simeq \Ext^{1}(E_{*})$ because the latter group is $p$-torsion. We deduce that the second one is an isomorphism using the five-lemma, which was our claim.
Now suppose that $N$ is an arbitrary dualizable comodule; by \cref{lemma:comodules_generated_by_extensions_of_e*} we can find a surjection $C \rightarrow N$ from a pure comodule. By taking duals, we can assume that we instead have a monomorphism $N \hookrightarrow C$, where $C$ is again pure by \cref{remark:pure_comodules_closed_under_taking_duals}. Since we already know $\Ext^{0}(C)$ and $\rmH^{0}(C)$ are free finitely generated over respectively $\mathbb{Z}_{(p)}$ and $\mathbb{Z}_{p}$, we deduce that the same is true for their submodules $\Ext^{0}(N)$ and $\rmH^{0}(N)$, proving $(1)$.
Let us denote the cokernel of $N \hookrightarrow C$ by $C^{\prime}$. We then have another commutative diagram
\[
\xymatrix{0 \ar[r] & \mathbb{Z}_{p} \otimes \Ext^{0}(N) \ar[r] \ar[d] & \mathbb{Z}_{p} \otimes \Ext^{0}(C) \ar[r] \ar[d] & \mathbb{Z}_{p} \otimes \Ext^{0}(C^{\prime}) \ar[d] \\
0 \ar[r] & \rmH^{0}(N) \ar[r] & \rmH^{0}(C) \ar[r] & \rmH^{0}(C^{\prime}).}
\]
By what we have proven above, the middle map is an isomorphism as well. Since the right one is injective, we deduce from the four-lemma that the left one is surjective, ending the argument.
\end{proof}
\begin{remark}
One would like to prove that \cref{theorem:ext_0_coincides_with_cohomology_after_p_completion_for_fp_comodules} holds for all finitely generated comodules, rather than only the finite projective ones. It is possible this more general result can be deduced formally from the projective case, but we were not able to do so.
The key ingredient in the above proof is the calculation of $\rmH_{\cts}^{0}(\mathbb{G}_{n}, E_{*})$; it seems likely that the version for arbitrary comodules would follow from a more general calculation of $\rmH_{\cts}^{0}(\mathbb{G}_{n}, E_{*} / I_{k})$ by Landweber filtration arguments.
\end{remark}
\subsection{The comodule completion of Morava $E$-theory}
We now come to the main theorem of this section. One may view the computation of this inverse limit at all heights and primes as uniform evidence for the validity for the algebraic chromatic splitting conjecture and hence of the chromatic splitting conjecture itself.
\begin{theorem}
\label{thm:zeroth_derived_limit_of_completion}
We have
\[
\textstyle\varprojlim_{E_{*}E} E_{*} / \mfrak^{k} \simeq \rmH_{\cts}^0(\mathbb{G}_n,E_*E) \simeq E_{*} \otimes _{\mathbb{Z}} \mathbb{Z}_{p},
\]
where the left hand side is the limit in the category of $E_{*}E$-comodules.
\end{theorem}
\begin{proof}
The first isomorphism is an instance of \cref{prop:inverselimascontcohom}. By \cref{corollary:continuous_cohomology_commutes_with_filtered_colimits}, if $M_{\alpha}$ is a filtered diagram of dualizable $E_{*}E$-comodules such that $\varinjlim M_{\alpha} \simeq E_{*}E$, then
\[
\rmH_{\cts}^0(\mathbb{G}_n,E_*E) \simeq \varinjlim \rmH_{\cts}^0(\mathbb{G}_n, M_{\alpha}) \simeq \Ext^{0}_{E_{*}E}(E_{*}, M_{\alpha}) \otimes_{\mathbb{Z}} \mathbb{Z}_{p} \simeq E_{*} \otimes_{\mathbb{Z}} \mathbb{Z}_{p}
\]
where the second isomorphism is \cref{theorem:ext_0_coincides_with_cohomology_after_p_completion_for_fp_comodules}. This ends the argument.
\end{proof}
\begin{remark}
Since $E_{*} S^{0}_{p} \simeq E_{*} \otimes _{MU_{*}} MU_{*}S^{0}_{p} \simeq E_{*} \otimes \mathbb{Z}_{p}$, one can rephrase \cref{thm:zeroth_derived_limit_of_completion} as saying that $\varprojlim^{0} E_{*} / \mfrak^{k} \simeq E_{*}S^{0}_{p}$, which relates this result to the topological chromatic splitting conjecture.
\end{remark}
\begin{remark}
There is a short exact sequence of $E_*E$-modules
\[
\xymatrix{0 \ar[r] & \Sigma^{|v_{n-1}|}E_*/I_{n-1} \ar[r]^-{v_{n-1}} & E_*/I_{n-1} \ar[r] & E_*/I_{n} \ar[r] & 0.}
\]
Applying the functor $\varprojlim_{\dcat(E_{*}E)}E_*/I_n^k\otimes-$ and considering the induced long exact sequence in cohomology, we deduce that $v_{n-1}$ acts on $\varprojlim_{\dcat(E_{*}E)}^sE_*/(p,\ldots,v_{n-2},v_{n-1}^i)$ injectively for $s=0$ and bijectively for $s>0$. This provides further evidence for \cref{conj:acsc}.
\end{remark}
\section{Cooperations at height one}\label{sec:heightone}
In the previous section, we discussed the importance of the derived functors of the limit in the category of comodules and its relation to the chromatic splitting conjecture. In \cref{prop:inverselimascontcohom}, we gave an explicit formula for these in the form of an isomorphism
\[
\textstyle\varprojlim^{s}_{E_{*}E} E_{*} / \mfrak^{k}E_{*} \simeq \rmH^{s}_{\cts}(\mathbb{G}_{n}, E_{*}E),
\]
where on the left hand side we have the $s$-th derived functor of the limit in $E_{*}E$-comodules and the right hand side we have continuous cohomology of the Morava stabilizer group. To make use of this, we would like to understand $E_{*}E$ as a $\mathbb{G}_{n}$-representation; in this section, we will do so for $n = 1$.
\begin{warning}
Let us warn the reader again that we are interested in the uncompleted cooperations; that is, $E_{*}E = \pi_{*}(E \otimes E)$. The description of the completed cooperations
\[
E^{\vee}_{*}E := \pi_{*}(L_{K}(E \otimes E)) \simeq (E_{*}E)^{\vee}_{\mfrak}
\]
is well-known, as we have $E^{\vee}_*E \simeq \map_{\cts}(\mathbb{G}_{n}, E_{*})$ at any height by a result of Strickland \cite{strickland2000gross} or \cite{hovey_operations}.
\end{warning}
\begin{notation}
Instead of the Honda formal group law, it will be convenient to work with the multiplicative formal group law over $\mathbb{F}_{p}$, which we will denote by $\Gamma$. This is of height one, so that the Lubin--Tate ring is isomorphic to the $p$-adics, with the universal deformation the multiplicative formal group law over $\mathbb{Z}_{p}$.
By uniqueness part of the Goerss--Hopkins--Miller theorem, the Lubin--Tate spectrum associated to this universal deformation is given by
\[
E := KU^{\vee}_{p},
\]
the $p$-complete complex $K$-theory. The action of $\mathbb{G}_{1}$ is given by Adams operations.
\end{notation}
\begin{lemma}
\label{lemma:ee_at_height_one_in_relation_to_eku}
There is an isomorphism
\[
E_{*}E \simeq E_{*}KU \otimes_{\mathbb{Z}} \mathbb{Z}_{p}
\]
of $\mathbb{G}_{1}$-representations, where on the right hand side the action is trivial on the $\mathbb{Z}_{p}$-factor.
\end{lemma}
\begin{proof}
We have $KU_{*} \simeq \mathbb{Z}[u^{\pm 1}]$ and $E_{*} \simeq KU_{*} \otimes \mathbb{Z}_{p}$, which is flat over $KU_*$. It then follows from flatness that we have an isomorphism
\[
E_{*}E \simeq E_{*}KU \otimes_{\mathbb{Z}} \mathbb{Z}_{p}
\]
induced by the multiplication map
\[
(E \otimes KU) \otimes_{KU} E \rightarrow E \otimes E
\]
Since this multiplication map is $\mathbb{G}_{1}$-equivariant with respect to the action on the left $E$ factor, this is in fact an isomorphism of representations.
\end{proof}
\begin{corollary}
\label{corollary:coh_of_ee_is_coh_of_eku_otimes_zp}
There is an isomorphism
\[
\rmH^{s}_{\cts}(\mathbb{G}_{1}, E_{*}E) \simeq \rmH^{s}_{\cts}(\mathbb{G}_{1}, E_{*}KU) \otimes_{\mathbb{Z}} \mathbb{Z}_{p}.
\]
\end{corollary}
\begin{proof}
Let us write $\rmH^{*}(M) := \rmH^{*}_{\cts}(\mathbb{G}_{1}, M)$ for brevity. As $\mathbb{Z}_{p}$ is flat, it can be written as a filtered colimit $\mathbb{Z}_{p} \simeq F_{\alpha}$ of finitely generated, free abelian groups by Lazard's theorem. We then have
\[
\rmH^{s}(E_{*}E) \simeq \rmH^{s}(E_{*}KU \otimes_{\mathbb{Z}} \mathbb{Z}_{p}) \simeq \rmH^{s}(\varinjlim E_{*}KU \otimes_{\mathbb{Z}} F_{\alpha}) \simeq \varinjlim \rmH^{s}(E_{*}KU \otimes_{\mathbb{Z}} F_{\alpha}),
\]
where the last equivalence used \cref{corollary:continuous_cohomology_commutes_with_filtered_colimits}, and further
\[
\varinjlim \rmH^{s}(E_{*}KU \otimes_{\mathbb{Z}} F_{\alpha}) \simeq \varinjlim \rmH^{s}(E_{*}KU) \otimes_{\mathbb{Z}} F_{\alpha} \simeq \rmH^{s}(E_{*}KU) \otimes_{\mathbb{Z}} \mathbb{Z}_{p},
\]
where the first equivalence used additivity of cohomology.
\end{proof}
It follows from \cref{corollary:coh_of_ee_is_coh_of_eku_otimes_zp} that to understand the cohomology with coefficients in $E_{*}E$, it is enough to understand $E_{*}KU$.
As we are working with coefficients in the homology of a spectrum rather than an arbitrary $\mathbb{G}_{1}$-module, we have a canonical family of invariants. To be more precise, we have a Hurewicz homomorphism
\[
KU_{*} \rightarrow E_{*}KU
\]
whose image necessarily consists of $\mathbb{G}_{1}$-invariants. It follows that the cohomology groups
\[
\rmH^{*}_{\cts}(\mathbb{G}_{1}, E_{*}KU)
\]
are $2$-periodic. As they're also even, as $E_{*}KU$ is, it is enough to determine the cohomology with coefficients in $E_{0}KU$.
\begin{lemma}
\label{lemma:completion_map_on_e0ku_is_injective}
The composite
\[
E_{0}KU \rightarrow E_{0}^{\vee}KU \simeq E_{0}^{\vee}E \simeq \map_{\cts}(\mathbb{G}_{1}, \mathbb{Z}_{p})
\]
is injective.
\end{lemma}
\begin{proof}
Note that by Landweber exactness we have
\[
E_{0}KU \simeq \mathbb{Z}_{p} \otimes _{\mathbb{Z}} KU_{0}KU.
\]
This is flat over $\mathbb{Z}_{p}$ and we deduce from \cref{proposition:collapse_of_k_based_ass_for_flat_e_modules} that the above composite can be identified with a $p$-completion map. By a result of Adams, Harris, and Switzer, $KU_{0}KU$ is free as an abelian group \cite{adams1971hopf}. It follows that the $p$-completion map
\[
\mathbb{Z}_{p} \otimes_{\mathbb{Z}} KU_{0}KU \rightarrow (\mathbb{Z}_{p} \otimes_{\mathbb{Z}} KU_{0}KU)^{\vee}_{p} \simeq (KU_{0}KU)^{\vee}_{p} \simeq E_{0}^{\vee}E
\]
is injective as needed.
\end{proof}
The completion map of \cref{lemma:completion_map_on_e0ku_is_injective} is $\mathbb{G}_{1}$-equivariant, and we deduce that $E_{0}KU$ is naturally a subrepresentation of $\map_{\cts}(\mathbb{G}_{1}, \mathbb{Z}_{p})$. On the latter, $\mathbb{G}_{1}$ acts by right multiplication on the source. Our goal is to describe the given subrepresentation.
By unwrapping the definitions, the map of \cref{lemma:completion_map_on_e0ku_is_injective} can be described explicitly. By standard results about weakly periodic Landweber exact homology theories, $KU_{0}KU$ is the algebra classifying automorphisms of the multiplicative formal group law. If $g \in \mathbb{G}_{1}$ is an automorphisms over the $p$-adics, then we can write
\[
g(x) = \sum_{i} b_{i}(g) x^{i},
\]
where the $b_{i}$ depend continuously on $g$. The above expression defines an automorphism of the multiplicative formal group law over $\map_{\cts}(\mathbb{G}_{1}, \mathbb{Z}_{p})$; this is the automorphism classified by the injective map
\[
KU_{0}KU \rightarrow \map_{\cts}(\mathbb{G}_{1}, \mathbb{Z}_{p}).
\]
Combined with \cref{lemma:completion_map_on_e0ku_is_injective}, this proves the following.
\begin{proposition}
\label{proposition:e0ku_as_a_subalgebra_generated_by_certain_functions}
The ring $E_{0}KU$ is the $\mathbb{Z}_{p}$-subalgebra of $\map_{\cts}(\mathbb{G}_{1}, \mathbb{Z}_{p})$ generated by the functions $b_{i}\colon \mathbb{G}_{1} \rightarrow \mathbb{Z}_{p}$ for $i \geq 1$ as well as $b_{1}^{-1}$.
\end{proposition}
As we work with the multiplicative formal group law, these functions $b_{i}$ can be made explicit. Assume first that $a \in \mathbb{Z}$, so that the corresponding multiple of the identity of the multiplicative formal group law is given by
\[
[a](x) = (1+x)^{a}-1 = \sum^{\infty}_{k > 0} {a \choose k} x^{k}.
\]
The last expression makes sense even if $a$ is a $p$-adic integer, where we take
\[
{a \choose k} = \frac{a(a-1)\ldots(a-k+1)}{k!},
\]
and the sum is now possibly infinite. One can check that these are automorphisms of the multiplicative formal group law even if $a$ is $p$-adic, so that the identification
\[
\mathbb{Z}_{p}^{\times} \simeq \mathbb{G}_{1}
\]
is given by $a \mapsto [a](x) = \sum^{\infty}_{k > 0} {a \choose k} x^{k}$. Thus, \cref{proposition:e0ku_as_a_subalgebra_generated_by_certain_functions} can be stated in the following more explicit form.
\begin{proposition}
\label{proposition:e0ku_as_subalgebra_generated_by_binomial_coefficients}
The algebra $E_{0}KU$ is the $\mathbb{Z}_{p}$-subalgebra of $\map_{\cts}(\mathbb{G}_{1}, \mathbb{Z}_{p})$ generated by the functions $b_{i}(a) = {a \choose i}$ for $i \geq 1$ as well as $b_{1}^{-1}(a) = \frac{1}{a}$.
\end{proposition}
\begin{remark}
The functions $b_{i}$ appearing in \cref{proposition:e0ku_as_subalgebra_generated_by_binomial_coefficients} are well-known to $p$-adic analysts. By a fundamental theorem of Mahler, any continuous $p$-adic function on $\mathbb{Z}_{p}$ can be written as an infinite convergent sum of the $b_{i}$-s in a unique way \cite{mahler1958interpolation}.
\end{remark}
This immediately implies the following calculation.
\begin{proposition}
\label{proposition:h0_of_eku}
We have $\rmH^{0}_{\cts}(\mathbb{G}_{1}, E_{0}KU) \simeq \mathbb{Z}_{p}$.
\end{proposition}
\begin{proof}
This is immediate from \cref{proposition:e0ku_as_subalgebra_generated_by_binomial_coefficients}, as the only functions $\mathbb{G}_{1} \rightarrow \mathbb{Z}_{p}$ invariant under translation are the constants.
\end{proof}
To compute higher cohomology groups, it is convenient to observe that they are necessarily rational.
\begin{lemma}
\label{lemma:higher_coh_groups_of_eku_are_rational}
The groups $\rmH^{i}_{\cts}(\mathbb{G}_{1}, E_{0}KU)$ are rational vector spaces for $i > 0$.
\end{lemma}
\begin{proof}
Since the inclusion $E_{0}KU \hookrightarrow \map_{\cts}(\mathbb{G}_{1}, \mathbb{Z}_{p})$ is an isomorphism after $p$-completion, as we observed in the proof of \cref{lemma:completion_map_on_e0ku_is_injective}, we deduce that it is an isomorphism modulo $p$. It follows that
\[
\rmH^{i}_{\cts}(\mathbb{G}_{1}, E_{0}KU \otimes_{\mathbb{Z}_{p}} \mathbb{F}_{p}) \simeq \rmH^{i}_{\cts}(\mathbb{G}_{1}, \map_{\cts}(\mathbb{G}_{1}, \mathbb{F}_{p}))
\]
and the right hand side vanishes for $i > 0$. The long exact sequence of cohomology associated to
\[
0 \rightarrow E_{0}KU \rightarrow E_{0}KU \rightarrow E_{0}KU \otimes_{\mathbb{Z}_{p}} \mathbb{F}_{p} \rightarrow 0,
\]
where the first map is multiplication by $p$, gives the required result.
\end{proof}
\begin{proposition}
\label{proposition:h1_coh_of_eku}
We have
\[
\rmH^{1}_{\cts}(\mathbb{G}_{1}, E_{0}KU) \simeq \mathbb{Q}_{p}
\]
and the higher cohomology groups vanish.
\end{proposition}
\begin{proof}
By \cref{lemma:higher_coh_groups_of_eku_are_rational}, all cohomology groups of positive degree are rational, so that we can replace $E_{0}KU$ by $\mathbb{Q}_{p} \otimes_{\mathbb{Z}_{p}} E_{0}KU$ without changing these. The cohomology with the coefficients in the latter can be computed directly.
As each binomial coefficient $b_{i}(a) = {a \choose i}$ is a polynomial in $a$ with rational coefficients, we deduce from \cref{proposition:e0ku_as_subalgebra_generated_by_binomial_coefficients} that $\mathbb{Q}_{p} \otimes_{\mathbb{Z}_{p}} E_{0}KU$ can be identified with the $\mathbb{Q}_{p}$-subalgebra of $\map_{\cts}(\mathbb{G}_{1}, \mathbb{Q}_{p})$ generated by $b_{1}(a) = a$ and its inverse.
It follows that there is an isomorphism of representations
\[
\mathbb{Q}_{p} \otimes_{\mathbb{Z}_{p}} E_{0}KU \simeq \mathbb{Q}_{p}[b^{\pm 1}],
\]
where $b = b_{1}$ and the action is determined by $g \cdot b = gb$ for $g \in \mathbb{G}_{1}$. This decomposes as a direct sum of character representations
\[
\mathbb{Q}_{p}[b^{\pm 1}] \simeq \bigoplus _{k \in \mathbb{Z}} \mathbb{Q}_{p}(k),
\]
where we identify $\mathbb{Q}_{p}(k)$ with the subrepresentation generated by $b^{k}$. It follows that
\[
\rmH^{i}_{\cts}(\mathbb{G}_{1}, \mathbb{Q}_{p}[b^{\pm 1}]) \simeq \bigoplus_{k \in \mathbb{Z}} \rmH^{i}_{\cts}(\mathbb{G}_{1}, \mathbb{Q}_{p}(k)).
\]
Let $U \subseteq \mathbb{G}_{1} \simeq \mathbb{Z}_{p}^{\times}$ be a normal subgroup of finite index such that $U \simeq \mathbb{Z}_{p}$. As finite groups have no rational cohomology, Lyndon--Hochschild--Serre spectral sequence collapses and gives an isomorphism
\[
\rmH^{i}(\mathbb{G}_{1}, \mathbb{Q}_{p}(k)) \simeq \rmH^{i}(U, \mathbb{Q}_{p}(k))^{\mathbb{G}_{1}/U}.
\]
It follows that the cohomology groups vanish above degree $1$, as $U \simeq \mathbb{Z}_{p}$ is of cohomological dimension one.
As $U \simeq \mathbb{Z}_{p}$ is cyclic as a profinite group, its cohomology can be computed using a two-step cochain complex, as we did in \cref{sec:somecomputations}. It follows that the only one-dimensional representation of $U$ over $\mathbb{Q}_{p}$ with non-trivial cohomology is the trivial one. As the characters $\mathbb{Q}_{p}(k)$ restrict to non-trivial representations for $k \neq 0$, we deduce that
\[
\rmH^{1}_{\cts}(\mathbb{G}_{1}, \mathbb{Q}_{p}[b^{\pm 1}]) \simeq \rmH^{1}_{\cts}(\mathbb{G}_{1}, \mathbb{Q}_{p}) \simeq \mathbb{Q}_{p}
\]
as claimed.
\end{proof}
\begin{theorem}
\label{theorem:cohomology_of_ee_at_height_one}
At height $n=1$ and any prime, we have
\[
\rmH_{\cts}^{s}(\mathbb{G}_{1}, E_{*}E) \simeq \begin{cases}
E_{*} \otimes_{\mathbb{Z}} \mathbb{Z}_{p} & \mbox{when } $s=0$;\\
E_{*} \otimes_{\mathbb{Z}} \mathbb{Q}_{p}& \mbox{when } $s=1$;\\
0 & \mbox{otherwise}.
\end{cases}
\]
\end{theorem}
\begin{proof}
As we observed previously that the cohomology with coefficients in $E_{*}E$ is concentrated in even internal degrees and $2$-periodic, it is enough to establish this in internal degree zero, where this is a combination of \cref{corollary:coh_of_ee_is_coh_of_eku_otimes_zp}, \cref{proposition:h0_of_eku} and \cref{proposition:h1_coh_of_eku}.
\end{proof}
\begin{remark}
As we observed at the beginning, the cohomology groups of \cref{theorem:cohomology_of_ee_at_height_one} coincide with derived functors of completion in comodules by \cref{prop:inverselimascontcohom}. Using the latter perspective, a sketch of calculation of the rationalization of the above groups at $p > 2$ appears in a talk by Hopkins, recorded in \cite[Talk 14]{report_on_e_theory_conjectures}. We were independently informed by Hopkins that he first made this calculation in the 1990s and that it was part of the motivation towards the algebraic chromatic splitting conjecture.
\end{remark}
\addtocontents{toc}{\vspace{5mm}}
|
\section{Introduction}
\subsection{Background}
\IEEEPARstart{T}{HE EMERGING} implementation of information techniques has reformed the power gird into a complex cyber-physical power system (CPPS), where the two-way real-time communication among multiple parties raises new risks in the grid
\cite{ten2008vulnerability}. Musleh \textit{et al.} \cite{musleh2019survey} reviewed seven recent cyber attacks in energy industry and spotted the related vulnerabilities in both physical and cyber layers. Recently, false data injection (FDI) attacks have been developed against power system state estimation (SE) by intruding through the Modbus/TCP protocol without being noticed by the bad data detector (BDD) at the control centre \cite{liu2011false,hug2012vulnerability, rahman2013false, ahmed2019unsupervised}. As the accurate state estimation is crucial for energy management system (EMS) activities, such as power system dispatch, contingency analysis, and fault diagnosis, states falsified by the FDI attacks can result in erroneous control action, causing economic losses, system instability, and safety violation \cite{gomez2018electric, tajer2017false, xie2011integrity}.
As the power system operates quasi-statically, the intruders have sufficient time to learn the system parameters and preparing the FDI attacks \cite{kim2014subspace, yu2015blind,lakshminarayana2021data}. As a result, it is crucial to invalidate the attacker's knowledge by proactively changing the system configuration. Moving target defence (MTD), which is firstly conceptualized for the information technology security, utilizes this proactive defence idea \cite{cho2020toward}. With the distributed flexible AC transmission system (D-FACTS) devices, the control centre can alter the reactances of the transmission lines to physically change the system parameters that
are unknown to the attackers.
\subsection{Related Work}
Initially, researches on MTD involve using random placement and reactance perturbations to expose FDI attacks \cite{morrow2012topology, davis2012power, rahman2014moving}. However, it has been shown that the so-called `naive' applications cannot guarantee an effective detection on stealthy FDI attacks. Therefore, \cite{liu2018reactance} and \cite{zhang2019analysis} demonstrate that the MTD effectiveness is dependent on the rank of the composite pre- and post- MTD measurement matrices. Furthermore, Liu, \textit{et al.} \cite{liu2020optimal} and Zhang \textit{et al} \cite{zhang2021strategic} investigate the D-FACTS devices placement in the planning stage to maximize the effectiveness while minimising the investment budget. The authors in \cite{lakshminarayana2021cost} analyze the MTD effectiveness using the metric of minimal principal angle and numerically show the relationship between the angle and the average detection rate, which can be used to design the MTD. Liu, \textit{et al.} \cite{liu2019joint} extends the MTD strategy in \cite{liu2018reactance} with sensor protections and Tian, \textit{et al.} \cite{tian2019moving} applies MTD to detect Stuxnet-like attack. Moreover, Higgins \textit{et.al.} \cite{higgins2020stealthy} suggests to perturb the reactance through Gaussian watermarking to prevent the attacker from inferring the new system parameters. However, majority of the above literature studies the effectiveness of MTD under DC and noiseless assumptions. As demonstrated by \cite{li2019feasibility}, the MTD detection rate is limited by the ratio between the attack strength and the noise level. As a result, there is no guarantee on the detection performance of existing MTD strategies against the unseen attacks in a noisy environment.
\subsection{Contributions}
With the attackers becoming more resourceful and intelligent, it becomes harder to anticipate the attackers' strategy. Therefore, it is critical for the system operator to determine and guarantee the lowest detection rate of MTD against all unknown attacks. In this context, this paper introduces the concept of \textit{robust MTD}, which aims to guarantee the worst-case MTD effectiveness against a given level of attack strength under noisy environment. The main contributions of this paper are summarized as follows.
\begin{itemize}
\item
This paper, for the first time, proposes the concept of robust MTD in a noisy environment. We theoretically proves that, for any given grid topology and MTD strategy, the minimal principal angle between the pre- and post-MTD Jacobian subspaces is directly linked with the worst-case performance against all potential attacks, which can be used as a new metric to represent the MTD effectiveness.
\item A novel MTD design algorithm is formulated to improve the worst-case detection rate by maximizing the minimal principal angle under the complete grid configuration. We then demonstrate that the worst-case detection rate of the grid with incomplete configuration cannot be improved. Therefore, an iterative algorithm is formulated to maximize the minimal nonzero principal angle while limits the chance of attacking on the subspace that cannot be detected.
\item Numerical simulations on IEEE case-6, 14, and 57 systems demonstrate the improved detection performance of the robust MTD algorithms against the worst-case, random, and single-state attacks, under both simplified and full AC models.
\end{itemize}
The rest of the paper is organized as follows. The preliminaries are summarised in Section \ref{sec:preliminary}; Analysis on MTD effectiveness is presented in Section \ref{sec:analysis_mtd_eff}; Problem formulation and the proposed robust algorithms are presented in Section \ref{sec:robust mtd algorithm};
Case studies are given in Section \ref{sec:simulation} with conclusions in Section \ref{sec:conclusion}.
\section{Preliminaries}\label{sec:preliminary}
\subsection{Notations}
In this paper, vectors and matrices are represented by bold lowercase and uppercase letters, respectively. The $p$-norm for $\bm{a}$ is written as $\|\bm{a}\|_p$. The column space of $\bm{A}$ is $\mathcal{A}=\text{Col}(\bm{A})$.
$\bm{P}_{\bm{A}} = \bm{A}(\bm{A}^T\bm{A})^{-1}\bm{A}^T$ represents the orthogonal projector to $\text{Col}(\bm{A})$ while $\bm{S}_{\bm{A}} = \bm{I} - \bm{P}_{\bm{A}}$ represents the orthogonal projector to
$\text{Kel}(\bm{A}^T)$. The set of singular values is $\sigma(\bm{A}) = \{\sigma_1(\bm{A}), \sigma_2(\bm{A}),\dots,\sigma_{\min\{m,n\}}(\bm{A})\}$. The spectral norm is $\|\bm{A}\|=\max_i \sigma_i(\bm{A})$ and the Frobenius norm is $\|\cdot\|_F$. We use $(\cdot)'$ symbol to indicate the quantities after MTD and $(\cdot)_a$ to indicate the quantities after the attack. Matrix operator $\circ$ represents Hadamard product. Other symbols and operators are defined in the paper whenever proper.
\subsection{System Model and State Estimation}
\label{sec:system_model}
\begin{figure}[]
\centering
\includegraphics[width=0.42\textwidth]{figure/EMS.pdf}
\caption{EMS with injection attacks and MTD in CPPS.}
\label{fig:ems}
\end{figure}
In this paper, the power system is modelled as a graph $\mathcal{G}(\mathcal{N},\mathcal{E})$ with $|\mathcal{N}| = n+1$ number of buses and $|\mathcal{E}| = m$ number of branches. For each bus, we denote its complex voltage as ${\bm{\nu}}=\bm{v}\angle\bm{\theta}$; and for each branch, we denote the admittance as $\bm{y}=\bm{g}+j\bm{b}$. The power system can be modelled by the nonlinear equations $\bm{z}=\bm{h}(\bm{\nu})+\bm{e}$ where $\bm{z}\in\mathbb{R}^{p}$ is the sensor measurement; $\bm{h}(\cdot)\in\mathbb{R}^{p}$ is the power balancing equation; $\bm{\nu}\in\mathbb{R}^{2(n+1)}$ is the system state composing of voltage magnitude and phase angle. The measurement noise vector $\bm{e}\in\mathcal{N}(\bm{0},\bm{R})$ follows independent Gaussian distribution with diagonal covariance matrix $\bm{R}=\text{diag}([\sigma_1^2,\sigma_2^2,\cdots,\sigma_p^2])$. In detail, $\bm{h}(\cdot)$ is considered as \cite{gomez2018electric}:
\begin{subequations}\label{eq:power_balance}
\begin{equation}
P_{i}=v_{i} \sum_{j=1}^{n} v_{j}\left(g_{i j} \cos \theta_{i j}+b_{i j} \sin \theta_{i j}\right)
\end{equation}
\begin{equation}
Q_{i}=v_{i} \sum_{j=1}^{n} v_{j}\left(g_{i j} \sin \theta_{i j}-b_{i j} \cos \theta_{i j}\right)
\end{equation}
\begin{equation}
P_{k: i \to j}=v_{i} v_{j}\left(g_{i j} \cos \theta_{i j}+b_{i j} \sin \theta_{i j}\right)-g_{i j} v_{i}^{2}
\end{equation}
\begin{equation}
Q_{k: i \to j}=v_{i} v_{j}\left(g_{i j} \sin \theta_{i j}-b_{i j} \cos \theta_{i j}\right)+ b_{i j}v_{i}^{2}
\end{equation}
\end{subequations}
where $P_i$ and $Q_i$ are the active and reactive power injections at bus $i$; and $P_{k:i\to j}$ and $Q_{k: i\to j}$ are the $k$-th active and reactive power flows from bus $i$ to $j$; $\theta_{ij} = \theta_i - \theta_j$ is the phase angle difference between bus $i$ and $j$;
As shown in Fig. \ref{fig:ems},
the control centre is equipped with state estimation (SE) serving as a bridge between remote terminal units (RTUs) and energy management system (EMS) \cite{gomez2018electric}. Given the measurements, the AC-SE is solved by the following weighted least-square problem using iterative algorithm, such as Gauss-Newton method \cite{abur2004power}:
\begin{equation}\label{eq:ac_se}
\min _{\hat{\bm{\nu}}} J(\hat{\bm{\nu}})=(\bm{z}-\bm{h}(\hat{\bm{\nu}}))^{T} \cdot \bm{R}^{-1} \cdot(\bm{z}-\bm{h}(\hat{\bm{\nu}}))
\end{equation}
where $\hat{\bm{\nu}}$ is the estimated state. Furthermore, the bad data detection (BDD) at the control centre detects any measurement error that violates a Gaussian prior. Given $\bm{\hat{\nu}}$, the residual vector is calculated as $\bm{r} = \bm{z}-\bm{h}(\bm{\hat{\nu}})$ and the residual is represented as $\gamma(\bm{z}) = \|\bm{R}^{-\frac{1}{2}}\bm{r}\|_2^2$. Let $\bm{e}$ be the random variable; then $\gamma$ approximately follows $\chi^2$ distribution with degree of freedom (DoF) $p-2n$ \cite{abur2004power}. The threshold $\tau_\chi(\alpha)$ of the $\chi^2$ detector can be defined probabilistically based on the desired False Positive Rate (FPR) $\alpha\in(0,1)$ by the system operator \cite{abur2004power}:
\begin{equation}\label{eq:chi_square_distribution}
\int_{\tau_\chi(\alpha)}^\infty g(u)du = \alpha
\end{equation}
where $g(u)$ is the p.d.f of the $\chi^2$ distribution and $\alpha$ is usually set as 1\%-5\%.
\subsection{Attack Assumptions}
\label{sec:fdia}
With the emerging implementation of information and communication techniques, standard protocols, such as Modbus can be vulnerable to FDI attacks. It has been shown that an FDI attack $\bm{z}_a=\bm{z}+\bm{a}$ can pass the BDD if $\bm{a} = \bm{h}({\bm{\nu}}+\bm{c}) - \bm{h}({\bm{\nu}})$ where $\bm{c}$ is the attack vector on the state vector.
In this case, the contaminated measurement becomes $\bm{z}_a=\bm{h}(\bm{\nu}+\bm{c})+\bm{e}$ whose residual follows the same $\chi^2$ distribution as the legit measurement $\bm{z}$.
It also implies that, an attacker against AC state estimation has to know the exact formulation of $\bm{h}(\cdot)$ and the value of the state $\bm{\nu}$ to avoid being detected. To successfully launch such attack, we assume the attacker's abilities as follows:
\textbf{\textit{Assumption 1}}: The attackers can access all the measurements and are aware of the admittance and topology of the grid to build $\bm{h}(\cdot)$. The exfiltration can be achieved by data-driven algorithms \cite{zhang2020topology, kim2014subspace, yu2015blind, lakshminarayana2021data}. However, the duration of data collection is much longer than a single state estimation time, implying that the attacker cannot immediately know the exact value of reactance changes \cite{lakshminarayana2021cost}. Meanwhile, the attackers are also aware of the exact state or state estimation from the previous measurements \cite{hug2012vulnerability, rahman2013false}.
\textbf{\textit{Assumption 2}}: The attackers can modify or replace all the eavesdropped measurements to achieve the attack purpose. However, as large instant measurement change may violate the temporal trends of the grid measurements and be detected \cite{zhao2017short, xu2020deep,liu2019joint, lakshminarayana2021cost}, the attack strength $\|\bm{a}\|_2$ is assumed to be limited.
Assumption 1-2 require the attacker's efforts to gain sufficient knowledge on the grid topology and operational conditions which may not be easy in practice. However, we assume a strong attack ability and study the defence algorithm against the most general and unpredictable FDI attacks.
\subsection{Moving Target Defence}
By using the D-FACTS devices, the system operator can proactively change the reactances to keep invaliding the attacker's knowledge on $\bm{h}(\cdot)$:
\begin{equation*}
\bm{h}_{\bm{x}}(\cdot)\xrightarrow{\text{D-FACTS}} \bm{h}_{\bm{x}'}(\cdot)
\end{equation*}
where $\bm{x}' = \bm{x} + \Delta \bm{x}$ is the reactance after triggering the D-FACTS devices. As illustrated by Fig. \ref{fig:ems}, the channels of D-FACTS devices are encrypted and MTD is implemented with period shorter than the reconnaissance time of the attacker (see Assumption 1). In addition, the reactance changed by the D-FACTS devices are limited physically:
\begin{subequations}
\begin{equation}\label{eq:reactance_limit_in}
-\tau \bm{x}_i \leq \Delta \bm{x}_i \leq \tau \bm{x}_i,\quad i\in\mathcal{E}_D
\end{equation}
\begin{equation}\label{eq:reactance_limit_out}
\Delta \bm{x}_i = 0, \quad i\in\mathcal{E}\setminus\mathcal{E}_D
\end{equation}
\end{subequations}
where $\tau$ represents the maximum perturbation ratio of D-FACTS devices. Typical values of $\tau$ are reported as $20\%-50\%$ in literature \cite{liu2018reactance,liu2020optimal,zhang2019analysis,lakshminarayana2021cost}; $\mathcal{E}_D$ represents the branch set equipped with the D-FACTS devices.
After implementing MTD, the residual vector becomes $\bm{r}_a' = \bm{h}'(\bm{x}) + \bm{h}(\bm{x}+\bm{c}) - \bm{h}(x) + \bm{e}$ which may not follow the $\chi^2$ distribution of the legit measurement any more and hence trigger the bad data detector.
\subsection{Model Simplification for MTD Design}
To design the MTD against FDI attacks, most of the literature relies on either DC or simplified AC power system models \cite{liu2018reactance, zhang2019analysis, liu2020optimal, lakshminarayana2021cost, zhang2021strategic, li2019feasibility, liu2019joint} and then verifies the performance on the full AC model. Here we adopt the simplified AC model based on the linearized measurement equation. Compared with the DC model, the simplified AC model can reflect on different state values with the branch resistance also considered.
In detail, the first-order Taylor expansion can be established around a stationary state $\bm{\nu}_0$:
\begin{equation}\label{eq:linear_equation}
\bm{z} = \bm{h}(\bm{\nu}_0) + \bm{J}_{\bm{\nu}_0}(\bm{\nu}-\bm{\nu}_0) + \bm{e}
\end{equation}
where the Jacobian matrix of $\bm{h}(\cdot)$ is found with respect to $\bm{\nu}_0$ as $\bm{J}_{\bm{\nu}_0} = \left[\left.\frac{\partial \bm{h}_k}{\partial \bm{\nu}_i} \right\vert_{\bm{\nu}=\bm{\nu}_0} \right]_{i,k}$. The state $\bm{\nu}_0$ can be simulated from security constrained AC-OPF \cite{gomez2018electric} around the estimated active and reactive loads before the real-time operation. Alternatively, the states estimated from the previous measurements or a flat state \cite{liu2019joint, liu2021optimalcoding} can also be used. Following the recent literature on MTD \cite{liu2018reactance,liu2020optimal,liu2019joint},
we consider the FDI attacks on the voltage phase angle and derive the defence strategies according to the power flow measurements at each branch. Therefore, the Jacobian matrix is considered as:
\begin{equation}\label{eq:jacobian}
\left[\left.\frac{\partial \bm{P}_{k:i\to j}}{\partial \bm{\theta}_i} \right\vert_{\bm{\theta}=\bm{\theta}_0} \right]_{k} = -\bm{V}\cdot\bm{G}\cdot\bm{A}_r^{\sin}+\bm{V}\cdot\bm{B}\cdot\bm{A}_r^{\cos}
\end{equation}
where $\bm{V} = \text{diag}\left((\bm{C}_f\bm{v})\circ(\bm{C}_t\bm{v})\right)$; $\bm{G} = \text{diag}(\bm{g})$; $\bm{B} = \text{diag}(\bm{b})$; $\bm{A}_{r}^{\sin} = \text{diag}(\sin{\bm{A\theta_0}})\bm{A}_r$; and $\bm{A}_{r}^{\cos} = \text{diag}(\cos{\bm{A\theta_0}})\bm{A}_r$. Moreover, $\bm{C}_f$ and $\bm{C}_t$ are the `from' and `to' -side incidence matrices; $\bm{A}_r$ is the reduced incidence matrix by removing the column representing the reference bus from incidence matrix $\bm{A}$. To simplify the notation, we omit the subscript $\bm{\theta}_0$ in the following discussion.
According to Assumption 2, as the attack strength is limited, the attack vector can also be linearized around $\bm{\theta}_0$ as \cite{liu2019joint}:
\begin{equation}\label{eq:linear_attack}
\bm{a} = \bm{h}(\bm{\theta}_0+\bm{c}) - \bm{h}(\bm{\theta}_0) = \bm{J}\bm{c}
\end{equation}
We design the MTD algorithm based on the simplified AC model \eqref{eq:linear_equation}-\eqref{eq:linear_attack} using active power flow measurements. The designed MTD will be then applied on the original AC model \eqref{eq:power_balance}-\eqref{eq:ac_se} to verify the performance.
\section{Analysis on MTD Effectiveness}
\label{sec:analysis_mtd_eff}
In this section, we firstly extend the concept of complete MTD in the literature from DC model to simplified AC model. We then define the MTD effectiveness in a probabilistic way and illustrate the need for new metric on effective MTD design in a noisy environment.
\subsection{Complete MTD}
Let $\bm{H}$ and $\bm{H}'$ be the DC measurement matrices. Under the noiseless condition, the \textit{complete MTD} can be designed to detect any FDI attack by keeping the composite matrix $[\bm{H},\bm{H}']$ full column rank \cite{liu2018reactance, zhang2019analysis, liu2020optimal, zhang2021strategic}. If the full rank condition cannot be achieved due to the sparse grid topology (e.g. $m<2n$) or limited D-FACTs devices, a \textit{max-rank incomplete MTD} can be designed to minimize the attack space. As the rank of the composite matrix is maximized in both complete and incomplete conditions, we refer the MTD strategies in \cite{liu2018reactance, zhang2019analysis, liu2020optimal, zhang2021strategic} as \textit{max-rank MTD}.
To help better define the problem, we extend the concept of complete and incomplete MTDs from the DC model to the simplified AC models in the following proposition:
\begin{proposition}\label{prop:com_incom}
The power system modelled by \eqref{eq:linear_equation} is with complete configuration against the FDI attack modelled by \eqref{eq:linear_attack} only if $m\geq 2n$.
\end{proposition}
\begin{proof}
Please refer to Appendix \ref{sec:proof_com_incom}
\end{proof}
As stated by Proposition \ref{prop:com_incom}, to have complete configuration $\text{rank}([\bm{J}_N,\bm{J}_N']) = 2n$, the number of branches should be at least one time larger than the number of buses. In addition, the max-rank incomplete MTD with $\text{rank}([\bm{J}_N,\bm{J}_N']) = m$ can be designed for the grid with incomplete configuration. In the following discussions, we refer the grid that can achieve complete MTD under certain topology and D-FACTS deployment as \textit{complete configuration}, otherwise as \textit{incomplete configuration}.
\subsection{$\beta$-Effective MTD}
Following the equation \eqref{eq:linear_equation}, denote $\bm{z} \triangleq \bm{z} - \bm{h}(\bm{\theta}_0)$ and $\bm{\theta} \triangleq \bm{\theta} - \bm{\theta}_0$. For the new system equation $\bm{z} = \bm{J}\bm{\theta} + \bm{e}$, the residual vector of the $\chi^2$ detector can be written as $\bm{r} = \bm{S}(\bm{J}\bm{\theta}+\bm{e}) = \bm{Se}$ where $\bm{S} = \bm{I} - \bm{J}(\bm{J}^T\bm{R}^{-1}\bm{J})^{-1}\bm{J}^T\bm{R}^{-1}$ is the weighted orthogonal projector on $\text{Ker}(\bm{J}^T)$. The residual $\gamma = \|\bm{R}^{-\frac{1}{2}}\bm{S}\bm{e}\|^2$ follows the $\chi^2$ distribution with DoF $m-n$. Referring to the simplified attack model \eqref{eq:linear_attack}, the residual vector after MTD under attack can be written as $\bm{r}_a' = \bm{S}'(\bm{J}\bm{c} + \bm{e})$ where $\bm{S}' = \bm{I} - \bm{J}'(\bm{J}'^T\bm{R}^{-1}\bm{J}')^{-1}\bm{J}'^T\bm{R}^{-1}$. As $\bm{a}$ is usually not in $\mathcal{J}'$ and $\bm{r}'_a$ is biased from zero,
the residual $\gamma_a' = \|\bm{R}^{-\frac{1}{2}}\bm{S}'(\bm{Jc} + \bm{e})\|^2$ follows the non-central $\chi^2$ distribution, i.e. $\gamma_a' \sim \chi_{m-n}^2(\lambda)$ with non-centrality parameter $\lambda = \|\bm{R}^{-\frac{1}{2}}{\bm{S}}'\bm{Jc}\|_2^2$ \cite{krishnamoorthy2006handbook}. Meanwhile, the mean and variance of the distribution are given as $\mathbf{E}(\gamma_a') = m-n +\lambda$ and $\mathbf{Var}(\gamma_a') = 2(m-n+2\lambda)$, respectively. It is clear that when a noisy environment is considered, the deterministic criteria can not be used any more to describe the effectiveness of MTD. A probabilistic criteria is hence defined below. For clear presentation, the matrices are normalized with respect to the measurement noises (see Appendix \ref{sec:app_normalize}).
Followed by \eqref{eq:chi_square_distribution}, for any given attack vector $\bm{a}$, we define an MTD as $\beta$-\textit{effective} ($\beta$-MTD in short) if the following inequality is satisfied:
\begin{equation}\label{eq:non_central_chi_square_distribution}
f(\lambda) = \int_{\tau_\chi(\alpha)}^\infty g_\lambda(u)du \geq \beta
\end{equation}
where $g_\lambda(u)$ is the p.d.f. of non-central $\chi^2$ distribution. When $\lambda$ increases from 0, the detection probability on $\bm{a}$ also increases as the mean and variance increase \cite{teixeira2010cyber}. Therefore, for a given $\beta$, there exists a minimum $\lambda$ such that \eqref{eq:non_central_chi_square_distribution} is satisfied, which is defined as critical $\lambda$ and denoted as $\lambda_c(\beta)$.
Consequently, the rank conditions in \cite{liu2018reactance, zhang2019analysis, liu2020optimal, zhang2021strategic, liu2019joint} cannot guarantee the detection performance as they are not directly linked with the increase of $\lambda$ to have $\beta$-MTD. Fig. \ref{fig:residual_illustration} illustrates the c.d.f. of $\gamma'$ on a random FDI attack using max-rank MTDs in a case-14 system. Without using MTD, the detection rate is 5\% which is consistent with the FPR. To have high detection rate, e.g. $\beta=95\%$, it is desirable to sufficiently shift the distribution as shown by the blue curve. For the max-rank MTDs that satisfy max-rank condition, they can shift the c.d.f. positively, but there is no guarantee on how much of such shift can be achieved and whether it
leads to the desired detection rates. This finding clearly calls for new criterion to guide the design of MTD in a noisy environment.
Moreover, as numerically shown by \cite{li2019feasibility}, not all attacks can be detected by the MTD with high detection rate. Therefore, we theoretically introduce the following necessary condition to have $\beta$-MTD which can be seen as the limitation of MTD against FDI attacks with small attack strength.
\begin{proposition}\label{prop:min_att}
An MTD is $\beta$-effective only if $\|\bm{a}_N\|_2\geq \sqrt{\lambda_c(\beta)}$.
\end{proposition}
\begin{proof}
Please refer to Appendix \ref{sec:app_min_att}.
\end{proof}
Proposition \ref{prop:min_att} can be further analyzed on $\bm{a}$ to have $\|\bm{a}\|_2\geq \sigma_{min}\sqrt{\lambda_c(\beta)}$ with $\sigma_{min} = \min_i\{\sigma_1,\sigma_2,\dots,\sigma_m\}$. This implies that $\beta$-MTD can be achieved only if the ratio between attack strength and measurement noise is higher than a certain value, which verifies the numerical results in \cite{li2019feasibility}.
\begin{figure}[]
\centering
\includegraphics[width=0.3\textwidth]{figure/case14_dc_residual_illustration.pdf}
\caption{Illustration of attack detection probability on IEEE case-14 system based on simplified AC model \eqref{eq:linear_equation}-\eqref{eq:linear_attack}. The more positively the c.d.f. is shifted, the higher averaged detection rate can be achieved.}
\label{fig:residual_illustration}
\end{figure}
\subsection{Max MTD}
While Proposition \ref{prop:min_att} establishes the theoretic limit on the detection probability for any given attack strength, in practise, the constraints on the D-FACTS devices \eqref{eq:reactance_limit_in}-\eqref{eq:reactance_limit_out} will further restrict such limit. In this context, the maximum detection rate on a known attack vector $\bm{a}_N$, with the limits of D-FACTS devices considered, can be found by the \textit{max-MTD} algorithm:
\begin{equation}\label{eq:max_mtd}
\begin{array}{cc}
\max_{\Delta \bm{x}} & {\|\bm{S}_N'\bm{a}_N\|_2^2} \\
\text{s.t.} & \eqref{eq:reactance_limit_in}-\eqref{eq:reactance_limit_out}
\end{array}
\end{equation}
\textit{max-MTD} can be used to evaluate the MTD effectiveness against any known attack vector. However, it is impossible to design $\Delta \bm{x}$ to achieve certain $\lambda_c(\beta)$ in advance when $\bm{a}_N$ is not known. Instead, \textit{max-MTD} can be regarded as the performance upper-bound for any MTD strategies with the same placement and perturbation limit.
\section{Robust MTD Algorithms}
In this section, we start by establishing the concept of robust MTD and its mathematical formulation. Then the robust MTD algorithms are formulated for the grid with complete and incomplete configurations, respectively.
\label{sec:robust mtd algorithm}
\subsection{Definition and Problem Formulation}
Instead of considering the average detection rate, this paper defines the robust MTD that can maximise the worst-case detection rate against all possible attacks. Firstly, we define the weakest point for a given MTD design as follows.
\begin{definition}\label{def:vul}
Given $\Delta \bm{x}$ and the corresponding pair of subspaces $(\mathcal{J}_N,\mathcal{J}_N')$, the weakest point of $(\mathcal{J}_N,\mathcal{J}_N')$ is defined as an unitary element $\bm{j}_N^*\in\mathcal{J}_N$ such that $\lambda(\Delta \bm{x}, {\bm{j}_N^*}) \leq \lambda(\Delta \bm{x},\bm{j}_N)$ for $\forall \bm{j}_N\in\mathcal{J}_N$, $\|\bm{j}_N\|_2=1$. The worst-case detection rate on attack strength $\|\bm{a}_N\|_2 = |a|\neq0$ is defined as $f(\lambda_{\text{min}})$ with $\lambda_{\text{min}} = \lambda(\Delta \bm{x}, a\bm{j}_N^*)$.
\end{definition}
According to the Definition \ref{def:vul}, the weakest point in $(\mathcal{J}_N,\mathcal{J}_N')$ satisfies $|a|\|\bm{S}_N'\bm{j}_N^*\|_2\leq |a|\|\bm{S}_N'\bm{j}_N\|_2$, $\forall \bm{j}_N\in\mathcal{J}_N, \|\bm{j}_N\|_2 = 1, a \neq 0$. Let $\bm{a}_N^* = a\bm{j}_N^*$ and $\bm{a}_N = a\bm{j}_N$, the detection rate on $\bm{a}_N^*$ is the lowest among all attacks with the same strength as $\|\bm{S}_N'\bm{a}_N^*\|_2\leq \|\bm{S}_N'\bm{a}_N\|_2, \forall \bm{a}_N\in\mathcal{J}_N, \|\bm{a}_N\|_2=|a| \neq 0$.
Note that the weakest point may not be unique, but all of them are with the same worst-case detection rate.
Based on the definition of MTD weakest point, the following robust max-min optimization problem can be formulated:
\begin{equation}\label{eq:robust_opt}
\begin{array}{cc}
\max_{\Delta \bm{x}} \min_{\|\bm{a}_N\|_2 = 1,\bm{a}_N \in \mathcal{J}_N} & {\|\bm{S}_N'\bm{a}_N\|_2^2} \\
\text{s.t.} & \eqref{eq:reactance_limit_in}-\eqref{eq:reactance_limit_out}
\end{array}
\end{equation}
The inner problem $\min_{\|\bm{a}_N\|_2 = 1,\bm{a}_N \in \mathcal{J}_N} {\|\bm{S}_N'\bm{a}_N\|_2^2}$ is the mathematical formulation of the weakest point in Definition \ref{def:vul} which is maximized over the outer programming. From a game theorem point of view, we can present this setting as an intelligent attacker aims to develop an FDI attack with the highest probability to bypass BDD and the system operator tries to improve his/her defence strategy against this intelligent attacker. In the following sections, we will show that the two-stage problem \eqref{eq:robust_opt} can be reduced into a single-stage minimization problem by analytically representing the weakest point using the principal angles between $\mathcal{J}_N$ and $\mathcal{J}_N'$.
\subsection{Robust MTD for the Grid with Complete Configuration }
Similar to the one-dimensional case where the angle between two unitary vectors $\bm{u}$ and $\bm{v}$ is defined as $\cos{\theta} = \bm{v}^T\bm{u}$, the minimal angle between $\mathcal{J}_N, \mathcal{J}_N'\subseteq \mathbb{R}^p$ is defined as $0 \leq \theta_{1} \leq \pi/2$ \cite{meyer2000matrix}:
\begin{equation}\label{eq:minimal_angle}
\cos \theta_{1}=\max_{\bm{u}\in \mathcal{J}_N, \bm{v} \in \mathcal{J}_N' \atop\| \bm{u}\left\|_{2}=\right\| \bm{v} \|_{2}=1} \bm{u}^{T} \bm{v} = \bm{u}_1^{T} \bm{v}_1
\end{equation}
Apart from $\theta_1$, a sequence of angles can be defined iteratively by finding the orthonormal basis of $\mathcal{J}_N$ and $\mathcal{J}_N'$ such that for $i=2,\dots,n$ \cite{meyer2000matrix}:
\begin{equation}\label{eq:principal_angle}
\cos \theta_{i}=\max _{\bm{u} \in \mathcal{J}_{N,i}, \bm{v} \in \mathcal{J}_{N,i}' \atop\|\bm{u}\|_{2}=\|\bm{v}\|_{2}=1} \bm{u}^{T} \bm{v}=\bm{u}_{i}^{T} \bm{v}_{i}
\end{equation}
where $\mathcal{J}_{N,i} = \bm{u}_{i-1}^{\perp} \cap \mathcal{J}_{N,i-1}$ and $\mathcal{J}_{N,i}'=\bm{v}_{i-1}^{\perp} \cap \mathcal{J}_{N,i-1}'$.
From \eqref{eq:minimal_angle}-\eqref{eq:principal_angle}, a sequence of angles $\Theta = \{\theta_1,\theta_2,\dots,\theta_n\}$ is defined as the principal angles between $\mathcal{J}_N$ and $\mathcal{J}_N'$. The sequence of $\theta_i$ can be seperated into three parts. Let $\Theta_{1} = \{\theta_i|\theta_i = 0\}$, $\Theta_{2} = \{\theta_i|0<\theta_i < \pi/2\}$, and $\Theta_{3} = \{\theta_i|\theta_i = \pi/2\}$ with cardinality equals to $k$, $r$, and $l$ respectively, and $n=k+r+l$.
The corresponding vectors $\bm{U} = \{\bm{u}_1,\bm{u}_2,\dots,\bm{u}_n\}$ and $\bm{V} = \{\bm{v}_1,\bm{v}_2,\dots,\bm{v}_n\}$ are called as principal vectors which are the orthonormal basis of $\mathcal{J}_N$ and $\mathcal{J}_N'$, respectively. Similarly, $\bm{U}$ and $\bm{V}$ can also be separated into $\bm{U}_{1}, \bm{V}_{1},\cdots$. Specifically, $\bm{U}_1=\bm{V}_1=\mathcal{J}_N'\cap\mathcal{J}_N$ represents the intersection subspace of dimension $k$ and $l$ is the dimension of orthogonality. Moreover, it is proved that there always exist semi-orthogonal matrices $\bm{U}$ and $\bm{V}$ for any $\mathcal{J}_N$ and $\mathcal{J}_N'$ such that the bi-orthogonality is satisfied \cite{galantai2008subspaces}:
\begin{equation}\label{eq:bithogonality}
\bm{U}^T\bm{V} = \text{diag}([\cos{\theta_1}, \cos{\theta_2}, \dots, \cos{\theta_n}]) = \Gamma
\end{equation}
Referring to \eqref{eq:minimal_angle}, the following proposition specifies that the weakest point with the lowest detection rate of $(\mathcal{J}_N,\mathcal{J}_N')$
is the first principal vector $\bm{u}_1$ associated to the minimal principal angle $\theta_1$.
\begin{proposition}\label{prop:min_eff}
Given a pair of $(\mathcal{J}_N,\mathcal{J}_N')$, the minimum non-centrality parameter under attack strength $\|\bm{a}_N\|_2 = |a| \neq 0$ is $\lambda_{\text{min}} = a^2\sin^2{\theta_1}$. Meanwhile, $\lambda_{\text{min}}$ is achieved when attacking on the first principal vector $\bm{u}_1$ of $\bm{J}_N$.
\end{proposition}
\begin{proof}
Please refer to Appendix \ref{sec:app_proof_1}.
\end{proof}
As $\sin{\cdot}$ is monotonically increasing in $[0,\pi/2]$, Proposition \ref{prop:min_eff} implies that \eqref{eq:robust_opt} can be equivalently solved by:
\begin{equation}\label{eq:robust_opt_2}
\begin{array}{cc}
\max_{\Delta\bm{x}} & \theta_1 \\
\text{s.t.} & \eqref{eq:reactance_limit_in}-\eqref{eq:reactance_limit_out}
\end{array}
\end{equation}
As the orthogonal projector is uniquely defined \cite{meyer2000matrix} and also by \eqref{eq:bithogonality}, rewriting $\bm{P}_N = \bm{U}\bm{U}^T$ and $\bm{P}_N' = \bm{V}\bm{V}^T$ gives
\begin{equation}\label{eq:com_svd}
\bm{P}_N\bm{P}_N' =\bm{U}\bm{U}^T\bm{V}\bm{V}^T =\bm{U}\Gamma \bm{V}^T
\end{equation}
Eq.\eqref{eq:com_svd} is the truncated singular value decomposition (t-SVD) on $\bm{P}_N\bm{P}_N'$ where the diagonal matrix $\Gamma$ contains the first $n$ largest singular values of $\bm{P}_N\bm{P}_N'$, and $\bm{U}$ and $\bm{V}$ are the first (left- and right-hand) $n$ singular vectors of $\bm{P}_N\bm{P}_N'$. As $\sigma(\bm{P}_N\bm{P}_N') = \{\bm{1}_k, \cos^2{\theta_{k+i}(i=1,\dots,r)},\bm{0}_{k+r+i}(i=1,\dots,l),\bm{0}_{n+i}(i=1,\dots,m-n)\}$, this t-SVD is an exact decomposition of $\bm{P}_N\bm{P}_N'$.
Based on the t-SVD, Algorithm \ref{alg:vulnerable_point} is proposed to find the weakest point and the worst-case detection rate. For the grid with complete configuration, the composite matrix can be full column rank so that $k=0$. Line 6 outputs the weakest point $\bm{u}_1$ while line 9 outputs the empty intersection subspace. The worst-case detection rate is calculated according to Proposition \ref{prop:min_eff} in line 7. Practically, once the MTD strategy is determined, the weakest point $\bm{u}_1$ of this strategy can be directly spotted. The system operator can therefore evaluate the worst-case detection rate with respect to a maximum tolerable attack strength $|a|$. In addition, when $\theta_1 = \pi/2$, Proposition \ref{prop:min_eff} implies that the minimum non-centrality parameter equals to $a^2$. As two subspaces $\mathcal{J}_N$ and $\mathcal{J}_N'$ are orthogonal if $\theta_1 = \pi/2$, Proposition \ref{prop:min_eff} is consistent with the maximum detection probability described in Theorem 1 of \cite{lakshminarayana2021cost}.
\begin{algorithm}[t]
\footnotesize
\SetKwInOut{Input}{Input}
\SetKwInOut{Output}{Output}
\Input{$\mathcal{G}(\mathcal{N},\mathcal{E})$, $\Delta \bm{x}$, and $|a|$}
\Output{$\bm{u}_{k+1}, \bm{U}_{1}, f_{min}$}
Construct the pre- and post- MTD measurement matrices $\bm{J}_N$ and $\bm{J}_N'$, respectively;
Find the orthogonal projectors $\bm{P}_N$ and $\bm{P}_N'$ on $\bm{J}_N$ and $\bm{J}_N'$. Then do t-SVD \eqref{eq:com_svd};
$rank = \text{rank}([\bm{J}_N,\bm{J}_N'])$; \tcc{Rank of the composite matrix.}
$k = 2n - rank$; \tcc{The dimension of $\mathcal{J}_N'\cap\mathcal{J}_N$.}
$\theta_{k+1} = \Sigma(k+1,k+1)$;
$\bm{u}_{k+1} = \bm{U}(k+1,k+1)$; \tcc{The weakest point in $\mathcal{J}_N\setminus(\mathcal{J}_N'\cap \mathcal{J}_N)$.}
$f_{\min} = f(|a|^2\sin^2(\theta_{k+1}))$; \tcc{The worst-case detection rate in $\mathcal{J}_N\setminus(\mathcal{J}_N'\cap \mathcal{J}_N)$.}
\eIf{$rank = 2n$}
{
\tcc{Complete MTD configuration.}
$\bm{U}_{1} = \varnothing$;
}
{
\tcc{Incomplete MTD configuration.}
$\bm{U}_{1} = \bm{U}(:,1:k)$;
}
\caption{Find the Weakest Point(s) and the Worst-Case Detection Rate}
\label{alg:vulnerable_point}
\end{algorithm}
To guarantee the robust detection performance on the weakest point, $\theta_1$ should be maximized. As $\cos{\theta_1}$ is the largest singular value of $\bm{P}_N\bm{P}_N'$ by t-SVD, the worst-case detection rate can be maximized by the \textit{robust MTD} algorithm for the grid with complete configuration as follows:
\begin{equation}\label{eq:complete_mtd}
\begin{array}{cc}
\min_{\Delta\bm{x}} & {\|\bm{P}_N\bm{P}_N'\|} \\
\text{s.t.} & \eqref{eq:reactance_limit_in}-\eqref{eq:reactance_limit_out}
\end{array}
\end{equation}
where the property $\|\bm{P}_N\bm{P}_N'\| = \sigma_{\max}(\bm{P}_N\bm{P}_N')$ is used and $\|\bm{P}_N\bm{P}_N'\|\in[0,1]$.
\begin{remark}
The robust MTD algorithm \eqref{eq:complete_mtd} requires sufficient D-FACTS devices placement (as a planning stage problem) to guarantee $k=0$, e.g., using the `D-FACTS placement for the complete MTD' algorithm in \cite{liu2020optimal}.
\end{remark}
\subsection{Robust MTD for the Grid with Incomplete Configuration}
The robust MTD in \eqref{eq:complete_mtd} is not tractable for power system with incomplete MTD configuration. As $k\neq0$, $\theta_1 \equiv 0$ and $\|\bm{P}_N\bm{P}_N'\| \equiv 1$ no matter how $\Delta \bm{x}$ is designed. Fig. \ref{fig:illustraion_3d} shows a three-dimensional incomplete-MTD case. The attack $\bm{a}_N$ in green shows a random attack attempt with nonzero $\lambda$. However, the weakest point $\text{Col}(\bm{u_1})$ is not trivial. As the attacker can possibly target on $\text{Col}(\bm{u}_1)$, the worst-case detection rate equals to FPR constantly. Apart from $\theta_1$, every attack in $\bm{U}_1$ is undetectable. The intersection can be regarded as the space of the weakest points, whose dimension is calculated as $k = 2n-\text{rank}([\bm{J}_N,\bm{J}_N'])\neq 0$. Therefore, the smallest nonzero principal angle (which also corresponds the weakest point in $\mathcal{J}_N\setminus (\mathcal{J}_N'\cap\mathcal{J}_N)$) can be found as $\theta_{k+1}$ in line 5 of Algorithm \ref{alg:vulnerable_point} with the minimum detection rate calculated in line 7. Meanwhile, $\bm{U}_1$, corresponding to the subspace that cannot be detected, is calculated in line 11.
\begin{figure}
\centering
\includegraphics[width=0.3\textwidth]{figure/illustration_3d.pdf}
\caption{An illustration on the grid with incomplete configuration, $\mathcal{J}_N,\mathcal{J}_N'\subset \mathbb{R}^3$.}
\label{fig:illustraion_3d}
\end{figure}
To solve the intractable problem for the grid with incomplete configuration, the following design principles are considered which can improve the robust performance of MTD with incomplete configuration:
\textbf{\textit{Principle 1}}: Minimize $k$, the dimension of intersection.
\textbf{\textit{Principle 2}}: The attacker shall not easily attack on the intersection subspace $\bm{U}_1$ by chance.
\textbf{\textit{Principle 3}}: Maximize $\theta_{k+1}$, the minimum nonzero principal angle in $(\mathcal{J}_N,\mathcal{J}_N')$.
\textbf{\textit{Principle 1:}} The idea of Principal 1 is to minimize the attack space that can never be detected by MTD so that the probability of detectable FDI attacks increases. Minimizing $k$ is a planning stage problem as the rank of the composite matrix is almost not related to the perturbation amount of the D-FACTS devices once they are deployed \cite{zhang2019analysis}.
In this paper, we propose a new D-FACTS device placement algorithm to achieve minimum $k$. Compared with the existing work \cite{liu2020optimal, zhang2019analysis, liu2018reactance}, our algorithm uses the BLOSSOM algorithm \cite{galil1986efficient} to find the maximum cardinality matching \cite{bondy2008graph} of $\mathcal{G}(\mathcal{N},\mathcal{E})$, which can reach all necessary buses with the smallest number of D-FACTS devices. More details are presented in Appendix \ref{sec:deployment}.
\textbf{\textit{Principle 2:}} From the robust consideration, the following lemma is derived for the attacks targeting on the weakest point(s) for the grid with incomplete MTD configuration.
\begin{lemma}\label{lemma:vul}
Let $\bm{U} = (\bm{U}_{1},\bm{U}_{2,3})$ where $\bm{U}_{2,3}$ is the collection of columns in $\bm{U}_2$ and $\bm{U}_3$. Let $\bm{a}_N = \bm{U}_{1}\bm{c}_{1} + \bm{U}_{2,3}\bm{c}_{2,3}$ with $\bm{c}_{1}\in\mathbb{R}^{k}$ and $\bm{c}_{2,3}\in\mathbb{R}^{r+l}$. The detection rate on $\bm{a}_N$ does not depend on the value of $\bm{c}_{1}$.
\end{lemma}
\begin{proof}
Please refer to Appendix \ref{sec:proof_vul}.
\end{proof}
Although the attackers cannot immediately know the exact $\bm{x}'$ (Assumption 1), Lemma \ref{lemma:vul} suggests that the MTD algorithm should be designed such that the attackers cannot easily attack on $\bm{U}_1$ by chance. Specifically, considering the attack targeting on a single state $i$, if $\text{Col}(\bm{J}_N(:,i))\subseteq \bm{U}_1$, the single-state attack on bus $i$ can bypass the MTD while any attack involving bus $i$ can be detected ineffectively. To avoid the ineffective MTD on single-state attack,
the following constraint is considered:
\begin{equation}\label{eq:idle_mtd}
\|\bm{P}_N^i\bm{P}_N'\| \geq \gamma_i, \quad \forall i\in\mathcal{N}^c
\end{equation}
where $\bm{P}_N^i = \left(\bm{J}_N(:,i)^T\bm{J}_N(:,i)\right)^{-1}\bm{J}_N(:,i)\bm{J}_N(:,i)^T$ is the orthogonal projector on $\text{Col}(\bm{J}_N(:,i))$. $\mathcal{N}^c$ represents the index set of buses that are included by at least a loop\footnote{As proved by \cite{zhang2020hiddenness}, if a bus is not included by any loop, the attacks on this bus cannot be detected regardless of the MTD strategies.} of $\mathcal{G}$. Since $\|\bm{P}_N^i\bm{P}_N'\|\in[0,1]$ and 1 is achieved when $\text{Col}(\bm{J}_N(:,i))\subseteq\bm{U}_1$, the threshold $\gamma_i$ can be set close but not equal to 1.
Notice that the constraint in \eqref{eq:idle_mtd} cannot eliminate the weakest point(s) nor improve the worst-case detection rate on $\bm{U}_1$, but it can restrict the attacker's knowledge on the weakest point(s) due to the following reason. Rewriting $\lambda$ as $\lambda = \|(\bm{I}-\bm{P}_N')\sum_{i=1}^n\bm{J}_N(:,i)\bm{c}(i)\|_2^2$, constraint \eqref{eq:idle_mtd} ensures that $(\bm{I}-\bm{P}_N')\bm{J}_N(:,i)\bm{c}(i)\neq0$, $\forall i \in \mathcal{N}^c$. To have low MTD detection rate, the attacker has to coordinate the attack strength on at least two bus to have low $\lambda$ which is only possible if $\bm{x}'$ is known. As long as the attacker cannot easily attack on $\bm{U}_1$, the probability to have the worst case is low and the MTD strategy is still
effective from robust point of view.
\begin{remark}
To fulfill constraint \eqref{eq:idle_mtd}, all buses in $\mathcal{N}^c$ should be incident to at least a branch equipped with D-FACTS devices, which can be achieved by the proposed D-FACTS devices placement algorithm in Appendix \ref{sec:deployment}.
\end{remark}
\textbf{\textit{Principle 3:}} Although the chance of the worst-case attack is minimized by Principle 1-2, it does not necessarily imply a high detection rate when $\bm{a}_N \notin \bm{U}_1$. Similar to \eqref{eq:robust_opt_2}, the minimum nonzero principal angle $\theta_{k+1}$, which represents the weakest point in subspace $\mathcal{J}_N\setminus(\mathcal{J}_N'\cap\mathcal{J}_N)$ should be maximized by
\begin{equation}\label{eq:incomplete_0}
\begin{array}{cl}
\min_{\Delta\bm{x}} & {\cos{\theta_{k+1}}} \\
\text{s.t.} & \eqref{eq:reactance_limit_in}-\eqref{eq:reactance_limit_out}, \eqref{eq:idle_mtd}
\end{array}
\end{equation}
where $\cos{\theta_{k+1}}$ is the $(k+1)$th largest singular value.
As far as we know, there is no direct method to solve \eqref{eq:incomplete_0} as finding the singular value at a certain position requires solving the SVD on $\bm{P}_N\bm{P}_N'$ and locating the $1$th to $k$th singular vectors. Therefore, we propose an iterative Algorithm \ref{alg:incomplete} to solve \eqref{eq:incomplete_0}. In line 1 of Algorithm \ref{alg:incomplete}, a warm start $\Delta \bm{x}^0$ is firstly found by minimizing the Frobenius norm $\|\cdot\|_F$, which is shown to be an upper bound to $\cos{\theta_{k+1}}$.
\begin{equation}\label{eq:incomplete_fro}
\begin{array}{cl}
\min_{\Delta\bm{x}} & {\|\bm{P}_N\bm{P}_N'\|_F} \\
\text{s.t.} & \eqref{eq:reactance_limit_in}-\eqref{eq:reactance_limit_out}, \eqref{eq:idle_mtd}
\end{array}
\end{equation}
For a given warm-start perturbation value $\Delta \bm{x}^0$, the intersection subspace $\bm{U}_1$ can be located by Algorithm \ref{alg:vulnerable_point}. Denoting $\bm{U}_1(\Delta \bm{x}^0)$ as $\bm{U}_1^0$, the t-SVD \eqref{eq:com_svd} can be rewritten as
\begin{equation*}
\begin{array}{rl}
\bm{P}_N\bm{P}_N' = & \left(\bm{U}_1^0, \bm{U}_{2,3}\right)\begin{pmatrix}
\bm{I} & \bm{0}\\
\bm{0} & \Gamma_{2,3}
\end{pmatrix}\begin{pmatrix}
\bm{V}_{1}^{0T}\\
\bm{V}_{2,3}^T
\end{pmatrix} \\
= & \bm{U}_{1}^0\bm{U}_{1}^{0T} + \bm{U}_{2,3} \Gamma_{2,3} \bm{V}_{2,3}^T
\end{array}
\end{equation*}
where $\bm{I}$ is the identity matrix of dimension $k$; $\Gamma_{2,3} = \text{diag}([\cos(\theta_{k+1}),\cdots,\cos(\theta_{n}]))$ with $\theta_{k+1}\neq0$. Note that $\bm{U}_1^0 = \bm{V}_1^0 = \mathcal{J}_N'\cap\mathcal{J}_N$.
Therefore, the following optimization problem can be formulated to minimize $\cos{\theta_{k+1}}$:
\begin{equation}\label{eq:incomplete_1}
\begin{array}{cl}
\min_{\Delta\bm{x}} & \|\bm{P}_N\bm{P}_N' - \bm{U}_{1}^0\bm{U}_{1}^{0T}\| \\
\text{s.t.} & \eqref{eq:reactance_limit_in}-\eqref{eq:reactance_limit_out}, \eqref{eq:idle_mtd}
\end{array}
\end{equation}
Denoting the optimal value of \eqref{eq:incomplete_1} as $\Delta \bm{x}^1$, a new intersection subspace $\bm{U}_1^1= \bm{U}_1(\Delta\bm{x}^1)$ can be located. As $\Delta\bm{x}^1$ is solved with fixed $\bm{U}_1^0$, $\bm{U}_1^1$ may not be the same as $\bm{U}_1^0$. After finding the new intersection subspace from $\Delta \bm{x}^1$, \eqref{eq:incomplete_1} can be iteratively solved until convergence as shown by line 3-11 in Algorithm \ref{alg:incomplete}.
\begin{algorithm}[t]
\footnotesize
\SetKwInOut{Input}{Input}
\SetKwInOut{Output}{Output}
\Input{$\mathcal{G}(\mathcal{N},\mathcal{E})$, $tol$, $max\_ite$}
\Output{$\Delta\bm{x}^1$}
Find the warm start point $\Delta \bm{x}^0$ by solving \eqref{eq:incomplete_fro};
Find the intersection subspace $\bm{U}_{1}^0$ by Algorithm \ref{alg:vulnerable_point};
\tcc{iteration until convergence.}
\While{$step<max\_ite$}
{
Find $\Delta\bm{x}^1$ by solving \eqref{eq:incomplete_1};
Find the intersection subspace $\bm{U}_{1}^1$ by Algorithm \ref{alg:vulnerable_point};
\eIf{$\|\bm{U}_1^1-\bm{U}^0_1\|_2 \leq tol$}
{
break; \tcc{converged.}
}
{
$\bm{U}_{1}^0:=\bm{U}_{1}^1$;
}
}
\caption{Robust MTD for the Grid with Incomplete Configuration}
\label{alg:incomplete}
\end{algorithm}
To sum up, Algorithm \ref{alg:incomplete} limits the chance of attacking on $\mathcal{J}_N'\cap\mathcal{J}_N$ (Principal 1-2) and guarantees the worst-case detection rate in $\mathcal{J}_N\setminus(\mathcal{J}_N'\cap\mathcal{J}_N)$ (Principal 3 and \eqref{eq:incomplete_fro}-\eqref{eq:incomplete_1}) for the grid with incomplete configuration.
\subsection{Discussions on Full AC Model Design}
\label{sec:discussion}
In the previous sections, we theoretically established the robust MTD algorithm based on the simplified AC model \eqref{eq:linear_equation}-\eqref{eq:linear_attack} and intend to apply the algorithms on the full AC model \eqref{eq:power_balance}-\eqref{eq:ac_se}. There exists similar concept on the weakest point in the original AC settings. Let $\bm{h}'^{-1}(\cdot)$ represent the result of state estimation in \eqref{eq:ac_se}. The estimated state on attacked measurement is written as $\hat{\bm{\nu}}_a' = \bm{h}'^{-1}(\bm{z}_a')$ and the residual is $\gamma_a'=\|\bm{R}^{-\frac{1}{2}}(\bm{z}_a' - \bm{h}'(\hat{\bm{\nu}}_a'))\|_2^2$. The weakest point can be defined as a unitary attack vector such that $\gamma_a'$ is minimized. However, there are several obstacles to analytically write its expression. Firstly, recall that $\bm{a} = \bm{h} (\bm{\nu}'+\bm{c}) - \bm{h} (\bm{\nu}')$ which is non-linearly dependent on post-MTD state $\bm{\nu}'$ and state attack vector $\bm{c}$. Note that $\bm{\nu}'$ is dependent on $\bm{x}'$ which cannot be determined in advance. Second, $\bm{h}'^{-1}(\cdot)$ requires iterative update, such as Gauss-Newton or Quasi-Newton algorithm. Although it is possible to reformulate AC-SE as semi-definite programming \cite{zhu2014power}, it lacks of analytical solution in general. Third, it is difficult to define the concept of angles between subspaces defined by two functions $\bm{h}(\cdot)$ and $\bm{h}'(\cdot)$. Consequently, we theoretically derived the robust algorithm based on the simplified AC model and numerically verify the performance on AC-FDI attacks in simulation. We found out that the MTD designed by the sufficient separation between the subspaces between the Jacobian matrices can provide effective detection in the full AC model.
\section{Simulation}
\label{sec:simulation}
\subsection{Simulation Set-ups}
We test the proposed algorithms\footnote{{Code is available at \url{https://github.com/xuwkk/Robust_MTD}}} on IEEE benchmarks case-6ww, case-14, and case-57 in MATPOWER \cite{zimmerman2011matpower}. AC-OPF is solved using python package PYPOWER 5.1.15. and the nonlinear optimization problems are solved using open-source library SciPy. More simulation set-ups are given as follows.
\subsubsection{Attack Pools and BDD Threshold}
Firstly, we define the attack strength with respect to the noise level as:
\begin{equation}\label{eq:att_rho}
\rho = \frac{\|\bm{a}\|_2}{\sqrt{\sum_i^m \sigma_i^2}}
\end{equation}
We consider three types of attacks for the simplified AC model. 1). \textbf{\textit{Worst-case attack}} where the attacker attacks on the weakest point $\bm{u}_{k+1}$ of a given MTD strategy according to Algorithm \ref{alg:vulnerable_point}; 2). \textbf{\textit{Single-state attack}} where the attacker only injects on single non-reference phase angle of the grid; and 3). \textbf{\textit{Random attack}} where the attack vector $\bm{a}$ is randomly generated as follows. Firstly, the attack state vector $\bm{c}$ is generated with a random number of attacked states $\|\bm{c}\|_0 = q, q = 1, 2, \dots, n$ and then sampled from multivariate Gaussian distribution with $q$ non-zero entries. Then the attack vector is found as $\bm{a}=\bm{J}\bm{c}$ and rescaled by different $\rho = 5,7,10,15,20$ according to \eqref{eq:att_rho}. To simplify the analysis, the measurement noise is set as $\sigma_i = 0.01p.u., \forall i$ in all the case studies. In this case, to have $\beta$-MTD, a necessary condition is $\rho\geq\sqrt{\lambda_c(\beta)/m}$ according to Proposition \ref{prop:min_att}.
The measurement in the original AC model consists of $P_i$, $Q_i$, $P_{k:i\to j}$, and $Q_{k:i\to j}$ \eqref{eq:power_balance}. To simulate the attack with different attack strength, we randomly sample $\bm{c}$ from uniform distribution and classify $\bm{a} = \bm{h}'(\bm{\nu}'+\bm{c}) - \bm{h}'(\bm{\nu}')$ into one of the ranges $\{[5,7), [7,10), [10,15), [15,20), [20,25), [25,\infty)\}$ by \eqref{eq:att_rho}.
We sample \texttt{no\_load}=50 load conditions on a uniform distribution of the default load profile in MATPOWER \cite{zimmerman2011matpower} for each grid. We then set the D-FACTS devices by robust MTD algorithms and simulate the real-time measurements. Under each load condition, we generate \texttt{no\_attack}=200 attack attempts for each of the attack types. The BDD threshold $\tau_\chi(\alpha)$ is determined to have $\alpha=5\%$ FPR according different DoFs in different systems.
\subsubsection{Metrics and Baselines}
The key metric to evaluate the MTD detection performance is the true positive rate, also known as the attack detection probability (ADP) which is the ratio between the number of attacks that are detected by the MTD detector and the total number of attacks.
The max-rank MTD algorithm modified from \cite{liu2018reactance, liu2020optimal, zhang2019analysis, zhang2021strategic} is compared as the baseline where the reactances are randomly changed with $\mu_{min}\bm{x}_i \leq |\Delta \bm{x}_i|\leq \mu_{max}\bm{x}_i$. Note that each reactance is perturbed by $\mu_{min}>0$ to fulfill max-rank condition on the composite matrix. For each one of the \texttt{no\_attack} attack attempt, we simulate \texttt{no\_maxrank}=20 number of max rank MTDs to evaluate its average detection performance.
\subsection{Verification of Theoretical Analysis on Simplified AC Model}
\begin{figure}
\centering
\begin{subfigure}[b]{0.24\textwidth}
\centering
\includegraphics[width=\textwidth]{figure/case6_dc_worst.pdf}
\caption{Worst-Case Attack}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.24\textwidth}
\centering
\includegraphics[width=\textwidth]{figure/case6_dc_random.pdf}
\caption{Random Attack}
\end{subfigure}
\caption{ADPs on simplified case-6 system. }
\label{fig:case6}
\end{figure}
\begin{figure}
\centering
\begin{subfigure}[b]{0.24\textwidth}
\centering
\includegraphics[width=\textwidth]{figure/case14_dc_worst.pdf}
\caption{Worst-Case Attack}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.24\textwidth}
\centering
\includegraphics[width=\textwidth]{figure/case14_dc_random.pdf}
\caption{Random Attack}
\end{subfigure}
\caption{ADPs on simplified case-14 system. }
\label{fig:case14}
\end{figure}
In the first case study, we verify the theoretical analysis of robust MTD algorithms and demonstrate its effectiveness on the simplified AC model \eqref{eq:linear_equation}-\eqref{eq:linear_attack}.
Firstly, the ADPs on case-6 with complete configuration are illustrated in Fig. \ref{fig:case6} for both worst-case attacks and random attacks. The reactances are changed with
$\tau = 0.2$. Meanwhile, $\mu_{min} = 0.05$ and $\mu_{max} = 0.2$ in the max-rank MTD. In Fig. \ref{fig:case6}(a), the simulation result on the ADPs of robust MTD is the same as the theoretic detection rate $f(\lambda_{min})$ calculated by Proposition \ref{prop:min_eff}, which verifies the theoretic analysis and the design criteria. In addition, the robust MTD algorithm shows much higher ADPs than the max-rank MTD on the worst-case attack. Although the max-rank MTD's performance may approach to the robust MTD in some cases, its average ADP is similar to the FPR as the worst-case performance cannot be explicitly considered under the noiseless setting.
In Fig. \ref{fig:case6}(b), the max MTD is added by solving \eqref{eq:max_mtd} with the assumption that the attack vector $\bm{a}_N$ is known, which represents the performance upper-bound of any MTD design. As shown by Fig. \ref{fig:case6}(b), the robust MTD algorithm, not only guarantees the worst case condition, but also outperforms the max-rank MTD by 10\%-45\% on random attacks with different $\rho$.
Moreover, the gap between robust MTD and max MTD algorithms is smaller than 25\% and approaches to zero when $\rho\geq15$. However, by comparing Fig. \ref{fig:case6}(a) and Fig. \ref{fig:case6}(b), it is worth noting that the major improvement of robust MTD over max-rank MTD still lays in the worst-case attacks.
Fig. \ref{fig:case14} investigates the performance on the case-14 system with incomplete configuration. By Algorithm \ref{alg:vulnerable_point}, the minimum $k$ equals to 6 and the worst point in $\mathcal{J}_N\setminus(\mathcal{J}_N'\cap\mathcal{J}_N)$ is at $\bm{u}_7$. Assume that all branches are equipped with D-FACTS devices and the maximum perturbation ratio is set as $\tau = 0.2$. Although the detection rates on attacks in $\bm{U}_1$ equal to $\alpha$ according to Lemma \ref{lemma:vul}, the ADP on $\bm{u}_7$ is nonzero by implementing Algorithm \ref{alg:incomplete} and increases as the attack strength increases. Similar to Fig. \ref{fig:case6}(a), although the max-rank MTD algorithm can, by chance, give high detection rate against the worst case attack, its average detection rate is extremely low. In Fig. \ref{fig:case14}(b), the gap between max MTD and robust MTD is low (5\%-30\%). The results demonstrate that the robust design can also effectively improve the detection performance for the grid with incomplete configuration.
To further investigate on the weakest points in $\bm{U}_1$, we generate single-bus attack with $\rho = 10$ and record the ADPs in Fig. \ref{fig:single_state_attack} with and without Principle 2 \eqref{eq:idle_mtd}. Firstly, attacks targeting on bus-8 can only be detected by 5\%. This is because bus-8 is a degree-one bus which is excluded by any loop. Second, with Principle 2 considered, the robust MTD can give more than 90\% ADPs for all buses. In contrast, there are attacks against certain buses, e.g. bus-7, 10, 11, and 13 can be barely detected without Principle 2.
Consequently, the simulation result verifies that Principle 2 can sufficiently reduce the chance of attacking on the weakest points.
\begin{figure}
\centering
\includegraphics[width=0.3\textwidth]{figure/case14_single.pdf}
\caption{ADPs on single-state attacks of case-14 system.}
\label{fig:single_state_attack}
\end{figure}
\subsection{Simulation Results on Full AC Model}
In this section, we verify the detection effectiveness of the proposed robust MTD algorithms on FDI attacks under the original AC settings \eqref{eq:power_balance}-\eqref{eq:ac_se}.
\subsubsection{Random Attack}
The ADPs on random attacks for case-6, case-14, and case-57 systems are summarised in Table \ref{tab:random_ac}. Similar to the studies on simplified AC models, the proposed robust algorithms can improve the ADPs by 10\%-40\% than the max-rank algorithm. In particular, for the cases with attack strength below 20, robust MTD can almost double the ADPs of max-rank MTD for all three systems. Therefore, the robust MTD designed by the principal angles between the subspaces of pre- and post- MTD Jacobian matrices can significantly improve the detection effectiveness for defending AC-FDI attacks. In addition, the
attacks with larger attack strength are more likely to be detected while the detection probability for different systems under the same attack strength is slightly different due to their different load levels, parameters (e.g. the reactance to resistance ratios), and topologies. For instance, case-57 system is harder to detect as the ADPs in both max-rank and robust MTDs are lower than the case-6 and case-14 systems.
To certify the detection performance, the residual distributions for case-6, case-14, and case-57 systems are summarized in Fig. \ref{fig:ac_result} where kernel density estimation (KDE) is used to smooth the histograms.
The result implies that the proposed algorithms can generalize well on the AC-FDI attacks by sufficiently shifting the distribution positively, which is shown to be a key property on effective MTD with the measurement noise considered in Fig. \ref{fig:residual_illustration}. For each sub-figure, it can be observed that the detection probability without MTD is around 5\% which is the same as the predefined FPR and on average, the max-rank MTD performs worse than the robust MTD.
\begin{table}
\footnotesize
\caption{Average ADPs on Random AC-FDI Attacks.}
\begin{tabular}{c|cccccc}
\hline
\multirow{2}{*}{$\rho$} &
\multicolumn{2}{c}{\textbf{case-6}} &
\multicolumn{2}{c}{\textbf{case-14}} &
\multicolumn{2}{c}{\textbf{case-57}} \\
& Max-Rk & Robust & Max-Rk & Robust & Max-Rk & Robust \\
\hline\hline
$[5,7)$ & 7.1\% & 13.7\% & 8.6\% & 18.1\% & 10.3\% & 30.3\% \\
\hline
$[7,10)$ & 12.6\% & 33.2\% & 14.4\% & 41.2\% & 15.2\% & 39.2\% \\
\hline
$[10,15)$ & 25.1\% & 67.3\% & 27.5\% & 63.1\% & 23.7\% & 55.9\% \\
\hline
$[15,20)$ & 44.5\% & 92.4\% & 43.4\% & 87.5\% & 36.0\% & 69.1\%\\
\hline
$[20,25)$ & 60.2\% & 98.2\% & 60.6\% & 94.5\% & 50.6\% & 81.6\%\\
\hline
\end{tabular}
\label{tab:random_ac}
\end{table}
\begin{figure}
\centering
\begin{subfigure}[b]{0.24\textwidth}
\centering
\includegraphics[width=\textwidth]{figure/case6_ac_residual_2.pdf}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.24\textwidth}
\centering
\includegraphics[width=\textwidth]{figure/case6_ac_residual_4.pdf}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.24\textwidth}
\centering
\includegraphics[width=\textwidth]{figure/case14_ac_residual_2.pdf}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.24\textwidth}
\centering
\includegraphics[width=\textwidth]{figure/case14_ac_residual_4.pdf}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.24\textwidth}
\centering
\includegraphics[width=\textwidth]{figure/case57_ac_residual_2.pdf}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.24\textwidth}
\centering
\includegraphics[width=\textwidth]{figure/case57_ac_residual_4.pdf}
\end{subfigure}
\caption{Residual distributions of AC FDI attacks. The first row: case-6 system; the second row: case-14 system; the third row: case-57 system; the first column: attacks in range $[10,15)$; the second column: attacks in range $[20,25)$.}
\label{fig:ac_result}
\end{figure}
\subsubsection{Impact of Different Placements and Perturbation Ratios of D-FACTs Devices}
Fig. \ref{fig:case14_random_compare} records the simulation results on AC random attacks under two D-FACTS devices placements and four perturbation ratio limits. In detail, `all' represents perturbing all branches whereas `part' represents perturbing on branch- 2, 3, 4, 12, 15, 18, and 20, which is the outcome of the `D-FACTS Devices Placement Algorithm' in Appendix \ref{sec:deployment}. Simulation result shows that $k=6$ is achieved and all buses are covered except bus 8 in `part' placement. As the maximum perturbation ratio is reported as 50\% in literature \cite{lakshminarayana2021cost}, $\tau$ is set as 0.2, 0.3, 0.4, and 0.5. As a result, the grey curve in Fig. \ref{fig:case14_random_compare} is simulated under the same settings as the robust MTD in Table \ref{tab:random_ac}. When the number of D-FACTS devices is limited, although the minimum $k$ is still fulfilled by Principle 1, the detection rate is significantly reduced. In order to attain higher detection rate, the perturbation limit should be further increased.
Notably, these findings on the dependency of ADP on different D-FACTS device placements and perturbation ratios can only be found when the sensor noise is considered.
\begin{figure}
\centering
\includegraphics[width = 0.3\textwidth]{figure/case14_compare.pdf}
\caption{ADPs under different D-FACTS devices parameters.}
\label{fig:case14_random_compare}
\end{figure}
\subsubsection{Computational Time}
The computational time of the proposed algorithms are summarized in Table \ref{tab:time}. While the computation time depends on the system scales, number of D-FACTS devices, and algorithms, they are acceptable for real-time applications. In practice, as the attackers spend time to learn the new parameters, the system operator can solve the robust MTD algorithms with period much larger than the state estimation time, e.g. several hours, or only change Jacobian matrix $\bm{J}_N$ when the loads are significantly changed. A flat state vector may also be a choice to construct the Jacobian matrix if the loads change slowly. Multi-run strategy is also applied to approach to the global optimum of the nonlinear optimization problem.
\begin{table}[h]
\centering
\footnotesize
\caption{Computational Time (averaged by \texttt{no\_load} runs). }
\begin{tabular}{ c|ccc }
\hline
\textbf{Case} & \textbf{No. D-FACTS} & \textbf{Algorithm} & \textbf{Time (s)} \\\hline\hline
case-6 & 11 & \eqref{eq:complete_mtd} & 0.022 \\\hline
\multirow{3}{4em}{case-14} & 20 & Algorithm \ref{alg:incomplete} & 1.925 \\
& 20 & Algorithm \ref{alg:incomplete} without \eqref{eq:idle_mtd} & 0.325 \\
& 7 & Algorithm \ref{alg:incomplete} & 0.532 \\\hline
case-57 & 78 & Algorithm \ref{alg:incomplete} & 9.357 \\
\hline
\end{tabular}
\label{tab:time}
\end{table}
\section{Conclusions}
\label{sec:conclusion}
In this paper, we address the real-time robust implementation of MTD against unknown FDI attacks. Using the concept of angles between subspaces, we theoretically prove that the weakest point for any given MTD strategy corresponds to the smallest principal angle and the worst-case detection rate is proportional to the sine of this angle, with the impact of measurement noises being explicitly considered. These novel findings can help evaluate the effectiveness of the MTD strategy to tackle the unseen attacks in CPPS. A robust MTD algorithm is proposed by increasing the worst-case detection rate for the grid with complete MTD configuration. We then demonstrate that the weakest point(s) in incomplete MTD always exist and cannot be improved. Therefore, robust MTD for the grid with incomplete configuration is proposed by refraining from the ineffective MTD operation and improving the worst-case detection rate in the detectable subspace. Simulation results on standard IEEE benchmarks verify the real-time detection effectiveness on AC-FDI attacks, compared with the baseline.
|
\section{Introduction}
Artificial intelligence (AI) has undergone significant and continuous progress in the past decade, resulting in the increased adoption of its algorithms (e.g., machine learning (ML) algorithms) for solving many problems, even those that were difficult to resolve in the past. However, these outstanding achievements are accompanied by increasing model complexity and utilizing black-box AI models that lack transparency. Therefore, it becomes necessary to come up with solutions that can contribute to addressing such a challenge, which could help expand utilizing AI systems in critical and sensitive domains (e.g., healthcare and security domains) where other criteria must be met besides the high accuracy.
Explainable artificial intelligence (XAI) has been proposed as a solution that can help to move towards more transparent AI and thus avoid limiting the adoption of AI in critical domains \cite{Adadi2018,arrieta2020}. Generally speaking, according to \cite{gunning2016broad}, XAI focuses on developing explainable techniques that empower end-users in comperhending, trusting, and efficiently managing the new age of AI systems. Historically, the need for explanations dates back to the early works in explaining expert systems and Bayesian networks \cite{biran2017explanation}. Deep learning (DL), however, has made XAI a thriving research area.
Every year, a large number of studies dealing with XAI are published. At the same time, various review studies are published covering a range of general or specific aspects of XAI. With many of these review studies, several challenges and research directions are discussed. While this has led to identifying challenges and potential research directions, however, they are scattered.
To the best of our knowledge, this is the first meta-survey that explicitly organizes and reports on the challenges and potential research directions in XAI. This meta-survey aims to provide a reference point for researchers interested in working on challenges and potential research directions in XAI.
The organization of the paper is as follows. In Section 2, we discuss the need for XAI from various perspectives. Following that, Section 3 tries to contribute to a better distinction between explainability and interpretability. The protocol used in planning and executing this systematic meta-survey is presented in Section 4. Afterward, Section 5 discusses the challenges and research directions in XAI. Lastly, final remarks are highlighted in Section 6.
\section{Why Explainable AI is Needed?}
\label{motivations}
Nowadays, we are surrounded by black-box AI systems utilized to make decisions for us, as in autonomous vehicles, social networks, and medical systems. Most of these decisions are taken without knowing the reasons behind these decisions.
According to \cite{Adadi2018}, not all black-box AI systems need to explain why they take each decision because this could result in many consequences such as reducing systems efficiency and increasing development cost. Generally, explainability/interpretability is not needed in two situations \cite{Doshi-Velez2017-yb}: (1) results that are unacceptable are not accompanied by severe consequences, (2) the problem has been studied in-depth and well-tested in practice, so the decision made by the black-box system is trustworthy, e.g., advertisement system and postal code sorting. Therefore, we should think about why and when explanations/interpretations can be helpful \cite{Adadi2018}.
Based on the retrieved surveys in this work, the need for XAI can be discussed from various perspectives as shown in Fig. \ref{perspectives}. The perspective groups below are to some extent based on the work in \cite{Gade2020-gt}:
\begin{figure}[t]
\centering
\includegraphics[width=0.9\linewidth]{perspectives.png}
\caption{The five main perspectives for the need for XAI.}
\label{perspectives}
\end{figure}
\begin{itemize}
\item Regulatory perspective: Black-box AI systems are being utilized in many areas of our daily lives, which could resulting in unacceptable decisions, especially those that may lead to legal effects. Thus, it poses a new challenge for the legislation. The European Union's General Data Protection Regulation (GDPR)\footnote{https://www.privacy-regulation.eu/en/r71.htm} is an example of why XAI is needed from a regulatory perspective. These regulations create what is called "right to explanation," by which a user is entitled to request an explanation about the decision made by the algorithm that considerably influences them \cite{GoodmanFlaxman2017}. For example, if an AI system rejects one’s application for a loan, the applicant is entitled to request justifications behind that decision to guarantee it is in agreement with other laws and regulations \cite{Samek2019-po}. However, the implementation of such regulations is not straightforward, challenging, and without an enabling technology that can provide explanations, the "right to explanation" is nothing more than a "dead letter" \cite{Samek2019-po,Payrovnaziri2020-xi,Guidotti2018-tx}.
\item Scientific perspective: When building black-box AI models, we aim to develop an approximate function to address the given problem. Therefore, after creating the black-box AI model, the created model represents the basis of knowledge, rather than the data \cite{Molnar2021-do}. Based on that, XAI can be helpful to reveal the scientific knowledge extracted by the black-box AI models, which could lead to discovering novel concepts in various branches of science.
\item Industrial perspective: Regulations and user distrust in black-box AI systems represent challenges to the industry in applying complex and accurate black-box AI systems \cite{Veiber2020-gv}. Less accurate models that are more interpretable may be preferred in the industry because of regulation reasons \cite{Veiber2020-gv}. A major advantage of XAI is that it can help in mitigating the common trade-off between model interpretability and performance \cite{arrieta2020}, thus meeting these common challenges. However, it can increase development and deployment costs.
\item Model's developmental perspective: Several reasons could contribute to inappropriate results for black-box AI systems, such as limited training data, biased training data, outliers, adversarial data, and model's overfitting. Therefore, what black-box AI systems have learned and why they make decisions need to be understood, primarily when they affect humans’ lives. For that, the aim will be to use XAI to understand, debug, and improve the black-box AI system to enhance its robustness, increase safety and user trust, minimize or prevent faulty behavior, bias, unfairness, and discrimination \cite{Confalonieri2021-yx}. Furthermore, when comparing models with similar performance, XAI can help in the selection by revealing the features that the models use to produce their decisions \cite{Samek2017-lb,arras2017relevant}. In addition, XAI can serve as a proxy function for the ultimate goal because the algorithm may be optimized for an incomplete objective \cite{Doshi-Velez2017-yb}. For instance, optimizing an AI system for cholesterol control with ignoring the likelihood of adherence \cite{Doshi-Velez2017-yb}.
\item End-user and social perspectives: In the literature of deep learning \cite{Guidotti2018-tx,szegedy2013intriguing,nguyen2015deep}, it has been shown that altering an image such that humans cannot observe the change can lead the model in producing a wrong class label. On the contrary, completely unrecognizable images of humans can be recognizable with high confidence using DL models. Such findings could raise doubts about trusting such black-box AI models \cite{Guidotti2018-tx}. The possibility to produce unfair decisions is another concern about black-box AI systems. This could happen in case black-box AI systems are developed using data that may exhibit human biases and prejudices \cite{Guidotti2018-tx}. Therefore, producing explanations and enhancing the interpretability of the black-box AI systems will help in increasing trust because it will be possible to understand the rationale behind the model’s decisions, and we can know if the system serves what it is designed for instead of what it was trained for \cite{Guidotti2018-tx,Burkart2021-tq}. Furthermore, the demand for the fairness of black-box AI systems' decisions, which cannot be ensured by error measures, often leads to the need for interpretable models \cite{Lipton2018-dx}.
\end{itemize}
The above list is far from complete, and there may be an overlap between these perspectives. However, these highlight the most critical reasons why XAI is needed.
\section{From Explainability to Interpretability}
\label{expvsintp}
In the literature, there seems to be no agreement upon what ``explainability'' or ``interpretability'' mean. While both terms are often used interchangeably in the literature, some examples from the selected papers distinguish them \cite{Markus2021-bp,arrieta2020,Akata2020-ct,Chakraborty,zhang2020survey,Chatzimparmpas}. To show that there are no agreement upon definitions, three different definitions from \cite{Markus2021-bp,arrieta2020,Akata2020-ct} are provided. In \cite{Markus2021-bp}, the authors stated that ``\emph{... we consider interpretability a property related to an explanation and explainability a broader concept referring to all actions to explain.}''. In another work \cite{arrieta2020}, interpretability is defined as ``\emph{the ability to explain or to provide the meaning in understandable terms to a human.}'', while ``\emph{explainability is associated with the notion of explanation as an interface between humans and a decision-maker that is, at the same time, both an accurate proxy of the decision-maker and comprehensible to humans...}''. Another distinction is drawn in \cite{Akata2020-ct}, in which authors stated that ``\emph{... In the case of interpretation, abstract concepts are translated into insights useful for domain knowledge (for example, identifying correlations between layers in a neural network for language analysis and linguistic knowledge). An explanation provides information that gives insights to users as to how a model came to a decision or interpretation.}''. It can be noticed from these distinctions that the authors have different definitions for these two terms. In addition, there is still considerable ambiguity in some of the given distinctions.
To contribute to a better distinction between explainability and interpretability, this paper attempts to present a distinction between these terms as follows:
\begin{quote}
Explainability provides \textbf{insights} to a \textbf{targeted audience} to fulfill a \textbf{need}, whereas interpretability is the degree to which the provided \textbf{insights} can make sense for the \textbf{targeted audience}'s domain knowledge.
\end{quote}
There are three components in the definition of explainability, as shown in the above distinction: \textbf{insights}, \textbf{targeted audience}, and \textbf{need}. \textbf{Insights} are the output from explainability techniques used (e.g., text explanation, feature relevance, local explanation). These insights are provided to a \textbf{targeted audience} such as domain experts (e.g., medical doctors), end-users (e.g., users affected by the model decision), modeling experts (e.g., data scientists). The \textbf{need} for the provided insights may be to handle any issues discussed in Section \ref{motivations} such as justifying decisions, discovering new knowledge, improving the black-box AI model, and ensuring fair decisions. That means explainability aims to help the targeted audience to fulfill a need based on the provided insights from the explainability techniques used.
As for interpretability, are the provided explanations consistent with the targeted audience’s knowledge? Do the explanations make sense to the targeted audience? Is the targeted audience able to reason/inference to support decision-making? Are the provided explanations reasonable for the model's decision?
Although the distinction is not ideal, we believe that it represents an initial step toward understanding the difference between explainability and interpretability. This paper will use this proposed distinction when discussing the challenges and research directions in XAI.
\section{Systematic review planning and execution}
\label{methodology}
This work is mainly based on a systematic literature review (SLR) introduced by Kitchenham and Charters \cite{keele2007}. We started our SLR by specifying the research question: \emph{What are the challenges and research directions in XAI reported in the existing survey studies?} The answer to this question will help researchers and practitioners to know the various dimensions that one can consider when working in the XAI research area.
Having the research question established, the search terms based on the research question are:
\begin{itemize}
\item XAI keywords: explainable, XAI, interpretable.
\item Review keywords: survey, review, overview, literature, bibliometric, challenge, prospect, agenda, trend, insight, opportunity, lesson, research direction
\end{itemize}
Relevant and important electronic databases were selected and used for searching the primary studies based on the search terms. These databases are:
\begin{enumerate}
\item Scopus
\item Web of Science
\item Science Direct
\item Institute of Electrical and Electronics Engineers Xplore Digital Library (IEEEXplore)
\item Springer Link
\item Association for Computing Machinery Digital Library (ACM)
\item Google Scholar
\item arXiv
\end{enumerate}
The last search using these databases was conducted on 16 Feb. 2021. After obtaining search results, all studies were analyzed individually to assess their relevance in the context of this SLR. Inclusion and exclusion criteria were used to select or discard the retrieved studies. The inclusion criteria are the following:
\begin{itemize}
\item The study presents a survey of explainable AI.
\item The study presents challenges and/or research directions for XAI.
\end{itemize}
On the other hand, the exclusion criteria are the following:
\begin{itemize}
\item The study is not written in English.
\item The study presents a survey of XAI without discussing any challenges or research directions.
\end{itemize}
The retrieved studies were first analyzed by their titles and abstracts to decide if the paper matched the first inclusion criterion. If matched, the paper was analyzed in detail in the second step. In the second step, the exclusion criteria and the second inclusion criterion were checked.
We reviewed the list of references of the selected studies to include other papers that may not be retrieved from the selected electronic databases, which resulted in retrieving eight non-survey papers that reported challenges and/or research directions in XAI \cite{Aurangzeb,Doshi-Velez2017-yb,fox2017explainable,gunning2019xai,Lipton2018-dx,Molnar2021-do,Preece2018,Ras2018}.
Overall, the total number of selected papers is 58. The majority of the selected papers were published in 2020, as shown in Table \ref{year}. In 2021, we found 10 papers. However, since the last search was in Feb 2021, we expect more publications to appear until the end of the year. As shown in Table \ref{type}, the primary outlet for the selected papers are journal articles followed by conference papers and arXiv papers. The distribution of the selected papers per publisher is shown in Table \ref{publisher}.
\begin{table}[t]
\centering
\caption{Distribution of selected papers per year.\label{year}}
\begin{tabular}{lp{2.5cm}}
\toprule
Year & Number of papers\\
\midrule
2017 & 4\\
2018 & 12\\
2019 & 10\\
2020 & 22\\
2021 & 10\\
Total & 58\\
\bottomrule
\end{tabular}
\\
\end{table}
\begin{table}[t]
\centering
\caption{Distribution of selected papers per publication type.\label{type}}
\begin{tabular}{lp{2.5cm}}
\toprule
Publication type & Number of papers\\
\midrule
Journal articles & 28\\
Conference papers & 12\\
arXiv & 12\\
Book chapter & 5\\
Book & 1\\
Total & 58\\
\bottomrule
\end{tabular}
\\
\end{table}
\begin{table}[!t]
\centering
\caption{Distribution of selected papers per publisher.\label{publisher}}
\begin{tabular}{lp{2.5cm}}
\toprule
Publisher & Number of papers\\
\midrule
Springer & 12\\
arXiv & 12\\
IEEE & 9\\
Elsevier & 7\\
ACM & 6\\
Frontiers & 2\\
MDPI & 1\\
Wiley & 1\\
Sage & 1\\
Others & 7\\
Total & 58\\
\bottomrule
\end{tabular}
\\
\end{table}
\section{Discussion}
This section discusses the challenges and research directions in XAI. In order to place them in a meaningful context, the discussion is organized into two main themes, as shown in Fig. \ref{ml_cycle}. The first theme focuses on the general challenges and research directions in XAI. The second theme is about the challenges and research directions of XAI based on the ML life cycle's phases. For simplicity, we divided the life cycle into three main phases: design, development, and deployment phases. The following subsections are shed light on these challenges and research directions.
\begin{figure}
\centering
\includegraphics[width=\linewidth]{ML_Cycle.png}
\caption{The proposed organization to discuss the challenges and research directions in XAI. For simplicity, the arrows that show the flow in the life cycle are removed.}
\label{ml_cycle}
\end{figure}
\begin{table}
\caption{A summary of the selected papers, categorized by phases as well as challenges and research directions.\label{challenges}}
\begin{tabular}{lp{8.7cm}p{5.3cm}}
\toprule
Phases & Challenges and research directions & Papers\\
\midrule
\multirow{2}{*}{General} &
\hyperlink{Towards_more_formalism}{Towards more formalism} & \cite{Adadi2018,Dosilovic2018-xu,Guidotti2018-tx,Molnar2020,Payrovnaziri2020-xi,Carvalho,Aurangzeb,arrieta2020,Reyes,Samek2019-po,Longo2020,Pocevi020,Mi2020,Burkart2021-tq,Li2020,nunes2017systematic,seeliger2019semantic}, \cite{xie2020explainable,lucieri2020achievements,mueller2019explanation,buhrmester2019analysis,islam2021explainable,Messina} $^*$\\
& \hyperlink{Multidisciplinary}{Multidisciplinary research collaborations} & \cite{MILLER20191,Longo2020,Burkart2021-tq,Kovalerchuk2021,Fan2021,Abdul2018}, \cite{islam2021explainable,vilone2020explainable,lucieri2020achievements} *\\
& \hyperlink{expertise}{Explanations and the nature of user experience and expertise} & \cite{gunning2019xai,Chatzimparmpas,Naiseh2020,Aurangzeb,nunes2017systematic,arrieta2020}, \cite{mueller2019explanation,xie2020explainable,lucieri2020achievements} $^*$\\
& \hyperlink{trustworthiness}{XAI for trustworthiness AI}& \cite{Pocevi020,Ras2018,Markus2021-bp,Chakraborty,islam2021explainable,belle2021principles, Chatzimparmpas,arrieta2020},\cite{xie2020explainable}*\\
& \hyperlink{trade-off}{Interpretability vs. performance trade-off}&
\cite{arrieta2020,gunning2019xai,Aurangzeb,danilevsky2020survey,Longo2020}\\
& \hyperlink{Causal}{Causal explanations} & \cite{Molnar2020,Aurangzeb,belle2021principles,Pocevi020,Moraffah2020}\\
& \hyperlink{Contrastive}{Contrastive and counterfactual explanations} & \cite{Stepin2021}, \cite{mueller2019explanation} $^*$\\
& \hyperlink{heterogeneous}{XAI for non-image, non-text, and heterogeneous data} & \cite{Reyes,Kovalerchuk2021}, \cite{yuan2020explainability,lucieri2020achievements}*\\
& \hyperlink{composition}{Explainability methods composition} & \cite{Aurangzeb,Adadi2018,belle2021principles,Ras2018}, \cite{lucieri2020achievements} $^*$\\
& \hyperlink{models/methods}{Challenges in the existing XAI models/methods} & \cite{Aurangzeb,Molnar2020,Samek2019-po}\\
& \hyperlink{generation}{Natural language generation} & \cite{lucieri2020achievements}*\\
& \hyperlink{models}{Analyzing models, not data}& \cite{Molnar2021-do}\\
& \hyperlink{uncertainties}{Communicating uncertainties} & \cite{Chatzimparmpas,Molnar2020}\\
& \hyperlink{Time}{Time constraints}
& \cite{xie2020explainable,Doshi-Velez2017-yb}*\\
& \hyperlink{Reproducibility}{Reproducibility} & \cite{Payrovnaziri2020-xi}\\
& \hyperlink{economics}{The economics of explanations} & \cite{Adadi2018}\\
\\
\multirow{2}{*}{Design} & \hyperlink{data_quality}{Communicating data quality} & \cite{Markus2021-bp,Burkart2021-tq} \\
& \hyperlink{sharing}{Data sharing} & \cite{Holzinger}\\
\\
\multirow{2}{*}{Development} & \hyperlink{infusion}{Knowledge infusion} & \cite{Choo8402187,Zhang2018,Li2020}, \cite{Messina,zhang2020survey}$^*$ \\
& \hyperlink{training_process}{Developing approaches supporting explaining the training process} & \cite{Choo8402187,Chatzimparmpas,gunning2019xai}\\
& \hyperlink{debugging}{Developing model debugging techniques} & \cite{Zhang2018}, \cite{xie2020explainable}$^*$ \\
& \hyperlink{comparison}{Using interpretability/explainability for models/architectures comparison} & \cite{Chatzimparmpas}\\
& \hyperlink{visual_analytics}{Developing visual analytics approaches for advanced DL architectures} & \cite{Chatzimparmpas,Choo8402187}\\
& \hyperlink{sparsity}{Sparsity of analysis} & \cite{dao2020demystifying}*\\
& \hyperlink{innovation}{Model innovation} &\cite{LIANG2021168,belle2021principles}\\
& \hyperlink{rules}{Rules extraction} & \cite{HE2020346,Abdul2018,Fan2021}\\
& \hyperlink{bayesian}{Bayesian approach to interpretability} & \cite{Chakraborty}\\
& \hyperlink{competencies}{Explaining competencies} & \cite{gunning2019xai}\\
\\
\multirow{2}{*}{Deployment} & \hyperlink{ontologies}{Improving explanations with ontologies} & \cite{Burkart2021-tq} \\
& \hyperlink{privacy}{XAI and privacy} & \cite{Longo2020,Holzinger} \\
& \hyperlink{security}{XAI and security} & \cite{arrieta2020,Tjoa2020,LIANG2021168,Ras2018}\\
& \hyperlink{safety}{XAI and safety} & \cite{HUANG2020100270,NaisehJiang2020,arrieta2020}\\
& \hyperlink{teaming}{Human-machine teaming} & \cite{seeliger2019semantic,Ras2018,gunning2019xai,NaisehJiang2020,Adadi2018,Abdul2018,Choo8402187,Chatzimparmpas}, \cite{islam2021explainable,Messina}* \\
& \hyperlink{agency}{Explainable agency} & \cite{Anjomshoae}\\
& \hyperlink{mexplanation}{Machine-to-machine explanation} & \cite{Preece2018,Weller2019,Adadi2018}\\
& \hyperlink{reinforcement}{XAI and reinforcement learning} & \cite{Lipton2018-dx,Guidotti2018-tx,Wells}, \cite{dao2020demystifying}*\\
& \hyperlink{planning}{Explainable AI planning (XAIP)} & \cite{Adadi2018},\cite{fox2017explainable}*\\
& \hyperlink{recommendation}{Explainable recommendation} & \cite{ZhangINR-066,NaisehJiang2020}\\
& \hyperlink{service}{XAI as a service}& \cite{Molnar2021-do}\\
\bottomrule
\end{tabular}
\\
\footnotesize{$^*$ A non-peer-reviewed paper from arXiv.}
\end{table}
\subsection{General Challenges and Research Directions in XAI}
In this section, we reported the general challenges and research directions in XAI.
\subsubsection{Towards more formalism}
\hypertarget{Towards_more_formalism}
It is one of the most raised challenges in the literature of XAI \cite{Adadi2018,Dosilovic2018-xu,Guidotti2018-tx,Molnar2020,Payrovnaziri2020-xi,xie2020explainable,Carvalho,Aurangzeb,mueller2019explanation,lucieri2020achievements,arrieta2020,Reyes,Samek2019-po,buhrmester2019analysis,Longo2020,Pocevi020,Mi2020,Burkart2021-tq,islam2021explainable,Li2020,nunes2017systematic,seeliger2019semantic,Messina}. It was suggested that more formalism should be considered in terms of systematic definitions, abstraction, and formalizing and quantifying \cite{Adadi2018}.
Starting from the need for systematic definitions, until now, there is no agreement on what an explanation is \cite{Guidotti2018-tx}. Furthermore, it has been found that similar or identical concepts are called by different names and different concepts are called by the same names \cite{Adadi2018,Dosilovic2018-xu}. In addition, without a satisfying definition of interpretability, how it is possible to determine if a new approach better explains ML models \cite{Molnar2020}? Therefore, to facilitate easier sharing of results and information, definitions must be agreed upon \cite{Adadi2018, Dosilovic2018-xu}.
With regards to the abstraction, many works have been proposed in an isolated way; thus, there is a need to be consolidated to build generic explainable frameworks that would guide the development of end-to-end explainable approaches \cite{Adadi2018}. Additionally, taking advantage of the abstraction explanations in identifying properties and generating hypotheses about data-generating processes (e.g., causal relationships) could be essential for future artificial general intelligence (AGI) systems \cite{Dosilovic2018-xu}.
Regarding the formalization and quantification of explanations, it was highlighted in \cite{Adadi2018} that some current works focus on a detailed problem formulation which becomes irrelevant as the method of interpretation or the explanation differs. Therefore, regardless of components that may differ, the expansibility problem must be generalized and formulated rigorously, and this will improve the state-of-the-art for identifying, classifying, and evaluating sub-issues of explainability \cite{Adadi2018}.
Establishing formalized rigorous evaluation metrics need to be considered as well \cite{Adadi2018}. However, due to the absence of an agreement on the definitions of interpretability/explainability, no established approach exists to evaluating XAI results \cite{Payrovnaziri2020-xi}. The lack of ground truth in most cases is the biggest challenge for rigorous evaluations \cite{xie2020explainable,Molnar2020}. So far, different evaluation metrics have been proposed, such as reliability, trustworthiness, usefulness, soundness, completeness, compactness, comprehensibility, human-friendly or human-centered, correctness or fidelity, complexity, generalizability \cite{xie2020explainable}. However, it seems that there are two main evaluation metrics groups: objective and human-centered evaluations \cite{Molnar2020}. The former is quantifiable mathematical metrics, and the latter relies on user studies \cite{Molnar2020}. Further progress are needed towards evaluating XAI techniques' performance and establishing objective metrics for evaluating XAI approaches in different contexts, models, and applications \cite{arrieta2020}. Achieving that may help in developing a model-agnostic framework that can suggest the most appropriate explanation taking into account problem domain, use case, and user's type \cite{Carvalho}.
\subsubsection{Multidisciplinary research collaborations}
\hypertarget{Multidisciplinary}
One area of research that can offer new insights for explainable methods is working closely with researchers from other disciplines such as psychology, behavioral and social sciences, human-computer interaction, physics, and neuroscience. Multidisciplinary research is therefore imperative to promote human-centric AI and expand utilizing XAI in critical applications \cite{islam2021explainable}.
Several studies, for instance \cite{Longo2020,Burkart2021-tq,lucieri2020achievements,Kovalerchuk2021,Fan2021,MILLER20191,Abdul2018}, have been suggested some potential multidisciplinary research works. In \cite{Longo2020}, it has been highlighted that approaching psychology discipline can help to get insights on both the structure and the attributes of explanations and the way they can influence humans. They also have suggested that defining the context of explanations is an important research direction. Here, it is essential to consider the domain of application, the users, type of explanations (e.g., textual, visual, combinations of solutions), and how to provide the explanations to the users. This research direction can form a connection with behavioral and social sciences. The paper in \cite{MILLER20191} also has shown that XAI can benefit from the work in philosophy, cognitive psychology/science, and social psychology. The paper summarizes some findings and suggests ways to incorporate these findings into work on XAI.
Approaching HCI studies are essential to XAI. However, few user experiments have been conducted in the area of explainability \cite{Burkart2021-tq}. Therefore, more should be conducted to study the topic adequately \cite{Burkart2021-tq}. Humans must be included in the process of creating and utilizing XAI models, as well as enhancing their interpretability/explainability \cite{Longo2020}. In \cite{lucieri2020achievements}, it has been highlighted that interactive tools may help users understand, test, and engage with AI algorithms, thereby developing new approaches that can improve algorithms' explainability. Furthermore, interactive techniques can help users to interpret predictions and hypothesis-test users' intuitions rather than relying solely upon algorithms to explain things for them. In \cite{Abdul2018}, it has been suggested drawing from the HCI research on interaction design and software learnability to improve the usability of intelligible or explainable interfaces. Additionally, HCI researchers can take advantage of the theoretical work on the cognitive psychology of explanations to make understandable explanations. They can also empirically evaluate the effectiveness of new explanation interfaces.
The advances in neuroscience should be of great benefit to the development and interpretation of DL techniques (e.g., cost function, optimization algorithm, and bio-plausible architectural design) owing to the close relationship between biological and neural networks \cite{Fan2021}. It is imperative to learn from biological neural networks so that better and explainable neural network architectures can be designed \cite{Fan2021}. Finally, connecting with physics and other disciplines that have a history of explainable visual methods might provide new insights for explainable methods \cite{Kovalerchuk2021}.
\subsubsection{Explanations and the nature of user experience and expertise}
\hypertarget{expertise}
Based on the nature of the application, users who use ML models can vary (e.g., data scientists, domain experts, decision-makers, and non-experts). The nature of user experience and expertise matters in terms of what kind of cognitive chunks they possess and the complexity they expect in their explanations \cite{Doshi-Velez2017-yb}. In general, users have varying backgrounds, knowledge, and communication styles \cite{Doshi-Velez2017-yb}. However, it seems that the current focus of explanation methods is tailored to users who can interpret the explanations based on their knowledge in the ML process \cite{Ras2018,xie2020explainable}.
The works in \cite{xie2020explainable,mueller2019explanation,gunning2019xai,Chatzimparmpas,Naiseh2020,Aurangzeb,nunes2017systematic} have highlighted what is needed to be considered with regards to explanations and the nature of user experience and expertise. In \cite{xie2020explainable}, user-friendly explanations have been suggested so users can interpret the explanations with less technical knowledge. Therefore, figuring out what to explain should follow the identification of the end-user. In \cite{mueller2019explanation}, it has been highlighted that previous works in explainable AI systems (e.g., expert systems) generally neglected to take into account the knowledge, goals, skills, and abilities of users. Additionally, the goals of users, systems, and explanations were not clearly defined. Therefore, clearly stating goals and purposes are needed to foster explanations testing within the appropriate context. In \cite{Naiseh2020}, the authors have discussed that identifying the users' goals and keeping up with their dynamic nature means collecting more data from them. It is also essential to develop changes detection approaches of goals and needs for the purpose of adapting these changes to end-users. For a deeper understanding of these dynamics, user studies (e.g., diary studies, interviews, and observation) can help develop guidelines for developing long-term explainable systems and determining which user data to gather to improve personalization.
In \cite{gunning2019xai}, it has been suggested that abstraction can be used to simplify the explanations. Understanding how abstractions are discovered and shared in learning and explanation is an essential part of the current XAI research. The work in \cite{Aurangzeb} has mentioned that the inclusion of end-users in the design of black-box AI models is essential, especially for specific domains, e.g., the medical domain. That would help to understand better how the end-users will use the outputs and interpret explanations. It is a way to educate them about the predictions and explanations produced by the system. In \cite{Chatzimparmpas}, the authors have discussed that utilizing users' previous knowledge is a significant challenge for visualization tools today. Customizing visualization tools for different user types can be useful at several stages of the ML model pipeline. However, to use prior users' knowledge in predictive models, it is important to establish processes to digitally capture and quantify their knowledge.
In \cite{lucieri2020achievements}, it has been mentioned that DL models often use concepts that are unintelligible to predict outcomes. Therefore, using systems that use such models requires human-centric explanations that can accurately explain a decision and make sense to the users (e.g., medical domain expert). An approach to come with human-centric explanations is examining the role of human-understandable concepts acquired by DL models. It is also essential to analyze the features used by the DL models in predicting correct decisions, but based on incorrect reasoning. Having an understanding of the model’s concepts would help reduce reliability concerns and develop trust when deploying the system, especially in critical applications. The authors also highlighted the importance of addressing the domain-specific needs of specific applications and their users when developing XAI methods. Finally, the work in \cite{arrieta2020} has discussed that XAI can facilitate the process of explaining to non-experts how a model reached a given decision, which can substantially increase information exchange among heterogeneous people regarding the knowledge learned by models, especially when working in projects with multi-disciplinary team.
To sum up, it is crucial to tailor explanations based on user experience and expertise. Explanations should be provided differently to different users in different contexts \cite{royal2019}. In addition, it is also essential to clearly define the goals of users, systems, and explanations. Stakeholder engagement and system design are both required to understand which explanation type is needed \cite{royal2019}.
\setcounter{footnote}{0}
\subsubsection{XAI for trustworthiness AI}
\hypertarget{trustworthiness}
Increasing the use of AI in everyday life applications will increase the need for AI trustworthiness, especially in situations where undesirable decisions may have severe consequences \cite{xie2020explainable}. The High-Level Expert Group in European Commission put seven essentials for achieving trustworthy AI \footnote{\url{https://ec.europa.eu/commission/presscorner/detail/en/IP\_19\_1893}}: (1)\emph{human agency and oversight}; (2) \emph{robustness and safety}; (3) \emph{privacy and data governance}; (4) \emph{transparency}; (5) \emph{diversity, non-discrimination and fairness}; (6) \emph{societal and environmental well-being}; and (7) \emph{accountability}. The discussion about privacy, security, and safety are given in \hyperlink{privacy}{XAI and Privacy} Section, \hyperlink{security}{XAI and Security} Section, and \hyperlink{safety}{XAI and Safety} Section, respectively. The discussion in this section is for what is reported in the selected papers regarding fairness and accountability.
With regards to fairness, ML algorithms must not be biased or discriminatory in the decisions they provide. However, with the increased usage of ML techniques, new ethical, policy, and legal challenges have also emerged, for example, the risk of unintentionally encoding bias into ML decisions \cite{Chakraborty}. Meanwhile, the opaque nature of data mining processes and the complexity of ML make it more challenging to justify consequential decisions \cite{Chakraborty}. The work in \cite{islam2021explainable} argues that data, algorithmic, and social biases need to be remedied in order to promote fairness. Further, it is imperative to be able to analyze AI systems to have trust in the model and its predictions, especially for some critical applications. Researchers started trying to form a definition of fairness and the meaning of fairness in an algorithm as discussed in \cite{Chakraborty}. According to \cite{Chakraborty}, it would also be necessary to devise new techniques for discrimination-aware data mining. It is also worth noting that when converting fairness into a computational problem, we need to keep the fairness measures fair \cite{Chakraborty}. The work in \cite{Ras2018} states that it is possible to visualize learned features using XAI methods and assess bias using methods other than explanation methods. On the other hand, regulations and laws are necessary for the suspicion about unfair outcomes \cite{Ras2018}.
Having accountability means having someone responsible for the results of AI decisions if harm occurs. In \cite{royal2019}, it has been mentioned that investigating and appealing decisions with major consequences for people is an important aspect of systems of accountability, and some current regulations also aim to achieve this. XAI can be an important factor in systems of accountability by providing users with the means to appeal a decision or modify their behavior in the future to achieve a better result. However, more work should be done to establish an environment that promotes individual autonomy and establish a system of accountability. It has also been discussed in \cite{Chakraborty} that developing procedures for testing AI algorithms for policy compliance is necessary so that we can establish whether or not a given algorithm adheres to a specific policy without revealing its proprietary information. It is also desirable for a model to specify its purposes and provide external verification of whether these goals are met and, if not, describe the causes of the predicted outcomes.
The use of XAI can enhance understanding, increase trust, and uncover potential risks \cite{islam2021explainable}. Therefore, when designing XAI techniques, it is imperative to maintain fairness, accountability, and transparency \cite{islam2021explainable}. On the other hand, it is necessary to highlight that not only black-box AI models are vulnerable to adversarial attacks, but also XAI approaches \cite{slack2020fooling}. There is also a risk that to promote trust in black-box AI models predictions; explainers may be more persuasive but misleading than informative, so users may become deceived, thinking the system to be trustworthy \cite{Pocevi020,royal2019}. It is possible to increase trust through explanations, but explanations do not always produce systems that produce trustworthy outputs or ensure that system implementers make trustworthy claims about its abilities \cite{royal2019}.
The work in \cite{Markus2021-bp} discusses measures to create trustworthy AI. It has been highlighted that before employing AI systems in practice, it is essential to have quantitative proxy metrics to assess explanation quality objectively, compare explanation methods, and complement them with human evaluation methods (e.g., data quality reporting, extensive testing, and regulation).
Finally, it is good to note that a further explore the idea of Responsible AI with a discussion about principles of AI, fairness, privacy, and data fusion can be found in \cite{arrieta2020}.
\subsubsection{Interpretability vs. performance trade-off}
\hypertarget{trade-off}
The belief that complicated models provide more accurate outcomes is not necessarily correct \cite{rudin2019stop}. However, this can be incorrect in cases when the given data is structured and with meaningful features \cite{rudin2019stop}. In a situation where the function being approximated is complex, that the given data is widely distributed among suitable values for each variable and the given data is sufficient to generate a complex model, the statement \emph{``models that are more complex are more accurate''} can be true \cite{arrieta2020}. In such a situation, the trade-off between interpretability and performance becomes apparent \cite{arrieta2020}.
When the performance is coupled with model complexity, model interpretability is in question \cite{arrieta2020}. Explainability techniques, however, could help in minimizing the trade-off \cite{arrieta2020}. However, according to \cite{Aurangzeb}, what determines this trade-off? and who determines it? The authors have highlighted the importance of discussing with end-users this trade-off so that they can be aware of the potential risks of misclassification or opacity. Another point that should be considered is the approximation dilemma: models need to be explained in enough detail and in a way that matches the audience for whom they are intended while keeping in mind that explanations reflect the model and do not oversimplify its essential features \cite{arrieta2020}. Even though studying the trade-off is essential, it is impossible to proceed without standardized metrics for assessing the quality of explanations \cite{danilevsky2020survey}.
Another possible solution for the trade-off is suggested in \cite{Longo2020} which is developing fully transparent models throughout the entire process of creation, exploitation, and exploration and can provide local and global explanations. In turn, this leads to using methods that embed learning capabilities to develop accurate models and representations. The methods should also be able to describe these representations in effective natural language consistent with human understanding and reasoning.
\subsubsection{Causal explanations}
\hypertarget{Causal}
Developing causal explanations for AI algorithms (i.e., why they made those predictions instead of how they arrived at those predictions) can help increasing human understanding \cite{Pocevi020}. In addition, causal explanations strengthen models' resistance to adversarial attacks, and they gain more value when they become part of decision-making \cite{Molnar2020}. However, there can be conflicts between predicting performance and causality \cite{Molnar2020}. For example, when the confounder, which is a variable that influences both the dependent variable and independent variable, is missing from the model \cite{Molnar2020}.
Causal explanations are anticipated to be the next frontier of ML research and to become an essential part of the XAI literature \cite{Aurangzeb,belle2021principles}. There is a need for further research to determine when causal explanations can be made from an ML model \cite{Molnar2020}. In addition, according to a recent survey on causal interpretability for ML \cite{Moraffah2020}, it has been highlighted the absence of ground truth data for causal explanations and verification of causal relationships make evaluating causal interpretability more challenging. Therefore, more research is needed to guide on how to evaluate causal interpretability models \cite{Moraffah2020}.
\subsubsection{Contrastive and counterfactual explanations}
\hypertarget{Contrastive}
Contrastive explanations describe why one event occurred but not another, while counterfactual explanations describe what is needed to produce a contrastive output with minimal changes in the input \cite{Stepin2021}. Questions in the contrastive form "Why x and not y?" and questions of the counterfactual form "What if?" and "What would happen if?" \cite{mueller2019explanation}.
In a recent survey of contrastive and counterfactual explanations \cite{Stepin2021}, it has been found that contrastive and counterfactual explanations help improve the interaction between humans and machines and personalize the explanation of algorithms. A further important point as observed by \cite{Stepin2021} that one of the significant barriers towards a fair assessment of new frameworks is the lack of standardization of evaluation methods. The theoretical frameworks are also found inadequate for applying to XAI as a result of the disconnect between the philosophical accounts of counterfactual explanation to scientific modeling as well as ML-related concepts. Furthermore, it has been found that different domains of science define counterfactual explanations differently, as do the approaches used to solve specific tasks.
In the light of possible research directions in this point, it has been suggested in \cite{Stepin2021} the importance of including end-users in the evaluation of generated explanations since these explanations are designed to be user-oriented. In addition, since contrastive and counterfactual explanations address causal and non-causal relationships, new horizons open to the XAI community by unifying causal and non-causal explanatory engines within a contfactually-driven framework. Furthermore, bringing together researchers from the humanities and the computational sciences could contribute to further development for contrastive and counterfactual explanations generation.
\subsubsection{XAI for non-image, non-text, and heterogeneous data}
\hypertarget{heterogeneous}
The focus of XAI works is mainly on image and text data. However, other data types exist but until now not well explained, such as sequences, graphs, and Spatio-temporal data.
Using visualization to transform non-image data into images creates opportunities to discover explanations through salient pixels and features \cite{Kovalerchuk2021}. However, this should not be the only way for explainability for non-image or non-text data. For example, existing explanation approaches for image or text data need to be adjusted to be used with graph data (without any transformation) \cite{yuan2020explainability}. Additionally, there is a need to develop new approaches for explaining the information that exists with non-image or non-text data, e.g., structural information for graph data \cite{yuan2020explainability}.
Finally, with the advent of AI systems that use various types of data, explainability approaches that can handle such heterogeneity of information are more promising \cite{Reyes}. For example, such systems can simulate clinicians' diagnostic processes in the medical domain where both images and physical parameters are utilized to make decisions \cite{lucieri2020achievements}. Thus, they can enhance the diagnostic effectiveness of the systems as well as explain phenomena more thoroughly \cite{lucieri2020achievements}.
\subsubsection{Explainability methods composition}
\hypertarget{composition}
For specific applications in healthcare (e.g., predicting disease progression), several types of explanations at different levels are needed (e.g., local and global explanations) \cite{Aurangzeb} in order to provide the most complete and diverse explanations we can \cite{lucieri2020achievements}. This is derived from the way clinicians communicate decisions utilizing visualizations and temporal coherence as well as textual descriptions \cite{lucieri2020achievements}.
Some overlap exists between explainability methods, but for the most part, each seems to address a different question \cite{belle2021principles}. According to \cite{Adadi2018}, combining various methods to obtain more powerful explanations is rarely considered. In addition, rather than using disparate methods separately, we should investigate how we can use them as basic components that can be linked and synergized to develop innovative technologies \cite{Adadi2018}. It is argued that enabling composability in XAI may contribute to enhancing both explainability and accuracy \cite{Adadi2018}. Furthermore, it could help to provide answers in a simple human interpretable language \cite{Ras2018}. First efforts, as cited in \cite{lucieri2020achievements}, have been recently made as in \cite{Park_2018_CVPR} where the authors proposed a model that can provide visual relevance and textual explanations.
\subsubsection{Challenges in the existing XAI models/methods}
\hypertarget{models/methods}
There are some challenges in the existing XAI models/methods that have been discussed in the literature. Starting with scalability, which is a challenge that exists in explainable models as discussed in \cite{Aurangzeb}. For example, each case requiring an explanation entails creating a local model using LIME explainable model \cite{ribeiro2016should}. The scalability can be an issue when there is a huge number of cases for which prediction and explanation are needed. Likewise, when computing Shapley values \cite{vstrumbelj2014explaining}, all combinations of variables must be considered when computing variable contributions. Therefore, such computations can be costly for problems that have lots of variables.
Feature dependence presents problems in attribution and extrapolation \cite{Molnar2020}. If features are correlated, attribution of importance and features effects becomes challenging. For sensitivity analyses that permute features, when the permuted feature has some dependence on another feature, the association breaks, resulting in data points outside the distribution, which could cause misleading explanations.
In \cite{Samek2019-po}, the authors discussed some limitations with heatmaps explanations. Heatmaps explanations visualize what features are relevant for making predictions. However, the relation between these features, either individually or in combination, remains unclear. Low abstraction levels of explanations are another limitation. Heatmaps highlight that specific pixels are significant without indicating how the relevance values relate to abstract concepts in the image, such as objects or scenes. The model's behavior can be explained in more abstract, more easily understood ways by meta-explanations that combine evidence from low-level heatmaps. Therefore, further research is needed on meta-explanations.
Model-based (i.e., ante-hoc models) and post-hoc explainability models have some challenges, as have been discussed in \cite{Murdoch22071}. When model-based methods cannot predict with reasonable accuracy, practitioners start the search for more accurate models. Therefore, one way to increase the usage of model-based methods is to develop new modeling methods that maintain the model's interpretability and render more accurate predictions. More details about this direction are provided in \cite{rudin2019stop}. Further, for model-based methods, there is a need to develop more tools for feature engineering. It is possible to achieve comparable predictive accuracy in some applications when the features are more informative and meaningful. Therefore, enhancing the possibility of model-based methods can be accomplished by producing more useful features. Two categories of works can help achieve that: improve tools for exploratory data analysis and improve unsupervised techniques. The former helps to understand the data, and domain knowledge could help to identify helpful features. The latter is needed because unsupervised techniques are often used to identify relevant structures automatically, so advances in unsupervised techniques may result in better features.
The authors in \cite{Murdoch22071} have also discussed some challenges for post-hoc explainability models. According to the authors, it is challenging to determine what format or combination of formats will adequately describe the model's behavior. Furthermore, there is uncertainty over whether the current explanation methods are adequate to capture a model’s behavior or novel methods still needed. Another challenge is if post-hoc explanations methods identify learned relationships by the model that practitioners know to be incorrect, is it possible that practitioners fix these relationships learned and increase the predictive accuracy? Further research in pos-thoc explanations can help exploit prior knowledge to improve the predictive accuracy of the models.
\subsubsection{Natural language generation}
\hypertarget{generation}
Explaining in natural language needs to be accurate, useful, and easy to understand \cite{reiter-2019-natural}. Furthermore, in order to produce good quality explanations, the generated explanations need to be tailored to a specific purpose and audience, be narrative and structured, and communicate uncertainty and data quality that could affect the system's output \cite{reiter-2019-natural}.
Four challenges that are crucial in generating good quality explanations have been discussed in \cite{reiter-2019-natural}:
\begin{itemize}
\item Evaluation challenge: Develop inexpensive but reliable ways of estimating scrutability, trust, etc. Do we have a chance to obtain reliable results if we ask users to read explanations and estimate, for example, scrupability? What experimental design gives the best results? Before we do these steps, should we make sure the explanations are accurate?
\item Vague Language challenge: Using vague terms in explanations is much easier to understand by humans because they think in qualitative terms \cite{van2012not}. However, how can vague language be used in explanations, such that the user does not interpret it in a way that will lead to a misunderstanding of the situation? In addition, setting the priority of messages based upon features and concepts that the user is aware of would be helpful. Furthermore, phrasing and terminology used should be intuitive to users.
\item Narrative challenge: Explaining symbolic reasoning narratively is more straightforward to comprehend than numbers and probabilities \cite{daniel2017thinking}. Therefore, we need to develop algorithms for creating narrative explanations to present the reasoning.
\item Communicating data quality challenge: Techniques should be developed to keep users informed when data problems affect results. We have discussed this issue in detail in \hyperlink{data_quality}{Communicating Data Quality} Section.
\end{itemize}
Another challenge has been discussed in \cite{lucieri2020achievements}. In some medical domains, it could be necessary for AI systems to generate long textual coherent reports to mimic the behavior of doctors. The challenge here is that after generating a few coherent sentences, language generation models usually start producing seemingly random words that have no connection to previously generated words. One of the solutions to this problem would be to use transformer networks \cite{vaswani2017attention} as language model decoders, which can capture word relationships in a longer sentence. In order to evaluate the generated reports, it is essential to compare them with human-generated reports. However, since human-generated reports are usually free-text reports (i.e., not following any specific template), it is important to first eliminate unnecessary information for the final diagnosis from human-generated reports then conduct the comparison.
\subsubsection{Analyzing models, not data}
\hypertarget{models}
The author in \cite{Molnar2021-do} has discussed that analyzing models instead of data is a possible future of ML interpretability. It has been mentioned that a possible way to extract knowledge from data is through interpretable ML. That is because an ML model can automatically identify if and how features are important to predicting outcomes and recognize how relationships are represented.
He added that there is a need to move from analyzing assumption-based data models to analyzing assumption-free black-box AI models. That's because making assumptions about the data (i.e., distribution assumptions) is problematic. Typically, they are wrong (do not follow the Gaussian distribution), hard to check, extremely inflexible, and difficult to automate. Further, assumption-based data models in many domains are typically less predictive than black-box AI models (i.e., generalization) when having lots of data, which is available due to digitization. Therefore, the author has argued that there should be a development of all the tools that statistics offer for answering questions (e.g., hypothesis tests, correlation measures, interaction measures) and rewrite them for black-box AI models. To some extent, this is already taking place. For example, in a linear model, the coefficients quantify the effects of an individual feature on the result. The partial dependent plot \cite{friedman2001greedy} represents this idea in a more generalized form.
\subsubsection{Communicating uncertainties}
\hypertarget{uncertainties}
Communicating uncertainty is an important research direction because it can help to inform the users about the underlying uncertainties in the model and explanations. According to \cite{Chatzimparmpas}, there are already inherent uncertainties in ML models; and model refinement efforts by developers may introduce new uncertainties (e.g., overfitting). Furthermore, some explanation methods such as permutation feature importance and Shapley value give explanations without measuring the uncertainty implied by the explanations \cite{Molnar2020}.
Quantifying uncertainty is an open research topic \cite{Chatzimparmpas}. However, some works exist towards quantifying uncertainty as discussed in \cite{Molnar2020}. The uncertainty surrounding ML models can take many forms and occur throughout the ML life phases \cite{Chatzimparmpas}. Therefore, in order to make progress, it is needed to become more rigorous in studying and reliably quantifying uncertainties at model's various phases and with the explanation methods and communicate them to the users, then users can respond accordingly \cite{Chatzimparmpas,Molnar2020}.
\subsubsection{Time constraints}
\hypertarget{Time}
Time is an essential factor in producing explanations and in interpretation. Some explanations must be produced promptly to let the user react to the decision \cite{xie2020explainable}. Producing explanations efficiently can save computing resources, thereby making it useful for industrial use or in environments with limited computing capability \cite{xie2020explainable}. In some situations (e.g., plant operation application), the provided explanations need to be understood quickly to help the end-user to make a decision \cite{Doshi-Velez2017-yb}. On the other hand, in some situations (e.g., scientific applications), users would likely be willing to devote considerable time understanding the provided explanation \cite{Doshi-Velez2017-yb}. Therefore, time is an essential factor considering the situation, available resources, and end-users.
\subsubsection{Reproducibility}
\hypertarget{Reproducibility}
In a recent review of XAI models based on electronic health records, it has been found that research reproducibility was not stressed well in the reviewed literature, though it is paramount \cite{Payrovnaziri2020-xi}. In order to facilitate comparisons between new ideas and existing works, researchers should use open data, describe the methodology and infrastructure they used, and share their code \cite{Payrovnaziri2020-xi}. In addition, it has been suggested that publication venues should establish reproducibility standards that authors must follow as part of their publication process \cite{Payrovnaziri2020-xi}.
\subsubsection{The economics of explanations}
\hypertarget{economics}
Research into the economic perspective of XAI is sparse, but it is essential \cite{Adadi2018}. With the pressures of social and ethical concerns about trusting black-box AI models, XAI has the potential to drive a real business value \cite{Adadi2018}. XAI, however, comes at a cost \cite{doshi2017accountability}.
Recently, the work in \cite{beaudouin2020flexible} identified costs of explanations in seven main categories (1) costs of explanations design, (2) costs of creating and storing audit logs, (3) costs of trade secrets violation (e.g., the forced disclosure of source code), (4) costs of slowing down innovation (e.g., increasing time-to-market), (5) costs of reducing decisional flexibility if the future situation does not justify the previous explanation, (6) cost of conflict with security and privacy matters, and (7) costs of using less efficient models for their interpretability. Therefore, costs associated with algorithmic explanations should be incurred when the benefits of the explanations outweigh the costs \cite{doshi2017accountability}.
Cost estimation is one of the issues that should be addressed by encouraging economic interpretations. Other issues include algorithms proprietary, revealing trade secrets, and predicting XAI market evolution \cite{Adadi2018}.
\subsection{Challenges and Research Directions of XAI in the Design Phase}
In this phase, the data is collected from at least one source. Then, the data preparation step is done to prepare the collected data for the training phase. By grouping what was collected from the selected papers, we identified two main challenges that need further research: communicating data quality and data sharing.
\subsubsection{Communicating data quality}
\hypertarget{data_quality}
The provided explanations for the AI system or its outcomes depend on the data used to build the system. Data bias, data incompleteness, and data incorrectness are issues that affect the quality of the data. Training AI systems using low-quality data will be reflected in their outcomes \cite{reiter-2019-natural}. For example, an AI system developed for lung cancer risks prediction using data from Americans may not accurately estimate risks for a resident of Delhi due to the differences in polluted environments in which they are living at \cite{reiter-2019-natural}. So, what can be of high quality for a particular purpose can be of low quality for another \cite{Markus2021-bp}. Reducing system accuracy is not the only consequence of building an AI system using low-quality data; producing unfair decisions and degrading the explainability of the AI system are other possible consequences.
With this in mind, it has been suggested to be aware of how data was collected and any limitations associated with the collected data \cite{royal2019}. Further, it has been highlighted the importance of clarifying any data issues that can reduce accuracy when producing explanations \cite{reiter-2019-natural}. However, how can we communicate data quality to users to let them know how the results are influenced by data used.
In \cite{ahmad2019challenge}, the authors discussed several issues that arise when producing explanations for AI models that use imputation of missing data. They recommended disclaimers accompanied by the derived explanations and educating end-users about the risks involved of incorrect explanations. Even though it is good to come with appropriate disclaimers, we believe that future studies should be undertaken to develop a practical and measurable way to communicate data quality to users. Proposing dimensions of data quality could be the basis for that. We recommend starting with the following questions which is inspired from the work in \cite{dama2020}:
\begin{itemize}
\item Which essential dimensions of data quality are wanted?
\item What are the definitions of those dimensions? and how to measure them?
\item How to deal with them to improve the AI model and hence its explanations?
\item How to communicate them (and highlight any possible risks)?
\end{itemize}
According to \cite{Burkart2021-tq}, there is a variety of data quality dimensions such as completeness, accuracy, and consistency. For an extensive list of dimensions of data quality that occur in information systems, the reader may refer to the research paper in \cite{dama2020}. The fairness dimension can also be included, which may include demographic parity differences. It is essential to highlight that the way that can be used to communicate data quality can vary based on the type of users.
\subsubsection{Data sharing}
\hypertarget{sharing}
Data privacy and data security are two major issues concerning XAI. Since AI is used as a data-driven method and because any requested explanations depend on data used to build AI systems, two main aspects related to data should be considered: data sharing and data preservation.
Data sharing in this context means making raw data available to be used by other partners \cite{Holzinger}. Data preservation is the retention of raw data, at least until we stop using the AI solution. The discussion here focuses on data sharing, which is related to the data collection and preparation phase. Challenges related to data preservation are discussed later in \hyperlink{privacy}{XAI and Privacy} Section.
Data sharing is a significant challenge in many data-driven solutions \cite{Holzinger}. Two common strategies are used for data sharing between the partners: share raw data directly or send it to a central analysis repository \cite{Holzinger}. At this phase, no explanations that may violate privacy are demanded because the solutions are still not developed yet. However, it is essential for users (e.g., patients) to feel confident that their data is secured and protected from unauthorized access and misuse, and any processes are limited to the part of data that they have consented to \cite{Holzinger}. According to \cite{Holzinger}, the implementation of watermarking or fingerprinting are typical reactive techniques used to deal with this issue. Watermarking techniques prove the authenticity and ownership of a dataset, while fingerprinting techniques help to identify the data leak because partners receive the same basic set but marked differently with their fingerprints \cite{Holzinger}.
Federated learning can be a possible solution to avoid raw data sharing. Federated learning allows building ML models using raw data distributed across multiple devices or servers \cite{mcmahan2017communication,Wahab}. As described in \cite{Wahab}, the training using federated learning starts by sending initial model parameters by the central server, which are obtained after a few training iterations, to a set of clients. Then, each client uses its resources to train an ML model locally on its own dataset using the shared parameters. Afterward, each client sends the server an updated version of the parameters. As a result of aggregating clients' parameters, the server creates a global model. As soon as the global model reaches a certain accuracy level, the training process is stopped.
Even though the data never leaves the user’s device, increasing the number of clients involved in a collaborative model makes it more susceptible to inference attacks intended to infer sensitive information from training data \cite{Federated1,Wahab}. Possible research directions to deal with privacy challenges of federated learning have been discussed in \cite{Wahab} such as privacy-preserving security assurance, defining optimal bounds of noise ratio, and proposing granular and adaptive privacy solutions.
\subsection{Challenges and Research Directions of XAI in the Development Phase}
There are three main types of learning in ML: supervised, unsupervised, and reinforcement learning. In supervised learning, a learning algorithm is used to train an ML model to capture patterns in the training data that map inputs to outputs. With unsupervised learning, which is used when only the input data is available, an ML model is trained to describe or extract relationships in the training data. For reinforcement learning, an ML model is trained to make decisions in a dynamic environment to perform a task to maximize a reward function. In the following subsections, we discuss the challenges and research directions during developing ML models.
\subsubsection{Knowledge infusion}
\hypertarget{infusion}
A promising research direction is incorporating human domain knowledge into the learning process (e.g., to capture desired patterns in the data). According to \cite{Messina}, understanding how experts analyze images and which regions of the image are essential to reaching a decision could be helpful to come with novel model architectures that mimic that process. Furthermore, our explanations can be better interpretable and more informative if we use more domain/task-specific terms \cite{zhang2020survey}.
Recently, the work in \cite{XIE2021101985} highlights various ways of incorporating approaches for medical domain knowledge with DL models such as transfer learning, curriculum learning, decision level fusion, and feature level fusion. According to that survey, it was seen that with appropriate integrating methods, different kinds of domain knowledge could be utilized to improve the effectiveness of DL models. A review focused on knowledge-aware methods for XAI is given by \cite{Li2020}. Based on the knowledge source, two categories are identified: knowledge methods and knowledge-based methods. Unstructured data is used as a knowledge source in knowledge methods, while knowledge-based methods use structured knowledge to build explanations. According to \cite{Li2020}, when we use external domain knowledge, we are able to produce explanations that identify important features and why they matter. As concluded in that survey, many questions remain unanswered regarding utilizing external knowledge effectively. For instance, in a vast knowledge space, how can relevant knowledge be obtained or retrieved? To demonstrate this point, let us take the Human-in-the-loop approach as an example. Typically, a user has a wide range of knowledge in multiple domains; thus, the XAI system must ensure that the knowledge provided to the user is desirable.
Recent works in the knowledge that can be incorporated during training ML are given in \cite{Fe-Fei,Gaur9357868,Zhang2018}. In \cite{Fe-Fei}, a one-shot learning technique was presented for incorporating knowledge about object categories, which may be obtained from previously learned models, to predict new objects when very few examples are available from a given class. Another work in \cite{Gaur9357868} has shown how knowledge graph is integrated into DL using knowledge-infused learning and presented examples on how to utilize knowledge-infused learning towards interpretability and explainability in education and healthcare. The work in \cite{Zhang2018} has mentioned that the middle-to-end learning of neural networks with weak supervision via human-computer interaction is believed to be a fundamental research direction in the future.
Based on all that, it can be seen that using XAI to explain the outcomes of the models (e.g., pointing which regions of the image were used to reach the decision) can help to understand better what was learned from the incorporated human knowledge. Thus, it would help to adjust the way used in incorporating the knowledge or come with innovations in model architectures. Furthermore, it could be used to confirm whether a model follows the injected knowledge and rules, especially with critical applications, e.g., autonomous driving model \cite{Choo8402187}. Therefore, more research is needed to investigate how experts can interact with ML models to understand them and improve their abilities, which would be a promising direction in which XAI can contribute.
\subsubsection{Developing approaches supporting explaining the training process}
\hypertarget{training_process}
Training ML models, especially DL, is a lengthy process that usually takes hours to days to finish, mainly because of the large datasets used to train the models \cite{Choo8402187}. Therefore, researchers and practitioners have contributed to developing systems that could help steer the training process and develop better models.
\setcounter{footnote}{0}
Examples of progressive visual analytics systems are cited in \cite{Choo8402187}. For example, DeepEyes \cite{Pezzotti} is an example of a progressive visual analytics system that enables advanced analysis of DNN models during training. The system can identify stable layers, identify degenerated filters that are worthless, identify inputs that are not processed by any filter in the network, reasons on the size of a layer, and it helps to decide whether more layers are needed or eliminate unnecessary layers. DGMTracker is another example \cite{Liu2018} which is developed for better understanding and diagnosing the training process of deep generative models (DGMs). In addition, big tech companies such as Google and Amazon have developed toolkits to debug and improve the performance of ML models such as Tensor-Board\footnote{\url{https://www.tensorflow.org/tensorboard}} and SageMaker Debugger\footnote{\url{https://aws.amazon.com/sagemaker/debugger/}}.
Future studies to deal with this challenge are therefore recommended in order to develop XAI approaches supporting the online training monitoring to get insights that could help to steer the training process by the experts, which could help in developing better models and minimizing time and resources \cite{Chatzimparmpas,Choo8402187}.
\subsubsection{Developing model debugging techniques}
\hypertarget{debugging}
The model is already trained at this stage, and we want to discover any problems that can limit its predictions. The debugging of ML models is paramount for promoting trust in the processes and predictions, which could result in creating new applications \cite{hall2019proposed,Zhang2018}, e.g., visual applications for CNN. A variety of debugging techniques exists, including model assertion, security audit, variants of residual analysis and residual explanation, and unit tests \cite{hall2019proposed}. According to \cite{xie2020explainable}, understanding what causes errors in the model can form the foundation for developing interpretable explanations. The next step is developing more model debugging techniques and combining them with explanatory techniques to provide insight into the model's behavior, enhance its performance, and promote trust \cite{hall2019proposed}.
\subsubsection{Using interpretability/explainability for models/architectures comparison}
\hypertarget{comparison}
It is widely known that the performance of ML models/architectures varies from one dataset/task to another \cite{Chatzimparmpas}. Usually, error performance metrics are used for the comparison to choose the suitable model/architecture for the given dataset/task and to decide how to combine models/architectures for better performance \cite{Chatzimparmpas}. However, even if the models may have the same performance, they can use different features to reach the decisions \cite{Samek2017-lb}. Therefore, the interpretability/explainability of models can be helpful for models/architectures comparison \cite{Samek2017-lb}. It could even be said that the better we understand models' behavior and why they fail in some situations, the more we can use those insights to enhance them \cite{Samek2017-lb}. In the future, it is expected that explanations will be an essential part of a more extensive optimization process to achieve some goals such as improving a model’s performance or reducing its complexity \cite{Samek2019-po}. Further, XAI can be utilized in models/architectures comparison.
\subsubsection{Developing visual analytics approaches for advanced DL architectures}
\hypertarget{visual_analytics}
While visual analytic approaches have been developed for basic DL architectures (e.g., CNNs and RNNs), advanced DL architectures have yet to be addressed in this way (e.g., ResNet \cite{he2016deep} and DenseNet \cite{huang2017densely}) \cite{Chatzimparmpas,Choo8402187}. Advanced DL architectures pose several challenges for visual analytic and information visualization communities due to their large number of layers, the complexity of network design for each layer, and the highly connected structure between layers \cite{Choo8402187}. Therefore, developing efficient visual analytics approaches for such architectures in order to increase their interpretability as well as the explainability of their results is needed \cite{Chatzimparmpas,Choo8402187}.
\subsubsection{Sparsity of analysis}
\hypertarget{sparsity}
Interpreting and validating the reasoning behind a neural network classifier requires examining the saliency maps of various samples from input data, which can be a challenging task if there are a vast number of samples \cite{dao2020demystifying}. Therefore, the number of visualizations that a user has to analyze should be as small as possible to reduce the sparsity of the analysis \cite{dao2020demystifying}. A way to achieve that can be developing novel methods to identify a meaningful subset of the entire dataset to interpret; then, by using this meaningful subset, it is needed to come up with an interpretation of the relationship between various samples and various subsets \cite{dao2020demystifying}.
\subsubsection{Model innovation}
\hypertarget{innovation}
By explaining DL models, we can gain a deeper understanding of their internal structure and can lead to the emergence of new models (e.g., ZFNet \cite{Zeiler}) \cite{LIANG2021168}. Therefore, in the future, the development of explanation methods for DL and new DL models are expected to complement each other \cite{LIANG2021168}.
Another research area is developing new hybrid models where the expressiveness of opaque models is combined with the apparent semantics of transparent models (e.g., combining a neural network with a linear regression) \cite{belle2021principles}. This research area can be helpful for bridging the gap between opaque and transparent models and could help in developing highly efficient explainable models \cite{belle2021principles}.
\subsubsection{Rules extraction}
\hypertarget{rules}
Historically, the need for explanations dates back to the early works in explaining expert systems and Bayesian networks \cite{biran2017explanation}. Rule extraction from ML models has been studied for a long time \cite{towell1993extracting,OMLIN199641,ANDREWS1995373}. However, there is still an increasing interest in utilizing rule extraction for explainability/interpretability \cite{Zilke,HE2020346}. Therefore, to discover methods that may work for explainability/interpretability, we should revisit the past research works \cite{Abdul2018}.
According to \cite{ANDREWS1995373,HE2020346}, there are three main approaches for rule extraction: (1) Decomposition approach based on the principle that the rules are extracted at the neuron level, such as visualizing a network's structure, (2) Pedagogical approach that extracts rules that map inputs directly to outputs regardless of their underlying structure, such as computing gradient, (3) Eclectics approach, which is the combination of both decompositional and pedagogical approaches.
Further research for rules extraction is needed, which has been discussed in \cite{HE2020346}. First, visualize neural networks' internal structure. Through visualizing each activated weight connection/neuron/filter from input to output, one can understand how the network works internally and produce the output from the input. Second, transform a complex neural network into an interpretable structure by pruning unimportant or aggregating connections with similar functions. By doing this, the overfitting issue can be reduced, and the model's structure becomes easier to interpret. Third, explore the correspondence between inputs and outputs, for example, by modifying the inputs and observing their effects on the output. Fourth, calculate the gradient of the output to the inputs to know their contributions. It has also been suggested to combine the best of DL and fuzzy logic toward an enhanced interpretability \cite{Fan2021}.
\subsubsection{Bayesian approach to interpretability}
\hypertarget{bayesian}
The work in \cite{Chakraborty} has discussed that there exist elements in DL and Bayesian reasoning that complement each other. Comparing Bayesian reasoning with DL, Bayesian reasoning offers a unified framework for modeling, inference, prediction, and decision making. Furthermore, uncertainty and variability of outcomes are explicitly accounted for. In addition, the framework has an "Occam's Razor" effect that penalizes overcomplicated models, which makes it robust to model overfitting. However, to ensure computational tractability, Bayesian reasoning is typically limited to conjugate and linear models.
In a recent survey on Bayesian DL (BDL) \cite{Wang2016} this complement observation has been exploited, and a general framework for BDL within a uniform probabilistic framework has been proposed. Further research is needed to be done to exploit this complement observation because it could improve model transparency and functionality \cite{Chakraborty}.
\subsubsection{Explaining competencies}
\hypertarget{competencies}
There is a need for users to gain a deeper understanding of the competencies of the AI system, which includes knowing what competencies it possesses, how its competencies can be measured, as well as whether or not it has blind spots (i.e., classes of solutions it never finds) \cite{gunning2019xai}. Through knowledge and competency research, XAI could play a significant role in society. Besides explaining to individuals, other roles include leveraging existing knowledge for further knowledge discovery and applications and teaching both agents and humans \cite{gunning2019xai}.
\subsection{Challenges and Research Directions of XAI in the Deployment Phase}
The following subsections are dedicated to challenges and research directions during the deployment of AI systems. The deployment phase starts from deploying ML solutions until we stop using the solutions (or maybe after that).
\subsubsection{Improving explanations with ontologies}
\hypertarget{ontologies}
An ontology is defined as \textit{"an explicit specification of a conceptualization"} \cite{GRUBER1993199}. The use of ontologies for representing knowledge of the relationships between data is helpful for understanding complex data structures \cite{Burkart2021-tq}. Therefore, the use of ontologies can help to produce better explanations as found in \cite{Panigutti,CONFALONIERI2021103471}.
The work in \cite{Burkart2021-tq} has discussed some recent works of the literature on this topic such as \cite{Panigutti,CONFALONIERI2021103471}. In \cite{Panigutti}, Doctor XAI was introduced as a model-agnostic explainer that focused on explaining the diagnosis prediction task of Doctor AI \cite{pmlr-v56-Choi16}, which is a black-box AI model that predicts the patient’s next visit time. It was shown that taking advantage of the temporal dimension in the data and incorporating the domain knowledge into the ontology helped improve the explanations' quality. Another work in \cite{CONFALONIERI2021103471} showed that ontologies can enhance human comprehension of global post-hoc explanations, expressed in decision trees.
It should be noted that ontologies are thought of as contributing a lot to explaining AI systems because they provide a user's conceptualization of the domain, which could be used as a basis for explanations or debugging \cite{tudorache2020ontology}. Toward that goal, new design patterns, new methodologies for creating ontologies that can support explainable systems, and new methods for defining the interplay between ontologies and AI techniques are needed \cite{tudorache2020ontology}. Furthermore, it is essential to conduct several user studies to determine the benefits of combining ontologies with explanations \cite{Burkart2021-tq}.
\subsubsection{XAI and privacy}
\hypertarget{privacy}
When individuals are affected by automated decision-making systems, two rights conflict: the right to privacy and the right to an explanation \cite{grant2020show}. At this stage, it could be a demand to disclose the raw training data and thus violate the privacy rights of the individuals from whom the raw training data came \cite{grant2020show}. Another legal challenge has been discussed in \cite{Longo2020}, which is the right to be forgotten \cite{VILLARONGA2018304}. By this right, individuals can claim to delete specific data so that they cannot be traced by a third party \cite{Longo2020}. Data preservation is another related issue because to use XAI to justify a decision reached by automated decision-making, the raw data used for training must be kept, at least until we stop using the AI solution.
One of the key challenges is establishing trust in the handling of personal data, particularly in cases where the algorithms used are challenging to understand \cite{Holzinger}. This can pose a significant risk for acceptance to end-users and experts alike \cite{Holzinger}. For example, end-users need to trust that their personal information is secured and protected as well as only their consented data is used, while experts need to trust that their input is not altered later \cite{Holzinger}.
Anonymization of data can be used to obscure the identity of people. However, privacy cannot always be protected by anonymization \cite{grant2020show}. According to \cite{grant2020show}, the more information in a data set, the greater the risk of de-anonymization, even if the information is not immediately visible. Asserting that anonymization helps conceal who supplied the data to train the automated decision-making system might be comforting for the individuals whom the training data came from, but this does not the case with individuals who are entitled to an explanation of the results produced by the system \cite{grant2020show}.
In order to address some issues with anonymization techniques, it is recommended that further research should be undertaken in privacy-aware ML, which is the intersection between ML and security areas \cite{Holzinger}. XAI can play an essential role in this matter because to develop new techniques to ensure privacy and security, it will be essential to learn more about the inner workings of the system they are meant to protect \cite{Holzinger}. In addition, in the future, to promote the acceptance of AI and increase privacy protection, XAI needs to provide information on how the personal data of a particular individual was utilized in a data analysis workflow \cite{Longo2020}. However, according to \cite{grant2020show}, what if it is needed to review the data of many individuals and they may not have consented to review their data in litigation. In such cases, a path to review data for which individuals have not consented would be demanded, but it would be difficult to find such a path \cite{grant2020show}.
\subsubsection{XAI and security}
\hypertarget{security}
Two main concerns have been discussed for XAI and security: confidentiality and adversarial attacks \cite{arrieta2020,LIANG2021168,Ras2018,Tjoa2020}. For the confidentiality concern, several aspects of a model may possess the property of confidentiality \cite{arrieta2020}. As an example given by \cite{arrieta2020}, think of a company invested in a multi-year research project to develop an AI model. The model's synthesized knowledge may be regarded as confidential, and hence if only inputs and outputs are made available, one may compromise this knowledge \cite{orekondy19knockoff}. The work in \cite{joon18iclr} presented the first results on how to protect private content from automatic recognition models. Further research is recommended to develop XAI tools that explain ML models while maintaining models' confidentiality \cite{arrieta2020}.
Turning now to the adversarial attacks concern, the information revealed by XAI can be utilized in generating efficient adversarial attacks to cause security violations, confusing the model and cause it to produce a specific output, and manipulation of explanations \cite{arrieta2020,Tjoa2020}. In adversarial ML, three types of security violations can be caused by attackers using adversarial examples \cite{Huang2011}: integrity attacks (i.e., the system identifies intrusion points as normal), availability attacks (i.e., the system makes multiple classification errors, making it practically useless), and privacy violation (i.e., violating the privacy of system users). Attackers can do such security violations because an AI model can be built based on training data influenced by them, or they might send carefully crafted inputs to the model and see its results \cite{Huang2011}. According to \cite{LIANG2021168}, existing solutions to handle perturbations still suffer from some issues, including instabilities and lack of variability. Therefore, it is necessary to develop new methods to handle perturbations more robustly \cite{LIANG2021168,Ras2018}.
The information uncovered by XAI can also be utilized in developing techniques for protecting private data, e.g., utilizing generative models to explain data-driven decisions \cite{arrieta2020}. Two recent research directions have been highlighted in this context \cite{arrieta2020}: using generative models as an attribution method to show a direct relationship between a particular output and its input variables \cite{baumgartner2018visual}. The second is creating counterfactuals through generative models \cite{Liu2019}. It is expected that generative models will play an essential role in scenarios requiring understandable machine decisions \cite{arrieta2020}.
\subsubsection{XAI and safety}
\hypertarget{safety}
Trust and acceptance are benefits of explainability/interpretability \cite{NaisehJiang2020}. However, focusing on benefits without considering the potential risks may have severe consequences (e.g., relying too much or too little on the advice provided by the prescription recommendation system) \cite{NaisehJiang2020}. Several studies have been conducted to evaluate the safety of processes that depend on model outputs because erroneous outputs can lead to harmful consequences in some domains \cite{arrieta2020}. Therefore, possible risks must be at the top priority when designing the presented explanations \cite{NaisehJiang2020}.
Many techniques have been proposed to minimize the risk and uncertainty of adverse effects of decisions made using model outputs \cite{arrieta2020}. As an example, the model’s output confidence technique can examine the extent of uncertainty resulting from lack of knowledge regarding the inputs and the corresponding output confidence of the model to notify the user and cause them to reject the output produced by the model \cite{arrieta2020}. In order to achieve this, explaining what region of the inputs was used by the model to arrive at the outcome can be used for separating out such uncertainty that may exist within the input domain \cite{arrieta2020}. Additionally, as has been suggested in \cite{NaisehJiang2020}, it is important to develop explanations that evolve with time, keeping in mind past explanations for long-term interactions with end-users and identifying ways to minimize risks. Developing evaluation metrics and questionnaires would be essential to integrate the user-centric aspects of explanations as well as evaluating error-proneness and any possible risks \cite{NaisehJiang2020}. Finally, in \cite{HUANG2020100270}, some major challenges have been discussed, including developing distance metrics that more closely reflect human perception, improvement to robustness by designing a set of measurable metrics for comparing the robustness of black-box AI models across various architectures, verification completeness using various verification techniques, scalable verification with tighter bounds, and unifying formulation of interpretability.
\subsubsection{Human-machine teaming}
\hypertarget{teaming}
Most provided explanations for AI systems are typically static and carry one message per explanation \cite{Abdul2018}. Explanations alone do not translate to understanding \cite{Adadi2018}. Therefore, for a better understanding of the system, users should be able to explore the system via interactive explanations, which is a promising research direction to advance the XAI field \cite{Abdul2018,Adadi2018}.
Even though there are already some works in this research direction as has been reported in \cite{Abdul2018}, much work is still needed to tailor interfaces to different audiences, exploit interactivity, and choose appropriate interactions for better visualization designs \cite{Abdul2018,Chatzimparmpas}. Various works have also been suggested to go beyond static explanations and enhance human-machine teaming. In \cite{Messina}, open-ended visual question answering (VQA) has been suggested to be used rather than providing a report with too many details. Here, an user queries (or make follow-up questions), and the system answers. Achieving that would provide better interaction between the system and the expert user. In another work \cite{Choo8402187}, it has been mentioned that generative models can allow for interactive DL steering because they allow for multiple answers. They highlighted that developing new DL models capable of adapting to various user inputs and generating outputs accordingly as well as developing visualization-based interfaces that enable effective interaction with DL systems are promising research areas in the future.
In \cite{Abdul2018}, rather than providing static explanations, the authors have suggested building on existing intelligibility work for context-aware systems (e.g., design space explorations, conceptual models for implicit interaction, and intelligible interfaces for various scenarios and using a variety of modalities). Additionally, they have highlighted a research area that is effectively interacting with AI augmentation tools. In \cite{Adadi2018}, it has been emphasized the importance of bridging HCI empirical studies with human sciences theories to make explainability models more human-centered models. In this way, adaptive explainable models would emerge by providing context-aware explanations that could be adapted to any changes in the parameters of their environment, such as user profile (e.g., expertise level, domain knowledge, cultural background, interests and preferences) and the explanation request setting (e.g., justification).
The authors in \cite{Chatzimparmpas} have mentioned that extracting, visualizing, and keeping track of the history of interaction data between users and systems can allow users to undo certain actions and examine them interactively would help to address some common challenges (e.g., hyperparameter exploration). Finally, the authors in \cite{NaisehJiang2020} have highlighted that user-friendliness and intelligent interface modalities need to take into account the type of explanations that meet users' goals and needs. For example, the system can ask for feedback from the users to know how good was the provided explanations (e.g., ``explain more'', ``redundant explanation'', or ``different explanation''). Such interaction can help to improve future explanations.
Taken together, it seems that different ways are needed to enhance human-machine teaming. Approaching HCI and other related studies can contribute to making explainability models more human-centered. In addition, humans can provide feedback on the provided explanations, which can help in improving future explanations.
\subsubsection{Explainable agency}
\hypertarget{agency}
Explainable agency refers to a general capability in which autonomous agents must provide explanations for their decisions and the reasons leading to these decisions \cite{Langley}. Based on the three explanation phases proposed in \cite{Neerincx}, the authors in \cite{Anjomshoae} presents a research roadmap for the explainable agency.
The first phase is explanation generation which is intended to explain why an action/result was taken/achieved \cite{Anjomshoae}. This phase of research focuses on the following key research directions: (1) there is a need to connect the internal AI mechanism of the agent/robot with the explanation generation module, (2) to produce dynamic explanations, new mechanisms are required for identifying relevant explanation elements, identifying its rationales, and combining these elements to form a coherent explanation.
The second phase is the explanation communication phase. Here, the focus is on what content end users will receive and how to present that content \cite{Neerincx}. According to \cite{Anjomshoae}, explainable agents/robots may be deployed in a variety of environments. Therefore, for some cases, multimodal explanation presentations (e.g., visual, audio, and expressive) could be a useful explanation communication approach for enabling efficient explainable agency communication.
For the last phase, explanation reception, the focus is on the human’s understanding of explanations. Some considerations should be taken into account to ensure an accurate reception \cite{Anjomshoae}. It is important to develop metrics to measure the explanations' effectiveness and the users' reaction to the provided explanations. In addition, the agent/robot should maintain a model of user knowledge and keep updating it based on the evolution of user expertise and the user's perception of the State of Mind (SoM) of the agent/robot, i.e., an internal representation of how the agent/robot treats the outer world.
\subsubsection{Machine-to-machine explanation}
\hypertarget{mexplanation}
A promising area of research is enabling machine-to-machine communication and understanding \cite{Weller2019}. Furthermore, it is an important research area because of the increasing adoption of the Internet of Things (IoT) in different industries. A growing body of research has begun exploring how multiple agents can efficiently cooperate and exploring the difference between explanations intended for humans and those intended for machines \cite{Preece2018,Weller2019}.
According to \cite{Preece2018}, future explainable approaches are likely to provide both human and machine explanations, especially adaptive explainable approaches \cite{Adadi2018}. For machine explanations, complex structures that are beyond the comprehension of humans may be developed \cite{Weller2019}. However, how is it possible to measure the success of ``transfer of understanding'' between agents? The work in \cite{Weller2019} has suggested a metric for that, which is measuring the improvement of agent B’s performance on a particular task, or set of tasks, as a result of the information obtained from agent A - though it will be crucial to determine some key details, such as the bandwidth constraints and already existing knowledge with agent A.
Based on what has been mentioned above, it is expected that much work is going to be done on how to construct machine explanations, how to communicate these explanations, and which metrics we need to measure as a success of the transfer of understanding between agents and how to measure them. With more research into how machines communicate/explain themselves, we will be able to understand intelligence better and creating intelligent machines \cite{Molnar2021-do}.
\subsubsection{XAI and reinforcement learning}
\hypertarget{reinforcement}
The use of DL by reinforcement learning (RL) has been applied successfully to many areas \cite{dao2020demystifying}. Through the explicit modeling of the interaction between models and environments, RL can directly address some of the interpretability objectives \cite{Lipton2018-dx}. Despite that, unexplained or non-understandable behavior makes it difficult to users to trust RL agents in a real environment, especially when it comes to human safety or failure costs \cite{dao2020demystifying}. Additionally, we lack a clear understanding of why an RL agent decides to perform an action and what it learns during training \cite{dao2020demystifying}. RL's interpretability can help in exploring various approaches to solving problems \cite{dao2020demystifying}. For instance, understanding why the RL AlphaFold system \cite{senior2020improved} is capable of making accurate predictions can assist bioinformatics scientists in understanding and improving the existing techniques in protein structures to speed produce better treatment before new outbreaks happen \cite{dao2020demystifying}.
Recently, the work in \cite{Wells} highlighted several issues that need to be addressed and potential research directions in the area of XAI for RL. The authors find that the selected studies used "toy" examples or case studies that were intentionally limited in scope mainly to prevent the combinatorial explosion problem in the number of combinations of states and actions. Therefore, more focus on real-world applications has been suggested. It has also been mentioned that there is a lack of new algorithms in the area. Therefore, the design of RL algorithms with an emphasis on explainability is essential. Symbolic representations can be utilized so RL agents can inherently be explained and verified. Another issue is highlighted, which is the lack of user testing with the existing approaches, which is in line with what was mentioned in \cite{MILLER20191}. As for the complexity of the provided explanations, it has been found that the current focus is presenting explanations for users with a background in AI. Therefore, it has been suggested to conduct further research to present the explanations for those who might interact with the agents, which may have no background in AI. For example, providing more visceral explanations, e.g., annotations in a virtual environment. Additionally, enriching visualization techniques by considering the temporal dimensions of RL and multi-modal forms of visualization, e.g., virtual or augmented reality. Lastly, it has been emphasized the importance of open-source code sharing for the academic community.
Another interesting point for consideration has been highlighted in \cite{Guidotti2018-tx}, which is learning from explanations. The work in \cite{Krening} provides a starting point, which presents an agent who trained to simulates the Mario Bros. game using explanations instead of prior play logs.
\subsubsection{Explainable AI planning (XAIP)}
\hypertarget{planning}
Existing literature focuses mainly on explainability in ML, though similar challenges apply to other areas in AI as well \cite{Adadi2018}. AI planning is an example of such an area that is important in applications where learning is not an option \cite{fox2017explainable}. Recent years have seen increased interest in research on explainable AI planning (XAIP) \cite{Hoffmann2019}. XAIP includes a variety of topics from epistemic logic to ML, and techniques including domain analysis, plan generation, and goal recognition \cite{Hoffmann2019}. There are, however, some major trends that have emerged, such as plan explanations, contrastive explanations, human factors, and model reconciliation \cite{Hoffmann2019}.
Recently, the work in \cite{fox2017explainable} has explored the explainability opportunities that arise in AI planning. They have provided some of the questions requiring explanation. They also have described initial results and a roadmap toward achieving the goal of generating effective explanations. Additionally, they have suggested several future directions in both plan explanations and executions. Temporal planning, for instance, can open up interesting choices regarding the order of achieving (sub)goals. It is also interesting to consider whether giving the planner extra time to plan would improve the performance. In addition, one of the challenges in plan execution is explaining what has been observed at the execution time that prompts the planner to make a specific choice. As with XAI, it is crucial to have a good metric for XAIP that defines what constitutes a good explanation. Finally, it is imperative that the existing works on XAIP be reconsidered and leveraged so that XAIP will be more effective and efficient when used in critical domains.
\subsubsection{Explainable recommendation}
\hypertarget{recommendation}
Explainable recommendation aims to build models that produce high quality recommendations as well as provide intuitive explanations that can help to enhance the transparency, persuasiveness, effectiveness, trustworthiness, and satisfaction of recommendation systems \cite{ZhangINR-066}.
The work in \cite{ZhangINR-066} conducted a comprehensive survey of explainable recommendations, and they discussed potential future directions to promote explainable recommendations. With regards to the methodology perspective, it has been suggested that (1) further research is needed to make deep models explainable for recommendations because we still do not fully understand what makes something recommended versus other options, (2) develop knowledge-enhanced explainable recommendation which allows the system to make recommendations based on domain knowledge, e.g., combine graph embedding learning with recommendation models, (3) use heterogeneous information for explainability such as multi-modal explanations, transfer learning over heterogeneous information sources, information retrieval and recommendation cross-domain explanations, and the impact that specific information modalities have on user receptiveness on the explanations, (4) develop context-aware explainable recommendations, (5) aggregate different explanations, (6) integrate symbolic reasoning and ML to make recommendations and explainability better by advancing collaborative filtering to collaborative reasoning, (7) further research is needed to help machines explain themselves using natural language, and (8) with the evolution of conversational recommendations powered by smart agent devices, users may ask “why” questions to get explanations when a recommendation does not make sense. Therefore, it is essential to answer the “why” in conversations which could help to improve system efficiency, transparency, and trustworthiness.
For the evaluation perspective, the authors in \cite{ZhangINR-066} have been suggested that the importance of developing reliable and easily implemented evaluation metrics for different evaluation perspectives (i.e., user perspective and algorithm perspective). Additionally, evaluating explainable recommendation systems using user behavior perspectives may be beneficial as well. Lastly, it has been highlighted that explanations should have broader effects than just persuasion. For example, investigate how explanations can make the system more trustworthy, efficient, diverse, satisfying, and scrutable.
In \cite{NaisehJiang2020}, the authors have presented several research challenges in delivery methods and modalities in user experience. As mentioned in that paper, for the delivery method, the current focus in the literature is on providing the explanation to the users while they are working on a task or looking for recommendations. However, more focus should be done on the long-term retrieval of such explanations, for example, through a digital archive, and their implications for accountability, traceability, and users' trust and adoption. That could increase the adoption of intelligent human-agent systems in critical domains. Another challenge is designing autonomous delivery capable of considering the context and situation in which users may need explanations and suitable explanations for them. It is worth mentioning that privacy matters should be taken into account when deriving the recommendations.
It has also been highlighted in \cite{NaisehJiang2020} that users' goals and needs would have to be met by user-friendly and intelligent interface modalities that provide appropriate explanations. Further, interaction with the system is needed and could help to improve future generated explanations. Finally, focusing on the benefits of explainability without considering the potential risks may have severe consequences. Therefore, when designing explanations, possible risks should be the first priority.
\subsubsection{XAI as a service}
\hypertarget{service}
There is an increasing trend in developing automated ML (AutoML) tools \cite{Molnar2021-do}. AutoML tool is an end-to-end pipeline starting with raw data and going all the way to a deployable ML model. Model-agnostic explanation methods are applicable to any ML model resulting from automated ML \cite{Molnar2021-do}. Similarly, we can automate the explanation step: calculate the importance of each feature, plot the partial dependence, construct a surrogate model, etc \cite{Molnar2021-do}. Some existing AutoML tools provide automatic generated explanations, e.g., AutoML H2O \cite{H2OAutoML20} and MLJAR AutoML \cite{mljar}. We expect that more Auto XAI tools will be available in the future, either incorporated with AutoML tools or as services.
\section{Conclusions}
In this systematic meta-survey paper, we present two main contributions to the literature of XAI. First, we propose an attempt to present a distinction between explainability and interpretability terms. Second, we shed light on the significant challenges and future research directions of XAI resulting from the selected 58 papers, which guide future exploration in the XAI area. Even though they are presented individually in 39 points, they can overlap and combine them based on researchers' backgrounds and interests, resulting in new research opportunities where XAI can play an important role. This meta-survey has three limitations. First, because we cannot ensure that the selected keywords are complete, we could miss some very recent papers. Second, to avoid listing the challenges and future research directions per each paper, we come up with the reported 39 points, which are the results of combining what was reported in the selected papers based on the authors' point of view. Third, we believe that more challenges and future research directions can be added where XAI can play an important role in some domains, such as IoT, 5G, and digital forensic. However, related surveys did not exist at the time of writing this meta-survey.
\bibliographystyle{unsrtnat}
|
\section{Acknowledgments}
\noindent This publication was supported by the European Union through the European Regional Development Fund (ERDF), by the French region of Brittany and Rennes M{\'e}tropole through the CPER Project SOPHIE/STIC \& Ondes. M. D. acknowledges the Institut Universitaire de France. C. F. acknowledges funding from the French ``Minist{\`e}re de la D{\'e}fense, Direction G{\'e}n{\'e}rale de l'Armement''.
\section{Data availability}
\noindent The data that support the plots within this paper and other findings of this study are available from the corresponding authors on reasonable request.
|
\section{Introduction}
Neural networks have been successful in handling various forms of data. Since some of the world's most interesting data is represented by graphs, Graph Neural Networks (GNNs) have achieved state-of-the-art performance in various fields such as quantum chemistry, physics, or social networks \citep{gilmer2017neural, sanchez2020learning, kipf2017semisupervised}. On the other hand, GNNs are also known to have severe limitations and are sometimes unable to recognize even simple graph structures.
In this paper, we present a new approach to increase the expressiveness of GNNs, called Dropout Graph Neural Networks (DropGNNs). Our main idea is to execute not one but \textit{multiple} different runs of the GNN. We then aggregate the results from these different runs into a final result.
In each of these runs, we remove (``drop out'') each node in the graph with a small probability $p$. As such, the different runs of an episode will allow us to not only observe the actual
extended neighborhood of a node for some number of layers $d$, but rather to observe various slightly perturbed versions of this $d$-hop neighborhood. We emphasize that this notion of dropouts is very different from the popular dropout regularization method; in particular, DropGNNs remove nodes during \emph{both training and testing}, since their goal is to observe a similar distribution of dropout patterns during training and testing.
This dropout technique increases the expressive power of our GNNs dramatically: even when two distinct $d$-hop neighborhoods cannot be distinguished by a standard GNN, their dropout variants (with a few nodes removed) are already separable by GNNs in most cases. Thus by learning to identify the dropout patterns where the two $d$-hop neighborhoods differ, DropGNNs can also distinguish a wide variety of cases that are beyond the theoretical limits of standard GNNs.
\subparagraph*{Our contributions.} We begin by showing several example graphs that are not distinguishable in the regular GNN setting but can be easily separated by DropGNNs. We then analyze the theoretical properties of DropGNNs in detail. We first show that executing $\widetilde{O}(\gamma)$ different runs is often already sufficient to ensure that we observe a reasonable distribution of dropouts in a neighborhood of size $\gamma$. We then discuss the theoretical capabilities and limitations of DropGNNs in general, as well as the limits of the dropout approach when combined with specific aggregation methods.
We validate our theoretical findings on established problems that are impossible to solve for standard GNNs. We find that DropGNNs clearly outperform the competition on these datasets. We further show that DropGNNs have a competitive performance on several established graph benchmarks, and they provide particularly impressive results in applications where the graph structure is really a crucial factor.
\section{Related Work}
GNNs apply deep learning to graph-structured data \citep{scarselli2008graph}. In GNNs, every node has an embedding that is shared over multiple iterations with its neighbors. This way nodes can gather their neighbors' features. In recent years, many different models have been proposed to realize how the information between nodes is shared~\citep{wu2020comprehensive}. Some approaches take inspiration from convolution~\citep{niepert2016learning, defferrard2016convolutional, hamilton2017inductive}, others from graph spectra~\citep{kipf2017semisupervised, bruna2014spectral}, others from attention~\citep{velickovic2018graph}, and others extend previous ideas of established concepts such as skip connections~\citep{xu2018jumping}.
Principally, GNNs are limited in their expressiveness by the \textit{Weisfeiler-Lehman test} (WL-test)~\citep{GIN}, a heuristic to the graph isomorphism problem. The work of \cite{GIN} proposes a new architecture, \textit{Graph Isomoprhism Networks} (GIN), that is proven to be exactly as powerful as the WL-test. However, even GINs cannot distinguish certain different graphs, namely those that the WL-test cannot distinguish. This finding~\citep{limits} motivated more expressive GNN architectures. These improvements follow two main paths.
The first approach augments the features of nodes or edges by additional information to make nodes with similar neighborhoods distinguishable. Several kinds of information have been used: inspired from distributed computing are port numbers on edges~\citep{ports}, unique IDs for nodes~\citep{loukas2020graph}, or random features on nodes~\citep{randomFeatures1, randomFeatures2}. Another idea is to use angles between edges~\citep{angles} from chemistry (where edges correspond to electron bonds).
However, all of these approaches have some shortcomings. For ports and angles, there are some simple example graphs that still cannot be distinguished with these extensions~\citep{limits}. Adding IDs or random features helps during training, but the learned models do not generalize: GNNs often tend to overfit to the specific random values in the training set, and as such, they produce weaker results on unseen test graphs that received different random values. In contrast to this, DropGNNs observe a similar distribution of embeddings during training and testing, and hence they also generalize well to test set graphs.
The second approach exploits the fact that running the WL-test on tuples, triples, or generally $k$-tuples keeps increasing its expressiveness. Thus a GNN operating on tuples of nodes has higher expressiveness than a standard GNN~\citep{morris2019weisfeiler, maron2019provably}. However, the downside of this approach is that even building a second-order graph blows up the graph quadratically. The computational cost quickly becomes a problem that needs to be to addressed, for example with sampling~\citep{morris2019weisfeiler}. Furthermore, second-order graph creation is a global operation of the graph that destroys the local semantics induced by the edges. In contrast to this, DropGNN can reason about graphs beyond the WL-test with only a small overhead (through run repetition), while also keeping the local graph structure intact.
Our work is also somewhat similar to the randomized smoothing approach \citep{cohen2019certified}, which has also been extended to GNNs recently \citep{bojchevski2020efficient}. This approach also conducts multiple runs on slightly perturbed variants of the data. However, in randomized smoothing, the different embeddings are combined in a smoothing operation (e.g. majority voting), which specifically aims to get rid of the atypical perturbed variants in order to increase robustness. In contrast to this, the main idea of DropGNNs is exactly to find and identify these perturbed special cases which are notably different from the original neighborhood, since these allow us to distinguish graphs that otherwise seem identical.
Finally, we note that removing nodes is a common tool for regularization in deep neural networks, which has also seen use in GNNs~\citep{rong2019dropedge, Grand}. However, as mentioned before, this is a different dropout concept where nodes are only removed during training to reduce the co-dependence of nodes.
\section{DropGNN}
\subsection{About GNNs}
Almost all GNN architectures~\citep{velickovic2018graph, kipf2017semisupervised, GIN, defferrard2016convolutional, wu2020comprehensive, hamilton2017inductive, xu2018jumping} follow the message passing framework~\citep{gilmer2017neural, battaglia2018relational}. Every node starts with an embedding given by its initial features. One round of message passing has three steps. In the first \textsc{message} step, nodes create a message based on their embedding and send this message to all neighbors. Second, nodes \textsc{aggregate} all messages they receive. Third, every node \textsc{update}s its embedding based on its old embedding and the aggregated messages. One such round corresponds to one GNN layer. Usually, a GNN performs $d$ rounds of message passing for some small constant $d$. Thus, the node's embedding in a GNN reflects its features and the information within its $d$-hop neighborhood. Finally, a \textsc{readout} method translates these final embeddings into predictions. Usually, \textsc{message}, \textsc{aggregate}, \textsc{update} and \textsc{readout} are functions with learnable parameters, for instance linear layers with activation functions.
This GNN paradigm is closely related to the WL-test for a pair of graphs, which is an iterative color refinement procedure. In rounds $1, ..., d$, each node looks at its own color and the multiset of colors of its direct neighbors, and uses a hash function to select a new color based on this information. As such, if the WL-test cannot distinguish two graphs, then a standard GNN cannot distinguish them either: intuitively, the nodes in these graphs receive the same messages and create the same embedding in each round, and thus they always arrive at the same final result.
\subsection{Idea and motivation}
The main idea of DropGNNs is to execute multiple independent runs of the GNN during both training and testing. In each run, every node of the GNN is removed with probability $p$, independently from all other nodes. If a node $v$ is removed during a run, then $v$ does not send or receive any messages to/from its neighbors and does not affect the remaining nodes in any way. Essentially, the GNN behaves as if $v$ (and its incident edges) were not present in the graph in the specific run, and no embedding is computed for $v$ in this run (see Figure \ref{fig:dropouts} for an illustration).
Over the course of multiple runs, dropouts allow us to not only observe the $d$-hop neighborhood around any node $u$, but also several slightly perturbed variants of this $d$-hop neighborhood. In the different runs, the embedding computed for $u$ might also slightly vary, depending on which node(s) are missing from its $d$-hop neighborhood in a specific run. This increases the expressive power of GNNs significantly: even when two different $d$-hop neighborhoods cannot be distinguished by standard GNNs, the neighborhood variants observed when removing some of the nodes are usually still remarkably different. In Section \ref{sec:examples}, we discuss multiple examples for this improved expressiveness.
Our randomized approach means that in different runs, we will have different nodes dropping out of the GNN. As such, the GNN is only guaranteed to produce the same node embeddings in two runs if we have exactly the same subset of nodes dropping out. Given the $d$-hop neighborhood of a node $u$, we will refer to a specific subset of nodes dropping out as a \textit{dropout combination}, or more concretely as a $k$-dropout in case the subset has size $k$.
\begin{figure}
\centering
\resizebox{0.9\textwidth}{!}{\input{pics/dropouts.tikz}}
\caption{Illustration of $4$ possible dropout combinations from an example $2$-hop neighborhood around $u$: a $0$-dropout, two different $1$-dropouts and a $2$-dropout.}
\label{fig:dropouts}
\end{figure}
In order to analyze the $d$-hop neighborhood of $u$, the reasonable strategy is to use a relatively small dropout probability $p$: this ensures that in each run, only a few nodes are removed (or none at all), and thus the GNN will operate on a $d$-hop neighborhood that is similar to the original neighborhood of $u$. As a result, $1$-dropouts will be frequent, while for a larger $k$, observing a $k$-dropout will be unlikely.
To reduce the effect of randomization on the final outcome, we have to execute multiple independent runs of our GNN; we denote this number of runs by $r$. For a successful application of the dropout idea, we have to select $r$ large enough to ensure that the set of observed dropout combinations is already reasonably close to the actual probability distribution of dropouts. In practice, this will not be feasible for $k$-dropouts with large $k$ that occur very rarely, but we can already ensure for a reasonably small $r$ that e.g. the frequency of each $1$-dropout is relatively close to its expected value.
\subsection{Run aggregation}
Recall that standard GNNs first compute a final embedding for each node through $d$ layers, and then they use a \textsc{readout} method to transform this into a prediction. In DropGNNs, we also need to introduce an extra phase between these two steps, called \textit{run aggregation}.
In particular, we execute $r$ independent runs of the $d$-layer GNN (with different dropouts), which altogether produces $r$ distinct final embeddings for a node $u$. Hence we also need an extra step to merge these $r$ distinct embeddings into a single final embedding of $u$, which then acts as the input for the \textsc{readout} function. This run aggregation method has to transform a multiset of embeddings into a single embedding; furthermore, it has to be a \textit{permutation-invariant} function (similarly to neighborhood aggregation), since the ordering of different runs carries no meaning.
We note that simply applying a popular permutation-invariant function for run aggregation, such as \texttt{sum} or \texttt{max}, is often not expressive enough to extract sufficient information from the distribution of runs. Instead, one natural solution is to first apply a transformation on each node embedding, and only execute \texttt{sum} aggregation afterward. For example, a simple transformation $x \rightarrow \sigma\left( W x + b \right)$, where $\sigma$ denotes a basic non-linearity such as a sigmoid or step function, is already sufficient for almost all of our examples and theoretical results in the paper.
\subsection{Motivational examples} \label{sec:examples}
We discuss several examples to demonstrate how DropGNNs are more expressive than standard GNNs. We only outline the intuitive ideas behind the behavior of the DropGNNs here; however, in Appendix \ref{app:examples}, we also describe the concrete functions that can separate each pair of graphs.
\subparagraph*{Example 1.} Figure \ref{fig:ex_cycles} shows a fundamental example of two different graphs that cannot be distinguished by the $1$-WL test, consisting of cycles of different length. This example is known to be hard for extended GNNs variants: the two cases cannot even be distinguished if we also use port numbers or angles between the edges \citep{limits}.
The simplest solution here is to consider a GNN with $d=2$ layers; this already provides a very different distribution of dropouts in the two graphs. For example, the $8$-cycle has $2$ distinct $1$-dropouts where $u$ retains both of its direct neighbors, but it only has $1$ neighbor at distance 2; such a situation is not possible in the $4$-cycle at all. Alternatively, the $4$-cycle has a $1$-dropout case with probability $p \cdot (1-p)^2$ where $u$ has $2$ direct neighbors, but no distance $2$ neighbors at all; this only happens for a $2$-dropout in the $8$-cycle, i.e. with a probability of only $p^2 \cdot (1-p)^2$. With appropriate weights, a GNN can learn to recognize these situations, and thus distinguish the two cases.
\subparagraph*{Example 2.} Figure \ref{fig:ex_WL} shows another example of two graphs that cannot be separated by a WL test; note that node features simply correspond to the degrees of the nodes. From an algorithmic perspective, it is not hard to distinguish the two graphs from specific $1$-dropout cases. Let $u$ and $v$ denote the two gray nodes in the graphs, and consider the process from $u$'s perspective. In both graphs, $u$ can recognize if $v$ is removed in a run since $u$ does not receive a ``gray'' message in the first round. However, the dropout of $v$ has a different effect in the two graphs later in the process: in the right-hand graph, it means that there is no gray neighbor at a $3$-hop distance from $u$, while in the left-hand graph, $u$ will still see a gray node (itself) in a $3$-hop distance.
Thus by identifying the $1$-dropout of $v$, an algorithm can distinguish the two graphs: if we observe runs where $u$ receives no gray message in the first round, but it receives an (aggregated) gray message in the third round, then $u$ has the left-hand neighborhood. This also means that a sufficiently powerful GNN which is equivalent to the $1$-WL test can also separate the two cases.
\subparagraph*{Example 3.} Note that using a \texttt{sum} function for neighborhood aggregation is often considered a superior choice to \texttt{mean}, since $u$ cannot separate e.g. the two cases shown in Figure \ref{fig:ex_mean} with \texttt{mean} aggregation \citep{GIN}. However, the \texttt{mean} aggregation of neighbors also has some advantages over \texttt{sum}; most notably, it the computed values do not increase with the degree of the node.
We show that dropouts also increase the expressive power of GNNs with \texttt{mean} aggregation, thus possibly making \texttt{mean} aggregation a better choice in some applications. In particular, a DropGNN with \texttt{mean} aggregation is still able to separate the two cases on Figure \ref{fig:ex_mean}.
Assume that the two colors in the figure correspond to feature values of $1$ and $-1$, and let $p=\frac{1}{4}$. In the left-hand graph, there is a $1$-dropout where $u$ ends up with a single neighbor of value $1$; hence mean aggregation yields a value of $1$ with probability $\frac{1}{4} \cdot \frac{3}{4} \approx 0.19$ in each run. However, in the right-hand graph, the only way to obtain a mean of $1$ is through a $2$-dropout or some $3$-dropouts; one can calculate that the total probability of these is only $0.06$ (see Appendix \ref{app:examples}). If we first transform all other values to $0$ (e.g. with $\sigma(x-0.5)$, where $\sigma$ is a step function), then run aggregation with \texttt{mean} or \texttt{sum} can easily separate these cases. Note that if we apply a more complex transformation at run aggregation, then separation is even much easier, since e.g. the mean value of 0.33 can only appear in the right-hand graph.
\begin{figure}
\centering
\begin{subfigure}[b]{0.24\textwidth}
\centering
\resizebox{1.0\textwidth}{!}{\input{pics/ex1_cycles.tikz}}
\caption{}
\label{fig:ex_cycles}
\end{subfigure}
\hspace{0.12\textwidth}
\begin{subfigure}[b]{0.26\textwidth}
\centering
\resizebox{1.0\textwidth}{!}{\input{pics/ex2_WL.tikz}}
\caption{}
\label{fig:ex_WL}
\end{subfigure}
\hspace{0.12\textwidth}
\begin{subfigure}[b]{0.22\textwidth}
\centering
\resizebox{1.0\textwidth}{!}{\input{pics/ex3_mean.tikz}}
\vspace{-4pt}
\caption{}
\label{fig:ex_mean}
\end{subfigure}
\caption{Several example graphs which show that DropGNNs are more expressive than standard GNNs in various cases. Different node colors correspond to different node features.}
\label{fig:example}
\end{figure}
\section{Theoretical analysis} \label{sec:theory}
\subsection{Required number of runs}
We analyze DropGNNs with respect to the \textit{neighborhood of interest} around a node $u$, denoted by $\Gamma$. That is, we select a specific region around $u$, and we want to ensure that the distribution of dropout combinations in this region is reasonably close to the actual probabilities. This choice of $\Gamma$ then determines the ideal choice of $p$ and $r$ in our DropGNN.
One natural choice is to select $\Gamma$ as the entire $d$-hop neighborhood of $u$, since a GNN will always compute its final values based on this region of the graph. Note that even for this largest possible $\Gamma$, the size of this neighborhood $\gamma:=|\Gamma|$ does not necessarily scale with the entire graph. That is, input graphs in practice are often sparse, and we can e.g. assume that their node degrees are upper bounded by a constant; this is indeed realistic in many biological or chemical applications, and also a frequent assumptions in previous works \citep{ports}. In this case, having $d=O(1)$ layers implies that $\gamma$ is also essentially a constant, regardless of the size of the graph.
However, we point out that $\Gamma$ can be freely chosen as a neighborhood of any specific size. That is, even if a GNN aggregates information within a distance of $d=5$ layers, we can still select $\Gamma$ to denote, for example, only the $2$-hop neighborhood of $u$. The resulting DropGNN will still compute a final node embedding based on the entire $5$-hop neighborhood of $u$; however, our DropGNN will now only ensure that we observe a reasonable distribution of dropout combinations in the $2$-hop neighborhood of $u$.
In this sense, the size $\gamma$ is essentially a trade-off hyperparameter: while a smaller $\gamma$ will require a smaller number of runs $r$ until the distribution of dropout combinations stabilizes, a larger $\gamma$ allows us to observe more variations of the region around $u$.
\subparagraph*{1-complete dropouts.} From a strictly theoretical perspective, choosing a sufficiently large $r$ always allows us to observe every possible dropout combination. However, since the number of combinations is exponential in $\gamma$, this approach is not viable in practice (see Appendix \ref{app:Chernoff} for more details).
To reasonably limit the number of necessary runs, we focus on the so-called \emph{$1$-complete case}: we want to have enough runs to ensure that at least every $1$-dropout is observed a few times. Indeed, if we can observe each variant of $\Gamma$ where a single node is removed, then this might already allow a sophisticated algorithm to reconstruct a range of useful properties of $\Gamma$. Note that in all of our examples, a specific $1$-dropout was already sufficient to distinguish the two cases.
For any specific node $v \in \Gamma$, the probability of a $1$-dropout for $v$ is $p \cdot (1-p)^{\gamma}$ in a run (including the probability that $u$ is not dropped out). We apply the $p$ value that maximizes the probability of such a $1$-dropout; a simple differentiation shows that this maximum is obtained at $p^*=\frac{1}{1+\gamma}$.
This choice of $p$ also implies that the probability of observing a specific $1$-dropout in a run is
\[ \frac{1}{1+\gamma} \cdot \left( \frac{\gamma}{1+ \gamma} \right)^{\gamma} \geq \frac{1}{1+\gamma} \cdot \frac{1}{e} \, . \]
Hence if we execute $r \geq e \cdot (\gamma+1) = \Omega(\gamma)$ runs, then the expected number of times we observe a specific $1$-dropout (let us denote this by $\mathbb{E}_1$) is at least $\mathbb{E}_1 \geq r \cdot \frac{1}{e} \cdot \frac{1}{1+\gamma} \geq 1$.
Moreover, one can use a Chernoff bound to show that after $ \Omega(\gamma \log \gamma)$ runs, the frequency of each $1$-dropout is sharply concentrated around $\mathbb{E}_1$. This also implies that we indeed observe each $1$-dropout at least once with high probability.
For a more formal statement, let us consider a constant $\delta \in [0,1]$ and an error probability $\frac{1}{t}<1$. Also, given a node $v \in \Gamma$ (or subset $S \subseteq \Gamma$), let $X_v$ (or $X_S$) denote the number of times this $1$-dropout ($|S|$-dropout) occurs during our runs.
\begin{theorem} \label{th:chernoff1}
If $r \geq \Omega \left( \gamma \log \gamma t \right)$, then with a probability of $1-\frac{1}{t}$, it holds that for each $v \in \Gamma$, we have $X_v \in [\, (1\!-\!\delta) \cdot \mathbb{E}_{1\,}, \, (1\!+\!\delta) \cdot \mathbb{E}_1\, ]$.
\end{theorem}
With slightly more runs, we can even ensure that each $k$-dropout for $k \geq 2$ happens less frequently than $1$-dropouts. In this case, it already becomes possible to distinguish $1$-dropouts from multiple-dropout cases based on their frequency.
\begin{theorem} \label{th:chernoff2}
If $r \geq \Omega \left( \gamma^2 + \gamma \log \gamma t \right)$, then with a probability of $1-\frac{1}{t}$ it holds that
\begin{itemize}
\setlength\itemsep{0.8pt}
\item for each $v \in \Gamma$, we have $X_v \in [\,(1\!-\!\delta) \cdot \mathbb{E}_{1\,} , \, (1\!+\!\delta) \cdot \mathbb{E}_1\, ]$,
\item for each $S \subseteq \Gamma$ with $|S| \geq 2$, we have $X_S < (1\!-\!\delta) \cdot \mathbb{E}_1$.
\end{itemize}
\end{theorem}
Since the number of all dropout combinations is in the magnitude of $2^{\gamma}$, proving this bound is slightly more technical. We discuss the proofs of these theorems in Appendix \ref{app:Chernoff}.
Note that in sparse graphs, where $\gamma$ is essentially a constant, the number of runs described in Theorems \ref{th:chernoff1} and \ref{th:chernoff2} is also essentially a constant; as such, DropGNNs only impose a relatively small (constant factor) overhead in this case.
Finally, note that these theorems only consider the dropout distribution around a specific node $u$. To ensure the same properties for all $n$ nodes in the graph simultaneously, we need to add a further factor of $n$ within the logarithm to the number of necessary runs in Theorems \ref{th:chernoff1} and \ref{th:chernoff2}. However, while this is only a logarithmic dependence on $n$, it might still be undesired in practice.
\subsection{Expressive power of DropGNNs} \label{sec:power}
In Section \ref{sec:examples}, we have seen that DropGNNs often succeed when a WL-test fails. It is natural to wonder about the capabilities and limits of the dropout approach in general; we study this question for multiple neighborhood aggregation methods separately.
We consider neighborhood aggregation with \texttt{sum} and \texttt{mean} in more detail; the proofs of the corresponding claims are discussed in Appendices \ref{app:express} and \ref{app:aggregate}, respectively. Appendix \ref{app:aggregate} also discusses briefly why \texttt{max} aggregation does not combine well with the dropout approach in practice.
\subparagraph*{Aggregation with \texttt{sum}.}
Previous work has already shown that \texttt{sum} neighborhood aggregation allows for an injective GNN design, which computes a different embedding for any two neighborhoods whenever they are not equivalent for the WL-test \citep{GIN}. Intuitively speaking, this means that \texttt{sum} aggregation has the same expressive power as a general-purpose $d$-hop distributed algorithm in the corresponding model, i.e. without IDs or port numbers. Hence to understand the expressiveness of DropGNNs in this case, one needs to analyze which embeddings can be computed by such a distributed algorithm from a specific (observed) distribution of dropout combinations.
It is already non-trivial to find two distinct neighborhoods that cannot be distinguished in the $1$-complete case. However, such an example exists, even if we also consider $2$-dropouts. That is, one can construct a pair of $d$-hop neighborhoods that are non-isomorphic, and yet they produce the exact same distribution of $1$- and $2$-dropout neighborhoods in a $d$-layer DropGNN.
\begin{theorem} \label{th:counter}
There exists a pair of neighborhoods that cannot be distinguished by $1$- and $2$-dropouts.
\end{theorem}
We illustrate a simpler example for only $1$-dropouts in Figure \ref{fig:counterexample}. For a construction that also covers the case of $2$-dropouts, the analysis is more technical; we defer this to Appendix \ref{app:express}.
We note that even these more difficult examples can be distinguished with our dropout approach, based on their $k$-dropouts for larger $k$ values. However, this requires an even higher number of runs: we need to ensure that we can observe a reliable distribution even for these many-node dropouts.
\begin{figure}
\centering
\resizebox{0.95\textwidth}{!}{\input{pics/counterexample.tikz}}
\caption{Example of two graphs not separable by $1$-dropouts (left side). In both of the graphs, for any of the $1$-dropouts, $u$ observes the same tree structure for $d=2$, shown on the right side.}
\label{fig:counterexample}
\end{figure}
On the other hand, our dropout approach becomes even more powerful if we combine it e.g. with the extension by port numbers introduced in \citep{ports}. Intuitively speaking, port numbers allow an algorithm to determine all paths to the removed node in a $1$-dropout, which in turn allows us to reconstruct the entire $d$-hop neighborhood of $u$. As such, in this case, $1$-complete dropouts already allow us to distinguish any two neighborhoods.
\begin{theorem} \label{th:ports}
In the setting of Theorem \ref{th:chernoff1}, a DropGNN with port numbers can distinguish any two non-isomorphic $d$-hop neighborhoods.
\end{theorem}
Finally, we note that the expressive power of DropGNNs in the $1$-complete case is closely related to the \textit{graph reconstruction problem}, which is a major open problem in theoretical computer science since the 1940s \citep{reconstruction}. We discuss the differences between the two settings in Appendix \ref{app:express}.
\subparagraph*{Aggregation with \texttt{mean}.} We have seen in Section \ref{sec:examples} that even with \texttt{mean} aggregation, DropGNNs can sometimes distinguish $1$-hop neighborhoods (that is, multisets $S_1$ and $S_2$ of features) which look identical to a standard GNN. One can also prove in general that a similar separation is possible in various cases, e.g. whenever the two multisets have the same size.
\begin{lemma} \label{th:mean}
Let $S_1 \neq S_2$ be two multisets of feature vectors with $|S_1| = |S_2|$. Then $S_1$ and $S_2$ can be distinguished by a DropGNN with \texttt{mean} neighborhood aggregation.
\end{lemma}
However, in the general case, \texttt{mean} aggregation does not allow us to separate any two multisets based on $1$-dropouts. In particular, in Appendix \ref{app:express}, we also describe an example of multisets $S_1 \cap S_2 = \emptyset$ where the distribution of means obtained from $0$- and $1$-dropouts is essentially identical in $S_1$ and $S_2$. This implies that if we want to distinguish these multisets $S_1$ and $S_2$, then the best we can hope for is a more complex approach based on multiple-node dropouts.
\section{Experiments}
In all cases we extend the base GNN model to a DropGNN by running the GNN $r$ times in parallel, doing mean aggregation over the resulting $r$ node embedding copies before the graph readout step and then applying the base GNN's graph readout. Additionally, an auxiliary readout head is added to produce predictions based on each individual run. These predictions are used for an auxiliary loss term which comprises $\frac{1}{3}$ of the final loss.
Unless stated otherwise, we set the number of runs to $m$ and choose the dropout probability to be $p=\frac{1}{m}$, where $m$ is the mean number of nodes in the graphs in the dataset. This is based on the assumption, that in the datasets we use the GNN will usually have the receptive field which covers the whole graph. We implement random node dropout by, in each run, setting all features of randomly selected nodes to $0$. See Appendix E for more details about the experimental setup and dataset statistics. The code is publicly available\footnote{\url{https://github.com/KarolisMart/DropGNN}}.
\subsection{Datasets beyond WL}
\begin{table*}[ht]
\centering
\resizebox{\textwidth}{!}{
\begin{tabular}{@{}l*{11}{S[table-format=-3.4]}@{}}
\toprule
& \multicolumn{2}{c}{GIN} & \multicolumn{2}{c}{+Ports} & \multicolumn{2}{c}{+IDs} & \multicolumn{2}{c}{+Random feat.} & \multicolumn{2}{c}{+Dropout}\\
\cmidrule(lr){2-3} \cmidrule(lr){4-5} \cmidrule(lr){6-7} \cmidrule(lr){8-9} \cmidrule(lr){10-11}
{Dataset} & {Train} & {Test} & {Train} & {Test} & {Train} & {Test} & {Train} & {Test} & {Train} & {Test} \\
\midrule
{\textsc{Limits 1} \citep{limits}} & \makebox{0.50 \raisebox{.2ex}{$\scriptstyle\pm$} 0.00} & \makebox{0.50 \raisebox{.2ex}{$\scriptstyle\pm$} 0.00} & \makebox{0.50 \raisebox{.2ex}{$\scriptstyle\pm$} 0.00} & \makebox{0.50 \raisebox{.2ex}{$\scriptstyle\pm$} 0.00} & \makebox{1.00 \raisebox{.2ex}{$\scriptstyle\pm$} 0.00} & \makebox{0.59 \raisebox{.2ex}{$\scriptstyle\pm$} 0.19} & \makebox{0.66 \raisebox{.2ex}{$\scriptstyle\pm$} 0.19} & \makebox{0.66 \raisebox{.2ex}{$\scriptstyle\pm$} 0.22} & \makebox{1.00 \raisebox{.2ex}{$\scriptstyle\pm$} 0.00} & \makebox{\textbf{1.00 \raisebox{.2ex}{$\scriptstyle\pm$} 0.00}}\\
{\textsc{Limits 2} \citep{limits}} & \makebox{0.50 \raisebox{.2ex}{$\scriptstyle\pm$} 0.00} & \makebox{0.50 \raisebox{.2ex}{$\scriptstyle\pm$} 0.00} & \makebox{0.50 \raisebox{.2ex}{$\scriptstyle\pm$} 0.00} & \makebox{0.50 \raisebox{.2ex}{$\scriptstyle\pm$} 0.00} & \makebox{1.00 \raisebox{.2ex}{$\scriptstyle\pm$} 0.00} & \makebox{0.61 \raisebox{.2ex}{$\scriptstyle\pm$} 0.26} & \makebox{0.72 \raisebox{.2ex}{$\scriptstyle\pm$} 0.17} & \makebox{0.64 \raisebox{.2ex}{$\scriptstyle\pm$} 0.19} & \makebox{1.00 \raisebox{.2ex}{$\scriptstyle\pm$} 0.00} & \makebox{\textbf{1.00 \raisebox{.2ex}{$\scriptstyle\pm$} 0.00}}\\
{\textsc{$4$-cycles} \citep{loukas2020graph}} & \makebox{0.50 \raisebox{.2ex}{$\scriptstyle\pm$} 0.00} & \makebox{0.50 \raisebox{.2ex}{$\scriptstyle\pm$} 0.00} & \makebox{1.00 \raisebox{.2ex}{$\scriptstyle\pm$} 0.01} & \makebox{0.84 \raisebox{.2ex}{$\scriptstyle\pm$} 0.07} & \makebox{1.00 \raisebox{.2ex}{$\scriptstyle\pm$} 0.00} & \makebox{0.58 \raisebox{.2ex}{$\scriptstyle\pm$} 0.07} & \makebox{0.75 \raisebox{.2ex}{$\scriptstyle\pm$} 0.05} & \makebox{0.77 \raisebox{.2ex}{$\scriptstyle\pm$} 0.05} & \makebox{0.99 \raisebox{.2ex}{$\scriptstyle\pm$} 0.03} & \makebox{\textbf{1.00 \raisebox{.2ex}{$\scriptstyle\pm$} 0.01}}\\
{\textsc{LCC} \citep{randomFeatures1}} & \makebox{0.41 \raisebox{.2ex}{$\scriptstyle\pm$} 0.09} & \makebox{0.38 \raisebox{.2ex}{$\scriptstyle\pm$} 0.08} & \makebox{1.0 \raisebox{.2ex}{$\scriptstyle\pm$} 0.00} & \makebox{0.39 \raisebox{.2ex}{$\scriptstyle\pm$} 0.09} & \makebox{1.00 \raisebox{.2ex}{$\scriptstyle\pm$} 0.00} & \makebox{0.42 \raisebox{.2ex}{$\scriptstyle\pm$} 0.08} & \makebox{0.45 \raisebox{.2ex}{$\scriptstyle\pm$} 0.16} & \makebox{0.46 \raisebox{.2ex}{$\scriptstyle\pm$} 0.08} & \makebox{1.00 \raisebox{.2ex}{$\scriptstyle\pm$} 0.00} & \makebox{\textbf{0.99 \raisebox{.2ex}{$\scriptstyle\pm$} 0.02}}\\
{\textsc{Triangles} \citep{randomFeatures1}} & \makebox{0.53 \raisebox{.2ex}{$\scriptstyle\pm$} 0.15} & \makebox{0.52 \raisebox{.2ex}{$\scriptstyle\pm$} 0.15} & \makebox{1.0 \raisebox{.2ex}{$\scriptstyle\pm$} 0.00} & \makebox{0.54 \raisebox{.2ex}{$\scriptstyle\pm$} 0.11} & \makebox{1.00 \raisebox{.2ex}{$\scriptstyle\pm$} 0.00} & \makebox{0.63 \raisebox{.2ex}{$\scriptstyle\pm$} 0.08} & \makebox{0.57 \raisebox{.2ex}{$\scriptstyle\pm$} 0.08} & \makebox{0.67 \raisebox{.2ex}{$\scriptstyle\pm$} 0.05} & \makebox{0.93 \raisebox{.2ex}{$\scriptstyle\pm$} 0.12} & \makebox{\textbf{0.93 \raisebox{.2ex}{$\scriptstyle\pm$} 0.13}}\\
{\textsc{Skip-circles} \citep{chen2019equivalence}} & \makebox{0.10 \raisebox{.2ex}{$\scriptstyle\pm$} 0.00} & \makebox{0.10 \raisebox{.2ex}{$\scriptstyle\pm$} 0.00} & \makebox{1.00 \raisebox{.2ex}{$\scriptstyle\pm$} 0.00} & \makebox{0.14 \raisebox{.2ex}{$\scriptstyle\pm$} 0.08} & \makebox{1.00 \raisebox{.2ex}{$\scriptstyle\pm$} 0.00} & \makebox{0.10 \raisebox{.2ex}{$\scriptstyle\pm$} 0.09} & \makebox{0.16 \raisebox{.2ex}{$\scriptstyle\pm$} 0.11} & \makebox{0.16 \raisebox{.2ex}{$\scriptstyle\pm$} 0.05} & \makebox{0.81 \raisebox{.2ex}{$\scriptstyle\pm$} 0.28} & \makebox{\textbf{0.82 \raisebox{.2ex}{$\scriptstyle\pm$} 0.28}}\\
\bottomrule
\end{tabular}}
\caption{Evaluation of techniques that increase GNN expressiveness on challenging synthetic datasets. We highlight the best test scores in bold. Compared to other augmentation techniques DropGNN (GIN +Dropout) achieves high training accuracy but also generalizes well to the test set.}
\label{tab:expressiveness}
\end{table*}
To see the capabilities of DropGNN in practice we test on existing synthetic datasets, which are known to require expressiveness beyond the WL-test. We use the datasets from~\citet{randomFeatures1} that are based on $3-$regular graphs. Nodes have to predict whether they are part of a triangle (\textsc{Triangles}) or have to predict their local clustering coefficient (\textsc{LCC}). We test on the two counterexamples \textsc{Limits 1} (Figure~\ref{fig:ex_cycles}) and \textsc{Limits 2} from ~\citet{limits} where we compare two smaller structures versus one larger structure. We employ the dataset by \citet{loukas2020graph} to classify graphs on containing a cycle of length 4 (\textsc{$4$-cycles}). We increase the regularity in this dataset by ensuring that each node has a degree of 2. Finally we experiment on circular graphs with skip links (\textsc{Skip-circles}) by~\citet{chen2019equivalence}, where the model needs to classify if a given circular graph has skip links of length $\{2, 3, 4, 5, 6, 9, 11, 12, 13, 16\}$.
For comparison, we try several other GNN modifications which increase expressiveness. For control, we run a vanilla GNN on these datasets. We then extend this base GNN with (i) ports~\citep{ports} (randomly assigned), (ii) node IDs~\citep{loukas2020graph}
(randomly permuted), and (iii) a random feature from the standard normal distribution~\citep{randomFeatures1}. The architecture for all GNNs is a 4-layer GIN with \texttt{sum} as aggregation and $\varepsilon=0$. For DropGNN $r=50$ runs are performed. For the \textsc{Skip-circles} dataset we use a 9-layer GIN instead, as the skip links can form cycles of up to 17 hops.
We train all methods for $1,000$ epochs and then evaluate the accuracy on the training set. We then test on a new graph (with new features). We report training and testing averaged across $10$ initializations in Table~\ref{tab:expressiveness}. We can see that DropGNN outperforms the competition.
\subsection{Sensitivity analysis}
We investigate the impact of the number of independent runs on the overall accuracy. Generally, we expect an increasing number of runs to more reliably produce informative dropouts. We train with a sufficiently high number of runs ($50$) with the same setting as before. Now, we reevaluate DropGNN but limit the runs to a smaller number. We measure the average accuracy over $10$ seeds with $10$ tests each and plot this average in Figure~\ref{fig:run_ablation} on three datasets: \textsc{Limits 1} (Figure~\ref{fig:run_ablation_limitsone}), \textsc{$4$-cycles} (Figure~\ref{fig:run_ablation_4cycles}), and \textsc{Triangles} (Figure~\ref{fig:run_ablation_triangles}). In all three datasets, more runs directly translate to higher accuracy.
\begin{figure}
\begin{subfigure}{0.3\textwidth}
\centering
\includegraphics[width=\textwidth]{plots/ablation_runs_limitsone.pdf}
\caption{\textsc{Limits 1}}
\label{fig:run_ablation_limitsone}
\end{subfigure}\hfill
\begin{subfigure}{0.3\textwidth}
\centering
\includegraphics[width=\textwidth]{plots/ablation_runs_fourcycles.pdf}
\caption{\textsc{$4-$cycles}}
\label{fig:run_ablation_4cycles}
\end{subfigure}\hfill
\begin{subfigure}{0.3\textwidth}
\centering
\includegraphics[width=\textwidth]{plots/ablation_runs_triangles.pdf}
\caption{\textsc{Triangles}}
\label{fig:run_ablation_triangles}
\end{subfigure}
\caption{Investigating the impact of the number of runs ($x-$axis) versus the classification accuracy ($y-axis$). In all three plots, having more runs allows for more stable dropout observations, increasing accuracy. The tradeoff is higher runtime since the model computes more runs.}
\label{fig:run_ablation}
\end{figure}
Next, we investigate the impact of the dropout probability $p$. We use the same setting as before, but instead of varying the number of runs in the reevaluation, we train and test with different probabilities $p$ on an exponential scale from $0.01$ to $0.64$. We also try $0$ (no dropout) and $0.95$ (almost everything is dropped). Figure~\ref{fig:p_ablation} shows the accuracy for each dropout probability, again averaged over $10$ seeds with $10$ tests each. Generally, DropGNN is robust to different values of $p$ until $p$ becomes very large.
\begin{figure}
\begin{subfigure}{0.3\textwidth}
\centering
\includegraphics[width=\textwidth]{plots/ablation_p_limitsone.pdf}
\caption{\textsc{Limits 1}}
\label{fig:p_ablation_limitsone}
\end{subfigure}\hfill
\begin{subfigure}{0.3\textwidth}
\centering
\includegraphics[width=\textwidth]{plots/ablation_p_fourcycles.pdf}
\caption{\textsc{$4-$cycles}}
\label{fig:p_ablation-4cycles}
\end{subfigure}\hfill
\begin{subfigure}{0.3\textwidth}
\centering
\includegraphics[width=\textwidth]{plots/ablation_p_triangles.pdf}
\caption{\textsc{Triangles}}
\label{fig:p_ablation_triangles}
\end{subfigure}
\caption{Investigating the impact of the dropout probability ($x-$axis) versus the classification accuracy ($y-$axis). DropGNN is robust to the choice of $p$ for decently small $p$. Choosing $p\approx \gamma^{-1}$ is a decent default that is shown by vertical black lines.}
\label{fig:p_ablation}
\end{figure}
\subsection{Graph classification}
\begin{table*}[ht]
\centering
\resizebox{0.9\textwidth}{!}{
\begin{tabular}{@{}l*{7}{S[table-format=-3.4]}@{}}
\toprule
{Model} & {Complexity} & {MUTAG} & {PTC} & {PROTEINS} & {IMDB-B} & {IMDB-M} & \\
\midrule
{WL subtree \citep{GIN, shervashidze2011weisfeiler}} & \makebox{$O(n)$} & \makebox{90.4 \raisebox{.2ex}{$\scriptstyle\pm$} 5.7} & \makebox{59.9 \raisebox{.2ex}{$\scriptstyle\pm$} 4.3} & \makebox{75.0 \raisebox{.2ex}{$\scriptstyle\pm$} 3.1} & \makebox{73.8 \raisebox{.2ex}{$\scriptstyle\pm$} 3.9} & \makebox{50.9 \raisebox{.2ex}{$\scriptstyle\pm$} 3.8}\\
{DCNN \citep{atwood2016diffusion}} & \makebox{$O(n)$} & \makebox{ - } & \makebox{ - } & \makebox{61.3 \raisebox{.2ex}{$\scriptstyle\pm$} 1.6} & \makebox{49.1 \raisebox{.2ex}{$\scriptstyle\pm$} 1.4} & \makebox{33.5 \raisebox{.2ex}{$\scriptstyle\pm$} 1.4}\\
{PatchySan \citep{niepert2016learning}} & \makebox{$O(n)$} & \makebox{89.0 \raisebox{.2ex}{$\scriptstyle\pm$} 4.4} & \makebox{62.3 \raisebox{.2ex}{$\scriptstyle\pm$} 5.7} & \makebox{75.0 \raisebox{.2ex}{$\scriptstyle\pm$} 2.5} & \makebox{71.0 \raisebox{.2ex}{$\scriptstyle\pm$} 2.3} & \makebox{45.2 \raisebox{.2ex}{$\scriptstyle\pm$} 2.8}\\
{DGCNN \citep{zhang2018end}} & \makebox{$O(n)$} & \makebox{85.8 \raisebox{.2ex}{$\scriptstyle\pm$} 1.7} & \makebox{58.6 \raisebox{.2ex}{$\scriptstyle\pm$} 2.5} & \makebox{75.5 \raisebox{.2ex}{$\scriptstyle\pm$} 0.9} & \makebox{70.0 \raisebox{.2ex}{$\scriptstyle\pm$} 0.9} & \makebox{47.8 \raisebox{.2ex}{$\scriptstyle\pm$} 0.9}\\
{GIN \citep{GIN}} & \makebox{$O(n)$} & \makebox{89.4 \raisebox{.2ex}{$\scriptstyle\pm$} 5.6} & \makebox{64.6 \raisebox{.2ex}{$\scriptstyle\pm$} 7.0} & \makebox{76.2 \raisebox{.2ex}{$\scriptstyle\pm$} 2.8} & \makebox{75.1 \raisebox{.2ex}{$\scriptstyle\pm$} 5.1} & \makebox{\textbf{52.3 \raisebox{.2ex}{$\scriptstyle\pm$} 2.8}}\\
{DropGIN (ours)} & \makebox{$O(rn), r\approx 20$} & \makebox{\textbf{90.4 \raisebox{.2ex}{$\scriptstyle\pm$} 7.0}} & \makebox{\textbf{66.3 \raisebox{.2ex}{$\scriptstyle\pm$} 8.6}} & \makebox{\textbf{76.3 \raisebox{.2ex}{$\scriptstyle\pm$} 6.1}} & \makebox{\textbf{75.7 \raisebox{.2ex}{$\scriptstyle\pm$} 4.2}} & \makebox{51.4 \raisebox{.2ex}{$\scriptstyle\pm$} 2.8}\\
\midrule
{1-2-3 GNN \citep{morris2019weisfeiler}} & \makebox{$O(n^4)$} & \makebox{86.1} & \makebox{60.9} & \makebox{75.5} & \makebox{\textbf{74.2}} & \makebox{49.5}\\
{PPGN \citep{maron2019provably}*} & \makebox{$O(n^3)$} & \makebox{\textbf{90.6 \raisebox{.2ex}{$\scriptstyle\pm$} 8.7}} & \makebox{\textbf{66.2 \raisebox{.2ex}{$\scriptstyle\pm$} 6.5}} & \makebox{\textbf{77.2 \raisebox{.2ex}{$\scriptstyle\pm$} 4.7}} & \makebox{73 \raisebox{.2ex}{$\scriptstyle\pm$} 5.8} & \makebox{\textbf{50.5 \raisebox{.2ex}{$\scriptstyle\pm$} 3.6}}\\
\bottomrule
\end{tabular}}
\caption{Graph classification accuracy (\%). The best performing model in each complexity class is highlighted in bold. *We report the best result achieved by either of the three versions of their model.}
\label{tab:graph_classification}
\end{table*}
We evaluate and compare our modified GIN model (DropGIN) with the original GIN model and other GNN models of various expressiveness levels on real-world graph classification datasets. We use three bioinformatics datasets (MUTAG, PTC, PROTEINS) and two social networks (IMDB-BINARY and IMDB-MULTI) \citep{yanardag2015deep}. Following \citep{GIN} node degree is used as the sole input feature for the IMDB datasets, while for the bioinformatics datasets the original categorical node features are used.
We follow the evaluation and model selection protocol described in \citep{GIN} and report the 10-fold cross-validation accuracies \citep{yanardag2015deep}. We extend the original 4-layer GIN model described in \citep{GIN} and use the same hyper-parameter selection as \citep{GIN}.
From Figure \ref{fig:p_ablation} we can see that it is usually safer to use a slightly larger $p$ than a slightly smaller one. Due to this, we set the node dropout probability to $p=\frac{2}{m}$, where $m$ is the mean number of nodes in the graphs in the dataset.
Our method successfully improves the results achieved by the original GIN model on the bioinformatics datasets (Table \ref{tab:graph_classification}) and is, in general, competitive with the more complex and computationally expensive expressive GNNs. Namely, the 1-2-3 GNN \citep{morris2019weisfeiler} which has expressive power close to that of 3-WL and $O(n^4)$ time complexity, and the Provably Powerful Graph Network (PPGN) \citep{maron2019provably} which has 3-WL expressive power and $O(n^3)$ time complexity. Compared to that, our method has only $O(rn)$ time complexity.
However, we do observe, that our approach slightly underperforms the original GIN model on the IMDB-M dataset.
Since the other expressive GNNs also underperform when compared to the original GIN model, it is possible that classifying graphs in this dataset rarely requires higher expressiveness. In such cases, our model can lose accuracy compared to the base GNN as many runs are required to achieve a fully stable dropout distribution.
\subsection{Graph property regression}
\begin{table*}[ht]
\centering
\resizebox{0.9\textwidth}{!}{
\begin{tabular}{@{}l*{8}{S[table-format=-3.4]}@{}}
\toprule
{Property} & \makebox{Unit} & {MPNN \citep{wu2018moleculenet}} & {1-GNN \citep{morris2019weisfeiler}} & {1-2-3 GNN \cite{morris2019weisfeiler}} & {PPGN \cite{maron2019provably}} & {DropMPNN} & {Drop-1-GNN}\\
\midrule
{$\mu$} & \makebox{Debye} & \makebox{0.358} & \makebox{0.493} & \makebox{0.473} & \makebox{0.0934} & \makebox{\textbf{0.059}*} & \makebox{0.453*}\\
{$\alpha$} & \makebox{$\text{Bohr}^3$} & \makebox{0.89} & \makebox{0.78} & \makebox{0.27} & \makebox{0.318} & \makebox{\textbf{0.173}*} & \makebox{0.767*}\\
{$\epsilon_{\text{HOMO}}$} & \makebox{Hartree} & \makebox{0.00541} & \makebox{0.00321} & \makebox{0.00337} & \makebox{\textbf{0.00174}} & \makebox{0.00193*} & \makebox{0.00306*}\\
{$\epsilon_{\text{LUMO}}$} & \makebox{Hartree} & \makebox{0.00623} & \makebox{0.00350} & \makebox{0.00351} & \makebox{0.0021} & \makebox{\textbf{0.00177}*} & \makebox{0.00306*}\\
{$\Delta\epsilon$} & \makebox{Hartree} & \makebox{0.0066} & \makebox{0.0049} & \makebox{0.0048} & \makebox{0.0029} & \makebox{\textbf{0.00282}*} & \makebox{0.0046*}\\
{$\langle R^2 \rangle$} & \makebox{$\text{Bohr}^2$} & \makebox{28.5} & \makebox{34.1} & \makebox{22.9} & \makebox{3.78} & \makebox{\textbf{0.392}*} & \makebox{30.83*}\\
{$\text{ZPVE}$} & \makebox{Hartree} & \makebox{0.00216} & \makebox{0.00124} & \makebox{0.00019} & \makebox{0.000399} & \makebox{\textbf{0.000112}*} & \makebox{0.000895*}\\
{$U_0$} & \makebox{Hartree} & \makebox{2.05} & \makebox{2.32} & \makebox{0.0427} & \makebox{\textbf{0.022}} & \makebox{0.0409*} & \makebox{1.80*}\\
{$U$} & \makebox{Hartree} & \makebox{2.0} & \makebox{2.08} & \makebox{0.111} & \makebox{\textbf{0.0504}} & \makebox{0.0536*} & \makebox{1.86*}\\
{$H$} & \makebox{Hartree} & \makebox{2.02} & \makebox{2.23} & \makebox{0.0419} & \makebox{\textbf{0.0294}} & \makebox{0.0481*} & \makebox{2.00*}\\
{$G$} & \makebox{Hartree} & \makebox{2.02} & \makebox{1.94} & \makebox{\textbf{0.0469}} & \makebox{0.24} & \makebox{0.0508*} & \makebox{2.12}\\
{$C_v$} & \makebox{cal/(mol K)} & \makebox{0.42} & \makebox{0.27} & \makebox{0.0944} & \makebox{\textbf{0.0144}} & \makebox{0.0596*} & \makebox{0.259*}\\
\bottomrule
\end{tabular}}
\caption{Mean absolute errors on QM9 dataset \citep{ramakrishnan2014quantum}. Best performing model is in bold and DropGNN versions that improve over the corresponding base GNN are marked with a *.}
\label{tab:graph_regression}
\end{table*}
We investigate how our dropout technique performs using different base GNN models on a different, graph regression, task. We use the QM9 dataset \citep{ramakrishnan2014quantum}, which consists of 134k organic molecules. The task is to predict 12 real-valued physical quantities for each molecule. In all cases, a separate model is trained to predict each quantity. We choose two GNN models to augment: MPNN~\citep{gilmer2017neural} and 1-GNN~\citep{morris2019weisfeiler}.
We set the DropGNN run count and node dropout probability the same way as done for graph classification. Following previous work \citep{morris2019weisfeiler, maron2019provably} the data is split into $80\%$ training, $10\%$ validation, and $10\%$ test sets. Both DropGNN model versions are trained for 300 epochs.
From Table~\ref{tab:graph_regression} we can see that Drop-1-GNN improves upon 1-GNN in most of the cases. In some of them, it even outperforms the much more computationally expensive 1-2-3-GNN, which uses higher-order graphs and has three times more parameters \citep{morris2019weisfeiler}. Meanwhile, DropMPNN always substantially improves on MPNN, often outperforming the Provably Powerful Graph Network (PPGN), which as you may recall scales as $O(n^3)$. This highlights the fact that while the DropGNN usually improves upon the base model, the final performance is highly dependent on the base model itself. For example, 1-GNN does not use skip connections, which might make retaining detailed information about the node's extended neighborhood much harder and this information is crucial for our dropout technique.
\section{Conclusion}\label{sec:conclusion}
We have introduced a theoretically motivated DropGNN framework, which allows us to easily increase the expressive power of existing message passing GNNs, both in theory and practice.
DropGNNs are also competitive with more complex GNN architectures which are specially designed to have high expressive power but have high computational complexity. In contrast, our framework allows for an arbitrary trade-off between expressiveness and computational complexity by choosing the number of rounds $r$ accordingly.
\subparagraph{Societal Impact.}
In summary, we proposed a model-agnostic architecture improvement for GNNs. We do not strive to solve a particular problem but to enhance the GNN toolbox. Therefore, we do not see an immediate impact on society. We found in our experiments that DropGNN works best on graphs with smaller degrees, such as molecular graphs. Therefore, we imagine that using DropGNN in these scenarios is interesting to explore further.
|
\section{Appendix}
\subsection{Hyperparameters}
We list our hyperparameters so that others can replicate our results. Nonetheless, our code has a \texttt{README.md} file with instructions on how to run our code per dataset.
\subsubsection{Implicit matrix hyperparameters}
\textbf{Planetoid datasets} \citep{planetoid} (Cora, CiteSeer, PubMed):
we
(implicitly) represent
$\widehat{\mathbf{M}}^\textrm{(NE)}$ with negative coefficient $\lambda=0.05$ and context window $C=3$; we construct $\widehat{\mathbf{M}}^\textrm{(NC)}$ with number of layers $L=15$. In one option, we use no dropout (first line of ``our models'' in Table \ref{table:resultsPlanetoid}) and in another (second line), we use dropout as described in \S\ref{sec:creative}.
\textbf{SNAP and node2vec datasets} (PPI, FB, AstroPh, HepTh): we (implicitly) represent $\widehat{\mathbf{M}}^\textrm{(NE)}$ with negative coefficient $\lambda=0.02$ and context window $C=10$.
\textbf{Stanford OGB Drug-Drug Interactions} (ogbl-DDI): we (implicitly) represent $\widehat{\mathbf{M}}^\textrm{(NE)}$ with negative coefficient $\lambda=1$ and context window $C=5$.
\textbf{Stanford OGB ArXiv} (ogbn-ArXiv): Unlike Planetoid datasets, we see in arXiv that
appending the decomposition of
$\widehat{\mathbf{M}}^\textrm{(NE)}$ onto $\mathbf{X}$ does not improve the validation performance. Nonetheless, we construct $\widehat{\mathbf{M}}^\textrm{(NC)}$ with $L=2$: we apply two forms of dropout, as explained in \S\ref{sec:exp_ogb}.
\textbf{Justifications}:
For Stanford OGB datasets, we make the selection based on the validation set. For the other (smaller) datasets, we choose them based on one datasets in each scenario (Cora for node classification and PPI for link prediction) and keep them fixed, for all datasets in the scenario -- hyperparameter tuning (on validation) per dataset may increase test performance, however, our goal is to make a general framework that helps across many GRL tasks, more-so than getting bold on specific tasks.
\subsubsection{Finetuning hyperparameters}
For Stanford OGB datasets (since competition is challenging), we use SVD to initialize a network that is finetuned on cross-entropy objectives.
For ogbl-DDI, we finetune model in \S\ref{sec:finetuneNE} using Adam optimizer with 1,000 positives and 10,000 negatives per batch, over one epoch, with learning rate of $1e^{-2}$.
For ogbn-ArXiv, we finetune model in \S\ref{sec:finetuneNC} using Adam optimizer using GTTF \citep{gttf} with \texttt{fanouts=[4, 4]} (i.e., for every labeled node in batch, GTTF samples 4 neighbors, and 4 of their neighbors). With batch size 500, we train with learning rate $=1e^{-3}$ for 7 epochs and with learning rate $=1e^{-4}$ for 8 epochs (\textit{i.e.}, 15 epochs total).
\subsection{SVD Implementation}
\label{sec:appendixsvdimplementation}
\definecolor{CommentColor}{HTML}{006619}
\algrenewcommand\algorithmicindent{1.0em}%
\begin{algorithm}[h]
\caption{
Computes rank-$k$ SVD of implicit matrix $\widehat{\mathbf{M}}$, defined symbolically as in \S\ref{sec:appendixsymbolic}.
Follows prototype of \citet{halko2009svd} but uses an alternative routine for orthonormalization.}
\label{alg:fsvd}
\begin{algorithmic}[1]
\State {\bfseries input:} SVD rank $k \in \mathbb{N}_+$; implicit matrix $\widehat{\mathbf{M}} \in \mathbb{R}^{r \times c}$ implementing functions in \S\ref{sec:appendixsymbolic}.
\Procedure{\textnormal iSVD}{$\widehat{\mathbf{M}}, k$}
\State $(r, c) \leftarrow \widehat{M}\texttt{.shape()}$
\State $\mathbf{Q} \leftarrow \hspace{0.05cm} \sim \mathcal{N}(0, 1)^{c \times 2k}$
\Comment{\textcolor{CommentColor}{Random matrix sampled from standard normal. Shape: $(c \times 2k)$}}
\For{$i \leftarrow 1$ {\bfseries to} \texttt{iterations}}
\State $\mathbf{Q} \leftarrow \texttt{orthonorm}(\widehat{\mathbf{M}}\texttt{.dot}(\mathbf{Q}))$
\Comment{\textcolor{CommentColor}{$(r \times 2k)$}}
\State $\mathbf{Q} \leftarrow \texttt{orthonorm}(\widehat{\mathbf{M}}\texttt{.T}()\texttt{.dot}(\mathbf{Q}))$
\Comment{\textcolor{CommentColor}{$(c \times 2k)$}}
\EndFor
\State $\mathbf{Q} \leftarrow \texttt{orthonorm}(\widehat{\mathbf{M}}\texttt{.dot}(\mathbf{Q}))$
\Comment{\textcolor{CommentColor}{$(r \times 2k)$}}
\State $\mathbf{B} \leftarrow \widehat{\mathbf{M}}\texttt{.T}()\texttt{.dot}(\mathbf{Q})^\top$
\Comment{\textcolor{CommentColor}{$(2k \times c)$}}
\State $\mathbf{U}, \mathbf{s}, \mathbf{V}^\top \leftarrow \texttt{tf.linalg.svd}(B)$
\State $\mathbf{U} \leftarrow \mathbf{Q} \times \mathbf{U}$
\Comment{\textcolor{CommentColor}{$(r \times 2k)$}}
\State \textbf{return } $\mathbf{U}[:, :k], \mathbf{s}[:k], \mathbf{V}[:, :k]$
\EndProcedure
\end{algorithmic}
\end{algorithm}
The prototype algorithm for SVD described by \citet{halko2009svd} suggests using QR-decomposition to for orthonormalizing\footnote{$\widehat{\mathbf{Q}} \leftarrow \texttt{orthonorm}(\mathbf{Q})$ yields $\widehat{\mathbf{Q}}$ such that $\widehat{\mathbf{Q}}^\top \widehat{\mathbf{Q}} = \mathbf{I}$ and $\texttt{span}(\widehat{\mathbf{Q}}) = \texttt{span}(\mathbf{Q})$.} but we rather use Cholesky decomposition which is faster to calculate (our framework is written on top of TensorFlow 2.0 \citep{tensorflow}). Fig.~\ref{fig:time} in the main paper shows the time comparison.
\begin{algorithm}[h]
\caption{Comparison of orthonorm routines. \textbf{Left}: \citet{halko2009svd} (via QR decomposition). \textbf{Right}: ours (via Cholesky decomposition).
}
\label{algs:orthonorm}
\begin{minipage}{0.48\linewidth}
\begin{algorithmic}[1]
\State {\bfseries input:} matrix $\mathbf{Q}$.
\Procedure{\textnormal{\texttt{orthonormHalko}}}{$\mathbf{Q}$}
\State $\widehat{\mathbf{Q}}, \mathbf{R} \leftarrow \texttt{QRdecomposition}(\mathbf{Q})$
\State \textbf{return} $\widehat{\mathbf{Q}}$
\EndProcedure
\end{algorithmic}
\end{minipage}
\hspace{0.03\linewidth}
\begin{minipage}{0.48\linewidth}
\begin{algorithmic}[1]
\State {\bfseries input:} matrix $\mathbf{Q}$.
\Procedure{\textnormal{\texttt{orthonormOurs}}}{$\mathbf{Q}$}
\State $\mathbf{L} \leftarrow \texttt{Choleskydecomposition}(\mathbf{Q}^\top \mathbf{Q})$
\State \textbf{return} $\widehat{\mathbf{Q}} \leftarrow \mathbf{Q} \times (\mathbf{L}^{-1})^\top$
\EndProcedure
\end{algorithmic}
\end{minipage}
\end{algorithm}
\subsection{Symbolic representation}
\label{sec:appendixsymbolic}
Each node is a python class instance, that inherits base-class \texttt{SymbolicPF} (where PF stands for \textit{product function}.
Calculating SVD of (implicit) matrix $\mathbf{M} \in \mathbb{R}^{r \times c}$ is achievable, without explicitly knowing entries of $\widehat{\mathbf{M}}$, if we are able to multiply $\widehat{\mathbf{M}}$ by arbitrary $\mathbf{G}$.
\begin{enumerate}[topsep=0pt,itemsep=0pt,leftmargin=15pt]
\item \texttt{shape()} property must return a tuple of the shape of the underlying implicit matrix e.g. $=(r, c)$.
\item \texttt{dot}($\mathbf{G}$) must right-multiply an arbitrary explicit matrix $\mathbf{G} \in \mathbb{R}^{c \times .}$ with (implicit) $\widehat{\mathbf{M}}$ as $\widehat{\mathbf{M}} \times \mathbf{G}$ returning explicit matrix $\in \mathbb{R}^{r \times .}$
\item \texttt{T()} must return an instance of \texttt{SymbolicPF} which is the (implicit) transpose of $\widehat{\mathbf{M}}$ i.e. with \texttt{.shape() = $(c, r)$}
\end{enumerate}
The names of the functions were purposefully chosen to match common naming conventions, such as of numpy. We now list our current implementations of \texttt{PF} classes used for symbolically representing the design matrices discussed in the paper ($\widehat{\mathbf{M}}^\textrm{(NE)}$ and $\widehat{\mathbf{M}}^\textrm{(NC)}$).
\subsubsection{Leaf nodes}
The leaf nodes \textbf{explicitly} hold the underlying matrix. Next, we show an implementation for a dense matrix (left), and another for a sparse matrix (right):
\newline
\begin{minipage}{0.49\linewidth}
\begin{lstlisting}[columns=fullflexible,language=Python]
class DenseMatrixPF(SymbolicPF):
"""(implicit) matrix is explicitly
stored dense tensor."""
def __init__(self, m):
self.m = tf.convert_to_tensor(m)
def dot(self, g):
return tf.matmul(self.m, g)
@property
def shape(self):
return self.m.shape
@property
def T(self):
return DenseMatrixPF(
tf.transpose(self.m))
\end{lstlisting}
\end{minipage}
\begin{minipage}{0.49\linewidth}
\begin{lstlisting}[columns=fullflexible,language=Python]
class SparseMatrixPF(SymbolicPF):
"""(implicit) matrix is explicitly
stored dense tensor."""
def __init__(self, m):
self.m = scipy_csr_to_tf_sparse(m)
def dot(self, g):
return tf.sparse.sparse_dense_matmul(
self.m, g)
@property
def shape(self):
return self.m.shape
@property
def T(self):
return SparseMatrixPF(
tf.sparse.transpose(self.m))
\end{lstlisting}
\end{minipage}
\subsubsection{Symbolic nodes}
Symbolic nodes \textbf{implicitly} hold a matrix. Specifically, their constructors (\texttt{\_\_init\_\_}) accept one-or-more other (leaf or) symbolic nodes and their implementations of \texttt{.shape()} and \texttt{\texttt{.dot()}} invokes those of the nodes passed to their constructor. Let us take a few examples:
\begin{minipage}{0.49\linewidth}
\begin{lstlisting}[columns=fullflexible,language=Python]
class SumPF(SymbolicPF):
def __init__(self, pfs):
self.pfs = pfs
for pf in pfs:
assert pf.shape == pfs[0].shape
def dot(self, m):
sum_ = self.pfs[0].dot(m)
for pf in self.pfs[1:]:
sum_ += pf.dot(m)
return sum_
@property
def T(self):
return SumPF([f.T for f in self.pfs])
@property
def shape(self):
return self.pfs[0].shape
\end{lstlisting}
\end{minipage}~
\begin{minipage}{0.49\linewidth}
\begin{lstlisting}[columns=fullflexible,language=Python]
class ProductPF(SymbolicPF):
def __init__(self, pfs):
self.pfs = pfs
for i in range(len(pfs) - 1):
assert (pfs[i].shape[1]
== pfs[i+1].shape[0])
@property
def shape(self):
return (self.pfs[0].shape[0],
self.pfs[-1].shape[1])
@property
def T(self):
return ProductPF(reversed(
[f.T for f in self.pfs]))
def dot(self, m, cache=None):
product = m
for pf in reversed(self.pfs):
product = pf.dot(product)
return product
\end{lstlisting}
\end{minipage}
\newline
Our attached code (will be uploaded to github, after code review process) has concrete implementations for other \texttt{PF}s. For example, \texttt{GatherRowsPF} and \texttt{GatherColsPF}, respectively, for (implicitly) selecting row and column slices; as well as, concatenating matrices, and multiplying by scalar. Further, the actual implementation has further details than displayed, for instance, for implementing lazy caching \S\ref{sec:lazycache}.
\subsubsection{Syntactic sugar}
Additionally, we provide a couple of top-level functions e.g. \texttt{fsvd.sum}(), \texttt{fsvd.gather}(), with interfaces similar to numpy and tensorflow (but expects implicit PF as input and outputs the same). Further, we override python operators (e.g., \texttt{+}, \texttt{-}, \texttt{*}, \texttt{**}, \textbf{@}) on the base \texttt{SymbolicPF} that instantiates symbolic node implementations, as:
\begin{lstlisting}[columns=fullflexible,language=Python]
class SymbolicPF:
# ....
def __add__(self, other):
return SumPF([self, other])
def __matmul__(self, other):
return ProductPF([self, other])
def __sub__(self, other):
return SumPF([self, TimesScalarPF(-1, other)])
def __mul__(self, scalar):
return TimesScalarPF(scalar, self)
def __rmul__(self, scalar):
return TimesScalarPF(scalar, self)
def __pow__(self, integer):
return ProductPF([self] * int(integer))
\end{lstlisting}
This allows composing implicit matrices using syntax commonly-used for calculating explicit matrices. We hope this notation could ease the adoption of our framework.
\subsubsection{Optimization by lazy caching}
\label{sec:lazycache}
One can come up with multiple equivalent mathematical expressions. For instance, for matrices $\mathbf{A}, \mathbf{B}, \mathbf{C}$, the expressions $\mathbf{A} \times \mathbf{B} + \mathbf{A}^2 \mathbf{C}$ and $\mathbf{A} \times (\mathbf{B} + \mathbf{A} \mathbf{C})$ are equivalent. However, the first one should be cheaper to compute.
At this point, we have not implemented such equivalency-finding: we do not
Substitute expressions with their more-efficient equivalents.
Rather,
we implement (simple) optimization by \textit{lazy-caching}. Specifically, when computing product $\widehat{\mathbf{M}}^\top \mathbf{G}$, as the \texttt{dot()} methods are recursively called, we populate a cache (python \texttt{dict}) of intermediate products with key-value pairs:
\begin{itemize}[itemsep=0pt,topsep=0pt]
\item key: ordered list of (references of) \texttt{SymbolicPF} instances that were multiplied by $\mathbf{G}$;
\item value: the product of the list times $\mathbf{G}$.
\end{itemize}
While this optimization scheme is suboptimal, and trades memory for computation, we observe up to 4X speedups on our tasks (refer to Fig.~\ref{fig:time}). Nonetheless, in the future, we hope to utilize an open-source \textit{expression optimizer}, such as TensorFlow's \citep{grappler}.
\subsection{Approximating the integral of Gaussian 1d kernel}
\label{appendix:integral}
The integral from Equation \ref{eq:gaussiankernel} can be approximated using softmax, as:
\begin{align*}
\int_\Omega \mathbf{S}^x \overline{\mathcal{N}}(x \mid \mu, s ) \,dx
= \frac{\int_\Omega \mathbf{S}^x \exp\left(-\frac{1}{2}\left(\frac{x - \mu}{s}\right)^2\right) \,dx}{\int_\Omega \exp\left(-\frac{1}{2}\left(\frac{y - \mu}{s}\right)^2\right) \,dy} &\approx \frac{\sum_{x \in \overline{\Omega}} \mathbf{S}^x \exp\left(-\frac{1}{2}\left(\frac{x - \mu}{s}\right)^2\right) \,dx}{\sum_{y \in \overline{\Omega}} \exp\left(-\frac{1}{2}\left(\frac{y - \mu}{s}\right)^2\right) \,dy} \\
& = \left\{\mathbf{S}^x \right\}_{x \in \overline{\Omega}} \otimes \mathop{\textrm{softmax}}_{x \in \overline{\Omega}}\left( - (x - \mu)^2 \exp(\overline{s}) \right)
\end{align*}
where the first equality expands the definition of truncated normal, i.e. divide by partition function, to make the mass within $\Omega$ sum to 1. In our experiments, we use $\Omega=[0.5, 2]$. The $\approx$ comes as we use discretized $\overline{\Omega}=\{0.5, 0.505, 0.51, 0.515, \dots, 2.0 \}$ (i.e., with 301 entries). Finally, the last expression contains a constant tensor (we create it only once) containing $\mathbf{S}$ raised to every power in $\overline{\Omega}$, stored along (tensor) axis which gets multiplied (via tensor product i.e. \textit{broadcasting}) against softmax vector (also of 301 entries, corresponding to $\overline{\Omega}$). We parameterize with two scalars $\mu, \overline{s} \in \mathbb{R}$ i.e. implying $\overline{s} = \log \frac{1}{2s^2}$
\subsection{Analysis, theorems and proofs}
\subsubsection{Norm regularization of wide models}
Proof of Theorem \ref{theorem:min_norm} is common in university-level linear algebra courses, but here for completeness. It imples that
if $\widehat{\mathbf{M}}^\textrm{(NC)}$ is too wide,
then we need \textbf{not} to worry much about \textit{overfitting}.
\begin{proof} of Theorem \ref{theorem:min_norm}
Assume $\mathbf{Y} =\mathbf{y}$ is a column vector (the proof can be generalized to matrix $\mathbf{Y}$ by repeated column-wise application\footnote{Minimizer of Frobenius norm is composed, column-wise, of minimizers $\mathop{\mathrm{argmin}}_{\widehat{\mathbf{M}} \mathbf{W}_{:, j} = \mathbf{Y}_{:, j}} ||\mathbf{W}_{:, j}||_2^2$, $\forall j$.}). $\textrm{SVD}_k(\widehat{\mathbf{M}})$ for $k\geq \textrm{rank}(\widehat{\mathbf{M}})$, recovers the solution:
\begin{equation}
\widehat{\mathbf{W}}^* = \left(\widehat{\mathbf{M}}\right)^\dagger \mathbf{y} = \widehat{\mathbf{M}}^\top \left(\widehat{\mathbf{M}} \widehat{\mathbf{M}}^\top \right)^{-1} \mathbf{y}.
\end{equation}
The \textit{Gram matrix} $\widehat{\mathbf{M}} \widehat{\mathbf{M}}^\top$ is nonsingular as the rows of $\widehat{\mathbf{M}}$ are linearly independent.
To prove the claim let us first verify that $\widehat{\mathbf{W}}^* \in \widehat{\mathcal{W}}^*$:
\begin{equation*}
\widehat{\mathbf{M}} \widehat{\mathbf{W}}^*
= \widehat{\mathbf{M}} \widehat{\mathbf{M}}^\top \left(\widehat{\mathbf{M}} \widehat{\mathbf{M}}^\top \right)^{-1} \mathbf{y} = \mathbf{y}.
\end{equation*}
Let $\widehat{\mathbf{W}}_p \in \widehat{\mathcal{W}}^*$.
We must show that $||\widehat{\mathbf{W}}^*||_2 \le ||\widehat{\mathbf{W}}_p||_2$.
Since $\widehat{\mathbf{M}} \widehat{\mathbf{W}}_p = \mathbf{y}$ and $\widehat{\mathbf{M}} \widehat{\mathbf{W}}^* = \mathbf{y}$, their subtraction gives:
\begin{equation}
\label{eq:nullM}
\widehat{\mathbf{M}} ( \widehat{\mathbf{W}}_p - \widehat{\mathbf{W}}^* ) = 0.
\end{equation}
It follows that $( \widehat{\mathbf{W}}_p - \widehat{\mathbf{W}}^* ) \perp \widehat{\mathbf{W}}^*$:
\begin{align*}
( \widehat{\mathbf{W}}_p - \widehat{\mathbf{W}}^* )^\top \widehat{\mathbf{W}}^* &= ( \widehat{\mathbf{W}}_p - \widehat{\mathbf{W}}^* )^\top \widehat{\mathbf{M}}^\top \left(\widehat{\mathbf{M}} \widehat{\mathbf{M}}^\top \right)^{-1} \mathbf{y} \\
&= \underbrace{(\widehat{\mathbf{M}} (\widehat{\mathbf{W}}_p - \widehat{\mathbf{W}}^* ))^\top}_{=0 \textrm{ due to Eq.~\ref{eq:nullM}}} \left(\widehat{\mathbf{M}} \widehat{\mathbf{M}}^\top \right)^{-1} \mathbf{y} = 0
\end{align*}
Finally, using Pythagoras Theorem (due to $\perp$):
\begin{align*}
||\widehat{\mathbf{W}}_p||_2^2 &= ||\widehat{\mathbf{W}}^* + \widehat{\mathbf{W}}_p - \widehat{\mathbf{W}}^*||_2^2 \\
&= ||\widehat{\mathbf{W}}^*||_2^2 + ||\widehat{\mathbf{W}}_p - \widehat{\mathbf{W}}^*||_2^2 \ge ||\widehat{\mathbf{W}}^*||_2^2
\end{align*}
\end{proof}
\subsubsection{At initialization, deep model is identical to linear (convexified) model}
\begin{proof}
of Theorem \ref{theorem:initialization}:
The layer-to-layer ``positive'' and ``negative'' weight matrices are initialized as: $ \mathbf{W}_\textrm{(p)}^{(l)} = -\mathbf{W}_\textrm{(n)}^{(l)} = \mathbf{I} $. Therefore, at initialization:
\begin{align*}
\mathbf{H}^{(l+1)} &= \left[\widehat{\mathbf{A}} \mathbf{H}^{(l)} \mathbf{W}_\textrm{(p)}^{(l)} \right]_+ - \left[\widehat{\mathbf{A}} \mathbf{H}^{(l)} \mathbf{W}_\textrm{(n)}^{(l)} \right]_+ = \left[\widehat{\mathbf{A}} \mathbf{H}^{(l)} \right]_+ - \left[ - \widehat{\mathbf{A}} \mathbf{H}^{(l)} \right]_+ \\
&= \mathds{1}_{\left[\widehat{\mathbf{A}} \mathbf{H}^{(l)} \ge 0 \right]} \circ \left(\widehat{\mathbf{A}} \mathbf{H}^{(l)} \right) - \mathds{1}_{\left[ - \widehat{\mathbf{A}} \mathbf{H}^{(l)} \ge 0 \right]} \circ \left(-\widehat{\mathbf{A}} \mathbf{H}^{(l)} \right) \\
&=\mathds{1}_{\left[\widehat{\mathbf{A}} \mathbf{H}^{(l)} \ge 0 \right]} \circ \left(\widehat{\mathbf{A}} \mathbf{H}^{(l)} \right) + \mathds{1}_{\left[ - \widehat{\mathbf{A}} \mathbf{H}^{(l)} \ge 0 \right]} \circ \left(\widehat{\mathbf{A}} \mathbf{H}^{(l)} \right) \\
&=\left( \mathds{1}_{\left[\widehat{\mathbf{A}} \mathbf{H}^{(l)} \ge 0 \right]} + \mathds{1}_{\left[ - \widehat{\mathbf{A}} \mathbf{H}^{(l)} \ge 0 \right]} \right) \circ \left(\widehat{\mathbf{A}} \mathbf{H}^{(l)} \right) \\
&= \widehat{\mathbf{A}} \mathbf{H}^{(l)},
\end{align*}
where the first line comes from the initialization; the second line is an alternative definition of relu: the indicator function $\mathds{1}$ is evaluated element-wise and evaluates to 1 in positions its argument is true and to 0 otherwise; the third line absorbs the two negatives into a positive; the fourth by factorizing; and the last by noticing that \textbf{exactly one} of the two indicator functions evaluates to 1 almost everywhere, except at the boundary condition i.e. at locations where $\widehat{\mathbf{A}}\mathbf{H}^{(l)} = 0$ but there the right-term makes the Hadamard product 0 regardless. It follows that, since $\mathbf{H}^{(0)} = \mathbf{X}$, then $\mathbf{H}^{(1)} = \widehat{\mathbf{A}} \mathbf{X}$, \hspace{0.3cm} $\mathbf{H}^{(2)} = \widehat{\mathbf{A}}^2 \mathbf{X}$, \dots, $\mathbf{H}^{(L)} = \widehat{\mathbf{A}}^L \mathbf{X}$.
The layer-to-output positive and negative matrices are initialized as: $\mathbf{W}_\textrm{(op)}^{(l)} = -\mathbf{W}_\textrm{(on)}^{(l)} = \widehat{\mathbf{W}}^*_{[dl \, : \, d(l+1)]}$. Therefore, at initialization, the final output of the model is:
\begin{align*}
\mathbf{H}=\sum_{l=0}^{l=L} \left[\mathbf{H}^{(l)} \mathbf{W}_\textrm{(op)}^{(l)} \right]_+ - \left[\mathbf{H}^{(l)} \mathbf{W}_\textrm{(on)}^{(l)} \right]_+
&= \sum_{l=0}^{l=L} \left[\mathbf{H}^{(l)} \widehat{\mathbf{W}}^*_{[dl \, : \, d(l+1)]} \right]_+ - \left[- \mathbf{H}^{(l)} \widehat{\mathbf{W}}^*_{[dl \, : \, d(l+1)]} \right]_+ \\
&= \sum_{l=0}^{l=L} \mathbf{H}^{(l)} \widehat{\mathbf{W}}^*_{[dl \, : \, d(l+1)]},
\end{align*}
where the first line comes from the definition and the initialization. The second line can be arrived by following exactly the same steps as above: expanding the re-writing the ReLu using indicator notation, absorbing the negative, factorizing, then lastly unioning the two indicators that are mutually exclusive almost everywhere. Finally, the last summation can be expressed as a block-wise multiplication between two (partitioned) matrices:
\begin{align*}
\mathbf{H} = \left[
\begin{matrix}[c;{1pt/1pt}c;{1pt/1pt}c;{1pt/1pt}c]
\mathbf{H}^{(0)} & \mathbf{H}^{(1)} & \dots & \mathbf{H}^{(L)} \end{matrix} \right] \left[
\begin{matrix}[l]
\widehat{\mathbf{W}}^*_{[0 \, : \, d]} \\
\widehat{\mathbf{W}}^*_{[d \, : \, 2d]} \\
\dots \\
\widehat{\mathbf{W}}^*_{[dL \, : \, d(L+1)]}
\end{matrix}
\right]
& =\left[
\begin{matrix}[c;{1pt/1pt}c;{1pt/1pt}c;{1pt/1pt}c]
\mathbf{X} & \widehat{\mathbf{A}} \mathbf{X} & \dots & \widehat{\mathbf{A}}^L \mathbf{X} \end{matrix} \right] \widehat{\mathbf{W}}^* \\
&= \widehat{\mathbf{M}}^\textrm{(NC)} \widehat{\mathbf{W}}^* = \mathbf{H}^\textrm{(NC)}_\textrm{linearized}
\end{align*}
\end{proof}
\subsubsection{Computational Complexity and Approximation Error}
\begin{theorem}
\label{theorem:linear_time}
{\normalfont (Linear Time)}
Implicit SVD (Alg.~\ref{alg:fsvd}) trains our convexified GRL models in time linear in the graph size.
\end{theorem}
\begin{proof} of Theorem \ref{theorem:linear_time} for our two model families:
\begin{enumerate}[itemsep=0pt,topsep=0pt,leftmargin=12pt]
\item For rank-$k$ SVD of $\widehat{\mathbf{M}}^\textrm{(NE)}$:
Let cost of running $\widehat{\mathbf{M}}^\textrm{(NC)}\texttt{.dot()}$ be $T_\textrm{mult}$. The run-time to compute SVD, as derived in Section 1.4.2 of \citep{halko2009svd}, is:
\begin{equation}
\mathcal{O}(k T_\textrm{mult} + (r + c) k^2).
\end{equation}
Since $\widehat{\mathbf{M}}^\textrm{(NE)}$ can be defined as $C$ (context window size) multiplications with sparse $n \times n$ matrix $\mathcal{T}$ with $m$ non-zero entries, then
running $\textrm{iSVD}_k(\widehat{\mathbf{M}}^\textrm{(NE)})$ costs:
\begin{equation}
\mathcal{O}(k m C + n k^2)
\end{equation}
\item For rank-$k$ SVD over $\widehat{\mathbf{M}}^\textrm{(NC)}$: Suppose feature matrix contains $d$-dimensional rows. One can calculate $\widehat{\mathbf{M}}^\textrm{(NC)} \in \mathbb{R}^{n \times L d}$ with $L$ sparse multiplies in $\mathcal{O}(L m d)$. Calculating and running SVD \citep[see Section 1.4.1 of][]{halko2009svd} on $\widehat{\mathbf{M}}^\textrm{(JKN)}$ costs total of:
\begin{equation}
\mathcal{O}(n d L \log(k) + (n + d L ) k ^ 2 + L m d).
\end{equation}
\end{enumerate}
Therefore, training time is linear in $n$ and $m$.
\end{proof}
Contrast with methods of WYS \citep{wys} and NetMF \citep{qiu2018network}, which require assembling a dense $n \times n$ matrix requiring $\mathcal{O}(n^2)$ time to decompose. One wonders: how far are we from the optimal SVD with a linear-time algorithm? The following bounds the error.
\begin{theorem}
\label{theorem:error}
{\normalfont (Exponentially-decaying Approx. Error)}
Rank-$k$ randomized SVD algorithm of \citet{halko2009svd} gives an approximation error that can be brought down, exponentially-fast, to no more than twice of the approximation error of the optimal (true) SVD.
\end{theorem}
\begin{proof} of Theorem \ref{theorem:error} is in Theorem 1.2 of \citet{halko2009svd}
\end{proof}
Consequently,
compared to $\widetilde{\textrm{NetMF}}$ of \citep{qiu2018network}, which incurs unnecessary estimation error as they sample the matrix entry-wise,
our estimation error can be brought-down exponentially by increasing the \texttt{iterations} parameter of Alg.~\ref{alg:fsvd}. In particular, as long as we can compute products against the matrix, we can decompose it, almost as good as if we had the individual matrix entries.
\section{Introduction}
Truncated Singular Value Decomposition (SVD)
provides solutions to a variety of mathematical problems, including computing a matrix rank, its pseudo-inverse, or mapping its rows and columns onto the \textit{orthonormal singular bases} for low-rank approximations.
Machine Learning (ML) software frameworks (such as TensorFlow) offer efficient SVD implementations, as SVD
can estimate solutions for a variaty of tasks, e.g., in
\textit{computer vision} \citep[][]{turk1991eigenfaces},
\textit{weather prediction} \citep{molteni1996ECMWF},
\textit{recommendation} \citep{koren2009recsys},
\textit{language} \citep{deerwester1990-lsi, levy2014-neural},
and more-relevantly, \textit{graph representation learning} (GRL) \citep{qiu2018network}.
SVD's benefits include training models, without calculating gradients, to arrive at globally-unique solutions, optimizing Frobenius-norm objectives (\S\ref{sec:prelimsvd}), without requiring hyperparameters for the learning process, such as the choice of the learning algorithm, step-size, regularization coefficient, etc.
Typically, one \textit{constructs} a
\textbf{design matrix} $\mathbf{M}$, such that, its decomposition provides a solution to a task of interest.
Unfortunately, existing popular ML frameworks \citep{tensorflow, pytorch}
cannot calculate the SVD of an arbitrary linear matrix given its computation graph:
they
compute the matrix (entry-wise)
then\footnote{TensorFlow can caluclate matrix-free SVD if one implements a \texttt{LinearOperator},
as such, our code could be re-implemented as a routine that can convert \texttt{TensorGraph} to \texttt{LinearOperator}.} its decomposition.
This limits the scalability of these libraries in several cases of interest, such as in GRL, when explicit calculation of the matrix is prohibitive due to memory constraints.
These limitations
render SVD as impractical for achieving state-of-the-art (SOTA) for tasks at hand.
This has been circumvented by \citet{qiu2018network} by sampling $\mathbf{M}$ entry-wise, but this produces sub-optimal estimation error and experimentally degrades the empirical test performance (\S\ref{sec:experiments}: Experiments).
We design a software library that allows \textbf{symbolic} definition of $\mathbf{M}$, via composition of matrix operations, and we implement an SVD algorithm that can decompose $\mathbf{M}$ from said symbolic representation, without need to compute $\mathbf{M}$.
This is valuable for many GRL tasks,
where the design matrix $\mathbf{M}$ is too large, \textit{e.g.}, quadratic in the input size. With our implementation, we show that SVD can perform learning, orders of magnitudes faster than current alternatives.
Currently, SOTA GRL models
are generally graph neural networks trained
to optimize cross-entropy objectives.
Their inter-layer non-linearities
place their (many) parameters onto a non-convex objective surface where convergence is rarely verified\footnote{Practitioners rarely verify that $\nabla_\theta J = 0$, where $J$ is mean train objective and $\theta$ are model parameters.}.
Nonetheless,
these models can be \textit{convexified} (\S\ref{sec:convexification}) and trained via SVD, \textbf{if} we remove nonlinearities between layers \textbf{and} swap the cross-entropy objective with Frobenius norm minimization.
Undoubtedly, such linearization incurs a drop of accuracy on empirical test performance.
Nonetheless, we show that the (convexified) model's parameters learned by SVD can provide initialization to deeper (non-linear) models, which then can be fine-tuned on cross-entropy objectives. The non-linear models are endowed with our novel Split-ReLu layer, which has twice as many parameters as a ReLu fully-connected layer, and behaves as a linear layer when its parameters reside on some hyperplane (\S\ref{sec:finetuneNC}).
Training on modest hardware (e.g., laptop) is sufficient for this learning pipeline (convexify $\rightarrow$ SVD $\rightarrow$ fine-tune) yet it trains much faster than current approaches, that are commonly trained on expensive hardware.
We summarize our contributions as:
\begin{enumerate}[itemsep=0pt, topsep=0pt,leftmargin=15pt]
\item We open-source a flexible python software library that allows symbolic definition of matrices and computes their SVD without explicitly calculating them.
\item
We linearize popular GRL models,
and train them via SVD of design matrices.
\item We show that fine-tuning a few parameters on-top of the SVD initialization sets state-of-the-art on many GRL tasks while, overall, training orders-of-magnitudes faster.
\end{enumerate}
\section{Preliminaries \& notation}
We denote a graph with $n$ nodes and $m$ edges with an \textit{adjacency matrix} $\mathbf{A} \in \mathbb{R}^{n\times n}$ and additionally, if nodes have ($d$-dimensional) features,
with a \textit{feature matrix} $\mathbf{X} \in \mathbb{R}^{n \times d}$.
If nodes $i, j \in [n]$ are connected then
$\mathbf{A}_{ij}$ is set to their edge weight and otherwise $\mathbf{A}_{ij} = 0$.
Further, denote the (row-wise normalized) \textit{transition matrix} as $\mathcal{T} = \mathbf{D}^{-1} \mathbf{A}$ and denote the symmetrically normalized adjacency with self-connections as
$\widehat{\mathbf{A}} = (\mathbf{D} + \mathbf{I})^{-\frac12} (\mathbf{A} + \mathbf{I}) (\mathbf{D} + \mathbf{I})^{-\frac12}$ where $\mathbf{I}$ is identity matrix.
We review model classes: (1) network embedding and (2) message passing that we define as follows.
The first inputs a graph ($\mathbf{A}$, $\mathbf{X}$) and outputs \textit{node embedding matrix} $ \mathbf{Z} \in \mathbb{R}^{n \times z}$ with $z$-dimensions per node.
$\mathbf{Z}$ is then used for an upstream task, \textit{e.g.}, link prediction.
The second
class utilizes a function $\mathbf{H} : \mathbb{R}^{n \times n} \times \mathbb{R}^{n \times d} \rightarrow \mathbb{R}^{n \times z}$ where the function $\mathbf{H}(\mathbf{A}, \mathbf{X})$ is usually directly trained on the upstream task, \textit{e.g.}, node classification.
In general, the first class is transductive while the second is inductive.
\subsection{Network embedding models based on DeepWalk \& node2vec}
\label{sec:prelimNE}
The seminal work of DeepWalk \citep{perozzi2014deepwalk} embeds nodes of a network using a two-step process: (i) simulate random walks on the graph -- each walk generating a sequence of node IDs then (ii) pass the walks (node IDs) to a language word embedding algorithm, e.g. word2vec \citep{word2vec}, as-if each walk is a sentence. This work was extended by node2vec \citep{grover2016node2vec} among others.
It has been shown by \citet{wys} that the learning outcome of the two-step process of DeepWalk is equivalent, in expectation, to optimizing a single objective\footnote{Derivation is in \citep{wys}. Unfortunately, matrix in Eq.~\ref{eq:wys} is dense with $\mathcal{O}(n^2)$ nonzeros.}:
\begin{equation}
\min_{\mathbf{Z} = \{\mathbf{L}, \mathbf{R}\} } \sum_{(i, j) \in [n] \times [n]} \left[ - \mathop{\mathbb{E}}_{q \sim Q}\left[ \mathcal{T}^q \right] \circ \log \sigma ( \mathbf{L} \mathbf{R}^\top ) - \lambda (1 - \mathbf{A}) \circ \log(1- \sigma ( \mathbf{L} \mathbf{R}^\top ) ) \right]_{ij},
\label{eq:wys}
\end{equation}
where $\mathbf{L}, \mathbf{R} \in \mathbb{R}^{n \times \frac{z}{2}}$ are named by word2vec as the \textit{input} and \textit{output} embedding matrices, $\circ$ is Hadamard product, and the $\log(.)$ and the standard logistic $\sigma(.) = (1+\exp(.))^{-1}$ are applied element-wise.
The objective above is weighted cross-entropy
where the (left) positive term weighs the
dot-product $\mathbf{L}_i^\top \mathbf{R}_j$ by the (expected) number of random walks simulated from $i$ and passing through $j$, and the (right) negative term weighs non-edges $(1-\mathbf{A})$ by scalar $\lambda \in \mathbb{R}_{+}$.
The \textit{context distribution} $Q$ stems from step (ii) of the process.
In particular, word2vec accepts hyperparameter \textit{context window size} $C$ for its stochasatic sampling:
when it samples a \textit{center token} (node ID), it then samples its \textit{context tokens} that are up-to distance $c$ from the center. The integer $c$ is sampled from a coin flip uniform on the integers $[1, 2, \dots, C]$ -- as detailed by Sec.3.1 of \citep{levy-goldberg}. Therefore, $P_Q(q \mid C) \propto \frac{C - q + 1}{C}$. Since $q$ has support on $[C]$, then $P_Q(q \mid C) = \left(\frac{2}{(C+1) C}\right) \frac{C - q + 1}{C}$.
\subsection{Message passing graph networks for (semi-)supervised node classification}
\label{sec:prelimMP}
We are also interested in a class of (message passing) graph network models taking the general form:
\begin{equation}
\label{eq:prelim_mp}
\textrm{ for $l=0, 1, \dots L$: \hspace{0.2cm} }\mathbf{H}^{(l+1)} = \sigma_l \left( g ( \mathbf{A} ) \mathbf{H}^{(l)} \mathbf{W}^{(l)} \right); \hspace{0.4cm} \mathbf{H}^{(0)} = \mathbf{X}; \hspace{0.4cm} \mathbf{H} = \mathbf{H}^{(L)};
\end{equation}
where $L$ is the number of layers,
$\mathbf{W}^{(l)}$'s are trainable parameters, $\sigma_l$'s denote element-wise activations (e.g. logistic or ReLu), and $g$ is some (possibly trainable) transformation of adjacency matrix. GCN \citep{kipf} set $g(\mathbf{A}) = \widehat{\mathbf{A}}$,
GAT \citep{gat} set $g(\mathbf{A}) = \mathbf{A} \circ \textrm{MultiHeadedAttention}$ and GIN \citep{GIN} as $g(\mathbf{A}) = \mathbf{A} + (1+\epsilon) \mathbf{I} $ with $\epsilon > 0$.
For node classification, it is common to set $\sigma_L = \textrm{softmax}$ (applied row-wise), specify the size of $\mathbf{W}_L$ s.t. $\mathbf{H} \in \mathbb{R}^{n \times y}$ where $y$ is number of classes, and optimize cross-entropy objective: \newline
\textcolor{white}{.}
\hspace{0.2cm}
$\min_{ \{\mathbf{W}_j\}_{j=1}^L} \left[- \mathbf{Y} \circ \log \mathbf{H} - (1-\mathbf{Y}) \circ \log (1-\mathbf{H})\right],$
\hspace{0.4cm}
where $\mathbf{Y}$ is a binary matrix with one-hot rows indicating node labels.
In semi-supervised settings where not all nodes are labeled, before measuring the objective, subset of rows can be kept in $\mathbf{Y}$ and $\mathbf{H}$ that correspond to labeled nodes.
\subsection{Truncated Singular Value Decomposition (SVD)}
\label{sec:prelimsvd}
SVD is an algorithm that approximates any matrix $\textbf{M} \in \mathbb{R}^{r \times c}$ as a product of three matrices:
\begin{equation*}
\textrm{SVD}_k(\mathbf{M}) \triangleq \argmin_{\mathbf{U}, \mathbf{S}, \mathbf{V}} || \mathbf{M} - \mathbf{U} \mathbf{S} \mathbf{V}^\top ||_\textrm{F} \hspace{0.2cm}\textrm{subject to}\hspace{0.2cm} \mathbf{U}^\top\mathbf{U} = \mathbf{V}^\top \mathbf{V} = \mathbf{I}_k; \hspace{0.1cm} \mathbf{S} = \textrm{diag}(s_1, \dots, s_k).
\end{equation*}
The \textit{orthonormal} matrices $\mathbf{U} \in \mathbb{R}^{r \times k}$ and $\mathbf{V} \in \mathbb{R}^{c \times k} $, respectively, are known as the left- and right-singular bases. The values along diagonal matrix $\mathbf{S} \in \mathbb{R}^{k \times k}$ are known as the \textit{singular values}.
Due to theorem of \citet{eckart1936lowrank}, SVD recovers the best rank-$k$ approximation of input $\mathbf{M}$, as measured by the Frobenius norm $||.||_\textrm{F}$.
Further, if $k\ge \textrm{rank}(\mathbf{M}) \Rightarrow ||.||_\textrm{F} =0$.
Popular SVD implementations follow Random Matrix Theory algorithm of \citet{halko2009svd}.
The prototype algorithm starts with a random matrix and repeatedly multiplies it by $\mathbf{M}$ and by $\mathbf{M}^\top$, interleaving these multiplications with orthonormalization.
Our SVD implementation (in Appendix) also follows the prototype of \citep{halko2009svd}, but with two modifications: (i) we replace the recommended orthonormalization step from QR decomposition to Cholesky decomposition, giving us significant computational speedups and (ii) our implementation accepts
symbolic representation of $\mathbf{M}$ (\S\ref{sec:symbolic}), in lieu of its explicit value (constrast to TensorFlow and PyTorch, requiring explicit $\mathbf{M}$).
In \S\ref{sec:convexification}, we derive linear first-order approximations of models reviewed in \S\ref{sec:prelimNE} \& \S\ref{sec:prelimMP} and explain how SVD can train them. In \S\ref{sec:finetuning}, we show how they can be used as initializations of non-linear models.
\section{Convex first-order approximations of GRL models}
\label{sec:convexification}
\subsection{Convexification of Network Embedding Models}
\label{sec:cvxNE}
We can interpret objective~\ref{eq:wys} as self-supervised learning, since node labels are absent.
Specifically,
given a node $i \in [n]$, the task is to predict its neighborhood as weighted by the row vector $\mathbb{E}_q[\mathcal{T}^q]_i$, representing the subgraph\footnote{$\mathbb{E}_q[\mathcal{T}^q]_i$ is a distribution on $[n]$: entry $j$ equals prob. of walk starting at $i$ ending at $j$ if walk length $\sim \mathcal{U}[C]$.} around $i$.
Another interpretation
is that Eq.~\ref{eq:wys} is a decomposition objective:
multiplying the tall-and-thin matrices, as $\mathbf{L} \mathbf{R}^\top \in \mathbb{R}^{n \times n}$,
should give a larger value at $(\mathbf{L} \mathbf{R}^\top)_{ij} = \mathbf{L}_j^\top \mathbf{R}_i$ when nodes $i$ and $j$ are well-connected but a lower value when $(i, j)$ is not an edge.
We propose a matrix such that its decomposition can incorporate the above interpretations:
\begin{equation}
\label{eq:ourNE}
\widehat{\mathbf{M}}^\textrm{(NE)} = \mathbb{E}_{q|C}[\mathcal{T}^q] - \lambda(1 - \mathbf{A}) = \left(\frac{2}{(C+1) C}\right) \sum_{q=1}^C \left(\frac{C-q+1}{C} \right) \mathcal{T}^q - \lambda ( 1 - \mathbf{A})
\end{equation}
If nodes $i, j$ are nearby, share a lot of connections, and/or in the same community, then entry $\widehat{\mathbf{M}}^\textrm{(NE)}_{ij}$ should be positive.
If they are far apart, then $\widehat{\mathbf{M}}^\textrm{(NE)}_{ij} = -\lambda$. To embed the nodes onto a low-rank space that approximates this information, one can
decompose $\widehat{\mathbf{M}}^\textrm{(NE)}$ into two thin matrices $(\mathbf{L}, \mathbf{R})$:
\begin{equation}
\label{eq:ourNEapprox}
\mathbf{L} \mathbf{R}^\top \approx \widehat{\mathbf{M}}^\textrm{(NE)} \Longleftrightarrow (\mathbf{L} \mathbf{R}^\top)_{i, j} = \langle \mathbf{L}_i, \mathbf{R}_j \rangle \approx \widehat{\mathbf{M}}^\textrm{(NE)}_{ij} \textrm{ \ \ \ for all \ } i, j \in [n].
\end{equation}
SVD gives low-rank approximations that minimize the Frobenius norm of error (\S\ref{sec:prelimsvd}).
The remaining challenge is computational burden: the right term ($1 - \mathbf{A}$), \textit{a.k.a}, graph compliment, has $\approx n^2$ non-zero entries and
the left term has non-zero at entry $(i, j)$ if nodes $i, j$ are within distance $C$ away, as $q$ has support on $[C]$ -- for reference
Facebook network has an average distance of 4 \citep{backstrom2012fourdegrees} i.e. yielding $\mathcal{T}^4$ with $\mathcal{O}(n^2)$ nonzero entries --
Nonetheless, Section \S\ref{sec:symbolic} presents a framework for decomposing $\widehat{\mathbf{M}}$ from its symbolic representation, without explicitly computing its entries. Before moving forward, we note that one can replace $\mathcal{T}$ in Eq.~\ref{eq:ourNE} by its symmetrically normalized counterpart $\widehat{\mathbf{A}}$, recovering a basis where $\mathbf{L} = \mathbf{R}$. This symmetric modeling might be emperically preferred for undirected graphs.
\underline{\textbf{Learning}} can be performed via SVD.
Specifically, the node at the $i^\textrm{th}$ row and the node at the $j^\textrm{th}$th column will be embedded, respectively, in $\mathbf{L}_i$ and $\mathbf{R}_j$ computed as:
\begin{equation}
\label{eq:svd_train_NE}
\mathbf{U}, \mathbf{S}, {\mathbf{V}} \leftarrow \textrm{SVD}_k(\widehat{\mathbf{M}}^\textrm{(NE)}); \hspace{0.6cm}
{\mathbf{L}} \leftarrow {\mathbf{U}} {\mathbf{S}}^{\frac12};
\hspace{0.6cm}
{\mathbf{R}} \leftarrow {\mathbf{V}} {\mathbf{S}}^{\frac{1}{2}}
\end{equation}
In this $k$-dim
space of rows and columns, Euclidean measures are plausible: \underline{\textbf{Inference}} of nodes' similarity at row $i$ and column $j$ can be modeled as $f(i, j) = \langle {\mathbf{L}}_i, {\mathbf{R}}_j \rangle = \mathbf{U}_i^\top \mathbf{S} \mathbf{V}_j \triangleq \langle {\mathbf{U}}_i, {\mathbf{V}}_j \rangle_{_{\mathbf{S}}} $.
\subsection{Convexification of message passing graph networks}
\label{sec:cvx:MP}
Removing all $\sigma_l$'s from Eq.~\ref{eq:prelim_mp} and setting $g(\mathbf{A}) = \widehat{\mathbf{A}}$ gives outputs of layers 1, 2, and $L$, respectively,
\begin{equation}
\textrm{as: \hspace{0.4cm}} \widehat{\mathbf{A}} \mathbf{X} \mathbf{W}^{(1)} \textrm{\hspace{0.5cm} , \hspace{0.5cm} } \widehat{\mathbf{A}}^2 \mathbf{X} \mathbf{W}^{(1)} \mathbf{W}^{(2)} \textrm{\hspace{0.5cm} ,\hspace{0.2cm} and \hspace{0.5cm} } \widehat{\mathbf{A}}^L \mathbf{X} \mathbf{W}^{(1)} \mathbf{W}^{(2)} \dots \mathbf{W}^{(L)}.
\end{equation}
Without non-linearities, adjacent parameter matrices can be absorbed into one another.
Further, the model output can concatenate all layers, like JKNets \citep{jknet}, giving final model output of:
\begin{equation}
\label{eq:h_jkn}
\mathbf{H}^\textrm{(NC)}_\textrm{linearized} =
\left[
\begin{matrix}[c;{1pt/1pt}c;{1pt/1pt}c;{1pt/1pt}c;{1pt/1pt}c]
\mathbf{X} & \widehat{\mathbf{A}} \mathbf{X} & \widehat{\mathbf{A}}^2 \mathbf{X} & \dots & \widehat{\mathbf{A}}^L \mathbf{X} \end{matrix} \right]
\widehat{\mathbf{W}}
\hspace{0.5cm}
\triangleq
\hspace{0.2cm}
\widehat{\mathbf{M}}^{\textrm{(NC)}} \widehat{\mathbf{W}} ,
\end{equation}
where the linearized model implicitly constructs design matrix $\widehat{\mathbf{M}}^{\textrm{(NC)}} \in \mathbb{R}^{n \times F}$ and multiplies it with parameter $\widehat{\mathbf{W}} \in \mathbb{R}^{F \times y}$ -- here, $F = d+dL$.
Crafting design matrices is a creative process (\S\ref{sec:creative}).
\underline{\textbf{Learning}} can be performed by minimizing the Frobenius norm: $||\mathbf{H}^{\textrm{(NC)}} - \mathbf{Y}||_\textrm{F} = || \widehat{\mathbf{M}}^\textrm{(NC)} \widehat{\mathbf{W}} - \mathbf{Y}||_\textrm{F}$.
Moore-Penrose Inverse (a.k.a, the psuedoinverse) provides one such minimizer:
\begin{equation}
\label{eq:wstar}
\widehat{\mathbf{W}}^{*} = \textrm{argmin}_{\widehat{\mathbf{W}}} \Big|\Big| \widehat{\mathbf{M}} \widehat{\mathbf{W}} - \mathbf{Y} \Big|\Big|_\textrm{F} =
\widehat{\mathbf{M}}^\dagger \mathbf{Y}
\approx \mathbf{V} \mathbf{S}^{+} \mathbf{U}^\top \mathbf{Y},
\end{equation}
with $\mathbf{U}, \mathbf{S}, \mathbf{V} \leftarrow \textrm{SVD}_k(\widehat{\mathbf{M}})$.
Notation $\mathbf{S}^{+}$ reciprocates non-zero entries of diagonal $\mathbf{S}$ \citep{golub1996matrix}.
Multiplications in the right-most term should, for efficiency, be executed right-to-left.
The pseudoinverse $\widehat{\mathbf{M}}^\dagger \approx \mathbf{V} \mathbf{S}^{+} \mathbf{U}^\top$ recovers the $\widehat{\mathbf{W}}^{*}$ with least norm (\S\ref{sec:analysis}, Theorem~\ref{theorem:min_norm}).
The $\approx$ becomes $=$ when $k \ge \textrm{rank}(\widehat{\mathbf{M}})$.
In semi-supervised settings, one can take rows subset
of either (i) $\mathbf{Y}$ and $\mathbf{U}$, or of (ii) $\mathbf{Y}$ and $\mathbf{M}$, keeping only rows that correspond to labeled nodes. Option (i) is supported by existing frameworks (e.g., \texttt{tf.gather()}) and our symbolic framework (\S\ref{sec:symbolic}) supports (ii) by \textit{implicit row (or column) gather} -- i.e., calculating SVD of submatrix of $\mathbf{M}$ without explicitly computing $\mathbf{M}$ nor the submatrix.
\underline{\textbf{Inference}} over a (possibly new) graph $(\mathbf{A}, \mathbf{X})$ can be calculated by (i) (implicitly) creating the design matrix $\widehat{\mathbf{M}}$ corresponding to $(\mathbf{A}, \mathbf{X})$ then (ii) multiplying by the explicitly calculated $\widehat{\mathbf{W}}^{*}$. As explained in \S\ref{sec:symbolic}, $\widehat{\mathbf{M}}$ need not to be explicitly calculated for computing multiplications.
\section{Symbolic matrix representation}
\label{sec:symbolic}
To compute the SVD of any matrix $\mathbf{M}$ using algorithm prototypes presented by \citet{halko2009svd}, \textbf{it suffices to provide functions that can multiply arbitrary vectors} with $\mathbf{M}$ and $\mathbf{M}^\top$, and \textbf{explicit calculation} of $\mathbf{M}$ is \textbf{not required}.
Our software framework can symbolically represent $\mathbf{M}$
as a directed acyclic graph (DAG) of computations.
On this DAG, each
node can be one of two kinds:
\begin{enumerate}[topsep=0pt, itemsep=0pt]
\item \textbf{Leaf node} (no incoming edges) that \textbf{explicitly} holds a matrix.
Multiplications against leaf nodes are directly executed via an underlying math framework (we utilize TensorFlow).
\item \textbf{Symbolic node} that only \textbf{implicitly}
represents a matrix
as as a function of other DAG nodes.
Multiplications are recursively computed,
traversing incoming edges, until leaf nodes.
\end{enumerate}
For instance, suppose leaf DAG nodes $\mathbf{M}_1$ and $\mathbf{M}_2$, respectively, explicitly contain row vector $\in \mathbb{R}^{1 \times n}$ and column vector $\in \mathbb{R}^{n \times 1}$. Then, their (symbolic) product DAG node $\mathbf{M} = \mathbf{M}_2 \texttt{@} \mathbf{M}_1$ is $\in\mathbb{R}^{n \times n}$.
Although storing $\mathbf{M}$ explicitly requires $\mathcal{O}(n^2)$ space,
multiplications against $\mathbf{M}$
can remain within $\mathcal{O}(n)$ space
if efficiently implemented as
$\left\langle \mathbf{M}, . \right\rangle = \left\langle \mathbf{M}_2, \left\langle \mathbf{M}_1, . \right\rangle \right\rangle$.
Figure \ref{fig:symbolic} shows code snippet for composing
DAG to represent symbolic node $\widehat{\mathbf{M}}^\textrm{(NE)}$ (Eq.~\ref{eq:ourNE}), from leaf nodes initialized with in-memory matrices.
Appendex lists symbolic nodes and their implementations.
\begin{figure}[t]
\begin{minipage}{0.465\textwidth}
\begin{lstlisting}[columns=fullflexible,language=Python]
a = scipy.sparse.csr_matrix(...)
d = scipy.sparse.diags(a.sum(axis=1))
t = (1/d).dot(a)
t, a = F.leaf(t), F.leaf(a)
row1 = F.leaf(tf.ones([1, a.shape[0]]))
q1, q2, q3 = np.array([3, 2, 1]) / 6.0
M = q1 * t + q2 * t@t + q3 * t@t@t
M -= lamda * (row1.T @ row1 - A)
\end{lstlisting}
\end{minipage}
~
\begin{minipage}{0.53\textwidth}
\includegraphics[height=2.8cm]{figs/uncompiled_dag.pdf}
~
\includegraphics[height=2.8cm]{figs/compiled_dag.pdf}
\end{minipage}
\caption{Symbolic Matrix Representation. \textbf{Left}: code using our framework to implicitly construct the design matrix $\mathbf{M}=\widehat{\mathbf{M}}^\textrm{(NE)}$ with our framework. \textbf{Center}: DAG corresponding to the code. \textbf{Right}: An equivalent automatically-optimized DAG (via lazy-cache, Fig.~\ref{fig:time}) requiring fewer floating point operations. The first 3 lines of code create explicit input matrices (that fit in memory): \underline{\textbf{a}}djacency $\mathbf{A}$, diagonal \underline{\textbf{d}}egree $\mathbf{D}$, and \underline{\textbf{t}}ransition $\mathcal{T}$. Matrices are imported into our framework with $\texttt{F.leaf}$ (depicted on computation DAGs in blue). Our classes overloads standard methods ($\texttt{+}$, $\texttt{-}$, $\texttt{*}$, $\texttt{@}$, $\texttt{**}$) to construct computation nodes (intermediate in grey). The output node (in red) needs not be exactly calculated yet can be efficiently multiplied by any matrix by recursive downward traversal.}
\label{fig:symbolic}
\end{figure}
\section{SVD initialization for deeper models fine-tuned via cross-entropy}
\label{sec:finetuning}
\subsection{Edge function for network embedding as a (1-dimensional) Gaussian kernel}
\label{sec:finetuneNE}
SVD provides decent solutions to link prediction tasks.
Computing $\mathbf{U}, \mathbf{S}, \mathbf{V} \leftarrow \textrm{SVD}(\mathbf{M}^\textrm{(NE)})$ is much faster
than training SOTA models for link prediction,
yet, simple edge-scoring function
$f(i,j) = \langle {\mathbf{U}}_i, {\mathbf{V}}_j \rangle_{_{\mathbf{S}}}$
yields competitive empirical (test) performance. We propose $f$ with $\theta = \{\mu, s\}$:
\begin{align}
\label{eq:gaussiankernel}
f_{\mu, s}(i, j) = {\mathbb{E}}_{x \sim \overline{\mathcal{N}}(\mu, s) } \langle\mathbf{U}_i, \mathbf{V}_j \rangle_{_{\mathbf{S}^x}}
& = \mathbf{U}_i^\top \mathbb{E}_x\left[\mathbf{S}^x \right] \mathbf{V}_j
= \mathbf{U}_i^\top \left(\int_\Omega \mathbf{S}^x \overline{\mathcal{N}}(x \mid \mu, s ) \,dx \right) \mathbf{V}_j,
\end{align}
where $\overline{\mathcal{N}}$ is the truncated normal distribution (we truncate to $\Omega = [0.5, 2]$). The integral can be
approximated by discretization and applying softmax (see \S A.4).
The parameters $\mu \in \mathbb{R}, \sigma \in \mathbb{R}_{>0}$ can be optimized on cross-entropy objective for link-prediction:
\begin{equation}
\label{eq:finetuneNE}
\textrm{min}_{\mu, s} -{\mathbb{E}}_{(i, j) \in \mathbf{A}} \left[ \log \left( \sigma(f_{\mu, s}(i, j)) \right) \right]
- k_\textrm{(n)} {\mathbb{E}}_{(i, j) \notin \mathbf{A}} \left[ \log \left(1 - \sigma(f_{\mu, s}(i, j)) \right) \right],
\end{equation}
where the left- and right-terms, respectively, encourage $f$ to score high for edges, and the low for non-edges.
$k_\textrm{(n)} \in \mathbb{N}_{>0}$ controls the ratio of negatives to positives per batch (we use $k_\textrm{(n)} = 10$).
If the optimization sets $\mu = 1$ and $s\approx 0$, then $f$ reduces to no-op. In fact,
we initialize it as such,
and we observe that
$f$ converges \textbf{within one epoch}, on graphs we experimented on.
If it converges as $\mu < 1$, VS $\mu > 1$, respectively, then $f$ would effectively squash, VS enlarge, the spectral gap.
\subsection{Split-ReLu (deep) graph network for node classification (NC)}
\label{sec:finetuneNC}
$\widehat{\mathbf{W}
}^{*}$ from SVD (Eq.\ref{eq:wstar}) can initialize an $L$-layer graph network with input: $\mathbf{H}^{(0)} = \mathbf{X}$, with:
\begin{align}
\textrm{message passing (MP) }\hspace{1cm}&\mathbf{H}^{(l+1)} = \left[\widehat{\mathbf{A}} \mathbf{H}^{(l)} \mathbf{W}_\textrm{(p)}^{(l)} \right]_+ - \left[\widehat{\mathbf{A}} \mathbf{H}^{(l)} \mathbf{W}_\textrm{(n)}^{(l)} \right]_+, \label{eq:splitrelu_mp} \\
\textrm{output }\hspace{1cm}&
\mathbf{H}=\sum_{l=0}^{l=L} \left[\mathbf{H}^{(l)} \mathbf{W}_\textrm{(op)}^{(l)} \right]_+ - \left[\mathbf{H}^{(l)} \mathbf{W}_\textrm{(on)}^{(l)} \right]_+ \label{eq:splitrelu_output} \\
\textrm{initialize MP }\hspace{1cm}& \mathbf{W}_\textrm{(p)}^{(l)} \leftarrow \mathbf{I}; \mathbf{W}_\textrm{(n)}^{(l)} \leftarrow -\mathbf{I}; \label{eq:splitrelu_mp_init} \\
\textrm{initialize output }\hspace{1cm}&
\mathbf{W}_\textrm{(op)}^{(l)} \leftarrow \widehat{\mathbf{W}}^*_{[dl \, : \, d(l+1)]};
\mathbf{W}_\textrm{(on)}^{(l)} \leftarrow - \widehat{\mathbf{W}}^*_{[dl \, : \, d(l+1)]};
\label{eq:splitrelu_output_init}
\end{align}
Element-wise $[.]_+=\max(0, .)$.
Further, $\mathbf{W}_{[i\, :\, j]}$ denotes rows from $(i)^\textrm{th}$ until $(j$$-$$1)^\textrm{th}$ of $\mathbf{W}$.
The deep network layers (Eq. \ref{eq:splitrelu_mp}\&\ref{eq:splitrelu_output}) use our Split-ReLu layer which we formalize as:
\begin{equation}
\textrm{SplitReLu}(\mathbf{X}; \mathbf{W}_\textrm{(p)}, \mathbf{W}_\textrm{(n)}) = \left[\mathbf{X} \mathbf{W}_\textrm{(p)} \right]_+ - \left[\mathbf{X} \mathbf{W}_\textrm{(n)} \right]_+,
\end{equation}
where the subtraction is calculated entry-wise. The layer has twice as many parameters as standard fully-connected (FC) ReLu layer. In fact, learning algorithms can recover FC ReLu from SplitReLu by assigning $\mathbf{W}_\textrm{(n)} = 0$. More importantly, the layer behaves as linear in $\mathbf{X}$ when $\mathbf{W}_\textrm{(p)} = - \mathbf{W}_\textrm{(n)}$. On this hyperplane, this linear behavior allows us to establish the equivalency:
the (non-linear) model $\mathbf{H}$ is equivalent to the linear $\mathbf{H}_\textrm{linearized}$ at initialization (Eq.~\ref{eq:splitrelu_mp_init}\&\ref{eq:splitrelu_output_init}) due to Theorem \ref{theorem:initialization}.
Following the initialization, model can be fine-tuned on cross-entropy objective as in \S\ref{sec:prelimMP}.
\subsection{Creative Add-ons for node classification (NC) models}
\label{sec:creative}
\textbf{Label re-use} (LR):
Let
$\widehat{\mathbf{M}}^\textrm{(NC)}_\textrm{LR} \triangleq \left[
\begin{matrix}[c;{1pt/1pt}c;{1pt/1pt}c]
\widehat{\mathbf{M}}^{^\textrm{(NC)}} &
(\widehat{\mathbf{A}} - (\mathbf{D} + \mathbf{I})^{{-1}}) \mathbf{Y}_{_\textrm{[train]}} &
(\widehat{\mathbf{A}} - (\mathbf{D} + \mathbf{I})^{{-1}})^{2} \mathbf{Y}_{_\textrm{[train]}} \end{matrix} \right]$.\newline
This follows the motivation of \citet{wang2020unifying, huang2021combining, wang2021bag} and their empirical results on ogbn-arxiv dataset,
where $\mathbf{Y}_\textrm{[train]} \in \mathbb{R}^{n \times y}$ contains one-hot vectors at rows corresponding to labeled nodes but contain zero vectors for unlabeled (test) nodes.
Our scheme is similar to concatenating $\mathbf{Y}_\textrm{[train]}$ into $\mathbf{X}$, but with care to prevent label leakage from row $i$ of $\mathbf{Y}$ to row $i$ of $\widehat{\mathbf{M}}$, as we zero-out the diagonal of the adjacency multiplied by $\mathbf{Y}_\textrm{[train]}$.
\textbf{Pseudo-Dropout} (PD):
Dropout \citep{srivastava14dropout}
reduces overfitting of models.
It can be related to \textit{data augmentation},
as each example is presented multiple times.
At each time, it appears with a different set of \textit{dropped-out features} -- input or latent feature values, chosen at random, get replaced with zeros.
As such, we can replicate the design matrix as:
${{\widehat{\mathbf{M}}}}^\top \leftarrow
\left[\begin{matrix}[c;{1pt/1pt}c]
\widehat{\mathbf{M}}^\top & \textrm{PD}(\widehat{\mathbf{M}})^\top
\end{matrix}\right]$.
This row-wise concatenation
maintains the width of $\widehat{\mathbf{M}}$
and therefore the number of model parameters.
In the above add-ons, concatenations, as well as PD, can be implicit or explicit (see \S A.3).
\section{Analysis \& Discussion}
\label{sec:analysis}
\begin{theorem}
\label{theorem:min_norm}
{\normalfont (Min. Norm)}
If system $\widehat{\mathbf{M}} \widehat{\mathbf{W}} = \mathbf{Y}$ is underdetermined\footnote{E.g., if the number of labeled examples i.e. height of $\mathbf{M}$ and $\mathbf{Y}$ is smaller than the width of $\mathbf{M}$.} with rows of $\widehat{\mathbf{M}}$ being linearly independent,
then
solution space $\widehat{\mathcal{W}}^* = \Big\{\widehat{\mathbf{W}} \ \Big| \ \widehat{\mathbf{M}} \widehat{\mathbf{W}} = \mathbf{Y} \Big\}$ has
infinitely many solutions.
Then, for $k\geq \textrm{rank}(\widehat{\mathbf{M}})$, matrix $\widehat{\mathbf{W}}^*$, recovered by Eq.\ref{eq:wstar} satisfies:
$
\widehat{\mathbf{W}}^* = \mathop{\mathrm{argmin}}_{\widehat{\mathbf{W}}\in \widehat{\mathcal{W}}^* } ||\widehat{\mathbf{W}}||_F^2
$.
\end{theorem}
Theorem \ref{theorem:min_norm} implies that, even though one can design a wide $\widehat{\mathbf{M}}^\textrm{(NC)}$ (Eq.\ref{eq:h_jkn}), \textit{i.e.}, with many layers, the recovered parameters with least norm should be less prone to overfitting. Recall that this is the goal of L2 regularization. Analysis and proofs are in the Appendix.
\begin{theorem}
\label{theorem:initialization}
{\normalfont (Non-linear init)} The initialization Eq.~\ref{eq:splitrelu_mp_init}\&\ref{eq:splitrelu_output_init} yields
$\mathbf{H}^\textnormal{(NC)}_\textnormal{linearized} = \mathbf{H} \big|_{\theta \leftarrow \textnormal{via Eq. \ref{eq:splitrelu_mp_init}\&\ref{eq:splitrelu_output_init} }} $.
\end{theorem}
Theorem \ref{theorem:initialization} implies that the deep (nonlinear) model is the same as the linear model, at the initialization of $\theta$ (per Eq.~\ref{eq:splitrelu_mp_init}\&\ref{eq:splitrelu_output_init}, using $\mathbf{\widehat{W}}^*$ as Eq.~\ref{eq:wstar}). Cross-entropy objective can then fine-tune $\theta$.
This end-to-end process, of (i) computing SVD bases and (ii) training the network $f_\theta$ on singular values, \textit{advances} SOTA on competitve benchmarks,
with
(i) converging (quickly) to a unique solution and (ii) containing merely a few parameters $\theta$ -- see \S\ref{sec:applications}.
\section{Applications \& Experiments}
\label{sec:applications}
\label{sec:experiments}
We download and experiment on 9 datasets summarized in
Table \ref{table:datasets}.
We attempt link prediction (LP) tasks on smaller graph datasets (< 1 million edges) of:
Protein-Protein Interactions (PPI) graph from \citet{grover2016node2vec};
as well as ego-Facebook (FB), AstroPh, HepTh from Stanford SNAP \citep{snapnets}.
For these datasets, we use the train-test splits of \citet{wys}.
We also attempt semi-supervised node classification (SSC) tasks on smaller graphs of Cora, Citeseer, Pubmed, all obtained from Planetoid \citep{planetoid}.
For these smaller datasets, we only train and test using the SVD basis (without finetuning)
Further, we attempt on slightly-larger datasets (> 1 million edges) from Stanford's Open Graph Benchmark \citep[OGB,][]{ogb}. We use the official train-test-validation splits and evaluator of OGB. We attempt LP and SSC, respectively, on
Drug Drug Interactions (ogbl-DDI) and
ArXiv citation network (ogbn-ArXiv). For these larger datasets, we use the SVD basis as an initialization that we finetune, as described in \S\ref{sec:finetuning}. For time comparisons, we train all models on Tesla K80.
\begin{table}[t]
\caption{Dataset Statistics}
\label{table:datasets}
\centering{
\begin{tabular}{r rl rl l c c c}
\toprule
\textbf{Dataset} & \multicolumn{2}{c}{\textbf{Nodes}} & \multicolumn{2}{c}{\textbf{Edges}} & \multicolumn{1}{c}{\textbf{Source}} & \multicolumn{1}{c}{\textbf{Task}} & \multicolumn{1}{c}{$\mathbf{X}$} \\
\cmidrule(lr){1-1}\cmidrule(lr){2-3}\cmidrule(lr){4-5}\cmidrule(lr){6-6}\cmidrule(lr){7-7}\cmidrule(lr){8-8}
PPI & 3,852 &\hspace{-0.3cm}proteins & 20,881 &\hspace{-0.3cm}chem. interactions & \href{https://snap.stanford.edu/node2vec/#datasets}{node2vec} & LP & \xmark \\
FB & 4,039 &\hspace{-0.3cm}users & 88,234 &\hspace{-0.3cm}friendships & SNAP & LP & \xmark \\
AstroPh & 17,903 &\hspace{-0.3cm}researchers & 197,031 &\hspace{-0.3cm}co-authorships & SNAP & LP & \xmark \\
HepTh & 8,638 &\hspace{-0.3cm}researchers & 24,827 &\hspace{-0.3cm}co-authorships & SNAP &LP & \xmark \\
Cora & 2,708 &\hspace{-0.3cm}articles & 5,429 &\hspace{-0.3cm}citations & Planetoid & SSC & \cmark\\
Citeseer & 3,327 &\hspace{-0.3cm}articles &\hspace{-0.3cm}4,732 &\hspace{-0.3cm}citations & Planetoid & SSC & \cmark \\
Pubmed & 19,717 &\hspace{-0.3cm}articles &\hspace{-0.3cm}44,338 &\hspace{-0.3cm}citations & Planetoid & SSC & \cmark\\
ogbn-ArXiv & 169,343&\hspace{-0.3cm}papers &1,166,243&\hspace{-0.3cm}citations & OGB & SSC & \cmark\\
ogbl-DDI & 4,267&\hspace{-0.3cm}drugs &1,334,889 &\hspace{-0.3cm}interactions & OGB & LP & \cmark\\
\bottomrule
\end{tabular}
}
\vspace{-0.4cm}
\end{table}
\begin{table}[t]
\caption{Test accuracy (\& train time) on citation graphs for task: \textit{semi-supervised node classification}.}
\label{table:resultsPlanetoid}
\centering{
\begin{tabular}{r lr lr lr lr}
\toprule
\multicolumn{1}{r}{\textbf{Graph dataset:}} \hspace{-1cm} & \multicolumn{2}{c}{\textbf{Cora}}
& \multicolumn{2}{c}{\textbf{Citeseer}}
& \multicolumn{2}{c}{\textbf{Pubmed}} \\
\cmidrule(lr){2-3}
\cmidrule(lr){4-5}
\cmidrule(lr){6-7}
\multicolumn{1}{l}{\hspace{-0.1cm}\textbf{Baselines:}} & acc & tr.time & acc & tr.time & acc & tr.time \\
\cmidrule(lr){1-1}
\cmidrule(lr){2-2}
\cmidrule(lr){3-3}
\cmidrule(lr){4-3}
\cmidrule(lr){5-4}
\cmidrule(lr){6-6}
\cmidrule(lr){7-7}
Planetoid & 75.7 & (13s) & 64.7 & (26s) & 77.2 & (25s) \\
GCN & 81.5 & (4s) & 70.3 & (7s) & 79.0 & (83s) \\
GAT & 83.2 & (1m) & 72.4 & (3m) & 77.7 & (6m) \\
MixHop & 81.9 & (26s) & 71.4 & (31s) & 80.8 & (1m) \\
GCNII & 85.5 & (2m) & 73.4 & (3m) & 80.3 & (2m) \\
\cmidrule(lr){1-1}
\cmidrule(lr){2-2}
\cmidrule(lr){3-3}
\cmidrule(lr){4-3}
\cmidrule(lr){5-4}
\cmidrule(lr){6-6}
\cmidrule(lr){7-7}
\multicolumn{1}{l}{\hspace{-0.1cm}\textbf{Our models:}} & acc & tr.time & acc & tr.time & acc & tr.time \\
\cmidrule(lr){1-1}
\cmidrule(lr){2-2}
\cmidrule(lr){3-3}
\cmidrule(lr){4-3}
\cmidrule(lr){5-4}
\cmidrule(lr){6-6}
\cmidrule(lr){7-7}
$\textrm{iSVD}_{100}({\widehat{\mathbf{M}}^\textrm{(NC)}})$
& 82.0 {\footnotesize $\pm$0.13} \hspace{-0.45cm} & (0.1s)
& 71.4 {\footnotesize $\pm$0.22} \hspace{-0.45cm} & (0.1s)
& 78.9 {\footnotesize $\pm$0.31} \hspace{-0.45cm} & (0.3s) \\
+ dropout (\S\ref{sec:creative})
& 82.5 {\footnotesize $\pm$0.46} \hspace{-0.45cm} & (0.1s)
& 71.5 {\footnotesize $\pm$0.53} \hspace{-0.45cm} & (0.1s)
& 78.9 {\footnotesize $\pm$0.59} \hspace{-0.45cm} & (0.2s) \\
\bottomrule
\end{tabular}{}
}
\vspace{-0.4cm}
\end{table}
\begin{table}[t]
\caption{Test ROC-AUC (\& train time) on Stanford SNAP graphs for task: \textit{link prediction}.}
\label{table:snapAUC}
\centering
{
\begin{tabular}{r lr lr lr lr}
\toprule
\multicolumn{1}{r}{\textbf{Graph dataset:}} \hspace{-1cm} & \multicolumn{2}{c}{\textbf{FB}}
& \multicolumn{2}{c}{\textbf{AstroPh}}
& \multicolumn{2}{c}{\textbf{HepTh}}
& \multicolumn{2}{c}{\textbf{PPI}}
\\
\cmidrule(lr){2-3}
\cmidrule(lr){4-5}
\cmidrule(lr){6-7}
\cmidrule(lr){8-9}
\multicolumn{1}{l}{\hspace{-0.1cm}\textbf{Baselines:}} & AUC & \hspace{-0.2cm} tr.time & AUC &\hspace{-0.2cm} tr.time & AUC & \hspace{-0.2cm} tr.time & AUC & \hspace{-0.2cm} tr.time \\
\cmidrule(lr){1-1}
\cmidrule(lr){2-2}
\cmidrule(lr){3-3}
\cmidrule(lr){4-3}
\cmidrule(lr){5-4}
\cmidrule(lr){6-6}
\cmidrule(lr){7-7}
\cmidrule(lr){8-8}
\cmidrule(lr){9-9}
WYS & 99.4 & (54s) & 97.9 & (32m) & 93.6 & (4m) & 89.8 & (46s) \\
n2v & 99.0 & (30s) & 97.8 & (2m) & 92.3 & (55s) & 83.1 & (27s) \\
NetMF & 97.6 & (5s) & 96.8 & (9m) & 90.5 & (72s) & 73.6 & (7s) \\
$\textrm{Net}\widetilde{\textrm{MF}}$ & 97.0 & (4s) & 81.9 & (4m) & 85.0 & (48s) & 63.6 & (10s) \\
\cmidrule(lr){2-3}
\cmidrule(lr){4-5}
\cmidrule(lr){6-7}
\cmidrule(lr){8-9}
\multicolumn{1}{l}{\hspace{-0.1cm}\textbf{Our models:}} & AUC & \hspace{-0.2cm} tr.time & AUC &\hspace{-0.2cm} tr.time & AUC & \hspace{-0.2cm} tr.time & AUC & \hspace{-0.2cm} tr.time \\
\cmidrule(lr){1-1}
\cmidrule(lr){2-2}
\cmidrule(lr){3-3}
\cmidrule(lr){4-3}
\cmidrule(lr){5-4}
\cmidrule(lr){6-6}
\cmidrule(lr){7-7}
\cmidrule(lr){8-8}
\cmidrule(lr){9-9}
\hspace{-0.2cm} $\textrm{iSVD}_{32}({\widehat{\mathbf{M}}^\textrm{(NE)}})$
& 99.1 {\footnotesize $\pm$1e-6}\hspace{-0.4cm} & (0.2s)
& 94.4 {\footnotesize $\pm$4e-4}\hspace{-0.4cm} & (0.5s)
& 90.5 {\footnotesize $\pm$0.1}\hspace{-0.4cm} & (0.1s)
& 89.3 {\footnotesize $\pm$0.01}\hspace{-0.4cm} & (0.1s) \hspace{-0.2cm} \\
\hspace{-0.2cm} $\textrm{iSVD}_{256}({\widehat{\mathbf{M}}^\textrm{(NE)}})$
& 99.3 {\footnotesize $\pm$9e-6}\hspace{-0.4cm} & (2s)
& 98.0 {\footnotesize $\pm$0.01}\hspace{-0.4cm} & (7s)
& 90.1 {\footnotesize $\pm$0.54}\hspace{-0.4cm} & (2s)
& 89.3 {\footnotesize $\pm$0.48}\hspace{-0.4cm} & (1s) \hspace{-0.2cm} \\
\bottomrule
\end{tabular}{}
}
\vspace{-0.4cm}
\end{table}
\begin{table}[t]
\vspace{-0.2cm}
\caption{Test Hits@20 for link prediction over Drug-Drug Interactions Network (ogbl-ddi).}
\label{table:ogb1-ddi}
\centering{
\begin{tabular}{r l l r }
\toprule
\multicolumn{4}{c}{\textbf{Graph dataset:} \hspace{1cm} \textbf{ogbl-DDI}} \\
\cmidrule(lr){1-4}
\multicolumn{2}{l}{ \textbf{Baselines:} } & \multicolumn{1}{c}{{HITS@20}} & \multicolumn{1}{c}{{tr.time}} \\
\cmidrule(lr){1-2}
\cmidrule(lr){3-3}
\cmidrule(lr){4-4}
DEA+JKNet & \hspace{-0.1cm}\citep{yang2021globallocal}\hspace{-0.5cm} & \hspace{-0cm}76.72 {\small$\pm$2.65} & (60m) \\
LRGA+n2v & \hspace{-0.1cm}\citep{hsu2021cs224w}\hspace{-0.5cm}& \hspace{-0cm}73.85 {\small$\pm$8.71} & (41m) \\
MAD & \hspace{-0.1cm}\citep{luo2021memoryassociated}\hspace{-0.5cm} &\hspace{-0cm}67.81 {\small$\pm$2.94} & (2.6h) \\
LRGA+GCN & \hspace{-0.1cm}\citep{puny2020global}\hspace{-0.3cm}& \hspace{-0cm}62.30 {\small$\pm$9.12} & (10m) \\
GCN+JKNet & \hspace{-0.1cm}\citep{jknet}\hspace{-0.3cm} & \hspace{-0cm}60.56 {\small$\pm$8.69} & (21m) \\
\cmidrule(lr){1-2}
\cmidrule(lr){3-3}
\cmidrule(lr){4-4}
\multicolumn{2}{l }{ \textbf{Our models:} } & \multicolumn{1}{c}{{HITS@20}} & \multicolumn{1}{c}{{tr.time}} \\
\cmidrule(lr){1-2}
\cmidrule(lr){3-3}
\cmidrule(lr){4-4}
\multicolumn{1}{l}{(a) $\textrm{iSVD}_{100}(\widehat{\mathbf{M}}^\textrm{(NE)})$} & \multicolumn{1}{r}{(\S\ref{sec:cvxNE}, Eq.~\ref{eq:svd_train_NE})} & 67.86 {\small$\pm$0.09} & (6s) \\
\multicolumn{1}{l}{(b)\hspace{0.1cm} + finetune $f_{\mu, s}(\mathbf{S})$ } & \multicolumn{1}{r}{(\S\ref{sec:finetuneNE}, Eq.~\ref{eq:gaussiankernel} \& \ref{eq:finetuneNE}; sets $\mu = 1.15$)} & 79.09 {\small$\pm$0.18} & (24s) \\
\multicolumn{1}{l}{(c)\hspace{0.1cm} + update $\mathbf{U}$,$\mathbf{S}$,$\mathbf{V}$ } & (on validation, keeps $f_{\mu, s}$ fixed) & 84.09 {\small$\pm$0.03} & (30s) \\
\bottomrule
\end{tabular}{}
}
\vspace{-0.5cm}
\end{table}
\subsection{Test performance \& runtime on smaller datasets from: Planetoid \& Stanford SNAP}
For SSC over Planetoid's datasets, both $\mathbf{A}$ and $\mathbf{X}$ are given.
Additionally, only a handful of nodes are labeled. The goal is to classify the unlabeled test nodes. Table \ref{table:resultsPlanetoid} summarizes the results.
For \textbf{baselines}, we download code of GAT \citep{gat}, MixHop \citep{mixhop}, GCNII \citep{GCNII} and re-ran them with instrumentation to record training time.
However, for baselines Planetoid \citep{planetoid} and GCN \citep{kipf}, we copied numbers from \citep{kipf}.
For \textbf{our models}, the row labeled $\textrm{iSVD}_{100}(\widehat{\mathbf{M}}^\textrm{(NC)})$, we run our implicit SVD twice per graph.
The first run incorporates structural information:
we (implicitly) construct $\widehat{\mathbf{M}}^\textrm{(NE)}$ with $\lambda=0.05$ and $C=3$, then obtain $\mathbf{L}, \mathbf{R} \leftarrow \textrm{SVD}_{64}(\widehat{\mathbf{M}}^\textrm{(NE)})$, per Eq.~\ref{eq:svd_train_NE}. Then, we concatenate
$\mathbf{L}$ and $\mathbf{R}$ into $\mathbf{X}$. Then, we PCA the resulting matrix to 1000 dimensions, which forms our new $\mathbf{X}$.
The second SVD run is to train the classification model parameters $\widehat{\mathbf{W}}^{*}$.
From the PCA-ed $\mathbf{X}$,
we construct the implicit matrix $\widehat{\mathbf{M}}^\textrm{(NC)}$ with $L=15$ layers and obtain
$\widehat{\mathbf{W}}^{*} = \mathbf{V} \mathbf{S}^{+} \mathbf{U}^\top \mathbf{Y}_\textrm{[train]} $
with $\mathbf{U}, \mathbf{S}, \mathbf{V} \leftarrow \textrm{SVD}_{100}(\widehat{\mathbf{M}}^\textrm{(NC)}_\textrm{[train]} )$, per in RHS of Eq.~\ref{eq:wstar}.
For our second ``+ dropout'' model variant, we (implicit) augment the data by
$ {\widehat{\mathbf{M}}^{\textrm{(NC)}^\top}} \leftarrow \big[{\widehat{\mathbf{M}}^{\textrm{(NC)}^\top} } \mid {\textrm{PD}( \widehat{\mathbf{M}}^{\textrm{(NC)}} )^\top } \big]$,
update indices ${[\textrm{train}]} \leftarrow \big[ \textrm{train} \mid \textrm{train} \big]^\top$ then similarly learn as: $\textrm{SVD}_{100}(\widehat{\mathbf{M}}^\textrm{(NC)}_\textrm{[train]}) \rightarrow \widehat{\mathbf{W}}^{*} = \mathbf{V} \mathbf{S}^{+} \mathbf{U}^\top \mathbf{Y}_\textrm{[train]} $
\textbf{Discussion:}
Our method is competitive yet trains faster than SOTA. In fact, the only method that reliably beats ours on all dataset is GCNII, but its training time is about one-thousand-times longer.
For LP over SNAP and node2vec datasets, the training adjacency $\mathbf{A}$ is given but not $\mathbf{X}$. The split includes test positive and negative edges, which are used to measure a ranking metric: ROC-AUC, where the metric increases when test positive edges are ranked above than negatives.
Table \ref{table:snapAUC} summarizes the results.
For
\textbf{baselines}, we download code of WYS \citep{wys};
we use the efficient implementation of PyTorch-Geometric \citep{pyg} for node2vec (n2v)
and we download code of \citet{qiu2018network} and run it with their two variants, denoting their
first variant as NetMF (for \textit{exact}, explicitly computing the design matrix) as and their second variant as $\textrm{Net}\widetilde{\textrm{MF}}$ (for \textit{approximate}, sampling matrix entry-wise)
-- their code runs SVD after computing either variant.
For the first of \textbf{our models}, we compute $\mathbf{U}, \mathbf{S}, \mathbf{V} \leftarrow \textrm{SVD}_{32}({\widehat{\mathbf{M}}^\textrm{(NE)}})$ and score every test edge as
$\mathbf{U}_i^\top \mathbf{S} \mathbf{V}_j$. For the second, we first run $\textrm{SVD}_{256}$ on \textbf{half} of the training edges, determine the ``\textit{best}'' rank $\in \{8, 16, 32, 128, 256\}$ by measuring the AUC on the remaining half of training edges, then using this best rank, recompute the SVD on the entire training set, then finally score the test edges.
\textbf{Discussion}: Our method is competitive on SOTA while training much faster.
Both NetMF and WYS explicitly calculate a dense matrix before factorization.
On the other hand, $\textrm{Net}\widetilde{\textrm{MF}}$ approximates the matrix entry-wise, trading accuracy for training time. In our case, we have the best of both worlds: using our symbolic representation and SVD implementation, we can decompose the design matrix while only implicitly representing it, as good as if we had explicitly calculated it.
\subsection{Experiments on Stanford's OGB datasets}
\label{sec:exp_ogb}
We summarize experiments ogbl-DDI and ogbn-ArXiv, respectively, in Tables \ref{table:ogb1-ddi} and
\ref{table:ogbn-arxiv}. For \textbf{baselines}, we copy numbers from the \href{https://ogb.stanford.edu/docs/leader_overview/}{public leaderboard}, where the competition is fierce.
We then follow links on the leaderboard to download author's code, that we re-run, to measure the training time.
\textbf{For our models on ogbl-DDI,}
we (a) first calculate $\mathbf{U}, \mathbf{S}, \mathbf{V} \leftarrow \textrm{SVD}_{100}(\widehat{\mathbf{M}}^\textrm{(NE)})$
built only from training edges
and score test edge $(i, j)$ using $\mathbf{U}_i^\top \mathbf{S} \mathbf{V}_j$.
Then, we (b) then finetune $f_{\mu, s}$ (per \S\ref{sec:finetuneNE}, Eq.~\ref{eq:gaussiankernel} \& \ref{eq:finetuneNE})) for \textbf{only a single epoch}
and score using $f_{\mu, s}(i, j)$.
Then, we (c) update the SVD basis to include edges from the validation partition and also score using $f_{\mu, s}(i, j)$. We report the results for the three steps. For the last step,
the rules of OGB allows using the validation set for training, but only after the hyperparameters are finalized.
The SVD has no hyperparameters (except the rank, which was already determined by the first step).
More importantly, this simulates a realistic situation:
it is cheaper to obtain SVD (of an implicit matrix) than back-propagate through a model.
For a time-evolving graph, one could run the SVD more often than doing gradient-descent epochs on a model.
For \textbf{our models on ogbn-ArXiv}, we (a) compute $\mathbf{U}, \mathbf{S}, \mathbf{V} \leftarrow \textrm{SVD}_{250}(\widehat{\mathbf{M}}^\textrm{(NC)}_\textrm{LR})$ where the implicit matrix is defined in \S\ref{sec:creative}.
We (b) repeat this process where we replicate $\widehat{\mathbf{M}}^\textrm{(NC)}_\textrm{LR}$ once: in the second replica, we replace the $\mathbf{Y}_\textrm{[train]}$ matrix with zeros (as-if, we drop-out the label with 50\% probability). We (c) repeat the process where we concatenate two replicas of $\widehat{\mathbf{M}}^\textrm{(NC)}_\textrm{LR}$ into the design matrix, each with different dropout seed. We
(d) fine-tune the last model over 15 epochs using stochastic GTTF \citep{gttf}. \textbf{Discussion}: Our method competes or sets SOTA, while training much faster.
\newline
\underline{\textbf{Time improvements}}: We replaced the recommended orthonormalization of \citep{halko2009svd} from QR decomposition, to Cholesky decomposition (\S A.2).
Further, we implemented \textit{caching} to avoid computing sub-expressions if already calculated (\S A.3.4). Speed-ups are shown in Fig.~\ref{fig:time}.
\begin{minipage}{0.64\linewidth}
\captionof{table}{Test classification accuracy over ogbn-arxiv.}
\label{table:ogbn-arxiv}
\centering{
\begin{tabular}{r l l r}
\toprule
\multicolumn{4}{c}{ \hspace{-1.3cm} \textbf{Graph dataset:} \hspace{1cm} \textbf{ogbn-ArXiv}} \\
\cmidrule(lr){1-4}
\multicolumn{2}{l }{ \textbf{Baselines:} } & \multicolumn{1}{c}{{accuracy}} & \multicolumn{1}{c}{{tr.time}} \\
\cmidrule(lr){1-2}
\cmidrule(lr){3-3}
\cmidrule(lr){4-4}
GAT+LR+KD & \hspace{-0.3cm}\citep{gat_kd}\hspace{-0.5cm} & 74.16 {\footnotesize$\pm$0.08} & (6h) \\
GAT+LR &\hspace{-0.3cm}\citep{gat_topoloss_ogb}\hspace{-0.5cm} & 73.99 {\small$\pm$0.12} & (3h) \\
AGDN &\hspace{-0.3cm}\citep{sun2020adaptive}\hspace{-0.1cm} & 73.98 {\footnotesize$\pm$0.09} & (50m) \\
GAT+C\&S & \hspace{-0.1cm}\citep{huang2021combining}\hspace{-0.1cm} & 73.86 {\small$\pm$0.14} & (2h) \\
GCNII &\hspace{-0.3cm}\citep{GCNII}\hspace{-0.5cm} & 72.74 {\footnotesize$\pm$0.16} & (3h) \\
\cmidrule(lr){1-2} \cmidrule(lr){3-3} \cmidrule(lr){4-4}
\multicolumn{2}{l }{ \textbf{Our models:} }
& \multicolumn{1}{c}{{accuracy}}
& \multicolumn{1}{c}{{tr.time}} \\
\cmidrule(lr){1-2}
\cmidrule(lr){3-3}
\cmidrule(lr){4-4}
\multicolumn{1}{l}{(a) $\textrm{iSVD}_{250}(\widehat{\mathbf{M}}^\textrm{(NC)}_\textrm{LR})$}\hspace{-2cm} &
\multicolumn{1}{r}{(\S\ref{sec:cvx:MP}, Eq.~\ref{eq:wstar})} & 68.90 {\footnotesize $\pm$0.02} & (1s) \\
\multicolumn{1}{l}{(b)\hspace{0.1cm} + dropout(LR)}\hspace{-2cm} &\multicolumn{1}{r}{(\S\ref{sec:creative})} & 69.34 {\footnotesize $\pm$0.02} & (3s) \\
\multicolumn{1}{l}{(c)\hspace{0.1cm} + dropout($\widehat{\mathbf{M}}^\textrm{(NC)}_\textrm{LR})$}\hspace{-2cm}& \multicolumn{1}{r}{(\S\ref{sec:creative})} & 71.95 {\footnotesize $\pm$0.03} & (6s) \\
\multicolumn{1}{l}{(d)\hspace{0.1cm} + finetune $\mathbf{H}$}\hspace{-2cm} &
\multicolumn{1}{r}{(\S\ref{sec:finetuneNC}, Eq.~\ref{eq:splitrelu_mp}-\ref{eq:splitrelu_output_init})} & 74.14 {\footnotesize $\pm$0.05} & (2m) \\
\bottomrule
\end{tabular}{}
}
\end{minipage}
\begin{minipage}{0.34\linewidth}
\centering{
\includegraphics[width=3cm]{figs/time_chart.pdf}
}
\captionof{figure}{SVD runtime configs of (lazy caching; orthonormalization) as a ratio of SVD's common default (QR decomposition)}
\label{fig:time}
\end{minipage}
\section{Related work}
\textbf{Applications:}
SVD was used to project rows \& columns of matrix $\mathbf{M}$ onto an embedding space. $\mathbf{M}$ can be the Laplacian of a homogenous graph \citep{eigenmaps}, Adjacency of user-item bipartite graph \citep{koren2009recsys}, or stats \citep{deerwester1990-lsi, levy2014-neural} for word-to-document.
We differ:
our $\mathbf{M}$ is a \textit{function} of (leaf) matrices -- useful when $\mathbf{M}$ is expensive to store (e.g., quadratic).
While \citet{qiu2018network} circumvents this by entry-wise sampling the (otherwise $n^2$-dense) $\mathbf{M}$, our SVD implementation could decompose exactly $\mathbf{M}$ without calculating it.
\textbf{Symbolic Software Frameworks:} including
Theano \citep{theano},
TensorFlow \citep{tensorflow} and PyTorch \citep{pytorch},
allow chaining
operations
to compose a computation (directed acyclic) graph (DAG).
They can efficiently \textbf{run the DAG upwards} by
evaluating (all entries of) matrix $\mathbf{M} \in \mathbb{R}^{r \times c}$ at any DAG node.
Our DAG differs:
instead of calculating $\mathbf{M}$,
it provides product function
$u_\mathbf{M}(.) = \mathbf{M} \times .$
--
The \textbf{graph is run downwards} (reverse direction of edges).
\textbf{Matrix-free SVD:} For many matrices of interest, multiplying against $\mathbf{M}$ is computationally cheaper than explicitly storing $\mathbf{M}$ entry-wise.
As such, many researchers implement $\textrm{SVD}(u_\mathbf{M})$, e.g. \citet{calvetti1994restarted, kesheng2000thick, bose2019terapca}. We differ in the programming flexibility: the earlier methods expect the practitioner to directly implement $u_\mathbf{M}$. On the other hand, our framework allows composition of $\mathbf{M}$ via operations native to the practitioner (e.g., @, +, concat), and $u_\mathbf{M}$ is automatically defined.
\textbf{Fast Graph Learning:} We have applied our framework for fast graph learning, but so as
\textit{sampling-based approaches} including \citep{fastgcn, clustergcn, graphsaint-iclr2020, gttf}. We differ in that ours can be used to obtain an initial closed-form solution (very quickly) and can be fine-tuned afterwards using any of the aforementioned approaches.
Additionally, Graphs shows one general application.
Our framework might be useful in other areas utilizing SVD.
\section{Conclusion, our limitations \& possible negative societal impact}
We develop a software framework for symbolically representing matrices and compute their SVD.
Without computing gradients,
this trains convexified models over
GRL tasks, showing empirical metrics competitive with SOTA while training
significantly faster.
Further,
convexified model parameters can
initialize (deeper) neural networks that can be fine-tuned with cross entropy.
Practitioners adopting our framework would now spend more effort in crafting design matrices instead of running experiments or tuning hyperparameters of the learning algorithm.
We hope our framework makes high performance graph modeling more accessible by reducing reliance on
energy-intensive computation.
\textbf{Limitations of our work}: From a representational prospective,
\textbf{our space of functions is smaller} than TensorFlow's,
as our DAG must be a linear transformation of its input matrices,
\textit{e.g.},
unable to encode element-wise transformations and hence demanding first-order linear approximation of models.
As a result, our \textbf{gradient-free learning} can be performed using SVD. Further, our framework only works when the leaf nodes (\textit{e.g.}, sparse adjacency matrix) fit in memory of one machine.
\textbf{Possible societal impacts}:
First, our convexified models directly learn parameters in the feature space i.e. they are more explainable than deeper counterparts. Explainability is a double-edged sword: it gives better ability to interpret the model's behavior, but also allows for malicious users, e.g., to craft attacks on the system (if they can replicate the model parameters).
Further, we apply our method on graphs. It is possible to train our models to detect sensitive attributes of social networks (e.g., ethnicity). However, such ethical concerns exists with any modeling technique over graphs.
\begin{ack}
This material is based upon work supported by the Defense Advanced
Research Projects Agency (DARPA) and the Army Contracting Command-
Aberdeen Proving Grounds (ACC-APG) under Contract Number
W911NF-18-C-0020.
\end{ack}
\bibliographystyle{plainnat}
|
\section{Introduction}
Physical systems are described by sets of partial differential equations
(PDEs). In general, these equations can be expressed as first-order partial
derivatives, leading us to focus on the study of first-order quasi-linear
PDEs. These equations (or systems) usually include gauge freedoms and
differential constraints (see \cite{geroch1996partial},
\cite{Hilditch:2013ila}). Namely, degrees of freedom (variables) that are not
determined by the equations and degrees of freedom that are restricted to a
lower-dimensional subspace, respectively. In this paper we do not consider
gauge freedoms, we assume that they have already been fixed if the system
includes them, and instead we concentrate on the study of the constraints.
Furthermore, in our development, we do not assume the presence of a background
metric, so the relativistic systems are included in our analysis but we do not
restrict exclusively to them.
To study the above mentioned physical systems, the space-time $M$ is foliated
as $M=\left[ 0,T\right] \times\underset{0\leq t\leq T}{\cup}\Sigma_{t}$ and
the PDEs are divided into two subsets, evolution and constraints equations.
The evolution equations determine how the variables change along the different
hypersurfaces $\Sigma_{t}$, with $0<t\leq T$, while the constraint equations
restrict the allowed values of the variables on each $\Sigma_{t}$. In order
for these two subsets to have predictive power, the associated Cauchy problem
should be well-posed (see \cite{gustafsson1995time}, \cite{kreiss2004initial},
\cite{sarbach2012continuum}). This means that given an initial data over
$\Sigma_{t}$, the associated solutions of the evolution equations exist, are
unique, continuous with respect to the initial data and satisfy the
constraints. Although, we emphasize that the evolution equations describing a
physical system are not unique, since they can be modified by adding
constraint terms to them. Such freedom can lead to well- or ill-posed
evolution equations \cite{kreiss2004initial}, \cite{abalos2017necessary}. In
this paper, we restrict our attention to one class of equations within the
well-posed ones, the strongly hyperbolic (SH). In this class, the recipe for
finding strongly hyperbolic evolution equations has been given in
\cite{Abalos:2018uwg}.
Several methods are used to find numerical and analytical solutions to the
PDEs. Among the most widely used is the free-evolution approach. In this
approach, one begins by verifying that when constraints are initially
satisfied they will remain preserved during the evolution, a property commonly
referred to as constraint preservation (or conservation). Then, one solves the
evolution equations for initial data which satisfy the constraints; and due to
the constraints preservation, this method automatically yields solutions of
the complete system. Some relevant numerical implementations of this approach
can be founded in reviews \cite{Frauendiener:2011zz, Lehner:2014asa,
sarbach2012continuum, shibata2015numerical, Palenzuela:2020tga} and references therein.
The present paper addresses the study of the constraint preservation from a
PDEs point of view, by establishing sufficient conditions that the equations
of a physical system have to satisfy to guarantee the constraint preservation.
The standard method to verify this conservation is: 1) deriving a set of
evolution equations for the constraints, 2) establishing that zero constraints
are a solution of this system and 3) checking that this system is strongly
hyperbolic. When these steps are satisfied, they analytically guarantee that
the vanishing solution is unique and therefore that the constraints are
preserved. Moreover, the strong hyperbolicity is used to find which boundary
conditions preserve the constraints (see \cite{Calabrese:2002xy},
\cite{Tarfulea:2013qjq}, \cite{sarbach2012continuum} \cite{Calabrese:2002xy}).
What do we refer to by "a set of evolution equations for the constraints"?. It
is a set of partial differential evolution equations whose variables are the
constraints. This set of equations is called subsidiary system (SS). \ Most
well-known physical systems have quasi-linear first-order partial differential
subsidiary systems. Some examples are: Maxwell \cite{Calabrese2004ARF}, and
Non-linear \cite{PhysRevD.92.084024} Electrodynamics, Einstein
\cite{Frittelli:1996nj}, \cite{Shinkai:2008yb}, Einstein-Christoffel
\cite{Calabrese:2002xy}\ ,BSSN \cite{Yoneda:2002kg}, \cite{Beyer2004OnTW}, ADM
\cite{Kidder:2001tz}, \cite{Yoneda:2001iy}, $n+1$ ADM \cite{Shinkai:2003vu},
$f\left( R\right) $-Gravity \cite{Paschalidis:2011ww},
\cite{Mongwane:2016qtz}, Bimetric Relativity \cite{Kocic:2018yvr} theories,
etc. The process for deriving the SS and verifying its strong hyperbolicity
is\ conducted separately for each physical theory and usually involves very
cumbersome calculations.To assess this problem, we present a theory that
simplifies and automatizes the process.
In \cite{reula2004strongly}, Reula studies the constant coefficient case,
assumes the existence of a first-order SS and explains how the characteristic
structure of this system connects with the characteristic structure of the
evolution equations of the system. Here, we focus on the quasi-linear case,
show which conditions guarantee the existence of a SS and explain how the
Reula result arises naturally from our results. Indeed, we show that the
principal symbol of the SS is a simple projection of some tensorial objects
$C_{A}^{a\Gamma}$ called here Geroch fields (see \cite{geroch1996partial}). We
also show that the resulting SSs are not unique. In fact, they are families of
equations with free parameters letting to choose the propagation velocities of
the constraints.
Finally, we restrict ourselves to the constant-coefficient case to simplify
the discussion and give sufficient conditions that the original set of PDEs
has to satisfy for having SH evolution equations and an associated SH
subsidiary system. To reach this result we give a detailed description of the
characteristic structure of both subsystems and explain how they should be
chosen to make them SH.
The constraints studied here are called first-class constraints in their
Hamiltonian version, see for example \cite{Henneaux:1992ig},
\cite{Giulini:2015qha}, \cite{Hilditch:2013ila}. We highlight the work of
Hilditch and Richter \cite{Hilditch:2013ila}, where a similar problem to the
one presented here is studied with a different approach.
There are widely used methods that introduce extra variables to the system and
avoid dealing with constraints. Some of the most popular ones are the
$\lambda-$systems \cite{brodbeck1999einstein}, the divergence cleaning
\cite{munz2000three}, \cite{dedner2002hyperbolic} for electrodynamics, the Z4
systems\ \cite{Bona:2002fq, Bona:2002ft, Bona:2003fj, Bona:2004yp,
Bona:2004ky, Alic:2011gg, Gundlach:2005eh, Bernuzzi:2009ex} and the modified
harmonic gauge \cite{Kovacs:2020ywu} for Einstein equations. However, for
these systems to work properly, the constraints conservation of the original
system has to be satisfied. Therefore, this aspect of the problem is crucial.
The outline of the paper is the following. In Section \ref{Seccion_Setting_1},
we introduce the Geroch fields, used to obtain the expressions for the
constraints; the reductions, that select the evolution equations; and the
integrability conditions of the system, which allow to obtain first-order
partial derivatives SSs. In section \ref{Teorema_1}, we introduce the
expressions of the SSs in the first main theorem and the proof to this
theorem. In sections \ref{Sec_Const_coeff_1} and \ref{teorema_2}, we present
the well-posed concepts and our second main theorem about the strong
hyperbolicity of the SS, respectively. This section include the subsection
\ref{Proof_Theorem_coef_const_2}, where we provide the proof of the second
main theorem. This proof is splited in many subsubsections which include the
analysis of the Kronecker decomposition of the principal symbol of the system
and of the SS (including the constraints of the constraints). In subsection
\ref{teo_3}, we discuss how to suppress one condition of the second main
theorem. In section \ref{Examples}, we present two examples of application of
the developed formalism: the Maxwell electrodynamics and the wave equation. In
section \ref{Conclusions}, we briefly discuss the results and make some
comments on future work. In appendix \ref{Ap_lemmas}, we introduce some lemmas
that are used to find the Kronecker decomposition of a pencil. They are used
in the proof of Theorem \ref{teorema_2}. Finally, in appendix
\ref{App_coordenadas}, we discuss the introduction of the lapse and shift
variables in the $n+1$ foliations. This is done in the simple system
$\nabla_{a}q^{b}=0$, in cases with and without a metric.
\section{Setting\label{Seccion_Setting_1}}
Following Geroch's notation, let $b\overset{\pi}{\rightarrow}M$ a fiber bundle
over a space-time $M$ with $\dim M=n+1$. We consider the cross-section
$\phi^{\alpha}:M\rightarrow b$, which defines the physics fields and satisfies
the equation of motion
\begin{equation}
E^{A}:=\mathfrak{N}_{~\alpha}^{Aa}\left( x,\phi\right) \nabla_{a
\phi^{\alpha}-J^{A}\left( x,\phi\right) =0. \label{eq_sys_1
\end{equation}
Here the Greek indices $\alpha,\beta,\gamma\,$represent field indices, lower
letters $a,b,c,...$ represent space-time indices and capital letters
$A,B,C,...$ represent multi-tensorial indices on the fiber space of equations
call $\Psi_{L}$. We are considering $\dim\left( A\right) =e$ and
$\dim\left( \alpha\right) =u$ such that $e\geq u$. The tensor fields
$\mathfrak{N}_{~\alpha}^{Aa}$ and $J^{A}$ are the principal symbol and the
source term respectively, they do not depend on derivatives of $\phi$. \ The
derivative $\nabla_{a}$ represent a partial derivative or a Levi Civitta
connection when the system has \ a metric.
Solving $E^{A}=0$, with a given initial condition for unknown fields
$\phi^{\alpha}$, is called the initial value problem of the system
(\ref{eq_sys_1}). As we commented in the introduction, this can be done
numerically or analytically using the free-evolution approach, i.e.,
separating $E^{A}=0$ into two sets of equations, the evolutions and the
constraints, finding the Subsidiary System (SS), showing the constraint
conservation and using the evolution equations to find the solutions.
Following this path, we present in the first two sections, and from a PDE
perspective, a theory that introduces the sufficient conditions by which a
system such as (\ref{eq_sys_1}) has a SS of first-order partial derivatives.
We begin, this section, introducing the main tools of the article, defining
the evolution, the constraints and the integrability conditions associated to
(\ref{eq_sys_1}).
\subsection{$n+1$ foliation\label{n+1_decomposition_sec}}
We introduce a foliation of $M$ given by the level surfaces of a function
$t:M\rightarro
\mathbb{R}
$, and call $\Sigma_{t_{0}}=\left\{ p\in M\mid t\left( p\right)
=t_{0}\right\} $ to these hypersurfaces. We also introduce coordinates
$x^{a}=\left( t,x^{i}\right) ,$ $i=1,..,n,$ with $x^{i}$ adapted to the
$\Sigma_{t_{0}}$'s. This is a local $n+1$ foliation of $M$ given by
$[0,T]\times\Sigma_{t_{0}}$ with $T\i
\mathbb{R}
.$ We call \textit{time coordinate} to $t$ and \textit{spatial coordinates} to
$x^{i}$; these names are just names to differentiate one coordinate from the
others since we are not assuming the presence of a background metric here. We
also consider the vector $t^{a}:=\left( \partial_{t}\right) ^{a}\dot
{=}\left( 1,0,...,0\right) $, and co-vector $n_{a}:=\nabla_{a}t\dot
{=}\left( 1,0,...,0\right) $, where the dot in the equal sign means "in this
coordinates the explicit expressions are". They satisfy the condition
$t^{a}n_{a}=1$, which allows to definite the projector
\begin{equation}
\eta_{b}^{a}:=\delta_{b}^{a}-t^{a}n_{b}\dot{=}\left[
\begin{array}
[c]{cccc
0 & 0 & 0 & 0\\
0 & 1 & 0 & 0\\
0 & 0 & ... & 0\\
0 & 0 & 0 & 1
\end{array}
\right] . \label{Eq_eta_1
\end{equation}
It has the properties $\eta_{b}^{a}t^{b}=0=\eta_{b}^{a}n_{a}$ and $\eta
_{b}^{a}\eta_{c}^{b}=\eta_{c}^{a}$. We use them to rewrite (\ref{eq_sys_1}),
as follow
\begin{align}
E^{A} & =\mathfrak{N}_{~\alpha}^{Ac}\left( n_{c}t^{b}+\eta_{c}^{b}\right)
\nabla_{b}\phi^{\alpha}-J^{A},\label{eq_split_E_1}\\
& =\mathfrak{N}_{~\alpha}^{Ac}n_{c}\nabla_{t}\phi^{\alpha}+\mathfrak{N
_{~\alpha}^{Ac}\eta_{c}^{b}\nabla_{b}\phi^{\alpha}-J^{A}=0
\label{eq_split_E_2
\end{align}
In this equation the derivatives are splitting in time and spatial derivatives
since the term $\eta_{c}^{b}\nabla_{b}$ has no derivatives in the $t^{c}$
direction ($t^{c}\eta_{c}^{b}\nabla_{b}=0$). Notice that if we use the
coordinates $\left( t,x^{i}\right) $ the last expression can be rewritten
as
\[
E^{A}\dot{=}\mathfrak{N}_{~\alpha}^{A0}\nabla_{t}\phi^{\alpha}+\mathfrak{N
_{~\alpha}^{Ai}\nabla_{i}\phi^{\alpha}-J^{A}=0,
\]
where $i=1,...,n$ and
\[
\mathfrak{N}_{~\alpha}^{A0}:=\mathfrak{N}_{~\alpha}^{Ac}n_{c}.
\]
In these expressions the temporal and spatial derivatives are explicit.
Without loss of generality, we will assume along this work that we are in
these coordinates. \ Nevertheless, to simplify the notation we will suppress
the dot in the symbol $\dot{=}$ and the use of $\left( t,x^{i}\right) $ will
be understood from the context.
\subsection{Geroch fields and constraints}
We introduce now the main objects of this article, the Geroch fields
$C_{A}^{a}\left( x,\phi\right) $, which are used to define the constraints
as we explain below (see \cite{geroch1996partial} too). These Geroch fields
are defined by the equation
\begin{equation}
C_{A}^{(a}\mathfrak{N}_{~\alpha}^{\left\vert A\right\vert b)}=0.
\label{Eq_C_K_nueva_1
\end{equation}
\bigskip At each point $\kappa=\left( x,\phi\right) \in b$, they define a
vector space\footnote{Notice that given $X_{A}^{a}$ and $Y_{A}^{a}$ satisfying
(\ref{Eq_C_K_nueva_1}), any linear combination $\alpha X_{A}^{a}+\beta
Y_{A}^{a}$ with $\alpha.\beta\i
\mathbb{R}
$ also satisfies (\ref{Eq_C_K_nueva_1}).}.
We introduce now a condition about these objects that we will explain after
the definition of the constraints. For every open $U$ of $M$, we assume that
it is possible to choos
\begin{equation}
c:=e-u \label{Eq_c_e_u_1
\end{equation}
Geroch fields $C_{A}^{\Gamma a}$, denoted by the index $\Gamma$, and such that
for all $x\in U$ and all $\phi$, the components
\[
C_{A}^{\Gamma0}:=C_{A}^{\Gamma a}n_{a
\]
are linearly independent. This field $C_{A}^{\Gamma0}$ can be thought of as a
linear transformation from the index $A$ to the index $\Gamma$, so the above
condition is equivalent to:
\begin{condition}
\label{cod_C0_rank_max_b}For each $\kappa=\left( x,\phi\right) \in b$,
\begin{equation}
C_{A}^{\Gamma0}\text{ has maximal rank\footnote{There is no $X_{\Gamma}$ such
that $X_{\Gamma}n_{a}C_{A}^{\Gamma a}=0$ to except of the trivial one
$X_{\Gamma}=0$}.} \label{cod_C0_rank_max
\end{equation}
\end{condition}
These Geroch fields $C_{A}^{\Gamma a}$ satisfy the equation
\begin{equation}
C_{A}^{\Gamma(a}\mathfrak{N}_{~\alpha}^{\left\vert A\right\vert b)}=0,
\label{eq_CK_1
\end{equation}
and define the \textit{constraints} of the system as follows
\[
\psi^{\Gamma}:=n_{a}C_{A}^{\Gamma a}E^{A}.
\]
These $\psi^{\Gamma}$ are called the constraints equations since they have not
time derivatives, they only have spatial derivatives as follow.
\begin{align}
\psi^{\Gamma} & =n_{a}C_{A}^{\Gamma a}E^{A}\nonumber\\
& =n_{a}C_{A}^{\Gamma a}\mathfrak{N}_{~\alpha}^{Ac}n_{c}\nabla_{t
\phi^{\alpha}+n_{a}C_{A}^{\Gamma a}\mathfrak{N}_{~\alpha}^{Ac}\eta_{c
^{b}\nabla_{b}\phi^{\alpha}-n_{a}C_{A}^{\Gamma a}J^{A}\nonumber\\
& =n_{a}C_{A}^{\Gamma a}\mathfrak{N}_{~\alpha}^{Ac}\eta_{c}^{b}\nabla_{b
\phi^{\alpha}-n_{a}C_{A}^{\Gamma a}J^{A}\label{eq_def_constr_1}\\
& =C_{A}^{\Gamma0}\mathfrak{N}_{~\alpha}^{Ai}\nabla_{i}\phi^{\alpha
-C_{A}^{\Gamma0}J^{A}=0.\nonumber
\end{align}
Where equation (\ref{eq_split_E_2}) was used in the first line and $n_{a
C_{A}^{\Gamma a}\mathfrak{N}_{~\alpha}^{Ac}n_{c}=0$ (follows directly from
(\ref{eq_CK_1})) in the fourth line. This means that they can be calculated
only with the information of $\phi^{\alpha}$ pullbacked to the hypersurfaces
$\Sigma_{t}$.
Notice that these constraints depend on the hypersurfaces considered through
$n_{a}$. In general and when we have a metric, these hypersurfaces are chosen
as spatial, however, there is no restriction on $n_{a}$ to be temporal so by
changing the choice of $n_{a}$ we could obtain constraints associated with any hypersurface.
The motivation of the introduction of condition \ref{cod_C0_rank_max} is that
it implies that the $\psi^{\Gamma}$ are algebraically independents and the
number of them is $c$. This is a natural condition in physical examples as
General Relativity. Recalling that $c$ is the difference between the number of
equations and the fields (see eq. (\ref{Eq_c_e_u_1})), we observe that there
remain $u$ equations in $E^{A}$, these will be the evolution equations.
On the other hand, we notice that assumption (\ref{cod_C0_rank_max}) is a
non-covariant request, however, we will explain how to recover the covariance
in subsection \ref{Sec_Vin_M}. It will be with the introduction of another
kind of Geroch fields.
\begin{remark}
A set of PDE's (as (\ref{eq_sys_1})) that includes constraints might not admit
Geroch fields. In this case, the system should accept another kind of fields
$C_{A}^{a_{1}a_{2}...a_{m}}$ used to define the constraints. Such systems are
not considered here and will be included in future work. However, this does
not represent a significant loss of generality since, as Geroch showed in
\cite{geroch1996partial}, most of the classical physical systems have only
Geroch fields.
\end{remark}
\subsection{Non algebraic constraints}
We also assume, in this work, that there is no $X_{A}$ such tha
\begin{equation}
X_{A}\mathfrak{N}_{~\alpha}^{Aa}=0. \label{eq_X_K_1
\end{equation}
It is equivalent to claim that the system (\ref{eq_sys_1}) has no algebraic
constraints since if such $X_{A}$ exists, then the expression
\begin{equation}
X_{A}E^{A}:=-X_{A}J^{A}=0 \label{eq_alg_con_1
\end{equation}
would be an algebraic constraint. In the cases in which the system has an
$X_{A}$ satisfying (\ref{eq_X_K_1}), we redefine the system to obtain a new
system without algebraic constraints. This is done by suppressing some
equations from (\ref{eq_sys_1}) and adding the derivatives of
(\ref{eq_alg_con_1}) to the system. We end-up with a new system of first order
in derivatives with more equations, equivalent to the original system, but
without the algebraic constraint. We assume that this new system only admits
Geroch fields.
\subsection{Evolution equations\label{Evo}}
We have introduced the constraints, now we introduce the evolution equations.
We begin with an additional assumption,
\begin{condition}
\label{cod_K0_rank_max_b} For each $\kappa\in b$,
\begin{equation}
\mathfrak{N}_{~\beta}^{A0}\text{ has maximal rank\footnote{This means there is
not $\delta\phi^{\alpha}\neq0$ such that $\mathfrak{N}_{~\beta}^{A0}\delta
\phi^{\alpha}=0.$}.} \label{cod_K0_rank_max
\end{equation}
\end{condition}
This condition guarantees that we can define evolution equations for each of
the unknown variables \ $\phi^{\alpha}$ as follow.
First, we introduce the \textit{reduction} tensor $h_{A}^{\alpha}$ such that
\begin{equation}
h_{A}^{\alpha}\mathfrak{N}_{~\beta}^{A0}=\delta_{\beta}^{\alpha},
\label{eq_hK0_1
\end{equation}
where $\delta_{\beta}^{\alpha}$ is the identity map. This reduction is always
possible to find when condition (\ref{cod_K0_rank_max}) holds.
Now, with the aid of the reduction $h_{A}^{\alpha}$, we choose the
\textit{evolution equations
\begin{equation}
e^{\alpha}:=h_{A}^{\alpha}E^{A}=t^{a}\nabla_{a}\phi^{\alpha}+h_{A}^{\alpha
}\mathfrak{N}_{~\beta}^{Ac}\eta_{c}^{b}\nabla_{b}\phi^{\beta}-h_{A}^{\alpha
}J^{A}=0, \label{eq_evol_1
\end{equation}
where expressions (\ref{eq_split_E_2}) and (\ref{eq_hK0_1}) were used. These
equations are called evolution equations since they include derivatives in the
$t^{a}$ direction for each $\phi^{\alpha}$, telling how the fields\ $\phi
^{\alpha}$ evolve outside of the hypersurfaces $\Sigma_{t}$ In general, when
the system has a background metric, $t^{a}$ is chosen temporal.
These evolution equations are a linear combination of the eq. (\ref{eq_sys_1})
and, of course, they are not unique. We can define other reductions $\tilde
{h}_{A}^{\alpha}$ as follo
\[
\tilde{h}_{A}^{\alpha}=h_{A}^{\alpha}+p_{A}^{\alpha},
\]
where $p_{A}^{\alpha}$ satisfies
\[
p_{A}^{\alpha}\mathfrak{N}_{~\beta}^{A0}=0,
\]
and produce another set of evolutions equations
\begin{equation}
\tilde{e}^{\alpha}:=\tilde{h}_{A}^{\alpha}E^{A}=t^{a}\nabla_{a}\phi^{\alpha
}+\left( h_{A}^{\alpha}\mathfrak{N}_{~\beta}^{Ac}+p_{A}^{\alpha
\mathfrak{N}_{~\beta}^{Ac}\right) \eta_{c}^{b}\nabla_{b}\phi^{\beta}-\left(
h_{A}^{\alpha}+p_{A}^{\alpha}\right) J^{A}=0. \label{Eq_evo_2
\end{equation}
In general, this freedom is used to obtain well-posed evolution equations (see
\cite{Abalos:2018uwg}).
\subsection{Evolution and constraints equations}
By requiring that system (\ref{eq_sys_1}) satisfies conditions
(\ref{cod_C0_rank_max}) and (\ref{cod_K0_rank_max}), we have decomposed
(\ref{eq_sys_1}) into two subsets of equations, the evolution \ ($e^{\alpha}$)
and constraint equations ($\psi^{\Delta}$) as follow
\begin{equation}
\left[
\begin{array}
[c]{c
e^{\alpha}\\
\psi^{\Delta
\end{array}
\right] :=\left[
\begin{array}
[c]{c
h_{A}^{\alpha}\\
C_{A}^{\Delta0
\end{array}
\right] E^{A}. \label{Eq_e_y_vin_Mat_E_1
\end{equation}
Now, we will show that for each $\kappa\in b$, the matrix $\left[
\begin{array}
[c]{c
h_{B}^{\alpha}\\
C_{B}^{\Delta0
\end{array}
\right] $ is invertible. This means that $e^{\alpha}$ and $\psi^{\Delta}$ are
linearly equivalent to the whole system.
\begin{lemma}
For each $\kappa\in b$, the matrix $\left[
\begin{array}
[c]{c
h_{B}^{\alpha}\\
C_{B}^{\Delta0
\end{array}
\right] $ is invertible, with inverse $\left[
\begin{array}
[c]{cc
\mathfrak{N}_{~\beta}^{B0} & h_{~\Gamma}^{B
\end{array}
\right] $, wher
\
\begin{array}
[c]{ccc
C_{B}^{\Delta0}h_{~\Gamma}^{B}=\delta_{\Gamma}^{\Delta}, & & h_{B}^{\alpha
}h_{~\Gamma}^{B}=0
\end{array}
\]
an
\begin{equation}
\left[
\begin{array}
[c]{cc
\mathfrak{N}_{~\alpha}^{A0} & h_{~\Delta}^{A
\end{array}
\right] \left[
\begin{array}
[c]{c
h_{B}^{\alpha}\\
C_{B}^{\Delta0
\end{array}
\right] =\mathfrak{N}_{~\alpha}^{A0}h_{B}^{\alpha}+h_{~\Delta}^{A
C_{B}^{\Delta0}=\delta_{B}^{A} \label{eq_Kh_hC_1
\end{equation}
\end{lemma}
\begin{proof}
To show that $\left[
\begin{array}
[c]{c
h_{B}^{\alpha}\\
C_{B}^{\Delta0
\end{array}
\right] $ is invertible, we will assume that it is not, and conclude an
absurd. So, if $\left[
\begin{array}
[c]{c
h_{B}^{\alpha}\\
C_{B}^{\Delta0
\end{array}
\right] $ is not invertible, there exists $\left[
\begin{array}
[c]{cc
X_{\alpha} & Y_{\Delta
\end{array}
\right] \neq0$ such that
\begin{equation}
X_{\alpha}h_{B}^{\alpha}+Y_{\Delta}C_{B}^{\Delta0}=0. \label{eq_ld_h_C0
\end{equation}
\sloppy
Multiplying this expression by $\mathfrak{N}_{~\beta}^{B0}$ and using the
equations (\ref{eq_hK0_1}) and $C_{A}^{\Gamma0}\mathfrak{N}_{~\alpha}^{A0}=0$
we conclude
\[
X_{\alpha}=0.
\]
This means that $Y_{\Delta}C_{B}^{\Delta0}=0$, but since $C_{B}^{\Delta0}$ has
only trivial kernel in the $\Delta$ index, it follows $Y_{\Delta}=0$. So, the
unique solution of eq. (\ref{eq_ld_h_C0}) is $\left[
\begin{array}
[c]{cc
X_{\alpha} & Y_{\Delta
\end{array}
\right] =0$, which is a contradiction. We conclude that $\left[
\begin{array}
[c]{c
h_{B}^{\alpha}\\
C_{B}^{\Delta0
\end{array}
\right] $ is invertible.
If $\left[
\begin{array}
[c]{cc
\mathfrak{N}_{~\beta}^{B0} & h_{~\Gamma}^{B
\end{array}
\right] $ is the inverse, the product
\[
\left[
\begin{array}
[c]{c
h_{B}^{\alpha}\\
C_{B}^{\Delta0
\end{array}
\right] \left[
\begin{array}
[c]{cc
\mathfrak{N}_{~\beta}^{B0} & h_{~\Gamma}^{B
\end{array}
\right] =\left[
\begin{array}
[c]{cc
h_{B}^{\alpha}\mathfrak{N}_{~\beta}^{B0} & h_{B}^{\alpha}h_{~\Gamma}^{B}\\
C_{B}^{\Delta0}\mathfrak{N}_{~\beta}^{B0} & C_{B}^{\Delta0}h_{~\Gamma}^{B
\end{array}
\right] ,
\]
should be equal to the identity matrix, then $C_{B}^{\Delta0}h_{~\Gamma
^{B}=\delta_{\Gamma}^{\Delta}$, $h_{B}^{\alpha}h_{~\Gamma}^{B}=0.$
On the other hand, commuting the matrices we obtain the following equatio
\begin{align*}
\delta_{B}^{A} & =\left[
\begin{array}
[c]{cc
\mathfrak{N}_{~\alpha}^{A0} & h_{~\Delta}^{A
\end{array}
\right] \left[
\begin{array}
[c]{c
h_{B}^{\alpha}\\
C_{B}^{\Delta0
\end{array}
\right] ,\\
& =\mathfrak{N}_{~\alpha}^{A0}h_{B}^{\alpha}+h_{~\Delta}^{A}C_{B}^{\Delta0}.
\end{align*}
\end{proof}
\subsection{Non unicity in the Geroch fields\label{Sec_Vin_M}}
Frequently, the constraints have certain differential relationships between
them on the hypersurfaces $\Sigma_{t}$. The canonical example is when a
second-order derivative system is reduced to first-order by defining the first
derivatives as new variables. This approach reduces the order of the
derivatives and at the same time introduces extra constraints to the system.
These extra constraints also satisfy certain differential relationships
between them, their cross-derivatives should vanish. The way we parameterize
this, and any other, differential relationship between the constraints is
through another class of Geroch fields that we call $M_{A}^{\tilde{\Delta
a}\left( x,\phi\right) $, where the $\tilde{\Delta}$ index numerates them.
For each $\kappa=\left( x,\phi\right) \in b$, these new extra fields satisfy
the equation (\ref{eq_CK_1}), which means
\begin{equation}
M_{A}^{\tilde{\Delta}(a}\mathfrak{N}_{~\alpha}^{\left\vert A\right\vert b)}=0,
\label{eq_M_K_1
\end{equation}
and also satisfy the following condition
\begin{equation}
M_{A}^{\tilde{\Delta}a}\mathfrak{N}_{~\alpha}^{A0}=0. \label{eq_M_K0_1
\end{equation}
Notice that, for each $\kappa,$ the equations (\ref{eq_M_K_1}) and
(\ref{eq_M_K0_1}) define a vector subspace within the Geroch field space.
To simplify the analysis, we add an assumption on the $M_{A}^{\tilde{\Delta
a}$ which also defines the $\tilde{\Delta}$ index.
\begin{condition}
\label{As_M_1_b}In every open $U\subset M$ and for all $\kappa=\left(
x,\phi\right) \in b$ with $x\in U$, there exist a finite number of fields
$M_{A}^{\tilde{\Delta}a}$ indexed by $\tilde{\Delta}$ such that any field
satisfying (\ref{eq_M_K_1}) and (\ref{eq_M_K0_1}) can be expanded as a linear
combination of these $M_{A}^{\tilde{\Delta}a}$. This means that any field
satisfying (\ref{eq_M_K_1}) and (\ref{eq_M_K0_1}) is a linear combination of
the $M_{A}^{\tilde{\Delta}a}$.
\end{condition}
This assumption is usually satisfied in physical systems in such a way that
the pair $\left( \Gamma,\tilde{\Delta}\right) $ define a tensor index. Thus,
the covariance of these Geroch tensors is recovered by considering the pair
$C_{A}^{\Gamma a}$ and $M_{A}^{\tilde{\Delta}a}$ together. We will comment
more about this at the end of this subsection.
Another result that arises from this assumption is:
\begin{lemma}
If the conditions \ref{cod_C0_rank_max}, \ref{cod_K0_rank_max} and
\ref{As_M_1_b} hold, then in every open $U\subset M$ and for all
$\kappa=\left( x,\phi\right) \in b$ with $x\in U$, every Geroch field
$X_{A}^{a}$ (satisfying the eq. (\ref{Eq_C_K_nueva_1})) is a linear
combination of $C_{A}^{\Gamma a}$ and $M_{A}^{\tilde{\Delta}a}$, i.e. there
exist always $S_{\Delta}$ and $S_{\tilde{\Delta}}$ such that
\[
X_{A}^{a}=S_{\Delta}C_{A}^{\Delta a}+S_{\tilde{\Delta}}M_{A}^{\tilde{\Delta
a}.
\]
\end{lemma}
\begin{proof}
Let $X_{A}^{a}$ be any Geroch field, we have to show that there exist
$S_{\Delta}$ and $S_{\tilde{\Delta}}$ such that $X_{A}^{a}=S_{\Delta
C_{A}^{\Delta a}+S_{\tilde{\Delta}}M_{A}^{\tilde{\Delta}a}.$ We propose
$S_{\Delta}=X_{B}^{0}h_{~\Delta}^{B}$ and point out that to conclude the
proof, we only need to show that
\begin{equation}
X_{A}^{a}-\left( X_{B}^{0}h_{~\Delta}^{B}\right) C_{A}^{\Delta a}
\label{eq_proy_a_M_1
\end{equation}
satisfies (\ref{eq_M_K_1}) and (\ref{eq_M_K0_1}). Since, if this is true, by
the condition \ref{As_M_1_b} the field (\ref{eq_proy_a_M_1}) should be
expanded by $M_{A}^{\tilde{\Delta}a}$. Showing (\ref{eq_M_K_1}) is trivial
from the definitions of Geroch fields, so it only remains to demonstrate
(\ref{eq_M_K0_1}). Multiplying (\ref{eq_proy_a_M_1}) by $\mathfrak{N
_{~\alpha}^{A0},$ we obtain
\begin{align*}
\left( X_{A}^{a}-\left( X_{B}^{0}h_{~\Delta}^{B}\right) C_{A}^{\Delta
a}\right) \mathfrak{N}_{~\alpha}^{A0} & =-\left( X_{A}^{0}-\left(
X_{B}^{0}h_{~\Delta}^{B}\right) C_{A}^{\Delta0}\right) \mathfrak{N
_{~\alpha}^{Aa},\\
& =-X_{B}^{0}\left( \delta_{A}^{B}-h_{~\Delta}^{B}C_{A}^{\Delta0}\right)
\mathfrak{N}_{~\alpha}^{Aa},\\
& =-X_{B}^{0}\mathfrak{N}_{~\alpha}^{B0}h_{A}^{\alpha}\mathfrak{N}_{~\alpha
}^{Aa},\\
& =0.
\end{align*}
We used $X_{A}^{a}\mathfrak{N}_{~\alpha}^{A0}=-X_{A}^{0}\mathfrak{N}_{~\alpha
}^{Aa}$ and $C_{A}^{\Delta a}\mathfrak{N}_{~\alpha}^{A0}=-C_{A}^{\Delta
0}\mathfrak{N}_{~\alpha}^{Aa}$ in the first equality, (\ref{eq_Kh_hC_1}) (only
valid assuming \ref{cod_C0_rank_max}, \ref{cod_K0_rank_max}) in the second
one, (\ref{eq_Kh_hC_1}) in the third one and $X_{B}^{0}\mathfrak{N}_{~\alpha
}^{B0}=0$ in the last one. This concludes the proof.
\end{proof}
We said that the $M_{A}^{\tilde{\Delta}a}$ parameterize the differential
relationships between the constraints, their explicit expressions can be found
in subsection (\ref{teorema_ec_ev_vin_1_a}), eq. (\ref{Eq_vinc_de_los_vinc_1
). It is important to mention that to reach those expressions we need the
integrability condition (\ref{eq_int_LE_2}) which appears in the next
subsection. From now on we will refer to (\ref{Eq_vinc_de_los_vinc_1})\ as the
\textit{constraints of the constraints.}
Using the identity (\ref{eq_Kh_hC_1}), $M_{A}^{\tilde{\Delta}a}$ can be
rewritten a
\begin{align}
M_{B}^{\tilde{\Delta}a} & =M_{A}^{\tilde{\Delta}a}\left[
\begin{array}
[c]{cc
\mathfrak{N}_{~\alpha}^{A0} & h_{~\Gamma}^{A
\end{array}
\right] \left[
\begin{array}
[c]{c
h_{B}^{\alpha}\\
C_{B}^{\Gamma0
\end{array}
\right] \nonumber\\
& =M_{A}^{\tilde{\Delta}a}h_{~\Gamma}^{A}C_{B}^{\Gamma0}, \label{Eq_M_h_C0_1
\end{align}
where the equation (\ref{eq_M_K0_1}) has been used. So, by redefining
\begin{equation}
M_{\Gamma}^{\tilde{\Delta}a}:=M_{A}^{\tilde{\Delta}a}h_{~\Gamma}^{A}
\label{Eq_M_n_1
\end{equation}
and using the eq. (\ref{eq_CK_1}) we derive an equation for $M_{\Gamma
}^{\tilde{\Delta}a}
\begin{equation}
M_{\Gamma}^{\tilde{\Delta}(a}C_{A}^{\left\vert \Gamma0\right\vert
}\mathfrak{N}_{~\alpha}^{\left\vert A\right\vert b)}=0. \label{eq_M_C0_K_1
\end{equation}
There is an isomorphism between $M_{\Gamma}^{\tilde{\Delta}a}$ and
$M_{B}^{\tilde{\Delta}a}$, in the sense that one of them completely define the
other. For each $M_{\Gamma}^{\tilde{\Delta}a}$ satisfying the eq.
(\ref{eq_M_C0_K_1}), $M_{B}^{\tilde{\Delta}a}$ can be defined by the equation
(\ref{Eq_M_h_C0_1}), and for each $M_{B}^{\tilde{\Delta}a}$ satisfying
(\ref{eq_M_K_1}) and (\ref{eq_M_K0_1}),\ \ $M_{\Gamma}^{\tilde{\Delta}a}$
\ can be defined by $M_{A}^{\tilde{\Delta}a}h_{~\Gamma}^{A}$. In addition,
considering the following expression,
\[
M_{A}^{\tilde{\Delta}a}\left[
\begin{array}
[c]{cc
\mathfrak{N}_{~\alpha}^{A0} & h_{~\Gamma}^{A
\end{array}
\right] =\left[
\begin{array}
[c]{cc
0 & M_{\Gamma}^{\tilde{\Delta}a
\end{array}
\right] ,
\]
where we have used that $M_{A}^{\tilde{\Delta}a}\mathfrak{N}_{~\alpha}^{A0
=0$; and recalling that $\left[
\begin{array}
[c]{cc
\mathfrak{N}_{~\alpha}^{A0} & h_{~\Gamma}^{A
\end{array}
\right] $ is invertible, the last expression means that any set of
$M_{\Gamma}^{\tilde{\Delta}a}$'s is linearly independent if and only if the
associated set of $M_{B}^{\tilde{\Delta}a}$'s is linearly independent too.
From now on, we will regard $M_{B}^{\tilde{\Delta}a}$ and $M_{\Gamma
^{\tilde{\Delta}a}$ as equivalent to simplify the discussion.
The $M_{\Gamma}^{\tilde{\Delta}0}$ has another interesting property. By
employing eq. (\ref{eq_M_C0_K_1}), we reach to the following lemma.
\begin{lemma}
\label{Lema_M0} $M_{\Gamma}^{\tilde{\Delta}0}=0$ (or $M_{A}^{\tilde{\Delta
0}=0$).
\end{lemma}
\begin{proof}
Considering the $0,b$ components of eq. (\ref{eq_M_C0_K_1}) and the equation
$C_{A}^{\Gamma0}\mathfrak{N}_{~\alpha}^{A0}=0$, we conclude
\[
M_{\Gamma}^{\tilde{\Delta}0}C_{A}^{\Gamma0}\mathfrak{N}_{~\alpha
^{Ab}=-M_{\Gamma}^{\tilde{\Delta}b}C_{A}^{\Gamma0}\mathfrak{N}_{~\alpha
^{A0}=0.
\]
Since we are working with systems that have not algebraic constraints, there
is not $X_{A}$ such that eq. (\ref{eq_X_K_1}) holds, so it should be valid
\[
M_{\Gamma}^{\tilde{\Delta}0}C_{A}^{\Gamma0}=0.
\]
On the other hand, the $C_{A}^{\Gamma0}$ has only trivial kernel in the
$\Gamma$ index, thus we finally conclude $M_{\Gamma}^{\tilde{\Delta}0}=0$.
\end{proof}
Notice that this result allows us to redefine the Geroch's fields
$C_{A}^{\Gamma a}$ in the following wa
\[
\tilde{C}_{A}^{\Gamma a}=C_{A}^{\Gamma a}+N_{\tilde{\Delta}}^{\Gamma
M_{A}^{\tilde{\Delta}a},
\]
leaving the $C_{A}^{\Gamma0}$ component unchanged and keeping the same
expressions for the constraints eq. (\ref{eq_def_constr_1}). Here, the
$N_{\tilde{\Delta}}^{\Gamma}$ fields can be freely chosen and the reader can
see that the Geroch fields have significant freedom in their definition. This
freedom will also appear in the evolution equations of the constraints.
Finally, we conclude the discussion of the covariance in the Geroch fields. So
far, it has been requested that $C_{A}^{\Delta0}$ has maximal rank and has
shown that $M_{A}^{\tilde{\Gamma}0}=0$, but these are not covariant conditions
since the $n_{a}$ direction is privileged among others in these two objects.
In general, one can recover the covariance, thinking that we have split the
real tensors (the ones which transform good under change of coordinates) into
two parts, the $C_{A}^{\Delta a}$ and $M_{B}^{\tilde{\Delta}a}$. We will not
pay attention to this non-covariant splitting in the future and we will call
fields or tensors to $C_{A}^{\Delta a}$ and $M_{B}^{\tilde{\Delta}a}$, even if
they do not transform as real ones.
We have said nothing about how many derivatives admit the objects considering
until here, so, we will assume that \ $\mathfrak{N}_{~\alpha}^{Ac}$, $J^{A}$,
$C_{A}^{\Delta a}$, $M_{B}^{\tilde{\Delta}a}$, $h_{B}^{\alpha}$ and
$h_{~\Gamma}^{A}$ admit at least one derivative in any direction for the next
conditions hold.
\subsection{Integrability conditions}
At this point, the Geroch fields $C_{A}^{\Gamma a}$ have been defined but only
the $C_{A}^{\Gamma0}$ component has been used to define the constraints. We
introduce now a set of \textit{off-shell identities} which include all the
components of $C_{A}^{\Gamma a}$ and $M_{\Gamma}^{\tilde{\Delta}d}$, they ar
\begin{equation}
\nabla_{d}\left( C_{A}^{\Gamma d}E^{A}\right) =L_{1A}^{\Gamma}\left(
x,\phi,\nabla\phi\right) E^{A}\left( x,\phi,\nabla\phi\right) ,
\label{eq_int_LE_1
\end{equation
\begin{equation}
\nabla_{d}\left( M_{\Gamma}^{\tilde{\Delta}d}C_{A}^{\Gamma0}E^{A}\right)
=L_{2A}^{\tilde{\Delta}}\left( x,\phi,\nabla\phi\right) E^{A}\left(
x,\phi,\nabla\phi\right) . \label{eq_int_LE_2
\end{equation}
We call \textit{integrability conditions }to these identities and we assume
they hold for the rest of this work. The motivation behind their inclusion
will be explained in the following paragraphs.
We begin by noticing, the integrability conditions specify that the
divergences of certain combinations of the equations of the system are
proportional to the system. These proportionality factors may depend on
$\left( x,\phi,\nabla\phi\right) $. \ In the case of the Einstein equations,
the eq. (\ref{eq_int_LE_1}) follows from the conservation law of the Einstein
tensor $\nabla_{a}G_{~b}^{a}=0$ (or from the 2nd Bianchi identity). This is
obtained by the Noether theorem as a conserved quantity from the Lagrangian
coordinate transformation invariance (see \cite{Wald:1984rg}). The eq.
(\ref{eq_int_LE_2}) appears when the Einstein equations are cast into
first-order form. In the case of the Maxwell equations (or Yang-Mills
equations), eq. (\ref{eq_int_LE_1}) follows from the charge conservation law,
this is obtained by the Noether theorem and associated to the Lagrangian gauge
invariance symmetry (see \cite{Brading:2002lzh}).This system does not have an
identity as (\ref{eq_int_LE_2}). In the case of the wave equations, the
identities (\ref{eq_int_LE_1}) and (\ref{eq_int_LE_2}) appear when the system
is reduced into first order in derivative adding extra constraints. The study
of the constraints propagation of the Maxwell and wave equations are discussed
in section \ref{Examples}, we there give explicit expressions for these identities.
It is likely that at least the identity (\ref{eq_int_LE_1}) always comes from
the Noether theorem (in there off-shell version) as a conserved quantity and
associated to some "gauge" symmetry of the system. As we will show in the next
section, these identities are needed to show the constraints conservation
since they are the evolutions equations of the constraints. Therefore, in
general, we beleive that the constraints conservation will be associated to
some symmetry of the system.
From the PDE point of view, it has sense that the Geroch fields appear in the
expressions (\ref{eq_int_LE_1}) and (\ref{eq_int_LE_2}). Considering only the
left-hand side of (\ref{eq_int_LE_1}) and replacing $E^{A}$ by its definition
(eq. (\ref{eq_sys_1})), we find tha
\begin{align}
& \nabla_{d}\left( C_{A}^{\Gamma d}E^{A}\right) \nonumber\\
& =\nabla_{d}\left( C_{A}^{\Gamma d}\left( \mathfrak{N}_{~\alpha
^{Aa}\nabla_{a}\phi^{\alpha}-J^{A}\right) \right) ,\\
& =\nabla_{d}\left( C_{A}^{\Gamma d}\mathfrak{N}_{~\alpha}^{Aa}\right)
\nabla_{a}\phi^{\alpha}+\left( C_{A}^{\Gamma\lbrack d}\mathfrak{N}_{~\alpha
}^{\left\vert A\right\vert a]}\right) \nabla_{d}\nabla_{a}\phi^{\alpha
}-\nabla_{d}\left( C_{A}^{\Gamma d}J^{A}\right) ,\\
& =\nabla_{d}\left( C_{A}^{\Gamma d}\mathfrak{N}_{~\alpha}^{Aa}\right)
\nabla_{a}\phi^{\alpha}+\frac{1}{2}\left( C_{A}^{\Gamma\lbrack d
\mathfrak{N}_{~\alpha}^{\left\vert A\right\vert a]}\right) R_{~\beta
da}^{\alpha}\phi^{\beta}-\nabla_{d}\left( C_{A}^{\Gamma d}J^{A}\right) .
\label{eq_N_CE_1
\end{align}
Where we have used the chain rule, equation (\ref{eq_CK_1}) and the following
definition of the curvature tensor
\[
\nabla_{\lbrack d}\nabla_{a]}\phi^{\alpha}=\frac{1}{2}R_{~\beta da}^{\alpha
}\phi^{\beta}.
\]
\sloppy
We note that the resulting expression (\ref{eq_N_CE_1}) does not include
second derivatives of the fields $\phi^{\beta}$, so it is reasonable that it
can be factored as a product of expressions in first derivatives of
$\phi^{\beta}$, i.e. as $L_{1A}^{\Gamma}\left( x,\phi,\nabla\phi\right)
E^{A}\left( x,\phi,\nabla\phi\right) $. The case for (\ref{eq_int_LE_2}) is
analogous, this is
\begin{align*}
& \nabla_{d}\left( M_{\Gamma}^{\tilde{\Delta}d}C_{A}^{\Gamma0}E^{A}\right)
,\\
& =\nabla_{d}\left( M_{\Gamma}^{\tilde{\Delta}d}C_{A}^{\Gamma0}\left(
\mathfrak{N}_{~\alpha}^{Aa}\nabla_{a}\phi^{\alpha}-J^{A}\right) \right) ,\\
& =\nabla_{d}\left( M_{\Gamma}^{\tilde{\Delta}d}C_{A}^{\Gamma0
\mathfrak{N}_{~\alpha}^{Aa}\right) \nabla_{a}\phi^{\alpha}+\frac{1}{2}\left(
M_{\Gamma}^{\tilde{\Delta}[d}C_{A}^{\left\vert \Gamma0\right\vert
}\mathfrak{N}_{~\alpha}^{\left\vert A\right\vert a]}\right) R_{~\beta
da}^{\alpha}\phi^{\beta},\\
& -\nabla_{d}\left( M_{\Gamma}^{\tilde{\Delta}[d}C_{A}^{\left\vert
\Gamma0\right\vert }J^{A}\right) ,\\
& =\nabla_{i}\left( M_{\Gamma}^{\tilde{\Delta}i}C_{A}^{\Gamma0
\mathfrak{N}_{~\alpha}^{Aj}\right) \nabla_{j}\phi^{\alpha}+\frac{1}{2}\left(
M_{\Gamma}^{\tilde{\Delta}[i}C_{A}^{\left\vert \Gamma0\right\vert
}\mathfrak{N}_{~\alpha}^{\left\vert A\right\vert j]}\right) R_{~\beta
ij}^{\alpha}\phi^{\beta},\\
& -\nabla_{i}\left( M_{\Gamma}^{\tilde{\Delta}[i}C_{A}^{\left\vert
\Gamma0\right\vert }J^{A}\right) ,
\end{align*}
with $i,j=1,...,n$ and where the eq. $M_{\Gamma}^{\tilde{\Delta}0}=0$ and
$C_{A}^{\Gamma0}\mathfrak{N}_{~\alpha}^{A0}=0$ have been used. \ We obtain
again a set of first derivative equations for the fields $\phi^{\beta}$.
We shall give a final justification of why equations (\ref{eq_int_LE_1}) and
(\ref{eq_int_LE_2}) make sense in the next subsection.
\section{First main Theorem: Subsidiary System (SS) \label{Teorema_1}}
As we have shown the set of equations $E^{B}$ can be decomposed into evolution
$e^{\alpha}\left( \phi\right) =h_{B}^{\alpha}E^{B}=0$ and constraints
$\psi^{\Delta}\left( \phi\right) =C_{B}^{\Delta0}E^{B}=0$ equations. One
mechanism to find solutions for these equations is the free-evolution
approach. That is, we give an initial data $\left. \phi\right\vert
_{\Sigma_{0}}=\phi_{0}$ satisfying the constraints $\left. \psi^{\Gamma
}\left( \phi_{0}\right) \right\vert _{\Sigma_{0}}=0$, and use the evolution
equations $e^{\alpha}\left( \phi\right) =0$ to find the solutions of the
system over the future hypersurfaces $\Sigma_{t}$ with $t>0$. However, these
found solutions may not satisfy the constraints for $t>0$, and therefore may
not be solutions of the complete system. For this reason, we need equations
that tell us how these constraints evolve for $t>0$ when $e^{\alpha}\left(
\phi\right) =0$. With these equations, we can determine if the constraint are
preserved or not. The standard mechanism to obtain these constraint evolution
equations is: take a time derivative of $\psi^{\Gamma}=n_{a}C_{A}^{\Gamma
a}\mathfrak{N}_{~\alpha}^{Ac}\eta_{c}^{b}\nabla_{b}\phi^{\alpha}-J^{A}$, which
translates into time derivatives of $\phi^{\alpha}$, plus other terms; use the
evolution equations $e^{\alpha}\left( \phi\right) =0$ to eliminate
$\partial_{t}\phi^{\alpha}$ from these expressions and (hopefully and after
many calculations) rearrange the resulting terms into expressions that only
include spatial derivatives of $\psi^{\Gamma}$ and lower order terms
proportional to $\psi^{\Gamma}$. As we shall see in the proof of the following
theorem, counting with the identities (\ref{eq_int_LE_1}) and
(\ref{eq_int_LE_2}) is equivalent to the mentioned process (since they include
the time derivatives of $\psi^{\Gamma}$). Actually, if these conditions are
not satisfied, the system has extra constraints that should be added to the
system and whose conservation has to be studied.
The following theorem, called \textit{Subsidiary System,} shows the explicit
form of the subsidiary system (SS) in the quasi-linear case.
\begin{theorem}
\label{teorema_ec_ev_vin_1}Considere the system (\ref{eq_sys_1}) such that the
following conditions hold:
(i) The principal symbol satisfies assumption \ref{cod_K0_rank_max_b}.
(ii) All constraints of the systems come from Geroch fields $C_{A}^{\Gamma a}$
as in eq. (\ref{eq_def_constr_1}) and these $C_{A}^{\Gamma a}$ satisfy
assumption \ref{cod_C0_rank_max_b}.
(iii) The system can admit (or not) extra Geroch fields $M_{A}^{\tilde{\Delta
}a}$ satisfying assumption \ref{As_M_1_b}.
(iv) The integrability conditions (\ref{eq_int_LE_1}) and (\ref{eq_int_LE_2})
are satisfied.
Then, the following off-shell identity is satisfie
\begin{align}
& \nabla_{0}\psi^{\Gamma}+\left( C_{A}^{\Gamma i}h_{~\Delta}^{A
+N_{\tilde{\Delta}}^{\Gamma}M_{\Delta}^{\tilde{\Delta}i}\right) \nabla
_{i}\psi^{\Delta}\nonumber\\
& =\left( L_{1A}^{\Gamma}\mathfrak{N}_{~\alpha}^{A0}-\nabla_{d}\left(
C_{A}^{\Gamma d}\mathfrak{N}_{~\alpha}^{A0}\right) \right) e^{\alpha
-C_{A}^{\Gamma d}\mathfrak{N}_{~\alpha}^{A0}\nabla_{d}e^{\alpha}\nonumber\\
& +\left( L_{1A}^{\Gamma}h_{~\Delta}^{A}-\nabla_{d}\left( C_{A}^{\Gamma
d}h_{~\Delta}^{A}\right) +N_{\tilde{\Delta}}^{\Gamma}\left( L_{2A
^{\tilde{\Delta}}h_{~\Delta}^{A}-\nabla_{i}\left( M_{\Delta}^{\tilde{\Delta
}i}\right) \right) \right) \psi^{\Delta}
\label{Eq_ide_ev_off_shell_const_1_a
\end{align}
Here the tensors $N_{\tilde{\Delta}}^{\Gamma}$ can be freely chosen. If the
system does not admit Geroch fields $M_{A}^{\tilde{\Delta}a}$, then
$N_{\tilde{\Delta}}^{\Gamma}=0$, $L_{2A}^{\tilde{\Delta}}=0$. In this case,
the SS is unique.
In the on-shell case i.e. $e^{\alpha}\left( \phi\right) =0$, the above
identity reduces to the subsidiary system
\begin{align}
& \nabla_{0}\psi^{\Gamma}+\left( C_{A}^{\Gamma i}h_{~\Delta}^{A
+N_{\tilde{\Delta}}^{\Gamma}M_{\Delta}^{\tilde{\Delta}i}\right) \nabla
_{i}\psi^{\Delta}\nonumber\\
& =\left( L_{1A}^{\Gamma}h_{~\Delta}^{A}-\nabla_{d}\left( C_{A}^{\Gamma
d}h_{~\Delta}^{A}\right) +N_{\tilde{\Delta}}^{\Gamma}\left( L_{2A
^{\tilde{\Delta}}h_{~\Delta}^{A}-\nabla_{i}\left( M_{\Delta}^{\tilde{\Delta
}i}\right) \right) \right) \psi^{\Delta} \label{eq_ev_on_shell_const_1_b
\end{align}
where $i=1,...,n.$
\end{theorem}
We notice that if $\left\vert \psi^{\Gamma}\right. _{\Sigma_{0}}=0,$ then
$\psi^{\Gamma}=0$ is a solution of equation (\ref{eq_ev_on_shell_const_1_b}).
If in addition, the equation (\ref{eq_ev_on_shell_const_1_b}) is well-posed,
the solution $\psi^{\Gamma}=0$ is unique and continue in the initial data.
It is important to emphasize that the set of constraint evolution equations is
not unique when the $M_{\Delta}^{\tilde{\Delta}i}$ fields exist. \ In other
words, constraints $\psi^{\Gamma}$ are solutions of a family of differential
equations (\ref{eq_ev_on_shell_const_1_b}). This family is parametrized by the
field $N_{\tilde{\Delta}}^{\Gamma}$ (which can be freely chosen), producing
the non-uniqueness. An example of this class of systems is the wave equation,
see section\ \ref{S_wave_equation}.
On the other hand, the well-posedness of the SS depends on the explicit form
of its principal symbol
\begin{equation}
B_{~\Delta}^{\Gamma i}:=C_{A}^{\Gamma i}h_{~\Delta}^{A}+N_{\tilde{\Delta
}^{\Gamma}M_{\Delta}^{\tilde{\Delta}i}. \label{Eq_simb_vinc_1
\end{equation}
Since, this symbol includes the field $h_{~\Delta}^{A}$, which is completely
determined by the choice of $h_{B}^{\alpha}$ (see eq. (\ref{eq_Kh_hC_1
))\footnote{Notice that the reverse is true too, $h_{B}^{\alpha}$ is
completely defined by choice of $h_{~\Delta}^{A}$.}, and the free field
$N_{\tilde{\Delta}}^{\Gamma}$, the well-posedness of the subsidiary systems is
given by $h_{B}^{\alpha}$ and $N_{\tilde{\Delta}}^{\Gamma}$. We focus on this
problem in section \ref{teorema_2}, where we provide a theorem on how to
apropriately chose $h_{B}^{\alpha}$ and $N_{\tilde{\Delta}}^{\Gamma}$ to
obtain a well-posed SS.
Finally, when there are no $M_{\Delta}^{\tilde{\Delta}i}$ fields present in
the system, the evolution equations of the constraints are unique, so the
analysis of well-posedness is highly simplified. The Maxwell equations are an
example of this case, see section \ref{S_Maxwell_eq}.
We will discuss the hyperbolicity of the SS in the next sections, focusing on
the constant-coefficient systems where we can give a closed answer.
\subsection{Invariance of the choice of $e^{\alpha}$ in the subsidiary system
\label{invariance_e}}
Given a particular physical theory, different reductions $h_{B}^{\alpha}$ and
$\tilde{h}_{B}^{\alpha}$ lead to different choices of the evolution equations
$e^{\alpha}$ and $\tilde{e}^{\alpha}$ (see subsection \ref{Evo}).\ Therefore,
to study the conservation of the constraints, the SS should be calculated for
each choice of $e^{\alpha}$ or $\tilde{e}^{\alpha}$. In addition, it might
happen that the SS does not exist\footnote{Meaning that it is not possible to
write a set of PDE's where the variables of the PDE's are the constraints.}
for a particular choice of the evolution equations. \ In our scheme, this
dilemma is solved in a quite elegant way. As we explained before, each
reduction $h_{B}^{\alpha}$ ($\tilde{h}_{B}^{\alpha}$) introduces a unique
field $h_{~\Delta}^{A}$ ($\tilde{h}_{~\Delta}^{A}$) given by the equation
(\ref{eq_Kh_hC_1}). \ This field $h_{~\Delta}^{A}$ ($\tilde{h}_{~\Delta}^{A}$)
appear in the equations of the SS (\ref{eq_ev_on_shell_const_1_b}), showing
that the SS always exists and how the information of the reductions
$h_{B}^{\alpha}$ ($\tilde{h}_{B}^{\alpha}$) is propagated to this equation.
The reason for this simple answer is that equation
(\ref{Eq_ide_ev_off_shell_const_1_a}) follows directly from the integrability
conditions (\ref{eq_int_LE_1}) and (\ref{eq_int_LE_2}), which do not depend on
any reduction $h_{B}^{\alpha}$. As it is shown in the proof of the theorem,
these reductions appear only as a trick to transform $C_{A}^{\Gamma a}E^{A}$
into $\left[
\begin{array}
[c]{cc
C_{A}^{\Gamma d}\mathfrak{N}_{~\alpha}^{A0} & C_{A}^{\Gamma d}h_{~\Delta}^{A
\end{array}
\right] \left[
\begin{array}
[c]{c
e^{\alpha}\\
\psi^{\Delta
\end{array}
\right] $, where the explicit form of $h_{B}^{\alpha}$ plays no role in this
proof, so that, eq. (\ref{Eq_ide_ev_off_shell_const_1_a}) can be obtained for
any $h_{B}^{\alpha}$.
It may happen that in some physical theories the integrability conditions
(\ref{eq_int_LE_1}) and (\ref{eq_int_LE_2}) depend on a specific reduction
$h_{B}^{\alpha}$ and consequently on its corresponding evolution equations
$e^{\alpha}$. In this class of systems, the eqs. (\ref{eq_int_LE_1}) and
(\ref{eq_int_LE_2}) could include 2-order or higher partial derivatives of the
evolution equations. \ This should not modify equation
(\ref{eq_ev_on_shell_const_1_b}) since $e^{\alpha}=0$ in the on-shell case,
but this system would be forced to evolve only with $e^{\alpha}=0$ since any
other choice of the evolution equations $\tilde{e}^{\alpha}$ would not produce
a SS as (\ref{eq_ev_on_shell_const_1_b}). Therefore, the constraint
conservation would not be guaranteed.
\subsection{Proof of Subsidiary System theorem \label{teorema_ec_ev_vin_1_a}}
The proof consists of showing how to go from the sum of (\ref{eq_int_LE_1})
and (\ref{eq_int_LE_2})
\begin{equation}
\nabla_{d}\left( C_{A}^{\Gamma d}E^{A}\right) +N_{\tilde{\Delta}}^{\Gamma
}\nabla_{d}\left( M_{\Gamma}^{\tilde{\Delta}d}C_{A}^{\Gamma0}E^{A}\right)
=L_{1A}^{\Gamma}E^{A}+N_{\tilde{\Delta}}^{\Gamma}L_{2A}^{\tilde{\Delta}}E^{A},
\label{Eq_com_conds_int_1
\end{equation}
to (\ref{Eq_ide_ev_off_shell_const_1_a}), where $N_{\tilde{\Delta}}^{\Gamma}$
can be freely chosen.
Using (\ref{Eq_e_y_vin_Mat_E_1}) (valid from the assumptions
\ref{cod_K0_rank_max_b} and \ref{cod_C0_rank_max_b}) and (\ref{eq_Kh_hC_1})
the system $E^{A}$ can be written as
\begin{align}
E^{A} & =\left[
\begin{array}
[c]{cc
\mathfrak{N}_{~\alpha}^{A0} & h_{~\Delta}^{A
\end{array}
\right] \left[
\begin{array}
[c]{c
h_{B}^{\alpha}\\
C_{B}^{\Delta0
\end{array}
\right] E^{B},\nonumber\\
& =\left[
\begin{array}
[c]{cc
\mathfrak{N}_{~\alpha}^{A0} & h_{~\Delta}^{A
\end{array}
\right] \left[
\begin{array}
[c]{c
e^{\alpha}\\
\psi^{\Delta
\end{array}
\right] . \label{Eq_E_e_phi_1
\end{align}
By replacing the latter equation in each of the terms of
(\ref{Eq_com_conds_int_1}), we obtai
\begin{align}
& \nabla_{d}\left( C_{A}^{\Gamma d}E^{A}\right) \nonumber\\
& =\nabla_{d}\left( \left[
\begin{array}
[c]{cc
C_{A}^{\Gamma d}\mathfrak{N}_{~\alpha}^{A0} & C_{A}^{\Gamma d}h_{~\Delta}^{A
\end{array}
\right] \left[
\begin{array}
[c]{c
e^{\alpha}\\
\psi^{\Delta
\end{array}
\right] \right) ,\nonumber\\
& =\nabla_{d}\left( C_{A}^{\Gamma d}\mathfrak{N}_{~\alpha}^{A0}\right)
e^{\alpha}+\nabla_{d}\left( C_{A}^{\Gamma d}h_{~\Delta}^{A}\right)
\psi^{\Delta}+C_{A}^{\Gamma d}\mathfrak{N}_{~\alpha}^{A0}\nabla_{d}e^{\alpha
}+C_{A}^{\Gamma d}h_{~\Delta}^{A}\nabla_{d}\psi^{\Delta},\nonumber\\
& =\nabla_{d}\left( C_{A}^{\Gamma d}\mathfrak{N}_{~\alpha}^{A0}\right)
e^{\alpha}+\nabla_{d}\left( C_{A}^{\Gamma d}h_{~\Delta}^{A}\right)
\psi^{\Delta}+C_{A}^{\Gamma d}\mathfrak{N}_{~\alpha}^{A0}\nabla_{d}e^{\alpha
}\label{Eq_nabla_C_E__e_phi_1}\\
& +\nabla_{0}\psi^{\Gamma}+C_{A}^{\Gamma i}h_{~\Delta}^{A}\nabla_{d
\psi^{\Delta}.\nonumber
\end{align}
Where we used $C_{A}^{\Gamma0}h_{~\Delta}^{A}=\delta_{\Delta}^{\Gamma}$ and
$i=1,...,n$ in the last equation. On the other hand
\begin{align}
L_{1A}^{\Gamma}E^{A} & =L_{1A}^{\Gamma}\left[
\begin{array}
[c]{cc
\mathfrak{N}_{~\alpha}^{A0} & h_{~\Delta}^{A
\end{array}
\right] \left[
\begin{array}
[c]{c
e^{\alpha}\\
\psi^{\Delta
\end{array}
\right] ,\nonumber\\
& =L_{1A}^{\Gamma}\mathfrak{N}_{~\alpha}^{A0}e^{\alpha}+L_{1A}^{\Gamma
}h_{~\Delta}^{A}\psi^{\Delta}. \label{Eq_L_e_phi_1
\end{align}
We conclude that (\ref{eq_int_LE_1}) can be written as
\begin{align*}
\nabla_{d}\left( C_{A}^{\Gamma d}E^{A}\right) & =L_{1A}^{\Gamma}E^{A}\\
\nabla_{d}\left( C_{A}^{\Gamma d}\mathfrak{N}_{~\alpha}^{A0}\right)
e^{\alpha}+\nabla_{d}\left( C_{A}^{\Gamma d}h_{~\Delta}^{A}\right)
\psi^{\Delta}+ & \\
+C_{A}^{\Gamma d}\mathfrak{N}_{~\alpha}^{A0}\nabla_{d}e^{\alpha}+\nabla
_{0}\psi^{\Gamma}+C_{A}^{\Gamma i}h_{~\Delta}^{A}\nabla_{d}\psi^{\Delta} &
=L_{1A}^{\Gamma}\mathfrak{N}_{~\alpha}^{A0}e^{\alpha}+L_{1A}^{\Gamma
}h_{~\Delta}^{A}\psi^{\Delta}.
\end{align*}
Considering now the divergences which involve $M_{\Gamma}^{\tilde{\Delta}d}$,
we obtain:
\begin{align}
\nabla_{d}\left( M_{\Gamma}^{\tilde{\Delta}d}C_{A}^{\Gamma0}E^{A}\right) &
=\nabla_{d}\left( M_{\Gamma}^{\tilde{\Delta}d}\psi^{\Gamma}\right)
,\nonumber\\
& =\nabla_{d}\left( M_{\Gamma}^{\tilde{\Delta}d}\right) \psi^{\Gamma
}+M_{\Gamma}^{\tilde{\Delta}d}\nabla_{d}\left( \psi^{\Gamma}\right)
,\nonumber\\
& =\nabla_{i}\left( M_{\Gamma}^{\tilde{\Delta}i}\right) \psi^{\Gamma
}+M_{\Gamma}^{\tilde{\Delta}i}\nabla_{i}\left( \psi^{\Gamma}\right) .
\label{Eq_vinc_vinc_1
\end{align}
Where in the first line we used the definition of $\psi^{\Gamma}=C_{A
^{\Gamma0}E^{A}$ and in the third line that $M_{\Gamma}^{\tilde{\Delta}0}$
vanishes (see Lemma \ref{Lema_M0}). The expression for $L_{2A}^{\tilde{\Delta
}}E^{A}$ is
\[
L_{2A}^{\tilde{\Delta}}E^{A}=L_{2A}^{\tilde{\Delta}}\mathfrak{N}_{~\alpha
}^{A0}e^{\alpha}+L_{2A}^{\tilde{\Delta}}h_{~\Delta}^{A}\psi^{\Delta},
\]
so the expression (\ref{eq_int_LE_2}) can be rewritten as
\begin{align*}
\nabla_{d}\left( M_{\Gamma}^{\tilde{\Delta}d}C_{A}^{\Gamma0}E^{A}\right) &
=L_{2A}^{\tilde{\Delta}}E^{A}\\
\nabla_{i}\left( M_{\Gamma}^{\tilde{\Delta}i}\right) \psi^{\Gamma
+M_{\Gamma}^{\tilde{\Delta}i}\nabla_{i}\left( \psi^{\Gamma}\right) &
=L_{2A}^{\tilde{\Delta}}\mathfrak{N}_{~\alpha}^{A0}e^{\alpha}+L_{2A
^{\tilde{\Delta}}h_{~\Delta}^{A}\psi^{\Delta}\text{.
\end{align*}
Recalling that this expression is an identity that holds for all $\phi
^{\alpha}$, it should not be possible that the right hand side of this
equation to contains time derivatives (on $e^{\alpha}$) while the left hand
side does not. This means that,
\[
L_{2A}^{\tilde{\Delta}}\mathfrak{N}_{~\alpha}^{A0}=0,
\]
hence
\begin{equation}
\nabla_{i}\left( M_{\Gamma}^{\tilde{\Delta}i}\right) \psi^{\Gamma
+M_{\Gamma}^{\tilde{\Delta}i}\nabla_{i}\left( \psi^{\Gamma}\right)
=L_{2A}^{\tilde{\Delta}}h_{~\Delta}^{A}\psi^{\Delta}.
\label{Eq_vinc_de_los_vinc_1
\end{equation}
Finally, replacing these results in eq. (\ref{Eq_com_conds_int_1}) gives
(\ref{Eq_ide_ev_off_shell_const_1_a}), which concludes the proof.
As a final comment, in section \ref{Sec_Vin_M} we said that the fields
$M_{\Gamma}^{\tilde{\Delta}d}$ parameterize the constraints of the constraint.
The explicit expression for these differential relationships between the
constraints is given by the last equation (\ref{Eq_vinc_de_los_vinc_1}).
\section{Constant coefficient and Strong Hyperbolicity (SH)
\label{Sec_Const_coeff_1}}
In this section we introduce a brief summary of paper \cite{Abalos:2018uwg},
we present the definitions and the main result of that work about strong
hyperbolicity (SH) for systems with constraints. This main result gives the
necessary and sufficient conditions under which the system
\begin{equation}
E^{A}:=\mathfrak{N}_{~\alpha}^{Aa}\partial_{a}\phi^{\alpha},
\label{Eq_sis_1_coef_cte_1
\end{equation}
has a strong hyperbolic set of evolution equations.
As in \cite{Abalos:2018uwg}, we do not consider the quasi-linear case
(\ref{eq_sys_1}) and focus on the constant coefficient case, where
$\mathfrak{N}_{~\alpha}^{Aa}$ does not depend on $x,\phi$ (i.e. $\nabla
_{c}\mathfrak{N}_{~\alpha}^{Aa}=0$). This simplification leads to a closed
pseudo-differential theorem about SH for the evolution equations (theorem
\ref{Theor_FyL_2}) and will allow us, in the next section, to derive a theorem
about the strong hyperbolicity of the SS (theorem \ref{Theorem_coef_const_2}).
Note also that the covariant derivatives $\nabla_{a}$ have been changed to
partial derivatives $\partial_{a}$ and that the lower order terms have been
suppressed because the SH is not affected by them.
As before, the initial value problem\ consists on solving
\begin{equation}
E^{A}\left( \phi\right) =0\text{ with initial data }\left. \phi^{\alpha
}\right\vert _{\Sigma_{0}}=f^{\alpha}\left( x\right) .
\label{Eq_ini_val_E_f_1
\end{equation}
To discuss its SH, we convert the problem to its Fourier space and present a
pseudo-differential analysis. Applying to (\ref{Eq_ini_val_E_f_1}) a Fourier
transformation on the spatial variables $x^{i}$, with $i=1,...,n$,\ we obtain
\begin{equation}
\tilde{E}^{A}:=\mathfrak{N}_{~\alpha}^{A0}\partial_{t}\tilde{\phi}^{\alpha
}+i\mathfrak{N}_{~\alpha}^{Ai}k_{i}\tilde{\phi}^{\alpha}=0,
\label{Eq_sis_1_pseudo_1
\end{equation}
with
\[
\left. \tilde{\phi}^{\alpha}\right\vert _{\Sigma_{0}}=\tilde{f}^{\alpha
}\left( x\right) .
\]
Where $k_{a}$ is the \textit{wave vector} such that $k_{a}t^{a}=0$ (i.e.
$k_{0}=0$) and $k_{a}\eta_{b}^{a}=k_{b}$ (i.e. $k_{a}=\left( 0,k_{i}\right)
$). \
Now we introduce the reduction $h_{~A}^{\beta}\left( k_{i}\right) $
(satisfying eq. (\ref{eq_hK0_1})) which may depend on the wave vector $k_{a}$
or not. \ Applying $h_{~A}^{\beta}\left( k_{i}\right) $\ to the equation
(\ref{Eq_sis_1_pseudo_1}), we obtain a set of evolution equations for
$\tilde{\phi}^{\beta}$
\begin{equation}
\tilde{e}^{\beta}=\partial_{t}\tilde{\phi}^{\beta}+ih_{~A}^{\beta}\left(
k_{i}\right) \mathfrak{N}_{~\alpha}^{Ai}k_{i}\tilde{\phi}^{\alpha}=0.
\label{Eq_evol_pseudo_1
\end{equation}
As before, the reduction aim is to combine the constraint and time derivative
equations, to produce systems of evolution equations for each field
$\tilde{\phi}^{\beta}$. The main difference is that now these evolution
equations are pseudo-differential and their solutions must be anti-transformed
to obtain solutions of the original system.
The set of evolution equations has to be well-posed to be predictive. This is
a property that depends on the choice of $h_{~A}^{\beta}$ since different
reductions may lead to ill-posed or well-posed systems. \ In particular, we
are considering a sub-class within the well-posed equations, namely the
strongly hyperbolic ones. This leads us to introduce the following definition.
\begin{definition}
\label{def_hyp_fuerte_1}Consider $n_{a}=\nabla_{a}t$ such that the assumption
\ref{cod_K0_rank_max_b} holds, we say that the system
(\ref{Eq_sis_1_coef_cte_1}) is strongly hyperbolic if there exists at least
one reduction $h_{~A}^{\beta}\left( k\right) $ satisfying (\ref{eq_hK0_1}),
such that for all $k_{i}$, with $\left\vert k\right\vert =1$, the principal
symbol of the evolution equations $A_{~\alpha}^{\beta i}k_{i}:=h_{~A}^{\beta
}\mathfrak{N}_{~\alpha}^{Ai}k_{i}$ is uniformly diagonalizable with real
eigenvalues. Namely, for all $k_{i}$ with $\left\vert k\right\vert =1,$ there
exists $\left( T\left( k\right) \right) _{~\gamma}^{\beta}$ such that
$h_{~A}^{\beta}\mathfrak{N}_{~\alpha}^{Ai}k_{i}=$ $\left( T\left( k\right)
\right) _{~\gamma}^{\beta}\Lambda_{~\theta}^{\gamma}\left( T^{-1}\left(
k\right) \right) _{~\alpha}^{\theta}$ with $\Lambda_{~\theta}^{\gamma}$
diagonal and real; and the diagonalization is uniform, which means that there
exists a constant $C>0$ such that
\begin{equation}
\left\vert T\left( k\right) \right\vert +\left\vert \left( T\left(
k\right) \right) ^{-1}\right\vert <C. \label{Eq_T_T_C_1
\end{equation}
\end{definition}
The norms $\left\vert \cdot\right\vert $, used in $\left\vert k\right\vert =1$
and in the eq. (\ref{Eq_T_T_C_1}), can be any $k_{a}$ independent, positive
definite norms. We will assume for the following definitions and theorems that
the wave vector $k_{i}$ is \textit{normalized to} $\left\vert k\right\vert =1$.
When the reductions $h_{~A}^{\beta}$ satisfy the definition, we call them
\textit{hyperbolizations.} Note that when the system has no constraints, the
reduction $h_{~A}^{\beta}$ is unique and defined by (\ref{eq_hK0_1}).
In the literature, the above definition is presented as a theorem and the
original definition of strong hyperbolicity is another one, but since this
section is only an introduction to the topic, we condense the discussion and
present it as a definition. For more details about the theory we suggest
\cite{sarbach2012continuum} and the reference therein.
We introduce now a set of definitions to conclude with the definition of
canonical angles and after it, the theorem \ref{Theor_FyL_2}.
We call $\Phi$ and $\Psi$ to the vector fibers that include the vectors
$\delta\phi^{\alpha}$ and $X_{A}$ respectively. These spaces contain the right
and left kernel subspaces of the principal symbol $\mathfrak{N}_{~\alpha
^{Aa}w_{a}$ (for a given $w_{a}$), whose elements satisfy
\begin{align*}
\mathfrak{N}_{~\alpha}^{Aa}w_{a}\delta\phi^{\alpha} & =0,\\
X_{A}\mathfrak{N}_{~\alpha}^{Aa}w_{a} & =0.
\end{align*}
respectively. We will use this notation for any other operator, a vector will
belong to the right (left) kernel when it contracts with the down (up) index
operator and the result vanishes.
We introduce the set of planes $S_{n_{a}}^
\mathbb{C}
}=\{l_{a}\left( \lambda\right) =-\lambda n_{a}+k_{a}=\left( -\lambda
,k_{i}\right) .$ for all $k_{a}$ not proportional to $n_{a}$, with
$\left\vert k\right\vert =1$ and $\lambda\i
\mathbb{C}
$ $\}$. They are complex planes for each fixed $k_{a}$ and they reduce to
lines when $\lambda\i
\mathbb{R}
$. We call \ $S_{n_{a}}$ to the set of these lines.
We consider the left and right kernel of the principal symbol on these planes
$\mathfrak{N}_{~\alpha}^{Aa}l_{a}\left( \lambda\right) $ with \ $l_{a
\left( \lambda\right) \in S_{n_{a}}^
\mathbb{C}
}$. \ For each $k_{a}$, there exist certain values of $\lambda$ called
generalized eigenvalues $\lambda_{i}\left( k\right) $ with $i\in D_{\left(
k\right) }$, $D_{\left( k\right) }:=\left\{ 1,2,...,q_{\left( k\right)
}\right\} $ and $\lambda_{1}\left( k\right) <...<\lambda_{q_{\left(
k\right) }}\left( k\right) $ such that $\mathfrak{N}_{~\alpha}^{Aa
l_{a}\left( \lambda_{i}\left( k\right) \right) $ has non-trivial right and
left kernel. We call $\Phi_{R}^{\lambda_{i}\left( k\right) }$ and $\Psi
_{L}^{\lambda_{i}\left( k\right) }$ to these subspaces of $\Phi$ and $\Psi$
respectively. Notice that the explicit form of these generalized eigenvalues
and the $q_{\left( k\right) }$ number of them is $k_{i}-$dependent. We call
$d_{\lambda_{i}\left( k\right) }$ to the geometric multiplicity of
$\lambda_{i}\left( k\right) $, thus $\dim\Phi_{R}^{\lambda_{i}\left(
k\right) }=d_{\lambda_{i}\left( k\right) }$.
The left kernel behaves differently since its dimension is larger, $\dim
\Psi_{L}^{\lambda_{i}\left( k\right) }=d_{\lambda_{i}\left( k\right) }+c.$
This difference is due to the fact that for all $k_{i}$ and all $\lambda$, the
dimension of the left kernel of $\mathfrak{N}_{~\alpha}^{Aa}l_{a}\left(
\lambda\right) $ is $c$ \footnote{Recall that $\mathfrak{N}_{~\alpha
^{Aa}l_{a}\left( \lambda\right) $ is an $e\times u$ matrix and $e=u+c$, this
means that $\mathfrak{N}_{~\alpha}^{Aa}l_{a}\left( \lambda\right) $ always
has a left kernel of dimension equal or greater than $c$.
\par
{}}, except when $\lambda=\lambda_{i}\left( k\right) $, where the left
kernel associated to the generalized eigenvalues increases the dimension to
$d_{\lambda_{i}\left( k\right) }+c$. A better explanation will be given in
the proof of theorem SH of the SS.
A necessary condition for the well-posedness of the system is that the
generalized eigenvalues be real. When the system satisfies this condition we
call it hyperbolic,
\begin{definition}
The system (\ref{Eq_sis_1_coef_cte_1}) is called hyperbolic, if there exists a
co-vector $n_{a}=\nabla_{a}t$ such that
(a) $\mathfrak{N}_{~\alpha}^{A0}:=\mathfrak{N}_{~\alpha}^{Aa}n_{a}$ has only
trivial right kernel.
(b) For each $l_{a}\left( \lambda\right) \in S_{n_{a}}^
\mathbb{C}
}$, if $\mathfrak{N}_{~\alpha}^{Aa}l_{a}\left( \lambda\right) $ has
non-trivial right kernel, then $\lambda\i
\mathbb{R}
.$
\end{definition}
Notice that condition (a) is condition \ref{cod_K0_rank_max}.
Now, we introduce over $\Phi$ a positive definite Hermitian form
$G^{\alpha\beta}$. This allows us to define the vector subspace $\Phi
_{L}^{\lambda_{i}\left( k\right) }$, as the subspace obtained by projecting
$\Psi_{L}^{\lambda_{i}\left( k\right) }$ with $\mathfrak{N}_{~\alpha
^{A0}G^{\alpha\beta}$. Since $\Phi_{L}^{\lambda_{i}\left( k\right) }$ and
$\Phi_{R}^{\lambda_{i}\left( k\right) }$ are vector subspaces of $\Phi$ we
can introduce the canonical angles $\theta_{j}^{\lambda_{i}\left( k\right)
}$ between them (see \ \cite{afriat1957orthogonal},
\cite{taslaman2014principal} for an introduction to the topic), where the
index $j$ runs from $1$ to $d_{\lambda_{i}\left( k\right) }$. These angles
measure the "separation distance" of these vector subspaces. As explained
below, the strong hyperbolicity theorem indicates that these distances should
be bounded for all $k_{i}.$
\begin{theorem}
\label{Theor_FyL_2} \cite{Abalos:2018uwg} \ The constant-coefficient system
(\ref{Eq_sis_1_coef_cte_1}) is strongly hyperbolic (admits at least one
hyperbolization) if and only if it is hyperbolic for some direction
$n_{a}=\nabla_{a}t$ and, for all $i\in D_{\left( k\right) }$ and all
normalized $k_{a}$ non-proportional to $n_{a}$, there is a constant maximum
angle $\vartheta<\frac{\pi}{2}$ between the canonical angles of $\Phi
_{L}^{\lambda_{i}\left( k\right) }$ and $\Phi_{R}^{\lambda_{i}\left(
k\right) }$.
This last condition is equivalent to: if $\theta_{j}^{\lambda_{i}\left(
k\right) }$ are the canonical angles between $\Phi_{L}^{\lambda_{i}\left(
k\right) }$ and $\Phi_{R}^{\lambda_{i}\left( k\right) }$, then there exists
$\vartheta<\frac{\pi}{2}$ such tha
\begin{equation}
\cos\theta_{j}^{\lambda_{i}\left( k\right) }\geq\cos\vartheta>0
\label{Eq_cota_ang_can_1
\end{equation}
for all normalized $k_{a}$ non-proportional to $n_{a}$, with $i\in D_{\left(
k\right) }$ and .$j=1,...,d_{\lambda_{i}\left( k\right) }$.
\end{theorem}
From the proof of the theorem, it follows how to build all the
hyperbolizations that the system admits, pseudo-differential or not (those
that do not depend on $k_{i}$). However, for simplicity, the hyperbolization
used in the proof is the one where the eigenvalues, adding by $h_{~A}^{\beta}$
to $A_{~\alpha}^{\beta i}k_{i}=h_{~A}^{\beta}\mathfrak{N}_{~\alpha}^{Ai}k_{i
$, are simple, different from each other and different from the generalized
eigenvalues for all $k_{i}$. \ When condition (\ref{Eq_cota_ang_can_1}) is
satisfied and this hyperbolization is chosen, it trivially guarantees the
definition \ref{def_hyp_fuerte_1} for $A_{~\alpha}^{\beta i}k_{i}$. We will
use this particular hyperbolization to prove our second main theorem.
\section{Second main theorem: Strong hyperbolicity of the SS \label{teorema_2
}
In this section, we continue considering the constant-coefficient case
(\ref{Eq_sis_1_coef_cte_1}), where $\mathfrak{N}_{~\alpha}^{Ab}$ is constant.
This means that $C_{A}^{\Gamma a}$ and $M_{\Delta}^{\tilde{\Delta}i}$ are
constants too since they are defined by the equations $C_{A}^{(a
\mathfrak{N}_{~\alpha}^{\left\vert A\right\vert b)}=0$ and $M_{A
^{\tilde{\Delta}(a}\mathfrak{N}_{~\alpha}^{\left\vert A\right\vert b)}=0$.
This simplification helps us to present the below closed theorem with simple hypotheses.
Theorem \ref{Theor_FyL_2} says nothing about the preservation or
non-preservation of the constraints during the evolution. For answering this
question, we assume valid all the hypotheses of section
\ref{Seccion_Setting_1} and show the sufficient conditions for the strong
hyperbolicity of the evolution equations of the constraints (eq.
(\ref{eq_ev_on_shell_const_1_b})). As already mentioned, these equations
(\ref{eq_ev_on_shell_const_1_b}) may not be unique since $N_{\tilde{\Delta
}^{\Gamma}$ can be freely chosen, i.e. $N_{\tilde{\Delta}}^{\Gamma}$ plays the
role of a reduction. So, we say that the SS is strongly hyperbolic when it is
possible to choose at least one reduction $N_{\tilde{\Delta}}^{\Gamma}$ (or
hyperbolization) such that the set of subsidiary equations
(\ref{eq_ev_on_shell_const_1_b}) is strongly hyperbolic. Following definition
\ref{def_hyp_fuerte_1} and considering eq. (\ref{eq_ev_on_shell_const_1_b}),
this is equivalent to requiring that there exists a reduction $N_{\tilde
{\Delta}}^{\Gamma}\left( k\right) $ such that the principal symbol of the
subsidiary system
\[
B_{~\Delta}^{\Gamma i}k_{i}:=C_{A}^{\Gamma i}h_{~\Delta}^{A}k_{i
+N_{\tilde{\Delta}}^{\Gamma}M_{\Delta}^{\tilde{\Delta}i}k_{i
\]
is uniformly diagonalizable with real eigenvalues. The following second main
theorem, called \textit{strong hyperbolicity of the subsidiary system},
explains under which conditions exists such hyperbolization.
\begin{theorem}
\label{Theorem_coef_const_2}Consider the system of constant coefficients
(\ref{Eq_sis_1_coef_cte_1}). This system admits a hyperbolization
$h_{~A}^{\beta}\left( k\right) $ and has\ at least one strongly hyperbolic
subsidiary system associated to the (strongly hyperbolic) evolution equations
$h_{~A}^{\beta}\left( k\right) \tilde{E}^{A}=0$ if the following conditions hold:
i) The system satisfies the hypotheses of Theorem \ref{Theor_FyL_2}.
ii) All the constraints of the systems come from Geroch fields $C_{A}^{\Gamma
a}$ as in the eq. (\ref{eq_def_constr_1}) and these $C_{A}^{\Gamma a}$ satisfy
the assumption \ref{cod_C0_rank_max_b}.
iii) The system can admit (or not) extra Geroch fields $M_{A}^{\tilde{\Delta
}a}$ which satisfy assumption \ref{As_M_1_b}.
iv) The integrability conditions (\ref{eq_int_LE_1}) and (\ref{eq_int_LE_2}).
are satisfied.
v) For each $k_{i}$ with $\left\vert k\right\vert =1$, the fields
$M_{A}^{\tilde{\Delta}i}k_{i}$ span the left kernel of $C_{A}^{\Gamma
0}\mathfrak{N}_{~\alpha}^{Aj}k_{j}.$
\end{theorem}
Notice that ii), iii) and iv) are the same conditions as in the subsidiary
system theorem, and that we have added one extra condition, the number v).
This has to be included to guarantee the existence of the hyperbolization
$N_{\tilde{\Delta}}^{\Gamma}\left( k\right) $, as we will show in the proof
of the theorem. However, this is not a condition that physical systems
necessarily satisfy. For example, there could exist $X_{\Gamma}\left(
k\right) $ defined only for a particular direction of $k_{i}$ and such that
$X_{\Gamma}\left( k\right) C_{A}^{\Gamma0}\mathfrak{N}_{~\alpha}^{Aj
k_{j}=0$, so $X_{\Gamma}\left( k\right) $ could not be expanded by the
$M_{A}^{\tilde{\Delta}i}k_{i}$ since the latter and any linear combination of
them are defined for all $k_{i}$. We will explain in subsection \ref{teo_3},
after the proof of the theorem, how to deal with this class of cases. These
results will only be valid in the pseudo-differential version, so we may not
be able to extrapolate them so directly to the non-pseudo-differential case.
The proof of this theorem is given in subsection
\ref{Proof_Theorem_coef_const_2}. This proof uses the fact that the
diagonalization bases of the principal symbol of the evolution equations are
related in a particular way to the diagonalization bases of the principal
symbol of the subsidiary system. This relationship was found by Reula
\cite{reula2004strongly}, assuming that there are subsidiary equations for the
constraints and that they are first-order in derivatives. The theorem
presented here completes these ideas since Reula's assumption is obtained as a
result when conditions (ii), (iii), (iv) are satisfied. On the other hand,
these conditions are the hypotheses of the SS theorem for the quasi-linear
case. Therefore, this relationship between the bases is also valid in the
quasi-linear case (\ref{eq_sys_1}), when the reductions $h_{~A}^{\beta}$ and
$N_{\tilde{\Delta}}^{\Gamma}$ cannot depend on the wave vector $k_{a}$. The
problem in these non-pseudo differential cases is that the propagation
velocities of the subsidiary system cannot be freely chosen. This complicates
a possible proof of a general theorem. Nevertheless, the steps of the proof
presented here can be adapted to each particular theory to conclude similar
results. Following all these ideas and to avoid a more complex discussion, we
present here a closed theorem with simple hypotheses for the
constant-coefficient case.
\subsection{Proof of theorem SH of the SS \label{Proof_Theorem_coef_const_2}}
We show the theorem assuming that the system admits non-trivial $M_{\Delta
}^{\tilde{\Delta}i}k_{i}$. The proof for the case where the system does not
admit $M_{\Delta}^{\tilde{\Delta}i}k_{i}$ will be trivial from the previous
case. We comment on this at the end.
By (i) and theorem \ref{Theor_FyL_2}, we know that there exists a family of
hyperbolizations $h_{B}^{\alpha}\left( k\right) $ of the system
(\ref{Eq_sis_1_coef_cte_1}). From this family, a particular hyperbolization
was used in \cite{Abalos:2018uwg} to prove theorem \ref{Theor_FyL_2}, we call
it \textit{hyperbolization 1}. We shall comment and use about it in the following.
The idea of the proof of our theorem is to show that if we choose the
\textit{hyperbolization 1} $h_{B}^{\alpha}\left( k\right) $, then there
exists $N_{\tilde{\Delta}}^{\Gamma}\left( k\right) $ such that the principal
symbol of the subsidiary system $B_{~\Delta}^{\Gamma i}k_{i}:=C_{A}^{\Gamma
i}h_{~\Delta}^{A}k_{i}+N_{\tilde{\Delta}}^{\Gamma}M_{\Delta}^{\tilde{\Delta
i}k_{i}$ satisfies the definition \ref{def_hyp_fuerte_1} and therefore the
subsidiary equations are strongly hyperbolic. Recall that $h_{B}^{\alpha
}\left( k\right) $ defines the evolution equations $\tilde{e}^{\alpha
}\left( \tilde{\phi}\right) =h_{B}^{\alpha}\left( k\right) \tilde{E
^{B}=0$ whose principal symbol is $\ A_{~\alpha}^{\beta i}k_{i}=h_{~A}^{\beta
}\mathfrak{N}_{~\alpha}^{Ai}k_{i}$.
The characteristic structure of $\ A_{~\alpha}^{\beta i}k_{i}$: the
eigenvalues of $A_{~\alpha}^{\beta i}k_{i}$ define the propagation velocities
of the system \ (see \cite{sarbach2012continuum}). These eigenvalues are
divided into two groups. Those we call the "\textit{physics}", which are
associated to the evolution of the physically relevant fields and do not
change no matter the chosen reduction $h_{~A}^{\beta}$ ; and the rest, which
we call the "\textit{constraints 1}", that depend on the chosen reduction. As
shown in \cite{Abalos:2018uwg}, the "constraints 1" can be freely chosen using
specific reductions. Particularly, by choosing the "hyperbolization 1", the
matrix $A_{~\alpha}^{\beta i}k_{i}$ becomes uniformly diagonalizable with real
eigenvalues. This hyperbolization satisfies that for each $k_{i}$ the
"constraints 1" are simple, non-degenerate, different from each other and
different from the "physics".
The characteristic structure of $\ B_{~\Delta}^{\Gamma i}k_{i}$: we will show
that the eigenvalues of $B_{~\Delta}^{\Gamma i}k_{i}$ are also divided into
two groups. The "constraints 1" (inherited from $A_{~\alpha}^{\beta i}k_{i}$),
which remain unchanged by any choice of $N_{\tilde{\Delta}}^{\Gamma}$; and the
other group which we call the "constraints 2", that depend on $N_{\tilde
{\Delta}}^{\Gamma}$.
Following these ideas, we will show, for each normalized $k_{i}$, that:
a) The set of "constrains 1" are all the generalized eigenvalues of the penci
\begin{equation}
\left[
\begin{array}
[c]{c
-\delta_{\Delta}^{\Gamma}\lambda+C_{A}^{\Gamma j}h_{~\Delta}^{A}k_{j}\\
M_{\Delta}^{\tilde{\Delta}j}k_{j
\end{array}
\right] \label{Eq_pencil_cons_1
\end{equation}
b) In the Kronecker decomposition (see \cite{gantmacher1992theory} and
\cite{gantmakher1998theory} for its definition) of this pencil, all its Jordan
blocks are of dimension 1 for all $k_{i}$. \ In other words, for each $k_{i}$,
these generalized eigenvalues are simple, non-degenerated and different from
each other. \
This implies that (\ref{Eq_pencil_cons_1}) satisfies the condition
(\ref{Eq_cota_ang_can_1}) for canonical angles. So we can use theorem
\ref{Theor_FyL_2} applied to the following pseudo-differential equations
\begin{equation}
\left[
\begin{array}
[c]{c
\delta_{\Delta}^{\Gamma}\partial_{t}+iC_{A}^{\Gamma j}h_{~\Delta}^{A}k_{j}\\
M_{\Delta}^{\tilde{\Delta}j}k_{j
\end{array}
\right] \psi^{\Delta}=0. \label{Eq_pencil_ cons_2
\end{equation}
\ Thus, using the thesis of this theorem, we conclude that there exists (a
hyperbolization) $N_{\tilde{\Delta}}^{\Gamma}\left( k\right) $ such that
$B_{~\Delta}^{\Gamma i}k_{i}=C_{A}^{\Gamma i}h_{~\Delta}^{A}k_{i
+N_{\tilde{\Delta}}^{\Gamma}M_{\Delta}^{\tilde{\Delta}i}k_{i}$ satisfies the
definition \ref{def_hyp_fuerte_1}. \ Here, $N_{\tilde{\Delta}}^{\Gamma}\left(
k\right) $ is chosen as the "hyperbolization 1" for (\ref{Eq_pencil_ cons_2})
and such that the eigenvalues "constraints 2" are simple, non-degenerate,
different from each other and different from "constraints 1".
This discussion says that we should show (a) and (b) from (i), (ii), (iii),
(iv) and (v) to complete the proof of our theorem. \ Although, we should also
justify the use of the pseudo-differential reductions $h_{B}^{\alpha}\left(
k\right) $, $h_{~\Delta}^{A}\left( k\right) $ and $N_{\tilde{\Delta
}^{\Gamma}\left( k\right) $.
\subsubsection{Subsidiary system in Fourier's form\label{Subsi_Fourier}}
We first reproduce the calculations of subsection \ref{teorema_ec_ev_vin_1_a}
until we arrive at the evolution equations for the constraints in their
Fourier version. We have to repeat these steps since the quasi-linear system
includes non-linearities that complicate the Fourier analysis and these
systems do not admit pseudo-differential reductions. The use of these
reductions is only valid in the constant coefficients case and after applying
the Fourier transform to the system.
Recalling that $C_{A}^{\Gamma a}$ and $M_{\Gamma}^{\tilde{\Delta}a}$ are
constants, it is easy to prove the identity
\[
\left( n_{a}\partial_{t}+ik_{a}\right) \left[
\begin{array}
[c]{c
C_{A}^{\Gamma a}\\
M_{\Delta}^{\tilde{\Delta}a
\end{array}
\right] \mathfrak{N}_{~\alpha}^{Ab}\left( n_{b}\partial_{t}+ik_{b}\right)
=0.
\]
Multiplying this expression by $\tilde{\phi}^{\alpha}$, we obtain
\begin{equation}
\left[
\begin{array}
[c]{c
C_{A}^{\Gamma0}\partial_{t}+iC_{A}^{\Gamma j}k_{j}\\
iM_{\Delta}^{\tilde{\Delta}j}k_{j}C_{A}^{\Delta0
\end{array}
\right] \tilde{E}^{A}=0, \label{eq_id_1
\end{equation}
where the expressions $C_{A}^{\Gamma0}=C_{A}^{\Gamma a}n_{a}$, $k_{a}=\left(
0,k_{i}\right) $ and $M_{\Delta}^{\tilde{\Delta}0}=0$ have been used.
We introduce the identity $\delta_{B}^{A}=\left[
\begin{array}
[c]{cc
\mathfrak{N}_{~\alpha}^{A0} & h_{~\Delta}^{A}\left( k\right)
\end{array}
\right] \left[
\begin{array}
[c]{c
h_{B}^{\alpha}\left( k\right) \\
C_{B}^{\Delta0
\end{array}
\right] $ in the above equation and separate the terms as follow
\begin{align}
& \left[
\begin{array}
[c]{c
C_{A}^{\Gamma0}\partial_{t}+iC_{A}^{\Gamma j}k_{j}\\
iM_{A}^{\tilde{\Delta}j}k_{j
\end{array}
\right] \left[
\begin{array}
[c]{cc
\mathfrak{N}_{~\alpha}^{A0} & h_{~\Delta}^{A}\left( k\right)
\end{array}
\right] \nonumber\\
& =\left[
\begin{array}
[c]{cc
-iC_{A}^{\Gamma0}\mathfrak{N}_{~\alpha}^{Aj}k_{j} & \delta_{\Delta}^{\Gamma
}\partial_{t}+iC_{A}^{\Gamma j}k_{j}h_{~\Delta}^{A}\\
0 & iM_{\Delta}^{\tilde{\Delta}j}k_{j
\end{array}
\right] , \label{eq_id_3
\end{align}
where we used that $M_{\Delta}^{\tilde{\Delta}0}=0$, $C_{B}^{\Gamma
0}h_{~\Delta}^{A}=\delta_{\Delta}^{\Gamma}$, $-C_{A}^{\Gamma0}\mathfrak{N
_{~\alpha}^{Aj}=C_{A}^{\Gamma i}\mathfrak{N}_{~\alpha}^{A0}$. An
\begin{align}
\left[
\begin{array}
[c]{c
\tilde{e}^{\beta}\\
\tilde{\psi}^{\Delta
\end{array}
\right] & :=\left[
\begin{array}
[c]{c
h_{B}^{\alpha}\left( k\right) \\
C_{B}^{\Delta0
\end{array}
\right] \tilde{E}^{B}\label{eq_id_4}\\
& =\left[
\begin{array}
[c]{c
\partial_{t}\tilde{\phi}^{\beta}+ih_{~A}^{\beta}\left( k_{i}\right)
\mathfrak{N}_{~\alpha}^{Ai}k_{i}\tilde{\phi}^{\alpha}\\
iC_{B}^{\Delta0}\mathfrak{N}_{~\alpha}^{Ai}k_{i}\tilde{\phi}^{\alpha
\end{array}
\right] , \label{eq_id_5
\end{align}
where we used that $h_{~A}^{\beta}\mathfrak{N}_{~\alpha}^{A0}=\delta_{\alpha
}^{\beta}$ and $C_{B}^{\Delta0}\mathfrak{N}_{~\alpha}^{A0}=0$. So, by
contracting eq. (\ref{eq_id_3}) with (\ref{eq_id_4}), the identity
(\ref{eq_id_1}) is rewritten as
\begin{equation}
\left[
\begin{array}
[c]{cc
-iC_{A}^{\Gamma0}\mathfrak{N}_{~\alpha}^{Aj}k_{j} & \delta_{\Delta}^{\Gamma
}\partial_{t}+iC_{A}^{\Gamma j}k_{j}h_{~\Delta}^{A}\\
0 & iM_{\Delta}^{\tilde{\Delta}j}k_{j
\end{array}
\right] \left[
\begin{array}
[c]{c
\tilde{e}^{\beta}\\
\tilde{\psi}^{\Delta
\end{array}
\right] =0. \label{eq_id_6
\end{equation}
In the on-shell case, when the evolution equations are satisfied $\tilde
{e}^{\beta}=0,$ we obtain the system (\ref{Eq_pencil_ cons_2}), that is,
\[
\left[
\begin{array}
[c]{c
\delta_{\Delta}^{\Gamma}\partial_{t}+iC_{A}^{\Gamma j}k_{j}h_{~\Delta}^{A}\\
iM_{\Delta}^{\tilde{\Delta}j}k_{j
\end{array}
\right] \tilde{\psi}^{\Delta}=0.
\]
The constraints $\tilde{\psi}^{\Delta}$ satisfy all these pseudo-differential
equations, then the final expression for the subsidiary system \ (on-shell
case $\tilde{e}^{\alpha}\left( \tilde{\phi}\right) =0$) is
\begin{equation}
\partial_{0}\psi^{\Gamma}+\left( C_{A}^{\Gamma i}h_{~\Delta}^{A
+N_{\tilde{\Delta}}^{\Gamma}M_{\Delta}^{\tilde{\Delta}i}\right) k_{i
\psi^{\Delta}=0. \label{Eq_sist_B_1
\end{equation}
Notice that the principal symbol of the subsidiary system is the same as in
the quasi-linear case (see eq. (\ref{Eq_simb_vinc_1})), the difference here is
that $h_{~\Delta}^{A}\left( k\right) $ and $N_{\tilde{\Delta}}^{\Gamma
}\left( k\right) $ can depend on the wave vector.
We also note that by contracting the eq. (\ref{eq_id_3}) with (\ref{eq_id_5})
the following identity is obtaine
\begin{equation}
\left[
\begin{array}
[c]{cc
-iC_{A}^{\Gamma0}\mathfrak{N}_{~\alpha}^{Aj}k_{j} & \delta_{\Delta}^{\Gamma
}\partial_{t}+iC_{A}^{\Gamma j}k_{j}h_{~\Delta}^{A}\\
0 & iM_{\Delta}^{\tilde{\Delta}j}k_{j
\end{array}
\right] \left[
\begin{array}
[c]{c
\partial_{t}\tilde{\phi}^{\beta}+ih_{~A}^{\beta}\left( k_{i}\right)
\mathfrak{N}_{~\alpha}^{Ai}k_{i}\tilde{\phi}^{\alpha}\\
iC_{B}^{\Delta0}\mathfrak{N}_{~\alpha}^{Ai}k_{i}\tilde{\phi}^{\alpha
\end{array}
\right] =0. \label{Eq_left_ker_K_1
\end{equation}
By condition v), we are assuming that $M_{\Delta}^{\tilde{\Delta}j}k_{j}$
expands the entire left kernel of $C_{B}^{\Delta0}\mathfrak{N}_{~\alpha
^{Ai}k_{i}$, otherwise, the above system would be modified as follows
\begin{equation}
\left[
\begin{array}
[c]{cc
-iC_{A}^{\Gamma0}\mathfrak{N}_{~\alpha}^{Aj}k_{j} & \delta_{\Delta}^{\Gamma
}\partial_{t}+iC_{A}^{\Gamma j}k_{j}h_{~\Delta}^{A}\\
0 & iM_{\Delta}^{\tilde{\Delta}j}k_{j}\\
0 & iX_{\Delta}^{\check{s}}\left( k\right)
\end{array}
\right] \left[
\begin{array}
[c]{c
\tilde{e}^{\beta}\\
\tilde{\psi}^{\Delta
\end{array}
\right] =0. \label{Eq_sin_v_1
\end{equation}
Where the vectors $X_{\Delta}^{\check{s}}\left( k\right) $ satisfy
$X_{\Delta}^{\check{s}}\left( k\right) C_{B}^{\Delta0}\mathfrak{N}_{~\alpha
}^{Ai}k_{i}=0$ for some $k_{i}$ directions and cannot be obtained from linear
combinations of $M_{\Delta}^{\tilde{\Delta}j}k_{j}$. The index $\check{s}$
numbers these vectors such that, for each $k_{i}$, $span\left\langle
X_{\Delta}^{\check{s}},M_{\Delta}^{\tilde{\Delta}j}k_{j}\right\rangle
=left\_\ker\left( C_{B}^{\Delta0}\mathfrak{N}_{~\alpha}^{Ai}k_{i}\right) $.
We will return to this discussion after the proof of this theorem and explain
how to obtain the strong hyperbolicity of the subsidiary system by suppressing
condition v).
\subsubsection{Relationship between the principal symbols}
As a second step, we study the relationship between the principal symbols of
the evolution equations and the subsidiary system. Every expression found in
this subsubsection is also valid in the quasi-linear case.
For each normalized $k_{a}=\left( 0,k_{i}\right) $, consider the lines
$l_{a}\left( \lambda\right) =-n_{a}\lambda+k_{a}\in S_{n_{a}}$. Using the
equations (\ref{Eq_C_K_nueva_1}) and (\ref{eq_M_K_1}) we can conclude the
following identities
\begin{align}
0 & =l_{a}\left( \lambda\right) M_{\Gamma}^{\tilde{\Delta}a}C_{A
^{\Gamma0}\mathfrak{N}_{~\alpha}^{Ab}l_{b}\left( \lambda\right)
,\label{Eq_M_C0_K_1}\\
& =k_{i}M_{\Gamma}^{\tilde{\Delta}i}\left( C_{A}^{\Gamma0}\mathfrak{N
_{~\alpha}^{Aj}k_{j}\right) , \label{Eq_M_C0_K_2
\end{align}
an
\begin{align*}
0 & =l_{a}\left( \lambda\right) C_{A}^{\Gamma a}\left[
\begin{array}
[c]{cc
\mathfrak{N}_{~\alpha}^{A0} & h_{~\Delta}^{A
\end{array}
\right] \left[
\begin{array}
[c]{c
h_{B}^{\alpha}\\
C_{B}^{\Delta0
\end{array}
\right] \mathfrak{N}_{~\beta}^{Bb}l_{b}\left( \lambda\right) ,\\
& =-\left( l_{a}C_{A}^{\Gamma0}\mathfrak{N}_{~\alpha}^{Aa}\right) \left(
h_{B}^{\alpha}\mathfrak{N}_{~\beta}^{Bb}l_{b}\right) +\left( l_{a
C_{A}^{\Gamma a}h_{~\Delta}^{A}\right) \left( C_{B}^{\Delta0}\mathfrak{N
_{~\beta}^{Bb}l_{b}\right) ,\\
& =-\left( k_{j}C_{A}^{\Gamma0}\mathfrak{N}_{~\alpha}^{Aj}\right) \left(
-\lambda\delta_{\beta}^{\alpha}+h_{~A}^{\alpha}\mathfrak{N}_{~\beta}^{Ai
k_{i}\right) +\left( -\lambda\delta_{\Delta}^{\Gamma}+C_{A}^{\Gamma
j}h_{~\Delta}^{A}k_{j}\right) \left( C_{B}^{\Delta0}\mathfrak{N}_{~\beta
}^{Bj}k_{j}\right) .
\end{align*}
Notice that these equations are exactly the rows of (\ref{Eq_left_ker_K_1}) by
replacing $\partial_{t}$ by $-i\lambda$ and dividing by $i$.\ We conclude the
following equation
\begin{equation}
\left( k_{j}C_{A}^{\Gamma0}\mathfrak{N}_{~\alpha}^{Aj}\right) \left(
-\lambda\delta_{\beta}^{\alpha}+h_{~A}^{\alpha}\mathfrak{N}_{~\beta}^{Ai
k_{i}\right) =\left( -\lambda\delta_{\Delta}^{\Gamma}+C_{A}^{\Gamma
j}h_{~\Delta}^{A}k_{j}\right) \left( C_{B}^{\Delta0}\mathfrak{N}_{~\beta
}^{Bj}k_{j}\right) . \label{Eq_simbs_1
\end{equation}
This equation was found in \cite{reula2004strongly}, assuming that the
subsidiary system was first order in derivatives. The latter equation is not
necessarily valid unless the system has the structure associated with the
Geroch fields presented here, i.e. hypotheses ii), iii) and iv) of the theorem.
From the eq. (\ref{Eq_M_C0_K_2}) we know that $k_{i}M_{\Gamma}^{\tilde{\Delta
}i}$ belongs to the left-hand kernel of $C_{A}^{\Gamma0}\mathfrak{N}_{~\alpha
}^{Aj}k_{j}$. Therefore, considerin
\[
B_{~\Delta}^{\Gamma j}k_{j}:=\left( C_{A}^{\Gamma j}h_{~\Delta}^{A
+N_{\tilde{\Delta}}^{\Gamma}M_{\Delta}^{\tilde{\Delta}j}\right) k_{j
\]
with free $N_{\tilde{\Delta}}^{\Gamma}$, the above equation can be rewritten
as
\begin{equation}
\left( k_{j}C_{A}^{\Gamma0}\mathfrak{N}_{~\alpha}^{Aj}\right) \left(
-\lambda\delta_{\beta}^{\alpha}+A_{~\beta}^{\alpha i}k_{i}\right) =\left(
-\delta_{\Delta}^{\Gamma}\lambda+B_{~\Delta}^{\Gamma j}k_{j}\right) \left(
C_{B}^{\Delta0}\mathfrak{N}_{~\beta}^{Bi}k_{i}\right) .
\label{Eq_rel_simbolos_1
\end{equation}
This last expression shows how the principal symbols of the evolution
equations for $\phi^{\alpha}$ are related to the principal symbol of the
subsidiary system.
\subsubsection{Left Kernel of $\mathfrak{N}_{~\alpha}^{Ab}l_{b}\left(
\lambda\right) $}
In this subsubsection, we shall choose a basis for the left kernel of
$\mathfrak{N}_{~\alpha}^{Ab}l_{b}\left( \lambda\right) $. This basis will
allow us to find the Kronecker structure of the pencil $\mathfrak{N}_{~\alpha
}^{Ab}l_{b}\left( \lambda\right) $ in the following subsubsection.
From the previous subsubsections, we know that
\[
0=\left( -\lambda n_{a}+k_{a}\right) \left[
\begin{array}
[c]{c
C_{A}^{\Gamma a}\\
M_{\Delta}^{\tilde{\Delta}a
\end{array}
\right] \mathfrak{N}_{~\alpha}^{Ab}\left( -\lambda n_{b}+k_{b}\right) .
\]
As before, using equation (\ref{eq_Kh_hC_1}) this expression can be rewritten
as
\[
0=\left[
\begin{array}
[c]{cc
-C_{A}^{\Gamma0}\mathfrak{N}_{~\alpha}^{Aj}k_{j} & -\lambda\delta_{\Delta
}^{\Gamma}+\left( C_{A}^{\Gamma j}h_{~\Delta}^{A}\right) k_{j}\\
0 & M_{\Delta}^{\tilde{\Delta}j}k_{j
\end{array}
\right] \left[
\begin{array}
[c]{c
-\lambda\delta_{\alpha}^{\beta}+h_{~A}^{\beta}\mathfrak{N}_{~\alpha}^{Ai
k_{i}\\
C_{B}^{\Delta0}\mathfrak{N}_{~\alpha}^{Ai}k_{i
\end{array}
\right] .
\]
Therefore, the basis we are looking for will result from the choice of a
subset of vectors of
\begin{equation}
\left[
\begin{array}
[c]{cc
-C_{A}^{\Gamma0}\mathfrak{N}_{~\alpha}^{Aj}k_{j} & -\lambda\delta_{\Delta
}^{\Gamma}+\left( C_{A}^{\Gamma j}h_{~\Delta}^{A}\right) k_{j}\\
0 & M_{\Delta}^{\tilde{\Delta}j}k_{j
\end{array}
\right] . \label{left_ker_1
\end{equation}
Notice that, for each $k_{j}$, the pencil
\begin{align}
\lambda I_{\alpha}^{A}+K_{\alpha}^{A} & :=\left[
\begin{array}
[c]{c
h_{B}^{\beta}\\
C_{B}^{\Delta0
\end{array}
\right] \mathfrak{N}_{~\alpha}^{Bb}l_{b}\left( \lambda\right) =\left[
\begin{array}
[c]{c
-\lambda\delta_{\alpha}^{\beta}+h_{~A}^{\beta}\mathfrak{N}_{~\alpha}^{Ai
k_{i}\\
C_{B}^{\Delta0}\mathfrak{N}_{~\alpha}^{Ai}k_{i
\end{array}
\right] \label{eq_pencil_h_C_K_1}\\
& =\lambda\left[
\begin{array}
[c]{c
-\delta_{\beta}^{\alpha}\\
0
\end{array}
\right] +\left[
\begin{array}
[c]{c
h_{~A}^{\beta}\mathfrak{N}_{~\alpha}^{Ai}k_{i}\\
C_{B}^{\Delta0}\mathfrak{N}_{~\alpha}^{Ai}k_{i
\end{array}
\right] \nonumber
\end{align}
has the form of the pencil (\ref{pencil_1}) in appendix \ref{Ap_lemmas}. So,
by making use of the lemma \ref{Lemma_1}) (in this appendix) and noticing that
\ $\lambda I_{\alpha}^{A}+K_{\alpha}^{A}$ and $\mathfrak{N}_{~\alpha
^{Bb}l_{b}\left( \lambda\right) $ are related by an invertible matrix
independent of $\lambda$, we conclude that
\[
\dim\left( left\_\ker\left( \mathfrak{N}_{~\alpha}^{Bb}l_{b}\left(
\lambda\right) \right) \right) =c
\]
for any $\lambda$ different from the generalized eigenvalues $\lambda
_{i}\left( k\right) $. This means that we have to choose $c$ linearly
independent vectors of (\ref{left_ker_1}) as the left kernel basis of
$\mathfrak{N}_{~\alpha}^{Bb}l_{b}\left( \lambda\right) $.
First, since $C_{A}^{\Gamma0}\mathfrak{N}_{~\alpha}^{Ai}k_{i}$ plays a very
important role in the rest of the proof, we introduce some definitions
associated with this operator. For each $k_{i}$, we call
\begin{equation}
d\left( k\right) :=\dim\left( right\_\ker\left( C_{A}^{\Gamma
0}\mathfrak{N}_{~\alpha}^{Ai}k_{i}\right) \right) \label{Eq_d_1
\end{equation}
to the dimension of the right kernel of $C_{A}^{\Gamma0}\mathfrak{N}_{~\alpha
}^{Ai}k_{i}$,
\begin{equation}
r\left( k\right) :=rank\left( C_{A}^{\Gamma0}\mathfrak{N}_{~\alpha
^{Ai}k_{i}\right) \label{Eq_r_1
\end{equation}
to its rank and
\begin{equation}
s\left( k\right) =:\dim\left( left\_\ker\left( C_{A}^{\Gamma0
\mathfrak{N}_{~\alpha}^{Ai}k_{i}\right) \right) \label{Eq_s_1
\end{equation}
to the dimension of its left kernel. \ We also recall that, by the
rank-nullity theorem,
\begin{align}
u & =r\left( k\right) +d\left( k\right) ,\label{Eq_u_r_d_1}\\
c & =r\left( k\right) +s\left( k\right) . \label{Eq_c_r_s_2
\end{align}
By hypothesis v), for each $k_{i}$, the vectors $M_{\Delta}^{\tilde{\Delta
j}k_{j}$ expand the left kernel of $C_{A}^{\Gamma0}\mathfrak{N}_{~\alpha
^{Ai}k_{i}$ (whose dimension is $s\left( k\right) $). Therefore, we choose
the following $s\left( k\right) $ linearly independent vectors
\[
m_{B}^{zj}k_{j}:=h_{\tilde{\Delta}}^{r}\left[
\begin{array}
[c]{cc
0 & M_{\Delta}^{\tilde{\Delta}j}k_{j
\end{array}
\right] \left[
\begin{array}
[c]{c
h_{B}^{\alpha}\\
C_{B}^{\Delta0
\end{array}
\right]
\]
as part of our basis. These are obtained from the rows of (\ref{left_ker_1}),
where $z=1,...,s\left( k\right) $ and $h_{\tilde{\Delta}}^{r}$ is the
projector representing our choice. \ These vectors belong to the left kernel
of $\mathfrak{N}_{~\alpha}^{Bb}l_{b}\left( \lambda\right) $ and have the
property that they do not depend on $\lambda$. Notice also that by the form of
(\ref{eq_pencil_h_C_K_1}) any other vector of the left kernel of
$\mathfrak{N}_{~\alpha}^{Bb}l_{b}\left( \lambda\right) $, linearly
independent from the $m_{B}^{zj}k_{j}$, will depend on $\lambda$.
On the other hand, since for each $k_{i}$, the rank of $C_{A}^{\Gamma
0}\mathfrak{N}_{~\alpha}^{Ai}k_{i}$ is $r\left( k\right) $ and since
$C_{A}^{\Gamma0}\mathfrak{N}_{~\alpha}^{Ai}k_{i}$ appears explicitly in the
first rows of (\ref{left_ker_1}), we complete our basis with the following
linearly independent $r\left( k\right) $ vector
\[
c_{B}^{wa}l_{a}\left( \lambda\right) :=h_{\Gamma}^{w}\left[
\begin{array}
[c]{cc
-C_{A}^{\Gamma0}\mathfrak{N}_{~\alpha}^{Aj}k_{j} & -\lambda\delta_{\Delta
}^{\Gamma}+\left( C_{A}^{\Gamma j}h_{~\Delta}^{A}\right) k_{j
\end{array}
\right] \left[
\begin{array}
[c]{c
h_{B}^{\alpha}\\
C_{B}^{\Delta0
\end{array}
\right] .
\]
Where $w=1,...,r\left( k\right) ,$ and $h_{\Gamma}^{w}$ is the projector
associated with our choice. It only remains to show that and $c_{B}^{wa
l_{a}\left( \lambda\right) $ are linearly independent between them. This is
shown by recalling that $h_{\Gamma}^{w}\left( -C_{A}^{\Gamma0}\mathfrak{N
_{~\alpha}^{Aj}k_{j}\right) $ has no left kernel (in the index $w$) and thus
the only $\left[
\begin{array}
[c]{cc
X_{w} & X_{r
\end{array}
\right] $ which cancels the first $u$ columns of
\[
\left[
\begin{array}
[c]{cc
h_{\Gamma}^{w}\left( -C_{A}^{\Gamma0}\mathfrak{N}_{~\alpha}^{Aj}k_{j}\right)
& h_{\Gamma}^{w}\left( -\lambda\delta_{\Delta}^{\Gamma}+\left( C_{A}^{\Gamma
j}h_{~\Delta}^{A}\right) k_{j}\right) \\
0 & h_{\tilde{\Delta}}^{r}\left( M_{\Delta}^{\tilde{\Delta}j}k_{j}\right)
\end{array}
\right] ,
\]
is the trivial ones.
We conclude by (\ref{Eq_c_r_s_2}) that our chosen basis $\left\{ m_{B
^{zj}k_{j},c_{B}^{wa}l_{a}\left( \lambda\right) \right\} $ has $c$ vectors,
where the $m_{B}^{zj}k_{j}$ are $s\left( k\right) $ vectors that do not
depend on $\lambda$ and the $c_{B}^{wa}l_{a}\left( \lambda\right) $ are
$r\left( k\right) $ vectors that depend linearly on $\lambda$. This is a
base of the left kernel of $\mathfrak{N}_{~\alpha}^{Ab}l_{b}\left(
\lambda\right) $ valid for $\lambda$ different to the generalized eigenvalues.
\subsubsection{Kronecker decomposition of $\mathfrak{N}_{~\beta}^{Bb
l_{b}\left( \lambda\right) $ \label{Kronecker_decomposition}}
The aim of this subsubsection is to give the Kronecker decomposition of the
principal symbol (pencil matrix
\begin{equation}
\mathfrak{N}_{~\beta}^{Bb}l_{b}\left( \lambda\right) =-\lambda
\mathfrak{N}_{~\beta}^{B0}+\mathfrak{N}_{~\beta}^{Bi}k_{i}.
\label{Eq_pencil_K_l_1
\end{equation}
where $l_{b}\left( \lambda\right) =-\lambda n_{a}+k_{a}\in S_{n_{a}}$. This
decomposition will guide us in the following subsubsections to complete the
proof of the theorem.
This pencil is an $e\times u$ matrix in
\mathbb{C}
$, although it makes no difference if we think of its components in
\mathbb{R}
$. The Kronecker decomposition consists of rewriting the pencil as
\[
\mathfrak{N}_{~\beta}^{Bb}l_{b}\left( \lambda\right) =Y_{~B}^{A}\left(
k\right) K_{~\alpha}^{B}\left( \lambda,k\right) W_{~\beta}^{\alpha}\left(
k\right) .
\]
Where $Y_{~B}^{A}\left( k\right) \i
\mathbb{C}
^{e\times e}$ and $W_{~\beta}^{\alpha}\left( k\right) \i
\mathbb{C}
^{u\times u}$ are two invertible matrices that depend on $k_{i}$ and not on
$\lambda$; and $K_{~\alpha}^{B}\left( \lambda,k\right) \i
\mathbb{C}
^{e\times u}$ is a block matrix (see \cite{gantmacher1992theory},
\cite{gantmakher1998theory} and \cite{Abalos:2018uwg} for details), which
depends on $\lambda$ and $k_{i}$. The matrices $Y_{~B}^{A}\left( k\right) $
and $W_{~\beta}^{\alpha}\left( k\right) $ can be thought of as
change-of-basis matrices, and $K_{~\alpha}^{B}\left( \lambda,k\right) $ is
the block matrix which we call the Kronecker structure.
We know from Lemmas 1 and 2 of \cite{Abalos:2018uwg} (whose hypotheses are
valid in this proof), that $K_{~\alpha}^{B}\left( \lambda,k\right) $ has all
their Jordan blocks of dimension 1. We also know from subsection 3.1 of
\cite{Abalos:2018uwg} that since $\mathfrak{N}_{~\beta}^{B0}$ has only trivial
right kernel, the remaining blocks appearing in this decomposition are the
$L_{m}^{T}$ and zero rows (called $L_{0}^{T}$ simplifying the notation). This
decomposition depends on $k_{i}$, i.e. the generalized eigenvectors, the
number of them, the block structure $L_{m}^{T}$ (how many and which ones) and
the number of zero rows can be modified for different $k_{i}$'s.
We say that the Kronecker structure of the $e\times u$ pencil $\mathfrak{N
_{~\beta}^{Bb}l_{b}\left( \lambda\right) $ has the following structure:
\begin{equation}
d\left( k\right) \times J_{1},~r\left( k\right) \times L_{1}^{T},~s\left(
k\right) \times L_{0}^{T}. \label{Kro_K_1
\end{equation}
Where for each $k_{i}$, the quantities $d\left( k\right) $, $r\left(
k\right) $, $s\left( k\right) $ are defined by (\ref{Eq_d_1}),
(\ref{Eq_r_1}) and (\ref{Eq_s_1}) respectively and they also satisfy the
equations (\ref{Eq_u_r_d_1}) and (\ref{Eq_c_r_s_2}). We justify each of the
terms of (\ref{Kro_K_1}) below.
The following analysis is valid for each $k_{i}$, so we assume $k_{i}$ is fixed.
$\bullet$ We begin by justifying the block structure $r\left( k\right)
\times L_{1}^{T},~s\left( k\right) \times L_{0}^{T}$.
Let us study the left kernel of $L_{m}^{T}=\left[
\begin{array}
[c]{ccccc
\lambda & 0 & 0 & 0 & 0\\
1 & \lambda & 0 & 0 & 0\\
0 & 1 & ... & 0 & 0\\
0 & 0 & ... & \lambda & 0\\
0 & 0 & 0 & 1 & \lambda\\
0 & 0 & 0 & 0 & 1
\end{array}
\right] \i
\mathbb{R}
^{m+1\times m}$. It has dimension 1 and is expanded by the vecto
\[
X=\left[ -1,\lambda,...,\left( -1\right) ^{m}\lambda^{m-1},\left(
-1\right) ^{m+1}\lambda^{m}\right] \i
\mathbb{R}
^{1\times m+1}.
\]
\ The coefficients of $X$ are polynomials in $\lambda$ whose major degree is
$m$. We can increase the degree of these coefficients, for example, by
considering the vector $\lambda X$, but we cannot reduce it without obtaining
rational functions. \ This allows us to introduce a method to detect the
$L_{m}^{T}$ blocks present in a given pencil. \ We first define the function
$gr$ which takes vectors with polynomial coefficients in $\lambda$ and returns
the greater polynomial degree between their coefficients (in our example,
$gr\left( X\right) =m$ and $gr\left( \lambda X\right) =m+1$). If we now
consider a pencil with different blocks $L_{m_{1}}^{T},...,L_{m_{v}}^{T}$ such
that $0\leq m_{1}\leq m_{2}\leq...\leq m_{v}$, there exists a left kernel
basis $\left\{ X_{i}\text{ with }i=1,...,v\right\} $ of this pencil such
that the values of $gr\left( X_{i}\right) =m_{i}$ identify the types of
$L_{m}^{T}$ blocks present in the pencil. Any other basis $\left\{
Z_{i}\right\} $, with $gr\left( Z_{i}\right) =z_{i}$ and ordered such that
$z_{1}\leq z_{2}\leq...\leq z_{p}$, has at least one $z_{i}$ such that
$m_{i}\leq z_{i}$ with $i$ between $1$ and $v$. \ This indicates that to find
the Kronecker structure of a pencil, we have to identify a left kernel basis
of the pencil that minimizes the $gr$ function for their elements. Notice
that, if we have a zero row $L_{0}^{T}$, its left kernel $F$ can be chosen
independent of $\lambda$, then $gr\left( F\right) =0$.
Applying this method to $\mathfrak{N}_{~\beta}^{Bb}l_{b}\left( \lambda
\right) $, using the basis $\left\{ m_{B}^{zj}k_{j}\text{, }c_{B}^{wa
l_{a}\left( \lambda\right) \right\} $ with $z=1,...,s\left( k\right) $,
$w=1,...,r\left( k\right) $ \ from the previous subsubsection, which
satisfies that $gr\left( m_{B}^{rj}k_{j}\right) =0$ and $gr\left(
c_{B}^{wa}l_{a}\left( \lambda\right) \right) =1$, we conclude that the
Kronecker structure of $\mathfrak{N}_{~\beta}^{Bb}l_{b}\left( \lambda\right)
$ has the blocks $r\left( k\right) \times L_{1}^{T},~s\left( k\right)
\times L_{0}^{T}$ and the rest of the structure are Jordan blocks.
$\bullet$ We have already explained that this pencil only has Jordan blocks of
dimension 1, it remains to explain that they are $d\left( k\right) $ blocks.
This result can be concluded by counting the number of $\lambda$'s appearing
in the columns of $K_{~\alpha}^{B}\left( \lambda,k\right) $. Since only one
$\lambda$ can be present per column and $K_{~\alpha}^{B}\left( \lambda
,k\right) $ has $u$ columns, there are $u$ $\lambda$'s in $K_{~\alpha
^{B}\left( \lambda,k\right) $. Notice that $L_{0}^{T}$ has no $\lambda$
since it is a row of zeros and that each $L_{1}^{T}$ has only one $\lambda$
per column, then in $r\left( k\right) \times L_{1}^{T},~s\left( k\right)
\times L_{0}^{T}$\ there are $r\left( k\right) $ $\lambda$'s. Finally, the
number of Jordan blocks of dimension 1 is $u-r\left( k\right) =d\left(
k\right) $ (see eq.(\ref{Eq_u_r_d_1})).
As a final comment, we note that the sum of the multiplicities of each of the
different $q\left( k\right) -$generalized eigenvalues has to be equal to
$d\left( k\right) $, that is,
\[
d_{\lambda_{1}}+d_{\lambda_{2}}+...+d_{\lambda_{q\left( k\right) }}=d\left(
k\right) .
\]
\subsubsection{Basis which diagonalize $A_{~\beta}^{\alpha i}k_{i}$
\label{Eigenbasis_1}}
In this subsubsection, we use the obtained information of the Kronecker
decomposition of $\mathfrak{N}_{~\beta}^{Bb}l_{b}\left( \lambda\right) $ to
find the bases that diagonalize $A_{~\beta}^{\alpha i}k_{i}$.
The eigenvectors which diagonalize $A_{~\beta}^{\alpha i}k_{i}$ are divided
into two groups:
$\bullet$ The generalized eigenvectors $\delta\phi_{\lambda_{i}\left(
k\right) }^{\beta}$, associated to the generalized eigenvalues $\lambda
_{i}\left( k\right) $, such that they satisf
\[
\mathfrak{N}_{~\beta}^{Bb}\left( -\lambda_{i}\left( k\right) n_{a
+k_{a}\right) \delta\phi_{\lambda_{i}\left( k\right) }^{\beta}=0,
\]
with $\lambda_{1}\left( k\right) \leq\lambda_{2}\left( k\right)
\leq...\leq\lambda_{d\left( k\right) }\left( k\right) $. We know from the
previous subsubsection that for each $\lambda_{i}\left( k\right) $ there is
a generalized eigenvector $\delta\phi_{\lambda_{i}\left( k\right) }^{\beta
$. Notice that to simplify the notation, we have changed how we denote the
generalized eigenvalues with respect to the section \ref{Sec_Const_coeff_1}.
From the above equation, it follows that
\[
\left[
\begin{array}
[c]{c
h_{B}^{\alpha}\\
C_{B}^{\Delta0
\end{array}
\right] \mathfrak{N}_{~\beta}^{Bb}\left( -\lambda_{i}\left( k\right)
n_{a}+k_{a}\right) \delta\phi_{\lambda_{i}\left( k\right) }^{\beta}=\left[
\begin{array}
[c]{c
-\lambda_{i}\left( k\right) \delta_{\beta}^{\alpha}+A_{\beta}^{\alpha
i}k_{i}\\
C_{B}^{\Delta0}\mathfrak{N}_{~\beta}^{Bi}k_{i
\end{array}
\right] \delta\phi_{\lambda_{i}\left( k\right) }^{\beta}=0,
\]
therefore, these generalized eigenvectors are eigenvectors of $A_{\beta
}^{\alpha i}k_{i}$ and belong to the right kernel of $C_{B}^{\Delta
0}\mathfrak{N}_{~\beta}^{Bi}k_{i}$.
Recalling that the dimension of the right kernel of \ $C_{B}^{\Delta
0}\mathfrak{N}_{~\beta}^{Bi}k_{i}$ is $d\left( k\right) $ (see eq.
(\ref{Eq_d_1})) and that the $\delta\phi_{\lambda_{i}\left( k\right)
}^{\beta}$ are $d\left( k\right) $ linearly independent vectors, we conclude
tha
\begin{equation}
right\_\ker\left( C_{B}^{\Delta0}\mathfrak{N}_{~\beta}^{Bi}k_{i}\right)
=span\left\langle \delta\phi_{\lambda_{i}\left( k\right) }^{\beta
}\right\rangle . \label{Eq_span_ker_der_C0_Ki_1
\end{equation}
We also recall that the set of generalized eigenvalues $\left\{ \lambda
_{i}\left( k\right) \right\} $ are those we called the "physical" at the
beginning of the proof.
$\bullet$ The eigenvectors $\delta\phi_{\pi_{i}\left( k\right) }^{\beta}$,
associated to the eigenvalues $\pi_{i}\left( k\right) $, such that they
satisfy
\begin{align}
\left( -\pi_{i}\left( k\right) \delta_{\beta}^{\alpha}+A_{\beta}^{\alpha
i}k_{i}\right) \delta\phi_{\pi_{i}\left( k\right) }^{\beta} &
=0,\label{eq_phi_pi_1}\\
C_{B}^{\Delta0}\mathfrak{N}_{~\beta}^{Bi}k_{i}\delta\phi_{\pi_{i}\left(
k\right) }^{\beta} & \neq0.\nonumber
\end{align}
We are considering that the reduction $h_{~A}^{\beta}\left( k\right) $ is
chosen such that, for each $k_{i}$, the $\left\{ \pi_{i}\left( k\right)
\right\} $ are simple, different from each other and different from the
$\left\{ \lambda_{i}\left( k\right) \right\} $. We order them in the
following way $\pi_{1}\left( k\right) <\pi_{2}\left( k\right)
<...<\pi_{r\left( k\right) }\left( k\right) $ and notice that since they
are simple, each $\pi_{i}\left( k\right) $ has its associated eigenvector
$\delta\phi_{\pi_{i}\left( k\right) }^{\beta}$. Notice that $\left\{
\lambda_{i}\left( k\right) ,\text{ }\pi_{i}\left( k\right) \right\} $ and
$\left\{ \delta\phi_{\lambda_{i}\left( k\right) }^{\beta},\text{
\delta\phi_{\pi_{i}\left( k\right) }^{\beta}\right\} $ are the sets of the
eigenvalues and eigenvectors of $A_{\beta}^{\alpha i}k_{i}$ and each of these
sets has $u=d\left( k\right) +r\left( k\right) $ elements. \ Finally, we
recall that the set of eigenvalues $\left\{ \pi_{i}\left( k\right)
\right\} $ was previously called the "constraints 1".
\subsubsection{Kronecker structure of the subsidiary system
(\ref{Eq_pencil_cons_1}) \label{Kro_sub}}
In this subsubsection, we show that the Kronecker structure of the pencil
(\ref{Eq_pencil_cons_1}) is
\begin{equation}
J_{1}\left( \pi_{1}\left( k\right) \right) ,...,J_{1}\left( \pi_{r\left(
k\right) }\left( k\right) \right) ,s\left( k\right) \times L_{1
^{T},y\left( k\right) \times L_{0}^{T}, \label{eq_sub_kron_1
\end{equation}
where $y\left( k\right) :=\dim\left( left\_\ker\left( M_{\Delta
^{\tilde{\Delta}j}k_{j}\right) \right) $.
$\bullet$ We begin by showing that $J_{1}\left( \pi_{1}\left( k\right)
\right) ,...,J_{1}\left( \pi_{r\left( k\right) }\left( k\right) \right)
$ is part of the Kronecker structure.
For this purpose, we will show that the set of $\pi_{i}\left( k\right) $ are
the generalized eigenvalues of (\ref{Eq_pencil_cons_1}), with their
corresponding linearly independent generalized eigenvectors
\begin{equation}
\delta\psi_{\pi_{i}\left( k\right) }^{\Delta}:=C_{B}^{\Delta0
\mathfrak{N}_{~\beta}^{Bi}k_{i}\delta\phi_{\pi_{i}\left( k\right) }^{\beta};
\label{Eq_chi_pi_1
\end{equation}
i.e., for each $i=1,...,r\left( k\right) $, it holds that
\begin{equation}
\left[
\begin{array}
[c]{c
-\delta_{\Delta}^{\Gamma}\pi_{i}\left( k\right) +C_{A}^{\Gamma j}h_{~\Delta
}^{A}k_{j}\\
M_{\Delta}^{\tilde{\Delta}j}k_{j
\end{array}
\right] \delta\psi_{\pi_{i}\left( k\right) }^{\Delta}=0.
\label{Eq_vin_chi_1
\end{equation}
First, recalling that $M_{\Delta}^{\tilde{\Delta}j}k_{j}$ expands the left
kernel of $C_{B}^{\Delta0}\mathfrak{N}_{~\beta}^{Bi}k_{i}$, we conclude tha
\[
M_{\Delta}^{\tilde{\Delta}j}k_{j}\delta\psi_{\pi_{i}\left( k\right)
}^{\Delta}=M_{\Delta}^{\tilde{\Delta}j}k_{j}C_{B}^{\Delta0}\mathfrak{N
_{~\beta}^{Bi}k_{i}\delta\phi_{\pi_{i}\left( k\right) }^{\beta}=0.
\]
Second, evaluating the expression (\ref{Eq_simbs_1}) at $\lambda=\pi
_{i}\left( k\right) $, multiplying by $\delta\phi_{\pi_{i}\left( k\right)
}^{\beta}$ and recalling the eq. (\ref{eq_phi_pi_1}),\ we obtain tha
\begin{align*}
0 & =\left( k_{j}C_{A}^{\Gamma0}\mathfrak{N}_{~\alpha}^{Aj}\right) \left(
-\pi_{i}\left( k\right) \delta_{\alpha}^{\beta}+A_{\alpha}^{\beta i
k_{i}\right) \delta\phi_{\pi_{i}\left( k\right) }^{\alpha},\\
& =\left( -\pi_{i}\left( k\right) \delta_{\Delta}^{\Gamma}+C_{A}^{\Gamma
j}h_{~\Delta}^{A}k_{j}\right) \left( C_{B}^{\Delta0}\mathfrak{N}_{~\alpha
}^{Bj}k_{j}\right) \delta\phi_{\pi_{i}\left( k\right) }^{\alpha},\\
& =\left( -\pi_{i}\left( k\right) \delta_{\Delta}^{\Gamma}+C_{A}^{\Gamma
j}h_{~\Delta}^{A}k_{j}\right) \delta\psi_{\pi_{i}\left( k\right) }^{\Delta
}.
\end{align*}
This shows that equations (\ref{Eq_vin_chi_1}) holds. Let us now demonstrate
that the $\delta\psi_{\pi_{i}\left( k\right) }^{\Delta}$ are linearly
independent. For this purpose, we assume that they are not, i.e. that there
exists $U^{\pi_{i}\left( k\right) }$ such that $0=U^{\pi_{i}\left(
k\right) }\delta\psi_{\pi_{i}\left( k\right) }^{\Delta}$ (the sum runs into
the $i$ index) and we conclude that $U^{\pi_{i}\left( k\right) }=0$. We note
tha
\[
0=U^{\pi_{i}\left( k\right) }\delta\psi_{\pi_{i}\left( k\right) }^{\Delta
}=U^{\pi_{i}\left( k\right) }C_{B}^{\Delta0}\mathfrak{N}_{~\beta}^{Bi
k_{i}\delta\phi_{\pi_{i}\left( k\right) }^{\beta}=C_{B}^{\Delta
0}\mathfrak{N}_{~\beta}^{Bi}k_{i}\left( U^{\pi_{i}\left( k\right)
\delta\phi_{\pi_{i}\left( k\right) }^{\beta}\right) ,
\]
thus $U^{\pi_{i}\left( k\right) }\delta\phi_{\pi_{i}\left( k\right)
}^{\beta}$ belongs to the right-hand kernel of $C_{B}^{\Delta0}\mathfrak{N
_{~\beta}^{Bi}k_{i}$ and, by (\ref{Eq_span_ker_der_C0_Ki_1}), should be a
linear combination of the $\left\{ \delta\phi_{\lambda_{i}\left( k\right)
}^{\beta}\right\} $. Since these $\left\{ \delta\phi_{\lambda_{i}\left(
k\right) }^{\beta}\right\} $ are linearly independent of the $\left\{
\delta\phi_{\pi_{i}\left( k\right) }^{\beta}\right\} $ by construction, it
should be that $U^{\pi_{i}\left( k\right) }=0$.
$\bullet$ \ It remains to show that $s\left( k\right) \times L_{1
^{T},y\left( k\right) \times L_{0}^{T}$ is the other part of the Kronecker
structure. This can be concluded directly from lemma \ref{Lemma_2} (in
appendix \ref{Ap_lemmas}) if we prove the condition (\ref{eq_d1_d2_d_1}). This
condition follows from recalling that
\[
rigth\_\ker\left( M_{\Delta}^{\tilde{\Delta}j}k_{j}\right) =\left\langle
\delta\psi_{\pi_{i}\left( k\right) }^{\Delta}\right\rangle .
\]
Furthermore, we conclude that the pairs $\left\{ \pi_{i}\left( k\right)
\right\} $,$\left\{ \delta\psi_{\pi_{i}\left( k\right) }^{\Delta}\right\}
$ are all the generalized eigenvalues and eigenvectors of the pencil.
Notice that we called the "constraints 1" to the eigenvalues $\left\{ \pi
_{i}\left( k\right) \right\} $, they are at the same time the generalized
eigenvalues of the system (\ref{Eq_pencil_cons_1}), i.e. they are the
"constraints 1" eigenvalues of $A_{\beta}^{\alpha i}k_{i}$ and the "physic"
eigenvalues for the pencil (\ref{Eq_pencil_cons_1}).
\subsubsection{Basis which diagonalize $B_{~\Delta}^{\Gamma j}k_{j}$
\ \label{Diag_sub}}
The previous subsubsection is the proof of a) and b) (stated at the beginning
of this proof). Therefore, there exists $N_{\tilde{\Delta}}^{\Gamma}$ such
that $B_{~\Delta}^{\Gamma j}k_{j}=C_{A}^{\Gamma j}h_{~\Delta}^{A
k_{j}+N_{\tilde{\Delta}}^{\Gamma}M_{\Delta}^{\tilde{\Delta}j}k_{j}$ is
uniformly diagonalizable. This concludes the proof of the theorem in the case
where $M_{\Delta}^{\tilde{\Delta}j}$ is non zero.
In this subsubsection, we will find the bases that diagonalize $B_{~\Delta
}^{\Gamma j}k_{j}$. We will use these bases with the bases that diagonalize
$A_{\beta}^{\alpha j}k_{j}$ to give a very simple expression of the equation
(\ref{Eq_rel_simbolos_1}). Finally, we will discuss the structure of the
latter equation in the cases where $A_{\beta}^{\alpha j}k_{j}$ and/or
$B_{~\Delta}^{\Gamma j}k_{j}$ are not diagonalizable.
Analogously to the case of $A_{\beta}^{\alpha i}k_{i}$, the eigenvectors which
diagonalize $B_{~\Delta}^{\Gamma j}k_{j}$ are also divided into two groups for
each $k_{i}$:
$\bullet$ The eigenvectors $\left\{ \delta\psi_{\pi_{i}\left( k\right)
}^{\Delta}\right\} $ with their corresponding eigenvalues $\pi_{i}\left(
k\right) $, with $i=1,...,r\left( k\right) $, founded in the previous
subsubsection (see (\ref{Eq_chi_pi_1})) and such that they satisf
\[
\left( -\delta_{\Delta}^{\Gamma}\pi_{i}\left( k\right) +C_{A}^{\Gamma
j}h_{~\Delta}^{A}k_{j}+N_{\tilde{\Delta}}^{\Gamma}M_{\Delta}^{\tilde{\Delta
j}k_{j}\right) \delta\psi_{\pi_{i}\left( k\right) }^{\Delta}=0.
\]
$\bullet$ And the eigenvectors $\delta\psi_{\rho_{i}\left( k\right)
}^{\Delta}$ associated to the eigenvalues $\rho_{i}\left( k\right) $, with
$i=1,..,s\left( k\right) $, such tha
\begin{align*}
\left( -\delta_{\Delta}^{\Gamma}\rho_{i}\left( k\right) +C_{A}^{\Gamma
j}h_{~\Delta}^{A}k_{j}+N_{\tilde{\Delta}}^{\Gamma}M_{\Delta}^{\tilde{\Delta
j}k_{j}\right) \delta\psi_{\rho_{i}\left( k\right) }^{\Delta} & =0\\
M_{\Delta}^{\tilde{\Delta}j}k_{j}\delta\psi_{\rho_{i}\left( k\right)
}^{\Delta} & \neq0.
\end{align*}
Where the set $\left\{ \rho_{i}\left( k\right) \right\} $ are simple,
different from each other and different from the $\left\{ \pi_{i}\left(
k\right) \right\} $. These $\left\{ \rho_{i}\left( k\right) \right\} $
was called "constraints 2".
The set $\left\{ \delta\psi_{\pi_{i}\left( k\right) }^{\Delta}\text{,
}\delta\psi_{\rho_{i}\left( k\right) }^{\Delta}\right\} $ uniformly
diagonalizes $B_{~\Delta}^{\Gamma j}k_{j}$ (since, for each $k_{i}$, all its
eigenvalues $\left\{ \pi_{i}\left( k\right) ,\rho_{i}\left( k\right)
\right\} $ are simple) showing that the system (\ref{Eq_sist_B_1}) is
strongly hyperbolic for this choice of $N_{\tilde{\Delta}}^{\Gamma}$.
Let us now rewrite the equation (\ref{Eq_rel_simbolos_1}) in the founded bases.
$\bullet$ For each $k_{i}$, we consider the co-basis $\left\{ \delta
\phi_{\beta}^{\pi_{i}\left( k\right) }\text{, }\delta\phi_{\beta
^{\lambda_{i}\left( k\right) }\right\} $ and $\left\{ \delta\psi_{\Delta
}^{\pi_{i}\left( k\right) }\text{, }\delta\psi_{\Delta}^{\rho_{i}\left(
k\right) }\right\} $ of the bases $\left\{ \delta\phi_{\pi_{i}\left(
k\right) }^{\beta}\text{, }\delta\phi_{\lambda_{i}\left( k\right) }^{\beta
}\right\} $and $\left\{ \delta\psi_{\pi_{i}\left( k\right) }^{\Delta
}\text{, }\delta\psi_{\rho_{i}\left( k\right) }^{\Delta}\right\} $
respectively. They satisfy
\begin{align*}
&
\begin{array}
[c]{ccc
\delta\phi_{\beta}^{\lambda_{i}\left( k\right) }\delta\phi_{\lambda
_{j}\left( k\right) }^{\beta}=\delta_{j}^{i} & & \delta\phi_{\beta
^{\pi_{i}\left( k\right) }\delta\phi_{\lambda_{j}\left( k\right) }^{\beta
}=0
\end{array}
\\
&
\begin{array}
[c]{ccc
\delta\phi_{\beta}^{\lambda_{i}\left( k\right) }\delta\phi_{\pi_{j}\left(
k\right) }^{\beta}=0 & & \delta\phi_{\beta}^{\pi_{i}\left( k\right)
}\delta\phi_{\pi_{j}\left( k\right) }^{\beta}=\delta_{j}^{i
\end{array}
\end{align*
\begin{align*}
&
\begin{array}
[c]{ccc
\delta\psi_{\Delta}^{\pi_{i}\left( k\right) }\delta\psi_{\pi_{j}\left(
k\right) }^{\Delta}=\delta_{j}^{i} & & \delta\psi_{\Delta}^{\rho_{i}\left(
k\right) }\delta\psi_{\pi_{j}\left( k\right) }^{\Delta}=0
\end{array}
\\
&
\begin{array}
[c]{ccc
\delta\psi_{\Delta}^{\pi_{i}\left( k\right) }\delta\psi_{\rho_{j}\left(
k\right) }^{\Delta}=0 & & \delta\psi_{\Delta}^{\rho_{i}\left( k\right)
}\delta\psi_{\rho_{j}\left( k\right) }^{\Delta}=\delta_{j}^{i
\end{array}
\end{align*}
In these bases, the equation (\ref{Eq_rel_simbolos_1}) reduces to
\begin{equation}
\left[
\begin{array}
[c]{cc
I & 0\\
0 & 0
\end{array}
\right] \left[
\begin{array}
[c]{cc
\Pi & 0\\
0 & \Lambda
\end{array}
\right] =\left[
\begin{array}
[c]{cc
\Pi & 0\\
0 & \Theta
\end{array}
\right] \left[
\begin{array}
[c]{cc
I & 0\\
0 & 0
\end{array}
\right] \label{Eq_bloques_1
\end{equation}
where \newline
$I=\left[
\begin{array}
[c]{ccc
1 & 0 & 0\\
0 & ... & 0\\
0 & 0 & 1
\end{array}
\right] \i
\mathbb{R}
^{r\left( k\right) \times r\left( k\right) }$, $\Pi=-\left[
\begin{array}
[c]{ccc
\lambda-\pi_{1}\left( k\right) & 0 & 0\\
0 & ... & 0\\
0 & 0 & \lambda-\pi_{r}\left( k\right)
\end{array}
\right] \i
\mathbb{R}
^{r\left( k\right) \times r\left( k\right) }$, $\Lambda=-\left[
\begin{array}
[c]{ccc
\lambda-\lambda_{1}\left( k\right) & 0 & 0\\
0 & ... & 0\\
0 & 0 & \lambda-\lambda_{d}\left( k\right)
\end{array}
\right] \i
\mathbb{R}
^{d\left( k\right) \times d\left( k\right) }$ \ and \newline $\ \Theta=-\left[
\begin{array}
[c]{ccc
\lambda-\rho_{1}\left( k\right) & 0 & 0\\
0 & ... & 0\\
0 & 0 & \lambda-\rho_{s}\left( k\right)
\end{array}
\right] \i
\mathbb{R}
^{s\left( k\right) \times s\left( k\right) }$.
If we consider the not strongly hyperbolic case, where $\Lambda$ has at least
one Jordan block $J_{m}$ with $m\geq2$. Eq. (\ref{Eq_bloques_1}) implies that
the matrix $\left[
\begin{array}
[c]{cc
\Pi & 0\\
0 & \Theta
\end{array}
\right] $ could still be diagonalizable and hence the subsidiary system would
be strongly hyperbolic.
On the other hand, in the case where $\Lambda$ shares some eigenvalue with
$\Pi$ forming a Jordan block, for example
\[
-\left[
\begin{array}
[c]{cc
\Pi & 0\\
0 & \Lambda
\end{array}
\right] =\left[
\begin{array}
[c]{cccccc
\lambda-\pi_{1}\left( k\right) & 0 & 0 & 0 & 0 & 0\\
0 & ... & 0 & 0 & 0 & 0\\
0 & 0 & \lambda-\lambda_{1}\left( k\right) & 0 & 0 & 0\\
0 & 0 & 1 & \lambda-\lambda_{1}\left( k\right) & 0 & 0\\
0 & 0 & 0 & 0 & ... & 0\\
0 & 0 & 0 & 0 & 0 & \lambda-\lambda_{d}\left( k\right)
\end{array}
\right] ,
\]
by the eq. (\ref{Eq_bloques_1}), it could still happen that the subsidiary
system is diagonalizable.
We concluded that by proposing different options for \ $\left[
\begin{array}
[c]{cc
\Pi & 0\\
0 & \Lambda
\end{array}
\right] $ and $\left[
\begin{array}
[c]{cc
\Pi & 0\\
0 & \Theta
\end{array}
\right] $ such that they satisfy (\ref{Eq_bloques_1}) we obtain all possible
cases where the evolution equations and the subsidiary system are well-posed
or not.
\subsubsection{Case without $M_{\Delta}^{\tilde{\Delta}j}$}
Let us now study the case where the system (\ref{Eq_sis_1_pseudo_1}) does not
admit $M_{\Delta}^{\tilde{\Delta}j}$, i.e., by condition v), $C_{A}^{\Gamma
0}\mathfrak{N}_{~\alpha}^{Ai}k_{i}$ only has trivial left kernel for any
$k_{i}$. In this case, $s\left( k\right) =0$ and $c=r\left( k\right) $ by
equation (\ref{Eq_c_r_s_2}).
The proof of the theorem is the same as presented above, but now we conclude
that the set $\left\{ \delta\psi_{\pi_{i}\left( k\right) }^{\Delta
:=C_{B}^{\Delta0}\mathfrak{N}_{~\beta}^{Bi}k_{i}\delta\phi_{\pi_{i}\left(
k\right) }^{\beta}\right\} $ uniformly diagonalizes the matrix $B_{~\Delta
}^{\Gamma j}k_{j}:=C_{A}^{\Gamma j}h_{~\Delta}^{A}k_{j}$. Therefore, system
(\ref{Eq_sist_B_1}) is strongly hyperbolic. This concludes the proof of the theorem.
\subsection{Comment about condition v) in theorem SH of the SS \label{teo_3}}
The condition v) of the theorem can be suppressed without losing the strong
hyperbolicity of the subsidiary system in the case of constant coefficients.
For this purpose, we have to change the pencil
\[
\left[
\begin{array}
[c]{c
-\lambda\delta_{\Delta}^{\Gamma}+\left( C_{A}^{\Gamma j}h_{~\Delta
^{A}\right) k_{j}\\
M_{\Delta}^{\tilde{\Delta}j}k_{j
\end{array}
\right] \rightarrow\left[
\begin{array}
[c]{cc
-C_{A}^{\Gamma0}\mathfrak{N}_{~\alpha}^{Aj}k_{j} & -\lambda\delta_{\Delta
}^{\Gamma}+C_{A}^{\Gamma j}h_{~\Delta}^{A}k_{j}\\
0 & M_{\Delta}^{\tilde{\Delta}j}k_{j}\\
0 & X_{\Delta}^{\check{s}}\left( k\right)
\end{array}
\right] ,
\]
as explained in subsubsection \ref{Subsi_Fourier} and carry on the same proof
as before but with this new pencil. This change is introduced since
$M_{\Delta}^{\tilde{\Delta}j}k_{j}$ no longer spans the $left\_\ker\left(
C_{B}^{\Delta0}\mathfrak{N}_{~\beta}^{Bi}k_{i}\right) $, so it is necessary
to add the vectors $X_{\Delta}^{\check{s}}\left( k\right) $ such that, for
each $k_{i}$, it holds that
\[
span\left\langle M_{\Delta}^{\tilde{\Delta}j}k_{j},X_{\Delta}^{\check{s
}\left( k\right) \right\rangle =left\_\ker\left( C_{B}^{\Delta
0}\mathfrak{N}_{~\beta}^{Bi}k_{i}\right) .
\]
The cost of this change is reflected in the principal symbol of the subsidiary
system, whose new form i
\begin{equation}
B_{~\Delta}^{\Gamma j}k_{j}=C_{A}^{\Gamma j}h_{~\Delta}^{A}k_{j
+N_{1\tilde{\Delta}}^{\Gamma}M_{\Delta}^{\tilde{\Delta}j}k_{j}+N_{2\check{s
}^{\Gamma}X_{\Delta}^{\check{s}}\left( k\right) . \label{eq_B_new_1
\end{equation}
This change the statement of the theorem.
\begin{theorem}
When all the hypotheses of theorem \ref{Theorem_coef_const_2} are satisfied
except condition v), it is possible to find $N_{1\tilde{\Delta}}^{\Gamma}$ and
$N_{2\check{s}}^{\Gamma}$ such that the principal symbol of the subsidiary
system eq. (\ref{eq_B_new_1}) is uniformly diagonalizable and therefore the
subsidiary system is strongly hyperbolic.
\end{theorem}
Since the $X_{\Delta}^{\check{s}}\left( k\right) $ do not come from Geroch
fields, they could be non-zero only for some particular $k_{i}$ or they could
have a non-linear dependence on $k_{i}$. This latter case would imply that the
Fourier anti-transform of the term $N_{2\check{s}}^{\Gamma}X_{\Delta
^{\check{s}}\left( k\right) $ include second or higher derivatives for any
$N_{2\check{s}}^{\Gamma}$ (except for $N_{2\check{s}}^{\Gamma}=0$). This
changes the final hyperbolic answer of the subsidiary system into a pure
pseudo-differential answer, which can not be directly extrapolated to the
quasi-linear case. Some known systems admit vectors $X_{\Delta}^{\check{s
}\left( k\right) $, but in general, it is possible to set $N_{2\check{s
}^{\Gamma}=0$ and find some $N_{1\tilde{\Delta}}^{\Gamma}$ such that
(\ref{eq_B_new_1}) is uniformly diagonalizable.
\section{Examples \label{Examples}}
In this section, we reproduce some known results about the constraint
propagations of two specific theories: Maxwell electrodynamics and the wave
equation. We use these theories to illustrate the results presented in the
previous sections.
We consider the systems on a space-time $M$ of $\dim M=3+1$ with a background
Lorentzian metric $g_{ab}$ \ (with signature $-,+,+,+$) and with their
equations in first-order derivatives (i.e. eqs. (\ref{eq_max_or_1
-\ref{eq_max_or_2}) and (\ref{Eq_W_E_1})).
We show that in the Maxwell case there are no $M_{A}^{\Delta a}$ fields and
the $C_{A}^{\Gamma a}$ fields are associated with the standard constraints
$\psi_{1}:=D_{a}E^{a}-\tilde{J}^{0}$ and $\psi_{2}=D_{a}B^{a}$. We present its
subsidiary equations and comment on its characteristic analysis. We also note
that it is commonly used in the literature that when the equations are coupled
to a source $J^{a}$, this has to satisfy an on-shell integrability condition
$\nabla_{a}J^{a}=0$ to preserve the constraints. However, we will show that
this condition is relaxed in the off-shell case, by choosing the divergence
proportional to the equations of the system (i.e. \ref{eq_max_id_dJ}) and
maintaining the preservation of the constraints.
On the other hand, for the wave equation, we find both $M_{A}^{\tilde{\Delta
}a}$ and $C_{A}^{\Gamma a}$ fields. They appear as a result of reducing the
system from second to a first-order derivative. As we explained in theorem
\ref{teorema_ec_ev_vin_1}, the non-uniqueness of the subsidiary system is
associated with the presence of $M_{A}^{\tilde{\Delta}a}$. Therefore, we
verify this non-uniqueness and comment on its characteristic analysis.
In both cases, we introduce (as in subsection \ref{n+1_decomposition_sec}) a
foliation of $M=\underset{t\i
\mathbb{R}
}{\cup}\Sigma_{t}$ associated to the function $t:M\rightarro
\mathbb{R}
$, with the spatial (with respect to the metric $g_{ab}$) hypersurfaces
$\Sigma_{t}$. In addition, following appendix \ref{App_coordenadas}, we
consider the definition
\begin{align*}
n_{a} & :=\nabla_{a}t,\\
\tilde{n}_{b} & :=-Nn_{b}\text{\ \ \ with \ }N:=\frac{1}{\sqrt{-\nabla
t.\nabla t}},\\
p^{a} & =\left( \partial_{t}\right) ^{a}-\beta^{a}\text{ \ \ with }\left(
\partial_{t}\right) ^{a}n_{a}=1\text{ and }\beta^{a}n_{a}=0\\
\tilde{m}^{a} & :=\tilde{n}^{a}=\frac{1}{N}p^{a},\text{ }\\
\tilde{n}^{a}\tilde{n}_{a} & =-1,
\end{align*}
and the expression (\ref{Eq_met_1}) for the metric $g_{ab}$. We also consider
the projection $\tilde{\eta}_{b}^{a}$ to the hypersurfaces $\Sigma_{t}$ a
\[
\tilde{\eta}_{b}^{a}:=\delta_{b}^{a}-p^{a}n_{b}=\delta_{b}^{a}+\tilde{n
^{a}\tilde{n}_{b}.
\]
\subsection{Maxwell electrodynamics\label{S_Maxwell_eq}}
We define the fields $Q_{1,2}^{d}$ as
\begin{align*}
Q_{1}^{d} & :=\nabla_{a}F^{ad}-J^{d},\\
Q_{2}^{d} & :=\nabla_{a}\ast F^{ad},
\end{align*}
where $F^{ad}$ is the electromagnetic (antisymmetric) tensor, $\ast
F^{ad}:=\frac{1}{2}\varepsilon_{cq}^{~\ \ \ ad}F^{cq}$ and $J^{d}$ is the
source of the system. Here, $J^{d}=J^{d}\left( F^{ad},\ast F^{ad
,x^{a}\right) $ may depend on $F^{ad}$, $\ast F^{ad}$ and $x^{a}\in M$ but it
can not depend on derivatives of $F^{ad}$ or $\ast F^{ad}$. In addition, we
assume that it\ satisfies the off-shell identity
\begin{equation}
\nabla_{d}J^{d}=L_{1d}Q_{1}^{d}+L_{2d}Q_{2}^{d}, \label{eq_max_id_dJ
\end{equation}
As $J^{d}$, the fields $L_{1,2d}\left( F^{ad},\ast F^{ad},x^{a}\right) $ do
not depend on derivatives of $F^{ad}$ or $\ast F^{ad}$.
We notice two extra off-shell identitie
\[
\nabla_{d}\nabla_{a}F^{ad}=0=\nabla_{d}\nabla_{a}\ast F^{ad},
\]
(both easy to verify). These expressions, in addition with (\ref{eq_max_id_dJ
),\ give the off-shell identities
\begin{align}
\nabla_{d}Q_{1}^{d} & =-L_{1d}Q_{1}^{d}-L_{2d}Q_{2}^{d
,\label{eq_max_id_dQ1}\\
\nabla_{d}Q_{2}^{d} & =0. \label{eq_max_id_dQ2
\end{align}
Multiplying by $-N$ and $N$, these equations can be rewritten as
\begin{equation}
\nabla_{d}\left( \left[
\begin{array}
[c]{c
-NQ_{1}^{d}\\
NQ_{2}^{d
\end{array}
\right] \right) +\left[
\begin{array}
[c]{cc
\nabla_{d}\left( N\right) +NL_{1d} & NL_{2d}\\
0 & -\nabla_{d}\left( N\right)
\end{array}
\right] \left[
\begin{array}
[c]{c
Q_{1}^{d}\\
Q_{2}^{d
\end{array}
\right] =0 \label{eq_max_id_mat_1
\end{equation}
These latter equations are the integrability conditions (\ref{eq_int_LE_1})
from which the evolution equations of the constraints are obtained as we
explain below.
The Maxwell equations are defined by $Q_{1,2}^{d}$ a
\begin{align}
Q_{1}^{d} & =\nabla_{a}F^{ad}-J^{d}=0,\label{eq_max_or_1}\\
Q_{2}^{d} & =\nabla_{a}\ast F^{ad}=0. \label{eq_max_or_2
\end{align}
We will use the electric $E^{c}$ and magnetic $B_{e}$ fields as the variables
of the system. For this purpose, we begin by rewriting $F^{cq}$ in terms of
$E^{c}$ and $B_{e}$
\[
F^{cq}=\tilde{m}^{c}E^{q}-\tilde{m}^{q}E^{c}+\varepsilon^{cqde}\tilde{n
_{d}B_{e}.
\]
These fields are defined by
\[
E^{c}:=F^{cq}\tilde{n}_{q}\text{, \ \ }B_{d}:=\ast F_{ld}\tilde{m}^{l
\]
and they belong to the tangent of $\Sigma_{t}$, since
\[
\tilde{n}_{d}E^{d}=0=\tilde{m}^{e}B_{e}.
\]
On the other hand, the dual $\ast F_{lm}$ can be written as
\[
\ast F_{lm}=-\tilde{n}_{l}B_{m}+\tilde{n}_{m}B_{l}+\varepsilon_{cqlm}\tilde
{m}^{c}E^{q}.
\]
Following the same steps as in appendix \ref{App_coordenadas} and the
definitions (\ref{Eq_apendix_T_3}-\ref{Eq_apendix_T_8}), we rewrite
$Q_{1,2}^{d}$ a
\begin{align}
Q_{1}^{d} & =\tilde{\eta}_{l}^{d}\frac{1}{N}\tilde{e}_{1}^{l}-\tilde{m
^{d}\psi_{1},\label{eq_Q1_E_B_1}\\
Q_{2}^{d} & =-\tilde{\eta}_{l}^{d}\frac{1}{N}\tilde{e}_{2}^{l}+\tilde{m
^{d}\psi_{2}, \label{eq_Q2_E_B_1
\end{align}
where
\begin{align}
\tilde{J}^{l} & :=\tilde{\eta}_{c}^{l}J^{c},\text{ \ \ \ \ }\tilde{J
^{0}=\tilde{n}_{c}J^{c},\label{eq_J_1}\\
\tilde{e}_{1}^{l} & :
\mathcal{L
_{p}E^{l}+\varepsilon^{clde}\tilde{n}_{d}D_{c}\left( NB_{e}\right)
-NKE^{l}-N\tilde{J}^{l},\label{eq_e1_1}\\
\tilde{e}_{2}^{l} & :
\mathcal{L
_{p}B^{l}-\varepsilon^{clde}\tilde{n}_{d}D_{c}\left( NE_{e}\right)
-NKB^{l},\label{eq_e1_2}\\
\psi_{1} & =D_{a}E^{a}-\tilde{J}^{0},\label{eq_phi1_1}\\
\psi_{2} & =D_{a}B^{a}. \label{eq_phi2_1
\end{align}
Here, the expressions (\ref{eq_J_1}-\ref{eq_phi2_1}) are quantities over
$\Sigma_{t}$ by definition, which means that $\tilde{\eta}_{l}^{d}\tilde
{e}_{1,2}^{l}=\tilde{e}_{1,2}^{d}$ and $\tilde{\eta}_{l}^{d}\tilde{J
^{l}=\tilde{J}^{d}$.
Notice that $\tilde{e}_{1,2}^{d}$ and $\psi_{1,2}$ are the evolution and the
constraints equations\ of the system respectively. $\psi_{1,2}$ act as
constraints since they have not derivatives in the $\partial_{t}$ direction.
We will see below that these constraints are preserved in the evolutions
$\tilde{e}_{1,2}^{l}=0$ when the initial data $\left. \phi\right\vert
_{\Sigma_{0}}$ is chosen such that $\left. \psi_{1,2}\right\vert _{\Sigma
_{0}}=0$.
Using the expressions (\ref{eq_Q1_E_B_1}, \ref{eq_Q2_E_B_1}, \ref{eq_e1_1
-\ref{eq_phi2_1}), we conclude that the principal symbol of the system i
\begin{align}
\left[
\begin{array}
[c]{c
Q_{1}^{s}\\
Q_{2}^{s
\end{array}
\right] & \approx\mathfrak{N}_{~\beta}^{Bq}\nabla_{q}\phi^{\beta
},\nonumber\\
& =\left[
\begin{array}
[c]{cc
2\delta_{b}^{[s}\tilde{m}^{q]} & \varepsilon_{~\ \ \ \ b}^{qsd}\tilde{n}_{d}\\
\varepsilon_{~\ \ \ \ b}^{qsd}\tilde{n}_{d} & -2\delta_{b}^{[s}\tilde{m}^{q]
\end{array}
\right] \nabla_{q}\left[
\begin{array}
[c]{c
E^{b}\\
B^{b
\end{array}
\right] . \label{eq_max_K_1
\end{align}
We are considering
\
\mathcal{L
_{p}E^{r}\approx\tilde{\eta}_{b}^{r}p^{q}\nabla_{q}E^{b},
\]
neglecting lower order terms.
Associated with this system, we have the Geroch fields (see eq. (\ref{eq_CK_1
)
\begin{equation}
C_{B}^{\Delta z}=\left[
\begin{array}
[c]{cc
-N\delta_{s}^{z} & 0\\
0 & N\delta_{s}^{z
\end{array}
\right] . \label{eq_max_Cz_1
\end{equation}
Contracting it with $n_{z}$, we obtain
\[
n_{z}C_{B}^{\Delta z}=C_{B}^{\Delta0}=\left[
\begin{array}
[c]{cc
\tilde{n}_{s} & 0\\
0 & -\tilde{n}_{s
\end{array}
\right] .
\]
This last expression gives the constraints $\psi_{1,2}$ when it is contracted
with $\left[
\begin{array}
[c]{c
Q_{1}^{s}\\
Q_{2}^{s
\end{array}
\right] $.
On the other hand, we choose the following reduction
\[
h_{B}^{\alpha}=\left[
\begin{array}
[c]{cc
N\tilde{\eta}_{s}^{l} & 0\\
0 & -N\tilde{\eta}_{s}^{l
\end{array}
\right] .
\]
This reduction leads to the standard symmetric hyperbolic Maxwell evolution
equations $e_{1,2}^{l}$ when it is contracted with $\left[
\begin{array}
[c]{c
Q_{1}^{s}\\
Q_{2}^{s
\end{array}
\right] $. Combining the last two expressions, we conclude that
\begin{equation}
\left[
\begin{array}
[c]{c
h_{B}^{\alpha}\\
C_{B}^{\Delta0
\end{array}
\right] =\left[
\begin{array}
[c]{cc
N\tilde{\eta}_{s}^{l} & 0\\
0 & -N\tilde{\eta}_{s}^{l}\\
\tilde{n}_{s} & 0\\
0 & -\tilde{n}_{s
\end{array}
\right] . \label{eq_max_h_C0_1
\end{equation}
Then by eqs. (\ref{eq_Q1_E_B_1}) and (\ref{eq_Q2_E_B_1}), we obtain
\begin{equation}
\left[
\begin{array}
[c]{c
h_{B}^{\alpha}\\
C_{B}^{\Delta0
\end{array}
\right] E^{B}=\left[
\begin{array}
[c]{cc
N\tilde{\eta}_{s}^{l} & 0\\
0 & -N\tilde{\eta}_{s}^{l}\\
\tilde{n}_{s} & 0\\
0 & -\tilde{n}_{s
\end{array}
\right] \left[
\begin{array}
[c]{c
Q_{1}^{s}\\
Q_{2}^{s
\end{array}
\right] =\left[
\begin{array}
[c]{c
\tilde{e}_{1}^{l}\\
\tilde{e}_{2}^{l}\\
\psi_{1}\\
\psi_{2
\end{array}
\right] . \label{eq_max_e_cons_1
\end{equation}
The inverse of $\left[
\begin{array}
[c]{c
h_{B}^{\alpha}\\
C_{B}^{\Delta0
\end{array}
\right] $ has the following form
\begin{equation}
\left[
\begin{array}
[c]{cc
\mathfrak{N}_{~\alpha}^{A0} & h_{~\Delta}^{A
\end{array}
\right] =\left[
\begin{array}
[c]{cccc
\frac{1}{N}\tilde{\eta}_{l}^{s} & 0 & -\tilde{m}^{s} & 0\\
0 & -\frac{1}{N}\tilde{\eta}_{l}^{s} & 0 & \tilde{m}^{s
\end{array}
\right] , \label{eq_max_inv_1
\end{equation}
and, of course, satisfies that
\begin{align}
\left[
\begin{array}
[c]{cc
\mathfrak{N}_{~\alpha}^{A0} & h_{~\Delta}^{A
\end{array}
\right] \left[
\begin{array}
[c]{c
h_{B}^{\alpha}\\
C_{B}^{\Delta0
\end{array}
\right] & =\delta_{B}^{A},\nonumber\\
\left[
\begin{array}
[c]{cccc
\frac{1}{N}\tilde{\eta}_{l}^{s} & 0 & -\tilde{m}^{s} & 0\\
0 & -\frac{1}{N}\tilde{\eta}_{l}^{s} & 0 & \tilde{m}^{s
\end{array}
\right] \left[
\begin{array}
[c]{cc
N\tilde{\eta}_{s}^{l} & 0\\
0 & -N\tilde{\eta}_{s}^{l}\\
\tilde{n}_{s} & 0\\
0 & -\tilde{n}_{s
\end{array}
\right] & =\left[
\begin{array}
[c]{cc
\delta_{s}^{l} & 0\\
0 & \delta_{s}^{l
\end{array}
\right] . \label{eq_p_p_1_1
\end{align}
Using this expression and the equations (\ref{eq_p_p_1_1}) and
(\ref{eq_max_e_cons_1}), we conclude that
\begin{align*}
\nabla_{d}\left( C_{B}^{\Delta d}E^{B}\right) & =\nabla_{d}\left( \left(
C_{A}^{\Delta d}\left[
\begin{array}
[c]{cc
\mathfrak{N}_{~\alpha}^{A0} & h_{~\Gamma}^{A
\end{array}
\right] \right) \left( \left[
\begin{array}
[c]{c
h_{B}^{\alpha}\\
C_{B}^{\Gamma0
\end{array}
\right] E^{B}\right) \right) \\
& \nabla_{d}\left( \left[
\begin{array}
[c]{cc
-N\delta_{s}^{d} & 0\\
0 & N\delta_{s}^{d
\end{array}
\right] \left[
\begin{array}
[c]{c
Q_{1}^{s}\\
Q_{2}^{s
\end{array}
\right] \right) \\
& =\nabla_{d}\left( \left[
\begin{array}
[c]{cccc
-\tilde{\eta}_{l}^{s} & 0 & N\tilde{m}^{s} & 0\\
0 & -\tilde{\eta}_{l}^{s} & 0 & N\tilde{m}^{s
\end{array}
\right] \left[
\begin{array}
[c]{c
\tilde{e}_{1}^{l}\\
\tilde{e}_{2}^{l}\\
\psi_{1}\\
\psi_{2
\end{array}
\right] \right) ,
\end{align*}
which allows to rewrite the identities (\ref{eq_max_id_mat_1}) as
\begin{align*}
0 &
\mathcal{L
_{p}\left( \left[
\begin{array}
[c]{c
\psi_{1}\\
\psi_{2
\end{array}
\right] \right) +\left[
\begin{array}
[c]{cc
-N-N\tilde{m}^{d}L_{1d} & N\tilde{m}^{d}L_{2d}\\
0 & -NK
\end{array}
\right] \left[
\begin{array}
[c]{c
\psi_{1}\\
\psi_{2
\end{array}
\right] \\
& +\left[
\begin{array}
[c]{c
-N\nabla_{d}\left( \frac{1}{N}e_{1}^{d}\right) +L_{1d}\tilde{e}_{1
^{d}-L_{2d}\tilde{e}_{2}^{d}\\
-N\nabla_{d}\left( \frac{1}{N}e_{2}^{d}\right)
\end{array}
\right] .
\end{align*}
On-shell (i.e., when $e_{1,2}^{d}=0$), these equations lead to the subsidiary
system
\begin{equation}
0
\mathcal{L
_{p}\left( \left[
\begin{array}
[c]{c
\psi_{1}\\
\psi_{2
\end{array}
\right] \right) +\left[
\begin{array}
[c]{cc
-N-N\tilde{m}^{d}L_{1d} & N\tilde{m}^{d}L_{2d}\\
0 & -NK
\end{array}
\right] \left[
\begin{array}
[c]{c
\psi_{1}\\
\psi_{2
\end{array}
\right] . \label{Eq_sub_max_1
\end{equation}
Clearly (since there are no spatial derivatives of $\psi_{1,2}$) this system
is strongly hyperbolic, so it has a unique solution for a given initial data.
Since the initial data is chosen such that $\left. \psi_{1,2}\right\vert
_{\Sigma_{0}}=0$ and $\psi_{1,2}=0$ is a solution of the subsidiary system, we
conclude (by the uniqueness of the solutions) that the constraints are
preserved during the evolutions $e_{1,2}^{d}=0$.
We now consider the relationship between the principal symbols of the
evolution equations and the principal symbols of the constraints equations.
Using equations (\ref{eq_max_K_1}), (\ref{eq_max_h_C0_1}), (\ref{eq_max_Cz_1})
and (\ref{eq_max_inv_1}), we obtain the matrices $\left[
\begin{array}
[c]{c
h_{B}^{\alpha}\\
C_{B}^{\Gamma0
\end{array}
\right] \mathfrak{N}_{~\beta}^{Bq}$ and $l_{z}C_{A}^{\Delta z}\left[
\begin{array}
[c]{cc
\mathfrak{N}_{~\alpha}^{A0} & h_{~\Gamma}^{A
\end{array}
\right] $, they are
\begin{equation}
\left[
\begin{array}
[c]{c
h_{B}^{\alpha}\\
C_{B}^{\Gamma0
\end{array}
\right] \mathfrak{N}_{~\beta}^{Bq}l_{q}\phi^{\beta}=\left[
\begin{array}
[c]{cc
\tilde{\eta}_{b}^{l}p^{q} & N\varepsilon_{~\ \ \ \ b}^{qld}\tilde{n}_{d}\\
-N\varepsilon_{~\ \ \ \ b}^{qld}\tilde{n}_{d} & \tilde{\eta}_{b}^{l}p^{q}\\
\tilde{\eta}_{b}^{q} & 0\\
0 & \tilde{\eta}_{b}^{q
\end{array}
\right] l_{q}\left[
\begin{array}
[c]{c
E^{b}\\
B^{b
\end{array}
\right] , \label{Eq_max_1
\end{equation}
an
\[
l_{z}C_{A}^{\Delta z}\left[
\begin{array}
[c]{cc
\mathfrak{N}_{~\alpha}^{A0} & h_{~\Gamma}^{A
\end{array}
\right] =l_{z}\left[
\begin{array}
[c]{cccc
-\tilde{\eta}_{l}^{z} & 0 & N\tilde{m}^{z} & 0\\
0 & -\tilde{\eta}_{l}^{z} & 0 & N\tilde{m}^{z
\end{array}
\right] .
\]
Recalling that
\[
l_{z}C_{A}^{\Delta z}\left[
\begin{array}
[c]{cc
\mathfrak{N}_{~\alpha}^{A0} & h_{~\Gamma}^{A
\end{array}
\right] \left[
\begin{array}
[c]{c
h_{B}^{\alpha}\\
C_{B}^{\Gamma0
\end{array}
\right] \mathfrak{N}_{~\beta}^{Bq}l_{q}=0,
\]
we concluded
\begin{align}
& l_{z}\left[
\begin{array}
[c]{cc
-\tilde{\eta}_{l}^{z} & 0\\
0 & -\tilde{\eta}_{l}^{z
\end{array}
\right] \left[
\begin{array}
[c]{cc
\tilde{\eta}_{b}^{l}p^{q} & N\varepsilon_{~\ \ \ \ b}^{qld}\tilde{n}_{d}\\
-N\varepsilon_{~\ \ \ \ b}^{qld}\tilde{n}_{d} & \tilde{\eta}_{b}^{l}p^{q
\end{array}
\right] l_{q}\nonumber\\
& =l_{z}\left[
\begin{array}
[c]{cc
p^{z} & 0\\
0 & p^{z
\end{array}
\right] \left[
\begin{array}
[c]{cc
-\tilde{\eta}_{b}^{q} & 0\\
0 & -\tilde{\eta}_{b}^{q
\end{array}
\right] l_{q}. \label{eq_mas_C_A_P_C_1
\end{align}
This expression is exactly the equation (\ref{Eq_simbs_1}), wher
\[
h_{B}^{\alpha}\mathfrak{N}_{~\beta}^{Bq}l_{q}=\left[
\begin{array}
[c]{cc
\tilde{\eta}_{b}^{l}p^{q} & N\varepsilon_{~\ \ \ \ b}^{qld}\tilde{n}_{d}\\
-N\varepsilon_{~\ \ \ \ b}^{qld}\tilde{n}_{d} & \tilde{\eta}_{b}^{l}p^{q
\end{array}
\right] l_{q
\]
is the principal symbol of $\left[
\begin{array}
[c]{c
e_{1}^{l}\\
e_{2}^{l
\end{array}
\right] $,
\[
l_{z}C_{A}^{\Delta z}h_{~\Gamma}^{A}=l_{z}\left[
\begin{array}
[c]{cc
p^{z} & 0\\
0 & p^{z
\end{array}
\right]
\]
is the principal symbol of the subsidiary system an
\[
C_{B}^{\Gamma0}\mathfrak{N}_{~\beta}^{Bq}l_{q}=\left[
\begin{array}
[c]{cc
\tilde{\eta}_{b}^{q} & 0\\
0 & \tilde{\eta}_{b}^{q
\end{array}
\right] l_{q}.
\]
Let us now study the characteristic structures of the system and the
subsidiary system. We will study the Kronecker decomposition of the pencils
$\left[
\begin{array}
[c]{c
h_{~A}^{\alpha}\mathfrak{N}_{~\alpha}^{Aq}\\
C_{A}^{\Gamma0}\mathfrak{N}_{~\alpha}^{Aq
\end{array}
\right] l_{q}$ and $C_{A}^{\Delta d}h_{~\Gamma}^{A}l_{d}$ with $l_{q
=-\lambda n_{a}+k_{a}$ and $k_{a}\left( \partial_{t}\right) ^{a}=0$. In
their pullback version to $\Sigma_{t}$, these pencils are
\begin{align}
& \left[
\begin{array}
[c]{c
h_{~A}^{\alpha}\mathfrak{N}_{~\alpha}^{Aq}\\
C_{A}^{\Gamma0}\mathfrak{N}_{~\alpha}^{Aq
\end{array}
\right] l_{q}\phi^{\alpha}\nonumber\\
& =\left[
\begin{array}
[c]{cc
\tilde{\eta}_{b}^{l}\left( -\lambda-\left( \beta.k\right) \right) &
N\varepsilon_{~\ \ \ \ b}^{dl}k_{d}\\
-N\varepsilon_{~\ \ \ \ b}^{dl}k_{d} & \left( -\lambda-\left( \beta
.k\right) \right) \tilde{\eta}_{b}^{l}\\
k_{b} & 0\\
0 & k_{b
\end{array}
\right] \left[
\begin{array}
[c]{c
E^{b}\\
B^{b
\end{array}
\right] , \label{Eq_Max_p_1
\end{align
\begin{equation}
C_{A}^{\Delta q}h_{~\Gamma}^{A}l_{q}\psi^{\Gamma}=\left[
\begin{array}
[c]{cc
\left( -\lambda-\left( \beta.k\right) \right) & 0\\
0 & \left( -\lambda-\left( \beta.k\right) \right)
\end{array}
\right] \left[
\begin{array}
[c]{c
\psi_{1}\\
\psi_{2
\end{array}
\right] . \label{Eq_Max_p_2
\end{equation}
Where $\varepsilon^{qwy}:=\tilde{n}_{d}\varepsilon^{dqwy}$ is the Levi-Civita
tensor over $\Sigma_{t}$; and all the lowercase indices $l,b,d$ (and any other
lowercase indices that appear until the end of this section) run from $1$ to
$3$. We also note the size of the matrices $\left[
\begin{array}
[c]{c
h_{~A}^{\alpha}\mathfrak{N}_{~\alpha}^{Aq}\\
C_{A}^{\Gamma0}\mathfrak{N}_{~\alpha}^{Aq
\end{array}
\right] l_{q}$ $\i
\mathbb{R}
^{8\times6}$, $C_{A}^{\Delta q}h_{~\Gamma}^{A}l_{q}\i
\mathbb{R}
^{2\times2}$, $h_{~A}^{\alpha}\mathfrak{N}_{~\alpha}^{Aq}l_{q}\i
\mathbb{R}
^{6\times6}$ and $C_{A}^{\Gamma0}\mathfrak{N}_{~\alpha}^{Aq}l_{q}\i
\mathbb{R}
^{2\times6}$\ in this pullbacked version.
Consider the matrix $\left( C_{B}^{\Delta0}\mathfrak{N}_{~\alpha
^{Bq}\right) l_{q}$ whose expression,
\[
C_{B}^{\Gamma0}\mathfrak{N}_{~\beta}^{Bq}l_{q}\phi^{\beta}=\left[
\begin{array}
[c]{cc
k_{b} & 0\\
0 & k_{b
\end{array}
\right] \left[
\begin{array}
[c]{c
E^{b}\\
B^{b
\end{array}
\right] .
\]
is obtained from (\ref{Eq_Max_p_1}). For each $k_{q}$, we hav
\begin{align}
\dim\left( left\_\ker\left( C_{B}^{\Gamma0}\mathfrak{N}_{~\beta}^{Bq
l_{q}\right) \right) & =0,\label{eq_max_lf_1}\\
rank\left( C_{B}^{\Gamma0}\mathfrak{N}_{~\beta}^{Bq}l_{q}\right) &
=2,\label{eq_max_r_1}\\
\dim\left( right\_\ker\left( C_{B}^{\Delta0}\mathfrak{N}_{~\alpha}^{Bq
k_{q}\right) \right) & =4.
\end{align}
This means that there are not Geroch fields $M_{\Gamma}^{\tilde{\Delta}z}$
such that $l_{z}M_{\Gamma}^{\tilde{\Delta}z}C_{B}^{\Gamma0}\mathfrak{N
_{~\beta}^{Bq}l_{q}=0$ and therefore the condition v) of the theorem
\ref{Theorem_coef_const_2} is satisfied.
This result allows us to give the Kronecker structure associated with the
Maxwell equation, i.e.
\[
2\times J_{1}\left( N\sqrt{k.k}-\beta.k\right) ,2\times J_{1}\left(
-N\sqrt{k.k}-\beta.k\right) ,2\times L_{1}^{T}.
\]
is the Kronecker structure of the pencil (\ref{Eq_Max_p_1}).
The $2\times L_{1}^{T}$ blocks are justified by (\ref{eq_max_lf_1}) and
(\ref{eq_max_r_1}) as explained in subsubsection \ref{Kronecker_decomposition
. The Jordan part is justified by giving explicitly the generalized
eigenvectors
\[
\left( \delta\phi_{\lambda_{1}}^{1}\right) ^{\beta}=\left[
\begin{array}
[c]{c
\frac{1}{\sqrt{\left( v.v\right) }}v^{b}\\
\frac{1}{\sqrt{\left( w.w\right) }}w^{b
\end{array}
\right] ,\text{ \ }\left( \delta\phi_{\lambda_{1}}^{2}\right) ^{\beta
}=\left[
\begin{array}
[c]{c
\frac{1}{\sqrt{\left( w.w\right) }}w^{b}\\
-\frac{1}{\sqrt{\left( v.v\right) }}v^{b
\end{array}
\right]
\
\[
\left( \delta\phi_{\lambda_{2}}^{1}\right) ^{\beta}=\left[
\begin{array}
[c]{c
\frac{1}{\sqrt{\left( v.v\right) }}v^{b}\\
-\frac{1}{\sqrt{\left( w.w\right) }}w^{b
\end{array}
\right] ,\text{ \ }\left( \delta\phi_{\lambda_{2}}^{2}\right) ^{\beta
}=\left[
\begin{array}
[c]{c
\frac{1}{\sqrt{\left( w.w\right) }}w^{b}\\
\frac{1}{\sqrt{\left( v.v\right) }}v^{b
\end{array}
\right]
\]
where $v^{l}$ and $w^{l}$ are linearly independent vectors, defined by
\begin{align*}
2v^{[l}w^{d]} & =\varepsilon^{qlb}k_{q}\\
v.k & =v.w=w.k=0.
\end{align*}
They are associated to the generalized eigenvalues $\lambda_{1}=N\sqrt
{k.k}-\beta.k$ and $\lambda_{2}=-N\sqrt{k.k}-\beta.k$. These eigenvalues were
called "the physical" eigenvalues in subsection
\ref{Proof_Theorem_coef_const_2}.
Now, let us study the characteristic structure of $h_{~A}^{\alpha
\mathfrak{N}_{~\alpha}^{Aq}l_{q}$, whose explicit expression is
\[
h_{~A}^{\alpha}\mathfrak{N}_{~\alpha}^{Aq}l_{q}\phi^{\alpha}=\left[
\begin{array}
[c]{cc
\tilde{\eta}_{b}^{l}\left( -\lambda-\left( \beta.k\right) \right) &
N\varepsilon_{~\ \ \ \ b}^{dl}k_{d}\\
-N\varepsilon_{~\ \ \ \ b}^{dl}k_{d} & \left( -\lambda-\left( \beta
.k\right) \right) \tilde{\eta}_{b}^{l
\end{array}
\right] \left[
\begin{array}
[c]{c
E^{b}\\
B^{b
\end{array}
\right] .
\]
Since $\left( \lambda_{1,2},\left( \delta\phi_{\lambda_{1,2}}\right)
^{\beta}\right) $ are the generalized eigenvalues and eigenvectors of
$\left[
\begin{array}
[c]{c
h_{~A}^{\alpha}\mathfrak{N}_{~\alpha}^{Aq}\\
C_{A}^{\Gamma0}\mathfrak{N}_{~\alpha}^{Aq
\end{array}
\right] l_{q}$,\ \ they are also the eigenvalues and eigenvectors of
$h_{~A}^{\alpha}\mathfrak{N}_{~\alpha}^{Aq}l_{q}$. Moreover, $h_{~A}^{\alpha
}\mathfrak{N}_{~\alpha}^{Aq}l_{q}$ has $2$ additional eigenvalues $\pi
_{1,2}\left( k\right) =-\beta.k$, these were called "the constraints 1"
eigenvalues in subsection \ref{Proof_Theorem_coef_const_2}. \ These
eigenvalues have the associated eigenvectors
\[
\left( \delta\phi_{\pi_{1}}\right) ^{\beta}=\left[
\begin{array}
[c]{c
k^{b}\\
0
\end{array}
\right] \text{ \ \ }\left( \delta\phi_{\pi_{2}}\right) ^{\beta}=\left[
\begin{array}
[c]{c
0\\
k^{b
\end{array}
\right] .
\]
To conclude this subsection, let us now study the Kronecker decomposition of
the pencil $C_{A}^{\Delta q}h_{~\Gamma}^{A}l_{q}$ \ (see eq. (\ref{Eq_Max_p_2
)). This is a square pencil and it is associated to the constraints as can be
seen from (\ref{Eq_sub_max_1}). Its Kronecker decomposition is given by
\[
2\times J_{1}\left( -\beta.k\right) .
\]
As it was shown in subsubsection \ref{Kro_sub}, the generalized eigenvectors
of this pencil are obtained by projecting $\left\{ \left( \delta\phi
_{\pi_{1}}\right) ^{\beta},\left( \delta\phi_{\pi_{2}}\right) ^{\beta
}\right\} $ with $C_{B}^{\Gamma0}\mathfrak{N}_{~\beta}^{Bq}k_{q}$, that is,
\begin{align*}
\left[ \delta\psi_{\pi_{1}}^{\Gamma},\delta\psi_{\pi_{2}}^{\Gamma}\right]
& :=C_{B}^{\Gamma0}\mathfrak{N}_{~\beta}^{Bq}l_{q}\left[ \left( \delta
\phi_{1}\right) ^{\beta},\left( \delta\phi_{\pi_{2}}\right) ^{\beta
}\right] ,\\
& =\left[
\begin{array}
[c]{cc
k_{b} & 0\\
0 & k_{b
\end{array}
\right] \left[
\begin{array}
[c]{cc
k^{b} & 0\\
0 & k^{b
\end{array}
\right] ,\\
& =\left[
\begin{array}
[c]{cc
\left( k.k\right) & 0\\
0 & \left( k.k\right)
\end{array}
\right] .
\end{align*}
Where the eigenvectors are the columns of this matrix and they are associated
to "the constraints 1" eigenvalues $\pi_{1,2,3}\left( k\right) =-\beta.k$.
This result follows from equation (\ref{Eq_simbs_1}) and can be easily
checked, since $C_{A}^{\Delta q}h_{~\Gamma}^{A}l_{q}=0$ when $\lambda
=-\beta.k$.
\subsection{Wave equation \label{S_wave_equation}}
Consider the wave equatio
\begin{equation}
g^{ab}\nabla_{a}\nabla_{b}\phi=0. \label{eq_wave_1
\end{equation}
We lead this equation to first order in derivatives. The wave equation in
first-order is
\begin{equation}
E=0,\text{ \ \ }E_{b}=0,\text{ \ \ }E_{ab}=0. \label{Eq_W_E_1
\end{equation}
where we have defined
\[
u_{b}:=\nabla_{b}\phi,
\]
and
\begin{align*}
E & :=g^{ab}\nabla_{a}u_{b},\\
E_{b} & :=\nabla_{b}\phi-u_{b},\\
E_{ab} & :=\nabla_{\lbrack a}u_{b]}.
\end{align*}
Notice that $E_{ab}$ is obtained from $E_{b}$ by taking an antisymmetric
derivative, i.e. $\nabla_{\lbrack a}E_{b]}=-E_{ab}$. Additionally, there is
another identity for $E_{ab}$, this is $\nabla_{\lbrack c}E_{ab]}=0$
\footnote{To show this result, you need to use the first Bianchi identity.}.
Thus, the off-shell identities of the system ar
\begin{align}
0 & =\nabla_{f}\left( \delta_{a}^{[f}\delta_{b}^{g]}E_{g}\right)
+E_{ab},\label{Eq_W_id_1}\\
0 & =\nabla_{c}\left( \delta_{f}^{[c}\delta_{g}^{a}\delta_{h}^{b]
E_{ab}\right) . \label{Eq_W_id_2
\end{align}
The following four projections of these equations
\[
0=\left[
\begin{array}
[c]{c
2N\tilde{m}^{a}\tilde{\eta}_{r}^{b}\left( \nabla_{f}\left( \delta_{a
^{[f}\delta_{b}^{g]}E_{g}\right) +E_{ab}\right) \\
3N\tilde{m}^{f}\tilde{\eta}_{g}^{d}\tilde{\eta}_{h}^{e}\nabla_{c}\left(
\delta_{f}^{[c}\delta_{d}^{a}\delta_{e}^{b]}E_{ab}\right) \\
\tilde{\eta}_{s}^{a}\tilde{\eta}_{r}^{b}\nabla_{f}\left( \delta_{a
^{[f}\delta_{b}^{g]}E_{g}\right) +E_{ab}\\
\tilde{n}_{d}\varepsilon^{dfgh}\nabla_{c}\left( \delta_{f}^{[c}\delta_{g
^{a}\delta_{h}^{b]}E_{ab}\right)
\end{array}
\right] ,
\]
can be rewritten as follows
\begin{align}
0 & =\nabla_{z}\left( \left[
\begin{array}
[c]{ccc
0 & 2N\tilde{m}^{[z}\tilde{\eta}_{r}^{g]} & 0\\
0 & 0 & 3N\tilde{m}^{[z}\tilde{\eta}_{g}^{a}\tilde{\eta}_{h}^{b]}\\
0 & \tilde{\eta}_{[s}^{z}\tilde{\eta}_{r]}^{g} & 0\\
0 & 0 & \tilde{n}_{d}\varepsilon^{dzab
\end{array}
\right] \left[
\begin{array}
[c]{c
E\\
E_{g}\\
E_{ab
\end{array}
\right] \right) \nonumber\\
& +\left[
\begin{array}
[c]{ccc
0 & -\nabla_{z}\left( 2N\tilde{m}^{[z}\tilde{\eta}_{r}^{g]}\right) &
2N\tilde{m}^{a}\tilde{\eta}_{r}^{b}\\
0 & 0 & -\nabla_{z}\left( 3N\tilde{m}^{[z}\tilde{\eta}_{g}^{a}\tilde{\eta
}_{h}^{b]}\right) \\
0 & -\nabla_{z}\left( \tilde{\eta}_{s}^{[z}\tilde{\eta}_{r}^{g]}\right) &
\tilde{\eta}_{s}^{[a}\tilde{\eta}_{r}^{b]}\\
0 & 0 & -\nabla_{z}\left( \tilde{n}_{d}\varepsilon^{dzab}\right)
\end{array}
\right] \left[
\begin{array}
[c]{c
E\\
E_{g}\\
E_{ab
\end{array}
\right] . \label{eq_wav_id_dCE_1
\end{align}
As we will show, these expressions are exactly the evolution equations of the
constraints and the constraints of the constraints of the system. The latter
is associated to the Geroch fields $M_{\Gamma}^{\tilde{\Delta}z}$.
We now introduce new variables
\begin{align*}
\tilde{u}^{0} & :=\tilde{n}_{b}u^{b},\\
\tilde{u}_{d} & :=\tilde{\eta}_{db}u^{b},
\end{align*}
where $\tilde{u}_{d}$ results from projecting $u^{b}$ onto $\Sigma_{t}$. Since
we are interested in describing the system with the variables $\left(
\tilde{u}^{0},\text{ }\phi,\text{ }\tilde{u}_{w}\right) $, we rewrite the
equations (\ref{Eq_W_E_1}) as follows
\begin{align}
E & =-\frac{1}{N}\tilde{e}_{1},\label{eq_wav_E_1}\\
E_{c} & =-\frac{1}{N}\tilde{n}_{c}\tilde{e}_{2}+\psi_{1c},\label{eq_wav_E_2
\\
E_{ac} & =\frac{1}{N}\tilde{e}_{3[a}\tilde{n}_{c]}+\psi_{2ac},
\label{eq_wav_E_3
\end{align}
where
\begin{align}
\tilde{e}_{1} & :
\mathcal{L
_{p}\tilde{u}^{0}-ND_{d}\tilde{u}^{d}-N\left( \tilde{u}^{w}S_{w}\right)
-N\tilde{u}^{0}K,\label{Eq_W_sis_1}\\
\tilde{e}_{2} & :
\mathcal{L
_{p}\phi-N\tilde{u}^{0},\label{Eq_W_sis_2}\\
\tilde{e}_{3a} & :
\mathcal{L
_{p}\tilde{u}_{a}-ND_{a}\tilde{u}^{0}-N\left( \tilde{u}^{r}K_{ar}+\tilde
{u}^{0}S_{a}\right) -N\tilde{u}^{f}K_{fa},\label{Eq_W_sis_3}\\
\psi_{1c} & :=D_{f}\phi-\tilde{u}_{f},\label{Eq_W_sis_4}\\
\psi_{2ac} & :=D_{[a}\tilde{u}_{c]}. \label{Eq_W_sis_5
\end{align}
Notice that (\ref{Eq_W_sis_1}-\ref{Eq_W_sis_5}) are projected onto $\Sigma
_{t}$, where $\tilde{e}_{1},$ $\tilde{e}_{2},$ $\tilde{e}_{3a}$ are the
evolution equations for our variables and $\psi_{1c},$ $\psi_{2ac}$ are the
constraints of the system.
Using the expressions (\ref{eq_wav_E_1}), (\ref{eq_wav_E_2}) and
(\ref{eq_wav_E_3}), we obtain the principal symbol of the system
\begin{align*}
\left[
\begin{array}
[c]{c
E\\
E_{c}\\
E_{ac
\end{array}
\right] & \approx\mathfrak{N}_{~\beta}^{Bq}\nabla_{q}\phi^{\beta}\\
& =\left[
\begin{array}
[c]{ccc
-\tilde{m}^{q} & 0 & \tilde{\eta}^{qw}\\
0 & \delta_{c}^{q} & 0\\
-\tilde{n}_{[c}\tilde{\eta}_{a]}^{q} & 0 & -\delta_{\lbrack c}^{q}\tilde{\eta
}_{a]}^{w
\end{array}
\right] \nabla_{q}\left[
\begin{array}
[c]{c
\tilde{u}^{0}\\
\phi\\
\tilde{u}_{w
\end{array}
\right]
\end{align*}
and the Geroch fields $C_{A}^{\Gamma z}$ and $M_{A}^{\tilde{\Delta}z}$
\begin{equation}
\left[
\begin{array}
[c]{c
C_{A}^{\Gamma z}\\
M_{A}^{\tilde{\Delta}z
\end{array}
\right] E^{A}=\left[
\begin{array}
[c]{ccc
0 & 2N\tilde{m}^{[z}\tilde{\eta}_{r}^{g]} & 0\\
0 & 0 & 3N\tilde{m}^{[z}\tilde{\eta}_{g}^{a}\tilde{\eta}_{h}^{b]}\\
0 & \tilde{\eta}_{[s}^{z}\tilde{\eta}_{r]}^{g} & 0\\
0 & 0 & \tilde{n}_{d}\varepsilon^{dzab
\end{array}
\right] \left[
\begin{array}
[c]{c
E\\
E_{g}\\
E_{ab
\end{array}
\right] . \label{Eq_W_C_M_1
\end{equation}
The first two lines are associated with $C_{A}^{\Gamma z}$ and the next two
with $M_{A}^{\tilde{\Delta}z}$. Notice that
\[
\left[
\begin{array}
[c]{c
C_{A}^{\Gamma z}\\
M_{A}^{\tilde{\Delta}z
\end{array}
\right] n_{z}=\left[
\begin{array}
[c]{ccc
0 & \tilde{\eta}_{w}^{c} & 0\\
0 & 0 & \tilde{\eta}_{w}^{[a}\tilde{\eta}_{y}^{c]}\\
0 & 0 & 0\\
0 & 0 & 0
\end{array}
\right]
\]
where $M_{A}^{\tilde{\Delta}0}:=M_{A}^{\tilde{\Delta}z}n_{z}=0$ and
\[
C_{A}^{\Gamma0}:=C_{A}^{\Gamma z}n_{z}=\left[
\begin{array}
[c]{ccc
0 & \tilde{\eta}_{w}^{c} & 0\\
0 & 0 & \tilde{\eta}_{w}^{[a}\tilde{\eta}_{y}^{c]
\end{array}
\right] .
\]
The constraints\ $\psi_{1c}$ and $\psi_{2ac}$ are obtained by contracting this
last expression with $E^{A}$.
On the other hand, we choose the following reduction $h_{~A}^{\alpha}$, such
tha
\[
h_{~A}^{\alpha}E^{A}=\left[
\begin{array}
[c]{ccc
-N & 0 & 0\\
0 & N\tilde{m}^{c} & 0\\
0 & 0 & -2N\tilde{\eta}_{q}^{[a}\tilde{m}^{c]
\end{array}
\right] \left[
\begin{array}
[c]{c
E\\
E_{c}\\
E_{ac
\end{array}
\right] .
\]
which gives the (symmetric hyperbolic) evolution equations $\tilde{e}_{1,2,3}$.
Combining the last two results, we have that
\begin{equation}
\left[
\begin{array}
[c]{c
h_{~A}^{\alpha}\\
C_{A}^{\Gamma0
\end{array}
\right] E^{A}=\left[
\begin{array}
[c]{ccc
-N & 0 & 0\\
0 & N\tilde{m}^{c} & 0\\
0 & 0 & -2N\tilde{\eta}_{q}^{[a}\tilde{m}^{c]}\\
0 & \tilde{\eta}_{w}^{c} & 0\\
0 & 0 & \tilde{\eta}_{w}^{[a}\tilde{\eta}_{y}^{c]
\end{array}
\right] \left[
\begin{array}
[c]{c
E\\
E_{c}\\
E_{ac
\end{array}
\right] =\left[
\begin{array}
[c]{c
\tilde{e}_{1}\\
\tilde{e}_{2}\\
\tilde{e}_{3q}\\
\psi_{1w}\\
\psi_{2wy
\end{array}
\right] . \label{eq_wav_e_c_1
\end{equation}
The principal symbol $\left[
\begin{array}
[c]{c
h_{~A}^{\alpha}\\
C_{A}^{\Gamma0
\end{array}
\right] \mathfrak{N}_{~\alpha}^{Aq}l_{q}$ of the latter expression is
\begin{equation}
\left[
\begin{array}
[c]{c
h_{~A}^{\alpha}\\
C_{A}^{\Gamma0
\end{array}
\right] \mathfrak{N}_{~\alpha}^{Aq}\nabla_{q}\left[
\begin{array}
[c]{c
\tilde{u}^{0}\\
\phi\\
\tilde{u}_{w
\end{array}
\right] =\left[
\begin{array}
[c]{ccc
N\tilde{m}^{q} & 0 & -N\tilde{\eta}^{qw}\\
0 & N\tilde{m}^{q} & 0\\
-N\tilde{\eta}_{s}^{q} & 0 & N\tilde{\eta}_{s}^{w}\tilde{m}^{q}\\
0 & \tilde{\eta}_{s}^{q} & 0\\
0 & 0 & \tilde{\eta}_{s}^{[q}\tilde{\eta}_{y}^{w]
\end{array}
\right] \nabla_{q}\left[
\begin{array}
[c]{c
\tilde{u}^{0}\\
\phi\\
\tilde{u}_{w
\end{array}
\right] . \label{Eq_W_pencil_1
\end{equation}
On the other hand, the inverse of $\left[
\begin{array}
[c]{c
h_{B}^{\alpha}\\
C_{B}^{\Delta0
\end{array}
\right] $ has the following for
\[
\left[
\begin{array}
[c]{cc
\mathfrak{N}_{~\alpha}^{B0} & h_{~\Gamma}^{B
\end{array}
\right] =\left[
\begin{array}
[c]{ccccc
-\frac{1}{N} & 0 & 0 & 0 & 0\\
0 & -\frac{1}{N}\tilde{n}_{c} & 0 & \tilde{\eta}_{c}^{w} & 0\\
0 & 0 & \frac{1}{N}\tilde{\eta}_{[a}^{q}\tilde{n}_{c]} & 0 & \tilde{\eta
_{[a}^{w}\tilde{\eta}_{c]}^{y
\end{array}
\right] ,
\]
and satisfy that
\[
\left[
\begin{array}
[c]{cc
\mathfrak{N}_{~\alpha}^{A0} & h_{~\Delta}^{A
\end{array}
\right] \left[
\begin{array}
[c]{c
h_{B}^{\alpha}\\
C_{B}^{\Delta0
\end{array}
\right] =\delta_{B}^{A},
\]
i.e
\begin{align}
& \left[
\begin{array}
[c]{ccccc
-\frac{1}{N} & 0 & 0 & 0 & 0\\
0 & -\frac{1}{N}\tilde{n}_{r} & 0 & \tilde{\eta}_{r}^{w} & 0\\
0 & 0 & \frac{1}{N}\tilde{\eta}_{[r}^{q}\tilde{n}_{s]} & 0 & \tilde{\eta
_{[r}^{w}\tilde{\eta}_{s]}^{y
\end{array}
\right] \left[
\begin{array}
[c]{ccc
-N & 0 & 0\\
0 & N\tilde{m}^{c} & 0\\
0 & 0 & -2N\tilde{\eta}_{q}^{[a}\tilde{m}^{c]}\\
0 & \tilde{\eta}_{w}^{c} & 0\\
0 & 0 & \tilde{\eta}_{w}^{[a}\tilde{\eta}_{y}^{c]
\end{array}
\right] \nonumber\\
& =\left[
\begin{array}
[c]{ccc
1 & 0 & 0\\
0 & \delta_{r}^{c} & 0\\
0 & 0 & \delta_{r}^{[a}\delta_{s}^{c]
\end{array}
\right] \label{eq_wav_p_p_1_1
\end{align}
Using the above expressions, we obtain
\begin{align*}
& \left[
\begin{array}
[c]{c
C_{A}^{\Delta d}\\
M_{A}^{\tilde{\Delta}z
\end{array}
\right] E^{A}\\
& =\left( \left[
\begin{array}
[c]{c
C_{A}^{\Delta d}\\
M_{A}^{\tilde{\Delta}z
\end{array}
\right] \left[
\begin{array}
[c]{cc
\mathfrak{N}_{~\alpha}^{A0} & h_{~\Gamma}^{A
\end{array}
\right] \right) \left( \left[
\begin{array}
[c]{c
h_{B}^{\alpha}\\
C_{B}^{\Gamma0
\end{array}
\right] E^{B}\right) ,\\
& =\left[
\begin{array}
[c]{ccccc
0 & -\tilde{\eta}_{r}^{z} & 0 & N\tilde{m}^{z}\tilde{\eta}_{r}^{w} & 0\\
0 & 0 & \tilde{\eta}_{g}^{[q}\tilde{\eta}_{h}^{z]} & 0 & N\tilde{m}^{z
\tilde{\eta}_{g}^{[w}\tilde{\eta}_{h}^{y]}\\
0 & 0 & 0 & \tilde{\eta}_{[s}^{z}\tilde{\eta}_{r]}^{w} & 0\\
0 & 0 & 0 & 0 & \tilde{n}_{d}\varepsilon^{dzab}\tilde{\eta}_{a}^{w}\tilde
{\eta}_{b}^{y
\end{array}
\right] \left[
\begin{array}
[c]{c
\tilde{e}_{1}\\
\tilde{e}_{2}\\
\tilde{e}_{3q}\\
\psi_{1w}\\
\psi_{2wy
\end{array}
\right] .
\end{align*}
Furthermore, $\left[
\begin{array}
[c]{c
C_{A}^{\Delta d}\\
M_{A}^{\tilde{\Delta}z
\end{array}
\right] E^{A}$ can be written as (\ref{Eq_W_C_M_1}) and its divergence as
(\ref{eq_wav_id_dCE_1}). This leads to the following identity
\begin{align}
& 0=\nabla_{z}\left( \left[
\begin{array}
[c]{ccccc
0 & -\tilde{\eta}_{r}^{z} & 0 & N\tilde{m}^{z}\tilde{\eta}_{r}^{w} & 0\\
0 & 0 & \tilde{\eta}_{g}^{[q}\tilde{\eta}_{h}^{z]} & 0 & N\tilde{m}^{z
\tilde{\eta}_{g}^{[w}\tilde{\eta}_{h}^{y]}\\
0 & 0 & 0 & \tilde{\eta}_{[s}^{z}\tilde{\eta}_{r]}^{w} & 0\\
0 & 0 & 0 & 0 & \tilde{n}_{d}\varepsilon^{dzab}\tilde{\eta}_{a}^{w}\tilde
{\eta}_{b}^{y
\end{array}
\right] \left[
\begin{array}
[c]{c
\tilde{e}_{1}\\
\tilde{e}_{2}\\
\tilde{e}_{3q}\\
\psi_{1w}\\
\psi_{2wy
\end{array}
\right] \right) \nonumber\\
& +\left[
\begin{array}
[c]{ccc
0 & -\nabla_{z}\left( 2N\tilde{m}^{[z}\tilde{\eta}_{r}^{g]}\right) &
2N\tilde{m}^{a}\tilde{\eta}_{r}^{b}\\
0 & 0 & -\nabla_{z}\left( 3N\tilde{m}^{[z}\tilde{\eta}_{g}^{a}\tilde{\eta
}_{h}^{b]}\right) \\
0 & -\nabla_{z}\left( \tilde{\eta}_{s}^{[z}\tilde{\eta}_{r}^{g]}\right) &
\tilde{\eta}_{s}^{[a}\tilde{\eta}_{r}^{b]}\\
0 & 0 & -\nabla_{z}\left( \tilde{n}_{d}\varepsilon^{dzab}\right)
\end{array}
\right] \left[
\begin{array}
[c]{c
-\frac{1}{N}\tilde{e}_{1}\\
-\frac{1}{N}\tilde{n}_{g}\tilde{e}_{2}+\psi_{1g}\\
\frac{1}{N}\tilde{e}_{3[a}\tilde{n}_{b]}+\psi_{2ab
\end{array}
\right] , \label{Eq_W_D_C_M_3
\end{align}
where it has been using the eqs. (\ref{eq_wav_E_1}), (\ref{eq_wav_E_2}) and
(\ref{eq_wav_E_3}) \ Finally, this expression can be rewritten a
\begin{align*}
0 & =\left[
\begin{array}
[c]{cc
\tilde{\eta}_{c}^{q} & 0\\
0 & \tilde{\eta}_{s}^{g}\tilde{\eta}_{r}^{h}\\
0 & 0\\
0 & 0
\end{array}
\right]
\mathcal{L
_{p}\left[
\begin{array}
[c]{c
\psi_{1q}\\
\psi_{2gh
\end{array}
\right] +\left[
\begin{array}
[c]{cc
0 & 0\\
0 & 0\\
\tilde{\eta}_{[s}^{f}\tilde{\eta}_{r]}^{q} & 0\\
0 & \tilde{n}_{d}\varepsilon^{dfgh
\end{array}
\right] D_{f}\left[
\begin{array}
[c]{c
\psi_{1q}\\
\psi_{2gh
\end{array}
\right] \\
& +\left[
\begin{array}
[c]{cc
0 & 0\\
0 & 0\\
0 & \tilde{\eta}_{[s}^{g}\tilde{\eta}_{r]}^{h}\\
0 & 0
\end{array}
\right] \left[
\begin{array}
[c]{c
\psi_{1q}\\
\psi_{2gh
\end{array}
\right] +\left[
\begin{array}
[c]{c
-2N\tilde{m}^{a}\tilde{\eta}_{c}^{b}\nabla_{\lbrack a}\left( \tilde{n
_{b]}\frac{1}{N}\tilde{e}_{2}\right) +\tilde{e}_{3c}\\
+3N\tilde{m}^{f}\tilde{\eta}_{g}^{a}\tilde{\eta}_{h}^{b}\nabla_{\lbrack
f}\left( \tilde{e}_{3a}\tilde{n}_{b]}\frac{1}{N}\right) \\
-\tilde{\eta}_{s}^{a}\tilde{\eta}_{r}^{c}\nabla_{\lbrack a}\left( \tilde
{n}_{c]}\frac{1}{N}\tilde{e}_{2}\right) \\
+\tilde{n}_{d}\varepsilon^{dfgh}\nabla_{f}\left( \tilde{e}_{3g}\tilde{n
_{h}\frac{1}{N}\right)
\end{array}
\right]
\end{align*}
In the on-shell case ($e_{1,2,3}=0$), these equations are the subsidiary
system (1st and 2nd line) and the constraints of the constraints (3rd and 4th
line) which are trivially satisfied in the evolution. Namely,
\begin{align}
0 & =\left[
\begin{array}
[c]{cc
\tilde{\eta}_{c}^{q} & 0\\
0 & \tilde{\eta}_{s}^{g}\tilde{\eta}_{r}^{h}\\
0 & 0\\
0 & 0
\end{array}
\right]
\mathcal{L
_{p}\left[
\begin{array}
[c]{c
\psi_{1q}\\
\psi_{2gh
\end{array}
\right] +\left[
\begin{array}
[c]{cc
0 & 0\\
0 & 0\\
\tilde{\eta}_{[s}^{f}\tilde{\eta}_{r]}^{q} & 0\\
0 & \tilde{n}_{d}\varepsilon^{dfgh
\end{array}
\right] D_{f}\left[
\begin{array}
[c]{c
\psi_{1q}\\
\psi_{2gh
\end{array}
\right] \label{Eq_sub_onda_1}\\
& +\left[
\begin{array}
[c]{cc
0 & 0\\
0 & 0\\
0 & \tilde{\eta}_{[s}^{g}\tilde{\eta}_{r]}^{h}\\
0 & 0
\end{array}
\right] \left[
\begin{array}
[c]{c
\psi_{1q}\\
\psi_{2gh
\end{array}
\right] .\nonumber
\end{align}
where
\mathcal{L
_{p}=\partial_{t}
\mathcal{L
_{\beta}$, since we are considering the coordinates $\left( t,x^{i}\right)
$. \ We note that the evolution equations obtained are quite simple. However,
by adding to these evolution equations terms proportional to the constraints
of the constraints, one can modify this system and obtain a family of
subsidiary systems. This family is obtained by contracting the expression
(\ref{Eq_sub_onda_1}) with the following reduction
\[
N_{\tilde{\Delta}}^{\Gamma}=\left[
\begin{array}
[c]{cccc
\tilde{\eta}_{w}^{c} & 0 & N_{1w}^{sr} & N_{2w}\\
0 & \tilde{\eta}_{n}^{s}\tilde{\eta}_{m}^{r} & N_{3nm}^{sr} & N_{4nm
\end{array}
\right] ,
\]
where \ $N_{1w}^{sr},$ $N_{2w},$ $N_{3nm}^{sr}$ and $N_{4nm}$ can be freely
chosen. Of course, different choices of $N_{\tilde{\Delta}}^{\Gamma}$ give
rise to ill/well-posed evolution equations.
Let us now study the characteristic structures of the system and the
subsidiary system. In other words, we will study the Kronecker decomposition
of the pencils $\left[
\begin{array}
[c]{c
h_{~A}^{\alpha}\mathfrak{N}_{~\alpha}^{Aq}\\
C_{A}^{\Gamma0}\mathfrak{N}_{~\alpha}^{Aq
\end{array}
\right] l_{q}$ and $\left[
\begin{array}
[c]{c
C_{A}^{\Delta d}h_{~\Gamma}^{A}\\
M_{A}^{\tilde{\Delta}z}h_{~\Gamma}^{A
\end{array}
\right] l_{d}$ with $l_{q}=-\lambda n_{a}+k_{a}$ and $k_{a}\left(
\partial_{t}\right) ^{a}=0$. In their pullback version to $\Sigma_{t}$, these
pencils are
\begin{align}
& \left[
\begin{array}
[c]{c
h_{~A}^{\alpha}\mathfrak{N}_{~\alpha}^{Aq}\\
C_{A}^{\Gamma0}\mathfrak{N}_{~\alpha}^{Aq
\end{array}
\right] l_{q}\phi^{\alpha}\nonumber\\
& =\left[
\begin{array}
[c]{ccc
\left( -\lambda-\left( \beta.k\right) \right) & 0 & -Nk^{w}\\
0 & \left( -\lambda-\left( \beta.k\right) \right) & 0\\
-Nk_{s} & 0 & \left( -\lambda-\left( \beta.k\right) \right) \tilde{\eta
}_{s}^{w}\\
0 & k_{s} & 0\\
0 & 0 & k_{[s}\tilde{\eta}_{y]}^{w
\end{array}
\right] \left[
\begin{array}
[c]{c
\tilde{u}^{0}\\
\phi\\
\tilde{u}_{w
\end{array}
\right] , \label{Eq_W_pencil_1_b
\end{align
\begin{equation}
\left[
\begin{array}
[c]{c
C_{A}^{\Delta q}h_{~\Gamma}^{A}\\
M_{A}^{\tilde{\Delta}q}h_{~\Gamma}^{A
\end{array}
\right] l_{q}\psi^{\Gamma}=\left[
\begin{array}
[c]{cc
\tilde{\eta}_{r}^{w}\left( -\lambda-\left( \beta.k\right) \right) & 0\\
0 & \tilde{\eta}_{g}^{[w}\tilde{\eta}_{h}^{y]}\left( -\lambda-\left(
\beta.k\right) \right) \\
k_{[s}\tilde{\eta}_{r]}^{w} & 0\\
0 & k_{q}\varepsilon^{qwy
\end{array}
\right] \left[
\begin{array}
[c]{c
\psi_{1w}\\
\psi_{2wy
\end{array}
\right] . \label{Eq_W_pencil_sub_1
\end{equation}
Where $\varepsilon^{qwy}:=\tilde{n}_{d}\varepsilon^{dqwy}$ is the Levi-Civita
tensor over $\Sigma_{t}$; and all the lowercase indices $w,s,y,g,q$, and any
other lowercase indices that appear until the end of this section, run from
$1$ to $3$. We also note that $\left[
\begin{array}
[c]{c
h_{~A}^{\alpha}\mathfrak{N}_{~\alpha}^{Aq}\\
C_{A}^{\Gamma0}\mathfrak{N}_{~\alpha}^{Aq
\end{array}
\right] l_{q}$ $\i
\mathbb{R}
^{11\times5}$, $\left[
\begin{array}
[c]{c
C_{A}^{\Delta d}h_{~\Gamma}^{A}\\
M_{A}^{\tilde{\Delta}z}h_{~\Gamma}^{A
\end{array}
\right] l_{d}\i
\mathbb{R}
^{10\times6}$, $h_{~A}^{\alpha}\mathfrak{N}_{~\alpha}^{Aq}l_{q}\i
\mathbb{R}
^{5\times5}$, $C_{A}^{\Gamma0}\mathfrak{N}_{~\alpha}^{Aq}l_{q}\i
\mathbb{R}
^{6\times5}$, $C_{A}^{\Delta d}h_{~\Gamma}^{A}l_{d}\i
\mathbb{R}
^{6\times6}$\ \ and $M_{A}^{\tilde{\Delta}z}h_{~\Gamma}^{A}l_{d}\i
\mathbb{R}
^{4\times6}$ in their pullbacked version.
Let us now study the matrix $\left( C_{B}^{\Delta0}\mathfrak{N}_{~\alpha
}^{Bq}\right) l_{q}$, whose expression
\[
\left( C_{B}^{\Delta0}\mathfrak{N}_{~\alpha}^{Bq}\right) l_{q}\phi^{\alpha
}=\left[
\begin{array}
[c]{ccc
0 & \tilde{\eta}_{s}^{q} & 0\\
0 & 0 & \tilde{\eta}_{s}^{[q}\tilde{\eta}_{y}^{w]
\end{array}
\right] k_{q}\left[
\begin{array}
[c]{c
\tilde{u}^{0}\\
\phi\\
\tilde{u}_{w
\end{array}
\right] ,
\]
is obtained from (\ref{Eq_W_pencil_1_b}).
For each $k_{q}$,
\begin{align*}
k_{z}\left( M_{A}^{\tilde{\Delta}z}h_{~\Delta}^{A}\right) \left(
C_{B}^{\Delta0}\mathfrak{N}_{~\alpha}^{Bq}\right) k_{q} & =0,\\
k_{z}\left[
\begin{array}
[c]{cc
\tilde{\eta}_{[f}^{z}\tilde{\eta}_{r]}^{s} & 0\\
0 & \tilde{n}_{d}\varepsilon^{dzsw
\end{array}
\right] \left[
\begin{array}
[c]{ccc
0 & \tilde{\eta}_{s}^{q} & 0\\
0 & 0 & \tilde{\eta}_{s}^{[q}\tilde{\eta}_{y}^{w]
\end{array}
\right] k_{q} & =0,
\end{align*}
i.e. $k_{z}\left( M_{A}^{\tilde{\Delta}z}h_{~\Delta}^{A}\right) $ expands
the left kernel of $\left( C_{B}^{\Delta0}\mathfrak{N}_{~\alpha}^{Bq}\right)
k_{q}$ and therefore the condition v) of the theorem
\ref{Theorem_coef_const_2} is satisfied. Furthermore, this shows that
\begin{align}
\dim\left( left\_\ker\left( C_{B}^{\Delta0}\mathfrak{N}_{~\alpha}^{Bq
k_{q}\right) \right) & =3,\label{Eq_W_lk_1}\\
rank\left( C_{B}^{\Delta0}\mathfrak{N}_{~\alpha}^{Bq}k_{q}\right) &
=3,\label{Eq_W_ra_1}\\
\dim\left( right\_\ker\left( C_{B}^{\Delta0}\mathfrak{N}_{~\alpha}^{Bq
k_{q}\right) \right) & =2. \label{Eq_W_rk_1
\end{align}
This result allows us to give the Kronecker structure associated to the Wave
equation, i.e., the Kronecker structure of the pencil (\ref{Eq_W_pencil_1_b}).
It i
\[
J_{1}\left( N\sqrt{k.k}-\beta.k\right) ,J_{1}\left( -N\sqrt{k.k
-\beta.k\right) ,3\times L_{1}^{T},3\times L_{0}^{T}.
\]
The $3\times L_{1}^{T},3\times L_{0}^{T}$ blocks are justified by
(\ref{Eq_W_lk_1}) and (\ref{Eq_W_ra_1}) as explained in subsubsection
\ref{Kronecker_decomposition}. The Jordan part is justified by giving
explicitly the generalized eigenvectors
\[
\left( \delta\phi_{\lambda_{1}}\right) ^{\beta}=\left[
\begin{array}
[c]{c
1\\
0\\
-\frac{k_{w}}{\sqrt{\left( k.k\right) }
\end{array}
\right] ,\text{ \ \ \ }\left( \delta\phi_{\lambda_{2}}^{2}\right) ^{\beta
}=\left[
\begin{array}
[c]{c
1\\
0\\
\frac{k_{w}}{\sqrt{\left( k.k\right) }
\end{array}
\right] .
\]
They are associated to the generalized eigenvalues $\lambda_{1}=N\sqrt
{k.k}-\beta.k$ and $\lambda_{2}=-N\sqrt{k.k}-\beta.k$. These eigenvalues were
called "the physical" eigenvalues in subsection
\ref{Proof_Theorem_coef_const_2}.
Now, let us study the characteristic structure of $h_{~A}^{\alpha
\mathfrak{N}_{~\alpha}^{Aq}l_{q}$, whose explicit expression is
\[
h_{~A}^{\alpha}\mathfrak{N}_{~\alpha}^{Aq}l_{q}\phi^{\alpha}=\left[
\begin{array}
[c]{ccc
\left( -\lambda-\left( \beta.k\right) \right) & 0 & -Nk^{w}\\
0 & \left( -\lambda-\left( \beta.k\right) \right) & 0\\
-Nk_{s} & 0 & \left( -\lambda-\left( \beta.k\right) \right) \tilde{\eta
}_{s}^{w
\end{array}
\right] \left[
\begin{array}
[c]{c
\tilde{u}^{0}\\
\phi\\
\tilde{u}_{w
\end{array}
\right] .
\]
Since $\left( \lambda_{1,2},\left( \delta\phi_{\lambda_{1,2}}\right)
^{\beta}\right) $ are the generalized eigenvalues and eigenvectors of
$\left[
\begin{array}
[c]{c
h_{~A}^{\alpha}\mathfrak{N}_{~\alpha}^{Aq}\\
C_{A}^{\Gamma0}\mathfrak{N}_{~\alpha}^{Aq
\end{array}
\right] l_{q}$,\ \ they are also the eigenvalues and eigenvectors of
$h_{~A}^{\alpha}\mathfrak{N}_{~\alpha}^{Aq}l_{q}$. Moreover, $h_{~A}^{\alpha
}\mathfrak{N}_{~\alpha}^{Aq}l_{q}$ has $3$ more eigenvalues $\pi
_{1,2,3}\left( k\right) =-\beta.k$, these were called "the constraints 1"
eigenvalues in subsection \ref{Proof_Theorem_coef_const_2}. \ These
eigenvalues have the associated eigenvectors
\begin{align*}
\left( \delta\phi_{\pi_{1}}\right) ^{\beta} & =\left[
\begin{array}
[c]{c
0\\
1\\
0
\end{array}
\right] \text{ \ \ }\left( \delta\phi_{\pi_{2}}\right) ^{\beta}=\left[
\begin{array}
[c]{c
0\\
0\\
v_{1w
\end{array}
\right] \\
\left( \delta\phi_{\pi_{2}}\right) ^{\beta} & =\left[
\begin{array}
[c]{c
0\\
0\\
v_{2w
\end{array}
\right] \text{\
\end{align*}
where $v_{1},v_{2}$ are linearly independent vectors, and they are defined by
the following conditions
\[
\left( v_{1}.v_{2}\right) =\left( v_{1,2}.k\right) =\left( v_{1,2
.\tilde{m}\right) =0.
\]
Let us now study the Kronecker decomposition of the pencil $\left[
\begin{array}
[c]{c
C_{A}^{\Delta q}h_{~\Gamma}^{A}\\
M_{A}^{\tilde{\Delta}q}h_{~\Gamma}^{A
\end{array}
\right] l_{q}$ \ (see eq. (\ref{Eq_W_pencil_sub_1})). This pencil is
associated to the constraints as can be seen from (\ref{Eq_W_D_C_M_3}) and
(\ref{Eq_sub_onda_1}); and its Kronecker decomposition is given by
\[
3\times J_{1}\left( -\beta.k\right) ,3\times L_{1}^{T},1\times L_{0}^{T},
\]
as we will show below.
We begin by studying the Jordan blocks. As it was shown in subsubsection
\ref{Kro_sub}, the generalized eigenvectors of this pencil are obtained by
projecting $\left\{ \left( \delta\phi_{\pi_{1}}\right) ^{\beta},\left(
\delta\phi_{\pi_{2}}\right) ^{\beta},\left( \delta\phi_{\pi_{3}}\right)
^{\beta}\right\} $ with $C_{B}^{\Gamma0}\mathfrak{N}_{~\beta}^{Bq}k_{q}$,
that is,
\begin{align*}
\left[ \delta\psi_{\pi_{1}}^{\Gamma},\delta\psi_{\pi_{2}}^{\Gamma},\delta
\psi_{\pi_{3}}^{\Gamma}\right] & :=C_{B}^{\Gamma0}\mathfrak{N}_{~\beta
}^{Bq}k_{q}\left[ \left( \delta\phi_{1}\right) ^{\beta},\left( \delta
\phi_{\pi_{2}}\right) ^{\beta},\left( \delta\phi_{\pi_{3}}\right) ^{\beta
}\right] ,\\
& =\left[
\begin{array}
[c]{ccc
0 & k_{s} & 0\\
0 & 0 & k_{[s}\tilde{\eta}_{y]}^{w
\end{array}
\right] \left[
\begin{array}
[c]{ccc
0 & 0 & 0\\
1 & 0 & 0\\
0 & v_{1w} & v_{2w
\end{array}
\right] ,\\
& =\left[
\begin{array}
[c]{ccc
k_{s} & 0 & 0\\
0 & k_{[s}v_{1y]} & k_{[s}v_{2y]
\end{array}
\right] .
\end{align*}
Where the eigenvectors are the columns of this matrix and they are associated
to the eigenvalues "the constraints 1" $\pi_{1,2,3}\left( k\right)
=-\beta.k$.
To complete the Kronecker decomposition, we need to find the
\[
\dim\left( left\_\ker\left( M_{A}^{\tilde{\Delta}z}h_{~\Delta}^{A
k_{z}\right) \right) ,
\]
where $M_{A}^{\tilde{\Delta}z}h_{~\Delta}^{A}k_{z}$ is given b
\[
M_{A}^{\tilde{\Delta}z}h_{~\Delta}^{A}k_{z}\psi^{\Delta}=\left[
\begin{array}
[c]{cc
\tilde{\eta}_{[s}^{z}\tilde{\eta}_{r]}^{w} & 0\\
0 & \varepsilon^{zwy
\end{array}
\right] k_{z}\left[
\begin{array}
[c]{c
\psi_{1w}\\
\psi_{2wy
\end{array}
\right] .
\]
Notice that
\[
span\left\langle \left[
\begin{array}
[c]{cc
\tilde{n}_{f}k_{f}\varepsilon^{fgsr} & 0
\end{array}
\right] \right\rangle =left\_\ker\left( M_{A}^{\tilde{\Delta}z}h_{~\Delta
}^{A}k_{z}\right)
\]
therefore
\[
\dim\left( left\_\ker\left( M_{A}^{\tilde{\Delta}z}h_{~\Delta}^{A
k_{z}\right) \right) =1.
\]
Using this result and the expressions (\ref{Eq_W_lk_1}) and
(\ref{eq_sub_kron_1}), we conclude that the rest of the Kronecker
decomposition is $3\times L_{1}^{T},1\times L_{0}^{T}$.
Finally, we give the basis that diagonalizes matrix $C_{A}^{\Gamma
z}h_{~\Delta}^{A}l_{z}$. This matrix
\[
C_{A}^{\Gamma z}h_{~\Delta}^{A}l_{z}\psi^{\Delta}=\left[
\begin{array}
[c]{cc
\tilde{\eta}_{r}^{w}\left( -\lambda-\left( \beta.k\right) \right) & 0\\
0 & \tilde{\eta}_{g}^{[w}\tilde{\eta}_{h}^{y]}\left( -\lambda-\left(
\beta.k\right) \right)
\end{array}
\right] \left[
\begin{array}
[c]{c
\psi_{1w}\\
\psi_{2wy
\end{array}
\right]
\]
can be read from (\ref{Eq_W_pencil_sub_1}). By construction, we know that
$\left\{ \delta\psi_{\pi_{1}}^{\Gamma},\delta\psi_{\pi_{2}}^{\Gamma
,\delta\psi_{\pi_{3}}^{\Gamma}\right\} $ are eigenvectors associated to the
eigenvalues $\pi_{1,2,3}\left( k\right) =-\beta.k$. We still need to find
$3$ more eigenvectors associated to the "Constraints 2" eigenvalues
$\rho_{1,2,3}\left( k\right) $. These eigenvectors $\left\{ \delta
\psi_{\rho_{1}}^{\Delta},\delta\psi_{\rho_{2}}^{\Delta},\delta\psi_{\rho_{3
}^{\Delta}\right\} $ are any $3$ vectors, linearly independent from $\left\{
\delta\psi_{\pi_{1}}^{\Gamma},\delta\psi_{\pi_{2}}^{\Gamma},\delta\psi
_{\pi_{3}}^{\Gamma}\right\} $.
\section{Conclusions and discussion \label{Conclusions}}
In this paper, we have considered generic systems of first-order partial
differential equations that include differential constraints. We have shown
sufficient conditions for these systems to have first-order partial
differential subsidiary system (SS) with a strongly hyperbolic evolution. This
guarantees the constraint preservation.
We have shown that if the constraints of the system are defined by the Geroch
fields $C_{A}^{a\Gamma}$ and the system admits the integrability conditions
(\ref{eq_int_LE_1}) and (\ref{eq_int_LE_2}), then the SS exists and it is a
set of first-order partial differential equations. Furthermore, we have shown
that when the system only admits Geroch fields $C_{A}^{a\Gamma}$ and
$M_{A}^{\tilde{\Delta}z}$, it implies that the Kronecker structures of the
principal symbol, of the system and of the subsidiary system, does not include
$L_{m}^{T}$ blocks with $m\geq2$. Since most-known physical systems have these
kind of Geroch fields, we conclude that they only have $L_{1}^{T}-$blocks and
vanishing rows in their Kronecker decomposition. On the other hand, this
connection between Geroch fields and the Kronecker structure indicates a
possible extension to first-order partial differential equations,\ of the
classifications presented in the cases of ordinary differential equations
\cite{gantmacher1992theory} and algebraic differential equations
\cite{kunkel2006differential}. In these latter classifications, the Kronecker
structure of the principal symbol is used to find the integrability conditions
(as we have done here) and the solutions of the system. Currently, we are
working on extending the results presented here to systems that admit other
kinds of Geroch fields.
The study of strong hyperbolicity of the SS is performed in the case of
constant coefficients and presented as a continuation of the work
\cite{Abalos:2018uwg}. We have given a complete analysis of the characteristic
structure of the system, showing how can be chosen the propagation velocities
of the physical fields and of the constraints. As in \cite{Abalos:2018uwg},
the analysis is algebraic and pseudo-differential, so the possible non
analyticity in the evolution equations as well as in the subsidiary systems
may introduce causality issues. However, the steps of the proofs presented
here can be readapted to each physical system (including the quasi-linear
ones) avoiding the non analytical pseudo-differential reductions. This
procedure will/may come at the cost of less freedom in the choice of the
propagation velocities of the constraints.
With the new tools developed here, it seems natural to continue with the study
of boundary conditions that guarantee the constraint preservation. We have
made great progress in this direction and we are currently writing an article
where we generalize the ideas of \cite{Calabrese:2001kj},
\cite{Calabrese:2002xy}, \cite{Tarfulea:2013qjq}, \cite{sarbach2012continuum}
to the variable and quasi-linear coefficient cases.
\section*{Acknowledgements}
\addcontentsline{toc}{section}{Acknowledgement}
I would like to thank Federico Carrasco for many interesting comments and
suggestions. I am also grateful to Carlos Olmos, David Hilditch and Oscar
Reula for the discussions and ideas exchanged throughout this work. This
research work was partially supported by CONICET.
\begin{appendices}
\section{Lemmas\label{Ap_lemmas}}
In this appendix, we introduce some lemmas about the Kronecker decomposition
of a pencil. These results are used in the proof of Theorem SH of the SS.
Consider the following matrix pencil
\begin{equation}
\lambda I_{\beta}^{A}+K_{\beta}^{A}=\lambda\left[
\begin{array}
[c]{c
-\delta_{\beta}^{\alpha}\\
0
\end{array}
\right] +\left[
\begin{array}
[c]{c
A_{\beta}^{\alpha}\\
C_{\beta}^{\Gamma
\end{array}
\right] , \label{pencil_1
\end{equation}
such that $\delta_{\beta}^{\alpha},A_{\beta}^{\alpha}\i
\mathbb{R}
^{u\times u}$, $C_{\beta}^{\Gamma}\i
\mathbb{R}
^{c\times u}$, $\delta_{\beta}^{\alpha}$ is the identity matrix, $I_{\beta
}^{A}=\left[
\begin{array}
[c]{c
-\delta_{\beta}^{\alpha}\\
0
\end{array}
\right] $ and $K_{\beta}^{A}=\left[
\begin{array}
[c]{c
A_{\beta}^{\alpha}\\
C_{\beta}^{\Gamma
\end{array}
\right] $.
Since $I_{\beta}^{A}$ has only trivial right kernel, the Kronecker
decomposition of this pencil may only include Jordan blocks (with $\lambda
_{i}$ as the generalized eigenvalues)
\[
J_{m}\left( \lambda_{i}\right) =\left[
\begin{array}
[c]{cccc
\left( \lambda-\lambda_{i}\right) & 1 & 0 & 0\\
0 & \left( \lambda-\lambda_{i}\right) & 1 & 0\\
0 & 0 & ... & 1\\
0 & 0 & 0 & \left( \lambda-\lambda_{i}\right)
\end{array}
\right] \i
\mathbb{C}
^{m\times m},
\]
$L_{m}^{T}$ block
\begin{equation}
L_{m}^{T}=\left[
\begin{array}
[c]{ccccc
\lambda & 0 & 0 & 0 & 0\\
1 & \lambda & 0 & 0 & 0\\
0 & 1 & ... & 0 & 0\\
0 & 0 & ... & \lambda & 0\\
0 & 0 & 0 & 1 & \lambda\\
0 & 0 & 0 & 0 & 1
\end{array}
\right] \i
\mathbb{C}
^{\left( m+1\right) \times m}, \label{bloq_LmT_1
\end{equation}
and zero rows (called here $L_{0}^{T}$).
\begin{lemma}
\label{Lemma_1}a) Let
\begin{equation}
J,m_{0}\times L_{0}^{T},m_{1}\times L_{1}^{T},m_{2}\times L_{2}^{T
,...,m_{n}\times L_{n}^{T} \label{estruc_1
\end{equation}
be the Kronecker structure of $\lambda I_{\beta}^{A}+K_{\beta}^{A}$, where
$J\i
\mathbb{C}
^{a\times a}$ includes only the Jordan blocks, the
\begin{align}
m_{0}+m_{1}+m_{2}+...+m_{n} & =c,\label{eq_mi_c_1}\\
a+m_{1}+2m_{2}+...+nm_{n} & =u. \label{eq_mi_c_2
\end{align}
b) For any $\lambda$ other than the generalized eigenvalues, it holds that
\begin{equation}
\dim\left( left\_\ker\left( \lambda I_{\beta}^{A}+K_{\beta}^{A}\right)
\right) =c. \label{eq_dim_left_ker_1
\end{equation}
c) Let (\ref{estruc_1}) with
\[
J=d_{1}\times J_{1}\left( \lambda_{1}\right) ,d_{2}\times J_{1}\left(
\lambda_{2}\right) ,...,d_{q}\times J_{1}\left( \lambda_{q}\right) ,
\]
and $\lambda_{1}<...<\lambda_{q}$ be the Kronecker structure of $\lambda
I_{\beta}^{A}+K_{\beta}^{A}$, then
\begin{align}
\dim\left( left\_\ker\left( \lambda I_{\beta}^{A}+K_{\beta}^{A}\right)
\right) & =c\text{ \ \ \ }\forall\lambda\neq\lambda_{i}\text{ with
}i=1,...,q,\label{eq_lk_pen_1}\\
\dim\left( left\_\ker\left( \left. \lambda I_{\beta}^{A}+K_{\beta
^{A}\right\vert _{\lambda=\lambda_{i}}\right) \right) & =c+d_{i}.
\label{eq_lk_pen_2
\end{align}
\end{lemma}
\begin{proof}
a) The pencil $\lambda I_{\beta}^{A}+K_{\beta}^{A}$ is a $(u+c)\times u$
matrix so its Kronecker matrix is of the same size. Therefore, the number of
columns of $J\i
\mathbb{C}
^{a\times a}$ plus that of the $L_{m}^{T}\i
\mathbb{C}
^{m+1\times m}$ blocks in (\ref{estruc_1}) has to be equal to $u$, i.e.,
\[
a+m_{1}+2m_{2}+...+nm_{n}=u.
\]
This concludes the proof of eq. (\ref{eq_mi_c_2}).
The same analysis on the rows shows that
\[
a+m_{0}+2m_{1}+3m_{2}+...+\left( n+1\right) m_{n}=u+c.
\]
Subtracting these last two expressions, we obtain the eq. (\ref{eq_mi_c_1})
and conclude the proof of (a). Notice that (\ref{eq_mi_c_1}) is independent of
the size of $J$.
b) Studying the left kernel of $\lambda I_{\beta}^{A}+K_{\beta}^{A}$ is
equivalent to studying the left kernel of its Kronecker structure, so we
assume that $\lambda I_{\beta}^{A}+K_{\beta}^{A}$ is already in its Kronecker form.
We begin by noting that the Jordan blocks only have non-trivial left kernel
when $\lambda=\lambda_{i}$. Since we are not considering these values of
$\lambda$, we conclude that to show (\ref{eq_dim_left_ker_1}), it is
sufficient to study the left kernel of the blocks $L_{m}^{T}$ and the zero
rows $L_{0}^{T}$.
The left kernel of each block $L_{m}^{T}$ (eq. (\ref{bloq_LmT_1})) has
dimension 1 and is expanded by the vector
\[
X=\left[ -1,\lambda,...,\left( -1\right) ^{m}\lambda^{m-1},\left(
-1\right) ^{m+1}\lambda^{m}\right] \i
\mathbb{R}
^{1\times m+1}.
\]
Therefore, if we assume without loss of generality, that the Kronecker
structure of the pencil is given by (\ref{estruc_1}) we conclude that the left
kernel of the pencil has dimension
\[
m_{0}+m_{1}+m_{2}+...+m_{n}.
\]
Using the equation (\ref{eq_mi_c_1}) of a) we conclude the proof of b).
c) The proof of equation (\ref{eq_lk_pen_1}) is the same as the proof of b),
since b) is independent of the form of $J$.
The proof of eq. (\ref{eq_lk_pen_2}) is followed by noting that all Jordan
blocks in $J$ are of $1\times1$ and that each generalized eigenvalue
$\lambda_{i}$ has degeneracy $d_{i}$; therefore, when $\lambda=\lambda_{i}$
there are $d_{i}$ extra vectors added to the left kernel of the pencil
justifying the expression (\ref{eq_lk_pen_2}).
\end{proof}
\bigskip
The following lemma reveals the connection between $C_{\beta}^{\Gamma}$ and
the Kronecker structure of the pencil (\ref{pencil_1}).
We call
\begin{align*}
d & :=\dim\left( right\_\ker\left( C_{\beta}^{\Gamma}\right) \right) ,\\
r & :=rank\left( C_{\beta}^{\Gamma}\right) ,\\
s & :=\dim\left( left\_\ker\left( C_{\beta}^{\Gamma}\right) \right) ,
\end{align*}
and recall that by the rank--nullity theore
\begin{align}
u & =r+d,\label{eq_u_r_d_1}\\
c & =r+s. \label{eq_c_r_s_1
\end{align}
\begin{lemma}
\label{Lemma_2}Let the pencil (\ref{pencil_1}) such that part of its Kronecker
structure i
\begin{equation}
d_{1}\times J_{1}\left( \lambda_{1}\right) ,d_{2}\times J_{1}\left(
\lambda_{2}\right) ,...,d_{q}\times J_{1}\left( \lambda_{q}\right) ,
\label{eq_d1_J1_1
\end{equation}
with
\begin{equation}
d_{1}+d_{2}+...+d_{q}=d. \label{eq_d1_d2_d_1
\end{equation}
Then the complete Kronecker structure of this pencil is
\begin{equation}
d_{1}\times J_{1}\left( \lambda_{1}\right) ,d_{2}\times J_{1}\left(
\lambda_{2}\right) ,...,d_{q}\times J_{1}\left( \lambda_{q}\right) ,s\times
L_{0}^{T},r\times L_{1}^{T}. \label{eq_penc_kro_1
\end{equation}
\end{lemma}
\begin{proof}
We begin by showing that if the conditions (\ref{eq_d1_J1_1}) and
(\ref{eq_d1_d2_d_1}) are satisfied, the system does not admit any other Jordan
blocks than those present in (\ref{eq_d1_J1_1}). Consider the set of
generalized eigenvectors $\delta\phi_{j}^{\beta\lambda_{i}}$ associated to the
structure (\ref{eq_d1_J1_1}), they compound the set $\left\{ \delta\phi
_{j}^{\beta\lambda_{i}}\right\} $ of $d$ linearly independent vectors with
$j=1,...,d_{i}$ and $i=1,...,q$ \ such that
\begin{equation}
\left( \lambda_{i}\left[
\begin{array}
[c]{c
-\delta_{\beta}^{\alpha}\\
0
\end{array}
\right] +\left[
\begin{array}
[c]{c
A_{\beta}^{\alpha}\\
C_{\beta}^{\Gamma
\end{array}
\right] \right) \delta\phi_{j}^{\beta\lambda_{i}}=0. \label{eq_aut_gen_1
\end{equation}
This means that
\[
C_{\beta}^{\Gamma}\delta\phi_{j}^{\beta\lambda_{i}}=0\text{ \ \ }\forall
j=1,..d\text{ and }\forall i=1,...,q.
\]
Notice that any other possible generalized eigenvector satisfying
(\ref{eq_aut_gen_1}) should also belong to the right kernel of $C_{\beta
}^{\Gamma}$. But since $d=\dim\left( right\_\ker\left( C_{\beta}^{\Gamma
}\right) \right) $, the entire right kernel of $C_{\beta}^{\Gamma}$ must be
expanded by the $d$ vectors $\left\{ \delta\phi_{j}^{\beta\lambda_{i
}\right\} $. In other words, the pencil cannot admit any extra Jordan block,
otherwise, we will reach a contradiction.
It only remains to show that the complete pencil structure is given by
(\ref{eq_penc_kro_1}). We call $J$ the part of the Jordan blocks
\[
J=d_{1}\times J_{1}\left( \lambda_{1}\right) ,d_{2}\times J_{1}\left(
\lambda_{2}\right) ,...,d_{q}\times J_{1}\left( \lambda_{q}\right)
\]
and we notice that $J$ is a $d\times d$ matrix, so using the equation
(\ref{eq_mi_c_2}) of the previous lemma, we obtain
\[
d+m_{1}+2m_{2}+...+nm_{n}=u.
\]
Since by (\ref{eq_u_r_d_1}), $u=r+d$, we conclude that
\begin{equation}
m_{1}+2m_{2}+...+nm_{n}=r. \label{eq_sum_r_1
\end{equation}
On the other hand, since $\dim\left( left\_\ker\left( C_{\beta}^{\Gamma
}\right) \right) =s$, the pencil $\lambda I_{\beta}^{A}+K_{\beta}^{A}$ has
$s$ linearly independent vectors $\left[
\begin{array}
[c]{cc
0 & v_{\Gamma}^{i
\end{array}
\right] $, with $i=1,...,s$, belonging to the left kernel of the pencil.
These vectors do not depend on $\lambda$ since $C_{\beta}^{\Gamma}$ does not.
Moreover, due to the explicit form of $\lambda I_{\beta}^{A}+K_{\beta}^{A}$,
every other vector of the left kernel, independent of $\lambda$ should be a
linear combination of the previous ones. Therefore, the set of vectors
$\left[
\begin{array}
[c]{cc
0 & v_{\Gamma}^{i
\end{array}
\right] $\ are associated to $s$ null rows in the Kronecker decomposition of
$\lambda I_{\beta}^{A}+K_{\beta}^{A}$, i.e.,
\[
m_{0}=s.
\]
Replacing this expression in (\ref{eq_sum_r_1}) and (\ref{eq_c_r_s_1}) we
obtain
\begin{align*}
c & =r+s\\
& =m_{1}+2m_{2}+...+nm_{n}+m_{0}.
\end{align*}
Using now the eq. (\ref{eq_mi_c_1}), we conclude
\[
m_{0}+m_{1}+m_{2}+...+m_{n}=c=m_{1}+2m_{2}+...+nm_{n}+m_{0
\]
and, therefore,
\[
0=m_{2}+2m_{3}+...+\left( n-1\right) m_{n}.
\]
Since $m_{i}\geq0$ the unique solution for this equation is
\[
m_{i}=0\text{ para }i\geq2.
\]
Therefore
\[
c=r+s=m_{1}+m_{0}=m_{1}+s,
\]
from which we conclude that $m_{1}=r$, ending the proof.
\end{proof}
\section{Lapse and shift\label{App_coordenadas}}
In this appendix, we consider a space-time $M$, with the foliation $\Sigma
_{t}$ described in subsection \ref{n+1_decomposition_sec}, and we study the
equation
\begin{equation}
\nabla_{a}q^{b}=0, \label{Eq_ap_sis_1
\end{equation}
where $\nabla_{a}$ is any covariant derivative without torsion. We present
this simple system as an example of the process of rewriting the equations in
their $n+1$ version. Any more complex system can be handled in the same way by
repeating the steps presented here. We perform this $n+1$ decomposition by
using a different projector $\tilde{\eta}_{b}^{a}$ to the one $\eta_{b}^{a}$
used in section \ref{Seccion_Setting_1} and assuming that the system does not
have a background metric. This new projector is parametrized by the lapse
function $N$ and the shift vector $\beta^{a}.$ Of course, we recover the
standard results in the cases with a\ background metric \
We remark that theorems \ref{teorema_ec_ev_vin_1} and
\ \ref{Theorem_coef_const_2} can be re-adapted to the type of projection used
here. This is showing in the examples of section \ref{Examples}.
We begin by considering the definitions introduced in subsection
\ref{n+1_decomposition_sec}, and defining the shift vector $\beta^{a}$ such
that
\[
\beta^{a}n_{a}=0.
\]
It allows us to define the projector
\begin{equation}
\tilde{\eta}_{b}^{a}:=\delta_{b}^{a}-p^{a}n_{b}, \label{Eq_proy_b_1
\end{equation}
with
\[
p^{a}=t^{a}-\beta^{a}.
\]
Notice that this projector reduces to the $\eta_{b}^{a}$ projector when
$\beta^{a}=0$.
Using the coordinates $\left( t,x^{i}\right) $ adapted to the foliation
introduced in subsection \ref{n+1_decomposition_sec} we obtain
\[
\tilde{\eta}_{b}^{a}=\left[
\begin{array}
[c]{cccc
0 & 0 & 0 & 0\\
\beta^{1} & 1 & 0 & 0\\
\beta^{2} & 0 & 1 & 0\\
\beta^{3} & 0 & 0 & 1
\end{array}
\right] ,
\]
where the indices $a$ and $b$ correspond to the columns and the rows of the
matrix respectively.
On the other hand, since
\[
p^{a}n_{a}=t^{a}n_{a}-\beta^{a}n_{a}=t^{a}n_{a}=1,
\]
it is easy to check that $\tilde{\eta}_{b}^{a}$ satisfies similar projector
properties as $\eta_{c}^{b}$, that is,
\
\begin{array}
[c]{ccccc
\tilde{\eta}_{b}^{a}\tilde{\eta}_{c}^{b}=\tilde{\eta}_{c}^{a}, & &
\tilde{\eta}_{b}^{a}p^{b}=0, & & \tilde{\eta}_{b}^{a}n_{a}=0.
\end{array}
\]
We introduce now the lapse function $N$ in the following way, consider the
projector
\begin{align}
\tilde{\eta}_{b}^{a} & :=\delta_{b}^{a}+\frac{1}{N}p^{a}\left(
-Nn_{b}\right) ,\nonumber\\
& =\delta_{b}^{a}+\tilde{m}^{a}\tilde{n}_{b}, \label{Eq_proy_b_3
\end{align}
where we have define
\begin{align*}
\tilde{m}^{a} & :=\frac{1}{N}p^{a},\\
\tilde{n}_{b} & :=-Nn_{b}\text{,
\end{align*}
such that
\begin{equation}
\tilde{m}^{a}\tilde{n}_{a}=-1. \label{Eq_nor_m_n_1
\end{equation}
With these definitions, we can project the eq. (\ref{Eq_ap_sis_1}) and its
variables as follows.
\begin{lemma
\begin{align}
\nabla_{a}q^{b} & =-\tilde{m}^{b}n_{a}\left(
\mathcal{L
_{p}\tilde{q}^{0}-N\left( \left( \tilde{q}^{w}S_{w}\right) +\tilde{q
^{0}\left( \tilde{m}^{d}Z_{d}\right) \right) \right) \label{Eq_apendix_T
\\
& +\tilde{\eta}_{w}^{b}n_{a}\left(
\mathcal{L
_{p}\tilde{q}^{w}-N\left( \tilde{q}^{r}K_{~r}^{w}+\tilde{q}^{0}S^{w}\right)
\right) \\
& +\tilde{\eta}_{w}^{b}\tilde{\eta}_{a}^{d}\left( D_{d}\tilde{q}^{w
+\tilde{q}^{0}K_{~d}^{w}\right) \\
& -\tilde{m}^{b}\tilde{\eta}_{a}^{d}\left( D_{d}\tilde{q}^{0}+\tilde{q
^{w}K_{wd}-\tilde{q}^{0}Z_{d}\right)
\end{align}
with $\tilde{q}^{w},S^{r},$ $S_{r}$, $K_{ba}$ and $K_{~a}^{b}$ tangents to
$\Sigma_{t}$ and such that
\begin{align}
\tilde{q}^{w} & :=\tilde{\eta}_{r}^{w}q^{r},\label{Eq_apendix_T_1}\\
\tilde{q}^{0} & :=\tilde{n}_{r}q^{r}, \label{Eq_apendix_T_2
\end{align
\begin{equation}
Z_{d}:=\tilde{n}_{w}\nabla_{d}\tilde{m}^{w}, \label{Eq_apendix_T_3
\end{equation
\begin{align}
S_{r} & :=\tilde{\eta}_{r}^{w}\tilde{m}^{d}\nabla_{d}\tilde{n
_{w},\label{Eq_apendix_T_4}\\
S^{r} & :=\tilde{\eta}_{w}^{r}\tilde{m}^{d}\nabla_{d}\tilde{m}^{w},
\label{Eq_apendix_T_5
\end{align
\begin{align}
-K_{~a}^{b} & :=\tilde{\eta}_{w}^{b}\tilde{\eta}_{a}^{d}\nabla_{d}\tilde
{m}^{w},\label{Eq_apendix_T_6}\\
-K_{ba} & :=\tilde{\eta}_{a}^{d}\tilde{\eta}_{b}^{r}\nabla_{d}\tilde{n}_{r}
\label{Eq_apendix_T_7
\end{align}
The covariant derivative $D_{d}$ is defined over $\Sigma_{t}$ in the standard
form
\begin{equation}
D_{d}\tilde{q}^{w}:=\tilde{\eta}_{r}^{w}\tilde{\eta}_{d}^{s}\nabla_{s
\tilde{q}^{w}, \label{Eq_apendix_T_8
\end{equation}
and the Lie derivative
\mathcal{L
_{p}$ i
\begin{align
\mathcal{L
_{p} &
\mathcal{L
_{\partial_{t}}
\mathcal{L
_{\beta},\label{Eq_apendix_T_9}\\
& =\partial_{t}
\mathcal{L
_{\beta},
\end{align}
where the last equation holds since we are considering the coordinates
$\left( t,x^{i}\right) .$
Moreover, it holds
\begin{equation}
S_{r}=\tilde{\eta}_{r}^{w}\left( D_{w}\left( \ln N\right) -Z_{w}\right)
\label{Eq_apendix_T_10
\end{equation}
and
\begin{align
\mathcal{L
_{\tilde{m}}\tilde{\eta}_{r}^{d} & =\tilde{m}^{d}D_{r}\left( \ln N\right)
,\label{Eq_apendix_T_11}\\%
\mathcal{L
_{p}\tilde{\eta}_{r}^{d} & =0 \label{Eq_apendix_T_12
\end{align}
\end{lemma}
\begin{proof}
We start by isolating $\delta_{a}^{d}$ from equations (\ref{Eq_proy_b_1}) and
(\ref{Eq_proy_b_3}) and rewriting (\ref{Eq_ap_sis_1}) as follows
\begin{align*}
\nabla_{a}q^{b} & =\delta_{a}^{d}\delta_{q}^{b}\nabla_{d}\left( \delta
_{c}^{b}q^{c}\right) ,\\
& =\left( \tilde{\eta}_{a}^{d}+p^{d}n_{a}\right) \left( \tilde{\eta
_{q}^{b}+p^{b}n_{q}\right) \nabla_{d}\left( \left( \tilde{\eta}_{c
^{b}-\tilde{m}^{b}\tilde{n}_{c}\right) q^{c}\right) .
\end{align*}
Defining
\begin{align*}
\tilde{q}^{0} & :=\tilde{n}_{c}q^{c},\\
\tilde{q}^{b} & :=\tilde{\eta}_{c}^{b}q^{c
\end{align*}
and introducing these definitions in the last expression we obtai
\begin{align*}
\nabla_{a}q^{b} & =-\tilde{m}^{b}n_{a}\left( p^{d}\tilde{n}_{w}\left(
\nabla_{d}\tilde{q}^{w}-\tilde{q}^{0}\nabla_{d}\tilde{m}^{w}-\tilde{m
^{w}\nabla_{d}\tilde{q}^{0}\right) \right) \\
& +\tilde{\eta}_{w}^{b}n_{a}\left( p^{d}\left( \nabla_{d}\tilde{q
^{w}-\tilde{q}^{0}\nabla_{d}\tilde{m}^{w}-\tilde{m}^{w}\nabla_{d}\tilde{q
^{0}\right) \right) \\
& +\tilde{\eta}_{w}^{b}\tilde{\eta}_{a}^{d}\left( \nabla_{d}\tilde{q
^{w}-\tilde{q}^{0}\nabla_{d}\tilde{m}^{w}-\tilde{m}^{w}\nabla_{d}\tilde{q
^{0}\right) \\
& -\tilde{m}^{b}\tilde{\eta}_{a}^{d}\left( \tilde{n}_{w}\left( \nabla
_{d}\tilde{q}^{w}-\tilde{q}^{0}\nabla_{d}\tilde{m}^{w}-\tilde{m}^{w}\nabla
_{d}\tilde{q}^{0}\right) \right) .
\end{align*}
Using now that
\begin{align*}
\tilde{\eta}_{w}^{b}\tilde{m}^{w} & =0,\\
\tilde{n}_{w}\tilde{q}^{w} & =0,\\
\tilde{n}_{w}\nabla_{d}\tilde{q}^{w} & =-\tilde{q}^{w}\nabla_{d}\tilde
{n}_{w},\\
\tilde{n}_{w}\tilde{m}^{w} & =-1,\\
\tilde{n}_{w}\nabla_{d}\tilde{m}^{w} & =-\tilde{m}^{w}\nabla_{d}\tilde
{n}_{w},
\end{align*}
we conclud
\begin{align*}
\nabla_{a}q^{b} & =-\tilde{m}^{b}n_{a}\left( -p^{d}\left( \left(
\tilde{q}^{w}-\tilde{q}^{0}\tilde{m}^{w}\right) \nabla_{d}\tilde{n
_{w}\right) +p^{d}\nabla_{d}\tilde{q}^{0}\right) \\
& +\tilde{\eta}_{w}^{b}n_{a}\left( p^{d}\left( \nabla_{d}\tilde{q
^{w}-\tilde{q}^{0}\nabla_{d}\tilde{m}^{w}\right) \right) \\
& +\tilde{\eta}_{w}^{b}\tilde{\eta}_{a}^{d}\left( \nabla_{d}\tilde{q
^{w}-\tilde{q}^{0}\nabla_{d}\tilde{m}^{w}\right) \\
& -\tilde{m}^{b}\tilde{\eta}_{a}^{d}\left( -\left( \left( \tilde{q
^{w}-\tilde{q}^{0}\tilde{m}^{w}\right) \nabla_{d}\tilde{n}_{w}\right)
+\nabla_{d}\tilde{q}^{0}\right) .
\end{align*}
Using also that
\begin{align*}
D_{a}\tilde{q}^{b} & :=\tilde{\eta}_{w}^{b}\tilde{\eta}_{a}^{d}\nabla
_{d}\tilde{q}^{w}=\tilde{\eta}_{w}^{b}\tilde{\eta}_{a}^{d}D_{d}\tilde{q
^{w},\\
D_{a}\tilde{q}^{0} & =\tilde{\eta}_{a}^{d}\nabla_{d}\tilde{q}^{0
=\tilde{\eta}_{a}^{d}D_{d}\tilde{q}^{0},\\%
\mathcal{L
_{p}\tilde{q}^{0} & =p^{d}\nabla_{d}\tilde{q}^{0},\\%
\mathcal{L
_{p}\tilde{q}^{w} & =p^{a}\nabla_{a}\tilde{q}^{w}-\tilde{q}^{a}\nabla
_{a}p^{w}\rightarro
\mathcal{L
_{p}\tilde{q}^{w}+\tilde{q}^{a}\nabla_{a}p^{w}=p^{a}\nabla_{a}\tilde{q}^{w},
\end{align*}
we conclud
\begin{align*}
\nabla_{a}q^{b} & =-\tilde{m}^{b}n_{a}\left( -p^{d}\left( \tilde{q
^{w}-\tilde{q}^{0}\tilde{m}^{w}\right) \nabla_{d}\tilde{n}_{w}
\mathcal{L
_{p}\tilde{q}^{0}\right) \\
& +\tilde{\eta}_{w}^{b}n_{a}\left(
\mathcal{L
_{p}\tilde{q}^{w}+\tilde{q}^{r}\nabla_{r}p^{w}-\tilde{q}^{0}p^{d}\nabla
_{d}\tilde{m}^{w}\right) \\
& +\tilde{\eta}_{w}^{b}\tilde{\eta}_{a}^{d}\left( D_{d}\tilde{q}^{w
-\tilde{q}^{0}\nabla_{d}\tilde{m}^{w}\right) \\
& -\tilde{m}^{b}\tilde{\eta}_{a}^{d}\left( -\left( \tilde{q}^{w}-\tilde
{q}^{0}\tilde{m}^{w}\right) \nabla_{d}\tilde{n}_{w}+D_{d}\tilde{q
^{0}\right) .
\end{align*}
Recalling that
\[
\tilde{m}^{w}=\frac{1}{N}p^{w},
\]
we rewrite $\nabla_{r}p^{w}$ as follows
\[
\nabla_{r}p^{w}=\tilde{m}^{w}\nabla_{r}N+N\nabla_{r}\tilde{m}^{w}.
\]
Replacing this expression in the previous development we arrive at the
following resul
\begin{align*}
\nabla_{a}q^{b} & =-\tilde{m}^{b}n_{a}\left( -\left( \tilde{q}^{w
-\tilde{q}^{0}\tilde{m}^{w}\right) p^{d}\nabla_{d}\tilde{n}_{w}
\mathcal{L
_{p}\tilde{q}^{0}\right) \\
& +\tilde{\eta}_{w}^{b}n_{a}\left(
\mathcal{L
_{p}\tilde{q}^{w}+\tilde{q}^{r}N\nabla_{r}\tilde{m}^{w}-\tilde{q}^{0
p^{d}\nabla_{d}\tilde{m}^{w}\right) \\
& +\tilde{\eta}_{w}^{b}\tilde{\eta}_{a}^{d}\left( D_{d}\tilde{q}^{w
-\tilde{q}^{0}\nabla_{d}\tilde{m}^{w}\right) \\
& -\tilde{m}^{b}\tilde{\eta}_{a}^{d}\left( -\left( \tilde{q}^{w}-\tilde
{q}^{0}\tilde{m}^{w}\right) \nabla_{d}\tilde{n}_{w}+D_{d}\tilde{q
^{0}\right) .
\end{align*}
Finally, introducing the definitions
\begin{align*}
-K_{~w}^{b} & :=\tilde{\eta}_{w}^{b}\tilde{\eta}_{a}^{d}\nabla_{d}\tilde
{m}^{w},\\
-K_{ew} & :=\tilde{\eta}_{w}^{d}\tilde{\eta}_{e}^{r}\nabla_{d}\tilde{n
_{r},\\
Z_{d} & :=\tilde{n}_{w}\nabla_{d}\tilde{m}^{w}=-\tilde{m}^{w}\nabla
_{d}\tilde{n}_{w},\\
S_{r} & :=\tilde{\eta}_{r}^{w}\tilde{m}^{d}\nabla_{d}\tilde{n}_{w},\\
S^{r} & :=\tilde{\eta}_{w}^{r}\tilde{m}^{d}\nabla_{d}\tilde{m}^{w},
\end{align*}
we obtai
\begin{align*}
\nabla_{a}q^{b} & =-\tilde{m}^{b}n_{a}\left(
\mathcal{L
_{p}\tilde{q}^{0}-N\left( \left( \tilde{q}^{w}S_{w}\right) +\tilde{q
^{0}\left( \tilde{m}^{d}Z_{d}\right) \right) \right) \\
& +\tilde{\eta}_{w}^{b}n_{a}\left(
\mathcal{L
_{p}\tilde{q}^{w}-\tilde{q}^{r}NK_{~r}^{w}-N\tilde{q}^{0}S^{w}\right) \\
& +\tilde{\eta}_{w}^{b}\tilde{\eta}_{a}^{d}\left( D_{d}\tilde{q}^{w
+\tilde{q}^{0}K_{~d}^{w}\right) \\
& -\tilde{m}^{b}\tilde{\eta}_{a}^{d}\left( D_{d}\tilde{q}^{0}+\tilde{q
^{w}K_{wd}-\tilde{q}^{0}Z_{d}\right)
\end{align*}
In addition, we show that
\begin{align*}
\tilde{m}^{d}\nabla_{d}\tilde{n}_{b} & =-\tilde{m}^{d}\nabla_{d}\left(
N\nabla_{b}t\right) =-\tilde{m}^{d}\nabla_{d}N\nabla_{b}t-\tilde{m
^{d}N\nabla_{d}\left( \nabla_{b}t\right) ,\\
& =\frac{1}{N}\tilde{n}_{b}\tilde{m}^{d}\nabla_{d}N-\tilde{m}^{d}N\nabla
_{b}\left( \nabla_{d}t\right) ,\\
& =\frac{1}{N}\tilde{n}_{b}\tilde{m}^{d}\nabla_{d}N+\tilde{m}^{d}N\nabla
_{b}\left( \frac{1}{N}\tilde{n}_{d}\right) ,\\
& =\frac{1}{N}\tilde{n}_{b}\tilde{m}^{d}\nabla_{d}N-N\nabla_{b}\left(
\frac{1}{N}\right) +\tilde{m}^{d}\nabla_{b}\tilde{n}_{d},\\
& =\frac{1}{N}\tilde{n}_{b}\tilde{m}^{d}\nabla_{d}N+\frac{1}{N}\nabla
_{b}N+\tilde{m}^{d}\nabla_{b}\tilde{n}_{d},\\
& =\frac{1}{N}\tilde{\eta}_{b}^{d}\nabla_{d}N+\tilde{m}^{d}\nabla_{b
\tilde{n}_{d},\\
& =D_{b}\left( \ln N\right) -Z_{b},
\end{align*}
and
\begin{align*
\mathcal{L
_{\tilde{m}}\tilde{\eta}_{r}^{d} & =\tilde{m}^{q}\nabla_{q}\tilde{\eta
_{r}^{d}-\tilde{\eta}_{r}^{q}\nabla_{q}\tilde{m}^{d}+\tilde{\eta}_{q
^{d}\nabla_{r}\tilde{m}^{q},\\
& =\tilde{m}^{q}\nabla_{q}\left( \tilde{m}^{d}\tilde{n}_{r}\right) -\left(
\delta_{r}^{q}+\tilde{m}^{q}\tilde{n}_{r}\right) \nabla_{q}\tilde{m
^{d}+\left( \delta_{q}^{d}+\tilde{m}^{d}\tilde{n}_{q}\right) \nabla
_{r}\tilde{m}^{q},\\
& =\tilde{n}_{r}\tilde{m}^{q}\nabla_{q}\tilde{m}^{d}+\tilde{m}^{d}\tilde
{m}^{q}\nabla_{q}\tilde{n}_{r}-\tilde{n}_{r}\tilde{m}^{q}\nabla_{q}\tilde
{m}^{d}+\tilde{m}^{d}\tilde{n}_{q}\nabla_{r}\tilde{m}^{q},\\
& =\tilde{m}^{d}\tilde{m}^{q}\nabla_{q}\tilde{n}_{r}+\tilde{m}^{d}\tilde
{n}_{q}\nabla_{r}\tilde{m}^{q},\\
& =\tilde{m}^{d}\left( \tilde{m}^{q}\nabla_{q}\tilde{n}_{r}+Z_{r}\right)
,\\
& =\tilde{m}^{d}D_{r}\left( \ln N\right) .
\end{align*}
Notice that
\mathcal{L
_{p}\tilde{\eta}_{r}^{d}$ can be obtained from the latter expression by taking
$N=1$, so that
\
\mathcal{L
_{p}\tilde{\eta}_{r}^{d}=0
\]
\end{proof}
\bigskip
Let us now consider the case when we have a Lorentzian metric $g_{ab}$, such
that in coordinates $\left( t,x^{i}\right) $ it has the following for
\begin{equation}
ds^{2}=\left( -N^{2}+\beta_{k}\beta^{k}\right) dt^{2}+2\beta_{i
dtdx^{i}+\gamma_{ij}dx^{i}dx^{j}. \label{Eq_met_1
\end{equation}
Here $N$ and $\beta^{k}$ have been defined as before. In its matrix mode
$g_{ab}$ and its inverse $g^{bc}$ are given b
\[
g_{ab}=\left[
\begin{array}
[c]{cc
-N^{2}+\beta_{k}\beta^{k} & \beta_{j}\\
\beta_{j} & \gamma_{ij
\end{array}
\right] ,
\
\[
g^{bc}=\left[
\begin{array}
[c]{cc
-\frac{1}{N^{2}} & \frac{\beta^{j}}{N^{2}}\\
\frac{\beta^{j}}{N^{2}} & \gamma^{ij}-\frac{\beta^{i}\beta^{j}}{N^{2}
\end{array}
\right] .
\]
With this ansatz, it is easy to check tha
\[
N=\frac{1}{\sqrt{-\nabla t.\nabla t}},
\]
and if we use the metric $g_{ab}$ (and its inverse) to raise and lower indices
we obtain that
\[
\tilde{m}^{b}=g^{bc}\tilde{n}_{c}.
\]
Notice that the latter expression is equivalent t
\[
\tilde{n}^{b}=\tilde{m}^{b
\]
and furthermore $\tilde{n}_{c}$ is temporal due to the eq. (\ref{Eq_nor_m_n_1
), i.e.
\[
\tilde{n}^{b}\tilde{n}_{b}=\tilde{m}^{b}\tilde{n}_{b}=-1.
\]
Moreover, we see that the projector $\tilde{\eta}_{b}^{a}$ (eq.
(\ref{Eq_proy_b_1})), takes the for
\begin{equation}
\tilde{\eta}_{b}^{a}:=\delta_{b}^{a}+\tilde{n}^{a}\tilde{n}_{b}.
\label{eq_proy_b
\end{equation}
When we lower the $a$ component with $g_{ac}$, the latter expression is a
Riemannian metric on $\Sigma_{t}$, it is
\begin{equation}
\tilde{\eta}_{ab}=g_{ab}+\tilde{n}_{a}\tilde{n}_{b}. \label{eq_3_metrica
\end{equation}
Notice also that when we do not consider a metric we have 2 unrelated types of
"extrinsic curvatures" $K_{~a}^{b}$ and $K_{ba}$. However, when we have a
metric like (\ref{Eq_met_1}) and we consider a Levi-Civita connection
$\nabla_{d}$ then
\begin{align*}
K_{ba} & =g_{bc}K_{~a}^{c},\\
S_{r} & =g_{rc}S^{c},
\end{align*}
furthermore
\[
Z_{d}:=\tilde{n}_{w}\nabla_{d}\tilde{m}^{w}=\tilde{n}_{w}\nabla_{d}\tilde
{n}^{w}=\frac{1}{2}\nabla_{d}\left( \tilde{n}.\tilde{n}\right) =0.
\]
\end{appendices}
\bigskip
\bibliographystyle{plain}
|
\section{Introduction}\label{introduction}}
Massive high-dimensional (HD) data sets are now common in many areas of
scientific inquiry. As new methods are developed for data analysis, a
fundamental challenge lies in designing and conducting simulation
studies to assess the operating characteristics of proposed methodology
--- such as false positive rates, statistical power, interval coverage,
and robustness. Further, efficient simulation empowers statistical
computing strategies, such as the parametric bootstrap (Chernick 2008)
to simulate from a hypothesized null model, providing inference in
analytically challenging settings. Such Monte Carlo (MC) techniques
become difficult for HD dependent data using existing algorithms and
tools. This is particularly true when simulating massive multivariate,
non-normal distributions, arising in many fields of study.
As others have noted, it can be vexing to simulate dependent,
non-normal/discrete data, even for low-dimensional (LD) settings (Madsen
and Birkes 2013; Xiao and Zhou 2019). For continuous non-normal LD
multivariate data, the well-known NORmal To Anything (NORTA) algorithm
(Cario and Nelson 1997) and other copula approaches (Nelsen 2007) are
well-studied and implemented in publicly-available software (Yan 2007;
Chen 2001). Yet these approaches do not scale in a timely fashion to HD
problems (X. Li et al. 2019). For discrete data, early simulation
strategies had major flaws, such as failing to obtain the full range of
possible correlations --- such as admitting only positive correlations
(Park, Park, and Shin 1996). While more recent approaches (Madsen and
Birkes 2013; Xiao 2017; Barbiero and Ferrari 2017) have remedied this
issue for LD problems, the existing tools are not designed to scale to
high dimensions.
Another central issue lies in characterizing dependence between
components in the HD random vector. The choice of correlation in
practice usually relates to the eventual analytic goal and
distributional assumptions of the data (e.g., non-normal, discrete,
infinite support, etc.). For normal data, the Pearson product-moment
correlation describes the dependence perfectly. However, simulating
arbitrary random vectors that match a target Pearson correlation matrix
is computationally intense (Chen 2001; Xiao 2017). On the other hand, an
analyst might consider use of nonparametric correlation measures to
better characterize monotone, non-linear dependence, such as Spearman's
\(\rho\) and Kendall's \(\tau\). Throughout, we focus on matching these
nonparametric dependence measures, as our aim lies in modeling
non-normal data and these rank-based measures possess invariance
properties enabling our proposed methodology. We do, however, implement
Pearson matching, but several layers of approximation are required.
With all this in mind, we present a scalable, flexible multivariate
simulation algorithm. The crux of the method lies in the construction of
a Gaussian copula in the spirit of the NORTA procedure. As we will
describe in more detail, the algorithm's design leverages useful
properties of nonparametric correlation measures, namely invariance
under monotone transformation and well-known closed-form relationships
between dependence measures for the multivariate normal (MVN)
distribution. For our method, we developed a high-performance
implementation: the \texttt{Bigsimr} Julia package, with R and Python
interfaces \texttt{bigsimr}.
This article proceeds by providing background information, including a
description of a motivating example application: RNA-sequencing
(RNA-seq) breast cancer data. Then we describe and justify our
simulation methodology and related algorithms. We proceed by providing
an illustrative LD \texttt{bigsimr} workflow. Next we conduct MC studies
under various bivariate distributional assumptions to evaluate
performance and accuracy. After the MC evaluations, we simulate random
vectors motivated by our RNA-seq example, evaluate the accuracy, and
provide example statistical computing tasks, namely MC estimation of
joint probabilities and evaluating HD correlation estimation efficiency.
Finally, we discuss the method's utility, limitations, and future
directions.
\hypertarget{background}{%
\section{Background}\label{background}}
The \texttt{Bigsimr} package presented here provides multiple
high-performance algorithms that operate with HD multivariate data. All
these algorithms were originally designed to support a single task: to
generate random vectors drawn from multivariate probability
distributions with given marginal distributions and dependency metrics.
Specifically, our goal is to efficiently simulate a large number, \(B\),
of HD random vectors \({\bf Y}=(Y_1, \ldots, Y_d)^\top\) with
\emph{correlated} components and heterogeneous marginal distributions,
described via cumulative distribution functions (CDFs)
\(F_i, i=1,\ldots,d\).
When designing this methodology, we developed the following properties
to guide our effort. We divide the properties into two categories: (1)
basic properties (BP) and `scalability' properties (SP). The BPs are
adapted from an existing criteria due to Nikoloulopoulos (2013). A
suitable simulation strategy should possess the following properties:
\begin{itemize}
\item
BP1: A wide range of dependences, allowing both positive and negative
values, and, ideally, admitting the full range of possible values.
\item
BP2: Flexible dependence, meaning that the number of dependence
parameters can be equal to the number of bivariate pairs in the
vector.
\item
BP3: Flexible marginal modeling, generating heterogeneous data ---
including mixed continuous and discrete marginal distributions.
\end{itemize}
Moreover, the simulation method must \emph{scale} to high dimensions:
\begin{itemize}
\item
SP1: Procedure must scale to high dimensions with practical
computation times.
\item
SP2: Procedure must scale to high dimensions while maintaining
accuracy.
\end{itemize}
\hypertarget{motivating-example-rna-seq-data}{%
\subsection{Motivating example: RNA-seq
data}\label{motivating-example-rna-seq-data}}
Simulating HD, non-normal, correlated data motivates this work, in
pursuit of modeling RNA-sequencing (RNA-seq) data (Wang, Gerstein, and
Snyder 2009; Conesa et al. 2016) derived from breast cancer patients.
The RNA-seq data-generating process involves counting how often a
particular form of messenger RNA (mRNA) is expressed in a biological
sample. RNA-seq platforms typically quantify the entire transcriptome in
one experimental run, resulting in HD data. For human-derived samples,
this results in count data corresponding to over 20,000 genes
(protein-coding genomic regions) or even over 77,000 isoforms when
alternatively-spliced mRNA are counted (Schissler et al. 2019).
Importantly, due to inherent biological processes, gene expression data
exhibit correlation (co-expression) across genes (Efron 2007; Schissler,
Piegorsch, and Lussier 2018).
We illustrate our methodology using the Breast Invasive Carcinoma (BRCA)
data set housed in The Cancer Genome Atlas (TCGA; see Acknowledgments).
For ease of modeling and simplicity of exposition, we only consider high
expressing genes. In turn, we begin by filtering to retain the top r d
of the highest-expressing genes (in terms of median expression) of the
over 20,000 gene measurements from \(N=r nrow(example_brca)\) patients'
tumor samples. This gives a great number of pairwise dependencies among
the marginals (specifically, \(r choose(d,2)\) correlation parameters).
Table \ref{tab:ch010-realDataTab} displays RNA-seq counts for three
selected high-expressing genes for the first five patients' breast tumor
samples. To help visualize the bivariate relationships for these three
selected genes across all patients, Figure \ref{fig:ch010-realDataFig}
displays the marginal distributions and estimated Spearman's
correlations.
\begin{table}
\caption{\label{tab:ch010-realDataTab}mRNA expression for three selected high-expressing genes, STAU1, FKBP1A, NME2, for the first five patients in the TCGA BRCA data set.}
\centering
\begin{tabular}[t]{lrrr}
\toprule
Patient ID & STAU1 & FKBP1A & NME2\\
\midrule
TCGA-A1-A0SB & 10440 & 11354 & 17655\\
TCGA-A1-A0SD & 21523 & 20221 & 14653\\
TCGA-A1-A0SE & 21733 & 22937 & 35251\\
TCGA-A1-A0SF & 11866 & 19650 & 16551\\
TCGA-A1-A0SG & 12486 & 12089 & 10434\\
\bottomrule
\end{tabular}
\end{table}
\begin{figure}
\centering
\includegraphics{ch010-realDataFig-1.pdf}
\caption{\label{fig:ch010-realDataFig}Pairwise scatterplots for three
example genes (Table 1), with estimated Spearman's correlations and
marginal density estimates displayed. The data possess outliers,
heavy-right tails, are discrete, and have non-trivial intergene
correlations. Modeling these data motivate our simulation methodology.}
\end{figure}
\hypertarget{measures-of-dependency}{%
\subsection{Measures of dependency}\label{measures-of-dependency}}
In multivariate analysis, an analyst must select a metric to quantify
dependency. The most widely-known is the Pearson correlation coefficient
that describes the linear association between two random variables \(X\)
and \(Y\), and, it is given by
\begin{equation}
\rho_P(X,Y) = \frac{E(XY) - E(X)E(Y)}{\left[ \mathrm{Var}(X)\mathrm{Var}(Y)\right]^{1/2}}.
\label{eq:pearson}
\end{equation}
As Madsen and Birkes (2013) and Mari and Kotz (2001) discuss, for a
bivariate normal \((X,Y)\) random vector, the Pearson correlation
completely describes the dependency between the components. For
non-normal marginals with monotone correlation patterns, however,
\(\rho_P\) suffers some drawbacks and may mislead or fail to capture
important relationships (Mari and Kotz 2001). Alternatively, analysts
often prefer rank-based correlation measures to describe the degree of
monotonic association.
Two nonparametric, rank-based measures common in practice are Spearman's
correlation (denoted \(\rho_S\)) and Kendall's \(\tau\). Spearman's
\(\rho_S\) has an appealing correspondence as the Pearson correlation
coefficient on \emph{ranks} of the values, thereby capturing nonlinear
yet monotone relationships. Kendall's \(\tau\), on the other hand, is
the difference in probabilities of concordant and discordant pairs of
observations, \((X_i, Y_i)\) and \((X_j, Y_j)\) (concordance meaning
that orderings have the same direction, e.g., if \(X_i < X_j\), then
\(Y_i < Y_j\)). Note that concordance is determined by the ranks of the
values, not the values themselves. Both \(\tau\) and \(\rho_S\) are
\emph{invariant under monotone transformations} of the underlying random
variates. As we will describe more fully in the
\protect\hyperlink{algorithms}{Algorithms} section, this property
enables matching rank-based correlations with speed (SP1) and accuracy
(SP2).
\emph{Correspondence among Pearson, Spearman, and Kendall correlations}
There is no closed form, general correspondence among the rank-based
measures and the Pearson correlation coefficient, as the marginal
distributions \(F_i\) are intrinsic in their calculation. For
\emph{bivariate normal vectors}, however, the correspondence is
well-known:
\begin{equation}
\label{eq:convertKendall}
\rho_{P} = \sin \left( \tau \times \frac{\pi}{2} \right),
\end{equation}
\noindent and similarly for Spearman's \(\rho\) (Kruskal 1958),
\begin{equation}
\label{eq:convertSpearman}
\rho_P = 2 \times \sin \left( \rho_S \times \frac{\pi}{6} \right).
\end{equation}
\emph{Marginal-dependent bivariate correlation bounds}
Given two marginal distributions, \(\rho_P\) is not free to vary over
the entire range of possible correlations \([-1,1]\). The
\emph{Frechet-Hoeffding bounds} are well-studied (Nelsen 2007; Barbiero
and Ferrari 2017). These constraints cannot be overcome through
algorithm design. In general, the bounds are given by
\begin{equation}
\label{eq:frechet}
\rho_P^{max} = \rho_P \left( F^{-1}_1 (U), F^{-1}_2 (U) \right), \quad \rho_P^{min} = \rho_P \left( F^{-1}_1 (U), F^{-1}_2 (1 - U) \right)
\end{equation}
\noindent where \(U\) is a uniform random variable on \((0,1)\) and
\(F^{-1}_1, F^{-1}_2\) are the inverse CDFs of \(X_1\) and \(X_2\),
respectively, defined as
\begin{equation}
F_{i}^{-1} = \inf\{y:F_{i}(y) \geq u \}
\label{eq:inverseCDF}
\end{equation}
when the variables are discrete.
\hypertarget{gaussian-copulas}{%
\subsection{Gaussian copulas}\label{gaussian-copulas}}
There is a connection of our simulation strategy to Gaussian
\emph{copulas} (Nelsen 2007). A copula is a distribution function on
\([0,1]^d\) that describes a multivariate probability distribution with
standard uniform marginals. This provides a powerful, natural way to
characterize joint probability. Consequently, the study of copulas is an
important and active area of statistical theory and practice. For any
random vector \({\bf X}=(X_1, \ldots, X_d)^\top\) with CDF \(F\) and
marginal CDFs \(F_i\) there is a copula function \(C(u_1, \ldots, u_d)\)
satisfying
\begin{equation}
F(x_1, \ldots, x_d) = {\mathbb P}(X_1\leq x_1, \ldots,X_d\leq x_d) = C(F_1(x_1), \ldots, F_d(x_d)), \quad x_i \in {\mathbb R}, i=1,\ldots,d.
\label{eq:copula}
\end{equation}
A Gaussian copula has marginal CDFs that are all standard normal,
\(F_i = \Phi, \forall \, i\), . This representation corresponds to a
multivariate normal (MVN) distribution with standard normal marginal
distributions and covariance matrix \(R_P\). Since the marginals are
standardized to have unit variance, this \(R_P\) is also Pearson
correlation matrix. If \(F_R\) is the CDF of such a multivariate normal
distribution, then the corresponding Gaussian copula \(C_R\) is defined
through
\begin{equation}
\label{eq:gauss}
F_R(x_1, \ldots, x_d) = C_R(\Phi(x_1), \ldots, \Phi(x_d)),
\end{equation}
where \(\Phi(\cdot)\) is the standard normal CDF.
Here, the copula function \(C_R\) is the CDF of the random vector
\((\Phi(X_1), \ldots, \Phi(X_d))\), where
\((X_1, \ldots, X_d) \sim N_d({\bf 0}, R_P)\).
Sklar's Theorem (Úbeda-Flores and Fernández-Sánchez 2017) guarantees
that given inverse CDFs \(F_i^{-1}\)s and a valid correlation matrix
(within the Frechet bounds) a random vector can be obtained via
transformations involving copula functions. For example, using Gaussian
copulas, we can construct a random vector
\({\bf Y} = (Y_1, \ldots, Y_d)^\top\) with \(Y_i \sim F_i\),
viz.~\(Y_i = F_i^{-1}(\Phi(X_i)), i=1, \ldots, d\), where
\((X_1, \ldots, X_d) \sim N_d({\bf 0}, R_P)\).
\hypertarget{other-multivariate-simulation-packages}{%
\subsection{Other Multivariate Simulation
Packages}\label{other-multivariate-simulation-packages}}
To place \texttt{Bigsimr}'s implementation and algorithms in context of
existing approaches, we qualitatively compare \texttt{Bbigsimr} to other
multivariate simulation packages. In \texttt{R}, the \texttt{MASS}
package's \texttt{mvnorm()} has long been available to produce random
vectors from multivariate normal (and \(t\)) distributions. This
procedure cannot simulate from arbitrary margins and dependency measures
and it fails to readily scale to high dimensions. SAS PROC SIMNORMAL and
Python's NumPy \texttt{random.multivariate\_normal} both do similar
tasks. Recently, the high-performance \texttt{R} package
\texttt{mvnfast} (Fasiolo 2016) has been released, which provides HD
multivariate normal generation. However, lacks flexibility in marginal
models and dependency modeling. For dependent discrete data, LD
\texttt{R} packages exist, such as \texttt{GenOrd}. For copula
computation, the full-featured \texttt{R} package \texttt{copula}
provides many different copulas and a fairly general LD random generator
\texttt{rCopula()}. High-dimesional random vector generation via
\texttt{copula} takes an impractical amount of computing time (X. Li et
al. 2019). The \texttt{nortaRA} R package matches Pearson coefficients
near exactly and is reasonable for LD settings. Finally, we know of no
other \texttt{R} package that can produce HD random vectors with
flexible margins and dependence structure. Table \ref{tab:compare-table}
summarizes the above discussion.
\begin{table}[h]
\centering
\caption{\label{tab:compare-table}Comparison of selected features of bigsimr to existing R packages for multivariate simulation.}
\begin{tabular}{@{}llll@{}}
Package & Scales to HD & Hetereogenous Margins & Flexible Dependence \\ \midrule
MASS & No & No & No \\
mvnfast & Yes & No & No \\
GenOrd & No & Discrete margins & Yes \\
copula & No & Yes & No \\
nortaRA & No & Yes & No \\
bigsimr & Yes & Yes & Yes
\end{tabular}
\end{table}
\hypertarget{algorithms}{%
\section{Algorithms}\label{algorithms}}
This section describes our methods for simulating a random vector
\(\bf Y\) with \(Y_i\) components for \(i=1,\ldots,d\). Each \(Y_i\) has
a specified marginal CDF \(F_i\) and its inverse \(F^{-1}_i\). To
characterize dependency, every pair \((Y_i, Y_j)\) has a given Pearson
correlation \(\rho_P\), Spearman correlation \(\rho_S\), and/or
Kendall's \(\tau\). The method can be described as a
\emph{high-performance Gaussian copula} (Equation \ref{eq:gauss})
providing a HD NORTA-inspired algorithm.
\hypertarget{normal-to-anything-norta}{%
\subsection{NORmal To Anything (NORTA)}\label{normal-to-anything-norta}}
The well-known NORTA algorithm (Cario and Nelson 1997) simulates a
random vector \(\bf Y\) with variance-covariance matrix
\(\Sigma_{\bf Y}\). Specifically, NORTA algorithm proceeds as follows:
\begin{enumerate}
\def\arabic{enumi}.{\arabic{enumi}.}
\item
Simulate a random vector \(\bf Z\) with \(d\) independent and
identically distributed (iid) standard normal components.
\item
Determine the input matrix \(\Sigma_{\bf Z}\) that corresponds with
the specified output \(\Sigma_{\bf Y}\).
\item
Produce a Cholesky factor \(M\) of \(\Sigma_{\bf Z}\) such that
\(M M^{\prime}=\Sigma_{\bf Z}\).
\item
Set \(X\) by \(X \gets MZ\).
\item
\(\text{Return} \; {\bf Y} \; \text{where} \; Y_i \gets F_i^{-1}[\Phi(X_i)], \; i=1,...,d\).
\end{enumerate}
With modern parallel computing, steps 1, 3, 4, 5 are readily implemented
as high-performance, multi-core and/or graphical-processing-unit (GPU)
accelerated algorithms --- providing fast scalability to HD.
Matching specified Pearson correlation coefficients exactly (step 2
above), however, is computationally costly. In general, there is no
convenient correspondence between the components of the input
\(\Sigma_{\bf Z}\) and target \(\Sigma_{\bf Y}\). Matching the
correlations involves evaluating or approximating \(\binom{d}{2}\)
integrals of the form
\begin{equation}
\mathrm{E}\left[Y_i Y_j\right] = \int_{-\infty}^{\infty} \int_{-\infty}^{\infty} F_i^{-1}\left[\Phi(z_i)\right] F_j^{-1}\left[\Phi(z_j)\right] \phi(z_i, z_j, \rho_z) dz_i dz_j,
\label{eq:pearsonIntegralRelation}
\end{equation}
where \(\phi(\cdot)\) is the joint probability density function of two
correlated standard normal variables. For HD simulation, exact
evaluation may become too costly to enable practical simulation studies.
Our remedy to enable HD Pearson matching in \texttt{bigsimr} is Hermite
polynomial approximation of these \(\binom{d}{2}\) double integrals, as
suggested by Xiao and Zhou (2019).
\emph{NORTA in higher dimensions}
Sklar's theorem provides a useful characterization of multivariate
distributions through copulas. In practice, however, the particulars of
the simulation algorithm affect which joint distributions may be
simulated. Even in LD spaces (e.g., \(d=3\)), there exist valid
multivariate distributions with \emph{feasible} Pearson correlation
matrices that NORTA cannot match exactly (S. T. Li and Hammond 1975).
This occurs when the bivariate transformations are applied to find the
input correlation matrix, yet when combined, the resultant matrix
becomes non-positive definite. These situations do occur, even using
exact analytic calculations. Such problematic target correlation
matrices are termed \emph{NORTA defective}.
Ghosh and Henderson (2002) conducted a study to estimate the probability
of encountering NORTA defective matrices while increasing the dimension
\(d\). They found that for what is now considered low-to-moderate
dimensions (\(d \approx 20\)), almost \emph{all} feasible matrices are
NORTA defective. This stems from the concentration of measure near the
boundary of the space of all possible correlation matrices as dimension
increases. Unfortunately, it is precisely near this boundary that NORTA
defective matrices reside.
There is hope, however, as Ghosh and Henderson (2002) also showed that
replacing an non-positive definite input correlation matrix with a close
proxy will give approximate matching to the target --- with adequate
performance for moderate \(d\). This provides evidence that our nearest
positive definite (PD) augmented approach has promise to provide
adequate accuracy if our input matching scheme returns an indefinite
Pearson correlation matrix.
\hypertarget{rand-vec-gen}{%
\subsection{Random vector generator}\label{rand-vec-gen}}
We now describe our algorithm to generate random vectors, which extends
NORTA to HD via nearest correlation matrix replacement and provides
rank-based dependency matching:
\begin{enumerate}
\def\arabic{enumi}.{\arabic{enumi}.}
\item
Mapping step, either
\begin{itemize}
\item
Convert the target Spearman correlation matrix \(R_S\) to the
corresponding MVN Pearson correlation \(R_X\). Alternatively,
\item
Convert the target Kendall \(\tau\) matrix \(R_K\) to the
corresponding MVN Pearson correlation \(R_X\). Alternatively,
\item
Convert the target Pearson correlation matrix to \(R_P\) to the
corresponding, approximate MVN Pearson correlation \(R_X\).
\end{itemize}
\item
Check correlation matrix admissibility and, if not, compute the
nearest correlation matrix.
\begin{itemize}
\item
Check that \(R_X\) is a correlation matrix, a positive definite
matrix with 1's along the diagonal.
\item
If the mapping produced an \(R_X\) that is a correlation matrix, we
describe the matrix as \textbf{admissible} in this scheme and retain
\(R_X\). Otherwise,
\item
Replace \(R_X\) with the nearest correlation matrix \(\tilde{R}_X\),
in the Frobenius norm and the result will be approximate.
\end{itemize}
\item
Gaussian copula
\begin{itemize}
\item
Generate \({\bf X}=(X_1, \ldots, X_d) \sim N_d({\bf 0}, R_X)\).
\item
Transform \({\bf X}\) to \({\bf U} = (U_1, \ldots, U_d)\)
viz.~\(U_i=\Phi(X_i)\), \(i=1, \ldots, d\).
\item
Return \({\bf Y} = (Y_1, \ldots, Y_d)\), where
\(Y_i=F_i^{-1}(U_i)\), \(i=1, \ldots, d\).
\end{itemize}
\end{enumerate}
\emph{Step 1: Mapping step.} We first employ the closed-form
relationships between \(\rho_S\) and \(\tau\) with \(\rho_P\) for
bivariate normal random variables via Equations \ref{eq:convertKendall}
and \ref{eq:convertSpearman}, respectively (implemented as
\texttt{cor\_covert()}). Initializing our algorithm to match the
nonparametric correlations by computing these equations for all pairs is
computationally trivial.
For the computationally expensive process to match Pearson correlations,
we approximate Equation \ref{eq:pearsonIntegralRelation} for all pairs
of margins. To this end, we implement the approximation scheme
introduced by (Xiao and Zhou 2019). Briefly, the many double integrals
of the form in \ref{eq:pearsonIntegralRelation} are approximated by
weighted sums of Hermite polynomials. Matching coefficients for pairs of
continuous distributions is made tractable by this method, but for
discrete distributions (especially discrete distributions with large
support sets or infinite support), the approximation is computationally
expensive. To remedy this, we further approximate discrete distributions
by a continuous distribution using Generalized S-Distributions (Muino,
Voit, and Sorribas 2006).
\emph{Step 2: Admissibility check and nearest correlation matrix
computation.} Once \(R_X\) has been determined, we check admissibility
of the adjusted correlation via the steps described above. If \(R_X\) is
not a valid correlation matrix, then we compute the nearest correlation
matrix. Finding the nearest correlation matrix is a common statistical
computing problem. The defacto function in \texttt{R} is
\texttt{Matrix::nearPD}, an alternating projection algorithm due to
Higham (2002). As implemented, the function fails to scale to HD.
Instead, we provide the quadratically-convergent algorithm based on the
theory of strongly semi-smooth matrix functions (Qi and Sun 2006). The
nearest correlation matrix problem can be written down as the following
convex optimization problem:
\(\mathrm{min} \quad \frac{1}{2} \Vert R_X - X \Vert^2, \quad \mathrm{s.t.} \quad X_{ii} = 1, \quad i = 1, \ldots , n, \quad X \in S_{+}^{n}\).
For nonparametric correlation measures, our algorithm allows the
generation of HD multivariate data with arbitrary marginal distributions
with a broad class of admissible Spearman correlation matrices and
Kendall \(\tau\) matrices. The admissible classes consist of the
matrices that map to a Pearson correlation matrix for a MVN. In
particular, if we let \(X\) be MVN with \(d\) components and denote
\(\Omega_P = \{ R_P : R_P \textrm{ is a Pearson correlation matrix for } X \}, \quad \Omega_K = \{ R_K : R_K \textrm{ is a Kendall correlation matrix for } X \}, \quad \Omega_S = \{ R_S : R_S \textrm{ is a Spearman correlation matrix for } X \}\).
There are 1-1 mappings between these sets. We conjecture that the sets
of admissible \(R_S\) and \(R_K\) are not highly restrictive. In
particular, \(R_P\) is approximately \(R_S\) for a MVN, suggesting that
the admissible set \(\Omega_S\) should be flexible as \(R_P\) can be any
PD matrix with 1's along the diagonal. We provide methods to check
whether a target \(R_S\) is an element of the \emph{admissible set}
\(\Omega_S\).
There is an increasing probability of encountering an inadmissible
correlation matrix as dimension increases. In our experience, the
mapping step for large \(d\) almost always produces a \(R_X\) that is
not a correlation matrix. In Section \ref{package}, we provide a basic
method of how to quantify and control the approximation error. Further,
the RNA-seq example in Section \ref{examples} provides an illustration
of this in practice.
\emph{Step 3: Gaussian copula.} The final step implements a
NORTA-inspired, Gaussian copula approach to produce the desired margins.
Steps 1 and 2 determine the MVN Pearson correlation values that will
eventually match the target correlation. Step 3 requires a fast MVN
simulator, a standard normal CDF, and well-defined quantile functions
for marginals. The MVN is transformed to a copula (distribution with
standard uniform margins) by applying the normal CDF \(\Phi(\cdot)\).
Finally, the quantile functions \(F_i^{-1}\) are applied across the
margins to return the desired random vector \({\bf Y}\).
\hypertarget{the-bigsimr-r-package}{%
\section{The bigsimr R package}\label{the-bigsimr-r-package}}
This section describes a bivariate random vector simulation workflow via
the \texttt{bigsimr} \texttt{R} package. This \texttt{R} package
provides an interface to the native code written in Julia (registered as
the \texttt{Bigsimr} Julia package). In addition to the native Julia
\texttt{Bigsimr} package and \texttt{R} interface \texttt{bigsimr}, we
also provide a Python interface \texttt{bigsimr} that interfaces with
the Julia \texttt{Bigsimr} package. The Julia package provides a
high-performance implementation of our proposed random vector generation
algorithm and associated functions (see Section \ref{algorithms}).
The subsections below describe the basic use of the \texttt{bigsimr} R
package by stepping through an example workflow using the data set
\texttt{airquality} that contains daily air quality measurements in New
York, May to September 1973 (Chambers et al. 1983). This workflow
proceeds from setting up the computing environment, to data wrangling,
estimation, simulation configuration, random vector generation, and,
finally, result visualization.
\hypertarget{bivariate-example}{%
\subsection{Bivariate example}\label{bivariate-example}}
We illustrate the use of \texttt{bigsimr} using the New York air quality
data set (\texttt{airquality}) included in the R \texttt{datasets}
package. First, we load the \texttt{bigsimr} library and a few other
convenient data science packages, including the syntactically-elegant
\texttt{tidyverse} suite of \texttt{R} packages. The code chunk below
prepares the computing environment:
\begin{Shaded}
\begin{Highlighting}[]
\FunctionTok{library}\NormalTok{(}\StringTok{"tidyverse"}\NormalTok{)}
\FunctionTok{library}\NormalTok{(}\StringTok{"bigsimr"}\NormalTok{)}
\CommentTok{\# Activate multithreading in Julia}
\FunctionTok{Sys.setenv}\NormalTok{(}\AttributeTok{JULIA\_NUM\_THREADS =}\NormalTok{ parallel}\SpecialCharTok{::}\FunctionTok{detectCores}\NormalTok{())}
\CommentTok{\# Load the Bigsimr and Distributions Julia packages}
\NormalTok{bs }\OtherTok{\textless{}{-}} \FunctionTok{bigsimr\_setup}\NormalTok{()}
\NormalTok{dist }\OtherTok{\textless{}{-}} \FunctionTok{distributions\_setup}\NormalTok{()}
\end{Highlighting}
\end{Shaded}
Here, we describe a minimal working example --- a bivariate simulation
of two airquality variables: \texttt{Temperature}, in degrees
Fahrenheit, and \texttt{Ozone} level, in parts per billion.
\begin{Shaded}
\begin{Highlighting}[]
\NormalTok{df }\OtherTok{\textless{}{-}}\NormalTok{ airquality }\SpecialCharTok{\%\textgreater{}\%} \FunctionTok{select}\NormalTok{(Temp, Ozone) }\SpecialCharTok{\%\textgreater{}\%} \FunctionTok{drop\_na}\NormalTok{()}
\FunctionTok{glimpse}\NormalTok{(df)}
\end{Highlighting}
\end{Shaded}
\begin{verbatim}
## Rows: 116
## Columns: 2
## $ Temp <int> 67, 72, 74, 62, 66, 65, 59, 61, 74, 69, 66, 68, 58, 64, 66, 5...
## $ Ozone <int> 41, 36, 12, 18, 28, 23, 19, 8, 7, 16, 11, 14, 18, 14, 34, 6, ...
\end{verbatim}
Figure \ref{fig:ch030-aq-joint-dist} visualizes the bivariate
relationship between \texttt{Ozone} and \texttt{Temperature}. We aim to
simulate random two-component vectors mimicking this structure. The
margins are not normally distributed; \texttt{Ozone} level exhibits a
strong positive skew.
\begin{figure}
\centering
\includegraphics{ch030-aq-joint-dist-1.pdf}
\caption{\label{fig:ch030-aq-joint-dist}Bivariate scatterplot of Ozone
versus Temp with estimated marginal densities. The Ozone data are
modeled marginally as log-normal and the Temperature data as normal.}
\end{figure}
Next, we specify the marginal distributions and correlation coefficient
(both type and magnitude). Here the analyst is free to be creative. For
this example, we avoid goodness-of-fit considerations to determine the
marginal distributions. It is sensible without domain knowledge to
estimate these quantities from the data, and \texttt{bigsimr} contains
fast functions designed for this task.
\hypertarget{specifying-marginal-distributions}{%
\subsection{Specifying marginal
distributions}\label{specifying-marginal-distributions}}
Based on the estimated densities in Figure
\ref{fig:ch030-aq-joint-dist}, we assume \texttt{Temp} is normally
distributed and \texttt{Ozone} is log-normally distributed, as the
latter values are positive and skewed. We use the classical unbiased
estimators for the normal distribution's parameters and maximum
likelihood estimators for the log-normal parameters:
\begin{Shaded}
\begin{Highlighting}[]
\NormalTok{df }\SpecialCharTok{\%\textgreater{}\%} \FunctionTok{select}\NormalTok{(Temp) }\SpecialCharTok{\%\textgreater{}\%}
\FunctionTok{summarise\_all}\NormalTok{(}\AttributeTok{.funs =} \FunctionTok{c}\NormalTok{(}\AttributeTok{mean =}\NormalTok{ mean, }\AttributeTok{sd =}\NormalTok{ sd))}
\end{Highlighting}
\end{Shaded}
\begin{verbatim}
## mean sd
## 1 77.87069 9.485486
\end{verbatim}
\begin{Shaded}
\begin{Highlighting}[]
\NormalTok{mle\_mean }\OtherTok{\textless{}{-}} \ControlFlowTok{function}\NormalTok{(x) }\FunctionTok{mean}\NormalTok{(}\FunctionTok{log}\NormalTok{(x))}
\NormalTok{mle\_sd }\OtherTok{\textless{}{-}} \ControlFlowTok{function}\NormalTok{(x) }\FunctionTok{mean}\NormalTok{( }\FunctionTok{sqrt}\NormalTok{( (}\FunctionTok{log}\NormalTok{(x) }\SpecialCharTok{{-}} \FunctionTok{mean}\NormalTok{(}\FunctionTok{log}\NormalTok{(x)))}\SpecialCharTok{\^{}}\DecValTok{2}\NormalTok{ ) )}
\NormalTok{df }\SpecialCharTok{\%\textgreater{}\%}
\FunctionTok{select}\NormalTok{(Ozone) }\SpecialCharTok{\%\textgreater{}\%}
\FunctionTok{summarise\_all}\NormalTok{(}\AttributeTok{.funs =} \FunctionTok{c}\NormalTok{(}\AttributeTok{meanlog =}\NormalTok{ mle\_mean, }\AttributeTok{sdlog =}\NormalTok{ mle\_sd))}
\end{Highlighting}
\end{Shaded}
\begin{verbatim}
## meanlog sdlog
## 1 3.418515 0.6966689
\end{verbatim}
Now, we configure the input marginals for later input into
\texttt{rvec}. The marginal distributions are specified using
\texttt{Julia}'s \texttt{Distributions} package and stored in a vector.
\begin{Shaded}
\begin{Highlighting}[]
\NormalTok{margins }\OtherTok{\textless{}{-}} \FunctionTok{c}\NormalTok{(dist}\SpecialCharTok{$}\FunctionTok{Normal}\NormalTok{(}\FunctionTok{mean}\NormalTok{(df}\SpecialCharTok{$}\NormalTok{Temp), }\FunctionTok{sd}\NormalTok{(df}\SpecialCharTok{$}\NormalTok{Temp)),}
\NormalTok{ dist}\SpecialCharTok{$}\FunctionTok{LogNormal}\NormalTok{(}\FunctionTok{mle\_mean}\NormalTok{(df}\SpecialCharTok{$}\NormalTok{Ozone), }\FunctionTok{mle\_sd}\NormalTok{(df}\SpecialCharTok{$}\NormalTok{Ozone)))}
\end{Highlighting}
\end{Shaded}
\hypertarget{specifying-correlation}{%
\subsection{Specifying correlation}\label{specifying-correlation}}
The user must decide how to describe correlation based on the
particulars of the problem. For non-normal data in our scheme, we
advocate the use of Spearman's \(\rho\) correlation matrix \(R_S\) or
Kendall's \(\tau\) correlation matrix \(R_K\). We also support Pearson
correlation coefficient matching, while cautioning the user to check the
performance for the distribution at hand (see
\protect\hyperlink{simulations}{Monte Carlo evaluations} below for
evaluation strategies and guidance). Note that these estimation methods
are classical approaches, not specifically designed for high-dimensional
correlation estimation.
\begin{Shaded}
\begin{Highlighting}[]
\NormalTok{(R\_S }\OtherTok{\textless{}{-}}\NormalTok{ bs}\SpecialCharTok{$}\FunctionTok{cor}\NormalTok{(}\FunctionTok{as.matrix}\NormalTok{(df), bs}\SpecialCharTok{$}\NormalTok{Spearman))}
\end{Highlighting}
\end{Shaded}
\begin{verbatim}
## [,1] [,2]
## [1,] 1.000000 0.774043
## [2,] 0.774043 1.000000
\end{verbatim}
\hypertarget{checking-target-correlation-matrix-admissibility}{%
\subsection{Checking target correlation matrix
admissibility}\label{checking-target-correlation-matrix-admissibility}}
Once a target correlation matrix is specified, first convert the
dependency type to Pearson to correctly specify the MVN inputs. Then,
check that the converted matrix \(R_X\) is a valid correlation matrix
(PD and 1s on the diagonal). For this bivariate example, \(R_X\) is a
valid correlation matrix. Note that typically in HD the resultant MVN
Pearson correlation matrix is indefinite and requires approximation (see
the HD examples in subsequent sections).
\begin{Shaded}
\begin{Highlighting}[]
\CommentTok{\# Step 1. Mapping}
\NormalTok{(R\_X }\OtherTok{\textless{}{-}}\NormalTok{ bs}\SpecialCharTok{$}\FunctionTok{cor\_convert}\NormalTok{(R\_S, bs}\SpecialCharTok{$}\NormalTok{Spearman, bs}\SpecialCharTok{$}\NormalTok{Pearson))}
\end{Highlighting}
\end{Shaded}
\begin{verbatim}
## [,1] [,2]
## [1,] 1.0000000 0.7885668
## [2,] 0.7885668 1.0000000
\end{verbatim}
\begin{Shaded}
\begin{Highlighting}[]
\CommentTok{\# Step 2. Check admissibility}
\NormalTok{bs}\SpecialCharTok{$}\FunctionTok{iscorrelation}\NormalTok{(R\_X)}
\end{Highlighting}
\end{Shaded}
\begin{verbatim}
## [1] TRUE
\end{verbatim}
Despite being a valid correlation matrix under this definition, marginal
distributions induce Frechet bounds on the possible Pearson correlation
values. To check that the converted correlation values fall within these
bounds, use \texttt{bigsimr::cor\_bounds} to estimate the pairwise lower
and upper correlation bounds. \texttt{cor\_bounds} uses the Generate,
Sort, and Correlate algorithm of Demirtas and Hedeker (2011). For the
assumed marginals, the Pearson correlation coefficient is not free to
vary in {[}-1,1{]}, as seen below. Our MC estimate of the bounds
slightly underestimates the theoretical bounds of \((-0.881, 0.881)\).
(See an analytic derivation presented in the Appendix). Since our single
Pearson correlation coefficient is within the theoretical bounds, the
correlation matrix is valid for our simulation strategy.
\begin{Shaded}
\begin{Highlighting}[]
\NormalTok{bs}\SpecialCharTok{$}\FunctionTok{cor\_bounds}\NormalTok{(margins[}\DecValTok{1}\NormalTok{], margins[}\DecValTok{2}\NormalTok{], bs}\SpecialCharTok{$}\NormalTok{Pearson, }\AttributeTok{n\_samples =} \FloatTok{1e6}\NormalTok{)}
\end{Highlighting}
\end{Shaded}
\begin{verbatim}
## Julia Object of type NamedTuple{(:lower, :upper),Tuple{Float64,Float64}}.
## (lower = -0.8829854580269612, upper = 0.882607249144551)
\end{verbatim}
In this 2D example, it is possible to verify the existence of a
bivariate distribution with the target Pearson correlation / margins
(See the {[}Appendix{]}). For general \(d-\)variate distributions, is it
not guaranteed that such a multivariate distribution exists with exact
the specified correlations and margins (Barbiero and Ferrari 2017).
Theoretical guidance is lacking in this regard, but, in practice, we
find that using bivariate bounds form rough approximations to the bounds
for the feasible region for the \(d-\)variate construction when
aggregated into a Pearson matrix. See the
\protect\hyperlink{examples}{RNA-seq data application} section for an
example discussing this, NORTA feasibility, and use of the nearest PD
matrix algorithm.
\hypertarget{simulating-random-vectors}{%
\subsection{Simulating random vectors}\label{simulating-random-vectors}}
Finally, we execute \texttt{rvec} to simulate the desired \(10,000\)
random vectors from the assumed bivariate distribution of \texttt{Ozone}
and \texttt{Temp}. Note that the target input is specified using the
pre-computed MVN correlation matrix. Figure \ref{fig:ch030-plot-sim}
plots the 10,000 simulated points.
\begin{Shaded}
\begin{Highlighting}[]
\NormalTok{x }\OtherTok{\textless{}{-}}\NormalTok{ bs}\SpecialCharTok{$}\FunctionTok{rvec}\NormalTok{(}\DecValTok{10000}\NormalTok{, R\_X, margins)}
\NormalTok{df\_sim }\OtherTok{\textless{}{-}} \FunctionTok{as.data.frame}\NormalTok{(x)}
\FunctionTok{colnames}\NormalTok{(df\_sim) }\OtherTok{\textless{}{-}} \FunctionTok{colnames}\NormalTok{(df)}
\end{Highlighting}
\end{Shaded}
\begin{figure}
\centering
\includegraphics{ch030-plot-sim-1.pdf}
\caption{\label{fig:ch030-plot-sim}Contour plot and marginal densities
for the simulated bivariate distribution of Air Quality Temperatures and
Ozone levels. The simulated points mimic the observed data with respect
to both the marginal characteristics and bivariate association.}
\end{figure}
\hypertarget{simulations}{%
\section{Monte Carlo evaluations}\label{simulations}}
In this section, we conduct MC studies to investigate method
performance. Marginal parameter matching is straightforward in our
scheme as it is a sequence of univariate inverse probability transforms.
This process is both fast and exact. On the other hand, accurate and
computationally efficient dependency matching presents a challenge,
especially in the HD setting. To evaluate our methods in those respects,
we design the following numerical experiments to first assess accuracy
of matching dependency parameters in bivariate simulations and then time
the procedure in increasingly large dimension \(d\).
\hypertarget{bivariate-experiments}{%
\subsection{Bivariate experiments}\label{bivariate-experiments}}
We select bivariate simulation configurations to ultimately simulate our
motivating discrete-valued RNA-seq example. Thus, we proceed by
increasing departure from normality, leading to a multivariate negative
binomial (MVNB) model for our motivating data. We begin with empirically
evaluating the dependency matching across all three supported
correlations --- Pearson, Spearman, and Kendall --- in identical,
bivariate marginal configurations. For each pair of identical margins,
we vary the target correlation across \(\Omega\), the set of possible
admissible values for each correlation type, to evaluate the
simulation's ability to generate all possible correlations. The
simulations progress from bivariate normal, to bivariate gamma
(non-normal yet continuous), and bivariate negative binomial (non-normal
and discrete).
Table \ref{tab:sims} lists our identical-marginal, bivariate simulation
configurations. We increase the simulation replicates \(B\) to visually
ensure that our results converge to the target correlations and gauge
efficiency. We select distributions beginning with a standard
multivariate normal (MVN) as we expect the performance to be exact (up
to MC error) for all correlation types. Then, we select a non-symmetric
continuous distribution: a standard (rate =1), two-component
multivariate gamma (MVG). Finally, we select distributions and marginal
parameter values that are motivated by our RNA-seq data, namely values
proximal to probabilities and sizes estimated from the data (see
\href{examples}{RNA-seq data application} for estimation details).
Specifically, we assume a MVNB with
\(p_1 = p_2 = 3\times10^{-4}, r_1 = r_2 = 4, \rho \in \Omega\).
\begin{table}[]
\centering
\caption{ \label{tab:sims} Identical margin, bivariate simulation configurations to evaluate matching accuracy.}
\begin{tabular}{@{}lcr@{}}
\toprule
Simulation Reps $B$ & Correlation Types & Identical-margin 2D distribution \\ \midrule
$1,000$ & Pearson ($\rho_P$) & ${\bf Y} \sim MVN( \mu= 0 , \sigma = 1, \rho_i ), i=1,\ldots,100$ \\
$10,000$ & Spearman ($\rho_S$) & ${\bf Y} \sim MVG( shape = 10, rate = 1, \rho_i ), i=1,\ldots,100$ \\
$100,000$ & Kendall ($\tau$) & ${\bf Y} \sim MVNB(p = 3\times10^{-4}, r = 4,\rho_i), i=1,\ldots,100$ \\ \bottomrule
\end{tabular}
\end{table}
For each of the unique 9 simulation configurations described above, we
estimate the correlation bounds and vary the correlations along a
sequence of 100 points evenly placed within the bounds, aiming to
explore \(\Omega\). Specifically, we set correlations
\(\{ \rho_1 = ( \hat{l} + \epsilon), \rho_2 = (\hat{l} + \epsilon) + \delta, \ldots, \rho_{100} = (\hat{u} - \epsilon) \}\),
with \(\hat{l}\) and \(\hat{u}\) being the estimated lower and upper
bounds, respectively, with increment value \(\delta\). The adjustment
factor, \(\epsilon=0.01\), is introduced to handle numeric issues when
the bound is specified exactly.
\begin{figure}
\centering
\includegraphics{ch040-biNormPlot-1.pdf}
\caption{\label{fig:ch040-biNormPlot}\texttt{bigsimr} recovers the
Pearson specified correlations for MVN.}
\end{figure}
\begin{figure}
\centering
\includegraphics{ch040-biGammaPlot-1.pdf}
\caption{\label{fig:ch040-biGammaPlot}\texttt{bigsimr} recovers the
Pearson specified correlations for Bivariate Gamma.}
\end{figure}
\begin{figure}
\centering
\includegraphics{ch040-biNegBinPlot-1.pdf}
\caption{\label{fig:ch040-biNegBinPlot}\texttt{bigsimr} recovers the
correlations for bivariate negative binomial only approximately for
Pearson but (nearly) exactly for the rank-based correlations.}
\end{figure}
Figure \ref{fig:ch040-bPlot} displays the aggregated bivariate
simulation results. Table \ref{tab:ch040-BiError} contains the mean
absolute error (MAE) in reproducing the desired dependency measures for
the three bivariate scenarios. Overall, the studies show that our
methodology is generally accurate across the entire range of possible
correlation for all three dependency measures. Our Pearson matching
performs nearly as well as Spearman or Kendall, except for a slight
increase in error for the negative binomial case.
\begin{table}
\caption{\label{tab:ch040-BiError}Average abolute error in matching the target dependency across the entire range of possible correlations for each bivariate marginal.}
\centering
\begin{tabular}[t]{lllr}
\toprule
No. of random vectors & Correlation type & Distribution & Mean abs. error\\
\midrule
1000 & Pearson & norm & 0.0156184\\
1000 & Pearson & gamma & 0.0159287\\
1000 & Pearson & nbinom & 0.0162465\\
\addlinespace
1000 & Spearman & norm & 0.0175440\\
1000 & Spearman & gamma & 0.0182176\\
1000 & Spearman & nbinom & 0.0151124\\
\addlinespace
1000 & Kendall & norm & 0.0130284\\
1000 & Kendall & gamma & 0.0114740\\
1000 & Kendall & nbinom & 0.0126014\\
\addlinespace
10000 & Pearson & norm & 0.0060226\\
10000 & Pearson & gamma & 0.0057667\\
10000 & Pearson & nbinom & 0.0058180\\
\addlinespace
10000 & Spearman & norm & 0.0062388\\
10000 & Spearman & gamma & 0.0056132\\
10000 & Spearman & nbinom & 0.0049256\\
\addlinespace
10000 & Kendall & norm & 0.0032618\\
10000 & Kendall & gamma & 0.0038067\\
10000 & Kendall & nbinom & 0.0033203\\
\addlinespace
1e+05 & Pearson & norm & 0.0017570\\
1e+05 & Pearson & gamma & 0.0016962\\
1e+05 & Pearson & nbinom & 0.0029582\\
\addlinespace
1e+05 & Spearman & norm & 0.0016607\\
1e+05 & Spearman & gamma & 0.0016408\\
1e+05 & Spearman & nbinom & 0.0015269\\
\addlinespace
1e+05 & Kendall & norm & 0.0010441\\
1e+05 & Kendall & gamma & 0.0011077\\
1e+05 & Kendall & nbinom & 0.0011976\\
\bottomrule
\end{tabular}
\end{table}
\begin{figure}
\centering
\includegraphics{ch040-bPlot-1.pdf}
\caption{\label{fig:ch040-bPlot}Bivariate simulations match target
correlations across the entire range of feasible correlations. The
horizontal axis plots the specified target correlations for each
bivariate margin. Normal margins are plotted in dark dark grey, gamma in
medium grey, and negative binomial in light grey. As the number of
simulated vectors \(B\) increases from left to right, the variation in
estimated correlations (vertical axis) decreases. The dashed line
indicates equality between the specified and estimated correlations.}
\end{figure}
\hypertarget{scale-up-to-high-dimensions}{%
\subsection{Scale up to High
Dimensions}\label{scale-up-to-high-dimensions}}
With information of our method's accuracy from a low-dimensional
perspective, we now assess whether \texttt{bigsimr} can scale to larger
dimensional problems with practical computation times. We ultimately
generate \(B=1,000\) random vectors for
\(d=\{100, 250, 500, 1000, 2500, 5000, 10000\}\) for each correlation
type, \{Pearson, Spearman, Kendall\} while timing the algorithm's major
steps. We begin by producing a synthetic ``data set'' by completing the
following steps:
\begin{enumerate}
\def\arabic{enumi}.{\arabic{enumi}.}
\item
Produce heterogeneous gamma marginals by randomly selecting the
\(j^{th}\) gamma shape parameter from
\(U_j \sim uniform(1,10), j=1,\ldots,d\) and the \(j^{th}\) rate
parameter from \(V_j \sim exp(1/5), j=1,\ldots,d\), with the constant
parameters determined arbitrarily.
\item
Produce a random full-rank Pearson correlation matrix via
\texttt{cor\_randPD} of size \(d \times d\).
\item
Simulate a ``data set'' of \(1,000 \times d\) random vectors via
\texttt{rvec}.
\end{enumerate}
With the synthetic data set in hand, we complete and time the following
four steps involved in a typical workflow (including correlation
estimation).
\begin{enumerate}
\def\arabic{enumi}.{\arabic{enumi}.}
\item
Estimate the correlation matrix from the ``data'' in the \emph{Compute
Correlation} step.
\item
Map the correlations to initialize the algorithm (Pearson to Pearson,
Spearman to Pearson, or Kendall to Pearson) in the \emph{Adjust
Correlation} step.
\item
Check whether the mapping produces a valid correlation matrix and, if
not, find the nearest PD correlation matrix in the \emph{Check
Admissibility} step.
\item
Simulate \(1,000\) vectors in the \emph{Simulate Data} step.
\end{enumerate}
The experiments are conducted on a MacBook Pro carrying a 2.4 GHz 8-Core
Intel Core i9 processor, with all 16 threads employed during
computation. Table \ref{tab:ch040-moderateDtab} displays the total
computation time for moderate dimesions (\(d \leq 500\)). In every
simullation setting, the 1,000 vectors are generated rapidly, executing
in under 2 seconds.
\begin{table}
\caption{\label{tab:ch040-moderateDtab}Total time to produce 1,000 random vectors with a random correlation matrix and hetereogeneous gamma margins.}
\centering
\begin{tabular}[t]{llr}
\toprule
Dimension & Correlation type & Total Time (Seconds)\\
\midrule
100 & Pearson & 0.081\\
100 & Spearman & 0.027\\
100 & Kendall & 0.075\\
\addlinespace
250 & Pearson & 0.378\\
250 & Spearman & 0.062\\
250 & Kendall & 0.375\\
\addlinespace
500 & Pearson & 1.425\\
500 & Spearman & 0.135\\
500 & Kendall & 1.495\\
\bottomrule
\end{tabular}
\end{table}
The results for \(d > 500\) show scalability to ultra-high dimensions
for all three correlation types, although the total times do become much
larger. Figure \ref{fig:ch040-largeDfig} displays computation times for
\(d=\{1000, 2500, 5000, 10000\}\). For \(d\) equal to 1000 and 2500, the
total time is under a couple of minutes. At \(d\) of 5000 and 10,000,
Pearson correlation matching in the \emph{Adjust Correlation} step
becomes costly. Interestingly, Pearson is actually faster than Kendall
for \(d=10,000\) due to bottlenecks in \emph{Compute Correlation} and
\emph{Check Admissibility}. Uniformly, matching Spearman correlations is
faster, with total times under 5 minutes for \(d=10,000\), making
Spearman the most computationally-friendly dependency type for our
package. With this in mind, we scaled the simulation to \(d=20,000\) for
the Spearman type and obtained the 1,000 vectors in under an hour (data
not shown). In principle, this would enable the simulation of an entire
human-derived RNA-seq data set. We note that for a given target
correlation matrix and margins, steps 1, 2, and 3 only need to be
computed once and the fourth step, \emph{Simulate Data}, is nearly
instaneous for all settings considered.
\begin{figure}
\centering
\includegraphics{ch040-largeDfig-1.pdf}
\caption{\label{fig:ch040-largeDfig}Computation times for HD
multivariate gamma simulation.}
\end{figure}
\emph{Limitations, conclusions, and recommendations}
In the bivariate studies, we chose arbitrary simulation parameters for
three distributions, moving from the Gaussian to discrete and non-normal
MVNB. Under these conditions, the simulated random vectors sample the
desired bivariate distribution across the entire range of pairwise
correlations for the three dependency measures. The simulation results
could differ for other choices of simulation settings. Specifying
extreme correlations near the boundary or Frechet bounds could result in
poor simulation performance. Fortunately, it is straightforward to
evaluate simulation performance by using strategies similar to those
completed above. We expect our random vector generation to perform well
for the vast majority of NORTA-feasible correlation matrices, but advise
to check the performance before making inferences/further analyses.
Somewhat surprisingly, Kendall estimation and nearest PD computation
scale poorly compared to Spearman and, even, approximate Pearson
matching. In our experience, Kendall computation times are sensitive to
the number of cores, benefiting from multi-core parallelization. This
could mitigate some of the current algorithmic/implementation
shortcomings. Despite this, Kendall matching is still feasible for most
HD data sets. Finally, we note that one could use our single-pass
algorithm \texttt{cor\_fastPD} to produce a `close' (not nearest PD) to
scale to even higher dimensions with some loss of accuracy.
\hypertarget{examples}{%
\section{RNA-seq data application}\label{examples}}
This section demonstrates how to simulate multivariate data using
\texttt{bigsimr}, aiming to replicate the structure of HD dependent
count data. In an illustration of our proposed methodology, we seek to
simulate RNA-sequencing data by producing simulated random vectors
mimicking the observed data and its generating process. Modeling RNA-seq
using multivariate probability distributions is natural as inter-gene
correlation occurs during biological processes (Wang, Gerstein, and
Snyder 2009). And yet, many models do not account for this, leading to
major disruptions to the operating characteristics of statistical
estimation, testing, and prediction. The following subsections apply
\texttt{bigsimr}'s methods to real RNA-seq data, including replicating
an estimated parametric structure, probability estimation, and
evaluation of correlation estimation efficiency.
\hypertarget{simulating-high-dimensional-rna-seq-data}{%
\subsection{Simulating High-Dimensional RNA-seq
data}\label{simulating-high-dimensional-rna-seq-data}}
\begin{Shaded}
\begin{Highlighting}[]
\NormalTok{d }\OtherTok{\textless{}{-}} \DecValTok{1000}
\NormalTok{brca1000 }\OtherTok{\textless{}{-}}\NormalTok{ example\_brca }\SpecialCharTok{\%\textgreater{}\%}
\FunctionTok{select}\NormalTok{(}\FunctionTok{all\_of}\NormalTok{(}\DecValTok{1}\SpecialCharTok{:}\NormalTok{d)) }\SpecialCharTok{\%\textgreater{}\%}
\FunctionTok{mutate}\NormalTok{(}\FunctionTok{across}\NormalTok{(}\FunctionTok{everything}\NormalTok{(), as.double))}
\end{Highlighting}
\end{Shaded}
We begin by estimating the structure of the TCGA BRCA RNA-seq data set.
Ultimately, we will simulate \(B=10,000\) random vectors
\({\bf Y}=(Y_1, \ldots, Y_d)^\top\) with \(d=1000\). We assume a MVNB
model as RNA-seq counts are often over-dispersed and correlated. Since
all \(d\) selected genes exhibit over-dispersion (data not shown), we
proceed to estimate the NB parameters \((r_i, p_i), i=1,\ldots,d\), to
determine the target marginal PMFs \(f_i\). To complete specification of
the simulation algorithm inputs, we estimate the Spearman correlation
matrix \(R_S\) to characterize dependency. With this goal in mind, we
first estimate the desired correlation matrix using the fast
implementation provided by \texttt{bigsimr}:
\begin{Shaded}
\begin{Highlighting}[]
\CommentTok{\# Estimate Spearman\textquotesingle{}s correlation on the count data}
\NormalTok{R\_S }\OtherTok{\textless{}{-}}\NormalTok{ bs}\SpecialCharTok{$}\FunctionTok{cor}\NormalTok{(}\FunctionTok{as.matrix}\NormalTok{(brca1000), bs}\SpecialCharTok{$}\NormalTok{Spearman)}
\end{Highlighting}
\end{Shaded}
Next, we estimate the marginal parameters. We use the method of moments
to estimate the marginal parameters for the multivariate negative
binomial model. The marginal distributions are from the same probability
family (NB), yet they are heterogeneous in terms of the parameters
probability and size \((p_i, n_i)\) for \(i,\ldots,d\). The functions
below support this estimation for later use in \texttt{rvec}.
\begin{Shaded}
\begin{Highlighting}[]
\NormalTok{make\_nbinom\_margins }\OtherTok{\textless{}{-}} \ControlFlowTok{function}\NormalTok{(sizes, probs) \{}
\NormalTok{ margins }\OtherTok{\textless{}{-}} \FunctionTok{lapply}\NormalTok{(}\DecValTok{1}\SpecialCharTok{:}\FunctionTok{length}\NormalTok{(sizes), }\ControlFlowTok{function}\NormalTok{(i) \{}
\NormalTok{ dist}\SpecialCharTok{$}\FunctionTok{NegativeBinomial}\NormalTok{(sizes[i], probs[i])}
\NormalTok{ \})}
\FunctionTok{do.call}\NormalTok{(c, margins)}
\NormalTok{\}}
\end{Highlighting}
\end{Shaded}
We apply these estimators to the highest-expressing r d genes across the
r nrow(brca1000) patients:
\begin{Shaded}
\begin{Highlighting}[]
\NormalTok{mom\_nbinom }\OtherTok{\textless{}{-}} \ControlFlowTok{function}\NormalTok{(x) \{}
\NormalTok{ m }\OtherTok{\textless{}{-}} \FunctionTok{mean}\NormalTok{(x)}
\NormalTok{ s }\OtherTok{\textless{}{-}} \FunctionTok{sd}\NormalTok{(x)}
\FunctionTok{c}\NormalTok{(}\AttributeTok{size =}\NormalTok{ m}\SpecialCharTok{\^{}}\DecValTok{2} \SpecialCharTok{/}\NormalTok{ (s}\SpecialCharTok{\^{}}\DecValTok{2} \SpecialCharTok{{-}}\NormalTok{ m), }\AttributeTok{prob =}\NormalTok{ m }\SpecialCharTok{/}\NormalTok{ s}\SpecialCharTok{\^{}}\DecValTok{2}\NormalTok{)}
\NormalTok{\}}
\NormalTok{nbinom\_fit }\OtherTok{\textless{}{-}} \FunctionTok{apply}\NormalTok{(brca1000, }\DecValTok{2}\NormalTok{, mom\_nbinom)}
\NormalTok{sizes }\OtherTok{\textless{}{-}}\NormalTok{ nbinom\_fit[}\StringTok{"size"}\NormalTok{,]}
\NormalTok{probs }\OtherTok{\textless{}{-}}\NormalTok{ nbinom\_fit[}\StringTok{"prob"}\NormalTok{,]}
\NormalTok{nb\_margins }\OtherTok{\textless{}{-}} \FunctionTok{make\_nbinom\_margins}\NormalTok{(sizes, probs)}
\end{Highlighting}
\end{Shaded}
Notably, the estimated marginal NB probabilities \(\{ \hat{p}_i \}\) are
small --- ranging in the interval \([r min(probs) , r max(probs)]\).
This gives rise to highly variable counts and, thus, less restriction on
potential pairwise correlation pairs. Given the marginals, we now
specify targets and check admissibility of the specified correlation
matrix.
\begin{Shaded}
\begin{Highlighting}[]
\CommentTok{\# 1. Mapping step first}
\NormalTok{R\_X }\OtherTok{\textless{}{-}}\NormalTok{ bs}\SpecialCharTok{$}\FunctionTok{cor\_convert}\NormalTok{(R\_S, bs}\SpecialCharTok{$}\NormalTok{Spearman, bs}\SpecialCharTok{$}\NormalTok{Pearson)}
\CommentTok{\# 2a. Check admissibility}
\NormalTok{(is\_valid\_corr }\OtherTok{\textless{}{-}}\NormalTok{ bs}\SpecialCharTok{$}\FunctionTok{iscorrelation}\NormalTok{(R\_X))}
\end{Highlighting}
\end{Shaded}
\begin{verbatim}
## [1] FALSE
\end{verbatim}
\begin{Shaded}
\begin{Highlighting}[]
\CommentTok{\# 2b. compute nearest correlation}
\ControlFlowTok{if}\NormalTok{ (}\SpecialCharTok{!}\NormalTok{is\_valid\_corr) \{}
\NormalTok{ R\_X\_pd }\OtherTok{\textless{}{-}}\NormalTok{ bs}\SpecialCharTok{$}\FunctionTok{cor\_nearPD}\NormalTok{(R\_X)}
\DocumentationTok{\#\# Quantify the error}
\NormalTok{ targets }\OtherTok{\textless{}{-}}\NormalTok{ R\_X[}\FunctionTok{lower.tri}\NormalTok{(R\_X, }\AttributeTok{diag =} \ConstantTok{FALSE}\NormalTok{)]}
\NormalTok{ approximates }\OtherTok{\textless{}{-}}\NormalTok{ R\_X\_pd[}\FunctionTok{lower.tri}\NormalTok{(R\_X\_pd, }\AttributeTok{diag =} \ConstantTok{FALSE}\NormalTok{)]}
\NormalTok{ R\_X }\OtherTok{\textless{}{-}}\NormalTok{ R\_X\_pd}
\NormalTok{\}}
\FunctionTok{summary}\NormalTok{(}\FunctionTok{abs}\NormalTok{(targets }\SpecialCharTok{{-}}\NormalTok{ approximates))}
\end{Highlighting}
\end{Shaded}
\begin{verbatim}
## Min. 1st Qu. Median Mean 3rd Qu. Max.
## 0.0000000 0.0001929 0.0004129 0.0005064 0.0007168 0.0069224
\end{verbatim}
As seen above, \(R_X\) is not strictly admissible in our scheme. This
can be seen from the negative result from bs\$iscorrelation(R\_X), which
checks if \(R_X\) is positive-definite with 1's along the diagonal.
However, the approximation is close with a maximum absolute error of r
max( abs(targets - approximates)) and average absolute error of r
mean(abs(targets - approximates)) across the r choose(d,2) correlations.
With the inputs configured, \texttt{rvec} is executed to produce a
synthetic RNA-seq data set:
\begin{Shaded}
\begin{Highlighting}[]
\NormalTok{sim\_nbinom }\OtherTok{\textless{}{-}}\NormalTok{ bs}\SpecialCharTok{$}\FunctionTok{rvec}\NormalTok{(}\DecValTok{10000}\NormalTok{, R\_X, nb\_margins) }
\end{Highlighting}
\end{Shaded}
Figure \ref{fig:ch050-simDataFig} displays the simulated counts and
pairwise relationships for our example genes from Table
\ref{tab:ch010-realDataTab}. Simulated counts roughly mimic the observed
data but with a smoother appearance due to the assumed parametric form
and with less extreme points than the observed data in Figure
\ref{fig:ch010-realDataFig}. Figure \ref{fig:ch050-figBRCA} compares the
specified target parameter (horizontal axis) with the corresponding
quantities estimated from the simulated data (vertical axis). The
evaluation shows that the simulated counts approximately match the
target parameters and exhibit the full range of estimated correlations
from the data.
\begin{figure}
\centering
\includegraphics{ch050-simDataFig-1.pdf}
\caption{\label{fig:ch050-simDataFig}Simulated data for three selected
high-expressing genes generally replicates the estimated data structure.
The data do not exhibit outlying points, but do possess the desired
Spearman correlations, central tendencies, and discrete values.}
\end{figure}
\begin{figure}
\includegraphics[width=0.8\linewidth]{ch050-figBRCA} \caption{\label{fig:ch050-figBRCA}Simulated random vectors from a multivariate negative binomial replicate the estimated structure from an RNA-seq data set. The dashed red lines indicate equality between estimated parameters from simulated data (vertical axes) and the specified target parameters (horizontal axes).}\label{fig:ch050-figBRCA}
\end{figure}
\emph{Limitations, conclusions, and recommendations}
The results show overall aquedate simulation performance for our choice
of parameters settings. Our settings were motivated by modeling
high-expressing genes from the TCGA BRCA data set. In general, the
ability to match marginal and dependence parameters depends on the
particular joint probability model. We recommend to evaluate and tune
your simulation until you can be assured of the accuracy.
\hypertarget{discussion}{%
\section{Conclusion and discussion}\label{discussion}}
We developed a general-purpose high-dimensional multivariate simulation
algorithm and provide a user-friendly, high-performance \texttt{R}
package \texttt{bigsimr}, Julia package \texttt{Bigsimr} and Python
package \texttt{bigsimr}. The random vector generation method is
inspired by NORTA (Cario and Nelson 1997) and Gaussian copula-based
approaches Xiao (2017). The major contributions of this work are methods
and software for flexible, scalable simulation of HD multivariate
probability distributions with broad data analytic applications for
modern, big-data statistical computing. For example, one could simulate
high-resolution time series data, such as those consistent with an
auto-regressive moving average model exhibiting a specified Spearman
structure. Our methods could also be used to simulate sparsely
correlated data, as many HD methods assume, via specifying a
\emph{spiked correlation matrix}.
There are limitations to the methodology and implementation. We could
only investigate selected multivariate distributions in our Monte Carlo
studies. There may be instances where the methods do not perform well.
Along those lines, we expect that correlation values close to the
boundary of the feasible region could result in algorithm failure.
Another issue is that for discrete distributions, we use continuous
approximations when the support set is large. This could limit the
scalability/accuracy for particular discrete/mixed multivariate
distributions. Our method would also benefit computationally from faster
Kendall estimation and more finely tuned nearest PD calculation for
non-admissible Kendall matrices.
\hypertarget{refs}{}
\begin{CSLReferences}{1}{0}
\leavevmode\vadjust pre{\hypertarget{ref-BF17}{}}%
Barbiero, Alessandro, and Pier Alda Ferrari. 2017. {``{An R package for
the simulation of correlated discrete variables}.''}
\emph{Communications in Statistics - Simulation and Computation} 46 (7):
5123--40. \url{https://doi.org/10.1080/03610918.2016.1146758}.
\leavevmode\vadjust pre{\hypertarget{ref-Cario1997}{}}%
Cario, Marne C., and Barry L. Nelson. 1997. {``{Modeling and generating
random vectors with arbitrary marginal distributions and correlation
matrix}.''} Technical Report.
\leavevmode\vadjust pre{\hypertarget{ref-Chambers1983}{}}%
Chambers, J. M., W. S. Cleveland, B. Kleiner, and P. A. Tukey. 1983.
\emph{{Graphical Methods for Data Analysis}}. Belmont, CA: Wadsworth
{\&} Brooks.
\leavevmode\vadjust pre{\hypertarget{ref-Chen2001}{}}%
Chen, Huifen. 2001. {``{Initialization for NORTA: Generation of random
vectors with specified marginals and correlations}.''} \emph{INFORMS
Journal on Computing} 13 (4): 312--31.
\leavevmode\vadjust pre{\hypertarget{ref-Chernick2008}{}}%
Chernick, Michael R. 2008. \emph{{Bootstrap Methods: A Guide for
Practitioners and Researchers}}. 2nd ed. Hoboken, NJ.
\leavevmode\vadjust pre{\hypertarget{ref-Conesa2016b}{}}%
Conesa, Ana, Pedro Madrigal, Sonia Tarazona, David Gomez-Cabrero,
Alejandra Cervera, Andrew McPherson, Michal Wojciech Szcześniak, et al.
2016. {``{A survey of best practices for RNA-seq data analysis}.''}
\url{https://doi.org/10.1186/s13059-016-0881-8}.
\leavevmode\vadjust pre{\hypertarget{ref-DH2011}{}}%
Demirtas, Hakan, and Donald Hedeker. 2011. {``{A practical way for
computing approximate lower and upper correlation bounds}.''}
\emph{American Statistician} 65 (2): 104--9.
\url{https://doi.org/10.1198/tast.2011.10090}.
\leavevmode\vadjust pre{\hypertarget{ref-BE07}{}}%
Efron, Bradley. 2007. {``{Correlation and large-scale simultaneous
significance testing}.''} \emph{Journal of the American Statistical
Association} 102 (477): 93--103.
\url{https://doi.org/10.1198/016214506000001211}.
\leavevmode\vadjust pre{\hypertarget{ref-Fasiolo2016}{}}%
Fasiolo, Matteo. 2016. \emph{{An introduction to mvnfast. R package
version 0.1.6.}} \url{https://cran.r-project.org/package=mvnfast}.
\leavevmode\vadjust pre{\hypertarget{ref-GH02}{}}%
Ghosh, Soumyadip, and Shane G Henderson. 2002. {``{Properties of the
Norta method in higher dimensions}.''} \emph{Proceedings of the 2002
Winter Simulation Conference}, 263--69.
\leavevmode\vadjust pre{\hypertarget{ref-higham2002computing}{}}%
Higham, Nicholas J. 2002. {``Computing the Nearest Correlation
Matrix---a Problem from Finance.''} \emph{IMA Journal of Numerical
Analysis} 22 (3): 329--43.
\leavevmode\vadjust pre{\hypertarget{ref-K58}{}}%
Kruskal, William H. 1958. {``{Ordinal measures of association}.''}
\emph{Journal of the American Statistical Association} 53 (284):
814--61. \url{https://doi.org/10.1080/01621459.1958.10501481}.
\leavevmode\vadjust pre{\hypertarget{ref-LH75}{}}%
Li, Shing Ted, and Joseph L. Hammond. 1975. {``{Generation of
pseudorandom numbers with specified univariate distributions and
correlation coefficients}.''} \emph{IEEE Transactions on Systems, Man
and Cybernetics} 5: 557--61.
\url{https://doi.org/10.1109/TSMC.1975.5408380}.
\leavevmode\vadjust pre{\hypertarget{ref-Li2019gpu}{}}%
Li, Xiang, A. Grant Schissler, Rui Wu, Lee Barford, Jr. Harris, Fredrick
C., and Frederick C. Harris. 2019. {``{A Graphical Processing Unit
accelerated NORmal to Anything algorithm for high dimensional
multivariate simulation}.''} \emph{Advances in Intelligent Systems and
Computing}, 339--45. \url{https://doi.org/10.1007/978-3-030-14070-0_46}.
\leavevmode\vadjust pre{\hypertarget{ref-MB13}{}}%
Madsen, L., and D. Birkes. 2013. {``{Simulating dependent discrete
data}.''} \emph{Journal of Statistical Computation and Simulation} 83
(4): 677--91. \url{https://doi.org/10.1080/00949655.2011.632774}.
\leavevmode\vadjust pre{\hypertarget{ref-MK01}{}}%
Mari, Dominique Drouet, and Samuel Kotz. 2001. \emph{{Correlation and
dependence}}. World Scientific.
\leavevmode\vadjust pre{\hypertarget{ref-muino2006gs}{}}%
Muino, JM, Eberhard O Voit, and Albert Sorribas. 2006.
{``GS-Distributions: A New Family of Distributions for Continuous
Unimodal Variables.''} \emph{Computational Statistics and Data Analysis}
50 (10): 2769--98.
\leavevmode\vadjust pre{\hypertarget{ref-Nelsen2007}{}}%
Nelsen, Roger B. 2007. \emph{{An Introduction to Copulas}}. 2nd ed. New
York: Springer Science {\&} Business Media.
\leavevmode\vadjust pre{\hypertarget{ref-Nik13a}{}}%
Nikoloulopoulos, Aristidis K. 2013. {``{Copula-based models for
multivariate discrete response data}.''} In \emph{Lecture Notes in
Statistics: Copulae in Mathematical and Quantitative Finance}, 213th
ed., 231--49. Heidelberg: Springer.
\leavevmode\vadjust pre{\hypertarget{ref-Park1996}{}}%
Park, Chul Gyu, Taesung Park, and Dong Wan Shin. 1996. {``{A simple
method for generating correlated binary variates}.''} \emph{American
Statistician} 50 (4): 306--10.
\url{https://doi.org/10.1080/00031305.1996.10473557}.
\leavevmode\vadjust pre{\hypertarget{ref-qi2006quadratically}{}}%
Qi, Houduo, and Defeng Sun. 2006. {``A Quadratically Convergent Newton
Method for Computing the Nearest Correlation Matrix.''} \emph{SIAM
Journal on Matrix Analysis and Applications} 28 (2): 360--85.
\leavevmode\vadjust pre{\hypertarget{ref-Schissler2019}{}}%
Schissler, Alfred Grant, Dillon Aberasturi, Colleen Kenost, and Yves A.
Lussier. 2019. {``{A Single-Subject Method to Detect Pathways Enriched
With Alternatively Spliced Genes}.''} \emph{Frontiers in Genetics} 10
(414). \url{https://doi.org/10.3389/fgene.2019.00414}.
\leavevmode\vadjust pre{\hypertarget{ref-Schissler2018}{}}%
Schissler, Alfred Grant, Walter W Piegorsch, and Yves A Lussier. 2018.
{``{Testing for differentially expressed genetic pathways with
single-subject N-of-1 data in the presence of inter-gene
correlation}.''} \emph{Statistical Methods in Medical Research} 27 (12):
3797--3813. \url{https://doi.org/10.1177/0962280217712271}.
\leavevmode\vadjust pre{\hypertarget{ref-Ubeda-Flores2017}{}}%
Úbeda-Flores, Manuel, and Juan Fernández-Sánchez. 2017. {``{Sklar's
theorem: The cornerstone of the Theory of Copulas}.''} In \emph{Copulas
and Dependence Models with Applications}.
\url{https://doi.org/10.1007/978-3-319-64221-5_15}.
\leavevmode\vadjust pre{\hypertarget{ref-Wang2009b}{}}%
Wang, Zhong, Mark Gerstein, and Michael Snyder. 2009. {``{RNA-Seq: A
revolutionary tool for transcriptomics}.''}
\url{https://doi.org/10.1038/nrg2484}.
\leavevmode\vadjust pre{\hypertarget{ref-Xia17}{}}%
Xiao, Qing. 2017. {``{Generating correlated random vector involving
discrete variables}.''} \emph{Communications in Statistics - Theory and
Methods} 46 (4): 1594--1605.
\url{https://doi.org/10.1080/03610926.2015.1024860}.
\leavevmode\vadjust pre{\hypertarget{ref-XZ19}{}}%
Xiao, Qing, and Shaowu Zhou. 2019. {``{Matching a correlation
coefficient by a Gaussian copula}.''} \emph{Communications in Statistics
- Theory and Methods} 48 (7): 1728--47.
\url{https://doi.org/10.1080/03610926.2018.1439962}.
\leavevmode\vadjust pre{\hypertarget{ref-Yan2007}{}}%
Yan, Jun. 2007. {``{Enjoy the joy of copulas: with a package copula}.''}
\emph{Journal of Statistical Software} 21 (4): 1--21.
\url{http://www.jstatsoft.org/v21/i04}.
\end{CSLReferences}
\bibliographystyle{unsrt}
|
\section{Poly-logarithmic Excess Risk and Confidence Band Recovery}
\label{sec:logT}
In this section we show how to achieve excess risk scaling poly-logarithmically in the number of iterations. While this is worse than the final bound we will show in Section~\ref{sec:TloglogT}, it will introduce many of the important steps in the final analysis and also yield a warm start for our estimate of the trajectory which we will subsequently refine in Section~\ref{sec:TloglogT} to get our final bound. Crucially, we show this algorithm can output a \emph{confidence band} which with high probability (over the entire data generating process) contains the true trajectory.
The main result of this section is the following:
\begin{theorem}\label{thm:main_TlogT}
For any $\eta \le 0.49$, there is a polynomial-time algorithm that, given the corrupted observations $\brc{\wt{y}_i}$, with probability $1 - \delta$ over the randomness of the input, outputs a trajectory $\brc{\wh{x}_i}$ and steps $\brc{\wh{w}_i}$ for which $\wh{x}_i = A\wh{x}_{i - 1} +\wh{w}_i$ for every $i\in [T]$, and for which,
\begin{multline}
\frac{1}{T}\left(\sum^{T-1}_{i=0}\left(a^*_i \norm{B\wh{x}_i - y_i}^2/\tau^2 + \norm{\wh{w}_i}^2/\sigma^2\right) + \|\wh x_0\|^2/R^2\right) - \mathsf{OPT} \\
\lesssim
\tau^{-2} \eta \cdot \brk*{E_{\mathsf{noise}} + \rho^2\left(\alpha +\norm{B}^2\sqrt{\log(dT/t\delta)/t}\right)\cdot \left(\frac{\rho^6 E_{\mathsf{noise}} t}{\kappa} + \frac{R^2}{T/t}(d + \log(1/\delta))\right)},
\end{multline}
where
\begin{align}
E_{\mathsf{noise}} &\triangleq \tau^2\left(m + \log(T/\delta)\right) + t\rho^2\norm{B}^2\sigma^2\left(d + \log(T/\delta)\right) \label{eq:Enoise_def} \\
t &\triangleq \Max{s}{\wt{\Theta}(\kappa^{-2}\rho^{12}\norm{B}^4\log(dT/\delta)))}. \label{eq:tdef}
\end{align}
\end{theorem}
\subsection{Sum-of-Squares Relaxation}
We now formulate the sum-of-squares program we work with in this section. We begin by introducing an important parameter, the so-called \emph{window size} $t$. Recall from Assumption~\ref{assume:obs} that we assume that the observability matrix $\calO_s$ is well-conditioned. We will take $t$ to be a sufficiently large multiple of $s$ such that, roughly speaking, the contribution to the observability matrix $\calO_t$ from the uncorrupted time steps is also well-conditioned. We defer the tuning of $t$ to later in the proof of Theorem~\ref{thm:main_TlogT}.
For convenience, given $0\le i < T$, let $\ell(i)\triangleq \floor{i/t}$ denote the index of the window to which iterate $i$ belongs.
At this point, we can define our sum-of-squares relaxation:
\begin{program}\label{program:sos}
Let $\brc{y_i}$ be the observations we are given, and let window size $t\in\mathbb{N}$ be a parameter to be tuned later. The program variables are $d$-dimensional vector-valued variables $\brc{x_i}$ (trajectory estimates) and $\brc{w_i}$ (process noise estimates), $m$-dimensional vector-valued variables $\brc{v_i}$ (observation noise estimates), and Boolean variables $\brc{a_i}$ (indicators for uncorrupted time steps), and the constraints are that for all $0 \le i < T$,
\vspace{0.05cm}
\noindent\emph{Boolean indicators for uncorrupted steps}
\begin{enumerate}
\item $a_i^2 = a_i$ \label{item:boolean}
\end{enumerate}
\noindent\emph{Trajectory estimate follow linear dynamics and fit $y_i$'s on uncorrupted steps}
\begin{enumerate}
\setcounter{enumi}{1}
\item $x_i = Ax_{i-1} + w_i$ \label{item:dynamics}
\item $a_i(y_i - Bx_i - v_i) = 0$ \label{item:measurements}
\end{enumerate}
\noindent\emph{Only $\eta$ fraction of timesteps corrupted}
\begin{enumerate}
\setcounter{enumi}{3}
\item $\sum_{i=0}^{T-1} a_i \ge (1-1.01\eta)T$ \label{item:many_inliers}
\end{enumerate}
\noindent\emph{Process and observation noise bounded}
\begin{enumerate}
\setcounter{enumi}{4}
\item $\norm{v_i}^2 \le O(\tau^2(m + \log(T/\delta))$\label{item:measurements_noise}
\item $\norm{w_i}^2 \le O(\sigma^2(d + \log(T/\delta))$\label{item:dynamics_noise}
\end{enumerate}
\noindent\emph{Random corruptions subsample observability matrix in each window}
\begin{enumerate}
\setcounter{enumi}{6}
\item $\sum^{t-1}_{j=0} (1 - a_{\ell t + j}) (A^j)^{\top} B^{\top} B A^j \preceq \eta\cdot \calO_{\t} + O\left(\rho^2\norm{B}^2\sqrt{\t\log(dT/\t\delta)}\right)\cdot \Id$ for all $0\le \ell < T/\t$ \label{item:subsample}
\end{enumerate}
\noindent\emph{Initial state bounded}
\begin{enumerate}
\setcounter{enumi}{7}
\item $\norm{x_0}^2 \le R^2(d + O(\log(1/\delta)))$ \label{constraint:bound}
\end{enumerate}
The program objective is to minimize
\begin{equation}
\min \frac{1}{T} \psE*{ \sum^{T-1}_{i=0} \left(a_i \norm{Bx_i - y_i}^2/\tau^2 + \norm{w_i}^2/\sigma^2\right) + \|x_0\|^2/R^2}\label{eq:obj}
\end{equation}
over degree-$4$ pseudoexpectations satisfying the above constraints.
\end{program}
\begin{remark}[Uncorrupted Case: Equivalence to Kalman Smoother]\label{rmk:uncorrupted-case}
Suppose that we know there are no corruptions: then we can set $\eta = 0$ in the above program and therefore eliminate the variables $a_i$ (they are all equal to $1$). Then, by a well-known folklore argument, the SoS program is equivalent to the corresponding convex program with actual variables $x_i \in \mathbb{R}^d, v_i \in \mathbb{R}^m$, etc. with the same set of constraints. (This is because, by SoS Cauchy Schwarz, replacing the pseudoexpectation $\psE*{\cdot}$ with the delta distribution over $\psE*{x}$ gives a valid pseudoexpectation with equal or better objective value.) Then the objective is the same as the MAP objective, and as argued below the constraints are satisfied with high probability by the unconstrained MAP solution (Kalman smoother), so our algorithm simply outputs the MAP.
\end{remark}
\subsection{Feasibility of Oracle Kalman Smoother}
In the following section, we show that the output of the oracle Kalman filter, i.e. the algorithm which knows precisely which time steps have been corrupted and runs the offline Kalman filter (Kalman smoother) on the uncorrupted steps to optimally estimate the trajectory, satisfies the constraints of the Program with high probability. In the proof of Lemma~\ref{lem:feasible}, we show how to do this by reducing to showing that the ground truth $x^*$ is feasible with high probability, which is more straightforward. The key fact which allows us to do this is knowledge that the posterior is a Gaussian centered at the output of the Kalman filter.
\begin{lemma}\label{lem:feasible}
Let $\brc{x_i}$ be the sequence of estimates given by running the Kalman smoother (i.e. offline Kalman filter) on the uncorrupted part of the trajectory, let $a_i = a^*_i$, let $w_i = x_i - A x_{i - 1}$ for all $T$, let $v_i = y_i - B x_i$ when $a^*_i = 1$ and otherwise $v_i = 0$. Let $E[\cdot]$ be the expectation with respect to the delta distribution at this point $(x_i,a_i,v_i,w_i)_{i = 1}^n$. Then $E[\cdot]$ is feasible for Program~\ref{program:sos} with probability at least $1 - \delta$.
\end{lemma}
\begin{proof}
It is immediate that Constraints~\ref{item:boolean}, \ref{item:dynamics}, and \ref{item:measurements} are satisfied.
Constraints~\ref{item:many_inliers} and \ref{item:subsample} only involve $a^*_i$ and we verify them in Lemma~\ref{lem:feasible-gt}. It remains to check Constraints \ref{item:measurements_noise} and \ref{item:dynamics_noise}.
For what follows, suppose $a^*_i$ is fixed.
We claim the following two distributions on $\brc{x^*_i}$ are equal:
\begin{enumerate}
\item Sample a trajectory $\brc{x^*_i}$ from the prior.
\item Sample a trajectory $\brc{x^0_i}$ from the prior, sample observations $y_i$ for times where $a^*_i = 1$ given this trajectory, and sample trajectory $\brc{x^*_i}$ from the resulting posterior on $\brc{x^0_i}$ given $y_i$.
\end{enumerate}
The equivalence of these two follows from the following basic fact: given a pair of random variables $(X,Y)$, it's equivalent to sample $X$ from its marginal law directly, or to first sample $Y$ from its marginal law, and then to sample $X$ conditional on $Y$. In the second case, the observations are the random variable $Y$ and the trajectory is $X$; the fact implies that $Y$ is sampled from its marginal law, which means that the marginal law of $\brc{x^*_i}$ is simply the prior on trajectories. This fact is sometimes called the Nishimori identity.
Recall that the Kalman smoother output is simply the posterior mean $\hat{x}_i = \E{x^*_i \mid \{y_i\}_{i : a^*_i = 1}}$ and that the posterior on trajectories is a multivariate Gaussian distribution. By Lemma~\ref{lem:feasible-gt}, we have that
\begin{align}
\|y_i - B x^*_i\|^2 &\le O(m\tau^2 + \tau^2\log(T/\delta)) \\
\|x^*_i - A x^*_{i - 1}\|^2 &\le O(d\sigma^2 + \sigma^2\log(T/\delta)) \label{eqn:calK}\\
\|x^*_0\|^2 &\le R^2(d + O(\log(1/\delta)))
\end{align}
uniformly over $i$ with probability at least $1 - \delta$, then by the law of total probability we know that for $\mathcal{K}$ the feasible set defined by the constraints above in \eqref{eqn:calK},
\[ \delta \ge \Pr{(x^*,y) \notin \mathcal{K}} = \E{\Pr{(x^*,y) \notin \mathcal{K} | y}} \]
so by Markov's inequality $\Pr{\Pr{(x^*,y) \notin \mathcal{K} | y} > 1/3} \le 3\delta$, i.e. $\Pr{\Pr{(x^*,y) \in \mathcal{K} | y} \ge 2/3} \ge 1 - 3\delta$, which by Lemma~\ref{lem:gaussian-convex} implies that $\Pr{(\E{x^*|y},y) \in \mathcal{K}} \ge 1 - 3\delta$ as well.
Adjusting the value of $\delta$ by constants proves the result.
\end{proof}
\begin{lemma}\label{lem:gaussian-convex}
Suppose that $\mathcal{K}$ is a closed convex set, $Z \sim N(\mu,\Sigma)$ is an arbitrary Gaussian random vector, and $\Pr{Z \in \mathcal K} \ge 0.5$. Then $\mu \in \mathcal{K}$.
\end{lemma}
\begin{proof}
First we show this when $\mathcal{K}$ is an affine halfspace, i.e. $\mathcal{K} = \{ x : \langle a, x \rangle \ge b \}$ for some $a$ and $b$ arbitrary.
The assumption gives that $\langle a, Z \rangle \ge b$ with probability greater than 50\%; since the marginal law of $\langle a, Z \rangle$ is $N(\langle a, \mu \rangle, a^T \Sigma a)$, and the Gaussian is symmetrical about its mean, it must be that $\langle a, \mu \rangle \ge b$ and so $\mu \in \mathcal{K}$. Now the result follows for arbitrary convex sets by writing them as intersections of affine halfspaces, since the above argument shows that $\mu$ will lie in each halfspace (since the probability of lying in each halfspace is at least as large as lying in the intersection), hence in the intersection of the halfspaces.
\end{proof}
\begin{lemma}\label{lem:matrix_conc}
For any $\delta > 0$, \begin{equation}
\norm*{\sum^{\t-1}_{i=0}a^*_i (A^i)^{\top}B^{\top}BA^i - (1 - \eta)\calO_{\t}} \le O(\rho^2\norm{B}^2 \sqrt{ \t \log(d/\delta)}) \label{eq:deviation}
\end{equation} with probability at least $1 - \delta$.
\end{lemma}
\begin{proof}
We apply the Matrix Hoeffding inequality (Lemma~\ref{lem:matrixhoeffding}),
using that $\norm{(A^i)^{\top} B^{\top}BA^i} \le \rho^2\norm{B}^2$ by uniform stability.
\iffalse so
\[ \Var{a^*_i}\norm{(A^i)^{\top} B^{\top}BA^i}^2 \le \eta[\rho^2\norm{B}^2]^2. \]
Therefore Matrix Bernstein gives an upper bound
\[ \norm*{\sum^{\t}_{i=0}a^*_i (A^i)^{\top}B^{\top}BA^i - (1 - \eta)\calO_{\t}} \le O\left(\rho^2\norm{B}^2 \sqrt{\eta \t \log(d/\delta)} + \rho^2\norm{B}^2 \log(d/\delta) \right)\]
with probability at least $1 - \delta$. Using the assumed lower bound on $t$ gives the result.
\fi
\end{proof}
\begin{lemma}\label{lem:feasible-gt}
With probability at least $1 - \delta$, the ground truth $(x^*_i,w^*_i, v^*_i,a^*_i)$ satisfies the constraints of Program~\ref{program:sos}
provided $T = \Omega(\log(2/\delta)/\eta)$.
\end{lemma}
\begin{proof}
Equality constraints~\ref{item:boolean}, \ref{item:dynamics}, and \ref{item:measurements} are satisfied by definition of the process. The remaining inequality constraints follow from a union bound as follows. The bound on Constraint~\ref{item:many_inliers} follows from Bernstein's inequality (see e.g. \cite{vershynin2018high}).
Constraint~\ref{item:measurements_noise} follows by standard Gaussian concentration with probability at least $1 - \delta$.
The same reasoning applies to Constraints~\ref{item:dynamics_noise} and \ref{constraint:bound}. Constraint 7 follow from Lemma~\ref{lem:matrix_conc} applied to every window $0\le \ell < T/t$.
\end{proof}
\subsection{Outer Argument}
In this section we reduce the problem of competing with $\mathsf{OPT}$ to getting good prediction error on the first iterate of every window.
\begin{lemma} \label{lem:outer}
Let $\psE{\cdot}$ be the solution to Program~\ref{program:sos}, assuming it is feasible. Let $\wh{x}_i \triangleq \psE{x_i}$ and $\wh{w}_i \triangleq \psE{w_i}$ for every $0\le i < T$. Provided the event of Lemma~\ref{lem:feasible} holds, then \begin{multline}
\frac{1}{T}\left(\sum^{T-1}_{i=0}\left(a^*_i \norm{B\wh{x}_i - y_i}^2/\tau^2 + \norm{\wh{w}_i}^2/\sigma^2\right) + \|\wh{x}_0\|^2/R^2\right) - \mathsf{OPT} \lesssim \\
\eta \left(E_{\mathsf{noise}} + \rho^2\left(\alpha +\norm{B}^2\sqrt{\log(dT/t\delta)/t}\right)\cdot \frac{1}{T/t}\sum^{T/t-1}_{\ell=0} \psE*{\norm{x_{\ell t} - x^*_{\ell t}}^2}/\tau^2\right). \label{eq:outer}
\end{multline}
where $E_{\mathsf{noise}}$ is defined in \eqref{eq:Enoise_def}.
\end{lemma}
Before proving this, we will need the following helper lemma which we will reuse with minor modifications later in Section~\ref{sec:TloglogT}.
\begin{lemma}\label{lem:kkm}
Let $\psE{\cdot}$ be the solution to Program~\ref{program:sos}, assuming it is feasible. Let $\wh{x}_i \triangleq \psE{x_i}$ and $\wh{w}_i \triangleq \psE{w_i}$ for every $0\le i < T$. Provided the event of Lemma~\ref{lem:feasible} holds, then
\begin{multline}
\frac{1}{T}\left(\sum_{i=0}^{T-1}\left(a^*_i \| B\wh{x}_i - y_i\|^2/\tau^2 + \norm{\wh{w}_i}^2/\sigma^2\right) + \|\hat{x}_0\|^2/R^2\right) - \mathsf{OPT} \le \\
\psE*{\frac{1}{T}\sum_{i=0}^{T-1} (1 - a_i)\| B(x_i - x^*_i)\|^2/\tau^2} + O(\eta\cdot (m + \log(T/\delta))).
\end{multline}
\end{lemma}
\begin{proof}
By Lemma~\ref{lem:psE-CS}, for any $0\le i < T$, $\norm{B\wh{x}_i - y_i}^2 \le \psE{\norm{Bx_i - y_i}^2}$ and $\norm{\wh{w}_i}^2 \le \psE{\norm{w_i}^2}$, so it suffices to prove that the pseudoexpectation of $\sum^{T-1}_{i=0}\left(a^*_i\norm{Bx_i - y_i}^2/\tau^2 + \norm{w_i}^2/\sigma^2\right) + \|x_0\|^2/R^2$ is sufficiently bounded using the constraints of Program~\ref{program:sos}.
First, by splitting up $a^*_i = a^*_ia_i + a^*_i(1-a_i)$, we have
\begin{align}
& \sum_{i=0}^{T-1} \left(a^*_i\| Bx_i - y_i\|^2/\tau^2 + \norm{w_i}^2/\sigma^2\right) \label{eq:main_obj}
\\
&= \sum_{i = 0}^{T-1}\left(\norm{w_i}^2/\sigma^2 + a^*_ia_i\| Bx_i - y_i\|^2/\tau^2 + a^*_i(1 - a_i)\| Bx_i - y_i\|^2/\tau^2\right) \\
&\leq \sum_{i = 0}^{T-1}\left(\norm{w_i}^2/\sigma^2 + a_i \| Bx_i - y_i\|^2/\tau^2 + 2a^*_i(1 - a_i)\left( \| B(x_i - x^*_i)\|^2/\tau^2 + \|v_i\|^2/\tau^2\right)\right) \label{eq:rewrite_obj}
\end{align}
where in the inequality we used the fact that $a^*_i \le 1$ and that for $i$ satisfying $a^*_i = 1$, $\norm{Bx_i - y_i}^2 = \norm{B(x_i - x^*_i) - v_i}^2 \le 2\norm{B(x_i - x^*_i)}^2 + 2\norm{v_i}^2$. Furthermore, note that
\begin{equation}
\sum^{T-1}_{i=0} a^*_i(1 - a_i)\norm{v_i}^2/\tau^2 \lesssim \eta\left(m + \log(T/\delta)\right)T, \label{eq:av}
\end{equation} by Constraints~\ref{item:many_inliers} and \ref{item:measurements_noise}. Putting \eqref{eq:rewrite_obj} and \eqref{eq:av} together allows us to upper bound the pseudo-expectation of $\sum^{T-1}_{i=0}\left(a^*_i\norm{Bx_i - y_i}^2/\tau^2 + \norm{w_i}^2/\sigma^2\right) + \|x_0\|^2/R^2$ by
\begin{equation}
\mathsf{OPT} + \psE*{\frac{1}{T}\sum_{i=0}^{T-1} (1 - a_i)\| B(x_i - x^*_i)\|^2/\tau^2} + O(\eta(m + \log(T/\delta))). \label{eq:firstopt}
\end{equation}
where we used the fact that $\psE{\cdot}$ minimizes the objective \eqref{eq:obj}, the fact that the oracle Kalman filter solution is feasible because the event of Lemma~\ref{lem:feasible} holds, as well as the fact that $a^*_i \le 1$.
\end{proof}
We now proceed with the proof of Lemma~\ref{lem:outer}.
\begin{proof}[Proof of Lemma~\ref{lem:outer}]
Lemma~\ref{lem:kkm} reduces upper bounding the excess risk achieved by $\brc{\wh{x}_i},\brc{\wh{w}_i}$ to bounding the main term $\psE{\frac{1}{T}\sum^{T-1}_{i=0} (1 - a_i) \norm{B(x_i - x^*_i)}^2/\tau^2}$ in \eqref{eq:firstopt}, which we do now. Using Fact~\ref{fact:unroll}, for any $i = \ell t + j$ we can write $B(x_i - x^*_i) = BA^j(x_{\ell t} - x^*_{\ell t}) + \sum^j_{s=0} BA^{j - s}(w_{\ell t + s} - w^*_{\ell t + s})$.
We thus have \begin{align}
\MoveEqLeft \frac{1}{T}\sum_{i=0}^{T-1} (1 - a_i)\| B(x_i - x^*_i)\|^2 \\
&= \frac{1}{T} \sum_{\ell,j} (1 - a_{\ell t + j})\norm*{BA^j(x_{\ell t} - x^*_{\ell t}) + \sum^j_{s=0} BA^{j-s}(w_{\ell t + s} - w^*_{\ell t + s})}^2 \\
&\le \frac{3}{T} \sum_{\ell, j} (1 - a_{\ell t + j}) \left(\norm*{BA^j(x_{\ell t} - x^*_{\ell t})}^2 + \norm*{\sum^j_{s=0}BA^{j-s}w_{\ell t + s}}^2 + \norm*{\sum^j_{s=0}BA^{j-s}w^*_{\ell t +s}}^2\right) \label{eq:subsample_and_noise}
\end{align}
We can control the two noise terms on the right by noting that for any $\ell,j$, \begin{equation}
\norm*{\sum^j_{s = 0} BA^{j-s} w_{\ell t + s}}^2 \le (j + 1)\sum^j_{s=0} \norm{BA^{j-s} w_{\ell t + s}}^2 \lesssim t\rho^2\norm{B}^2\sigma^2(d + \log(T/\delta)),
\end{equation} where in the last step we used Constraint~\ref{item:dynamics_noise}.
Because the true process noise $\brc{w^*_i}$ is part of a feasible solution to Program~\ref{program:sos}, from Constraint~\ref{item:many_inliers} we conclude that
\begin{equation}
\frac{1}{T}\sum_{\ell,j}(1 - a_{\ell t + j}) \left(\norm*{\sum_s BA^{j-s}w_{\ell t + s}}^2 + \norm*{\sum_s BA^{j-s}w^*_{\ell t + s}}^2\right) \lesssim \eta t\rho^2\norm{B}^2\sigma^2(d + \log(T/\delta))
\end{equation}
For the remaining terms in \eqref{eq:subsample_and_noise}, we invoke Constraint~\ref{item:subsample} and the bound on $\norm{\calO_t}$ in \eqref{eq:opnormOt} to get
\begin{equation}
\frac{1}{T}\sum_{\ell,j}(1 - a_{\ell t + j})\norm{BA^j(x_{\ell t} - x^*_{\ell t})}^2 \le 1.01\eta\rho^2\left(\alpha + O\left( \norm{B}^2\sqrt{\log(dT/t\delta)/t}\right)\right) \frac{1}{T/t}\sum^{T/t-1}_{\ell=0} \norm{x_{\ell t} - x^*_{\ell t}}^2
\end{equation}
from which the lemma follows by substituting the two estimates above into \eqref{eq:subsample_and_noise}.
\end{proof}
\subsection{Decay of Unobservable Subspace}
\label{sec:old_decay}
In this section we show how to bound our prediction error on the first iterate of every window. Towards proving this, the main result of this subsection is to show that our error in estimating these first iterates decays exponentially over time provided that a certain matrix concentration event holds in every window.
We begin by describing this event. Let $\Pi$ denote the projection to the \emph{observable subspace}, that is, to the subspace of $v\in\R^d$ for which $v^{\top} \calO_t v \ge \zeta$ for $\zeta \triangleq \frac{\kappa \t}{40000\rho^4}$, where the window size $t$ will be optimized at the end of this section. The matrix concentration that we need to hold in every window is the following:
\begin{lemma}\label{lem:anticonc}
Suppose $t= \wt{\Omega}\left(\kappa^{-2} \rho^{12} \norm{B}^4\log(dT/\delta)\right)$. Then with probability at least $1 - \delta$ over the randomness of $\brc{a^*_i}$, we have that for all windows $0\le \ell < T/t$, there is a degree-2 SoS proof of the psd inequality
\begin{equation}
\sum^{t-1}_{i = 0} a^*_{\ell t + i} a_{\ell t + i} \Pi (A^i)^{\top} B^{\top} B A^i \Pi \succeq \frac{1}{100}\Pi\calO_t \Pi \label{eq:psdlb}
\end{equation}
using the constraints of Program~\ref{program:sos}.
\end{lemma}
\begin{proof}
We will focus on $\ell = 0$ and apply a union bound over $\ell$ at the end. Recall from Lemma~\ref{lem:matrix_conc} that with probability at least $1 - \delta/(2T/t)$ we have \begin{equation}
\sum^{t-1}_{i=0} a^*_i \Pi (A^i)^{\top} B^{\top} B A^i \Pi \succeq (1 - \eta)\cdot \Pi\calO_t\Pi - O\left(\rho^2\norm{B}^2\sqrt{t\log(dT/t\delta)}\right)\cdot \Pi.
\end{equation}
Condition on this event. Write $a^*_i a_i = a^*_i - a^*_i(1 - a_i) \ge a^*_i - (1 - a_i)$, where in the inequality we use Constraint~\ref{item:boolean}. By Constraint~\ref{item:subsample}, we have \begin{equation}
\sum^{t-1}_{i=0} (1 - a_i) \Pi (A^i)^{\top} B^{\top} B A^i \Pi \preceq \eta\cdot \Pi\calO_t \Pi + O\left(\rho^2\norm{B}^2\sqrt{t\log(dT/t\delta)}\right) \cdot \Pi,
\end{equation}
so because $\Pi \preceq \zeta^{-1} \cdot \Pi\calO_t\Pi$ by definition of $\Pi$, we have a degree-2 SoS proof of the inequality
\begin{equation}
\sum^{t-1}_{i = 1} a^*_{\ell t + i} a_{\ell t + i} \Pi (A^i)^{\top} B^{\top} B A^i \Pi \succeq \left(1 - 2\eta - O\left(\zeta^{-1}\rho^2\norm{B}^2\sqrt{t\log(dT/t\delta)}\right)\right) \Pi\calO_t \Pi.
\end{equation}
We would like $t$ to be large enough that the factor on the right-hand side exceeds $1/100$. As $\eta \le 0.49$, it suffices for $\zeta \ge O(\rho^2 \norm{B}^2\sqrt{t\log(dT/t\delta)})$. Recalling that $\zeta = \frac{\kappa \t}{40000\rho^4}$, we need to take $t \ge \wt{\Omega}\left(\kappa^{-2} \rho^{12} \norm{B}^4 \log(dT/\delta)\right)$. The proof follows by union bounding over $0 \le \ell < T/t$.
\end{proof}
We now turn to showing the main result of this section, namely that provided the event of Lemma~\ref{lem:anticonc} holds, our prediction error on the first iterate of every window decays exponentially over time.
\begin{lemma}\label{lem:error_decay}
Let pseudoexpectation $\psE{\cdot}$ be the solution to Program~\ref{program:sos}, assuming it is feasible. Provided the event of Lemma~\ref{lem:anticonc} holds, we have
\[ \psE{\|x_{\ell t} - x^*_{\ell t}\|^2} \le \frac{1}{2}\psE{\|x_{(\ell - 1) t} - x^*_{(\ell - 1) t}\|^2} + O(\rho^6 E_{\mathsf{noise}} t/\kappa), \]
where $E_{\mathsf{noise}}$ is defined in \eqref{eq:Enoise_def}.
\end{lemma}
Before proving Lemma~\ref{lem:error_decay}, we first show how to use it to conclude the proof of Theorem~\ref{thm:main_TlogT}.
\begin{proof}[Proof of Theorem~\ref{thm:main_TlogT}]
Take $t$ as in \eqref{eq:tdef}. The events of Lemma~\ref{lem:feasible} and Lemma~\ref{lem:anticonc} hold with probability at least $1 - 2\delta$. By summing the conclusion of Lemma~\ref{lem:error_decay} over the time windows, we get \begin{equation}
\frac{1}{T/t}\sum^{T/t-1}_{\ell=0} \psE*{\norm{x_{\ell t} - x^*_{\ell t}}^2} \le \frac{1}{T/t}\psE*{\norm{x_0 - x^*_0}^2} + O(\rho^6 E_{\mathsf{noise}} t/\kappa). \label{eq:corgeo}
\end{equation}
Recall that $x^*_0\sim\calN(0,R^2\cdot\Id)$, so by standard concentration, $\norm{x^*_0}^2 \le R^2(d + O(\log(1/\delta)))$ with probability at least $1 - \delta$. $\norm{x^*_0}^2$ is similarly bounded by Constraint~\ref{constraint:bound}. We can thus bound $\psE*{\norm{x_0 - x^*_0}^2}$ by $2R^2(d + O(\log(1/\delta)))$, so plugging this into \eqref{eq:corgeo} and invoking Lemma~\ref{lem:outer}, we conclude the proof of Theorem~\ref{thm:main_TlogT}.
\end{proof}
We now proceed to the proof of Lemma~\ref{lem:error_decay}. The first step is an averaging argument to show that applying $A^t$ to a vector in the unobservable subspace is guaranteed to decrease its norm.
\begin{restatable}{lemma}{averaging}\label{lem:unobservable-decay}
For any vector $x\in\R^d$, $\norm{A^t \Pi^{\perp} x}^2 \le \frac{1}{40000\rho^2}\norm{\Pi^{\perp} x}^2$.
\end{restatable}
\begin{proof}
We have that \begin{align}
\frac{1}{\t/s}\sum^{\t/s-1}_{j=0} \norm{A^{j s}\Pi^{\perp} x}^2 &= \frac{1}{\t/s}\sum^{\t/s-1}_{j=0} x^{\top}\Pi^{\perp} (A^{j s})^{\top} A^{j s} \Pi^{\perp} x \\
&\le \frac{1}{\kappa \t} \sum^{\t/s-1}_{j=0} x^{\top} \Pi^{\perp}(A^{j s})^{\top} \calO_s A^{js}\Pi^{\perp} x \\
&= \frac{1}{\kappa \t} x^{\top}\Pi^{\perp}\calO_t \Pi^{\perp} x \le \frac{1}{40000\rho^4} \norm{\Pi^{\perp} x}^2,
\end{align} where the third step follows by the first part of Fact~\ref{fact:epochs} and the last step follows by the definition of $\Pi^{\perp}$ and $\zeta = \frac{\kappa \t}{40000\rho^4}$. By averaging, there exists there some $0 \le j < t/s$ for which $\norm{A^{js}\Pi^{\perp} x}^2 \le \frac{1}{40000\rho^4}\norm{\Pi^{\perp}x}^2$. The lemma follows by uniform stability.
\end{proof}
We will eventually take $x$ to be the difference between our estimate of an iterate at the beginning of a window and the ground truth. Informally, this will tell us that over the course of a window of size $t$, the component of the error that started in the unobservable subspace has decayed.
What about the component of the error that started in the \emph{observable} subspace? By uniform stability, it cannot increase by too much, but unlike the unobservable component, it need not decay. This brings us to the win-win argument at the core of the proof of Lemma~\ref{lem:observable-case-analysis}: when the observable component does not decay, we can still relate it to the observable component in the previous time window just by uniform stability, and then bound this by a tiny fraction of the {unobservable component in the previous time window}!
\begin{lemma}\label{lem:pi-piperp-ineq}
With probability at least $1 - \delta$, the following holds true for for every window $0\le \ell < T/t$, for $q \triangleq x_{\ell t} - x^*_{\ell t}$, all for all $i < t$.
There is a degree-4 SoS proof from the constraints in Program~\ref{program:sos} that
\begin{equation}
a^*_{\ell t + i} a_{\ell t + i}\| BA^i \Pi q\|^2 \leq 4a^*_{\ell t + i} a_{\ell t + i}\| BA^i \Pi^\perp q\|^2 + O(E_{\mathsf{noise}}),
\end{equation}
where recall that $E_{\mathsf{noise}}$ is defined in \eqref{eq:Enoise_def}.
\end{lemma}
\begin{proof}
Without loss of generality we can assume $\ell = 0$.
For any $i < t$, we have the following sequence of inequalities in degree-4 SoS \begin{align}
a_i a^*_i \norm*{BA^i(\Pi + \Pi^{\perp})q}^2 &= a_i a^*_i \norm*{BA^iq}^2 \\
&= a_i a^*_i\norm*{(y_i - Bx^*_i) - (y_i - Bx_i) + (Bx^*_i - BA^i x^*_0) - (Bx_i - BA^i x_0)}^2 \\
&\le 3\norm{v^*_i}^2 + 3\norm{v_i}^2 + 3\norm*{\sum^i_{s=1} BA^{i - s} (w_s - w^*_s)}^2 \le E_{\mathsf{noise}}, \label{eq:pipiperp}
\end{align}
where we used the constraints and event of Lemma~\ref{lem:feasible} in the last step (which holds with probability at least $1 - \delta$).
So the lemma follows by applying Fact~\ref{fact:sos_simple} to $\epsilon \triangleq E_{\mathsf{noise}}$, $v_1 = a^*_i a_i BA^i \Pi q $, and $v_2 = a^*_i a_i BA^i \Pi^\perp q $.
\end{proof}
\begin{lemma}\label{lem:observable-case-analysis}
Let pseudoexpectation $\psE{\cdot}$ be the solution to Program~\ref{program:sos}, assuming it is feasible. Provided the events of Lemma~\ref{lem:anticonc} and Lemma~\ref{lem:pi-piperp-ineq} hold, then at least one of the following holds for every window $0\le \ell < T/t$ for $q \triangleq x_{\ell t} - x^*_{\ell t}$:
\begin{enumerate}
\item (Observable component decays)
\begin{equation}
\psE*{\norm{A^{t}\Pi q}^2}
\le \frac{1}{10}\psE*{\| \Pi q \|^2}.
\end{equation}
\item (Observable error bounded by unobservable error)
\begin{equation}
\psE*{\norm{\Pi q}^2} \le \frac{1}{10\rho^2}\psE*{\norm{\Pi^{\perp} q}^2} + O\left(E_{\mathsf{noise}}\,\rho^4/\kappa\right).
\end{equation}
\end{enumerate}
\end{lemma}
\begin{proof}
Without loss of generality we can assume $\ell = 0$. In addition to the lower bound of \eqref{eq:psdlb}, we also have a degree-2 SoS proof of the upper bound
\begin{equation}\label{eqn:unobs-psd}
\sum_{i=0}^{T-1} a^*_i a_i \Pi^\perp (A^i)^T B^T B A^i \Pi^\perp \preceq \Pi^\perp \mathcal{O}_t \Pi^\perp \preceq \zeta \cdot I,
\end{equation}
where in the first step we used that $a_i a^*_i \le 1$ by Constraint~\ref{item:boolean} and in the second step we used the definition of $\Pi$.
For convenience, define $q\triangleq x_0 - x^*_0$. We proceed by casework on whether there is a gap between $\psE*{(\Pi q)^{\top} \calO_t (\Pi q)}$ and $\psE{\zeta\norm{\Pi q}^2}$:
\noindent{\textbf{Case 1}}: $\psE*{\| \Pi q \|^2} \ge \psE*{\frac{1}{4000\rho^2\zeta}\sum_{i=0}^{t-1} \| BA^i \Pi q\|^2}$.
The analysis for this case is very similar to the analysis in Lemma~\ref{lem:unobservable-decay}. We have
\begin{align*}
\psE*{\| \Pi q \|^2}
&\geq \psE*{\frac{1}{4000\rho^2\zeta}\sum_{i=0}^{t-1} \| BA^i \Pi q\|^2} \\
&= \psE*{\frac{1}{4000\rho^2\zeta}\sum_{j = 0}^{t/s-1} q^{\top}\Pi {A^{js}}^{\top} \calO_s A^{js} \Pi q} \geq \psE*{\frac{10\rho^2 s}{t}\sum_{j = 0}^{t/s-1} \|A^{js} \Pi q\|^2},
\end{align*}
where in the last step we used the definition of $\zeta$ and the assumption that $\lambda_{\min}(\calO_s) \ge \kappa s$. Rearranging, we obtain
\begin{equation}
\frac{1}{10\rho^2} \psE{\| \Pi q \|^2} \geq \frac{1}{t/s}\sum_{j=0}^{t/s-1} \psE*{\|A^{js} \Pi q\|^2}.
\end{equation}
Therefore, there exists some index $0 \le j < t/s$ for which $\psE*{\norm{A^{js}\Pi q}^2} \le \frac{1}{10\rho^2}\psE*{\norm{\Pi q}^2}$. By uniform stability, we obtain the first desired outcome in the lemma statement.
\noindent{\textbf{Case 2}}: $ \psE*{\| \Pi q \|^2} \le \psE*{\frac{1}{4000\rho^2\zeta}\sum_{i=0}^{t-1} \| BA^i \Pi q\|^2}$.
In this case we invoke \eqref{eq:psdlb} to obtain
\begin{equation} \label{eqn:fix-main}
\psE*{\| \Pi q \|^2} \leq \psE*{\frac{1}{4000\rho^2\zeta}\sum_{i=0}^{t-1}\| BA^i \Pi q\|^2} \leq \psE*{\frac{1}{40\rho^2\zeta}\sum_{i=0}^{t-1} a^*_i a_i\| BA^i \Pi q\|^2}.
\end{equation}
\iffalse
For any $i < t$, we have the following sequence of inequalities in degree-4 SoS \begin{align}
a_i a^*_i \norm*{BA^i(\Pi + \Pi^{\perp})q}^2 &= a_i a^*_i \norm*{BA^iq}^2 \\
&= a_i a^*_i\norm*{(y_i - Bx^*_i) - (y_i - Bx_i) + (Bx^*_i - BA^i x^*_0) - (Bx_i - BA^i x_0)}^2 \\
&\le 3\norm{v^*_i}^2 + 3\norm{v_i}^2 + 3\norm*{\sum^i_{s=1} BA^{i - s} (w_s - w^*_s)}^2 \\
&\le O\left(m\tau^2 + \tau^2\log(T/\delta) + t\norm{B}^2\rho^2 \cdot (d\sigma^2 + \sigma^2\log(T/\delta))\right), \label{eq:pipiperp}
\end{align}
so by applying Fact~\ref{fact:sos_simple} to $\epsilon = O\left(m\tau^2 + \tau^2\log(T/\delta) + t\norm{B}^2\rho^2 \cdot (d\sigma^2 + \sigma^2\log(T/\delta))\right)$, $v_1 = a^*_i a_i BA^i \Pi q $, and $v_2 = a^*_i a_i BA^i \Pi^\perp q $,
\fi
Recall from Lemma~\ref{lem:pi-piperp-ineq} that we have a degree-4 SoS proof of
\begin{equation}
a^*_i a_i\| BA^i \Pi q\|^2 \leq 4a^*_i a_i\| BA^i \Pi^\perp q\|^2 + O(E_{\mathsf{noise}}).
\end{equation}
Summing this inequality over $i < t$ and taking pseudo-expectations, we get
\[\psE*{\sum_{i=0}^{t-1} a^*_i a_i\| BA^i \Pi q\|^2} \le \psE*{4\sum_{i=0}^{t-1}a^*_i a_i\| BA^i \Pi^\perp q\|^2} + O(E_{\mathsf{noise}}t). \]
Substituting this back into the main bound \eqref{eqn:fix-main}, we get
\begin{align}
\psE*{\| \Pi q \|^2} &\leq \psE*{\frac{1}{10\rho^2\zeta}\sum_{i=0}^{t-1} a^*_i a_i\| BA^i \Pi^\perp q\|^2} + O\left(\frac{E_{\mathsf{noise}} t}{30\zeta}\right) \\
&\le \psE*{\frac{1}{10\rho^2}\norm{\Pi^{\perp} q}^2} + O\left(\frac{E_{\mathsf{noise}} t}{30\zeta}\right),
\end{align} where in the last step we used \eqref{eqn:unobs-psd}. Unpacking the definition of $\zeta$, we arrive at the second desired bound.
\end{proof}
We are now ready to prove Lemma~\ref{lem:error_decay}:
\begin{proof}[Proof of Lemma~\ref{lem:error_decay}]
By the SoS triangle inequality,
\begin{align*}
\MoveEqLeft\|x_{\ell t} - x^*_{\ell t}\|^2 \\
&\le 2 \|A^t(x_{(\ell - 1)t} - x^*_{(\ell - 1)t})\|^2 + O(t^2 \sigma^2(d + \log(T/\delta))) \\
&\le 4 \|A^t\Pi(x_{(\ell - 1)t} - x^*_{(\ell - 1)t})\|^2 + 4\|A^t\Pi^{\perp} (x_{(\ell - 1)t} - x^*_{(\ell - 1)t})\|^2 + O(t^2 \sigma^2(d + \log(T/\delta))) \\
&\le 4 \|A^t\Pi(x_{(\ell - 1)t} - x^*_{(\ell - 1)t})\|^2 + (1/10000) \|\Pi^{\perp} (x_{(\ell - 1)t} - x^*_{(\ell - 1)t})\|^2 + O(t^2 \sigma^2(d + \log(T/\delta)))
\end{align*}
where we used Constraint~\ref{item:dynamics_noise} and triangle inequality in the first inequality, SoS triangle inequality in the second inequality, and Lemma~\ref{lem:unobservable-decay} in the third inequality.
Now based on Lemma~\ref{lem:observable-case-analysis} applied to $\ell - 1$, we consider the following two cases:
\noindent\textbf{Case 1}: Observable component decays, that is, we have
\[ \psE*{\norm{A^{t}\Pi (x_{(\ell - 1) t} - x^*_{(\ell -1) t})}^2} \leq \frac{1}{10} \psE*{\| \Pi (x_{(\ell -1) t} - x^*_{(\ell - 1) t}) \|^2}. \]
Then we can argue that the error at time $\ell t$ is a small fraction of the error at time $(\ell - 1)t$ because both the observable and unobservable components of the error at time $(\ell - 1)t$ have decayed over $t$ steps. Formally:
\begin{align*}
\MoveEqLeft\psE{\|x_{\ell t} - x^*_{\ell t}\|^2} \\
&\le \psE*{4 \|A^t\Pi(x_{(\ell - 1)t} - x^*_{(\ell - 1)t})\|^2 + (1/10000) \|\Pi^{\perp} (x_{(\ell - 1)t} - x^*_{(\ell - 1)t})\|^2} + O(t^2 \sigma^2(d + \log(T/\delta))) \\
&\le \psE*{(2/5)\|\Pi (x_{(\ell - 1)t} - x^*_{(\ell - 1)t})\|^2 + (1/10000) \|\Pi^{\perp} (x_{(\ell - 1)t} - x^*_{(\ell - 1)t})\|^2} + O(t^2 \sigma^2(d + \log(T/\delta))) \\
&\le (1/2) \psE*{\|x_{(\ell - 1)t} - x^*_{(\ell - 1)t}\|^2} + O(t^2 \sigma^2(d + \log(T/\delta))).
\end{align*}
where in the last step we used the Pythagorean Theorem.
\noindent\textbf{Case 2}: Observable error bounded by unobservable error, that is
\begin{equation}
\psE*{\norm{\Pi q}^2} \le \frac{1}{10\rho^2}\psE*{\norm{\Pi^{\perp} q}^2} + O(E_{\mathsf{noise}}\rho^4/\kappa) \label{eq:usecase2}
\end{equation}
where $q \triangleq x_{(\ell - 1)t} - x^*_{(\ell - 1)t}$. Then we can argue that the error at time $\ell t$ is a small fraction of the error at time $(\ell - 1)t$ as follows. As in Case 1, the unobservable error at time $(\ell - 1)t$ has decayed. As discussed above, the observable error at time $\ell t$ might even be bigger than the observable error at time $(\ell - 1)t$, but it can't be much bigger because of uniform stability. On the other hand, the latter is bounded by a small fraction of the \emph{unobservable} error at time $(\ell - 1)t$. This lets us conclude that the overall error at time $\ell t$ is bounded even by the unobservable error at time $(\ell - 1)t$. Formally,
\begin{align*}
\MoveEqLeft \psE{\|x_{\ell t} - x^*_{\ell t}\|^2}
\\
&\le \psE*{4 \|A^t\Pi(x_{(\ell - 1)t} - x^*_{(\ell - 1)t})\|^2 + (1/10000) \|\Pi^{\perp} (x_{(\ell - 1)t} - x^*_{(\ell - 1)t})\|^2} + O(t^2\sigma^2(d + \log(T/\delta))) \\
&\le 4 \rho^2 \|\Pi(x_{(\ell - 1)t} - x^*_{(\ell - 1)t})\|^2 + (1/10000) \|\Pi^{\perp} (x_{(\ell - 1)t} - x^*_{(\ell - 1)t})\|^2 + O(t^2\sigma^2(d + \log(T/\delta))) \\
&\le \frac{1}{2}\psE*{\norm{q}^2} + O(\rho^6 E_{\mathsf{noise}} t/\kappa).
\end{align*}
where $C < 1$ is an absolute constant and in the last step we used \eqref{eq:usecase2} and absorbed $O(t^2\sigma^2(d + \log(T/\delta)))$ into $O(\rho^6 E_{\mathsf{noise}}t/\kappa)$.
Since we showed the desired conclusion in both cases, the proof is complete.
\end{proof}
\subsection{Confidence Band Recovery}
Here we note that as a consequence of Theorem~\ref{thm:main_TlogT}, our estimate $\brc{\psE{x_i}}$ of the trajectory is actually pointwise $O(\log T)$-close to the true trajectory at all time steps, except for a $o(1)$ proportion of time close to time zero. This will be useful in the next section when we use this as a warm start for a second sum-of-squares relaxation that will achieve excess risk \emph{doubly logarithmic} in $T$.
\begin{corollary}\label{cor:warmstart}
Let pseudoexpectation $\psE{\cdot}$ be the solution to Program~\ref{program:sos}, assuming it is feasible. Then provided the event of Lemma~\ref{lem:anticonc} holds, for all $0\le i < T$ the estimates $\brc{\psE{x_i}}$ satisfy
\begin{equation}
\norm{\psE{x_i} - x^*_i} \lesssim \frac{\rho}{2^{\ell(i)/2}}R\left(\sqrt{d} + \sqrt{\log(1/\delta)}\right) + O(\rho^4 E^{1/2}_{\mathsf{noise}}t^{1/2}/\kappa^{1/2})
\end{equation}
\end{corollary}
\begin{proof}
By unrolling Lemma~\ref{lem:error_decay} and using the fact that $\psE{\norm{x_0 - x^*_0}^2} \le 2R^2(d + O(\log(1/\delta)))$, we conclude that
\begin{equation}
\psE{\|x_{\ell t} - x^*_{\ell t}\|^2} \le \frac{1}{2}\psE{\|x_{(\ell - 1) t} - x^*_{(\ell - 1) t}\|^2} + O(\rho^6 E_{\mathsf{noise}}t/\kappa) \le \frac{1}{2^{\ell - 1}}\cdot 2R^2(d + O(\log(1/\delta))) + O(\rho^6 E_{\mathsf{noise}}t/\kappa),
\end{equation}
so by Lemma~\ref{lem:psE-CS}, we have that
\begin{equation}
\|\psE{x_{\ell t}} - x^*_{\ell t}\|^2 \le 4R^2(d + O(\log(1/\delta)))/2^{\ell} + O(\rho^6 E_{\mathsf{noise}}t/\kappa).
\end{equation}
It remains to bound the error on the iterates \emph{within} each window. For any $0 < i < t$, Constraint~\ref{item:dynamics} and Fact~\ref{fact:unroll} imply that
\begin{align}
\MoveEqLeft\norm*{\psE{x_{\ell t + i}} - x^*_{\ell t +i}} \\
&= \norm*{A^i (\psE{x_{\ell t + i}} - x^*_{\ell t + i}) + \sum^i_{j=1} A^{i-j} (\psE{w_{\ell t + j}} - w^*_{\ell t + j})} \\
&\le \rho\norm{\psE{x_{\ell t +i}} - x^*_{\ell t +i}} + \sum^i_{j=1} \rho\norm*{\psE{w_{\ell t + j}} - w^*_{\ell t+j}} \\
&\lesssim \rho \left(\frac{1}{2^{\ell/2}}R\left(\sqrt{d} + \sqrt{\log(1/\delta)}\right) + O(\rho^3 E^{1/2}_{\mathsf{noise}}/\kappa^{1/2})\right) + t\rho\left(\sigma\sqrt{d} + \sigma\sqrt{\log(T/\delta)}\right). \\
&\lesssim \rho \left(\frac{1}{2^{\ell/2}}R\left(\sqrt{d} + \sqrt{\log(1/\delta)}\right) + O(\rho^3 E^{1/2}_{\mathsf{noise}}/\kappa^{1/2})\right)
\end{align}
as desired.
\end{proof}
\section{Achieving \texorpdfstring{$\log\log T$}{loglog T} Excess Risk}
\label{sec:TloglogT}
In this section, we complete the proof of our main theorem by showing how to refine the poly-logarithmic excess risk guarantee of the previous section to get doubly logarithmic excess risk. At a very high level, the bottleneck in the analysis from the previous section arises when we argue that for most windows, the uncorrupted time steps subsample the observability matrix. The issue is that for any window, with probability $\delta$ the error from matrix concentration corresponding to the right-hand side of \eqref{eq:deviation} will scale worse than $\sqrt{\log(d/\delta)}$. In the previous section, we simply took $\delta = O(1/T)$ so that by a union bound, with high probability this will not happen for any window, hence our poly-logarithmic excess risk bound.
In this section, we instead take $\delta = O(1/\mathsf{polylog} T)$ so that matrix concentration will now fail on $O(1/\mathsf{polylog} T)$ fraction of the windows, but over the windows where matrix concentration holds, the risk scales with $\log\log T$ instead of $\log T$. To handle the ``bad'' windows where matrix concentration fails, we will exploit the fact that by Corollary~\ref{cor:warmstart}, the estimate from the previous section is pointwise $O(\mathsf{polylog} T)$-close to the true trajectory (i.e. it forms a high probability confidence band). By folding this into a new sum-of-squares relaxation and introducing additional indicator variables corresponding to the events that matrix concentration holds in particular windows, we are able to bound the contribution of the bad windows to the risk by $O(\mathsf{polylog}(T)/\mathsf{polylog}(T)) = O(1)$. Formally, we show the following bound:
\begin{theorem}\label{thm:main_TloglogT}
For any $\eta \le 0.49$, there is a polynomial-time algorithm that, given the corrupted observations $\brc{\wt{y}_i}$, outputs a trajectory $\brc{\wh{x}_i}$ and steps $\brc{\wh{w}_i}$ for which $\wh{x}_i = A\wh{x}_{i - 1} +\wh{w}_i$ for every $i\in [T]$, and for which
\begin{multline}
\frac{1}{T}\sum^{T-1}_{i=0}\left(a^*_i \norm{B\wh{x}_i - y_i}^2/\tau^2 + \norm{\wh{w}_i}^2/\sigma^2\right) + \frac{\|\wh{x}_0\|^2}{R^2 T} - \mathsf{OPT} \\
\lesssim \poly(s,\alpha,1/\kappa,\rho,\|B\|,\log d,\log(1/\delta)) \brk*{ \eta\log(1/\eta)\left(m + d(\sigma^2/\tau^2)\log\log T\right) + \eta^{1/2}(R^2 d/\tau^2)T^{-0.49}}
\end{multline}
provided $T \ge \Max{\wt{\Omega}(\eta^{-1}\kappa^{-2}\rho^{12}\norm{B}^4\log d)}{\Omega(t^2\log^k(1/\delta))})$.
\end{theorem}
\subsection{Sum-of-Squares Relaxation}
In this section we define a new sum-of-squares program to exploit Corollary~\ref{cor:warmstart}. Recall that the estimate from the previous section, call it $\brc{x'_i}$, satisfies
\begin{equation}
\norm{x'_i - x^*_i}^2 \lesssim \frac{\rho^2}{2^{\ell(i)}}R^2(d + \log(1/\delta)) + \epsilon_{\mathsf{geo}},
\end{equation}
where $\epsilon_{\mathsf{geo}}$ satisfies
\begin{equation}
\epsilon_{\mathsf{geo}} \triangleq O\left(\frac{\rho^8 t_{\mathsf{pre}}}{\kappa}\cdot \brk*{\tau^2\left(m + \log(T/\delta)\right) + \sigma^2 \left(d +\log(T/\delta)\right)\cdot \rho^2 t_{\mathsf{pre}} \norm{B}^2}\right) \label{eq:eps_geo2}
\end{equation}
for $t_{\mathsf{pre}}\triangleq \Max{s}{\wt{\Theta}(\kappa^{-2}\rho^{12}\norm{B}^4\log(dT/\delta))}$ from \eqref{eq:tdef} denoting the window size parameter from the previous section.
Note that $t_{\mathsf{pre}}$ scales logarithmically in $T$ and $\epsilon_{\mathsf{geo}}$ therefore scales as $\log^3 T$. We now define our program:
\begin{program}\label{program:sosv2}
Let $\brc{y_i}$ be the observations we are given, and let $\delta_1 > 0$ and window size $t\in\mathbb{N}$ be parameters to be tuned later. The program also takes in as parameters a sequence $\brc{x'_0,\ldots,x'_{T-1}}$ of vectors in $\R^d$, corresponding to an estimate of the trajectory.
The program variables are $\brc{x_i}$, $\brc{w_i}$, $\brc{v_i}$, and $\brc{a_i}$ as before, as well as Boolean variables $\brc{b_{\ell}}^{T/t-1}_{\ell = 0}$ (indicators for windows where matrix concentration holds), and the constraints are that for all $0 \le i < T$ and $0 \le \ell < T/t$:
\vspace{0.3cm}
\noindent\emph{Constraints from Program~\ref{program:sos}}
\begin{enumerate}
\item $a_i^2 = a_i$\label{constraint2:a_boolean}
\item $x_i = Ax_{i-1} + w_i$ \label{constraint2:dynamics}
\item $a_i(y_i - Bx_{i-1} - v_i) = 0$ \label{constraint2:measurements}
\item $\sum_{i=0}^{T-1} a_i \ge (1-1.01\eta)T$ \label{constraint2:many_inliers}
\item $\norm{x_0}^2 \le R^2(d + O(\log(1/\delta)))$ \label{constraint2:bound}
\end{enumerate}
\noindent\emph{Random windows where subsampling succeeds/fails}
\begin{enumerate}
\setcounter{enumi}{5}
\item $b_{\ell}^2 = b_{\ell}$ \label{constraint2:b_boolean}
\item $\sum_{\ell=0}^{T/t-1} b_{\ell} \ge (1 - \delta_1)\cdot T/t$ \label{constraint2:many_b}
\item $\sum_{i=0}^{T-1} (1 - b_{\ell(i)})(1 - a_i) \le \eta\delta_1$\label{constraint2:many_ab}
\end{enumerate}
\noindent\emph{Confidence band given by $\brc{x'_i}$}
\begin{enumerate}
\setcounter{enumi}{8}
\item $\norm{x_i - x'_i}^2 \le \epsilon_{\mathsf{geo}} + O(\rho^2 R^2(d + \log(1/\delta))/2^{\ell(i)})$ \label{constraint2:funnel}
\end{enumerate}
\noindent\emph{Process and observation noise bounded on average}
\begin{enumerate}
\setcounter{enumi}{9}
\item $\frac{1}{T}\sum_{\ell<T/\t, j<\t} (1 - a_{\ell t + j}) \norm{\sum^j_{i=1}BA^{j - i}w_{\ell\t+i}}^2 \le O\left(\eta^{1-2/k}\t\alpha\sigma^2\rho^2 m k\right)$ \label{constraint2:totalnoise}
\item $\frac{1}{T}\sum_{\ell<T/\t} \norm{\sum^t_{i=1}A^{t-i}w_{\ell\t+i}}^2 \le O\left(\sigma^2\rho^2 d\right)$ \label{constraint2:totalnoise_dumb}
\item $\frac{1}{T}\sum_{i = 0}^{T-1} \norm{v_i}^2 = O(\tau^2(m + \log(2/\delta)/T))$\label{constraint2:measurements_noise}
\item $\sum_{i=0}^{T-1} (1 - a_i)\norm{v_i}^2 \le O(mk\tau^2\cdot\eta^{1-2/k})$ \label{constraint2:measurements_noise_corrupted}
\end{enumerate}
\noindent\emph{Subsampling with confidence $1 - \delta_1$ in each window}
\begin{enumerate}
\setcounter{enumi}{13}
\item $b_{\ell} \sum^{\t- 1}_{j=0} (1 - a_{\ell t + j}) (A^j)^{\top} B^{\top} B A^j \preceq b_{\ell}\cdot \left(\eta\cdot \calO_{\t} + O\left(\rho^2\norm{B}^2\sqrt{\t\log(d/\delta_1)}\right)\cdot \Id\right)$ \label{constraint2:subsample}
\end{enumerate}
The program objective is to minimize
\begin{equation}
\min\psE*{\frac{1}{T} \sum^{T-1}_{i=0} \left(a_i \norm{Bx_i - y_i}^2 / \tau^2 + \norm{w_i}^2 / \sigma^2\right)}\label{eq:obj2}
\end{equation}
over degree-$4$ pseudoexpectations satisfying the above constraints.
\end{program}
As Program~\ref{program:sosv2} is considerably more involved than Program~\ref{program:sos} (note in particular the new Constraints~\ref{constraint2:b_boolean}-\ref{constraint2:subsample}), here we spend some time to highlight the key differences.
First notice that we now have an additional set of Boolean variables $\brc{b_{\ell}}$ indicating the windows whose uncorrupted time steps the program deems to have properly subsampled the observability matrix. In Constraint~\ref{constraint2:subsample} we choose the error in subsampling to scale with $\sqrt{\log(d/\delta_1)}$ for a parameter $\delta_1$ which we will take to scale logarithmically in $T$ so that $1 - \delta_1$ fraction of windows will incur at most this level of subsampling error (Constraint~\ref{constraint2:many_b}). We will additionally insist that the fraction of time steps which are simultaneously corrupted and belong to a window where matrix concentration fails is at most roughly $\eta\delta_1$ (Constraint~\ref{constraint2:many_ab}).
Also note that we have introduced an additional set of parameters, namely the sequence $\brc{x'_0,\ldots,x'_{T-1}}$ which we will ultimately take to be the trajectory estimate from the algorithm in Section~\ref{sec:logT}. In Constraint~\ref{constraint2:funnel} we encode the fact that this estimate is pointwise $\log T$-close to the true trajectory by Corollary~\ref{cor:warmstart}.
Finally, in lieu of constraining $\norm{v_i}^2,\norm{w_i}^2$ to be bounded for all $i$, we simply constrain them to be small on average in an appropriate sense captured by Constraints~\ref{constraint2:totalnoise}, \ref{constraint2:totalnoise_dumb}, \ref{constraint2:measurements_noise}, \ref{constraint2:measurements_noise_corrupted}. These four constraints may appear somewhat \emph{ad hoc}, but at a high level the reason for their inclusion is that in various parts in the proof of Theorem~\ref{thm:main_TloglogT}, we need to upper bound some sum over noise terms across all time steps or all time windows. In Section~\ref{sec:logT} we handled these sums by constraining their individual summands to be bounded, but this incurred a factor of $\log T$ through the union bound. Here we are being more careful and simply constraining the full sums to be bounded, leading to constraints in Program~\ref{program:sosv2} that are somewhat harder to parse but yield better error bounds.
We give a full description of our algorithm in Algorithm~\ref{alg:sos} below.
\begin{algorithm2e}
\DontPrintSemicolon
\caption{\textsc{SoSKalman}($\brc{y_i},\delta$)}
\label{alg:sos}
\KwIn{Observations $y_0,\ldots,y_{T-1}$, failure probability $\delta$}
\KwOut{Estimate $\wh{x}_0,\ldots,\wh{x}_{T-1}$ for the trajectory}
$\delta_1 \gets \Theta(\log(1/\delta)/\log^3 T)$.\;
$t_{\mathsf{pre}} \gets \Max{s}{\wt{\Theta}(\kappa^{-2}\rho^{12}\norm{B}^4\log(dT/\delta)))}$.\;
$t \gets \Max{s}{\wt{\Theta}(\kappa^{-2}\rho^{12}\norm{B}^4\log(d/\delta_1)))}$.\;
Let $\wt{\mathbb{E}}_{\mathsf{pre}}[\cdot]$ be the pseudoexpectation solving Program~\ref{program:sos} with window size $t_{\mathsf{pre}}$.\;
$x'_i \gets \wt{\mathbb{E}}_{\mathsf{pre}}[x_i]$ for every $0 \le i < T$.\;
Let $\psE{\cdot}$ be the pseudoexpectation solving Program~\ref{program:sosv2} with window size $t$ and sequence $\brc{x'_0,\ldots,x'_{T-1}}$.\;
$\wh{x}_i\gets \psE{x_i}$ for every $0 \le i < T$.\;
Output $\brc{\wh{x}_0,\ldots,\wh{x}_{T-1}}$.\;
\end{algorithm2e}
\subsection{Feasibility of Oracle Kalman Smoother}
The goal for the this subsection will be to prove that Program~\ref{program:sosv2} has a feasible solution corresponding the output of the oracle Kalman smoother.
To show this, it suffices by the argument of Lemma~\ref{lem:feasible} to show that the ground truth $x^*$ is feasible with high probability:
\begin{lemma}\label{lem:feasible2}
With probability at least $1 - \delta$, the ground truth $(x^*_i, w^*_i, v^*_i, a^*_i, b^*_{\ell})$ is feasible for Program~\ref{program:sosv2}
provided $T \ge \Omega(\Max{\brc*{(t/\eta\delta_1)\log(1/\delta)}}{\brc*{t^2\log^k(1/\delta)}})$.
\end{lemma}
To prove this, we first collect a number of tail bounds. We introduce random variables $b^*_{\ell}$ which measure failure of concentration for the ground truth. Formally, we define
\[ b^*_{\ell} = \bone*{\sum^{\t - 1}_{j=0} (1 - a^*_{\ell t + j}) (A^j)^{\top} B^{\top} B A^j \preceq \eta\cdot \calO_{\t} + O\left(\rho^2\norm{B}^2 \sqrt{ \t \log(d/\delta_1)}\right)\cdot \Id}. \]
Next we establish concentration for the sum of the $b^*_{\ell}$, showing that there are many ``good'' windows with high probability. Note that in the following Lemma, we require $\delta_1$ is at least size $\Omega((t/T) \log(1/\delta))$ for this concentration argument to work, but we actually will apply the Lemma with a larger value of $\delta_1$ in our analysis. Informally, picking a relatively large $\delta_1$ means that concentration fails for more windows (which we can handle using our confidence band technique) but we get tighter concentration bounds for the remaining ``good'' windows.
\begin{lemma}\label{lem:bstar_conc}
With probability at least $1 - \delta$,
\[ \sum_{\ell = 0}^{T/t-1} b^*_{\ell} \ge (1 - \delta_1) (T/t) \]
provided $T = \Omega((t/\delta_1) \log(1/\delta))$.
\end{lemma}
\begin{proof}
First, observe that the $b^*_{\ell}$ are i.i.d. Bernoulli random variables, since the only randomness in their definition is the $a^*_j$ which are i.i.d. and the windows do not overlap. Next, observe that
\begin{equation}
\E{b^*_{\ell}} = \Pr*{\sum^{\t - 1}_{j=0} (1 - a^*_{\ell t + j}) (A^j)^{\top} B^{\top} B A^j \preceq \eta\cdot \calO_{\t} + O\left(O(\rho^2\norm{B}^2 \sqrt{ \t \log(d/\delta_1)})\right)\cdot \Id} \ge 1 - \delta_1/2 \label{eq:bexp}
\end{equation} for any $\ell < T/t$
by Lemma~\ref{lem:matrix_conc}. Therefore by Bernstein's inequality,
\[ \Pr*{\sum_{\ell = 0}^{T/t - 1} b^*_{\ell} < (1 - \delta_1) (T/t)} \le \exp\left(-\Omega\left(\frac{\delta_1^2 (T/t)^2}{\delta_1 (T/t) + (1/3) \delta_1 (T/t)}\right)\right) = \exp\left(-\Omega(\delta_1 T/t)\right) \le \delta, \]
provided that $\delta_1 = \Omega((t/T) \log(1/\delta))$.
\end{proof}
\begin{lemma}\label{lem:many_ab}
With probability at least $1 - \delta$,
\begin{equation}
\frac{1}{T}\sum^{T-1}_{i=0} (1 - b^*_{\ell(i)})(1 - a^*_i) \le \eta\delta \label{eq:avab}
\end{equation}
provided $T = \Omega((t/\eta\delta_1)\log(1/\delta))$.
\end{lemma}
\begin{proof}
For any $\ell < T/t$, consider the random variable $Z_{\ell} \triangleq \frac{1}{t}\cdot (1-b^*_{\ell})\sum^{t-1}_{j=0} (1 - a^*_{\ell t + j})$. Note that $|Z_{\ell}| \le 1$ with probability 1, and $\E{Z_{\ell}} \le \eta\delta_1/2$ by \eqref{eq:bexp}. Furthermore \begin{equation}
\Var{Z_{\ell}} \le \E{Z_{\ell}^2} \le \left(\eta + 1/t\right)\eta\delta_1/2.
\end{equation}
As the left-hand side of \eqref{eq:avab} is $Z\triangleq \frac{1}{T/t}\sum^{T/t-1}_{\ell = 0} Z_{\ell}$, by Bernstein's we have
\begin{equation}
\Pr*{\frac{1}{T/t}\sum^{T/t-1}_{\ell = 0} Z_{\ell} > \eta\delta} \le \exp\left(-\Omega\left(\frac{\eta^2\delta_1^2(T/t)}{\eta\delta_1 + \eta\delta_1(\eta + 1/t)} \right)\right) \le \exp\left(-\Omega(\eta\delta_1T/t)\right) \le \delta,
\end{equation}
provided $T = \Omega((t/\eta \delta_1)\log(1/\delta))$.
\end{proof}
The next three lemmas will allow us to show that the ground truth trajectory $\brc{x^*_i}$ and noise $\brc{w^*_i, v^*_i}$ satisfy Constraints~\ref{constraint2:totalnoise} and \ref{constraint2:totalnoise_dumb}.
\begin{lemma}\label{lem:triplesum}
For any $\delta > 0$, if $T \ge \t^2 \log^k(1/\delta)$ then
\begin{equation}
\left(\frac{1}{T}\sum^{T/\t - 1}_{\ell = 0}\sum^{t-1}_{j = 0} \norm*{\sum^j_{i=1}BA^{j-i} w^*_{\ell\t+i}}^k\right)^{2/k} \le O\left(\t\alpha\sigma^2\rho^2 m k\right)
\end{equation} with probability at least $1 - \delta$.
\end{lemma}
\begin{proof}
We first bound the expectation \begin{align}
\E*{\norm*{\sum^j_{i=1} BA^{j-i} w^*_{\ell\t + i}}^k} &= \sigma^k \cdot \E[g\sim\calN(0,\calO_{j-1})]{\norm{g}^k} \\
&\le \sigma^k \cdot \E[g\sim\calN(0,\calO_{\t})]{\norm{g}^k}\\
&\le (\t\alpha\sigma^2\rho^2)^{k/2}\cdot\E[g\sim\calN(0,\Id_m)]{\norm{g}^k} \\
&\le O(\t\alpha\sigma^2\rho^2 m k)^{k/2} \label{eq:firstmoment},
\end{align} where the second and third steps follow by the fact that $\calO_{j-1} \preceq \calO_{\t}\preceq \t\alpha\rho^2$ (by Fact~\ref{fact:epochs}) so that the norm of a sample from $\calN(0,\calO_{j-1})$ is stochastically dominated by that of one from $\calN(0,\t\alpha\sigma^2\rho^2\cdot \Id)$.
Next for any $\ell$ we crudely upper bound the variance of $\sum^{\t-1}_{j=0}\norm{\sum^j_{i=1}BA^{j-i}w^*_{\ell\t+i}}^k$ by its second moment:
\begin{equation}
\E*{\left(\sum^{\t-1}_{j=0}\norm*{\sum^j_{i=1} BA^{j-i} w^*_{\ell\t + i}}^k\right)^2} \le \sigma^{2k}\t\sum^{\t-1}_{j=0}\E*{\norm*{\sum^j_{i=1}BA^{j-i}w^*_{\ell\t + i}}^{2k}} \le t^2 \cdot O(\t\alpha\sigma^2\rho^2 m k)^k, \label{eq:secondmoment}
\end{equation} where in the last step we used \eqref{eq:firstmoment}.
Putting \eqref{eq:firstmoment} and \eqref{eq:secondmoment}, we conclude that for the random variable $Z \triangleq \frac{1}{T}\sum_{\ell ,j}\norm*{\sum^j_{i=1} BA^{j-i} w^*_{\ell\t + i}}^k$, we have \begin{equation}
\E{Z} \le O(\t\alpha\sigma^2\rho^2 dk)^{k/2} \qquad \Var{Z} \le \frac{\t^2}{T}\cdot O(\t\alpha\sigma^2 \rho^2 dk)^k,
\end{equation}
so by Lemma~\ref{lem:polyconc} we conclude that $Z \le (1 + \t\log^{k/2}(1/\delta)/\sqrt{T})\cdot O(\t\alpha \sigma^2 \rho^2 d k)^{k/2}$ with probability at least $1 - \delta$. The desired bound follows from the assumed bound on $T$.
\end{proof}
We now use Lemma~\ref{lem:triplesum} to verify that the ground truth satisfies Constraint~\ref{constraint2:totalnoise} with high probability.
\begin{corollary}\label{cor:totalnoise_feasible}
For any $\delta > 0$, if $T \ge \Omega(\Max{\brc*{ \Omega(\log(2/\delta)/\eta)}}{\brc*{t^2\log^k(1/\delta)}})$, then
\begin{equation}
\frac{1}{T}\sum^{T/\t - 1}_{\ell = 0}\sum^{t-1}_{j = 0} a^*_{\ell t + j} \norm*{\sum^j_{i=1}BA^{j-i} w^*_{\ell\t+i}}^2 \le O\left(\eta^{1-2/k}\t\alpha\sigma^2\rho^2 m k\right)
\end{equation} with probability at least $1 - \delta$.
\end{corollary}
\begin{proof}
By H\"{o}lder's, we can bound the left-hand side by
\begin{equation}
\left(\frac{1}{T}\sum^{T/t-1}_{\ell=0}\sum^{t-1}_{j=0}a^*_{\ell t + j}\right)^{1 - 2/k}\left(\sum^{T/t-1}_{\ell = 0}\sum^{t-1}_{j = 0}\norm*{\sum^j_{i=1} BA^{j-i}w^*_{\ell t + i}}^k\right)^{2/k} \le O\left(\eta^{1-2/k}\t\alpha\sigma^2\rho^2 m k\right),
\end{equation}
where the last step follows by Bernstein's applied to $\brc{a_i}$ and by Lemma~\ref{lem:triplesum}.
\end{proof}
Using a proof similar to that of Lemma~\ref{lem:triplesum}, we now verify that the ground truth satisfies Constraint~\ref{constraint2:totalnoise_dumb} with high probability.
\begin{lemma}\label{lem:triplesum_dumb}
For any $\delta > 0$, if $T \ge \log^2(1/\delta)$ then \begin{equation}
\frac{1}{T}\sum^{T/\t - 1}_{\ell = 0} \norm*{\sum^t_{i=1}A^{t-i} w^*_{\ell\t+i}}^2 \le O\left(\sigma^2\rho^2 d\right)
\end{equation} with probability at least $1 - \delta$.
\end{lemma}
\begin{proof}
Let $M \triangleq \sum^t_{i = 1} (A^{t - i})^{\top} A^{t - i}$. We first bound the expectation \begin{equation}
\E*{\norm*{\sum^t_{i=1} A^{t-i} w^*_{\ell\t + i}}^2} = \sigma^2 \cdot \E[g\sim\calN(0,M)]{\norm{g}^2} \le \t\sigma^2\rho^2 d \label{eq:firstmoment2},
\end{equation} where the second step follows by the fact that $\norm{M} \le \sum^t_{i = 1}\norm{(A^{t - i})^{\top} A^{t - i}} \le t\cdot\rho^2$
so that the norm of a sample from $\calN(0,M)$ is dominated by that of one from $\calN(0,\t\rho^2\cdot \Id_d)$.
Next for any $\ell$ we upper bound the variance of $\norm{\sum^b_{i=1}A^{b-i}w^*_{\ell\t+i}}^2$ by its second moment:
\begin{equation}
\E*{\norm*{\sum^t_{i=1} A^{t-i} w^*_{\ell\t + i}}^4} = \sigma^4 \cdot \E[g\sim\calN(0,M)]{\norm{g}^4} O(\t\sigma^2\rho^2 d)^2, \label{eq:secondmoment2}
\end{equation} where in the last step we used \eqref{eq:firstmoment2}.
Putting \eqref{eq:firstmoment2} and \eqref{eq:secondmoment2}, we conclude that for the random variable $Z \triangleq \frac{1}{T}\sum_{\ell}\norm*{\sum^t_{i=1} BA^{t-i} w^*_{\ell\t + i}}^2$, we have \begin{equation}
\E{Z} \le O(\sigma^2\rho^2 d) \qquad \Var{Z} \le \frac{1}{T}\cdot O(\sigma^2 \rho^2 d)^2,
\end{equation}
so by Lemma~\ref{lem:polyconc} we conclude that $Z \le (1 + \log(1/\delta)/\sqrt{T})\cdot O(\sigma^2 \rho^2 d)$ with probability at least $1 - \delta$. The desired bound follows from the assumed bound on $T$.
\end{proof}
\begin{lemma}\label{lem:feasible_corrupt_measurement_noise}
For any even $k > 0$, if $T \ge \Omega(\Max{\brc*{ \Omega(\log(2/\delta)/\eta)}}{\brc*{\log^k(1/\delta)}})$, then
\begin{equation}
\sum^{T-1}_{i = 0} (1 - a^*_i)\norm{v^*_i}^2 \le O(mk\tau^2\cdot \eta^{1-2/k}).
\end{equation} with probability at least $1 - \delta$.
\end{lemma}
\begin{proof}
By H\"{o}lder's, \begin{equation}
\frac{1}{T} \sum^{T-1}_{i=0} (1 - a^*_i) \norm{v^*_i}^2 \le \left(\frac{1}{T}\sum^{T-1}_{i=0} (1-a^*_i)\right)^{1 - 2/k}\left(\frac{1}{T}\sum^{T-1}_{i=0} \norm{v^*_i}^k\right)^{2/k}. \label{eq:holdersav}
\end{equation}
By Bernstein's, we have that $\frac{1}{T}\sum^{T-1}_{i=0} (1 - a^*_i) \le 1.01\eta$. To control the random variable $Z\triangleq \frac{1}{T}\sum^{T-1}_{i=0} \norm{v^*_i}^k$, note that this is a degree-$k$ polynomial in an $mT$-dimensional Gaussian with covariance $\tau^2\cdot\Id$. We have $\E{Z} \le (mk\tau^2)^{k/2}$ and $\Var{Z}\le \E{Z^2} \le \frac{1}{T}(2mk\tau^2)^k$, so by Lemma~\ref{lem:polyconc},
\begin{equation}
\Pr*{Z > O(mk\tau^2)^{k/2}(1 + \log^{k/2}(1/\delta)/\sqrt{T})} \le \delta.
\end{equation}
By the assumed bound on $T$, we conclude that $Z \le O(mk\tau^2)^{k/2}$ with probability at least $1 - \delta$. The lemma follows by \eqref{eq:holdersav}.
\end{proof}
We can finally complete the proof that Program~\ref{program:sosv2} is feasible:
\begin{proof}[Proof of Lemma~\ref{lem:feasible2}]
The new constraints that we need to verify are Constraints~\ref{constraint2:many_b}, \ref{constraint2:funnel}, \ref{constraint2:totalnoise}, \ref{constraint2:totalnoise_dumb}, \ref{constraint2:measurements_noise}, \ref{constraint2:subsample}. Constraints~\ref{constraint2:many_b} and \ref{constraint2:subsample} follow from Lemma~\ref{lem:bstar_conc}, while Constraint~\ref{constraint2:many_ab} follows from Lemma~\ref{lem:many_ab}. Constraint~\ref{constraint2:funnel} follows from Corollary~\ref{cor:warmstart}. Constraint~\ref{constraint2:totalnoise} follows Corollary~\ref{cor:totalnoise_feasible}. Constraint~\ref{constraint2:totalnoise_dumb} follows from
Lemma~\ref{lem:triplesum_dumb}. Constraint~\ref{constraint2:measurements_noise} follows from Lemma~\ref{lem:normbound}. Constraint~\ref{constraint2:measurements_noise_corrupted} follows from Lemma~\ref{lem:feasible_corrupt_measurement_noise}.
\end{proof}
\subsection{Outer Argument}
In this section we show, analogously to Lemma~\ref{lem:outer} in Section~\ref{sec:logT}, that in order to compete with $\mathsf{OPT}$, it suffices for our estimate of the trajectory to be sufficiently accurate on the first iterate of every window. The difference between the following lemma and Lemma~\ref{lem:outer} is that here, we do not assume that matrix concentration holds in every window, but we do exploit the fact that the algorithm from the previous section yields a confidence band around the true trajectory.
\begin{lemma} \label{lem:outer2}
Let $\psE{\cdot}$ be the solution to Program~\ref{program:sosv2}, assuming it is feasible. Let $\wh{x}_i \triangleq \psE{x_i}$ and $\wh{w}_i \triangleq \psE{w_i}$ for every $0 \le i < T$. Provided the event of Lemma~\ref{lem:feasible2} holds, then
\begin{align*}
\MoveEqLeft\frac{1}{T}\sum^{T-1}_{i=0}\left(a^*_i \norm{B\wh{x}_i - y_i}^2/\tau^2 + \norm{\wh{w}_i}^2/\sigma^2\right) + \frac{\|\wh{x}_0\|^2}{R^2 T} - \mathsf{OPT} \\
&\lesssim \tau^{-2}\left\{\eta \rho^2 \left(\alpha +\norm{B}^2\sqrt{\log(t/\delta_1)/t}\right)\cdot \frac{1}{T/t}\sum^{T/t - 1}_{\ell=0} \psE*{\norm{x_{\ell t} - x^*_{\ell t}}^2} \right.\\
&\quad \left. + k\eta^{1-2/k}\cdot \left(\t\alpha\sigma^2\rho^2 m + m\tau^2\right) + \norm{B}^2\left(\epsilon_{\mathsf{geo}}\eta\delta_1 + \rho^2R^2(d+\log(1/\delta))\cdot \left(\frac{\eta\delta_1}{T/t}\right)^{1/2}\right)\right\}
\end{align*}
\end{lemma}
If we ignore the $\frac{\eta\delta_1}{T/t}$ term as it is vanishing in $T$, we see that the excess risk can be upper bounded in terms of the error on the first iterates of every window, some term that vanishes in $\eta$, and the quantity $\norm{B}^2\epsilon_{\mathsf{geo}}\eta\delta_1$. Recall that $\epsilon_{\mathsf{geo}}$ scales polylogarithmically in $T$, so because we will ultimately take $1/\delta_1$ polylogarithmic in $T$, the contribution of this term will be negligible.
To prove Lemma~\ref{lem:outer2}, we will use the following analogue of Lemma~\ref{lem:kkm}.
\begin{lemma}\label{lem:kkm2}
Let $\psE{\cdot}$ be the solution to Program~\ref{program:sosv2}, assuming it is feasible. Let $\wh{x}_i \triangleq \psE{x_i}$ and $\wh{w}_i \triangleq \psE{w_i}$ for every $0\le i < T$. Provided the event of Lemma~\ref{lem:feasible2} holds, then
\begin{multline}
\frac{1}{T}\left(\sum_{i=0}^{T-1}\left(a^*_i \| B\wh{x}_i - y_i\|^2/\tau^2 + \norm{\wh{w}_i}^2/\sigma^2\right) + \|\hat{x}_0\|^2/R^2\right) - \mathsf{OPT} \le \\
\psE*{\frac{1}{T}\sum_{i=0}^{T-1} (1 - a_i)\| B(x_i - x^*_i)\|^2/\tau^2} + O(mk\cdot \eta^{1-2/k}). \label{eq:firstopt2}
\end{multline}
\end{lemma}
\begin{proof}
The proof is essentially identical to that of Lemma~\ref{lem:kkm}. The only difference is in the step where we carry out the bound in \eqref{eq:av}. Previously in the proof of Lemma~\ref{lem:kkm}, we used Constraints~\ref{item:many_inliers} and \ref{item:measurements_noise} from Program~\ref{program:sos}. Instead, in Program~\ref{program:sosv2} we have directly included the desired bound on $\sum^{T-1}_{i=0} (1-a_i)\norm{v_i}^2$ via Constraint~\ref{constraint2:measurements_noise_corrupted}.
\end{proof}
We now proceed with the proof of Lemma~\ref{lem:outer2}.
\begin{proof}[Proof of Lemma~\ref{lem:outer2}]
Lemma~\ref{lem:kkm2} reduces upper bounding the excess risk achieved by $\brc{\wh{x}_i}$, $\brc{\wh{w}_i}$ to bounding the main term $\psE{\frac{1}{T}\sum^{T-1}_{i=0}(1 - a_i)\norm{B(x_i - x^*_i)}^2/\tau^2}$ in \eqref{eq:firstopt2}, which we do now. Writing $1 - a_i = b_{\ell(i)}\cdot (1 - a_i) + (1 - b_{\ell(i)})\cdot (1 - a_i)$, we have \begin{equation}
\sum^{T-1}_{i=0} (1 - a_i)\norm{B(x_i - x^*_i)}^2 = \sum^{T-1}_{i=0} \brk*{b_{\ell(i)}(1 - a_i)\norm{B(x_i - x^*_i)}^2 + (1 - b_{\ell(i)})(1-a_i)\norm{B(x_i - x^*_i)}^2}. \label{eq:breakup_bs}
\end{equation}
We first control the latter sum on the right-hand side of \eqref{eq:breakup_bs} using Constraint~\ref{constraint2:funnel} and the confidence band guarantee of Corollary~\ref{cor:warmstart}.
For every $0\le i < T$, we have
\begin{equation}
\norm{x_i - x^*_i}^2 \le 2\norm{x_i - x'_i}^2 + 2\norm{x'_i - x^*_i}^2 \lesssim \epsilon_{\mathsf{geo}} + \rho^2 R^2(d + \log(1/\delta))/2^{\ell(i)},
\end{equation}
so
\begin{multline}
\sum^{T-1}_{i=0} (1 - b_{\ell(i)})(1-a_i)\norm{B(x_i - x^*_i)}^2 \le \\ \norm{B}^2\epsilon_{\mathsf{geo}}\sum^{T-1}_{i=0}(1 - b_{\ell(i)})(1 - a_i) + \norm{B}^2\rho^2 R^2(d +\log(1/\delta))\sum^{T-1}_{i=0} (1 - b_{\ell(i)})(1 - a_i)/2^{\ell(i)/2}. \label{eq:abB}
\end{multline}
By Constraint~\ref{constraint2:many_ab}, we can bound the first term on the right-hand side by $\norm{B}^2\epsilon_{\mathsf{geo}}\eta\delta_1 T$. For the second term, note that
\begin{equation}
\frac{1}{T}\sum^{T-1}_{i=0} (1 - b_{\ell(i)})(1 - a_i)/2^{\ell(i)/2} = (\eta\delta_1)^{1/2}\cdot \left(\frac{1}{T}\sum^{T-1}_{i=0} 2^{-\ell(i)}\right)^{1/2} \le \left(\frac{\eta\delta_1}{T/t}\right)^{1/2}.
\end{equation}
We may thus upper bound \eqref{eq:abB} by
\begin{equation}
\norm{B}^2\epsilon_{\mathsf{geo}} \eta\delta_1 + \norm{B}^2\rho^2 R^2(d + \log(1/\delta))\cdot \left(\frac{\eta\delta_1}{T/t}\right)^{1/2}.
\end{equation}
We now turn to the first sum on the right-hand side of in \eqref{eq:breakup_bs}. For these terms, where $b_{\ell(i)} = 1$, we will eschew the guarantee of Corollary~\ref{cor:warmstart} in favor of a finer estimate.
Using Fact~\ref{fact:unroll}, for any $i = \ell t + j$ we can write $B(x_i - x^*_i) = BA^j(x_{\ell t} - x^*_{\ell t}) + \sum^j_{s=1} BA^{j - s}(w_{\ell t + s} - w^*_{\ell t + s})$, so by triangle inequality we can upper bound the first sum on the right-hand side of \eqref{eq:breakup_bs} by \begin{align}
3\sum_{\ell=0}^{T/t-1}b_{\ell} \sum_{j=0}^{t-1} (1 - a_{\ell t + j}) \left(\norm*{BA^j(x_{\ell t} - x^*_{\ell t})}^2 + \norm*{\sum^j_{s=0}BA^{j-s}w_{\ell t + s}}^2 + \norm*{\sum^j_{s=0}BA^{j-s}w^*_{\ell t +s}}^2\right) \label{eq:subsample_and_noise2}
\end{align}
We can control the two noise terms on the right by using $b_{\ell} \le 1$, Constraint~\ref{constraint2:totalnoise}, and Corollary~\ref{cor:totalnoise_feasible} to get
\begin{multline}
\frac{1}{T}\sum^{T/t - 1}_{\ell = 0}b_{\ell} \sum^{t-1}_{j=0}(1 - a_{\ell t + j})\left(\norm*{\sum^j_{s=0}BA^{j-s}w_{\ell t + s}}^2 + \norm*{\sum^j_{s=0}BA^{j-s}w^*_{\ell t +s}}^2\right) \\
\le O\left(\eta^{1-2/k}\cdot \t\alpha\sigma^2\rho^2 m k \right). \label{eq:noise_term}
\end{multline}
For the remaining sum in \eqref{eq:subsample_and_noise2}, we invoke Constraint~\ref{constraint2:subsample} and the bound on $\norm{\calO_t}$ in \eqref{eq:opnormOt} to get
\begin{equation}
\frac{1}{T}\sum^{T/t-1}_{\ell=0}b_{\ell}\sum^{t-1}_{j=0}(1 - a_{\ell t + j})\norm{BA^j(x_{\ell t} - x^*_{\ell t})}^2 \le \eta\left(\alpha\rho^2 + O\left( \rho^2\norm{B}^2\sqrt{\log(t/\delta_1)/t}\right)\right)\cdot \frac{1}{T/t}\sum^{T/t-1}_{\ell=0} b_{\ell} \norm{x_{\ell t} - x^*_{\ell t}}^2,
\end{equation}
from which the lemma follows by substituting this and \eqref{eq:noise_term} into \eqref{eq:subsample_and_noise2}.
\end{proof}
\subsection{Decay of Unobservable Subspace}
\label{sec:decay}
In this section we revisit the analysis from Section~\ref{sec:decay}. It turns out that essentially all of the steps from that section carry over to our setting once we incorporate the $b_{\ell}$ variables. As the proofs here are otherwise essentially identical to their counterparts in Section~\ref{sec:decay}, we will only state the relevant lemmas and defer the proofs to Appendix~\ref{app:defer}.
As in Section~\ref{sec:old_decay}, let $\Pi$ denote the projection to the \emph{observable subspace}, that is, to the subspace of $v\in\R^d$ for which $v^{\top} \calO_t v \ge \zeta$ for $\zeta \triangleq \frac{\kappa \t}{40000\rho^4}$.
We will use the following matrix concentration bound which is the analogue of Lemma~\ref{lem:anticonc}:
\begin{lemma}\label{lem:anticonc_2}
Suppose $\eta < 0.49$ and $t\ge \wt{\Omega}\left(\kappa^{-2} \rho^{12} \norm{B}^4\log(d/\delta_1)\right)$. Then with probability at least $1 - \delta$ over the randomness of $\brc{a^*_i}$, we have that for all windows $0 \le \ell < T/t$, there is a degree-4 SoS proof of the psd inequality
\begin{equation}
b_\ell b^*_{\ell} \sum^{t-1}_{i = 0} a^*_{\ell t + i} a_{\ell t + i} \Pi (A^i)^{\top} B^{\top} B A^i \Pi \succeq b_\ell b^*_\ell \frac{1}{100}\Pi\calO_t \Pi. \label{eq:psdlb2}
\end{equation}
using the constraints of Program~\ref{program:sosv2}.
\end{lemma}
\noindent The main result of this section, analogous to Lemma~\ref{lem:error_decay}, says that provided the event of Lemma~\ref{lem:anticonc} holds, our prediction error on the first iterate of every window decays exponentially over time:
\begin{lemma}\label{lem:error_decay2}
Let pseudoexpectation $\psE{\cdot}$ be the solution to Program~\ref{program:sosv2}, assuming it is feasible. Provided the event of Lemma~\ref{lem:anticonc_2} holds, we have
\[ \psE*{b_{\ell - 1} b^*_{\ell - 1} \|x_{\ell t} - x^*_{\ell t}\|^2} \le \psE*{b_{\ell - 1} b^*_{\ell - 1} \cdot \frac{1}{2}\norm{\Pi^{\perp} q}^2 + 4\gamma_{\ell} + 4\gamma^*_{\ell} + \frac{2}{15\rho^2\zeta}\sum^t_{i=1}(\epsilon_{\ell,i} + \epsilon^*_{\ell,i})}, \]
where
\begin{equation}
\gamma_{\ell} \triangleq \norm*{\sum^t_{i=1} A^{t-i} w_{\ell t + i}}^2, \qquad \gamma^*_{\ell} \triangleq \norm*{\sum^t_{i=1} A^{t-i} w^*_{\ell t + i}}^2,
\end{equation}
\begin{equation}
\epsilon_{\ell,i} \triangleq \left(3\norm{v_{\ell t +i}}^2 + 6\norm{\sum^i_{s=1}BA^{i-s}w_{\ell t + s}}^2\right), \qquad \epsilon^*_{\ell,i} \triangleq \left(3\norm{v^*_{\ell t +i}}^2 + 6\norm{\sum^i_{s=1}BA^{i-s}w^*_{\ell t + s}}^2\right).
\end{equation}
\end{lemma}
To prove this, we will use Lemma~\ref{lem:unobservable-decay} from Section~\ref{sec:logT} which is just a statement about vectors and the observable subspace. The other key ingredient, analogous to Lemma~\ref{lem:observable-case-analysis}, is the following:
\begin{lemma}\label{lem:observable-case-analysis2}
Let pseudoexpectation $\psE{\cdot}$ be the solution to Program~\ref{program:sosv2}, assuming it is feasible. Provided the event of Lemma~\ref{lem:anticonc_2} holds, then at least one of the following holds for every window $0 \le \ell < T/t$ for $q \triangleq x_{\ell t} - x^*_{\ell t}$:
\begin{enumerate}
\item (Observable component decays)
\begin{equation}
\psE*{b_{\ell} b^*_{\ell}\norm{A^{t}\Pi q}^2}
\le \frac{1}{10}\psE*{b_{\ell} b^*_{\ell} \| \Pi q) \|^2}.
\end{equation}
\item (Observable error bounded by unobservable error)
\begin{equation}
\psE*{b_{\ell} b^*_{\ell} \norm{\Pi q}^2} \le \psE*{\frac{1}{10\rho^2}b_{\ell} b^*_{\ell} \norm{\Pi^{\perp} q}^2 + \frac{1}{30\rho^2\zeta}\sum^t_{i=1}(\epsilon_{\ell,i} + \epsilon^*_{\ell,i})}.
\end{equation}
\end{enumerate}
\end{lemma}
\subsection{Error on First Iterates in Each Window}
\label{sec:first_iters_loglogT}
Recall from Lemma~\ref{lem:outer2} that we need to upper bound the pseudoexpectation of the average error on the first iterate in each window. In this section we prove the following bound:
\begin{lemma}\label{lem:error-first-iterate-sos2}
Provided $T \ge \Omega(\Max{\brc*{(t/\eta\delta_1)\log(1/\delta)}}{\brc*{t^2\log^k(1/\delta)}})$, with probability at least $1 - \delta$ we have
\iffalse
\begin{multline}
\frac{1}{T/t}\sum^{T/t-1}_{\ell =0} \psE*{\norm{x_{\ell t} - x^*_{\ell t}}^2} \lesssim \t\sigma^2\rho^2 m k\cdot \left(\Max{1}{(1/\sqrt{T})^{2/k}\log(1/\delta)}\right) + \\ \alpha\kappa^{-1}\rho^4\sigma^2 m \cdot \left(\Max{1}{(\t/\sqrt{T})\log(1/\delta)}\right) + \delta_1\epsilon_{\mathsf{geo}} +\frac{\rho^2 R^2(d + \log(1/\delta))}{T/t}.\label{eq:firstiters_2}
\end{multline}
\fi
\begin{align*}
\frac{1}{T/t}\sum^{T/t - 1}_{\ell = 0} \psE*{\norm{x_{\ell t} - x^*_{\ell t}}^2} \lesssim
\kappa^{-1} \t \alpha \sigma^2 \rho^4m + \kappa^{-1}\tau^2\rho^2 m + \t\sigma^2\rho^2 d +
\delta_1\epsilon_{\mathsf{geo}} +\frac{\rho^2 R^2(d + \log(1/\delta))}{T/t}.\label{eq:firstiters_2}
\end{align*}
\end{lemma}
\begin{proof}
We will use a naive bound on the first summand $\frac{1}{T/t}\psE{\norm{x_0 - x^*_0}^2}$ by using the fact that $x^*_0 \sim\calN(0,R^2\cdot\Id_d)$ and proceed to upper bound the contribution of windows $1$ to $T/t-1$.
For these terms, we will use the elementary inequality
\begin{equation}
1 \le b_{\ell - 1}b^*_{\ell-1} + \frac{3}{2}(1 - b_{\ell - 1}) + \frac{3}{2}(1 - b^*_{\ell - 1})
\end{equation} and upper bound the left hand side of our desired inequality by three sums.
For the first sum, we recall from Lemma~\ref{lem:error_decay2}
that for any $\ell > 1$,
\[ \psE*{b_{\ell - 1} b^*_{\ell - 1} \|x_{\ell t} - x^*_{\ell t}\|^2} \le \psE*{b_{\ell - 1} b^*_{\ell - 1} \cdot \frac{1}{2}\norm{\Pi^{\perp} q}^2 + 4\gamma_{\ell} + 4\gamma^*_{\ell} + \frac{2}{15\rho^2\zeta}\sum^t_{i=1}(\epsilon_{\ell,i} + \epsilon^*_{\ell,i})}, \]
where $q = x_{(\ell - 1)t} - x^*_{(\ell - 1)t}$, to get
\begin{align}
\MoveEqLeft \psE*{\frac{1}{T/t}\sum^{T/t-1}_{\ell =1} b_{\ell - 1} b^*_{\ell - 1} \norm{x_{\ell t} - x^*_{\ell t}}^2} \\
&\le \psE*{\frac{1}{T/t}\sum^{T/t-1}_{\ell =1} b_{\ell - 1} b^*_{\ell - 1} \cdot \frac{1}{2}\norm{\Pi^{\perp} q}^2 + \frac{1}{T/t}\sum^{T/t-1}_{\ell =1} \left(4\gamma_{\ell} + 4\gamma^*_{\ell} + \frac{2}{15\rho^2\zeta}\sum^t_{i=1}(\epsilon_{\ell,i} + \epsilon^*_{\ell,i})\right)} \\
&\le \psE*{\frac{1}{T/t}\sum^{T/t-1}_{\ell =2} \frac{1}{2}\norm{q}^2 + \frac{1}{T/t}\sum^{T/t-1}_{\ell =1} \left(4\gamma_{\ell} + 4\gamma^*_{\ell} + \frac{2}{15\rho^2\zeta}\sum^t_{i=1}(\epsilon_{\ell,i} + \epsilon^*_{\ell,i})\right)}.\label{eq:bb}
\end{align}
For the second sum, we have
\begin{align}
\MoveEqLeft\psE*{\frac{1}{T/t}\sum^{T/t-1}_{\ell =1} (1 - b_{\ell - 1}) \norm{x_{\ell t} - x^*_{\ell t}}^2} \\
&\le \psE*{\frac{2}{T/t}\sum^{T/t-1}_{\ell =1}(1 - b_{\ell - 1})\left(\norm{x_{\ell t} - x'_{\ell t}}^2 + \norm{x'_{\ell t} - x^*_{\ell t}}^2\right)} \\
&\le \psE*{\frac{2}{T/t}\sum^{T/t-1}_{\ell =1}(1 - b_{\ell - 1})\cdot 2(\epsilon_{\mathsf{geo}} + O(\rho^2R^2(d+\log(1/\delta))/2^{\ell(i)}))} \\
&\lesssim \delta_1 \epsilon_{\mathsf{geo}} + \frac{\rho^2 R^2(d + \log(1/\delta))}{T/t},\label{eq:oneminusb}
\end{align} where in the last step we used Constraint~\ref{constraint2:funnel} and Corollary~\ref{cor:warmstart}. We can bound the third sum in an identical fashion, as $\frac{1}{T/t}\sum^{T/t-1}_{\ell =1}(1 - b^*_{\ell - 1}) \le \delta_1$ by Bernstein's. Specifically, we get
\begin{equation}
\psE*{\frac{1}{T/t}\sum^{T/t-1}_{\ell =1} (1 - b^*_{\ell - 1}) \norm{x_{\ell t} - x^*_{\ell t}}^2} \lesssim \delta_1 \epsilon_{\mathsf{geo}} + \frac{\rho^2 R^2(d + \log(1/\delta))}{T/t} \label{eq:oneminusbstar}
\end{equation}
Putting \eqref{eq:bb}, \eqref{eq:oneminusb}, and \eqref{eq:oneminusbstar} together, we conclude that
\begin{multline}
\frac{1}{T/t}\sum^{T/t-1}_{\ell =1} \psE*{\norm{x_{\ell t} - x^*_{\ell t}}^2} \le O\left(\delta_1 \epsilon_{\mathsf{geo}} + \frac{\rho^2 R^2(d + \log(1/\delta))}{T/t}\right) + \psE*{\frac{1}{T/t}\sum^{T/t-1}_{\ell =1} \frac{1}{2} \norm{x_{(\ell - 1)t} - x^*_{(\ell - 1)t}}^2} \\
+ \psE*{ \frac{1}{T/t}\sum^{T/t-1}_{\ell =0} \left(4\gamma_{\ell} + 4\gamma^*_{\ell} + \frac{2}{15\rho^2\zeta}\sum^t_{i=1}(\epsilon_{\ell,i} + \epsilon^*_{\ell,i})\right)},
\end{multline}
so rearranging we get
\begin{multline}
\frac{1}{T/t}\sum^{T/t-1}_{\ell =1}\frac{1}{2}\norm{x_{\ell t} - x^*_{\ell t}}^2 \le \frac{1}{2}\cdot\frac{1}{T/t}\norm{x_0 - x^*_0}^2 + O\left(\delta_1 \epsilon_{\mathsf{geo}} + \frac{\rho^2 R^2(d + \log(1/\delta))}{T/t}\right) \\
+ \psE*{ \frac{1}{T/t}\sum^{T/t-1}_{\ell =0} \left(4\gamma_{\ell} + 4\gamma^*_{\ell} + \frac{2}{15\rho^2\zeta}\sum^t_{i=1}(\epsilon_{\ell,i} + \epsilon^*_{\ell,i})\right)} \label{eq:firstiters_intermed}
\end{multline}
It remains to control the $\epsilon$'s and $\gamma$'s. Taking $k = 2$ in Lemma~\ref{lem:triplesum}, we get
\[ \frac{1}{T}\sum^{T/\t - 1}_{\ell = 0}\sum^{t-1}_{b = 0} \norm*{\sum^b_{i=1}BA^{b-i} w^*_{\ell\t+i}}^2 \le O\left(\t\alpha\sigma^2\rho^2 m\right), \]
so we have from this, Constraint~\ref{constraint2:totalnoise}, Constraint~\ref{constraint2:measurements_noise}, and Lemma~\ref{lem:normbound} that
\begin{align}
\MoveEqLeft\frac{1}{T} \sum_{\ell = 0}^{T/t - 1} \sum^t_{i = 1} (\epsilon_{\ell,i} + \epsilon^*_{\ell,i}) \\
&= \frac{1}{T} \sum_{\ell = 0}^{T/t - 1} \sum^t_{i = 1} \left(3\norm{v_{\ell t +i}}^2 + 6\norm*{\sum^i_{s=1}BA^{i-s}w_{\ell t + s}}^2 + 3\norm*{v^*_{\ell t +i}}^2 + 6\norm{\sum^i_{s=1}BA^{i-s}w^*_{\ell t + s}}^2\right) \\
&\le O\left(\t\alpha\sigma^2\rho^2 m + \tau^2(m + \log(2/\delta)/T)\right)\label{eq:epsbound}
\end{align}
To bound the $\gamma$'s, we invoke Lemma~\ref{lem:triplesum_dumb}
to get that
\begin{equation}
\frac{1}{T}\sum^{T/\t - 1}_{\ell = 0} \gamma^*_{\ell} \le O\left(\sigma^2\rho^2 d \right), \label{eq:upsbound}
\end{equation}
and by Constraint~\ref{constraint2:totalnoise_dumb}, we get the same bound for $\frac{1}{T}\sum^{T/\t - 1}_{\ell = 0} \gamma_{\ell}$.
Putting together \eqref{eq:epsbound} and \eqref{eq:upsbound} we get
\begin{multline}
\psE*{ \frac{1}{T/t}\sum^{T/t-1}_{\ell =0} \left(4\gamma_{\ell} + 4\gamma^*_{\ell} + \frac{2}{15\rho^2\zeta}\sum^t_{i=1}(\epsilon_{\ell,i} + \epsilon^*_{\ell,i})\right)} \\
\lesssim \frac{t}{\rho^2\zeta}\cdot(\t\alpha \sigma^2 \rho^2m + \tau^2(m + \log(2/\delta)/T)) + \t\sigma^2\rho^2 d \lesssim \kappa^{-1} \t \alpha \sigma^2 \rho^4m + \kappa^{-1}\tau^2\rho^2 m + \t\sigma^2\rho^2 d
\end{multline}
where we recalled the definition of $\zeta \triangleq \frac{\kappa \t}{40000\rho^4}$ and used the assumed lower bound on $T$.
Combining with \eqref{eq:firstiters_intermed}, we obtain the claimed bound.
\end{proof}
\subsection{Putting Everything Together}
We can now complete the proof of Theorem~\ref{thm:main_TloglogT} by applying the bounds in Section~\ref{sec:first_iters_loglogT} to Lemma~\ref{lem:outer2}.
\begin{proof}[Proof of Theorem~\ref{thm:main_TloglogT}]
Let \begin{equation}
t = \Max{s}{\wt{\Theta}\left(\kappa^{-2} \rho^{12} \norm{B}^4\log(d/\delta_1)\right)} \label{eq:loglogTtdef}
\end{equation} as in Lemma~\ref{lem:anticonc_2}.
First, recall from Lemma~\ref{lem:outer2} that
\begin{align*}
\MoveEqLeft\frac{1}{T}\sum^{T-1}_{i=0}\left(a^*_i \norm{B\wh{x}_i - y_i}^2/\tau^2 + \norm{\wh{w}_i}^2/\sigma^2\right) + \frac{\|\wh{x}_0\|^2}{R^2 T} - \mathsf{OPT} \\
&\lesssim \tau^{-2}\left\{\eta \rho^2 \left(\alpha +\norm{B}^2\sqrt{\log(t/\delta_1)/t}\right)\cdot \frac{1}{T/t}\sum^{T/t - 1}_{\ell=0} \psE*{\norm{x_{\ell t} - x^*_{\ell t}}^2} \right.\\
&\quad \left. + k\eta^{1-2/k}\cdot \left(\t\alpha\sigma^2\rho^2 m + m\tau^2\right) + \norm{B}^2\left(\epsilon_{\mathsf{geo}}\eta\delta_1 + \rho^2R^2(d+\log(1/\delta))\cdot \left(\frac{\eta\delta_1}{T/t}\right)^{1/2}\right)\right\}
\end{align*}
and recall from Lemma~\ref{lem:error-first-iterate-sos2} that
\begin{equation*}
\frac{1}{T/t}\sum^{T/t-1}_{\ell =0} \psE*{\norm{x_{\ell t} - x^*_{\ell t}}^2} \lesssim \kappa^{-1} \t \alpha \sigma^2 \rho^4m + \kappa^{-1}\tau^2\rho^2 m + \t\sigma^2\rho^2 d + \delta_1\epsilon_{\mathsf{geo}} +\frac{\rho^2 R^2(d + \log(1/\delta))}{T/t}
\end{equation*}
so combining the two bounds gives
\begin{multline}
\frac{1}{T}\sum^{T-1}_{i=0}\left(a^*_i \norm{B\wh{x}_i - y_i}^2/\tau^2 + \norm{\wh{w}_i}^2/\sigma^2\right) + \frac{\|\wh{x}_0\|^2}{R^2 T} - \mathsf{OPT} \\
\lesssim \tau^{-2}\Bigg\{\eta \rho^2 \left(\alpha +\norm{B}^2\sqrt{\log(t/\delta_1)/t}\right)\cdot \Bigg[ \kappa^{-1} \t \alpha \sigma^2 \rho^4m + \kappa^{-1}\tau^2\rho^2 m + \t\sigma^2\rho^2 d \\
+\delta_1\epsilon_{\mathsf{geo}} +\frac{\rho^2 R^2(d + \log(1/\delta))}{T/t}\Bigg] + k\eta^{1-2/k}\cdot \left(\t\alpha\sigma^2\rho^2 m + m\tau^2\right) \\
+ \norm{B}^2\left(\epsilon_{\mathsf{geo}}\eta\delta_1 + \rho^2R^2(d+\log(1/\delta))\cdot \left(\frac{\eta\delta_1}{T/t}\right)^{1/2}\right)\Bigg\} \label{eq:ugly}
\end{multline}
Note that for the value of $\t$ we chose in \eqref{eq:loglogTtdef}, we have
\begin{align}
\log(t/\delta_1)/t
&= \kappa^2 \rho^{-12} \norm{B}^{-4} \frac{\log(\kappa^{-2} \rho^{12} \norm{B}^4) + \log \log(d/\delta_1) + \log(1/\delta_1)}{\log(d/\delta_1)} \\
&\le \kappa^2 \rho^{-12} \norm{B}^{-4}(2 + \log(\kappa^{-2} \rho^{12} \norm{B}^4)/\log(d/\delta_1)) \le O(\kappa^2\rho^{-12}\norm{B}^{-4}).\label{eq:logtovertbound}
\end{align}
The requirement for Lemmas~\ref{lem:bstar_conc} and \ref{lem:many_ab} is that $\delta_1 =\Omega((t/\eta T) \log(1/\delta))$, which by our choice of $t$ translates to $\delta_1 = \Omega(\kappa^{-2} \rho^{12} \|B\|^4 \log(d/\delta_1)\log(1/\delta)/\eta T)$, or equivalently
\begin{equation}
\delta_1/\log(d/\delta_1) = \Omega(\kappa^{-2} \rho^{12} \|B\|^4 \log(1/\delta)/\eta T) \label{eq:delta1def}
\end{equation}
and then taking \begin{equation}
\delta_1 = \Theta(\log(1/\delta)/\log^3 T)
\end{equation} satisfies this condition provided
$T = \wt{\Omega}(\eta^{-1}\kappa^{-2}\rho^{12}\norm{B}^4\log d)$. Note that this choice of $\delta_1$ yields \begin{equation}
\t = \Max{s}{\wt{\Theta}(\kappa^{-2} \rho^{12} \|B\|^4 \log(d\log(T)/\log(1/\delta)))}. \label{eq:finaltbound}
\end{equation}
In particular, $\t$ is \emph{doubly logarithmic} in $T$.
Plugging in \eqref{eq:logtovertbound} and \eqref{eq:delta1def} into \eqref{eq:ugly} gives
\begin{align*}
&\frac{1}{T}\sum^{T-1}_{i=0}\left(a^*_i \norm{B\wh{x}_i - y_i}^2/\tau^2 + \norm{\wh{w}_i}^2/\sigma^2\right) + \frac{\|\wh{x}_0\|^2}{R^2 T} - \mathsf{OPT} \\
&\lesssim \tau^{-2}\left\{\eta \rho^2 \left(\alpha + \kappa \rho^{-6} \right)\cdot \left( \kappa^{-1} \t \alpha \sigma^2 \rho^4m + \kappa^{-1}\tau^2\rho^2 m + \t\sigma^2\rho^2 d + \frac{\log(1/\delta)}{\log^3 T} \epsilon_{\mathsf{geo}} +\frac{\rho^2 R^2(d + \log(1/\delta))}{T/t}\right)\right.\\
&\quad \left. + k\eta^{1-2/k}\cdot \left(\t\alpha\sigma^2\rho^2 m + m\tau^2\right) + \norm{B}^2\left(\epsilon_{\mathsf{geo}}\eta\frac{\log(1/\delta)}{\log^3 T} + \rho^2R^2(d+\log(1/\delta))\cdot \left(\frac{\eta\delta_1}{T/t}\right)^{1/2}\right)\right\}.
\end{align*}
Recall from \eqref{eq:eps_geo2} that
\begin{equation}
\epsilon_{\mathsf{geo}} = O\left(\frac{\rho^8 t_{\mathsf{pre}}}{\kappa}\cdot \brk*{\tau^2\left(m + \log(T/\delta)\right) + \sigma^2 \left(d +\log(T/\delta)\right)\cdot \rho^2 t_{\mathsf{pre}} \norm{B}^2}\right)
\end{equation}
and $t_{\mathsf{pre}} = \Max{s}{\wt{\Theta}(\kappa^{-2}\rho^{12}\norm{B}^4\log(dT/\delta))}$. Crucially, we see that the factor of $\log^3 T$ in $\epsilon_{\mathsf{geo}}$ is canceled out by the $1/\log^3 T$ in our bound coming from the choice of $\delta_1$. We see that
\begin{multline}
\frac{1}{T}\sum^{T-1}_{i=0}\left(a^*_i \norm{B\wh{x}_i - y_i}^2/\tau^2 + \norm{\wh{w}_i}^2/\sigma^2\right) + \frac{\|\wh{x}_0\|^2}{R^2 T} - \mathsf{OPT} \\
\lesssim \poly(s,\alpha,1/\kappa,\rho,\|B\|,\log d,\log(1/\delta)) \brk*{ k\eta^{1-2/k}\left(m + d(\sigma^2/\tau^2)\log\log T\right) + \eta^{1/2}(R^2 d/\tau^2)T^{-0.49}}
\end{multline}
By taking $k = \floor{2\log(1/\eta)}$ so that $k\eta^{1-2/k} = 2e\eta\log(1/\eta)$, we obtain the claimed bound.
\end{proof}
\section{Deferred Proofs from Section~\ref{sec:TloglogT}}
\label{app:defer}
In this section we collect deferred proofs from Section~\ref{sec:TloglogT}.
\subsection{Proof of Lemma~\ref{lem:anticonc_2}}
\begin{proof}
Without loss of generality we can assume $\ell = 0$. By definition of $b^*_0$, we have \begin{equation}
b_0 b^*_0 \sum^t_{i = 1} a^*_i \Pi (A^i)^{\top} B^{\top} B A^i \Pi \succeq b_0 b^*_0 \left((1 - \eta)\cdot \Pi\calO_t\Pi - O\left(\rho^2\norm{B}^2\sqrt{t\log(d/\delta_1)}\right)\cdot \Pi\right),
\end{equation} and similarly by Constraint~\ref{item:subsample}, we have that \begin{equation}
b_0 b^*_0\sum^t_{i = 1} (1 - a_i) \Pi (A^i)^{\top} B^{\top} B A^i \Pi \preceq b_0 b^*_0\left(\eta\cdot \Pi\calO_t \Pi + O\left(\rho^2\norm{B}^2\sqrt{t\log(t/\delta_1)}\right) \cdot \Pi\right).
\end{equation} Writing $a^*_i a_i = a^*_i - a^*_i(1 - a_i) \ge a^*_i - (1 - a_i)$ and subtracting the above two equations gives
\begin{equation}
b_0 b^*_0 \sum^t_{i = 1} a^*_ia_i \Pi (A^i)^{\top} B^{\top} B A^i \Pi \succeq b_0 b^*_0 \left((1 - 2\eta)\cdot \Pi\calO_t\Pi - O\left(\rho^2\norm{B}^2\sqrt{t\log(t/\delta_1)}\right)\cdot \Pi\right).
\end{equation}
Recalling that $\Pi \preceq \zeta^{-1} \cdot \Pi\calO_t\Pi$ by definition of $\Pi$, we would like $t$ to be large enough that \begin{equation}
1 - 2\eta - O\left(\zeta^{-1}\rho^2\norm{B}^2\sqrt{t\log(t/\delta_1)}\right) \ge 1/3.
\end{equation}
As $\eta < 0.49$, it suffices for $\zeta \ge O(\rho^2 \norm{B}^2\sqrt{t\log(t/\delta_1)})$.
Recalling that $\zeta = \frac{\kappa \t}{40000\rho^4}$, it's sufficient that $t \ge \wt{\Omega}\left(\kappa^{-2} \rho^{12} \norm{B}^4 \log(d/\delta1)\right)$.
\end{proof}
\subsection{Proof of Lemma~\ref{lem:observable-case-analysis2}}
\begin{proof}
Without loss of generality we can assume $\ell = 0$. In addition to the lower bound of \eqref{eq:psdlb2}, we also have a degree-2 SoS proof of the upper bound
\begin{equation}\label{eqn:unobs-psd2}
\sum_{i=1}^t a^*_i a_i \Pi^\perp (A^i)^T B^T B A^i \Pi^\perp \preceq \Pi^\perp \mathcal{O}_t \Pi^\perp \preceq \zeta \cdot I,
\end{equation}
where in the first step we used that $a_i a^*_i \le 1$ by Constraint~\ref{item:boolean} and in the second step we used the definition of $\Pi$.
For convenience, define $q\triangleq x_0 - x^*_0$. We proceed by casework on whether there is a gap between $\psE*{(\Pi q)^{\top} \calO_t (\Pi q)}$ and $\psE{\zeta\norm{\Pi q}^2}$:
\noindent{\textbf{Case 1}}: $\psE*{b_0 b^*_0\| \Pi q \|^2} \ge \psE*{\frac{1}{4000\rho^2\zeta}b_0 b^*_0\sum_{i=1}^t \| BA^i \Pi q\|^2}$.
The analysis for this case is very similar to the analysis in Lemma~\ref{lem:unobservable-decay}. We have
\begin{align*}
\psE*{b_0b^*_0\| \Pi q \|^2}
&\geq \psE*{b_0b^*_0\frac{1}{4000\rho^2\zeta}\sum_{i=1}^t \| BA^i \Pi q\|^2} \\
&= \psE*{b_0b^*_0\frac{1}{4000\rho^2\zeta}\sum_{j = 1}^{t/s} q^T\Pi {A^{js}}^T \calO_s A^{js} \Pi q} \\
&\geq \psE*{b_0b^*_0\frac{10\rho^2 s}{t}\sum_{j = 1}^{t/s} \|A^{js} \Pi q\|^2},
\end{align*}
where in the last step we used the definition of $\zeta$ and the assumption that $\sigma_{\min}(\calO_s) \ge \kappa s$. Rearranging, we obtain
\begin{equation}
\frac{1}{10\rho^2} \psE{b_0 b^*_0\| \Pi q \|^2} \geq \frac{1}{t/s}\sum_{j=1}^{t/s} \psE*{b_0 b^*_0 \|A^{js} \Pi q\|^2}.
\end{equation}
Therefore, there exists some index $j\in[t/s]$ for which $\psE*{b_0 b^*_0\norm{A^{js}\Pi q}^2} \le \frac{1}{10\rho^2}\psE*{b_0 b^*_0 \norm{\Pi q}^2}$. By uniform stability, we obtain the first desired outcome in the lemma statement.
\noindent{\textbf{Case 2}}: $ \psE*{b_0 b^*_0 \| \Pi q \|^2} \le \psE*{\frac{1}{4000\rho^2\zeta} b_0 b^*_0\sum_{i=1}^t \| BA^i \Pi q\|^2}$.
We have
\begin{equation} \label{eqn:fix-main2}
\psE*{b_0b^*_0\| \Pi q \|^2} \leq \psE*{\frac{1}{4000\rho^2\zeta}b_0 b^*_0\sum_{i=1}^t\| BA^i \Pi q\|^2} \leq \psE*{b_0b^*_0\frac{1}{40\rho^2\zeta}\sum_{i=1}^t a^*_i a_i\| BA^i \Pi q\|^2}.
\end{equation}
where in the last step we invoked Lemma~\ref{lem:anticonc_2}.
We would like to upper bound $b_0b^*_0\sum^t_{i=1} a^*_i a_i \norm{BA^i\Pi q}^2$ in terms of $b_0 b^*_0\sum^t_{i = 1} a^*_i a_i \norm{BA^i \Pi^{\perp} q}^2$.
For any $i \in[t]$, we have the following sequence of inequalities in degree-4 SoS \begin{align}
a_i a^*_i \norm*{BA^i(\Pi + \Pi^{\perp})q}^2 &= a_i a^*_i \norm*{BA^iq}^2 \\
&= a_i a^*_i\norm*{(y_i - Bx^*_i) - (y_i - Bx_i) + (Bx^*_i - BA^i x^*_0) - (Bx_i - BA^i x_0)}^2 \\
&\le 3\norm{v^*_i}^2 + 3\norm{v_i}^2 + 3\norm*{\sum^i_{s=1} BA^{i - s} (w_s - w^*_s)}^2\\
&\le \left(3\norm{v^*_i}^2 + 6\norm*{\sum^i_{s=1}BA^{i-s}w^*_s}^2\right) + \left(3\norm{v_i}^2 + 6\norm*{\sum^i_{s=1}BA^{i-s}w_s}^2\right)
\end{align}
so by applying Fact~\ref{fact:sos_simple} to $\epsilon = \epsilon_{0,i} + \epsilon^*_{0,i}$, $v_1 = a^*_i a_i BA^i \Pi q $, and $v_2 = a^*_i a_i BA^i \Pi^\perp q $, we have a degree-4 SoS proof of
\begin{equation}
a^*_i a_i\| BA^i \Pi q\|^2 \leq 4a^*_i a_i\| BA^i \Pi^\perp q\|^2 + \frac{4}{3}(\epsilon_{0,i} + \epsilon^*_{0,i}).
\end{equation}
Summing this inequality over $i\in[t]$ and taking pseudo-expectations, we get
\[\psE*{\sum_{i=1}^t a^*_i a_i\| BA^i \Pi q\|^2} \le \psE*{4\sum_{i=1}^ta^*_i a_i\| BA^i \Pi^\perp q\|^2 + \frac{4}{3}\sum^t_{i=1}(\epsilon_{0,i} + \epsilon^*_{0,i})}. \]
Substituting this back into the main bound \eqref{eqn:fix-main2}, we get
\begin{align}
\psE*{b_0 b^*_0 \| \Pi q \|^2} &\leq \psE*{\frac{1}{10\rho^2\zeta}b_0b^*_0 \sum_{i=1}^t a^*_i a_i\| BA^i \Pi^\perp q\|^2 + \frac{1}{30\rho^2\zeta}b_0 b^*_0\sum^t_{i=1}(\epsilon_{0,i} +\epsilon^*_{0,i})} \\
&\le \psE*{\frac{1}{10\rho^2}b_0b^*_0\norm{\Pi^{\perp} q}^2 + \frac{1}{30\rho^2\zeta}\sum^t_{i=1} (\epsilon_{0,i} +\epsilon^*_{0,i})},
\end{align} where in the last step we used \eqref{eqn:unobs-psd2} and also the fact that $b_0 b^*_0 \le 1$. Unpacking the definition of $\epsilon,\zeta$, we arrive at the second desired bound.
\end{proof}
\subsection{Proof of Lemma~\ref{lem:error_decay2}}
\begin{proof}
Define $q \triangleq x_{(\ell - 1)t} - x^*_{(\ell - 1)t}$. Recalling Fact~\ref{fact:unroll}, we have by SoS triangle inequality that:
\begin{align*}
\|x_{\ell t} - x^*_{\ell t}\|^2
&\le 4 \|A^t\Pi q\|^2 + 4\|A^t\Pi^{\perp} q\|^2 + 4\gamma_{\ell} + 4\gamma^*_{\ell} \\
&\le 4 \|A^t\Pi q\|^2 + (1/10000) \|\Pi^{\perp} q\|^2 + 4\gamma_{\ell} + 4\gamma^*_{\ell}
\end{align*}
where we used triangle inequality in the first inequality, SoS triangle inequality in the second inequality, and Lemma~\ref{lem:unobservable-decay} in the third inequality.
Based on Lemma~\ref{lem:observable-case-analysis2} applied to $\ell - 1$ we either have that:
\begin{enumerate}
\item (Observable component decays.) We have
\[ \psE*{b_{\ell - 1}b^*_{\ell - 1}\norm{A^{t}\Pi q}^2} \leq \frac{1}{10} \psE*{b_{\ell - 1}b^*_{\ell - 1} \| \Pi q \|^2}. \]
Then
\begin{align*}
\MoveEqLeft \psE{b_{\ell - 1} b^*_{\ell - 1}\|x_{\ell t} - x^*_{\ell t}\|^2} \\
&\le \psE*{b_{\ell - 1} b^*_{\ell - 1}\cdot\left(4 \|A^t\Pi q\|^2 + (1/10000) \|\Pi^{\perp} q\|^2 \right) + 4\gamma_{\ell} + 4\gamma^*_{\ell}} \\
&\le \psE*{b_{\ell - 1} b^*_{\ell - 1}\cdot \left(\frac{2}{5}\|\Pi q\|^2 + (1/10000) \|\Pi^{\perp} q\|^2\right) + 4\gamma_{\ell} + 4\gamma^*_{\ell}} \\
&\le \psE*{b_{\ell - 1} b^*_{\ell - 1}\cdot \frac{2}{5}\|q\|^2 + 4\gamma_{\ell} + 4\gamma^*_{\ell}}
\end{align*}
where in the last step we used the Pythagorean Theorem.
\item (Observable error bounded by unobservable error.) We have
\[ \psE*{b_{\ell - 1}b^*_{\ell - 1} \norm{\Pi q}^2} \le \psE*{b_{\ell - 1}b^*_{\ell - 1}\cdot \frac{1}{10\rho^2}\norm{\Pi^{\perp} q}^2 + \frac{1}{30\rho^2\zeta}\sum^t_{i=1}(\epsilon_{\ell,i} + \epsilon^*_{\ell,i})} \]
Then
\begin{align*}
\psE{b_{\ell-1} b^*_{\ell-1}\|x_{\ell t} - x^*_{\ell t}\|^2}
&\le \psE*{b_{\ell - 1} b^*_{\ell - 1} \left( 4\|A^t\Pi q\|^2 + (1/10000) \|\Pi^{\perp} q\|^2\right) + 4\gamma_{\ell} + 4\gamma^*_{\ell}} \\
&\le \psE*{b_{\ell - 1} b^*_{\ell - 1}\left(4\rho^2\|\Pi q\|^2 + (1/10000) \|\Pi^{\perp} q\|^2\right) + 4\gamma_{\ell} + 4\gamma^*_{\ell}} \\
&\le \psE*{b_{\ell - 1} b^*_{\ell - 1} \left(\frac{1}{2}\norm{\Pi^{\perp} q}^2 + 4\gamma_{\ell} + 4\gamma^*_{\ell} + \frac{2}{15\rho^2\zeta}\sum^t_{i=1}(\epsilon_{\ell,i} + \epsilon^*_{\ell,i})\right)}.
\end{align*}
Since we showed the desired conclusion in both cases, the proof is complete.
\end{enumerate}
\end{proof}
\section{Some Instructive Examples}\label{apdx:examples}
In this section we present several examples that illustrate important subtleties in the model we consider, and which drove many of the insights underlying our analysis.
\subsection{Noisily Observed Simple Random Walk}
\label{subsec:srw}
Let $m = d = 1$ so that $A$ and $B$ are scalars. Let $A = 1$, $B = 1$, and $\sigma^2 = \tau^2 = 1$ so that the trajectory $\brc{x_i}$ is given by a random walk over $\R$ with Gaussian increments. Note that this model is completely observable with $s = 1$ and uniformly stable with $\rho = 1$. Moreover, observe that the final iterate $x_{T-1}$ will typically satisfy $\|x_{T-1}\| \approx \sqrt{T}$. This illustrates two important phenomena:
\paragraph{Oblivious outlier removal fails.} The most intuitively obvious way to deal with large corruptions is to perform some kind of simple outlier removal, like deleting all observations $y_t$ with $\|y_t\| \ge R$ for some threshold $R$, and then running a standard algorithm like Kalman filtering on whatever remains. In this example, $R$ would have to be at least $\Omega(\sqrt{T})$ in order to avoid removing a large portion of the ground truth observations. It can be straightforwardly checked, via the Kalman filter recursion, that an adversary making $\eta T$ corruptions of size $\Theta(\sqrt{T})$ can cause the Kalman filter to incur excess risk scaling as $\Omega(\eta T)$, which is much larger than our guarantee of $\wt{O}(\eta)\cdot\log\log T$.
The fact that simple outlier removal procedures fail is the key difficulty with dealing with marginally stable dynamical systems like the above. In contrast, in Appendix~\ref{apdx:wiener} we show that if the system is \emph{strictly} stable (i.e. all eigenvalues of $A$ are bounded away from one) then a fairly simple truncated filter will actually have reasonable robustness guaranteees.
\paragraph{Impossibility of handling adversarial corruptions in adversarial locations.}
Suppose that instead of the adversary controlling a random fraction of the timesteps, the adversary is allowed to choose which $\eta T$ timesteps it gets to corrupt. Now suppose that for the timesteps $T - 2\eta T$ up to $T - 1$, the adversary samples an independent random walk $x'_{T - 2\eta T},\ldots,x'_{T-1}$ with Gaussian increments conditioned on $x_{T - 2\eta T} = x'_{T - 2\eta T}$, chooses a random subset of $\eta T$ coordinates from $[T - 2\eta T, T]$ to corrupt, and for each corrupted step $i$ sets $y_i = x'_i + \calN(0,1)$ instead of $y^*_i = x_i + \calN(0,1)$. By construction, it is information-theoretically impossible to distinguish from its observations which of the following is the ground truth: the actual true trajectory $(x_0,\ldots,x_{T-1})$, or the ``fake'' trajectory $(x_0,\ldots,x_{T - 2\eta T}, x'_{T - 2 \eta T + 1},\ldots, x'_{T-1})$. Note that any choice of estimate $\hat{x}_i$ will be distance at least $|x_i - x'_i|/2$ from one of these two trajectories.
By linearity of expectation, the algorithm will pay mean squared error at least $$\frac{1}{T}\sum_{i = T - 2\eta T}^T \frac{1}{2} \mathbb{E} \left(\frac{x_i - x'_i}{2}\right)^2 = \Omega(\eta\cdot \eta T) = \Omega(\eta^2 T)$$ to estimate the trajectory. This is again much larger than the $\wt{O}(\eta)\cdot\log\log T$ guarantee we give in the setting with randomly located corruptions.
Based on this example, we see that it is information-theoretically necessary for the corruption to be \emph{spread out randomly}, and otherwise it is not possible to obtain strong estimation guarantees. It also demonstrates that the assumption $\eta < 1/2$ is required, because the attack above succeeds even with access to random locations when $\eta = 1/2$.
\subsection{Coordinate Cycling Dynamics}
\label{subsec:switching}
The definition of complete observability ensures that with $s$ consecutive uncorrupted observations $y_0,\ldots,y_{s-1}$, we can estimate $x_0$ nontrivially well. Here we give an example showing that if some of these observations are corrupted, one needs to go out to $\omega(s)$ time steps to be able to estimate $x_0$ nontrivially well.
Consider a dynamics where $A(x_1,\ldots,x_d) = (x_d,x_1,\ldots,x_2)$, i.e. $A$ cyclically permutes the coordinates of the state, and the observations are given by the 1-dimensional projection $B(x_1,\ldots,x_d) = x_1$. It is easy to see that this model is observable after $s = d$ steps.
However, in the presence of corruptions with corruption fraction $\eta > 1/d$, we generally cannot estimate the state from just $s = d$ observations. This is because if $\eta > 1/d$, there is a significant probability that one of the first $s$ observations was corrupted, and so we do not have an accurate estimate of $x_0$ in the coordinate corresponding to that timestep. In fact, for any coordinate $j\in[d]$, the distribution over the earliest cycle of $d$ timesteps in which we get a clean observation of $j$ is geometric with parameter $1 - \eta$, so if $\eta$ is a constant, we need to make $\Theta(d\log d)$ observations in order to guarantee that we make a clean observation of every coordinate at least once with probability $2/3$.
\subsection{Subspace That Is Hard to Robustly Observe}
\label{sec:3dexample}
We consider a variant of the previous example which illustrates some additional properties relevant to our analysis. Let
\[ A = \begin{pmatrix}
0 & 1 & 0\\
1 & 0 & 0\\
1 & 1 & 1/2
\end{pmatrix}\]
and define $B:\R^3\to\R^2$ by $B(x_1,x_2,x_3) = (x_1,x_3)$. Observe that
$BA^k (x_1,x_2,x_3)$ is either $(x_1, (1 + 1/2 + \cdots)(x_1 + x_2) + x_3/2^k)$ or $(x_2, (1 + 1/2 + \cdots)(x_1 + x_2) + x_3/2^k)$ depending on the parity of $k$. In particular, observe that $\|BA^k e_3\|^2 = 1/2^{2k}$ which shrinks exponentially fast with $k$, so $\sum_{k = 0}^{\infty} \|BA^k e_3\|^2 = \Theta(1)$ and the sum is dominated by its first few terms. The significance of this is if $x^*_0 = (x^*_{01},x^*_{02},x^*_{03})$
and the first few observations $\tilde{y}_0,\tilde{y}_1,\tilde{y}_2,...$ are corrupted (which can happen with decent probability if $\eta$ is a constant), then almost all of the information about $x^*_{03}$ is lost, even though this is system is observable with parameter $s=2$.
Without delving into more details, we think of the $z$ coordinate as an unobservable subspace i.e a subspace of the state space that can't be estimated once a few timesteps are corrupted.
\paragraph{Unobservable Subspace Can Contain Unbounded Error.}
\label{subsec:unobservableunbounded}
Using the same dynamics $A$ and observation $B$, we have established that starting from $(x_1,x_2,x_3)$ that if the first few observations are corrupted then measurement of $x_3$ is lost at an exponential rate. This would not be a problem if we anticipated the $z$ coordinate to converge quickly to zero so that over a long trajectory it suffices to estimate the $x$ and $y$ coordinates alone. However, this is not true. If $(x_{1}, x_{2}, x_{3}) = (R,R,R)$ the $z$ coordinate can have size scaling with $R$ over the entire length of the trajectory. That is to say, the state vector can have an unbounded component in the unobservable subspace.
\paragraph{Estimating Unobservable Error of the Present from Observable Error of the Past.}
\label{subsec:past2present}
Of course, this does not appear to be an intractable problem for the specific dynamics $A$ and measurement $B$ that we have chosen here. We can give up on estimating the state in the $z$ coordinate directly. Instead, if we can estimate the $x$ and $y$ coordinates at time $t$ we can build a good estimate of the $z$ coordinate at the next timestep $t+1$. In this manner, it is possible to estimate the state in the unobservable subspace, $z$, using estimates of the observable subspace $(x,y)$ of the past.
At first glance, the property that enables us to estimate the state of the unobservable subspace of the present from the observable subspace of the past seems to impose significant constraints on the dynamics $A$ and measurement $B$. Surprisingly, this is not the case. In fact, as our analysis in Sections~\ref{sec:old_decay} and \ref{sec:decay} show, this analysis can be carried out for \emph{any} linear dynamical system that is uniformly stable and completely observable.
\subsection{Lower Bound When \texorpdfstring{$\tau\to 0$}{tau->0}}
\label{sec:sigbytau}
Recall that our guarantee in Theorem~\ref{thm:sos_informal} has a term depending on the ratio $\sigma^2/\tau^2$, suggesting that the excess risk achieved by our estimator blows up as $\tau\to 0$ when the other parameters are fixed. This turns out to be unavoidable. Consider the dynamics given by $A(x,y) = (0,x)$ with $B = \Id$. We may observe conflicting information about the first coordinate of initial state $x_0$ from the first two observations $y_0$ and $y_1$ (and no information about the first coordinate of $x_0$ from subsequent observations). We thus have no way of knowing which of the two observations was corrupted, and so an error of size $\Omega(R^2/\tau^2)$ is unavoidable. For a similar reason, the guarantee needs to blow up if $\tau \to 0$ with $\sigma,\eta > 0$ fixed.
\subsection{Dimension Dependence is Necessary}
\label{app:ddepend}
In this section we describe a simple example showing that the dependence of the excess risk on the ambient dimension $d$ in our main guarantee is unavoidable.
\begin{lemma}
Consider the linear dynamical system where $A = 0$, $B = \Id$, and $\sigma^2 = \tau^2 = R^2 = 1$. If the corrupted observations are given by drawing an independent sample from $\calN(0,2\Id)$, then for any algorithm that takes as input the corrupted observations $\brc{y_i}$ and outputs a trajectory $\brc{\wh{x}_i}$, we have
\begin{equation}
\E{L(\wh{x})}\ge \E{\mathsf{OPT}} + \Omega(\eta d),
\end{equation}
where the expectation is over the randomness of the algorithm, which observations were corrupted, and the process and observation noise.
\end{lemma}
\begin{proof}
In this case, every observation $y_i$ is just a fresh draw from $\calN(0,2\Id)$ obtained either from sampling $x^*_i\sim\calN(0,\Id)$ and adding $v^*_i\sim\calN(0,\Id)$, or from sampling independently of the trajectory. Note that $L(\wh{x})$ can be expressed as
\begin{equation}
L(\wh{x}) = \frac{1}{T}\left(\sum^{T-1}_{i=0}\left(a^*_i\norm{\wh{x}_i - y_i}^2 + \norm{\wh{x}_i}^2\right) + \norm{\wh{x}_0}^2\right).
\end{equation}
We first compute what $\mathsf{OPT}$ would be, i.e. what value one could achieve if the indices $i$ for which $a^*_i = 1$ were known. In that case we would take
\begin{equation}
\wh{x}_i = \begin{cases}
y_i/3 & a^*_i = 1 \ \text{and} \ i = 0 \\
y_i/2 & a^*_i = 1 \ \text{and} \ i > 0 \\
0 & a^*_i = 0
\end{cases}
\end{equation} and conclude that in expectation over the randomness of which observations were corrupted,
\begin{equation}
\E{\mathsf{OPT}} = \E*{\frac{1}{T}\left( \frac{1}{6}a^*_0\norm{y_0}^2 + \sum^{T-1}_{i=0}\frac{1}{2}a^*_i\norm{y_i}^2\right)} = \frac{1-\eta}{6T}\norm{y_0}^2 + \frac{1-\eta}{2}\cdot \frac{1}{T}\sum^{T-1}_{i=0}\norm{y_i}^2. \label{eq:eOPT}
\end{equation}
Now consider any algorithm that doesn't know the indices of the corrupted observations. From their perspective, $\brc{y_i}$ is an i.i.d. sequence of draws from $\calN(0,2\Id)$, so we may without loss of generality assume that it forms each $\wh{x}_i$ independently of all other timesteps. Suppose that in timestep $i$, it uses randomized algorithm $\calA_i:\R^d\to\R^d$ to take in observation $y_i$ and outputs $\wh{x}_i$, where $\calA_i$ is oblivious to $a^*_i$. As $\norm{\wh{x}_i - y_i}^2$ and $\norm{\wh{x}_i}^2$ are both minimized when $\wh{x}_i$ has no component orthogonal to $y_i$, we may assume without loss of generality that $\calA(y)_i = \zeta_i\cdot y_i$ for some random $\zeta_i$. Then the sequence $\brc{\calA(y_i)}$ incurs clean posterior negative log likelihood at least
\begin{equation}
\frac{1}{T}\sum^{T-1}_{i=0} \left(a^*_i(\zeta_i - 1)^2 + \zeta_i^2\right)\norm{y_i}^2, \label{eq:subopt}
\end{equation}
where we have lower bounded $\frac{1}{T}\norm{\wh{x}_0}^2$ by zero. Now because the $\zeta_i$'s are independent of the $a^*_i$'s, the expectation of \eqref{eq:subopt} over the randomness of the algorithm and the $a^*_i$'s is
\begin{equation}
\frac{1}{T}\sum^{T-1}_{i=0} \E[\zeta_i]{(1 - \eta)(\zeta_i - 1)^2 + \zeta_i^2}\cdot\norm{y_i}^2.
\end{equation} The coefficient for each summand achieves its minimum of $\frac{1 - \eta}{2 - \eta}$ when $\zeta_i$ is deterministically $\frac{1 - \eta}{2-\eta}$, so any algorithm oblivious to the $a^*_i$'s incurs clean posterior negative log likelihood $L(\wh{x})$ at least $\frac{1-\eta}{2-\eta}\cdot\frac{1}{T}\sum^{T-1}_{i=0} \norm{y_i}^2$ in expectation. Contrasting this with \eqref{eq:eOPT} and noting that $\frac{1 - \eta}{2 - \eta} - \frac{1 - \eta}{2} = \Omega(\eta)$ for $0\le \eta < 1/2$, we conclude that in expectation over the randomness of the algorithm and the $a^*_i$'s,
\begin{equation}
\E{L(\wh{x})}\ge \E{\mathsf{OPT}} + \Omega(\eta)\frac{1}{T}\sum^{T-1}_{i=0} \norm{y_i}^2 - \frac{1 - \eta}{6T}\norm{y_0}^2.
\end{equation}
Taking a further expectation over the $y_i$'s concludes the proof.
\end{proof}
\section{Introduction}
\subsection{Background}
In this work, we revisit the classic problem of linear quadratic estimation, i.e. estimating the trajectory of a linear dynamical system from noisy measurements. First we review the setup:
\begin{enumerate}
\item[(1)] There are known matrices $A\in\R^{d\times d}$ and $B\in\R^{m\times d}$, and there is an unknown initial state $x^*_0\in\R^d$ drawn from $\mathcal{N}(0, R^2\, \Id)$.
\item[(2)] The trajectory $\brc{x^*_0,\ldots,x^*_{T-1}}$ and the observations $\brc{y^*_0,\ldots,y^*_{T-1}}$ are generated according to the following model:
\begin{align*}
x^*_i &= Ax^*_{i-1} + w^*_i \\
y^*_i &= Bx^*_i + v^*_i
\end{align*}
\noindent where $\brc{w^*_i}$ and $\brc{v^*_i}$ are \emph{process noise} and \emph{observation noise} and are vectors in dimensions $d$ and $m$ respectively. Furthermore $\brc{w^*_i}$ and $\brc{v^*_i}$ are all independently drawn from $\mathcal{N}(0, \sigma^2\, \Id)$ and $\mathcal{N}(0, \tau^2 \,\Id)$ respectively.
\end{enumerate}
The goal is to estimate the trajectory from the observations in either an offline or online sense, and to minimize the sum of squares of the error. The celebrated Kalman smoother and Kalman filter solve these two problems optimally. The main idea is that when the initialization and noise distributions are all Gaussian, at any time the posterior distribution on the trajectory given the observations is a Gaussian process. It suffices to estimate the mean of the posterior distribution and this can be done by finding the least squares solution to a structured regression problem that depends on $A$ and $B$ and the observations. It turns out that there is an even more compact formulation in terms of equations that define the Kalman filter.
The Kalman filter \cite{kalman1959general, kalman1961new} is one of the crowning achievements in control theory. It has wide-ranging applications in robotics, navigation, signal processing and econometrics. It is also a key building block in algorithms for estimating the model parameters of a linear dynamical system, as well as in change-point detection and building optimal controllers. Famously, the Kalman filter provided navigation estimates that helped guide the landing of the Apollo 11 lunar module in the Sea of Tranquility. Rudolf Kalman was awarded the National Medal of Science in 2008.
{\em But how brittle is the Kalman filter to assumptions of Gaussianity?} This is by no means a new question. If we relax the distributional assumptions but instead restrict the disturbances $\brc{w^*_i}$ and $\brc{v^*_i}$ to have bounded norm, then the minimax optimal filter can be found by dynamic programming. The solution is called the $H_\infty$ filter and has wide-ranging applications in its own right \cite{kailath2000linear}. However in many settings the assumption that the disturbances are bounded in norm is not reasonable either. In such cases, it has often been repeated that the Kalman filter can fail catastrophically. This is an unfortunate state-of-affairs because it means even though we can find the optimal filter when the noise is nice and Gaussian, it can break down badly with even a single badly outlying observation.
There are many natural heuristics that have been employed for dealing with outliers. However we are aware of only one work that gives rigorous guarantees in the presence of outliers. In 1994, Schick and Mitter \cite{schick1994robust} studied a model where the observation noise is drawn from a distribution
\begin{equation}
(1-\eta)\mathcal{N}(0, \sigma^2 \Id) + \eta \mathcal{H} \label{eq:sm}
\end{equation}
where $\mathcal{H}$ is a sufficiently regular distribution, but is allowed to be heavy-tailed. This is essentially the Huber contamination model. They derived provable guarantees but under a number of strong assumptions: First, they assumed that the distribution $\mathcal{H}$ is known to the filter designer. Thus the filter can use information about $\mathcal{H}$ to correct for the exact ways in which the noise is non-Gaussian. Second, their guarantees are asymptotic in nature in the sense that they only hold in the limit as $\eta \rightarrow 0$. As Schick and Mitter \cite{schick1994robust} discussed, as $T \rightarrow \infty$ for fixed $\eta > 0$, the number of outliers (i.e. timesteps where the observation noise is sampled from $\mathcal{H}$) goes to infinity. However their analysis relies on the exponential asymptotic stability of the Kalman filter, whereby outside of a window around the current timestep, the influence of older observations is significantly attenuated. Thus as $\eta \rightarrow 0$ the number of outliers in any window can be bounded, even if the total number of outliers cannot. In fact, while their estimator is nonlinear, as is necessary for handling heavy-tailed noise, it is constructed through a bank of Kalman filters. Each filter ignores one of the observations, assuming that it is the lone outlier. The filters are then combined in a natural way.
\paragraph{In this work we seek a truly robust filter.} We want to build a filter without any knowledge of $\mathcal{H}$. Moreover we want the outliers to not merely be sampled from a heavy-tailed distribution, but allow for their values to be chosen, possibly adaptively, by an unbounded adversary. For instance, this captures situations where the process generating the outliers is non-stationary. Moreover we want to prove guarantees that hold for fixed noise rates, as opposed to guarantees that only hold in the $\eta\to 0$ limit. Finally, we will want our filter to be a robustification of the Kalman filter itself in the sense that when $\eta = 0$ we want to achieve the same exact guarantees as before. Thus our filter competes with the Kalman filter in a strong sense, but gracefully degrades in performance as we move away from the precise distributional assumptions underlying the classic theory.
\subsection{Our Results}
\paragraph{The Corruption Model.} Here we explain our corruption model (see Section~\ref{sec:generative} for a formal description). Let $0 \le \eta < 1/2$ be the \emph{corruption fraction}. We will assume that for every timestep $i$, with probability $\eta$ the observation $y^*_i$ falls under the control of an all-powerful adversary. The adversary is allowed to replace all of the observations under his control with arbitrary values. Now let $\brc{y_0,\ldots,y_{T-1}}$ denote the sequence of observations that the learner ultimately receives. We emphasize that exactly which of these have been corrupted is unknown to the learner.
Note that the corrupted timesteps are \emph{randomly chosen}, just like in the Huber contamination model. As a special case this captures the setting \eqref{eq:sm} studied by Schick and Mitter. Moreover, because the adversary gets to coordinate his corruptions, our model also allows dependencies and captures situations where the observation noise is non-stationary over time. As we show in Appendix~\ref{subsec:srw}, in the stronger corruption model where the adversary gets to choose which timesteps to corrupt, there are strong impossibility results. Thus our corruption model seems to be one of the strongest where we can still hope for meaningful guarantees.
\paragraph{The Objective.}
As we discussed above, the Kalman filter can be thought of as computing the mean of the posterior distribution on the trajectory given the observations so far. When the noise is non-Gaussian, it is no longer true that the posterior distribution is itself a Gaussian process. It can be much more complex. {\em So how can we even define an optimization problem that generalizes that solved by Kalman filtering, if the posterior is non-Gaussian and, even worse, depends on $\mathcal{H}$ which is unknown to the filter and possibly changing over time?}
Our main idea is to compete with a strong oracle that knows which measurements are corrupted and which are not. Let $a^*_i\in\brc{0,1}$ denote the indicator variable for whether round $i$ is clean in the sense that its measurement error came from a Gaussian, rather than coming from $\mathcal{H}$ or being chosen by an adversary. When the $a^*_i$'s are known to the filter, the optimal estimator in a Bayesian sense is to estimate the mean of the posterior using only information from the uncorrupted observations. This leads us to the following objective:
\begin{equation}
L(\wh{x}) = \frac{1}{T} \left(\sum_{i = 0}^{T-1} (a^*_i \|B \wh{x}_i - y_i\|^2/\tau^2 + \norm{\wh{w}_i}^2/\sigma^2) + \|\wh{x}_0\|^2/R^2\right) \label{eqn:loss}
\end{equation}
where the steps $\wh{w}_i$ are defined in terms of the trajectory $\wh{x}_i$, i.e. $\wh{x}_i = A\wh{x}_{i-1} + \wh{w}_i$ for all $i > 0$.
This is the {\em clean posterior negative log likelihood}.
The best possible error we can achieve is given by
\begin{equation}
\mathsf{OPT} = \min_{\brc{\wh{x}_i}} L(\wh{x})
\label{eq:OPT_def_pre}
\end{equation}
which is attained by the maximum a posteriori (MAP) estimator. This is the same as the posterior mean and can be explicitly computed from a Kalman smoother that knows the observations that come from clean rounds, i.e. $y_i$ for the rounds $i$ for which $a^*_i = 1$. For general $\brc{\wh{x}_i}$, we refer to $L(\wh{x}) - \mathsf{OPT}$ as the \emph{excess risk}.
We take a moment to explain the differences in our approach compared to the usual approaches in algorithmic robust statistics, for example in robust mean estimation \cite{huber1964robust,diakonikolas2017being,lai2016agnostic}. Usually in a robust statistics problem, the objective of the estimator is to recover the \emph{ground truth}, e.g. the true mean in the example of mean estimation, and the goal in the robust setting is to recover the ground truth perfectly as $\eta \to 0$ while optimizing the dependence on the corruption fraction $\eta$. This makes sense for mean estimation because when $\eta = 0$, the mean can always be estimated consistently by taking more samples. It does not make as much sense in the case of Kalman filtering: even without corruptions, we only ever get one observation per timestep, so we cannot hope to recover the ground truth trajectory arbitrarily well. Uncertainty about the true trajectory is unavoidable because the complexity of the trajectory grows with the number of observations we get to make, and also because information about the past is washed away over time. For this reason, we need to pick our measure of success carefully.
From a Bayesian perspective, the clean posterior mean represents the \emph{best possible} estimate we can make of the ground truth given the clean observations {\em given additional information about which rounds have been corrupted}. Thus it gives a natural way to quantify the distance of our estimator from optimality.
\iffalse
Note that the objective function here corresponds to the negative log-density of the (Gaussian) posterior of the trajectory given the clean observations, up to additive constants and multiplicative factors.
In particular the minimum in \eqref{eq:OPT_def_pre} (i.e. the posterior MAP as well as the posterior mean, since the posterior is Gaussian) is achieved by iterates $\wh{x}_i$ and steps $\wh{w}_i$ given by running the offline Kalman filter (a.k.a. \emph{Kalman smoother}) on the part of the trajectory indexed by $i$'s for which $a^*_i = 1$. Since the algorithm does not know which times are uncorrupted (have $a^*_i = 1$), we cannot hope to exactly match the performance of $\mathsf{OPT}$. However, we use it as a benchmark and bound the amount of excess error our algorithms make compared to this oracle.
\fi
\paragraph{Main Results.} We show how to design an estimator which is robust to corruptions and competes with the optimum in the clean posterior log likelihood. In the clean case $\eta = 0$, our estimator gets asymptotically optimal posterior log likelihood, including the correct constant factor, and so its guarantee essentially matches the posterior mean/Kalman smoother. In fact, with high probability our estimator will be exactly the same as the Kalman smoother (see Remark~\ref{rmk:uncorrupted-case}).
\begin{theorem}[Informal, see Theorem~\ref{thm:main_TloglogT}]\label{thm:sos_informal}
For $\eta \le 0.49$, and for a uniformly stable and completely observable dynamical system,\footnote{Uniform stability and complete observability are standard assumptions from the control theory literature, which we introduce in Section~\ref{sec:ctrlbasics}.} there is a polynomial-time algorithm which takes as input the corrupted observations $\brc{y_i}$ and outputs a trajectory $\brc{\wh{x}_i}$ achieving excess risk \begin{equation}
L(\wh{x}) - \mathsf{OPT} \le C_d \,\eta\log(1/\eta)\left(m + d(\sigma^2/\tau^2)\log\log T\right) + o(1), \label{eq:Lbound}
\end{equation} with high probability, where the steps $\wh{w}_i$ are defined by $\wh{x}_i = A\wh{x}_{i-1} +\wh{w}_i$, where $C_d$ is a constant which is polynomially bounded in $\log d$ and the parameters of the system (see Section~\ref{sec:generative}), and where $o(1)$ is a quantity vanishing polynomially in $T$.
\end{theorem}
\noindent
To compare, we remind the reader of the performance of the obvious baselines: the Kalman filter can have unbounded error if there is even a single corruption, and oblivious outlier removal makes error $\Theta(T)$, see Appendix~\ref{subsec:srw}.
Recall that $\sigma^2$ is the variance of the process noise and $\tau^2$ is the variance of the observation noise. The dependence on the ratio $\sigma^2/\tau^2$ is unavoidable (Appendix~\ref{sec:sigbytau}), and the dependence on the dimensions $d$ and $m$ in \eqref{eq:Lbound} of Theorem~\ref{thm:sos_informal} is also unavoidable (Appendix~\ref{app:ddepend}).
Notably, we can obtain strong provable guarantees for any $\eta < 1/2$ (we wrote $\eta < 0.49$ above only to simplify the statement). Thus our estimator has an information-theoretically optimal breakdown point. Also, our result can handle the case where the eigenvalues of $A$ are on or near the unit circle, e.g. $A = \Id$, a situation where the system is marginally stable but not strictly stable (see e.g. \cite{simchowitz2018learning} for discussion of this terminology). This is an important distinction, because when the eigenvalues of $A$ are all small, a relatively simple method based on truncating the Kalman filter can work (we prove this in Appendix~\ref{apdx:wiener}), but the performance of such a heuristic will degrade badly as the eigenvalues approach the unit circle, whereas the algorithm of Theorem~\ref{thm:sos_informal} will still work (see Section~\ref{sec:overview} and Appendix~\ref{apdx:examples} for more discussion)
The above result works in an offline setting. But what happens when our measurements come in an online fashion and we need to estimate the position at time $i$ from only the observations up to that point? In this sense we want the filter to be {\em causal}.
Fortunately, at a small loss in our overall guarantees, we are able to make our approach online too. We need to change the definition of the objective slightly to handle the fact that predictions are made online: for the online case, we look at the suboptimality in predicting the next state compared to the oracle Kalman filter.
\begin{theorem}[Informal version of Theorem~\ref{thm:online}]\label{thm:online_informal}
There exists a polynomial time and causal estimator $\hat{\hat{x}}_{i + 1 | i}$ which satisfies the following guarantee with probability at least $1 - \delta$. With $\hat{x}_{i + 1 | x}$ denoting the output of the oracle Kalman filter, we have
\begin{equation*}
\frac{1}{T}\sum_{i = 1}^T \|\hat{\hat{x}}_{i + 1 | i} - x^*_{i + 1}\|^2
- \frac{1}{T} \sum_{i = 1}^T \|\hat{x}_{i + 1 | i} - x^*_{i + 1}\|^2 \le \frac{r''}{1 - \delta} (\eta + O(\sqrt{\eta \log(2/\delta)/T} + \log(2/\delta)/T))
\end{equation*}
where $r'' = C_d\, \eta \left(m + d(\sigma^2/\tau^2)\log T\right) + o(1)$ and $C_d$ is a constant which is polylogarithmic in $d,m$ and allowed to depend on other system parameters.
\end{theorem}
Our analysis combines our convex programming approach with standard properties of the Kalman filter, such as its exponential stability, building on an idea of Schick and Mitter \cite{schick1994robust}.
\subsection{Further Related Work}
\label{subsec:related}
\noindent\textbf{Robust Statistics and Sum-of-Squares.} Our main algorithm for robust filtering is based on the Sum of Squares hierarchy \cite{parrilo2000structured}, which has broad applications to both control theory (see e.g. \cite{prajna2005sostools,henrion2005positive}) and to algorithmic robust statistics (see e.g. \cite{klivans2018efficient, hopkins2018mixture, kothari2018robust,bakshi2020outlier,diakonikolas2020robustly,cherapanamjeri2020optimal}). It also builds upon a line of recent work in algorithmic robust statistics using both SoS and non-SoS methods (see e.g. \cite{diakonikolas2017being,diakonikolas2018list,lai2016agnostic,charikar2017learning,chinot2020erm,dalalyan2019outlier}). One of the techniques we use, introducing a positive semi-definiteness constraint utilizing matrix concentration bounds (see Technical Overview), is conceptually related to the main technical ingredient in \cite{chen2020online} for robust linear regression. The recent work \cite{banks2021local} also considers robust statistics in a Bayesian setting, namely community detection in the stochastic block model: Their method recovers the optimal detection threshold as the corruption level goes to zero, but not the optimal performance of the Bayes estimator (they only establish results for detection and not recovery).
\vspace{0.5em}
\noindent\textbf{Practical Approaches.} Algorithms based on minimizing loss functions that are less sensitive to outliers, e.g. the Huber loss, have been widely applied in practical works on robust filtering (see e.g. \cite{karlgaard2015nonlinear,mattingley2010real}) but thus far lack strong theoretical guarantees.
This is because, unlike in other contexts like robust/heavy-tailed mean estimation and regression \cite{huber1964robust,huber1973robust,chinot2020erm,mendelson2018learning}, the observations come from a non-stationary generative model and the length of the trajectory and number of observations are linked, so the existing proof techniques do not apply.
There are a number of similar ad hoc methods which have been applied in practice to handle outliers, especially by downweighting the Kalman filter updates when the innovation exceeds some threshold.
\vspace{0.5em}
\noindent\textbf{Other Notions of Robustness for Kalman Filtering} A popular variant of the standard Kalman filtering setup is to allow the variables $y_i$ to be dropped independently with some probability $p$ (analogous to our $\eta$); see e.g. \cite{sinopoli2004kalman,mo2008characterization,mo2011kalman}. This is like having an adversarial corruption model where the location of corruptions are known (and hence can be ignored); in this version of the problem, the Kalman filter remains optimal and the focus has been on understanding aspects of its behavior, e.g. in unstable systems.
We mention some other works in the direction of making the Kalman filter more robust. Some take the approach of assuming a particular parametric model for the noise that is non-Gaussian, such as a Student's $t$-distribution \cite{roth2013student} or L\'{e}vy distribution \cite{sornette2001kalman}, with the corresponding caveat that their results are limited to the model they assume. A major line of work in the control theory and filtering literature deals with robustness to uncertainty in the parameters of the system (e.g. $A$ and $B$), which is quite different from the problem of handling outliers. For example \cite{poor1980robust}, studies a version of this question in the context of the Wiener filter: the goal there is to choose the best linear filter given the uncertainty set, and so it cannot handle outlying observations where a nonlinear filter is required. Similarly, in recent works such as \cite{tsiamis2020online,simchowitz2020improper} the authors study online methods for control theory problems where the goal is to compete with the best policy in a certain class (e.g. compete with the Kalman filter when the system is unknown, or compete with the best of the $H_{\infty}$ controller and the $H_2$ controller without knowing the disturbance model), but in our setting the only filters which perform well are nonlinear and so competing with a class of linear filters like the Kalman filter is not sufficient.
\paragraph{Roadmap}
In Section~\ref{sec:overview} we provide an overview of the key ingredients in our proof of Theorems~\ref{thm:sos_informal} and \ref{thm:online_informal}. In Section~\ref{sec:prelims} we give technical preliminaries including a formal description of the generative model we consider as well as sum-of-squares basics. In Section~\ref{sec:logT} we give an algorithm that recovers a confidence band around the true trajectory, and in Section~\ref{sec:TloglogT} we use this to obtain a refined estimate achieving the excess risk bound of Theorem~\ref{thm:sos_informal}. In Section~\ref{sec:online} we prove the online guarantee of Theorem~\ref{thm:online_informal}. In Appendix~\ref{apdx:wiener}, we show that a simple variant of the Kalman filter works when $A$ has eigenvalues in the interior of the unit disk. In Appendix~\ref{apdx:examples} we present a number of examples elucidating why various components of our final bound and our analysis are necessary.
\section{Technical Overview}\label{sec:overview}
In this section, we first review basic concepts from control theory like observability and stability that will play an important role in the subsequent discussion. We then give an overview of the main challenges of handling adversarial corruptions when estimating a linear dynamical system and explain the techniques we develop to obtain our main results.
\subsection{Control Theory Basics}
\label{sec:ctrlbasics}
\paragraph{Observability.} Before we describe how to estimate a linear trajectory from corrupted observations, we first review how to do so in the absence of corruptions. Note that without additional assumptions on $A$ and $B$, this may not be possible \emph{a priori}. For instance, $A$ might only act nontrivially on some subspace of $\R^d$, and the rows of $B$ might simply be completely orthogonal to this subspace, in which case we can't hope to recover the trajectory.
In control theory, the standard way to ensure that the linear dynamical system at hand is not degenerate in this fashion is to assume it is \emph{observable}, as originally defined by Kalman in \cite{kalman1959general}. Formally, given a parameter $s\in\mathbb{N}$, define the \emph{observability matrix}
\begin{equation}
\calO_s \triangleq \sum^{s-1}_{i=0} (A^i)^{\top}B^{\top}B A^i.
\end{equation}
To motivate this object, suppose momentarily that there were no observation or process noise, so that the trajectory is given by $x^*_i = Ax^*_{i-1}$ and the observations are given by $y^*_i = Bx^*_i$. Then our observations up to the $s$-th timestep are given by $y^*_i = BA^i x^*_0$ for $0\le i < s$. Now note that
\begin{equation}
{x^*_0}^{\top}\calO_s x^*_0 = \sum^{s-1}_{i=0} \norm{BA^i x^*_0}^2 = \sum^s_{i=0} \norm{y^*_i}^2.
\end{equation}
In particular, if $\calO_s$ had nonzero kernel and $x^*_0$ were an element of this kernel, then all of the observations up to time $s$ would be zero, and we would get no information about $\brc{x^*_0,\ldots,x^*_{s-1}}$. Conversely, if $\calO_s$ were full rank, then one can recover $\brc{x^*_0,\ldots,x^*_{s-1}}$ given $\brc{y^*_0,\ldots,y^*_{s-1}}$ by solving the appropriate linear system. In other words, non-degeneracy of the observability matrix $\calO_s$ is a \emph{necessary and sufficient condition} for being able to recover the trajectory up to time $s$ from observations regardless of where the trajectory started.
More generally when there is observation and process noise, the natural quantitative analogue of non-degeneracy of $\calO_s$ is an upper bound on its condition number (see e.g. \cite{muller1972analysis,fetzer1975observability}):
\begin{assumption}[Complete observability\--- informal, see Assumption~\ref{assume:obs}] \label{assume:obs-no-noise}
For some $s\in\mathbb{N}$, $\calO_s \triangleq \sum^{s-1}_{i=0} (A^i)^{\top}B^{\top}B A^i$ is well-conditioned.
\end{assumption}
\paragraph{Stability}
We will focus on models which satisfy the following weak stability assumption, often made in the control theory literature (including the work of Schick and Mitter \cite{schick1994robust}):
\begin{assumption}[Uniform stability\--- informal, see Assumption~\ref{assume:uniform}]\label{assume:uniform_overview}
There is a constant $\rho \ge 1$ such that for any $i\in\mathbb{N}$, $\norm{A^i} \le \rho$ (here $\norm{\cdot}$ denotes the operator norm).
\end{assumption}
Intuitively, uniform stability ensures that the system initialized at any point will not eventually blow up at some time in the future. In contrast, if $A$ has an eigenvalue larger than one, the system is called \emph{explosive} or \emph{unstable} and the state will blow up at an exponential rate. Although Kalman filtering has also been studied in the case where $A$ is unstable, we know from the work on \emph{intermittent Kalman filtering} (see Section~\ref{subsec:related}) that even the oracle Kalman filter, which knows the location of the corruptions, will diverge if the corruption level in the unstable case is above some critical value \cite{mo2008characterization}. Since the setting we consider is strictly more difficult, there is no hope of closely tracking the trajectory in our setting.
\iffalse
On its own, Assumption~\ref{assume:obs-no-noise} is still insufficient to obtain meaningful guarantees. For instance, imagine $A$ were a large multiple of the identity. In this case the trajectory will explode at an exponential rate, and even without corrupted observations, no algorithm will be able to achieve bounded posterior negative log likelihood (see Appendix~\ref{app:stable} for details). As a result, a standard assumption that is often made in this literature (including the aforementioned work of Schick and Mitter \cite{schick1994robust}) is that of \emph{uniform stability}:
\begin{assumption}[Uniform stability\--- informal, see Assumption~\ref{assume:uniform}]\label{assume:uniform_overview}
There is a constant $\rho \ge 1$ such that for any $i\in\mathbb{N}$, $\norm{A^i} \le \rho$ (here $\norm{\cdot}$ denotes the operator norm).
\end{assumption}
Intuitively, uniform stability ensures that the system initialized at any point will not eventually blow up at some time in the future. \TODO{do we want to say more?}
\fi
\subsection{Our Techniques}
\label{sec:techniques}
\paragraph{Corruptions Degrade Observability.}
The first complication that arises in our setting is that corruptions can degrade the observability of the system. To see this, again consider the setup where there is no process or observation noise, but now some of the observations have been corrupted. We're essentially given a linear system $\brc{y_i = BA^i x^*_0}_{0<i<s}$ where some unknown subset $S_{\mathsf{bad}}\subseteq\brc{0,\ldots,s-1}$ of equations have been altered adversarially. If we knew $S_{\mathsf{bad}}$, then we could remove the corresponding equations and try solving for $x^*_0$ with the rest. Then the matrix that we need to be non-degenerate is no longer $\calO_s$ but rather
\begin{equation}
\calO'_s\triangleq \sum_{i\not\in S_{\mathsf{bad}}} (A^i)^{\top}B^{\top}B A^i. \label{eq:subO}
\end{equation}
Of course, $\calO_s$ being non-degenerate does not guarantee $\calO'_s$ is as well (see Appendix~\ref{subsec:switching}). One might wonder then whether Assumption~\ref{assume:obs-no-noise} must be significantly strengthened to ensure that the trajectory can be recovered from corrupted observations. As we illustrate in Appendix~\ref{subsec:switching}, this is indeed the case if the corruptions arrived at arbitrary timesteps. But if the corruptions arrive in a random fashion, we will demonstrate that no additional assumptions need to be made.
\paragraph{Corruptions Subsample the Observability Matrix.}
To get a sense for how this could be possible, note that if $S_{\mathsf{bad}}$ comprises a random $\eta$ fraction of the indices up to time $s$, then the expectation of $\calO'_s$ is exactly $(1 - \eta)\cdot \calO_s$. If we could argue that $\calO'_s$ also concentrates around its expectation, then because $\calO'_s$ is spectrally close to $(1 - \eta)\cdot \calO_s$ and $\calO_s$ is non-degenerate/well-conditioned by assumption, the corruptions don't actually impact the observability of the dynamical system in the presence of Huber contamination.
As the summands $(A^i)^{\top} B^{\top} B A^i$ are bounded in norm by Assumption~\ref{assume:uniform_overview}, we can carry through this matrix concentration argument as long as $s$ is sufficiently large. In Assumption~\ref{assume:obs-no-noise} however, we make no assumptions about how large $s$ is. Instead, we note that regardless of how large $s$ in Assumption~\ref{assume:obs-no-noise} is, by observing the system $t$ steps at a time rather than $s$ steps at a time for a large multiple $t$ of $s$, we find that Assumption~\ref{assume:obs-no-noise} also holds for $t$ in place of $s$.
More precisely, if we consider the observability matrix $\calO_t \triangleq \sum^{t-1}_{i=0} (A^i)^{\top}B^{\top} B A^i$ for some moderately large $t$, then
$\calO_t\succeq \calO_s$, and one can also easily check that $\norm{\calO_t} \le (t\rho^2/s)\norm{\calO_s}$. In other words, the condition number of $\calO_t$ is at most $t\rho^2/s$ worse than that of $\calO_s$. So even if $s$ from Assumption~\ref{assume:obs-no-noise} is too small for $\calO'_s$ to concentrate sufficiently around its expectation, it would appear that we can simply take $t$ large enough that $\calO'_t$ concentrates sufficiently around its expectation.
\paragraph{Key Complication: Observable/Unobservable Subspaces.}
There is one essential wrinkle in the above argument: the fluctuations for matrix concentration for $\calO'_t$ are of order $\sqrt{t}$, so at some point they may exceed the smallest singular value of $\calO_t$. In particular, it would appear that because of Huber contamination, we lose all control over how well we can estimate the component of the trajectory that lives in the span of the small singular vectors of $\calO_t$, and so no matter what value of $t$ we choose, matrix concentration alone fails to show we can estimate the state successfully.
We now sketch a way to get around this problem.
First note that in spite of the issue posed by the small singular vectors of $\calO_t$, the preceding discussion on matrix concentration does ensure that the projection of $\calO'_t$ to the \emph{large} singular vectors of $\calO_t$ is sufficiently nondegenerate with high probability (see Lemma~\ref{lem:anticonc}). One might therefore hope to be able to estimate the trajectory within this subspace. For this reason, we will refer to the span of the large singular vectors of $\calO_t$ as the \emph{observable subspace} and its orthogonal complement as the \emph{unobservable subspace}; denote projectors to these subspaces by $\Pi$ and $\Pi^{\perp}$ respectively. (The unobservable subspace should \emph{not} be thought of as some kind of ``invisible'' subspace which doesn't affect the observations; based on the discussion above, it represents directions which are difficult to recover locally based on partially corrupted observations.)
So what do we do about the unobservable subspace? Here is the key idea: while we do lose control of the trajectory's component inside the unobservable subspace \emph{within any fixed window of $t$ steps}, we can consolidate information \emph{across windows} to learn what goes on in the unobservable subspace. Before we can discuss how to implement this, we need to introduce our estimator.
\paragraph{An Inefficient Estimator.}
To motivate the design of our estimator, we will construct a system of constraints capturing salient features of the model. We begin by introducing vector-valued variables $\brc{x_0,\ldots,x_{T-1}}$ corresponding to our estimate for the trajectory, as well as variables $\brc{w_1,\ldots,w_{T-1}}$ and $\brc{v_0,\ldots,v_{T-1}}$ corresponding to our estimates for the process and observation noise. We also introduce Boolean variables $a_0,...,a_{T-1} \in \{0,1\}$, where $a_i$ corresponds to our guess for whether the $i$-th observation was uncorrupted. Of course, the true values of the quantities that these variables represent are unknown to us, but there are a number of basic constraints they must satisfy. Firstly, because each observation is independently corrupted with probability $\eta$, we know by Chernoff that
\begin{equation}
\frac{1}{T}\sum^{T-1}_{i=0} a_i \geq (1 - \eta) - o(1).
\end{equation}
Secondly, we know that the trajectory is given by a linear dynamical system, and in any uncorrupted timestep $i$, the observation $y_i$ is a noisy linear measurement of the trajectory at that time, so
\begin{equation}
x_i = A x_{i - 1} + w_i \label{eq:follow_dynamics}
\end{equation}
\begin{equation}
a_i (y_i - B x_{i - 1} - v_i) = 0. \label{eq:fit_clean_ys}
\end{equation}
Additionally, we know the dynamics and observation noise is bounded with high probability:
\begin{equation}
\|w_i\|^2 = O(\sigma^2 d), \quad \|v_i\| = O(\tau^2 m). \label{eq:noise_bound_informal}
\end{equation}
Thus far the constraints have been fairly straightforward. We now describe a key constraint capturing the preceding discussion on matrix concentration.
Recall that because the corruptions arrive in a random fashion, by matrix concentration the uncorrupted timesteps will ``subsample'' the observability matrix $\calO_t$ in each window. In other words, in every window $\brc{\ell t,\ldots,(\ell + 1)t - 1}$ of $t$ timesteps, the following spectral lower bound holds with high probability
\[\sum^{t - 1}_{j=0} (1 - a_{\ell t + j}) (A^j)^{\top} B^{\top} B A^j \preceq \eta\cdot \calO_{\t} + O(\sqrt{t}) \cdot \Id\]
Finally, with all of our constraints in place, it is straightforward to define an objective to optimize, in light of \eqref{eqn:loss}. We want to minimize the ``clean'' negative log-likelihood achieved by $\brc{x_i}$, where ``clean'' is defined with respect to the variables $\brc{a_i}$ instead of the true indicators $\brc{a^*_i}$:
\begin{equation}
\min \frac{1}{T} \left(\sum_{i = 1}^T (a_i \|Bx_i - y_i\|^2/\tau^2 + \norm{w_i}^2/\sigma^2) + \|x_0\|^2/R^2\right). \label{eqn:program_loss}
\end{equation}
Based on the discussion above, we might guess that $\|\Pi(x_i - x_i^*)\|^2$, the error from the observable subspace, can be bounded in the above program based on some kind of matrix concentration argument. It turns out this question itself is subtle, because we would need to rule out cancellations between the observable and unobservable subspace. However, even if we did argue that by itself, this is definitely not enough to make our ultimate objective \eqref{eqn:loss}, the clean negative posterior likelihood, small: for some of the clean observations ($i$ with $a^*_i = 1$), they will be dependent on the information in the unobservable subspace, and the true state $x_i^*$ can be very large in the unobservable subspace (see Appendix~\ref{sec:3dexample} for an illustrative example). The key difficulty to making the objective \eqref{eqn:loss} small will be to argue that $\|x_i - x^*_i\|$, i.e. the error in the \emph{whole} space, can be bounded.
\iffalse
As alluded to above, at this point it is natural to postulate that the estimator $\{x_i\}$ given by solving this optimization problem can at least achieve small error in the observable subspace. That is to say, one might hope to be able to bound $\|\Pi(x_i - x_i^*)\|^2$; we will often refer to this as the \emph{observable error} at time $i$. Unfortunately, it turns out to be difficult to bound the observable error without also bounding the unobservable error. The intuition is that it's easy to construct linear dynamical systems where error in either subspace can rotate into one another repeatedly over time (see Appendix~\ref{sec:3dexample} for an illustrative example). Inability to control unobservable error in the past translates into inability to control observable error in the present.
The key takeaway from this should thus be that it's important to reason about how the observable and unobservable error of timestep $i$ jointly relate to the observable and unobservable error at timestep $i-t$. With this picture in mind, we make good on our earlier promise to integrate information across windows, by elucidating a property of completely observable linear dynamical systems which enables us to relate errors in the present to errors in the past.
\fi
\paragraph{Decomposing the Error.}
Roughly speaking, we will prove this by showing that the effect of errors in the past decays exponentially fast for our estimator. Intuitively, this argument will show that we do not make large errors in the unobservable subspace because the estimator will successfully propagate information from the past. Formally, we prove the following key inequality:
\begin{equation}\label{eq:error-contraction}
\| x_{i+t} - x_{i+t}^*\|^2 \leq \frac{1}{2}\| x_{i} - x_{i}^*\|^2 + [\text{noise}],
\end{equation}
where here and throughout the rest of this overview, we use $[\text{noise}]$ to denote a small quantity that is polynomially bounded in $t$ and in the variance of the observation and process noise (see Lemma~\ref{lem:error_decay}).
The proof of \eqref{eq:error-contraction} starts with the following decomposition of the error:
\begin{equation}
x_{i+t} - x_{i+t}^* = \left[(x_{i + t} - x_{i + t}^*) - A^t(x_i - x_i^*)\right] + A^t \Pi(x_{i} - x_i^*) + A^t \Pi_{V_\perp}(x_{i} - x_i^*) \label{eqn:error-decomposition}
\end{equation}
The first term is the amount of new noise introduced between steps $i$ and $i + t$: it will be small because we are only taking $t$ steps, the process noise $\brc{w^*_i}$ is small, and the corresponding program variables $\brc{w_i}$ are also constrained to be small.
The remaining two terms in \eqref{eqn:error-decomposition} account for the error propagated from the past: the second term $A^t \Pi(x_{i} - x_i^*)$ represents the error propagated from the observable error in the past, and the third term $A^t \Pi^{\perp}(x_{i} - x_i^*)$ represents the error propagated from the unobservable error in the past. We now show how both of those terms can be bounded, starting with the last of these terms (unobservable error).
\paragraph{Unobservable Error from the Past.}
We identify a simple but critical fact about observable linear dynamical systems: any vector in the unobservable subspace decreases in norm when it evolves forward by $t$ timesteps. More formally, we can show that for any vector $x\in\R^d$,
\begin{equation}
\norm{A^t\Pi^{\perp}x}^2 \le c\cdot \norm{\Pi^{\perp}x}^2, \label{eq:unobservable_decay}
\end{equation}
for some small constant $0 < c < 1$ provided $t$ is sufficiently large relative to $s$ (see Lemma~\ref{lem:unobservable-decay}).
In particular, by applying this to the vector given by the unobservable error in the past, we conclude that the error $\norm{A^t\Pi(x_i - x^*_i)}^2$ from propagating the past unobservable error $\norm{\Pi(x_i - x^*_i)}^2$ can be upper bounded by a small fraction of $\norm{\Pi(x_i - x^*_i)}^2$.
\begin{remark}
We must caution the reader that \eqref{eq:unobservable_decay} does \emph{not} mean that the observable component any vector $x$ decays after $x$ evolves over $t$ timesteps, which would instead correspond to a bound of the form
\begin{equation}
\norm{\Pi_{V_\perp} A^{t} x}^2 \le C\cdot \norm{\Pi_{V_\perp} x}^2. \label{eq:point_decay_informal}
\end{equation} for some $C < 1$.
Of course, if \eqref{eq:point_decay_informal} were true, it would make life much easier: by taking $x$ to be any iterate in the trajectory, we would include that over time, the trajectory barely lives in the unobservable subspace at all! Unfortunately, this is not the case, as it is easy to construct linear dynamical systems
with a significant portion of the state in the unobservable subspace (see Appendix~\ref{subsec:unobservableunbounded}).
\end{remark}
\paragraph{Observable Error from the Past.}
We now discuss how to handle the error $\norm{A^t\Pi(x_i - x^*_i)}^2$ propagated from the observable error in the past. It is tempting to try the same approach as for the unobservable error here, namely to argue that $\norm{A^t\Pi(x_i - x^*_i)}^2$ is a small fraction of the observable error in the past. But this is too much to hope for: together with \eqref{eq:unobservable_decay} this would imply that $\norm{A^t} < 1$, whereas we only assume that $\norm{A^t} \le \rho$ for some $\rho \ge 1$ (as we show in Appendix~\ref{apdx:wiener}, it is quite easy to handle linear dynamical systems for which the former holds).
In the absence of an analogue of \eqref{eq:unobservable_decay}, our key insight is that we can instead relate $\norm{A^t\Pi(x_i - x^*_i)}^2$ to the \emph{unobservable} error in the past! Informally, any large errors in estimating the observable part of the state $x^*_i$ must be explained by a large amount of interference from the unobservable part of the state.
Essentially, in Lemma~\ref{lem:observable-case-analysis} we show that for some small constant $0 < c' < 1$,
\begin{equation}\label{eq:overview-observable}
\|\Pi(x_i - x_i^*)\|^2 \leq c'\cdot \| \Pi^{\perp}(x_i - x_i^*) \|^2 + [\text{noise}].
\end{equation}
The proof is rather involved but can be distilled into two main ingredients. Firstly, as alluded to earlier, the fact that the random corruptions subsample the observability matrix allows us to relate the error in estimating the state to the error in fitting the observations. Secondly, we can bound the latter using the following inequality (see Lemma~\ref{lem:pi-piperp-ineq}):
\begin{equation}\label{eqn:pi-piperp-intro}
a^*_i a_i\| BA^i \Pi (x_{0} - x^*_{0})\|^2 \leq 4a^*_i a_i\| BA^i \Pi^\perp (x_0 - x^*_0)\|^2 + [\text{noise}].
\end{equation}
Inequality \eqref{eqn:pi-piperp-intro} formalizes the following idea: because our estimator must match the observation at time $i$, any errors coming from the past are constrained to cancel between the observable and unobservable parts of the space, so if one is large the other is large as well.
Unfortunately, combining the two ingredients above naively would establish \eqref{eq:overview-observable} with too large of a constant on the right hand side.
It turns out however that we can greatly improve the constant by appealing to the aforementioned decay property of the unobservable subspace from \eqref{eq:unobservable_decay}.
The details here are a bit subtle, and we refer to the proof of Lemma~\ref{lem:error_decay} for the formal argument.
\paragraph{Contraction of Error over Time.}
From \eqref{eq:unobservable_decay} and \eqref{eq:overview-observable}, it is straightforward to deduce \eqref{eq:error-contraction} and conclude that the error incurred by the filter is decaying exponentially quickly. Recalling that $\norm{(x_{i + t} - x_{i + t}^*) - A^t(x_i - x_i^*)}^2$ will be some small noise term, we have by triangle inequality that
\begin{align*}
\| x_{i+t} - x_{i+t}^*\|^2
&\le [\text{noise}] + 3\| A^t \Pi(x_{i} - x_i^*)\|^2 + 3\|A^t \Pi^{\perp}(x_{i} - x_i^*)\|^2 \\
&\le [\text{noise}] + 3c'\|\Pi^{\perp}(x_{i} - x_i^*)\|^2 + 3c\| \Pi^{\perp}(x_{i} - x_i^*)\|^2 < [\text{noise}] + \frac{1}{2}\|x_{i} - x_i^*\|^2,
\end{align*}
which proves \eqref{eq:error-contraction}. So we have established that over the course of $t$ timesteps, the error of our estimate essentially decays geometrically. We remark that this doesn't give us any control over our error on the iterates \emph{within} these timesteps, and in particular it remains to be shown how to use all of the tools we've introduced to compete with the Bayes-optimal predictor. It turns out however that by virtue of the program constraints ensuring that our estimate of the trajectory follows the linear dynamics \eqref{eq:follow_dynamics} and tries to fit the clean observations \eqref{eq:fit_clean_ys} while minimizing the objective \eqref{eqn:program_loss}, the excess risk introduced \emph{within} any window of $t$ is polynomially bounded by $t$ and the variance of the noise (see Lemma~\ref{lem:outer}). In this overview, we do not delve into the details of this as this particular argument is reminiscent of existing analyses in the robust statistics literature.
\paragraph{Confidence Bands and Achieving $\log\log T$ Excess Risk.}
There is an important catch in the above discussion which we now need to address: We assumed that every noise vector $u^*_i$ and $w^*_i$ is bounded, and that subsampling holds in every window of size $t$.
The trouble is in order to ensure that these events hold across the entire trajectory of length $T$, we would need to take $t$ to be logarithmic in $T$. This would translate into incurring a logarithmic overhead in the excess risk bound. {\em We are able to avoid paying the full price for this union bound.}
First, it turns out that our analysis only requires that the observation and process noise are bounded \emph{on average} over the trajectory in an appropriate sense (see Constraints~\ref{constraint2:totalnoise}-\ref{constraint2:measurements_noise_corrupted}). The more serious issue is: How can we avoid assuming that subsampling holds in each window? This is a key component to being able to integrate information across different time windows. As a starting point, we observe that our initial estimator actually meets a stronger guarantee: It actually outputs a trajectory which is \emph{pointwise} close to the true trajectory by a distance that scales polynomially in variance of the noise and polylogarithmically in $T$ (see Corollary~\ref{cor:warmstart}). In other words, it allows us to form a \emph{confidence band} around the estimator of radius, and this logarithmic scaling is essentially optimal for any guarantee of this form.
We show how to exploit this confidence band. In particular, we engineer a second system of constraints which incorporates the output of our initial estimator and refines it to achieve our final $\log\log T$ excess risk bound in Theorem~\ref{thm:sos_informal}. The main idea is because the noise that accumulates over a window scales polynomially in the window size, we can consider windows over \emph{shorter timescales} scaling doubly logarithmically rather than logarithmically in $T$. The goal is to achieve higher accuracy on shorter windows whenever subsampling holds. This might seem counterproductive: By shortening the windows, we are only making it more likely that subsampling will fail in any given window! Indeed, if we are now taking windows of doubly logarithmic length, then in roughly a $1/\mathsf{polylog}(T)$ fraction of the windows, the random corruptions will fail to properly subsample the observability matrix. But this is where the confidence band comes in: Over these bad windows, we already know how to estimate those iterates pointwise to error $\mathsf{polylog}(T)$, so the total contribution of the bad windows to the excess risk scales as $(1/\mathsf{polylog}(T))\cdot \mathsf{polylog}(T) = O(1)$ !
The key complication is that the algorithm designer doesn't actually know which windows subsampling failed in. Instead, we will set up a system of constraints similar to the one for our earlier estimator but with additional Boolean variables, one per window, corresponding to our guess for whether the random corruptions in that window successfully subsampled the observability matrix (see Program~\ref{program:sosv2}). We show how to integrate information \emph{across the windows on which we correctly guessed that subsampling succeeded} to achieve our final guarantee. As the argument here is rather involved and we defer the details to Section~\ref{sec:decay} and Lemma~\ref{lem:error-first-iterate-sos2}.
\paragraph{Efficient Algorithm via Sum-of-Squares.}
While the estimators we have described appear to be inefficient as they require solving certain systems of polynomial constraints, our proofs that the solutions to these systems satisfy the guarantees of Theorem~\ref{thm:sos_informal} are simple in the sense that they can be implemented in the degree four sum-of-squares proof system \cite{Shor87,parrilo2000structured}. So instead of solving these polynomial systems which would \emph{a priori} incur an exponential runtime, it suffices to output a \emph{pseudo-distribution} over solutions and round it to an integral solution in a straightforward way. These tools have become a mainstay in algorithm design in robust statistics more broadly \cite{hopkins2018mixture, kothari2018robust, bakshi2020outlier, diakonikolas2018robustly, bakshi2021robust, liu2021settling, bakshi2020robustly}. We will explain the basics of the sum-of-squares proof system in Section~\ref{sec:sos_basics}.
\paragraph{Two-stage filter.} Thus far we have been discussing the offline problem, where the estimator at some time instant is allowed to depend on the entire observation sequence $y_1,\ldots,y_T$ \--- in other words, we designed a robust Kalman {\em smoother}. Our techniques can be used to solve the online filtering problem, where the prediction for $x_t$ is only allowed to depend on $y_1,\ldots,y_{t - 1}$. It turns out that the transformation to online guarantees is simple: We use the offline smoother developed above, in particular the confidence band it outputs, as an outlier removal algorithm and combine it with the standard Kalman filter run on the observations $\tilde{y}_1,\ldots,\tilde{y}_T$ which are not deleted by outlier removal. After the outlier removal, only small corruptions remain, so following an idea of Schick and Mitter \cite{schick1989robust,schick1994robust}, we can then use stability estimates for the Kalman filter to establish accuracy guarantees on its prediction for the next state.
\iffalse
tells us that with probability at least $1 - \delta$,
\[ \norm*{\sum^{\t}_{i=0}a^*_i (A^i)^{\top}B^{\top}BA^i - (1 - \eta)\calO_{\t}} \le O(\rho^2\norm{B}^2 \sqrt{ \t \log(d/\delta)}) \]
where $\calO_{\t}$ is the \emph{observability matrix} defined before.
\fi
\section{Preliminaries}
\label{sec:prelims}
Given matrix $M$, let $\lambda_{\min}(M)$ and $\lambda_{\max}(M)$ denote its bottom and top singular values, and let $\kappa(M) \triangleq \frac{\lambda_{\min}(M)}{\lambda_{\max}(M)}$. We will sometimes also denote $\lambda_{\max}(M)$ by $\norm{M}$.
\subsection{Generative Model}
\label{sec:generative}
For known matrices $A\in\R^{d\times d}$, $B\in\R^{m\times d}$, the underlying trajectory $\brc{x^*_i}$ and uncorrupted observations $\brc{y^*_i}$ are given by $x^*_0 \sim\calN(0,R^2\cdot\Id_d)$ and
\begin{align}
x^*_i &= Ax^*_{i-1} + w^*_i \ \text{for all} \ i>0 \label{eq:dynamics} \\
y^*_i &= Bx^*_i + v^*_i \ \text{for all} \ i\ge 0\label{eq:measure}
\end{align}
where the \emph{dynamics noise} $w^*_i$ is independently sampled from $\calN(0,\sigma^2 \cdot \Id_d)$ and and the \emph{observation noise} $v^*_i$ is independently sampled from $\calN(0,\tau^2 \cdot \Id_m)$, i.e. both types of noise in the system are isotropic Gaussian up to scaling. The assumption that the noise is isotropic simplifies notation greatly, and is largely without loss of generality in the following sense: if the noise covariance matrices are full rank, a change of basis will make the noise isotropic. Our results can also be extended to handle the case where the noise covariance is rank degenerate.
We now describe how the corrupted observations are formed. After the trajectory $\brc{x^*_i}$ and observations $\brc{y^*_i}$ have generated, an independent $\mathrm{Ber}(1 - \eta)$ coin is flipped for every $0\le i < T$; let $a^*_i\in\brc{0,1}$ denote the outcome at time $i$. For all $i$ for which $a^*_i = 1$, define $y_i = y^*_i$. For all $i$ for which $a^*_i = 0$, a computationally unbounded adversary is allowed to set $y_i$ arbitrarily. We can assume this adversary has full knowledge of the system, e.g. the full trajectory $\brc{x^*_0,\ldots,x^*_{T_1}}$, the full sequence of true observations $\brc{y^*_0,\ldots,y^*_{T-1}}$, etc.
The baseline estimation error which we will try to achieve approximately is
\begin{equation}
\mathsf{OPT} = \min_{\brc{\wh{x}_i}} \frac{1}{T} \left(\sum_{i = 0}^{T-1} (a^*_i \|B \wh{x}_i - y_i\|^2/\tau^2 + \norm{\wh{w}_i}^2/\sigma^2) + \|\wh{x}_0\|^2/R^2\right) \label{eq:OPT_def}
\end{equation}
where the minimum ranges over all trajectories $\wh{x}_i$ with steps $\wh{w}_i$ satisfying $\wh{x}_i = A\wh{x}_{i-1} + \wh{w}_i$ for all $i > 0$.
Note that the objective function here corresponds to the negative log-density of the (Gaussian) posterior of the trajectory given the clean observations, up to additive constants and multiplicative factors.
In particular the minimum in \eqref{eq:OPT_def} (i.e. the posterior MAP as well as the posterior mean, since the posterior is Gaussian) is achieved by iterates $\wh{x}_i$ and steps $\wh{w}_i$ given by running the offline Kalman filter (a.k.a. \emph{Kalman smoother}) on the part of the trajectory indexed by $i$'s for which $a^*_i = 1$. Since the algorithm does not know which times are uncorrupted (have $a^*_i = 1$), we cannot hope to exactly match the performance of $\mathsf{OPT}$. However, we use it as a benchmark and bound the amount of excess error our algorithms make compared to this oracle.
We make the following assumptions which are standard in the filtering/control literature (see e.g. \cite{anderson2007optimal,anderson2012optimal,schick1994robust}):
\begin{assumption}[Complete Observability] \label{assume:obs}
There exist constants $\alpha,\kappa > 0$ and $s\in\mathbb{N}$ for which \begin{equation}
\calO_s \triangleq \sum^{s-1}_{i=0} (A^i)^{\top}B^{\top}B A^i
\end{equation} satisfies $\lambda_{\min}(\calO_s) \ge \kappa s$ and $\lambda_{\max}(\calO_s) \le \alpha s$.
\end{assumption}
\begin{assumption}[Uniform stability] \label{assume:uniform}
There is a constant $\rho \ge 1$ for which $\norm{A^j} \le \rho$ for all $j\in\mathbb{N}$.
\end{assumption}
The following elementary manipulations will be useful:
\begin{fact}\label{fact:epochs}
For $t\in\mathbb{N}$ divisible by $s$, $\calO_t = \sum^{t/s - 1}_{r=0} (A^{rs})^{\top}\calO_s A^{rs}$. In particular, \begin{equation}
\norm{\calO_t} \le t\cdot \alpha \cdot \rho^2. \label{eq:opnormOt}
\end{equation}
\end{fact}
\begin{fact}\label{fact:unroll}
$x^*_t - A^t x^*_0 = \sum^t_{j=1} A^{t-j} w^*_j$
\end{fact}
\subsection{Sum-of-Squares Basics}
\label{sec:sos_basics}
Here we give a quick review of the sum-of-squares algorithm; for a more thorough treatment, we refer the reader to \cite{o2013approximability,barak2014sum,hopkins2018statistical}.
\paragraph{Pseudoexpectations.} The sum-of-squares SDP hierarchy is a series of increasingly tight SDP relaxations for solving polynomial systems $\calP \triangleq \{p_i(x) \geq 0\}_{i=1}^N$. Although it is in general NP-hard to solve polynomial systems, the level-$\ell$ SoS SDP attempts to approximately solve $\calP$ with increasing accuracy as $\ell$ increases by adding more constraints to the SDP. This improvement in approximation naturally comes at the expense of increasing runtime and space.
In particular, one can think of the SoS SDP as outputting a "distribution" $\mu$ over solutions to $\calP$. However, there are two important caveats. Firstly, one can only access the degree-$\ell$ moments of the "distribution" and secondly there may be no true distribution with the corresponding degree $\ell$ moments. Thus we refer to $\mu$ as a \emph{pseudodistribution}.
\begin{definition}
A degree $\ell$ pseudoexpectation $\wt{\mathbb{E}} : \R[x]_{\leq \ell} \to \R$ {\it satisfying $\calP$} is a linear functional over polynomials of degree at most $\ell$ satisfying
\begin{enumerate}
\item(Normalization) $\psE*{1} = 1$,
\item(Constraints of $\calP$) $\psE*{p(x) a^2(x)} \geq 0$ for all $p \in \calP$ and polynomials $a$ with $\deg(a^2 \cdot p) \le \ell$,
\item(Non-negativity on square polynomials) $\psE*{q(x)^2} \ge 0$ whenever $\deg(q^2) \le \ell$.
\end{enumerate}
\end{definition}
For any fixed $\ell \in \mathbb{N}$, given a polynomial system, one can efficiently compute a degree $\ell$ pseudo-expectation in polynomial time.
\begin{fact} (\cite{Nesterov00,parrilo2000structured,Lasserre01,Shor87}). For any $n$, $\ell \in \Z^+$, let $\wt{\mathbb{E}}_{\zeta}$ be degree $\ell$ pseudoexpectation satisfying a polynomial system $\calP$. Then the following set has a $n^{O(\ell)}$-time weak
separation oracle (in the sense of \cite{GLS1981}):
\begin{align*}
& \brc*{\wt{\mathbb{E}}_\zeta\left[(1, x_1, x_2, . . . , x_n)^{\otimes \ell}\right]: \text{ degree } \ell \text{ pseudoexpectations } \wt{\mathbb{E}}_{\zeta} \text{ satisfying }\calP}
\end{align*}
Using this separation oracle, the ellipsoid algorithm finds a degree $\ell$ pseudoexpectation in time $n^{O(\ell)}$. We call this algorithm the \emph{degree $\ell$ sum-of-squares algorithm}.
\end{fact}
To reason about the properties of pseudo-expectations, we turn to the dual object of sum-of-squares proofs.
\paragraph{Sum-of-Squares Proofs}
For any nonnegative polynomial $p(x): \R^d \rightarrow \R$, one could hope to prove its nonnegativity by writing $p(x)$ as a sum of squares of polynomials $p(x) = \sum_{i=1}^m q_i(x)^2$ for a collection of polynomials $\{q_i(x)\}_{i=1}^m$. Unfortunately, there exist nonnegative polynomials with no sum-of-squares proof even for $d = 2$. Nevertheless, there is a generous class of nonnegative polynomials that admit a proof of positivity via a proof in the form of a sum of squares. The key insight of the sum-of-squares algorithm is that these sum-of-squares proofs of nonnegativity can be found efficiently provided the degree of the proof is not too large.
\begin{definition} (Sum-of-Squares Proof)
Let $\mathcal{A}$ be a collection of polynomial inequalities $\{p_i(x) \geq 0\}_{i=1}^m$. A sum-of-squares proof that a polynomial $q(x) \geq 0$ for any $x$ satisfying the inequalities in $\mathcal{A}$ takes on the form
\[
\left(1+ \sum_{k \in [m']} b_k^2(x)\right) \cdot q(x) = \sum_{j\in [m'']} s_j^2(x) + \sum_{i \in [m]} a_i^2(x) \cdot p_i(x)
\]
where $\{s_j(x)\}_{j \in [m'']},\{a_i(x)\}_{i \in [m]}, \{b_k(x)\}_{i \in [m']}$ are real polynomials. If such an expression were true, then $q(x) \geq 0$ for any $x$ satisfying $\mathcal{A}$. We call these identities sum-of-squares proofs, and the degree of the proof is the largest degree of the involved polynomials $\max \{\deg(s_j^2), \deg(a_i^2 p_i)\}_{i,j}$. Naturally, one can capture polynomial equalities in $\mathcal{A}$ with pairs of inequalities. We denote a degree $\ell$ sum-of-squares proof of the positivity of $q(x)$ from $\calA$ as $\calA \sststile{\ell}{x} \{q(x) \geq 0\}$ where the superscript over the turnstile denote the formal variable over which the proof is conducted. This is often unambiguous and we drop the superscript unless otherwise specified.
\end{definition}
A number of basic inequalities like Cauchy-Schwarz and H\"{o}lder's admit sum-of-squares proofs (see e.g. Appendix A of \cite{hopkins2018mixture}).
Sum-of-squares proofs can also be strung together and composed according to the following convenient rules.
\begin{fact}
For polynomial systems $\calA$ and $\calB$, if $\calA \sststile{d}{x} \{p(x) \geq 0\}$ and $\calB \sststile{d'}{x} \{q(x)\geq 0\}$ then $\calA \cup \calB \sststile{\max(d,d')}{x}\{p(x) + q(x) \geq 0\}$. Also $\calA \cup \calB \sststile{dd'}{x} \{p(x)q(x) \geq 0\}$
\end{fact}
Sum of squares proofs yield a framework to reason about the properties of pseudo-expectations, that are returned by the SoS SDP hierarchy.
\begin{fact} (Informal Soundness)
If $\calA \sststile{r}{x} \{q(x) \ge 0\}$ and $\psE{\cdot}$ is a degree-$\ell$ pseudoexpectation operator for the polynomial system defined by $\calA$, then $\psE{q(x)} \ge 0$.
\end{fact}
The following standard fact which follows by ``SoS Cauchy-Schwarz'' (see e.g. Lemma A.5 of \cite{barak2014rounding}) will allow us to convert from pseudodistributions over solutions to a polynomial systems to integral solutions.
\begin{lemma}\label{lem:psE-CS}
For any vector $w^*$ and degree-2 pseudoexpectation $\psE{\cdot}$ over vector-valued variable $w$, we have that \begin{equation}
\norm{\psE{w} - w^*}^2 \le \psE{\norm{w - w^*}^2}.\label{eq:soscs}
\end{equation}
\end{lemma}
\begin{proof}
By the dual definition of $L_2$ norm, the left-hand side of \eqref{eq:soscs} can be written as \begin{equation}
\sup_{v\in\S^{d-1}}\iprod{\Sig v,\psE{w} - w^*}^2.
\end{equation} For any $v\in\S^{d-1}$, \begin{equation}
\iprod*{\Sig v,\psE{w} - w^*}^2 = \left(\psE{\iprod{\Sig v,w - w^*}}\right)^2 \le \psE{\iprod{\Sig v,w-w^*}^2} \le \psE{\norm{w - w^*}^2_{\Sig}},
\end{equation} where the first inequality follows by the pseudoexpectation version of SoS Cauchy-Schwarz (see e.g. Lemma A.5 of \cite{barak2014rounding}). Therefore, taking the maximum over all $v \in S^{d - 1}$ proves the inequality.
\end{proof}
Finally, we will need the following elementary but crucial inequality which admits a degree-2 sum-of-squares proof. Roughly speaking, it captures the fact that if the sum of two vectors is small in norm, then either vector must have norm upper bounded in terms of the norm of the other vector:
\begin{fact}\label{fact:sos_simple}
Let $v_1,v_2$ be $d$-dimensional vector-valued indeterminates. There is a degree-2 sum-of-squares proof of the inequality $\norm{v_1}^2 \le 4\norm{v_2}^2 + \frac{4}{3}\epsilon$ from the constraint $\norm{v_1 + v_2}^2 \le \epsilon$.
\end{fact}
\begin{proof}
By expanding out the hypothesis, we have
\[\|v_1\|^2 + 2\langle v_1, v_2 \rangle + \|v_2\|^2 \leq \epsilon.\] By Cauchy-Schwarz, we also have
\[-2\langle v_1, v_2 \rangle \leq \frac{1}{4}\|v_1\|^2 + 4\|v_2\|^2.\]
Adding these two inequalities together and rearranging gives the desired inequality.
\end{proof}
\subsection{Concentration Inequalities}
\iffalse
\begin{theorem}[Matrix Freedman Inequality, Corollary 4.2 of \cite{tropp2011user}]\label{thm:matrix-freedman}
Suppose $\vY_1,...,\vY_n\in\R^{d\times d}$ is a symmetric matrix martingale adapted to filtration $\mathcal{F}_t$, whose associated difference sequence $\brc{\vX_t}$ satisfies $\norm{\vX_t} \le 1$ almost surely for all $t$. Let $\vW = \sum_t \E{\vX_t^2 | \mathcal{F}_{t- 1}}$, then
\[ \Pr{\|\vY_n\| \ge t \text{ and } \|\vW\| \le \sigma_n^2} \le 2d \cdot \exp\left(\frac{-t^2/2}{\sigma_n^2 + t/3}\right). \]
\end{theorem}
\begin{corollary}\label{corr:matrix-bernstein}
In the same setting as Theorem~\ref{thm:matrix-freedman}, suppose that for some $\sigma \le 1$, $\E{\vX_t^2 | \mathcal{F}_{t - 1}} \preceq \sigma^2$ almost surely. Then
\[ \Pr{\|(1/n) \cdot \vY_n\| \ge u} \le 2d \cdot \exp\left(-\frac{n u^2/2}{\sigma^2 + u/3}\right) \]
\end{corollary}
\begin{proof}
Apply Theorem~\ref{thm:matrix-freedman} with $t = nu$ and $\sigma_n^2 = n \sigma^2$.
\end{proof}
\fi
\begin{lemma}[Matrix Hoeffding, see e.g. Theorem 1.3 in \cite{tropp}]\label{lem:matrixhoeffding}
For any $\delta > 0$, given symmetric random matrices $M_1,\ldots,M_T \in\R^{d\times d}$ satisfying $\norm{M_t} \le 1$ almost surely for all $t$, if $M \triangleq \sum_t M_t$, then \begin{equation}
\Pr{\norm{M - \E{M}} \ge \sqrt{8T\log(d/\delta)}} \le \delta.
\end{equation}
\end{lemma}
\begin{lemma}[see e.g. \cite{vershynin2018high}]\label{lem:normbound}
If $v\sim\calN(0,\Sig)$ for some $\Sig\in\R^{d\times d}$, then with probability at least $1 - \delta$, \begin{equation}
\norm{v} \le O\left(\left(\sqrt{d} + \sqrt{\log(1/\delta)}\right)\norm{\Sig}^{1/2}\right)
\end{equation}
\end{lemma}
We use concentration for Gaussian polynomials, which is a consequence of Gaussian hypercontractivity.
\begin{lemma}[see e.g. \cite{o2014analysis}]\label{lem:polyconc}
For degree-$d$ polynomial $p: \R^m\to\R$, if $x\sim\calN(0,\Id)$, then \begin{equation}
\Pr*{\abs*{p(x) - \E{p}} > t\cdot \sqrt{\Var{p}}} \le \exp(-\Omega(t^{2/d})).
\end{equation}
\end{lemma}
\section{Online Filtering}
\label{sec:online}
The main technical results proved earlier in this paper are for the problem of robust \emph{offline}/noncausal filtering, where we are allowed to look at future observations when estimating the state at time $t$. Of course, in many settings we would also like the ability to make online predictions.
In this section, we describe a two-stage filter which combines a first-stage filter based on offline filtering described above with a second stage of Kalman filtering. The main result of this section is that combining these two approaches gives good results for online robust filtering. The strategy is to use the first (offline) filter as a way to get a rough estimate of the state, which allows us to correct for large outliers in the sequence of observations. Given these corrected values, the success of the Kalman filter follows from classical stability results, which show that small deviations from the true response model do not cause the filter to explode. We note that the idea of using the stability of the Kalman filter as a tool in the analysis of robust filtering appeared in the work \cite{schick1994robust} referenced in the Introduction.
\subsection{Background: Stability of the Kalman Filter}
We start by stating the Kalman filter recursion in our notation. Recall from \eqref{eq:dynamics} and \eqref{eq:measure} that the uncorrupted version of our model is
\begin{align}
x^*_t &= Ax^*_{t-1} + w^*_t \\
y^*_t &= Bx^*_t + v^*_t
\end{align}
where $w^*_t\sim\calN(0,\sigma^2 \cdot \Id_d), v^*_t\sim\calN(0,\tau^2 \cdot \Id_m)$. Given this notation, the Kalman filter equations are
\begin{align*}
\hat{x}_{t | t } &= \hat{x}_{t | t - 1} + K_{t} \tilde y_{t} \\
\hat{x}_{t | t - 1} &= A \hat x_{t - 1 | t - 1} \\
\tilde y_{t} &= y^*_t - B \hat{x}_{t | t - 1} \\
S_t &= B P_{t | t - 1}B^T + \tau^2 \Id_m \\
K_{t} &= P_{t | t - 1} B^T S_t^{-1} \\
P_{t | t} &= (I - K_t B) P_{t | t - 1} \\
P_{t | t - 1} &= A P_{t - 1 | t - 1} A^T + \sigma^2 \Id_d
\end{align*}
where $\hat{x}_{t | t}$ is the \emph{a posteriori} state estimate, $\hat{x}_{t | t - 1}$ is the \emph{a priori} state estimate, $\tilde y_t$ is the \emph{innovation}, $S_t$ the covariance of the innovation, $K_t$ is the optimal Kalman gain, $P_{t | t}$ is the \emph{a posteriori} state estimate covariance, and $P_{t | t - 1}$ is the \emph{a priori} state estimate covariance.
\begin{theorem}[Exponential Asymptotic Stability of the Kalman Filter, \cite{moore1980coping}, Theorem 4.1 of \cite{schick1994robust} and \cite{schick1989robust}]\label{thm:kalman-stability}
Suppose that the matrices $A,B$ satisfy observability and $\sigma > 0$.
Then if
\[ \tilde x_{t | t} = (I - K_t B)A\tilde \tilde x_{t - 1 | t - 1}, \]
there exists $\lambda > 0$ and $\delta \in (0,1)$ depending on the system parameters such that
\[ \|\tilde x_{t | t}\| < \lambda \delta^t \|\tilde x_{0 | 0}\| \]
for any $t \ge 1$.
\end{theorem}
This result is shown using Lyapunov's method and more precise dependence of $\lambda,\delta$ on system parameters is given in \cite{schick1989robust},
based on the argument in \cite{moore1980coping}. We note that in the more general case where the system noise covariance is rank-degenerate, the result needs an additional (standard) assumption called \emph{controllability}.
The significance of $\tilde x_t$ can be seen from the equation
\[ \hat{x}_{t | t} = A \hat{x}_{t - 1 | t - 1} + K_t (y^*_t - B A\hat{x}_{t - 1 | t - 1}) = (I - K_t B) A\hat{x}_{t - 1 | t - 1} + K_t y^*_t = \tilde{x}_{t | t} + K_t y^*_t \]
which follows by expanding the Kalman filter equations; the Theorem shows that if we add a perturbation $\tilde x_{0 | 0}$ to the true value of $\hat{x}_{0 | 0}$, the effect of this perturbation disappears at an exponential rate. As a consequence we obtain the following result.
\begin{corollary}[\cite{schick1989robust,schick1994robust,moore1980coping}]\label{corr:stability}
Suppose that the assumptions of Theorem~\ref{thm:kalman-stability} are satisfied, and let $(y_t)_{t = 1}^T$ be an arbitrary sequence of random variables. Let $\hat{x}_{t | t}$ be defined by the Kalman filter equations above with arbitrary initialization, and let $\hat{ \hat{x}}_{t | t}$
be defined by the Kalman filter equations with input $y^*_t$ replaced by $y_t$ and the same initialization. Then for any $t$,
\[ \|\hat{\hat{x}}_{t | t} - \hat{x}_{t | t}\| \le \lambda \sum_{s = 1}^t \delta^{t - s} \|K_s(y_s - y^*_s)\|. \]
\end{corollary}
\subsection{Two-Stage Filter}
The following Lemma describes the two-stage filter we propose. Given access to a sequence of $y'_1,\ldots,y'_T$ which is close to the true (noisy) observations $y^*_1,\ldots,y^*_T$, the Kalman filter is applied to the sequence $y''_1,\ldots,y''_T$ which corresponds to $\tilde{y}_1,\ldots,\tilde{y}_T$ with all gross outliers removed. Note that the sequence $y'_1,\ldots,y'_T$ is allowed to dependence arbitrarily on $y_1,\ldots,y_T$, and so can be produced by an offline filter, while the final prediction for time $T + 1$ is produced online. The result has accuracy close to that of the Kalman filter on the clean data for predicting (in an online/causal fashion) the next state $x^*_{T + 1}$. The first Lemma below states this result for a single timestep, and the Theorem gives a version averaged over time steps.
\begin{lemma}\label{lem:filter1}
Let $\mathcal{F}_T$ is the filtration generated by all random variables from the generative model up to time $T$ (i.e. $a^*_t,y^*_t,x^*_t$).
Suppose that the collection of random variables $y'_1,\ldots,y'_T$ are measurable with respect to $\mathcal{F}_T$
Furthermore, suppose that there exists $r \ge 0$ such that
\begin{equation}\label{eqn:guaranteed-corruption}
\max_{t = 1}^T |y^*_t - y'_t| \le r.
\end{equation}
For $1 \le t \le T$, define $y''_t = \tilde{y}_t$ if $|\tilde{y}_t - y'_t| \le r$ and $y''_t = y'_t$ otherwise. Let $\hat{\hat{x}}_{T + 1 | T}$ denote the output of the Kalman filter applied to input $y''_t$ and $\hat{x}_{T + 1 | T}$ denote the output of the Kalman filter applied to the (unknown) ground truth $y^*_1,\ldots,y^*_{T}$, then
\[ \E{\|\hat{\hat{x}}_{T + 1 | T} - x^*_{T + 1}\|^2 \mid \mathcal{F}_T} = \E{\|\hat{x}_{T + 1 | T} - x^*_{T + 1}\|^2 \mid \mathcal{F}_T} + \|A\| \lambda r \sum_{s = 1}^T \|K_s\| \delta^{T - s} (1 - a^*_s). \]
\end{lemma}
\begin{proof}
Recall that $\hat{\hat{x}}_{T + 1 | T} = A \hat{\hat{x}}_{T | T}$ and likewise for $\hat{x}_T$. It follows that
\begin{align*}
\|\hat{\hat{x}}_{T + 1 | T} - \hat{x}_{T + 1 |T}\|
&\le \|A\|\|\hat{\hat{x}}_{T | T} - \hat{x}_{T | T}\| \\
&\le \|A\| \lambda \sum_{s = 1}^T \delta^{T - s} \|K_s(y^*_s - y''_s)\| \\
&\le \|A\| \lambda \sum_{s = 1}^T \delta^{T - s} (1 - a^*_s) \|K_s(y^*_s - y'_s)\| \\
&\le \|A\| \lambda \sum_{s = 1}^T \delta^{T - s} (1 - a^*_s) \|K_s\| r
\end{align*}
where the second inequality follows from Corollary~\ref{corr:stability} and the last inequality follows from the fact that $y''_t = y^*_t$ unless $\tilde y_t$ is guaranteed to be a corruption, i.e. $a^*_s = 0$, based on \eqref{eqn:guaranteed-corruption}.
\end{proof}
\begin{theorem}
Let $\delta > 0$ be arbitrary.
Let $\mathcal{F}_i$ is the filtration generated by all random variables from the generative model up to time $i$ (i.e. $a^*_t,y^*_t,x^*_t$).
Suppose that the collection of random variables $y'_{i1},\ldots,y'_{ii}$ are measurable with respect to $\mathcal{F}_i$.
Suppose\footnote{Such an upper bound can be derived from the other assumptions we make, see e.g. \cite{schick1994robust,moore1980coping}.} that $\|K_i\| \le K$ for all $i$.
Furthermore, suppose that there exists $r \ge 0$ such that
\begin{equation}
\max_{t = 1}^T |y^*_t - y'_t| \le r.
\end{equation}
with probability at least $1 - \delta/2$. Then with $\hat{\hat{x}}_{i + 1 | i}$ defined as in Lemma~\ref{lem:filter1} (in terms of $y'_{i1},\ldots,y'_{ii}$), we have
\begin{align*}
&\frac{1}{T}\sum_{i = 1}^T \E{\|\hat{\hat{x}}_{i + 1 | i} - x^*_{i + 1}\|^2 \mid \mathcal{F}_i}
- \frac{1}{T} \sum_{i = 1}^T \E{\|\hat{x}_{i + 1 | i} - x^*_{i + 1}\|^2 \mid \mathcal{F}_i} \\
&\qquad \le \frac{\|A\| \lambda r K}{1 - \delta} (\eta + O(\sqrt{\eta \log(2/\delta)/T} + \log(2/\delta)/T)).
\end{align*}
with probability at least $1-\delta$.
\end{theorem}
\begin{proof}
Observe that
\begin{align*}
\sum_{i = 1}^T \sum_{s = 1}^i \|K_i\| \delta^{i - s} (1 - a^*_s)
&\le K \sum_{i = 1}^T \sum_{s = 1}^i \delta^{i - s} (1 - a^*_s) \\
&= K \sum_{s = 1}^T (1 - a^*_s) \sum_{i = s}^T \delta^{i - s} \\
&\le \frac{K}{1 - \delta} \sum_{s = 1}^T (1 - a^*_s)
\end{align*}
and
\[ \sum_{s = 1}^T (1 - a^*_s) \le \eta T + O(\sqrt{\eta T \log(2/\delta)} + \log(2/\delta)) \]
with probability at least $1 - \delta/2$ by Bernstein's inequality. Therefore the result follows from Lemma~\ref{lem:filter1} and the union bound.
\end{proof}
Now we can apply the two-stage approach by combining it with the SoS filter.
In this case, $y'_{i1},\ldots,y'_{ii}$ are simply the output of the offline SoS run at time $i$. Because the key input to the previous Theorem is the confidence band assumption, we only need to run the first (simpler) SoS program in \textsc{SoSKalman}, although running the second step won't hurt.
\begin{theorem}\label{thm:online}
Suppose as described above that $y'_{i1},\ldots,y'_{ii}$ are the output of rounding Program~\ref{program:sos} as in \textsc{SoSKalman}, and the assumptions of Corollary~\ref{cor:warmstart} are satisfied. Suppose (for simplicity) that the initialization radius $R = O(r''/\sqrt{d})$ where $r''$ is as defined below.
With probability at least $1 - \delta$,
\begin{align*}
&\frac{1}{T}\sum_{i = 1}^T \E{\|\hat{\hat{x}}_{i + 1 | i} - x^*_{i + 1}\|^2 \mid \mathcal{F}_i}
- \frac{1}{T} \sum_{i = 1}^T \E{\|\hat{x}_{i + 1 | i} - x^*_{i + 1}\|^2 \mid \mathcal{F}_i} \\
&\qquad \le \frac{\|A\| \lambda r'' K}{1 - \delta} (\eta + O(\sqrt{\eta \log(2/\delta)/T} + \log(2/\delta)/T))
\end{align*}
where $r'' = O(\|B\|\rho \epsilon_{\mathsf{geo}}^{1/2})$ as in Corollary~\ref{cor:warmstart}.
\end{theorem}
\begin{remark}
The above version of the result is stated in terms of conditional expectations, which means we integrate over the conditional randomness of $x^*_{i + 1}$ (i.e. over the Gaussian posterior) when evaluating the prediction. This is analogous to looking at pseudoregret vs. regret in bandits, see e.g. \cite{bubeck2012regret}. By a standard martingale concentration argument, this can be converted into a high probability upper bound on $\frac{1}{T}\sum_{i = 1}^T \|\hat{\hat{x}}_{i + 1 | i} - x^*_{i + 1}\|^2
- \frac{1}{T} \sum_{i = 1}^T \|\hat{x}_{i + 1 | i} - x^*_{i + 1}\|^2$ if desired.
\end{remark}
Note that for simplicity, we assumed that the width $R$ of the prior in the initialization $x_0 \sim N(0,R^2 I_d)$ was not too large, to simplify the final bound. It is possible to modify the above in a straightforward way and get a more complex statement showing the dependence on $R$; just as in the SoS filter, the the term involving $R$ will decay exponentially fast as time goes on.
\iffalse
As an example, we can combine the two-stage approach with the Huber filter. In this case, $y'_{i1},\ldots,y'_{ii}$ are simply the output of the offline \textsc{HuberFilter} run at time $i$.
\begin{theorem}
Suppose as described above that $y'_{i1},\ldots,y'_{ii}$ are the output of \textsc{HuberFilter}, and the assumptions of Corollary~\ref{corr:main_huber} are satisfied.
With probability at least $1 - \delta$,
\begin{align*}
&\frac{1}{T}\sum_{i = 1}^T \E{\|\hat{\hat{x}}_{i + 1 | i} - x^*_{i + 1}\|^2 \mid \mathcal{F}_i}
- \frac{1}{T} \sum_{i = 1}^T \E{\|\hat{x}_{i + 1 | i} - x^*_{i + 1}\|^2 \mid \mathcal{F}_i} \\
&\qquad \le \frac{\|A\| \lambda r'' K}{1 - \delta} (\eta + O(\sqrt{\eta \log(2/\delta)/T} + \log(2/\delta)/T))
\end{align*}
where $r''$ is as defined in Corollary~\ref{corr:main_huber}.
\end{theorem}
As mentioned before, applying the same approach with the SoS filter instead gives a very similar bound with weaker assumptions on the observability and larger breakdown point guarantee.
\fi
\section{Stationary Case: Robust Wiener Filter}\label{apdx:wiener}
In this Appendix, we consider the case where the matrix $A$ in the linear dynamical system is \emph{stable}, i.e. has all eigenvalues strictly within the unit ball of radius one. Equivalently \cite{kailath2000linear} the linear dynamical system is \emph{exponentially stable}:
\begin{equation}\label{eqn:exp-stable}
\|A^t\| \le \rho e^{-\gamma t}
\end{equation}
for some positive constants $\rho,\gamma$ dependent on $A$.
In this case, the state of the dynamical system, stared at any initialization, will converge to a stationary measure $N(0, \Sigma)$
with
\[ \Sigma = \sigma^2(I + AA^T + (A^2)(A^2)^T + \cdots) = \sigma^2 \sum_{k = 0}^{\infty} (A^k)(A^k)^T \]
and the convergence of the infinite matrix sum follows from \eqref{eqn:exp-stable}.
Because the system will converge to the stationary measure from any initialization, it makes sense (and is conventional) to focus on the case of a \emph{stationary} system where time ranges from $-\infty$ to $\infty$. In this case the Kalman filter simplifies to the \emph{Wiener filter}; we recall the relevant facts here. (The precise statement of Theorem 8.4.1 in \cite{kailath2000linear} is more detailed and explains how to solve for $K$ as the unique stabilizing solution of a Discrete-time Algebraic Ricatti Equation.)
\begin{theorem}[Theorem 8.4.1 of \cite{kailath2000linear}]
In the time-invariant stationary state space model described above,
the minimal-squared error predictor $\hat{x}_{t + 1}$ of the next state $x_{t + 1}$
satisfies the recursion
\begin{align*}
\hat{x}_{t + 1} := \E{x_{t + 1} \mid (y_j)_{t = -\infty}^j} = A \hat{x}_t + K(y_t - B \hat{x}_t) = (A - KB) \hat{x_{t}} + K y_t
\end{align*}
for some gain matrix $K$ which is determined by the system parameters but does not depend on $t$, and such that $A - KB$ is stable (i.e. has all eigenvalues within the open unit disk).
\end{theorem}
Equivalently, we can write the Wiener filter as
\[ \hat{x}_{t} = \sum_{s = 1}^{\infty} (A - KB)^{s - 1} K y_{t - s}. \]
We now show that in this stationary setting, a truncated version of the Wiener has reasonable robustness guarantees. This contrasts with the much more difficult setting which we focus on in the main text (uniformly but not strictly stable dynamics), where naively truncating the Kalman filter makes total error $\Omega(T^2)$, see Technical Overview.
Given parameters $h \ge 1, \tau \ge 0$ we formally define the truncated filter by
\[ \hat{x}_{t,h,\tau} := f_{\tau}\left(\sum_{s = 1}^{h} (A - KB)^{s - 1} K y_{t - s}\right) \]
where
\[ f_{\tau}(x) = x \cdot \bone{\|x\| \le \tau} \]
zeroes out its input if its input $x$ has norm larger than $\tau$.
\iffalse
\begin{lemma}
As a function of $y_{t - 1},y_{t - 2},\ldots$ we have
\[ \|\hat{x}_{t,h,\tau}(y_{t - 1},\ldots) - \hat{x}_{t,h,\tau}(y'_{t - 1},\ldots)\| \le ? \]
where
\end{lemma}
\fi
\begin{lemma}\label{lem:wiener-truncation}
With the above notation,
\begin{align*}
\MoveEqLeft \sqrt{\E{\|\hat x_t - \hat x_{t,h,\tau}\|^2}} \\
&\le \|A - KB\|^h \left(\sum_{s = 1}^{\infty} \|A - K B\|^{s - 1}\right) \sqrt{\E{\|K y_0\|^2}} + \frac{1}{\tau} \left(\sum_{s = 1}^{\infty} \|A - K B\|^{s - 1}\right)^2 \sqrt{\E{\|K y_0\|^4}}.
\end{align*}
\end{lemma}
\begin{proof}
By the $L_2$ triangle inequality,
\begin{equation}\label{eqn:truncation-error}
\sqrt{\E{\|\hat x_t - \hat x_{t,h,\tau}\|^2}} \le \sqrt{\E{\|\hat x_t - \sum_{s = 1}^{h} (A - KB)^{s - 1} K y_{t - s}\|^2}} + \sqrt{\E{\|\hat x_{t,h,\tau} - \sum_{s = 1}^{h} (A - KB)^{s - 1} K y_{t - s}\|^2}}.
\end{equation}
To bound the first term on the right hand side, observe that
\begin{align*}
\E*{\|\hat x_t - \sum_{s = 1}^{h} (A - KB)^{s - 1} K y_{t - s}\|^2}
&= \E*{\left\|\sum_{s = h + 1}^{\infty} (A - KB)^{s - 1} K y_{t - s}\right\|^2} \\
&\le \E*{\left(\sum_{s = h + 1}^{\infty} \|A - KB\|^{s - 1} \|K y_{t - s}\|\right)^2} \\
&\le \left(\sum_{s = h + 1}^{\infty} \|A - K B\|^{s - 1}\right) \E*{\sum_{s = h + 1}^{\infty} \|A - K B\|^{s - 1} \|K y_{t - s}\|^2} \\
&\le \|A - K B\|^{2h} \left(\sum_{s = 1}^{\infty} \|A - K B\|^{s - 1}\right)^2 \E{\|K y_0\|^2}
\end{align*}
where the second inequality is Cauchy-Schwarz and in the last step we used stationarity.
To upper bound the second term on the right hand side of \eqref{eqn:truncation-error}, observe
\begin{align*}
\MoveEqLeft \E*{\|\sum_{s = 1}^{h} (A - KB)^{s - 1} K y_{t - s}\|^2 \bone*{\|\sum_{s = 1}^{h} (A - KB)^{s - 1} K y_{t - s}\|^2 > \tau^2}} \\
&\le \sqrt{\E*{\|\sum_{s = 1}^{h} (A - KB)^{s - 1} K y_{t - s}\|^4} \Pr*{\|\sum_{s = 1}^{h} (A - KB)^{s - 1} K y_{t - s}\|^2 > \tau^2}} \\
&\le \frac{1}{\tau^2} \E*{\|\sum_{s = 1}^{h} (A - KB)^{s - 1} K y_{t - s}\|^4}
\end{align*}
where the first inequality is Cauchy-Schwarz and the second inequality is Markov's inequality. Observe by Holder's inequality that
\begin{align*}
\|\sum_{s = 1}^{h} (A - KB)^{s - 1} K y_{t - s}\|^4
&\le \left(\sum_{s = 1}^h \|A - K B\|^{s - 1} \|K y_{t - s}\|\right)^4 \\
&\le \left(\sum_{s = 1}^h \|A - K B\|^{s - 1}\right)^3\sum_{s = 1}^h \|A - K B\|^{s - 1} \|K y_{t - s}\|^4
\end{align*}
so using stationarity and linearity of expectation
\[
\E*{\|\sum_{s = 1}^{h} (A - KB)^{s - 1} K y_{t - s}\|^4} \le \left(\sum_{s = 1}^{\infty} \|A - K B\|^{s - 1}\right)^4 \E{\|K y_0\|^4}.
\]
\end{proof}
\begin{theorem}
Suppose that $\eta$ fraction of responses are arbitrarily corrupted, let $\hat x_t$ denote the Wiener filter run on the uncorrupted responses and let $x'_{t,h,\tau}$ denote the truncated Wiener filter run on the corrupted responses. Then taking $h = \Theta_{A,K,B}(\log(1/\eta))$ and $\tau = \Theta_{A,K,B}([(1/\eta)\log(1/\eta)]^{1/3})$,
we have that
\[ \E{\|\hat{x}_0 - x'_{0,h,\tau}\|^2} \le C \eta^{1/3} \log(1/\eta)^{1/3} \]
where $C$ is a constant independent of $\eta$ but allowed to depend on all other system constants.
\end{theorem}
\begin{proof}
By the union bound, there is a probability of at most $\eta h$ that one of observations in the $h$ previous time steps is corrupted. Since the probability of corruption is independent of the process, the expected error conditional on a corruption occuring is at most $O(\tau^2 + E \|x_0\|^2)$. Conditional on no corruption in the last $h$ timesteps, the error is upper bounded by the previous Lemma~\ref{lem:wiener-truncation}. Therefore the error is upper bounded by
\begin{multline} O\Bigg(\eta h (\tau^2 + E \|x_0\|^2) + \|A - KB\|^h \left(\sum_{s = 1}^{\infty} \|A - K B\|^{s - 1}\right) \sqrt{\E{\|K y_0\|^2}} \\
+ \frac{1}{\tau} \left(\sum_{s = 1}^{\infty} \|A - K B\|^{s - 1}\right)^2 \sqrt{\E{\|K y_0\|^4}}\Bigg)
\end{multline}
Note that provided $\eta h \tau^2 \to 0$, $h \to \infty$, and $1/\tau \to 0$ the right hand side goes to zero as $\eta \to 0$. Taking $h = \Theta_{A,K,B}(\log(1/\eta))$ and $\tau = \Theta_{A,K,B}([(1/\eta)\log(1/\eta)]^{1/3})$ gives the result.
\end{proof}
We note that both $C$ and the dependence on $\eta$ can easily be further optimized, but we have not done this for simplicity's sake. Also, note that this argument indeed relied strongly on the fact that the matrix $A$ is (strictly) stable; when $A$ has eigenvalues on the unit circle, there is no constant threshold $\tau$ which will work in the argument, since the dynamical system generally won't stay near the origin.
\iffalse
we write this filter as
\[ \hat{x}_t = \sum_{s = 1}^{\infty} W_s \hat{y}_{t - s} \]
where $W_t$ are a sequence of matrices determined by the system parameters (see \cite{kailath2000linear} for a more explicit description of the filter). (TODO: need that they are bounded which they should be, look at textbook equation 8.4.7.)
We observe that truncating the sum has only a small effect on the filter accuracy:
\begin{lemma}
Define
\[ \hat{x}_{t,h} := \sum_{s = 1}^{h} W_s \hat{y}_{t - s} \]
then
\[ \lim_{h \to \infty} \E*{(\hat{x}_t - \hat{x}_{t,h})^2} = 0. \]
\end{lemma}
\begin{proof}
\end{proof}
Next, we define a version of the filter which is also truncated in its output:
\[ \hat{x}_{t,h,\tau} := \hat{x}_{t,h} \bone(|\hat{x}_{t,h}| \le \tau). \]
\begin{lemma}
For any $h$,
\[ \lim_{\tau \to \infty} \E*{(\hat{x}_{t,h} - \hat{x}_{t,h,\tau})^2} = 0. \]
\end{lemma}
TODO: construct robust filter by truncation.
\fi |
\section{Introduction}
In the past six years, more than $50$ binary \rev{compact object (CO)} mergers were detected during the first three observing runs of Advanced LIGO and Virgo gravitational-wave (GW) interferometers \citep{VIRGOdetector,LIGOdetector,abbottGW150914,abbottastrophysics,abbottO1,abbottGW170817,abbottO2,abbottO2popandrate,abbottO3a,abbottO3popandrate,abbottGW190425,abbottGW190412,abbottGW190814,LVK2021BHNS,gwtc-2.1,gwtc-3}. So far, the sample includes the merger of 47 binary black holes (BBHs), 2 double neutron stars and two black hole -- neutron star (BHNS) binaries. Understanding the formation and the merger of binary COs and searching for distinctive signatures of different formation scenarios is of utmost importance to help us interpret the current and future detections of the LIGO-Virgo-KAGRA collaboration (LVK).
Among the main proposed formation channels for merging binary COs, we find: pairing of primordial black holes (BHs) \citep[e.g.,][]{carr1974,carr2016,bird2016,scelfo2018,deluca2021}, isolated binary evolution via common envelope \citep[e.g.,][]{tutukov1973, bethe1998,portegieszwart1998,belczynski2002, belczynski2008,dominik2013, belczynski2016,eldridge2016,stevenson2017,mapelli2017,mapelli2018,mapelli2019, klencki2018,ablimit2018,kruckow2018,spera2019,neijssel2019,eldridge2019}, via stable mass transfer \citep[e.g.][]{kinugawa2014,kinugawa2020,inayoshi2017,vandenheuvel2017,tanikawa2021a,tanikawa2022}, or via chemically homogeneous mixing \citep[e.g.,][]{marchant2016,demink2016,mandel2016,dubuisson2020}, \rev{ dynamical perturbations in the field \citep{micheaely2019,michaely2020}}, dynamical formation in young star clusters (YSCs, e.g. \citealt{banerjee2010,ziosi2014,mapelli2016,askar2017,banerjee2017, rastello2018,banerjee2018,dicarlo2019,dicarlo2020a,dicarlo2020b,kumamoto2019,kumamoto2020,rastello2020,banerjee2021a,trani2021,rastello2021}), globular clusters (GCs, e.g. \citealt{portegieszwart2000,downing2010,tanikawa2013a,rodriguez2015,rodriguez2016,rodriguez2018,samsing2014,samsing2018,zevin2019,antonini2020}), nuclear star clusters (NSCs, e.g. \citealt{oleary2009,miller2009,antonini2012,prodan2015,antonini2016,petrovich2017,gondan2018,rasskazov2019,arcasedda2018,arcasedda2019,arcasedda2020a,arcasedda2020b}) and AGN discs \citep[e.g.,][]{mckernan2012,mckernan2018,bartos2017,stone2017,yang2019,tagawa2020}.
\rev{CO mergers} from hierarchical triple systems were investigated in the context of field triples \citep{antonini2017,silsbee2017,toonen2018,rodriguez2018,vignagomez2021} and triples from binary-binary interactions in globular clusters \citep{antonini2016b,martinez2020,martinez2021,arcasedda2021}.
In this work we study the evolution of triple systems formed in low-mass young star clusters, focusing on triples composed of COs, either BHs, white dwarfs (WDs) or neutron stars (NSs).
Unlike previous studies, here we select triples formed self-consistently from $N$-body simulations, which include up-to-date stellar evolution \citep{mapelli2017} and regularized integration scheme for close encounters \citep{kus65}. Our focus is on triples from low-mass ($300$--$1000{\,\rm M_\odot}$) star clusters, which rapidly dissolve within $100 {\,\rm Myr}$. All the triples we consider have survived the dissolution of their parent cluster. Our study can be considered complementary to the ones of \citet{kimpson2016,britt2021}, who estimated the merger rate of in-cluster mergers of triples from open clusters.
In Section~\ref{sec:tripleprop} we describe the population of triples that we obtained from our direct-$N$ body simulations of low-mass young star clusters. Section~\ref{sec:GWmergers} discusses the numerical setup we use to follow the dynamical evolution of the triples. Section~\ref{sec:mergobj} presents the properties of merging BBH, BHNS and black hole -- white dwarfs (BHWD) binaries, including their merger rate density and their the mass distribution. Finally, we discuss and summarize our results in Section~\ref{sec:conc}.
\section{Triples' properties}\label{sec:tripleprop}
We select our hierarchical triples from the simulations presented in \citet{rastello2020} who performed a suite of $N$-body simulations using the direct-summation $N$-body code \texttt{NBODY6++GPU} \citep{wang2015} coupled with the population synthesis code \texttt{MOBSE}{} \citep{mapelli2017,giacobbo2018,giacobbo2018b,giacobbo2018c}.
\subsection{Low-mass star cluster simulations}
\texttt{NBODY6++GPU} is the GPU parallel version of \textsc{nbody6} \citep{aarseth2003} that implements a 4th-order Hermite integrator, Kustaanheimo-Stiefel regularization of close encounters
\citep{stiefel1965,kschain} and individual block time--steps \citep{makino1992}. No post-newtonian terms are included in the version of the code used in \cite{rastello2020}. \texttt{MOBSE}{} \citep{mapelli2017,giacobbo2018,giacobbo2018b}, is an upgrade of \texttt{BSE} \citep{hurley2000,hurley2002}, including up-to-date prescriptions for core-collapse supernovae, electron capture, stellar winds, pair instability and pulsational pair instability.
\rev{BH natal kicks are randomly drawn from a Maxwell-Boltzmann distribution with a root mean square velocity of $15 {\,\rm km/s}$. The natal kick velocity is reduced by $1-f_{\rm fb}$, where $f_{\rm fb}$ is the fraction of the fallback mass \citep{fry12}. The wind-mass loss rate for massive stars depends on the electron-scattering Eddington ratio, and considers the increase of the mass-loss rate when a star is close to the Eddington limit \citep{grafner2008,chen2015}. We do not take into account rotationally enhanced mass loss. Pulsational pair-instability supernovae and pair-instability supernovae are treated as in \cite{spera2017} and \cite{mapelli2020}.
All the remaining processes, such as tides, mass transfer, common
envelope and GW orbital decay, are implemented as in \citet{hurley2002}.
The assumptions of our stellar population synthesis model are summarized in Table~\ref{tab:bse}.
The above assumptions crucially affect the evolution of binaries, and therefore the formation and stability of triples in our simulations. In particular, higher-speed natal kicks would likely hinder the formation of stable triples in various ways. First, they would lead to a lowered retention of binaries and COs in the clusters, preventing them to form stable hierarchical triples. Second, natal kicks in stellar triples may lead to dynamical instability, resulting in the disintegration of the triple \citep{pijloo2012,perets2012a,lu2019}.
}
\citet{rastello2020} performed $100002$ direct $N$-body simulations of low-mass young star clusters exploring three different metallicities: $Z = 0.02$, $0.002$ and $0.0002$ ($33334$ simulations per each metallicity). The young star clusters have masses in the range $300\leq{} m_{\rm SC}/{\rm M}_\odot <1000$ sampled from a power-law distribution $dN/dm_{\rm SC}\propto m_{\rm SC}^{-2}$, reminiscent of the distribution of young star clusters in Milky-Way like galaxies \citep{lada2003}. The initial star cluster half mass radius $r_{\rm h}$ is chosen according to \cite{markskroupa12}:
\begin{equation}
r_{\rm h}=0.10^{+0.07}_{-0.04}\,{}{\rm pc}\,{} \left(\frac{m_{\mathrm{SC}}}{{\rm M}_{\odot}}\right)^{0.13\pm 0.04}
\end{equation}
Stellar masses are extracted from a Kroupa \citep{kroupa2001} initial mass function in the mass range $0.1 \le{} m \le{} 150 {\,\rm M_\odot}$.
The orbital parameters of original binaries are set following the distributions of \cite{sana2012}: the binary eccentricities $e$ are randomly drawn from a distribution $p(e)\propto{}e^{-0.42}$ with $0\leq{}e<1$ while the orbital periods $P$ follows the distribution $p(\Pi)\propto{}\Pi^{-0.55}$, where $\Pi\equiv{}\log_{10}(P/\mathrm{days})$ and $0.15\leq{}\Pi\leq{}6.7$.
The simulations have been performed adopting the rapid core-collapse supernova model \citep{fryer2012}, which prevents the formation of COs in the mass range $2-5$ M$_\odot$.
The simulated young star clusters initially host 40$\%$ original binaries \footnote{Here and in the following, \emph{original binaries} are stars already bound in a binary in the initial conditions.}.
Stars are randomly paired by using a distribution $\mathcal{P}(q)\propto{}q^{-0.1}$, where $q=m_2/m_1$ is the ratio of the mass between the secondary and the primary star according to \cite{sana2012}.
Hence, all the stars with mass $\,m\,\ge{}5 {\,\rm M_\odot}$ are members of binary systems, while stars with mass $m\,<\,5$~M$_\odot ${} are randomly paired until the imposed total binary fraction $f_{\mathrm{bin}}=0.4$ is reached. The result of this method is that the most massive stars (down to $5{\,\rm M_\odot}$) are all binary members, while the fraction of binaries falls to lower values for lighter stars, in agreement with \cite{moe2017}.
The simulated young star clusters are embedded in a solar neighbourhood-like static external tidal field and we put them on a circular orbit around the centre of the Milky Way at a distance $8\,\mathrm{kpc}$ \citep{wang2016}.
Each young star cluster is integrated for a maximum time $t=100\,\mathrm{Myr}$.
\begin{table}
\begin{tabular}{|lc|}
\hline
Wind mass loss: $\dot{M}(Z) \propto M^{\beta(Z)}$ & \citet{chen2015} \\
Supernovae model: rapid core-collapse & \citet{fryer2012} \\
Pair-instability supernovae & \citet{spera2017} \\
\multirow{2}*{Common envelope: $\alpha \lambda$ model} & \citet{webbink1984} \\
& $\alpha = 5$, $\lambda$: \citet{claeys2014} \\
Natal kicks: $\sigma = 15 \rm\, km/s$ + fallback & \citet{giacobbo2018b} \\
\hline
\end{tabular}
\caption{Properties of our binary stellar population synthesis model.}\label{tab:bse}
\end{table}
There are no primordial triples in the initial clusters, meaning that all the hierarchical triples we find in our clusters are dynamically formed through $4$+body encounters.
We select those triples that have escaped, meaning they have reached a distance from the star cluster's centre larger than twice its tidal radius.
Moreover, by 100 Myr the clusters' velocity dispersion has lowered down to $0.26$--$0.86 {\,\rm km/s}$, and the triples that remain in the clusters have an average velocity of $1.5 {\,\rm km/s}$. For this reason, we include in our analysis also the triples that have survived until the end of the integration at 100 Myr.
Consequently, all our triples have the same age as the clusters, 100 Myr.
By that time all of the most massive stars have already collapsed into BHs, and most mass loss by stellar winds has already occurred.
\rev{{We integrated the dynamics of the triples with the direct N-body code \texttt{Nbody6++GPU}. Furthermore, we fully took into account the orbital changes induced by stellar and binary evolution of the inner binary and the outer star, because we used our custom version of \texttt{Nbody6++GPU} coupled with \texttt{MOBSE}{} and \texttt{MOSSE} \citep[see][for more details]{dicarlo2019}, respectively. Therefore, important processes relevant to the evolution of triples, such as mass loss and dynamical instability, are treated self-consistently.}
Other processes specific to triples, such as triple common envelope \citep{glanz2021a}, tertiary tides \citep{gao2020} or tertiary mass transfers are not taken into account by \texttt{Nbody6++GPU}. However, these processes are important only for close triples \citep{toonen2020}, while the triples in our sample are very wide.}
We refer to \cite{rastello2020} and \cite{rastello2021} for further details on the star cluster simulations.
\subsection{Demography of triples with an inner-CO binary}
In the following, orbital quantities such as semimajor axis $a$ and eccentricity $e$ have the subscripts $_1$ and $_2$ when referred to the inner and outer orbit of the hierarchical triple. Quantities pertaining to the individual bodies, such as the mass $m$, have subscripts $_1$ and $_2$ when referring to the inner binary members, and $_3$ when referring to the outer body.
We first only select triples whose inner binary members are either BHs, NSs or WDs. Table~\ref{tab:triplesum} summarizes the number of triples with an inner BBH, BHWD or BHNS binary for each metallicity set.
Triples with an inner double neutron star are particularly interesting because of the possibility of characterizing the triple via neutron star pulsations \citep{suzuki2019,suzuki2021}. Unfortunately, we do not find inner double neutron stars in our sample.
The percentage of triples with an inner CO binary are $0.86\%$, $0.48\%$ and $0.15\%$ at $Z = 0.01, 0.1$ and 1 ${\,\rm Z_\odot}$, respectively. \rev{For brevity, hereafter we call ``inner-CO triples'' all the triples with an inner CO binary, and ``CO triples'' all the triples exclusively composed of COs.}
In about $25\%$ of the \rev{inner-CO triples}, the outer object is a NS or a BH. The frequency of the remaining stellar types are shown in Figure~\ref{fig:triplek3}, divided per metallicity set. Here, we label as main sequence (MS) objects with \textsc{BSE} type 0 or 1, white dwarf (WD) for \textsc{BSE} types 10, 11 and 12, and evolved star (EV) for \textsc{BSE} types 2, 3, 4, 5 and 6.
Most of the \rev{inner-CO triples} in our sample have a MS outer companion. In the Universe, such triples may be detected via astrometry measurements of the outer companion \citep{mashian2017,breivik2018,yamaguchi2018,yalinewich2018,shao2019,shikauchi2020,wiktorowicz2020} and subsequently confused as MS--BH binaries. Even so, radial-velocity monitoring might break the observational degeneracy between these two classes of objects \citep{hayashi2020a,hayashi2020b}. However, because the period of the outer binaries is $1.5 \times 10^4 {\,\rm yr}$ on average, triples formed in low-mass star clusters are too wide to be detected through astrometry. The same consideration applies to \rev{inner-CO triples} with an outer pulsar, which could be detected via pulsar arrival time analysis \citep{hayashi2021}.
\rev{On the other hand, depending on the local environment, wide triples may experience perturbations from flybys and the galactic potential. These may destabilize triples and trigger GW mergers \citep{michaely2020}. We leave this issue to future investigations, and focus on the evolution of the triples in isolation.
}
Figure~\ref{fig:tripleic} shows the masses and orbital parameters of all the \rev{CO triples} in our sample. The median semimajor axes for the inner and outer binary are $\langle a_1 \rangle \simeq 86 {\,\rm au}$ and $\langle a_2 \rangle \simeq 2700 {\,\rm au}$, with a median semimajor axis ratio of $\langle a_2/a_1 \rangle = 30$.
\rev{An indication of the dynamical origin of our triples is the eccentricity distribution of the outer orbit (Figure~\ref{fig:tripleic}, second panel). At low eccentricity, the distribution grows as a thermal distribution, which is the typical outcome of dynamical interactions \citep{antognini2016,leigh2016c,geller2019}. The cut-off at high eccentricities is linked to the stability of triples: if the outer orbit is too eccentric, the outer star will pass too close to the inner binary, destabilizing the inner orbit and leading to the disruption of the triple.
}
\rev{
In 98\% of all the triples the inner binary was an original binary at the beginning of the $N$-body simulations, indicating that the outer companion was acquired later. Restricting the sample to CO triples only, we find that in 48.5\% of the systems the inner binary was an original binary, in 49.5\% there is no relation between the triples' members, and in the remaining 2\% the outer object and one inner binary member were originally part of an original binary. This indicates that the CO triples, and triples in general, do not form via ``democratic'' binary-binary encounters, but rather via some different mechanisms. In fact, if triples were formed through ``democratic'' encounters, the two original binary members would have similar probabilities to become the inner binary or to break up and one star becoming the outer object. A possibility is that the outer object may be captured through a mechanism analogue to the capture of wide-orbit planets in dispersing clusters \citep{perets2012b}. Another possibility is that that most of the few-body encounters that are producing the triples are not ``resonant'' \citep{hut83a} or ergodic \citep{mon76a,mon76b}. This is unexpected but not unlikely, because the three-body problem is divided into chaotic motion and regular motion \citep{shevchenko2010}, and the statistical theories based on ergodicity can reproduce the results of numerical experiments only after discarding the latter (i.e. flybys and prompt interactions, see for example \citealt{stone2019,manwadkar2020,manwadkar2021,kol2021,ginat2021}).
Lastly, half of the CO triples do not contain both members of an original binary in any configuration. Inspecting some of such triples, we find that more than one member results from the merger of an original binary. This suggests that such triples may be formed during 4-body encounters between dynamically formed binaries, or even through more complex 4-body encounters not involving any original binary. Because here we focus mainly on GW sources from triples, we leave the investigation of triple formation mechanisms to future works.
}
\rev{
The presence of ${\sim}50\%$ of inner original binaries in the CO triples raises potentially interesting implications for the spins of COs and their alignment. Specifically, BH spins at merger might not be entirely uncorrelated with the orbital orientation, which has an impact on the spin effective parameter that can be measured from the GW signal. However, in this paper we abstain from making considerations on spin-orbit misalignment for the following two reasons. First, \texttt{MOBSE}{} does not follow the orientation of the spins in space. While stellar dynamics does not change the spin orientation, it can change the orientation of the binary orbital plane \citep[see][]{trani2021}. Furthermore, the version of \texttt{MOBSE}{} we used for these simulations does not model the spin of the CO remnants at birth.
Second, the inner-CO binaries in our sample are very wide. Because of this, tidal spin-up is inefficient, so any correlation between the binary spins would be due to `primordial' binary spin, whose evolution we cannot follow for the aforementioned reasons.
}
We expect all the triples obtained from the $N$-body simulations to be dynamically stable. We double checked the dynamical stability of the triples using the \citet{mard01} criterion:
\begin{equation}\label{eq:mardaars}
\frac{a_2}{a_1} > \frac{2.8}{1 - e_2} \left[(1+\frac{1}{q_2}) \frac{1 + e_2}{\sqrt{1 - e_2}}\right]^{2/5} \left(1 - 0.3 \frac{i_\mathrm{mut}}{\pi} \right)
\end{equation}
where $q_2 = (m_1 + m_2)/m_3$ is the mass ratio of the outer orbit. As expected, all the triples were found to be dynamically stable according to the above equation.
If the \rev{inner-CO} binaries were to merge only via GW radiation, their merger time can be estimated using the following expression from \citet{peters1964}:
\begin{equation}\label{eq:delay}
t_\mathrm{gw} = \frac{15 c^5}{304 G^3} \frac{a_1^4}{(m_1+m_2)\,m_1\,m_2}\,f(e_1)
\end{equation}
where $f(e)$ is a factor that takes into account the orbital eccentricity, which we evaluate numerically as:
\begin{equation}\label{eq:fe}
f(e_1) = \frac{(1-e_1^2)^4}{e_1^{\frac{48}{19}}(e_1^2 + \frac{304}{121})^{\frac{3480}{2299}}} \int^{e_1}_0 \frac{x^{\frac{29}{19}} (1 + \frac{121}{304}x^2)^{\frac{1181}{2299}}}{(1-x^2)^{3/2}} dx
\end{equation}
The median GW merger time for the inner binaries is ${\approx} 2 \times 10^{10} {\,\rm Gyr}$, that is an exceedingly long time with respect to the age of the Universe. Only about $2\times 10^{-3}$ of the triples would merge within 13.3 Gyr.
Fortunately, the outer object can shorten the merger timescale of the inner binaries via secular gravitational interactions, which we consider in the next Section.
\begin{table}
\centering
\caption{Summary of the triple systems we obtain from the \texttt{NBODY6++GPU} simulations.}
\label{tab:triplesum}
\begin{tabular}{lrrrr}
\hline
$Z$ & $N_\mathrm{tot}$ & $N_\mathrm{BBH}$ & $N_\mathrm{BHNS}$ & $N_\mathrm{BHWD}$ \\
\hline
$0.01{\,\rm Z_\odot}$ & 67793 & 396 & 101 & 81\\
$0.1{\,\rm Z_\odot}$ & 70372 & 247 & 47 & 43\\
$1{\,\rm Z_\odot}$ & 71661 & 76 & 3 & 29\\
\hline
\end{tabular}
\begin{flushleft}
\justify
Column~1: metallicity ($Z$);
column~2: total number of triples ($N_\mathrm{tot}$);
column~3: number of triples with an inner BBH ($N_\mathrm{BBH}$);
column~4: number of triples with an inner BHNS ($N_\mathrm{BHNS}$);
column~5: number of triples with an inner BHWD ($N_\mathrm{BHWD}$);
\end{flushleft}
\end{table}
\begin{figure}
\includegraphics[width=\columnwidth]{fractiontype.pdf}
\caption{Frequency of the stellar types of the outer object in triples with an \rev{inner-CO} binary. Panels from top to bottom: sets with $Z = 0.01, 0.1$ and $1 {\,\rm Z_\odot}$. MS: main sequence stars. BH: black holes. WD: white dwarfs. EV: evolved giant stars. NS: neutron stars. Refer to the main text for the precise \textsc{BSE} type each label corresponds to.}
\label{fig:triplek3}
\end{figure}
\begin{figure}
\includegraphics[width=\columnwidth]{icorbpars.pdf}
\caption{Distributions of initial orbital parameters for the \rev{CO triples}. The dark-shaded histograms indicate the distribution obtained from the original \texttt{NBODY6++GPU} simulations, while the light-shaded histograms were obtained from the Bayesian Gaussian mixture model described in Section~\ref{sec:numsetup}. From top to bottom: semimajor axis of the inner and outer orbit, $a_1$ and $a_2$; eccentricity of the inner and outer orbit, $e_1$ and $e_2$; mutual inclination $i_\mathrm{mut}$; masses $m_1$, $m_2$ and $m_3$. The distributions include all the three sets of metallicities combined.}
\label{fig:tripleic}
\end{figure}
\section{Evolution of hierarchical triple systems}\label{sec:GWmergers}
A hierarchical triple system evolves via secular exchanges of angular momentum between the inner and outer orbits, which drive cyclic oscillations in eccentricities and mutual inclination. This mechanism was originally studied by \citet{lid62} and \citet{koz62} with applications to Earth-orbiting satellites and asteroids in the solar system, respectively. Recently, \citet{ito2019} pointed out a long forgotten work by \citet{zeipel1910} that investigated this mechanism over 50 years before the works of Lidov and Kozai. Therefore, we refer to the secular exchanges of angular momentum in hierarchical triple systems as the von~Zeipel-Lidov-Kozai (ZLK) mechanism \citep[for a review and a book on the ZLK mechanism, see][]{naoz2016,shev2017}.
The ZLK mechanism is particularly important for our problem, because it can drive the eccentricity of the inner binary to extreme values. Together with GW radiation, the eccentricity increase may trigger the coalescence of the inner binary.
To qualify the role of the ZLK mechanism for our triples, we estimate the ZLK timescale as
\begin{equation}\label{eq:kozai}
T^\mathrm{quad}_\mathrm{ZLK} = \frac{P^2_2}{P_1} \frac{m_1 + m_2 + m_3}{m_3} (1 - e_2^2)^{3/2}
\end{equation}
where $P_1$ and $P_2$ are the periods of inner and outer orbit \citep[e.g.,][]{antognini2015}.
The median ZLK timescale is $T^\mathrm{quad}\approx10 {\,\rm Myr}$, which tells us that ZLK oscillations may play an important role in driving the evolution of our triples.
Equation~\ref{eq:kozai} refers only to the quadrupole-level interactions, which correspond to the second order term in the expansion of the three-body Hamiltonian. The next order in this approximation is the octupole-level interaction; this term can cause the inner binary to flip its orientation from prograde to retrograde. During these orbital flips, the inner binary reaches extremely high eccentricity \citep{naoz2013a}.
Therefore, octupole-level interactions can be crucial in triggering the GW coalescence of the inner binary. We estimate the timescale of octupole-level ZLK oscillations as in \citet{antognini2015}:
\begin{equation}\label{eq:kozai-lidov-octupole}
T^\mathrm{oct}_\mathrm{ZLK} = \frac{T^\mathrm{quad}_\mathrm{ZLK}}{\sqrt{\epsilon^\mathrm{oct}}}
\end{equation}
where $\epsilon^\mathrm{oct}$ is the ratio of the octupole-to-quadrupole level interaction terms:
\begin{equation}\label{eq:octupleterm}
\epsilon^\mathrm{oct} = \frac{m_1 - m_2}{m_1 + m_2} \frac{a_1}{a_2} \frac{e_2}{1-e^2_2}
\end{equation}
The octupole-level oscillations that are associated to inner orbit flips occur on a longer timescale with respect to the quadrupole-level oscillations. For our triples, $T^\mathrm{oct}_\mathrm{ZLK}\approx 3 {\,\rm Gyr}$, which is still less than the age of the Universe.
\subsection{Numerical setup}\label{sec:numsetup}
We study the evolution of all triples composed of BHs, WDs or NSs. In this way, we can safely neglect stellar evolution and focus only on triple dynamics.
The sample of \rev{CO triples} that we obtained from the $N$-body simulations is not sufficiently large to obtain a satisfying statistics on the number of GW mergers. We therefore resample the initial conditions using a Bayesian Gaussian mixture model with a Dirichlet process prior \citep{bishop2006}. We apply the model to 7 parameters of the triples: $m_1$, $m_2$, $m_3$, $a_1$, $e_1$, $a_2$, $e_2$. \rev{The mixture model is multivariate and allows us to keep the correlations among these 7 parameters.} \rev{We then sample the arguments of pericenter $\omega_1$, $\omega_2$ and the mutual inclinations $i_\mathrm{mut}$ uniformly in $\cos{(i_\mathrm{mut})}$ between $1$ and $-1$} to obtain the full\footnote{We also sample the longitude of the ascending node $\Omega_1$ in the $[0, 2\pi)$ interval, but its value does not affect the evolution of the triple.} set of parameters needed study the secular evolution of the triples. In this way we are able to preserve the correlations between the properties of the triples.
While resampling the triples, we make sure that they satisfy the stability criterion of Equation~\ref{eq:mardaars}. We also truncate the mass distributions according to the original mass upper limit, to avoid unrealistic mass values that may arise from the tails of the Gaussian mixture.
We generate $10^5$ realizations of \rev{CO triples} per metallicity, for a total of $3\times10^5$ triples.
Figure~\ref{fig:tripleic} compares the marginal distributions obtained from the Bayesian Gaussian mixture model with the original distributions. The model well reproduces the original distributions, including the low mass gap between BHs and the population of WDs and NSs.
We evolve each triple with the secular evolution code \textsc{okinami}. \textsc{okinami} evolves the double-average, octupole-level equations of motion derived from the 3-body Hamiltonian in Delaunay variables. The equations are integrated with a 7th order Runge-Kutta-Fehlberg integrator with adaptive timestep. \textsc{okinami} includes general relativity precession due to the post-Newtonian term PN1 and GW radiation from the post-Newtonian term PN2.5.
\rev{Adopting the secularly averaged equations allows us to integrate a larger number of triples for a longer time, in contrast with $N$-body methods. However, the secularly averaged equations cannot capture non-secular effects that might be important in modeling CO mergers. Specifically, triples with a weak hierarchy might undergo the so-called quasi-secular evolution, during which the binaries undergo oscillations on a timescale shorter than the secular timescale. This effect is related to the well-known problem of Lunar evection in celestial mechanics, but only recently it was incorporated in the more modern ZLK formalism \citep[see][]{cuk2004,luo2016}. Recent works have shown that quasi-secular evolution may underestimate the binary eccentricities during the GW inspiral \citep{antonini2012,antonini2016}. Consequently, this may overestimate the GW merger times and possibly underestimate the GW merger rates \citep{grishin2018,toonen2018}. The quasi-secular corrections terms for the double-averaged equations have been derived only in a specific reference frame \citep[e.g.][]{luo2016} or in the test-particle approximation \citep{cuk2004}; it is beyond the scope of this paper to derive and implement the correction terms in terms of Delaunay coordinates. For these reasons, our estimates on the merger rates, and especially on the rate of eccentric mergers, should be regarded as a lower limit.
}
We integrate our triples until either a merger occurs or the total integration time reaches 15 Gyr. A collision happens when $a_1 (1-e_1) < R_1 + R_2$, where $R_1$,$R_2$ are the radii of the inner binary members. For BHs, the radius is set to 50 times the Schwarzschild radius; for NSs we adopt a fixed radius of $10\,\rm km$; for WDs, we use equation 91 from \citet{hurley2000}.
It may occur that secular evolution brings the system out of dynamical stability. At each timestep we monitor the stability of the triple using Equation~\ref{eq:mardaars}. If the system does not satisfy the stability condition, we stop \textsc{okinami} and continue the integration using the few-body code \textsc{tsunami} \citep[see e.g.,][]{trani19b}. In order to convert from secular Keplerian orbital elements to Cartesian positions and velocities, we randomly sample the mean anomalies of the inner and outer binaries uniformly in $[0, 2\pi)$.
We stop the $N$-body integration when the chaotic triple breaks up into an unbound binary-single, or when a merger occurs. In the case of triple breakup, we sum the triple breakup time to the GW merger timescale of the escaping binary, and take it as the merger time.
\section{Properties of mergers}\label{sec:mergobj}
\subsection{Merger times and rates}\label{sec:mergrate}
Figure~\ref{fig:mergdist} shows the merger time (also called delay time) distribution for the triples with an inner BBH. The shape of the distribution is very similar for all triples at all metallicities, and is strongly peaked at ${<} 1 {\,\rm Gyr}$. From the merger time distribution, we can expect that triples born at higher redshifts will not contribute much to the local merger rate.
We find that the distributions of delay times are well represented by a mixture model composed of a Weibull distribution plus a flat component:
\begin{equation}\label{eq:mergertimedist}
p(x) = w_1 \,\frac{\alpha}{\beta} x^{\alpha-1} \exp{\left(-\left(\frac{x}{\beta}\right)^\alpha\right)} + w_0
\end{equation}
The curves in Figure~\ref{fig:mergdist} show the fit to Equation~\ref{eq:mergertimedist} performed with PyMC3, assuming a flat prior on ($\alpha$,$\beta$) and a Dirichlet distribution prior with unitary concentrations on $(w_1, w_2)$.
\begin{figure}
\includegraphics[width=\columnwidth]{mergdist_all.pdf}
\caption{Probability density distribution of the merger time for triples with an inner BBH (top panel), inner BHNS (middle panel) and inner BHWD (bottom panel). The histograms indicate the distribution obtained from the simulations run with \textsc{okinami} and \textsc{tsunami}. The solid lines are the Monte Carlo Markov Chain fit to Equation~\ref{eq:mergertimedist}. Triples merging with the ZLK mechanism have a very short ($<1{\,\rm Gyr}$) delay time.}
\label{fig:mergdist}
\end{figure}
We calculate the merger rate density as a function of redshift using the same approach as \citet{santoliquido2020}, in order to compare our results with those of \cite{rastello2020} and \cite{rastello2021}. In particular, the merger rate density in the comoving frame is
\begin{eqnarray}\label{eq:cosmorate}
\Gamma_i(z) = f_{\rm YSC} \frac{\rm d\quad{}\quad{}}{{\rm d}t(z)}\int_{z_{\rm max}}^{z}\psi(z')\,{}\frac{{\rm d}t(z')}{{\rm d}z'}\,{}{\rm d}z' \,{} \nonumber \\
\int_{Z_{\rm min}(z')}^{Z_{\rm max}(z')}\eta{}(Z)\,{}\mathcal{F}(z',z, Z)\,{}{\rm d}Z,
\end{eqnarray}
where the index $i$ indicates BBHs, BHNSs or BHWDs, $\psi(z')$ is the star formation rate density at redshift $z'$, $f_{\rm YSC}$ is the fraction of star formation rate that happens in low-mass young star clusters, $t(z)$ is the look-back time at redshift $z$, $Z_{\rm min}(z')$ and $Z_{\rm max}(z')$ are the minimum and maximum metallicity of stars formed at redshift $z'$, $\eta{}_i(Z)$ is the merger efficiency at metallicity $Z$, and $\mathcal{F}_i(z', z, Z)$ is the fraction of BBHs, BHNSs or BHWDs that form at redshift $z'$ from stars with metallicity $Z$ and merge at redshift $z$, normalized to all BBHs, BHNSs or BHWDs that form from stars with metallicity $Z$. To calculate the look-back time we take the cosmological parameters from \cite{planck2016}.
The merger efficiency is the total number of BBHs, BHNSs or BHWDs with delay time shorter than the Hubble time, divided by the total initial mass of their host star clusters. For the cosmic star formation rate density, we use the fit from \cite{madau2017}:
\begin{equation}\label{eq:madau}
\psi{}(z)=0.01\,{}\frac{(1+z)^{2.6}}{1+[(1+z)/3.2]^{6.2}}~\text{M}_\odot\,{}\text{Mpc}^{-3}\,{}\text{yr}^{-1}.
\end{equation}
Finally,
\begin{equation}\label{eq:Fz}
\mathcal{F}_i(z',z,Z)=\frac{\mathcal{N}_i(z',z,Z)}{\mathcal{N}_{\text{TOT\,{}i}}(Z)}\,{}p(z', Z),
\end{equation}
where $\mathcal{N}_i(z',z,Z)$ is the total number of BBHs, BHNSs or BHWDs that form at redshift $z'$ with metallicity $Z$ and merge at redshift $z$, $\mathcal{N}_{\text{TOT,\,{}i}}(Z)$ is the total number of BBHs, BHNSs or BHWDs with progenitor's metallicity $Z$, and
\begin{equation}
\label{eq:pdf}
p(z', Z) = \frac{1}{\sqrt{2 \pi\,{}\sigma_{\rm Z}^2}}\,{} \exp\left\{{-\,{} \frac{\left[\log{(Z(z')/{\rm Z}_\odot)} - {\langle{}\log{Z(z')/Z_\odot}\rangle{}}\right]^2}{2\,{}\sigma_{\rm Z}^2}}\right\},
\end{equation}
is the stellar metallicity distribution at a given redshift. We take the average metallicity $\langle{}\log{Z(z')/Z_\odot}\rangle{}$ from \cite{santoliquido2021} and assume a metallicity spread $\sigma_{\rm Z}=0.2$.
The resulting BBH merger rate density as a function of look-back time (or redshift) is shown in Figure~\ref{fig:rateshift}. As expected from the short delay time, the merger rate closely follows the evolution of the star formation rate density, which peaks at $11 {\,\rm Gyr}$ ($z\simeq2$). The largest contribution comes from triples at $0.1 {\,\rm Z_\odot}$, which have only a moderate merger fraction of $f^\mathrm{BBH}_\mathrm{merg} = 0.028$, but can still form at smaller redshifts. The local merger rate density of BBH is therefore:
\begin{equation}\label{eq:bbhmerg}
\Gamma_\mathrm{BBH} \simeq 0.60^{+0.84}_{-0.37} \,f_{\rm YSC} \,{\,\rm yr}^{-1} \,\rm Gpc^{-3}
\end{equation}
while for BHNSs and BHWDs, we find
\begin{equation}\label{eq:bhnsmerg}
\Gamma_\mathrm{BHNS} \simeq 0.11^{+0.23}_{-0.06} \,f_{\rm YSC} \, {\,\rm yr}^{-1} \,\rm Gpc^{-3}
\end{equation}
and
\begin{equation}\label{eq:bhwdmerg}
\Gamma_\mathrm{BHWD} \simeq 0.50^{+0.59}_{-0.27} \,f_{\rm YSC} \, {\,\rm yr}^{-1} \,\rm Gpc^{-3}
\end{equation}
respectively.
These merger rates can directly be compared to the ones from dynamically formed binaries, obtained from the same clusters. The binary channel has an expected local merger density rate of ${\sim} 28 {\,\rm yr}^{-1} \rm\,Gpc^{-3}$ for BHNSs \citep{rastello2020} and $88^{+34}_{-26} {\,\rm yr}^{-1} \rm\,Gpc^{-3}$ for BBHs \citep{rastello2021}, which are about 100 times higher than what we have estimated from triples.
\rev{
As stated earlier, the final rates depend on the precise value of $f_{\rm YSC}$, which is the fraction of star formation that occurs in clusters similar to the ones we have considered. This value is uncertain, but given an initial cluster mass function we can provide some rough estimates. Assuming a log-uniform mass distribution between 50 and 1000, $f_{\rm YSC} \approx 0.4$ \citep{lada2003}. On the one hand, the final rates may be obtained substituting $f_{\rm YSC} = 0.4$ in (\ref{eq:bbhmerg}), (\ref{eq:bhnsmerg}), and (\ref{eq:bhwdmerg}). On the other hand, this implicitly assumes that (a) the star formation in higher-mass cluster is negligible, and (b) clusters with $50$--$100 {\,\rm M_\odot}$ do not form CO triples, or their merger efficiency is nil.
}
The rates for triples are lower than those of binaries because of the lower merging efficiency. For example, the merger efficiency of dynamically formed BBHs at $Z=0.01{\,\rm Z_\odot}$ is about $1.4×10^{-5} {\,\rm M_\odot}^{-1}$, compared to $2.5×10^{-7} {\,\rm M_\odot}^{-1}$ for triples. The ratio of merger efficiencies between dynamically formed binaries and triples is about 100:1 at all metallicities, for both BBHs and BHNS mergers.
The merger rate might be underestimated because we only simulated triples where all members are COs. In fact, triples with an outer CO only account for 29\% of all the \rev{inner-CO triples} (Figure~\ref{fig:triplek3}). Assuming that the triples with an outer star have a similar evolution as the ones with an outer CO would increment the BBH merger rate to $\Gamma_\mathrm{BBH} \simeq 2.05 \,f_{\rm YSC} \, {\,\rm yr}^{-1} \,\rm Gpc^{-3}$. However, it is unlikely that triples with an outer MS star would lead to a similar merger fraction, because MS stars are less massive than the inner-CO binary, and also because possible mass loss (by stellar winds or mass transfer) would increase the semimajor axis of the outer orbit.
Our local merger rate densities are similar to those estimated in previous studies on field triples and triples from globular clusters. Specifically, the BBH local merger rate for field triples has been estimated to be $0.14$--$6$ \citep{silsbee2017}, $0.3$--$1.3$ \citep{antonini2017}, and $2$--$23 {\,\rm yr}^{-1} \,\rm Gpc^{-3}$ \citep{rodriguez2018}, while the local merger rate for triples in globular clusters is $0.4$--$1$ \citep{antonini2016} and $0.35$ ${\,\rm yr}^{-1} \,\rm Gpc^{-3}$ \citep{martinez2020}.
The various discrepancies among the above studies are to be attributed to different physical ingredients, most importantly the prescriptions for BH natal kicks, which can greatly affect the survival of \rev{CO triples}. Furthermore, some studies do not consider ZLK evolution during the progenitor stars' lifetime, which likely induces stellar mergers before the \rev{inner binary members can become COs}.
\begin{figure}
\includegraphics[width=\columnwidth]{rateshift_cr.pdf}
\caption{Merger rate density of BBHs (black), BHWDs (yellow) and BHNSs (red) as a function of redshift and lookback time. The shaded area indicates the 90\% confidence interval.
The merger rate density closely follows the star formation rate density because of the short delay time.}
\label{fig:rateshift}
\end{figure}
\subsection{Mass distributions}\label{sec:mergprop}
Even though the local merger rates from the triple channel are about 100 times smaller than those from the binary channel, the two channels might be disentangled using other merger properties, such as the masses and the mass ratios.
Figure~\ref{fig:m1_q_dist} shows the distribution of primary mass $m_\mathrm{p}$ and mass ratio $q = m_\mathrm{s} / m_\mathrm{p}$ for merging binaries.
The distribution of primary masses in merging binaries largely resembles the initial distribution, with small selection effects due to the ZLK mechanism. The primary masses can be substantially higher with respect to binary mergers from the same clusters \citep{rastello2021}. Furthermore, mergers from triples lack the primary mass peak at ${\sim}10 {\,\rm M_\odot}$, which is present in binary mergers and also matches the latest analysis of the GW Transient Catalogue data \citep{gwtc2020b,gwtc-3pop}.
Overall, binaries with massive primaries are more likely to merge than binaries with lower mass primaries, which is not surprising given that the GW coalesce timescale decreases for increasing masses. In addition, binaries with more massive primaries can have smaller mass ratios, which will increase the strength of the octupole ZLK mechanism\footnote{Sometimes referred as the eccentric ZLK mechanism \citep{naoz2016}.}.
This effect can be clearly seen in the bottom panel of Figure~\ref{fig:m1_q_dist}, which shows that mergers occur preferentially at low mass ratios. The importance of the octupole ZLK term for low mass ratio systems was also recently pointed out by \citet{suyubu2021}.
The comparison between the initial and the merging populations shows clearly that the merger fraction increases at low $q$, with a peak at $q=0.3$. This trend is opposite compared to the distribution of star cluster binaries, which instead follows the initial distribution, which decreases at low $q$.
Our result are in contrast with the results of \citet{martinez2020}, who find no difference in the mass distributions between globular cluster triples and globular cluster binaries. One possible explanation is that triples from globular clusters are more compact, and therefore the inner binaries can merge by GW radiation alone without the need of ZLK mechanism. Therefore, the selection effect on low $q$ and high $m_\mathrm{p}$ is less evident.
Another possible cause of this discrepancy is that our triples originate self-consistently from direct-$N$ body simulations that included stellar evolution and regularized close encounters \citep{wang2015}, while the triples in \citet{martinez2020} are the result of binary-binary scattering in isolation and without stellar evolution effects \citep{fregeau2003}. Additionally, our $N$-body simulations allow us to follow the evolution of in-cluster triples together with the rest of the cluster, while in \citet{martinez2020} the triples do not interact with the rest of the cluster, because of the limitations of the \textsc{cluster monte carlo} code \citep{cmcpaper}.
\begin{figure}
\includegraphics[width=\columnwidth]{m1_q_dist.pdf}
\caption{Distributions of primary mass $m_1$(top) and mass ratio $q = m_\mathrm{s} / m_\mathrm{p}$ (bottom) for merging \rev{CO triples}. The different contributions from the three metallicity are weighted according to the local merger rate density. Blue histogram: distributions for merging BBHs. Purple dashed line: distribution for merging BHNSs and BHWDs ($m_1$ distribution only). Grey lines: initial distributions. Red dot-dashed lines: distributions of merging BBH binaries from low-mass young star clusters from \citep{rastello2021}. Thin dotted lines: initial distributions of the original triples from \texttt{NBODY6++GPU}. Each distribution is normalized to unity. The ZLK mechanism in triples favors mergers with lower mass ratios, compared to mergers from binaries.}
\label{fig:m1_q_dist}
\end{figure}
\subsection{Eccentric mergers}
Given the extremely high eccentricity during the ZLK oscillations, merging binaries might retain some eccentricity when entering the observable GW bands. While no clear evidence for eccentric mergers exists to date, the LVK network at design sensitivity will be able to distinguish between eccentric and circular inspirals \citep{lower2018,huerta2018,gondan2019}. This will provide a new fundamental piece of information to discriminate among the astrophysical formation scenarios of GWs.
We investigate this by analyzing the last in-spiral phase of the inner binaries in our triple. Figure~\ref{fig:eccfreq} shows the eccentricity of inspiralling binaries as a function of the GW peak frequency, calculated as in \citet{wen2003}. Even though the inspiral phase begins with extremely high eccentricities, most binaries have substantially circularized before entering the LVK band at 10 Hz. At lower frequencies the eccentricity is much higher, especially towards the DECi-hertz Interferometer Gravitational-wave Observatory (DECIGO, peak sensitivity at 0.1 Hz, \citealt{decigo2011,decigo2018}) and the Laser Interferometer Space Antenna (LISA, peak sensitivity at 0.01 Hz) bands, where the eccentricity is close to 1. Note that binaries can enter the LISA band several times before merging, due to repeated ZLK oscillations \citep[e.g.][]{antonini2017,hoang2019,gupta2020}, but here we just display the last part of the inspiral.
The distribution of eccentricities at 10 Hz is shown in the top panel of Figure~\ref{fig:ecchz}, divided per binary type. The eccentricity at 10 Hz of BBHs is comparable to that of field triples and triples from globular clusters, whose range is $e_{\rm 10\,Hz} \simeq 10^{-4}$--$10^{-2})$. This also implies that only 7\% of the BBH mergers will have detectable eccentricities in the LVK band. As a caveat, here we may have missed about $10\%$ of highly eccentric inspirals ($1 - e\simeq 10^{-4}$), which emerge when using $N$-body methods, rather than the secular equations \citep{antonini2016b}.
Over 60\% BHNS have eccentricity greater than $e = 0.014$ at 10 Hz. The eccentricity of BHNS binaries is significantly higher, as expected if their merger was driven by the octupole-level interactions. In fact, the BHNS mergers from our cluster triples have similar eccentricities to BBH mergers from in-cluster captures \citep{rodriguez2018b}. This can constitute an important diagnostic to distinguish BHNS mergers from hierarchical triples and those from three-body encounters in young star clusters \citep{rastello2020}. The radius of WDs is much larger than NSs, so that BHWD binaries merge before reaching the 10 Hz band, but can be detected at lower frequencies.
At lower frequencies, the eccentricity distribution shifts to higher values. At ${\sim}1$ Hz, the eccentricities of BBH, BHNS and BHWD is comparable, lying in the range $e_{\rm 1\,Hz} \simeq 10^{-3}$--$0.5$. As expected, the eccentricity of high mass ratio binaries like BHNS and BHWD is higher than that of BBH.
This is more manifest at 0.01 Hz, where the eccentricity distribution of all populations becomes bimodal, separated into a low-eccentricity component at $e_{\rm 0.01\,Hz} \simeq 10^{-3}$--$0.5$ and a high-eccentricity component at $1-e_{\rm 0.01\,Hz} \simeq 10^{-3}$--$10^{-5}$.
The eccentricity of merging BHWDs and BHNSs determines more than just the GW waveform, because it affects the properties of the merger remnant and of the possible electromagnetic counterpart \citep{fernandez2016,zenati2020}. If the eccentricity at merger is sufficiently high, the merger becomes essentially a head-on collision. For example, head-on collisions between WDs in hierarchical triples have been proposed as a detonation mechanism for type Ia supernovae \citep{raskin2009,hawley2012,katz2012,papish2016}. However, the consequences of low-impact parameter BH--WD and BH--NS collisions have not been explored yet.
Low eccentricity mergers can instead produce tidal disruption events (TDEs) of WDs. Such WD TDEs might appear as high-energy transients associated with gamma-ray emissions \citep{krolik2011,ioka2016,fragione2020}, and thermonuclear transients \citep{rosswog2009,tanikawa2017,kawana2018,anninos2018}.
\begin{figure}
\includegraphics[width=\columnwidth]{eccfreq.pdf}
\caption{Orbital eccentricity as a function of GW peak frequency for the merging binaries in our simulation. The dots indicate the beginning of the evolutionary track, whose colour depends on whether they are BBHs (blue), BHNSs (green) or BHWDs (orange). As the binaries spiral in, they circularize due to GW radiation.}
\label{fig:eccfreq}
\end{figure}
\begin{figure}
\includegraphics[width=\columnwidth]{ecchz.pdf}
\caption{Distributions of orbital eccentricities at different GW peak frequencies. From top to bottom: 10 , 1 , 0.1 and 0.01 Hz. The colours indicate type of binary: BBH (blue), BHNS (green) or BHWD (orange). Note that BHWD binaries merge before reaching the 10 Hz band. The vertical lines in the top panel indicate the minimum detectable eccentricity for LIGO+Virgo (LV), Einstein Telescope (ET) and Cosmic Explorer (CE) as estimated by \citet{lower2018}.}
\label{fig:ecchz}
\end{figure}
\section{Summary and conclusions}\label{sec:conc}
Hierarchical triple systems are composed of a binary orbited by an outer companion star. Such systems are ubiquitous in the Universe, whether they are formed in the field or through dynamical interactions.
The presence of an outer companion gives rise to secular exchanges of angular momentum between the inner and the outer binary. These exchanges of angular momentum manifest themselves as periodic oscillations in the eccentricity and inclination of the inner binary -- the so called ZLK oscillations.
The increase in the inner binary eccentricity can be the key to explain a variety of observable phenomena. Here, we studied the case in which the increase of the eccentricity can lead to the rapid coalescence of the inner binary via GW emission.
In this work, we have considered the evolution of hierarchical triple systems with an inner BBH, BHNS or BHWD. The triples were formed self-consistently via dynamical interactions in stellar clusters, which were modeled with direct $N$-body simulations that included up-to-date stellar evolution and accurate integration of close encounters. In particular, we focused on triples from low-mass young star clusters with an initial mass between $300$ and $10^3 {\,\rm M_\odot}$ at three different metallicities ($Z=1$, $0.1$ and $0.01{\,\rm Z_\odot}$, \citealt{rastello2020,rastello2021}). The clusters were evolved up to 100 Myr, at which point they were entirely disrupted by the Galactic tidal field. We then selected the triples composed of COs that were stable and isolation. To obtain a better statistics on CO mergers, we resampled the distributions of the triple's properties using a Bayesian Gaussian mixture model. We finally evolved the triple systems for 15 Gyr using the secularly averaged equations at the octupole-level expansion, including PN1 and PN2.5 corrections for the inner orbit.
We find that ZLK oscillations are a crucial mechanism to trigger the merger of the inner binaries in triples from low-mass young star clusters: only 0.2\% of the binaries would have merged within 13.3 Gyr by GW radiation alone. In contrast, ZLK oscillations cause the inner binaries to merge very quickly: about 90\% of the mergers occur within ${\sim} 1 {\,\rm Gyr}$ (Figure~\ref{fig:mergdist}). Because of the short delay times, the merger rate density closely traces the star formation density history (Figure~\ref{fig:rateshift}).
We derive a local merger rate density of $0.60^{+0.84}_{-0.84} \,f_{\rm YSC}{\,\rm yr}^{-1} \,\rm Gpc^{-3}$ for BBHs, $0.11^{+0.23}_{-0.23} \,f_{\rm YSC} {\,\rm yr}^{-1} \,\rm Gpc^{-3}$ for BHNSs and $0.50^{+0.59}_{-0.59} \,f_{\rm YSC} {\,\rm yr}^{-1} \,\rm Gpc^{-3}$ for BHWDs. The rates for BBHs and BHNSs are about 100 times lower than those of binary mergers from the same clusters. \rev{The origin for this difference stems from the merger efficiency of triple systems, which is about $\sim{100}$ times lower than that of binaries.}
Compared to BBH mergers from open cluster binaries, BBH mergers from triples have more massive primaries, with the $m_\mathrm{p}$ distribution peaked at around ${\sim}30 {\,\rm M_\odot}$ rather than ${\sim}10 {\,\rm M_\odot}$ -- the latter value being more consistent with the latest observational data \citep{gwtc2020b,gwtc-3pop}.
Another distinctive trait of BBH mergers from triples is the distribution of mass ratios $q=m_\mathrm{s}/m_\mathrm{p}$. In contrast to the cluster binaries pathway, which favours equal mass binaries, the mass ratio distribution for cluster triples peaks at $q = 0.3$ (Figure~\ref{fig:m1_q_dist}). This is caused by the ZLK mechanism, whose eccentricity-pumping effect is enhanced at low mass ratios.
Finally, we show that many BBHs, BHWDs and BHNSs merging through this mechanism will have detectable eccentricities in the LVK, ET and LISA bands. We expect the eccentricity of merging BHNSs in the LVK band to be higher than that of BBHs in the triple scenario (Figure~\ref{fig:ecchz}). The eccentricities in the observable frequencies might be even higher than our estimate, because we assumed secular approximation, which has been shown to underestimate the eccentricity at merger.
\rev{ Another possible way to identify CO mergers from triple system is from the possible electromagnetic counterpart of BHNS and BHWD mergers. We show that BHWD mergers can occur at both high and low eccentricities, giving rise to TDEs and head-on collisions. The outcome of head-on collisions in BHNS and BHWDs binaries remains to be investigated. }
In conclusion, gravitational interactions in hierarchical triple systems are an important pathway to CO mergers. We have shown that triple systems formed in open clusters can contribute, albeit in a minor part, to the observed GW event rate. Here, we have presented the main properties -- merger rates, masses and eccentricities -- that might help disentangle the origin of present and future events.
In our upcoming work, we will extend our analysis to the \rev{non-CO triples} in our sample, and compare them with field triples.
\section*{Acknowledgements}
We thank Hagai Perets for
a constructive and well thought review.
This work received support from JSPS KAKENHI Grant Numbers 17H06360, 19K03907 and 21K13914.
MM, UNDC, SR and FS acknowledge financial support from the European Research Council for the ERC Consolidator grant DEMOBLACK, under contract no. 770017. AAT would like to thank Naoki Yoshida and Yasushi Suto for insightful discussions. SR thanks Giuliano Iorio for suggestions on data exploration.
\section*{Data Availability}
The \textsc{okinami} code, the initial conditions and the simulation data underlying this article will be shared on reasonable request to the corresponding author.
\bibliographystyle{mnras}
|
\section{Summary}
Acoustic emission (AE) is one of the most promising non-destructive techniques for its ability to detect changes in the integrity of materials at the microscale. When a damage occurs, an elastic wave is released which propagates onto the surface creating displacements ranging from picometers to nanometers. The displacements are detected by highly sensitive piezoelectric sensors permanently attached onto a material. Data are continuously collected at sampling rate around 5 MHz. The obtained AE data stream (time-series) is then processed by algorithms to detect AE signals related to damages.
AE is used in many laboratories for materials characterization \cite{app11156718} and in industrial applications for real time monitoring of manufacturing process \cite{shevchik2019deep} or storage facilities \cite{ai2021source}. Despite many publications on the topic, the precise identification of the source of AE signals remains a challenge. Indeed, AE source identification is an inverse problem which is difficult to solve due to the sensitivity of the sensors which provide many AE signals corrupted by noise, as well as due to the effects of wave propagation in damaged materials which creates wave scattering \cite{Schubert}.
The lack of prior knowledge on AE sources compels AE users to interpret AE by means of unsupervised learning algorithms which include the following steps:
\begin{enumerate}
\item AE signals detection: Also called wave-picking (like in seismology \cite{li2018machine}), this step aims at processing the AE data stream to keep only relevant signals, defined as being above a given threshold (minimum amplitude). A preprocessing step can be included using wavelet denoising in noisy cases \cite{Pomponi2015110,kharrat2016signal}.
\item AE signals clustering: The signals obtained in the previous step fed a clustering method which aims at creating groups (clusters) of AE signals \cite{zhou2018cluster}.
\item Clustering validation: The validation of the results obtained in the previous step is performed in a subjective manner because of the lack of prior knowledge of AE sources. To circumvent this problem, AE users apply different clustering strategies in the previous step and then compute clustering validity indices \cite{vendramin2013combination,vendramin2010relative} which provide a scalar value that allows end-users to compare different results. The results can also be visually analyzed, like in interactive clustering \cite{bae2020interactive}, to evaluate whether the clusters can be related to AE sources with a physical meaning \cite{sendrowicz2021challenges}.
\end{enumerate}
{{An alternative to these three-steps strategy is based on anomaly detection \cite{martin2017online,nasir2021review}. A baseline is considered to train a one-class supervised learning algorithm which can then be used to generate a health indicator for online monitoring \cite{FarrarBook}. Anomaly detection was also exploited for clustering in \cite{RAMASSO2020103478}.}}
In real applications or for complex materials, AE signals detection (step 1), clustering (step 2) or validation (step 3), can be difficult. In that case, specific algorithms are developed. However, there is no common AE data sets {{that can serve as references (benchmark)}} to compare the methods of the literature. This work presents a {{benchmark}} data set that can be used for this purpose.
The experiment was designed to reproduce the loosening phenomenon observed in aeronautics, automotive or civil engineering structures where parts are assembled together by means of bolted joints. {{During the service life of a bolted structures, the torque can evolve according to operational conditions such as loading, in particular when these structures are submitted to vibrations. Monitoring the loosening condition in bolted structures during operation remains challenging because contact and friction in bolted joints induce a nonlinear stochastic behavior \cite{chen2022quantitative,li2020monitoring,nikravesh2017review,zhang2016quantitative,asamene2012analysis,amerini2011structural}. A common solution consists in measuring the pretightening force using a strain gauge bonded on the top surface of a bolt head \cite{wang2020bolt}. This is a local approach since we need to instrument every bolt. On the other hand, global techniques make use of only a few sensors for bolt condition monitoring. There are two categories of global techniques: active and passive approaches. Active approaches such as lamb waves were used in the past and consists in using a transmitter of high harmonics waveforms tuned to be sensitive to the changes in the contacts. The receiver is used to collect the waveforms after their propagation in the bolted structure. Machine learning can then be used to build an anomaly detector like in \cite{FarrarBook} and \cite{nasir2021review}. Conversely, passive approaches do not use an external source, but the transmitter is actually the structure itself (if a damage occurs). Therefore, a continuous reading of the piezoelectric sensors is necessary in order to not miss any damage or event. The use of AE sensors for condition monitoring of bolts is not frequent. For example, in recent works \cite{zhang2019continuous,imac_bolts}, it was shown that high sensitivity piezoelectric sensors used in this technique are able to detect AE signals despite low signal-to-noise ratio in bolted joints under vibration. }}
In the context of stationary and harmonic vibration tests, the bolts can be subject to self-loosening under vibrations, eventually leading to a critical failure of the assembly. Therefore, it is of paramount importance to develop sensing strategies and algorithms for early loosening estimation. The main characteristics of this data set are the following:
\begin{itemize}
\item During the tests, the tightening level of one of the bolts was controlled to reach seven different values, from highly tightened to loosened. Therefore, the end-user knows that the tightening level has been set to a specific value for some periods of the data stream. This reference can be used either to train supervised (deep/transfer) learning methods (with possibly 7 classes) and for unsupervised or semi supervised learning. This reference can thus be helpful to validate steps 2 and 3 in the aforementioned methodology.
\item The data set is made of raw acoustic emission data streams acquired by three different sensors and a laser vibrometer. Multisensor data fusion can thus be performed and evaluated using the reference. This can be helpful to improve algorithms in steps 2 and 3.
\item The vibrometer data also allows use users to evaluate the performance of wave-picking algorithms with low signal-to-noise ratio. This can be helpful to validate step 1 in the aforementioned methodology.
\end{itemize}
\section{Data Description}
The data set was obtained on a test rig called ORION \cite{imac_bolts,teloli2022good}. It is constituted of a jointed structure, dynamically loaded with a vibration shaker and monitored with acoustic emission, force, and velocity sensors.
For each series of measurements, four sensors were used simultaneously: a laser vibrometer and three different AE sensors (micro80, F50A, micro200HF), each with a given frequency band, {{all sampled at 5 MHz}}, yielding approximately between 1.4 and 1.9 GB of data. A total of five series of measurements were performed denoted as {\it measurementSeries\_B}, {\it measurementSeries\_C}, {\it measurementSeries\_D}, {\it measurementSeries\_E} and {\it measurementSeries\_F}.
The organization of repositories and files is depicted in Figure \ref{lkdmedkzmedze} for series of measurements B. There is one folder for each series of measurements. The ORION-AE data set is thus composed of five folders. For each series of measurements, there are seven subfolders corresponding to seven tightening levels: $60$, $50$, $40$, $30$, $20$, $10$, $05$~cNm, except for {\it measurementSeries\_C} for which $20$~cNm is missing.
\begin{figure}[!ht]
\centering
\includegraphics[width=0.9\linewidth]{repofiles-crop.pdf}
\caption{Tree structure of the repositories and files exemplified for series of measurements "B". {{The root directory on Dataverse also contains two additional repositories: One with images illustrating the content of the data in each level and campaign, and one with the sensors datasheet. Finally a Matlab script is provided to read the files and correctly extract the timestamps.}}
\label{lkdmedkzmedze}}
\end{figure}
It is important to \textbf{read the subfolders in this precise order}, otherwise the physical meaning of the data becomes wrong. Th upper bolt have been untightened with a precise sequence. For example, reading from $5$ cNm to $60$ cNm (reverse order) does \textit{not} correspond to the "tightening" of bolts since the process is not reversible in terms of acoustic emission characteristics.
Each subfolder is made of {\it .mat} files generated using MATLAB 2016b. There is about one file per second. The files in a subfolder are named according to the timestamps (time of recording) and must \textbf{be read in the correct order}. Each file has the following format:
\begin{center}
\textsf{salves\_out\_\textbf{X}cNm\_\textbf{Y}\_Fs5MHz\_Tf1s\_2019-04-02-\textbf{HH}-\textbf{MM}-\textbf{SS}\_1.mat}
\end{center}
\noindent where $\mathbf{X} \in \{5,10,20,30,40,50,60\}$ representing the tightening level, $\mathbf{Y} \in \{B,C,D,E,F\}$ for the series of measurements. The production date is "2019-04-02" and the timestamp began at HH-MM-SS. For example,
\begin{center}
\textsf{salves\_out\_05cNm\_B\_Fs5MHz\_Tf1s\_2019-04-02-16-22-09\_1.mat}
\end{center}
\noindent represents AE data for 5 cNm in series of measurements B taken at 16:22:09. Each {\it .mat} file is composed of vectors of data with the following names:
\begin{itemize}
\item Variable "A" corresponding to micro80 sensor data (sampling frequency: 5MHz).
\item Variable "B" corresponding to F50A sensor data (sampling frequency: 5MHz).
\item Variable "C" corresponding to micro200HF sensor data (sampling frequency: 5MHz).
\item Variable "D" corresponding to laser velocimeter data (sampling frequency: 5MHz).
\end{itemize}
{{The number of samples is about $5\times 4$ millions per file (each file corresponds to about 1 second), with all sensors data sampled at 5 MHz.}}
The labels corresponding to the sensors must not be confused with the labels corresponding to the test campaigns. Illustrations of the data are given in Figure \ref{fig:ejizefiozjifojzB2} for sensor "F50A" (variable "B" in a file) in {\it measurementSeries\_B}. About 6 periods are depicted (using the red curve representing the displacement of the upper beam).
\begin{figure}[hbtp]
\subfigure[5 cNm \label{fig:f1seqA1}]{\hspace{-0.5cm}\includegraphics[width=0.8\textwidth]{sample_zoom_sensorB_level_05cNm_measurementSeries_B-eps-converted-to.pdf}}\\
\subfigure[10 cNm \label{fig:f1seqA2}]{\hspace{-0.5cm}\includegraphics[width=0.8\textwidth]{sample_zoom_sensorB_level_10cNm_measurementSeries_B-eps-converted-to.pdf}}\\
\caption{\emph{Cont}. }
\end{figure}
\unskip
\begin{figure}[hbtp]\ContinuedFloat
\setcounter{subfigure}{2}
\subfigure[20 cNm \label{fig:f1seqA3}]{\hspace{-0.5cm}\includegraphics[width=0.8\textwidth]{sample_zoom_sensorB_level_20cNm_measurementSeries_B-eps-converted-to.pdf}}
\subfigure[30 cNm \label{fig:f2seqA}]{\hspace{-0.5cm}\includegraphics[width=0.8\textwidth]{sample_zoom_sensorB_level_30cNm_measurementSeries_B-eps-converted-to.pdf}}
\subfigure[40 cNm \label{fig:f1seqA4}]{\hspace{-0.5cm}\includegraphics[width=0.8\textwidth]{sample_zoom_sensorB_level_40cNm_measurementSeries_B-eps-converted-to.pdf}}
\caption{\emph{Cont}. }
\end{figure}
\unskip
\begin{figure}[hbtp]\ContinuedFloat
\setcounter{subfigure}{5}
\subfigure[50 cNm \label{fig:f2seqA5}]{\hspace{-0.5cm}\includegraphics[width=0.8\textwidth]{sample_zoom_sensorB_level_50cNm_measurementSeries_B-eps-converted-to.pdf}}
\subfigure[60 cNm \label{fig:f1seqA6}]{\hspace{-0.5cm}\includegraphics[width=0.8\textwidth]{sample_zoom_sensorB_level_60cNm_measurementSeries_B-eps-converted-to.pdf}}
\caption{{({a}--{g})} {Sample of the data in measurement} series ``B'' with sensor ``F50A'' for each tightening level. Blue curve: AE signal; red curve: laser vibrometer data (harmonic vibration signal at 100 Hz with amplitude control).
\label{fig:ejizefiozjifojzB2}
}
\end{figure}
Figure \ref{fzefzefzef33133} depicts the acoustic emission and laser vibrometer data, superimposed with the labels (tightening levels) evolving from 1 to 7, for measurement $B$. This figure (red curve) shows that the \textbf{control of the displacement of the beam by a feedback from the laser vibrometer} (as explained in the "Method" Section) leads to a similar displacement during a whole test (about $70$ s). In this figure, the green curve representing the acoustic emission shows that the amplitude can not be used as a feature to discriminate the levels.
\begin{figure}[!ht]
\centering
\includegraphics[width=0.9\linewidth]{resultsmeasurementSeries_B.pdf}
\caption{Tightening levels, acoustic emission and laser vibrometer data superimposed for measurements "B" and sensor micro-200-HF (variable C). Due to the amount of data involved, a subsampling by a factor of 10 was applied for each file. The x-axis is here represented using the time of test (about $70$ s), starting from $60$ cNm from the left (around between $t \in [0,10]$ s) to $5$ cNm on the right (from $t > 60$ s).
\label{fzefzefzef33133}}
\end{figure}
It is important to note that the real timestamps in filenames must {not} be taken as the timestamps of the AE signals in a data stream. The real timestamps must be calculated from the exact number of samples in each file, starting from $t=0 \,s$ for $60$ cNm. There are about $10$~s of continuous recording of data per level and for the four sensors. The exact duration of a period can be found according to the number of files in each subfolder, the number of points per file and the sampling frequency. The precise values are given in Table \ref{dlkmefeez}. A Matlab code (\textsf{ORION\_AE\_sample\_read\_files.m}) is provided in the data directory to read the files and reproduce the duration of each period.
\begin{table}[hbtp]
\centering
\begin{tabular}{|c||c|c|c|c|c|c|c|}
\hline
Meas. series & 60 cNm & 50 cNm & 40 cNm & 30 cNm & 20 cNm & 10 cNm & 5 cNm \\
\hline\hline
B & 10.0000 & 10.0000 & 10.0000 & 9.0170 & 10.0000 & 10.7725 & 9.2275 \\
\hline
C & 10.0000 & 10.0000 & 10.0000 & 10.0000 & N.A. & 10.0000 & 10.0000 \\
\hline
D & 10.0000 & 10.0000 & 10.8119 & 9.1881 & 10.0000 & 10.0000 & 10.0000 \\
\hline
E & 10.0000 & 10.0000 & 10.0000 & 10.0000 & 10.0000 & 10.0000 & 10.0000 \\
\hline
F & 10.0000 & 10.0000 & 10.0000 & 10.0000 & 10.0000 & 10.0000 & 10.0000 \\
\hline
\end{tabular}
\caption{Duration in seconds of each period for a given tightening level. The exact duration of a period can be found according to the number of files in each subfolder, the number of points per file and the sampling frequency. This table can be reproduced using a A Matlab code (\textsf{ORION\_AE\_sample\_read\_files.m}) provided in the data directory. \label{dlkmefeez}}
\end{table}
\clearpage
\section{Methods}
The ORION test rig (Figure \ref{fig:setup}) has been designed for the study of vibration damping and makes the vibration tests highly reproducible as shown in \cite{imac_bolts,teloli2022good}. It is a jointed structure made of two plates manufactured with 2024 aluminium alloy, linked together by three M4 bolts (Figure \ref{fig:setup}). Contact patches are machined overlays which have been added at each bolt connection to retain the contact between both beams in a small area, minimizing uncertainties on the structure's response and enhancing the repeatability between measurements. The contact patches have an area of $12\times 12$ mm$^2$ and are $2$ mm thick. The central bolt is dedicated to "static" functions, i.e. to ensure structural integrity and provide resistance to dynamic loads without substantial stiffness changes. The external bolts, in turn, perform "damping" functions, i.e. to increase energy dissipation due to frictional contact. It is worth noticing that these two functions have to be monitored in order to ensure the integrity (central bolt) and the damping (lateral bolts).
\begin{figure}[!h]
\centering
\includegraphics[width=0.9\linewidth]{setupORIONv2-crop.pdf
\caption{Setup description: part dimensions, sensors positions, bolts positions}
\label{fig:setup}
\end{figure}
The harmonic excitation load was applied using a Tira TV51120, electromagnetic shaker which can deliver a $200 N$ force. The structure was submitted to a $100$ Hz harmonic excitation force during about $10$~s. The force was measured using a PCB 208C02 piezoelectric force sensor.
The vibration magnitude was measured using a Polytec PSV500Xtra, laser vibrometer. The obtained laser data allowed us to make sure that the amplitude of the displacement of the top of the upper plate remains constant for all tightening levels.
Seven tightening levels, corresponding to seven operating conditions of the structure under vibration, were applied on the upper bolt. The tightening level was first set to $60$~cNm with a torque screwdriver. After about a $10$~s vibration test, the shaker was stopped and this vibration test was repeated after a torque modification at $50$ cNm. Then torque modifications at $40$, $30$, $20$, $10$ and $5$~cNm were applied. The change in the tightening level was made using a CDI torque screwdrivers. The torque was checked three times at each change.
The acoustic emission sensors used were a "micro-200-HF", a "micro-80" and a "F50A", made by Euro-Physical Acoustics. They were connected to a preamplifer set to 60 dB (model 2/4/6 preamplifier made by Europhysical acoustics). Their detailed characteristics, such as dimensions and frequency bands, are provided in the datasheets enclosed in the data set, with a summary of the main ones given in Table \ref{mldkomdkezd}. The sensors were attached onto the lower plate (5 cm above the end of the plate) using a silicone grease. All data were sampled at 5 MHz using a Picoscope 4824 (20 MHz bandwidth, low noise, 12-bit resolution, 256 MS buffer memory, USB~3) connected to a linux PC with MATLAB\textsuperscript{\tiny\textregistered} 2016b and the Data Acquisition\textsuperscript{\tiny\textregistered} toolbox. The datasheet of the Picoscope is provided in the data set.
\begin{table}[hbtp]
\begin{tabular}{|c||c|c|c|c|}
\hline
& Units & F50A & micro80 & micro200HF \\
\hline
\hline
Peak Sensitivity, Ref V/(m/s) & dB & 65 & 57 & 62 \\
\hline
Operating Frequency Range (kHz) & kHz & 200-800 kHz & 200-900 & 500-4500 \\
\hline
Resonant Frequency, Ref V/(m/s) & kHz & 100 kHz & 250 & 2500 \\
\hline
\end{tabular}
\caption{Frequency and sensitivity of the acoustic emission sensors used. The data sheets are provided in the data set. \label{mldkomdkezd} }
\end{table}
\section{User notes}
\subsection{About the ground truth}
In ORION-AE data set, we know that, during each period of $10$~s, the tightening level has been set to a specific value. The tightening level can be considered as constant during each period, and therefore can be used as a reference.
However, during each cycle of the vibration test for a given tightening level, different AE sources can generate AE signals and those sources may be activated or not, depending on the tribological conditions which are not known. The tightening level may thus slightly change during the period of $10$ seconds. The tribological conditions may also be different from a campaign to another. Despite this variability, preliminary results shown in \cite{clustersSequence} demonstrated that the content of AE signals in those series of measurements is sufficient to discriminate between the periods, which seem to corroborate that, first, the AE sources have indeed a signature depending on the tightening level, and second, that the tightening level and the amplitude of the vibrations (controlled by a closed-loop using the laser vibrometer) remains constant during the period of $10\,s$.
\subsection{ORION-AE for AE signal detection} %
In addition to machine learning tasks, this data set can be used for signal processing and in particular \textit{wave-picking} (step 1, presented in introduction). As for seismic wave detection \cite{baillard2014automatic}, this step aims at finding relevant signals in a data stream. Wave-picking is one the most important task in acoustic emission analysis because it paves the way for feature extraction, classification, localization of acoustic emission sources and interpretation by machine learning.
In ORION-AE, the AE sources are related to tribological phenomenon characterized by low amplitude and with low signal-to-noise ratio. Denoising methods are thus generally used as a preprocessing step. Figure \ref{dejfdlkjzelkjzkej11} depicts the result of wavelet denoising with 14 levels of decomposition and using a Daubechies "db45" wavelet (the same settings as \cite{kharrat2016signal}). {{This was done for two sensors in the same chunk of data as shown in Fig.~\ref{dsefefe1111} and \ref{dsefefe2222}. It can be observed that the responses of the sensors differ since they have different characteristics (in particular the sensitivity and frequency band). We can also observe the presence of AE signals that occur at each cycle with some regularity.
We took the starting time, called onsets, of each signal detected after wavelet denoising (from Figure \ref{dejfdlkjzelkjzkej11} using \cite{kharrat2016signal}). Onsets were then reported onto the vibrometer data in Figure \ref{dejfdlkjzelkjzkej333}. Each blue circle corresponds to one AE signal detected. The vertical position represents the amplitude of the displacement (measured by the vibrometer) when the signal was detected. We can observe that the amplitude is very similar between cycles (there is about 1 second of test represented on this figure), which means that the activation of the AE sources is highly reproducible between cycles. Similar results have been observed on other levels and considering measurements B to F. However, we also observed that several AE sources can occur in some levels, and that the average amplitude of displacement can change. It means that the instant of activation of the AE sources (related to the signals) seems to depend on the tightening levels. These observations can be dependent on the wave-picking algorithm used (here \cite{kharrat2016signal}) and should be confirmed in the future. }}
\begin{figure}[!h]
\centering
\subfigure[After wavelet denoising: Measurement B, first file in 60 cNm, sensor "F50A" (sensor 2). \label{dsefefe1111}]{\includegraphics[width=0.90\textwidth]{vibSens2_Denoised-crop-eps-converted-to.pdf}}
\subfigure[After wavelet denoising: Measurement B, first file in 60 cNm, sensor "micro-200-HF" (sensor 3). \label{dsefefe2222}]{\includegraphics[width=0.90\textwidth]{vibSens3_Denoised-crop-eps-converted-to.pdf}}
\caption{Presence of signals in each cycle after wavelet denoising using two sensors (on the same period of time). For a given tightening level, the displacement measured by the vibrometer is quite close for each signal found at different cycles showing the reproducibility of the test rig. \label{dejfdlkjzelkjzkej11}}
\end{figure}
\begin{figure}[!h]
\centering
{\includegraphics[width=0.88\textwidth]{measureB_reglageGros_onsetAEsignalVSvibro_micro200-crop-eps-converted-to.pdf}}
\caption{Acoustic emission onsets found by \cite{kharrat2016signal} and positioned onto vibrometer data for measurement B, first file in 60 cNm, sensor "micro-200-HF". {{The vertical position of the blue circle corresponds to the amount of displacement required to activate the AE sources. A similar amount is found in each cycle which illustrate the reproducibility of the test rig.}} \label{dejfdlkjzelkjzkej333}}
\end{figure}
\subsection{Possible challenges}
\noindent ORION-AE data set can be used for different tasks.
\begin{description}
\item[Train supervised learning methods]
{{The classification task can consider up to 7 classes in order to discriminate between the levels using various machine and deep learning methods such as support vector machines or neural networks~\cite{bishop06a,hesser2021identification}. By considering only the tightened levels, such as 60 cNm and 50 cNm, it can also be used to evaluate anomaly detectors based on autoencoders, self-organizing maps or one-class classifiers \cite{FarrarBook,nasir2021review}. In both cases, it is possible to use convolutionnal neural networks (CNN) taking as inputs the RGB images computed by the continuous wavelet transform (CWT) of the raw signals in each cycle. Cycles can be precisely obtained by using a zero-crossing detector on the vibrometer data. Illustration of such CWT images are depicted in Figure \ref{zefzfzefzef4e54fz5} using two measurements series (B, D). }}
\item[{{Data normalization}}] {{aims at compensating time-varying operational variations which can occur in a system. According to our experience on this data set, it seems that, for a given tightening level, the signature of the signals accross campaigns seems to change. This can be due to the tribological conditions in the contacts that change between levels and campaigns. The classification algorithms (using deep learning and autoencoders) that we tested seem to work well when we mix levels and campaigns. However, the classification performance drastically drops when we distinguish the campaigns (for example training on B, C, D, E and testing on F supposed unknown as in a SHM task). A data normalization is thus needed (starting from data in the first levels, for example $60$ cNm or $\{60,50\}$ cNm). This is an important topic in SHM as presented in \cite[Chap. 12]{FarrarBook}.}}
\item[Validate unsupervised learning] (clustering) using the tightening levels as a reference to be compared with, using, for example, the Adjusted Rand Index \cite{ari}. Initial results on this data set can be found in \cite{clustersSequence} using various algorithms.
\item[Challenge wave-picking algorithms] by trying to find either one set of signals per cycle or a set of several signals per cycle which should occur at similar displacement level along cycles for a given tightening level. Some results have been presented in the previous paragraphs using the method \cite{kharrat2016signal}.
\end{description}
{{The generalization capabilities of the methods can be evaluated by considering different campaigns for training and testing. This is a critical point for structural health monitoring tasks. }}
\begin{figure}[hbtp]
{\includegraphics[width=0.99\textwidth]{sampledimagesCWT_measBD_sensorB-crop.pdf}}
\caption{Continuous wavelet transform (CWT) images extracted for some cycles in sensor "F50A" (vector B) in measurement B (top figures) and in measurement D (bottom figures). Extraction of cycles can be automatically made by detecting zero-crossing in the vibrometer data (vector D). \label{zefzfzefzef4e54fz5}}
\end{figure}
\subsection{Potential sources of error or variability}
{{
According to our experience on this data set, the low signal-to-noise ratio represents a real challenge to solve the three aforementioned tasks. Some preliminary results have been shown in \cite{clustersSequence} with a good performance of various clustering methods. These results were obtained by considering campaigns independently. Therefore, this study shown that the information hold by the data are relevant for estimating the tightening levels.
The variability accross campaigns can represent a difficulty for simple classification methods, as discussed in the previous section. To tackle this issue, a proper data normalization is needed.
In terms of error in the data set, we are aware that one level is missing: 20 cNm in campaign C. It means that this campaign is made of six classes.
This data set was already used by Master students in 2021 and 2022 for machine learning illustration and a SHM course, at both the University of Franche-Comté and National Engineering School of Mechanics and Microtechnics of Besan\c con, who reported only one issue concerning the presence of a directory "5cNm" (made of one file) in addition to "05cNm" (made of 9 files). This folder appears after the downloading but not on the online directory. We recommend to delete the former after having transfered the file in the latter. This "phantom" directory is a small issue reported to the support of Harvard Dataverse and still under consideration at the time of writing the present work.
Finally, we acknowledge that this data set is not representative of all possible data sets that can be collected using the AE technique. The purpose of this data set is simply to provide a common reference to researchers in the AE community interested in developing new machine / deep learning or pattern recognition methods or new algorithms dedicated for AE data interpretation. With the expansion of deep learning, we think that this data set can be useful for benchmarking, which can give a hand in {discovering new and better ways} of processing and interpreting acoustic emission data.}}
\section{Data availability}
\noindent Directory name: Harvard Dataverse.
\noindent Data identification number: 10.7910/DVN/FBRDU0.
\noindent Direct URL to data: \url{https://doi.org/10.7910/DVN/FBRDU0}, comprising a Matlab code to read the files. Codes and main results will be updated on \url{https://github.com/emmanuelramasso/ORION_AE_acoustic_emission_multisensor_datasets_bolts_loosening}.
\section*{Acknowledgement}
This work was partly carried out in the framework of: EIPHI Graduate school (contract ANR-17-EURE-0002), project RESEM-COALESCENCE funded by "Institut de Recherche Technologique" Mat\'eriaux M\'etallurgie Proc\'ed\'es (IRT M2P) and "Agence Nationale de la Recherche" (ANR); project CLIMA funded by the "Fond Unique Interministériel" (FUI 19). The authors are also thankful to MIFHySTO and AMETISTE platforms.
\clearpage
\bibliographystyle{plain
|
\subsection{Complex Constraints}
\label{sec:aggregates}
As mentioned, we aim at accommodating complex language constructs as aggregates in order to leverage the full expressiveness of ASP's
modeling languages.
For instance, we may replace lines~9 and~10 in Listing~\ref{lp:ham} by
\begin{lstlisting}[firstnumber=9]
:- { hc(U,V) } >= 2, node(U).
:- { hc(U,V) } >= 2, node(V).
\end{lstlisting}
without violating its stratification.
More generally,
a rule with an aggregate
`\(
\#\mathit{op}\{l_1,\dots, l_m\}\prec k
\)'
in the consequent can be represented with choice rules along with an integrity constraint,
as shown in~\cite{siniso02a}.
That is, we can replace any rule of form
\[
\#\mathit{op}\{l_1,\dots, l_m\}\prec k\leftarrow l_{m+1}\wedge\dots\wedge l_n
\]
by\footnote{In practice, a set of such choice rules can be represented by a single one of form
$\{l_1,\dots,l_m\}\leftarrow l_{m+1}\wedge\dots\wedge l_n$.}
\begin{align*}
\{l_i\}&\leftarrow l_{m+1}\wedge\dots\wedge l_n &\text{ for }i=1,\dots,m\text{ and}\\
\bot&\leftarrow \neg(\#\mathit{op}\{l_1,\dots, l_m\}\prec k)\wedge l_{m+1}\wedge\dots\wedge l_n \ .&
\end{align*}
This allows us to integrate aggregate literals into easy logic programs without sacrificing expressiveness.
In fact, many encodings build upon restricted choices that are easily eliminated by such a transformation.
A very simple example is graph coloring.
Assume a problem instance is given in terms of facts
\lstinline{node/1},
\lstinline{edge/2}, and
\lstinline{color/1}.
A corresponding encoding is given by the following two rules:
\lstinputlisting{encoding/color.lp}
Note that the aggregate in the consequent of Line~1 is a shortcut for a \textit{\#count} aggregate.
To eliminate the restricted choice from the consequent in Line~1,
we may apply the above transformation to obtain the following easy encoding:
\lstinputlisting{encoding/color.ez}
Given some set of facts, $F$, this encoding amounts to the easy logic programs \linebreak
\(
(F,\{\mathrm{1}\},\{\mathrm{2}\},\{\mathrm{3}\}).
\)
The decomposition into a choice and its restriction may appear unnecessary to the experienced ASP modeler.
However, we feel that such a separation adds clarity and is preferable to language constructs combining several aspects,
at least for ASP beginners.
Also, it may be worth noting that this decomposition is done anyway by an ASP system and hence brings about no performance loss.
Two further examples of easy logic programs are given in Listing~\ref{lst:queens} and~\ref{lst:toh},
solving the Queens and the Tower-of-Hanoi puzzles both with parameter~\texttt{n}.
\footnote{This parameter is either added from the command line via option \texttt{--const} or a default added via directive \texttt{\#const}
(see~\cite{PotasscoUserGuide} for details).}
\lstinputlisting[float=ht,caption={An easy logic program for the $\mathtt{n}$-Queens puzzle},label=lst:queens]{encoding/queens-diag.ez}
\lstinputlisting[float=ht,caption={An easy logic program for a Towers-of-Hanoi puzzle (for plans of length $\mathtt{n}$)},label=lst:toh]{encoding/toh.ez}
While the easy logic program for the $\mathtt{n}$-Queens puzzle has the format
\[
(\emptyset,\{\mathrm{1}\},\{\mathrm{3,4}\},\{\mathrm{6,7}\},\{\mathrm{9,10}\}),
\]
the one for the Tower-of-Hanoi puzzle can be partitioned into
\[
(\{\mathrm{1,2,3,4}\},\{\mathrm{6}\},\{\mathrm{8}\},\{\mathrm{10,11,12}\},\{\mathrm{14,15}\},\{\mathrm{17,19,20,21,23}\})
\ .
\]
\section{Approach}\label{sec:approach}
\section{Background}\label{sec:background}
The logical foundations of ASP rest upon the logic of Here-and-There (HT~\cite{heyting30a})
along with its non-monotonic extension, Equilibrium Logic~\cite{pearce06a}.
We start by defining the monotonic logic of \emph{Here-and-There} (HT).
Let $\ensuremath{\mathcal{A}}} % {\ensuremath{\mathit{At}}$ be a set of atoms.
A \emph{formula} $\varphi$ over \ensuremath{\mathcal{A}}} % {\ensuremath{\mathit{At}}\ is an expression built with the grammar:
\[
\varphi ::= a \mid \bot \mid \varphi \wedge \varphi \mid \varphi \vee \varphi \mid \varphi \to \varphi
\]
for any atom $a\in \ensuremath{\mathcal{A}}} % {\ensuremath{\mathit{At}}$.
We also use the abbreviations:
\(
\neg \varphi \mathbin{\stackrel{\mathrm{def}}{=}} (\varphi \to \bot)
\),
\(
\top \mathbin{\stackrel{\mathrm{def}}{=}} \neg \bot
\),
and
\(
\varphi \leftrightarrow \psi \mathbin{\stackrel{\mathrm{def}}{=}} (\varphi \to \psi) \wedge (\psi \to \varphi)
\).
Given formulas $\varphi, \alpha$ and $\beta$,
we write $\varphi[\alpha/\beta]$ to denote the uniform substitution of all occurrences of formula $\alpha$ in $\varphi$ by~$\beta$.
This generalizes to the replacement of multiple formulas in the obvious way.
As usual, a \emph{theory} over \ensuremath{\mathcal{A}}} % {\ensuremath{\mathit{At}}\ is a set of formulas over \ensuremath{\mathcal{A}}} % {\ensuremath{\mathit{At}}.
We sometimes understand finite theories as the conjunction of their formulas.
An \emph{interpretation} over \ensuremath{\mathcal{A}}} % {\ensuremath{\mathit{At}}\ is a pair $\tuple{H,T}$ of atoms (standing for ``here'' and ``there'', respectively)
satisfying $H \subseteq T \subseteq \ensuremath{\mathcal{A}}} % {\ensuremath{\mathit{At}}$.
An interpretation is \emph{total} whenever $H=T$.
An interpretation $\tuple{H,T}$ \emph{satisfies} a formula $\varphi$, written $\tuple{H,T} \models \varphi$,
if the following conditions hold:
\[
\begin{array}{rcll}
\tuple{H,T} & \models & p & \text{if } p \in H\\
\tuple{H,T} & \models & \varphi \wedge \psi & \text{if } \tuple{H,T} \models \varphi \text{ and } \tuple{H,T} \models \psi\\
\tuple{H,T} & \models & \varphi \vee \psi & \text{if } \tuple{H,T} \models \varphi \text{ or } \tuple{H,T} \models \psi\\
\tuple{H,T} & \models & \varphi\! \to \! \psi & \text{if } \tuple{H',T} \not\models \varphi \text{ or } \tuple{H',T} \models \psi \text{ for both } H'\in\{H,T\}
\end{array}
\]
A formula $\varphi$ is \emph{valid}, written $\models \varphi$,
if it is satisfied by all interpretations.
An interpretation $\tuple{H,T}$ is a \emph{model} of a theory $\Gamma$, written $\tuple{H,T} \models \Gamma$, if $\tuple{H,T}\models \varphi$ for all $\varphi \in \Gamma$.
Classical entailment is obtained via the restriction to total models.
Hence, we define the classical satisfaction of a formula $\varphi$ by an interpretation $T$,
written $T\models\varphi$, as $\tuple{T,T}\models\varphi$.
A total interpretation $\tuple{T,T}$ is an \emph{equilibrium model} of a theory $\Gamma$ if $\tuple{T,T}$ is a model of $\Gamma$ and there is no other model $\tuple{H,T}$ of $\Gamma$ with $H \subset T$.
In that case, we also say that $T$ is a \emph{stable model} of~$\Gamma$.
We denote the set of all stable models of~$\Gamma$ by $\SM{\Gamma}$
and use $\SMR{\Gamma}{V} \, \mathbin{\stackrel{\mathrm{def}}{=}} \, \{\, T \cap V \mid T \in \SM{\Gamma}\,\}$
for their projection onto some vocabulary $V \subseteq \ensuremath{\mathcal{A}}} % {\ensuremath{\mathit{At}}$.
Since ASP is a non-monotonic formalism,
it may happen that two different formulas share the same equilibrium models
but behave differently in different contexts.
The concept of \emph{strong equivalence} captures the idea that two such formulas have the same models regardless of any context.
More precisely,
given two theories $\Gamma$ and $\Pi$ and a set $V \subseteq \ensuremath{\mathcal{A}}} % {\ensuremath{\mathit{At}}$ of atoms,
we say that $\Gamma$ and $\Pi$ are \emph{\mbox{$V$-strongly} equivalent}~\cite{agcafapepevi19a},
written $\Gamma \cong_V \Pi$,
if
$\SMR{\Gamma \cup \Delta}{V} = \SMR{\Pi \cup \Delta}{V}$
for any theory $\Delta$ over $\ensuremath{\mathcal{A}}} % {\ensuremath{\mathit{At}}'$ such that $\ensuremath{\mathcal{A}}} % {\ensuremath{\mathit{At}}' \subseteq V$.
For formulas~$\varphi$ and~$\psi$,
we write~$\varphi \cong_V \psi$
if $\{ \varphi \} \cong_V \{ \psi \}$.
A \emph{rule} is a (reversed) implication of the form
\begin{align}\label{rule}
l_1\vee\dots\vee l_m\leftarrow l_{m+1}\wedge\dots\wedge l_n
\end{align}
where each $l_i$ is a literal, that is, either an atom or a negated atom, for $1\leq i\leq n$.
If $n=1$, we refer to the rule as a \emph{fact} and write it as $l_1$ by dropping the trailing implication symbol.
A rule is said to be \emph{normal} whenever $m=1$ and $l_1$ is an atom.
A negation-free normal rule is called \emph{definite}.
An \emph{integrity constraint} is a rule with $m=0$ and equivalent to
\(
\bot\leftarrow l_{m+1}\wedge\dots\wedge l_n
\).
Finally, the law of the excluded middle $a\vee\neg a$ is often represented as $\{a\}$ and called a \emph{choice}.
Accordingly, a rule with a choice on the left-hand side is called a \emph{choice rule}.
A \emph{logic program} is a set of rules.
It is called \emph{normal}, if it consists only of normal rules and integrity constraints,
and \emph{definite} if all its rules are definite.
\subsection{Debugging}
\label{sec:debugging}
Another application of optimization statements is debugging easy logic programs.
As mentioned, a special feature of austere logic programs (as well as the ones admitting stratified negation only)
is that inconsistencies are confined to the violation of integrity constraints
(rather than caused by odd loops spread over the program).
We can exploit this for debugging in the following way.
Replace each integrity constraint
\[
\bot\leftarrow l_{m+1}\wedge\dots\wedge l_n
\quad\text{ by }\quad
\mathit{ic}(i)\leftarrow l_{m+1}\wedge\dots\wedge l_n
\]
where $\mathit{ic}$ is a new predicate symbol and $i$ an identifier for the integrity constraint at hand.
And add a minimization statement
\[
\#\mathit{minimize}\{\mathit{ic}(I)\}
\]
to output only solution candidates with a minimum number of violations.
As an example, let us replace lines~8 to~10 in Listing~\ref{lp:ham} by
\begin{lstlisting}[firstnumber=8]
ic(unreached(V)) :- node(V), not reached(V).
ic(two_out(V)) :- hc(V,U), hc(V,W), U!=W.
ic(two_in(V)) :- hc(U,V), hc(W,V), U!=W.
#minimize{ ic(I) }.
\end{lstlisting}
The purpose of the optimization directive in Line~12 is to restrict the stable models to the ones violating the fewest constraints in lines~8-10.
Hence, if there are proper solutions satisfying all integrity constraints, they are output and no others.
Note that the above introduction of new consequents yields normal yet stratified rules.
\section{Conclusion}\label{sec:discussion}
We have revisited an old idea from the literature on logic programming under stable model semantics
and elaborated upon it in several ways.
We started by tracing it back to the principle of extension by definition.
The resulting formalization in the setting of the logic of Here-and-there provides us with a logical framework
that can be instantiated in various ways.
Along these lines,
we have shown that normal logic programs can be reduced to choices, definite rules, and integrity constraints,
while keeping the same expressiveness as the original program.
A major advantage of this austere format is that it confines non-determinism and incoherence to well-defined spots in the program.
The resulting class of austere logic programs could play a similar role in ASP as formulas in conjunctive normal form in classical logic.
Drawing on the properties observed on austere logic program,
we put forward the modeling methodology of ezASP.
The idea is to compensate for the lacking guarantees provided by the restricted format of austere programs
by following a sequential structure when expressing a problem in terms of a logic program.
This makes use of the well-known concept of stratification to refine ASP's traditional guess-define-check methodology.
Although the ordering of rules may seem to disagree with the holy grail of full declarativeness,
we evidence its great value in introducing beginners to ASP.
Also, many encodings by experienced ASP users follow the very same pattern.
Moreover, the ezASP paradigm aligns very well with that of \emph{achievements}~\cite{lifschitz17a}
that aims not only at easily understandable but moreover provably correct programs.
To this end, formal properties are asserted in between a listing of rules to express what has been achieved up to that point.
Extending ezASP with achievements and automatically guiding the program development with ASP verifiers, like \sysfont{anthem}~\cite{lilusc19a},
appears to us as a highly interesting avenue of future research.
In this context, it will also be interesting to consider the components of an easy logic program as modules with an attached input-output
specification, so that the meaning of the overall program emerges from the composition of all components.
This would allow for successive refinements of programs' components, while maintaining their specification.
\section{Good modeling outside our setting.}
\begin{itemize}
\item
\textbf{Example of choice rules in the middle. }
\item
Let $G$ be a graph defined by atoms of predicates \verb|node/1| and \verb|edge/2|,
and let $p$ be a positive integer.
Consider the problem of finding a set
of $p$ non-conflicting paths that visit all the nodes of the graph.
The problem can be represented naturally by this encoding.
\begin{verbatim}
path(1..p).
{ start(P,N) : node(N) } = 1 :- path(P).
visit(P,N) :- start(P,N).
{ visit(P,N1) : edge(N1,N2) } 1 :- visit(P,N2).
:- node(N), { visit(P,N) } != 1.
\end{verbatim}
Every stable model of the program corresponds to a solution of the problem,
and for every solution of the problem there are $p!$ stable models
that correspond to it.
The encoding uses naturally a choice rule with a body.
Furthermore, the rule is recursive.
\item
This example cannot be nicely represented in IDP, and
does not fit in the Tarskian semantics that Denecker et al. propose.
\item
In general, a choice rule of this form:
\begin{verbatim}
L { a(X) : b(X) } U :- B.
\end{verbatim}
can be replaced by:
\begin{verbatim}
{ choice(a(X)) } :- domain(a(X)).
a(X) :- B, b(X), choice(a(X)).
:- not L { a(X) : b(X) } U, B.
\end{verbatim}
and if we want to have a 1-to-1 correspondence among the stable models,
we need the constraints:
\begin{verbatim}
:- choice(a(X)), #false: B, b(X).
:- not choice(a(X)), a(X), B, b(X).
\end{verbatim}
In the example, we would have this:
\begin{verbatim}
path(1..p).
{ start(P,N) : node(N) } = 1 :- path(P).
visit(P,N) :- start(P,N).
{ choice(visit(P,N)) } :- path(P), node(N).
visit(P,N1) :- visit(P,N2), edge(N1,N2), choice(visit(P,N1)).
: not { visit(P,N1) : edge(N1,N2) } 1, visit(P,N2).
:- node(N), { visit(P,N) } != 1.
\end{verbatim}
and the constraints for 1-to-1 correspondence:
\begin{verbatim}
:- choice(visit(P,N1)), #false: visit(P,N2), edge(N1,N2).
:- not choice(visit(P,N1)), visit(P,N), visit(P,N2), edge(N1,N2).
\end{verbatim}
\item
A very similar encoding is used in your paper on Train Scheduling with Hybrid ASP.
This is Listing 1.1:
\begin{verbatim}
1 { visit(T,V) : start(T,V) } 1 :- train(T).
1 { route(T,(V,V’)) : edge(T,V,V’) } 1 :- visit(T,V), not end(T,V).
visit(T,V’) :- route(T,(V,V’)).
\end{verbatim}
If you did not need elsewhere the \verb|route/2| predicate,
then it could be simplified to:
\begin{verbatim}
1 { visit(T,V) : start(T,V) } 1 :- train(T).
1 { visit(T,V’) : edge(T,V,V’) } 1 :- visit(T,V), not end(T,V).
\end{verbatim}
that is quite similar to the example above.
\item
\textbf{Example of recursion through negation. }
\item
The example is the solution to the frame problem in logic programming,
\comment{JR: Refer to Lifschitz's paper on the frame problem?}
i.e., the representation of the law of inertia.
For two-valued fluents we have either:
\begin{verbatim}
holds(F,T) :- holds(F,T-1), not -holds(F,T).
-holds(F,T) :- nholds(F,T-1), not holds(F,T).
\end{verbatim}
or:
\begin{verbatim}
holds(F,V,T) :- holds(F,V,T-1), not holds(F,V',T), value(V'), V!=V'.
\end{verbatim}
For multi-valued fluents we have:
\begin{verbatim}
holds(F,V,T) :- holds(F,V,T-1), not holds(F,V',T) : value(V'), V != V'.
\end{verbatim}
\item
The encoding involves recursion over predicate holds.
This happens at the first order level, and also shows up in
the result of intelligent grounding.
\item
If the indirect effects of the corresponding action theory are stratified,
then for every initial situation and sequence of actions
the whole logic program will have a total well founded model and hence a
unique stable model.
\item
But if they are not stratified, we can see here all the effects of the
cycles of ASP.
\comment{JR: The other day in our conversation I said that this could not happen.
That was wrong.}
This example leads to a kind of even cycle that generates a choice:
\begin{verbatim}
holds(f,t,T) :- holds(g,f,T).
holds(g,t,T) :- holds(f,f,T).
\end{verbatim}
and this one leads to a kind of odd cycle that may eliminate a model:
\begin{verbatim}
holds(f,t,T) :- holds(f,f,T).
\end{verbatim}
\item
To understand how that could happen,
this may be helpful.
The stable models of a logic program with rules of the form:
\begin{verbatim}
a :- b1, ..., bm, not c1, ..., not cn.
\end{verbatim}
correspond 1-to-1 to the stable models of the program
\begin{verbatim}
holds(a,t,T) :- time(T),
holds(b1,t,T), ..., holds(bm,t,T),
holds(c1,f,T), ..., holds(cn,f,T).
\end{verbatim}
together with the facts
\begin{verbatim}
holds(x,f,0).
time(1).
\end{verbatim}
This translation implies that all (good and bad) we can do in ASP
can also be done using the indirect effects.
\item
Maybe M. Gelfond has some more examples similar to this in his latest book.
\item
The version from J. Lee uses choices in the middle:
\begin{verbatim}
{ holds(F,V,T) } :- holds(F,V,T-1).
:- not { holds(F,V,T) } = 1, fluent(V), time(T).
\end{verbatim}
\item
We can get rid of the choice in the middle using the method above.
We can also have a natural representation like this:
\begin{verbatim}
{ persist(F,T) } :- fluent(F), time(T).
holds(F,V,T) :- persist(F,T), holds(F,V,T-1).
:- not { holds(F,V,T) } = 1, fluent(V), time(T).
\end{verbatim}
with the next constraint to get the 1-to-1 correspondence:
\begin{verbatim}
:- not persist(F,T), holds(F,V,T), holds(F,V,T-1).
\end{verbatim}
Another option is to add \verb|noop| actions:
\begin{verbatim}
{ occ(noop(F),T) } :- fluent(F), time(T).
holds(F,V,T) :- occ(noop(F),T), holds(F,V,T-1).
:- not { holds(F,V,T) } = 1, fluent(V), time(T).
\end{verbatim}
with the constraint to get the 1-to-1 correspondence:
\begin{verbatim}
:- not occ(noop(F),T), holds(F,V,T), holds(F,V,T-1).
\end{verbatim}
Personally, I find this last option really surprising:
reifying the truth values and using \verb|noop|'s solves
the frame problem!
\end{itemize}
\section{Easy Answer Set Programming}
Austere logic programs provide a greatly simplified format that reflects ASP's \emph{guess-define-check} methodology~\cite{lifschitz02a}
for writing encodings.
Their simple structure allows for translating the methodology into
an intuitive process that consists of
making non-deterministic choices, followed by a deterministic bottom-up computation, and a final consistency check.
In what follows,
we want to turn the underlying principles into a modeling methodology for ASP that aims at laypersons.
To this end, we leave the propositional setting and aim at full-featured input languages of ASP systems like \sysfont{clingo}~\cite{gekakasc17a} and \sysfont{dlv}~\cite{dlv03a}.
Accordingly, we shift our attention to predicate symbols rather than propositions and let the terms `logic program', `rule', etc.\ refer to these languages
without providing a technical account (cf.~\cite{gehakalisc15a,cafageiakakrlemarisc19a}).
Moreover, we allow for normal rules instead of definite ones as well as aggregate literals in bodies
in order to accommodate the richness of existing ASP modeling languages.
The admission of normal rules comes at the expense of losing control over the origin of multiple or non-existing stable models as well as
over a deterministic development of guessed solutions.
In fact,
the idea of \emph{Easy Answer Set Programming} (ezASP) is to pursue the principles underlying austere logic programs
without enforcing them through a severely restricted syntax.
However,
rather than having the user fully absorb the loss in control,
we shift our focus to a well-founded development of ASP encodings,
according to which predicates are defined on top of previously defined predicates
(or facts).
This parallels the structure and the resulting operational semantics of austere logic programs.
To this end,
we start by capturing dependencies among predicates~\cite{apblwa87a}.
\begin{definition}
Let $P$ be a logic program.
\begin{itemize}
\item A predicate symbol $p$ \emph{depends} upon a predicate symbol $q$,
if there is a rule in $P$ with $p$ on its left-hand side and $q$ on its right-hand side.
If $p$ depends on $q$ and $q$ depends on $r$, then $p$ depends on $r$, too.
\item The \emph{definition} of a predicate symbol $p$ is the subset of $P$
consisting of all rules with $p$ on their left-hand side.
\end{itemize}
\end{definition}
We denote the definition of a predicate symbol $p$ in $P$ by $\mathit{def}(p)$
and view integrity constraints as rules defining $\bot$.
Our next definition makes precise what we mean by a well-founded development of a logic program.
\footnote{The term \emph{stratification} differs from the one used in the literature~\cite{apblwa87a}.}
\begin{definition}\label{def:stratification}
Let $P$ be a logic program.
We define a partition
\(
(P_1,\dots, P_n)
\)
of $P$ as a \emph{stratification} of $P$,
if
\begin{enumerate}
\item $\mathit{def}(p)\subseteq P_i$ for all predicate symbols $p$ and some $i\in\{1,\dots,n\}$
and
\item if
$p$ depends on $q$,
$\mathit{def}(p)\subseteq P_i$,
and
$\mathit{def}(q)\subseteq P_j$ for some $i,j\in\{1,\dots,n\}$,
then
\begin{enumerate}
\item $i>j$ unless $q$ depends on $p$, and
\item $i=j$ otherwise
\end{enumerate}
\end{enumerate}
\end{definition}
Any normal logic program has such a stratification.
One way to see this is that mutually recursive programs can be trivially stratified via a single partition.
For instance, this applies to both programs
\(
\{a\leftarrow b, b\leftarrow a\}
\)
and
\(
\{a\leftarrow \neg b, b\leftarrow \neg a\}
\)
in which $a$ and $b$ mutually depend upon each other.
Accordingly, similar recursive structures in larger programs are confined to single partitions, as required by (\emph{2b}) above.
With it, we are ready to give shape to the concept of an \emph{easy logic program}.
\begin{definition}[Easy logic program]\label{def:easy:program}
An easy logic program is a logic program
having stratification
\(
(F,C,D_1,\dots,D_n,I)
\)
such that
$F$ is a set of facts,
$C$ is a set of choice rules,
$D_i$ is a set of normal rules for $i=1,\dots,n$, and
$I$ is a set of integrity constraints.
\end{definition}
\comment{JR: I don't understand the definition: normal logic programs have no choice rules.
We say that they can be expressed by normal rules, but
strictly speaking choice rules are not normal rules, and
they require non-stratified negation (that we eliminate later) to be expressed
by normal programs.\par
T: partly fixed}
As in traditional ASP, we often divide a logic program into facts representing a problem instance and the actual encoding of the problem class.
For easy programs, this amounts to separating $F$ from $(C,D_1,\dots,D_n,I)$.
Clearly, an austere logic program is also an easy one.
Thus, the program in Listing~\ref{lp:ham} is also an easy logic program having the stratification
\[
(\{\mathrm{1},\mathrm{2},\mathrm{3}\},\{\mathrm{5}\},\{\mathrm{6},\mathrm{7}\},\{\mathrm{8},\mathrm{9},\mathrm{10}\})
\]
where each number stands for the rules in the respective line.
Predicates \texttt{node/1}, \texttt{edge/2}, and \texttt{start/1} are only used to form facts or occur in rule bodies.
Hence, they do not depend on any other predicates and can be put together in a single component, $F$.
This makes sense since they usually constitute the problem instance.
Putting them first reflects that the predicates in the actual encoding usually refer to them.
The choices in $C$ provide a solution candidate that is checked by means of the rules in the following components.
In our case, the guessed extension of predicate \texttt{hc/2} in Line~5 is simply a subset of all edges provided by predicate \texttt{edge/2}.
Tests for being a path or even a cycle are postponed to the \emph{define-check} part:
The rules in $\{\mathrm{6},\mathrm{7}\}$, that is, $D_1$, define the auxiliary predicate \texttt{reached/1},
and aim at analyzing and/or extending our guessed solution candidate by
telling us which nodes are reachable via the instances of \texttt{hc/2} from the start node.
The actual checks are then conducted by the integrity constraints, $I$, in the final partition $\{\mathrm{8},\mathrm{9},\mathrm{10}\}$.
At this point, the solution candidate along with all auxiliary atoms are derived and ready to be checked.
Line~8 tests whether each node is reached in the solution at hand,
while lines~9 and~10 make sure that a valid cycle never enters or leaves any node twice.
Finally, it is instructive to verify that strata $\{\mathrm{5}\}$ and $\{\mathrm{6},\mathrm{7}\}$ cannot be reversed or merged.
We observe that
\begin{itemize}
\item \texttt{hc/2} depends on \texttt{edge/2} only,
\end{itemize}
while
\begin{itemize}
\item \texttt{reached/1} depends on \texttt{hc/2}, \texttt{edge/2}, \texttt{start/1}, and itself,
\end{itemize}
and no other dependencies.
The rules defining \texttt{hc/2} and \texttt{reached/1} must belong to the same partition, respectively, as required by (\emph{2a}) above.
Thus,
$\{\mathrm{5}\}\subseteq P_i$
and
$\{\mathrm{6},\mathrm{7}\}\subseteq P_j$
for some $i,j$.
Because \texttt{reached/1} depends on \texttt{hc/2} and not vice versa, we get $i<j$.
This dependency rules out an inverse arrangement,
and the fact that it is not symmetric excludes a joint membership of both definitions in the same partition,
as stipulated by (\emph{2b}) above.
\input{methodology}
\input{stratified}
\input{aggregates}
\input{limitations}
\section{Logical Foundations}
We begin by investigating the logical underpinnings of the simple format of logic programs discussed in the introductory section.
Although the discussion of the exemplary logic program has revealed several characteristic properties,
not all of them can be captured in a logical setting, such as order related features.
What remains is the division of the encoding into facts, rules, choices, and integrity constraints.
In logical terms,
the first two amount to negation-free formulas,
choices are instances of the law of the excluded middle,
and finally integrity constraints correspond to double-negated formulas in HT.
While the first two types of formulas are arguably simpler because of their restricted syntax,
the latter's simplicity has a semantic nature and is due to the fact that in HT double negated formulas can be treated as in classical logic.
In what follows,
we show that any formula can be divided into a conjunction of corresponding subformulas.
This conjunction is strongly equivalent (modulo the original vocabulary) to the original formula and the translation can thus also be applied to substitute subformulas.
Interestingly, the resulting conjunction amounts to a conservative extension of the original formula and
the underlying translation can be traced back to the logical principle of extension by definition,
as we show below.
To this end,
we associate with each formula~$\varphi$ over \ensuremath{\mathcal{A}}} % {\ensuremath{\mathit{At}}\ a new propositional atom~\auxatom{\varphi}.
We then consider defining axioms of the form
\(
(\auxatom{\varphi} \leftrightarrow \varphi)
\).
We can now show that replacing any subformula $\varphi$ by \auxatom{\varphi} while adding a corresponding defining axiom amounts to a conservative
extension of $\psi$.
\begin{proposition}\label{prop:extension:definition}
Let $\psi$ and $\varphi$ be formulas over \ensuremath{\mathcal{A}}} % {\ensuremath{\mathit{At}}\ and $\auxatom{\varphi} \not \in \ensuremath{\mathcal{A}}} % {\ensuremath{\mathit{At}}$.
Then,
\(
\psi \cong_{\ensuremath{\mathcal{A}}} % {\ensuremath{\mathit{At}}} (\psi[\varphi/\auxatom{\varphi}] \wedge (\varphi \leftrightarrow \auxatom{\varphi}))
\).
\end{proposition}
\comment{T: comment on short proof and relation to Vladimir}
Moreover, we get a one-to-one correspondence between the stable models of both formulas.
\begin{proposition}\label{prop:extension:definition:one:one}
Let $\psi$ and $\varphi$ be formulas over \ensuremath{\mathcal{A}}} % {\ensuremath{\mathit{At}}\ and $\auxatom{\varphi} \not \in \ensuremath{\mathcal{A}}} % {\ensuremath{\mathit{At}}$.
\begin{enumerate}
\item If $T\subseteq\ensuremath{\mathcal{A}}} % {\ensuremath{\mathit{At}}$ is a stable model of $\psi$,
then $T \cup \{\auxatom{\varphi} \mid T \models \varphi \}$ is a stable model of
$(\psi[\varphi/\auxatom{\varphi}] \wedge (\varphi \leftrightarrow \auxatom{\varphi}))$.
\item If $T\subseteq (\ensuremath{\mathcal{A}}} % {\ensuremath{\mathit{At}} \cup \{\auxatom{\varphi}\})$ is a stable model of
$(\psi[\varphi/\auxatom{\varphi}] \wedge (\varphi \leftrightarrow \auxatom{\varphi}))$,
then $T\cap\ensuremath{\mathcal{A}}} % {\ensuremath{\mathit{At}}$ is a stable model of $\psi$.
\end{enumerate}
\end{proposition}
Clearly, the above results generalize from replacing and defining a single subformula $\varphi$ to several such subformulas.
With this, we can now turn our attention to negated subformulas:
Given a formula~$\psi$, let $\negocc{\psi}$ stand for the set of all maximal negated subformulas occurring in $\psi$.
This leads us to the following variant of Proposition~\ref{prop:extension:definition}.
\begin{corollary}\label{cor:extension:definition:negated}
Let $\psi$ be a formula over \ensuremath{\mathcal{A}}} % {\ensuremath{\mathit{At}}\ and $\auxatom{\varphi} \not \in \ensuremath{\mathcal{A}}} % {\ensuremath{\mathit{At}}$.
Then,
\(
\psi
\cong_{\ensuremath{\mathcal{A}}} % {\ensuremath{\mathit{At}}}
\psi\big[\varphi/\auxatom{\varphi}\mid\varphi\in\negocc{\psi}\big]
\wedge
\bigwedge_{\varphi\in\negocc{\psi}}(\varphi\leftrightarrow\auxatom{\varphi})
\).
\end{corollary}
Given that we exclusively substitute negated subformulas,
we can actually treat the defining axiom as in classical logic.
This is because in HT, we have
\(
\tuple{H,T} \models\neg \varphi
\)
iff
(classically)
\(
T \models\neg \varphi
\).
The classical treatment of the defining axiom is then accomplished by replacing
\(
(\varphi\leftrightarrow\auxatom{\varphi})
\)
by
\(
\neg \neg (\varphi\leftrightarrow\auxatom{\varphi})
\)
and
\(
(\neg\auxatom{\varphi}\vee\auxatom{\varphi})
\).
This results in the following decomposition recipe for formulas.
\begin{definition}\label{def:translation}
Let $\psi$ be a formula over \ensuremath{\mathcal{A}}} % {\ensuremath{\mathit{At}}\ and $\auxatom{\varphi} \not \in \ensuremath{\mathcal{A}}} % {\ensuremath{\mathit{At}}$.
Then, we define
\
\translation{\psi}
=
\psi\big[\varphi/\auxatom{\varphi}\mid\varphi\in\negocc{\psi}\big]
\wedge
\bigwedge_{\varphi\in\negocc{\psi}}(\neg\auxatom{\varphi}\vee\auxatom{\varphi})
\wedge
\bigwedge_{\varphi\in\negocc{\psi}}\neg \neg (\varphi\leftrightarrow\auxatom{\varphi})
\ .
\]
\end{definition}
\begin{example}\label{ex:lp:translation}
Let $\psi$ be $\neg a\to b\vee\neg \neg(c\wedge\neg d)$.
Then,
\begin{align*}
\negocc{\psi} =\ & \{\neg a,\neg\neg(c\wedge\neg d)\}\\
\translation{\psi} =\ & (\auxatom{\neg a}\to b\vee\auxatom{\neg\neg (c\wedge\neg d)})\; \wedge\\
& (\auxatom{\neg a}\vee\neg\auxatom{\neg a}) \wedge
(\auxatom{\neg\neg(c\wedge\neg d)}\vee\neg\auxatom{\neg\neg (c\wedge\neg d)}) \\
& \neg\neg({\neg a} \leftrightarrow\auxatom{\neg a}) \wedge
\neg\neg({\neg\neg(c\wedge\neg d)}\leftrightarrow\auxatom{\neg\neg(c\wedge\neg d)})
\end{align*}
\end{example}
With the translation from Definition~\ref{def:translation}, we obtain an analogous conservative extension result as above.
\begin{theorem}\label{thm:extension:definition:negated}
Let $\psi$ be a formula over \ensuremath{\mathcal{A}}} % {\ensuremath{\mathit{At}}.
Then, we have
\(
\psi \cong_{\ensuremath{\mathcal{A}}} % {\ensuremath{\mathit{At}}}\translation{\psi}
\).
\end{theorem}
In analogy to Proposition~\ref{prop:extension:definition:one:one},
we get a one-to-one correspondence between the stable models of both formulas.
\begin{theorem}\label{thm:extension:definition:negated:one:one}
Let $\psi$ be a formula over \ensuremath{\mathcal{A}}} % {\ensuremath{\mathit{At}}.
\begin{enumerate}
\item If $T\subseteq\ensuremath{\mathcal{A}}} % {\ensuremath{\mathit{At}}$ is a stable model of $\psi$,
then $T \cup \{\auxatom{\varphi} \mid \varphi \in \negocc{\psi} \text{ and }T \models \varphi \}$
is a stable model of $\translation{\psi}$.
\item If $T\subseteq (\ensuremath{\mathcal{A}}} % {\ensuremath{\mathit{At}} \cup \{\auxatom{\varphi} \mid \varphi \in \negocc{\psi}\})$ is a stable model of $\translation{\psi}$,
then $T\cap\ensuremath{\mathcal{A}}} % {\ensuremath{\mathit{At}}$ is a stable model of $\psi$.
\end{enumerate}
\end{theorem}
For instance, $\{b\}$ is a stable model of the formula $\psi=\neg a\to b\vee\neg \neg(c\wedge\neg d)$ from Example~\ref{ex:lp:translation}.
From Theorem~\ref{thm:extension:definition:negated}, $\{\auxatom{\neg a}, b\}$ is a stable model of $\translation{\psi}$.
Conversely, from the stable model $\{\auxatom{\neg a}, b\}$ of $\translation{\psi}$,
we get the stable model $\{b\}$ of $\psi$ by dropping the new atoms.
\section{First Section}
\subsection{A Subsection Sample}
Please note that the first paragraph of a section or subsection is
not indented. The first paragraph that follows a table, figure,
equation etc. does not need an indent, either.
Subsequent paragraphs, however, are indented.
\subsubsection{Sample Heading (Third Level)} Only two levels of
headings should be numbered. Lower level headings remain unnumbered;
they are formatted as run-in headings.
\paragraph{Sample Heading (Fourth Level)}
The contribution should contain no more than four levels of
headings. Table~\ref{tab1} gives a summary of all heading levels.
\begin{table}
\caption{Table captions should be placed above the
tables.}\label{tab1}
\begin{tabular}{|l|l|l|}
\hline
Heading level & Example & Font size and style\\
\hline
Title (centered) & {\Large\bfseries Lecture Notes} & 14 point, bold\\
1st-level heading & {\large\bfseries 1 Introduction} & 12 point, bold\\
2nd-level heading & {\bfseries 2.1 Printing Area} & 10 point, bold\\
3rd-level heading & {\bfseries Run-in Heading in Bold.} Text follows & 10 point, bold\\
4th-level heading & {\itshape Lowest Level Heading.} Text follows & 10 point, italic\\
\hline
\end{tabular}
\end{table}
\noindent Displayed equations are centered and set on a separate
line.
\begin{equation}
x + y = z
\end{equation}
Please try to avoid rasterized images for line-art diagrams and
schemas. Whenever possible, use vector graphics instead (see
Fig.~\ref{fig1}).
\begin{figure}
\includegraphics[width=\textwidth]{fig1.eps}
\caption{A figure caption is always placed below the illustration.
Please note that short captions are centered, while long ones are
justified by the macro package automatically.} \label{fig1}
\end{figure}
\begin{theorem}
This is a sample theorem. The run-in heading is set in bold, while
the following text appears in italics. Definitions, lemmas,
propositions, and corollaries are styled the same way.
\end{theorem}
\begin{proof}
Proofs, examples, and remarks have the initial word in italics,
while the following text appears in normal font.
\end{proof}
For citations of references, we prefer the use of square brackets
and consecutive numbers. Citations using labels or the author/year
convention are also acceptable. The following bibliography provides
a sample reference list with entries for journal
articles~\cite{ref_article1}, an LNCS chapter~\cite{ref_lncs1}, a
book~\cite{ref_book1}, proceedings without editors~\cite{ref_proc1},
and a homepage~\cite{ref_url1}. Multiple citations are grouped
\cite{ref_article1,ref_lncs1,ref_book1},
\cite{ref_article1,ref_book1,ref_proc1,ref_url1}.
\section{Inconsistency in Austere Logic Programs}
\begin{itemize}
\item
In normal logic, inconsistency is caused by odd loops.
\item
An inconsistent logic program:
\begin{align*}
a\leftarrow& \\
e\leftarrow& \\
b\leftarrow& c \wedge e \\
c\leftarrow& a\wedge \neg d \\
d\leftarrow& b
\end{align*}
Here the inconsistency is caused by the odd loop formed due to the last three rules.
\item
But not odd loops cause inconsistency.
%
Some odd loops are used to remove unwanted answer sets like the integrity constraint in the last rule of the program below.
\begin{align*}
a\leftarrow& \neg b \\
b\leftarrow& \neg a \\
c\leftarrow& b \wedge d \\
d\leftarrow& a \wedge \neg c
\end{align*}
Integrity constraints are used for this purpose.
\item
We see that the odd loops are important but for ensuring consistency, we can not limit the checking process to a specific segment of the program like in case of austere programs.
\item
Instead austere programs can be used to express odd loops and in this case we can limit our consistency checking to the integrity constraints in the program.
\end{itemize}
\section{Introduction}\label{sec:introduction}
Many people like Answer Set Programming (ASP~\cite{lifschitz02a})
because its declarative approach frees them from expressing any procedural information.
In ASP,
neither the order of rules nor
the order of conditions in rule antecedents or consequents matter
and thus leave the meaning of the overall program unaffected.
Although this freedom is usually highly appreciated by ASP experts,
sometimes laypersons seem to get lost without any structural guidance
when modeling in ASP.
We address this issue in this (preliminary) paper and
develop a methodology for ASP modeling that targets laypersons,
such as biologists, economists, engineers, and alike.
As a starting point, we explore an idea put forward by Ilkka Niemel{\"{a}} in~\cite{niemela08b}, although already present in~\cite{felomisu93a,grsaza01a} as well as the neighboring area of Abductive Logic Programming~\cite{eshkow89a,denkak02a}.
To illustrate it, consider the logic program encoding a Hamiltonian circuit problem in Listing~\ref{lp:ham}.
\lstinputlisting[float=ht,caption={A logic program for a Hamiltonian circuit problem},label=lp:ham]{encoding/Hamiltonian_Circuit.lp}
Following good practice in ASP,
the problem is separated into the specification of the problem instance in lines~1-3
and the problem class in lines~5-10.
This strict separation, together with the use of facts for problem instances,
allows us to produce
uniform\footnote{A problem encoding is \textit{uniform}, if it can be used to solve all its problem instances.}
and
elaboration tolerant\footnote{A formalism is \emph{elaboration tolerant} if it is convenient
to modify a set of facts expressed in the formalism to take into account new phenomena or changed circumstances~\cite{mccarthy98a}.}
specifications.
Building upon the facts of the problem instance,
the actual encoding follows the guess-define-check methodology of ASP.
A solution candidate is guessed in Line~5,
analyzed by auxiliary definitions in Line~6 and~7,
and finally checked through integrity constraints in lines~8-10.
A closer look reveals even more structure in this example.
From a global perspective,
we observe that the program is partitioned into
facts, choices, rules, and integrity constraints,
and in this order.
From a local perspective,
we note moreover that the predicates in all rule antecedents are defined beforehand.
This structure is not arbitrary and simply follows the common practice
that concept formation is done linearly by building concepts on top of each other.
Moreover, it conveys an intuition on how a solution is formed.
Importantly, such an arrangement of rules is purely methodological and has no impact on the meaning
(nor the performance\footnote{Shuffling rules in logic programs has an effect on performance since it affects tie-breaking during search;
this is however unrelated to the ordering at hand.}) of the overall program.
From a logical perspective,
it is interesting to observe that the encoding refrains from using negation explicitly,
except for the integrity constraints.
Rather this is hidden in Line~5, where the choice on \lstinline{hc(V,U)} amounts
to the disjunction \lstinline[mathescape]{hc(V,U)$\,\vee\,\neg$hc(V,U)},
an instance of the law of the excluded middle.
Alternatively, \lstinline{hc(V,U)} can also be regarded as an abducible that may or may not be added to a program,
as common in Abductive Logic Programming.
Presumably motivated by similar observations,
Ilkka Niemelä already argued in~\citeyear{niemela08b} in favor of an ASP base language based on choices, integrity constraints, and stratified negation.
\footnote{This concept eliminates the (problematic) case of recursion through negation.}
We also have been using such an approach
when initiating students to ASP as well as teaching laypersons.
Our experience has so far been quite positive and we believe that a simple and more structured approach helps to
get acquainted with posing constraints in a declarative setting.
We elaborate upon this idea in complementary ways.
First of all, we lift it to a logical level to investigate its foundations and identify its scope.
Second, we want to draw on this to determine a syntactically restricted\comment{irreducible} subclass of logic programs
that still warrants the full expressiveness of traditional ASP.
Such a subclass can be regarded as a normalform for logic programs in ASP.
This is also interesting from a research perspective since it allows scientists to initially develop their theories in a restricted setting without
regarding all corner-cases emerging in a full-featured setting.
And last but not least, inspired by this,
we want to put forward a simple and more structured modeling methodology for ASP that aims at beginners and laypersons.
\subsection{Limitations}
\label{sec:limitations}
The methodology of ezASP has its limits.
For instance, sometimes it is convenient to make choices depending on previous choices.
Examples of this are the combination of routing and scheduling, as in train scheduling~\cite{abjoossctowa19a},
or the formalization of frame axioms in (multi-valued) planning advocated in~\cite{leliya13a}.
Another type of encodings escaping our methodology occurs in meta programming,
in which usually a single predicate, like \texttt{holds}, is used and atoms are represented as its arguments.
Thus, for applying the ezASP methodology, one had to refine the concept of stratification to access the term level
in order to capture the underlying structure of the program.
And finally, formalizations of planning and reasoning about actions involve
the formalization of effect and inertia laws that are usually self-referential on the predicate level
(sometimes resolved on the term level, through situation terms or time stamps).
A typical example of circular inertia laws is the following:
\begin{lstlisting}[numbers=none]
holds(F,T) :- holds(F,T-1), not -holds(F,T).
-holds(F,T) :- -holds(F,T-1), not holds(F,T).
\end{lstlisting}
Here, `\texttt{-}' denotes classical negation, and \texttt{F} and \texttt{T} stand for (reified) atoms and time points.
On the other hand, the sophistication of the given examples illustrates that they are usually not addressed by beginners but rather experts in
ASP for which the strict adherence to ezASP is less necessary.
\section{Austere Answer Set Programming}
In this section,
we restrict the application of our formula translation to logic programs.
Although we focus on normal programs,
a similar development with other classes of logic programs, like disjunctive ones, can be done accordingly.
For simplicity, we write $\auxatm{a}$ instead of $\auxatom{\neg a}$ for $a\in \ensuremath{\mathcal{A}}} % {\ensuremath{\mathit{At}}$ and let $\{\auxatm{a}\}$ stand for $\auxatm{a}\vee\neg\auxatm{a}$.
Note that, for a rule $r$ as in \eqref{rule}, the set $\negocc{r}$ consists of negative literals only.
The next two definitions specialize our translation of formulas to logic programs.
\begin{definition}\label{def:rule:translation}
Let $r$ be a rule over \ensuremath{\mathcal{A}}} % {\ensuremath{\mathit{At}}\ as in \eqref{rule} with $m\geq 1$.
Then, we define
\begin{align*}
\translation{r}
&\textstyle =
r\big[\neg a/\auxatm{a}\mid \neg a \in \negocc{r}]
\cup
\bigcup_{\neg a \in \negocc{r}}
\left\{\{\auxatm{a}\} \leftarrow\right\}
\cup
\bigcup_{\neg a \in \negocc{r}}
\left\{
\begin{array}{c@{\ }l}
\leftarrow& a\wedge \auxatm{a} \\
\leftarrow& \neg a\wedge \neg\auxatm{a}
\end{array}
\right\}
\end{align*}
\end{definition}
\begin{definition}\label{def:program:translation}
Let $P$ be a logic program over \ensuremath{\mathcal{A}}} % {\ensuremath{\mathit{At}}.
Then,
\(
\translation{P} = \bigcup_{r \in P} \translation{r}
\).
\end{definition}
This translation substitutes negated literals in rule bodies with fresh atoms and
adds a choice rule along with a pair of integrity constraints providing an equivalence between
the eliminated negated body literals and the substituted atoms.
By applying the above results in the setting of logic programs,
we get that
a logic program and its translation have the same stable models when restricted to the original vocabulary.
\begin{corollary}\label{cor:extension:definition:program}
Let $P$ be a logic program over \ensuremath{\mathcal{A}}} % {\ensuremath{\mathit{At}}.
Then, we have
\(
P \cong_{\ensuremath{\mathcal{A}}} % {\ensuremath{\mathit{At}}} \translation{P}
\)
\end{corollary}
In other words,
every stable model of a logic program can be extended to a stable model of its translation
and vice versa.
\begin{corollary}\label{cor:extension:definition:program:one:one}
Let $P$ be a logic program over \ensuremath{\mathcal{A}}} % {\ensuremath{\mathit{At}}.
\begin{enumerate}
\item If $T\subseteq\ensuremath{\mathcal{A}}} % {\ensuremath{\mathit{At}}$ is a stable model of $P$,
then $T \cup \{\auxatm{a} \mid \neg a \in \negocc{P} \text{ and } a \not\in T \}$ is a stable model of \translation{P}.
\item $T\subseteq(\ensuremath{\mathcal{A}}} % {\ensuremath{\mathit{At}}\cup \{\auxatm{a} \mid \neg a \in \negocc{P} \}$ is a stable model of \translation{P},
then $T\cap\ensuremath{\mathcal{A}}} % {\ensuremath{\mathit{At}}$ is a stable model of $P$.
\end{enumerate}
\end{corollary}
For illustration,
consider the following example.
\begin{example}\label{ex:program:translation}
Consider the normal logic program $P$:
\[
\begin{array}{rcl}
a &\leftarrow& \\
b &\leftarrow& \neg c\\
c &\leftarrow& \neg b\\
d &\leftarrow& a \wedge \neg c
\end{array}
\]
Then, \translation{P} is:
\[
\begin{array}{rclrclrcl}
a &\leftarrow& & \{\auxatm{b} \} &\leftarrow& & \{\auxatm{c} \} &\leftarrow& \\
b &\leftarrow& \auxatm{c} & &\leftarrow& b\wedge\auxatm{b} & &\leftarrow& c\wedge\auxatm{c}\\
c &\leftarrow& \auxatm{b} & &\leftarrow& \neg b\wedge \neg \auxatm{b} & &\leftarrow& \neg c\wedge \neg \auxatm{c} \\
d &\leftarrow& a \wedge \auxatm{c}
\end{array}
\]
The stable models of $P$ are $\{a,b,d\}$ and $\{a,c\}$ and the ones of \translation{P} are $\{a,b,d,\auxatm{c}\}$ and $\{a,c,\auxatm{b}\}$, respectively.
\end{example}
The example underlines that our translation maps normal rules to definite ones along with choices and pairs of integrity constraints.
In other words,
it can be seen as a means for expressing normal logic programs in the form of programs with facts, definite rules, choice rules and integrity
constraints over an extended vocabulary.
We call this class of programs \emph{austere logic programs}, and further elaborate upon them in the following.
\subsection{Austere Logic Programs}
We define austere logic programs according to the decomposition put forward in the introduction.
\begin{definition}[Austere logic program]\label{def:austere:program}
An \emph{austere logic program} is a quadruple \linebreak
\(
(F,C,D,I)
\)
consisting of
a set $F$ of facts,
a set $C$ of choices,\footnote{That is, choice rules without body literals.}
a set $D$ of definite rules, and
a set $I$ of integrity constraints.
\end{definition}
A set of atoms is a stable model of an austere logic program, if it is a stable model of the union of all four components.
In view of the above results, austere logic programs can be regarded as a normalform for normal logic programs.
\comment{JR: This sounds a bit strange. Typically a normal form of some language is a subset of that language.
But here the austere programs have choice rules, that do not belong to normal logic programs.
In other words, the normal form of a normal logic program is not a normal logic program.\par
We could mention that it is useful for theory and practice to resort to such a form.
\par T: Right, we may have to talk about "even loops", sigh!
\par J: we must be careful then when we move to stratified negation, because there we forbid these loops.\par
Another option is simply to say that austere programs provide a normal form for normal logic programs with choice rules, or something like that.}
\begin{corollary}\label{cor:austere:normalform}
Every normal logic program can be expressed as an austere logic program and vice versa.
\end{corollary}
The converse follows from the fact that choice rules are expressible by a pair of normal rules~\cite{siniso02a}.
In fact, the (instantiation of) Listing~\ref{lp:ham} constitutes an austere logic program.
To see this observe that
\begin{itemize}
\item
lines~1-3 provide facts, $F$, capturing the problem instance,
here giving the specification of a graph;
\item
Line~5 provides choices, $C$, whose instantiation is derived from facts in the previous lines.
%
Grounding expands this rule to several plain choice rules with empty bodies;
\item
lines 5-6 list definite rules, $D$, defining (auxiliary) predicates used in the integrity constraints;
\item
finally, integrity constraints, $I$, are given in lines 7-9, stating conditions that solutions must satisfy.
\end{itemize}
This example nicely illustrates a distinguishing feature of austere logic programs,
namely, the \emph{compartmentalization} of the program parts underlying ASP's guess-define-check encoding methodology
(along with its strict separation of instance and encoding):
The problem instance is described by means of
\begin{itemize}
\item the facts in $F$
\end{itemize}
and the problem encoding confines
\begin{itemize}
\item non-deterministic choices to $C$,
\item the deterministic extension of the taken decisions to $D$, and
\item the test of the obtained extension to $I$.
\end{itemize}
This separation also confines the sources of multiple or non-existing stable models to well-defined locations, namely, $C$ and $I$, respectively
(rather than spreading them over several circular rules; see below).
As well,
the rather restricted syntax of each compartment gives rise to a very simple operational semantics of austere logic programs,
as we see in the next section.
\subsection{Modeling Methodology}
The backbone of easy ASP's modeling methodology is the structure imposed on its programs in Definition~\ref{def:easy:program}.
This allows us to encode problems by building concepts on top of each other.
Also, its structure allows for staying in full tune with ASP's \emph{guess-define-check} methodology~\cite{lifschitz02a}
by offering well-defined spots for all three parts.
Easy logic programs tolerate normal rules
in order to encompass full-featured ASP modeling languages.
Consequently, the interplay of the guess, define, and check parts of an easy logic program defies any control.
To tame this opening, we propose to carry over Observation~\ref{obs:uniqueness} to easy logic programs:
For any easy logic program
\(
(F,C,D_1,\dots,D_n,I)
\)
and $C' \subseteq F_C$,
the logic program $F\cup C'\cup D_1\cup\dots\cup D_n$ should have a unique stable model.
Even better if this can be obtained in a deterministic way.
This leads us to the following advice on easy ASP modeling:
\begin{enumerate}
\item Compartmentalize a logic program into
facts, $F$,
choice rules, $C$,
normal rules, $D_1\cup\dots\cup D_n$, and
integrity constraints $I$,
such that the overall logic program has stratification $(F,C,D_1,\dots,D_n,I)$.
\item
Aim at defining one predicate per stratum $D_i$
and
avoid cycles within each $D_i$ for $i=1,\dots,n$.
\item Ensure that $F\cup C'\cup D_1\cup\dots\cup D_n$ has a unique stable model for any $C' \subseteq F_C$.
\end{enumerate}
While the first two conditions have a syntactic nature and can thus be checked automatically,
the last one refers to semantics
and, to the best of our knowledge, has only sufficient but no known necessary syntactic counterparts.
One is to restrict $D_1\cup\dots\cup D_n$ to definite rules as in austere programs,
the other is to use stratified negation, as proposed in~\cite{niemela08b} and detailed in the next section.
Our favorite is to stipulate that
\(
{F\cup C'\cup D_1\cup\dots\cup D_n}$ has a total well-founded model~\cite{gerosc91a} for any $C' \subseteq F_C
\)
but unfortunately, we are unaware of any syntactic class of logic programs warranting this condition beyond the ones mentioned above.
\subsection{Operational Semantics}
In our experience,
a major factor behind the popularity of the approach sketched in the introductory section
lies in the possibility to intuitively form stable models along the order of the rules in a program.
In fact,
the simple nature of austere logic programs provides a straightforward scheme for computing stable models
by means of the well-known immediate consequence operator,
whose iteration mimics this proceeding.
Moreover,
the simplicity of the computation provides the first evidence of the value of austere logic programs as a normalform.
The operational semantics of austere logic programs follows ASP's \emph{guess-define-check} methodology.
In fact,
the only non-determinism in austere logic programs is comprised of choice rules.
Hence,
once choices are made,
we may adapt well-known deterministic bottom-up computation techniques for computing stable models.
However,
the results of this construction provide merely candidate solutions
that still need to satisfy all integrity constraints.
If this succeeds, they constitute stable models of the austere program.
Let us make this precise for an austere logic program $(F,C,D,I)$ in what follows.
To make choices and inject them into the bottom-up computation,
we translate the entire set of choices, $C$, into a set of facts:
\[
F_C = \{{a \leftarrow}\mid{\{a\} \leftarrow }\in C\}
\]
A subset of $F_C$, the original facts $F$, along with the definite program $D$ are then passed to a
corresponding consequence operator that determines a unique stable model candidate.
More precisely,
the $T_P$ operator of a definite program $P$ is defined for an interpretation $X$ as follows~\cite{lloyd87}:
\[
T_P(X) = \{ l_1 \mid (l_1\leftarrow l_{m+1}\wedge\dots\wedge l_n) \in P,\ X \models l_{m+1}\wedge\dots\wedge l_n \}
\]
With this, the candidate solutions of an austere program can be defined.
\begin{definition}\label{def:candidate:stable:model}
Let $(F,C,D,I)$ be an austere logic program over \ensuremath{\mathcal{A}}} % {\ensuremath{\mathit{At}}.
We define a set $X\subseteq\ensuremath{\mathcal{A}}} % {\ensuremath{\mathit{At}}$ of atoms as a candidate stable model of $(F,C,D,I)$,
if $X$ is the least fixpoint of $T_{F\cup C'\cup D}$ for some $C' \subseteq F_C$.
\end{definition}
The existence of the least fixpoint is warranted by the monotonicity of $T_{F\cup C'\cup D}$~\cite{lloyd87}.
Similar to traditional ASP,
several candidate models are obtained via the different choices of $C'$.
While the choice of $C'$ constitutes the \emph{guess} part and the definite rules in $D$ the \emph{define} part of the approach,
the \emph{check} part is accomplished by the integrity constraints in~$I$.
\begin{proposition}\label{prop:austere:stable:models}
Let $(F,C,D,I)$ be an austere logic program over \ensuremath{\mathcal{A}}} % {\ensuremath{\mathit{At}}\
and $X\subseteq\ensuremath{\mathcal{A}}} % {\ensuremath{\mathit{At}}$.
Then, $X$ is a stable model of $(F,C,D,I)$
iff
$X$ is a candidate stable model of $(F,C,D,I)$ such that $X \models I$.
\end{proposition}
We illustrate the computation of stable models of austere logic programs in the following example.
\begin{example}
Consider the austere logic program $P$
\[
\begin{array}{rcl}
a &\leftarrow& \\
\{b\} &\leftarrow& \\
c &\leftarrow& b\\
&\leftarrow& a \wedge \neg c\\
\end{array}
\]
We get the candidate stable models $\{a,b,c\}$ and $\{a\}$ from the first three rules depending on whether
we choose $b$ to be true or not,
that is, whether we add the fact $b \leftarrow$ or not.
%
Then,
on testing them against the integrity constraint expressed by the fourth rule,
we see that $\{a,b,c\}$ is indeed a stable model, since it satisfies the integrity constraint,
while set~$\{a\}$ is not a stable model since checking the integrity constraint fails.
\end{example}
A major intention of austere logic programs is to confine the actual guess and check of an encoding to dedicated components,
namely, the choices in $C$ and constraints in $I$.
The definite rules in $D$ help us to analyze and/or extend the solution candidate induced by the facts $F$ and the actual choices in $C'$.
The emerging candidate is then evaluated by the integrity constraints in $I$.
This stresses once more the idea that the extension of a guessed solution candidate should be deterministic;
it elaborates the guess but refrains from introducing any ambiguities.
This is guaranteed by the definite rules used in austere programs.
\begin{observation}\label{obs:uniqueness}
For any austere logic program $(F,C,D,I)$ and $C' \subseteq F_C$,
the logic program $F\cup C'\cup D$ has a unique stable model.
\end{observation}
This principle is also in accord with~\cite{niemela08b}, where stratified logic programs are used instead of definite ones (see below).
\subsection{Optimization}
\label{sec:optimization}
Strictly speaking, optimization statements do not belong to a logic program and are rather instructions to an ASP system.
Since they take predicates as argument but refrain from deriving any predicates,
it is natural to state them at the end of a program.
For instance, we can turn the formalization of the Hamiltonian circuit problem in Listing~\ref{lp:ham} into one of the Traveling salesperson problem
by adding edge costs to the facts in lines 1-3, such as
\begin{lstlisting}[firstnumber=2]
cost(1,2,3). cost(2,3,5). cost(2,4,6). cost(3,1,4).
cost(3,4,7). cost(4,1,5). cost(4,3,7).
\end{lstlisting}
along with an optimization directive of the form:
\begin{lstlisting}[firstnumber=12]
#minimize { C,X,Y : hc(X,Y), cost(X,Y,C) }.
\end{lstlisting}
This statement tells the solver to search for solutions having a cost minimal Hamiltonian circuit;
it naturally extends the stratification of the easy program, since it only refers to predicates defined in previous lines,
and does not contribute to the formation but only the selection of stable models, just like integrity constraints.
\section{Proofs}
\input{proofs/ht_model_correspondence}
\input{proofs/pse_formula_extension}
\input{proofs/extension_definition_one_one}
\input{proofs/entail_proof}
\input{proofs/formula_se}
\input{proofs/pse_translation}
\input{proofs/sm_correspondence_translation}
\input{proofs/program_sm_correspondence}
\input{proofs/prog_ext}
\input{proofs/csm_to_sm}
\section{Related Work}
\label{sec:related}
Apart from advocating the idea illustrated in the introduction,
Ilkka Niemel{\"{a}} also showed in~\cite{niemela08b} that
negative body literals can be replaced by a new atom for which
a choice needs to be made whether to include it in the model or not;
and
such that a model cannot contain both the new atom and the atom of the replaced literal
but one of them needs to be included.
This technique amounts exactly to the transformation in Definition~\ref{def:rule:translation}
and traces back to Abductive logic programming~\cite{eshkow89a,denkak02a}.
Indeed, it was already shown in \cite{grsaza01a} that for DATALOG queries the expressive power of stable model semantics can be achieved via stratified negation and choices.
We elaborated upon this idea in several ways.
First, we have shown that the full expressiveness of normal logic programs can even be achieved with definite rules rather than normal ones with
stratified negation.
Second, we have provided a strong equivalence result that allows for applying the transformation in Definition~\ref{def:rule:translation} to selected
rules only.
Third, we have generalized the idea by means of the logic of Here-and-There, which made it applicable to other fragments of logic programs.
And finally, this investigation has revealed that the roots of the idea lie in the logical principle of extension by definition.
Over the last decades many more related ideas were presented in the literature.
For instance, in \cite{felomisu93a}, normal logic programs are translated into positive disjunctive programs by introducing new atoms for negative literals.
Also, strong negation is usually compiled away via the introduction of new atoms along with integrity constraints excluding that both the original atom
and the atom representing its strong negation hold~\cite{gellif90a}.
The principle of extension by definition was also used in~\cite{ferlif02a} to prove properties about programs with nested expressions.
EzASP is closely related to the paradigm of IDP~\cite{debobrjade18a},
where the program parts $F$, $C$ and $I$ are expressed in first-order logic,
while the $D_i$'s form inductive definitions.
Finally, in~\cite{delitrve19a},
the authors propose an informal semantics for logic programs based on the guess-define-check methodology,
that are similar to the easy logic programs that we introduce in this paper.
\subsection{Stratified Negation}
\label{sec:stratified:negation}
The purpose of stratified negation is to eliminate the (problematic) case of recursion through negation.
What makes this type of recursion problematic is that it may eliminate stable models and that the actual source may be spread over several rules.
To give some concise examples, consider the programs
\(
\{a\leftarrow \neg a\}
\)
and
\(
\{
a\leftarrow \neg b,
b\leftarrow \neg c,
c\leftarrow \neg a
\}
\)
admitting no stable models.
Following the dependencies in both examples,
we count one and three dependencies, respectively, all of which pass through negated body literals.
More generally, cyclic dependencies traversing an odd number of negated body literals (not necessarily consecutively) are known sources of incoherence.
Conversely, an even number of such occurrences on a cycle is not harmful but spawns alternatives, usually manifested in multiple stable models.
To see this, consider the program
\(
{\{a\leftarrow \neg b, b\leftarrow \neg a\}}
\)
producing two stable models.
Neither type of rule interaction is admitted in austere logic programs.
Rather the idea is to confine the sources of multiple and eliminated stable models to dedicated components,
namely, choices and integrity constraints.
The same idea was put forward by Niemel{\"{a}} in~\cite{niemela08b}
yet by admitting a more general setting than definite rules
by advocating the concept of stratified negation.
To eliminate the discussed cyclic constellations,
stratified negation imposes an additional constraint on the stratification of a logic program:
Given the prerequisites of Definition~\ref{def:stratification}, we define:
\begin{enumerate}\setcounter{enumi}{2}
\item If a predicate symbol $q$ occurs in a negative body literal of a rule in $P_i$,
then $\mathit{def}(q)\subseteq P_j$ for some $j<i$.
\end{enumerate}
In other words,
while the definitions of predicates appearing positively in rule bodies may appear in a lower or equal partition,
the ones of negatively occurring predicates are restricted to lower components.
Although this condition tolerates positive recursion as in
\(
\{a\leftarrow b, b\leftarrow a\}
\),
it rules out negative recursion as in the above programs.
Since using programs with stratified negation rather than definite programs generalizes austere logic programs,
their combination with choices and integrity constraints is also as expressive as full ASP~\cite{niemela08b}.
An example of stratified negation can be found in Listing~\ref{lst:toh}.
The negative literal in Line~5 refers to a predicate defined --- beforehand --- in Line~8.
An attractive feature of normal logic programs with stratified negation is that they yield a unique stable model,
just as with austere programs (cf.~Observation~\ref{obs:uniqueness}).
Hence, they provide an interesting generalization of definite rules
maintaining the property of deterministically extending guessed solution candidates.
|
\section{Introduction}
Without symmetries to prevent it, scalar fields that are realised in nature will inevitably couple to any sector of a particle physics model that knows about gravity. This might, e.g., be through Higgs portals \cite{Silveira:1985rk, McDonald:1993ex, Burgess:2000yq, Davoudiasl:2004be, Schabinger:2005ei,Patt:2006fw}, through neutrino portals \cite{Falkowski:2009yz,GonzalezMacias:2015rxl} or through direct couplings to curvature invariants, such as the Ricci scalar in the case of theories of Brans-Dicke type~\cite{Brans:1961sx}. However, the ability to make Weyl rescalings in metric theories of gravity makes it difficult to disentangle these various ways that scalar fields can couple to matter~\cite{Burrage:2018dvt, Millington:2019jbx}. Theories of gravity in which scalar fields couple directly to curvature terms are often referred to as scalar-tensor theories \cite{FM}.
In fact, as far as quantum effects are concerned, any theory of gravity plus a scalar field should be regarded as a scalar-tensor theory: If, e.g., we tune away a direct coupling to the Ricci scalar at tree level or at a given energy scale, it will be regenerated by loop corrections or the renormalization group running to another scale (see, e.g., Refs.~\cite{Herranen:2014cua, Markkanen:2018bfx}). Hence, if we want to treat quantum field theories that know about gravity and involve scalar fields, we should treat them in the so-called Jordan frame (or Jordan-like frames), wherein all of the allowed couplings are present, be they to the gravitational or non-gravitational sectors. The import is that the so-called screening mechanisms (see, e.g.,~\cite{Joyce:2014kja, Burrage:2017qrf}) that can allow scalar-tensor theories to evade local tests of gravity can similarly apply to Higgs-portal theories~\cite{Burrage:2018dvt, Brax:2021rwk}. We will not take screening into account in this work, but examples include the chameleon~\cite{Khoury:2003aq, Khoury:2003rn}, symmetron~\cite{Hinterbichler:2010es, Hinterbichler:2011ca} and Vainshtein~\cite{Vainshtein:1972sx} mechanisms, and their variants. For recent reviews on experimental and observational constraints (on screened) fifth forces, see Refs.~\cite{AlvesBatista:2021gzc,Burrage:2017qrf,Brax:2021wcv}.
One symmetry that can affect the way that additional scalar degrees of freedom from the gravitational sector couple to matter is scale symmetry (or, in the local case, Weyl symmetry). In general, these scalars will mediate fifth forces between matter sources, and new long-range forces of nature are heavily constrained \cite{Burrage:2017qrf}. However, if the matter sector is scale invariant, the classical fifth forces are suppressed. In the case of the Standard Model (SM) of particle physics, the scale breaking term that plays a dominant role in allowing gravitational scalars to couple to matter is the quadratic term of the SM Higgs potential. This becomes clear when one works in the Einstein frame, as was described in Ref.~\cite{Burrage:2018dvt}. On the other hand, the role played by the same operator is less obvious when considered in the Jordan frame.
A popular example of a model that exploits scale symmetry to prevent the emergence of fifth forces is the Higgs-dilaton theory~\cite{Wetterich:1987fm, Buchmuller:1988cj, Shaposhnikov:2008xb, Shaposhnikov:2008xi, Blas:2011ac, Garcia-Bellido:2011kqb, Garcia-Bellido:2012npk, Bezrukov:2012hx, Henz:2013oxa, Rubio:2014wta, Karananas:2016grc, Ferreira:2016vsc, Ferreira:2016kxi, Casas:2017wjh, Ferreira:2018qss}. Therein, both the Higgs field and an additional gauge-singlet scalar are non-minimally coupled to the Ricci scalar. The action has a global Weyl symmetry, which is broken when the fields acquire vacuum expectation values (vevs) along a particular trajectory in field space. Note that the structure of the Higgs potential remains to be probed experimentally~\cite{Maltoni:2018ttu,Borowka:2018pxx,ATLAS:2019pbo,CMS:2018ipl,Tian:2016qlk,Abramowicz:2017,Agrawal:2019bpm,Nielsen:2020uvd}. However, due to the dynamical origin of the scale breaking, the massless mode does not couple to the matter sector, leaving a theory devoid of fifth forces~\cite{Ferreira:2016kxi,Brax:2014baa, Ferreira:2016vsc}. The role of explicit versus spontaneous (in this dynamical sense) scale breaking in the emergence of fifth forces was described in Ref.~\cite{Burrage:2018dvt}, where the analysis was performed in the Einstein frame. In this paper, we will show explicitly how the same arguments are borne out directly in the Jordan frame.
The merits of dealing with scalar-tensor theories in the Jordan frame are threefold: First, there are theories for which an Einstein frame does not, in general, exist, such as Horndeski \cite{Horndeski:1974wa, Kobayashi:2019hrl}, beyond Horndeski \cite{Traykova:2019oyx, Gleyzes:2013ooa} and DHOST \cite{Langlois:2015cwa, Langlois:2018dxi} theories. Second, as discussed above, in an interacting quantum field theory, the Einstein frame may exist only at a particular energy scale, with loop corrections and the renormalization of couplings regenerating Jordan-like frames. Third, the conformal transformation to the Einstein frame, and the subsequent field redefinitions needed to bring the theory as close to being canonically normalized as possible (notwithstanding any curvature of the field-space metric) must be performed on a model-by-model basis and may not be easily automated.
In order to proceed directly in the Jordan frame, we will need to work with linearized gravity, which cannot be decoupled (as it can be by construction if we work in the Einstein frame) due to the presence of the non-minimal coupling. By this means, we are able to isolate the kinetic mixings between the graviton and non-minimally coupled scalar field that can give rise to fifth forces. Most importantly, we find that the non-minimal coupling, through its effects on the geodesics of the spacetime, necessitates an update to the covariant derivative. Focusing on the (would-be) harmonic gauge, we determine the fifth-force potential by means of the non-relativistic limit of the scalar-mediated scattering matrix elements. Moreover, our results agree with those obtained in the Einstein frame, as per Ref.~\cite{Burrage:2018dvt}.
This paper is intended to be both explicit and pedagogical. As such, we begin in Sec.~\ref{second} with a review of previous studies on the relation between scale-symmetry breaking and fifth forces. After properly identifying the symmetries of Brans-Dicke-type \cite{Brans:1961sx} scalar-tensor theories in Sec.~\ref{thirdS}, we define a so-called scalar-harmonic gauge, by updating the harmonic gauge from Einstein gravity, and perform the subsequent linearization of the gravitational sector. From there, we consider the fifth-force contributions to the M\o{}ller scattering in Sec.~\ref{fourth} and obtain the Yukawa potential from its non-relativistic limit, concentrating on systems where the scale symmetry is broken either explicitly or dynamically. Our conclusions are presented in Sec.~\ref{fifth}. Additional details are provided in the Appendix.
\section{Fifth forces and scale symmetry in the Einstein frame}\label{second}
In this section, we review the Einstein-frame description of fifth forces, emphasizing the key role of explicit sources of scale breaking in the matter sector in allowing fifth forces to couple to matter fields. What follows is based heavily on Ref.~\cite{Burrage:2018dvt}.
In the Jordan frame, the equations of motion for the class of scalar-tensor theories on which we will focus can be derived from the following action:
\begin{equation}
S=\int \textnormal{d}^4{x} \sqrt{-g} \left[\frac{1}{2}F(X)R - \frac{1}{2}Z(X)g^{\mu\nu}\partial_\mu X \partial_\nu X - U(X)\right] + S_{\rm{m}}[ g_{\mu \nu},\{\psi\}],
\end{equation}
where the real scalar field $X$ couples non-minimally to gravity through the function $F(X)$ and evolves subject to the potential $U(X)$. $Z(X)$ allows for a non-canonical kinetic term. In addition, $R$ is the Ricci scalar, defined in terms of the Jordan-frame metric $g_{\mu\nu}$, and $S_{\rm{m}} = \int \textnormal{d}^4{x} \sqrt{-g} {\cal L}_m$ is the matter action, containing the set of matter fields $\{\psi\}$. Throughout this article, we employ the mostly plus signature convention $(-,+,+,+)$.
We now have a choice:~we can either work directly in the Jordan frame, as we will do later in this article, or we can eliminate the direct coupling of the field $X$ to gravity, by making a Weyl rescaling of the metric, and work in the Einstein frame. Therein, the gravity sector is of canonical Einstein-Hilbert form, and we may instead have direct couplings of the field $X$ to the matter sector. While calculations in the Jordan frame are complicated by the need to treat gravity dynamically, those in the Einstein frame are complicated by the Weyl rescaling itself and the need to rescale the matter fields, which must be done on a model-by-model basis, as we will now describe.
The Weyl rescaling takes the form
\begin{equation}
\label{eq:Weyl_rescaling}
{g}_{\mu \nu}=A^2(X)\tilde{g}_{\mu \nu},
\end{equation}
where $\tilde{g}_{\mu\nu}$ is the Einstein-frame metric, $A^2(X)=\tilde M_{\rm Pl}^2F^{-1}(X)$ is the squared coupling function and $\tilde M_{\rm Pl}$ is the reduced Planck mass of the Einstein frame.
After some algebra, the transformation~\eqref{eq:Weyl_rescaling} leads to the Einstein-frame action
\begin{equation}
\label{eq:EF_action_general}
S=\int \textnormal{d}^4{x} \sqrt{-\tilde{g}} \left[\frac{1}{2}\tilde M_{\rm Pl}^2\tilde{R} - \frac{1}{2}\tilde{g}^{\mu\nu}\partial_\mu \tilde{\chi} \partial_\nu \tilde{\chi} - \tilde{A}^4(\tilde{\chi})\tilde{U}(\tilde{\chi})\right] + S_{\rm{m}}[ \tilde{A}^2(\tilde{\chi})\tilde{g}_{\mu \nu},\{\psi\}],
\end{equation}
where $\tilde R$ is the Ricci scalar built with the Einstein-frame metric $\tilde{g}_{\mu\nu}$. In addition, $\tilde{A}(\tilde{\chi})\equiv A(X(\tilde{\chi}))$ and $\tilde{U}(\tilde{\chi})\equiv U(X(\tilde{\chi}))$ are respectively the coupling function and potential expressed in terms of the canonically normalized field
\begin{equation}
\tilde{\chi}(X)\equiv \tilde M_{\rm Pl}\int^{X}_{X_0}\textnormal{d}{\hat{X}} \sqrt{\frac{{Z}(\hat{X})}{{F}(\hat{X})}+3[{F'}(\hat{X})]^2},
\label{O2}
\end{equation}
where the prime in $F'(\hat{X})$ indicates the derivative with respect to the argument. We see from Eq.~\eqref{eq:EF_action_general} that, in the Einstein frame, matter fields still move on geodesics of the Jordan-frame metric, potentially giving rise to deviations from Einstein gravity that can be interpreted as a fifth force due to the exchange of fluctuations in the $\tilde{\chi}$ field.
If $\tilde{A}^2$ can be expanded about unity, such that $\tilde{A}^2(\tilde{\chi})=1+\frac{1}{n!}\frac{\tilde{\chi}^n}{M^n}+\dots$, we can similarly expand the matter action to show that
\begin{equation}
\label{eq:actionexpansion}
S_{\rm{m}}[\tilde{A}^{2}(\tilde{\chi}) \tilde{g}_{\mu \nu},\{\psi\}]= S_{\rm{m}}\left[\tilde{g}_{\mu \nu},\{\psi\}\right]+\frac{1}{n!}T\frac{\tilde{\chi}^{n}}{M^{n}}+\dots,
\end{equation}
wherein we see that the field $\tilde{\chi}$ couples to the trace of the energy-momentum tensor of the matter sector $T=\tilde{g}_{\alpha\beta}T^{\alpha\beta}$. Expressing the fifth-force coupling in this way, however, obscures the role that scale-symmetry breaking plays in the emergence of fifth forces~\cite{Burrage:2018dvt}, as we will describe in the next subsection.
\subsection{Fifth forces in the Einstein frame}
The fifth forces that arise in scalar-tensor theories can be studied in multiple ways. For instance, one can solve the classical equations of motion and extract the corrections to the Newtonian potential from the spatial gradient of the solution (see, e.g., Ref.~\cite{Joyce:2014kja, Burrage:2017qrf}). Equivalently, we can work with the tree-level matrix element for the $t$-channel exchange of the scalar mediator, from which we can then extract the non-relativistic potential (see, e.g., Refs.~\cite{Burrage:2018dvt, Banks:2020gpu}).
We consider the following toy model, as constructed in Ref.~\cite{Burrage:2018dvt} and written in terms of the Jordan-frame metric $g_{\mu \nu}$:
\begin{align}
S_{\rm{m}}=\int \textnormal{d}^4{x} \sqrt{-g}& \left[-\frac{1}{2}g^{\mu \nu}\partial_\mu \Phi \partial_\nu \Phi -\frac{1}{2}g^{\mu \nu}\partial_\mu \Theta \partial_\nu \Theta -\bar{\psi}i{\stackrel{\leftrightarrow}{\slashed{\partial}}}\psi - y\bar{\psi}\Phi\psi\right.\nonumber\\ &\qquad \left.-U(\Phi, \Theta)+\frac{1}{2} \mu_{\theta}^{2} A^{-2}(X) \Theta^{2}-\frac{\lambda_{\theta}}{4 !} \Theta^{4}-\frac{3}{2} \frac{\mu_{\theta}^{4}}{\lambda_{\theta}} A^{-4}(X)\right],
\label{sm}
\end{align}
where
\begin{equation}
U(\Phi, \Theta)=\frac{\lambda}{4 !}\left(\Phi^{2}-\frac{\beta}{\lambda} \Theta^{2}\right)^{2}-\frac{1}{2} \mu^{2}\left(\Phi^{2}-\frac{\beta}{\lambda} \Theta^{2}\right)+\frac{3}{2} \frac{\mu^{4}}{\lambda} .\label{comb_pot}
\end{equation}
The matter sector contains a would-be Higgs field $\Phi$ and a Dirac fermion $\psi$ (a proxy for the SM electron). The fermion kinetic term has been antisymmetrized, with ${\stackrel{\leftrightarrow}{\slashed{\partial}}}\equiv e^{\mu}_{a}\gamma^a{\stackrel{\rightarrow}{\partial}}_\mu-{\stackrel{\leftarrow}{\partial_\mu}} e^{\mu}_{a}\gamma^a$ (where we make use of the vierbein $e^{\mu}_{a}$), so that we can omit the spin connection from the action (see refs.~\cite{Burrage:2018dvt, Ferreira:2016kxi}). These two fields interact through a Yukawa coupling, which gives mass to the fermion after the would-be Higgs sector undergoes symmetry breaking. The additional $\Theta$ field has been introduced so that we can move smoothly between two scenarios of scale breaking, and independently of the non-minimally coupled field $X$ and its dynamics:~the first ($\beta \to 0$) in which the scale breaking is explicit, due to the presence of the dimensionful parameter $\mu$ in the potential of the matter fields; the second ($\mu \to 0$) in which the scale breaking is dynamical, arising along a particular trajectory in the $\Phi-\Theta$ field space without explicit scale-breaking terms appearing in the Lagrangian. The specific choice of couplings between the $\Theta$ and $X$ fields has been tuned so that these fields do not have a mass mixing in the Einstein frame [see, e.g., Eq.~\eqref{eq:EFrameL_with_theta} below], while also allowing us to establish a hierarchy between the masses of the three physical modes (see Ref.~\cite{Burrage:2018dvt}).
The two limiting cases for the distinct sources of scale breaking are as follows:
\paragraph*{\bf Pure explicit breaking (prototype SM Higgs sector) $\boldsymbol{\beta\to 0}$:}
In the limit in which $\beta\to0$, the mixings between $\Phi$ and $\Theta$ vanish, decoupling $\Theta$ from the matter Lagrangian. We are then left with the following potential in the Jordan frame [Eq.~\eqref{comb_pot}]:
\begin{equation}
U(\Phi)=\frac{\lambda}{4 !}\Phi^{4}-\frac{1}{2} \mu^{2}\Phi^{2}+\frac{3}{2} \frac{\mu^{4}}{\lambda},
\end{equation}
which is just the prototype of the SM Higgs potential, where we can see that the constant term in the potential $U$ ensures that the vacuum has zero energy density in the symmetry-broken phase of the would-be Higgs sector. In this case, the term quadratic in $\Phi$ provides an explicit source of scale breaking through the dimensionful mass parameter $\mu$. As we will show below, it is this term that plays the key role in allowing fifth forces to couple to the fermion field $\psi$.
\paragraph*{\bf Pure dynamical scale breaking (prototype Higgs-dilaton model) $\boldsymbol{\mu\to0}$:}
In this limit, all the sources of explicit scale breaking vanish from $U(\Phi,\Theta)$, leaving a scale-invariant potential. We therefore do not expect the conformal field $X$ to couple to this potential in the Einstein frame, leaving the fermionic sector free of fifth forces. The potential is reduced to
\begin{equation}
U(\Phi, \Theta)=\frac{\lambda}{4 !}\left(\Phi^{2}-\frac{\beta}{\lambda} \Theta^{2}\right)^{2},
\end{equation}
and analogous potentials appear in Higgs-dilaton theories~\cite{Wetterich:1987fm, Buchmuller:1988cj, Shaposhnikov:2008xb, Shaposhnikov:2008xi, Blas:2011ac, Garcia-Bellido:2011kqb, Garcia-Bellido:2012npk, Bezrukov:2012hx, Henz:2013oxa, Rubio:2014wta, Karananas:2016grc, Ferreira:2016vsc, Ferreira:2016kxi, Casas:2017wjh, Ferreira:2018qss}. In those scenarios, however, both $\Phi$ and $\Theta$ are non-minimally coupled to the Ricci scalar in the Jordan frame (and the dilaton is the light degree of freedom with the potential to mediate long-range forces), whereas we will take only the additional field $X$ to be non-minimally coupled.
The classical scale symmetry of this model is broken when the scalar fields $\Phi$ and $\Theta$ obtain non-vanishing vevs, leading to the emergence of a scale.
Since this scale appears indirectly through the stabilization of the fields, we will refer to this as dynamical scale breaking~\cite{Ferreira:2018itt}.
In contrast to the case of explicit scale breaking, the fifth force mediated by $X$ will not couple to the matter fields $\psi$. In what follows, we will show this by explicit calculation of the matrix elements. Alternatively, this can be understood in terms of the existence of a conserved dilatation current (see, e.g., Ref.~\cite{Ferreira:2016kxi, Brax:2014baa}).
Let us now turn our attention to an explicit calculation of the fifth forces for this model. First, we need to express the matter action in terms of the Einstein-frame metric $\tilde{g}_{\mu \nu}$. To do so, we must perform the conformal transformation defined previously in Eq.~(\ref{eq:Weyl_rescaling}). This gives
\begin{align}
S_{\rm{m}}&=\int \textnormal{d}^4{x} \sqrt{-\tilde{g}} \left[-\frac{1}{2}\tilde{A}^2(\tilde{\chi})\tilde{g}^{\mu \nu}\partial_\mu \Phi \partial_\nu \Phi -\frac{1}{2}\tilde{A}^2(\tilde{\chi})\tilde{g}^{\mu \nu}\partial_\mu \Theta \partial_\nu \Theta-\tilde{A}^3(\tilde{\chi})\bar{\psi}i\stackrel{\leftrightarrow}{\tilde{\slashed{\partial}}}\psi \right.\nonumber\\ &\qquad \left.- y \tilde{A}^4(\tilde{\chi})\bar{\psi}\Phi\psi-\tilde{A}^4(\tilde{\chi})U(\Phi, \Theta)+\frac{1}{2}\tilde{A}^2(\tilde{\chi}) \mu_{\theta}^{2} \Theta^{2}-\frac{\lambda_{\theta}}{4 !} \tilde{A}^4(\tilde{\chi})\Theta^{4}-\frac{3}{2} \frac{\mu_{\theta}^{4}}{\lambda_{\theta}} \right],
\label{sm1}
\end{align}
where $\tilde{\chi}\equiv \tilde{\chi}(X)$ is the canonically normalized field [cf.~Eq.~\eqref{O2}] and $\stackrel{\leftrightarrow}{\tilde{\slashed{\partial}}}\equiv \tilde{e}^{\mu}_{a}\gamma^a{\stackrel{\rightarrow}{\partial}}_\mu-{\stackrel{\leftarrow}{\partial_\mu}} \tilde{e}^{\mu}_{a}\gamma^a=\tilde{A}(\tilde{\chi})e^{\mu}_{a}\gamma^a{\stackrel{\rightarrow}{\partial}}_\mu-{\stackrel{\leftarrow}{\partial_\mu}} e^{\mu}_{a}\gamma^a\tilde{A}(\tilde{\chi})$. To leave the matter sector as close to being canonically normalized as possible, we redefine the fields according to their classical scaling dimensions, such that
\begin{align}
\tilde{\phi} \equiv \tilde{A}(\tilde{\chi}) \Phi, \qquad\tilde{\theta} \equiv \tilde{A}(\tilde{\chi}) \Theta, \qquad \tilde{\psi} \equiv \tilde{A}^{3 / 2}(\tilde{\chi}) \psi.
\label{scasca}
\end{align}
With this, the Lagrangian becomes
\begin{align}
\tilde{\mathcal{L}}_{\rm{m}}=& -\frac{1}{2} \tilde{g}^{\mu \nu} \partial_{\mu} \tilde{\phi} \partial_{\nu} \tilde{\phi}+\tilde{g}^{\mu \nu} \tilde{\phi} \partial_{\mu} \tilde{\phi} \partial_{\nu} \ln \tilde{A}(\tilde{\chi})-\frac{1}{2} \tilde{g}^{\mu \nu} \tilde{\phi}^{2} \partial_{\mu} \ln \tilde{A}(\tilde{\chi}) \partial_{\nu} \ln \tilde{A}(\tilde{\chi}) \nonumber\\
& -\frac{1}{2} \tilde{g}^{\mu \nu} \partial_{\mu} \tilde{\theta} \partial_{\nu} \tilde{\theta}+\tilde{g}^{\mu \nu} \tilde{\theta} \partial_{\mu} \tilde{\theta} \partial_{\nu} \ln \tilde{A}(\tilde{\chi})-\frac{1}{2} \tilde{g}^{\mu \nu} \tilde{\theta}^{2} \partial_{\mu} \ln \tilde{A}(\tilde{\chi}) \partial_{\nu} \ln \tilde{A}(\tilde{\chi}) \nonumber\\
& +\tilde{U}(\tilde{\phi},\tilde{\theta},\tilde{\chi})-\bar{\tilde{\psi}}i\stackrel{\leftrightarrow}{\tilde{\slashed{\partial}}}\tilde{\psi}-y \bar{\tilde{\psi}} \tilde{\phi} \tilde{\psi}-\frac{1}{2} \mu_{\theta}^{2} \tilde{\theta}^{2}+\frac{\lambda_{\theta}}{4 !} \tilde{\theta}^{4}+\frac{3}{2} \frac{\mu_{\theta}^{4}}{\lambda_{\theta}},\label{mat_lag}
\end{align}
where
\begin{align}
\label{eq:EFrameL_with_theta}
\tilde{U}(\tilde{\phi}, \tilde{\theta}, \tilde{\chi})=\frac{\lambda}{4 !}\left(\tilde{\phi}^{2}-\frac{\beta}{\lambda} \tilde{\theta}^{2}\right)^{2}-\frac{1}{2} \mu^{2}\tilde{A}^2(\tilde{\chi})\left(\tilde{\phi}^{2}-\frac{\beta}{\lambda} \tilde{\theta}^{2}\right)
+\frac{3}{2} \tilde{A}^4(\tilde{\chi})\frac{\mu^{4}}{\lambda}.
\end{align}
Thus, we can see that the redefinitions from Eq.~(\ref{scasca}) eliminate all the couplings of $\tilde{\chi}$ in the fermionic sector and in the pure $\Theta$ potential from the last line of Eq.~\eqref{mat_lag}. However, the same does not apply to $\tilde{U}(\tilde{\phi}, \tilde{\theta}, \tilde{\chi})$, since it contains dimensionful parameters. Moreover, as explained before, the only terms coupling to $\tilde{\chi}$ in the Einstein frame are the ones which break the scale symmetry explicitly.
Keeping the calculation as generic as possible, we now expand the coupling function as
\begin{align}
\tilde{A}^{2}(\tilde{\chi})=a+b \frac{\tilde{\chi}}{\tilde{M}}+c \frac{\tilde{\chi}^{2}}{\tilde{M}^{2}}+\mathcal{O}\left(\frac{\tilde{\chi}^{3}}{\tilde{M}^{3}}\right),
\label{para}
\end{align}
where $\tilde{M}$ is an energy scale, and $a$, $b$ and $c$ are dimensionless constants, which will be defined for specific models. After including the original kinetic energy term for $\tilde{\chi}$ from Eq.~\eqref{eq:EF_action_general} and making a further redefinition $\partial_\mu \tilde{\chi} \to \sqrt{1+\tilde{\theta}^2 + \tilde{\phi}^2}\; \partial_{\mu} \ln \tilde{A}(\tilde\chi)$, the non-gravitational part of the Einstein-frame Lagrangian can be written up to second order in $\tilde{M}^{-1}$ as
\begin{align}
\tilde{\mathcal{L}}=& -\frac{1}{2} \tilde{g}^{\mu \nu} \partial_{\mu} \tilde{\chi} \partial_{\nu} \tilde{\chi}-\frac{1}{2} \tilde{g}^{\mu \nu} \partial_{\mu} \tilde{\phi} \partial_{\nu} \tilde{\phi}+\frac{1}{2} \tilde{g}^{\mu \nu} \frac{\tilde{\phi}}{\tilde{M}}\left(b+2 a c \frac{\tilde{\chi}}{\tilde{M}}-b^{2} \frac{\tilde{\chi}}{2 \tilde{M}}\right) \partial_{\mu} \tilde{\phi} \partial_{\nu} \tilde{\chi} \nonumber\\
-&\frac{1}{2} \tilde{g}^{\mu \nu} \partial_{\mu} \tilde{\theta} \partial_{\nu} \tilde{\theta}+\frac{1}{2} \tilde{g}^{\mu \nu} \frac{\tilde{\theta}}{\tilde{M}}\left(b+2 a c \frac{\tilde{\theta}}{\tilde{M}}-b^{2} \frac{\tilde{\chi}}{2 \tilde{M}}\right) \partial_{\mu} \tilde{\theta} \partial_{\nu} \tilde{\chi} -\tilde{U}(\tilde{\phi}, \tilde{\theta}, \tilde{\chi})\nonumber\\
-&\bar{\tilde{\psi}} i\stackrel{\leftrightarrow}{\tilde{\slashed{\partial}}}\tilde{\psi}-y \bar{\tilde{\psi}} \tilde{\phi} \tilde{\psi}-\frac{1}{2} \mu_{\theta}^{2} \tilde{\theta}^{2}+\frac{\lambda_{\theta}}{4 !} \tilde{\theta}^{4}+\frac{3}{2} \frac{\mu_{\theta}^{4}}{\lambda_{\theta}}\cdots,
\label{genein}
\end{align}
where
\begin{align}
\tilde{U}(\tilde{\phi}, \tilde{\theta}, \tilde{\chi})=&\frac{\lambda}{4 !}\left(\tilde{\phi}^{2}-\frac{\beta}{\lambda} \tilde{\theta}^{2}\right)^{2}-\frac{1}{2} \mu^{2}\left(\tilde{\phi}^{2}-\frac{\beta}{\lambda} \tilde{\theta}^{2}\right)\left(a+b \frac{\tilde{\chi}}{\tilde{M}}+c \frac{\tilde{\chi}^{2}}{\tilde{M}^{2}}\right) \nonumber\\
+&\frac{3}{2} \frac{\mu^{4}}{\lambda}\left(a+2ab \frac{\tilde{\chi}}{\tilde{M}}+(2ac+b^2) \frac{\tilde{\chi}^{2}}{\tilde{M}^{2}}\right).
\label{potential_comb}
\end{align}
We see that there are both kinetic and mass mixings of $\tilde{\phi}$ and $\tilde{\theta}$ with $\tilde{\chi}$. However, fifth forces arising through kinetic mixings when these involve a field with non-zero mass are suppressed due to the additional momentum dependence ($\propto q^2$) that occurs for each insertion into the matrix element of the kinetic mixing operator.
As a result, the mass mixing will provide the dominant fifth force.
Moreover, since we are free to choose $\mu_{\theta}^{2} \gg \mu^{2}$, the $\tilde{\theta}$ field can be decoupled from the long-range fifth forces. We can then focus solely on the mixing between the would-be Higgs field $\tilde{\phi}$ and the conformally coupled scalar $\tilde{\chi}$, as we will do in the next subsection.
\subsection{Fifth forces}
As a concrete example, we will now specialize to the Brans-Dicke theory \cite{Brans:1961sx}, whose Jordan-frame action is
\begin{equation}\label{BDactionX}
S=\int \mathrm{d}^{4} x \sqrt{-g}\left[\frac{X}{2} R-\frac{\omega(X)}{2 X} g^{\mu \nu} \partial_{\mu} X \partial_{\nu} X\right]+S_{\rm{m}}[g_{\mu \nu}, \{\psi\}] .
\end{equation}
After performing the conformal transformation to the Einstein frame and canonically normalizing the fields (the explicit calculation can be found in Ref.~\cite{Burrage:2018dvt}), we find that the coupling function takes the form
\begin{align}
A^{2}(X(\tilde{\chi}))=&\frac{M_{\rm{Pl}}^2}{X}=\exp\left[2\frac{\tilde{\chi}}{\tilde{M}}\right],
\label{parcham}
\end{align}
wherein we have taken $\omega(X)=\omega$ to be a constant and defined
\begin{equation}
\tilde{M}^2=2(2\omega +3)\tilde{M}_{\rm{Pl}}^2.
\end{equation}
Thus, the Brans-Dicke model amounts to taking $a=1$, $b=2$ and $c=2$ in Eq.~(\ref{para}). It then follows from Eq.~(\ref{potential_comb}) that the Einstein-frame potential is
\begin{align}
\tilde{U}(\tilde{\phi}, \tilde{\theta}, \tilde{\chi})=&\frac{\lambda}{4 !}\left(\tilde{\phi}^{2}-\frac{\beta}{\lambda} \tilde{\theta}^{2}\right)^{2}-\frac{1}{2} \mu^{2}\left(\tilde{\phi}^{2}-\frac{\beta}{\lambda} \tilde{\theta}^{2}\right)\left(1+2 \frac{\tilde{\chi}}{\tilde{M}}+2 \frac{\tilde{\chi}^{2}}{\tilde{M}^{2}}\right) \nonumber\\
+&\frac{3}{2} \frac{\mu^{4}}{\lambda}\left(1+4 \frac{\tilde{\chi}}{\tilde{M}}+8 \frac{\tilde{\chi}^{2}}{\tilde{M}^{2}}\right)-\frac{1}{2} \mu_{\theta}^{2} \tilde{\theta}^{2}+\frac{\lambda_{\theta}}{4 !} \tilde{\theta}^{4}+\frac{3}{2} \frac{\mu_{\theta}^{4}}{\lambda_{\theta}}.
\end{align}
The fields acquire the vevs
\begin{align}
&v_{\tilde{\phi}}=\pm\left(\frac{6 \mu^{2}+\beta v_{\tilde{\theta}}^{2}}{\lambda}\right)^{1 / 2},& &v_{\tilde{\theta}}=\pm'\left(\frac{6 \mu_{\theta}^{2}}{\lambda_{\theta}}\right)^{1 / 2},& &v_{\chi}=0,&
\label{vevss}
\end{align}
where the $\prime$ indicates that the choice of sign for the two non-vanishing vevs is independent. Expanding around the vevs ($\tilde{\phi} \rightarrow v_{\tilde{\phi}}+\tilde{\phi}$, $\tilde{\theta} \rightarrow v_{\tilde{\theta}}+\tilde{\theta}$ and $\tilde\chi \rightarrow v_{\tilde{\chi}}+\tilde{\chi}$) and recalling that the main contribution to the fifth forces is given by the mass mixing between $\tilde{\phi}$ and $\tilde{\chi}$, the operator of interest from Eq.~\eqref{genein} is given by
\begin{equation}
\label{eq:massmixing}
\tilde{\mathcal{L}} \supset \alpha_{\rm{M}} \tilde\phi \tilde\chi=2 \mu^{2} \frac{v_{\tilde{\phi}}}{\tilde{M}} \tilde{\phi} \tilde{\chi}.
\end{equation}
We are now in the position to calculate the matrix elements for the fifth force.
We proceed by considering the scalar contributions to the M\o{}ller scattering ($e^{-}e^{-} \to e^{-}e^{-}$) for our fermion $\psi$. These arise from the series of diagrams shown in Fig.~\ref{fig:massmixing}. The external fermions couple only to the would-be Higgs field, represented by a continuous line, which then oscillates into a $\tilde \chi$ particle (dashed line) via the mass mixing term from the effective Lagrangian [Eq.~\eqref{eq:massmixing}]. The ellipsis represents the infinite series of insertions of the mass mixing.
\begin{figure}[t]
\centering
\includegraphics[scale=0.28]{phi_sigma.png}
\caption{Diagrammatic representation of the infinite series of diagrams contributing to the M\o{}ller scattering in the Einstein frame.}
\label{fig:massmixing}
\end{figure}
The resulting matrix element is given by
\begin{align}
i \mathcal{M}\left(e^{-} e^{-} \rightarrow e^{-} e^{-}\right) &\supset \bar{u} \left(\mathbf{p}_{1}, s_{1}\right)(-i y) u\left(\mathbf{p}_{3}, s_{3}\right)\nonumber \\
& \times \frac{i}{t-m_{\tilde{\phi}}^{2}}\left[\sum_{n=0}^{\infty}\left(i \alpha_{\rm{M}}\right)^{2 n}\left(\frac{i}{t-m_{\tilde{\phi}}^{2}}\right)^{n}\left(\frac{i}{t}\right)^{n}\right] \nonumber\\
& \times \bar{u}\left(\mathbf{p}_{2}, s_{2}\right)(-i y) u\left(\mathbf{p}_{4}, s_{4}\right).
\end{align}
Since we assume the scattering fermions to be distinguishable, we need only consider the $t$-channel exchange, where $t = -(p_1 - p_3)^2$ is the usual Mandelstam variable. Also, $u(\mathbf{p},s)$ and $\bar{u}(\mathbf{p},s)$ are respectively the Dirac four-spinor and its Dirac conjugate, with spin projection $s$. The resulting non-relativistic potential is given by
\begin{equation}
\tilde{V}(r) =-y^{2} \int \frac{\mathrm{d}^{3} \mathbf{Q}}{(2 \pi)^{3}} e^{i \mathbf{Q} \cdot \mathbf{x}} \frac{\mathbf{Q}^{2}}{\mathbf{Q}^{2}\left(\mathbf{Q}^{2}+m_{\tilde{\phi}}^{2}\right)-\alpha_{\rm{M}}^{2}}
\approx-\frac{y^{2}}{4 \pi}\left(1-\frac{\alpha_{\rm{M}}^{2}}{m_{\tilde{\phi}}^{4}}\right) \frac{e^{-m_{\rm{h}} r}}{r}-\frac{y^{2}}{4 \pi} \frac{\alpha_{\rm{M}}^{2}}{m_{\tilde{\phi}}^{4}} \frac{1}{r},
\end{equation}
where $m_{\rm{h}}$ is the mass of the would-be Higgs boson and the potential has been expanded to leading order in $\alpha^2_{\rm{M}}$.
Isolating the fifth-force contribution and plugging in $\alpha_{\rm{M}}$, as extracted from Eq.~\eqref{eq:massmixing}, we find
\begin{equation}
\tilde{V}_5(r) =-\frac{1}{4 \pi r} \frac{m_{e}^{2}}{\tilde M^2_{\rm{Pl}}2(2\omega+3)} \frac{4\mu^4}{m^4_{\tilde{\phi}}}, \label{solu}
\end{equation}
where we have chosen the fermions to represent electrons with mass $m_e$. Notice that, since the fifth-force mediator is massless, the potential has a similar form to the usual Newtonian gravitational potential.
To study how the different mechanisms of scale breaking affect the modification to the Yukawa potential [Eq.~\eqref{solu}], we need only recall that the mass of the $\tilde{\phi}$ field is given by
\begin{equation}
m^2_{\tilde{\phi}}=2\mu^2 + \frac{\beta v_{\tilde{\theta}}^2}{3}.
\end{equation}
\paragraph*{\bf Pure explicit scale breaking (SM toy model) $\boldsymbol{\beta\to 0}$:} In this limit, the mass of the $\tilde{\phi}$ field reduces to
\begin{equation}
m^2_{\tilde{\phi}}=2\mu^2,
\end{equation}
agreeing with the numerator of the fraction in the potential~\eqref{solu}. Hence, the modification to the Yukawa potential becomes independent of the Higgs mass, and we find
\begin{equation}
\tilde{V}_5(r) =-\frac{1}{4 \pi r} \frac{m_{e}^{2}}{\tilde M^2_{\rm{Pl}}2(2\omega+3)}.
\label{exp_eins}
\end{equation}Such a contribution to the non-relativistic potential can lead to significant deviations in the inferred gravitational force. A comprehensive review on the different tests and bounds on fifth forces can be found in Ref.~\cite{Burrage:2017qrf}, where the authors mainly focus on the chameleon model, which applies to our results given that we consider only up to second order in the field fluctuations. In particular, the most stringent constraint at Solar System scales is given by the Cassini spacecraft~\cite{Bertotti:2003rm}, setting a bound on $\omega \gg40,000$. Bounds at cosmological scales are less stringent\footnote{Even though Solar System scale tests are more constraining than cosmological ones, they are more affected by higher-order terms, making it possible to avoid the bounds through screening mechanisms.}, such as those based on Cosmic Microwave Background data from Planck~\cite{Avilez:2013dxa}, which are consistent with $\omega>692$ at the $99\%$ confidence level. Therefore, in the absence of any screening mechanism, we can see that a fine tuning of the value of $\omega$ in the case of pure explicit scale breaking is necessary to achieve an agreement with experiments.
\paragraph*{\bf Pure dynamical scale breaking (Higgs-dilaton model) $\boldsymbol{\mu\to0}$:}
In this case, the numerator of the modified Yukawa potential [Eq.~\eqref{solu}] tends to zero, whereas the denominator tends to
\begin{equation}
m^2_{\tilde{\phi}}=\frac{\beta v_{\tilde{\theta}}^2}{3}.
\end{equation}
Hence, even though classically scale-invariant theories might break the scale symmetry dynamically, the fifth forces still do not couple to the fermionic sector. It is important to remark that the vev of the would-be Higgs field $\tilde{\phi}$ field [Eq.~(\ref{vevss})] does not vanish in the limit $\mu\to0$, such that the mass-generation mechanism for the elementary fermions is preserved (with $m_e=yv_{\tilde{\phi}}$). More generally, we see that the fifth-force coupling is proportional to the ratio $\mu/m_{\tilde{\phi}}$, such that we can suppress fifth forces by combining explicit and dynamical scale-breaking mechanisms~\cite{Burrage:2018dvt}.
For this tree-level example, the transformation to the Einstein frame and the subsequent calculation of the matrix elements were easily tractable. This may not be the case, in general, however. In the next section, we will describe in detail how we can proceed directly in the Jordan frame (or Jordan-like frames), without performing the conformal transformation and subsequent field rescalings.
\section{Linearized scalar-tensor gravity}\label{thirdS}
If we wish to proceed directly in the Jordan frame, it is necessary to treat the metric dynamically and to work with linearized gravity, as we can in the weak-field limit. We will first review the linearization of Einstein gravity, which has been widely studied (see, e.g., Refs.~\cite{Fierz:1939ix,Donoghue:1995cz,Donoghue:2017pgk}). We will then generalize this procedure to actions with non-minimal couplings of Brans-Dicke type, taking care to consider the differing symmetries of this class of scalar-tensor theories.
\subsection{Standard gravity}
Gauge symmetries reflect the redundancies of a theory, so we need to break these symmetries through a gauge fixing term to fully describe the field theory we are working with. We therefore take the Einstein-Hilbert action to be of the form
\begin{equation}
S_{\rm EH}=\int \textnormal{d}^4{x} \sqrt{-g} \left[ \frac{{M_{\rm{Pl}}}^2}{2}R + \mathcal{L}_{\rm{gf}}\right],
\label{two}
\end{equation}
where we have included the gauge fixing terms via $\mathcal{L}_{\rm{gf}}$.
As we will justify later, we choose to work with the harmonic gauge, since it can be expressed at the level of the Lagrangian. The harmonic gauge is defined such that
\begin{equation}
\label{HG-constraint}
\nabla^\mu\nabla_\mu=\partial^\mu\partial_\mu,
\end{equation}
where $\nabla_\mu$ is the usual covariant derivative in General Relativity. This constraint is uniquely satisfied by setting
\begin{equation}\label{HG_condition}
\Gamma^\mu=g^{\alpha \beta}\Gamma^\mu_{\alpha \beta}=0,
\end{equation}
where $\Gamma^\mu_{\alpha \beta}$ are the Christoffel symbols. Therefore, the gauge fixing term for the harmonic gauge becomes
\begin{equation}
\mathcal{L}_{\rm{gf}}=-\frac{{M_{\rm{Pl}}}^2}{4}g_{\alpha \beta}\Gamma^\alpha\Gamma^\beta.
\label{gf}
\end{equation}
To linearize gravity, we must make small perturbations of the metric around a constant background. In this paper, we will assume this background to be flat, such that
\begin{equation}
g_{\mu \nu}=\eta_{\mu \nu} + h_{\mu \nu},
\end{equation}
where $\eta_{\mu\nu}={\rm diag}(-1,1,1,1)$ is the Minkowski metric. We then arrive at the following weak-field expansions:
\begin{subequations}
\begin{align}
\sqrt{-g}\approx&\, 1 + \frac{1}{2}\eta_{\mu \nu}h^{\mu \nu}, \\
R_{\mu \nu}^{(1)}=&\,\frac{1}{2}\left( \partial^\rho \partial_\mu h_{\nu \rho} + \partial^\rho \partial_\nu h_{\mu \rho} -\Box h_{\mu \nu} -\partial_\mu \partial_\nu h\right), \label{E1}\\
R_{\mu \nu}^{(2)}=&\,\frac{1}{2}h^{\rho \sigma} \partial_\mu \partial_\nu h_{\rho \sigma} -h^{\rho \sigma} \partial_\mu {\partial_{(\nu}} h_{\rho) \sigma} +\frac{1}{4}\partial_\mu h^{\rho \sigma} \partial_\nu h_{\rho \sigma}\nonumber \\
&\,+\partial^\sigma h^{\rho}_{\nu} {\partial_{[\sigma}} h_{\rho] \mu} + \frac{1}{2}\partial_\sigma(h^{\sigma \rho}\partial_\rho h_{\mu \nu}) - \frac{1}{4}\partial^\rho h \partial_{\rho} h_{\mu \nu} \nonumber\\
&\,-(\partial_\sigma h^{\sigma \rho} - \frac{1}{2}\partial^\rho h){\partial_{(\mu}} h_{\nu) \rho}, \\
\Gamma^{\mu(1)}_{\alpha \beta} =&\, \frac{1}{2}\eta^{\mu \lambda}(\partial_\alpha h_{\lambda \beta} + \partial_\beta h_{\alpha \lambda} - \partial_\lambda h_{\alpha \beta}),\label{E2}
\end{align}
\end{subequations}
where the exponent in parenthesis shows the order in the metric fluctuations $h$. With these ingredients, we can then determine the expansion of the terms in Eq.~(\ref{two}) up to second order:
\begin{subequations}
\begin{align}
\sqrt{-g}\frac{{M_{\rm{Pl}}}^2}{2}R=&\frac{{M_{\rm{Pl}}}^2}{2}\left(1 + \frac{1}{2}\eta_{\mu\nu}h^{\mu\nu}\right)\left[R_{\mu\nu}(\eta^{\mu\nu} -h^{\mu\nu})\right]\nonumber\\
=&\frac{{M_{\rm{Pl}}}^2}{2}\left(R^{(1)} - R_{\mu\nu}^{(1)}h^{\mu \nu} + \frac{1}{2}R^{(1)}\eta_{\mu \nu}h^{\mu \nu}+R^{(2)}\right), \\
\sqrt{-g}\mathcal{L}_{\rm{gf}}=&-\sqrt{-g}\frac{{M_{\rm{Pl}}}^2}{4}g_{\alpha \beta}\Gamma^\alpha\Gamma^\beta = -\frac{{M_{\rm{Pl}}}^2}{4}\eta_{\alpha \beta}\eta^{\mu\nu}\Gamma^{\alpha(1)}_{\mu \nu}\eta^{\sigma \rho}\Gamma^{\beta(1)}_{\sigma \rho}, \\
\sqrt{-g}\mathcal{L}_{\rm{m}}(g_{\mu \nu})=&\frac{1}{2}h^{\mu \nu}T_{\mu \nu} + \mathcal{L}_{\rm{m}}(\eta_{\mu \nu})\label{R2},
\end{align}
\end{subequations}
where we have re-introduced the matter sector, giving rise to the contributions from its energy-momentum tensor $T_{\mu\nu}$.
Making use of the results in Eqs.~(\ref{E1}) to (\ref{E2}) and integrating by parts, we obtain the following expression for the Lagrangian up to second order in $h_{\mu \nu}$:
\begin{equation}
\mathcal{L}=\frac{M_{\rm{Pl}}^2}{4}\left[ \frac{1}{4}\partial_\mu h \partial^\mu h -\frac{1}{2}\partial_\rho h_{\mu \nu}\partial^\rho h^{\mu \nu}\right] + \frac{M_{\rm{Pl}}^2}{2}\left[\partial^\mu\partial^\nu h_{\mu \nu} - \Box h\right] +\frac{1}{2}h^{\mu\nu}T_{\mu\nu} + \mathcal{L}_{\rm{m}}(\eta_{\mu\nu}).
\end{equation}
The first term corresponds to the kinetic energy of the graviton and can be canonically normalized by rescaling it such that $h_{\mu \nu} \to 2h_{\mu \nu}/M_{\rm{Pl}}$. We draw attention to the second term, containing second derivatives of the graviton field, which can be removed on integrating by parts and setting the boundary terms to zero. In contrast, the analogous term for the scalar-tensor theory will not vanish, and it will be the source of the fifth forces in the Jordan frame, as we will show in the next subsection.
\subsection{Jordan frame}
We now repeat the linearization for Brans-Dicke-type scalar-tensor theories. The action in this case will have a non-minimal coupling to the Ricci scalar, such that
\begin{equation}\label{Action-BD-GF}
S=\int \textnormal{d}^4{x} \sqrt{-g} \left[\frac{F(X)}{2}R + \mathcal{L}'_{\rm{gf}} - \frac{Z(X)}{2}\partial_\mu X \partial^\mu X - U(X)\right] + S_{\rm{m}}[ g_{\mu \nu},\{\psi\}],
\end{equation}
where, once again, we have included a gauge fixing term $\mathcal{L}'_{\rm{gf}}$. Most importantly, this term will not be the same as in standard gravity [cf.~Eq.~(\ref{gf})], since the scalar-tensor action does not have the same symmetries as the Einstein-Hilbert one, given that the non-minimal coupling affects the (non-null) geodesics. This requires us to upgrade the usual harmonic gauge condition.
The first change is to replace the constant prefactor ${M_{\rm{Pl}}}^2$ by the non-minimal coupling, such that
\begin{equation}\label{gf_mod1}
-\frac{{M_{\rm{Pl}}}^2}{4}g_{\alpha \beta}\Gamma^\alpha\Gamma^\beta\to -\frac{F(X)}{4}g_{\alpha \beta}\Gamma^\alpha\Gamma^\beta.
\end{equation}
Second, we want to make sure that the gauge condition reflects the symmetries of the action, which are modified relative to the Einstein-Hilbert case\footnote{We might be tempted to work with only the first modification in Eq.~\eqref{gf_mod1} for the updated gauge fixing term. While doing so does not affect the non-relativistic limit for the fifth forces, this choice significantly complicates the calculations and leads to apparent deviations from the corresponding Einstein-frame results in the relativistic limit.}. Thus, the scalar-tensor equivalent of the harmonic condition must be defined via
\begin{equation}\label{DDjf}
D^\mu D_\mu=\partial^\mu \partial_\mu,
\end{equation}
where $D^\mu$ is the covariant derivative as constructed with respect to the Jordan-frame action. This is the focus of the next subsection.
\subsubsection{Updating the covariant derivative}
We know that General Relativity must be symmetric under diffeomorphisms. Thus, we will study which conditions the covariant derivative must satisfy in order to leave the action invariant under transformations of the form
\begin{equation}\label{isometry}
g^{\mu \nu}\to g^{\mu \nu} + D^\mu \xi^\nu +D^\nu \xi^\mu,
\end{equation}
where $D_\mu$ is generically defined when acting on a four vector $Y^\nu$ as
\begin{equation}
D^\mu Y^\nu= \partial^\mu Y^\nu + g^{\mu\rho}\Gamma_{\rho \sigma}^\nu Y^\sigma+ C^\mu Y^\nu,
\label{newcov}
\end{equation}
and $C^\mu$ contains the possible contributions of $X$ to the isometries of the spacetime.
To find whether the action is symmetric under diffeomorphisms, we first need to vary the gravitational part of the action [Eq.~\eqref{Action-BD-GF}] with respect to the metric, leading to
\begin{equation}
\delta S= -\int \textnormal{d}^4{x} \sqrt{-g} \frac{F(X)}{2} G_{\mu \nu} \delta g^{\mu\nu},
\end{equation}
where $G_{\mu \nu}$ is the Einstein tensor and we take the small variation of the metric $\delta g^{\mu\nu}$ to be given by the isometry from Eq.~\eqref{isometry}. Since we expect this transformation to be a symmetry of the action, $\delta S$ should vanish, i.e.,
\begin{equation}
\delta S= -\int \textnormal{d}^4{x} \sqrt{-g} F(X) G_{\mu \nu} D^{\left(\mu\right.} \xi^{\left. \nu\right)}=0,
\end{equation}
where the parentheses indicate the symmetrization of indices, such that
\begin{equation}
D^{\left(\mu\right.} \xi^{\left. \nu\right)}=\frac{1}{2}(D^\mu\xi^\nu +D^\nu \xi^\mu).
\end{equation}
Inserting the covariant derivative from Eq.~\eqref{newcov} and integrating by parts, we find
\begin{equation}
\delta S= \int \textnormal{d}^4{x} \sqrt{-g} [\nabla^\nu (F(X)G_{\mu \nu}) - F(X) G_{\mu\nu}C^\nu ]\xi^\mu=0,
\end{equation}
where
\begin{equation}
\nabla_\mu Y^\nu= \partial_\mu Y^\nu + \Gamma_{\mu \sigma}^\nu Y^\sigma
\end{equation}
is the covariant derivative of standard gravity. Since both $\nabla^\nu$ and $G_{\mu\nu}$ only depend on the metric, they will still satisfy the Bianchi identity, leading to
\begin{equation}
\nabla^\nu G_{\mu \nu}=0.
\end{equation}
However, as $\nabla^\nu$ also acts on $F(X)$, the perturbed action will take the following form
\begin{equation}
\delta S= \int \textnormal{d}^4x \sqrt{-g}[\nabla^\nu F(X) - F(X) C^\nu]G_{\mu\nu}\xi^\mu=0.
\end{equation}
Note that, for a constant $F(X)$, this equation would vanish, showing that the geodesics of the Einstein-Hilbert action are in agreement with those from standard gravity, encoded in $\nabla_\mu$. However, to have a vanishing perturbed action in the Jordan frame, we need
\begin{equation}
C^\nu=\frac{\partial^\nu F(X)}{F(X)}.
\end{equation}
Hence, we see that the non-minimal coupling of $X$ alters the isometries of the spacetime.
The new contribution $C^\mu$ to the covariant derivative can be also expressed as a modification of the Christoffel symbols [i.e., as a $(1,2)$-form]. While this full expression is not needed to update the harmonic gauge condition, it is useful for explicit calculations of geodesics. While the derivation of the $(1,2)$-form can be done by studying the symmetries of the Jordan-frame action, one can instead (since we know that it exists) use the Einstein frame as a shortcut, where standard gravity is recovered. Since we already know the symmetries of the Einstein-frame action, we can fix the gauge in that frame and then undo the coordinate transformation. In this way, we will get the corresponding set of symmetries for the Jordan frame.
For instance, Weyl transforming the covariant derivative of the vector $Y_\nu$ from the Einstein to the Jordan frame, we find that
\begin{equation}
\tilde\nabla_\mu Y_\nu \to D_\mu Y_\nu=\partial_\mu Y_\nu - \Gamma^\rho _{\mu \nu} Y_\rho + C^\rho_{\mu \nu} Y_\rho,
\end{equation}
where
\begin{equation}
C_{\mu \nu}^\sigma =-\frac{1}{2F(X)}\left[2\delta^\sigma_{\left(\mu\right.}\partial_{\nu\left.\right)}F(X) -g_{\mu \nu}\partial^\sigma F(X)\right].
\end{equation}
This is the correction to the Christoffel symbols coming from the breaking of the weak equivalence principle. Moreover, taking the trace of this term, we find that it is consistent with our previous result:
\begin{equation}
g^{\mu\nu} C_{\mu \nu}^\sigma=\frac{\partial^\sigma F(X)}{F(X)}=C^\sigma.
\label{traceC}
\end{equation}
Thus, in contrast to standard gravity, where the geometry of spacetime is exclusively modified through the Christoffel symbols, the non-minimal coupling of the field $X$ alters the isometries, accounting for the breaking of the weak equivalence principle.
\subsubsection{Gauge fixing and obtaining the generic linearized Jordan-frame Lagrangian}
Having derived the scalar-tensor covariant derivative, we can now continue upgrading the harmonic gauge from Eq.~\eqref{gf_mod1}. Returning to Eq.~\eqref{DDjf}, the harmonic condition imposes
\begin{equation}
D^\mu D_\mu=g^{\mu\nu}\left[\partial_\nu \partial_\mu-\Gamma ^\sigma_{\mu \nu}\partial_\sigma + C_{\mu \nu}^\sigma \partial_\sigma \right]=\partial^{\mu}\partial_{\mu}.
\end{equation}
Thus, to satisfy the scalar-tensor harmonic gauge, it follows that
\begin{equation}
\Gamma^\mu - \frac{\partial^\mu F(X)}{F(X)}=0.
\end{equation}
Following the same logic as in Eq.~\eqref{HG-constraint}, we now replace each $\Gamma^\alpha$ from Eq.~\eqref{gf_mod1} by the new constraint. With this, we introduce the {\it scalar-harmonic gauge} via
\begin{equation}
\mathcal{L}'_{\rm{gf}}=-\frac{F(X)}{4}g_{\mu\nu}\left[ \Gamma^\mu -\frac{\partial^\mu F(X)}{F(X)}\right]\left[\Gamma ^\nu -\frac{\partial^\nu F(X)}{F(X)}\right].
\label{gfjf1}
\end{equation}
It is important to point out that this is not the first time this gauge has been used. For instance, it is mentioned by Fuji and Maeda~\cite{FM} and used in a number of papers where the authors employ the background covariant DeWitt condition \cite{DeWitt:1964mxt,Barvinsky:1985an,Steinwachs:2011zs}. However, here, we have presented a different way to find the gauge symmetries of the Jordan frame. Moreover, unlike in previous papers, we define the gauge fixing term using the complete metric, which allows us to perturb consistently to higher orders in the fluctuations.
Note also that we have been able to define this Jordan-frame gauge fixing term because we already had the standard gravity condition defined at the Lagrangian level [Eq.~\eqref{gf}]. This might be problematic when working with other gauges that are defined only after gravity has been linearized. For instance, the Newtonian gauge does not have a full metric expression; it is defined by setting certain modes of the graviton to zero. Therefore, we cannot apply our gauging method in this case, which could lead to misinterpreted results when going to higher orders in the perturbations, since additional terms should appear because of the new symmetries of the action.
For the gauge fixing term in Eq.~\eqref{gfjf1}, it proves convenient to divide it into three parts:
\begin{itemize}
\item [(i)] a contribution to the graviton kinetic energy:
\begin{equation}
\mathcal{L}_{{\rm{G}}}=-\frac{F(X)}{4}g_{\mu\nu}\Gamma^\mu\Gamma^\nu;
\end{equation}
\item [(ii)] a contribution to the kinetic mixing:
\begin{equation}
\mathcal{L}_{\rm{KM}}=\frac{1}{2}\Gamma^\mu\partial_\mu F(X) ;
\end{equation}
\item [(iii)] a contribution to the $X$ kinetic energy:
\begin{equation}
\mathcal{L}_{\rm{SF}}=-\frac{1}{4F(X)}g_{\mu \nu}\partial^\mu F(X) \partial^\nu F(X);
\end{equation}
\end{itemize}
such that
\begin{equation}
\mathcal{L}'_{\rm{gf}}=\mathcal{L}_{\rm{G}}+\mathcal{L}_{\rm{KM}}+\mathcal{L}_{\rm{SF}}.
\end{equation}
Inserting this gauge fixing term into Eq.~\eqref{Action-BD-GF} and linearizing the action similarly to the standard gravity case, we obtain the following expansion up to second order for the associated Lagrangian:
\begin{align}
\mathcal{L}=&\frac{F(X)}{4}\left[ \frac{1}{4}\partial_\mu h \partial^\mu h -\frac{1}{2}\partial_\rho h_{\mu \nu}\partial^\rho h^{\mu \nu}\right]
+ \frac{F(X)}{2}\left[\partial^\mu\partial^\nu h_{\mu \nu} - \Box h\right]\nonumber\\& + \frac{F'(X)}{2}\left[\partial^\lambda h_{\lambda\mu} - \frac{1}{2}\partial_{\mu}h\right]\partial^\mu X
-\frac{1}{2}\left[Z(X)+\frac{F'(X)}{2F(X)}^2\right]g^{\mu\nu}\partial_\mu X\partial_\nu X - U(X)\nonumber\\& + \frac{1}{2}h^{\mu\nu}T_{\mu\nu}+ \mathcal{L}_{\rm{m}}(\eta_{\mu\nu}).
\end{align}
The would-be total derivative term in the case of standard gravity, corresponding to the terms involving second derivatives of the metric fluctuations, no longer vanishes on integration by parts. Moreover, it creates an additional kinetic mixing term that, once added to $\mathcal{L}_{\rm KM}$, yields
\begin{align}
\mathcal{L}=&\frac{F(X)}{4}\left[ \frac{1}{4}\partial_\mu h \partial^\mu h -\frac{1}{2}\partial_\rho h_{\mu \nu}\partial^\rho h^{\mu \nu}\right]
+ \frac{F'(X)}{4}\eta^{\mu\nu}\partial_\mu h \partial_\nu X \nonumber\\
-&\frac{1}{2}\left[Z(X)+\frac{F'(X)}{2F(X)}^2\right]\eta^{\mu\nu}\partial_\mu X\partial_\nu X - U(X) \label{4}
+ \frac{1}{2}h^{\mu\nu}T_{\mu\nu}+ \mathcal{L}_{\rm{m}}\{\eta_{\mu\nu}\}.
\end{align}
This is the generic effective Lagrangian of the linearized theory in the Jordan frame. When the function $F(X)$ is constant, the expression reduces to that of General Relativity, as expected. Thus, the main difference is the new kinetic mixing term, and it is this which can lead to fifth forces between matter fields.
In contrast to the Einstein frame, the main contribution to the fifth-force coupling, as analysed in the Jordan frame, is via the kinetic mixing and not a mass mixing. This comes from the fact that the graviton propagator ($\propto 1/q^2$) cancels the momentum dependence of the mixing vertex ($\propto q^2$) in every oscillation between the field $X$ and the graviton, such that, unlike the case of a massive field, there is no additional momentum suppression in the non-relativistic limit. In the next section, we will calculate explicitly the fifth forces that arise through this kinetic mixing in the Jordan frame.
\section{Fifth forces in the Jordan frame}\label{fourth}
Having derived the general expression for the Lagrangian up to second order in the metric fluctuations for the Brans-Dicke-type scalar-tensor theories in Eq.~\eqref{4}, we now turn our attention to specific models to see how the fifth forces arise from the kinetic mixings between the graviton and the non-minimally coupled field $X$. We will show that the results agree with those obtained previously in the Einstein frame.
The action corresponding to Eq.~\eqref{BDactionX}, with a matter sector given by Eqs.~\eqref{sm} and \eqref{comb_pot}, is
\begin{align} \label{BDpureparam}
S=\int \textnormal{d}^4{x} \sqrt{-g} \, &\bigg{[}\frac{X}{2}R +\mathcal{L}'_{\rm{gf}} -\frac{\omega(X)}{2X}g^{\mu \nu}\partial_{\mu} X \partial_{\nu} X - \frac{1}{2}g^{\mu \nu}\partial_{\mu} \Phi \partial_{\nu} \Phi \nonumber\\
-& \frac{1}{2} g^{\mu \nu} \partial_{\mu} \Theta \partial_{\nu} \Theta+\frac{1}{2} \mu_{\theta}^{2} \frac{X}{\tilde{M}_{\rm{Pl}}} \Theta^{2}-\frac{\lambda_{\theta}}{4 !} \Theta^{4}-\frac{3}{2} \frac{\mu_{\theta}^{4}}{\lambda_{\theta}} \frac{X^2}{\tilde{M}_{\rm{Pl}}^2}\, \nonumber\\
-&\bar{\psi}i{\stackrel{\leftrightarrow}{\slashed{\partial}}}\psi - y\bar{\psi}\Phi\psi -U(\Phi,\Theta)\bigg{]},
\end{align}
where $U$ is given by Eq.~\eqref{comb_pot}. Note that we have already substituted in Eq.~\eqref{BDpureparam} for the Brans-Dicke model functions~\cite{Brans:1961sx}, with
\begin{equation}
F(X)=X,\qquad Z(X)=\frac{\omega(X)}{X}.
\label{5}
\end{equation}
We assume that $\omega(X)$ is a slowly varying function that can be taken effectively constant.
We now proceed to linearize the Lagrangian, making use of the results from the preceding section. Substituting Eq.~\eqref{5} into Eq.~\eqref{4}, we thus find
\begin{align}
\mathcal{L}=&\frac{X}{4}\left[ \frac{1}{4}\partial_\mu h \partial^\mu h -\frac{1}{2}\partial_\rho h_{\mu \nu}\partial^\rho h^{\mu \nu}\right] - \frac{1}{2}\frac{2\omega+1}{2X}\eta^{\mu\nu}\partial_\mu X\partial_\nu X- \frac{1}{2}\eta^{\mu\nu}\partial_\mu \Phi\partial_\nu \Phi \nonumber \\
+& \frac{1}{4}\eta^{\mu\nu}\partial_\mu h \partial_\nu X -U(\Phi,\Theta)
- \frac{1}{2}\eta^{\mu\nu}\partial_\mu \Theta\partial_\nu\Theta + \frac{1}{2} \mu_{\theta}^{2} \frac{X}{\tilde{M}_{\rm{Pl}}} \Theta^{2}-\frac{\lambda_{\theta}}{4 !} \Theta^{4}-\frac{3}{2} \frac{\mu_{\theta}^{4}}{\lambda_{\theta}} \frac{X^2}{\tilde{M}_{\rm{Pl}}^2}\nonumber\\
+& \frac{1}{2}h^{\mu\nu}T_{\mu\nu}-\bar{\psi}i{\stackrel{\leftrightarrow}{\slashed{\partial}}}\psi - y\bar{\psi}\Phi\psi+ \cdots.
\end{align}
where the ellipsis indicates terms higher than second order in $h_{\mu \nu}$.
When linearizing around the background solution for $X$, namely $v_X$, it is possible to canonically normalize all the fields, including the graviton. To do so, we assume the background value of $X$ to vary very slowly compared to the other fields. This is a reasonable assumption, given that the tightest constraints from the analysis of the Moon's orbit~\cite{Muller:2007zzb} set the Planck mass to be almost constant at late times, a result that can be obtained by considering the impact of the late-time Hubble friction on the evolution of $X$. Defining
\begin{equation}
X=\frac{\chi^2}{2(2\omega + 1)}
\label{mplsb}
\end{equation}
and making the replacement $h_{\mu\nu}\to2h_{\mu\nu}/M_{\rm{Pl}}$,
where
\begin{equation}\label{Eq:M_pl}
M^2_{\rm{Pl}}=\frac{v_\chi^2}{2(2\omega +1)},
\end{equation}
the Lagrangian then takes the form
\begin{align}
\mathcal{L}=& \frac{1}{4}\partial_\mu h \partial^\mu h -\frac{1}{2}\partial_\rho h_{\mu \nu}\partial^\rho h^{\mu \nu} - \frac{1}{2}\eta^{\mu\nu}\partial_\mu \chi\partial_\nu \chi - \frac{1}{2}\eta^{\mu\nu}\partial_\mu \Phi\partial_\nu \Phi+\frac{1}{\sqrt{2(2\omega + 1)}}\eta^{\mu\nu}\partial_\mu h \partial_\nu \chi \nonumber \\
-& \frac{1}{2}\eta^{\mu\nu}\partial_\mu \Theta\partial_\nu \Theta+U(\Phi, \Theta)+\frac{1}{2} \mu_{\theta}^{2} \frac{\chi^2}{\tilde{M}'^2} \Theta^{2}-\frac{\lambda_{\theta}}{4 !} \Theta^{4}-\frac{3}{2} \frac{\mu_{\theta}^{4}}{\lambda_{\theta}} \frac{\chi^4}{\tilde{M}'^4}\label{laggrav}\nonumber\\
+& \frac{1}{M_{\rm{Pl}}}h^{\mu\nu}T_{\mu\nu}-\bar{\psi}i{\stackrel{\leftrightarrow}{\slashed{\partial}}}\psi - y\bar{\psi}\Phi\psi+ \cdots,
\end{align}
where
\begin{equation}
\tilde{M}'^2=\tilde{M}_{\rm{Pl}}^2 2({2\omega +1}).
\end{equation}
Note that $M^2_{\rm{Pl}}$ is the effective gravitational coupling in the Jordan frame, while $\tilde{M}_{\rm{Pl}}^2$ is the one defined in the Einstein frame. Even though they belong to different frames, the conformal transformations appear to have forced them into the same Lagrangian. However, as we will see, this is not the case as the $\tilde{M}_{\rm{Pl}}^2$ dependence cancels out, leading to a final result in terms of $M^2_{\rm{Pl}}$, as expected. The next step is to diagonalize the mass matrix so that we can isolate the expected massless mode that can mediate any long-range fifth force.
\subsection{Diagonalization}
By considering the mass mixing terms from Eq.~(\ref{laggrav}), we can construct the following mass matrix:
\begin{equation}
m^2=
\begin{pmatrix}
m^2_\Phi & -Am_\Phi & 0\\
-Am_\Phi & m^2_\Theta & -Bm_\chi \\
0 & -Bm_\chi & m^2_\chi
\end{pmatrix},
\label{massmat}
\end{equation}
where
\begin{subequations}
\begin{gather}
m^2_\Phi=\frac{\lambda v_\Phi^2}{3},\qquad m_\Theta^2 =\frac{\beta^2}{2\lambda}v^2_\Theta + \frac{\mu^2_\theta}{ \tilde{M}'^2}v_\chi^2, \qquad m_\chi^2=\frac{\mu^2_\theta}{ \tilde{M}'^2}v_\Theta^2,\\
v^2_\Phi=\frac{6\mu^2 +\beta{v_\Theta}^2}{\lambda}, \qquad v_\Theta^2 =\frac{3\mu^2_\theta }{\lambda_\theta \tilde{M}'^2}v_\chi^2, \qquad v_\chi^2=\frac{\lambda_\theta \tilde{M}'^2}{6\mu^2_\theta}v_\Theta^2,\\
A^2=\frac{\beta^2}{2\lambda}v^2_\Theta,\qquad B^2= \frac{\mu^2_\theta}{ \tilde{M}'^2}v_\chi^2.
\label{AB}
\end{gather}
\end{subequations}
Diagonalizing this mass matrix, we obtain a new set of fields $ \phi$, $ \theta$ and $\sigma$, whose squared mass eigenvalues are
\begin{equation}
m^2_{{\phi},{\theta}}=\frac{m^2_\Phi+m^2_\Theta+m^2_\chi\pm\sqrt{(-m^2_\Phi -A^2 +B^2 +m^2_\chi)^2 +4A^2B^2}}{2},\qquad \label{mpt}
m_\sigma^2=0,
\end{equation}
wherein we see the anticipated massless mode $\sigma$.
To determine how the original fields depend on these three modes, we need to find the eigenvectors of the mass matrix~(\ref{massmat}). After some algebra, we can show that
\begin{align}
&\phi={N_{\phi}}
\begin{pmatrix}
\frac{\beta v_\Theta v_\Phi}{3(m^2_\Phi -C +D)} \\
1 \\
\frac{\mu^2_\theta v_\Theta v_\chi}{\tilde{M}'^2(m^2_\chi -C +D)}
\end{pmatrix}&
&\theta={N_{\theta}}
\begin{pmatrix}
\frac{\beta v_\Theta v_\Phi}{3(m^2_\Phi -C -D)} \\
1 \\
\frac{\mu^2_\theta v_\Theta v_\chi}{\tilde{M}'^2(m^2_\chi -C-D)}
\end{pmatrix}&
& \sigma={N_\sigma}
\begin{pmatrix}
\frac{\beta v_\Theta}{\lambda v_\Phi} \\
1 \\
\frac{v_\chi}{v_\Theta}
\end{pmatrix},&
\label{evec}
\end{align}
where $N_{\phi}$, $N_{\theta}$ and $N_\sigma$ are normalization factors, and
\begin{align}
C=\frac{m^2_\Phi+m^2_\Theta+m^2_\chi}{2},\qquad D=\frac{\sqrt{(-m^2_\Phi -A^2 +B^2 +m^2_\chi)^2 +4A^2B^2}}{2}.
\end{align}
For the fifth-force contribution to the M\o{}ller scattering, we need only expand the $\chi$ and $\Phi$ fields in terms of the massless eigenmode, since they are the only ones coupling to the fermion and graviton directly.
The relevant expansions take the forms
\begin{equation}
\chi= \frac{a}{N_{{\phi}}}{\phi}+\frac{b}{N_{{\theta}}}{\theta}+\frac{c}{N_\sigma}\sigma,\qquad
\Phi=\frac{a'}{N_{{\phi}}}{\phi}+\frac{b'}{N_{{\theta}}}{\theta}+\frac{c'}{N_\sigma}\sigma, \label{eq:field_exp}
\end{equation}
where $\{a,b,...\}$ are constant coefficients, which should not be confused with those appearing in Eq.~\eqref{para}. Since we are interested in the massless mode, we only need to determine $c$ and $c'$, and, after some algebra, we have
\begin{subequations}
\begin{align}
c&=\frac{\theta_3-\phi_3}{(\theta_1 -\phi_1)(\sigma_3 - \phi_3) + (\sigma_1 -\phi_1)(\phi_3-\theta_3)},
\label{c}\\
c'&=-\frac{\theta_1-\phi_1}{(\theta_1 -\phi_1)(\sigma_3 - \phi_3) + (\sigma_1 -\phi_1)(\phi_3-\theta_3)},
\label{cp}
\end{align}
\end{subequations}
where the subscripts refer to each component of the eigenvectors defined in Eq.~\eqref{evec}, without the corresponding normalizing factor $N_{\{\phi,\theta,\sigma\}}$.
We are now in a position to derive an expression for the effective Lagrangian in terms of the massless mode and subsequently calculate its contribution to the M\o{}ller scattering.
\subsection{Non-relativistic fifth-force potential}
After diagonalizing the mass terms in the Lagrangian, we have found all the different ways that the long-range fifth forces can couple to the matter fields. From the linearization of scalar-tensor gravity, the fifth forces arise through the kinetic mixing between the graviton and the $\sigma$ field. In addition, after diagonalizing the mass terms, a new coupling between the massless mode and the fermion field appears as a result of their Yukawa interaction with the $\Phi$ field. Thus, there are four distinct Feynman diagrams contributing to the M\o{}ller scattering, and these are shown in Fig.~\ref{fig:sigmakinetic2}.
The terms in the Lagrangian relevant to the fifth force are as follows:
\begin{align}
\mathcal{L}_{\textrm{JF}}=& \frac{1}{4}\partial_\mu h \partial^\mu h-\frac{1}{2}\partial_\rho h_{\mu \nu}\partial^\rho h^{\mu \nu}-\frac{1}{2}\eta^{\mu\nu}\partial_\mu \sigma\partial_\nu \sigma \nonumber \\
+& \frac{cN^{-1}_\sigma}{\sqrt{2(2\omega + 1)}}\eta^{\mu\nu}\partial_\mu h \partial_\nu \sigma - yc'N^{-1}_\sigma \bar{\psi}\sigma\psi
+ \frac{1}{M_{\rm{Pl}}}h^{\mu\nu}T_{\mu\nu} + \mathcal{L}_{\rm{m}}(\eta_{\mu\nu}),\label{SHLag}
\end{align}
and the resulting Feynman rules are summarized in Fig.~\ref{fig:FeynExp}.
\begin{figure}
\centering
\includegraphics[scale=0.5]{grav_sigma.png}
\caption{The diagrams that contribute to the M\o{}ller scattering in the Jordan frame. }
\label{fig:sigmakinetic2}
\end{figure}
Since the structure of all the diagrams is very similar, we will describe only the contribution from Fig.~\ref{fig:sigmakinetic2}(a) in detail. The matrix element for this process is
\begin{figure}
\fbox{
\begin{tabular}{c c}
\begin{minipage}[b]{0.4\linewidth}
$\bullet$ Graviton propagator \cite{Donoghue:2017pgk}
\begin{equation}
\begin{gathered}
\includegraphics[scale=0.28]{grav_prop.png}
\end{gathered}
= \frac{i P^{\mu \nu \sigma \rho}}{k^2}
\nonumber
\end{equation}
\vspace{-7mm}
\begin{equation}
P^{\mu \nu \sigma \rho}=\frac{1}{2}\left(\eta^{\mu \sigma}\eta^{\nu \rho} + \eta^{\nu \sigma}\eta^{\mu \rho} -\eta^{\mu \nu}\eta^{\rho \sigma}\right)
\label{propp}
\nonumber
\end{equation}\\
\vspace{11mm}
\end{minipage} &
\begin{minipage}[b]{0.4\linewidth}
\vspace{8mm}
$\bullet$ $\sigma$ field propagator
\begin{equation}
\begin{gathered}
\includegraphics[scale=0.28]{sigma_prop.png}
\end{gathered}
=-\frac{i}{k^2}
\nonumber
\end{equation}
\vspace{3mm}
$\bullet$ Kinetic mixing
\begin{equation}
\begin{gathered}
\includegraphics[scale=0.22]{kinmixing.png}
\end{gathered}
=i \frac{\eta_{\mu \nu} k^2}{\sqrt{2(2\omega+1)}}
\nonumber
\end{equation}\\
\end{minipage} \\
\begin{minipage}{0.4\linewidth}
\vspace{9mm}
$\bullet$ Gravitational interaction \cite{Olyaei:2018asy}
\begin{equation}
\begin{gathered}
\includegraphics[scale=0.21]{grav_inter.png}
\end{gathered}
= i \frac{\eta^{\mu\nu}\tau_{\mu \nu}}{M_{\rm{Pl}}}
\nonumber
\end{equation}
\vspace{-5mm}
\begin{equation}
\tau_{\mu \nu}= \frac{1}{4}\left[(p+q)_\mu \gamma_\nu + \gamma_\mu (p+q)_\nu - 2\eta_{\mu\nu}\left(\slashed{q}+\slashed{p} -2m_e\right)\right]
\nonumber
\end{equation}
\vspace{1mm}
\end{minipage} &
\begin{minipage}{0.4\linewidth}
$\bullet$ Fermion-fermion-$\sigma$ interaction
\vspace{-2mm}
\begin{equation}
\begin{gathered}
\includegraphics[scale=0.21]{sigma_inter.png}
\end{gathered}
= i \frac{yc'v_\chi}{cM_{\rm{Pl}}\sqrt{2(2\omega+3)}}\nonumber
\end{equation}
\vspace{5mm}
\end{minipage}\\
\end{tabular}\qquad\quad}
\caption{Feynman rules for the Lagrangian [Eq.~\eqref{SHLag}] with an explicitly broken scale symmetry, where $\gamma_{\mu}$ are the gamma matrices. To a good approximation, we can take $cN^{-1}_\sigma\approx1$, since $\tilde{M}\gg1$. \label{fig:FeynExp}}
\end{figure}
\begin{align}
i\mathcal{M}_{\rm{\bf{(a)}}}&= \bar{u}(\mathbf{p}_1,s_1)\left( i\frac{\tau_{\mu \nu}}{{M}_{\rm{Pl}}}\right)u(\mathbf{p}_3,s_3)
\left(i\frac{P^{\mu \nu a b}}{t}\right)\left(i \eta_{a b}\alpha_{\rm{K}} t\right)\left(\frac{i}{t}\right) \nonumber\\
& \times \left[ \sum_{n=0}^{\infty} (i\alpha_{\rm{K}} t)^n\left(\frac{i\eta_{c d}P^{c d e f}\eta_{e f}}{t}\right)^n(i\alpha_{\rm{K}} t)^n \left(\frac{i}{t}\right)^n\right] \nonumber\\
& \times (i\alpha_{\rm{K}} \eta_{g h}t)\left(\frac{iP^{g h \sigma \rho}}{t}\right) \bar{u}(\mathbf{p}_2,s_2)\left(i\frac{\tau_{\sigma \rho}}{{M}_{\rm{Pl}}}\right)u(\mathbf{p}_4,s_4),\label{amplitude_JF}
\end{align}
where, as before, $t=-(p_1 -p_3)^2$, $u(\mathbf{p},s)$ and $\bar{u}(\mathbf{p},s)$ are respectively the Dirac four-spinor and its Dirac conjugate, with spin projection $s$. Note that, for clarity, we have isolated each vertex and propagator with parentheses. For convenience, we have also defined the parameter
\begin{equation}
\alpha_{\rm{K}}=\frac{1}{\sqrt{2(2\omega+1})}.
\end{equation}
Equation~(\ref{amplitude_JF}) can be simplified by making use of the following identities for $P^{\mu \nu \sigma \rho}$:
\begin{align}
\eta_{\mu \nu}P^{\mu \nu \sigma \rho}=-\eta^{\sigma \rho},\qquad \eta_{\mu \nu}P^{\mu \nu \sigma \rho}\eta_{\sigma \rho}=-4,
\label{id}
\end{align}
and we find that we only have vertices involving the trace of $\tau_{\mu\nu}$, as we would have expected from Eq.~\eqref{eq:actionexpansion}.
Working in the non-relativistic limit and choosing the fermions to represent electrons with mass $m_e$, such that $p^{\mu}\sim q^{\mu}\approx (m_e,\vec{0})$, the spinors satisfy
\begin{equation}
\bar{u}(\mathbf{p},s)\gamma_\mu u(\mathbf{q},s')=2m_e \delta_{\mu 0}\delta_{s s'},
\end{equation}
in which case, using the expression for $\tau=\eta^{\mu\nu}\tau_{\mu\nu}$ extracted from Fig.~\ref{fig:FeynExp}, we have
\begin{equation}
\bar{u}(\mathbf{p},s)\tau u(\mathbf{q},s')=-2m_e^2.
\end{equation}
The matrix element then reduces to
\begin{equation}
\mathcal{M}_{\rm{\bf{(a)}}}=-\frac{1}{{M}^2_{\rm{Pl}}} \frac{4m^4_e\alpha^2_{\rm{K}}}{t}\left[\sum_{n=0}^{\infty} \left(-4\alpha_{\rm{K}}^2\frac{1}{t}\right)^n\right]\delta_{s_1 s_3}\delta_{s_2 s_4} =-\frac{1}{{M}^2_{\rm{Pl}}} \frac{4m^4_e\alpha^2_{\rm{K}}}{(1+4\alpha^2_{\rm{K}})t}\delta_{s_1 s_3}\delta_{s_2 s_4}.
\end{equation}
To extract the non-relativistic potential, we take $t=-\mathbf{Q}^2$ (where $\mathbf{Q}$ is the exchange momentum), and the contribution to the Yukawa potential is
\begin{equation}
V_{\rm{\bf{(a)}}}(r)=-\frac{1}{{M}^2_{\rm{Pl}}}\frac{m^2_e}{(4+\alpha^{-2}_{\rm{K}})}\int\frac{\textnormal{d}^3{\mathbf{Q}}}{(2\pi)^3}e^{i\mathbf{Q}\cdot\mathbf{x}} \frac{1}{\mathbf{Q}^2}
=-\frac{1}{4\pi r}\frac{m^2_e}{{M}^2_{\rm{Pl}}2(2\omega +3)}.
\label{Vcham}
\end{equation}
The contributions from the remaining processes in Fig.~\ref{fig:sigmakinetic2} are
\begin{subequations}
\label{V_other}
\begin{align}
V_{\rm{\bf{(b)}}}(r)=& V_{\rm{\bf{(c)}}}(r)=-\frac{1}{4\pi r}\left(\frac{c'v_\chi}{cv_\Phi}\right)\frac{m^2_e}{{M}^2_{\rm{Pl}}2(2\omega +3)},\\
V_{\rm{\bf{(d)}}}(r)=&-\frac{1}{4\pi r}\left(\frac{c'v_\chi}{cv_\Phi}\right)^2\frac{m^2_e}{{M}^2_{\rm{Pl}}2(2\omega +3)},
\end{align}
\end{subequations}
and the sum of all the contributions to the Yukawa potential is
\begin{equation}
V_5(r)=-\frac{m_e^2}{4\pi rM_{\rm{Pl}}^2}\frac{\left(1+\frac{v_\chi \gamma}{v_\Phi}\right)^2}{2(2\omega+3)}.
\label{Veffe}
\end{equation}
After some algebra, we can show that
\begin{equation}
\label{eq:gammadef}
\gamma=\frac{c'}{c}
=-\frac{\beta v^2_\Theta}{\lambda v_\chi v_\Phi},
\end{equation}
and, using the fact that $v_\Theta^2=(\lambda v_\Phi^2-6\mu^2)/\beta$, we obtain the following final expression
\begin{equation}
V_5(r)=-\frac{1}{4\pi r}\frac{m_e^2}{M^2_{\rm{Pl}}2(2\omega +3)}\frac{4\mu^4}{m_\Phi^4},
\label{eq:V potential_spontaneous}
\end{equation}
where we recall that
\begin{equation}
m_\Phi^2=2\mu^2+\frac{\beta v_\Theta^2}{3}.
\end{equation}
This is in perfect agreement with the result in the Einstein frame\footnote{Since conformal transformations modify the rulers used to measure distances, we must compare dimensionless quantities, which are unaffected by coordinate transformations. This could, e.g., be the ratio of the fifth-force potential to the standard Newtonian potential. This is to say that the expressions for the potentials should match but with $\tilde{M}_{\rm{Pl}}$ for the Einstein frame and $M_{\rm{Pl}}$ for the Jordan frame.} [Eq.~\eqref{solu}].
Notice therefore that we also find that the fifth force vanishes in the absence of explicit scale breaking ($\mu\to0$), as we did in the Einstein frame.
Before concluding this work, we consider in the next subsection the full expressions for the matrix element of the M\o{}ller scattering away from the non-relativistic limit.
\subsection{M\o{}ller scattering for purely explicit scale breaking ($\beta \to 0$)}
\label{calc:inconsistent}
For simplicity, we will work in the purely explicit scale-breaking limit (i.e., $\beta \to 0$). We consider the spin-averaged squared matrix element
\begin{equation}
\overline{|\mathcal{M}|^{2}}=\frac{1}{4}\sum_{ \text {spins }}|\mathcal{M}|^{2},
\end{equation}
After some algebra, we find the $t$-channel contribution
\begin{align}
\overline{|\mathcal{M}_{\rm{JF}}|^{2}}=&\frac{\tilde{M}^4}{M^4}\overline{|\mathcal{M}_{\rm{EF}}|^{2}}\nonumber\\=&\frac{1}{4{M}^4t^2} {\rm Tr} \left\lbrace \left[\frac{3}{2}(\slashed{p}_1+\slashed{p}_3)-4m_e\right](\slashed{p}_3+m_e)\left[\frac{3}{2}(\slashed{p}_1+\slashed{p}_3)-4m_e\right](\slashed{p}_1+m_e) \right\rbrace \nonumber\\
&\times {\rm Tr} \left\lbrace \left[\frac{3}{2}(\slashed{p}_2+\slashed{p}_4)-4m_e\right](\slashed{p}_4+m_e)\left[\frac{3}{2}(\slashed{p}_2+\slashed{p}_4)-4m_e\right](\slashed{p}_2+m_e) \right\rbrace,
\label{CS_SH}
\end{align}
showing the same frame covariance as in the Yukawa potential in the sense that the Jordan- and Einstein-frame results differ only by $M\to \tilde{M}$.\footnote{Had we not used the updated harmonic gauge condition in Eq.~\eqref{gfjf1}, we would have found agreement up to an additional numerical multiplicative factor.} The mass scales are related to the Planck masses of each theory through ${M}^2=2(2\omega +3) M_{\rm{Pl}}^2$.
Including the $u$-channel contribution ($u=-(p_4-p_1)^2$), and taking the ultra-relativistic limit (i.e., $m_e\to0$), we obtain:\footnote{To calculate the traces of the product of the gamma matrices, we used the Mathematica package FeynCalc~\cite{Shtabovenko:2020gxv,Shtabovenko:2016sxi,Mertig:1990an}.}
\begin{equation}
\overline{|\mathcal{M}_{\rm{JF}}|^{2}}=\frac{\tilde{M}^4}{M^4}\overline{|\mathcal{M}_{\rm{EF}}|^{2}}=\frac{81\left(t^2 + u^2 \right)}{16M^4}.
\end{equation}
We thus see that the scalar-harmonic gauge from Eq.~\eqref{gfjf1} leads to a perfect agreement between frames.
\section{Conclusion}\label{fifth}
In this paper, we have studied the fifth forces that can arise in scalar-tensor theories of gravity by considering the tree-level matrix elements directly in the Jordan frame. To do so, we had to perturb both the metric and the matter fields, requiring a consistent linearization of the modified gravity sector.
We have shown that the fifth forces arise through a kinetic mixing between the graviton and the non-minimally coupled field in the Jordan frame. For the specific model described in this work, we have illustrated how the diagonalization of the mass matrix of the scalar sector yields an additional direct coupling between the massless fifth-force mediator and the fermion through the original Yukawa coupling of the would-be Higgs field. By this means, we were able to illustrate from the Jordan-frame perspective the role played by sources of explicit scale breaking in the matter sector in allowing fifth forces to couple to matter fields.
In addition, we have shown that a full evaluation of the symmetries of the modified gravitational action is crucial for consistently updating the usual gauge fixing conditions. By imposing diffeomorphism invariance on the gravitational action in the knowledge that it can be mapped to Einstein gravity, we obtained a redefinition of the covariant derivative, wherein the breaking of the weak equivalence principle is manifest. In this way, we were led to an update of the usual harmonic gauge --- the so-called scalar-harmonic gauge [see Eq.~\eqref{gfjf1}] --- providing results in perfect agreement with those found in the Einstein frame.
While this gauge fixing term is not completely new, we have been able to define it at the level of the full metric. Up to second order, one can equivalently define the gauge using the DeWitt background condition~\cite{DeWitt:1964mxt, Barvinsky:1985an,Steinwachs:2011zs}. However, there are additional operators at order three and four in the graviton and additional scalar fields, which are captured by our full metric expression.
This work forms a basis for considering the consistent gauging and performing perturbative analyses of fifth forces in other modified theories of gravity, including and especially those for which an Einstein frame does not exist. This may form the focus of future works.
\section*{Note added}
While preparing the arXiv preprint of this work for journal submission, a thesis was uploaded to the arXiv \cite{Riva:2021zli}, where similar techniques are used for the derivation of the scalar-harmonic gauge [Eq.\eqref{gfjf1}].
\begin{acknowledgements}
The authors thank Clare Burrage for helpful discussions. This work was supported by the Science and Technology Facilities Council (STFC) Consolidated Grant [Grant No.\ ST/T000732/1] and STFC studentship [Grant No.\ ST/V506928/1]; a United Kingdom Research Innovation (UKRI) Future Leaders Fellowship [Grant No.\ MR/V021974/1]; a Nottingham Research Fellowship from the University of Nottingham; and a Leverhulme Research Fellowship [Grant No.\ RF-2021-312].
\end{acknowledgements}
|
\section{Introduction}
The goal of speech enhancement (SE) is to convert low-quality speech signals to ones with improved quality and intelligibility. SE serves as an important regression task in the speech-processing field and has been widely used for a pre-processor in speech-related applications, such as speech coding~\cite{li2011comparative}, automatic speech recognition (ASR)~\cite{li2015robust}, speaker recognition~\cite{michelsanti2017conditional}, and assistive hearing devices~\cite{wang2017deep, lai2016deep}. Recent advances in machine learning have made significant progress to the SE technology. Generally, learning-based SE approaches estimate a transformation to characterize the mapping function from noisy to clean speech signals in the training phase~\cite{wang2018supervised}. The estimated transformation converts noisy speech signals to generate clean-like signals in the testing phase. Various neural network models have been used to characterize noisy-to-clean transformations. Well-known examples of such models include fully connected neural network~\cite{xu2014regression}, deep denoising autoencoder~\cite{lu2013speech}, convolutional neural network~\cite{fu2016snr}, long-short-term memory~\cite{chen2015speech}, and Transformer~\cite{koizumi2020speech}. To effectively handle diverse noisy conditions, we usually prepare a considerable amount of training data that cover various noise types to train SE models. However in real-application scenarios, the noise types in the testing data may not always be involved in the training set. Consequently, the noisy-to-clean transformation learned from the training data cannot be suitably applied to handle the testing noise, resulting in limited enhancement performance. This training-testing mismatch is generally called a domain mismatch issue for SE. An effective solution is required to perform domain adaptation to adjust the SE models with formulating a precise noisy-to-clean transformation that matches the testing conditions. Most existing domain adaptation methods rely on at least one of the following adaptation mechanisms: aligning domain-invariant features~\cite{tzeng2014deep, sun2016deep, morerio2017minimal} and adversarial training, where a discriminator is introduced during training as a domain classifier~\cite{ganin2015unsupervised, ganin2016domain,tzeng2017adversarial}.
This study aims to solve the unsupervised domain adaptation problem for SE by introducing optimal transport (OT). In particular, we consider the circumstance where SE is tested on a target domain with completely unlabeled data, and only labeled data from the source domain is available for reference. Generally speaking, OT theory compares two (probability) distributions and considers all possible transportation plans in between to find one with a minimal displacement cost. The concept of OT can be applied to minimize domain mismatch and consequently achieve unsupervised domain adaptation. Even with the mathematical characteristics offered by OT, obstacles to excellent SE performance persist due to the complex structure possessed by human speech. To further overcome the obstacles, another concept from Generative Adversarial Network (GAN) is integrated to assist attaining sophisticated SE domain adaptation. Although an existing domain transition technique ``domain adversarial training'' and our proposal share similarity in names, the fundamental constructions are substantially different. A key element in our method lies in a discriminator utilized to examine speech output characteristics, instead of a \textit{domain classifier}. More precisely, the discriminator in our method was employed to govern the output speech quality by learning the probability distribution of the source labels. This novel approach was designed especially for the unsupervised SE domain adaptation to exhibit excellent performance, which was verified on the VoiceBank and TIMIT datasets.
\paragraph{Contributions}
We proposed a novel method designed specifically for unsupervised domain adaptation in a regression setting. This area of study still has very limited results; moreover, the existing methods often require additional classification of source domains or may not yet be supported by strong regression applications. Conversely, our approach does not require any additional input information other than the source samples, source labels, and target samples. Our approach was applied to two standardized SE tasks, namely VoiceBank-DEMAND and TIMIT, and achieved superior adaptation performance in terms of both Perceptual Evaluation of Speech Quality (PESQ) and Short-Time Objective Intelligibility (STOI) scores. Furthermore, owing to the simple input requirements, we can easily investigate the effect of target sample complexity on our method by increasing the number of noise types allowed in the target domain, which has not been reported by previous literature to the best of our knowledge.
\section{Related work}
\paragraph{Adversarial domain adaptation:}
The main objective of Domain Adversarial Training (DAT) is to train a deep model (from the source domain) capable of adapting to other similar domains by leveraging a considerable amount of unlabeled data from the target domain~\cite{ganin2015unsupervised, ganin_book}. A conventional DAT system consists of three parts, deep feature extractor, label predictor, and domain classifier. By using a gradient reversal layer, the extracted deep features are discriminative for the main learning task and invariant with shifts between the source and target domains. The DAT approach has been applied and confirmed to effectively compensate for the mismatch of source (training time) and target (testing time) conditions in numerous tasks, such as speech signal processing~\cite{liao2018noise, serdyuk2016invariant}, image processing~\cite{ganin2015unsupervised, tang2020srda}, and wearable sensor signal processing~\cite{liu2020domain}. A later development in Multisource Domain Adversarial Networks (MDAN)~\cite{zhao2018adversarial} extended the original DAT to lift the constraint of single domain transition, utilizing multiple domain classifiers to extract discriminative deep features for the main learning task while being invariant to multiple domain shifts~\cite{michelsanti2017conditional, richard2020unsupervised}.
\paragraph{Optimal transport for domain adaptation:}
Hitherto, OT ~\cite{villani2008optimal, peyre2019computational} has been utilized to domain adaptation~\cite{courty2014domain, courty2016optimal} with related analytical results~\cite{redko2017theoretical}. Furthermore, the concepts of OT have proved itself even more useful under a joint distribution framework~\cite{courty2017joint, damodaran2018deepjdot}. More recently, to improve the sensitivity of OT to outliers, Robust OT was proposed~\cite{balaji2020robust}. Moreover, a method for combining the notion of adversarial domain adaptation with OT and margin separation has been proposed~\cite{dhouib2020margin}. However, almost all experiments performed in these studies were classification problems, unlike the SE task that is the focus of our study.
\paragraph{Domain adaptation in speech enhancement:}
Existing domain adaptation in SE approaches can be divided into two categories: supervised and unsupervised. For supervised domain adaptation, paired noisy and clean speech signals for the testing conditions are available to adjust the parameters in the SE models. In~\cite{xu2014cross, wang2015transfer}, transfer-learning-based approaches have been proposed to adapt SE models to alleviate corpus mismatches. To combat the catastrophic forgetting issue, Lee et al., proposed a SERIL algorithm that combines curvature-based regularization and path optimization augmenting strategies when preforming domain adaptation on SE models~\cite{lee2020seril}. Conversely, for unsupervised domain adaptation, only noisy speech signals are provided, and the corresponding clean counterparts are not accessible. Generally, unsupervised domain adaptation has good applicability to real-world scenarios. In~\cite{wang2020cross}, unsupervised domain adaptation for SE was performed by
minimizing the Kullback-Leibler divergence between posterior probabilities produced by teacher and student senone classifiers without paired noisy-clean adaptation data. In~\cite{liao2018noise, hou2019domain}, the DAT approach was used to adapt SE models to new noisy conditions.
Despite yielding promising performance, the existing unsupervised domain adaptation approaches require additional information, such as word labels, language models, and noise-type labels. In this paper, we propose a new approach: discriminator-constrained OT network (DOTN) to perform unsupervised domain adaptation on SE. In contrast to related works, DOTN does not require additional label information when adapting the original SE models to match new noisy conditions. Our experiments show that DOTN can effectively adapt the SE models to new testing conditions, and that it achieves better adaptation performance than previous adversarial domain adaptation methods, which require additional noise type information.
\section{Method}
\subsection{Problem setting and notation }
Consider a source domain with paired data $(\mathbf{X}^s, \mathbf{Y}^s)= \left\{(\mathbf{x}^s_i, \mathbf{y}^s_i ) \right\}_{i=1}^{N_s}$, where $\mathbf{x}^s_i \in \mathbb{R}^n$, $\mathbf{y}^s_i \in \mathbb{R}^m$ stand for the input and the corresponding label of sample $i$. The unsupervised domain adaptation assumes that another target domain exists containing only data unlabeled, $\mathbf{X}^t=\left\{\mathbf{x}^t_i \in \mathbb{R}^n \right\}_{i=1}^{N_s}$ . The goal is to seek a ground truth estimator (or statistical hypothesis) $f: \mathbb{R}^n \to \mathbb{R}^m$ for the target labels $\mathbf{Y}^t=\left\{\mathbf{y}^t_i\right\}_{i=1}^{N_t}$ (exist but not known), based on the knowledge provided by the source domain.
The probability distribution of a dataset $\mathcal{D}$ is denoted by $\mathbb{P}_{\mathcal{D}}$, where $\mathcal{D}$ is either $\mathbf{X}^s$, $\mathbf{Y}^s$, $\mathbf{X}^t$ or $\mathbf{Y}^t$ in our discussion. Our problem is to find a function $f$ such that $f$ induces a probability distribution $\mathbb{P}_{f(\mathbf{X}^t)}$ in $\mathbf{Y}^t$ with $\mathbb{P}_{f(\mathbf{X}^t)} \rightarrow \mathbb{P}_{\mathbf{Y}^t}$ under certain measure. We propose using the concept of OT to solve this problem.
\subsection{Proposed model: Discriminator-Constrained Optimal Transport Network (DOTN)}
Given a pair of distributions $\mathbb{P}_{\mathcal{D}_1}$ and $\mathbb{P}_{\mathcal{D}_2}$ and a displacement cost matrix $C \geq 0$, OT solves for the transportation plan $\gamma\in\prod(\mathbb{P}_{\mathcal{D}_1},\mathbb{P}_{\mathcal{D}_2})$ that minimizes the total cost (in a discrete setting)
\begin{equation}\label{E:total cost}
\min_{\gamma\in\prod(\mathbb{P}_{\mathcal{D}_1},\mathbb{P}_{\mathcal{D}_2})} \langle C, \gamma \rangle_F,
\end{equation}
where $\prod(\mathbb{P}_{\mathcal{D}_1},\mathbb{P}_{\mathcal{D}_2})$ denotes the space of joint distributions with marginal $\mathbb{P}_{\mathcal{D}_1}$ and $\mathbb{P}_{\mathcal{D}_1}$. $\langle \cdot ,\,\cdot \rangle_F$ is the Frobenius product, and the entry $C_{ij}$ of $C$ represents the displacement cost of the $i^{th}$ and $j^{th}$ samples. It can be proven that the minimum of this problem is a distance and called the Wasserstein distance when the corresponding cost is a norm~\cite{villani2008optimal, peyre2019computational}.
Our proposed method consists of two parts: OT alignment and Wasserstein Generative Adversarial Network (WGAN) training~\cite{arjovsky2017wasserstein, gulrajani2017improved}. Both steps are based on OT; however, they are considered with two different pairs of distributions and employ different algorithms.
\paragraph{Adaptation by Joint Distribution Optimal Transport}
Our adaptation mechanism relies on the alignment between the joint distributions of source and target domains (for the target domain, the label is estimated label). In particular, we approximate $f$ by minimizing the OT loss between the joint distributions $\mathbb{P}_{\mathbf{X}^s}\times \mathbb{P}_{\mathbf{Y}^s}$ and $\mathbb{P}_{\mathbf{X}^t}\times \mathbb{P}_{f(\mathbf{X}^t)} $, with a chosen cost matrix,
\begin{equation}\label{cost matrix}
C_{ij} = \alpha \, \| \mathbf{x}^s_i - \mathbf{x}^t_j \|^2 + \beta \, \| \mathbf{y}^s_i - f( \mathbf{x}^t_j ) \|^2, \quad \qquad (\alpha, \beta > 0)
\end{equation}
By aligning the joint distributions of the source and target domains, noise adaptation is naturally achieved as OT seeks the source sample that is the most ``similar'' for each target sample.
Although the OT provides accurate estimates for each sample, the effect of each estimation error could accumulate in the training process and mislead $f$ to a convenient local minimum without preserving the speech data structure. To avoid this situation, we employ Wasserstein Generative Adversarial Network (WGAN) training to complement and enhance our adaptation system.
\paragraph{Discriminative training on outputs and source labels}
Distinct from the adaptation where we consider the joint distributions of the inputs and labels, we focus on WGAN training for the source label distribution $\mathbb{P}_{\mathbf{Y}^s}$ and output distribution $\mathbb{P}_{f(\mathbf{X}^t)}$. In the terminology of generative adversarial training, we consider $f$ a generator and introduce a convolutional neural network-based discriminator, $h$, as the 'critic'. In general, we use the discriminator to decide whether the outputs of $f$ are 'similar' to the source labels. Formally, the WGAN algorithm solves
\begin{equation}\label{Eq: WGAN}
\min_f\max_{h\in \mathcal{L}} \Big\{\mathbb{E}_{y\sim \mathbb{P}_{\mathbf{Y}^s}}(h(y)) - \mathbb{E}_{x\sim \mathbb{P}_{\mathbf{X}^t}}(h(f(x))) \Big\}
\end{equation}
by Kantorovich-Rubinstein duality~\cite{villani2008optimal}, where $\mathcal{L}$ is the set of 1-Lipschitz functions. In this case, under an optimal discriminator minimizing the value function with respect to the generator
parameters minimizes the Wasserstein distance between the distributions $\mathbb{P}_{\mathbf{Y}^s}$ and $\mathbb{P}_{f(\mathbf{X}^t)}$.
This discriminative training complements our alignment and provides additional constraints from the explicit relation between the source labels and estimations of target labels. These constraints support the joint distribution alignment and provides further guidance in the gradient descent training process. The performance of the experiments is considerably improved when our discriminative training supports the joint distribution alignment.
\subsection{Loss functions and the proposed algorithm}
Our domain alignment can be achieved by solving the following optimization problem:
\begin{equation}\label{E:OT}
\min_{\gamma, f} \mathcal{L}_1+\mathcal{L}_2= \min_{\gamma, f}\frac{1}{N^s}\sum_{i} \|\mathbf{y}_i^s-f(\mathbf{x}_i^s)\|^2+\sum_{i,j}\gamma_{ij} \Big( \alpha \, \|\mathbf{x}^s_i-\mathbf{x}^t_j\|^2 +\beta \, \|\mathbf{y}^s_i -f( \mathbf{x}^t_j)\|^2 \Big),
\end{equation}
where $\alpha,\beta>0$ are the parameters chosen for balance. Notably, the first term emphasizes the knowledge from the source domain is not to be forgotten during training, which has been revealed in several works ~\cite{damodaran2018deepjdot, li2017learning, shmelkov2017incremental}. Without this emphasis, the source domain knowledge cannot be well-maintained, and thus the overall performance may degrade. Such phenomenon was also observed in the SE experiments. The second term is intended for domain alignment. To show some intuitions, consider the ideal case where Eq.~(\ref{E:OT}) is completely minimized to zero, which leads to
\begin{equation}
\|\mathbf{x}^s_i - \mathbf{x}^t_j \|^2 \equiv 0 \,\, \mbox{and} \,\, \| \mathbf{y}^s_i -f( \mathbf{x}^t_j) \|^2 \equiv 0 \quad \Rightarrow \quad \mathbf{x}^s_i = \mathbf{x}^t_j \,\, \mbox{and} \,\, \mathbf{y}^s_i = f( \mathbf{x}^t_j)
\end{equation}
for all $i, j$. This indicates that for each given target domain sample $\mathbf{x}^t_j$, an identical sample $\mathbf{x}^s_i$ from the source domain is found, and the unknown target label is then constructed by the corresponding source label.
Even though practically the ideal case of zero loss is unlikely to happen, the OT loss aims to search for the most ``similar'' correspondence, which entails the intuition of domain alignment for Eq.~(\ref{E:OT}). From this point of view, it is noted that although the term $\|\mathbf{x}^s_i-\mathbf{x}^t_j\|$ in $C_{ij}$ (in Eq. (\ref{cost matrix})) is not directly related to the network backpropagations of $f$ and $h$, it may not be ignored as the discard of this term will result in a wrong transportation plan $\gamma_{ij}$ and eventually lead to an undesired alignment.
For discriminative training, the discriminator $h$ is trained by the discriminator loss function $\mathcal{L}_h=\frac{1}{m}\sum_{i=1}^m h(\mathbf{y}_i^s)-h(f(\mathbf{x}^t_i)),$ and $f$ follows the generator loss function $\mathcal{L}_f= - \frac{1}{m}\sum_{i=1}^m h(f(\mathbf{x}^t_i))$ where $m$ is the batch size. As there are multiple sets of parameters $\gamma$, $h$ and $f$ in our framework, one set of parameters is updated each time, while the other sets of parameters are fixed.
\begin{figure}
\centering
\includegraphics[width=0.68 \columnwidth]{./figures/network_structure.pdf}
\caption{DOTN network structure.}
\end{figure}
\begin{algorithm}
\caption{DOTN, proposed algorithm}
\begin{algorithmic}[1]
\Require $\mathbf{x}^s$, source domain inputs. $\mathbf{y}^s$, source domain labels. $\mathbf{x}^t$, target domain inputs. $c$, the clipping parameter. $m$, the batch size. $n_f, n_h, n_s$: number of iterations of OT per generator training, discriminator training, and source domain training, respectively. $n$, number of iterations.
\Require $\theta_f$, initial parameters of estimator $f$. $\theta_h$, initial parameters of discriminator $h$.
\For{each batch of source samples $(\mathbf{x}^s, \mathbf{y}^s)$ and target samples $(\mathbf{y}^t)$}
\State fix $\theta_f$, solve for $\gamma$ in Eq.~(\ref{E:OT}) by OT.
\State fix $\gamma$, $\theta_f\leftarrow $Adam$(\nabla_{\theta_f}\mathcal{L}_2,\theta_f, \theta_h)$.
\If{$n$ mod $n_{s}==0$ }
\State $\theta_f\leftarrow$Adam$( \nabla_{\theta_f}\mathcal{L}_1, \theta_f, \theta_h)$.
\EndIf
\If{$n$ mod $n_{f}==0$ }
\State $\theta_f\leftarrow$Adam$( \nabla_{\theta_f}\mathcal{L}_f, \theta_f, \theta_h)$.
\EndIf
\If{$n$ mod $n_{h}==0$ }
\State $\theta_h\leftarrow $Adam$(\nabla_{\theta_h}\mathcal{L}_h, \theta_f, \theta_h)$.
\State $\theta_h\leftarrow$clip$(\theta_h, -c,c).$
\EndIf
\EndFor
\end{algorithmic}
\end{algorithm}
\section{Experiments}
We evaluated our method in SE on two datasets: Voice Bank corpus~\cite{veaux2013voice} and TIMIT~\cite{garofolo1988getting}. Although there is abundant literature on unsupervised domain adaptation, very limited methods have been successfully applied to SE or regression problems. Closely relevant methods often require additional input structures or domain label. Unlike previous methods, our approach does \textbf{not} rely on additional data information. Nevertheless, we compared our results with two most relevant adversarial domain adaptation methods: DAT~\cite{liao2018noise} and MDAN~\cite{zhao2018adversarial}. The original MDAN was designed for classification problems and cannot be directly applied on SE. Preserving the fundamental ideas, we modified the MDAN structure for regressions, so that SE experiments can be performed for comparison. The implementations are summarized in the supplementary material and codes are available on Github\footnote{\href{https://github.com/hsinyilin19/Discriminator-Constrained-Optimal-Transport-Network}{https://github.com/hsinyilin19/Discriminator-Constrained-Optimal-Transport-Network} }.
\subsection{Comparisons to DAT and MDAN}
As both DAT\footnote{DAT consists of three components, a deep feature extractor $\mathcal{E}: \mathcal{X} \to \mathcal{Z}$, a (task) label predictor $F_Y: \mathcal{Z} \to \mathcal{Y}$, and a domain classifier $F_D: \mathcal{Z} \to \mathcal{K}$, to form two functional pairs: $F_D \circ \mathcal{E}$ and $F_Y \circ \mathcal{E}$. Here $\mathcal{X}$, $\mathcal{Y}$ are the input space and (task) label space respectively. $\mathcal{Z}$ denotes the invariant (latent) feature space; $\mathcal{K} = \{0:\mbox{source}, 1:\mbox{target}\}$ as the domain label classes. The pair $F_D \circ \mathcal{E}: \mathcal{X} \to \mathcal{K}$ formed by $F_D$, $\mathcal{E}$ works against each other, by a Gradient Reversal Layer, to derive domain invariant features in $\mathcal{Z}$. Another pair $F_Y \circ \mathcal{E}: \mathcal{X} \to \mathcal{Y}$ demands that the invariant features in $\mathcal{Z}$ encode sufficient information for (main task) label classifications at the same time. Via adversarial training, the two pairs eventually reach a balance completing the main learning task as well as eliminating the domain mismatch.} and MDAN\footnote{MDAN inherits the DAT architecture with an extension to $K$ source domains and $K$ domain classifiers $F_{D_i}: \mathcal{Z} \to \{0, 1\}$, $i=1,\ldots K$ with similar adversarial training applied.} utilize a domain classifier to derive domain invariant features, domain labels become essential in their adaptation mechanism. Thus, both these methods are considered \textit{weakly supervised}. Specifically, they request inputs of the form $\left\{\mathbf{x}_i^s, \mathbf{y}^s_i, c_i\right\}_i$ and $\left\{\mathbf{x}_j^t, c_j\right\}_j$ with $c_i, c_j \in \mathcal{K}$, where $\mathcal{K}$ is an index set specifying the origin of data domains. For DAT, $\mathcal{K}=\{0, 1\}$ to indicate whether a sample belongs to the source or target domain, while for MDAN $\mathcal{K} = \{0, 1, \ldots, K\}$ to indicate that there are $K$ distinct source domains and one target domain. The requirement on domain labels poses restrictions in certain circumstances, such as a new target sample may not always fall into any existing categories, or the data origin is simply unknown. In contrast, the proposed method DOTN is not bounded by specifications of data origin and simply receives inputs of the form $\left\{\mathbf{x}_i^s, \mathbf{y}^s_i, \mathbf{x}_j^t\right\}_{i,j}$. With less input requirement, DOTN is more flexible to be applied in various scenarios, more approachable to real-world applications.
\subsection{Voice Bank with DEMAND noise database}
\paragraph{Dataset} In the first set of experiments, the pre-selected subset of Voice Bank provided by~\cite{valentini2016investigating} was used to test the proposed DOTN. For source domain data, 14 male speakers and 14 female speakers were randomly selected out of totally 84 speakers (42 male and 42 female), and each speaker pronounced around 400 sentences. As a result, the clean data in source domain contained 5,724 male utterances and 5,848 female utterances, amounting to 11,572 utterances in total. We then mixed the 11,572 clean utterances with noise from DEMAND~\cite{thiemann_joachim_2013_1227121}, in transportation category: ``Bus'', ``Car'', ``Metro'' at 7 Signal-to-noise ratio (SNR) levels (-9, -6, -3, 0, 3, 6, and 9 dB). Accordingly, for this set of source domain data, both noisy speech signals and the corresponding clean references are prepared.
A target domain data contained $5,768$ noisy utterances mixed by 824 clean ones from two random speakers (1 male, 1 female) (followed the design in ~\cite{pascual2017segan}) with one of the three noise types from DEMAND, in STREET category: ``Traffic'', ``Cafe'', or ``Public square'' and 7 SNR ratios (-9, -6, -3, 0, 3, 6, and 9 dB). No clean labels were given under target domain. That is, for the target domain data, only noisy speech signals are provided and the corresponding clean references are not accessible.
All source samples (both noisy utterances and the corresponding clean references, prepared from the three transportation noise types) and the target samples (only noisy utterances without the corresponding clean references, prepared from the one street noise type) were included in our training set. We conducted the experiments under single-target noise-type circumstances and compared the results with DAT and MDAN. More specifically, we ran this setting for all three cases, where the target domain contained either cafe, public square, or traffic noise.
\paragraph{Results}
Table~\ref{VBD_table_pesq} and Table~\ref{VBD_table_stoi} list the PESQ and STOI scores, respectively, of DAT, MDAN and DOTN under three noise types at 7 SNR levels. ``Avg'' denotes the averaged scores over 7 SNR levels. From the PESQ scores reported in Table~\ref{VBD_table_pesq}, we note that the proposed DOTN outperforms both DAT and MDAN consistently over different noise types and SNR levels, except for the Cafe noise type at -9 dB SNR. This might be owning to a potential limitation of DOTN, which will be detailed in the next section. Next, from Table~\ref{VBD_table_stoi}, we note that STOI scores show very similar trends to that of PESQ scores, as listed in Table~\ref{VBD_table_pesq}.
\begin{table}
\caption{PESQ scores for VoiceBank-DEMAND}
\label{VBD_table_pesq}
\centering
\resizebox{\columnwidth}{!}{%
\begin{tabular}{cccccccccc}
\toprule
noise type & \multicolumn{3}{c}{\textbf{Traffic}}& \multicolumn{3}{c}{\textbf{Cafe}}& \multicolumn{3}{c}{\textbf{Public square}} \\ [0.5ex]
\cmidrule(r){2-10}
SNR(dB)/model & \textbf{DAT}~\cite{liao2018noise} &\textbf{MDAN}~\cite{zhao2018adversarial} &\textbf{DOTN} &\textbf{DAT} &\textbf{MDAN} & \textbf{DOTN} &\textbf{DAT} &\textbf{MDAN} &\textbf{DOTN}\\
\cmidrule(r){1-10}
-9 &1.307 &1.670 &\textbf{1.863} &1.058 &\textbf{1.539} &1.497 &1.436 &1.929 &\textbf{2.037} \\
-6 &1.446 &1.920 &\textbf{2.182} &1.184 &1.735 &\textbf{1.853} &1.655 &2.119 &\textbf{2.258} \\
-3 &1.718 &2.153 &\textbf{2.395} &1.362 &1.949 &\textbf{2.089} &1.939 &2.318 &\textbf{2.439} \\
0 &2.081 &2.366 &\textbf{2.591} &1.599 &2.153 &\textbf{2.332} &2.268 &2.512 &\textbf{2.601} \\
3 &2.381 &2.535 &\textbf{2.740} &1.865 &2.345 &\textbf{2.490} &2.575 &2.670 &\textbf{2.761} \\
6 &2.712 &2.708 &\textbf{2.888} &2.189 &2.534 &\textbf{2.661} &2.867 &2.824 &\textbf{2.889} \\
9 &3.016 &2.854 &\textbf{3.015} &2.493 &2.695 &\textbf{2.783} &3.120 &2.966 &\textbf{3.057} \\
Avg &2.094 &2.315 &\textbf{2.525} &1.679 &2.136 &\textbf{2.244} &2.266 &2.477 &\textbf{2.577} \\
\bottomrule
\end{tabular}}
\end{table}
\begin{table}
\caption{STOI scores for VoiceBank-DEMAND}
\label{VBD_table_stoi}
\centering
\resizebox{\columnwidth}{!}{%
\begin{tabular}{cccccccccccc}
\toprule
noise type & \multicolumn{3}{c}{\textbf{Traffic}}& \multicolumn{3}{c}{\textbf{Cafe}}& \multicolumn{3}{c}{\textbf{Public square}} \\ [0.5ex]
\cmidrule(r){2-10}
SNR(dB)/model &\multicolumn{1}{c}{ \textbf{DAT}~\cite{liao2018noise}}&\multicolumn{1}{c}{ \textbf{MDAN}~\cite{zhao2018adversarial}}&\multicolumn{1}{c}{ \textbf{DOTN}}&\multicolumn{1}{c}{ \textbf{DAT} }&\multicolumn{1}{c}{ \textbf{MDAN} }&\multicolumn{1}{c}{ \textbf{DOTN}}&\multicolumn{1}{c}{ \textbf{DAT} }&\multicolumn{1}{c}{ \textbf{MDAN}}&\multicolumn{1}{c}{ \textbf{DOTN}}\\
\cmidrule(r){1-10}
-9 &0.584 &0.708 &\textbf{0.721} &0.557 &\textbf{0.643} &0.633 &0.667 &0.747 &\textbf{0.765} \\
-6 &0.659 &0.761 &\textbf{0.790} &0.616 &0.702 &\textbf{0.723} &0.725 &0.792 &\textbf{0.815} \\
-3 &0.728 &0.810 &\textbf{0.833} &0.687 &0.760 &\textbf{0.779} &0.771 &0.830 &\textbf{0.850} \\
0 &0.785 &0.849 &\textbf{0.871} &0.741 &0.805 &\textbf{0.831} &0.815 &0.859 &\textbf{0.877} \\
3 &0.824 &0.874 &\textbf{0.894} &0.785 &0.842 &\textbf{0.861} &0.845 &0.881 &\textbf{0.900} \\
6 &0.860 &0.896 &\textbf{0.914} &0.825 &0.871 &\textbf{0.887} &0.872 &0.898 &\textbf{0.915} \\
9 &0.885 &0.910 &\textbf{0.927} &0.854 &0.893 &\textbf{0.905} &0.893 &0.914 &\textbf{0.931} \\
Avg &0.761 &0.830 &\textbf{0.850} &0.724 &0.788 &\textbf{0.803} &0.798 &0.846 &\textbf{0.865} \\
\bottomrule
\end{tabular}}
\end{table}
\subsection{TIMIT}
\paragraph{Dataset}
For the second part of experiments, TIMIT was used to prepare the source and target samples. The \textit{clean speech} of the source domain $\{ \mathbf{y}^s_i \}_{i=1}^{N_s}$ for training consists of utterances, contributed by 48 male speakers and 24 female speakers from 8 dialect regions. Each speaker had 8 sentences, including 5 SX (phonetically compact sentences) and 3 SI (phonetically diverse sentences), according to the official suggestion of TIMIT. The number of the speakers selected was to maintain the balance of the original data.
The above clean utterances were used to mix with 5 \textit{stationary} noise types (Car, Engine, Pink, Wind, and Cabin) at 9 SNR levels (-12, -9, -6, -3, 0, 3, 6, 9, and 12 dB), amounting to 25,920 noisy utterances, to be the \textit{noisy speech} of the source domain $\{\mathbf{x}^s_i\}_{i=1}^{N_s}$ with $N_s = 25,920$.
For the target domain, a total of 24 speakers suggested by TIMIT core test set was all used to have 192 clean utterances for the target domain $\mathbf{y}^t_i$, which were subsequently mingled with one of the four \textit{non-stationary} noise types: ``Helicopter'', ``Cafeteria'', ``Baby-cry'', or ``Crowd-party'' under 7 SNRs (-9, -6, -3, 0, 3, 6, and 9 dB) as target inputs $\{\mathbf{x}^t_i\}_{i=1}^{N_t}$ with $N_t = 1,344$. The choice of noise types for the source and target domain was to let the learning algorithms adapt from distinguished environments in the real-world.
\paragraph{Results}
Table~\ref{TIMIT_table} lists the PESQ and STOI scores of the DAT, MDAN, and DOTN under four noise types at seven SNR levels. From the table, we first note that DOTN consistently outperforms DAT and MDAN in terms of both ``Avg'' PESQ and STOI scores among the four noise types. With a more careful investigation, the DOTN achieves higher PESQ and STOI scores over DAT and MDAN for all SNR conditions in the Helicopter noise. For the Crowd-party and Cafeteria noises, DOTN outperforms DAT and MDAN in most higher SNR conditions for both PESQ and STOI scores. However, for the Baby-cry noise, DOTN outperforms DAT and MDAN in STOI but underperforms MDAN in PESQ. Note that Cafeteria, Crowd-party, and Baby-cry noise types involved clear human speech components, which may cause confusions when DOTN tries to retrieve the target speech (clean reference). Thus, DOTN yields sub-optimal performance when dealing with these noise types, especially under very low SNR conditions, where background speech components might overwhelm the target speech. Nevertheless, the overall average PESQ and STOI scores of DOTN (P=1.838; Q=0.7238) are still higher than that of DAT (P=1.5728; Q=0.6298) and MDAN (P=1.8001; Q=0.7038), where P and Q denote the PESQ and STOI scores, respectively, over the 4 noise types and 9 SNR levels. Please also note that DAT and MDAN require additional domain label information~\cite{liao2018noise} while DOTN performs domain adaptation in a purely unsupervised manner.
To show the advantages and flexibility of DOTN, we further explored and applied DOTN to the circumstances when multiple noise types are included in the target domain. Fig~\ref{TIMIT_PESQ} and Fig~\ref{TIMIT_STOI}, respectively, present the PESQ and STOI results over four SNR levels on multiple target noise types; here H, Ca, Cr, and B denote Helicopter, Cafeteria, Crowd-party, and Baby-cry noise types, respectively. In Fig~\ref{TIMIT_PESQ}, when the testing noise is ``Helicopter'', the PESQ scores of using two noise types (H+Ca, H+Cr, H+B) and three noise types (H+Cr+B, H+Cr+Ca, H+B+Ca) are comparable. Further, the PESQ scores of these six systems are also comparable to that of using one noise type (H). The results confirmed that the DOTN has minor effects on catastrophic forgetting, which is a common issue in domain adaptation approaches~\cite{choy2006neural, goodfellow2013empirical}. In the case of 4 noise types (the system sequentially learned Helicopter, Baby-cry, Cafeteria, and Crowd-party noise types and was tested on the Helicopter noise), the achieved performance drops moderately. We observe very similar trends for the other three noise types (baby-cry, cafeteria, and crowd-party) in Fig~\ref{TIMIT_PESQ}.
From Fig~\ref{TIMIT_STOI}, similar trends as those from Fig~\ref{TIMIT_PESQ} are observed: (1) The seven results using 1, 2, and 3 noise types are comparable, thereby confirming that the DOTN has minor effects on catastrophic forgetting. (2) When there are 4 noise types for sequential learning, the achievable STOI scores start to decrease.
\begin{table}
\caption{TIMIT results}
\label{TIMIT_table}
\centering
\resizebox{\columnwidth}{!}{%
\begin{tabular}{ccccccccccccc}
\toprule
noise type &\multicolumn{6}{c}{\textbf{Helicopter}}&\multicolumn{6}{c}{ \textbf{Crowd-party}}\\
\cmidrule(r){2-13}
model &\multicolumn{2}{c}{\textbf{DAT}~\cite{liao2018noise}}&\multicolumn{2}{c}{\textbf{MDAN}~\cite{zhao2018adversarial} }&\multicolumn{2}{c}{\textbf{DOTN}}&\multicolumn{2}{c}{ \textbf{DAT} }&\multicolumn{2}{c}{\textbf{MDAN} }&\multicolumn{2}{c}{\textbf{DOTN}}\\
\cmidrule(r){2-13}
SNR(dB) & PESQ & STOI & PESQ & STOI
& PESQ & STOI & PESQ & STOI & PESQ & STOI & PESQ & STOI\\ [0.5ex]
\cmidrule(r){1-13}
-9 &1.031 &0.392 &1.252 & 0.517 &\textbf{1.455} &\textbf{0.577} &\textbf{1.483} &\textbf{0.544} & 1.150 &0.440 &1.056 &0.451\\
-6 &1.015 &0.431 &1.443 &0.594 &\textbf{1.669} &\textbf{0.649} &\textbf{1.484} &\textbf{0.560} &1.356 &0.516 &1.302 &0.538\\
-3 &1.094 &0.497 &1.664 &0.670 &\textbf{1.890} &\textbf{0.716} &1.528 &0.592 &\textbf{1.560} &0.600 &1.559 & \textbf{0.621}\\
0 &1.268 &0.566 &1.902 &0.742 &\textbf{2.104} &\textbf{0.775} &1.596 &0.636 &1.776 &0.684 &\textbf{1.816} & \textbf{0.709}\\
3 &1.518 &0.637 &2.134 &0.801 &\textbf{2.289} &\textbf{0.822} &1.736 &0.690 &1.986 &0.762 &\textbf{2.042} & \textbf{0.782}\\
6 &1.779 &0.701 &2.363 &0.849 &\textbf{2.497} &\textbf{0.865} &1.953 &0.750 &2.179 &0.823 &\textbf{2.236} & \textbf{0.838}\\
9 &2.094 &0.759 &2.563 &0.884 &\textbf{2.677} &\textbf{0.895} &2.200 &0.801 &2.355 &0.865 &\textbf{2.447} &\textbf{0.885} \\
Avg &1.400 &0.569 & 1.903& 0.722& \textbf{2.083} & \textbf{0.757} &1.711 &0.653 &1.766 &0.670 &\textbf{1.780} &\textbf{0.690} \\
\cmidrule(r){1-13}
noise type &\multicolumn{6}{c}{ \textbf{Cafeteria}}&\multicolumn{6}{c}{\textbf{Baby-cry}}\\
\cmidrule(r){2-13}
model &\multicolumn{2}{c}{\textbf{DAT}}&\multicolumn{2}{c}{\textbf{MDAN}} &\multicolumn{2}{c}{\textbf{DOTN}}&\multicolumn{2}{c}{\textbf{DAT}}&\multicolumn{2}{c}{\textbf{MDAN}}&\multicolumn{2}{c}{\textbf{DOTN}}\\
\cmidrule(r){2-13}
SNR(dB) & PESQ & STOI & PESQ & STOI
& PESQ & STOI & PESQ & STOI & PESQ & STOI & PESQ & STOI\\ [0.5ex]
\hline
-9 &1.196 & 0.440 & \textbf{1.248}& \textbf{0.471}& 1.185 &0.436 & 1.109 & 0.580 &\textbf{1.116}& 0.561& 0.984 & \textbf{0.597}\\
-6 &1.206 & 0.471 &1.395& \textbf{0.535}& \textbf{1.419}& 0.528 &\textbf{1.313}& 0.630 &\textbf{1.313}& 0.636& 1.180 & \textbf{0.668}\\
-3 &1.244 & 0.519 &1.608& 0.614& \textbf{1.631} &\textbf{0.620} &1.495 & 0.665 &\textbf{1.500}& 0.699& 1.431& \textbf{0.732}\\
0 &1.408 & 0.579 &1.827& 0.692& \textbf{1.859}& \textbf{ 0.699} & 1.621& 0.707 &\textbf{1.734}& 0.768& 1.665 & \textbf{0.789}\\
3 &1.631 & 0.651 &2.031& 0.764& \textbf{2.075} & \textbf{0.769} & 1.801 & 0.746 & \textbf{1.909}& 0.809& 1.889 & \textbf{0.835}\\
6 &1.915 & 0.719 & 2.244& 0.822& \textbf{2.271} & \textbf{0.826} &1.973 & 0.782 &\textbf{2.112}& 0.849& 2.105 & \textbf{0.871}\\
9 &2.216 & 0.782 & 2.422& 0.864& \textbf{2.458} &\textbf{0.873} & 2.133 & 0.814 &2.274& 0.880& \textbf{2.277} & \textbf{0.890} \\
Avg &1.545 &0.594 & 1.825& \textbf{0.680}& \textbf{1.843} & 0.679 & 1.635& 0.703 &\textbf{1.708}& 0.743& 1.647 &\textbf{0.769} \\
\bottomrule
\end{tabular}}
\end{table}
\begin{figure}
\centering
\includegraphics[width=0.9\columnwidth]{./figures/PESQ.pdf}
\caption{Comparisons in PESQ average for cases with multiple noise types in target domain, where H, Ca, Cr, and B denote Helicopter, Cafeteria, Crowd-party, and Baby-cry noise types.}
\label{TIMIT_PESQ}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.9\columnwidth]{./figures/STOI.pdf}
\caption{Comparisons in STOI average for cases with multiple noise types in target domain.}
\label{TIMIT_STOI}
\end{figure}
\section{Conclusion}
In this study, we proposed a novel DOTN method, which was designed specifically for unsupervised domain adaptation in regression setting. Our approach skillfully fuses OT and generative adversarial frameworks to achieve unsupervised learning in target domain based on the information provided from the source domain, requiring no additional structure or inputs, such as multi-source domain and noise type labels. Our experiments show that the proposed method is capable of superior adaptation performance in SE by outperforming other adversarial domain adaptation methods on both PESQ and STOI scores for the VoiceBank-DEMAND and TIMIT datasets. Further, we show that when moderately increasing the complexity of the target samples (by increasing the number of noise types in the target domain), only a small degree of degradation was observed. This suggests that our method is robust to sample complexity in the target domain.
\bibliographystyle{unsrtnat}
|
\section{Introduction}
Cosmological galaxy surveys, both current and planned, span increasingly large volumes, eg. DESI \cite{desi}, EUCLID \cite{euclid}, DES \cite{des3} and LSST \cite{lsst}. In the previous decade, surveys such as the Baryon Oscillation Spectroscopic Survey \cite{boss} have used bright red galaxies as tracers of large scale structure. To make larger maps, the current and next generation of surveys rely on utilizing fainter galaxies. Spectroscopic surveys specifically are planning to observe galaxies targeted by the presence of star formation within them. Such galaxies, some of them Emission Line Galaxies (ELGs; \cite{Raichoor_2017}), are the cosmological workhorse of the next decade.
An important issue with creating bigger large-scale structure with smaller galaxies is being able to simulate such maps computationally. ELGs reside in halos at or below the mass scale of our own Milky Way \cite{Guo_2019, Gonzalez_Perez_2017}. Thus, simulations of such galaxy samples require not just large volume but also higher mass resolution than previously needed. A single simulation with equal volume of the DESI ELG sample that properly resolves all halos in which ELGs could form would require upwards of $9500^3$ particles, assuming a volume of 122 Gpc and a resolution of halos at around $1000$ particles for $10^{11} \ \textup{M}_{\Sun} $ halos. We often need a large suite of such simulations to explore various parameters or build up statistics. This will quickly become computationally expensive.
A promising solution to this problem is the use of deep learning networks. Various works in recent years applied convolutional neural networks to cosmological applications successfully, generally increasing small scale structure information (eg. galaxy distributions in \cite{Yip2019}, mapping dark matter to galaxies in \cite{kasmanoff2020dm2gal,Yip2019}, extracting cosmological information while marginalizing over baryonic effects in \cite{villaescusanavarro2020neural}, or for general simulation output \cite{He2019}, \cite{deoliveira2020fast} and \cite{Kodi_Ramanah_2020}). In the case of creating cosmological simulations spanning large volumes, a simple first approach is to use DM-only density fields.
In this paper, we propose a novel method to predict the $z=0$ density field of a high resolution simulation simply by giving it the $z=0$ density field of a low resolution simulation of the same volume, run using same cosmologies and initial conditions. High and low resolution here refers to the simulation particle's mass and the number of particles used inside the simulation volume. Once a network is trained successfully it takes little computational effort to produce new output from low resolution density fields. This would allow one to run a computationally less expensive, lower resolution simulation spanning Gpc/h in volume, and then using the resulting density field to predict the structure at higher resolution using the CNN. The halos and subhalos in this high-resolution output can be populated with galaxies in any number of ways, as required by the specific goal of the mock galaxy catalog (see, \cite{TinkerWechsler} for a review on the galaxy-halo connection).
This paper provides a first proof of work that uses density fields to super resolve dark matter halos.
Previous works either focus on other statistics such as void abundance \cite{Kodi_Ramanah_2020}, or use particle displacement fields that assumes lattice pre-initial conditions (pre-ICs, the Lagrangian particle positions) \cite{YinLi2021} and therefore cannot be applied to many of the state-of-the-art simulations that use glass pre-ICs.
In particular, our presented method here can be applied to dark matter simulations with glass initial conditions \cite{white1996} (here after: IC) while previous super-resolution techniques such as \cite{YinLi2021_2} cannot. Many state of the art simulations use glass IC, such as Illustris project \cite{Vogelsberger2014b} while there is currently no satisfactory way to interpolate those glass IC particles onto a grid in order to compute the displacement field needed by previous super-resolution methods such as \cite{YinLi2021}. Any kind of interpolation technique will always create voxels that contain 0 particles, thus making displacement field calculation impossible for that pixel.
Rather than wasting computational resources on new costly simulations for super-resolution tasks,
we develop the alternative Eulerian method to exploit the existing state-of-the-art cosmological simulations that started with glass IC.
\section{Method}
\subsection{Dataset}
\label{sec:dataset}
In this work, we used redshift=0 data from the Illustris-2-Dark (high-res) and Illustris-3-Dark (low-res) dark matter only simulations \cite{Vogelsberger2014b} for training and testing.
Illustris-2-Dark has a box size of $75$ Mpc/h, and contains $910^3$ ($m_{\text{dm}} = 6 \times 10^7 \ \textup{M}_{\Sun} $, high-res) and Illustris-3-Dark contains $455^3$ ($m_{\text{dm}} = 4.8 \times 10^8 \ \textup{M}_{\Sun}$, low-res) dark matter particles.
This results in both simulation boxes having the same mean mass density, meaning if more particles are simulated in the higher resolution simulation (Illustris-2-Dark in this case) they have to be lighter.
The cosmology for these simulations is consistent with WMAP-9 measurements \cite{Bennett_2013} and the glass tiled initial conditions are described in detail in \cite{Vogelsberger2014a}. Similar work done by \cite{YinLi2021} used displacement fields created from particle positions for training, and resolved only higher mass halos (lowest physical halo mass resolved at $z=0$ is $10^{11} \ \textup{M}_{\Sun}$, containing 120 tracer particles). However, displacement fields are only naturally defined for lattice initial conditions, and are difficult to produce from glass tiled initial conditions, making this work ideal for simulations that do not have initial conditions that can be turned into lattice like configuration.
Most state of the art hydro simulations (like Illustris) use glass ICs to improve statistics at high redshift snapshots, thus the underlying DM-only simulations are usually initialized using glass tiles, which are simply copied next to each other to fill the simulation volume. There is no simple way to interpolate those initial particles onto a grid in order to then use the displacement field super-resolution method pioneered in \cite{YinLi2021}. Any kind of interpolation technique will always have some residual pixels contain zero or 2 particles. Finding a general, automatic, way to fill those remaining pixels with exactly one particle in a coherent way is very impractical if not impossible, and has so far not been done. Additionally using displacement fields also fixes the number of pixels a given simulation has to be divided into. If a simulation contains for example $512^3$ particles, exactly one pixel per spatial coordinate and particle needs to be created for training and testing, complicating general applicability on any given simulation.
In our proposed method, we do however use grids to create density grids, but we do not need to create a displacement field (which requires having at least one particle per pixel in order to find the displacement field).
Our deep learning method uses convolution layers, which are typically applied imaging data that has translational symmetry.
Convolutional neural networks (CNNs) generally work well with equidistantly gridded n-dimensional data. In our case, a fixed spatial resolution is perfect for a convolutional neural network.
We prepare dark matter density maps, with a $2048^3$ grid, creating a number count density field where each voxel of side-length $\approx 36.6$ kpc/h contains information about the number of particles that lie inside that small region.
Each of the data sets is then subsequently divided into eight sub-cubes of equal size (37.5 Mpc/h on the side), from which six are used for training, one for validation, and the last one for testing. This results in the low resolution input (Illustris-3-Dark) and high resolution target (Illustris-2-Dark) simulations being divided into $32768$ 3-dimensional $64^3$ pixel cubes, of which an eighth is used for testing.
\subsection{Data pre-processing}
\label{sec:preprocessing}
Neural networks benefit from certain input and target data distributions more than others, in particular, they should not have a very long tail or an extremely large dynamic range. This makes the following pre-processing steps necessary for a successful and efficient training process.
\subsubsection{Mesh painting}
To achieve a data PDF shape which is continuous and roughly resembles a Gaussian distribution, the Illustris simulations' catalogs are painted onto a mesh using a TSC (Triangular Shaped Cloud) window kernel, since most voxels are empty, this helps to "smooth" out the integer number count values especially in low density voxels which importantly helps smoothing out the individual training cubes' PDF curves and removes large discontinuities.
\subsubsection{Normalization}
The ranges of feature distributions vary a lot among different training cubes and are also dependent on the training cube's position inside the simulation. To speed up and stabilize the training process, we normalize all training cubes as we, for example, do not want to learn to predict certain features better than others.
The following simple normalization $N$ is used:
\begin{equation}
N_{\text{log1p}}(x) = \text{log}\left(\epsilon + 1 + x\right),
\end{equation}
where the not strictly necessary factor $\epsilon$ is set to $\epsilon \equiv 10^{-7}$.
\subsubsection{Cube division}
The simulations painted on 3D meshes are too big to fit in a single GPU memory all at once. We then choose to use sub-volumes with $112^3$ voxels
as the input low-res data to the networks, and $64^3$ as the target high-res data.
These include periodically padded regions of width of 24 voxels,
and require no further ``paddings'' once inside the networks.
By doing so we fully preserve the translational symmetry, which is crucial for generating outputs without artifacts near their boundaries.
Using the model described in \ref{sec:models and training}, this results in a generated output cube size of $64^3$ voxels.
\subsection{Models and training}
\label{sec:models and training}
At the $36.6$ kpc/h pixel scale, dark matter in the low and high resolution simulations are already spatially shifted from each other, due to the addition of the high-frequency modes in the high-resolution simulation.
The high density regions can shift by a few pixels.
This makes the simpler training approach of supervised learning (e.g.\ using a mean squared error loss function) impossible as the model does not have enough information
to predict this shift and will simply blur the output.
Recent works (e.g.\ on displacement fields not density fields in \cite{YinLi2021}) showed that an unsupervised approach to deep learning is a possible pathway forward given the limitation described here.
First proposed by Ian Goodfellow in 2014 \cite{Goodfellow2014} Generative Adversarial Networks (GANs) were introduced as a new
class of generative models.
GANs are able to produce output from input sampled from a random noise distribution and generally do not rely on information from the input data to make predictions.
GANs consist of two different networks that are trained in parallel to each other: the generator G and the discriminator D.
For a given set of real training samples, the generator learns to produce fake samples that look similar to the real ones, while the discriminator tries to learn to differentiate between real and fake samples. Formally (as in \cite{Goodfellow2014}): we can define a mapping to data space $G(z, \theta_g)$ where $z$ is taken from a prior, and usually random, distribution $p_z(z)$, whereas $\theta_g$ are G's learnable parameters. Similarly, we define a mapping $D(x, \theta_d)$ that outputs a single scalar representing the probability that $x$ was taken from the data distribution ($p_{\text{data}}$) and not from $p_g$ (the "fake" distribution produced by G). Training both D and G means maximizing the probability of D correctly identifying fake and real data, while simultaneously training G to minimize $\log \left(1 - D(G(z) \right) $. This means that D and G play a minmax game with value function:
\begin{equation}
\begin{aligned}
\min_G \max_D V(G,D) = \ &\mathbb{E}_{x \sim p_{\text{data}}(x)} \left[ \log D(x) \right] + \\
&\mathbb{E}_{z \sim p_z(z)} \left[ \log (1 - D(G(z))) \right],
\end{aligned}
\end{equation}
where $\mathbb{E}$ denotes the expectation value, which in practice is replaced by the
average over a batch with stochastic gradient descent.
As both D and G improve in their training, the desired end result would be that G produces fake samples that D is no longer to differentiate, thus having a 0.5 probability to guess correctly.
This state is also called Nash-equilibrium \cite{Nash1950}. Of course, the training of actual models will stop before this happens, but if the training was stable, G might be trained well enough to produce valuable results before Nash-equilibrium is reached.
Nevertheless, in this work G needs to generate output not from a randomly sampled input distribution, but from the low resolution training cubes and from white noise fields in the noise layers. In order to help the discriminator differentiate between real and fake data, the low-res input is concatenated to the generated output and high-res target before passing the data through D. The training process thus is no longer purely unsupervised, and our method is a conditional GAN, an adapted version of the standard approach first introduced by \cite{mirza2014}, which utilized a different architecture.\\
This results in the following loss functions for the discriminator and generator in this work:
\begin{align}
\begin{aligned}
\mathcal{L}_D = \ &-\mathbb{E}_{\hat{x},x} \left[ \log D(\hat{x},x) \right] \\
&- \mathbb{E}_{\hat{x}, z} \left[ \log (1 - D(\hat{x},G(\hat{x}, z))) \right] \\
&- \gamma \ \mathbb{E}_{\hat{x},x} [ \left\Vert \nabla D(\hat{x},x) \right\Vert^2 ]
\label{eqn:d_loss}
\end{aligned}
\end{align}
\begin{align}
\begin{aligned}
\mathcal{L}_G = \ &- \mathbb{E}_{\hat{x}, z} \left[ \log (D(\hat{x},G(\hat{x},z))) \right],
\end{aligned}
\end{align}\\
where $x$ is sampled from the high-res distribution $p_{\text{high-res}}$, $\hat{x}$ is sampled from the low-res distribution $p_{\text{low-res}}$ and $z$ is sampled from a random white noise distribution $p_z$. The third term in equation \ref{eqn:d_loss} is the $R_1$ regularization (introduced in \cite{Mescheder2018}), which is applied for real data only. The chosen penalty weight $\gamma=5$ is constant.
\subsubsection{Models}
\paragraph*{Discriminator:}
Training is not too sensitive on the exact discriminator architecture, as long as the main idea of progressively extracting higher-level features by down-sampling convolutions is followed. Figure~\ref{fig:discrim} depicts a qualitative approximation of the architecture as not all shapes and sizes are completely accurate. It is important to understand that each convolution changes the shape of the output feature map slightly, but since this would barely be visible in the graph it is neglected here. Additionally, in this case the boxes indicate the operator acting on the feature map of each stage. Meaning that the actual data being passed through the network follows the path indicated by the arrows and is reshaped by the operators along the way. For the exact model architecture, please refer to the data availability section.
\begin{figure}
\centering
\includegraphics[width=0.8\columnwidth]{./Images/Tex_models/D_linear_shape.pdf}
\caption{A plot showing the rather simple architecture of the discriminator used for GAN training. Of course, every stage applies 3D operations on a 3D input, but the visualization works better in 2D. The net's stages are indicated as: \textit{light orange}: a 3D convolution; \textit{red}: LeakyReLU activation; \textit{green}: Residual Block: consisting of two 3D convolutions each followed by a LeakyReLU activation, another 3D convolution and a linear down-sampling layer. At the end D outputs a $4^3$ voxel cube which is averaged over before being passed to a sigmoid inside the BCELoss function. For further details please refer to the data availability section.}
\label{fig:discrim}
\end{figure}
\paragraph*{Generator:}
The model's exact architecture is extremely important to a successful training process. The model, itself not too complicated or large, is a shallower adaptation of the often used U-Net (introduced in \cite{Ronneberger2015}, adapted in eg. \cite{Giusarma2019,He2019} and \cite{chen2020learning}). As explained in \cite{Shelhamer2017}, convolutional neural nets are built on translational equivariance since their component stages only depend on relative spatial positions as they operate on local input regions. Thus at all blocks each voxel's receptive field has the same size, independent on how large or small the initial input is. This translational equivariance is very important as it ensures that the net will produce good output on a variety of different inputs. The most important change from a conventional U-Net is that we replace the transposed upsampling convolution by a tri-linear interpolation followed by a usual convolution (detailed implementation in figure \ref{fig:gen} and the data availability section, explained further in \cite{odena2016}). This greatly reduced the checkerboard effect produced by the transposed convolutions, thus helping the training process immensely. \\
In order to help the generator predict the highly non-linear structure of the high-res target, uniformly sampled random noise is added before each convolution. This provides the basis upon which the generator can build its small scale prediction on. Similar to the hierarchical structure formation of dark matter, the generator's U-Net structure causes the layers producing small scale structure to be conditioned on the large scale features that are passed through G. This inherent similarity makes U-Nets an obvious choice for working with cosmological simulation data.
\begin{figure}
\centering
\includegraphics[width=0.8\columnwidth]{./Images/Tex_models/Unet_ushape.pdf}
\caption{The U-Net used as a generator in this work. Each stage's output's number of channels is depicted by the small number below. The spatial dimensions of the feature maps at each stage are approximately depicted by the width and height of each stage. The net's stages are indicated as (different colors as for D in figure \ref{fig:discrim}): \textit{light red}: a 3D convolution; \textit{light orange}: a noise layer followed by another 3D convolution; \textit{dark orange / brown}: LeakyReLU activation; \textit{green}: noise layer followed by a 3D (down-)convolution; \textit{blue}: noise layer followed by a linear up-sampling layer, a 3D convolution, a LeakyReLU activation, another noise layer, another 3D convolution and lastly another activation; \textit{cat}: concatenation of two outputs from two layers along the channel dimension. For more exact details please refer to the data availability section.}
\label{fig:gen}
\end{figure}
\subsubsection{Training}
\label{sec:training}
Each epoch of the training process was divided into training with the training data set, followed by validation on the entire validation data set. To increase the training set's distribution variability, we augmented the dataset by flipping and permuting axes of the training cubes. This helps the network learn discrete rotations and parity symmetries which are inherent in cosmological simulations. Such augmentation was not done on the validation set for deterministic results at every epoch.
In the first epoch of training, only the generator was trained using the standard supervised approach using MSE Loss. This helped the generator learn large scale shapes of the low-res input and high-res target which was beneficial to the later epochs of GAN training.
With the AdamW optimizer \cite{adamw2019} and small weight-decay ($\lambda = 10^{-4}$), we adopt different learning rates for the two networks: $1 \times 10^{-5}$ for the discriminator and $5 \times 10^{-5}$ for the generator, with a ratio similar to the one suggested in \cite{ttur2017}.
It is important to notice that these learning rates are not universally applicable and will need to be adjusted for different training applications to reach an equilibrium training state. Furthermore a $R_1$ regularization gradient penalty scheme \cite{Mescheder2018} was used every 16 batches to penalize the discriminator upon deviating from equilibrium on real data alone, thus increasing training stability.
\section{Results}
The sections below will summarize the quality of the network's output after training it for 600 epochs using the method described in section \ref{sec:training}. Not every epoch produces equally good results, although at later stages of training the output quality is stable. This fluctuation is due to both networks (D and G) hovering around a local loss minimum, only incrementally improving. Most statistics are more or less unaffected by this, but for example the halo two-point function for small mass halos is very sensitive to the exact output. The best performing epoch of the last 10 is selected, in this case: epoch 599. Best performing in this case refers to good accordance with all target statistics. Epoch 599 was not the only one producing good results across all statistics, but was subjectively selected for providing the best results. It should also be mentioned that extending that run further until epoch 1200 did not increase the quality of the output as the training seemed to hover around a local minimum.
The testing dataset consists, as described in section \ref{sec:dataset}, of one eighth of the entire simulation region. Thus all testing cubes combine, stitched back together, a volume of $37.5$ Mpc/h. All plots will show comparisons between the Illustris-3 low-res input, the generator's output and the high-res Illustris-2 target, labeled as \textit{low-res}, \textit{generated} and \textit{high-res}.
Some statistics (eg. two-point function, see below) require the data to be stored as particle catalogs and not as number-count-field pixel values. In order to go back to catalog format in the most simple and physical way, two steps are taken:
First, all voxels are Poisson sampled as follows:
\begin{equation}
x_\text{poisson} = \begin{cases} x &\text{if} \ r > x - \floor{x}\\
x+1 & \text{else},
\end{cases}
\end{equation}
where $r$ is uniformly drawn between $0$ and $1$. This results in the net output's voxels being converted to integers only. Secondly, if a voxel's Poisson sampled value is eg. 4, four particles are randomly placed inside the voxel's volume and depending on that voxel's relative position inside the simulation, a position-catalog entry is made for each particle.
This method obviously completely erases any information about particle positions below the pixel-size scale, but this information is already lost upon creation of the training data. Thus this kind of pixelization poses fundamental limits and challenges to using number count or density fields for this training method, as compared to e.g.\ displacement field training (which however does not apply to simulations with glass initial conditions such as Illustris here).
This effect will also become more apparent in section \ref{sec:stats}, while we look at statistical properties of the output.
\subsection{Visual Comparison}
\begin{figure*}
\includegraphics[width=\textwidth]{./Images/Projection/lowres_proj_plot_zoomzoom_all_599_boldtext.jpg}
\caption{\textit{left:} low-res catalog, \textit{middle:} generated output catalog, \textit{right:} high-res catalog. A projection on the z-axis of each catalog's particles, where particles which are part of FoF halos, are highlighted in orange. The generated output plot is indistinguishable from the high-res plot by eye on large scales. At smaller scales, the shapes and positions of small halos look statistically consistent but vary a bit between the two as expected, whereas they are completely missing in the low-res plot. The depicted testing-region box is not periodic as it only spans one eighth in volume of the entire corresponding Illustris simulation, which spans $(75 \, \textrm{Mpc}/h)^3$.
This figure was created with the help of Yueying Ni.}
\label{fig:projection}
\end{figure*}
The visual differences between the low-res input and generated output / high-res target simulation become clear immediately in figure \ref{fig:projection}, especially when looking at the smaller mass halos that are missing in the low-res simulation data.
As expected, the generated output match the high-res simulation on large scales thanks to the conditional GAN training, while on small scales the generated fine structures are different from that in the simulated ones but look statistically consistent.
For quantitative comparison, below we look at various matter and halo statistics which provide a more solid and consistent way to quantify model performances.
\subsection{Statistics}
\label{sec:stats}
There are numerous different statistical properties that can be looked at, but most of them, eg. the power spectrum, don't provide any new information, as both the low-res input (Illustris-3) and high-res target (Illustris-2) simulation are already at a quite high resolution and thus are non-differentiable at this pixel-size scale. But, crucially, some statistics do show clear differences, proving the performance of the trained network.
\subsubsection{Power spectrum and 2PCF}
Figure \ref{fig:power_2pcf} depicts the particle 3D power spectrum and particle auto-correlation function of the entire testing box, defined as in for example \cite{movandenboschwhite2010}. They are compared to the power spectrum and two-point function calculated from the "true" Illustris simulation particle catalogs over the same region. The power spectrum $P(k)$ is calculated using a FFT over a cic-painted meshgrid whereas the two-point function $\xi_2(x)$ is calculated by counting particle pairs in binned distance-regions.
\begin{align}
\xi_2(x) &\equiv \left\langle \delta(\bm{x_1})\delta(\bm{x_2}) \right\rangle, \quad x = |\bm{x_1} - \bm{x_2}|\\
\xi_2(x) &= \frac{1}{(2\pi)^3 } \int d^3k \ P(k) e^{ikx},
\end{align}
where the power $P(k)$ is simply defined as the Fourier transform of the two-point correlation function and $\delta(\bm{x_i}) = \frac{\rho(\bm{x_i})}{\bar{\rho}} - 1$ is the matter overdensity. The effect of cic-compensation (as described in eq 18 of \cite{Jing_2005}), which aims to reduce the statistical effect that painting particles onto a mesh has on the power, is negligible and does not reduce the difference between the pixelated and true power spectra.
\begin{figure*}
\centering
\includegraphics[width=0.8\textwidth]{./Images/fftpower_2pcf.pdf}
\caption{\textit{left}: The particle power spectrum comparison between the catalogs obtained from Poisson sampling the training data and the "true" (unpixelized) Illustris simulation catalogs. It is important to also compare to the "true" Illustris statistics, as this will give an idea of the effects the pixelization has on the low and high-res testing data. The divergence at small scales due this pixelation becomes clearly visible. The low-res and and high-res simulation's power agree extremely well with each other up to the Nyquist frequency, thus making the comparison futile as a way to prove the neural net's performance. \textit{right}: The particle two-point correlation function up to bin sizes of 2 Mpc/h, calculated using the Landy-Szalay estimator. As with the power spectrum, the low and high resolution simulations coincide at scales above the pixelsize of ${\sim}36.6$ kpc/h, only proving that the network can at least mirror the input roughly in that regard. Small differences between low-res input and generated output originate from the fact that the net does not simply learn to copy the low-res input. The "true" Illustris two-point correlations differ slightly from the testing data as an effect of the pixelation.
}
\label{fig:power_2pcf}
\end{figure*}
For both, power and two-point function, the output agrees very well with the low-res input / high-res target. The power differences at small physical scales between testing data / generated output and the true simulation catalogs, are caused by the pixelation and the ensuing loss of information below pixel-size scales. The autocorrelation plot only covers a smaller range of distances, thus doesn't show this diversion. Since both the power and two-point function of input and target are equal at these scales, it comes as no surprise that the output will match these statistics reasonably well. It should also be stated that even though the low-res input distribution matches the high-res target's in these two statistics, the generator does not produce them exactly, as it does not simply learn to produce a one to one mapping between input and target.
\subsubsection{Halo mass function}
All halos in figure \ref{fig:hmf} where calculated using nbodykit's \cite{nbodykit} FoF (Friends of Friends) halo finder using a relative linking length of $0.2$. Again, the pixelation effect becomes visible at smaller mass halos below and around $10^{10}$ M$_{\Sun}$, as the HMF found from pixelized data starts to deviate from the halos found using the true simulation catalog. At smaller/larger pixel sizes this effect will shift to smaller/larger halo masses.
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{./Images/hmf.pdf}
\caption{The halo mass function obtained from FoF halos (linking length $0.2$) inside the testing region. Illustris halos are taken from the "true" high- and low-res $z=0$ simulation snapshots, which importantly show the effect the randomization of particle positions inside each voxel has on FoF halo finding. This effect becomes clearly visible for halos below $10^{10}$ M$_{\Sun}$ as the output and target halo mass functions start to diverge from the "true" Illustris-2 halos, whereas the "true" Illustris-3 halos still coincide with the halos found from the low-res testing data as they are more massive and thus unaffected by the pixelation.}
\label{fig:hmf}
\end{figure}
\subsubsection{Halo two-point function}
The halo auto-correlation function is calculated by using the halo's central positions. Figure \ref{fig:halo_2pcf_1} and \ref{fig:halo_2pcf_2} shows the two point function of different halo mass bins. Importantly for halos between $1 \times 10^{10}$ M$_{\Sun}$ and $4 \times 10^{10}$ M$_{\Sun}$, a clear difference between the low resolution and high resolution simulation can be seen. The network manages to match the high resolution simulation perfectly, proving it's performance. It also matches the two-point functions for higher mass halos, which is to be expected as those halos are larger in size and easier to predict. The two-point function for halo masses below $10^{10}$ M$_{\Sun}$ becomes less meaningful, as the FoF halos found there will be affected due to the pixelation. Nevertheless the two point function of the (pixelized) generated output and high-res target still coincide with the two-point function calculated from FoF halos found by using true Illustris-2 simulation particles.
\begin{figure*}
\centering
\includegraphics[width=0.8\textwidth]{./Images/halo_2pcf_0.pdf}
\caption{The halo two-point function comparison between the testing data and the "true" un-pixelated Illustris simulation FoF halos of the testing region, calculated using the Landy-Szalay estimator. Only halos between $1 \times 10^{10}$ M$_{\Sun}$ and $6 \times 10^{10}$ M$_{\Sun}$ are plotted. The network correctly predicts the high-res target's (and also "true" Illustris-2) correlation of halo positions, which importantly differs from the low-res input's. This plot is the most critical one, as it indicates that the method learns to predict the high-resolution simulation and not simply up-scales the low-resolution density fields. It also becomes clear from the comparison to the "true" Illustris halo two-point correlation functions, that pixelation only has a minor effect on the testing data in this regard.}
\label{fig:halo_2pcf_1}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=0.8\textwidth]{./Images/halo_2pcf_1.pdf}
\caption{Halo two-point functions of higher mass bins. Here low-res input and high-res target start to coincide again so the net's performance becomes less meaningful at these mass scales.}
\label{fig:halo_2pcf_2}
\end{figure*}
\subsection{Generalizing to another simulation}
Ideally, the above model would be perfectly applicable to any low resolution DM-only simulation's density field of similar pixel size and find new halos. The extend of this performance is tested in this section by using a set of the Illustris-TNG DM-only simulations \cite{Springel_2017}, which are the latest improvement on the previous Illustris simulations. They use almost the same cosmology, but of course are run using different initial conditions. By dividing both the TNG-300-3-Dark (low-res) and TNG-300-2-Dark (high-res) simulations into $5632^3$ density voxels, a similar pixelsize as before of ${\sim}36.6$ kpc/h is fixed for the data. For memory efficiency, only an eighth of the entire simulation volume of $205^3$ Mpc/h is used for testing again. Remarkably, even though the particle masses are different compared to the Illustris simulations (high-res TNG-300-2: $m_{\text{dm}} = 5.6 \times 10^8 \ \textup{M}_{\Sun} $, low-res TNG-300-3: $m_{\text{dm}} = 4.5 \times 10^9 \ \textup{M}_{\Sun} $), the network still manages to predict new halos successfully which can again be seen in both the halo mass function and halo two-point function in figure \ref{fig:tng_halo2pcf}. This is an important indication that convolutional neural networks are, once trained successfully, even applicable outside their initial training and testing region. Of course once the pixel size differs too much from the initial pixel size the network was trained on, deteriorating results are to be expected.
\begin{figure*}
\centering
\includegraphics[width=0.8\textwidth]{./Images/TNG/TNG_hmf_halo2pcf.pdf}
\caption{\textit{left}: The halo mass function of FoF halos of the TNG testing region, similar to figure \ref{fig:hmf}. Clearly the net performs weaker than on the Illustris data, but still manages to predict new lower mass halos, compared to the low-res input (TNG-300-3) simulation. Due to higher particle masses in the simulations, only higher mass FoF halos, compared to the Illustris halo mass function, are found. This also leads to the pixelation effect in figure \ref{fig:hmf} not yet to take effect, as it only affects FoF halos below $10^{10}$ M$_{\Sun}$ at this pixelsize.
\textit{right}: Halo two-point functions of halos between $1 \times 10^{11}$ M$_{\Sun}$ and $4 \times 10^{11}$ M$_{\Sun}$. Here low-res input and high-res target differ similarly to the Illustris simulations, but at higher halo masses, since the simulation's particle masses are different. The network makes the correct prediction without ever training on the TNG data set, indicating that it is applicable to a wide range of simulations. Again it becomes clear from the comparison to the "true" TNG-300 halo two-point correlation functions, that pixelation only has a minor effect on the testing data in this regard.}
\label{fig:tng_halo2pcf}
\end{figure*}
\section{Conclusions}
We apply GAN techniques to improve cosmological simulations to successfully generate high resolution outputs from low resolution counterparts, mimicking authentic high-res simulation results in all evaluated statistics (crucially in figure \ref{fig:halo_2pcf_1}).
This work pushes the resolution limit (from for example \cite{YinLi2021_2} and \cite{doogesh19}), while also being more generally applicable to general DM-only simulations than displacement field training solutions, as it only relies on knowing the particle positions at snapshot $z=0$. It provides an important first step to building a full suite going from low-res dark matter simulations, to ultimately producing ELG mock catalogs over Gpc in scale, by predicting new halos at mass ranges relevant to ELGs.
In the case of simulations where initial conditions are set using a glass tile configuration (like the Illustris simulations) it becomes almost impossible to gain information about the dark matter displacement field, as the initial particle positions cannot easily be placed on a grid together with their IDs.
This work allows to generate high-res density fields by only investing the CPU time it takes to run the low-res simulation instead of actually running the entire high-res simulation, as testing only took around 26 CPU hours for this work.
It is also naturally scaleable to much larger testing regions by simply providing a larger low-res input.
A natural downside to using density fields for training is mainly the fact that all information below pixel-size scales is lost naturally due to pixelization. This poses natural limits to the lowest mass of physical halos to be found when training on a given pixel-size. Nonetheless this work proves that the method was successful down to scales where the low and high resolution statistics start to deviate, and suggests that the same will also hold up for training on smaller pixelsizes. This will also lead to a better resolution of halos and subsequently also subhalos (identified using density information), which will lead to the initial goal of creating large scale high resolution ELG mocks from a low resolution simulation with almost no additional computational effort.
\section{Outlook}
Future improvements of this method might make use of a smaller pixel-size, eg. ${\sim}18.3$ kpc/h, to improve smaller mass halo predictions and enable subhalo predictions. In principle not much except for the learning rates will have to be adjusted, although the amount of training data will increase significantly, making the entire training process longer and more difficult to tune.
Additionally it will also be beneficial to train the network on multiple redshifts, to make it more generally applicable to DM-only simulation output.
Recently similar work using displacement fields and super resolution methods \cite{YinLi2021, YinLi2021_2} showed that training for higher redshift data is an easier task.
This also makes intuitively sense, as $z=0$ simulations went through the maximum amount of non-linear structure development, making it harder to learn the mapping accurately.
Another useful improvement to the training would be to add velocity fields to the input, containing information about the particles' bulk velocities inside each pixel-volume, helping to infer halo field phase-space distributions. The networks architecture supports training on multiple fields natively, so this change could be easily executed.
It will also be interesting to provide the net with completely different simulations, with different cosmological parameters and at similar mass resolutions and pixel-scales during training, to gain a more universally applicable network.
As we ultimately want to produce data comparable to observations, one will also have to infer redshift-space distributions from our real-space data, which will be subject of future work.
With the end goal in mind of creating a network that does not have to be retrained and can easily enhance low resolution simulation density output, this work provides an important foundation on the way to easily accessible large scale ELG mock catalogs.
\section*{Acknowledgements}
We thank Yueying Ni for sharing her plotting scripts.
This work was supported in part through the NYU IT High Performance Computing resources, services, and staff expertise. We would also like to thank AMD for the donation of critical hardware and support resources from its HPC Fund, that also made this work possible. The Flatiron Institute is supported by the Simons Foundation. JLT acknowledges support of NSF grant number 2009291.
\section*{Data Availability}
All code and models used and created in this work are publicly available on GitHub under the following url: \url{https://github.com/dschaurecker/dl_halo}. Please refer to the ReadMe file in the \texttt{dl\_halo} repository for a very in-depth guide through the code and the process of training and testing.
The training code builds upon the \texttt{map2map} code repository, available on GitHub as well (\url{https://github.com/eelregit/map2map}).
It allows for general training of $n$ arbitrary input fields to $n$ arbitrary output fields using custom models, normalizations and loss functions.
Furthermore some utilities from \texttt{nbodykit} an "Open-source, Massively Parallel Toolkit for Large-scale Structure" \cite{nbodykit} were used in pre-processing and statistical evaluation of this work.
\section{Some examples and best-practices}
\label{sec:intro}
Here follow some examples of common features that you may wanto to use
or build upon.
For internal references use label-refs: see section~\ref{sec:intro}.
Bibliographic citations can be done with cite: refs.~\cite{a,b,c}.
When possible, align equations on the equal sign. The package
\texttt{amsmath} is already loaded. See \eqref{eq:x}.
\begin{equation}
\label{eq:x}
\begin{split}
x &= 1 \,,
\qquad
y = 2 \,,
\\
z &= 3 \,.
\end{split}
\end{equation}
Also, watch out for the punctuation at the end of the equations.
If you want some equations without the tag (number), please use the available
starred-environments. For example:
\begin{equation*}
x = 1
\end{equation*}
The amsmath package has many features. For example, you can use use
\texttt{subequations} environment:
\begin{subequations}\label{eq:y}
\begin{align}
\label{eq:y:1}
a & = 1
\\
\label{eq:y:2}
b & = 2
\end{align}
and it will continue to operate across the text also.
\begin{equation}
\label{eq:y:3}
c = 3
\end{equation}
\end{subequations}
The references will work as you'd expect: \eqref{eq:y:1},
\eqref{eq:y:2} and \eqref{eq:y:3} are all part of \eqref{eq:y}.
A similar solution is available for figures via the \texttt{subfigure}
package (not loaded by default and not shown here).
All figures and tables should be referenced in the text and should be
placed at the top of the page where they are first cited or in
subsequent pages. Positioning them in the source file
after the paragraph where you first reference them usually yield good
results. See figure~\ref{fig:i} and table~\ref{tab:i}.
\begin{figure}[tbp]
\centering
\includegraphics[width=.45\textwidth,trim=0 380 0 200,clip]{example-image}
\hfill
\includegraphics[width=.45\textwidth,angle=180]{example-image}
\caption{\label{fig:i} Always give a caption.}
\end{figure}
\begin{table}[tbp]
\centering
\begin{tabular}{|lr|c|}
\hline
x&y&x and y\\
\hline
a & b & a and b\\
1 & 2 & 1 and 2\\
$\alpha$ & $\beta$ & $\alpha$ and $\beta$\\
\hline
\end{tabular}
\caption{\label{tab:i} We prefer to have borders around the tables.}
\end{table}
We discourage the use of inline figures (wrapfigure), as they may be
difficult to position if the page layout changes.
We suggest not to abbreviate: ``section'', ``appendix'', ``figure''
and ``table'', but ``eq.'' and ``ref.'' are welcome. Also, please do
not use \texttt{\textbackslash emph} or \texttt{\textbackslash it} for
latin abbreviaitons: i.e., et al., e.g., vs., etc.
\section{Sections}
\subsection{And subsequent}
\subsubsection{Sub-sections}
\paragraph{Up to paragraphs.} We find that having more levels usually
reduces the clarity of the article. Also, we strongly discourage the
use of non-numbered sections (e.g.~\texttt{\textbackslash
subsubsection*}). Please also see the use of
``\texttt{\textbackslash texorpdfstring\{\}\{\}}'' to avoid warnings
from the hyperref package when you have math in the section titles
|
\section{Application: Transference for removal lemmas}
\label{sec:application}
Hypergraph removal lemma states that for every $\alpha > 0$ there exists $\gamma > 0$ such that any $k$-graph with $n$ vertices and at most $\gamma n^{v(H)}$ copies of some fixed $k$-graph $H$ can be made $H$-free by removing at most $\alpha n^k$ edges. This innocent looking statement has striking implications, the most prominent being Szemer\'edi's theorem \cite{szemeredi1975sets} on arithmetic progressions in dense sets of natural numbers. For the history of the lemma, its applications and recent developments, we refer the reader to the survey by Conlon and Fox \cite{conlon13removalsurvey}.
As an application of Theorem \ref{thm:containers}, and a demonstration that it is not hindered by the existence of a small number of subsets for which neither \ref{prop:i} nor \ref{prop:ii} apply, we prove the following version of a removal lemma for random hypergraphs. To state it concisely, let us denote with $\mathcal{G}^{(k)}(n,p)$ a subgraph of $\ensuremath{\mathcal K}_n^{(k)}$, the complete $k$-graph with $n$ vertices, obtained by taking each hyperedge with probability $p$, independently of all other edges, for some $p \in (0, 1]$.
\begin{theorem} \label{thm:random_removal}
For any $k$-graph $H$ and $\alpha > 0$ there exists $C, \gamma > 0$ such that if $p \ge Cn^{-1/m_k(H)}$, where
$$
m_k(H) = \max\left\{ \frac{e(H') - 1}{v(H') - k} \colon H' \subseteq H, v(H') > k \right\},
$$
then with high probability $\Gamma = \mathcal{G}^{(k)}(n,p)$ has the following property: Every subgraph $\Gamma' \subseteq \Gamma$ with at most $\gamma n^{v(H)} p^{e(H)}$ copies of $H$ can be made $H$-free by removing at most $\alpha n^{k} p$ edges.
\end{theorem}
It should be noted that Theorem \ref{thm:random_removal} does not have nearly as striking implications as the original hypergraph removal lemma. Nonetheless, it is an interesting statement which in some way quantifies the distribution of the copies of $H$ in $\mathcal{G}^{(k)}(n,p)$. The graph case ($k=2$) of Theorem \ref{thm:random_removal} was originally obtained by Conlon, Gowers, Samotij, and Schacht \cite{conlon2015klr} using an adaption of the regularity method for random graphs. Conlon and Gowers \cite{conlon2016combinatorial} have proved the general case under the assumption that $m_k(H)$ is obtained uniquely for $H' = H$, that is, when $H$ is \emph{strictly balanced}. Here we prove the statement without any additional assumption.
It should be noted that the proof of the removal lemma for random graphs using the regularity method very closely follows the original proof of the regularity lemma for graphs. In contrast, the transference result presented here as well as the one by Conlon and Gowers \cite{conlon2016combinatorial} uses the the hypergraph removal lemma, a notoriously difficult result, as a black box. That being said, instead of directly proving Theorem \ref{thm:random_removal} we prove a more general result which abstracts out the removal property, from which Theorem \ref{thm:random_removal} then follows as a straightforward corollary.
Given $s, r \in \mathbb{N}$, we say that a hypergraph $\ensuremath{\mathcal H}$ is \emph{$(s, r)$-removable} if for every $I \subseteq V(\ensuremath{\mathcal H})$ with $e(I) < s$ there exists $X \subseteq I$ of size $|X| \le r$ such that $I \setminus X$ is an independent set in $\ensuremath{\mathcal H}$.
\begin{theorem} \label{thm:removal_transference}
For every $k \in \mathbb{N}$ and $B, \alpha > 0$ there exists $\gamma_0, C > 0$ such that the following holds. Let $\ensuremath{\mathcal H}$ be a $(\gamma e(\ensuremath{\mathcal H}), \alpha N)$-removable $k$-graph, for some $\gamma \le \gamma_0$, and suppose $m_0 \in \mathbb{N}$ is such that
$$
\Delta_{\ensuremath{\mathcal H}}(m_0) \le B \mu_{\ensuremath{\mathcal H}}(m_0)^2/m_0.
$$
Let $W \subseteq V(\ensuremath{\mathcal H})$ be a random subset formed by taking each element with probability $p \ge C m_0 / N$, independently of all other elements. Then with probability at least $1 - \exp(-\Theta(Np))$ the induced hypergraph $\ensuremath{\mathcal H}[W]$ is $(\gamma \alpha^k e(\ensuremath{\mathcal H}) p^k, 3 \alpha N p)$-removable.
\end{theorem}
\begin{proof}
Let $\gamma_0 > 0$ be as given by Theorem \ref{thm:containers} for some sufficiently small $\varepsilon = \varepsilon(\alpha)$. Suppose $\ensuremath{\mathcal H}$ is $(\gamma e(\ensuremath{\mathcal H}), \alpha N)$-removable for some $0 < \gamma \le \gamma_0$.
By Chernoff's inequality, we have $Np/2 < |W| < 2 Np$ with probability $1 - \exp(-\Theta(Np))$. Moreover, a simple union-bound gives the following upper bound on the probability that an $m$-element subset excluded by Theorem \ref{thm:containers}, for $\alpha Np \le m < 2 Np$, is a subset of $W$:
$$
\sum_{m = \alpha Np}^{2Np} \varepsilon^m \binom{N}{m} p^m < N \left( \frac{\varepsilon e N}{\alpha N p} p \right)^m < e^{-N p},
$$
for sufficiently small $\varepsilon = \varepsilon(\alpha)$. Therefore, from now on we can assume that for every $m$-element subset $I \subseteq W$ with
\begin{equation} \label{eq:bound_I_gamma0}
e(I) < \gamma_0 \alpha^k e(\ensuremath{\mathcal H})p^k \le \gamma_0 e(\ensuremath{\mathcal H})(m/N)^k,
\end{equation}
for $\alpha Np \le m < 2Np$, the property \ref{prop:ii} of Theorem \ref{thm:containers} holds.
Suppose there exists $I \subseteq W$ which satisfies \eqref{eq:bound_I_gamma0} such that $I \setminus X$ is not independent for every $X \subseteq I$ of size $|X| \le 3 \alpha Np$. In particular, this implies $|I| = m > 3 \alpha Np$. Then, by \ref{prop:ii}, $I \subseteq F \cup f_m(\mathbf{F})$ for some $F \subseteq I$ of size $|F| = \varepsilon m$ and $\mathbf{F} \in \ensuremath{\mathcal T}_{k-1}(F)$. As $\ensuremath{\mathcal H}$ is $(\gamma e(\ensuremath{\mathcal H}), \alpha N)$-removable and $e(f_m(\mathbf{F})) < \gamma e(\ensuremath{\mathcal H})$, there exists a subset $R(\mathbf{F}) \subset V(\ensuremath{\mathcal H})$ of size $|R(\mathbf{F})| \le \alpha N$ such that $f_m(\mathbf{F}) \setminus R(\mathbf{F})$ is an independent set. If $|W \cap R(\mathbf{F})| < 2 \alpha N p$ then we can make $I$ to be an independent set by removing $F \cup (W \cap R(\mathbf{F}))$ which, for $\varepsilon < \alpha / 2$, is of size at most $3\alpha N p$. Therefore, we must have $|W \cap R(\mathbf{F})| > 2 \alpha N p$.
To summarise, to show the theorem it suffices to bound the probability of an event that $|W \cap R(\mathbf{F})| \ge 2 \alpha N p$ for some $\mathbf{F} \in \ensuremath{\mathcal T}_{k-1}(F)$ where $F \subseteq W$ is of size $|F| = s$ with $\varepsilon \alpha Np \le s \le 2 \varepsilon Np$. The probability of this happening for one particular $\mathbf{F}$ is, by Chernoff's inequality, at most $e^{-\alpha N p}$, thus a union-bound over all possible $\mathbf{F}$ gives the following:
$$
\sum_{s = \alpha \varepsilon Np}^{2 \varepsilon Np} \binom{N}{s} 2^{ks} \cdot p^s \cdot e^{-\alpha N p} \le e^{-\alpha Np} \sum_{s = \alpha \varepsilon Np}^{2 \varepsilon Np} \left( \frac{e 2^k N p}{s} \right)^s \le e^{-\alpha Np} \cdot N \left( \frac{e2^k Np}{2\varepsilon Np} \right)^{2\varepsilon Np} = e^{-\Theta(Np)},
$$
for $\varepsilon$ sufficiently small with respect to $\alpha$. This concludes the proof.
\end{proof}
Note that the proof works just the same using the hypergraph containers theorem of Saxton and Thomason \cite{saxton2015containers}. For comparison, a variant of Balogh, Morris, and Samotij \cite{balogh2015indset} does not suffice as it only gives $e(I) > 0$ in \ref{prop:i}. Theorem \ref{thm:random_removal} now follows as a straightforward application of Theorem \ref{thm:removal_transference} and the hypergraph removal lemma.
\begin{proof}[Proof of Theorem \ref{thm:random_removal}]
Let $\ensuremath{\mathcal H}$ be an $e(H)$-graph whose vertices correspond to edges in $\ensuremath{\mathcal K}_n^{(k)}$ and $e(H)$ vertices in $\ensuremath{\mathcal H}$ form an edge if the corresponding hyperedges in $\ensuremath{\mathcal K}_n^{(k)}$ form a copy of $H$. By the hypergraph removal lemma, for a given $\alpha > 0$ there exists $\gamma_r > 0$ such that $\ensuremath{\mathcal H}$ is $(\gamma_r e(\ensuremath{\mathcal H}), \alpha N)$-removable. Furthermore, it is well known (and easy to verify) that $\Delta_{\ensuremath{\mathcal H}}(m_0)$ satisfies the condition of Theorem \ref{thm:removal_transference} for $m_0 = n^{k - 1/m_k(H)}$ and some $B > 0$. The theorem now follows for $\gamma = \min\{\gamma_r, \gamma_0\}$, where $\gamma_0$ is as given by Theorem \ref{thm:removal_transference} for $\alpha$ and $B$.
\end{proof}
\section{Concluding remarks} \label{sec:concluding}
\section{Introduction}
Let $\ensuremath{\mathcal H}$ be a $k$-uniform hypergraph, a \emph{$k$-graph} for short, for some $k \in \mathbb{N}$. Throughout the paper we use $N$ to denote the number of vertices in $\ensuremath{\mathcal H}$. What is the probability that an $m$-element subset $I \subseteq V(\ensuremath{\mathcal H})$ chosen uniformly at random is an independent set in $\ensuremath{\mathcal H}$? This question is addressed by Janson's inequality:
\begin{equation} \label{eq:janson}
\Pr[I \text{ is independent}] \le e^{- \mu_{\ensuremath{\mathcal H}}(m)^2 / (2\hat \Delta_{\ensuremath{\mathcal H}}(m))},
\end{equation}
where
$$
\mu_{\ensuremath{\mathcal H}}(m) = e(\ensuremath{\mathcal H}) (m / N)^{k}
$$
roughly corresponds to the expected number of edges in $\ensuremath{\mathcal H}[I]$, $\hat \Delta_{\ensuremath{\mathcal H}}(m) = \mu_{\ensuremath{\mathcal H}}(m) + \Delta_{\ensuremath{\mathcal H}}(m)$, and
\begin{equation} \label{eq:delta_def}
\Delta_{\ensuremath{\mathcal H}}(m) = \sum_{e \sim e'} (m/N)^{|e \cup e'|},
\end{equation}
where the sum goes over all pairs $(e, e')$ of edges of $\ensuremath{\mathcal H}$ such that $e \cap e' \neq \emptyset$ and $e \neq e'$. Note that the standard version of Janson's inequality is stated for \emph{binomial} random subsets, that is, when $I$ is formed by taking each element with some probability $p$, independently of all other elements (e.g.\ see \cite{alon_spencer}). The presented statement easily follows by standard concentration bounds, and we refer the reader to \cite[Lemma 5.2]{alon14sumfree} for details.
Consider the following two examples of $3$-graphs, $\ensuremath{\mathcal T}_n$ and $\ensuremath{\mathcal A}_n$. The vertices of $\ensuremath{\mathcal T}_n$ correspond to the edges of $K_n$, a complete graph with $n$ vertices, and three vertices form a hyperedge if the corresponding edges in $K_n$ form a triangle. The vertices of $\ensuremath{\mathcal A}_n$ are integers $\{1, \ldots, n\}$, and three vertices (that is, numbers) form a hyperedge if they form a $3$-term arithmetic progression. It is an easy exercise to show that in both cases we have $\Delta_{\ensuremath{\mathcal H}}(m) = \Theta(\mu_{\ensuremath{\mathcal H}}(m)^2 / m)$ for $m \ge \sqrt{N}$. Thus in both cases Janson's inequality bounds the probability of a random $m$-element subset being independent by $e^{-\Theta(m)}$. In the case where $\ensuremath{\mathcal H} = \ensuremath{\mathcal T}_n$ this exponent is correct, however for $\ensuremath{\mathcal H} = \ensuremath{\mathcal A}_n$ it drifts away from the truth as $m$ moves away from $\sqrt{N}$. In particular, it was shown in \cite{balogh2015indset} that the probability of a randomly chosen $m$-element set being independent in $\ensuremath{\mathcal A}_n$ decays as $e^{-\omega(m)}$ for $m = \omega(\sqrt{N})$ (see \cite{nenadov2021small} for a self-contained proof of this, and \cite{balogh2017kap} for further refinements).
The previous examples show that we cannot hope for a better (smaller) probability of the sampled set being independent than what is given by \eqref{eq:janson}, at least not based solely on parameters $\mu$ and $\Delta$. These two examples also hint that the probability of sampling an independent set, at least in the range for which \eqref{eq:janson} gives $e^{-\Theta(m)}$, is governed by the size of the largest independent set in $\ensuremath{\mathcal H}$. The size of the largest independent set certainly provides a lower bound on the probability of sampling an independent set, and in these two examples this bound is actually correct: the largest independent set is of size roughly $N/2$ in the case of $\ensuremath{\mathcal T}_n$ (Mantel's theorem) and $o(N)$ in the case of $\ensuremath{\mathcal A}_n$ (Roth's theorem), which implies that a sampled set is a subset of such (fixed) independent sets with probability $e^{-\Theta(m)}$ and $e^{-\omega(m)}$ respectively.
Our main result formalises this intuition. It states that with overwhelmingly high probability a sampled set $I$ is either not independent or is a subset of an (almost) independent set in $\ensuremath{\mathcal H}$. This would be rather trivial -- after all, if $I$ is an independent set then $I \subseteq I$ makes the previous statement vacuously hold -- if it was not for the additional fact that such an almost-independent set can be constructed by looking only at some carefully chosen $o(m)$ vertices of $I$. Consequently, the previously discussed lower bound coming from the probability that we sample a subset of a fixed independent set in $\ensuremath{\mathcal H}$ is almost correct: instead of being a subset of one fixed independent set, our result implies that it is at most the probability that the sampled set is a subset of one of the few carefully selected almost-independent sets.
To state the result concisely we need a few more definitions. Given a subset $V' \subseteq V(\ensuremath{\mathcal H})$, we use $e(V')$ as shorthand for $e(\ensuremath{\mathcal H}[V'])$, the number of hyperedges in the subgraph of $\ensuremath{\mathcal H}$ induced by $V'$. Given $\varepsilon > 0$, denote with $\ensuremath{\mathcal I}_{\varepsilon}(\ensuremath{\mathcal H})$ the family of all subsets $S \subseteq V(\ensuremath{\mathcal H})$ with $e(S) \le \varepsilon e(\ensuremath{\mathcal H})$. Finally, given a set $F$ and $k \in \mathbb{N}$, let $\ensuremath{\mathcal T}_k(F)$ denote the family of all $k$-tuples $(F_1, \ldots, F_k)$ of subsets of $F$, not necessarily disjoint.
\begin{theorem}[Probabilistic Hypergraph Containers] \label{thm:containers}
For every $k \in \mathbb{N}$ and $\varepsilon, B > 0$, there exists $\gamma, T > 0$ such that the following holds. Let $\ensuremath{\mathcal H}$ be an $N$-vertex $k$-graph, and suppose $m_0 \in \mathbb{N}$ is such that
\begin{equation} \label{eq:delta_m0}
\Delta_{\ensuremath{\mathcal H}}(m_0) \le B \mu_\ensuremath{\mathcal H}^2(m_0) / m_0.
\end{equation}
Then for every $m \ge Tm_0$, there exists a function $f_m \colon \ensuremath{\mathcal T}_{k-1}(V(\ensuremath{\mathcal H})) \rightarrow \ensuremath{\mathcal I}_\varepsilon(\ensuremath{\mathcal H})$ such that an $m$-element $I \subseteq V(\ensuremath{\mathcal H})$, chosen uniformly at random, with probability at least $1 - \varepsilon^m$ satisfies at least one of the following properties:
\begin{enumerate}[(P1)]
\item \label{prop:i} $e(I) \ge \gamma e(\ensuremath{\mathcal H}) (m/N)^k$, or
\item \label{prop:ii} There exists $F \subseteq I$, $|F| = \varepsilon m$, such that
$$
I \subseteq F \cup f_m(\mathbf{F})
$$
for some $\mathbf{F} \in \ensuremath{\mathcal T}_{k-1}(F)$.
\end{enumerate}
\end{theorem}
For $m \ge m_0$, where $m_0$ satisfies \eqref{eq:delta_m0}, Janson's inequality implies that $I$ is not an independent set with probability at least $1 - e^{-\Theta(m)}$. Theorem \ref{thm:containers}, in comparison, states that with probability superexponentially close to $1$, either $I$ is not independent or $I \setminus F$ is contained in one of at most $\binom{N}{\varepsilon m} 2^{k \cdot \varepsilon m}$ sparse subsets of $\ensuremath{\mathcal H}$, completely determined by $F$ (that is, a $(k-1)$-tuple stemming from $F$). In other words, by looking at an arbitrarily small fraction of the elements of $I$ we know the likely structure of the remaining $\ensuremath{\mathcal H}[I \setminus F]$.
The power of \ref{prop:ii} is best demonstrated in an example: If $\ensuremath{\mathcal H}$ is such that $e(S) < \varepsilon e(\ensuremath{\mathcal H})$ implies $|S| \le \beta v(\ensuremath{\mathcal H})$, where $\beta \rightarrow 0$ as $\varepsilon \rightarrow 0$, then $I$ is an independent set with probability $e^{-\omega(m)}$ as $\varepsilon \rightarrow 0$. This is, for example, the case for $\ensuremath{\mathcal H} = \ensuremath{\mathcal A}_n$ due to Roth's theorem, and for hypergraphs constructed from longer arithmetic progressions due to Szemer\'edi's theorem.
A reader familiar with recent developments in extremal and probabilistic combinatorics will notice that Theorem \ref{thm:containers} is, in some sense, a weaker version of the result of Balogh, Morris, and Samotij \cite{balogh2015indset} and Saxton and Thomason \cite{saxton2015containers}, colloquially called \emph{hypergraph container theorems}. In particular, both of these results state that the described property holds with probability exactly $1$, that is, for every $I$ which does not satisfy \ref{prop:i} the part \ref{prop:ii} holds. It is worth noting that: (i) Theorem \ref{thm:containers} is stronger than the main result from \cite{balogh2015indset} in the sense that the latter considers only $e(I) > 0$ in \ref{prop:i}, and (ii) the bound on $m_0$ here is arguably more natural (and potentially lower) than the one from \cite{balogh2015indset,saxton2015containers}, at least having Janson's inequality in mind. One could also argue that allowing for a super-exponentially small failure probability is a natural feature of the probabilistic treatment of these theorems.
Hypergraph container theorems and the method that has developed along the way can safely be considered one of the major breakthroughs in probabilistic combinatorics in the last decades. With very little effort they imply almost all known extremal results in random graphs, some of which were originally proven in breakthroughs by Schacht \cite{schacht2016extremal} and Conlon and Gowers \cite{conlon2016combinatorial}. All of these results also follow from Theorem \ref{thm:containers}, and we demonstrate its use to prove one such new result in Section \ref{sec:application}. Many counting results, such as the celebrated K\L R conjecture \cite{kohayakawa1997k}, can also be derived from Theorem \ref{thm:containers} in exactly the same way as using the original container theorems. However, two prominent examples where Theorem \ref{thm:containers} does not suffice are the results of Balogh and Solymosi \cite{balogh18solymosi} and Ferber, McKinley, and Samotij \cite{ferber20super}. Both of these results require iterative application of the containers, for which Theorem \ref{thm:containers} is not suited. For a thorough introduction to the method, we refer the reader to a survey by Balogh, Morris, and Samotij \cite{balogh2018survey}.
The proofs from \cite{balogh2015indset} and \cite{saxton2015containers} are roughly along the same lines and differ mainly in the analysis of an otherwise very similar algorithm for finding a subset $F \subseteq I$ and constructing containers. A few other proofs have been obtained since then, including the recent work by Balogh and Samotij \cite{balogh2020efficient} which provides almost optimal dependency of parameters, a short proof by Bernshteyn, Delcourt, Towsner, and Tserunyan \cite{bernshteyn2019short}, and another simple proof by Saxton and Thomason \cite{saxton2016simple} of a variant which only applies to linear hypergraphs. The main value of our new proof of (a variant of) container theorems lies in the simplicity and transparency of the ideas which, in our view, exploit the very essence of why the existence of the containers (i.e.\ the existence of $F$ and $f$ satisfying \ref{prop:ii}) is not surprising: The bound on $\Delta_{\ensuremath{\mathcal H}}(m_0)$ implies a `typical' set $I$ of size $m \ge m_0$ has the property that $\Theta(N)$ vertices in $V(\ensuremath{\mathcal H})$ form an edge with some $k-1$ vertices from $I$. This is also inherited by a `typical' subset $F \subseteq I$ of size, say, $|F| = \varepsilon m$. If $I$ is an independent set then $I \setminus F$ has to avoid all the vertices which form an edge with some $k-1$ vertices from $F$, thus significantly limiting where $I \setminus F$ has to live. Repeating a similar argument, we eventually narrow down the choices to a subset of vertices which induce very few edges.
\section{Preliminaries}
\begin{lemma} \label{lemma:deletion_general}
Let $Y$ be a set with $N$ elements, and $\ensuremath{\mathcal S}$ be a family of $s$-element subsets of $Y$, for some integers $s$ and $N$. Let $k < N/4$ be an integer and $T > 1$. Then, for all but at most
$$
T^{-k/s} \binom{N}{m}
$$
subsets $Y_m \subseteq Y$ of size $m$, there exists a subset $X \subseteq Y_m$ of size $|X| \le k$ such that $Y_m \setminus X$ contains at most
$$
2^s T |\ensuremath{\mathcal S}| (m/N)^{s}
$$
sets from $\ensuremath{\mathcal S}$.
\end{lemma}
\section{Proof of the probabilistic containers}
Given subsets $D, W \subseteq V(\ensuremath{\mathcal H})$ and $k' \in [k-1]$, let $e_{k'}(D, W)$ denote the number of edges in $\ensuremath{\mathcal H}[D \cup W]$ which intersect $W$ in at least $k'$ vertices. Similarly, for a vertex $v \in D$ let $\deg_{k'}(v, D, W)$ denote the number of edges in $\ensuremath{\mathcal H}[D \cup W]$ which contain $v$ and intersect $W \setminus \{v\}$ in at least $k'$ vertices.
\begin{definition} \label{def:saturating}
Let $\ensuremath{\mathcal H}$ be a $k$-graph and $k' \in [k-1]$. A subset $W \subseteq V(\ensuremath{\mathcal H})$ is \emph{$(k', \alpha, t)$-saturating} for $D \subseteq V(\ensuremath{\mathcal H})$, for some $\alpha, t > 0$, if
\begin{equation} \label{eq:saturated_def}
\left| \left\{ v \in D \colon \deg_{k'}(v, D, W) \ge t / N \right\} \right| \ge \alpha N.
\end{equation}
\end{definition}
The following lemma is the main building block in the proof of Theorem \ref{thm:containers}. We remark it is crucial that $\alpha$ does not depend on $\xi$.
\begin{lemma} \label{lemma:saturating}
For every $k \in \mathbb{N}$ and $\xi, \gamma, \beta, B > 0$, there exists $\alpha = \alpha(k, \gamma, \beta, B), \lambda, T > 0$ such that the following holds. Let $\ensuremath{\mathcal H}$ be a $k$-graph and $m_0 \in \mathbb{N}$ be such that
$$
\Delta_{\ensuremath{\mathcal H}}(m_0) \le B \mu_{\ensuremath{\mathcal H}}^2(m_0)/m_0.
$$
Let $m \ge T m_0$ and $k' \in [k-1]$. Then there exists a family $\ensuremath{\mathcal B}_m^{k'}$ of $m$-element subset of $V(\ensuremath{\mathcal H})$, $|\ensuremath{\mathcal B}_m^{k'}| < \beta^m \binom{N}{m}$, such that every $m$-element $I \subseteq V(\ensuremath{\mathcal H})$ which does not belong to $\ensuremath{\mathcal B}_m^{k'}$ has the following property: There exists $X \subseteq I$ of size $|X| \le \beta m$ such that if some $I' \subseteq I \setminus X$, $|I'| \ge m/2$, and $D \subseteq V(\ensuremath{\mathcal H})$ satisfy
\begin{equation} \label{eq:many_edges}
e_{k'}(D, I') \ge \gamma e(\ensuremath{\mathcal H}) q^{k'},
\end{equation}
where $q = m / N$, then $I'$ contains a $(k', \alpha, \lambda e(\ensuremath{\mathcal H}) q^{k'})$-saturating set $W$ for $D$ of size $|W| \le \xi m$.
\end{lemma}
Let us give a brief intuition behind the statement of Lemma \ref{lemma:saturating}. First, if $D = V(\ensuremath{\mathcal H})$ and $I'$ is chosen uniformly at random, then $e(\ensuremath{\mathcal H}) q^{k'}$ roughly corresponds to the expected number of edges which intersect $I'$ in at least $k'$ elements. If \eqref{eq:many_edges} holds and the edges intersecting $I'$ are nicely distributed, then we expect many vertices $v \in D$ to satisfy the degree condition in \eqref{eq:saturated_def} with $t = \Omega(e(\ensuremath{\mathcal H}) q^{k'})$ and $I'$ having the role of $W$. By choosing $W \subseteq I'$ uniformly at random we can hope that it satisfies a scaled down version of \eqref{eq:many_edges} and inherits the distribution of the edges intersecting it. Consequently, $W$ is saturating for $D$. The key part of the lemma is that properties which are sufficient to guarantee such a distribution of edges hold for all but at most $\beta^m$-fraction of $m$-element subsets. The proof relies on Chebyshev's and Paley-Zygmund inequalities, and the R\"odl--Ruci\'nski deletion method. We postpone it for the next section, and instead now prove Theorem \ref{thm:containers}.
\begin{proof}[Proof of Theorem \ref{thm:containers}]
Let us start by fixing the constants. Set $\beta = \varepsilon / (2k^2)$, $\lambda_0 = \varepsilon$, and for each $k' \in [k]$ set\footnote{For $k' = k$ we only set $\gamma_{k}$.}
$$
\gamma_{k'} = \lambda_{k'-1} \varepsilon / (2k^2), \;
\alpha_{k'} = \alpha_{\ref{lemma:saturating}}(\gamma_{k'}, \beta), \;
\xi_{k'} = \alpha_{k'} \varepsilon / (4k^2), \;
\lambda_{k'} = \lambda_{\ref{lemma:saturating}}(\xi_{k'}, \gamma_{k'}, \beta). \;
$$
Note that it is crucial here that $\alpha$ in Lemma \ref{lemma:saturating} does not depend on $\xi$, as otherwise we would get a circular dependency. Finally, set $T = \max_{k' \in [k-1]} T_{\ref{lemma:saturating}}(\xi_{k'}, \gamma_{k'}, \beta)$. We prove the statement for $\gamma = \gamma_k$. Throughout the proof we use $q = m / N$.
For each $k' \in [k-1]$, let $\ensuremath{\mathcal B}_m^{k'}$ be the family of $m$-element subsets given by Lemma \ref{lemma:saturating} for $k'$, $\xi_{k'}$ (as $\xi$), $\gamma_{k'}$ (as $\gamma$) and $\beta$. Take $\ensuremath{\mathcal B}_m$ to be the collection of all the sets from these families. Then $|\ensuremath{\mathcal B}_m| < \varepsilon^m \binom{v(\ensuremath{\mathcal H})}{m}$, with room to spare. We show that the statement of Theorem \ref{thm:containers} holds for all $m$-element sets not in $\ensuremath{\mathcal B}_m$.
Let us start with the function $f_m$. Given a $(k-1)$-tuple $\mathbf{F} = (F_1, \ldots, F_{k-1})$ of subsets of $V(\ensuremath{\mathcal H})$, set $D_k := V(\ensuremath{\mathcal H})$ and iteratively define $D_{k'}$ for $k' = k-1, \ldots, 0$ as follows:
\begin{equation} \label{eq:D_k}
D_{k'} = \left\{v \in D_{k' + 1} \colon \deg_{k'}(v, D_{k' + 1}, F_{k'}) < t_{k'} / v(\ensuremath{\mathcal H}) \right\},
\end{equation}
where
$$
t_{k'} = \lambda_{k'} e(\ensuremath{\mathcal H}) q^{k'}.
$$
In the case $k'=0$ we slightly abuse the notation by using non-defined $F_0$, which is actually irrelevant for the definition of $D_0$. Set $f_m(\mathbf{F}) = D_0$ and note that $e(D_0) \le \lambda_0 e(\ensuremath{\mathcal H}) = \varepsilon e(\ensuremath{\mathcal H})$ thus $f_m(\mathbf{F}) \in \ensuremath{\mathcal I}_\varepsilon(\ensuremath{\mathcal H})$, as required.
Consider some $m$-element $I \not \in \ensuremath{\mathcal B}_m$, and suppose \ref{prop:i} does not hold. Let $X \subseteq I$ be the union of the sets $X$ promised by Lemma \ref{lemma:saturating} for all $k' \in [k-1]$, with parameters associated to each $k'$ as stated above. Then $|X| < \varepsilon m/ (2k)$ by the choice of $\beta$. Set $I' = I \setminus X$. Next, we show there exist subsets $F_{k-1}, \ldots, F_1 \subseteq I'$, each of size $|F_i| \le \varepsilon m / (2k)$, such that
\begin{equation} \label{eq:D_k_upper}
e_{k'}(D_{k'}, I') < \gamma_{k'} e(\ensuremath{\mathcal H}) q^{k'}
\end{equation}
for every $k' \in \{k-1, \ldots, 1\}$, where $D_{k'}$ is as defined in \eqref{eq:D_k} with respect to these $F_i$'s. As $I$ does not satisfy \ref{prop:i} we also have $e_k(D_k, I') = e(I') < \gamma_k e(\ensuremath{\mathcal H}) q^k$.
We find such sets iteratively for $k' \in \{k-1, \ldots, 1\}$: Set $F_{k'} = \emptyset$ and $D_{k'} = D_{k'+1}$, where $D_{k'+1} = V(\ensuremath{\mathcal H})$ if $k'+1=k$ and otherwise it is given by \eqref{eq:D_k} with respect to already defined $F_{k-1}, \ldots, F_{k'+1}$. As long as \eqref{eq:D_k_upper} is not satisfied pick a $(k', \alpha_{k'},t_{k'})$-saturating set $W \subseteq I'$ for $D_{k'}$ of size $|W| \le \xi_{k'} m$, and update $F_{k'} = F_{k'} \cup W$ and $D_{k'}$, again as defined in \eqref{eq:D_k} (with respect to the new $F_{k'}$). As in each step the set $D_{k'}$ decreases by at least $\alpha_{k'} N$, this procedure is bound to terminate after at most $\lceil 1 / \alpha_{k'} \rceil$ steps. This produces a desired subset $F_{k'}$ of size at most $\varepsilon m / (2k)$, owing to the choice of $\xi_{k'}$.
It remains to show that $D_0$, which is equal to $f(\mathbf{F})$ for $\mathbf{F} = (F_1, \ldots, F_{k-1})$, contains almost all the elements from $I'$. To this end, let $R \subseteq I'$ denote the set of all elements which are not contained in $D_0$. Suppose, towards a contradiction, that $|R| \ge \varepsilon m / 2$. Then for some $k' \in \{0, \ldots, k-1\}$ we have $|R_{k'}| \ge \varepsilon m / (2k)$, where $R_{k'} = R \cap (D_{k'+1} \setminus D_{k'})$. This implies
\begin{align*}
e_{k'+1}(D_{k'+1}, I') &\ge \frac{1}{k} \sum_{v \in R_{k'}} \deg_{k'}(v, D_{k'+1}, I') \ge \frac{1}{k} \sum_{v \in R_{k'}} \deg_{k'}(v, D_{k'+1}, F_{k'}) \\
&\ge \frac{\varepsilon m}{2k^2} \cdot \frac{t_{k'}}{N} = \gamma_{k'+1} e(\ensuremath{\mathcal H}) q^{k'+1},
\end{align*}
which contradicts \eqref{eq:D_k_upper}. Therefore, $|R| < \varepsilon m / 2$. All together, the set $F = X \cup R \cup F_1 \cup \ldots \cup F_{k-1}$ is of size $|F| \le \varepsilon m$ and $I \subseteq F \cup f_m(\mathbf{F})$. By adding arbitrary elements to $F$ if needed, we can assume $|F| = \varepsilon m$. Note that the $(k-1)$-tuple $\mathbf{F} = (F_1, \ldots, F_{k-1}) \in \ensuremath{\mathcal T}_{k-1}(F)$ ignores the set $X$ and all these added vertices, and potentially also some vertices from $R$.
\end{proof}
\newpage
\section{Saturating sets} \label{sec:second_moment}
The proof of Lemma \ref{lemma:saturating} revolves around the following idea. Let $I, D \subseteq V(\ensuremath{\mathcal H})$ be such that $e_{k'}(D, I) \ge t$, for some $t$ and $k' \in [k-1]$. To say something about the distribution of the edges counted in $e_{k'}(D, I)$, it suffices to bound the second moment of the random variable $\deg_{k'}(v, D, I)$ with respect to uniformly chosen $v \in V(\ensuremath{\mathcal H})$. More precisely, if
$$
\mathbb{E}_v[\deg_{k'}^2(v, D, I)] \le Z(t/N)^2,
$$
for some $Z > 1$, then by the Paley-Zygmund inequality we have
$$
\textrm{Pr}_v \left( \deg_{k'}(v, D, I) \ge t/(2N) \right) \ge \frac{\mathbb{E}_v[\deg_{k'}(v, D, I)]^2}{4\mathbb{E}_v[\deg_{k'}^2(v, D, I)]} = 1 / (4Z).
$$
In other words, $I$ is $(k', 1/(4Z), t/2)$-saturating for $D$.
Most of the remaining work in the proof of Lemma \ref{lemma:saturating} goes into showing that $I$ has properties which allow us to apply the above idea on some small subset $W \subseteq I$ rather than on the whole $I$. The following lemma, known as the R\"odl-Ruci\'nski deletion method, is used to show that these properties hold with probability at least $1 - \beta^m$. The lemma originally appears in \cite[Lemma 4]{rodl1995threshold}, and the version stated here more closely follows \cite[Lemma A.1]{nenadov2021klr}.
\begin{lemma} \label{lemma:deletion_general}
Let $V$ be a set with $N$ elements and let $\ensuremath{\mathcal S}$ be a family of $s$-element subsets of $V$. For every $\beta > 0$ there exists $T > 1$, such that for every $m$ there exists a family $\ensuremath{\mathcal B}$ of $m$-element subsets of $V$, $|\ensuremath{\mathcal B}| \le \beta^m \binom{N}{m}$, with the following property: Every $m$-element subset $I \subseteq V$, $I \notin \ensuremath{\mathcal B}$, contains $X \subseteq I$ of size $|X| \le \beta m$ such that at most
$$
T \cdot |\ensuremath{\mathcal S}| \left( \frac{m}{N} \right)^{s'}
$$
sets from $\ensuremath{\mathcal S}$ intersect $I \setminus X$ in at least $s'$ elements, for every $1 \le s' \le s$.
\end{lemma}
We are now ready to prove Lemma \ref{lemma:saturating}.
\begin{proof}[Proof of Lemma \ref{lemma:saturating}]
For each $k \le s \le 2k$, let
$$
\ensuremath{\mathcal S}_s = \{S \subset V(\ensuremath{\mathcal H}) \colon |S| = s \text { and } S = e \cup e' \text{ for some } e, e' \in \ensuremath{\mathcal H} \}.
$$
Given a set $Q \subseteq V(\ensuremath{\mathcal H})$ and $s' \le s$, let
$$
\ensuremath{\mathcal S}_{s}(Q, s') = \{ S \in \ensuremath{\mathcal S}_{s} \colon |S \cap Q| \ge s'\}.
$$
By Lemma \ref{lemma:deletion_general} applied with $\beta/(4k)^2$ and union over all $s \in [k, 2k]$, there exists a family $\ensuremath{\mathcal B}_m$ of size $|\ensuremath{\mathcal B}_m| \le \beta^m \binom{N}{m}$ with the property that every $m$-element $I \subseteq V(\ensuremath{\mathcal H})$, $I \not \in \ensuremath{\mathcal B}_m$, contains $X \subseteq I$ of size $|X| \le \beta m$ such that
\begin{equation} \label{eq:bound_S}
|\ensuremath{\mathcal S}_s(I \setminus X, s')| \le T |\ensuremath{\mathcal S}_s| q^{s'}
\end{equation}
for each $s \in [k, 2k]$ and $1 \le s' \le s$, where $q = m / N$. We show that every such $I$ satisfies the property of the lemma.
Consider one such $I$ and suppose $e_{k'}(D, I') \ge \gamma e(\ensuremath{\mathcal H}) q^{k'}$ for some $I' \subseteq I \setminus X$, $|I'| \ge m/2$, and $D \subseteq V(\ensuremath{\mathcal H})$. Form a subset $W \subseteq I \setminus X$ by taking each element in $I \setminus X$ with probability $\xi/2$, independently of all other elements. By Markov's inequality, we have $|W| \le \xi m$ with probability at least $1/2$. We claim that with probability larger than $1/2$ we also have the following two inequalities:
\begin{equation} \label{eq:saturating_forcing}
e_{k'}(D, W) \ge \gamma e(\ensuremath{\mathcal H}) (q \xi / 2)^{k'}/2 \quad \text{ and } \quad \frac{1}{N} \sum_{v \in V(\ensuremath{\mathcal H})} \deg_{k'}^2(v, D, W) \le Z \left( \frac{\gamma e(\ensuremath{\mathcal H}) (q\xi)^{k'}}{N} \right)^2,
\end{equation}
where $Z$ is some constant that will be specified later and which, importantly, does not depend on $\xi$. This implies that there (deterministically) exists $W \subseteq I \setminus X$ of size $|W| \le \xi m$ which satisfies \eqref{eq:saturating_forcing}. Now we follow the idea from the beginning of the section: Consider $v \in V(\ensuremath{\mathcal H})$ chosen uniformly at random and let $Y = \deg_{k'}(v, D, W)$ (for $v \not \in D$ we have $\deg_{k'}(v, D, W) = 0$). Then $\mathbb{E}_v[Y] \ge e_{k'}(D, W) / N$, thus \eqref{eq:saturating_forcing} implies $\mathbb{E}[Y^2] \le 2^{2k} Z \cdot \mathbb{E}[Y]^2$. Applying the Paley-Zygmund inequality, we obtain
$$
\textrm{Pr}_v \left( Y \ge \mathbb{E}_v[Y]/2 \right) \ge \frac{\mathbb{E}_v[Y]^2}{4 \mathbb{E}_v[Y^2]} \ge 1/(4 \cdot 2^{2k} Z) := \alpha.
$$
In other words, for at least $\alpha N$ elements $v \in V(\ensuremath{\mathcal H})$ we have
$$
\deg_{k'}(v, D, W) \ge \lambda e(\ensuremath{\mathcal H}) q^{k'} / N,
$$
where $\lambda = \gamma (\xi/2)^{k'}/4$, hence $W$ is $(k',\alpha,\lambda e(\ensuremath{\mathcal H})q^{k'})$-saturating for $D$.
It remains to show that \eqref{eq:saturating_forcing} holds with probability larger than $1/2$. In the rest of the proof we frequently rely on the following two facts: if $x \ge m_0$ then
\begin{equation} \label{eq:sum_Delta}
\sum_{s = k+1}^{2k-1} |\ensuremath{\mathcal S}_s| (x/N)^s \le 3^{2k} \Delta_{\ensuremath{\mathcal H}}(x)
\end{equation}
and
\begin{equation} \label{eq:Delta_x}
\Delta_{\ensuremath{\mathcal H}}(x) \le B \mu_{\ensuremath{\mathcal H}}(x)^2 / x.
\end{equation}
The first inequality follows directly from the definition of $\Delta_{\ensuremath{\mathcal H}}(x)$, and perhaps the only thing worth mentioning is that $3^{2k}$ is used as an upper bound on the number of pairs of subsets $(e, e')$ of a given set $S$, $|S| \le 2k$, such that $S = e \cup e'$. Inequality \eqref{eq:Delta_x} can be obtained as follows:
\begin{align*}
\Delta_{\ensuremath{\mathcal H}}(x) = \sum_{e \sim e'} \left(\frac{x}{N}\right)^{|e \cup e'|} &= \sum_{e \sim e'} \left(\frac{m_0}{N}\right)^{|e \cup e'|} \left(\frac{x}{m_0}\right)^{|e \cup e'|} \stackrel{(x \ge m_0)}{\le} \sum_{e \sim e'} \left(\frac{m_0}{N}\right)^{|e \cup e'|} \left(\frac{x}{m_0}\right)^{2k-1} \\
&\le \left(\frac{x}{m_0} \right)^{2k-1} \cdot B\frac{(e(\ensuremath{\mathcal H}) (m_0/N)^{k})^{2}}{m_0} = B \mu_{\ensuremath{\mathcal H}}^2(x) / x.
\end{align*}
Let us start with the first inequality in \eqref{eq:saturating_forcing}. Note that $\mathbb{E}[e_{k'}(D, W)] \ge \gamma e(\ensuremath{\mathcal H}) (q \xi / 2)^{k'}$, thus it suffices to show
\begin{equation} \label{eq:real_e_k'}
\Pr\left( e_{k'}(D, W) \ge \mathbb{E}[e_{k'}(D, W)] / 2 \right) \ge 4/5.
\end{equation}
Our plan is to show $\Var[ e_{k'}(D, W)] = o(\mathbb{E}[e_{k'}(D, W)]^2)$, in which case the previous bound follows from Chebyshev's inequality, with room to spare. To this end, given a set $Q \subseteq V(\ensuremath{\mathcal H})$ and $s \in \mathbb{N}$, consider the following family of pairs of edges:
$$
\Lambda_{s}(Q,k') = \{ (e, e') \in \ensuremath{\mathcal H} \times \ensuremath{\mathcal H} \colon |e \cup e'| = s,\; |e \cap Q|, |e' \cap Q| \ge k', e \neq e' \text{ and } e \cap e' \cap Q \neq \emptyset \}.
$$
It is a standard fact (see, e.g., \cite[p.48]{alon_spencer}) that
\begin{equation} \label{eq:cheb_var}
\var[e_{k'}(D, W)] \le \mathbb{E}[e_{k'}(D, W)] + \sum_{s = k+1}^{2k-1} |\Lambda_{s}(W,k')|.
\end{equation}
On the one hand, for each pair $(e, e') \in \Lambda_{s}(W, k')$ we have $e \cup e' \in \ensuremath{\mathcal S}_s(W, \ell(s))$ for $\ell(s) = s + 2k' - 2k$. Consequently, as $W \subseteq I \setminus X$ we also have $e \cup e' \in \ensuremath{\mathcal S}_s(I \setminus X, \ell(s))$. For $s < 2k - 2k'$ we have $\ell(s) < 0$ which can be certainly be improved to at lest $0$, but it nonetheless suffices and makes calculations easier. On the other hand, each set in $\ensuremath{\mathcal S}_s(I \setminus X, \ell(s))$ corresponds to at most $3^{2k}$ pairs in $\Lambda_{s}(W, k')$. Therefore, we have
\begin{multline*}
\sum_{s = k + 1}^{2k-1} |\Lambda_{s}(W,k')| \le 3^{2k} \sum_{s = k + 1}^{2k-1} |\ensuremath{\mathcal S}_s(I \setminus X, \ell(s))|
\stackrel{\eqref{eq:bound_S}}{\le} 3^{2k}T \sum_{s = k + 1}^{2k-1} |\ensuremath{\mathcal S}_s| q^{\ell(s)} \\
\stackrel{\eqref{eq:sum_Delta}}{\le} 9^{2k} T q^{2k' - 2k} \Delta_{\ensuremath{\mathcal H}}(m) \stackrel{\eqref{eq:Delta_x}}{\le} 9^{2k} T B \frac{(e(\ensuremath{\mathcal H}) q^{k'})^2}{m} \le \frac{9^{2k} TB}{(\xi/2)^{2k'} \gamma^2 m} \mathbb{E}[e_{k'}(D, W)]^2.
\end{multline*}
Plugging this into \eqref{eq:cheb_var} and applying Chebyshev's inequality verifies \eqref{eq:real_e_k'}.
Next, we show that
\begin{equation} \label{eq:deg_k'_2}
\sum_{v \in V(\ensuremath{\mathcal H})} \deg_{k'}^2(v, D, W) \le Z \frac{(\gamma e(\ensuremath{\mathcal H}) (q\xi)^{k'})^2}{N}
\end{equation}
also holds with probability at least $4/5$. Note that the sum on the left hand side counts the number of pairs of edges $(e, e')$ such that $W \setminus \{v\}$ intersects both $e$ and $e'$ in at least $k'$ elements, for some $v \in e \cap e'$. For every such pair we have $e \cup e' \in \ensuremath{\mathcal S}_s(W,\ell'(s))$, where $s = |e \cup e'|$ and $\ell'(s) = s + 2k' + 1 - 2k$. Similarly as in the previous case, each set in $\ensuremath{\mathcal S}_s(W, \ell'(s))$ corresponds to at most $3^{2k}$ such pairs, thus
\begin{equation} \label{eq:bound_deg2}
\sum_{v \in V(\ensuremath{\mathcal H})} \deg_{k'}^2(v, D, W) \le 3^{2k} \sum_{s = k + 1}^{2k-1} |\ensuremath{\mathcal S}_s(W, \ell'(s))|.
\end{equation}
For each $e \in \ensuremath{\mathcal S}_s(I \setminus X, \ell'(s))$ we have $e \in \ensuremath{\mathcal S}_s(W, \ell'(s))$ with probability at most $2^{2k} \xi^{\ell'(s)}$, with room to spare.
As before, while this is well defined for every $s$, it does not make much sense for $s < 2k - 2k' - 1$. Nonetheless, it suffices and makes calculations easier.
By the linearity of expectation we have
$$
\mathbb{E}[|\ensuremath{\mathcal S}_s(W, \ell'(s))|] \le 2^{2k} |\ensuremath{\mathcal S}_s(I \setminus X, \ell'(s))| \xi^{\ell'(s)},
$$
thus
\begin{align*}
\mathbb{E}\left[\sum_{s=k+1}^{2k-1}|\ensuremath{\mathcal S}_s(W, \ell'(s))|\right] &\stackrel{\eqref{eq:bound_S}}{\le} 2^{2k} T \cdot \sum_{s = k+1}^{2k-1} |\ensuremath{\mathcal S}_s| (q \xi)^{\ell'(s)} \stackrel{\eqref{eq:sum_Delta}}{\le} 6^{2k} T (q\xi)^{-2k + 2k' + 1} \Delta_\ensuremath{\mathcal H}(\xi m) \\
&\le 6^{2k} TB \cdot q\xi \frac{(e(\ensuremath{\mathcal H}) (q\xi)^{k'})^2}{\xi m} = 6^{2k} TB \frac{(e(\ensuremath{\mathcal H}) (q\xi)^{k'})^2}{N}.
\end{align*}
By Markov's inequality, with probability at least $4/5$ the right hand side of \eqref{eq:bound_deg2} is at most
$$
Z \frac{(\gamma e(\ensuremath{\mathcal H})(q\xi)^{k'})^2}{N},
$$
where $Z = 5 \gamma^{-2} 18^{2k} TB$, thus \eqref{eq:deg_k'_2} follows. Note that it is crucial here that $Z$ does not depend on $\xi$. This concludes the proof.
\end{proof}
\newpage
\section{Constructing containers -- overview}
Consider some $I$ which not independent. Let $F_{k-1} \subseteq I$ be a subset of size $\varepsilon m / k$ chosen uniformly at random, and denote with $D_{k-1} \subset V(\ensuremath{\mathcal H})$ the set of all elements $v \in V(\ensuremath{\mathcal H})$ which do not form an edge with some $k-1$ elements from $F_{k-1} \setminus \{v\}$. As $I$ is independent, we necessarily have $I \subseteq D_{k-1}$. This already narrows down $I$ to a potentially smaller set, albeit not necessarily with few edges. As $F_{k-1}$ was chosen at random, we can hope that not only no element from $D_{k-1}$ forms an edge with elements from $F_{k-1}$, but also that it forms very few edges with any $k-1$ elements from $I$. Somewhat wishfully (and clearly wrong), let us suppose that it actually forms no edge with $k-1$ elements from $I$. Now repeat the whole step with this new information in mind. Pick a random $F_{k-2} \subseteq I$ of size $\varepsilon m / k$, and let $D_{k-2} \subseteq D_{k-1}$ be the set of elements which do not form an edge with some $k-2$ elements from $F_{k-2}$ and another element from $D_{k-1}$. As no element from $D_{k-1}$ forms an edge with $k-1$ elements from $I$, we necessarily have $I \subseteq D_{k-2}$. Like in the previous case, since $F_{k-2}$ was chosen at random we can be very optimistic and hope that not only no two elements from $D_{k-2}$ form an edge with $k-2$ elements from $F_{k-2}$, but also that they do not form an edge with $k-2$ elements from $I$, and so on.
To summarise, after step $i$ we obtain a set $D_{k-i} \subseteq V(\ensuremath{\mathcal H})$ such that: (i) $I \subseteq D_{k-i}$, (ii) $D_{k-i}$ can be computed as a function of the tuple $(F_{k-1}, \ldots, F_{k-i})$, and (iii) no $k-i$ elements from $I$ form an edge with $i$ elements from $D_{k-i}$. For $i = k$, this implies that $D_0$ is an independent set which contains $I$. This is stronger that what we promise in Theorem \ref{thm:containers} and, as anticipated, it is too good to be true implying that our wishful assumption cannot hold. However, a variation of it -- where we do not quite get that no $k-i$ elements from $I$ form an edge with some $i$ elements from $D_{k-i}$ but rather that the number of such edges is sufficiently small -- is true and suffices for a variant of the argument to go through.
One thing that we did not address here is the condition \eqref{eq:delta_m0}. For now, it suffices to keep in mind that this condition forces relatively uniform distribution of the edges in the following sense: Given a `typical' $m$-element subset $I$ such that there are $t = \Omega(e(\ensuremath{\mathcal H}) (m/N)^{k-1})$ edges in $\ensuremath{\mathcal H}$ with $k-1$ elements in $I$, then $\Theta(N)$ vertices in $V(\ensuremath{\mathcal H})$ are contained in $\Omega(t/N)$ such edges. A statement of this type is crucial in implementing described ideas for building containers, and it is made precise in Definition \ref{def:saturating} and Lemma \ref{lemma:saturating}. |
\section{Introduction}
In recent years, distributed optimization techniques over a multi-agent network have attracted considerable attention since they play an essential role in engineering problems in distributed control \cite{BCM, CYRC}, signal processing \cite{Boyd Gossip, QT} and the machine learning problems \cite{BCN, FCG, RB}.
In distributed optimization, many agents have their own local cost and try to find a minimizer of the sum of those local cost functions in a collaborative way that each agent only uses the information from its neighboring agents where the neighborhood structure is depicted as a graph, often undirected or directed.
There has been a significant interest in consensus-based distributed gradient method. One fundamental work is \cite{Ned DGD} which developed the distributed gradient descent on undirected graph. This algorithm consists of local gradient step and consensus step based on communication between neighboring agents. The convergence property of the algorithm has been studied in the works \cite{Ned DGD, R Ned V - stochastic, I, YLY}. There are also various distributed algorithms containing the distributed dual averaging method \cite{DAW}, consensus-based dual decomposition \cite{FMGP, SJ}, and the alternating direction method of multipliers (ADMM) based algorithms \cite{MJ, SLYW}.
These algorithms work with doubly-stochastic matrix associated with the undirected graph.
The gradient-push algorithm was introduced in \cite{Ned Push sum} to solve the distributed optimization for directed graph which utilizes push-sum algorithms \cite{KDG, TLR2}. The communication of this algorithm is represented by column stochastic matrix, which requires each agent to know its out-degree at each time, without having the information of the number of agents. This algorithm has influenced a significant impact of later works. The work \cite{NO1} studied
the algorithm with gradient having a noise. The time varying distributed optimization was also considered \cite{AGL} using the gradient-push algorithm. Recently, stochastic gradient-push algorithm was designed for large scale deep learning problem \cite{ALBR}. This work was also extended in \cite{TMHP} further to quantized communication settings. We also refer to \cite{AR, ZY} where the authors studied the asynchronous version of the gradient-push algorithm.
Regardless of the types of graphs, these distributed algorithms require each agent to communicate with their neighbors at every iteration, which leads to a overhead in restricted environments. Power consumption by communication may become more significant than that by computation of control inputs or optimization algorithms \cite{SHCA}. Recently, the event-triggering approach has appeared as a promising paradigm to reduce the communication load in distributed systems. In the distributed detection problem over sensor network \cite{AVJ, AMM}, each sensor censors its local data and send the updated data to the fusion center only when the data is informative. For distributed control problems, agents send their coordinate information only when a triggering condition is satisfied \cite{MT, DFJ}.
For the distributed optimization problems, recent works \cite{KHT, LL, MUA, WYLY, ZC} developed distributed optimization algorithms with event-triggered communication to overcome the communication overhead of distributed systems.
Lu-Li \cite{LL} designed the distributed gradient descent with event-triggered communication
for the distributed optimization on the whole space, and it was further studied in
Li-Mu \cite{LM} to establish a convergence rate.
For the distributed optimization on bounded
domain, Kajiyama et al \cite{KHT} designed the projected distributed gradient descent with
event-triggered communication. Liu et al \cite{LLSX} extended the work to the case with constant
step-size. Cao-Basar \cite{CB} studied the online distributed problem using the distributed
event-triggered gradient method.
In these algorithms, each agent sends its state only when the difference between the current state and the latest sent state is larger than a threshold,
therefore reducing possible unnecessary network utilization.
The consensus-based distributed optimization algorithms with event-triggering communication mentioned above have been proposed for the undirected graph. In this work, we are interested in
developing a distributed optimization on directed graph involving the even-triggered communication. Precisely we propose the gradient-push algorithm incorporating the event-triggered communication.
In the proposed algorithm, each agent only sends its state information when the differences between the latest sent states and the current states are larger than a triggering thresholds. We prove that the algorithm solves the distributed optimization under suitable decays and summability conditions on the stepsize and the triggering thresholds. The numerical experiments are given for the proposed algorithm, supporting the theoretical results.
The rest of the paper is organized as follows. In Section 2, we state the problem and introduce the algorithm with its convergence results. Section 3 is devoted to provide a consensus estimate, which is essentially used in Section 4 to prove the convergence results. In section 5, we present numerical results of the proposed algorithm.
Before ending this section, we state several notations used in this paper. For a matrix $A \in \mathbb{R}^{n \times m}$, $a_{ij}$ or $[A]_{ij}$ denotes the $(i,j)$th entry of A. For a vector $x\in\mathbb{R}^d$, $\|x\|=\sqrt{x^Tx}$ denotes the standard Eculidean norm. In addition, for $X \in \mathbb{R}^{m\times d}$ given by $X=[x_1; x_2; \cdots ; x_m]$ with row vector $x_k \in \mathbb{R}^{d}$, we define the mixed norm $\|X\|_1$ by $\|X\|_1 = \sum_{k=1}^m \|x_k\|$ nad the maximum norm $\|X\|_{\infty} = \max_{1\leq k\leq m} \|x_k\|$. Also we use $\bar{x}$ to denote $\bar{x} = \frac{1}{m}\sum^{m}_{k=1} x_k$.
\section{Problem, algorithm, and main results}
\subsection{Problem statement} We consider the distributed optimization problem, which consists of $m$ agents connected by a network that collaboratively minimize a global cost function given by the sum of local private cost functions. Formally, the problem is described by
\begin{equation}\label{problem}
\min_{x\in\mathbb{R}^d}f(x)=\sum^m_{i=1}f_i(x),
\end{equation}
where $f_i:\mathbb{R}^d\rightarrow\mathbb{R}$ is a local convex cost function only known to agent $i\in \mathcal{V}=\{1,2,\cdots m\}$. We let $f^*$ be the optimal value of problem \eqref{problem} and denote by $X^*$ the set of optimal solutions, i.e.,
$$
X^*=\{x \in \mathbb{R}^d~:~ f(x) = f^*\},
$$
which is assumed to be nonempty. We make the following standard assumption on the local cost functions.
\begin{ass}\label{GB}
For each $i \in \{1,\cdots, m\}$, there exists $D_i>0$ such that
\begin{equation}\label{eq-2-1}
\|\nabla f_i(x)\| \leq D_i \quad \forall x \in \mathbb{R}^d.
\end{equation}
We set $D = \max_{1\leq i \leq m} D_i$.
\end{ass}
The communication pattern among agents in \eqref{problem} at each time $t \in \mathbb{N}\cup \{0\}$ is characterized by a directed graph $\mathcal{G}(t)=(\mathcal{V},\mathcal{E}(t))$, where each node in $\mathcal{V}$ represents each agent, and each directed edge $(i,j)\in \mathcal{E}(t)$ means that $i$ can send messages to $j$. In this work, we consider a sequence of graphs $\{\mathcal{G}(t)\}_{t\geq0}$ satisfying the following assumption.
\begin{ass}\label{graph}
The sequence of graph $\{\mathcal{G}(t)\}_{t\geq0}$ is uniformly strongly connected, i.e., there exists a value $B \in \mathbb{N}$ such that the graph with edge set $\cup_{i=kB}^{(k+1)B -1} \mathcal{E}(i)$ is strongly connected for any $k \geq 0$.
\end{ass}
We define in-neighbors and out-neighbors of node $i$, respectively, as $N_i^{\text{in}}(t)=\{j|(j,i)\in\mathcal{E}(t)\}\cup \{i\}$ and $N_i^{\text{out}}(t)=\{j|(i,j)\in\mathcal{E}(t)\}\cup \{i\}$. Also the out-degree of node $i$ is defined as $d_i^{\text{out}}(t)=|N_i^{\text{out}}(t)|$. Define the matrix A(t) such that
$[A(t)]_{ij} = a_{ij}(t)$, where
$$
a_{ij}(t)=\begin{cases}
1/d_j^{\text{out}}(t),\ \text{if $j\in N_i^{\text{in}}(t)$},\\
0, \ \text{otherwise}.
\end{cases}
$$
The matrix $A(t) \in \mathbb{R}^{m\times m}$ is column stochastic and we recall some useful properties of this matrix from \cite[Corollary 2]{Ned Push sum}.
\begin{lem}[\cite{Ned Push sum}, $Corollary$ 2]\label{lemma2}
Suppose that the graph sequence $\{G(t)\}$ is uniformly strongly connected. Then, the following statements are valid.
\begin{enumerate}
\item For each integer $s\geq 0$, there is a stochastic vector $\phi(s)$ such that for all $i,j$ and $t\geq s$
\begin{equation}\label{MatrixEstimate}
|[A(t:s)]_{ij}-\phi_i(t)| \leq C_0 \lambda^{t-s}
\end{equation}
for some values $C_0 \geq 1$ and $\lambda\in(0,1)$ depending on the graph sequence.
\item The following inequality holds.
\begin{equation}\label{Q}
Q := \inf_{t=0,1,\cdots}{\min_{1\leq i \leq m}{[A(t:0)\mathbf{1}]_i}} \geq 1/n^{nB}.
\end{equation}
\end{enumerate}
Here we denote by $A(t:s)$ the matrix given as
\begin{equation*}
A(t:s) = A(t) A(t-s) \cdots A(s) \quad \textrm{for all}~ t \geq s \geq 0.
\end{equation*}
\end{lem}
\begin{algorithm}
\begin{algorithmic}[1]
\caption{Distributed Event-Triggered gradient-push algorithm}\label{algo}
\REQUIRE Initialize $x_i(0)$ arbitraily, $y_i(0)=1$ for all $i\in\{1,\cdots,m\}$. Set $\hat{x}_i(0)=x_i(0)$
\FOR{$t=0,1,\cdots$,}
\STATE Compute the new action as
\begin{align}
\widehat{w}_i(t+1) &= \sum^m_{j=1} a_{ij}(t) \hat{x}_j(t), \\
y_i(t+1) &= \sum^m_{j=1} a_{ij}(t) \hat{y}_j(t),\label{hatyy} &\\
\widehat{z}_i(t+1)&=\frac{\widehat{w}_i(t+1)}{y_i(t+1)}, \label{hatzz}
\end{align}
\begin{equation}\label{update}
x_i(t+1) = \widehat{w}_i(t+1) - \alpha(t+1) \nabla f_i(\widehat{z}_i(t+1)).
\end{equation}
\IF{$\|x_i(t+1)-\hat{x}_i(t)\|\geq \tau(t+1)$}
\STATE Send $x_i(t+1)$ to neighbors $\mathcal{N}_i$. Set $\hat{x}_i(t+1) = x_i(t+1)$
\ELSE
\STATE Do not send and set $\hat{x}_i(t+1) = \hat{x}_i(t)$
\ENDIF
\IF{$|y_i(t+1)-\hat{y}_i(t)|\geq \zeta(t+1)$}
\STATE Send $y_i(t+1)$ to neighbors $\mathcal{N}_i$. Set $\hat{y}_i(t+1) = y_i(t+1)$
\ELSE
\STATE Do not send and set $\hat{y}_i(t+1) = \hat{y}_i(t)$
\ENDIF
\ENDFOR
\end{algorithmic}
\end{algorithm}
In addition, we consider the following assumptions on the stepsize and the thresholds for the trigger conditions.
\begin{ass}\label{stepsize}
The sequence of stepsize $\{\alpha(t+1)\}_{t\geq 0}$ is monotonically non-increasing and satisfies
$$
\sum^\infty_{t=0}\alpha(t+1)=\infty, \ \sum^\infty_{t=0}\alpha(t+1)^2<\infty.
$$
\end{ass}
\begin{ass}\label{event-triggered}
The sequence of event-triggering thresholds $\{\tau(t)\}_{t\geq 0}$ is monotonically non-increasing and satisfies
$$
\sum^\infty_{t=1}\tau(t)<\infty.
$$
\end{ass}
\begin{ass}\label{y-trigger}
The sequence of event-triggering thresholds $\{\zeta(t)\}_{t\geq 0}$ is monotonically non-increasing and satisfies
\begin{equation*}
\sum^\infty_{t=1} t^{3/2}\zeta(t) < \infty, \quad \sum^\infty_{t=1} \zeta(t) <1.
\end{equation*}
\end{ass}
Note that $\sum^\infty_{t=1} t^{3/2}\zeta(t) < \infty$ implies that there exists a finite $M$ such that $\sum^\infty_{t=0} \zeta(t) = M$. If we set a new sequence $\{\tilde{\zeta}(t)\}_{t\in \mathbb{N}}$ by $\tilde{\zeta}(t) = \zeta(t)/(M+1)$, then it satisfy $\sum^\infty_{t=1} \tilde{\zeta}(t)<1$. Hence if we have a sequence $\{\zeta(t)\}_{t\in \mathbb{N}}$ satisfying $\sum^\infty_{t=1} t^{3/2}\zeta(t) < \infty$, then we may divide the sequence by a positive constant to satisfy Assumption \ref{y-trigger}. One example of the sequence that satisfies Assumption \ref{y-trigger} is $\zeta(t) = \frac{1}{3t^3}$.
For $T \geq 0$ we set the following variables
\begin{equation}\label{eq-2-8}
E_{\tau}(T) = \sum_{t=1}^T \tau (t), \quad E_{\tau,2} (T) = \sum_{t=1}^T \tau(t)^2, \quad E_{\tau}= \sum_{t=1}^{\infty} \tau (t)
\end{equation}
and
\begin{equation}\label{eq-2-8}
F_{\zeta}(T) = \sum_{t=1}^T \zeta (t), \quad F_{\zeta}= \sum_{t=1}^{\infty} \zeta (t), \quad F_{\zeta_{3/2}} = \sum_{t=1}^\infty t^{3/2}\zeta(t).
\end{equation}
These variables will appear in the statements and proofs of the convergence result for Algorithm 1.
\
Before finishing this subsection, we define the following constant
\begin{equation}\label{delta}
\delta:= \min_{1\leq i\leq m }\inf_{t\in\mathbb{N}} y_i(t).
\end{equation}
whose positivity is proved in Lemma \ref{lowerbound} under the Assumption \ref{y-trigger}.
\subsection{Main results}
Our first result establishes the convergence of $\hat{z}_i(t)$ to the optimal solutions for an arbitrary stepsize $\alpha(t)$ satisfying Assumption \ref{stepsize}, and event-triggering thresholds $\tau(t)$ and $\zeta(t)$ satisfying Assumption \ref{event-triggered} and \ref{y-trigger}.
\begin{thm}\label{maintheorem}
Suppose that Assumptions \ref{GB} - \ref{y-trigger} hold. Then the sequence $\{\hat{z}_i (t)\}_{t \in \mathbb{N}}$ for $1\leq i \leq n$ of the Algorithm \ref{algo} satisfies the following property:
$$
\lim_{t\rightarrow \infty} \hat{z}_i(t) = x^* \ \text{for all $i$ and for some $x^*\in X^*$}.
$$
\end{thm}
Next we consider the Algorithm \ref{algo} with specific stepsize $\alpha(t)=1/\sqrt{t}$. This stepsize does not satisfy Assumption \ref{stepsize}, but we may obtain an explicit convergence rate as in the following result. Before stating the result, we give some notations which are used throughout the paper. We define the constants $m_{\zeta}$ and $B_{\zeta}$ by
\begin{equation}\label{mzeta}
m_{\zeta} = 1_m^T y(0) + \sum_{s=1}^{\infty} 1_m^T \theta (s) = m+ \sum_{s=1}^{\infty} 1_m^T \theta (s)
\end{equation}
and
\begin{equation}\label{bzeta}
B_{\zeta} = \frac{m_{\zeta}}{m}.
\end{equation}
These constants are well-defined if $\{\zeta(s)\}_{s \geq 0}$ is summable since we have the inequality $|1_m^T \theta (s)| \leq m \zeta (s)$ from the triggering condition.
\begin{thm}\label{case2_main}
Suppose that Assumption \ref{GB}, \ref{graph}, \ref{event-triggered} and \ref{y-trigger} hold. Let $\alpha(t)=\frac{1}{\sqrt{t}}$ for $t \geq 0$. Define $H(-1)=1$ and $H(t) : =\prod_{k=0}^t (1+\tau(k))$. Moreover, suppose that every node $i$ maintains the variable $\tilde{z}_i(t)\in \mathbb{R}^d$ initialized at time $t=0$ with $\tilde{z}_i(0)\in\mathbb{R}^d$ and updated by
$$
\tilde{z}_i(t+1) = \frac{\frac{\alpha(t+1)}{H(t)}\hat{z}_i(t+1)+S(t)\tilde{z}_i(t)}{S(t+1)},
$$
where $S(0)=0$ and $S(t)=\sum^{t-1}_{k=0}\frac{\alpha(k+1)}{H(k)}$ for $t\geq 1$. Then we have for each $T\geq 0$ and $i=1,\cdots,m$, the following estimate
\begin{equation*}
\begin{split}
&f(\tilde{z}_i(T+1))-f(x^*)
\\
&{\leq \frac{me^{E_{\tau}}}{2\sqrt{T+1}}J_1 (T) + \frac{3mDe^{E_{\tau}}}{\delta \sqrt{T+1}} J_2 (T) + \frac{3mDe^{E_\tau}}{ \delta\sqrt{T+1}} J_3 (T),}
\end{split}
\end{equation*}
where
\begin{align*}
&\medmath{ J_1 (T)= \frac{\|\bar{x}(0)-x\|^2}{B_{\zeta}}+\bigg[2D^2 \Big(1+ \ln{(T+1)}\Big)+2E_{\tau,2}(T)+E_{\tau}(T)\Big)\bigg]B_{\zeta}}
\\
&J_2 (T)= \medmath{\bigg(\frac{C_0 }{ (1-\lambda)} \bigg)\|x(0)\|_1 + \frac{4mC_0 E_{\tau}(T)}{ (1-\lambda)} +\bigg( \frac{C_0mD}{ (1-\lambda)}\bigg)(1+\ln(T))}
\\
&J_3 (T)=\medmath{ \sum_{t=0}^T K(t) \alpha (t+1) \Big[ \|x(0)\|_1 + \sum_{s=0}^{t-1} (\alpha (s+1) D + \tau (s))\Big],}
\end{align*}
and $x^*\in X^*$. In addition, the right hand side is bounded by $O(\log (T+1)/ \sqrt{T+1})$. Here $K(t) = \beta(t)/m_{\zeta}$ denotes a bound for asymptotic behavior of $y_i (t)$ stated in \mbox{Lemma \ref{lem-3-4}.}
\end{thm}
\section{Properties of the sequence $\{y_i(t)\}_{t\in\mathbb{N}}$ and Disagreement in Agent Estimates}\label{sec3}
\subsection{Properties of the sequence $\{y_i(t)\}_{t\in\mathbb{N}}$}
A convergence property of $\{y_i(t)\}_{t\in\mathbb{N}}$ and their positive uniform lower bound are key points in proving our main results. Let us first look at the case without event-triggering $(\zeta(t)=0)$, which means all in-neighbors of agent $i$ share the $y_i(t)$ with this agent for every time step. In this case, since $y(t)=\hat{y}(t)$ for all $t\in \mathbb{N}$, it holds that
\begin{equation}\label{yyy}
y(t) = A(t-1:0)\mathbf{1}_m
\end{equation}
by \eqref {hatyy} in Algorithm \ref{algo}. Hence we can directly show that $y(t)$ converges to $\phi(t)$ and has a uniform lower bound $Q$ using Lemma \ref{lemma2}.
In the event-triggered case, $y(t)$ can be written as
\begin{equation}\label{yyyy}
y(t) = A(t-1:0) \mathbf{1}_m + \sum^{t-1}_{s=1}A(t-1:s)\theta(s),
\end{equation}
where $\theta(s) = \hat{y}(s) - y(s)$.
Therefore the convergence and uniform lower boundedness property may not hold due to the additional term
$$
\sum^{t-1}_{s=1}A(t-1:s)\theta(s).
$$
The following lemmas shows that $y(t)$ has a positive uniform lower bound $\delta$ and converges to $m_\zeta\phi(t)$ instead of $\phi(t)$ under Assumption \ref{y-trigger}.
\begin{lem}\label{lem-3-4}
Suppose that Assumption \ref{y-trigger} holds. Then we have
\begin{equation}\label{m_lowerbound}
m_\zeta \geq (1-F_\zeta) m
\end{equation}
and the following estimate holds:
\begin{equation}\label{y_consensus}
\Big\| y(t+1) - m_{\zeta}\phi(t)\Big\|_{\infty} \leq \beta(t) \quad \forall~t \in \mathbb{N},
\end{equation}
where
\begin{equation*}
\beta (t) =m\bigg(\Big(F_\zeta-F_{\zeta}(t)\Big) +C_0 \lambda^{t}+C_0\lambda^{t/2} F_\zeta(t) + \frac{\zeta([t/2]+1)}{1-\lambda}\bigg).
\end{equation*}
In addition, we have
\begin{equation*}
\lim_{t \rightarrow \infty} t^{3/2} \beta (t) =0.
\end{equation*}
\end{lem}
\begin{proof}
Observe that $|\theta_i (s)| =|\hat{y}_i (s) - y_i (s)| \leq \zeta (s)$ for $s \geq 1$ by the event-triggering condition, and so
\begin{equation*}
\Big|\sum_{s=1}^{\infty}1_m^T \theta (s) \Big|\leq m \sum_{s=1}^{\infty} \zeta(s) = m F_\zeta.
\end{equation*}
Using this in \eqref{mzeta}, we get
\begin{equation*}
m_{\zeta} = m + \sum_{s=1}^{\infty} 1_m^T \theta(s) \geq (1-F_\zeta)m,
\end{equation*}
which proves \eqref{m_lowerbound}.
Next we prove \eqref{y_consensus}. For each $s \geq 0$, by definition we have
\begin{equation*}
y(s+1) = A(s) \hat{y}(s) = A(s) (y(s) + \theta(s)),
\end{equation*}
where we have set $\theta (0)=0$. Using this iteratively gives the following formula
\begin{equation*}
\begin{split}
&y(t+1)
\\
& = A(t:0)y(0) + \sum_{s=1}^{t} A(t:s) \theta(s)
\\
& = \phi (t) \Big[ \mathbf{1}_m^T y(0) + \sum_{s=1}^{t} \mathbf{1}_m^T \theta (s) \Big]
\\
&\quad + (A(t:0) - \phi (t) 1_m^T) y(0) + \sum_{s=1}^{t} \Big[ (A(t:s) - \phi (t) \mathbf{1}_m^T) \theta (s) \Big].
\end{split}
\end{equation*}
Since $|\theta_j (s)| \leq \zeta(s)$, we find
\begin{equation*}
\begin{split}
\sum_{s=t+1}^{\infty} |\theta_j (s)| &\leq \sum_{s=t+1}^{\infty}\zeta(s) = F_\zeta-F_{\zeta}(t).
\end{split}
\end{equation*}
Using the above inequality, we obtain
\begin{equation}\label{eq-3-310}
\Big|m_{\zeta} - 1_m^T y(0) - \sum_{s=0}^{t}1_m^T \theta (s) \Big|= \Big| \sum_{s=t+1}^{\infty} 1_m^T\theta (s) \Big| \leq \Big(F_\zeta-F_{\zeta}(t)\Big)m.
\end{equation}
Hence we have
\begin{equation}\label{eq-3-340}
\begin{split}
&\Big\| y(t+1) - m_\zeta \phi(t)\Big\|_\infty
\\
&\leq \Big(F_\zeta-F_{\zeta}(t)\Big)m + \Big\|(A(t:0) - \phi (t) 1_m^T) y(0)\Big\|_\infty\\
&\quad + \Big\|\sum_{s=1}^{t} \Big[ (A(t:s) - \phi (t) 1_m^T) \theta (s) \Big]\Big\|_{\infty}.
\end{split}
\end{equation}
Now we estimate the second and third terms in the right hand side of \eqref{eq-3-340}. Using \eqref{MatrixEstimate} we have
\begin{equation}\label{eq-3-320}
\Big\|(A(t:0) - \phi (t) 1_m^T) y(0)\Big\|_\infty \leq m C_0 \lambda^{t}
\end{equation}
and
\begin{equation}\label{eq-3-330}
\begin{split}
&\Big\|\sum_{s=1}^{t} \Big[ (A(t:s) - \phi (t) 1_m^T) \theta (s) \Big]\Big\|_{\infty}
\\
&\leq mC_0 \sum_{s=1}^{t}\lambda^{t-s} \zeta (s)
\\
&\leq mC_0 \bigg(\sum^{[t/2]}_{s=1} \lambda^{t-s}\zeta(s) + \sum^t_{s=[t/2] +1} \lambda^{t-s} \zeta(s)\bigg)\\
&\leq mC_0\bigg( \lambda^{t/2} F_\zeta(t) + \zeta([t/2]+1) \sum^t_{s=[t/2] +1} \lambda^{t-s}\bigg)\\
& \leq mC_0\bigg(\lambda^{t/2} F_\zeta(t) + \frac{\zeta([t/2]+1)}{1-\lambda} \bigg).
\end{split}
\end{equation}
Putting the estimates \eqref{eq-3-320} and \eqref{eq-3-330} in \eqref{eq-3-340}, we get
\begin{equation*}
\begin{split}
&\Big\| y(t+1) - \phi(t) m_{\zeta}\Big\|_{\infty}
\\
& \leq m\bigg(\Big(F_\zeta-F_{\zeta}(t)\Big) +C_0 \lambda^{t}+C_0\lambda^{t/2} F_\zeta(t) + \frac{\zeta([t/2]+1)}{1-\lambda}\bigg).
\end{split}
\end{equation*}
This proves the second assertion of the lemma.
Now we shall show that $\lim_{t \rightarrow \infty} t^{3/2} \beta (t) = 0$. Since $\lambda \in (0,1)$, it suffices to show that
\begin{equation*}
\lim_{t \rightarrow \infty} t^{3/2} (F_{\zeta} - F_{\zeta}(t) + \zeta([t/2]))=0.
\end{equation*}
This fact follows directly from the fact that $\sum_{t=0}^{\infty} t^{3/2} \zeta (t) < \infty$ and the following inequality
\begin{equation*}
t^{3/2} (F_{\zeta} -F_{\zeta}(t)) = t^{3/2} \sum_{s=t+1}^{\infty}\zeta (s) \leq \sum_{s=t+1}^{\infty}s^{3/2} \zeta (s).
\end{equation*}
The proof is done.
\end{proof}
\begin{lem}\label{lowerbound}
Suppose that Assumption \ref{y-trigger} holds. Then the value $\delta \in \mathbb{R}$ defined in \eqref{delta} is positive.
\end{lem}
\begin{proof}
Note that from \eqref{MatrixEstimate} and \eqref{Q}, we have
\begin{equation*}
\begin{split}
m \phi_i (t) & = \sum_{j=1}^m [A(t:0)]_{ij} + \sum_{j=1}^m \Big( \phi_i (t) - [A(t:0)]_{ij}\Big)
\\
&\geq Q - mC_0 \lambda^t.
\end{split}
\end{equation*}
Using the above inequality and Lemma \ref{lem-3-4}, we deduce for each $1\leq i \leq m$ the following estimate
\begin{equation*}
\begin{split}
y_i (t+1) &\geq m_{\zeta}\phi_i (t) -\beta(t)
\\
& \geq (m_{\zeta}/m)Q - m_{\zeta} C_0 \lambda^t - \beta (t).
\end{split}
\end{equation*}
Since $\beta(t)$ converges to zero as $t$ goes to infinity and $\lambda \in (0,1)$, there exists a time $T\in \mathbb{N}$ and a constant $\tilde{\delta}>0$ such that for any $t\geq T$,
\begin{equation}\label{ddd}
y_i(t+1) \geq \tilde{\delta}.
\end{equation}
Note that by Assumption \ref{graph}, each matrix $A(t)$ has no zero row. This fact, together with the definition of $\hat{y}$ and \eqref{hatyy}, for any $t \in \mathbb{N}$ we have
\begin{equation}\label{eq-3-35}
\min_{1\leq i\leq m} y_i (t) >0.
\end{equation}
Therefore, combining \eqref{ddd} with \eqref{eq-3-35}, we conclude that $\delta$ defined in \eqref{delta} satisfies
\begin{equation*}
\delta \geq \min_{1\leq i \leq m}{\{y_i(0), y_i(1), \cdots, y_i(T), \tilde{\delta}\}}>0.
\end{equation*}
The proof is done.
\end{proof}
\subsection{Disagreement in Agent Estimates}
In this subsection, we derive a bound of the disagreement in agent estimates $\{\hat{z}_i (t)\}_{i=1}^m$ that will be used in the proofs of the main theorems. In the case without event-triggering $(\tau(t)=\zeta(t)=0)$, the paper \cite{Ned Push sum} proved that $\|\hat{z}_i(t+1) -\bar{x}(t)\|$ converges to zero for the stepsize satisfying Assumption \ref{stepsize} as $t$ goes to infinity. For the event-triggered case, the following proposition shows that the values $\{\hat{z}_i(t)\}_{i=1}^m$ approach $B_{\zeta}\bar{x}(t)$ instead of $\bar{x}(t)$ as $t$ goes to infinity due to the effect of the threshold $\zeta(t)$ for the triggering condition of $\{y_i(t)\}_{i=1}^m$.
\begin{prop}\label{mean}
For any $t \geq 1$ we have
\begin{equation*}
\begin{split}
&\|\hat{z}_i(t+1)-B_{\zeta}\bar{x}(t)\|\\
&\leq \frac{1}{\delta}\bigg(\Big( C_0 \lambda^t +K(t)\Big) \|x(0)\|_1
\\
&\quad + m\sum_{s=0}^{t-1} \Big[ C_0 \lambda^{t-s-1} +K(t)\Big] \Big( \alpha (s+1) D + \tau (s)\Big)\bigg) \\
&\qquad +\frac{1}{\delta}\bigg( d_i(t)\tau(t)\bigg),
\end{split}
\end{equation*}
where $K(t) = \beta(t)/m_{\zeta}$ and for $t=0$ we have
\begin{align*}
\|\hat{z}_i(1)-\bar{x}(0)\|\leq \frac{2C_0 }{\delta} \|x(0)\|,
\end{align*}
where the constant $\delta>0$ satisfies $y_i(t)>\delta$ for all $t>0$.
\end{prop}
To prove Proposition \ref{mean}, we consider a variable $w_i(t+1) \in \mathbb{R}^d$ which is a companion to the variable $\hat{w}_i (t+1) \in \mathbb{R}^n$ defined as
\begin{equation}\label{eq-3-10}
w_i(t+1) = \sum^m_{j=1} a_{ij} (t) x_j(t),
\end{equation}
and their difference
\begin{equation}\label{error}
e_i(t+1) = \hat{w}_i(t+1) - w_i(t+1).
\end{equation}
Then we may rewrite the gradient step \eqref{update} as
\begin{equation}\label{PEalgorithm}
\begin{split}
x_i(t+1)& = w_i(t+1) - \alpha(t+1) \nabla f_i(\hat{z}_i(t+1)) + e_i(t+1).
\end{split}
\end{equation} Summing up \eqref{PEalgorithm} for $1\leq i \leq m$ and using that $A(t)$ is column-stochastic, we have
\begin{equation}\label{eq-2-1a}
\bar{x}(t+1) = \bar{x}(t) - \frac{\alpha(t+1)}{m}\sum^m_{i=1}\nabla f_i(\hat{z}_i(t+1)) + \frac{1}{m}\sum^m_{i=1}e_i(t+1).
\end{equation}
Now we find a bound of $e_i(t+1)$ which is the difference between $w_i(t+1)$ and $\hat{w}_i(t+1)$ associated to the event-triggering $\tau(t)$.
\begin{lem}
The quantity $e_i(t+1)$ defined in \eqref{error} satisfies
\begin{equation}\label{error_estimate}
\|e_i(t+1)\| \leq d_i (t) \tau(t),
\end{equation}
where $d_i (t)= \sum_{j=1}^m a_{ij}(t)$. In addition, we have
\begin{equation}\label{error_estimate_2}
\sum_{i=1}^m \|e_i (t+1) \| \leq m \tau (t).
\end{equation}
\end{lem}
\begin{proof}
By using the triggering condition, we have
\begin{align*}
\|e_i(t+1)\| &\leq \|\hat{w}_i(t+1) -w_i(t+1)\| \\
&\leq \bigg\| \sum^m_{j=1} a_{ij}(t)(\hat{x}_j(t)-x_j(t)) \bigg\| \\
&\leq \sum^m_{j=1}a_{ij}(t) \|\hat{x}_j(t)-x_j(t)\| \leq d_i (t) \tau(t),
\end{align*}
which proves \eqref{error_estimate}. Summing this over $1\leq i \leq m$ and using that $A(t)$ is column stochastic, we find
\begin{equation*}
\begin{split}
\sum_{i=1}^m \|e_i (t+1) \| &\leq \sum_{i=1}^m \sum_{j=1}^m \Big(a_{ij} (t) \tau (t)\Big)
\\
&= \sum_{j=1}^m \Big( \sum_{i=1}^m a_{ij}(t)\Big) \tau (t) = m \tau (t).
\end{split}
\end{equation*}
The proof is finished.
\end{proof}
Now we are ready to prove Proposition \ref{mean}.
\begin{proof}[Proof of Proposition \ref{mean}]
We regard $x_k (t)$ as a row vector in $\mathbb{R}^{1 \times d}$ and define the variables $x(t)\in \mathbb{R}^{m \times d}$, $\nabla f(\hat{z}(t))\in \mathbb{R}^{m \times d}$, and $e (t) \in \mathbb{R}^{m \times d}$ as
\begin{equation*}
\medmath{
x(t) = \begin{pmatrix} x_1 (t) \\ \vdots \\ x_m (t) \end{pmatrix},~ \nabla f(\hat{z}(t)) = \begin{pmatrix} \nabla f_1 (\hat{z}_1 (t)) \\ \vdots \\ \nabla f_m (\hat{z}_m (t))\end{pmatrix},~ e (t) = \begin{pmatrix} e_1 (t) \\ \vdots \\ e_m (t) \end{pmatrix}.}
\end{equation*}
Note that by \eqref{error} and \eqref{hatzz}, we have
$$
\hat{z}_i(t+1) = \frac{w_i(t+1) +e_i(t+1)}{y_i(t+1)}.
$$
Also we see from definition \eqref{bzeta} that
$$
B_{\zeta}\bar{x}(t) =\frac{m\bar{x}(t)}{m_\zeta}=\frac{\mathbf{1}_m^Tx(t)}{m_\zeta}.
$$
Using these formulas and \eqref{eq-3-10} we have
\begin{equation}\label{eq-3-51}
\begin{split}
&\hat{z}_i (t+1) -B_{\zeta} \bar{x}(t)
\\
& = \frac{w_i (t+1) + e_i (t+1)}{y_i (t+1)} - \frac{\mathbf{1}_m^T x(t)}{m_{\zeta}}
\\
& = \frac{1}{y_i (t+1)} \bigg( [A(t)x(t)]_i - y_{i} (t+1) \frac{1_m^T x(t)}{m_{\zeta}}\bigg) + \frac{e_i (t+1)}{y_i (t+1)}.
\end{split}
\end{equation}
To estimate the first term on the right hand side of the last equality, we rewrite \eqref{PEalgorithm} as
\begin{equation*}
{x}(t+1) = A(t) {x}(t) - \alpha(t+1) \nabla f(\hat{z}(t+1)) + e(t+1).
\end{equation*}
Using this formula recursively, for $t \geq 1$ we have
\begin{equation}\label{Ax}
A (t) {x}(t)=A(t:0) {x}(0)-\sum^{t-1}_{s=0}A(t:s+1)\varepsilon(s),
\end{equation}
where we have let
$$\varepsilon(s)=\alpha(s+1)\nabla f(\hat{z}(s+1))-e(s+1).$$
Using Assumption \ref{GB} and \eqref{error_estimate_2} we have the following bound
\begin{equation}\label{eq100}
\|\varepsilon(s)\|_1 \leq m\Big(\alpha(s+1)D +\tau (s)\Big).
\end{equation}
Since $A(t)$ is column stochastic we have $1_m^T A(t) = 1_m^T$, and combine this with \eqref{Ax} to have
\begin{equation}\label{barx}
\mathbf{1}_m^T x(t) = \mathbf{1}_m^T x(0) -\sum^{t-1}_{s=0} \mathbf{1}_m^T \varepsilon(s).
\end{equation}
Combining \eqref{Ax} and \eqref{barx} yields
\begin{equation}\label{important}
\begin{split}
A(t) x(t)&= \phi(t)\mathbf{1}_m^T x(t)+\Big(A(t:0)-\phi (t)\mathbf{1}_m^T\Big)x(0)
\\
&\quad -\sum^{t-1}_{s=0}\Big(A(t:s+1)-\phi (t)\mathbf{1}_m^T\Big)\varepsilon (s),
\end{split}
\end{equation}
where $\phi (t)$ is the stochastic vector satisfying \eqref{MatrixEstimate}. By Lemma \ref{lem-3-4}, for $y(t):= (y_1 (t), \cdots, y_m (t))^T \in \mathbb{R}^{m \times 1}$ we have
\begin{equation*}
y(t+1) = m_{\zeta}\phi(t) + r(t),
\end{equation*}
where $r(t)$ satisfies $\|r(t)\|_{\infty} \leq \beta(t)$.
Combining this with \eqref{important}, we obtain
\begin{equation*}
\begin{split}
& [A(t)x(t)]_i - y_i (t+1) \frac{1_m^T x(t)}{m_{\zeta}}
\\
& = \phi_i (t) 1_m^T x(t) + [(A(t:0)- \phi (t)1_m^T) x(0)]_i
\\
&\qquad - \sum_{s=0}^{t-1} \Big[(A(t:s+1) - \phi (t) 1_m^T) \varepsilon (s)\Big]_i
\\
&\qquad -[m_{\zeta} \phi_i (t) +r_i (t)] \frac{1_m^T x(t)}{m_{\zeta}}
\\
& = [(A(t:0)- \phi (t)\mathbf{1}_m^T) x(0)]_i
\\
&\qquad- \sum_{s=0}^{t-1} \Big[(A(t:s+1) - \phi (t) \mathbf{1}_m^T)\varepsilon (s)\Big]_i - r_i (t) \frac{1_m^T x(t)}{m_{\zeta}}.
\end{split}
\end{equation*}
By applying \eqref{MatrixEstimate} here, we deduce
\begin{equation}\label{eq-3-50}
\begin{split}
&\Bigg\| [A(t)x(t)]_i -y_i (t+1) \frac{1_m^T x(t)}{m} \Bigg\|
\\
&\leq C_0 \lambda^t \|x(0)\|_1 + \sum_{s=0}^{t-1} C_0 \lambda^{t-s-1} \|\varepsilon (s)\|_1 + K(t) \|1_m^T x(t)\|,
\end{split}
\end{equation}
where $K(t) = \beta(t)/m_{\zeta}$.
From \eqref{barx} we find the following estimate
\begin{equation*}
\|1_m^T x(t)\| \leq \|x(0)\|_1 + \sum_{s=0}^{t-1} \|\varepsilon (s)\|_1.
\end{equation*}
Combining this with \eqref{eq-3-50} and using \eqref{eq100}, we obtain
\begin{equation}\label{eq-3-52}
\begin{split}
&\bigg\| [A(t)x(t)]_i -y_i (t+1) \frac{1_m^T x(t)}{m} \bigg\|
\\
&\leq C_0 \lambda^t \|x(0)\|_1 + \sum_{s=0}^{t-1} C_0 \lambda^{t-s-1} \|\varepsilon (s)\|_1
\\
&\quad+ K(t) \Big( \|x(0)\|_1 + \sum_{s=0}^{t-1} \|\varepsilon (s)\|_1\Big)
\\
&\leq \Big( C_0 \lambda^t + K(t)\Big) \|x(0)\|_1
\\
&\quad + m\sum_{s=0}^{t-1} \Big[ C_0 \lambda^{t-s-1} + K(t)\Big] \Big( \alpha (s+1) D + \tau (s)\Big).
\end{split}
\end{equation}
By applying Lemma \ref{lowerbound}, \eqref{error_estimate} and the above inequality to the norm of \eqref{eq-3-51}, we obtain
\begin{align*}
&\|\hat{z}_i(t+1)-B_{\zeta}\bar{x}(t)\|
\\
&\leq \frac{1}{\delta}\bigg(\Big( C_0 \lambda^t +K(t)\Big) \|x(0)\|_1
\\
&\quad + m\sum_{s=0}^{t-1} \Big[ C_0 \lambda^{t-s-1}+K(t) \Big] \Big( \alpha (s+1) D + \tau (s)\Big)\bigg)\\
&\qquad +\frac{1}{\delta}\bigg( d_i(t)\tau(t)\bigg).
\end{align*}
It remains to estimate the case $t=0$. By the algorithm, we have
\begin{equation*}
\begin{split}
\hat{z}_i (1) - \bar{x}(0) & = \frac{\hat{w}_i (1)}{y_i (1)} - \bar{x}(0) = \frac{\sum_{j=1}^m a_{ij} (0) x_j (0)}{\sum_{j=1}^m a_{ij}(0)} - \bar{x}(0).
\end{split}
\end{equation*}
Using this we find
\begin{equation*}
\begin{split}
\|z_i (1) -\bar{x}(0)\| &\leq \frac{1}{\delta} \|x(0)\|_1 + \frac{1}{m} \|x(0)\|_1
\\
& \leq 2 \|x(0)\|_1 \leq \frac{2C_0 }{\delta} \|x(0)\|.
\end{split}
\end{equation*}
The proof is finished.
\end{proof}
By utilizing Proposition \ref{mean}, we analyze the relation between $\hat{z}_i(t)$ and $B_{\zeta}\bar{x}(t)$ under the assumptions on $\{\alpha (t)\}_{t \in \mathbb{N}}$, $\{\tau(t)\}_{t \in \mathbb{N}}$ and $\{\zeta(t)\}_{t \in \mathbb{N}}$ of the main theorems. To do this, we first recall a useful lemma from \cite{R Ned V - stochastic}.
\begin{lem}[\cite{R Ned V - stochastic}, $Lemma$ 3.1]\label{lemma5}
If $\lim_{k\rightarrow \infty}\gamma_k=\gamma$ and $0<\beta<1$, then
$$
\lim_{k\rightarrow \infty} \sum^k_{l=0}\beta^{k-l}\gamma_l = \frac{\gamma}{1-\beta}
$$
\end{lem}
\begin{cor}\label{convergent}
Suppose that the triggering function $\tau$ satisfies Assumption \ref{event-triggered}. Also, assume that the stepsize $\alpha (t)$ satisfies Assumption \ref{stepsize} or $\alpha (t) =1/\sqrt{t}$. Then we have
$$
\lim_{t\rightarrow \infty} \|\hat{z}_i(t+1)-B_{\zeta}\bar{x}(t)\| = 0 \ \text{for all $i$}.
$$
\end{cor}
\begin{proof}
We recall from Proposition \ref{mean} the following inequality
\begin{equation}\label{eq-3-20}
\begin{split}
&\|\hat{z}_i(t+1)-B_{\zeta}\bar{x}(t)\|\\
&\leq \frac{1}{\delta}\bigg(\Big( C_0 \lambda^t +K(t)\Big) \|x(0)\|_1
\\
&\quad + m \sum_{s=0}^{t-1} \Big[ C_0 \lambda^{t-s-1} +K(t)\Big] \Big( \alpha (s+1) D + \tau (s)\Big)\bigg) \\
&\qquad +\frac{1}{\delta}\bigg(d_i(t)\tau(t)\bigg).
\end{split}
\end{equation}
We notice that $\lim_{t \rightarrow \infty} t^{3/2} K(t) = 0$ by Lemma \ref{lem-3-4}. From this and the boundedness of $\alpha (s)$ and $\tau (s)$, it easily follows that
\begin{equation*}
\lim_{t \rightarrow \infty} \frac{1}{\delta}K(t) \|x(0)\|_1 + m \sum_{s=0}^{t-1} K(t) \Big(\alpha (s+1) D + \tau (s)\Big) = 0.
\end{equation*}
In addition, by Assumptions \ref{stepsize}, \ref{event-triggered} and \ref{y-trigger}, we know that $
\lim_{s\rightarrow \infty}\alpha(s+1) =0$, $\lim_{s\rightarrow \infty} \tau(s)=0
$ and $\lim_{s\rightarrow \infty} \zeta(s)=0$. Using this fact with Lemma \ref{lemma5} in the right hand side of \eqref{eq-3-20}, we deduce
$$
\lim_{t\rightarrow \infty} \|\hat{z}_i(t+1)-B_{\zeta}\bar{x}(t)\| = 0,
$$
which completes the proof.
\end{proof}
\begin{cor}\label{specific} If Assumptions 2.4 holds, and the stepsize is chosen as $\alpha(t)= \frac{1}{\sqrt{t}}$, then we have
\begin{align*}
&\sum^T_{t=0}\alpha(t+1)\|\hat{z}_i(t+1)-B_{\zeta}\bar{x}(t)\|\\
& \leq \frac{C_0 }{\delta(1-\lambda)} \|x(0)\|_1 + \frac{4mC_0 E_{\tau}(T)}{\delta(1-\lambda)} + \frac{C_0mD}{\delta(1-\lambda)} (1+\ln(T))
\\
&~ + \frac{1}{\delta}\sum_{t=0}^T K(t) \alpha (t+1) \Big[ \|x(0)\|_1 + \sum_{s=0}^{t-1} (\alpha (s+1) D + \tau (s))\Big].
\end{align*}
\end{cor}
\begin{proof}
By Proposition \ref{mean}, we have
\begin{equation}\label{eq-3-1}
\begin{split}
&\delta \sum_{t=0}^T \alpha (t+1) \|\hat{z}_i (t+1) -B_{\zeta} \bar{x}(t)\|
\\
&\leq \|x(0)\|_1 \sum_{t=0}^T (C_0 \lambda^t +K(t)) \alpha (t+1)
\\
&\quad + m \sum_{t=0}^T \alpha (t+1) \sum_{s=0}^{t-1} (C_0 \lambda^{t-s-1} +K(t)) (\alpha (s+1) D + \tau (s))
\\
&\quad + \sum_{t=0}^T \alpha (t+1) (d_i (t) \tau (t)).
\end{split}
\end{equation}
The terms involving $K(t)$ are fit to the inequality of the lemma. Let us estimate each summation not involving $K(t)$ in the right hand side. Using that $\alpha (t) \leq 1$, the first term is bounded with
\begin{equation}\label{eq-3-30}
\sum^T_{t=0} \alpha(t+1)\lambda^t \leq \sum^T_{t=0} {\lambda^t} \leq \frac{1}{1-\lambda}.
\end{equation}
The fourth term is bounded using
\begin{equation}\label{eq-3-31}
\sum_{t=0}^{T}\alpha (t+1) \frac{d_i(t)\tau (t)}{\delta} \leq \frac{m}{\delta} \sum_{t=0}^{T} \tau (t) = \frac{mE_{\tau}(T)}{\delta}.
\end{equation}
We estimate the second term using
\begin{equation}\label{eq-3-32}
\begin{split}
\sum^T_{t=0} \alpha(t+1)\sum^{t-1}_{s=0}\lambda^{t-s-1}\alpha(s+1) &= \sum^{T+1}_{t=1}\frac{1}{\sqrt{t}}\sum^{t}_{s=1}\lambda^{t-s} \frac{1}{\sqrt{s}}\\
&\leq \sum^{T+1}_{t=1}\sum^t_{s=1}\lambda^{t-s}\frac{1}{s}\\
&=\sum^{T+1}_{s=1}\frac{1}{s}\sum^{T+1}_{t=s}\lambda^{t-s}\\
&\leq \frac{1+\ln{(T+1)}}{1-\lambda}
\end{split}
\end{equation}
In order to estimate the third term, we estimate
\begin{equation*}
\begin{split}
&\sum_{s=0}^{t-1} \lambda^{t-s-1}\tau (s)
\\
&= \sum_{s=0}^{[(t-1)/2]} \lambda^{t-s-1} \tau (s) + \sum_{s=[(t-1)/2]+1}^{t-1} \lambda^{t-s-1} \tau (s)
\\
& \leq \lambda^{(t-1)/2} \sum_{s=0}^{[(t-1)/2]} \tau (s) + \tau ([t/2] ) \sum_{s=[(t-1)/2]+1}^{t-1} \lambda^{t-s-1}
\\
&\leq \lambda^{(t-1)/2} E_{\tau} (T) + \frac{\tau ([t/2])}{1-\lambda},
\end{split}
\end{equation*}
where $[a]$ denotes the largest integer not larger than $a \in \mathbb{R}$.
Using this we derive
\begin{equation}\label{eq-3-33}
\begin{split}
&\sum_{t=1}^{T}\alpha (t+1) \Big[ \sum_{s=0}^{t-1} \lambda^{t-s-1} \tau (s)\Big]
\\
& \leq E_{\tau}(T) \sum_{t=1}^T \frac{\lambda^{(t-1)/2}}{\sqrt{t+1}} +\frac{1}{1-\lambda}\sum_{t=1}^{T} \frac{\tau ([t/2])}{\sqrt{t+1}}
\\
& \leq \frac{E_{\tau}(T)}{1-\sqrt{\lambda}} + \frac{E_{\tau}(T)}{1-\lambda} < \frac{3E_{\tau}(T)}{1-\lambda}.
\end{split}
\end{equation}
Putting the above estimates \eqref{eq-3-30}-\eqref{eq-3-33} in \eqref{eq-3-1}, we obtain
\begin{align*}
&\sum^T_{t=0}\alpha(t+1)\|\hat{z}_i(t+1)-B_{\zeta}\bar{x}(t)\|\\
& \leq \frac{C_0 }{\delta(1-\lambda)} \|x(0)\|_1 + \frac{4mC_0 E_{\tau}(T)}{\delta(1-\lambda)} + \frac{C_0mD}{\delta(1-\lambda)} (1+\ln(T))
\\
&~ + \frac{1}{\delta}\sum_{t=0}^T K(t) \alpha (t+1) \Big[ \|x(0)\|_1 + \sum_{s=0}^{t-1} (\alpha (s+1) D + \tau (s))\Big].
\end{align*}
which finishes the proof.
\end{proof}
\section{Convergence estimates}
In this section we prove our main results, namely Theorems \ref{maintheorem} and \ref{case2_main}. In Section 3, we obtained the bound of the disagreement in agent estimates. Especially, Corollary \ref{convergent} and \ref{specific} investigate the difference between the state $\hat{z}_i(t)$ in the Algorithm \ref{algo} and $B_{\zeta}\bar{x}(t)$ in \eqref{eq-2-1a}. Based upon these results, Theorem \ref{maintheorem} and \ref{case2_main} can be proved by comparing the cost values computed at the points $B_{\zeta}\bar{x}(t)$ and $x^*$.
\begin{lem}\label{submain}
Suppose Assumption \ref{GB} holds. Then for any $t \geq 0$ and $x \in \mathbb{R}^d$ we have
\begin{align}\label{combine}
&\sum_{i=1}^m \big(f_i(B_{\zeta}\bar{x}(t))-f_i(x)\big) \\
\nonumber&\leq \frac{m}{2\alpha(t+1)B_{\zeta}}(\|B_{\zeta}\bar{x}(t)-x\|^2-\|B_{\zeta}\bar{x}(t+1)-x\|^2)
\\
\nonumber&~+ \frac{B_{\zeta}m}{2\alpha(t+1)}\big(2\alpha(t+1)^2 D^2 + 2 \tau(t)^2\big)\\
\nonumber&~ + \frac{m}{\alpha(t+1)}\|B_{\zeta}\bar{x}(t)-x\|\tau(t)+ 2D\sum_{i=1}^{m} \|\hat{z}_i (t+1) -B_{\zeta}\bar{x}(t)\|.
\end{align}
\end{lem}
\begin{proof}
By convexity, we have
\begin{align*}
&f_i(\hat{z}_i(t+1))
\\
&\leq f_i(x) + (\hat{z}_i(t+1)-x)\nabla f_i(\hat{z}_i(t+1))\\
&= f_i(x) + (B_{\zeta}\bar{x}(t)-x)\nabla f_i(\hat{z}_i(t+1))
\\
&\quad + (\hat{z}_i(t+1)-B_{\zeta}\bar{x}(t))\nabla f_i(\hat{z}_i(t+1))\\
&= f_i(x) + \frac{1}{\alpha(t+1)} (B_{\zeta}\bar{x}(t)-x)(w_i(t+1)-x_i(t+1)
\\
&\quad +e_i(t+1))+ (\hat{z}_i(t+1)-B_{\zeta}\bar{x}(t))\nabla f_i(\hat{z}_i(t+1)),
\end{align*}
where \eqref{eq-3-10} is used in the last equality.
Summing up the above inequality from $i=1$ to $i=m$, we find that
\begin{align*}
&\sum_{i=1}^m f_i(\hat{z}_i(t+1))-f_i(x)
\\
&\leq \underbrace{\frac{m}{\alpha(t+1)}(B_{\zeta}\bar{x}(t)-x)(\bar{x}(t)-\bar{x}(t+1))}_{\uppercase\expandafter{\romannumeral1}}\\
&\quad + \underbrace{\frac{1}{\alpha(t+1)}(B_{\zeta}\bar{x}(t)-x)\sum^m_{i=1}e_i(t+1)}_{\uppercase\expandafter{\romannumeral2}} \\
&\quad + \underbrace{\sum^m_{i=1}(\hat{z}_i(t+1)-B_{\zeta}\bar{x}(t))\nabla f_i(\hat{z}_i(t+1))}_{\uppercase\expandafter{\romannumeral3}}.
\end{align*}
Now we estimate each term in the right hand side. First using the equality $\langle a,b\rangle=\frac{1}{2}(\|a\|^2+\|b\|^2-\|a-b\|^2)$ for $a,b \in \mathbb{R}^d$, we have
\begin{equation*}
\begin{split}
\uppercase\expandafter{\romannumeral1}& = \frac{m}{2\alpha(t+1)B_{\zeta}}(\|B_{\zeta}\bar{x}(t)-x\|^2-\|B_{\zeta}\bar{x}(t+1)-x\|^2
\\
&\quad \qquad+\|B_{\zeta}\bar{x}(t+1)-B_{\zeta}\bar{x}(t)\|^2).
\end{split}
\end{equation*}
Using \eqref{eq-2-1a} along with \eqref{error_estimate} and \eqref{eq-2-1}, we estimate the right-most term as
\begin{align*}
&\|\bar{x}(t+1)-\bar{x}(t)\|^2
\\
&\leq 2\bigg\|\frac{\alpha(t+1)}{m}\sum^m_{i=1}\nabla f_i(\hat{z}_i(t+1)) \bigg\|^2 + 2\bigg\|\frac{1}{m}\sum^m_{i=1}e_i(t+1) \bigg\|^2\\
&\leq 2 \alpha(t+1)^2 D^2 +2\tau(t)^2.
\end{align*}
We apply \eqref{error_estimate} again to estimate
\begin{align*}
\uppercase\expandafter{\romannumeral2} \leq \frac{m}{\alpha(t+1)}\|B_{\zeta}\bar{x}(t)-x\|\tau(t),
\end{align*}
and use \eqref{eq-2-1} to deduce
\begin{equation*}
\uppercase\expandafter{\romannumeral3} \leq D\sum_{i=1}^{m} \|\hat{z}_i (t+1) -B_{\zeta}\bar{x}(t)\|.
\end{equation*}
Combining the above estimates on $\uppercase\expandafter{\romannumeral1}$,$\uppercase\expandafter{\romannumeral2}$ and $\uppercase\expandafter{\romannumeral3}$, we have
\begin{equation}\label{eq-4-3}
\begin{split}
&\sum_{i=1}^m \big(f_i(\hat{z}_i(t+1))-f_i(x)\big) \\
\nonumber&\leq \frac{m}{2\alpha(t+1)B_{\zeta}}(\|B_{\zeta}\bar{x}(t)-x\|^2-\|B_{\zeta}\bar{x}(t+1)-x\|^2)
\\
\nonumber&\quad + \frac{mB_{\zeta}}{2\alpha(t+1)}\big(2\alpha(t+1)^2 D^2 + 2 \tau(t)^2\big)\\
\nonumber&\quad + \frac{m}{\alpha(t+1)}\|B_{\zeta}\bar{x}(t)-x\|\tau(t)+ D\sum_{i=1}^{m} \|\hat{z}_i (t+1) -B_{\zeta}\bar{x}(t)\|.
\end{split}
\end{equation}
Finally we observe that \eqref{eq-2-1} gives us the estimate
\begin{equation*}
\sum_{i=1}^m \Big( f_i (B_{\zeta}\bar{x}(t)) -f_i (\hat{z}_i (t+1))\Big) \leq D\sum_{i=1}^m \|B_{\zeta}\bar{x}(t) - \hat{z}_i (t+1)\|.
\end{equation*}
Summing up the above two inequalities, we obtain the desired estimate.
\end{proof}
\subsection{Proof of Theorem \ref{maintheorem}}
We recall the following lemma for proving Theorem \ref{maintheorem}.
\begin{lem}[\cite{Ned Push sum} Lemma 7]\label{lemma7}
Consider a minimization problem $\min_{x\in\mathbb{R}^d}f(x)$, where $f:\mathbb{R}^d\rightarrow \mathbb{R}$ is a continuous function. Assume that the solution $X^*$ of the problem is nonempty. Let $\{x(t)\}$ be a sequence such that for all $x\in X^*$ and for all $t\geq 0$,
$$
\|x(t+1)-x\|^2\leq (1+b(t))\|x(t) - x \|^2- a(t)(f(x(t))-f(x))+c(t)
$$
where $b(t)\geq 0$, $a(t)\geq 0$ and $c(t)\geq 0$ for all $t\geq 0$ with $\sum^\infty_{t=0}b(t) < \infty$,$\sum^\infty_{t=0}a(t) = \infty$ and $\sum^\infty_{t=0}c(t)<\infty$. Then the sequence $\{x(t)\}$ converges to some solution $x^*\in X^*$
\end{lem}
By manipulating the estimate in Lemma \ref{submain}, we obtain the following estimate which is suitable for applying Lemma \ref{lemma7}.
\begin{cor}\label{main}
Under the same assumptions in Lemma \ref{submain} we have
\begin{equation*}
\begin{split}
&\|B_{\zeta}\bar{x}(t+1) -x\|^2 \leq (1+\tau(t))\|B_{\zeta}\bar{x}(t) -x\|^2
\\
&\quad -\frac{2\alpha(t+1)B_{\zeta}}{m} (f(B_{\zeta}\bar{x}(t)) - f(x)) + c(t)+d(t),
\end{split}
\end{equation*}
where
\begin{equation*}
c(t) = \Big[2\alpha(t+1)^2 D^2 + 2 \tau(t)^2 + \tau(t)\Big]B_{\zeta},
\end{equation*}
and
\begin{equation*}
d(t) = \frac{4\alpha (t+1)B_{\zeta}D}{m}\sum_{i=1}^m \|\hat{z}_i (t+1) -\bar{x}(t)\|.
\end{equation*}
\end{cor}
\begin{proof}
We use Young's inequality to find
\begin{equation*}
\|B_{\zeta}\bar{x}(t)-x\|\tau(t) \leq \frac{\|B_{\zeta}\bar{x}(t)-x\|^2 \,\tau(t) }{2B_{\zeta}} + \frac{\tau(t)B_{\zeta}}{2}.
\end{equation*}
Applying this to \eqref{combine}, we get
\begin{align}\label{eq-3-2}
&\quad\sum_{i=1}^m \big(f_i(B_{\zeta}\bar{x}(t))-f_i(x)\big) \\
\nonumber&\leq \frac{m}{2\alpha(t+1)B_{\zeta}}(1+\tau(t))\|B_{\zeta}\bar{x}(t)-x\|^2
\\
\nonumber &\quad -\frac{m}{2\alpha (t+1)B_{\zeta}}\|B_{\zeta}\bar{x}(t+1)-x\|^2 \\
\nonumber&\quad + \frac{mB_{\zeta}}{2\alpha(t+1)}\big(2\alpha(t+1)^2 D^2 + 2 \tau(t)^2\big)+ \frac{B_{\zeta}m \tau(t)}{2\alpha(t+1)}
\\
\nonumber&\quad + 2D \sum_{i=1}^m \|\hat{z}_i (t+1) - B_{\zeta}\bar{x}(t)\|.
\end{align}
Dividing both sides by $\frac{m}{2\alpha (t+1) B_{\zeta}}$, it follows that
\begin{align}\label{eq-3-4}
&\frac{2\alpha (t+1)B_{\zeta}}{m}\sum_{i=1}^m \big(f_i(B_{\zeta}\bar{x}(t))-f_i(x)\big) \\
\nonumber&\leq (1+\tau(t))\|B_{\zeta}\bar{x}(t)-x\|^2- \|B_{\zeta}\bar{x}(t+1)-x\|^2 \\
\nonumber&\quad + B_{\zeta}^2\Big[ 2\alpha(t+1)^2 D^2 + 2 \tau(t)^2 + \tau(t) \Big]
\\
\nonumber&\quad + \frac{4\alpha (t+1)DB_{\zeta}}{m} \sum_{i=1}^m \|\hat{z}_i (t+1) - B_{\zeta}\bar{x}(t)\|.
\end{align}
Rearranging this we obtain the desired estimate.
\end{proof}
Now we are ready to prove Theorem \ref{maintheorem}.
\begin{proof}[Proof of Theorem \ref{maintheorem}]
By Lemmas \ref{lemma7} and Corollary \ref{main} it is enough to prove $\sum_{t=1}^{\infty}(c(t) +d(t)) <\infty$, where
\begin{equation*}
c(t) = 2\alpha(t+1)^2 D^2 + 2 \tau(t)^2 + \tau(t),
\end{equation*}
and
\begin{equation*}
d(t) = \frac{4\alpha (t+1)DB_{\zeta}}{m}\sum_{i=1}^m \|\hat{z}_i (t+1) -B_{\zeta}\bar{x}(t)\|.
\end{equation*}
It follows that $\sum^\infty_{t=0}c(t)<\infty$ by Assumptions \ref{stepsize} and \ref{event-triggered}.
Next we will show that $\sum^\infty_{t=0} d(t) <\infty.$ By Proposition \ref{mean}, it suffices to show that
\begin{equation}\label{eq-4-10}
\sum^\infty_{t=1}\alpha(t+1)\bigg( \lambda^t +\tau(t) + \sum^{t-1}_{s=0}\lambda^{t-s-1}\alpha(s+1) + \sum^{t-1}_{s=0}\lambda^{t-s-1}\tau(s) \bigg)<\infty,
\end{equation}
and
\begin{equation*}
\sum_{t=1}^{\infty}\alpha (t+1)\Big[K(t)+\sum_{s=0}^{t-1} K(t) \Big(\alpha (s+1) D + \tau (s)\Big)\Big]<\infty.
\end{equation*}
The latter one is proved in Lemma \ref{lem-4-4} below. We proceed to prove \eqref{eq-4-10}.
Using the Cauchy-Schwarz inequality, we have
\begin{align*}
\sum^\infty_{t=1}\alpha(t+1)\lambda^t \leq \frac{1}{2}\sum^\infty_{t=1}\alpha(t+1)^2 + \frac{1}{2}\sum^\infty_{t=1}\lambda^{2t} <\infty.
\end{align*}
By rearranging and using the decreasing property of $\alpha (t)$ in Assumption \ref{stepsize}, we find
\begin{equation*}
\begin{split}
&\sum^{\infty}_{t=1}\alpha(t+1)\sum^{t-1}_{s=0} \lambda^{t-s-1}\alpha(s+1)
\\
& = \sum_{s=0}^{\infty} \sum_{t=s+1}^{\infty} \lambda^{t-s-1} \alpha (t+1) \alpha (s+1)
\\
&\leq \sum_{s=0}^{\infty} \Big( \sum_{t=s+1}^{\infty} \lambda^{t-s-1} \Big) \alpha (s+1)^2
\\
& = \frac{1}{1-\lambda} \sum_{s=0}^{\infty} \alpha (s+1)^2 < \infty.
\end{split}
\end{equation*}
Similarly, due to Assumption \ref{event-triggered}, the last term is bounded as
\begin{equation*}
\begin{split}
\sum^\infty_{t=0}\alpha(t+1)\sum^{t-1}_{s=0}\lambda^{t-s-1}\tau(s) & = \sum_{s=0}^{\infty}\sum_{s<t} \lambda^{t-s-1} \alpha (t+1) \tau (s)
\\
&\leq \sum_{s=0}^{\infty} \sum_{s<t} \lambda^{t-s-1} \alpha (s+1) \tau (s)
\\
& = \frac{1}{1-\lambda} \sum_{s=0}^{\infty} \alpha (s+1) \tau (s) <\infty.
\end{split}
\end{equation*}
Gathering the above estimates, we find that $\sum_{t=1}^{\infty}(c(t) + d(t)) <\infty$. Hence by Lemma \ref{lemma7}, the sequence $\{B_{\zeta}\bar{x}(t)\}$ converges to some solution $x^*\in X^*$. Finally, we apply Corollary \ref{convergent} to conclude that each sequence $\{\hat{z}_i(t)\}$, $i=1,\cdots,n,$ converges to the same solution $x^*$. The proof is done.
\end{proof}
\begin{lem}\label{lem-4-4}Suppose that Assumption \ref{y-trigger} and Assumption \ref{event-triggered} hold. Then for the stepsize $\{\alpha(t+1)\}_{t\geq 0}$ satisfying Assumption \ref{stepsize} or $\alpha (t) =1/\sqrt{t}$, we have
\begin{equation*}
\sum_{t=1}^{\infty}\alpha (t+1) \Big[ K(t) + \sum_{s=0}^{t-1} K(t) \Big( \alpha (s+1) + \tau (s)\Big) < \infty.
\end{equation*}
\end{lem}
\begin{proof}
From Lemma \ref{lem-3-4}, we know that $\lim_{t \rightarrow \infty} t^{3/2} K(t) = 0$. Using this fact the summability of $\tau (s)$, it easily follows that
\begin{equation*}
\sum_{t=1}^{\infty} \alpha (t+1)\Big[ K(t) + \Big( \sum_{s=0}^{t-1} \tau (s)\Big) K(t)\Big] < \infty.
\end{equation*}
Next, for $\alpha (t)$ satisfying Assumption \ref{stepsize}, we observe that
\begin{equation*}
\begin{split}
\sum_{t=1}^{\infty}\alpha (t+1) K(t) \sum_{s=0}^{t-1} \alpha (s+1) & \leq \sum_{t=1}^{\infty} K(t) \sum_{s=0}^{t-1} \alpha (s+1)^2 <\infty.
\end{split}
\end{equation*}
For $\alpha (t) = 1/\sqrt{t}$, by using that $\sum_{s=0}^{t-1} 1/\sqrt{s+1} \leq 2 \sqrt{t}$, we deduce
\begin{equation*}
\begin{split}
\sum_{t=1}^{\infty}\alpha (t+1) K(t) \sum_{s=0}^{t-1} \alpha (s+1) & \leq 2\sum_{t=1}^{\infty} K(t) <\infty.
\end{split}
\end{equation*}
The proof is done.
\end{proof}
\subsection{Proof of Theorem \ref{case2_main}}
We now turn to the proof of Theorem \ref{case2_main}. we recall that
\begin{equation*}
H(-1) = 1\quad \textrm{and}\quad H(t) = \prod_{k=0}^{t} (1+\tau(k))\quad \textrm{for}\quad t\geq 1
\end{equation*}
and
\begin{equation*}
S(0) = 0\quad \textrm{and}\quad S(t) =\sum_{s=0}^{t-1} \frac{\alpha (s+1)}{H(s)}\quad \textrm{for}\quad t\geq 1.
\end{equation*}
First we find the boundedness of $H(t)$ and $S(t)$.
\begin{lem}\label{s(t+1)}
Let $\alpha(t)=\frac{1}{\sqrt{t}}$ and Assumption 2.4 hold. Then we have
\begin{equation}\label{eq-4-13}
\sup_{t\geq 0} H(t) < e^{E_{\tau}}\quad \textrm{and}\quad S(t)\geq e^{-E_{\tau}}\sqrt{t},
\end{equation}
where $E_{\tau} = \sum_{t=0}^{\infty}\tau (t) <\infty$.
\end{lem}
\begin{proof}
By applying the inequality $1+x \leq e^x$ for $x \geq 0$ we estimate $H(t)$ as
\begin{equation*}
\sup_{t\geq 0} H(t) <\exp{(\sum^\infty_{t=0}\tau(t))}=e^{E_{\tau}}.
\end{equation*}
Using this inequality, we deduce the following estimate
\begin{equation*}
\begin{split}
S(t)& = \sum_{s=0}^{t-1} \frac{\alpha (s+1)}{H(s)} \geq e^{-E_{\tau}} \sum_{s=1}^{t} \frac{1}{\sqrt{s}} \geq e^{-E_{\tau}} \sqrt{t}\quad \forall~ t \geq 1.
\end{split}
\end{equation*}
The proof is done.
\end{proof}
To prove Theorem \ref{case2_main}, we first modify Lemma \ref{submain} which states the boundedness of $\sum_{i=1}^m \big(f_i(\bar{x}(t))-f_i(x)\big)$, by replacing $\bar{x}(t)$ to $\frac{\sum^T_{t=0}\frac{\alpha(t+1)}{H(t)}B_\zeta\bar{x}(t)}{S(T+1)}$.
\begin{lem}\label{case2_submain}
Suppose that all the conditions are same as in Theorem \ref{case2_main}. Then we have
\begin{equation*}
\begin{split}
&f\bigg(\frac{\sum^T_{t=0}\frac{\alpha(t+1)}{H(t)}B_{\zeta}\bar{x}(t)}{S(T+1)}\bigg) - f(x)
\\
&\leq \frac{me^{E_{\tau}}}{2\sqrt{T+1}}J_1 (T)+ \frac{2mDe^{E_{\tau}}}{\delta \sqrt{T+1}}J_2 (T) + \frac{2mDe^{E_\tau}}{ \delta\sqrt{T+1}}J_3 (T)
\end{split}
\end{equation*}
for any $T \in \mathbb{N}$, where $J_1 (T), J_2 (T)$, and $J_3 (T)$ are defined in Theorem \ref{case2_main}
\end{lem}
\begin{proof}
We recall from \eqref{eq-3-2} the following inequality
\begin{align}\label{eq-3-2}
&\quad\sum_{i=1}^m \big(f_i(B_{\zeta}\bar{x}(t))-f_i(x)\big) \\
\nonumber&\leq \frac{m}{2\alpha(t+1)B_{\zeta}}(1+\tau(t))\|B_{\zeta}\bar{x}(t)-x\|^2
\\
\nonumber&\quad -\frac{m}{2\alpha (t+1)B_{\zeta}}\|B_{\zeta}\bar{x}(t+1)-x\|^2 \\
\nonumber&\quad + \frac{mB_{\zeta}}{2\alpha(t+1)}\big(2\alpha(t+1)^2 D^2 + 2 \tau(t)^2\big)
\\
\nonumber &\quad+ \frac{B_{\zeta}m \tau(t)}{2\alpha(t+1)} + 2D \sum_{i=1}^m \|\hat{z}_i (t+1) - B_{\zeta}\bar{x}(t)\|.
\end{align}
Dividing both sides of \eqref{eq-3-2} by $\frac{mH(t)}{2\alpha(t+1)}$, we get
\begin{align}\label{eq-3-3}
&\frac{2\alpha (t+1)}{mH(t)}\sum_{i=1}^m \big(f_i(B_{\zeta}\bar{x}(t))-f_i(x)\big) \\
\nonumber&\leq \frac{\|B_{\zeta}\bar{x}(t)-x\|^2}{B_{\zeta}H(t-1)}- \frac{\|B_{\zeta}\bar{x}(t+1)-x\|^2}{B_{\zeta}H(t)} \\
\nonumber&\quad + \frac{B_{\zeta}}{H(t)}\bigg(2\alpha(t+1)^2 D^2 + 2 \tau(t)^2+\tau(t)\Bigg)
\\
\nonumber&\quad + \frac{4\alpha (t+1)D}{mH(t)} \sum_{i=1}^m \|\hat{z}_i (t+1) -B_{\zeta} \bar{x}(t)\|.
\end{align}
Summing this from $t=0$ to $t=T$ we obtain
\begin{align}\label{eq415}
&\sum_{t=0}^{T}\Bigg[\frac{2\alpha (t+1)}{mH(t)}\sum_{i=1}^m \big(f_i(B_{\zeta}\bar{x}(t))-f_i(x)\big) \Bigg]\\
\nonumber&\leq \frac{\|\bar{x}(0)-x\|^2}{B_{\zeta}H(-1)}- \frac{\|\bar{x}(T+1)-x\|^2}{B_{\zeta}H(T)} \\
\nonumber&\quad + \sum_{t=1}^T \frac{B_{\zeta}}{H(t)}\bigg(2\alpha(t+1)^2 D^2 + 2 \tau(t)^2+\tau(t)\Bigg)
\\
\nonumber &\quad + \sum_{t=1}^T \frac{4\alpha (t+1)D}{mH(t)} \sum_{i=1}^m \|\hat{z}_i (t+1) - B_{\zeta}\bar{x}(t)\|.
\end{align}
This, together with the fact that $H(-1)=1$ and $H(t) \geq 1$, gives
\begin{equation}\label{eq-3-40}
\begin{split}
&\sum_{t=0}^{T}\Bigg[\frac{2\alpha (t+1)}{mH(t)}\sum_{i=1}^m \big(f_i(B_{\zeta}\bar{x}(t))-f_i(x)\big) \Bigg]\\
&\leq \frac{\|\bar{x}(0)-x\|^2}{B_{\zeta}} + \sum_{t=0}^T \bigg(2\alpha(t+1)^2 D^2 + 2 \tau(t)^2+\tau(t) \bigg)B_{\zeta}
\\
&\quad + \sum_{t=0}^T \bigg( \frac{4\alpha (t+1)D}{m} \sum_{i=1}^m \|\hat{z}_i (t+1) -B_{\zeta} \bar{x}(t)\|\bigg).
\end{split}
\end{equation}
We find
\begin{equation}\label{eq-3-41}
\sum^T_{t=0} 2\alpha(t+1)^2 D^2 = 2D^2 \sum^{T+1}_{t=1}\frac{1}{t}\leq 2D^2 \Big(1+ \ln{(T+1)}\Big),
\end{equation}
and by definition \eqref{eq-2-8} we have
\begin{equation}\label{eq-3-42}
\sum^T_{t=0} \Big(2\tau(t)^2 +\tau (t)\Big) = 2E_{\tau,2}(T)+E_{\tau}(T).
\end{equation}
Finally we estimate the last term of the right hand side of \eqref{eq-3-40} using Corollary \ref{specific} as follows:
\begin{equation*}
\begin{split}
&\sum_{t=0}^T \bigg( \frac{4\alpha (t+1)D}{m} \sum_{i=1}^m \|\hat{z}_i (t+1) - B_{\zeta}\bar{x}(t)\|\bigg) \\
&\medmath{\leq4D \Big( \frac{C_0 }{\delta(1-\lambda)} \|x(0)\|_1 + \frac{4mC_0 E_{\tau}(T)}{\delta(1-\lambda)} + \frac{C_0mD}{\delta(1-\lambda)} (1+\ln(T))\Big)}
\\
&\quad \medmath{+\frac{4D}{\delta} \sum_{t=0}^T K(t) \alpha (t+1) \Big[ \|x(0)\|_1 + \sum_{s=0}^{t-1} (\alpha (s+1) D + \tau (s))\Big].}
\end{split}
\end{equation*}
Putting this estimate, \eqref{eq-3-41} and \eqref{eq-3-42} in \eqref{eq-3-40}, we achieve the following estimate
\begin{align*}
&\sum_{t=0}^{T}\Bigg[\frac{2\alpha (t+1)}{mH(t)}\sum_{i=1}^m \big(f_i(B_{\zeta}\bar{x}(t))-f_i(x)\big) \Bigg]\\
\nonumber&\leq J_1 (T)+ \frac{4D}{\delta} J_2 (T) +\frac{4D}{\delta} J_3 (T).
\end{align*}
Now we set $S(T) = \sum_{t=0}^{T-1} \frac{\alpha (t+1)}{H(t)}$ for $T \in \mathbb{N}$ and divide the both sides by $\frac{2S(T+1)}{m}$. Then we apply the convexity of $f_i$ in the left hand side and use the lower bound $S(T+1) \geq e^{-E_{\tau}}{\sqrt{T+1}}$ to the right hand side, which leads to
\begin{equation*}
\begin{split}
&f\Bigg(\frac{\sum^T_{t=0}\frac{\alpha(t+1)}{H(t)}B_{\zeta}\bar{x}(t)}{S(T+1)}\Bigg) - f(x)
\\
&\leq \frac{me^{E_{\tau}}}{2\sqrt{T+1}}J_1 (T)+ \frac{2mDe^{E_{\tau}}}{\delta \sqrt{T+1}}J_2 (T) + \frac{2mDe^{E_\tau}}{ \delta\sqrt{T+1}}J_3 (T).
\end{split}
\end{equation*}
The proof is finished.
\end{proof}
Now we are ready to give the proof of Theorem \ref{case2_main}.
\begin{proof}[Proof of Theorem \ref{case2_main}]
Using the definition of $\tilde{z}_i$ and Assumption 2.1, we find
\begin{equation*}
\begin{split}
&f(\tilde{z}_i(T+1))-f\bigg(\frac{\sum^T_{t=0}\frac{\alpha(t+1)}{H(t)}B_{\zeta}\bar{x}(t)}{S(T+1)}\bigg)
\\
& =f\bigg(\frac{\sum_{t=0}^T \frac{\alpha (t+1)}{H(t)} \hat{z}_i (t+1)}{S(T+1)} \bigg)-f\bigg(\frac{\sum^T_{t=0}\frac{\alpha(t+1)}{H(t)}B_{\zeta}\bar{x}(t)}{S(T+1)}\bigg)
\\
&\leq \frac{mD}{S(T+1)}\sum^T_{t=0}\frac{\alpha(t+1)}{H(t)}\|\hat{z}_i(t+1)-B_{\zeta}\bar{x}(t)\|.
\end{split}
\end{equation*}
Then by Corollary \ref{specific} with the fact that $H(t) \geq 1$ and \eqref{eq-4-13}, we have
\begin{align*}
&f(\tilde{z}_i(T+1))-f\bigg(\frac{\sum^T_{t=0}\frac{\alpha(t+1)}{H(t)}B_{\zeta}\bar{x}(t)}{S(T+1)}\bigg)
\\
&\quad\leq \frac{mDe^{E_{\tau}}}{\delta \sqrt{T+1}}J_2 (T)
+ \frac{mDe^{E_{\tau}}}{\delta \sqrt{T+1}} J_3 (T).
\end{align*}
Combining this inequality with Lemma \ref{case2_submain}, we obtain
\begin{align*}
&f(\tilde{z}_i(T+1))-f(x^*) \\
&\leq \frac{me^{E_{\tau}}}{2\sqrt{T+1}}J_1 (T) + \frac{3mDe^{E_{\tau}}}{\delta \sqrt{T+1}}J_2 (T) + \frac{3mDe^{E_\tau}}{ \delta\sqrt{T+1}}J_3 (T).
\end{align*}
which is the desired estimate. Moreover, we see that the right hand side is bounded by $O(\log (T+1)/ \sqrt{T+1})$ using Lemma \ref{lem-4-4}.
\end{proof}
\section{Simulations}
In this section, we present simulation results of the proposed event-triggered gradient-push method to demonstrate that the theoretical results can be realized in practice. We consider the decentralized least squares problem:
\begin{equation*}
\min_{x\in\mathbb{R}^d}~\sum_{i=1}^m f_i (x)\quad \textrm{with}\quad f_i (x) =\|q_i - p^T_ix\|^2,
\end{equation*}
where, each agent $i$ in $\mathcal{V}= \{1,\cdots,m\}$ is given the local cost function $f_i$. The variable $p_i\in \mathbb{R}^{d\times p}$ is the input data and the variable $q_i\in\mathbb{R}^{p}$ is the output data. The data are generated according to the linear regression model $q_i = p_i^T \tilde{x} + \varepsilon_{i}$ where $\tilde{x}\in \mathbb{R}^d$ is the true weight vector, $p_i$'s are uniformly random values between 0 and 1, and $\varepsilon_i$'s are jointly Gaussian and zero mean. The initial points $x_i(0)$ are independent random variables, generated by a standard Gaussian distribution. In this simulation, we set the problem dimensions and the number of agents as $d=5 $, $p=1 $, and $m=50 $. We use connected directed graph where every node has four out-neighbors.
\medskip
\noindent \textbf{Test 1.} Here we fix $\alpha (t) = 1/t^{0.52}$ which satisfies Assumption \ref{stepsize} and $\zeta(t) = 1/(3t^3)$ which satisfies Assumption \ref{y-trigger}, and consider various choices of $\tau (t)$. We measure the relative distance between the states $z_i (t)$ and the optimal point $x^*$ the value
\begin{equation*}
R_d (t) = \frac{\sum_{i=1}^m \|z_i (t) - x^*\|}{\sum_{i=1}^m \|z_i (0) - x^*\|},
\end{equation*}
We set the termination time $k_f$ as the first time $k \in \mathbb{N}$ when $R_d (k) < 10^{-2}$. And we let $N_x$ and $N_y$ be the average of total number of triggers for all agents until the termination time associated with $\tau(t)$ and $\zeta(t)$, respectively. Table \ref{known results} indicates the average of those values depending on $\tau(t)$ and $\zeta(t)$ in 100 trials.
\begin{table*}
\begin{tabularx}{\textwidth}{@{}l*{10}{C}c}
\toprule
$\tau(t)$ & 0 & 0 &$1/ t^{1.1}$ & $1/t^{1.1}$ & $1/t^{1.3}$ & $1/t^{1.3}$ & $1/t^{1.5}$ & $1/t^{1.5}$ & $1/t^{1.7}$ &$1/t^{1.7}$ \\
$\zeta(t)$ & 0 & $1/ (3t^3)$ & 0 & $1/ (3t^3)$ & 0 & $1/ (3t^3)$ & 0 & $1/ (3t^3)$ & 0 & $1/ (3t^3)$ \\
\midrule
$N_x$ & 11425 & 11305 & 3125 & 3152 & 3299 & 3288 & 8860 & 8767 & 15537 & 11177 \\
\addlinespace
$N_y$ & 11425 & 26 & 72705 & 32 & 15696 & 27 & 11644 & 26 & 15572 & 26 \\
\addlinespace
$k_f$ & 11425 & 11305 & 72705 & 72757 & 15696 & 15572 & 11644 & 11514 & 15572 & 11207 \\
\bottomrule
\end{tabularx}
\vspace{0.1cm}
\caption{The number of triggers and termination time depending on $\tau(t)$ and $\zeta(t)$}
\label{known results}
\end{table*}
We first look at the effect of $\zeta(t)$, the threshold for variables $y_i(t)$. Table \ref{known results} shows that an existence of the threshold $(\zeta \neq 0)$ does not bring a big difference in the termination time if we compare the cases $\zeta (t) =0$ and $\zeta (t) = 1/(3t^3)$ with same $\tau (t)$, but there is a big improvement in the number of triggers for $y_i(t)$. Next we discuss the values $N_x$ and $k_f$ of Table \ref{known results} in terms of $\tau(t)$, the threshold for variables $x_i(t)$. As in Table \ref{known results}, some cases give us similar or worse results compared to the cases $\tau(t)=0$. For $\tau(t)=1/t^{1.1}$, the number of triggers is decreased by more than $70\%$, and the termination time increased by more than $530\%$. For $\tau(t) = 1/t^{1.7}$, both the termination time and the number of triggers increased by almost $36\%$ when $\zeta(t)=0$ and remain similar when $\zeta(t)=1/(3t^3)$ compared to the cases $\tau(t)=0$. For $\tau(t)=1/t^{1.5}$, the termination time is almost same, the number of triggers decreased by more than $20\%$. These results show that the proposed gradient-push with event-triggered communication with proper $\tau(t)$ and $\zeta(t)$ can diminish the number of communications to achieve the convergence compared to the gradient-push algorithm without triggering.
The threshold functions $\tau(t)$ and $\zeta(t)$ should be chosen carefully considering the characteristics of the given optimization problem.
\medskip
\begin{figure}[!htbp]
\centering
\includegraphics[width=7.4cm]{ft_error.png}
\hfil
\caption{The values of $R_f (t)$ with different choices of $\tau (t)$.}
\label{figure1}
\end{figure}
\begin{figure}[!htbp]
\centering
\includegraphics[width=7.4cm]{con_error.png}
\hfil
\caption{The values of $R_c (t)$ with different choices of $\tau (t)$.}
\label{figure1}
\end{figure}
\noindent \textbf{Test 2.} Here we fix $\alpha (t) =1/\sqrt{t}$ and take several choices of $\tau (t)$ and $\zeta(t)$. We measure the relative cost error and the consensus error given by
\begin{equation*}
R_f (t) = \sum_{i=1}^m ( f(\tilde{z}_i (t)) -f^*),\quad
R_c (t) = \max_{i,j \in V} \|z_i (t) - z_j (t)\|.
\end{equation*}
For $\tau(t)$, we consider two cases where $\tau(t) = 1/t^{1.5}$ satisfying Assumption \ref{event-triggered} and $\tau(t) = 1/t^{0.75}$ not satisfying Assumption \ref{event-triggered}. And for $\zeta(t)$, we consider also consider two cases where $\zeta(t) = 1/(3t^{3})$ satisfying Assumption \ref{y-trigger} and $\zeta(t) = 1/t$ not satisfying Assumption \ref{y-trigger}. Additionally, we test two constant cases $\tau (t), \zeta(t) =0$ and $\tau (t),\zeta(t) =0.05$. Figure 1 depicts the graph of the values of $R_f (t)$ versus the iteration time. The result shows that the cost error decreases to zero when $\sum_{t=1}^{\infty} \tau (t) < \infty$ while it does not converge to zero when $\sum_{t=1}^{\infty} \tau (t) = \infty$. This agrees with the convergence result of Theorem \ref{case2_main}. Figure 2 illustrates the consensus error $R_c (t)$. The result shows that the consensus error decreases to zero for any choices $\tau (t)$ except the case $\tau (t) =0.05$. This numerical result supports the theoretical result obtained in Corollary \ref{convergent}.
\section{Conclusion}
In this work, we considered the gradient-push algorithm with event-triggered communication for distributed optimization problem whose agents are connected by directed graphs. We showed that by the algorithm each agent's state converges to a common minimizer under a diminishing and summability condition on the stepsize and the triggering function. Numerical simulations have been conducted to support the convergence results. It would be interesting further about how to choose an optimal triggering function for reducing the communication burden, reflecting various elements such as the connectivity of graphs and the number of agents.
|
\section{Introduction}
There are both theoretical and observational motivations for the construction
of gravitational modifications, namely of extended theories of gravity that
possess general relativity as a particular limit, but which in general exhibit
a richer structure \cite{CANTATA:2021ktz}. The first is based on the fact that
since general relativity is non-renormalizable one could hope that more
complicated extensions of it would improve the renormalizability properties
\cite{Stelle:1976gc,Addazi:2021xuf}. The second motivation is related to the
observed
features of the Universe, and in particular the need to describe its two
accelerated phases, namely one at early times (inflation) and one at late times
(dark energy era). The usual approach in the construction of gravitational
modifications is to start from the Einstein-Hilbert action and extend it in
various ways \cite{Capozziello:2011et}. Nevertheless, one can
start from the equivalent torsional formulation of gravity, and in particular
from the Teleparallel Equivalent of General Relativity (TEGR)
\cite{Unzicker:2005in, Aldrovandi:2013wha,Shirafuji:1997wy, Maluf:2013gaa}, and modify it
accordingly, obtaining $f(T)$ gravity \cite{Cai:2015emx,
Bengochea:2008gz, Linder:2010py}, $f(T,T_G)$ gravity \cite{Kofinas:2014owa},
$f(T,B)$ gravity \cite{Bahamonde:2015zma, Karpathopoulos:2017arc,
Boehmer:2021aji}, scalar-torsion theories \cite{Geng:2011aj, Hohmann:2018rwf,
Bahamonde:2019shr}, etc. Torsional gravity, can lead to interesting cosmological
phenomenology and hence it has attracted a large amount of research
\cite{Cai:2015emx,Zheng:2010am,ElHanafy:2015egm,
Bamba:2010wb, Capozziello:2011hj,Wei:2011aa, Amoros:2013nxa,
Otalora:2013dsa, Bamba:2013jqa, Li:2013xea,
Malekjani:2016mtm, Farrugia:2016qqe,Nashed:2014sea,Saridakis:2017eqb, Qi:2017xzl,Cai:2018rzd,
Abedi:2018lkr, El-Zant:2018bsc, Anagnostopoulos:2019miu,
Cai:2019bdh, Yan:2019gbw,Awad:2017ign,ElHanafy:2019zhr, Wang:2020zfv,
ElHanafy:2020pek,Nashed:2009hn,Hashim:2020sez, Ren:2021tfi}.
Additionally, torsional and $f(T)$ gravity exhibit
novel and interesting
black hole and spherically symmetric solutions too \cite{Boehmer:2011gw,
Gonzalez:2011dr, Capozziello:2012zj,Ferraro:2011ks, Wang:2011xf,
Atazadeh:2012am,Nashed:2005kn,
Rodrigues:2013ifa, Nashed:2013bfa, Nashed:2014iua, Junior:2015fya,
Kofinas:2015hla, Das:2015gwa,Awad:2017sau, Nashed:2018efg,Rani:2016gnl, Rodrigues:2016uor, Mai:2017riq,
Singh:2019ykp, Nashed:2020kjh, Bhatti:2018fsc, Ashraf:2020yyo,
Ditta:2021wfl,Ren:2021uqb}.
In particular, spherically symmetric solution with a constant torsion
scalar $T$ have been studied in
\cite{2011PhRvD..84h3518F,Nashed:uja,2013PhRvD..88j4034N}, while cylindrically
charged black holes solutions using quadratic and cubic forms of $f(T)$ have
been derived \cite{Nashed:2019zmy,Nashed:2018cth,Awad:2017tyz,Nashed:2016tbj}.
Moreover, by using the Noether's symmetry approach, static spherically black
hole solutions have been investigated in \cite{Paliathanasis:2014iva}.
In similar lines, the research of static spherically symmetric solutions
using $f(T)$ corrections on TEGR was the focus of interest in many studies
using the perturbative approach
\cite{DeBenedictis:2016aze,Bahamonde:2020bbc,Bahamonde:2019zea,
2015PhRvD..91j4014R, Nashed:2021rgn}, while vacuum
regular BTZ black hole solutions in Born-Infeld gravity have been extracted in
\cite{Bohmer:2019vff,Bohmer:2020jrh}.
Although spherically symmetric solutions in $f(T)$ gravity
has been investigated in many works, the stability of motion around
them has not been examined in
detail. This issue is quite crucial, having in mind that modifications of
general relativity are known to present various instabilities is various
regimes of the parameter space. Hence, in this work we aim to derive charged
spherically symmetric
solution in $f(T)$ gravity using the perturbative approach, and then examine
the stability of motion and thermodynamic properties.
The arrangement of the manuscript is as follows: in Section \ref{S2} we
extract the charged black-hole solutions for $f(T)$ gravity, using the
perturbative approach, for two charge profiles.
In Section
\ref{S4} we study the properties of the extracted perturbative
solutions, and in particular their asymptotic forms, the invariants, and their
energy. In Section \ref{S6336b} we proceed to the investigation of
the stability of motion around the
solutions, by extracting and analyzing the geodesic deviation. Moreover, in
Section \ref{S5} we study in detail the thermodynamic properties, focusing on
the temperature, entropy, heat capacity and Gibbs free energy.
The final Section \ref{Conclusions} is reserved for
conclusions and discussion.
\section{Charged black hole solutions in $f(T)$ gravity}\label{S2}
Let us extract charged black hole solutions following the perturbative
approach. As usual, in torsional gravity as the dynamical field we use the
orthonormal tetrad, whose components are $h_a{}^\mu$, with Latin
indices (from 0 to 3) denoting the tangent space and
Greek indices (from 0 to 3) marking the coordinates
on the manifold. The relation between the tetrad and the manifold metric is
$g_{\alpha \beta}=\eta_{i j} h^i{}_\alpha h^j{}_\beta$, with
$\eta_{i j}$ the Minkowski metric $\eta_{i
j}=diag.(-1,+1,+1,+1)$. The torsion tensor is given as
$
T^a{}_{\mu\nu}:= \partial_{\mu}h^a{}_{\nu}-\partial_{\nu}h^a{}_{\mu}$.
The action of $f(T)$ gravity, alongside a minimally coupled elecromagnetic
sector, is \cite{Gonzalez:2011dr,Capozziello:2012zj}
\begin{align}\label{eq:fT1}
S_{f(T)} = \int d^4x\ |h| \left( \frac{1}{2\kappa^2} f(T) + {\cal
F}\right)\,,
\end{align}
with $\kappa=8\pi G$ the gravitational constant, and $|h|=det
(h^a{}_\mu)=\sqrt{-g}$. The torsion scalar $T$
is written as
$ T = T^a{}_{\mu\nu}S_a{}^{\mu\nu} $
in terms of the superpotential
$S_a{}^{\mu\nu} = \frac{1}{2}(K^{\mu\nu}{}_a - h_a{}^\mu
T_\lambda{}^{\lambda\nu} + h_a{}^\nu T_\lambda{}^{\lambda\mu})$,
with the contortion tensor being
$
{K^{\mu \nu}}_\alpha:=
-\frac{1}{2}\left({T^{\mu \nu}}_\alpha-{T^{\nu
\mu}}_\alpha-{T_\alpha}^{\mu \nu}\right)$.
Additionally, ${\cal {\cal F}}$ is the gauge-invariant
Lagrangian of electromagnetism given as ${\cal F} =
\frac{1}{4}{\cal F}_{\alpha \beta}{\cal F}^{\alpha \beta}$
\cite{plebanski1970lectures}.
Variation of action (\ref{eq:fT1}) with respect to the
tetrad yields the field
equations~\cite{Krssak:2015oua}:
\begin{eqnarray}\label{eq:fT}
&&
\zeta_a{}^\mu\equiv\frac{1}{4}f(T) h_a{}^\mu + f_T \left[ T^b{}_{\nu
a} S_b{}^{\mu \nu } +
\frac{1}{h}\partial_{\nu}(h S_a{}^{\mu \nu }) \right] + f_{TT}\
S_a{}^{\mu\nu} \partial_\nu T- \frac{1}{2}\kappa^2 \Theta_a{}^\mu=0\,,
\end{eqnarray}
with $f_T\equiv \partial f/\partial T$ and $f_{TT}\equiv\partial^2 f/\partial
T^2$, and where the elecromagnetic stress-energy tensor
is
\begin{equation} \label{max1}
\Theta_a{}^\mu\:={\cal F}_{a \alpha} {\cal F}^{\mu \alpha}-\frac{1}{4}
\delta_a{}^\mu {\cal F}_{\alpha \beta} {\cal F}^{\alpha \beta}. \end{equation}
Moreover, variation of action (\ref{eq:fT1}) with respect to the Maxwell
field gives
\begin{equation} \label{maxf}
\partial_\nu \left( \sqrt{-g} {\cal F}^{\mu \nu} \right)=0.\end{equation}
We can rewrite equation (\ref{eq:fT}) purely in terms of
spacetime indices by contracting with $g_{\mu\rho}$ and $h^a{}_\sigma$,
resulting to
\begin{align}
\label{eq:fT2}
H_{\sigma\rho} = \frac{1}{2}\kappa^2 \Theta_{\sigma\rho}\,.
\end{align}
The symmetric part of (\ref{eq:fT2}) was sourced by the energy-momentum tensor
(\ref{max1}), while their anti-symmetric part is a vacuum constraint for the
considered matter models. The latter is equal to the variation of the action
with respect to the flat spin-connection components \cite{Golovnev:2017dox,
Hohmann:2017duq}, namely
\begin{align}
H_{(\sigma\rho)} = \frac{1}{2}\kappa^2 \Theta_{(\sigma\rho)}, \quad H_{[\sigma\rho]} =0\,.
\end{align}
The explicit forms of these equations can be seen in Eqs. (26) and (30) of
\cite{Hohmann:2018rwf} by setting the scalar field $\phi$ to zero, however we
do not display them here since we will derive the spherically
symmetric field equations directly from \eqref{eq:fT}.
We proceed by focusing on spherically symmetric solutions.
Employing the spherical coordinates $(t,r,\theta, \phi)$ we write the suitable
spherically symmetric tetrad space as:
\begin{equation}
\!b^a{}_{\mu}\!=\!\left(
\begin{array}{cccc}\!\!
\sqrt{a} & 0 & 0 & 0 \\
0 & \frac{1}{\sqrt{b}} \cos (\phi ) \sin (\theta ) & r \cos (\phi ) \cos (\theta ) & -r \sin (\phi ) \sin (\theta ) \\
0 & \frac{1}{\sqrt{b}} \sin (\phi ) \sin (\theta ) & r \sin (\phi ) \cos (\theta ) & r \cos (\phi ) \sin (\theta ) \\
0 & \frac{1}{\sqrt{b}} \cos (\theta ) & -r \sin (\theta ) & 0 \\
\end{array}
\right)\label{tet},
\end{equation}
where $a\equiv a(r)$ and $b\equiv b(r)$ are two positive $r$-dependent
functions. The above tetrad corresponds to the usual metric
\begin{equation}
ds^2=-a(r) \,dt^2+\frac{dr^2}{b(r)}+r^2d\Omega^2,\label{met1}
\end{equation}
with $ d\Omega^2=(d\theta^2+\sin^2\theta d\phi^2)$.
Using (\ref{tet}) the torsion scalar becomes
\begin{align}\label{scaletorsion}
T= \frac{2 \left[1-\sqrt{b(r)}\right] \left[r a'(r)-a(r)
\sqrt{b(r)}+a(r)\right]}{r^2 a(r) b(r)}\,.
\end{align}
Note that $T$ becomes zero in the case $a=b\rightarrow 1$.
Inserting the above tetrad choice into the field equations (\ref{eq:fT}) we
acquire
\begin{eqnarray}
\zeta_t{}^t &=& \frac{1}{4} f+\frac{b^{3/2}(ra'+2a)+rab'-rba'-2ba}{2 r^2 a
b^2}f_T + \frac{(\sqrt{b}-1) }{r b}T'
f_{TT} -\frac{\mathbb{Q}'^2}{2ab}=0\,,\label{Eq1}\\
\zeta_r{}^r&=& \frac{\sqrt{b}(ra'+2a)-2(ra'+a)}{2 r^2 ab}f_{T} + \frac{f}{4}-\frac{\mathbb{Q}'^2}{2ab}=0\,,\label{Eq2}\\
\zeta_\theta{}^\theta&=&\zeta_\phi{}^\phi=\frac{2a\sqrt{b}-(ra'+2a) }{4 r a
b}T'f_{TT}+ \frac{f}{4}+\frac{\mathbb{Q}'^2}{2ab}\nonumber\\
&&\ \ \
\ \ \ \
\ \, +\frac{b(r^2a'^2-6raa'-2r^2aa''-4a^2)+a[(2a+ra')(rb'+4b^{3/2})-4ab^2]}{8
r^2 a^2 b^2}f_{T}
=0\,,\label{Eq3}
\end{eqnarray}
where primes denote derivatives with respect to $r$.
In the above equations we have introduced the components of the electric
field $\mathbb{Q}_{\mu}=[\mathbb{Q}(r),0,0,0]$, where
${\cal
F}_{\mu\nu}=\mathbb{Q}_{\mu,\nu}-\mathbb{Q}_{\nu,\mu}$. Hence, the
non-vanishing components of the Maxwell field are
\begin{equation}
\frac{\mathbb{Q}'[a(rb'-4b)+rba']-2rba\mathbb{Q}''}{2ra^2b^2}=0.\label{Eq4}
\end{equation}
Note that equations (\ref{Eq1})-(\ref{Eq4}) coincide
with those of \cite{Bahamonde_2019} when $\mathbb{Q}=0$.
In the following we solve the above equations to first-order
expansion around the Reissner-Nordstr\"om background, which allows us to
extract analytical solutions (since in general the torsion scalar is not a
constant, in which case one has the simple
Reissner-Nordstr\"om solution). Hence, we assume the perturbative general
vacuum charged solution as
\begin{align}
a(r)&= 1 - \frac{2M}{r}+\frac{s^2}{r^2}+\epsilon a_1(r)\,,\label{h}\\
b(r)&= 1 - \frac{2M}{r}+\frac{s^2}{r^2}+\epsilon b_1(r)\,,\label{k}\\
\mathbb{Q}(r)&= -\frac{s}{r}+\epsilon \mathbb{Q}_1(r)\label{q}\,.
\end{align}
Finally, concerning the $f(T)$ function we will
consider the power-law form
\begin{eqnarray}
f(T)=T+\frac{1}{2}\alpha \epsilon\, T^2\,,\label{tor}
\end{eqnarray}
with $\alpha$ the usual parameter of
the $T^2$ term and $\epsilon<<1$ the small tracking parameter used to
quantify
the expansion in a consistent way \cite{DeBenedictis:2016aze,Bahamonde_2019}.
The above expression in the limit $\epsilon\rightarrow0$ recovers Teleparallel
Equivalent of General Relativity, and it is known to be a
good approximation for every realistic $f(T)$ gravity
\cite{Nesseris:2013jea,Nunes:2016qyp,Li:2018ixg}, since the extra term
quantifies the deviation from General Relativity.
Substituting (\ref{h})-(\ref{tor}) into
(\ref{Eq1})-(\ref{Eq3}),
keeping $\epsilon$ terms up to first order, we obtain:
\begin{widetext}
\begin{eqnarray}
\zeta_t{}^t &=&\frac{\epsilon}{r^8\varrho^2}\Big\{
\alpha(\varrho-1)\left[
(10\varrho^2+5\varrho+1)(\varrho-1)^2r^4+2s^2(8\varrho^2+4\varrho+1)(\varrho-1)r
^2+s^4(3\varrho+1)\right]
-r^4s^2a_1-r^7\varrho^6b'_1
\nonumber\\
&&+(\varrho^2 r^2-2r^2+s^2 )r^4\varrho^4b_1+2\varrho^2r^6
s\mathbb{Q}_1'\Big\}=0 \,,
\nonumber\\
\zeta_r{}^r&=&\frac{\epsilon\left\{\alpha(\varrho-1)
[(\varrho-1)^2r^2+s^2(3\varrho-1)][(\varrho-1)r^2+s^2]+r^7\varrho^2a'_1{}
^2+(\varrho^2-1)r^6a_1+r^6\varrho^4b_1+
2\varrho^2r^2s\mathbb{Q}_1'\right\}}{r^8\varrho^2}=0\,,
\nonumber\\
\zeta_\theta{}^\theta&=&
\zeta_\phi{}^\phi=\frac{\epsilon}{4r^{10}\varrho^4}\Big\{2\alpha\varrho\left[
(5\varrho^2+4\varrho+1)(\varrho-1)^4r^6+s^2(\varrho-1)^2
(8\varrho^3-11\varrho^2-6\varrho-3)r^4+s^4r^2(3-10\varrho^3+7\varrho^2)-s^6
\right]\nonumber\\
&&+2\varrho^4r^{10}a''_1-8\varrho^4r^8s\mathbb{Q}_1'
+\varrho^2r^7[(1-3\varrho^2)r^2+s^2]a'_1-\varrho^6r^7[(\varrho^2+1)r^2-s^2]b'_1
+a_1[r^8(1-\varrho^4)-2r^6s^2+r^4s^4]\nonumber\\
&&+\varrho^4 b_1[(\varrho^4-1)r^8+2r^6s^2-r^4s^4]\Big\}=0\,,\label{Eq7}
\end{eqnarray}
\end{widetext}
while the Maxwell field equation (\ref{Eq4}) becomes
\begin{eqnarray}
&&
\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!
2\varrho^4r^5\mathbb{Q}_1''+4v\varrho^4r^4\mathbb{Q}_1'-s\left\{
\varrho^2r^3a'_1-\varrho^6r^3b'_1
-[(\varrho^2-1)r^2+s^2](a_1-\varrho^2b_1)\right\} =0\,,\label{Eq8}
\end{eqnarray}
where $\varrho=\sqrt{1-\frac{2M}{r}+\frac{s^2}{r^2}}$.
We solve the above equations separately in the cases where
$\mathbb{Q}_1(r)=0$ and $\mathbb{Q}_1(r)\neq 0$, namely the cases with or
without a perturbative correction in the charge profile.
\begin{itemize}
\item \underline{Case I: $\mathbb{Q}_1(r)=0$:}\\
In this case differential equations (\ref{Eq7})
and (\ref{Eq8}) admit the solution:
\begin{eqnarray} \label{Eqs0}
&&
\!\!\!\!\!\!\!
a_1(r)=\frac{1}{s^6r^6\varrho\varrho_1}
\Bigg\{4sr^5\alpha \varrho\Big\{
5\varrho_1{}^{3/2}r^2\varrho^2\tan^{-1}\Theta
+s\left[
rM(5s^2-6M^2)+2M^4
+s^2(M^2-2s^2)\right]\tan^{-1}\Theta_1\Big\}\nonumber\\
&&
\ \,\ \ \ \
+\varrho_1 \Big\{4\alpha
s^2r^5\varrho\ln\varrho\left[(s^2-3M^2)r+M(M^2+s^2)\right]\nonumber\\
&&
\ \,\ \ \ \ \ \ \ \,\ \ \ +
\varrho
\Big\{2r^6\alpha\ln\varrho_1(2M^4+M^2s^2-2s^4)
-s^2\Big\{8r^5\alpha\ln(r)\Big[
(s^2-3M^2)r+M^3 +s^2M\Big]
\nonumber\\
&&\ \ \ \ \ \,\ \ \ \ \ \ \ \ \ \ \ \ -s^2\left[s^2c_2r^6-(12\alpha
M+s^2c_1)r^5
+\alpha(4r^4\varrho_1{}^2-4/3
s^4 r^2+3/5 s^2)\right]\Big\}\Big\}\nonumber\\
&& \ \,\ \ \ \ \ \ \ \,\ \ \
-20r^2s^2\varrho\alpha\Big[r^3(4/3s^2-2M^2) +2/15Ms^4
+r^2M(M^2-1/3s^2)
+rs^2/3(M^2-1/5s^2)\Big]\Big\}\Bigg\}\,,
\end{eqnarray}
\begin{eqnarray}
&&
\!\!\!\!\!\!\!
b_1(r)=-\frac{4}{s^6r^2\varrho^{7/2}\varrho_1{}^2}\Bigg\{sr^4\alpha
\varrho^{3/2}
\varrho_1^{5/2}\Big\{5\varrho_1{}^{3/2}[s^2(2M+r)-3rM^2]\tan^{-1}\Theta
\nonumber\\
&& \ \,\ \ \ \ \ \ \ \,\ \ \ \ \,\ \ \ \ \ \ \ \,\ \ \ \ \,\ \ \ \ \
\ \ \,\ \ \ \ \,\ \ \ \ \ \ \ \
+2s\Big
[ (2r+5M)s^4-M^2s^2(6M+11r) +10rM^4\Big]\tan^{-1}\Theta_1\Big\}
\nonumber\\
&& \ \ \ \ \ \
+\varrho_1^2\Bigg\{2s^2r^4\alpha
\varrho^{3/2}[3M(M+r)s^2-s^4-5rM^3]\ln
\varrho+r^4\varrho^{3/2}(2M^4+2M^2s^2-s^4)(2Mr-s^2)\alpha
\ln\varrho_1{}^2\nonumber\\
&&
\ \ \ \ \ \ \ \ \ \ \
+s^2\bigg\{4r^4\varrho^{3/2}\alpha \ln(r)[5M^3r-3Ms^2(M+r)+s^4]-8/3\alpha
r s^{10}+1/10\alpha s^8(23\varrho^{3/2}+40Mr^2+100r^3)\nonumber\\
&&
\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \, \ \
+rs^6[28\alpha r^4-r^3/2(\varrho^{3/2}c_2+168\alpha M)+6\alpha
r^2M^2-8/3r\alpha\varrho^{3/2}-4M\alpha \varrho^{3/2}]
\nonumber\\
&& \ \ \ \ \ \
+r^3s^4\Big\{46/3\alpha
r^4-128Mr^3\alpha +r^2[(Mc_2-c_1/2)\varrho^{3/2}+188M^2\alpha]+8\alpha
r(65M^3/3+4\varrho^{3/2})+
16M\alpha \varrho^{3/2}\Big\}
\nonumber\\
&& \ \ \ \ \ \
-
2r^4Ms^2\alpha\Big[2M^2r^2-11Mr^3+M\varrho^{3/2}
+3r^4+16M^3r\Big]+6r^6M^3\alpha (r-2M)^2\bigg\}\Bigg\}\Bigg\},
\label{Eqs0b}
\end{eqnarray}
where $\Theta=\frac{Mr-s^2}{s\varrho}$, $\Theta_1=\frac{M-r}{\varrho_1}$ and
$\varrho_1=\sqrt{M^2-s^2}$. Expressions (\ref{Eqs0}),(\ref{Eqs0b}) are the
solution of the
field equations (\ref{eq:fT}) and (\ref{maxf}) up to ${\mathcal{O}}(\epsilon)$.
\item \underline{Case II: $\mathbb{Q}_1(r)\neq 0$:}\\
In this case the solution of (\ref{Eq7}) and (\ref{Eq8}) in the case
$\mathbb{Q}_1(r)\neq 0$ for the metric functions is
\begin{eqnarray} \label{Eqa}
&&
\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!
a_1(r)=c_4+
\frac{1}{r^2\alpha^{5/2}}\Bigg\{\int\frac{1}{r^6}\Bigg\{\alpha\Big[\varrho^{1/2}
r^2(s^2r^4\!-\!r^6\!-\!s^6\!+\!r^2s^2)+2r\varrho(3r^4s^2\!-\!2r^6\!-\!s^6)\!+\!r
^2 \varrho^ { 3/2 } (5s^4\!-\!3r^4\!+\!6r^2s^2)
\nonumber\\
&& \ \ \ \ \ \ \ \ \ \ \ \ \ \ \
\ \ \ \ \ \ \ \ \
+2r\varrho^2(10s^4\!+\!8r^4\!-\!22r^2s^2)
+\varrho^{5/2}
(18s^4\!+\!15r^4\!-\!23r^2s^2)\!+\!2r\varrho^3(10r^2\!-\!9s^2)\!+\!3r^2\varrho^{
7/2} \Big ] \nonumber\\
&& \ \ \ \ \ \ \ \ \ \ \ \ \ \ \
\ \ \ \ \ \ \ \ \,
-2r^7s\varrho^ { 5/2}\mathbb{Q}''_1\Bigg \}d r+c_3\alpha^{5/2}\Bigg\}\,,
\label{Eqbbb}
\end{eqnarray}
\begin{eqnarray} \label{Eqb}
&& b_1(r)= \frac{1}{r^4\varrho^{5/2}}\Big\{\alpha r^2\varrho^{5/2}+4r\alpha
\varrho^2(s^2-r^2)+\varrho^{3/2}[2r^6s\mathbb{Q}'_1+r^7a'_1+r^6a_1+6\alpha
r^4-10r^2s^2\alpha+3\alpha s^4]\nonumber\\
&&
\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \
-4r\alpha \varrho(r^2-s^2)^2 -r^2\varrho^{1/2}[2r^2\alpha
s^2+r^6a_1-\alpha s^4-\alpha r^4]\Big\}\,,
\end{eqnarray}
while inserting these into (\ref{Eq8}) we finally acquire
\begin{eqnarray}
&&
\!\!\!\!\!\!\!\!\!
\mathbb{Q}_1(r)=
\frac{1}{15r^{5}\varrho \varrho_1 s^6}\Bigg\{30\alpha r^4s\varrho
\varrho_1 [4M^3r-2Mrs^2-3M^2s^2+s^4]\ln(r^2\varrho^2) \nonumber\\
&&
\ \ \ \ \ \ \
+30\alpha r^4s\varrho
\tanh^{-1}\left(\frac{M-r}{\varrho_1}\right)[(8M^4
-8M^2s^2+s^4)r+Ms^2(5s^2-6M^2)]\nonumber\\
&&
\ \ \ \ \ \ \
+75\alpha r^4\varrho\varrho_1{}^{5/2}
\tanh^{-1}\left(\frac{Mr-s^2}{sr\varrho}
\right)(5M^2r-4Ms^2-s^2r)+300\alpha r^4\varrho_1{}^2\ln2
[(5M^2-\epsilon^2)r-4M\epsilon^2]\nonumber\\
&&
\ \ \ \ \ \ \
-15r^5M\alpha s\varrho
\varrho_1\ln\varrho_1{}^2(6M^2-5s^2) -60r^4\alpha s \varrho
\varrho_1\ln\,r\Big[4M^3r-2Mrs^2 -3M^2s^2+s^4\Big]
\nonumber\\
&& \ \ \ \ \ \ \
+s\varrho_1\Bigg\{
\varrho\left[
15(rc_4-c_3)r^4s^5+2s^2\alpha(3s^6+15r^4s^2-10s^4r^2+30Ms^2r^3-60M^2r^4)\right]
\nonumber\\
&&\ \ \ \ \,\ \ \ \ \ \ \ \ \ \ \
+5s^4r\alpha(25r^4-125Mr^3-2s^4+23s^2r^2-15r^2M^2-6s^2rM) \nonumber\\
&&\ \ \ \ \,\ \ \ \ \ \ \ \ \ \ \ +25\alpha
r^5M^3(29r-30M)+25Ms^2r^4\alpha(29M^2-13r^2
-19Mr)\Bigg\}\Bigg\}\,.
\end{eqnarray}
\end{itemize}
Hence, we have extracted the spherically symmetric solutions in the case
of a quadratic deviation from Teleparallel Equivalent of General Relativity,
which as we mentioned above is the first correction in every realistic $f(T)$
gravity. We stress here that all the above expressions in the limit
$\epsilon\rightarrow0$ recover the Schwarzschild results, hence our solutions
provide the corrections on the latter brought about the $f(T)$ gravity.
\section{ Properties of the solutions }\label{S4}
In this section we examine the properties of the extracted perturbative
solutions, and in particular their asymptotic forms, the invariants, and their
energy.
\subsection{Asymptotic forms}
In the case $\mathbb{Q}_1(r)=0$, the asymptotic form of the solutions
(\ref{Eqs0}),(\ref{Eqs0b}), up to ${\mathcal{O}}(\epsilon)$ become
\begin{eqnarray} \label{mets1}
&&
\!\!\!\!\!\!\!\!\!\!\!
a(r)\approx 1-\frac{2M}{r}+\frac{s^2}{r^2}
\nonumber\\
&& \ \ \ \,
+\epsilon\Bigg\{c_2-\frac{c_1}{r}
-\frac { \alpha}{r}\Bigg[\frac{68}{3M}+\frac{10M^2}{s^2{\varrho_1}}
-\frac{6}{{\varrho_1}}-\frac{76M}{3s^2}+
\frac{20\ln(2M-2s)}{s}-\frac{40M^2\ln(2M-2s)}{s^3}\Bigg]+{\mathcal{O}}\Big(\frac
{1}{r^5}\Big)\Bigg\}\,,
\end{eqnarray}
\begin{eqnarray}
&&
\!\!\!\!\!\!\!\!\!\!\!
b(r)\approx1-\frac{2M}{r}+\frac{s^2}{r^2}
+\epsilon
\Bigg\{\frac{c_1}{r}\!-\!\frac{
2Mc_2}{r}\!+\!\frac{\alpha}{r}\Bigg[\frac{68}{3M}\!+\!\frac{10M^2}{s^2{\varrho_1
}}
\!-\!\frac{6}{{\varrho_1}}\!-\!\frac{76M}{3s^2}+
\frac{20\ln(2M\!-\!2s)}{s}\!-\!\frac{40M^2\ln(2M\!-\!2s)}{s^3}\Bigg]\nonumber\\
&& \ \ \ \,
-\frac{4Mc_1\!+\!(s^2\!-\!8M)s^2}{r^2}\!-\!\frac{8\alpha}{3s^3r^2\varrho_1}
\left[ 30M(2M^2\!-\!s^2)\varrho_1\ln(2\varrho_1{}^2)\!-\!Ms(15M^2\!-\!9s^2)
\!-\!2s\varrho_1(17s^2\!-\!18M^2)\right]\Bigg\}\,,
\label{mets1b}
\end{eqnarray}
and thus the metric (\ref{met1}) becomes Minkowski for $r\rightarrow
\infty$.
On the other hand, in the case $\mathbb{Q}_1(r)\neq0$, the asymptotic forms of
the solutions (\ref{Eqbbb}),(\ref{Eqb}) up to ${\mathcal{O}}(\epsilon)$
become
\begin{eqnarray} \label{mets2}
&&
\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!
a(r)=1-\frac{2M}{r}+\frac{s^2}{r^2}
\nonumber\\
&& \!\!\!\!\!\!\!\!\!
+\epsilon\Bigg\{c_3\Big(1+\frac{2}{s}\Big)+\frac{c_4}{r}-\frac{\alpha}{r}
\Bigg[\frac{8M^3\ln(r)}{s^4}-\frac{8M\ln(r)}{s^2}
-\frac{136M^3}{3s^2\varrho_1{}^2}+\frac{20M^5}{s^4\varrho_1{}^2}+\frac{76M}{
3\varrho_1{}^2}\nonumber\\
&&
\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \,
+2\tanh^{-1}\Big(\frac{M}{s}\Big) \Bigg(\frac{32M^2}{s^3}-\frac{20M^4}{
s^5}-\frac{12}{s}
\Bigg)\Bigg]+\frac{2sc_3}{r^2}+{\mathcal{O}}\left(\frac{1}{r^3}
\right)\Bigg\}\,,
\end{eqnarray}
\begin{eqnarray}
&&
\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!
b(r)=1-\frac{2M}{r}
+\frac{s^2}{r^2}\nonumber\\
&& \!\!\!\!
-5\epsilon\Bigg\{\frac{12Mc_3+3c_4s}{rs}-\frac
{\alpha}{r}\Bigg[\frac{60M^3}{s^4}+\frac{24M^3\ln(r)}{s^4}
-\frac{76M}{s^2}-4\tanh^{-1}\Big(\frac{M}{s}\Big)\Bigg(\frac{24M^2}{s^3}
+\frac{9}{s}+\frac{15M^4}{s^5}
\Bigg)\Bigg]
\nonumber\\
&& \ \ \, \ \ \ +\frac{\alpha}{3s^5r^2}\Bigg[
240sM^4-208M^2s^3-120M^2s^3\ln(r)+36s^4-48\tanh^{-1}
\Big(\frac{M}{s}\Big)\Big(3Ms^4-10M^3s^2+5M^5
\Big)\Bigg] \nonumber\\
&& \ \ \, \ \ \
-\frac{(16M^2-s^3)c_3+4c_4sM}{sr^2}+{\mathcal{O}}\left(\frac{1}{r^3}
\right)\Bigg\} \ ,
\label{mets2bb}
\end{eqnarray}
which also become Minkowski in the limit
$r\rightarrow \infty$.
\subsection{Invariants }
Let us examine the behavior of various invariants in the obtained solutions.
For the case $\mathbb{Q}_1(r)=0$, and inserting the asymptotic forms
(\ref{mets1}),(\ref{mets1b}) into the tetrad (\ref{tet}) and metric
(\ref{met1}), and then
into the various tensor definitions we respectively acquire the following
expressions for the torsion tensor square, the torsion vector square, the
torsion scalar, the Kretschmann scalar, the Ricci tensor square, and the Ricci
scalar:
\begin{eqnarray}\label{inv1}
&&
\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!
T^{\mu \nu \lambda}T_{\mu \nu \lambda}
=\frac{16M+8\epsilon(c_1-2Mc_2)}{r^3}-\frac{16s^2-4M^2+\epsilon[68M^2\epsilon
c_2{}^2-16s^2c_2-4Mc_1+15c_1{}^2\epsilon+8M^2c_2-64Mc_1c_2\epsilon]}{2r^4}
\nonumber\\
&& \ \ \ \ \ \ \
+{\mathcal{O}}\Big(\frac{1}{r^5}\Big)\,,
\end{eqnarray}
\begin{equation}
\!\!\! \!\!\! \!\!\! \!\!\! \!\!\! \!\!\! \!\!\! \!\!\! \!\!\!
\!\!\! \!\!\! \!\!\! \!\!\! \!\!\! \!\!\! \!\!\! \!\!\! \!\!\! \!\!\!
\!\!\! \!\!\! \!\!\! \!\!\! \!\!\! \!\!\! \!\!\! \!\!\! \!\!\! \!\!\!
\!\!\! \!\!\! \!\!\! \!\!\! \!\!\! \!\!\! \!\!\! \!\!\! \!\!\! \!\!\! \!\!\!
\!\!\! \!\!\! \!\!\! \!\!\! \!\!\! \!\!\!
\!\!\! \!\!\! \!\!\! \!\!\! \!\!\!
\!\!\! \!\!\! \!\!\!
\!\!\! \!\!\! \!\!\! \!\!\! \!\!\! \!\!\! \!\!\!
T^\mu T_\mu =
\frac {4(2M+\epsilon [c_1-2
Mc_2+\epsilon c_1c_2-2M\epsilon
c_2{}^2])}{r^3}+{\mathcal{O}}\Big(\frac{1}{r^4}\Big)\,,
\end{equation}
\begin{eqnarray}
&& \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!
T(r)=\frac{4c_2\epsilon^2(2Mc_2-c_1)}{r^3} +
\frac{4M^2-8s^2-\epsilon(8M^2c_2-8s^2c_2-4Mc_1)+\epsilon^2[9c_1{}^2+76M^2c_2{}
^2-16s^2c_2{}^2-56Mc_1c_2]}{2r^4}
\nonumber\\
&& \ \ \ \ \ \ \
+{\mathcal{O}}\Big(\frac{1}{r^5}\Big)\,,
\end{eqnarray}
\begin{eqnarray}
&&\!\!\!
\! \!\!\! \!\!\! \!\!\! \!\!\! \!\!\! \!\!\! \!\!\! \!\!\!
\!\!\! \!\!\! \!\!\!
\!\!\! \!\!\! \!\!\! \!\!\!
\!\!\! \!\!\! \!\!\! \!\!\! \!
\!\!\! \!\!\! \!\!\! R^{\mu \nu \lambda \rho}R_{\mu \nu \lambda \rho}=
\frac{48M^2+48\epsilon\,M(c_1-2c_2M)+12\epsilon^2(c_1{}^2-6Mc_1c_2+16M^2c_2{}^2)
}{r^6}+{\mathcal{O}}\Big(\frac{1}{r^7}\Big)\,,
\end{eqnarray}
\begin{equation}\!\!\!
R^{\mu \nu}R_{\mu
\nu}=\frac{4\epsilon^2c_2s^2[c_1-2Mc_2]}{r^7}+\frac{2s^2[2s^2-4s^2\epsilon\,
c_2-\epsilon^2(2c_1{}^2-13Mc_1c_2-8s^2c_2{}^2+18M^2c_2{}^2)}
{r^8}+{\mathcal{O}}\Big(\frac{1}{r^9}\Big)\,, \end{equation}
\begin{eqnarray} &&\!\!\!\!\!
R=-\frac{\epsilon^2(10M^2c_2{}^2-2s^2c_2{}^2-9Mc_1c_2+2c_1{}^2)}{r^4}-\frac{
\epsilon^2(48M^3c_2{}^2-24s^2Mc_2{}^2-44M^2c_1c_2+10Mc_1{}^2+
11s^2c_1c_2)}{r^5}
\nonumber\\
&&
\frac{-\epsilon^2(168M^4c_2{}^2-128s^2Mc_2{}^2+10s^2c_2{}
^2-156M^3c_1c_2+81Ms^2c_1c_2+36M^2c_1{}^2-10s^2c_1{}^2)}{r^6} \nonumber\\
&&
-\frac{\epsilon^2}{r^7}\biggl(112M^3c_1{}^2+512M^5c_2{}
^2+372M^2s^2c_1c_2-16M^3\alpha c_2+60Ms^2c_2{}^2+96Ms^4c_2{}^2 \nonumber\\
&&
-31s^4c_1c_2-520M^3q^2c_2{}^2-480M^4c_1c_2\biggr)+\frac{16\epsilon\alpha
M^3}{r^7}+{\mathcal{O}}\Big(\frac{1}{r^8}\Big)\,.
\end{eqnarray}
Similarly, for the case $\mathbb{Q}_1(r)\neq0$ we obtain the same
expressions, and the only difference is in the torsion tensor square, which
now becomes
\begin{eqnarray}\label{inv1}
&& \!\!\! \!\!\! \!\!\! \!\!\! \!\!\!
\!\!\! \!\!\! \!\!\! \!\!\! \!\!\! \!\!\! T^{\mu \nu
\lambda}T_{\mu \nu \lambda} =
\frac{32\epsilon \alpha (3s^4-8M^2s^2+5M^4)\tanh^{-1}
\Big(\frac{M}{s}\Big) }{ r^3s^5}\nonumber\\
&&
+ \frac{ 24s^4[\epsilon(2Mc_1+c_2)-2M]+32\epsilon
M(3c_1s^3+19\alpha s^2-15\alpha M^2)}{3r^3s^4}
+{\mathcal{O}}\Big(\frac{1}{r^4}\Big)\,.
\end{eqnarray}
The above invariants reveal the presence of the singularity at $r=0$ as
expected, which is more mild than the case of simple TEGR, a known feature of
higher-order torsional theories \cite{Cai:2015emx, Nashed:2020kjh,
Bahamonde_2019,Ren:2021uqb}.
\subsection{Energy}
One of the advantages of teleparallel formulation of gravity is the easy
handling of the energy calculations, which is not the case in usual curvature
formulation \cite{Cai:2015emx}. We start with the gravitational
energy-momentum, $P^a$, which in integral form in
four dimensions is \cite{ULHOA_2013}
\begin{eqnarray} \label{en}
P^a=-\int_V d^3x \partial_i\Pi^{ai}f_T,\end{eqnarray}
where $V$ is the three-dimensional volume and $\Pi^{ai}=-4\pi
S^{a 0 i}$ is expressed in terms of the superpotential components. In the TEGR
limit, namely for $f_T=1$, the above expression reduces to the form given in
\cite{PhysRevD.65.124001}.
We start with the case $\mathbb{Q}_1(r)=0$. Inserting the tetrad functions
(\ref{mets1}),(\ref{mets1b}) into the tetrad (\ref{tet}) we can calculate the
involved
superpotential component as
\begin{equation}\label{sup}
S^{001}=\frac{6M^2s^3(\epsilon c_2-1)+\epsilon sM(76M\alpha-3c_1s^2-68\alpha
s^2)+60M\epsilon \alpha
\ln(2(M-s)(2M^2-s^2)}{6Mr^2s^3(1+c_2\epsilon)}+\frac{6\epsilon\alpha(3s^2-5M^2)}
{s^2(1+\epsilon c_2)\varrho_1}\,,
\end{equation}
which substituted into (\ref{en}) leads to
\begin{eqnarray} \label{en11}
&&
\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!
\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!
{P^0=E\approx
M+\epsilon\alpha\left\{M\frac{15M-38\varrho_1}{3s^2\varrho_1}-20\frac{
(2M^2-s^2)\ln[2(M-s)]}{s^3}\right\}-\frac{M^2+s^2}{2r}}\nonumber\\
&&
\!\!\!\!\!\!
{- 5\epsilon\alpha\left\{M\frac{15M-38\varrho_1}{3s^2\varrho_1 r
}+20\frac{(2M^2-s^2)\ln[2(M-s)]}{r s^3}\right\}={\cal
M}+O\left(\frac{1}{r}\right)}\,,
\end{eqnarray}
where ${{\cal M}\approx M-\epsilon\alpha \frac{38M}{3s^2}}$ is the
Arnowitt-Deser-Misner (ADM) mass
that contains $M$ and $ \epsilon$ up to first order.
In the case $\mathbb{Q}_1(r)\neq 0$, the above procedure leads to
\begin{equation}\label{sup1}
S^{001}=-\frac{M}{r^2}+\epsilon
\left[\frac{12\alpha(5M^4-8M^2s^2+3s^4)
\tanh^{-1}\left(\frac{M}{s}\right)+12Mc_3s^4(s+2)+4sM\alpha(17s^2-15M^2)
+3c_4s^5}{s^5r^2}\right] ,
\end{equation}
and then to
{\small{
\begin{eqnarray} \label{en1}
&&
\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!
P^0=E\approx M+\epsilon\left[\frac{12\alpha
\tanh^{-1}\left(\frac{M}{s}\right)(5M^4-8M^2s^2+3s^4)+
3Mc_3s^4(s+2)
+4sM\alpha(17s^2-15M^2)+3c_4s^5}{3s^5}\right]-\frac{M^2+s^2}{2r}
\nonumber\\
&&\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!
-\epsilon\left[\frac{12M\alpha\tanh^{-1}\left(\frac{M}{s}\right)\!
(10M^4\!-\!16M^2s^2\!+\!6s^4)+
3c_3s^4(3M^2s\!+\!6M^2\!-\!2s^2\!+\!s^3)
+8sM^2\alpha(19s^2\!-\!15M^2)
+6c_4Ms^5}{6rs^5}\right],
\end{eqnarray}}}
\begin{eqnarray}
&&
\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!
\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!
\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!
\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!
\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!
\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!
\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!
\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!
\!\!\!\!\!\!\!\!\!\!\!\!\!\!
{ E\approx {\cal M}_1+O\left(\frac{1}{r}\right)}\,,
\end{eqnarray}
where ${ {\cal
M}_1=M+\epsilon\left[\frac{68\alpha\,M+3Mc_3s(s+2)}{3s^2}\right]}$. Note that
for $\epsilon=0$, i.e. in
the TEGR limit, we recover the well-know energy expression of
Reissner-Nordstr\"om spacetime \cite{NASHED_2007}.
\section{ Geodesic deviation and stability of motion}\label{S6336b}
In this section we proceed to the examination of the stability of
motion around the obtained
black hole solutions, investigating the geodesic deviation.
The geodesic equations of a test particle in the gravitational field are given by
\begin{equation}\label{ge}
{d^2 x^\alpha \over d\lambda^2}+ \Bigl\{^\alpha_{ \beta \rho} \Bigr \}
{d x^\beta \over d\lambda}{d x^\rho \over d\lambda}=0,
\end{equation}
where $s$ denotes the affine connection parameter and $\Bigl\{^\alpha_{ \beta
\rho} \Bigr \}$ the Levi-Civita connection. The
geodesic deviation equations acquire the form
\cite{1992ier..book.....D}
\begin{equation} \label{ged}
{d^2 \psi^\sigma \over d\lambda^2}+ 2\Bigl\{^\sigma_{ \mu \nu} \Bigr \}
{d x^\mu \over d\lambda}{d \psi^\nu \over d\lambda}+
\Bigl\{^\sigma_{ \mu \nu} \Bigr \}_{,\ \rho}
{d x^\mu \over d\lambda}{d x^\nu \over d\lambda}\psi^\rho=0\,,
\end{equation}
where $\psi^\rho$ is the 4-vector deviation.
In the case of the spherically symmetric ansatz (\ref{met1}) the above
expressions give
\begin{eqnarray}
&&
{d^2 t \over d\lambda^2}=0,\nonumber\\
&& {1 \over 2} a'(r)\left({d t \over
d\lambda}\right)^2-r\left({d \phi \over d\lambda}\right)^2=0, \nonumber\\
&&{d^2
\theta \over d\lambda^2}=0,\nonumber\\
&& {d^2 \phi \over d\lambda^2}=0,
\end{eqnarray}
and therefore for the geodesic deviation we finally obtain
\begin{eqnarray}\label{ged11}
&& {d^2 \psi^1 \over d\lambda^2}+b(r)a'(r) {dt
\over d\lambda}{d
\psi^0 \over d\lambda}-2r b(r) {d \phi \over d\lambda}{d \psi^3 \over
d\lambda}+\left\{{1 \over 2}\left[a'(r)b'(r)+b(r) a''(r)
\right]\left({dt \over d\lambda}\right)^2-\left[b(r)+rb'(r)
\right] \left({d\phi \over d\lambda}\right)^2 \right\}\psi^1=0\,, \nonumber\\
&& {d^2 \psi^0 \over
d\lambda^2}+{b'(r) \over b(r)}{dt \over d\lambda}{d \psi^1 \over
d\lambda}=0\,,\nonumber\\
&& {d^2 \psi^2 \over d\lambda^2}+\left({d\phi \over d\lambda}\right)^2
\psi^2=0\,, \nonumber\\
&& {d^2 \psi^3 \over d\lambda^2}+{2 \over r}{d\phi \over d\lambda} {d
\psi^1 \over d\tau}=0\,.
\end{eqnarray}
Using
the circular orbit
$\theta={\pi \over 2}$, $
{d\theta \over d\lambda}=0$, and ${d r \over d\lambda}=0,
$
we acquire
$
\left({d\phi \over d\lambda}\right)^2={a'(r)
\over r[2a(r)-ra'(r)]}$ and
$
\left({dt \over
d\lambda}\right)^2={2 \over 2a(r)-ra'(r)}$, and thus equations (\ref{ged11})
can be rewritten as
\begin{eqnarray}
\label{ged22} &&
{d^2 \psi^1 \over d\phi^2}+a(r)a'(r) {dt \over
d\phi}{d \psi^0 \over d\phi}-2r a(r) {d \psi^3 \over
d\phi} +
\left\{{1 \over 2}\left[a'^2(r)+a(r) a''(r)
\right]\left({dt \over d\phi}\right)^2-\left[a(r)+ra'(r)
\right] \right\}\psi^1=0\,, \nonumber\\
&& {d^2 \psi^2 \over d\phi^2}+\psi^2=0\,, \nonumber\\
&& {d^2 \psi^0 \over d\phi^2}+{a'(r) \over
a(r)}{dt \over d\phi}{d \psi^1 \over d\phi}=0\,, \nonumber\\
&& {d^2 \psi^3 \over
d\phi^2}+{2 \over r} {d \psi^1 \over
d\phi}=0\,. \end{eqnarray}
The second equation of (\ref{ged22}) corresponds to a simple harmonic motion,
which indicates that the plane $\theta=\pi/2$ is stable.
Moreover, the other equations of (\ref{ged22}) have solutions of the
form:
\begin{equation} \label{ged33}
\psi^0 = \zeta_1 e^{i \sigma \phi}\,, \qquad \psi^1= \zeta_2e^{i
\sigma
\phi}\,, \qquad {\text{and}} \qquad \psi^3 = \zeta_3 e^{i \sigma
\phi}\,,
\end{equation}
where $\zeta_1, \zeta_2$ and $\zeta_3$ are constants. Substituting
(\ref{ged33}) into
(\ref{ged22}), we extract
the stability of motion condition as:
\begin{equation} \label{con1}
\frac{3abb'-\sigma^2a b'-2rb^{3/2}a'^{3/2}-rab'^2+rab'a'+rab a''}{a b'}>0.
\end{equation}
Equation (\ref{con1}) has the following solution in terms of the metric
potentials
\begin{equation} \label{stab1}
\sigma^2= \frac{3a bb'-2rb^{3/2}a'^{3/2}-rab'^2+rab'a'+rab a''}{a^2
b'^2}>0.\end{equation}
Hence, in order to conclude on the stability of motion around the
obtained black-hole
solutions, for the case $\mathbb{Q}_1(r)=0$ in the above expressions we
insert
$a(r)$ and $b(r)$ from (\ref{mets1}),(\ref{mets1b}), while for the case
$\mathbb{Q}_1(r)\neq 0$ from (\ref{mets2bb}).
\begin{figure}[ht]
\centering
\includegraphics[scale=0.3]{JMPFBPstt.eps}
\includegraphics[scale=0.3]{JMPFBPstn.eps}
\caption{ {\it{ The black-hole stability of motion parameter $\sigma^2$
versus
$r$. Left graph: $\mathbb{Q}_1(r)=0$ with $M=9$, $c_1=c_2=s=1$ and
various choices of the model parameters $\alpha$ and $\epsilon$ in Planck mass
units. Right graph: $\mathbb{Q}_1(r)\neq 0$ with $M=0.01$, $c_1=c_2=1$ and
various choices of the model parameters $\alpha$, $\epsilon$ and $s$ in Planck
mass
units.}}}
\label{Fig:1}
\end{figure}
In order to present the above results in a more transparent way, in Fig.
\ref{Fig:1} we depict the behavior of $\sigma^2$ for various choices of the
model parameters, for the two cases $\mathbb{Q}_1(r)=0$ and
$\mathbb{Q}_1(r)\neq 0$ separately. Note that for $\mathbb{Q}_1(r)=0$ we
always obtain stability of motion as expected, while for $\mathbb{Q}_1(r)\neq
0$ we find
potentially unstable regions.
\section{Thermodynamics}\label{S5}
In this section we perform an analysis of the thermodynamic properties of the
obtained black-hole solutions. Since the nature of the solutions and especially
their thermodynamic features change for $\mathbb{Q}_1(r)= 0$ and
$\mathbb{Q}_1(r)\neq 0$, in the following we examine the two cases separately.
\subsection{Thermodynamics of the black hole solution with $\mathbb{Q}_1(r)=
0$}\label{S5a}
We start by investigating the black-hole solution of the case $\mathbb{Q}_1(r)=
0$ given in (\ref{mets1}),(\ref{mets1b}).
In the left graph of Fig. \ref{Fig:2} we display the metric
potentials $a(r)$ and $b(r)$. As we can see, $a(r)$ may exhibit two horizons
while $b(r)$ does not. In the right graph of Fig. \ref{Fig:2} we focus on
$a(r)$, in order to make more transparent the behavior of its possible two
horizons, acquired by solving $a(r) = 0$, namely
$r_-$ which denotes the inner Cauchy horizon of the black
hole and $r_+$ which is the outer event horizon. In particular, for small
$\alpha$ values, namely small deviations from general relativity,
we obtain two horizons, however as $\alpha$ increases there is a specific
value in which the two horizons become degenerate ($r_{-}=r_{+}=r_d $), while
for larger values the horizon disappears and the central singularity
becomes a naked one. This is a known feature of torsional gravity,
namely for some regions of the parameter space naked singularities appear
\cite{Gonzalez:2011dr,Capozziello:2012zj,Nashed:2018cth}.
Finally, let us calculate the total mass contained within the event horizon
$r_+$. We find the mass-radius expression as
\begin{eqnarray} \label{hor-mass-rad1a}
&& {{{\cal{M}_+}}\equiv{{\cal{M}(r_+)}} \approx \frac{r_+}{2}\,,}
\end{eqnarray}
where ${{\cal{M}_+}}$ is given by Eq. (\ref{en11}) for $r_+$ in place of $r$.
\begin{figure}[ht]
\centering
\includegraphics[scale=0.3]{JRFMMMBRefab.eps}
\includegraphics[scale=0.3]{JMPFBPhort.eps}
\caption{{\it{Left graph: The two metric potentials $g_{tt}\equiv a(r)$ and
$g_{rr}\equiv b(r)$ given in
(\ref{mets1}),(\ref{mets1b}) versus $r$, for the black hole solution with
$\mathbb{Q}_1(r)=
0$, for $M=9$, $c_1=c_2=1$, $\alpha=0.1$,
$\epsilon=0.1$, and
$s=4$, in Planck
mass units. Right graph: The metric potential $a(r)$ versus $r$,
for $M=9$, $c_1=1$, $s=4$ and various values of the model parameters $c_2$,
$\alpha$ and $\epsilon$, in Planck
mass units. $r_-$ and $r_+$ are the inner and outer horizons
respectively, while $r_d $ is the degenerate horizon in which the above two
coincide. }}}
\label{Fig:2}
\end{figure}
We proceed by examining the temperature.
The Hawking black-hole temperature is defined as
\cite{PhysRevD.86.024013,Sheykhi:2010zz,Hendi:2010gq,PhysRevD.81.084040}
\begin{equation}
T_+ \equiv T(r_+)= \frac{a'(r_+)}{4\pi},
\label{tempdef}
\end{equation}
with $r = r_+$ the event horizon, which satisfies
$a'(r_+)\neq 0$. Additionally, in the framework of $f(T)$ gravity, the
black-hole entropy is given by
\cite{Miao:2011ki,PhysRevD.84.023515,Zheng:2018fyn}
\begin{equation}\label{ent}
S_+ \equiv S(r_+)=\frac{A}{4 f_{T}(r_+)},
\end{equation}
where $A$ is the area. Inserting the $f(T)$ form (\ref{tor}) and the
solution (\ref{mets1}),(\ref{mets1b}) into the above definitions we find
\begin{equation}\label{m44-ee}
{T_+} \simeq\frac{3r_+{}^2s^3+\epsilon\alpha[30s^2r_++11s^3+30r_+{}
^3\ln(2/r_+{})+45sr_+{}^2]+3\epsilon s^3r_+{}^2c_2}{12\pi r_+{}^3}\,,
\end{equation}
and
\begin{eqnarray} \label{ent1}
{S_+}\simeq \pi
r_+{}^2\left[1+2\epsilon \alpha
\left(\frac{4}{r_+{}^2}+ \frac{\cal{M}_+{}^2-2s^2}{r_+{}^4}\right)\right]\,.
\end{eqnarray}
These expressions
indicate that for $\epsilon=0$ we recover the standard
general-relativity temperature and entropy.
In Fig. \ref{Fig:3} we depict the temperature and entropy versus the
horizon, for various values of the model parameters. As we can see the entropy
is always positive and exhibits a quadratic behavior, while the temperature is
always positive when $\epsilon>0$ but for vanishing
$\epsilon$ it is positive only for $r_d>r_+$.
\begin{figure}[ht]
\centering
\includegraphics[scale=0.3]{JMPFBPtemptr.eps}
\includegraphics[scale=0.3]{JMPFBPenttr.eps}
\caption{{\it{ The temperature (left graph) and entropy (right
graph) versus the horizon,
for the black hole solution with
$\mathbb{Q}_1(r)=
0$, for $c_1=1$, $c_2=5$ and $s=4$ and for various values of the model
parameters $\alpha$ and $\epsilon$, in Planck
mass units.
}}}
\label{Fig:3}
\end{figure}
We now focus on the heat capacity, which is a crucial quantity concerning the
thermodynamic stability \cite{Nashed:2003ee,Myung_2011,Myung:2013oca}, since our
perturbative approach to the black-hole solution allows for an easy calculation.
The heat capacity at the event horizon
is defined as \cite{Nouicer:2007pu,DK11,Chamblin:1999tk}:
\begin{equation}\label{heat-capacity}
C_+\equiv C(r_+)\simeq\frac{\partial \cal{M}_+}{\partial T_+}=\frac{\partial
\cal{M}_+}{\partial r_+} \left(\frac{\partial T_+}{\partial r_+}\right)^{-1}\, ,
\end{equation}
and positive heat capacity implies thermodynamic stability.
Substituting (\ref{hor-mass-rad1a}) and (\ref{m44-ee}) into
(\ref{heat-capacity}) we obtain the heat capacity as
\begin{equation}\label{heat-cap1a}
C_+ \simeq\frac{2\pi r_+^2
}{3s^3}\Big[\epsilon \alpha (30 r_++7s-60r\ln2)-3s^3\Big] \, .
\end{equation}
Expression (\ref{heat-cap1a}) implies that $C_+$ does not diverge and
thus we do not have a second-order phase transition. In the left graph of
Fig. \ref{Fig:3b} we depict $C_+$ as a function of the horizon. As we can
see, in the $\epsilon=0$ case we have $C_+<0$ due to the negative derivative
of the temperature, as expected for the the Reissner Nordstr\"om
black hole. Nevertheless, for $\epsilon>0$ we obtain positive heat
capacity. This is one of the main results of the present work, namely that
$f(T)$ modifications
improve the thermodynamic stability. Note that this is not the case in other
gravitational modifications, since for instance in $f(R)$ gravity the heat
capacity is positive
only conditionally \cite{Elizalde:2020icc,Nashed:2020kdb,Nashed:2020mnp}.
We close this subsection by the examination of the Gibb's free energy. In terms
of the the mass, temperature and entropy at
the event horizon this is defined as
\cite{Zheng:2018fyn,Kim:2012cma}:
\begin{equation}
\label{enr1}
G(r_+)={\cal{M}}(r_+)-T(r_+)S(r_+)\,.
\end{equation}
Inserting (\ref{hor-mass-rad1a}), (\ref{m44-ee})
and (\ref{ent1}) into (\ref{enr1}), we obtain
\begin{equation} \label{m77}
{G_+}\equiv G(r_+) =\frac{3s^3(3s^2+r_+{}^2)+\epsilon\{\alpha[30r_+{}
^3ln(2r_+)+31sr_+{}^2+3s^3r_+{}^2-30s^2r_+-231s^3]\}+3\epsilon
s^3r_+(r_+c_2-c_1)}{12r_s{}^3}.
\end{equation}
In the right graph of Fig. \ref{Fig:3b} we depict the behavior of Gibb's free
energy.
As we observe it is always positive, for both $\epsilon=0$ and
$\epsilon>0$.
\begin{figure}[ht]
\centering
\includegraphics[scale=0.33]{JMPFBPheatcaptr.eps}
\includegraphics[scale=0.33]{JMPFBPGibbtr.eps}
\caption{{\it{
The heat capacity (left graph) and the Gibb's energy (right
graph) versus the horizon,
for the black hole solution with
$\mathbb{Q}_1(r)=
0$, for $c_1=1$, $c_2=5$ and $s=4$ and for various values of the model
parameters $\alpha$ and $\epsilon$, in Planck
mass units.}}}
\label{Fig:3b}
\end{figure}
\subsection{Thermodynamics of the black hole solution with $\mathbb{Q}_1(r)\neq
0$}\label{S5b}
In this subsection we repeat the above thermodynamic analysis in the case of
the black hole solution for $\mathbb{Q}_1(r)\neq
0$ given in (\ref{mets2}),(\ref{mets2bb}).
In the left graph of Fig. \ref{Fig:4} we depict the metric
potentials $a(r)$ and $b(r)$, and as we observe $a(r)$ may exhibit two horizons
while $b(r)$ does not. In the right graph of Fig. \ref{Fig:4} we present
$a(r)$. Similarly to the previous subsection, we see that for small
$\alpha$ values, namely small deviations from general relativity,
we obtain two horizons, however as $\alpha$ increases there is a specific
value in which the two horizons become degenerate ($r_{-}=r_{+}=r_d $), while
for larger values the horizon disappears and the central singularity
becomes a naked one. However, the interesting feature is that for the
same $\alpha$ value, the parameter $s$ that quantifies the charge profile also
affects the horizon structure, and in particular larger $s$ leads to the
appearance of the naked singularity.
\begin{figure}[ht]
\centering
\includegraphics[scale=0.3]{JRFMMMBRefabn.eps}
\includegraphics[scale=0.3]{JMPFBPhorn.eps}
\caption{{\it{Left graph:The two metric potentials $g_{tt}\equiv a(r)$ and
$g_{rr}\equiv b(r)$ given in
(\ref{mets2}),(\ref{mets2bb}) versus $r$, for the black hole solution with
$\mathbb{Q}_1(r)\neq
0$, for $M=9$, $c_1=c_2=1$, $\alpha=0.1$,
$\epsilon=0.1$, and
$s=4$, in Planck
mass units. Right graph: The metric potential $a(r)$ versus $r$,
for $M=9$, $c_3=1$, $c_4=1$ and various values of the model parameters
$\alpha$, $\epsilon$ and $s$, in Planck
mass units. $r_-$ and $r_+$ are the inner and outer horizons
respectively, while $r_d $ is the degenerate horizon in which the above two
coincide. }}}
\label{Fig:4}
\end{figure}
The mass-radius relation takes the form
\begin{eqnarray} \label{hor-mass-rad1ac}
{M_+}
=\frac{s^2+r_+{}^2}{2r_+}+\epsilon\left[\frac{3sc_3(2s^2+2r_+{}^2+sr_+{}^2)
+3s^2c_4r_++56\alpha (s^2+r_+{}^2)}{s^2r_+}\right]\,,
\end{eqnarray}
and it is
plotted in Fig. \ref{Fig:4mass}, where we can verify that
$M_+$ is always positive.
\begin{figure}[ht]
\centering
\includegraphics[scale=0.3]{JMPFBPmasstrefn.eps}
\caption{{\it{ The mass-radius relation (\ref{hor-mass-rad1ac}) for the black
hole solution with
$\mathbb{Q}_1(r)\neq
0$, for $c_3=1$, $c_4=0$, and $s=4$,
and various values of the model parameters
$\alpha$, $\epsilon$ in Planck
mass units. }}}
\label{Fig:4mass}
\end{figure}
For the temperature (\ref{tempdef}) we obtain
\begin{eqnarray} \label{m44ec}
{T_+} \simeq\frac{3s^2(r_+{}^2-s^2)+\epsilon [r_+{}^2(56\alpha
+3s^2c_3+6c_3s)+6s^2(4\alpha-sc_3)]}{12\pi s^2r_+{}^3}\,.
\end{eqnarray}
Moreover, for the entropy (\ref{ent}) we find
\begin{eqnarray} \label{ent1c}
&&{S_+}_ \simeq\frac{\pi}{r_+{}^3}\left[r_+{}^5-2\epsilon \alpha
M(r_+M-2s^2)
\right]\,,
\end{eqnarray}
which again for $\epsilon=0$ recovers the general relativity result.
In Fig. \ref{Fig:5} we depict the temperature and entropy versus the
horizon, for various values of the model parameters. We mention that in this
case both temperature and entropy may acquire negative values, however the
entropy, which is always quadratically increasing, is positive when
$r_+>r_d$.
\begin{figure}[ht]
\centering
\includegraphics[scale=0.3]{JMPFBPtempn.eps}
\includegraphics[scale=0.3]{JMPFBPentn.eps}
\caption{{\it{ The temperature (left graph) and entropy (right
graph) versus the horizon,
for the black hole solution with
$\mathbb{Q}_1(r)\neq
0$, for $M=9$, $c_3=c_4=1$ and $s=4$,
and various values of the model parameters
$\alpha$, $\epsilon$ in Planck
mass units.
}}}
\label{Fig:5}
\end{figure}
For the heat capacity $
C_+ =\frac{\partial
M_+}{\partial r_+} \left(\frac{\partial T_+}{\partial r_+}\right)^{-1}
$, using
(\ref{hor-mass-rad1ac}) and (\ref{m44ec}), we acquire
\begin{equation}\label{heat-cap1ac}
{C_+} \simeq-2\pi(2s^2+r_+{}^2-2s^2\epsilon c_3-8-\epsilon \alpha)
\, .
\end{equation}
Expression (\ref{heat-cap1ac}) implies that $C_+$ does not diverge and
therefore we do not have a second-order phase transition. In the left graph of
Fig. \ref{Fig:5b} we present $C_+$ as a function of the horizon. As we can
see, in the $\epsilon=0$ case we have $C_+<0$ due to the negative derivative
of the temperature, as expected for the the Reissner Nordstr\"om
black hole. Nevertheless, for $\epsilon>0$, namely in the case where
the $f(T)$ correction is switched on, we may obtain positive heat
capacity.
Finally, for the Gibb's free energy $G(r_+)=M(r_+)-T(r_+)S(r_+)\,$,
using (\ref{hor-mass-rad1ac}), (\ref{m44ec}) and (\ref{ent1c}), we find
\begin{eqnarray} \label{m77c}
&&{G_+} \simeq\frac{3s^2(r_+{}^2+s^2)+\epsilon[3
c_3s(2r_+{}^2+sr_+{}^2+6s^2)+6s^2r_+c_4+8\alpha(7r_+{}^2+11s^2)]}{12s^2r_+}.
\end{eqnarray}
In the right graph of Fig. \ref{Fig:5b} we present Gibb's free
energy as a function of the horizon.
As we can see for both $\epsilon=0$ and $\epsilon>0$ it is always positive.
\begin{figure}[ht]
\centering
\includegraphics[scale=0.33]{JMPFBPheatcapn.eps}
\includegraphics[scale=0.33]{JMPFBPGibbn.eps}
\caption{{\it{
The heat capacity (left graph) and the Gibb's energy (right
graph) versus the horizon,
for the black hole solution with
$\mathbb{Q}_1(r)\neq
0$, for $M=9$, $c_3=c_4=1$ and $s=4$,
and various values of the model parameters
$\alpha$, $\epsilon$ in Planck
mass units.}}}
\label{Fig:5b}
\end{figure}
\section{Conclusions and discussion}
\label{Conclusions}
We investigated the stability of motion and thermodynamics in the case
of spherically symmetric
solutions in $f(T)$ gravity using the perturbative approach. In particular, we
considered small deviations from teleparallel equivalent of general relativity
and we extracted charged black hole solutions for two charge profiles, namely
with or without a perturbative correction in the charge distribution.
Firstly, we examined their asymptotic behavior showing that for large distances
they become Minkowski. Then we extracted various torsional and curvature
invariants, which revealed the presence of the central singularity as expected.
Moreover, we calculated the energy and the mass of the solutions. As we showed,
all
results recover the general relativity ones in the case where the $f(T)$
deviation goes to zero.
As a next step we investigated the stability of motion around the
obtained
black hole solutions, by extracting and studying the geodesic
deviation of a test particle in their gravitational
field. Assuming a secular orbit, we extracted the corresponding stability of
motion
condition in terms of the metric potentials. As we saw, in the case where the
perturbative correction to the charge profile is absent the solution is always
stable, however in the case where it is present we obtained unstable regimes in
the parameter space.
Additionally, we performed a detailed analysis of the thermodynamic properties
of the black hole solutions. In particular, we extracted the inner (Cauchy) and
outer (event) horizons, the mass profile, the temperature, the entropy, the
heat capacity and the Gibb's free energy. As we showed, for
small $\alpha$ values, namely small deviations from general relativity,
we obtain the two horizons, however as $\alpha$ increases there is a
specific value in which the two horizons become degenerate, and
for larger values the horizon disappears and the central singularity
becomes a naked one, a known feature of torsional
gravity. Furthermore, we saw that for the
same $\alpha$ value, the parameter $s$ that quantifies the charge
profile also affects the horizon structure, and in
particular larger $s$ leads to the appearance of the naked singularity.
Concerning the temperature and entropy, we showed that although there are
regimes in which they become negative, for $r_+>r_d$ they are always
positive definite. Concerning the heat
capacity we saw that it does not diverge and
thus we do not have a second-order phase transition. However, the most
interesting result is that it becomes positive for larger
deviations from general relativity, which shows that $f(T)$ modifications
improve the thermodynamic stability, which is not the case in other
gravitational modifications.
Finally, for the Gibb's free energy, we showed that it is always positive, for
all torsional additions and for both charge-profile cases.
In summary, the present work indicates that torsional modification of gravity
may have an advantage comparing to other gravitational modification classes,
when stability issues are raised, which may serve as an additional motivation
for the corresponding investigations. One particular interesting issue is to
investigate in detail whether torsional modified gravity leads to smoother
(weaker) central singularities comparing to general relativity or curvature
modified gravity. This issue will be the focus of interest of a separate
project.
|
\section{Implementation details}
\label{sec:impl}
Sec.~\ref{sec:ad} details the technical aspects of training the framework by means of gradient-descent and the numerical libraries that have been employed.
Sec.~\ref{sec:withtime} reviews the extensions of the framework enabling the optimization of both the control functions and control times.
Finally, in Sec.~\ref{sec:hp} we describe how the hyperparameters of the framework are selected and provide details of the NNs which have been used for the results reported in Table~I.
\subsection{Training with gradient descent using the automatic differentiation toolbox}
\label{sec:ad}
Recall from the main text that, the training of the framework consists in optimizing the weights of the neural network (NN), denoted $\phi_{NN}$, to produce controls $f_{\paramstgt}(t)$ which minimize the deviation $\mathcal{I}$ between the induced propagators $U_{\alpha}$ and the targeted ones $U^{tgt}_{\alpha}$.
$\mathcal{I}$ is characterized by the average of the individual infidelities $\mathcal{I}_{\alpha}$ (Eq.~(2)), each corresponding to a fixed set of target parameters $\alpha$, taken over the domain of $\alpha$.
Optimization of $\phi_{NN}$ to minimize $\mathcal{I}$ is performed by gradient-descent, and we now describe the steps of implementation required.
For fixed parameters $\alpha$ but variable time $t$, the NN (as sketched in Fig.~1) outputs the time-dependent control values $f_{\paramstgt}(t)$ used to construct the system Hamiltonian $H_\alpha(t)$ (\emph{e.g.}, Eq.~(1)).
Dynamical evolution of the system under this Hamiltonian is performed by numerically solving the corresponding ordinary differential equation (ODE), which for closed quantum systems is the Schr\"{o}dinger equation.
Finally, the individual infidelity $\mathcal{I}_{\alpha}$ is evaluated based on the propagator $U_{\alpha} = U_{\alpha}(t = T)$, obtained at the end of the numerical evolution of the system, and its corresponding target $U_{\alpha}^{tgt}$.
Having detailed how $\mathcal{I}_{\alpha}$ is evaluated, it remains to detail how its gradients $\nabla_{\phi_{NN}}\mathcal{I}_{\alpha}$, with respects to the weights of the NN, are obtained.
Evaluations of these gradients necessitate to combine intermediate gradients resulting from each of the steps of computation discussed in the previous paragraph.
To make such gradient composition as effortless as possible, we resort to the automatic differentiation (AD) toolbox.
In a nutshell, AD provides the machinery to systematically extend arbitrary numerical computations to simultaneously compute its derivatives, at the cost of only a modest computational overhead~\cite{BARTHOLOMEWBIGGS2000171,baydin2018automatic}.
This concept is central to the training of many machine learning algorithms, where it is most famously known as back-propagation.
Still, AD is not limited to the training of NNs but applies to virtually any computational routines.
In particular, recent advances in AD popularized in~\cite{NEURIPS2018_69386f6b} permit the efficient differentiation over the trajectory of a system which dynamics is obtained by numerically solving an ODE.
For the implementation of these techniques we resort to the package~\cite{torchdiffeq} built on top of the widely-adopted machine learning library pytorch~\cite{NEURIPS2019_9015}.
Equipped with these libraries, it is only required to define how $\mathcal{I}_{\alpha}$ is evaluated (as was detailed in the previous paragraph), but, it is left to the underlying numerical routines to manage the intricacies of gradient computation.
Finally, the individual infidelities and corresponding gradients need to be averaged over the continuous domain of $\alpha$.
This average is approximated by means of Monte Carlo sampling over a finite number of $128$ values of $\alpha$ randomly drawn.
Given the stochastic nature of the resulting gradients, we resort to the stochastic gradient optimization routine Adam~\cite{kingma2014adam}.
\subsection{Learning control functions and control durations}
\label{sec:withtime}
In addition to the derivatives with respect to the control values, the techniques in~\cite{NEURIPS2018_69386f6b} (see Appendix.C) can also evaluate derivatives $\nabla_T\mathcal{I}_{\alpha}$ of the infidelities with respect to the duration $T$ of the control tasks.
Hence, this duration can be treated as a variable to be optimized over, rather than a fixed quantity.
Optimizing the cost $\mathcal{I_{\alpha}} + \mu \times T$, which now includes the penalty term $\mu \times T$ (with $\mu > 0$),
permits us to minimize the duration $T$ common to all the gates to be realized. This methodology is followed for the results presented in Table~I.
Going further, it is also possible to consider durations $T_\alpha$ which depend on $\alpha$, that is, target-dependent durations.
Given that the exact nature of this dependency is unknown beforehand it needs to be learnt.
To this intent, a second NN, taking $\alpha$ as an input and returning $T_{\alpha}$ as an output, is trained concurrently with the original NN to minimize the cost $\mathcal{I_{\alpha}} + \mu \times T_{\alpha}$.
Results of this approach were illustrated in Fig.~3.
In both these scenarios, the value of $\mu$ can be varied to put more or less emphasis on the objective of accurate control or on the objective of fast operations.
The value of $\mu=10^{-2}$ was found to be a good compromise, and was used for the optimization results presented.
Finally, given that the controls produced by the (original) NN should change depending on the control duration $T$ (or $T_{\alpha}$),
it is extended to also take the durations $T$ (or $T_{\alpha}$) as input. Note that this additional input is only required when the durations vary during the course of the optimization, that is, is not needed when $T$ is taken to be fixed (as was the case for the results presented in Fig.~2).
\subsection{Choice of the hyperparameters}
\label{sec:hp}
It is known that the quality of the solutions found by a NN can be affected by several design choices commonly referred to as hyperparameters~\cite{claesen2015hyperparameter}.
In particular, the choice of the size of the NN, the way its weights are initialized, and the learning rates employed during its optimization are often considered to be of particular importance.
As the optimal hyperparameters values are not known beforehand, they need to be identified by means of trial-and-error.
For instance, it is common to randomly sample~\cite{bergstra2012random} several hyperparameters configurations and to select a posteriori the ones resulting in optimal training. As each of these random configurations is evaluated independently, this approach has the benefit of being straightforward to parallelize.
To accelerate this search, it is based on partial training of the NNs with only a limited number $\sim 10$ of iterations, such as to quickly discover the most promising hyperparameters values.
Once identified, these configurations are used to perform a complete training of the framework, consisting of a number of $400$ iterations reported in the main text.
The hyperparameters used for the results of Table.~I are reported in Table.~\ref{tab:hp}.
The NNs employed are systematically taken to contain a number $n_L \in [4,10]$ of layers with, $n_N\in[150, 300]$ nodes per layer.
The inner layers are chosen to be ReLU layers which are known to limit issues of gradient vanishing when training NNs~\cite{pmlr-v15-glorot11a}.
The output layer, however, is taken to be a sigmoid layer which has bounded output values in the range $[0, 1]$, that are further shifted and rescaled to be in the desired range $[-1, 1]$ allowed for the control amplitudes.
Learning rates used for the optimization of the weights of the NNs with Adam, are considered in the range $l\in [10^{-4} , 10^{-2}]$, a common choice when training NNs.
Finally, we find the amplitudes of the initial weights of the NN to play an important role in the successful training of the framework. In particular, the default distribution of the initial weights yields small initial control amplitudes, which are prone to quickly converge to $0$. To avoid such effect, the initial weights are rescaled by a factor $\beta \in [1.8, 2.2]$ such that the control values initially produced by the NNs are different enough from $0$.
\setlength{\tabcolsep}{5pt}
\begin{table}[t]
\caption{Details of the hyperparameters used for the $7$ families of unitaries studied in Table.~I.
These include the number $n_L$ of NN layers and the number $n_U$ of units per layer, the learning rate $l$ of the Adam optimizer and the scales $\beta$ of the initial weights of the NN.
}
\centering
\begin{tabular}{ll ccc}
\hline\hline
& $U^{tgt}_{\alpha}$ & $n_L \times n_U$ & $l (10^{-3})$ & $\beta$\\
\hline
(i) & $\ket{0}\bra{0}\otimes I+\ket{1}\bra{1}\otimes \exp(-i\alpha_1\sigma^{(2)}_z)$ & $6 \times 150$ & $1$ & $1.8$ \\
(ii) & $\exp\left(-i\alpha_1 \sigma^{(1)}_z\sigma^{(2)}_z\right)$ & $6 \times 150$ & $1$ & $2.1$ \\
(iii) & $\ket{0}\bra{0}\otimes I+\ket{1}\bra{1}\otimes U_{1}(\vec\alpha)$ & $8 \times 300$ & $3$ & $1.9$\\
(iv) &
$\exp(-i\sum_{j}\alpha_j\sigma_j^{(1)}\sigma_j^{(2)})$ & $8 \times 300$ & $1$ & $1.9$\\
(v) & $\exp(-i \alpha_1 \sigma_z^{(1)}\sigma_z^{(2)}\sigma_z^{(3)})$ & $10 \times 200$ & $4$ & $2.2$\\
(vi) & $\exp(-i\sum_{j}\alpha_j\sigma_j^{(1)}\sigma_j^{(2)}\sigma_j^{(3)})$ & $10 \times 300$ & $2$ & $2.0$\\
(vii) & $\left(I-\ket{11}\bra{11}\right)\otimes I+\ket{11}\bra{11}\otimes U_{1}(\vec\alpha)$ & $10 \times 300$ & $4$ & $2.1$\\
\hline
\end{tabular}\label{tab:hp}
\end{table}
\section{Evaluation of the times entailed by gate decomposition}
\label{sec:dec}
In Table~I are reported the ratios $R$ between the times $T^{dec}$ entailed by gate decomposition of the unitaries considered and the control times $T$ required by the framework. We now detail the methodology used to evaluate $T^{dec}$.
Gate decompositions are performed with respect to the set of elementary gates which are supported by the Hamiltonian in Eq.~(1), that is, the gate set
\begin{equation}
\mathcal{G}=\{R_\sigma(\theta)=\exp(-i \frac{\theta}{2} \sigma)\}
\end{equation}
of rotations generated by the operators $\sigma \in \{\sigma^{(i)}_y,\sigma^{(i)}_z, \sigma^{(i,i'\neq i)}_{xx}\}$ acting on qubits $i,i'=1,\hdots,n$.
To systematize such decompositions, we rely on established automated routines and resort to the qiskit compilation algorithm.
Among the gate sets supported by qiskit, we select the gateset composed of two qubits rotations $R_{\sigma_{xx}}$ (between arbitrary pairs of qubits) and single qubit rotations
\begin{equation}
U_3(\theta, \phi, \lambda) = \begin{pmatrix}
\cos (\frac{\theta}{2}) & - e^{i \lambda} \sin (\frac{\theta}{2})\\
e^{i\phi} \sin (\frac{\theta}{2}) & e^{i (\lambda+\phi)} \cos (\frac{\theta}{2})
\end{pmatrix},
\end{equation}
which can be further broken down as elements of $\mathcal{G}$ as $R_{\sigma_z}(-\lambda)R_{\sigma_y}(\theta)R_{\sigma_z}(-\phi)$, up to a global phase.
Given the resulting gate decomposition of a unitary $U^{tgt}_{\alpha}$,
it remains to assess the time $T^{dec}_{\alpha}$ that would be needed for its implementation.
For the range of values $[-1,1]$ permitted for the controls of the Hamiltonian in Eq.~(1), it follows that the minimal time required for the implementation of any gate $R_{\sigma}(\theta)$ is $|\theta / 2|$.
Hence, a sequential application of $l$ gates $\prod_l R_{\sigma_l}(\theta_l)$ necessitates an implementation time of $T^{dec}=\sum_l |\theta_l / 2|$.
For instance, a gate $U_3(\theta, \psi, \lambda)$ requires a time $T^{dec}=(|\theta| + |\psi| +|\lambda|)/2$.
It should be noted that distinct sets of angles $(\theta, \psi, \lambda)$ can yield the same unitary $U_3(\theta, \psi, \lambda)$ but could entail different implementation times; the fastest implementation is systematically considered.
Finally, the times $T^{dec}_{\alpha}$ effectively depend on the individual targets $U^{tgt}_{\alpha}$ to be realized. As such, for each of the $7$ families of gates presented in Table.~I, these times are averaged over an ensemble of $100$ gates randomly drawn.
|
\section{Introduction}
Nowadays, word-based neural language models (LMs) consistently give better perplexities than count-based language models \cite{sundermeyer2012lstm, irie19:trafolm}, and are commonly used for second-pass rescoring or first-pass decoding of automatic speech recognition (ASR) outputs \cite{6854535, beck2019lstm, beck2020lvcsr}.
One challenge to train such LMs, especially when the vocabulary size is large, is the traversal over the full vocabulary in the softmax normalization.
During both training and testing, this brings inefficiencies and calls for sampling-based methods to ease the calculation burdens.
Previously, many sampling-based training criteria are proposed and investigated.
Some prominent examples include: hierarchical softmax \cite{DBLP:journals/corr/abs-1301-3781}, negative sampling \cite{DBLP:conf/nips/MikolovSCCD13}, importance sampling (IS) \cite{pmlr-vR4-bengio03a}, and noise contrastive estimation (NCE) \cite{pmlr-v9-gutmann10a}.
While NCE is one of the most popular choices for language modeling \cite{DBLP:conf/icml/MnihT12, goldberger-melamud-2018-self, 7179005}, recently Gao et al. \cite{gao21b_interspeech} show that other sampling-based criteria can also perform well, as long as an extra correction step is done, where the intended class posterior probability is recovered from the raw model outputs.
The motivation of this work begins from there.
We notice that by introducing simple modifications to the original IS training criteria, the additional correction step can be omitted and the class posteriors can be directly obtained from the model outputs.
In other words, LMs trained with such criteria are also self-normalized - straightforwardly, we call this type of training: self-normalized importance sampling.
Compared to Gao et al. \cite{gao21b_interspeech}, our method is simpler in the sense that no additional correction is required, and the LMs are directly trained to give the intended class posteriors.
Compared to NCE, which is also a self-normalized criterion, our method is comparable and competitive, which we further show through extensive experiments on both research-oriented and production-oriented ASR tasks.
\section{Related Work}
Neural network LMs are shown to bring consistent improvements over count-based LMs \cite{sundermeyer2015:lstm, sundermeyer2012lstm, irie19:trafolm}.
These neural LMs are then used either in second-pass lattice rescoring or first pass decoding for ASR \cite{6854535, beck2019lstm, beck2020lvcsr, kumar2017lattice, li2021parallelizable}.
To mitigate the problem of having to traverse over the full vocabulary in the softmax normalization, various sampling-based training criteria are proposed and investigated \cite{DBLP:journals/corr/abs-1301-3781, DBLP:conf/nips/MikolovSCCD13, pmlr-vR4-bengio03a, pmlr-v9-gutmann10a, DBLP:conf/icml/MnihT12, goldberger-melamud-2018-self, 7179005}.
These methods mainly deal with how to build up a contrastive loss such that the model is able to tell true target words from noisy ones.
Related but originated from a different motivation, that is, to explicitly encourage normalization during training and hope for self-normalization during testing, variance regularization is also discussed in literature \cite{7533441}.
Recently, in \cite{gao21b_interspeech}, the authors identify the relationships between model optimums under various sampling-based training criteria and the contextual posterior probabilities.
They show that with one additional correction step, the intended posteriors can be recovered.
This work can be thought of as a follow-up and extention of \cite{gao21b_interspeech}, where we propose several simple modifications to the IS training criterion to directly enable self-normalized models, without having to do the additional correction step.
\section{Methodology}
In this section, we first describe several training criteria, including Binary Cross Entropy (BCE), NCE and IS.
Then, we pinpoint the reason why IS is not self-normalized by default.
Afterwards, we move on to discuss three modes of modifications to enable self-normalization of IS.
In the following discussions, we define $c$ as the running index in the target vocabulary size $C$, $x$ as the history for next word prediction and $n$ as the running index of word position $N$.
\label{sec:methodology}
\subsection{Binary Cross Entropy (BCE)}
BCE is a traditional training criterion, which only requires the model outputs $q_\theta(x,c)$ to be bounded in $(0,1)$, and not necessarily a normalized distribution.
\begin{equation}
F_{\text{BCE}} = \frac{1}{N} \sum_{n=1}^N \{ \log q_\theta(x_n,c_n) + \sum_{c \neq c_n} \log ( 1 - q_\theta(x_n,c)) \}
\label{eq: BCE}
\end{equation}
Given enough amount of training pairs $(x_n,c_n)$, it can be shown that the optimum $\hat{q}_\theta(x,c)$ when maximizing $F_{\text{BCE}}$ is the empirical posterior distribution $p(c|x)$, which indicates that the model outputs are self-normalized.
\subsection{Fundamentals of Sampling-Based Training Criteria}
In sampling-based training criteria, for each training pair $(x_n,c_n)$, given the number of samples $K$ and the noisy distribution $D$ for sampling, the summation of some function $f(x_n,\Tilde{c}_k,\mathbb{E}(\mathbb{C}_{K}(c)))$ over $K$ random samples $\Tilde{c}_k$, can approximate the summation of $f$ over all the classes. The expected count $\mathbb{E}(\mathbb{C}_{K}(c))$ is the expectation of the number of the appearance of class $c$ in all samples. For instance, if we do sampling with replacement, the expected count would be $ KD(c)$. The term $\mathbb{E}(\mathbb{C}_{K}(c))$ in function $f$ is usually used to control the variance, i.e. with the increase of $K$, the variance goes down and the approximation is more precise. In the following discussions, to simplify the notation, we define
\begin{equation}
\mathbb{E}(\mathbb{C}_{K}(c)) \equiv E_c
\end{equation}
Omitting the summation over $n$ for simplicity, the approximation of the sampling method is:
\begin{equation}
\begin{aligned}
\sum_{k=1}^K f(x_n,\Tilde{c}_k,&E_{\Tilde{c}_k}) \approx \sum_{c=1}^C E_c \cdot f(x_n,c,E_c)
\end{aligned}
\label{eq: sum_estimator}
\end{equation}
\subsection{Noise Contrastive Estimation (NCE)}
NCE is a popular BCE-style criterion, with the function $f$ in the following form:
\begin{equation}
f(x_n,\Tilde{c}_k,E_{\Tilde{c}_k}) = \log( 1 - \frac{q_\theta(x_n,\Tilde{c}_k)}{q_\theta(x_n,\Tilde{c}_k) + E_{\Tilde{c}_k}})
\end{equation}
More specifically, the NCE criterion can be written as:
\begin{equation}
\begin{aligned}
F_{\text{NCE}} &= \frac{1}{N} \sum_{n=1}^N \{ \log \frac{q_\theta(x_n,c_n)}{q_\theta(x_n,c_n) + E_{c_n}} \\
&\quad + \sum_{k=1}^K \log( 1 - \frac{q_\theta(x_n,\Tilde{c}_k)}{q_\theta(x_n,\Tilde{c}_k) + E_{\Tilde{c}_k}}) \}
\end{aligned}
\end{equation}
Here the first term in the summation over $n$ corresponds to the target class $c_n$ and the second term corresponds to the approximation shown in Equation (\ref{eq: sum_estimator}). With enough amount of training data and samples $K$, the optimum of model output is $\hat{q}_\theta(x,c) \approx p(c|x)$.
\subsection{Importance Sampling (IS)}
IS is another sampling-based BCE-style criterion.
\begin{equation}
F_{\text{IS}} = \frac{1}{N} \sum_{n=1}^N \{\log q_\theta(x_n,c_n)+ \sum_{k=1}^K \frac{\log (1-q_\theta(x_n,\Tilde{c}_k))}{E_{\Tilde{c}_k}} \}
\label{eq: IS}
\end{equation}
Here, with enough amount of training data and samples $K$, the model optimum is not normalized.
\begin{equation}
\hat{q}_\theta(x,c) \approx \frac{p(c|x)}{1+p(c|x)}
\end{equation}
Nonetheless, as described in \cite{gao21b_interspeech}, $\hat{q}_\theta(x,c)$ can be transformed to $p(c|x)$ with an additional correction step.
In IS, the function $f$ has a simpler form, which is the second log term in BCE criterion devided by the expected count.
\begin{equation}
f(x_n,\Tilde{c}_k,E_{\Tilde{c}_k}) = \frac{\log (1-q_\theta(x_n,\Tilde{c}_k))}{E_{\Tilde{c}_k}}
\end{equation}
According to Equation (\ref{eq: sum_estimator}), the summation over samples is actually an approximation of the summation of the second log term in BCE over all classes,
\begin{equation}
\sum_{k=1}^K f(x_n,\Tilde{c}_k,E_{\Tilde{c}_k}) \approx \sum_{c=1}^C \log(1-q_{\theta}(x_n,c))
\label{eq: is_sum}
\end{equation}
which makes IS very similar to the original BCE criterion, except that the sum is over all classes rather than excluding the target class (notice the summation over $c$ in Equation (\ref{eq: is_sum}) and (\ref{eq: BCE})). Thus, our motivation is to modify the original IS criterion to have an approximation of the original BCE criterion, in order to directly obtain $p(c|x)$ as the optimum output, hence achieving self-normalized models.
\subsection{Self-Normalized Importance Sampling}
\subsubsection{Sampling including the Target Class}
\label{sec: Mode1}
When the target class is allowed to be sampled, from the discussion above, the model is not self-normalized. Therefore, an intuitive way to exclude the target class from the summation is to simply substract the target term $\log (1-q_\theta(x_n,c_n))$.
\begin{equation}
\begin{aligned}
F_{\text{Mode1}} &= \frac{1}{N} \sum_{n=1}^N \{\log q_\theta(x_n,c_n)+ \sum_{k=1}^K \frac{\log (1-q_\theta(x_n,\Tilde{c}_k))}{E_{\Tilde{c}_k}} \\
&\quad -\log(1-q(x_n,c_n))\} \\
& \approx F_{\text{BCE}}
\end{aligned}
\end{equation}
With this simple modification, the sampling-based criterion is exactly an approximation of BCE criterion and the model optimum is now at $p(c|x)$. However, if the target class is not sampled, the additional $-\log(1-q(x_n,c_n))$ term can result in a large gradient. When target class $c_n$ does not appear in $\Tilde{c}_k$, we can consider the gradient with respect to $\log q(x_n,c_n)$ of a single sentence $n$.
\begin{equation}
\frac{\partial F_{\text{Mode1}, n}}{\partial \log q(x_n,c_n)} = 1 + \frac{q(x_n,c_n)}{1 - q(x_n,c_n)}
\end{equation}
Suppose that $p(c_n|x_n)$ is a large number (close to one) and $q_\theta(x_n,c_n)$ is close to its optimum $p(c_n|x_n)$, the gradient is large and makes it hard to converge. In the following experimental results, we show that when $K$ is small, which means there is less chance to sample the target class out, this can lead to a bad performance.
\subsubsection{Sampling excluding the Target Class}
Another way to obtain the self-normalized output is to find some proper distribution or function $f$ to directly approximate the summation without target class $c_n$.
\begin{equation}
\sum_k f(x_n,\Tilde{c}_k,E_{\Tilde{c}_k} ) \approx \sum_{c \neq c_n} \log (1 - q_\theta(x_n,c))
\end{equation}
Considering equation (\ref{eq: sum_estimator}), we propose two approaches: letting the expected count of $c_n$ be zero, i.e. $E_{c_n}=0$, or having a function $f_{c_n}$ that is specific for each target class and letting the value of the function be zero when the summation index $c$ is $c_n$, i.e. $ f_{c_n}(x_n,c_n, E_{c_n})=0$.
For the first approach (Mode2), $E_{c_n}=0$ is equivalent to $D(c_n)=0$, i.e. the probability to sample the target class being zero. To this end, we utilize different distributions $D_{c_n}(c)$ with $D_{c_n}(c_n)=0$ for different target classes during training. Compared to IS, Mode2 has the same formula, but different noise distributions for sampling.
\begin{equation}
\begin{aligned}
F_{\text{Mode2}} &= \frac{1}{N} \sum_{n=1}^N \{\log q(x_n,c_n)+ \sum_{k=1}^K \frac{\log (1-q(x_n,\Tilde{c}_k))}{E_{\Tilde{c}_k}} \}\\
\end{aligned}
\end{equation}
Because in this way it is not possible to sample the target class $c_n$, there is no devision-by-zero problem in the criterion.
For the second approach, we simply set $f_{c_n}$ to zero when the target class gets sampled out.
\begin{equation}
F_{\text{Mode3}} = \frac{1}{N}\sum_{n=1}^N \{\log q(x_n,c_n)+ \sum_{k=1}^K f_{c_n}(x_n,\Tilde{c}_k,E_{\Tilde{c}_k})\}
\end{equation}
where
\begin{equation}
f_{c_n}(x_n,\Tilde{c}_k,E_{\Tilde{c}_k}) = \left \{\begin{array}{rl}
0 & \text{if} \quad \Tilde{c}_k = c_n \\
\frac{\log (1-q(x,\Tilde{c}_k))}{E_{\Tilde{c}_k}} & \text{else}
\end{array} \right .
\end{equation}
Compared to Mode2, Mode3 is more efficient since it can use one distribution for all training pairs $(x_n,c_n)$. For Mode2, on the other hand, it is necessary to use different distributions for different target classes. For Mode3, it is safer to do sampling without replacement, since if replacement is allowed, in extreme cases there can be many samples where $f$ evaluates to zero, which adds no contribution to the summation and thus can influence the precision of the approximation.
\section{Experimental Results}
\subsection{Experimental Setup}
We conduct experiments on public datasets including Switchboard and Librispeech. Additionally, we experiment with one in-house dataset: 8kHz US English (EN). The detailed statistics of these three datasets can be found in Table \ref{tab:data_stat}. For EN, 7473M in Table \ref{tab:data_stat} is the total amount of running words used for baseline count-based LM training, while roughly 694M running words are used for neural network training.
We follow the setup in \cite{gao21b_interspeech} in terms of model architectures. For Switchboard and EN, we use LSTM LMs with 2 layers, 2048 hidden dimension and 128 embedding dimension. For Librispeech, we use Transformer \cite{vaswani2017attention} LMs, with 42 layers, 2048 feed-forward hidden dimension, 512 key/query/value dimension, 8 attention heads, and 512 input embedding dimension, following the architecture in \cite{irie19:trafolm}. The acoustic models are hybrid hidden Markov models, with 7 bidirectional LSTM layers and 500 dimension. We refer the reader to \cite{kitza2019cumulative} and \cite{luscher2019rwth} for more details.
We evaluate our models with perplexities (PPLs) and second-pass rescoring word error rates (WERs). For Switchboard and EN, the lattices for rescoring are generated with 4-gram Kneser-Ney LMs \cite{kitza2019cumulative}. For librispeech, the lattice is generated with a LSTM LM \cite{beck2019lstm}. All the model outputs are normalized properly when computing PPLs in order to have a fair comparison. In rescoring, the outputs of models trained with normalized or self-normalized criteria are used as is and no additional renormalization is done. For Importance Sampling, as the output cannot directly be used as scores, we obtain scores by a proper mapping of outputs \cite{gao21b_interspeech}. We report rescoring results on clean and other test sets for Librispeech and on Switchboard (SWB) and CallHome (CH) Hub5'00 test sets for Switchboard. For EN, we use an in-house test set.
For sampling-based criteria, we use log-uniform distributions as noise distributions, following \cite{gao21b_interspeech}. If not mentioned otherwise, the number of samples $K$ is $8000$. We do sampling with replacement, except Mode3. During training, each batch shares the same sampled classes for efficiency \cite{zoph2016simple, jozefowicz2016exploring}, except Mode2. For Mode2, to build up the desired distribution for each training pair $(x_n,c_n)$, we use $C-1$ as the number of labels in log-uniform, and map the sampled indices $\Tilde{c}'_k$ to $\Tilde{c}_k$.
\begin{equation}
\Tilde{c}_k = \left \{\begin{array}{rl}
\Tilde{c}'_k & \text{if} \quad \Tilde{c}'_k < c_n \\
\Tilde{c}'_k + 1 & \text{else}
\end{array} \right .
\end{equation}
The average training time per batch is evaluated on NVIDIA GeForce GTX 1080 Ti GPUs, with a batch size of 32 for Switchboard, 64 for LibriSpeech and EN.
\begin{table}[]
\centering
\caption{Data statistics of the three datasets.}
\begin{tabular}{|l|r|r|}
\hline
\multicolumn{1}{|c|}{corpus} & \multicolumn{1}{c|}{vocabulary size} & \multicolumn{1}{c|}{running words} \\ \hline \hline
Switchboard & 30k & 24M \\ \hline
Librispeech & 200k & 812M \\ \hline
EN & 250k & 7473M \\ \hline
\end{tabular}
\label{tab:data_stat}
\end{table}
\subsection{Main Results}
We first compare three modes of self-normalized IS on Switchboard, employing NCE as the baseline. As shown in Table \ref{tab: swb_compare}, for Mode1, when $K=100$, the performance is much worse than other methods. When $K=8000$, on the other hand, the performance is on par with others, which verifies our statement about the big gradient problem in Section \ref{sec: Mode1}. For both $K=100$ and $K=8000$ in Mode2 and Mode3, the performance is comparable with NCE. Considering that Mode3 is more efficient than Mode2 in terms of the noise distribution, we choose Mode3 to be our main method. For both Mode2 and Mode3, we contribute the relatively small difference in PPLs to the small vocabulary size.
\begin{table}[h!]
\center
\caption{PPLs of NCE and Mode\{1,2,3\} on Switchboard.}
\begin{tabular}{|c|c|c|}
\hline
criterion & K & PPL \\ \hline \hline
\multirow{2}{*}{NCE} & 100 & 54.4 \\ \cline{2-3}
& 8000 & 52.7 \\ \hline
\multirow{2}{*}{Mode1} & 100 & 87.2 \\ \cline{2-3}
& 8000 & 53.0 \\ \hline
\multirow{2}{*}{Mode2} & 100 & 53.9 \\ \cline{2-3}
& 8000 & 53.5 \\ \hline
\multirow{2}{*}{Mode3} & 100 & 54.0 \\ \cline{2-3}
& 8000 & 53.0 \\ \hline
\end{tabular}
\label{tab: swb_compare}
\end{table}
We further investigate the influence of the number of samples $K$ on Switchboard in Mode3. Figure \ref{fig:swb_sweepk} shows that when $K$ is relatively small, the increase of $K$ almost does not influence the training speed, as the computation effort for the sum over $K$ is too small to influence the whole training procedure. When $K$ is large enough, increasing $K$ exponentially will slow down the training speed dramatically. On the other hand, when $K$ is relatively small, the increase of $K$ brings huge improvements in PPL. Because when $K$ is small, the sampling without replacement is similar to the sampling with replacement case, where the variance is proportional to $1/K$ and thus goes down fast with the increase of $K$, leading to a better approximation of the summation and fast decrease of PPL. When $K$ is large enough, the approximation is already good, so with the increase of $K$, the PPL only goes down slowly, and gradually converges. One can find the best trade-off between training speed and PPL according to the specific task. Surely, the trade-off also depends on the model architecture and the vocabulary size, and thus we do not do further experiments to investigate this relationship on other datasets.
\begin{figure}[h!]
\centering
\includegraphics[scale=0.15]{sweep_k_font5.png}
\caption{The influence of number of samples $K$.}
\label{fig:swb_sweepk}
\end{figure}
Table \ref{tab: swbmainresult} shows the comparison of different criteria on Switchboard. In this task, LSTM LMs are interpolated with count-based LM for evaluation. In terms of PPLs, the CE baseline is slightly better than others, because the model outputs are well-normalized during training, which is more consistent with the evaluation on PPL. All other methods have similar performance. In terms of WER, CE baseline is slightly better, but the advantage is not large enough to draw the conclusion that CE outperforms sampling-based methods. In terms of training speed, sampling-based methods are faster than the traditional methods, since for sampling-based methods, the summation is over $K$, which is less than $C$. However, since the vocabulary size of Switchboard is not so large, the training speedup is limited. Compared to the original IS, the output of self-normalized IS (Mode3) can be used as scores directly in rescoring without correction, giving comparable training speed and performance.
\begin{table}[h!]
\center
\caption{WERs of different criteria on Switchboard.}
\begin{tabular}{|c|c|c|c|c|c|}
\hline
\multirow{2}{*}{criterion} & \multirow{2}{*}{\begin{tabular}[c]{@{}c@{}}train time\\ (s/batch)\end{tabular}} & \multirow{2}{*}{PPL} & \multicolumn{3}{c|}{WER} \\ \cline{4-6}
& & & All & SWB & CH \\ \hline \hline
CE & 0.100 & 49.9 & 10.1 & 6.8 & 13.4 \\ \hline
BCE & 0.107 &52.3 & 10.3 & 6.9 & 13.7 \\ \hline
IS & 0.079 & 51.5 & 10.3 & 7.0 & 13.7 \\ \hline
NCE & 0.079 & 51.4 & 10.2 & 6.9 & 13.6 \\ \hline
Mode3 & 0.090 & 51.7 & 10.2 & 6.9 & 13.6 \\ \hline
\end{tabular}
\label{tab: swbmainresult}
\end{table}
Table \ref{tab: librimainresult} exhibits similar results on Librispeech. Since the vocabulary size is much larger than Swichboard, the training speedup gained from sampling is more significant, while the performance are still similar to traditional methods. With different model architectures and larger vocabulary sizes, we show that Mode3 is still competitive compared to NCE.
\begin{table}[h!]
\center
\caption{WERs of different criteria on Librispeech.}
\begin{tabular}{|c|c|c|c|c|}
\hline
\multirow{2}{*}{criterion} & \multirow{2}{*}{\begin{tabular}[c]{@{}c@{}}train time\\ (s/batch)\end{tabular}} & \multirow{2}{*}{PPL} & \multicolumn{2}{c|}{WER} \\ \cline{4-5}
& & & clean & other \\ \hline \hline
CE & 0.302& 57.7 & 2.5 & 5.4 \\ \hline
BCE & 0.358 & 58.5 & 2.5 & 5.4 \\ \hline
IS & 0.206 & 58.4 & 2.6 & 5.5 \\ \hline
NCE & 0.206 & 57.9 & 2.5 & 5.4 \\ \hline
Mode3 & 0.216 &58.3 & 2.5 & 5.4 \\ \hline
\end{tabular}
\label{tab: librimainresult}
\end{table}
Experimental results on EN are presented in Table \ref{tab: in_house_data}. Here, the baseline LM used for first-pass decoding is count-based. With LSTM LM rescoring, we observe a great improvement in WER, which is consistent with the common knowledge about neural network LMs outperforming count-based LMs for ASR.
Besides, we observe that Mode3 has similar performance to NCE for both $K=100$ and $K=8000$ on the production-oriented ASR dataset, validating our claim that the proposed self-normalized IS training criterion is a competitive sampling-based training criterion for neural LMs.
\begin{table}[]
\caption{WERs of different criteria on EN.}
\begin{tabular}{|c|c|c|c|c|c|}
\hline
corpus & criterion & K & \begin{tabular}[c]{@{}c@{}}train speed\\ (s/batch)\end{tabular} & PPL & WER \\ \hline \hline
\multirow{5}{*}{EN} & baseline & - & - & 82.3 &13.7 \\ \cline{2-6}
& \multirow{2}{*}{NCE} & 100 & 0.092 &65.9&13.3 \\ \cline{3-6}
& & 8000 & 0.098 &55.9&13.1 \\ \cline{2-6}
& \multirow{2}{*}{Mode3} & 100 & 0.089 &68.0& 13.3 \\ \cline{3-6}
& & 8000 & 0.114 &55.8&13.1\\ \hline
\end{tabular}
\label{tab: in_house_data}
\end{table}
\section{Conclusion}
We propose self-normalized importance sampling for training neural language models.
Previously, noise contrastive estimation is a popular choice for sampling-based training criterion.
In our recent work, we show that other sampling-based training criteria, including importance sampling, can perform on par with noise contrastive estimation.
However, one caveat in that work is an additional correction step to recover the posterior distribution.
In this work, we completely negate the need for this step by modifying the importance sampling criterion to be self-normalized.
Through extensive experiments on both research-oriented and production-oriented datasets, we obtain competitive perplexities as well as word error rates with our improved method.
\section{Acknowledgements}
This work has received funding from the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme (grant agreement No 694537, project ”SEQCLAS”). The work reflects only the authors’ views and none of the funding parties is responsible for any use that may be made of the information it contains. This work was partially supported by the project HYKIST funded by the German Federal Ministry of Health on the basis of a decision of the German Federal Parliament (Bundestag) under funding ID ZMVI1-2520DAT04A. We thank Christoph Lüscher for providing the LibriSpeech acoustic model and Eugen Beck for the lattices, and Markus Kitza for the SwitchBoard lattices.
\bibliographystyle{IEEEtran}
|
\section{Introduction}\label{introduction}
Let $X \cong \AA^6_\mathbb{R}$ be the affine space parametrizing homogeneous degree $5$ polynomials $F \in \mathbb{R}[x,y]$. Let the variety $X_0 \subset X$ parametrize polynomials with distinct roots, and $X_s \subset X$ polynomials with roots of multiplicity at most two (i.e. stable in the sense of geometric invariant theory).
The principal goal of this paper is to study the moduli space of stable \textit{real binary quintics}
$$
\overline{\mr M}_{\mathbb{R}} : = \textnormal{GL}_2(\mathbb{R}) \setminus X_s(\mathbb{R}) \supset \textnormal{GL}_2(\mathbb{R}) \setminus X_0(\mathbb{R}) =:{\mr M}_\mathbb{R}.
$$
If $P_s \subset \mathbb{P}^1(\mathbb{C})^5$ is the set $5$-tuples $(x_1, \dotsc, x_5)$ such that no three $x_i \in \mathbb{P}^1(\mathbb{C})$ coincide (c.f. \cite{MR0437531}), and $P_0 \subset P_s$ the subset of $5$-tuples all whose coordinates are distinct, then
$$
{\mr M}_\mathbb{R} \cong \textnormal{PGL}_2(\mathbb{R}) \setminus (P_0/\mf S_5)(\mathbb{R}) \;\;\; \;\;\; \text{ and } \;\;\; \;\;\; \overline{\mr M}_\mathbb{R} \cong \textnormal{PGL}_2(\mathbb{R}) \setminus (P_s/\mf S_5)(\mathbb{R}).
$$
In other words, ${\mr M}_\mathbb{R}$ is the space of subsets $S \subset \mathbb{P}^1(\mathbb{C})$ of cardinality $|S| = 5$ stable by complex conjugation modulo real projective transformations, and in $\overline{\mr M}_\mathbb{R}$ one or two pairs of points are allowed to collapse. For $i = 0,1,2$, we define $\mr M_{i}$ to be the connected component of ${\mr M}_\mathbb{R}$ parametrizing $5$-tuples in $\mathbb{P}^1(\mathbb{C})$ with $2i$ complex and $5 - 2i$ real points.
\\
\\
There is a natural period map that defines an isomorphism between $\textnormal{GL}_2(\mathbb{C}) \setminus X_s(\mathbb{C})$
and a certain arithmetic ball quotient $P\Gamma \setminus \mathbb{C} H^2$ \cite{shimuratranscendental}, \cite{DeligneMostow}. Moreover, one can prove that strictly stable quintics correspond to points in a hyperplane arrangement $\mr H \subset \mathbb{C} H^2$ (Proposition \ref{prop:stableperiodshyperplane}). Investigating the equivariance of the period map with respect to a suitable set of anti-holomorphic involutions $\alpha_i : \mathbb{C} H^2 \to \mathbb{C} H^2$, we obtain the following real analogue:
\begin{theorem} \label{th:theorem01}
For each $i \in \{0,1,2\}$, the period map induces an isomorphism of real analytic orbifolds $\mr M_i \cong P \Gamma_i \setminus \left(\mathbb{R} H^2 - \mr H_i \right)$. Here $\mathbb{R} H^2$ is the real hyperbolic plane, $\mr H_i$ a union of geodesic subspaces in $ \mathbb{R} H^2$ and $P\Gamma_i$ an arithmetic lattice in $\textnormal{PO}(2,1)$. Moreover, the $P\Gamma_i$ are projective orthogonal groups attached to explicit quadratic forms over $\mathbb{Z}[\zeta_5 + \zeta_5^{-1}]$, see (\ref{eq:explicitquadraticforms}).
\end{theorem}
In particular, Theorem \ref{th:theorem01} endows each component $\mr M_i$ with a natural hyperbolic metric. Since one can deform the topological type of a $\textnormal{Gal}(\mathbb{C}/\mathbb{R})$-stable five-element subset of $\mathbb{P}^1(\mathbb{C})$ by allowing two points to collide, the compactification $\overline{\mr{M}}_\mathbb{R}\supset \mr M_{\mathbb{R}}$ is connected.
One may thus wonder whether the metrics on the components $\mr M_i$ extend to a metric on the whole of $\overline{\mr M}_\mathbb{R}$, and if so, what the resulting space looks like at the boundary. Our main result is the following:
\begin{theorem} \label{th:theorem02}
There exists a complete hyperbolic metric on $\overline{\mr{M}}_\mathbb{R}$ that restricts to the metrics on $\mr M_i$ induced by Theorem \ref{th:theorem01}. With respect to it, $\overline{\mr{M}}_\mathbb{R}$ is isometric to the hyperbolic triangle of angles $\pi/3, \pi /5, \pi/10$. Thus, if $P\Gamma_\mathbb{R} = \langle \alpha_1, \alpha_2, \alpha_3 | \alpha_i^2 = (\alpha_1\alpha_2)^3 = (\alpha_1\alpha_3)^5 = (\alpha_2\alpha_3)^{10} = 1 \rangle$, there is a homeomorphism $\overline{\mr{M}}_\mathbb{R} \cong P\Gamma_\mathbb{R} \setminus \mathbb{R} H^2$ extending the isomorphisms in Theorem \ref{th:theorem01}.
\end{theorem}
\blfootnote{\textcolor{blue}{$^\ast$}\footnotesize{\'Ecole normale sup\'erieure, 45 rue d'Ulm, Office T17, 75230 Paris, \href{mailto:<EMAIL>@ens.fr}.}}
\blfootnote{\textcolor{blue}{$^\ast$}\textit{Date:} \today $\;
- This project has received funding from the European Union's Horizon 2020 research and innovation programme under the Marie Sk\l{}odowska-Curie grant agreement N\textsuperscript{\underline{o}} 754362.$\;$\img{EU}}
\restoregeometry
\pagestyle{plain}
\newgeometry{margin=0.5in, top=0.01in, bottom=1in, footskip=0.5in
\hspace*{-3cm}\includegraphics[scale=0.20]{triangle15transparent.png}
\textit{Figure 1: $\overline{\mr M}_\mathbb{R}$ as the hyperbolic triangle $\Delta_{3,5,10} \subset \mathbb{R} H^2$. Here $\lambda = \zeta_5 + \zeta_5^{-1}$ and $\omega = \zeta_3$.}
\begin{remark} \label{remark:takeuchi}
The lattice $P\Gamma_\mathbb{R} \subset \textnormal{PO}(2,1)$ is \textit{non-arithmetic}, as follows from \cite{takeuchi}.
\end{remark}
Equivariant period maps arise often in real algebraic geometry as a method to obtain real uniformization of the connected components of the moduli space of smooth varieties. For instance, this works for abelian varieties \cite{grossharris}, algebraic curves \cite{seppalasilhol2}, K3 surfaces \cite{Nikulin1980} and quartic curves \cite{heckman2016hyperbolic}. Only recently, Allcock, Carlson and Toledo have shown that in the cases of cubic surfaces \cite{realACTsurfaces} and binary sextics \cite{realACTnonarithmetic}, \cite{realACTbinarysextics}, the real ball quotient components can be glued along the hyperplane arrangement in order to uniformize the moduli space of real stable varieties. Binary quintics provide the first new example of this phenomenon.
\\
\\
To prove Theorem \ref{th:theorem02}, we construct a method of gluing together real arithmetic ball quotients, generalizing the work of Allcock, Carlson and Toledo cited above.
The quotient spaces that we glue are real loci of various real structures on the complex orbifold ball quotient attached to a hermitian lattice of hyperbolic signature over the ring of integers of CM field. Let us outline this construction.
\\
\\
Let $K$ be a CM field of degree $2g$ over $\mathbb{Q}$ with ring of integers $\mathcal{O}_K$, and let $\Lambda$ be a finite free $\mathcal{O}_K$-module equipped with a hermitian form $h: \Lambda \times \Lambda \to \mathcal{O}_K$. Suppose that $h$ has signature $(n,1)$ with respect to a fixed embedding $\tau: K \to \mathbb{C}$ and is definite for the other infinite places of $K$. Let $\mathbb{C} H^n$ be the space of negative lines in $\Lambda \otimes_{\mathcal{O}_K, \tau} \mathbb{C}$ and $P\Gamma = \textnormal{Aut}(\Lambda, h) / \mu_K$ where $\mu_K \subset \mathcal{O}_K^\ast$ is the group of finite units $\zeta \in \mathcal{O}_K^\ast$. Define $P\mr A$ to be the quotient of the set of anti-unitary involutions $\alpha: \Lambda \to \Lambda$ by $\mu_K$.
\\
\\
Consider the hyperplane arrangement $\mr H = \cup_{h(r,r) = 1} \langle r_\mathbb{C} \rangle ^\perp \subset \mathbb{C} H^n$ and assume $(\textcolor{blue}{\ast})$: different hyperplanes intersect orthogonally or not at all, c.f. \cite{orthogonalarrangements}.
For example, this holds
under a condition on the CM field $K$ (see Theorem \ref{th:conditionsimplyhypothesis})
satisfied when $K$ is cyclotomic or quadratic (see Lemma \ref{lemma:discr}). (In fact, condition $(\textcolor{blue}{\ast})$ is \textit{always} satisfied if one is willing to adapt the definition of $\mr H$, see Remark \ref{remark:avoidcondition}.)
\restoregeometry
We claim that there is a canonical way to glue the different copies $\mathbb{R} H^n_\alpha: = \left(\mathbb{C} H^n\right)^\alpha \subset \mathbb{C} H^n$ of the real hyperbolic space $\mathbb{R} H^n$ along the hyperplane arrangement $\mr H$. See Remark \ref{rem:gluing} for the precise formulation of the equivalence relation. This gives a topological space which we denote by $Y$, acted upon by $P\Gamma$. Define $P\Gamma_\alpha \subset P\Gamma$ to be stabilizer of $\mathbb{R} H^n_\alpha$. The result is as follows:
\begin{theorem} \label{th:theorem03}
The topological space $P\Gamma \setminus Y$ admits a metric that makes it a complete path metric space such that $P\Gamma \setminus Y \to P\Gamma \setminus \mathbb{C} H^n$ is a local isometry.
This metric induces a real hyperbolic orbifold structure on $P\Gamma \setminus Y$ such that $ \coprod_{\alpha \in P\Gamma \setminus P\mr A} [P\Gamma_\alpha \setminus \left(\mathbb{R} H^n_\alpha - \mr H \right)] \subset P\Gamma \setminus Y$ is an open suborbifold. Moreover, for each connected component $C \subset P\Gamma \setminus Y$ there exists a lattice $P\Gamma_C \subset \textnormal{PO}(n,1)$ and an isomorphism of real hyperbolic orbifolds $C \cong [P\Gamma_C \setminus \mathbb{R} H^n]$.
\end{theorem}
For some moduli stacks of hypersurfaces $\ca M$ one can then apply Theorem \ref{th:theorem03} to the hermitian lattice $\Lambda$ that arises as the cohomology of the cover of projective space ramified along a member of the moduli space. Let $\ca M_s$ be the stack of GIT stable hypersurfaces. If the discriminant $\Delta = \ca M_s(\mathbb{C}) - \ca M(\mathbb{C})$ is a normal crossings divisor and the period map induces an isomorphism of analytic spaces $\ca M_s(\mathbb{C}) \cong P\Gamma \setminus \mathbb{C} H^n$ identifying $\Delta$ with $P\Gamma \setminus \mr H$, then there is a real period homeomorphism $\ca M_s(\mathbb{R}) \cong P\Gamma \setminus Y$. For cubic surfaces and binary sextics, this is the content of \cite{realACTsurfaces}, \cite{realACTnonarithmetic} and for binary quintics, this yields Theorem \ref{th:theorem02} (see Theorem \ref{th:realstableperiod}).
\begin{remark}
The lattice $P\Gamma_C$ attached to a component $C \subset P\Gamma \setminus Y$ can be non-arithmetic. Indeed, such is the case for $K = \mathbb{Q}(\zeta_5)$ and $h = \textnormal{diag}(1, 1, \frac{1 - \sqrt{5}}{2})$ by Remark \ref{remark:takeuchi} and Theorem \ref{th:calculatemonodromyshimura}, and for $K = \mathbb{Q}(\zeta_3)$ and $h = \textnormal{diag}(1, \dotsc, 1, -1)$ for $n = 3$ \cite{realACTnonarithmetic} and $n = 4$ \cite{realACTsurfaces}.
\end{remark}
\begin{remark} Our gluing construction relies on condition $(\textcolor{blue}{\ast})$, saying that $\mr H \subset \mathbb{C} H^n$ is an \textit{orthogonal arrangement} in the sense of \cite{orthogonalarrangements}. Such arrangements are interesting in their own right. Indeed,
if $n > 1$ then the orbifold fundamental $\pi_1^\textnormal{orb} \left( P\Gamma \setminus \left( \mathbb{C} H^n - \mr H \right) \right)$ is not a lattice in any Lie group with finitely many connected components [\textit{loc. cit.}, Theorem 1.2]. In particular, neither $\pi_1 \left( P_0 / \mf S_5 \right)$ nor $\pi_1^\textnormal{orb}\left( \mr M_\mathbb{C} \right)$ is a lattice in any Lie group with finitely many connected components. The hyperplane arrangement $\mr H \subset \mathbb{C} H^n$ is an orthogonal arrangement under the following condition $(\textcolor{blue}{\ast\ast})$, satisfied by quadratic and cyclotomic CM fields: the different ideal $\mf D_K \subset \mathcal{O}_K$ is generated by an element $\eta \in \mathcal{O}_K - \mathcal{O}_F$ such that $\eta^2 \in \mathcal{O}_F$. See Section \ref{unitaryshimura}
\end{remark}
\begin{remark}
In fact, there is always a canonical orthogonal arrangement $\ca H \subset \mathbb{C} H^n$ attached to $h$ in such a way that $\ca H = \mr H$ when condition $(\textcolor{blue}{\ast\ast})$ holds, see Remark \ref{remark:avoidcondition}. Moreover, one can glue the different copies $\bb R H^n_\alpha$ of real hyperbolic $n$-space along the hyperplane arrangement $\ca H$
obtaining a complete hyperbolic orbifold as in Theorem \ref{th:theorem03}, but we will not prove this.
\end{remark}
\begin{remark}
The gluing construction of Gromov and Piatetski-Shapiro \cite{gromovshapiro} seems close to our construction. However, there are differences: see \cite[Section 13, Remark (1)]{realACTsurfaces}. In a different way, it might be interesting to compare Shimura's study of real points $V(\mathbb{R})$ of an arithmetic quotient $V(\mathbb{C})$ of a bounded symmetric domain, see \cite{shimurarealpoints}.
\end{remark}
\subsection{Acknowledgements}
I thank my thesis advisor Olivier Benoist for his great guidance and support. Moreover, I thank Romain Branchereau, Samuel Bronstein and Nicolas Tholozan for answering questions on hyperbolic geometry. I thank Frans Oort for a useful discussion, and Nicolas Bergeron for pointing me to Takeuchi's paper on hyperbolic triangle groups which led to Remark \ref{remark:takeuchi}. Finally, I would like to stress that the gluing construction, taking up the first part of this paper, is inspired by and based on work of Allcock-Carlson-Toledo \cite{realACTnonarithmetic}, \cite{realACTbinarysextics}, \cite{realACTsurfaces}.
\newpage
\section{Outline of the paper}
This paper is organized as follows. We start with the gluing construction in Section \ref{realshimura}, which explains how to obtain the complete real hyperbolic orbifold $P\Gamma \setminus Y$ from a free hermitian lattice $\Lambda$ over the ring of integers over a CM field $K$. Then, in Section \ref{unitaryshimura}, we prove that the complex ball quotient $P\Gamma \setminus \mathbb{C} H^n$ is a moduli space for abelian varieties, and that the abelian varieties in $P\Gamma \setminus \mr H$ contain a CM abelian subvariety. In Section \ref{complexperiodmaps}, we define the period for complex binary quintics, prove its compatibility with \cite{DeligneMostow} and show that the monodromy group is the group $P\Gamma$ attached to the hermitian lattice of item (5) in the list of \cite{shimuratranscendental}. In Section \ref{realperiodmaps} we define the real period map for stable binary quintics and prove that it is a homeomorphism.
In Section \ref{modulifivepoints}, we calculate the orbifold structure of $\overline{\mr M}_\mathbb{R}$ and finish the proof of
Theorem \ref{th:theorem02}.
\section{Gluing Real Hyperbolic Quotient Spaces} \label{realshimura}
\subfile{Xgluing} \label{gluing}
\section{Unitary Shimura Varieties} \label{unitaryshimura}
\subfile{unitaryshimura}
\section{The Moduli Space of Complex Binary Quintics as a Ball Quotient} \label{complexperiodmaps}
\subfile{Vcomplexperiodmaps}
\section{The Period Map for Real Binary Quintics} \label{realperiodmaps}
\subfile{realperiodmaps}
\section{The Moduli Space of Real Binary Quintics as Hyperbolic Triangle} \label{modulifivepoints}
The goal of this section is to prove Theorem \ref{th:triangle}, which states that the moduli space $\overline{\mr M}_\mathbb{R} = G(\mathbb{R}) \setminus X_s(\mathbb{R})$ of stable binary quintics equipped with the metric of Corollary \ref{cor:theorem2} is isometric to the triangle $\Delta_{3,5,10} \subset H^2$ of angles $\pi/3, \pi/5$ and $\pi/10$ in the real hyperbolic plane $\mathbb{R} H^2$.
\subsection{Classification of the stabilizers groups of $\overline{\mr M}_{\mathbb{R}}$} \label{sec:1}
The aim of this section is then to describe all possible elements $x = [\alpha_1, \dotsc, \alpha_5] \in (P_s/\mf S_5)(\mathbb{R})$ whose stabilizer $H_x:=\textnormal{PGL}_2(\mathbb{R})_x$ is non-trivial, and also to calculate $H_x$ in these cases:
\begin{proposition} \label{prop:allstabilizergroups}
All stabilizer groups $H_x \subset \textnormal{PGL}_2(\mathbb{R})$ for points $x \in (P_s/ \mf S_5)(\mathbb{R})$ are among $\mathbb{Z}/2, D_3, D_5$. For $n \in \{3,5\}$, there is a unique $\textnormal{PGL}_2(\mathbb{R})$-orbit in $(P_s/ \mf S_5)(\mathbb{R})$ of points $x$ with stabilizer $D_n$.
\end{proposition}
\begin{proof}
We have an injection
$
(P_s/\mf S_5)(\mathbb{R}) \hookrightarrow P_s/\mf S_5$ which is equivariant for the embedding $\textnormal{PGL}_2(\mathbb{R}) \hookrightarrow \textnormal{PGL}_2(\mathbb{C})$. In particular, $H_x \subset \textnormal{PGL}_2(\mathbb{C})_x$ for every $x \in (P_s/\mf S_5)(\mathbb{R})$. The groups $\textnormal{PGL}_2(\mathbb{C})_x$ for equivalence classes of distinct points $x \in P_0/\mf S_5$ are calculated in \cite[Theorem 22]{wu2019moduli}, and such a group is isomorphic to $\mathbb{Z}/2, D_3, \mathbb{Z}/4$ or $D_5$. None of these have subgroups isomorphic to $D_2 = \mathbb{Z}/2 \rtimes \mathbb{Z}/2$ or $D_4 = \mathbb{Z}/2 \rtimes \mathbb{Z}/4$.
Define an involution
\begin{equation} \label{eq:nu}
\nu: = (z \mapsto 1/z) \in \textnormal{PGL}_2(\mathbb{R}).
\end{equation}
\begin{comment}
\begin{lemma}
Let $x = (x_1, \dotsc, x_5) \in (P_s/\mf S_5)(\mathbb{R})$. Let $\tau \in \textnormal{PGL}_2(\mathbb{R})$ such that $\tau(x) = x$. Then $\tau^n = \textnormal{id}$ for some $n \leq 5$.
\end{lemma}
\begin{proof}%
We may assume that $\tau \neq \textnormal{id}$. First suppose that $x_i \neq x_j$ for every $i\neq j$. Then either $\tau(x_i) \neq x_i$ for every $i$, in which case $\tau^5(x_i) = x_i$ for every $i$ hence $\tau^5 = \textnormal{id}$; or $\tau$ permutes $4$ elements freely, in which case $\tau^4 = \textnormal{id}$, or $\tau$ permutes three elements in which case $\tau^3 = \textnormal{id}$; or $\tau$ permutes two $2$-cycles and fixes one element, in which case $\tau^2 = \textnormal{id}$. Now if $x_1 = x_2$ and $x_j \neq x_i$ for $i \neq j > 2$, then $\tau(x_1) = x_1$. Either $\tau$ permutes $(x_3, x_4, x_5)$ freely in which case $\tau^3 = \textnormal{id}$, or $\tau(x_3) = x_4$ and $\tau(x_5) = x_5$. In the latter case, $\tau^2 = \textnormal{id}$ by Lemma \ref{lemma:involution}.
If $x_1 = x_2$ and $x_3 = x_4$ then $\tau(x_5) = x_5$. Consequently, $\tau(x_1) = x_3$ (otherwise $\tau = \textnormal{id}$) hence $\tau^2 = \textnormal{id}$.
\end{proof}
\end{comment}
\begin{lemma} \label{lemma:involution}
Consider an element $\tau \in \textnormal{PGL}_2(\mathbb{R})$ and three distinct elements $x,y,z \in \mathbb{P}^1(\mathbb{C})$ such that $S = \{x,y,z\}$ is stabilized by complex conjugation, and such that $\tau(x) = x$, $\tau(y) = z$ and $\tau(z) = y$. Then there is a transformation $g \in \textnormal{PGL}_2(\mathbb{R})$ that maps $S$ to either $\{-1, 0, \infty\}$ or $\{-1, i, -i\}$, and that satisfies $g \tau g^{-1} = \nu = (z \mapsto 1/z) \in \textnormal{PGL}_2(\mathbb{R})$. In particular, $\tau^2 = \textnormal{id}$.
\end{lemma}
\begin{proof}
This follows readily from the fact that two transformations $g,h \in \textnormal{PGL}_2(\mathbb{C})$ that satisfy $g(x_i) = h(x_i)$ for three different points $x_1, x_2, x_3 \in \mathbb{P}^1(\mathbb{C})$ are necessarily equal.
\end{proof}
\begin{lemma}
There is no $x \in (P_s/\mf S_5)(\mathbb{R})$ stabilized by an element $\phi \in \textnormal{PGL}_2(\mathbb{R})$ of order $4$.
\end{lemma}
\begin{proof}
By \cite[Theorem 4.2]{beauvillePGL2}, all subgroups $G \subset \textnormal{PGL}_2(\mathbb{R})$ that are isomorphic to $\mathbb{Z}/4$ are conjugate to each other. Since the transformation $I: z \mapsto (z-1)/(z+1)$ is of order $4$, it gives a representative $G_I = \langle I \rangle$ of this conjugacy class. Hence, assuming there exists $x$ and $\phi$ as in the lemma,
possibly after replacing $x$ by $gx$ for some $g \in \textnormal{PGL}_2(\mathbb{R})$, we may and do assume that $\phi = I$. On the other hand, it is easily shown that $I$ cannot fix any $x \in (P_s/\mf S_5)(\mathbb{R})$.
\end{proof}
Define
$$
\rho \in \textnormal{PGL}_2(\mathbb{R}), \;\;\; \rho(z) = \frac{-1}{z+1}.
$$
\begin{lemma} \label{lemma:D3}
Let $x = (x_1, \dotsc, x_5) \in (P_s/\mf S_5)(\mathbb{R})$. Suppose $\phi(x) = x$ for an element $\phi \in \textnormal{PGL}_2(\mathbb{R})$ of order $3$. Then there is a transformation $g \in \textnormal{PGL}_2(\mathbb{R})$ mapping $x$ to $z = (-1, \infty, 0, \omega, \omega^2)$ with $\omega$ a primitive third root of unity, and the stabilizer of $x$ to the subgroup of $\textnormal{PGL}_2(\mathbb{R})$ generated by $\rho$ and $\nu$. In particular, the stabilizer $H_x$ is isomorphic to $D_3$.
\end{lemma}
\begin{proof}
It follows from Lemma \ref{lemma:involution} that there must be three elements $x_1, x_2, x_3$ which form an orbit under $\phi$. Since complex conjugation preserves this orbit, one element in it is real; since $g$ is defined over $\mathbb{R}$, they are all real. Let $g \in \textnormal{PGL}_2(\mathbb{R})$ such that $g(x_1) = -1$, $g(x_2) = \infty$ and $g(x_3) = 0$. Define $\kappa = g \phi g^{-1}$. Then $\kappa^3 = \textnormal{id}$, and $\kappa$ preserves $\{-1, \infty, 0\}$ and sends $-1$ to $\infty$ and $\infty$ to $0$. Consequently, $\kappa(0) = -1$, and it follows that $\kappa = \rho$. Hence $x$ is equivalent to an element of the form
$z = (-1, \infty, 0, \alpha, \beta)$. Moreover, $\beta = \bar \alpha$ and $\alpha^2 + \alpha + 1 = 0$.
\end{proof}
Recall that $\zeta_5 = e^{2i\pi/5} \in \mathbb{P}^1(\mathbb{C})$ and define
$$
\lambda = \zeta_5 + \zeta_5^{-1} \in \mathbb{R}, \;\;\;\white \gamma(z) = \frac{(\lambda + 1)z - 1}{z + 1} \in \textnormal{PGL}_2(\mathbb{R})
$$
\begin{lemma}\label{lemma:D5}
Let $x = (x_1, \dotsc, x_5) \in (P_s/\mf S_5)(\mathbb{R})$. Suppose $x$ is stabilized by a subgroup of $\textnormal{PGL}_2(\mathbb{R})$ of order $5$. Then there is a transformation $g \in \textnormal{PGL}_2(\mathbb{R})$ mapping $x$ to $z = (0, -1, \infty, \lambda+1, \lambda)$ and identifying the stabilizer of $x$ with the subgroup of $\textnormal{PGL}_2(\mathbb{R})$ generated by $\gamma$ and $\nu$. In particular, the stabilizer $H_x$ of $x$ is isomorphic to $D_5$.
\end{lemma}
\begin{proof}
Let $\phi\in H_x$ be an element of order $5$.
Using Lemma \ref{lemma:involution} one shows that $x$ must be smooth, i.e. all $x_i$ are distinct, and $x_i = \phi^{i-1}(x_1)$. Since there is one real $x_i$ and $\phi$ is defined over $\mathbb{R}$, all $x_i$ are real. Now note that $z = (0, -1, \infty, \lambda + 1, \lambda)$ is the orbit of $0$ under $\gamma: z \mapsto ((\lambda+1)z - 1)/(z+1)$. The reflection $\nu: z \mapsto 1/z$ preserves $z$ as well: if $\zeta = \zeta_5$ then $\lambda = \zeta + \zeta^{-1}$ hence $\lambda + 1 = - (\zeta^2 + \zeta^{-2}) = - \lambda^2 + 2$, so that $\lambda(\lambda + 1) = 1$. So we have $H_z \cong D_5$. But then, by \cite[Theorem 22]{wu2019moduli}, the point $z$ with its stabilizer $H_z$ must be equivalent under $\textnormal{PGL}_2(\mathbb{C})$ to the point $(1, \zeta, \zeta^2, \zeta^3, \zeta^4)$ with its stabilizer $\langle x \mapsto \zeta x, x \mapsto 1/x \rangle$.
Consequently, there exists an element $g \in \textnormal{PGL}_2(\mathbb{C})$ such that $g(x_1) = 0$, $g(x_2) = -1$, $g(x_3) = \infty$, $g(x_4) = \lambda + 1$ and $g(x_5) = \lambda$, and such that $gH_xg^{-1} = H_z$. Since all $x_i$ and $z_i \in z$ are real, we see that $\bar g(x_i) = z_i$ for every $i$, hence $g$ and $\bar g$ coincide on more than $2$ points, hence $g = \bar g \in \textnormal{PGL}_2(\mathbb{R})$.
\end{proof}
This finishes the proof of Proposition \ref{prop:allstabilizergroups}.
\end{proof}
\subsection{Points of $(P_s/\mf S_5)(\mathbb{R})$ with stabilizer $\mathbb{Z}/2 \hookrightarrow \textnormal{PGL}_2(\mathbb{R})$} \label{sec:2}
The goal of this section is to prove that there are no cone points in the orbifold $\textnormal{PGL}_2(\mathbb{R}) \setminus (P_s/ \mf S_5)(\mathbb{R})$, i.e. orbifold points whose stabilizer group is $\mathbb{Z}/n$ for some $n$ acting on the orbifold chart by rotations. By Proposition \ref{prop:allstabilizergroups}, this fact will follow from the following:
\begin{proposition} \label{prop:zmod2stabilizer}
Let $x=(x_1, \dotsc, x_5) \in (P_s/\mf S_5)(\mathbb{R})$ be such that its stabilizer group $H_x = \langle \tau \rangle $ has order $2$. Then there is a $H_x$-stable open neighborhood $U \subset (P_s/\mf S_5)(\mathbb{R})$ of $x$ such that $H_x \setminus U \to \overline{\mr M}_\mathbb{R}$ is injective, and a homeomorphism $\phi: (U,x) \to (B,0)$ for $0 \in B \subset \mathbb{R}^2$ an open ball, such that $\phi$ identifies $H_x$ with $\mathbb{Z}/2$ acting on $B$ by reflections in a line through $0$.
\end{proposition}
\begin{proof}
Using Lemma \ref{lemma:involution}, it is a routine exercise to check that the only possibilities for the element $x=(x_1, \dotsc, x_5) \in (P_s/\mf S_5)(\mathbb{R})$ are as follows:
1. $
x = (-1, 0, \infty, \beta, \beta^{-1}) \textnormal{ with } \beta \in \AA^1(\mathbb{R})\setminus \{-1,0,1\} \textnormal{ or } \beta \in \bb S^1\setminus \{-1,1\} \subset \AA^1(\mathbb{C}), H_x = \langle \nu \rangle.
$
2.
$
x = (-1, i, -i, \beta, \beta^{-1}) \textnormal{ with } \beta \in \AA^1(\mathbb{R})\setminus \{-1,1\}\textnormal{ or } \beta \in \bb S^1\setminus\{-1,1,i,-i\} \subset \AA^1(\mathbb{C}),$ $H_x = \langle \nu \rangle$.
3.
$
x = (-1, -1, \beta, 0, \infty) \textnormal{ with } \beta \in \AA^1(\mathbb{R})\setminus \{-1,0,1\}, H_x = \langle \nu \rangle.
$
4.
$
x = (-1, -1, \beta, i, -i) \textnormal{ with } \beta \in \AA^1(\mathbb{R})\setminus \{-1,1\}, H_x = \langle \nu \rangle.
$
5.
$
x = (0,0, \infty, \infty, -1), H_x = \langle \nu \rangle$.
6.
$
x = (-1, i, i, - i, -i), H_x = \langle \nu \rangle.
$
\end{proof}
\subsection{Comparing the orbifold structures} \label{sec:triangle} \label{sec:3}
The aim of this section is to prove the following:
\begin{proposition} \label{prop:conesreflectors}
Let $\overline{\mr M}_\mathbb{R}$ be the moduli space of real stable binary quintics. The hyperbolic orbifold structure of $\overline{\mr M}_\mathbb{R}$, induced by the homeomorphism of Corollary \ref{cor:theorem2}, has no cone points and three corner reflectors, whose orders are $\pi/3, \pi/5$ and $\pi/10$.
\end{proposition}
\begin{proof}
There are two topological orbifold structures on $\overline{\mr M}_\mathbb{R}$: one via its definition as $[G(\mathbb{R}) \setminus X_s(\mathbb{R})]$ and one via the isomorphism $\overline{\mr M}_{\mathbb{R}} \cong P\Gamma \setminus Y$ of Theorem \ref{th:realstableperiod} and the orbifold structure on $P\Gamma \setminus Y$ given by Theorem \ref{theorem1}. We claim that these topological orbifold structures differ only at the moduli point $(\infty, i, i, -i, -i)$.
Indeed, this can be deduced from Proposition \ref{localmodel}. The notation of that proposition was as follows: for $f \in Y \cong G(\mathbb{R}) \setminus \ca F_s(\mathbb{R})$ (see Theorem \ref{th:realstableperiod}) the group $A_f \subset P\Gamma$ is the stabilizer of $f \in K$; if $x \in \ca F_s(\mathbb{R})$ represents $f$ and if $F = [x] \in X_s(\mathbb{R})$ has $k = 2a + b$ nodes, then the image $y \in \mathbb{C} H^2$ lies on $k$ orthogonal hyperplanes $H_r$, with $a$ pairs of complex conjugate hyperplanes and $b$ real hyperplanes.
If $F$ has no nodes ($k = 0$), then $G(\textbf r)$ is trivial by Proposition \ref{localmodel}\textcolor{blue}{.1} and $G_F = A_f = \Gamma_f$. If $F$ has only real nodes, then $B_f = G(\textbf r)$ hence $G_F = A_f/G(\textbf r) = A_f / B_f = \Gamma_f$.
Now suppose that $a = 1$ and $b = 0$: the equation $F$ defines a pair of complex conjugate nodes. In other words, the zero set of $F$ defines a $5$-tuple $x= (\alpha_1, \dotsc, \alpha_5) \in \mathbb{P}^1(\mathbb{C})$, well-defined up to the $\textnormal{PGL}_2(\mathbb{R}) \times \mf S_5$ action on $\mathbb{P}^1$, where $\alpha_1 \in \mathbb{P}^1(\mathbb{R})$ and $\alpha_3 = \bar \alpha_2 = \alpha_5 = \bar \alpha_4 \in \mathbb{P}^1(\mathbb{C}) \setminus \mathbb{P}^1(\mathbb{R})$. So we may write $x = (\rho, \alpha, \bar \alpha, \alpha, \bar \alpha)$ with $\rho \in \mathbb{P}^1(\mathbb{R})$ and $\alpha \in \mathbb{P}^1(\mathbb{C}) \setminus \mathbb{P}^1(\mathbb{R})$. Then there is a unique $T \in \textnormal{PGL}_2(\mathbb{R})$ such that $T(\rho) = \infty$ and $T(\alpha) = i$. But this gives $T(x) = (\infty, i , -i, i, -i)$ hence $F$ is unique up to isomorphism. As for the stabilizer $G_F = A_f / G(\textbf r)$, we have $G(\textbf r) \cong (\mathbb{Z}/10)^2$. Since there are no real nodes, $B_f$ is trivial. By Proposition \ref{localmodel}\textcolor{blue}{.3}, $K_f$ is the union of $10$ copies of $\bb B^2(\mathbb{R})$ meeting along a common point $\mathbb{B}^0(\mathbb{R})$. In fact, in the local coordinates $(t_1, t_2)$ around $f$, the $\alpha_j: \mathbb{B}^2(\mathbb{C}) \to \mathbb{B}^2(\mathbb{C})$ are defined by $(t_1, t_2) \mapsto (\bar t_2 \zeta^j , \bar t_1 \zeta^j)$, for $j \in \mathbb{Z}/10$, and so the fixed points sets are given by the equations $\mathbb{R} H^2_j = \{t_2 = \bar t_1 \zeta^j\}\subset \mathbb{B}^2(\mathbb{C})$, $j \in \mathbb{Z}/10$. Notice that the subgroup $G \subset G(\textbf r)$ that stabilizes $\mathbb{R} H^2_j$ is the cyclic group of order $10$ generated by the transformations $(t_1, t_2) \mapsto (\zeta t_1, \zeta^{-1} t_2)$.
On the other hand, there is only one non-trivial transformation $T \in \textnormal{PGL}_2(\mathbb{R})$ that fixes $\infty$ and sends the subset $\{i, -i\} \subset \mathbb{P}^1(\mathbb{C})$ to itself, and $T$ is of order $2$. Hence $G_F = \mathbb{Z}/2$ so that we have an exact sequence $
0 \to \mathbb{Z}/10 \to \Gamma_f \to \mathbb{Z}/2 \to 0$ and this splits since $G_F$ is a subgroup of $\Gamma_f$.
We are done by Propositions \ref{prop:allstabilizergroups} and \ref{prop:zmod2stabilizer}.
\end{proof}
\subsection{The real moduli space of $5$ unordered points on $\mathbb{P}^1$ is a hyperbolic triangle}
\label{sec:triangle2}
In the sequal, we shall consider the topological space $\overline{\mr M}_{\mathbb{R}} = G(\mathbb{R}) \setminus X_s(\mathbb{R})$ as a hyperbolic orbifold via the metric induced by the homeomorphism $G(\mathbb{R}) \setminus X_s(\mathbb{R}) \cong P\Gamma \setminus K$ of Theorem \ref{th:realstableperiod}, the metric and hyperbolic orbifold structure on $P\Gamma \setminus K$ of Theorem \ref{theorem1}.1. The goal of Section \ref{sec:triangle2} will be to show that $\overline{\mr M}_{\mathbb{R}}$, as a hyperbolic orbifold, is isomorphic to the triangle $\Delta_{3,5,10}$ in the real hyperbolic plane $H^2$ with angles $\pi/3, \pi/5$ and $\pi/10$, by which we mean that there exists an isometry $\phi: \overline{\mr M}_{\mathbb{R}} \xrightarrow{\sim} \Delta_{3,5,10}$ that preserves the orbifold structures.
\\
\\
The results in the above Sections \ref{sec:1}, \ref{sec:2} and \ref{sec:3} give the orbifold singularities of $\overline{\mr M}_{\mathbb{R}}$ together with their stabilizer groups. In order to complete determine the hyperbolic orbifold structure of $\overline{\mr M}_{\mathbb{R}}$, however, we shall also need to know the underlying topological space of $\overline{\mr M}_{\mathbb{R}}$. This will be our first goal. The first observation is that $\overline{\mr M}_{\mathbb{R}}$ is compact. Indeed, it is classical that the topological space $\overline{\mr M}_{\mathbb{C}} = G(\mathbb{C}) \setminus X_s(\mathbb{C})$ parametrizing complex stable binary quintics is compact - for example because is homeomorphic to $\overline{M}_{0,5}(\mathbb{C})/ \mf S_5$ and the stack of stable $5$-pointed curves $\overline{M}_{0,5}$ is proper \cite{Knudsen1983} or because it is homeomorphic to a compact ball quotient \cite{shimuratranscendental} - and the map $\overline{\mr M}_{\mathbb{R}} \to \overline{\mr M}_{\mathbb{C}}$ is proper. The second observation is that $\overline{\mr M}_{\mathbb{R}}$ is connected: $X_s(\mathbb{R})$ is obtained from the manifold $X(\mathbb{R}) = \{F \in \mathbb{R}[x,y]: F \text{ is homogeneous of degree } 5\}$ by removing a subspace of codimension two.
We can prove more:
\begin{proposition} \label{prop:simplyconnected}
The moduli space $\overline{\mr M}_{\mathbb{R}}$ of real stable binary quintics is simply connected.
\end{proposition}
\begin{proof}
We shall prove that the closures of the three connected components of ${\mr M}_{\mathbb{R}}$ in $\overline{\mr M}_{\mathbb{R}}$ are homeomorphic to the closed disc $D$ in $\mathbb{R}^2$, and that they are glued together to form $\overline{\mr M}_{\mathbb{R}}$ in the way that the components $\mr M_0$ and $\mr M_2$ are glued to $\mr M_1$ in Figure \textcolor{blue}{1}.
Let $M_{0,5}(\mathbb{R})$ (resp. $\overline{M}_{0,5}(\mathbb{R})$) be the moduli space of real smooth (resp. stable) genus zero curves with five real marked points \cite{Knudsen1983}. Recall that $P_0 \subset \mathbb{P}^1(\mathbb{C})$ is the subset of distinct five-tuples, and $P_s \subset \mathbb{P}^1(\mathbb{C})$ the subset of five-tuples where no three coordinates coincides.
We have $M_{0,5}(\mathbb{R}) = \textnormal{PGL}_2(\mathbb{R}) \setminus P_0(\mathbb{R})$ and $\overline{M}_{0,5}(\mathbb{R}) = \textnormal{PGL}_2(\mathbb{R}) \setminus P_s(\mathbb{R})$.
Let $\sigma_1: \mathbb{P}^1(\mathbb{C})^5 \to \mathbb{P}^1(\mathbb{C})^5$ be the anti-holomorphic involution
$
(x_1, x_2, x_3, x_4, x_5) \mapsto (\bar x_1, \bar x_2, \bar x_3, \bar x_5, \bar x_4),
$
and let $\sigma_2: \mathbb{P}^1(\mathbb{C})^5 \to \mathbb{P}^1(\mathbb{C})^5$ be the anti-holomorphic involution
$
(x_1, x_2, x_3, x_4, x_5) \mapsto (\bar x_1, \bar x_3, \bar x_2, \bar x_5, \bar x_4).
$
Then define
$$
P_0^1(\mathbb{R}) = P_0^{\sigma_1}, \;\;\; P_s^1(\mathbb{R}) = P_1^{\sigma_1}, \;\;\;
P_0^2(\mathbb{R}) = P_0^{\sigma_2}, \;\;\; P_s^2(\mathbb{R}) = P_1^{\sigma_2}.
$$
The action of $\mf S_3 \times \mf S_2$ on $\mathbb{P}^1(\mathbb{C})^5$ commutes with $\sigma_1$ and hence $S_3 \times S_2$ preserves the sets $P_0^1(\mathbb{R})$ and $P_s^1(\mathbb{R})$. Similarly, $\mf S_2 \times \mf S_2$ preserves $P_0^2(\mathbb{R})$ and $P_s^2(\mathbb{R})$. On the other hand, $\textnormal{PGL}_2(\mathbb{R})$ acts on all the $P_0^i(\mathbb{R})'s$ and $P_s^i(\mathbb{R})'s$, and this commutes with the $\mf S_3 \times \mf S_2$ and $\mf S_2 \times \mf S_2$ actions.
\\
\\
Recall that $\mr M_i \subset \mr M_\mathbb{R}$ was the connected component of unordered five-tuples $(x_1, \dotsc, x_5) \in \textnormal{PGL}_2(\mathbb{R}) \setminus (P_s/\mf S_5)(\mathbb{R})$ that have $i$ pairs of complex conjugate points $x_i \neq \bar x_i \in x$. One readily observes that
$
\mr M_0 = \textnormal{PGL}_2(\mathbb{R}) \setminus P_0(\mathbb{R}) / \mf S_5, \mr M_1 = \textnormal{PGL}_2(\mathbb{R}) \setminus P_0^1(\mathbb{R}) / \mf S_3 \times \mf S_2, \mr M_2 = \textnormal{PGL}_2(\mathbb{R}) \setminus P_0^2(\mathbb{R}) / \mf S_2 \times \mf S_2$, and we define $\overline{\mr M}_0 = \textnormal{PGL}_2(\mathbb{R}) \setminus P_s(\mathbb{R}) / \mf S_5$,
\begin{align*}
\overline{\mr M}_1 = \textnormal{PGL}_2(\mathbb{R}) \setminus P_s^1(\mathbb{R}) / \mf S_3 \times \mf S_2, \;\;\;
\overline{\mr M}_2 = \textnormal{PGL}_2(\mathbb{R}) \setminus P_s^2(\mathbb{R}) / \mf S_2 \times \mf S_2.
\end{align*}
Note that the natural maps $\overline{\mr M}_i \to \overline{\mr M}_\mathbb{R} $ are closed embeddings of topological spaces; we identify $\overline{\mr M}_i$ with its image in $\overline{\mr M}_{\mathbb{R}}$. Moreover, we have
$
\overline{\mr M}_{\mathbb{R}} = \overline{\mr M}_0 \cup \overline{\mr M}_1 \cup \overline{\mr M}_2.
$
The first step is to show that each $\overline{\mr M}_i$ is homeomorphic to the closed disc $D$ in $\mathbb{R}^2$. We start with $\overline{\mr M}_0$. Let $\Gamma \subset \mathbb{P}^1(\mathbb{R}) \times \mathbb{P}^1(\mathbb{R})$ be the union of the lines
$$
l_1 = \mathbb{P}^1(\mathbb{R}) \times \{0\}, l_2 = \{0\} \times \mathbb{P}^1(\mathbb{R}), k_1 =
\mathbb{P}^1(\mathbb{R}) \times \{1\}, k_2 = \{1\} \times \mathbb{P}^1(\mathbb{R}), $$
$$
m_1 = \mathbb{P}^1(\mathbb{R}) \times \{\infty\}, m_2 = \{\infty\} \times \mathbb{P}^1(\mathbb{R}),
\delta = \Delta = \{(x,x): x \in \mathbb{P}^1(\mathbb{R})\}.
$$
Then
$
\textnormal{PGL}_2(\mathbb{R}) \setminus P_0(\mathbb{R}) \cong \left(\mathbb{P}^1(\mathbb{R}) \times \mathbb{P}^1(\mathbb{R})\right) - \Gamma.
$
Hence $\textnormal{PGL}_2(\mathbb{R}) \setminus P_0(\mathbb{R})$ has $12$ components $C \subset \textnormal{PGL}_2(\mathbb{R}) \setminus P_0(\mathbb{R})$. If $N = \text{Stab}_{\mf S_5}(C)$ is the stabilizer in $\mf S_5$ of one of them, then $|N| = 120/12 = 10$ hence $N = D_5$. On the other hand,
$
\overline{M}_{0,5}(\mathbb{R}) = \textnormal{PGL}_2(\mathbb{R}) \setminus P_s(\mathbb{R})
$
is homeomorphic to the blow-up of $\mathbb{P}^1(\mathbb{R}) \times \mathbb{P}^1(\mathbb{R})$ in three points $0, 1, \infty$. Let $\bar C$ be the closure of $C$ in $\overline{M}_{0,5}(\mathbb{R})$. Then $\bar C$ is simply connected as well as equivariant under $N \subset \mf S_5$. Since the only orbifold point of $\overline{\mr M}_{\mathbb{R}}$ with stabilizer $D_5$ is $\alpha_5:=(0, -1, \infty, \lambda + 1, \lambda)$ with $\lambda = \zeta_5 + \zeta_5^{-1} \in \mathbb{P}^1(\mathbb{R})$ by Lemma \ref{lemma:D5}, and $\alpha_5 \in \overline{M}_{0,5}(\mathbb{R}) / \mf S_5$, there is a $D_5$-equivariant homeomorphism between $\bar C$ and $D \subset \mathbb{R}^2$, mapping $\alpha_5$ to $0 \in D$, with $D_5$ acting on $(D,0)$ in the natural way. Hence
$$
\overline{\mr M}_0 = \overline{M}_{0,5}(\mathbb{R})/ \mf S_5 = \bar C / N = D/D_5
$$
is simply connected, and homeomorphic to the closed disc $D \subset \mathbb{R}^2$.
\\
\\
Next, we treat the case $\overline{\mr M}_1$. Let $\overline{M}_{0,5}^1(\mathbb{R}) = \textnormal{PGL}_2(\mathbb{R}) \setminus P_s^1(\mathbb{R})$ and $M_{0,5}^1(\mathbb{R}) = \textnormal{PGL}_2(\mathbb{R}) \setminus P_0^1(\mathbb{R})$. Now $M_{0,5}^1(\mathbb{R}) = \textnormal{PGL}_2(\mathbb{R}) \setminus P_0^1(\mathbb{R}) \cong$
and $\overline{M}^1_{0,5}$ is the real blow-up
$
B = \textnormal{Bl}_{0, 1, \infty}\mathbb{P}^1(\mathbb{C})
$ of $\mathbb{P}^1(\mathbb{C})$ in the three points $0, 1, \infty$, with $M_{0,5}^1(\mathbb{R}) \hookrightarrow \overline{M}_{0,5}^1(\mathbb{R})$ corresponding to the natural inclusion $\mathbb{P}^1(\mathbb{C}) \setminus \mathbb{P}^1(\mathbb{R}) \hookrightarrow B$. Let $C$ be any of the two connected components of $\mathbb{P}^1(\mathbb{C}) \setminus \mathbb{P}^1(\mathbb{R})$. Then $C$ is a manifold and an open subset of $B$, whose closure $\bar C$ in $B$ is a the simply connected manifold with corners diffeomorphic to the hexagon. Moreover, action of the subgroup $\mf S_2$ of $\mf S_3 \times \mf S_2$ on $\mathbb{P}^1(\mathbb{C}) \setminus \mathbb{P}^1(\mathbb{R})$ interchanges the connected components, hence the stabilizer $N = \text{Stab}_{\mf S_3 \times \mf S_2}(C)$ is $\mf S_3$, and $\bar C$ is stabilized by $N$. By Lemma \ref{lemma:D3}, the only orbifold point of $\overline{\mr M}_{\mathbb{R}}$ with stabilizer group $D_3$ is given by the point $\alpha_3:=(-1,\infty, 0, \omega, \omega^2)$ with $\omega$ a primitive third root of unity. Hence $(\bar C, \alpha_3)$ is homeomorphic to the closed disc $(D,0) \subset (\mathbb{R}^2,0)$ with $N = \mf S_3$ acting on it in the natural way. In particular,
$$
\overline{\mr M}_1 = \overline{M}^1_{0,5} / \mf S_3 \times \mf S_2 = \bar C / N = D / D_3
$$
is homeomorphic to the closed disc $D \subset \mathbb{R}^2$. Finally, consider $\overline{\mr M}_2$. Let $\overline{M}_{0,5}^2(\mathbb{R}) = \textnormal{PGL}_2(\mathbb{R}) \setminus P_s^2(\mathbb{R})$ and $M_{0,5}^2(\mathbb{R}) = \textnormal{PGL}_2(\mathbb{R}) \setminus P_0^2(\mathbb{R})$. Now $M_{0,5}^2(\mathbb{R}) = \textnormal{PGL}_2(\mathbb{R}) \setminus P_0^2(\mathbb{R}) \cong \mathbb{P}^1(\mathbb{C}) \setminus \left(\mathbb{P}^1(\mathbb{R}) \cup \{\infty, i, -i\}\right)$
and $\overline{M}^2_{0,5}$ is the real blow-up
$
B' = \textnormal{Bl}_{\infty}\mathbb{P}^1(\mathbb{C})
$ of the sphere $\mathbb{P}^1(\mathbb{C})$ in the point $\infty$, with $M_{0,5}^2(\mathbb{R}) \hookrightarrow \overline{M}_{0,5}^2(\mathbb{R})$ corresponding to the natural inclusion of $\mathbb{P}^1(\mathbb{C}) \setminus \left(\mathbb{P}^1(\mathbb{R}) \cup \{i, -i\}\right)$ in $B'$. Let $C$ be any of the two connected components of $\mathbb{P}^1(\mathbb{C}) \setminus \mathbb{P}^1(\mathbb{R})$. Then $C$ is a manifold, homeomorphic to the punctured unit disc, and forms an open subset of $B$ whose closure $\bar C$ in $B$ is the simply connected manifold-with-corners that has $2$ two corners.
The subgroup $N$ of $\mf S_2 \times \mf S_2$ that stabilizes $C$ is isomorphic to $\mathbb{Z}/2$, and $\bar C$ is equivariant under $N$. The induced action of $\mathbb{Z}/2$ on $D'$ is reflection in a line through the origin; in particular,
$$
\overline{\mr M}_2 = \overline{M}^2_{0,5} / \mf S_2 \times \mf S_2 = \bar C / N = D' / \mathbb{Z}/2
$$
is homeomorphic to the closed disc $D \subset \mathbb{R}^2$.
Finally, it is clear that $\overline{\mr M}_\mathbb{R} =\overline{\mr M}_0 \cup \overline{\mr M}_1 \cup \overline{\mr M}_2$ is topologically the gluing of three closed discs in the prescribed way, i.e. as in Figure \textcolor{blue}{1}.
\end{proof}
\begin{theorem} \label{th:triangle}
The moduli space $\overline{\mr M}_\mathbb{R} = G(\mathbb{R}) \setminus X_s(\mathbb{R})$ of stable binary quintics equipped with the metric given by Theorem \ref{th:realstableperiod} is isometric to the triangle $\Delta_{3,5,10} \subset \mathbb{R} H^2$ of angles $\pi/3, \pi/5$ and $\pi/10$ in the real hyperbolic plane $\mathbb{R} H^2$.
\end{theorem}
\begin{proof}
To any closed $2$-dimensional orbifold $O$ one can associate a set of natural numbers $S_O = \{n_1, \dotsc, n_k; m_1, \dotsc, m_l\}$ by letting $k$ be the number of cone points of $X_O$, $l$ the number of corner reflectors, $n_i$ the order of the $i$-th cone point and $2m_j$ the order of the $j$-th corner reflector. A closed $2$-dimensional orbifold $O$ is then determined, up to orbifold-structure preserving homeomorphism, by its underlying space $X_O$ and the set $S_O$. By Proposition \ref{prop:simplyconnected}, $\overline{\mr M}_\mathbb{R}$ is simply connected. By Proposition \ref{prop:conesreflectors}, $\overline{\mr M}_\mathbb{R}$ has no cone points and three corner reflectors whose orders are $\pi/3, \pi/5$ and $\pi/10$. This implies $\overline{\mr M}_\mathbb{R}$ and $\Delta_{3,5,10}$ are isomorphic as topological orbifolds. Consequently, the orbifold fundamental group of $\overline{\mr M}_\mathbb{R}$ is abstractly isomorphic to the group $\Gamma_{3,5,10} := \langle a,b,c \vert a^2 = b^2 = c^2 = (ab)^3 = (ac)^5 = (bc)^{10} = 1 \rangle$. Now let $\phi: \Gamma_{3,5,10} \hookrightarrow \text{PSL}_2(\mathbb{R})$ be \textit{any} embedding such that $X:=\phi\left(\Gamma_{3,5,10}\right) \setminus \mathbb{R} H^2$ is a hyperbolic orbifold; we claim that there is a fundamental domain $\Delta$ of $X$ isometric to $\Delta_{3,5,10}$. Consider the generator $a \in \Gamma_{3,5,10}$. Since $\phi(a)^2 = 1$, there exists a geodesic $L_1 \subset \mathbb{R} H^2$ such that $\phi(a) \in \text{PSL}_2(\mathbb{R}) = \text{Isom}(\mathbb{R} H^2)$ is the reflection across $L_1$. Next, consider the generator $b \in \Gamma_{3,5,10}$. Again, there exists a geodesic $L_2 \subset \mathbb{R} H^2$ such that $\phi(b)$ is the reflection across $L_2$. One easily shows that $L_2 \cap L_1 \neq \emptyset$.
Let $x \in L_1 \cap L_2$. Then $\phi(a)\phi(b)$ is an element of order three that fixes $x$, hence $\phi(a)\phi(b)$ is a rotation around $x$. Therefore, one of the angles between $L_1$ and $L_2$ must be $\pi/3$. Finally, we know that $\phi(c)$ is an element of order $2$ in $\textnormal{PGL}_2(\mathbb{R})$, hence a reflection across a line $L_3$. By the previous arguments, $L_3 \cap L_2 \neq \emptyset$ and $L_3 \cap L_1 \neq \emptyset$. It also follows that $x \in L_3 \cap L_2 \cap L_1 = \emptyset$.
Consequently,
the three geodesics $L_i \subset \mathbb{R} H^2$ enclose a hyperbolic triangle; the orders of $\phi(a)\phi(b)$, $\phi(a)\phi(c)$ and $\phi(b)\phi(c)$ imply that the three interior angles of the triangle are $\pi/3$, $\pi/5$ and $\pi/10$.
\end{proof}
\section{Monodromy Groups of Complex and Real Binary Quintics} \label{sec:monodromygroups}
In this section, we describe the monodromy group $P\Gamma$ attached to the universal $\mr C \to X_0(\mathbb{C})$ family of cyclic quintic covers $C \to \mathbb{P}^1_\mathbb{C}$ ramified along a smooth binary quintic, as well as the groups $P\Gamma_\alpha$ appearing in Proposition \ref{prop:realsmoothperiods}, i.e. the groups that make $P\Gamma_\alpha \setminus \left( \mathbb{R} H^2 - \mr H \right)$ isomorphic to a connected component $\mr M_i$ of the moduli space of smooth binary quintics $\mr M_\mathbb{R}$.
\\
\\
We start with an explicit description of $P\Gamma$. Let $\mr C \to X_0(\mathbb{C})$ be the universal family of cyclic quintic covers of $\mathbb{P}^1_\mathbb{C}$ ramified along a smooth binary quintic. Let $K = \mathbb{Q}(\zeta_5)$ and let $\Lambda$ be the $\mathcal{O}_K$-module $H^1(C(\mathbb{C}), \mathbb{Z})$ for the cyclic quintic $C \to \mathbb{P}^1_\mathbb{C}$ ramified along the smooth binary quintic defined by $F_0 \in X_0(\mathbb{C})$. Let $\mf h$ be the hermitian form of Equation (\ref{eq:hermitianformonbinaryquinticlattice}), define $\Gamma = \textnormal{Aut}_{\mathcal{O}_K}(\Lambda)$, $P\Gamma = \Gamma/\mu_K$ and consider the monodromy representation $\rho: \pi_1(X_0(\mathbb{C}), F_0) \to P\Gamma$ defined in Equation (\ref{eq:monodromy}). Recall that $\rho$ is surjective by Corollary \ref{cor:surjectivemon}. Moreover, we have:
\begin{theorem}[Shimura] \label{th:calculatemonodromyshimura}
There is an isomorphism $\left( \Lambda, \mf h \right) \cong \left( \mathcal{O}_K^3, \textnormal{diag}(1,1, \frac{1 - \sqrt 5}{2}) \right)$.
\end{theorem}
\begin{proof}
See \cite[Section 6]{shimuratranscendental} as well as item (5) in the table on page 1.
\end{proof}
So let us simply write $\Lambda = \mathcal{O}_K^3$ and $\mf h = \textnormal{diag}(1,1, \frac{1 - \sqrt 5}{2})$ in the remaining part of Section \ref{sec:monodromygroups}. Write $\alpha = \zeta_5 + \zeta_5^{-1} = \frac{\sqrt{5} - 1}{2}$. Recall that $\theta = \zeta_5 - \zeta_5^{-1}$ and observe that $|\theta|^2 = \frac{\sqrt{5} + 5}{2}$. Define three quadratic forms $q_0$, $q_1$ and $q_2$ on $\mathbb{Z}[\alpha]^3$ as follows:
\begin{equation} \label{eq:explicitquadraticforms}
\begin{split}
q_0(x_0, x_1, x_2) & = x_0^2 + x_1^2 - \alpha x_2^2
\\
q_1(x_0, x_1, x_2) & = |\theta|^2 x_0^2 + x_1^2 - \alpha x_2^2
\\
q_2(x_0, x_1, x_2) & = |\theta|^2 x_0^2 + |\theta|^2 x_1^2 - \alpha x_2^2
\end{split}
\end{equation}
We consider $\mathbb{Z}[\alpha]$ as a subring of $\mathbb{R}$ via the standard embedding.
\begin{theorem} \label{th:explicitquadratic}
Consider the quadratic forms $q_j : \mathbb{Z}[\alpha]^3 \to \mathbb{Z}[\alpha]$. There is a union geodesic subspaces $\mr H_j \subset \mathbb{R} H^2$ for each $j \in \{0,1,2\}$ and an isomorphism of real hyperbolic orbifolds
\begin{equation}
\mr M_\mathbb{R} \cong \coprod_{j = 0}^2 \textnormal{PO}(q_j,\mathbb{Z}[\alpha]) \setminus \left(\mathbb{R} H^2 - \mr H_j \right).
\end{equation}
\begin{proof}
Recall that $\theta = \zeta_5 - \zeta_5^{-1}$; we consider the $\bb F_5$-vector space $W$ equipped with the quadratic form $q = \mf h \mod \theta$.
We consider three explicit anti-involutions $\alpha_j$ on the $\ca O_K$-lattice $\Lambda$:
\begin{equation} \label{chi}\begin{split}
\alpha_0: & (x_0,x_1,x_2) \mapsto (\;\;\;\bar x_0, \;\;\;\bar x_1,\bar x_2) \\
\alpha_1: & (x_0,x_1,x_2) \mapsto (-\bar x_0, \;\;\;\bar x_1, \bar x_2) \\
\alpha_2: & (x_0,x_1,x_2) \mapsto (-\bar x_0, -\bar x_1, \bar x_2).
\end{split}
\end{equation}
For isometries $\alpha: W \to W$, the dimension and determinant of the fixed space $(W^\alpha, q|_{W^\alpha})$ are conjugacy-invariant. Using this, one easily shows that an anti-unitary involution of $\Lambda$ is $\Gamma$-conjugate to exactly one of the $\pm \alpha_j$, hence
$C\mr A$ has cardinality $3$ and is represented by $\alpha_0, \alpha_1, \alpha_2$ of (\ref{chi}). By Proposition \ref{prop:realsmoothperiods}, we obtain $\mr M_\mathbb{R} \cong \coprod_{j = 0}^2 P\Gamma_{\alpha_j} \setminus (\mathbb{R} H^2_{\alpha_j} - \mr H)$ where each hyperbolic quotient $P\Gamma_{\alpha_j} \setminus (\mathbb{R} H^2_{\alpha_j} - \mr H)$ is connected. Next, consider the fixed lattices
\begin{equation}\label{eq:fixedlattices} \begin{split}
\Lambda_0 := \Lambda^{\alpha_0} = \mathbb{Z}[\alpha] \oplus \mathbb{Z}[\alpha] \oplus \mathbb{Z}[\alpha] \\
\Lambda_1:= \Lambda^{\alpha_1} = \theta \mathbb{Z}[\alpha] \oplus \mathbb{Z}[\alpha] \oplus \mathbb{Z}[\alpha] \\
\Lambda_2:= \Lambda^{\alpha_2} = \theta \mathbb{Z}[\alpha] \oplus \theta\mathbb{Z}[\alpha] \oplus \mathbb{Z}[\alpha].
\end{split}
\end{equation}
One easily shows that $P\Gamma_{\alpha_j}=N_{P\Gamma}(\alpha_j)$ for the normalizer $N_{P\Gamma}(\alpha_j)$ of $\alpha_j$ in $P\Gamma$. Moreover, if $h_j$ denotes the restriction of $\mf h$ to $\Lambda^{\alpha_j}$, then there is a natural embedding
\begin{equation}
\iota: N_{P\Gamma}(\alpha_j) \hookrightarrow \textnormal{PO}(\Lambda_j,h_j, \mathbb{Z}[\alpha]).
\end{equation}
We claim that $\iota$ is actually an isomorphism. Indeed, this follows from the fact that the natural homomorphism
$
\pi: N_\Gamma(\alpha_j) \to O(\Lambda_j, h_j)
$
is surjective, where $N_\Gamma(\alpha_j) = \{g \in \Gamma: g \circ \alpha_j = \alpha_j \circ g \}$ is the normalizer of $\alpha_j$ in $\Gamma$. The surjectivity of $\pi$ follows in turn from the equality
$$
\Lambda = \ca O_K \cdot \Lambda_j + \ca O_K \cdot \theta \Lambda_j^\vee \subset K^3
$$
which can be deduced from (\ref{eq:fixedlattices}). Since $\textnormal{PO}(\Lambda_j,h_j, \mathbb{Z}[\alpha]) = \textnormal{PO}(q_j,\mathbb{Z}[\alpha])$, we are done.
\end{proof}
\end{theorem}
\printbibliography
\end{document}
\subsection{The Jacobian of a cyclic quintic cover of $\mathbb{P}^1$}
We begin with the following:
\begin{lemma} \label{lemma:refinedhodge}
Let $Z \subset \bb P^1_\mathbb{C}$ be a smooth quintic hypersurface. Let $ \bb P^2_\mathbb{C} \supset C \to \mathbb{P}^1_\mathbb{C}$ be the quintic cover of $\bb P^1$ ramified along $Z$. Then $C$ has the following refined Hodge numbers:
\begin{align}
h^{1,0}(C)_{\zeta} = 3, \;\;\; h^{1,0}(C)_{\zeta^2} = 2, \;\;\;
h^{1,0}(C)_{\zeta^3} = 1, \;\;\;
h^{1,0}(C)_{\zeta^{4}} = 0 \\
h^{0,1}(C)_{\zeta} = 0, \;\;\; h^{0,1}(C)_{\zeta^2} = 1, \;\;\;
h^{0,1}(C)_{\zeta^3} = 2, \;\;\;
h^{0,1}(C)_{\zeta^{4}} = 3.
\end{align}
\end{lemma}
\begin{proof}
This follows from the Hurwitz-Chevalley-Weil formula, see \cite[Proposition 5.9]{Moonen2011TheTL}. Alternatively, see \cite[Section 5]{carlsontoledomonodromy}.
\end{proof}
Now fix a point $F_0 \in X_0(\mathbb{C})$ and let
\begin{equation} \label{eq:coverp1}
C = \{z^5 = F_0(x,y) \} \subset \mathbb{P}^2_\mathbb{C}
\end{equation}
be the corresponding cyclic cover of $\mathbb{P}^1_\mathbb{C}$.
Let $A = J(C) = \text{Pic}^0(C)$ be the Jacobian of $C$. Then $A$ is a principally polarized abelian variety of dimension $ 6$ equipped with a homomorphism $\iota: \mathcal{O}_K = \mathbb{Z}[\zeta] \to \textnormal{End}(A)$. Write $\Lambda = H_1(A(\mathbb{C}), \mathbb{Z})$ and consider the Hodge decomposition $\Lambda \otimes_\mathbb{Z} \mathbb{C} = H^{-1,0} \oplus H^{0,-1}$.
\begin{comment}
For $i \in \mathbb{Z}/5$, define $$H^{-1,0}_{(i)} = \{x \in H^{-1,0}: \iota(\zeta)(x) = \zeta^ix \}.$$
Let $h^{-1,0}_i = \dim_\mathbb{C} H^{-1,0}_{(i)}$. Since
$
H^{-1,0} = \textnormal{Lie}(A) = H^1(C, \mathcal{O}_C)$, we have $h^{-1,0}_{(i)} = i-1$ for $1 \leq i \leq 4$.
\end{comment}
Define $\tau_i: K \to \mathbb{C}, i \in \{1,2\}$ by $
\tau_1(\zeta) = \zeta^3, \tau_2(\zeta) = \zeta^4.
$
Since
$
H^{-1,0} = \textnormal{Lie}(A) = H^1(C, \mathcal{O}_C) = H^{0,1}(C)$, Lemma \ref{lemma:refinedhodge} implies that
\begin{equation} \label{eq:signaturefivepoints2}
\dim_\mathbb{C} H^{-1,0}_{\tau_1} = 2, \;\;\; \dim_\mathbb{C} H^{-1,0}_{\tau_1\sigma} = 1, \;\;\; \dim_\mathbb{C} H^{-1,0}_{\tau_2} = 3, \;\;\; \dim_\mathbb{C} H^{-1,0}_{\tau_2\sigma} = 0.
\end{equation}
Define $\eta = 5/(\zeta - \zeta^{-1})$. Then $\mf D_K = (\eta)$ (see Lemma \ref{lemma:discr}). Let $E: \Lambda \times \Lambda \to \mathbb{Z}$ be the alternating form corresponding to the polarization of the abelian variety $A$. Then for $a \in \mathcal{O}_K$ and $x,y \in \Lambda$, we have $E(\iota(a)x,y) = E(x, \iota(a^\sigma)y)$. Let $T: \Lambda \times \Lambda \to \mf D_K^{-1}$ be the skew-hermitian form corresponding to $E$ via Lemma \ref{lemma:equivalentforms}. Then
$T(x,y) = \frac{1}{5}\sum_{j = 0}^{4}\zeta^jE\left( x, \iota(\zeta)^j y \right)$ by Example \ref{examplesunitary}.\ref{ex:unittwo}. Putting this together, we obtain a hermitian form on the free $\mathcal{O}_K$-module $\Lambda$ as follows:
\begin{equation} \label{eq:hermitianformonbinaryquinticlattice}
\mf h: \Lambda \times \Lambda \to \mathcal{O}_K, \;\;\;
\mf h(x,y) = \eta T(x,y) = (\zeta - \zeta^{-1})^{-1}\sum_{j = 0}^{4}\zeta^jE\left( x, \iota(\zeta)^j y \right)
\end{equation}
Observe that the hermitian lattice $(\Lambda, \mf h)$ is unimodular because $(\Lambda, E)$ is unimodular:
$$
\{x \in \Lambda_\mathbb{Q}: \mf h(x, \Lambda) \subset \mathcal{O}_K\} =
\{x \in \Lambda_\mathbb{Q}: T(x, \Lambda) \subset \mf D_K^{-1}\} =
\{x \in \Lambda_\mathbb{Q}: E(x, \Lambda) \subset \mathbb{Z}\} = \Lambda.
$$
For each embedding $\varphi: K \to \mathbb{C}$, the restriction of the hermitian form $\varphi(\eta)\cdot E_\mathbb{C}(x, \bar y)$ on $\Lambda_\mathbb{C}$
to $(\Lambda_\mathbb{C})_{\varphi} \subset \Lambda_\mathbb{C}$ coincides with
$
\mf{h}^{\varphi}$
by Lemma \ref{lemma:agree}.
Since $\Im (\tau_i(\zeta - \zeta^{-1})) < 0$ for $i = 1,2$,
the signature of $\mf{h}^{\tau_i}$ is $(h^{-1,0}_{\tau_1}, h^{0,-1}_{\tau_1}) = (2,1)$ for $i = 1$ and
$(h^{-1,0}_{\tau_2}, h^{0,-1}_{\tau_2}) = (3,0)$ for $i = 2$.
\subsection{Marked binary quintics} \label{sec:markedbinary}
Let $\pi: \mr C \to X_0$ be the universal family of cyclic covers $C \to \mathbb{P}^1$ ramified along a smooth binary quintic $\{F = 0\} \subset \mathbb{P}^1$. Let $\phi: J \to X_0$ be the relative Jacobian of $\pi$; it is an abelian scheme of relative dimension $6$ over $X_0$ with $\mathcal{O}_K$-action of signature $\{(2,1), (3,0)\}$ with respect to $\Psi = \{\tau_1, \tau_2\}$. Let $\bb V$ be the local system of hermitian $\mathcal{O}_K$-modules $R^1\pi_\ast \mathbb{Z}$ underlying the abelian scheme $J/X_0$. It corresponds to a monodromy representation $\rho': \pi_1(X_0(\mathbb{C}), F_0) \to \Gamma$, where $\Gamma = \textnormal{Aut}_{\mathcal{O}_K}(\Lambda, \mf h)$, whose with the quotient map $\Gamma \to P\Gamma = \Gamma / \mu_K$ defines a homomorphism
\begin{equation} \label{eq:monodromy}
\rho: \pi_1(X_0(\mathbb{C}), F_0) \to P\Gamma.
\end{equation}
We shall see that $\rho$ is surjective, see Corollary \ref{cor:surjectivemon} below.
\\
\\
Let $F \in X_0(\mathbb{C})$ and let $Z_F = \{F = 0\} \subset \mathbb{P}^1_\mathbb{C}$ be the associated hypersurface. A \textit{marking} of $F$ is a ring isomorphism $m: H^0(Z_F(\mathbb{C}), \mathbb{Z}) \xrightarrow{\sim} \mathbb{Z}^5$. Clearly, this is the same thing as a labelling of the points $p \in Z_F(\mathbb{C})$.
Let $\ca M_0$ be the space of marked binary quintics. Let $\psi: \ca Z \to X_0(\mathbb{C})$ be the universal complex binary quintic, and consider the local system $H = \psi_\ast \mathbb{Z}$ of stalk $H_F = H^0(Z_F(\mathbb{C}), \mathbb{Z})$ for $F \in X_0(\mathbb{C})$. Then $H$ corresponds to a monodromy representation
\begin{equation}\label{eq:monodromy3}
\tau: \pi_1(X_0(\mathbb{C}), F_0) \to \mf S_5.
\end{equation}
It can be shown that $\tau$ is surjective using the results in \cite{beauvillemonodromie}. This implies that $\ca M_0 \to X_0(\mathbb{C})$ is covering space, i.e. $\ca M_0$ is connected. If we choose a marking $m_0: H^0(Z_{F_0}(\mathbb{C}), \mathbb{Z}) \cong \mathbb{Z}^5$ lying over our base point $F_0 \in X_0(\mathbb{C})$, we obtain an embedding
$
\pi_1 \left( \ca M_0, m_0 \right) \hookrightarrow \pi_1(X_0(\mathbb{C}), F_0)$,
whose composition with $\rho$ in (\ref{eq:monodromy}) defines a homomorphism
\begin{equation}\label{eq:monodromy2}
\mu: \pi_1(\ca M_0, m_0) \to P\Gamma.
\end{equation}
Define $\theta = \zeta - \zeta^{-1}$ and consider the $3$-dimensional $\bb F_5$ vector space $\Lambda/\theta\Lambda$ and the quadratic space $W : = \left(\Lambda/\theta\Lambda,q\right)$ where $q$ is the quadratic form obtained by reducing $\mf h$ modulo $\theta\Lambda$. Let $\Gamma_\theta$ be the kernel of $\Gamma \to \textnormal{Aut}(W)$; then $P\Gamma_\theta = \textnormal{Ker}\left( P\Gamma \to P\textnormal{Aut}(W) \right) \subset \text{PU}(2,1)$.
Remark that the composition $\ca M_0 \to X_0(\mathbb{C}) \to X_s(\mathbb{C})$ admits an essentially unique \textit{completion} $\ca M_s \to X_s(\mathbb{C})$, see \cite{fox}. Here $\ca M_s$ a manifold and $\ca M_s \to X_s(\mathbb{C})$ is a ramified covering space.
\begin{proposition} \label{prop:commutativemonodromy}
The image of $\mu$ in (\ref{eq:monodromy2}) is the group $P\Gamma_\theta$, and the induced homomorphism $ \pi_1(X_0(\mathbb{C}), F_0)/ \pi_1 \left( \ca M_0, m_0 \right)= \mf S_5 \to P\Gamma /P\Gamma_\theta$ is an isomorphism. In other words, we obtain the following commutative diagram with exact rows:
\begin{equation} \label{eq:commutativemonodromy}
\xymatrix{
0 \ar[r] & \pi_1(\ca M_0, m_0) \ar@{->>}[d]^\mu\ar[r] & \pi_1(X_0(\mathbb{C}), F_0)\ar[d]^\rho \ar[r]^{\;\;\;\;\;\; \tau} & \mf S_5\ar[d]^*[@]{\sim}_\gamma \ar[r] & 0 \\
0 \ar[r] & P\Gamma_\theta \ar[r] & P\Gamma \ar[r] & P\textnormal{Aut}(W) \ar[r] & 0.
}
\end{equation}
\begin{proof}
Consider the quotient $Q = G \setminus \ca M_0 = \textnormal{PGL}_2(\mathbb{C}) \setminus P_0$ and let $0 \in Q$ be the image of $m_0 \in \ca M_0$. In \cite{DeligneMostow}, Deligne and Mostow define a hermitian space bundle $B_Q \to Q$ over $Q$ whose fiber over $0 \in Q$ is $\mathbb{C} H^2$. Consequently, writing $V_1 = \Lambda \otimes_{\mathcal{O}_K, \tau_1} \mathbb{C}$, this gives a monodromy representation $\pi_1(Q,0) \to \textnormal{PU}(V_1, \mf h^{\tau_1}) \cong \textnormal{PU}(2,1)$ whose image we denote by $\Gamma_{\text{DM}}$. Kondō has shown that in fact, $\Gamma_{\text{DM}} = P\Gamma_\theta$ \cite[Theorem 7.1]{kondo5points}. Since $\ca M_0 \to Q$ is a covering space (the action of $G$ on $\ca M_0$ being free) we have an embedding $\pi_1(\ca M_0, m_0) \hookrightarrow \pi_1(Q, 0)$ whose composition with $\pi_1(Q,0) \to\textnormal{PU}(2,1)$ is the map $\mu:\pi_1(\ca M_0, m_0) \to P\Gamma \subset \textnormal{PU}(2,1)$.
\\
\\
To prove that the image of $\mu$ is $P\Gamma_\theta$, it suffices to give a section of the map $\ca M_0 \to Q$. Indeed, this induces a retraction of $\pi_1(\ca M_0, m_0) \hookrightarrow \pi_1(Q, 0)$ so that the images of both groups in $\textnormal{PU}(2,1)$ are the same. To define such a section, observe that if $\Delta \subset \mathbb{P}^1(\mathbb{C})^5$ is the union of all hyperplanes $\{x_i = x_j\} \subset \mathbb{P}^1(\mathbb{C})^5$ for $i \neq j$, then
$$
Q = \textnormal{PGL}_2(\mathbb{C}) \setminus P_0 = \textnormal{PGL}_2(\mathbb{C}) \setminus \left( \mathbb{P}^1(\mathbb{C})^5 - \Delta \right) \cong \{(x_4,x_5) \in \AA^1(\mathbb{C})^2 : x_i \neq 0,1 \textnormal{ and } x_1 \neq x_2 \}.
$$
The section $Q \to \ca M_0$ may then be defined by sending $(x_4, x_5)$ to the binary quintic $F(X,Y) = X(X-Y)Y(X-x_4Y)(X-x_5Y) \in X_0(\mathbb{C})$, marked by the labelling of its roots $\{0,1, \infty, x_4, x_5\}$. It remains to prove that the homomorphism $\gamma: \mf S_5 \to P\Gamma/P\Gamma_\theta$ appearing on the right in (\ref{eq:commutativemonodromy}) is an isomorphism.
We use Theorem \ref{th:calculatemonodromyshimura}, proven by Shimura in \cite{shimuratranscendental}, which says that $( \Lambda, \mf h ) \cong ( \mathcal{O}_K^3, \textnormal{diag}(1,1, \frac{1 - \sqrt 5}{2}))$. It follows that $P\Gamma/P\Gamma_\theta = P\textnormal{Aut}(W) \cong \textnormal{PO}_3(\bb F_5) \cong \mf S_5$. Next, consider the manifold $\ca M_s$. Remark that $\mf S_5$ embeds into $\textnormal{Aut}(G \setminus \ca M_s)$, and that the natural map $\ca P: G \setminus \ca M_s \to P\Gamma_\theta \setminus \mathbb{C} H^2$ (see (\ref{eq:periodframed})) is an isomorphism, see \cite{DeligneMostow}, \cite{kondo5points}. The composition $\mf S_5 \subset \textnormal{Aut}(G \setminus \ca M_s) \cong \textnormal{Aut}(P\Gamma_\theta \setminus \mathbb{C} H^2)$ coincides with the composition $\mf S_5 \to P\Gamma/P\Gamma_\theta \subset \textnormal{Aut}(P\Gamma_\theta \setminus \mathbb{C} H^2)$ by equivariance of $\ca P$ with respect to $\gamma$, hence $\gamma$ is injective.
\end{proof}
\end{proposition}
\begin{corollary} \label{cor:surjectivemon}
The monodromy representation $\rho$ in (\ref{eq:monodromy}) is surjective. $\hfill \qed$
\end{corollary}
\subsection{Framed binary quintics}
By a \textit{framing} of a point $F \in X_0(\mathbb{C})$ we mean a projective equivalence class $[f]$, where $f: \mathbb{V}_F = H^1(C_F(\mathbb{C}), \mathbb{Z}) \to \Lambda$ is an $\mathcal{O}_K$-linear isometry: two such isometries are in the same class if and only if they differ by an element in $\mu_K$. Let $\ca F_0$ be the collection of all framings of all points $x \in X_0(\mathbb{C})$. The set $\ca F_0$ is naturally a complex manifold.
Note that Corollary \ref{cor:surjectivemon} implies that $\ca F_0$ is connected, hence $\ca F_0 \to X_0(\mathbb{C})$ is a covering space with Galois group $P\Gamma$.
\begin{lemma} \label{lemma:isomorphiccoveringspaces}
The spaces $P\Gamma_\theta \setminus \ca F_0$ and $ \ca M_0$ are isomorphic as covering spaces of $X_0(\mathbb{C})$. In particular, there is a covering map $\ca F_0 \to \ca M_0$ with Galois group $P\Gamma_\theta$.
\end{lemma}
\begin{proof}
The quotients $P\Gamma \to P\Gamma/P\Gamma_\theta$ and $P\Gamma \to \mf S_5$ are isomorphic by Proposition \ref{prop:commutativemonodromy}.
\end{proof}
\begin{lemma} \label{lemma:irreduciblenormal}
The variety $\Delta := X_s - X_0$ is an irreducible normal crossings divisor in $X_s$.
\end{lemma}
\begin{proof}
The proof is similar to the proof of Proposition 6.7 in \cite{beauvillecubicsurfaces}.
\end{proof}
\begin{lemma} \label{lemma:monodromy}
The local monodromy transformations of $\ca F_0 \to X_0(\mathbb{C})$ around every $x \in \Delta(\mathbb{C})$ are of finite order. More precisely, if $x \in \Delta(\mathbb{C})$ lies on the intersection of $k$ local components of $\Delta(\mathbb{C})$, then the local monodromy group around $x$ is isomorphic to $(\mathbb{Z}/10)^k$.
\end{lemma}
\begin{proof}
See \cite[Proposition 9.2]{DeligneMostow} or \cite[Proposition 6.1]{carlsontoledomonodromy} for the generic case, i.e. when a quintic $Z = \{F = 0\} \subset \mathbb{P}^1_\mathbb{C}$ aquires only one node. In this case, the local equation of the singularity is $x^2 = 0$, hence the curve $C_F$ acquires a singularity of the form $y^5 + x^2 = 0$. If the quintic acquires two nodes, then $C_F$ acquires two such singularities; the vanishing cohomology splits as an orthogonal direct sum hence the local monodromy transformations commute.
\end{proof}
\begin{corollary} \label{cor:framedquintics}
There is a manifold $\ca F_s$, unique up to isomorphism, with a branched cover $\pi: \ca F_s \to X_s(\mathbb{C})$ extending $\ca F_0 \to X_0(\mathbb{C})$ such that for any $x \in \Delta(\mathbb{C})$ and any open neighborhood $x \in U \subset X_s(\mathbb{C})$ with $U \cong D^6$ and $U \cap X_0(\mathbb{C}) \cong (D^\ast)^k \times D^{6-k}$,
any component of $\pi^{-1}(U) \subset \ca F_s$ is isomorphic to $D^k \times D^{6-k}$, mapping to $U$ by $(z_1, \dotsc, z_6) \mapsto (z_1^{r_1}, \dotsc, z_k^{r_k}; z_{k+1}, \dotsc, z_6)$.
\end{corollary}
\begin{proof}
See \cite[Lemma 7.2]{beauvillecubicsurfaces}. See also \cite{fox} and \cite[Section 8]{DeligneMostow}.
\end{proof}
The group $G = \textnormal{GL}_2(\mathbb{C})/\mu_K$ acts on $\ca F_0$ over its action on $X_0$. Explicitly, $g \in G$ and if $[\phi], \phi: \mathbb{V}_{t} \to \Lambda$ is a framing of $t \in X_0(\mathbb{C})$, then $[\phi \circ g^\ast], \phi \circ g^\ast: \mathbb{V}_{gt} \to \Lambda$ is a framing of $gt \in X_0(\mathbb{C})$. This is a left action.
Note that $P\Gamma$ also acts on $\ca F_0$ from the left, and that the actions of $P\Gamma$ and $G$ on $\ca F_0$ commute. By the naturality of the Fox completion, the action of $G$ on $\ca F_0$ extends to an action of $G$ on $\ca F_s$.
\begin{lemma} \label{cor:freeaction}
The group $G = \textnormal{GL}_2(\mathbb{C})/\mu_K$ acts freely on $\ca F_s$.
\end{lemma}
\begin{proof}
The universality of the Fox completion gives an action of $G$ on $\ca M_s$ such that, by Lemma \ref{lemma:isomorphiccoveringspaces}, $P\Gamma_\theta \setminus \ca F_s$ and $\ca M_s$ are $G$-equivariantly isomorphic as branched covering spaces of $X_s(\mathbb{C})$. In particular, it suffices to show that $G$ acts freely on $\ca M_s$. Note that $\ca M_0$ admits a natural $\mathbb{G}_m$-covering map $\ca M_0 \to P_0$ where $P_0 \subset \mathbb{P}^1(\mathbb{C})^5$ is the space of distinct ordered $5$-tuples in $\mathbb{P}^1(\mathbb{C})$. Consequently, there is a $\mathbb{G}_m$-quotient map $\ca M_s \to P_s$, where $P_s$ is the space of stable ordered $5$-tuples, and this map is equivariant for the homomorphism $\textnormal{GL}_2(\mathbb{C}) \to \textnormal{PGL}_2(\mathbb{C})$.
But it is clear that $\textnormal{PGL}_2(\mathbb{C})$ acts freely on $P_s$. Therefore, if $g \in \textnormal{GL}_2(\mathbb{C})$ and $x \in \ca M_s$ are such that $gx = x$, then $g = \lambda \in \mathbb{C}^\ast$. Let $F \in X_s(\mathbb{C})$ be the image of $x \in \ca M_s$; then $gF(x,y) = F(g^{-1}(x,y)) = F(\lambda^{-1}x, \lambda^{-1}y) = \lambda^{-5}F(x,y)$. The equality $gF = F$ implies that $\lambda^{5} = 1 \in \mathbb{C}$; consequently, $\lambda \in \mu_K$. We conclude that $[g] = [\textnormal{id}] \in G$ and the lemma follows.
\end{proof}
\subsection{Complex uniformization}
Consider the hermitian space $V_1 = \Lambda \otimes_{\mathcal{O}_K, \tau_1} \mathbb{C}$; define $\mathbb{C} H^2$ to be the space of negative lines in $V_1$. Using Proposition \ref{prop:canonicalbijection} we see that the abelian scheme $J \to X_0$ induces a $G$-equivariant morphism
$
\ca P: \ca F_0 \to \mathbb{C} H^2$. Explicitly, if $(F, [f]) \in \ca F_0$ is the framing $[f: H^1(C_F(\mathbb{C}), \mathbb{Z}) \xrightarrow{\sim}\Lambda]$ of the binary quintic $F \in X_0(\mathbb{C})$, and $A_F$ is the Jacobian of the curve $C_F$, then
\begin{equation} \label{eq:periodframed}
\ca P: \ca F_0 \to \mathbb{C} H^2, \;\;\; \;\;\;
\ca P \left( F, [f] \right) = f \left( H^{0,-1}(A_F){\tau_1} \right) = f \left(H^{1,0}(C_f)_{\zeta^3} \right) \in \mathbb{C} H^2.
\end{equation}
By general considerations, $\ca P$ is holomorphic.
It descends to a morphism of analytic spaces
\begin{equation} \label{eq:period}
\mr M_\mathbb{C} =G \setminus X_0(\mathbb{C}) \to P\Gamma \setminus \mathbb{C} H^2.
\end{equation}
By Riemann extension, (\ref{eq:periodframed}) extends to a $G$-equivariant holomorphic map
$
\overline{\ca P}: \ca F_s \to \mathbb{C} H^2.
$
\begin{theorem}[Deligne-Mostow] \label{th:delignemostow}
The map $\overline{\ca P}$ defines an isomorphism of complex manifolds
$
\overline{\mr M}_\mathbb{C}^f:=G \setminus \ca F_s \cong \mathbb{C} H^2.
$
Taking $P\Gamma$-quotients gives an isomorphism of complex analytic spaces
\begin{equation} \label{eq:isomstablefivepoints}
\overline{\mr M}_\mathbb{C} = G \setminus X_s(\mathbb{C}) \cong P\Gamma \setminus \mathbb{C} H^2.
\end{equation}
\end{theorem}
\begin{proof}
In \cite{DeligneMostow}, Deligne and Mostow define $\widetilde Q \to Q$ to be the covering space corresponding to the monodromy representation $\pi_1(Q,0) \to \textnormal{PU}(2,1)$; since the image of this homomomorphism is $P\Gamma_\theta$ (see the proof of Proposition \ref{prop:commutativemonodromy}), it follows that $G \setminus \ca F_0 \cong \widetilde Q$ as covering spaces of $Q$. Consequently, if $\widetilde Q_{\textnormal{st}}$ is the Fox completion of the spread $\widetilde Q \to Q \to Q_{\textnormal{st}} := G \setminus \ca M_s = \textnormal{PGL}_2(\mathbb{C}) \setminus P_s$, then there is an isomorphism $G \setminus \ca F_s \cong \widetilde Q_{\textnormal{st}}$ of branched covering spaces of $ Q_{\textnormal{st}}$. Consequently, we obtain commutative diagrams, where the lower right morphism uses (\ref{eq:commutativemonodromy}):
$$
\xymatrixcolsep{5pc}
\xymatrix{
G \setminus \ca F_s \ar[r]^{\sim \;\;\;\;} \ar[d] & \widetilde{Q}_{\textnormal{st}}\ar[d] \ar[r] & \mathbb{C} H^2 \ar[d] \\
G\setminus \ca M_s \ar[r]^{\sim \;\;\;\;} \ar[d] & Q_{\textnormal{st}} \ar[r] \ar[d]& P\Gamma_{\theta} \setminus \mathbb{C} H^2\ar[d]\\
G \setminus X_s(\mathbb{C}) \ar[r]^{\sim \;\;\;\;} & Q_{\textnormal{st}}/\mf S_5 \ar[r] & P\Gamma \setminus \mathbb{C} H^2.
}
$$
Now the map $\widetilde{Q}_{\textnormal{st}} \to \mathbb{C} H^2$ is an isomorphism by \cite[(3.11)]{DeligneMostow}. Therefore, we are done if the composition $G \setminus \ca F_0 \to \widetilde Q \to \mathbb{C} H^2$ is nothing but the period map $\ca P: G \setminus \ca F_0 \to \mathbb{C} H^2$ that we defined in Equation (\ref{eq:periodframed}). This follows from \cite[(2.23) and (12.9)]{DeligneMostow}.
\end{proof}
\begin{proposition} \label{prop:stableperiodshyperplane}
The isomorphism (\ref{eq:isomstablefivepoints}) induces an isomorphism of complex analytic spaces
\begin{equation}
\mr M_\mathbb{C} = G \setminus X_0 \cong P\Gamma \setminus \left(\mathbb{C} H^2 - \mr H \right).
\end{equation}
\end{proposition}
\begin{proof}
We have $\overline{\ca P}(\ca F_0) \subset \mathbb{C} H^2 - \mr H$ by Proposition \ref{prop:HcorrespondsNonSimple} because the Jacobian of a smooth curve cannot contain an abelian subvariety whose induced polarization is principal. Therefore $\overline{\ca P}^{-1}(\mr H) \subset \ca F_s - \ca F_0$. Since $\ca F_s$ is irreducible (it is smooth (Corollary \ref{cor:framedquintics}) and connected (Corollary \ref{cor:surjectivemon})), $\overline{\ca P}^{-1}(\mr H)$ is a divisor; but $\ca F_s - \ca F_0$ is also a divisor (Corollary \ref{cor:framedquintics}) hence $\overline{\ca P}^{-1}(\mr H) = \ca F_s - \ca F_0$ and we are done. (Alternatively,
let $H_{0,5}$ be the moduli space of degree $5$ covers of $\mathbb{P}^1$ ramified along five distinct marked points \cite[\S 2.G]{Harris1998ModuliOC}. The period map $H_{0,5}(\mathbb{C}) \to P\Gamma \setminus \mathbb{C} H^2$, mapping a curve $C \to \mathbb{P}^1$ to its Jacobian $J(C)$ with $\mathbb{Z}[\zeta_5]$-action, extends to the stable compactification $\overline{H}_{0,5}(\mathbb{C}) \supset H_{0,5}(\mathbb{C})$ because the curves in the limit are of compact type. Since the divisor $\mr H \subset \mathbb{C} H^2$ parametrizes abelian varieties that are products of lower dimensional ones (Proposition \ref{prop:HcorrespondsNonSimple}), the image of the boundary is exactly
$P\Gamma \setminus \mr H$.)
\end{proof}
\end{document}
\subsection{Notation and assumptions}
\begin{itemize}
\item Let $K$ be a CM field of degree $2g$ over $\mathbb{Q}$ and let $F \subset K$ be its totally real subfield. Let $\mathcal{O}_K$ be the ring of integers of $K$ and denote the non-trivial element in $\textnormal{Gal}(K/F)$ by $\sigma$. Fix a set of embeddings $\Psi = \{\tau_i : K \to \mathbb{C}\}_{1 \leq i \leq g}$ such that $\Psi \cup \Psi \sigma = \{\tau_i, \tau_i \sigma\}_{1 \leq i \leq g} = \Hom(K, \mathbb{C})$.
\item Let $\Lambda$ be a free $\mathcal{O}_K$-module of rank $n+1$ equipped with a hermitian form $h: \Lambda \times \Lambda \to \mathcal{O}_K$ of signature $(r_i,s_i)$ with respect to $\tau_i$. In other words, $h$ is linear in its first argument and $\sigma$-linear in its second, and the complex vector space $\Lambda \otimes_{\mathcal{O}_K, \tau_i} \mathbb{C}$ admits a basis $\{e_i\}$ such that $(h^{\tau_i}(e_i,e_j))_{ij}$ is a diagonal matrix with $r_i$ diagonal entries equal to $1$ and $s_i$ diagonal entries equal to $-1$. Here $h^{\tau_i}$ is the hermitian form on $\Lambda \otimes_{\mathcal{O}_K, \tau_i} \mathbb{C}$ induced by $h$.
\item Consider $\tau = \tau_1: K \to \mathbb{C}$ and define $V = \Lambda \otimes_{\mathcal{O}_K, \tau} \mathbb{C}$.
We suppose that $r_1 = n$ and $r_2 = 1$.
\item For $m \in \mathbb{Z}$, define $\zeta_m = e^{2 \pi i /m} \in \mathbb{C}$. Denote by $\mu_K \subset \ca O_K^\ast$ be the group of roots of unity in $K$. If $m$ is the largest positive integer for which the $m$-th cyclotomic field $\mathbb{Q}(\zeta_m)$ can be embedded in $K$, then $\mu_K= \langle \zeta \rangle$ for a primitive $m$-th root of unity $\zeta \in K$.
Let $\Gamma = U(\Lambda)(\mathcal{O}_K) = \textnormal{Aut}_{\mathcal{O}_K}(\Lambda, h)$ and define $P\Gamma$ to be the quotient group $\Gamma/\mu_K$.
\item Define $\mathbb{C} H^n$ to be the space of negative lines in $V$.
\item For $r \in V$ and $i \in \mathbb{Z}/m$, define an isometry $h_r^i: V \to V$ of as follows:
\begin{equation*}
h^i_r(x) = x-(1-\zeta^i)\frac{h(x,r)}{h(r,r)}r.
\end{equation*}
Then $h^i_r \in \Gamma$ is called the \textit{$\zeta^i$-reflection} in $r$; we write $h_r = h_r^1$.
A norm $1$ vector of $\Lambda$ is called a \textit{short root};
then for $i \in (\mathbb{Z}/m)^\ast$, the $\zeta^i$-reflections in short roots are isometries $\Lambda$.
\item Let $\mr R \subset \Lambda$ be the set of short roots, define $H_r = \{x \in \mathbb{C} H^n: h(x,r) = 0 \}$ for $r \in \mr R$, and
\begin{equation*}
\mr H = \bigcup_{r \in \mr R} H_r \;\;\;
\subset \;\;\; \mathbb{C} H^n.
\end{equation*}
\item Remark that the family of hyperplanes $(H_r)_{r \in \mr R}$ is locally finite by \cite[Lemma 5.3]{beauvillecubicsurfaces}, so that the hyperplane arrangement $\mr H \subset \mathbb{C} H^n$ is a divisor of $\mathbb{C} H^n$.
\item Define a $\ca O_F$-linear map $\alpha: \Lambda \to \Lambda$ to be \textit{anti-unitary} if for all $x,y \in \Lambda$ and $\lambda \in \mathcal{O}_K$, one has $\alpha(\lambda x) = \sigma(\lambda)\alpha(x)$ and $h(\alpha(x), \alpha(y)) = \sigma(h(x,y)) \in \mathcal{O}_K$. Let $\mr A$ be the set of anti-unitary involutions.
It is easy to see that $\mu_K$ acts on $\mr A$ by multiplication; define $P\mr A = \mu_K \setminus \mr A$. Moreover, $P\Gamma$ acts on $P\mr A$ by conjugation; define $C\mr A = P\Gamma \setminus P\mr A$.
\item Any $\alpha \in P\mr A$ defines an anti-holomorphic involution $\alpha: \mathbb{C} H^n \to \mathbb{C} H^n$. We define $\mathbb{R} H^n_{\alpha} \subset \mathbb{C} H^n $ to be its fixed-point set. If $\alpha \in \mr A$ is a representative for $\alpha$, then
the lattice $V^\alpha$ is hyperbolic and $\mathbb{R} H^n_\alpha = \mathbb{R} H(V^\alpha)$, the space of negative real lines in $V^\alpha= \Lambda^\alpha \otimes_{\mathcal{O}_F} \mathbb{R}$. Therefore $\mathbb{R} H^n_\alpha \subset \mathbb{C} H^n$ is isometric to the $n$-dimensional real hyperbolic space $\mathbb{R} H^n$.
\item Finally, let $P\Gamma_\alpha = \textnormal{Stab}_{P\Gamma}(\mathbb{R} H^n_\alpha) \subset P\Gamma$ be the stabilizer of $\mathbb{R} H^n_\alpha$ in $P\Gamma$.
\end{itemize}
\begin{comment}
We claim that $\text{rank}_{\mathbb{Z}}(\Lambda^\alpha) = \dim_{\mathbb{R}}W^\alpha$, in other words, that $\Lambda^\alpha \otimes_{\mathbb{Z}} \mathbb{R} = W^\alpha$. Indeed, $W$ admits a complex structure for which $\alpha$ is anti-holomorphic, and for $G = \textnormal{Gal}(K/F)$ acting on $\Lambda$ via $\alpha$, the exact sequence
$$
0 \to \Lambda^\alpha \to W^\alpha \to (W/\Lambda)^\alpha \to H^1(G, \Lambda) \to 0
$$
induces an exact sequence $0 \to W^\alpha/\Lambda^\alpha \to (W/\Lambda)^\alpha \to H^1(G, \Lambda) \to 0$. Since any homomorphism $\pi: G \to H$ from a Lie group $G$ to a finite group $H$ with the discrete topology is continuous, $W^\alpha/\Lambda^\alpha$ is closed in $W/\Lambda$, hence comprealbaact, from which the claim follows.
\end{comment}
\begin{comment}
We shall be interested in the real hyperbolic orbifold
$
\mf A_\mathbb{R} := \coprod_{\alpha \in C\mr A} [P\Gamma_\alpha \setminus H^n_\alpha].
$ See also its modular interpretation (Proposition \ref{prop:stackmodelOF}).
\end{comment}
\begin{comment}
\textbf{Claim}: If a holomorphic function $\phi: \mathbb{B}^n(\mathbb{C}) \to \mathbb{B}^n(\mathbb{C})$ with $\phi \left( \mathbb{B}^n(\mathbb{R}) \right) = \mathbb{B}^n(\mathbb{R})$ is the identity on $\mathbb{B}^n(\mathbb{R})$, then $\phi$ is the identity. \textit{Proof}: Consider the holomorphic function $\psi = \phi - \textnormal{id} : \mathbb{B}^n(\mathbb{C}) \to \mathbb{C}$, then $\psi |_{\mathbb{B}^n(\mathbb{R})} = 0$ implies that $\psi = 0$, hence indeed, $\phi = \textnormal{id}$.
\end{comment}
\subsection{The gluing procedure}
We assume that the following condition is satisfied:
\begin{condition} \label{cond:orthogonal}
If $r, t \in \mr R$ are such that $H_r \neq H_t$ and $H_r \cap H_t \neq \emptyset$, then $h(r,t) = 0$.
\end{condition}
\begin{example}
Theorem \ref{th:conditionsimplyhypothesis} of Section \ref{unitaryshimura} shows that Condition \ref{cond:orthogonal} holds if the following conditions are satisfied: \textcolor{blue}{(1)} the different ideal $\mf D_K \subset \mathcal{O}_K$ is generated by a single element $\eta \in \mathcal{O}_K$ such that $\sigma(\eta) = -\eta$ and $\Im(\tau_i(\eta)) > 0$ for every $i$, \textcolor{blue}{(2)} $(r_i,s_i) = (n+1,0)$ for $2 \leq i \leq g$, and \textcolor{blue}{(3)} the CM type $(K, \Psi)$ is primitive.
Note that Condition \textcolor{blue}{(1)} is satisfied for quadratic and cyclotomic CM fields $K$, see Lemma \ref{lemma:discr}.
\end{example}
\noindent
Let $x \in \mr H \subset \mathbb{C} H^n$ and $\textbf r = (r_1, \dotsc, r_k) \in \mr R^k$ be such that $x \in H_{r_1} \cap \dotsc \cap H_{r_k}$ with $H_{r_i} \neq H_{r_j}$ for all $r_i \neq r_j$. The hyperplanes are orthogonal by Condition \ref{cond:orthogonal} hence $H_{r_1} \cap \dotsc \cap H_{r_k} \cong \bb R H^{n-k}$, a totally geodesic subspace of codimension $k$.
The $\zeta$-reflections $h_{r_i} \in P\Gamma$ define a group $ \langle h_{r_1} , \dotsc, h_{r_k} \rangle \subset \Gamma$ which is isomorphic to $(\mathbb{Z}/m)^k$, and the composition $\langle h_{r_1}, \dotsc, h_{r_k} \rangle \to \Gamma \to P\Gamma$ is injective.
Define $G(\textbf r) =\langle h_{r_1}, \dotsc, h_{r_k} \rangle \subset P\Gamma$.
\begin{definition} \label{def:conditions}
Set $\widetilde Y = \coprod_{\alpha \in P\mr A} \mathbb{R} H^n_\alpha$. Define a relation $R \subset \widetilde Y \times \widetilde Y$: an element
$\left( (x, \alpha), (y, \beta) \right) \in \widetilde Y \times \widetilde Y$ is an element of $R$ if the following conditions are satisfied:
\begin{enumerate}
\item \label{item1} We have $x = y \in \mathbb{C} H^n$.
\item \label{item2} If $\alpha \neq \beta$, then $x = y \in \mr H$ and
$\beta = \phi \circ \alpha$ with $\phi \in G(\textbf r)$ for some $\textbf r = (r_i)_i \in \mr R^k$ such that $x \in \cap_{i = 1}^k H_{r_i}$.
\end{enumerate}
\end{definition}
\begin{remark} \label{rem:gluing}
Conditions \ref{item1} and \ref{item2} in Definition \ref{def:conditions} say that we are identifying points of $\mathbb{R} H^n_\alpha \cap \mr H$ and $\mathbb{R} H^n_\beta \cap \mr H$ that have the same image in $\mathbb{C} H^n$. But we do not glue all such points: the real structures $\alpha$ and $\beta$ are required to differ by complex reflections in the hyperplanes through $x$.
In fact, we will see below (see Lemma \ref{intersection}) that the gluing rules can be rephrased as follows: we glue $\mathbb{R} H^n_\alpha$ and $\mathbb{R} H^n_\beta$ along their intersection, provided that this intersection is contained in $\mr H$ in such a way that for some (equivalently, any) $x \in \mathbb{R} H^n_\alpha \cap \mathbb{R} H^n_\beta$, the real structures $\alpha$ and $\beta$ differ by reflections in hyperplanes that pass through $x$.
\end{remark}
\begin{definition} \label{def:fullset}
Let $x \in \mathbb{C} H^n$. We say that a set of hyperplanes $\{H_{r_1}, \dotsc, H_{r_k} \}$ attached to roots $r_i \in \mr R$ is a \textit{maximal set of hyperplanes through $x$} if $x \in \cap_{i = 1}^k H_{r_i}$, $H_{r_i} \neq H_{r_j}$ for $i \neq j$ and if whenever $t \in \mr R$ is such that $x \in H_t$, then $H_r = H_{r_i}$ for some $i$.
\end{definition}
\begin{lemma} \label{eqrel}
$R$ is an equivalence relation.
\end{lemma}
\begin{proof}
Let $(x,\alpha), (y,\beta), (z,\gamma) \in \widetilde Y$; the fact that $(x, \alpha) \sim (x, \alpha)$ is clear. Suppose that $(x , \alpha) \sim (y, \beta)$. If $\alpha = \beta$ then $(x , \alpha) = (y, \beta) \in \widetilde Y$ hence $(y, \beta) \sim (x , \alpha)$. If $\alpha \neq \beta$ then $x = y \in \mr H \subset \mathbb{C} H^n$, and $\beta = \phi \circ \alpha$ for $\phi \in G(\textbf r)$ as in Definition \ref{def:conditions}.
Since $\alpha = \phi^{-1} \circ \beta$ with $\phi^{-1} \in G(\textbf r)$, this shows that $(y, \beta) \sim (x, \alpha)$.
Finally, suppose that $(x, \alpha) \sim (y, \beta)$ and $(y, \beta) \sim (z, \gamma)$; we claim that $(x, \alpha) \sim (z, \gamma)$. We may and do assume that $\alpha, \beta$ and $\gamma$ are different, which implies that $x = y = z \in \mr H$. Let $\phi, k, \textbf r$ and $\psi, l, \textbf t$ be as in Definition \ref{def:conditions}: we have $\beta = \phi \circ \alpha$ with $\phi \in G(\textbf r)$ for some $\textbf r = (r_i)_i \in \mr R^k$ such that $x \in \cap H_{r_i}$, and $\gamma = \psi \circ \beta$ with $\psi \in G(\textbf t)$ for some $\textbf t = (t_i)_i \in R^l$ such that $y \in \cap H_{t_i}$. But then $x = y \in \left(\cap_i H_{r_i}\right) \cap \left( \cap_i H_{t_i}\right)$. We may assume that the sets of hyperplanes $\{H_{r_i}\}$ and $\{H_{t_j}\}$ through $x$ and $y$ are maximal, which implies that $k = l$, and there exists an element $\sigma \in S_k$ such that $H_{t_i} = H_{r_{\sigma(i)}}$ for each $i$. By Lemma \ref{lemma:finitereflectionorders} below, we have $G(\textbf r) = G(\textbf t) \subset P\Gamma$. Therefore, $\psi \circ \phi \in G(\textbf r)$, and
we conclude that $(x, \alpha) \sim (z, \gamma)$. \end{proof}
\begin{lemma} \label{lemma:finitereflectionorders}
Let $\phi: \mathbb{C} H^n \to \mathbb{C} H^n$ be an isometry such that $\phi^m = \textnormal{id}$ and such that $\phi$ is the identity when restricted to $H_r \subset \mathbb{C} H^n$ for some $r \in \mr R$. Then $\phi = h_r^i$ for some $i \in \mathbb{Z}/m$.
\end{lemma}
\begin{proof}
It suffices to show that $\phi \in G(\textbf r)$.
Let $\bb H^n_\mathbb{C}$ be the hyperbolic space attached to the standard hermitian space $\mathbb{C}^{n,1}$ of dimension $n+1$. It is easy to see that
$$
\text{Stab}_{U(n,1)}(\bb H^{n-1}_\mathbb{C}) = U(n) \times U(1)
$$
hence any element $f \in U(n,1)$ that fixes $\bb H^{n-1}_\mathbb{C}$ pointwise is contained in the subgroup $U(1) = \{z \in \mathbb{C}^\ast : |z|^2 = 1\}$. If $f^m = \textnormal{id}$ then it is contained in the unique order $m$ subgroup of $U(1)$.
\end{proof
\begin{definition}
Define $Y$ to be the quotient of $\widetilde Y$ by the equivalence relation $R$, and equip it with the quotient topology. We shall prove (Lemma \ref{lemma:pgammaaction}) that the group $P\Gamma$ acts on $Y$. We call $P\Gamma \setminus Y$ the \textit{hyperbolic gluing} of the hermitian $\mathcal{O}_K$-lattice $(\Lambda, h)$.
\end{definition}
\begin{lemma} \label{lemma:pgammaaction}
The action of $P\Gamma$ on $\mathbb{C} H^n$ induces an action of $P\Gamma$ on $\widetilde Y$ which is compatible with the equivalence relation $R$, so that $P\Gamma$ acts on $Y$. Moreover, $P\Gamma \setminus \widetilde Y = \coprod_{\alpha \in C\mr A} P\Gamma \setminus \mathbb{R} H^n_\alpha$.
\end{lemma}
\begin{proof}
If $\phi \in P\Gamma$, then $
\phi \left( \mathbb{R} H^n_\alpha \right) = \mathbb{R} H^n_{\phi \alpha \phi^{-1}}$ hence $P\Gamma$ acts on $\widetilde Y = \coprod_{\alpha \in P\mr A} \mathbb{R} H^n_\alpha$, and
$$
P\Gamma \setminus \widetilde Y = P\Gamma \setminus \coprod_{\alpha \in P\mr A} \mathbb{R} H^n_\alpha = \coprod_{\alpha \in C\mr A} P\Gamma \setminus \mathbb{R} H^n_\alpha.
$$
Now suppose that $(x, \alpha) \sim (y, \beta) \in \widetilde Y$ and $f \in P\Gamma$. We claim that $(f(x), f \alpha f^{-1}) \sim (f(y), f \beta f^{-1})$. We may and do assume that $(x, \alpha) \neq (y, \beta)$, hence $x = y \in \mr H \subset \mathbb{C} H^n$. Let $r= (r_i)_i \in \mr R^k$ with $x \in \cap_i H_{r_i}$ and $\beta = \phi \circ \alpha$ as in Definition \ref{def:conditions}. Observe that
$f(H_{r_i}) = H_{f(r_i)}$. Therefore $f(x) \in \cap_i H_{f(r_i)}$,
and we have $
f \beta f^{-1}
= f\phi f^{-1} \circ f \alpha f^{-1}$ and $f \phi f^{-1} \in f G(\textbf r) f^{-1} = \langle f h_{r_i} f^{-1} \rangle = \langle h_{f(r_i)} \rangle = G(f \textbf r)$.
\end{proof}
We are now in position to state the main theorem of Section \ref{gluing}:
\begin{theorem} \label{theorem1}
\begin{enumerate}
\item
The hyperbolic gluing $P\Gamma \setminus Y$ admits a metric that makes it a complete path metric space. With respect to it, the natural map $P\Gamma \setminus Y \to P\Gamma \setminus \mathbb{C} H^n$ is a local isometry. Each point $x \in P\Gamma \setminus Y$ admits an open neighborhood $U \subset P\Gamma \setminus Y$ which is isometric to the quotient of an open subset $V \subset \mathbb{R} H^n$ by a finite group of isometries. Hence $P\Gamma \setminus Y$ has a real hyperbolic orbifold structure.
\item One has $O:= \coprod_{\alpha \in C\mr A} [P\Gamma_\alpha \setminus \left(\mathbb{R} H^n_\alpha - \mr H \right)] \subset P\Gamma \setminus Y$ as an open suborbifold.
\item The connected components of the real-hyperbolic orbifold $P\Gamma \setminus Y$ are uniformized by $\mathbb{R} H^n$: for each component $C \subset P\Gamma \setminus Y$ there exists a lattice $P\Gamma_C \subset \textnormal{PO}(n,1)$ and an isomorphism of real hyperbolic orbifolds $C \cong [P\Gamma_C \setminus \mathbb{R} H^n]$. In other words,
\begin{equation*}
P\Gamma \setminus Y \cong \coprod_{C\in \pi_0(P\Gamma\setminus K)} \left[ P\Gamma_C \setminus \mathbb{R} H^n \right].
\end{equation*}
\end{enumerate}
\end{theorem}
\noindent
The remaining part of Section \ref{realshimura} will be devoted to the proof of Theorem \ref{theorem1}. It may very well happen that $P\Gamma \setminus Y$ is connected: such is the case when $K = \mathbb{Q}(\zeta_3)$ and $h = \text{diag}(1,1,1,1,-1)$, see \cite{realACTsurfaces}. When $K = \mathbb{Q}(\zeta_3)$ and $h = \text{diag}(1,1,1,-1)$, then $P\Gamma \setminus Y$ has two components, see \cite[Remark 6]{realACTbinarysextics}. See also Section \ref{realperiodmaps}: for some $d$ and $n$, there is a homeomorphism between $P\Gamma \setminus Y$ and the moduli space of GIT-stable hypersurfaces of degree $d$ in $\mathbb{P}^n_\mathbb{R}$, restricting to an isomorphism of orbifolds between $O$ and the moduli of smooth hypersurfaces (Theorem \ref{th:realstableperiod}).
\subsection{The path metric on the hyperbolic gluing} \label{prooftheorem}
We start with a lemma. We will need it in the proof of Lemma \ref{localisometry} below, which will in turn be used to define a path metric on $P\Gamma \setminus Y$ making it locally isometric to quotients of $\mathbb{R} H^n$ by finite groups of isometries. But it can also serve as a sanity check: it says that once there exists an element $x \in \mathbb{R} H^n_\alpha \cap \mathbb{R} H^n_\beta$ such that $(x, \alpha) \sim (x, \beta)$, then one glues the entire copy $\mathbb{R} H^n_\alpha$ to the copy $\mathbb{R} H^n_\beta$ along their intersection in $\mathbb{C} H^n$.
\begin{lemma} \label{intersection}
\begin{enumerate}
\item Let $\phi = \prod_{i = 1}^l h_{r_i}^{j_i} \in \Gamma$ for some set $\{r_i\} \subset \mr R$ of mutually orthogonal short roots $r_i$, where $j_i \neq 0 \mod m$ for each $i$. Then $\left(\mathbb{C} H^n\right)^\phi \subset \cap_{i = 1}^l H_{r_i}$.
\item
Let $\alpha, \beta \in P\mr A$ and $x \in \mathbb{R} H^n_\alpha \cap \mathbb{R} H^n_\beta$ such that $(x, \alpha) \sim (x, \beta)$. Then $(y, \alpha) \sim (y, \beta)$ for every $y \in \mathbb{R} H^n_\alpha \cap \mathbb{R} H^n_\beta$.
\item The natural map $\widetilde Y \to \mathbb{C} H^n$ descends to a $P\Gamma$-equivariant map $\mr P: Y \to \mathbb{C} H^n$.
\end{enumerate}
\end{lemma}
\begin{proof}
1. Let $y \in V$ be representing an element in $\left(\mathbb{C} H^n\right)^\phi$. Since the $r_i$ are orthogonal, and $\phi(y) = \lambda$ for some $\lambda \in \mathbb{C}^\ast$, we have
\begin{align}
\phi(y) = \prod_{i = 1}^l h_{r_i}^{j_i}(y) = y - \sum_{i = 1}^l \left(1 - \zeta^{j_i}\right) h(y, r_i)r_i = \lambda y,
\end{align}
hence $(1-\lambda)y = \sum_{i = 1}^l \left(1 - \zeta^{j_i}\right) h(y, r_i)r_i \in V$. But $y$ spans a negative definite subspace of $V$ while the $r_i$ span a positive definite subspace, so that we must have $1 - \lambda = 0 = \sum_{i = 1}^l \left(1 - \zeta^{j_i}\right) h(y, r_i)r_i$. Since the $r_i$ are mutually orthogonal, they are linearly independent; since $\zeta^{j_i} \neq 1$ we find $h(y, r_i) = 0$ for each $i$.
\\
\\
2. Let $\textbf r = (r_1, \dotsc, r_k) \in \mr R^k$ such that $x \in \cap_i H_{r_i}$ for a maximal set of hyperplanes $\{H_{r_i}\}$ through $x$. We have $\beta = \phi \circ \alpha$ for some $\phi \in G(\textbf r)$. Let $I \subset \{1, \dotsc, k\}$ be a subset such that $\phi = \prod_{i \in I} h_{r_i}^{j_i}$ with $j_i \neq 0 \mod m$. If $I = \emptyset$ then there is nothing to prove, so suppose the contrary. Notice that
$\phi(y) = y$.
Part 1 implies that $y \in \cap_{i \in I} H_{r_i}$. Now let $\textbf t = (t_1, \dotsc, t_l) \in \mr R^l$ such that $y \in \cap_j H_{t_j}$ as in Definition \ref{def:conditions}. Then for each $i$ there is a $j$ such that $H_{r_i} = H_{t_j}$. By Lemma \ref{lemma:finitereflectionorders}, we have $h_{r_i} \in G(t_j)$, so that $
\phi \in G( \textbf{r} ) \subset G(\textbf t).
$
\\
\\
3. If $(x, \alpha) \sim (y, \beta)$, then $x = y \in \mathbb{C} H^n$.
\end{proof}
Consequently, we obtain continuous maps $\mr P : K \to \mathbb{C} H^n$ and $\bar{\mr P}: P\Gamma \setminus Y \to P\Gamma \setminus \mathbb{C} H^n$. Our next goal is to prove that each point $x \in Y$ has a neighbourhood $V \subset Y$ that maps homeomorphically onto a finite union $\cup_{i = 1}^l \mathbb{R} H^n_{\alpha_i} \subset \mathbb{C} H^n$. Hence $x$ has an open neighourhood $x \in U \subset V$ that identifies with an open set in a union of copies of $\mathbb{R} H^n$ in $\mathbb{C} H^n$ under the map $\mr P$. This allows us to define a metric on $Y$ by pulling back the metric on $\mathbb{C} H^n$.
\\
\\
Fix a point $f \in Y$ and a point $(x, \alpha) \in \widetilde Y$ lying above $f$. Let $\alpha_1, \dotsc, \alpha_l$ be the elements in $P\mr A$ such that $(x, \alpha_i) \sim (x, \alpha)$ for each $i$. Let $p: \widetilde Y \to Y$ be the quotient map. Define
\begin{equation} \label{Kf}
Y_f = p\left(\coprod_{i = 1}^l \mathbb{R} H^n_{\alpha_i}\right) \subset Y.
\end{equation}
\begin{lemma} \label{lemma:compactfinite}
Each compact set $Z \subset \mathbb{C} H^n$ meets only finitely many $\mathbb{R} H^n_\alpha$, $\alpha \in P\mr A$.
\end{lemma}
\begin{proof}
If $P\Gamma' \subset \textnormal{Isom}(\mathbb{C} H^n)$ is the group of all $\mu_K$-orbits of isometries and anti-isometries $\phi: \Lambda \to \Lambda$, then $P\Gamma'$ acts properly discontinuously on $\mathbb{C} H^n$. So if $S$ is the set of $\alpha \in P\mr A$ such that $\alpha Z \cap Z \neq \emptyset$, then $S$ is finite. In particular, $Z$ meets only finitely many $\mathbb{R} H^n_\alpha$.
\end{proof}
It follows that $Y$ is locally isometric to opens in unions of real hyperbolic subspaces of $\mathbb{C} H^n$:
\begin{lemma} \label{localisometry}
\begin{enumerate}
\item The set $Y_f$ is closed in $Y$.
\item We have $\mr P\left( Y_f \right) = \cup_{i = 1}^l \mathbb{R} H^n_{\alpha_i} \subset \mathbb{C} H^n$, and the map
\begin{equation*}
\mr P_f: Y_f \to \cup_{i = 1}^l \mathbb{R} H^n_{\alpha_i}
\end{equation*}
induced by $\mr P$ is a homeomorphism.
\item The set $Y_f \subset Y$ contains an open neighborhood $U_f$ of $f$ in $Y$.
\end{enumerate}
\end{lemma}
\begin{proof}
1. Indeed, one has
\begin{equation*}
p^{-1} \left( Y_f \right) = p^{-1} \left( p\left(\coprod_{i = 1}^l \mathbb{R} H^n_{\alpha_i}\right) \right) = \bigcup_{i = 1}^l p^{-1} \left( p\left( \mathbb{R} H^n_{\alpha_i}\right) \right)
\end{equation*}
so that it suffices to show that $p^{-1} \left( p\left( \mathbb{R} H^n_{\alpha_i}\right) \right)$ is closed in $Y$. But notice that $(x,\beta) \in p^{-1} \left( p\left( \mathbb{R} H^n_{\alpha}\right) \right)$ if and only if $x \in \mathbb{R} H^n_\alpha$ and $(x, \alpha) \sim (x, \beta)$, which implies (Lemma \ref{intersection}) that $\mathbb{R} H^n_\alpha \cap \mathbb{R} H^n_\beta \subset p^{-1} \left( p\left( \mathbb{R} H^n_{\alpha}\right) \right)$. Hence for any $\alpha \in P\mr A$, one has
$
p^{-1} \left( p\left( \mathbb{R} H^n_{\alpha}\right) \right) = \coprod_{\beta \sim \alpha} \mathbb{R} H^n_\alpha \cap \mathbb{R} H^n_\beta$, where $\beta \sim \alpha$ if and only if there exists $x \in \mathbb{R} H^n_\alpha \cap \mathbb{R} H^n_\beta$ such that $(x, \alpha) \sim (x, \beta)$. It follows that $p^{-1} \left( p\left( \mathbb{R} H^n_{\alpha}\right) \right) \cap \mathbb{R} H^n_\beta$ is closed in $\mathbb{R} H^n_\beta$ for every $\beta \in P\mr A$. But the $\mathbb{R} H^n_\beta$ are open in $\widetilde Y$ and cover $\widetilde Y$, so that $p^{-1} \left( p\left( \mathbb{R} H^n_{\alpha}\right) \right)$ is closed in $\widetilde Y$.
\\
\\
2. We have
\begin{equation*}
\mr P_f(Y_f) = \mr P\left( p\left(\coprod_{i = 1}^l \mathbb{R} H^n_{\alpha_i}\right)\right) = \tilde{\mr P} \left( \coprod_{i = 1}^l \mathbb{R} H^n_{\alpha_i}\right) = \cup_{i = 1}^l \mathbb{R} H^n_{\alpha_i} \subset \mathbb{C} H^n.
\end{equation*}
To prove injectivity, let $(x, \alpha_i), (y, \alpha_j) \in \widetilde Y$ and suppose that $x = y \in \mathbb{C} H^n$. Then indeed, $(x, \alpha_i) \sim (y, \alpha_j)$ because $\sim$ is an equivalence relation by Lemma \ref{eqrel}.
\\
\\
Let $Z \subset \mathbb{C} H^n$ be a compact set. Write $\tilde {\mr P}: \widetilde Y \to \mathbb{C} H^n$ for the canonical map.
Remark that $Z$ meets only finitely many of the $\mathbb{R} H^n_\alpha$ for $\alpha \in P\mr A$, see Lemma \ref{lemma:compactfinite}. Each $Z \cap \mathbb{R} H^n_\alpha$ is closed in $Z$ since $\mathbb{R} H^n_\alpha$ is closed in $\mathbb{C} H^n$, so each $Z \cap \mathbb{R} H^n_\alpha$ is compact. We conclude that ${\tilde{\mr P}}^{-1}(Z) = \coprod Z \cap \mathbb{R} H^n_\alpha$ is compact. In particular, $\tilde{\mr P}$ is closed \cite[Theorem A.57]{MR2954043}.
\\
\\
Finally, we prove that $\mr P_f$ is closed. Let $Z \subset Y_f$ be a closed set. Then $Z$ is closed in $Y$ by part 1, hence $p^{-1}(Z)$ is closed in $\widetilde Y$, hence $\tilde{ \mr P} \left( p^{-1}(Z) \right) )$ is closed in $\mathbb{C} H^n$, so that
\begin{equation*}
\mr P_f(Z) = \mr P(Z) = \tilde{\mr P} \left( p^{-1} \left( Z \right) \right) = \left(\tilde{\mr P} \left( p^{-1} \left( Z \right) \right)\right) \cap \left(\cup_{i = 1}^l \mathbb{R} H^n_{\alpha_i} \right)
\end{equation*}
is closed in $\cup_{i = 1}^l \mathbb{R} H^n_{\alpha_i} $.
\\
\\
3. Let $x = \mr P(f) \in \mathbb{C} H^n$. Since $\mathbb{C} H^n$ is locally compact, there exists a compact set $Z \subset \mathbb{C} H^n$ and an open set $U \subset \mathbb{C} H^n$ with $x \in U \subset Z$. Since $Z$ is compact, it meets only finitely many of the $\mathbb{R} H^n_\beta \subset \mathbb{C} H^n$ (Lemma \ref{lemma:compactfinite}). Consequently, the same holds for $U$; define $V = \mr P^{-1}(U) \subset Y$. Let $\mr B = \{\beta \in P\mr A: U \cap \mathbb{R} H^n_\beta \neq \emptyset \}$. Also define, for $\beta \in P\mr A$, $H_\beta = p\left(\mathbb{R} H^n_\beta \right) \subset Y$. Then
\begin{equation*}
f \in V \subset \bigcup_{\beta \in \mr B} H_\beta = \bigcup_{\substack{\beta \in \mr B \\ \beta(x) = x}} H_\beta \bigcup_{\substack{\beta \in \mr B \\ \beta(x) \neq x}} H_\beta.
\end{equation*}
Since each $H_\beta$ is closed in $Y$ by the proof of part 1, there is an open subset $V' \subset V$ with
\begin{equation*}
f \in V' \subset \bigcup_{\substack{\beta \in \mr B \\ \beta(x) = x}} H_\beta =
\bigcup_{\substack{\beta \in \mr B \\ \beta(x) = x \\ (x, \beta) \sim (x, \alpha)}} H_\beta
\bigcup_{\substack{\beta \in \mr B \\ \beta(x) = x \\ (x, \beta) \not \sim (x, \alpha)}} H_\beta
\end{equation*}
Hence again there exists an open subset $V'' \subset V'$ with
\begin{equation*}
f \in V'' \subset \bigcup_{\substack{\beta \in \mr B \\ \beta(x) = x \\ (x, \beta) \sim (x, \alpha)}} H_\beta \subset
\bigcup_{\substack{\beta \in P \mr A \\ \beta(x) = x \\ (x, \beta) \sim (x, \alpha)}} H_\beta = Y_f.
\end{equation*}
Therefore, $U_f:=V'' \subset Y$ satisfies the requirements.
\end{proof}
We need one further lemma:
\begin{lemma}
The topological space $Y$ is Hausdorff.
\end{lemma}
\begin{proof}
Let $f , g \in Y$ be elements such that $f \neq g$. First suppose that $f \not \in Y_g$. Since $Y_g$ is closed in $Y$ by Lemma \ref{localisometry}, there is an open neighbourhood $U$ of $f$ such that $U \cap U_g \subset U \cap Y_g = \emptyset$. Next, suppose that $f \in Y_g$. Lift $f$ and $g$ to elements $(x,\alpha), (y,\beta) \in \widetilde Y$. Assume first that $x = y$. This means that $\mr P(f) = \mr P(g)$. Since $\mr P: Y_g \to \mathbb{C} H^n$ is injective, this implies that $f = g$, contradiction. So we have $x \neq y \in \mathbb{C} H^n$. But $\mathbb{C} H^n$ is Hausdorff, so there are open subsets $U \subset \mathbb{C} H^n$, $V \subset \mathbb{C} H^n$ such that $x \in U$, $y \in V$ and $U \cap V = \emptyset$. Then $\mr P^{-1}(U) \cap \mr P^{-1}(V) = \emptyset$.
\end{proof}
We then obtain:
\begin{proposition} \label{metric}
$Y$ is naturally a path metric space which is piecewise isometric to $\mathbb{R} H^n$.
\end{proposition}
\begin{proof}
We conclude that for each $f \in Y$ there exists an open neighborhood $f \in U_f \subset Y$ such that $\mr P$ induces a homeomorphism $Y \supset U_f \xrightarrow{\sim} \mr P(U_f) \subset \mathbb{C} H^n$. Pull back the metric on $\mr P(U_f)$ to obtain a metric on $U_f$. Then define a metric on $Y$ as the largest metric which is compatible with the metric on each open set $U_f$ and which preserves the lengths of paths.
\end{proof}
\begin{proposition} \label{prop:pathmetricquotient}
The path metric on $Y$ descends to a path metric on $P\Gamma \setminus Y$.
\end{proposition}
\begin{proof}
The metric on $Y$ descends in any case to a pseudo-metric on $P\Gamma \setminus Y$, which will be a metric in case $P\Gamma$ acts by isometries on $Y$ with closed orbits \cite[Chapter 1]{Gromov2007}.
The fact that $P\Gamma$ acts isometrically on $Y$ comes from the $P\Gamma$-equivariance of $\mr P: Y \to \mathbb{C} H^n$ (Lemma \ref{intersection}) together with the construction of the metric on $Y$ (Proposition \ref{metric}). To check that the $P\Gamma$-orbits are closed in $Y$, let $f \in Y$ with representative $(x, \alpha) \in \widetilde Y$. By equivariance of $p: \widetilde Y \to Y$, we have $p^{-1} \left( P\Gamma \cdot f \right) = P \Gamma \cdot \left(p^{-1}f \right)$, so since $p$ is a quotient map, it suffices to show that
$
P \Gamma \cdot \left(p^{-1}f \right) = P\Gamma \cdot \cup_{(x, \beta) \sim (x, \alpha)}(x,\beta) = \cup_{(x, \beta) \sim (x, \alpha)} P\Gamma \cdot (x, \beta)
$
is closed in $\widetilde Y$, thus that each orbit $P\Gamma \cdot (x, \beta)$ is closed in $\widetilde Y$.
Since $P\Gamma$ is discrete, it suffices to show that $P\Gamma$ acts properly on $\widetilde Y$. So let $Z \subset \widetilde Y$ be any compact set: we claim that $\{g \in P\Gamma: g Z \cap Z \neq \emptyset \}$ is a finite set. Indeed, for each $g \in P\Gamma$, one has $\tilde{\mr P} \left( g Z \cap Z \right) \subset g \tilde{\mr P}(Z) \cap \tilde{\mr P}(Z)$, and the latter is non-empty for only finitely many $g \in P\Gamma$, by properness of the action of $P\Gamma$ on $\mathbb{C} H^n$.
Since the metric on $Y$ is a path metric, so is the metric on $P\Gamma \setminus Y$ \textit{[loc. cit.]}.
\end{proof}
\subsection{The orbifold structure on the hyperbolic gluing}
The next step is to prove that $P\Gamma \setminus Y$ is locally isometric to quotients of open sets in $\mathbb{R} H^n$ by finite groups of isometries.
\\
\\
Let $f \in Y$ with representative $ (x, \alpha) \in \widetilde Y$ as before. Let $\textbf r = (r_1, \dotsc, r_k) \in \mr R^k$ be a vector of short roots such that $x \in \cap_{i = 1}^kH_{r_i}$, $H_{r_i} \neq H_{r_j}$ for $i \neq j$ and $k$ the order of $x$. We say that \textit{$f$ has $k$ nodes}, and call the $H_{r_i}$ the \textit{nodes of $f$}. Note that the anti-unitary involution $\alpha: \mathbb{C} H^n \to \mathbb{C} H^n$ interchanges the hyperplanes $H_{r_i}$ because $\alpha(x) = x$, $k$ is the order of $x$ and $\alpha H_{r_i} = H_{\alpha(r_i)}$.
If $\alpha(H_{r_i}) = H_{r_i}$ then we call $H_{r_i}$ a \textit{real node of $f$}. If $\alpha(H_{r_i}) = H_{r_j}$ for some $j$ with $H_{r_i} \neq H_{r_j}$ then we call the pair $(H_{r_i}, H_{r_j})$ a \textit{a pair of complex conjugate nodes of $f$}. Write $k = 2a + b$, with $a$ the number of pairs of complex conjugate nodes of $f$, and $b$ the number of real nodes of $f$. We relabel the $r_i$ so that they satisfy the following condition: $\alpha(H_{r_i}) = H_{r_{i+1}}$ for $i$ odd and $i \leq 2a$, $\alpha(H_{r_i}) = H_{r_{i-1}}$ for $i$ even and $i \leq 2a$, and $\alpha(H_{r_i}) = H_{r_{i}}$ for $i > 2a$.
\begin{lemma} \label{lemma:localcoordinates}
Let $f \in Y$ with representative $(x, \alpha) \in \widetilde Y$ be as above.
\begin{enumerate}
\item If $\beta \in P\mr A$ is such that $(x, \beta) \sim (x, \alpha)$, then $\beta = \prod_{i = 1}^a \left( h_{r_{2i-1}} \circ h_{r_{2i}} \right)^{j_i} \circ \prod_{i = 2a+1}^k h_{r_i}^{j_i} \circ \alpha$ for some $j_1, \dotsc, j_a, j_{2a+1}, \dotsc, j_k \in \mathbb{Z}$. In particular, there are $m^{a + b}$ such $\beta$.
\item There is an isometry $\mathbb{C} H^n \xrightarrow{\sim} \mathbb{B}^n(\mathbb{C})$ identifying $x$ with the origin, $h_{r_i}$ with \begin{equation*}
\mathbb{B}^n(\mathbb{C}) \to \mathbb{B}^n(\mathbb{C}), \;\;\; (t_1, \dotsc,t_i, \dotsc, t_n) \mapsto (t_1, \dotsc, \zeta t_i, \dotsc, t_n)
\end{equation*}
and $\alpha$ with the map defined by
\begin{equation} \label{eq:alpha}
t_i \mapsto
\begin{cases}
\bar t_{i+1} \;\;\; \textnormal{for $i$ odd and $i \leq 2a$}\\
\bar t_{i-1} \;\;\; \textnormal{for $i$ even and $i \leq 2a$}\\
\bar t_i \;\;\;\white \textnormal{for $i > 2a$}.
\end{cases}
\end{equation}
\end{enumerate}
\end{lemma}
\begin{proof}
We have $\beta = \phi \circ \alpha$ for $\phi \in G( \textbf r)$ by Definition \ref{def:conditions}; write $\beta = \prod_{i=1}^k h_{r_i}^{n_i} \circ \alpha$. But then
\begin{equation*}
\beta^2 = \left(\prod_{i=1}^k h_{r_i}^{n_i} \circ \alpha\right)\circ \left(\prod_{i=1}^k h_{r_i}^{n_i} \circ \alpha\right) = \prod_{i=1}^k h_{r_i}^{n_i} \circ \prod_{i=1}^k h_{\alpha(r_i)}^{-n_i} = \prod_{i=1}^k h_{r_i}^{n_i} \circ h_{\alpha(r_i)}^{-n_i} = \prod_{i = 1}^{2a}h_{r_i}^{n_i} \circ h_{\alpha(r_i)}^{-n_i}
\end{equation*}
\begin{equation*}
= \prod_{i = 1, odd}^{2a}h_{r_i}^{n_i} \circ h_{r_{i+1}}^{-n_i}
\prod_{i = 2, even}^{2a}h_{r_i}^{n_i} \circ h_{r_{i-1}}^{-n_i}
=
\prod_{i=1, odd}^{2a} h_{r_i}^{n_i} \circ h_{r_i}^{-n_{i+1}} \circ
\prod_{i=2, even}^{2a} h_{r_i}^{n_i} \circ h_{r_i}^{-n_{i-1}}
\end{equation*}
\begin{equation*}
= \textnormal{id} \;\;\; \iff \;\;\; n_i = n_{i+1} \textnormal{ for $i \leq 2a$ odd, and } n_i = n_{i-1} \textnormal{ for $i \leq 2a$ even}
\end{equation*}
$\iff n_i = n_{i+1}$ for $i \leq 2a$ odd, which proves part 1. For part 2, since the $H_{r_i}$ are orthogonal by Condition \ref{cond:orthogonal}, and their intersection contains $x$, we can find coordinates $t_0, \dotsc, t_n$ on $V$ that identify $(V,h)$ with $\mathbb{C}^{n+1}$ with the hermitian form $h(x,x) = \sum_{i = 0}^{n-1}|x_i|^2-|x_n|^2$, such that $H_{r_i} \subset V$ is identified with the hyperplane $\{t_i = 0\} \subset \mathbb{C}^{n+1}$ and $x \in \cap_i H_{r_i}$ with the point $(0,0, \dotsc, 0, 1)$. But the decomposition $V = \langle x \rangle \oplus_{i = 1}^k H_{r_i} \oplus W$ where $W$ is the orthogonal complement of the left hand side is such that $\alpha \left(\langle x \rangle \oplus_{i = 1}^k H_{r_i} \right) = \left(\langle x \rangle \oplus_{i = 1}^k H_{r_i} \right)$ hence $\alpha W = W$ as well. This implies that $\alpha$ is in the coordinate system $t_i$ given by a matrix
\begin{equation*}
\alpha = \begin{pmatrix}
\alpha_0 & 0 & \ldots & 0&&&&\\
0 & 0 & \alpha_1 & 0&&&&\\
0 & \alpha_2 & 0 & 0&&&&\\
0 & 0 & 0 & 0&&&&\\
\vdots & 0 & \ddots & \vdots&&&&\\
\vdots & \vdots & 0 & \vdots&&&\alpha_{n-1}& \\
0 & 0 &\ldots & 0 &&&&\alpha_n
\end{pmatrix}
\end{equation*}
where every $\alpha_i$ is an anti-holomorphic involution $\mathbb{C} \to \mathbb{C}$, and where $\alpha_i = \alpha_{i+1}$ for $i < 2a$. For $i = 0, \dotsc, n$ we then choose $\phi_i \in \mathbb{C}^\ast$ such that $\phi_i\alpha_i\phi_i^{-1} = $ complex conjugation on $\mathbb{C}$; the matrix $\text{diag}(\phi_i)$ then gives an isometry $\psi: \mathbb{C}^{n+1} \to \mathbb{C}^{n+1}$ such that $\psi \alpha \psi^{-1}$ is as in Equation (\ref{eq:alpha}). The end result is an isometry $V \xrightarrow{\sim} \mathbb{C}^{n,1}$ which induces the required $\mathbb{C} H^n \xrightarrow{\sim} \mathbb{B}^n(\mathbb{C})$.
\end{proof}
Define $A_f = \textnormal{Stab}_{P\Gamma}(f)$ to be the subgroup of $P\Gamma$ fixing $f \in Y$. This contains the group $G(\textbf r) \cong (\mathbb{Z}/m)^k$. We shall also consider $B_f$, the subgroup of $G(\textbf r)$ generated by the order $m$ complex reflections associated to the real nodes of $f$, rather than all the nodes.
Hence $B_f = \langle h_{r_i} \rangle_{i > 2a} \cong (\mathbb{Z}/m)^b$. Observe that the following holds:
\begin{lemma}
The stabilizer $A_f$ of $f \in Y$ preserves the subset $Y_f \subset Y$.
\end{lemma}
\begin{proof}
Let $\psi \in A_f$, with $f = [x, \alpha] \in Y$, $x \in \cap_i H_{r_i}$. Then $( \psi(x), \psi \alpha \psi^{-1}) \sim (x, \alpha) \in Y$. Now let $(y, \beta) \in Y_f$. Then $\beta(x) = x$ and $(x, \alpha) \sim (x, \beta)$. Hence $(x, \alpha ) \sim \psi \cdot (x, \alpha) \sim \psi \cdot (x, \beta) = (\psi(x), \psi \beta \psi^{-1})$. This implies $\psi \beta \psi^{-1} \circ \alpha \in G(\textbf r)$, so that $(\psi(y), \psi \beta \psi^{-1}) \in Y_f$.
\end{proof}
We shall need the following lemma. Write $m = 2^ak$ with $k \neq 0 \mod 2$.
\begin{lemma} \label{lemma:T/G}Let $T = \{t \in \mathbb{C}: t^m \in \mathbb{R}\}$. Then $G = \langle \zeta_m \rangle$ acts on $T$ by multiplication,
and each element in $T/G$ has a unique representative of the form $\zeta_{2^{a+1}}^\epsilon \cdot r$, with $r \geq 0$ and $\epsilon \in \{0,1\}$.
\end{lemma}
\begin{proof}
Since $m$ is the largest integer such that $\mathbb{Q}(\zeta_m)$ embeds in $K$, we have $2|m$. Indeed, suppose the contrary: then $-\zeta_m = \zeta_{2m} \in K$, contradiction. Therefore, we have $a \geq 1$.
Next, observe that $t = r \zeta_{2m}^j$ for some $j \in \mathbb{Z}$ and $r \in \mathbb{R}$ if and only if $t^m \in \mathbb{R}$.
One easily shows that since $\gcd(2,k) = 1$, we have $\zeta_{2^{a+1}} \cdot \zeta_{2^ak} = (\zeta_{2^{a+1}k})^{k+2}$. Raising both sides to the power $b = (k+2)^{-1} \in (\mathbb{Z}/m)^\ast$
gives $\zeta_{2m} = \zeta_{2^{a+1}}^b \cdot \zeta_m^b$.
Consequently, $t^m \in \mathbb{R}$
if and only if $t = r \cdot \zeta_{2^{a+1}}^{bj} \cdot \zeta_{m}^{bj}$ for some $r \in \mathbb{R}$.
Finally, $\zeta_{2^{a+1}}^u \cdot \zeta_{2^{a}}^v = \zeta_{2^{a+1}}^{u + 2v}$ hence $\langle\zeta_{2^{a+1}}\rangle / \langle \zeta_{2^{a}} \rangle \cong \mathbb{Z}/2$.
\end{proof}
We obtain:
\begin{proposition} \label{localmodel}
Keep the above notations, and let $Y_f \subset Y$ be as in Equation (\ref{Kf}). \begin{enumerate}
\item If $f$ has no nodes, then $G(\textbf r) = B_f$ is trivial, and $Y_f = \mathbb{R} H^n_\alpha \cong \mathbb{B}^n(\mathbb{R})$.
\item If $f$ has only real nodes, then $B_f \setminus Y_f$ is isometric to $\mathbb{B}^n(\mathbb{R})$.
\item If $f$ has $k = 2a$ complex nodes and no other singularities, then $B_f\setminus Y_f = Y_f$ is the union of $m^a$ copies of $\mathbb{B}^n(\mathbb{R})$, any two of which meet along a $\mathbb{B}^{2c}(\mathbb{R})$ for some integer $c$ with $0 \leq c \leq a$.
\item If $f$ has $2a$ complex conjugate nodes and $b$ real nodes, then there is an isometry between $B_f \setminus Y_f$ and the union of $m^a$ copies of $\mathbb{B}^n(\mathbb{R})$ identified along common $\mathbb{B}^{2c}(\mathbb{R})'s$, that is, the set $Y_f$ of case 3 above.
\item In each case, $A_f$ acts transitively on the indicated $\mathbb{B}^n(\mathbb{R})^{'}s$. If $\mathbb{B}^n(\mathbb{R})$ is any one of them, and $\Gamma_f = (A_f/B_f)_{\mathbb{B}^n(\mathbb{R})}$ its stabilizer, then the natural map
\begin{equation*}
\Gamma_f \setminus \mathbb{B}^n(\mathbb{R}) \to \left(A_f/B_f\right) \setminus \left(B_f \setminus Y_f \right) = A_f \setminus Y_f
\end{equation*}
is an isometry of path metrics.
\end{enumerate}
\end{proposition}
\begin{proof}
1. This is clear.
\\
\\
2. Suppose then that $f$ has $k$ real nodes. Then in the local coordinates $t_i$ of Lemma \ref{lemma:localcoordinates}\textcolor{blue}{.2}, we have that $\alpha: \mathbb{B}^n(\mathbb{C}) \to \mathbb{B}^n(\mathbb{C})$ is defined by $\alpha(t_i) = \bar t_i$. Part \textcolor{blue}{1} of the same lemma shows that any $\beta \in P\mr A$ fixing $x$ such that $(x,\alpha) \sim (x, \beta)$ is of the form
\begin{equation*}
\mathbb{B}^n(\mathbb{C}) \to \mathbb{B}^n(\mathbb{C}), \;\;\; (t_1, \dotsc,t_i, \dotsc, t_n) \mapsto (\bar t_1 \zeta^{j_1}, \dotsc, \bar t_k \zeta^{j_k}, \bar t_{k+1}, \dotsc, \bar t_n).
\end{equation*}
Since $f$ has $k$ real nodes and no complex conjugate nodes, we have (writing $j = (j_1, \dotsc, j_k)$ and $\alpha_j = \prod_{i = 1}^k h_{r_i}^{j_i} \circ \alpha$):
\begin{equation*}
Y_f \cong \bigcup_{j_1, \dotsc, j_k = 1}^m \mathbb{R} H^n_{\alpha_j} \cong \left\{ (t_1, \dotsc, t_n) \in \mathbb{B}^n(\mathbb{C}): t_1^m, \dotsc, t_k^m, t_{k+1}, \dotsc, t_n \in \mathbb{R} \right\}.
\end{equation*}
Each of the $2^k$ subsets
\begin{equation*}
K_{f, \epsilon_1, \dotsc, \epsilon_k} = \left\{ (t_1, \dotsc, t_n) \in \mathbb{B}^n(\mathbb{C}): \zeta_{2^{a+1}}^{-\epsilon_1}t_1, \dotsc, \zeta_{2^{a+1}}^{-\epsilon_k}t_k \in \mathbb{R}_{\geq 0} \textnormal{ and } t_{k+1}, \dotsc, t_n \in \mathbb{R} \right\},
\end{equation*}
indexed by $\epsilon_1, \dotsc, \epsilon_k \in \{0,1\}$, is isometric to the closed region in $\mathbb{B}^n(\mathbb{R})$ bounded by $k$ mutually orthogonal hyperplanes. By Lemma \ref{lemma:T/G}, their union $U$ is a fundamental domain for $B_f$ in the sense that it maps homeomorphically and piecewise-isometrically onto $B_f \setminus Y_f$. Under its path metric, $U = \cup K_{f, \epsilon_1, \dotsc, \epsilon_k} $ is isometric to $\mathbb{B}^n(\mathbb{R})$ by the following map:
\begin{equation*}
U \to \mathbb{B}^n(\mathbb{R}), \;\;\; (t_1, \dotsc, t_k) \mapsto \left((-\zeta_{2^{a+1}})^{-\epsilon_1}t_1, \dotsc, (-\zeta_{2^{a+1}})^{-\epsilon_k}t_k, t_{k+1}, \dotsc, t_n\right).
\end{equation*}
This identifies $B_f \setminus Y_f$ with the standard $\mathbb{B}^n(\mathbb{R}) \subset \mathbb{B}^n(\mathbb{C})$.
\\
\\
3. Now suppose $f$ has $k = 2a$ nodes $H_{r_1}, \dotsc, H_{r_{2a}}$. There are now $m^{a}$ anti-isometric involutions $\alpha_{j_i}$ fixing $x$ and such that $(x, \alpha_{j_i}) \sim (x, \alpha)$: they are given in the coordinates $t_i$ as follows, taking $j = (j_1, \dotsc, j_a) \in (\mathbb{Z}/m)^a$:
\begin{equation*}
\alpha_j: (t_1, \dotsc, t_n) \mapsto (\bar t_2 \zeta^{j_1}, \bar t_1 \zeta^{j_1}, \dotsc, \bar t_{2a} \zeta^{j_{a}}, \bar t_{2a-1} \zeta^{j_a}, \bar t_{2a+1}, \dotsc, \bar t_n).
\end{equation*}
So any fixed-point set $\mathbb{R} H^n_{\alpha_j}$ is identified with
\begin{equation*}
\mathbb{B}^n(\mathbb{R})_{\alpha_j} =
\left\{ (t_1, \dotsc, t_n) \in \mathbb{B}^n(\mathbb{C}): t_{i} = \bar t_{i-1} \zeta^{j_i} \textnormal{ for $1 \leq i \leq 2a $ even, } t_i \in \mathbb{R} \textnormal{ for $i > 2a$.} \right\}
\end{equation*}
All these $m^a$ copies of $\mathbb{B}^n(\mathbb{R})$ meet at the origin of $\mathbb{B}^n(\mathbb{C})$; in fact, for $j \neq j'$, the space $\mathbb{B}^n(\mathbb{R})_{\alpha_{j}}$ meets the space $\mathbb{B}^n(\mathbb{R})_{\alpha_{j'}}$ in a $\mathbb{B}^{2c}(\mathbb{R})$ if $c$ is the number of pairs $(j_i, j'_i)$
with $j_i = j'_i$.
\\
\\
4. Now we treat the general case. In the local coordinates $t_i$, any anti-unitary involutions fixing $x$ and equivalent to $\alpha$ is of the form
\begin{equation*}
\alpha_j: (t_1, \dotsc, t_n) \mapsto
(\bar t_2 \zeta^{j_1}, \bar t_1 \zeta^{j_1}, \dotsc, \bar t_{2a} \zeta^{j_{a}}, \bar t_{2a-1} \zeta^{j_a}, \bar t_{2a+1}\zeta^{j_{2a+1}}, \dotsc, \bar t_k \zeta^{j_{k}}, \bar t_{k+1}, \dotsc, \bar t_n)
\end{equation*}
for some $j = (j_1, \dotsc, j_a, j_{2a+1}, \dotsc, j_{k}) \in (\mathbb{Z}/m)^{a+b}$. We now have $B_f \cong (\mathbb{Z}/m)^b$ acting by multiplying the $t_i$ for $2a+1 \leq i \leq k$ by powers of $\zeta$, and there are $m^{a+b}$ anti-unitary involutions $\alpha_j$. We have $Y_f \cong $
\begin{equation*}
\bigcup_{j_1, \dotsc, j_k = 1}^m \mathbb{R} H^n_{\alpha_j} \cong
\left\{ (t_1, \dotsc, t_n) \in \mathbb{B}^n(\mathbb{C}): t_2^{m} = \bar t_1^{m}, \dotsc, t_{2a}^{m} = \bar t_{2a-1}^{m}, t_{2a+1}^{m}, \dotsc, t_k^{m}, t_{k+1}, \dotsc, t_n \in \mathbb{R} \right\}.
\end{equation*}
We look at subsets $K_{f, \epsilon_1, \dotsc, \epsilon_k} \subset Y_f$ again, this time defined as $ K_{f, \epsilon} = K_{f, \epsilon_1, \dotsc, \epsilon_k} =$
\begin{equation*}
\left\{ (t_1, \dotsc, t_n) \in \mathbb{B}^n(\mathbb{C}):
t_i^m = \bar t_{i-1}^m \textnormal{ $i \leq 2a$ even, }
\zeta_{2^{a+1}}^{-\epsilon_i}t_i \in \mathbb{R}_{\geq 0} \textnormal{ $2a<i\leq k$, }
t_i \in \mathbb{R}, i > k \right\}.
\end{equation*}
As before, we have that the natural map
$
U:= \bigcup_{\epsilon} K_{f, \epsilon} \to B_f \setminus Y_f
$ is an isometry. Define
\begin{equation*}
\widetilde Y_f =
\left\{ (t_1, \dotsc, t_n) \in \mathbb{B}^n(\mathbb{C}):
t_i^m = \bar t_{i-1}^m \textnormal{ for $i \leq 2a$ even, }
t_i \in \mathbb{R}, \textnormal{ for } i > 2a \right\}.
\end{equation*}
Under its path metric, $U = \cup_\epsilon K_{f, \epsilon_1, \dotsc, \epsilon_k}$ is isometric to $\widetilde Y_f$
by the following map:
\begin{equation*}
U \to \widetilde Y_f, \;\;\;
(t_1, \dotsc, t_k) \mapsto \left(t_1, \dotsc, t_{2a}, (-\zeta_{2^{a+1}})^{-\epsilon_1}t_{2a+1}, \dotsc, (-\zeta_{2^{a+1}})^{-\epsilon_k}t_k, t_{k+1}, \dotsc, t_n\right).
\end{equation*}
Hence $B_f \setminus Y_f \cong \widetilde Y_f$; but since $\widetilde Y_f$ is what $Y_f$ was in case \textcolor{blue}{3}, we are done.
\\
\\
5. The transitivity of $A_f$ on the $\mathbb{B}^n(\mathbb{R})^{'}s$ follows from the fact that $G(\textbf r) \subset A_f$ contains transformations multiplying $t_1, \dotsc, t_{2a}$ by powers of $\zeta$, hence $t_i \mapsto \zeta^ut_i, t_{i-1} \mapsto t_{i-1}$ maps those $t_{i-1}, t_i$ with $t_i = \bar t_{i-1}\zeta^{j_i}$ to those $t_{i-1}, t_i$ with $t_i = \bar t_{i-1}\zeta^{j_1+u}$. So if $B$ is any one of the $\mathbb{B}^n(\mathbb{R})^{'}s$ and $G = (A_f/B_f)_H$ its stabilizer, then it remains to prove that $G \setminus B \to A_f \setminus Y_f$ is an isometry. Surjectivity follows from the transitivity of $A_f$ on the $\mathbb{B}^n(\mathbb{R})^{'}s$. It is a piecewise isometry so we only need to prove injectivity.
\begin{lemma}
Let $X$ be a set on which a group $G$ acts, let $Y$ be a set and $\phi_i: Y \hookrightarrow X$ embeddings, write $Y_i = \phi_i(Y)$ and suppose that $X = \cup_i Y_i$. Let $Y_0 \subset X$ be any one of the $Y_i's$. Let $H \subset G$ be the stabilizer of $Y_0$.
Suppose that for all $y \in X$, the stabilizer of $y$ in $G$ acts transitively on the $Y_i's$ containing $y$.
Then $H\setminus Y_0 \to G \setminus X$ is injective.
\end{lemma}
\begin{proof}
Let $x,y \in Y_0$ and $g \in G$ such that $g \cdot x = y$. Then $y = gx \in gY_0$. Since also $y \in Y_0$, there is an element $h \in \text{Stab}_G(y)$ such that $hgY_0 = Y_0$ and $hg(x) = h(y) = y$. Let $f = hg$; then $f \in H$ and $f \cdot x = y$, which proves what we want.
\end{proof}
Now let us use the lemma: suppose that $y \in B_f \setminus Y_f$. We need to prove that $\text{Stab}_{A_f/B_f}(y)$ acts transtivitely on the $\mathbb{B}^n(\mathbb{R})'s$ containing $y$. There exists a tuple
\begin{equation*}
j = (j_1, \dotsc, j_a, j_{2a+1}, \dotsc, j_k) \in (\mathbb{Z}/m)^{a+b}
\end{equation*} such that $y = (t_1, \dotsc, t_n)$ with $t_{i} = \bar t_{i-1} \zeta^{j_i}$ for $i \leq 2a$ even, $t_{i} = \bar t_{i-1} \zeta^{j_i}$ for $2a < i \leq k$, and $t_i \in \mathbb{R}$ for $i > k$. If all $t_i$ are non-zero, then $y \in \cup_{j'} \mathbb{R} H^n_{\alpha_{j'}}$ is only contained in $\mathbb{R} H^n_{\alpha_j}$, so there is nothing to prove. Let us suppose that $t_1 = t_2 = 0$ and the other $t_i$ are non-zero. Then $y$ is contained in all the $\mathbb{R} H^n_{\alpha_{j'}}$ with $j_i' = j_i$ for $i \geq 2$; there are $m$ of them. The stabilizer of $y$ multiplies $t_1$ and $t_2$ by powers of $\zeta$ and leaves the other $t_i$ invariant; it acts transitively on the $\mathbb{R} H^n_{\alpha_{j'}}$ containing $y$ for if $t_2 = \bar t_1 \zeta^{j_1'}$ then $\zeta^{(j_1''-j_1')}t_2 = \bar t_1 \zeta^{j_1''}$. The general case is similar.
\end{proof}
Finally, we can prove Theorem \ref{theorem1}.
\begin{proof}[Proof of Theorem \ref{theorem1}]
1. The path metric on $P\Gamma \setminus Y$ is given by Proposition \ref{prop:pathmetricquotient}. Note that the map $\mr P: Y \to \mathbb{C} H^n$ is a local embedding by Lemma \ref{localisometry}, which was used to define the metric on $Y$ (Proposition \ref{metric}) - thus almost by definition, $\mr P$ is a local isometry. For each $f \in Y$ we can find a $P\Gamma_f$-invariant open neighborhood $U_f \subset Y_f \subset Y$ such that $P\Gamma_f \setminus U_f \subset P\Gamma \setminus Y$, with $U_f$ mapping bijectively onto an open subset $V_f$ in the closed subset $\mr P(Y_f) = \cup_i\mathbb{R} H^n_{\alpha_i} \subset \mathbb{C} H^n$. By $P\Gamma$-equivariance of $\mr P$, the set $V_f$ is $P\Gamma_f$-invariant, and we have $P\Gamma_f \setminus V_f \subset P\Gamma \setminus \mathbb{C} H^n$.
Thus $\bar{\mr P}: P\Gamma \setminus Y \to P\Gamma \setminus \mathbb{C} H^n$ is also a local isometry.
\\
\\
Note that the map $\mr P: Y \to \mathbb{C} H^n$ is proper because any compact set in $\mathbb{C} H^n$ meets only finitely many ${\mathbb{R} H^n_\alpha}^{'}s$, $\alpha \in P\mr A$ (Lemma \ref{lemma:compactfinite}), and $\mr P$ carries each $H_\alpha = p\left(\mathbb{R} H^n_\alpha \right)$ homeomorphically onto $\mathbb{R} H^n_\alpha$. Since $P\Gamma \setminus \mathbb{C} H^n$ is complete, the space $P\Gamma \setminus Y$ is complete as well.
\\
\\
Finally, each point $[f] \in P\Gamma \setminus Y$ has an open neighborhood isometric to the quotient of an open set $W$ in $\mathbb{R} H^n$ by a finite group of isometries $\Gamma_f$. Indeed, take $Y_f \subset Y$ as in Equation (\ref{Kf}), and $f \in U_f \subset Y_f$ as in Lemma \ref{localisometry}\textcolor{blue}{.2}. We let $A_f = P\Gamma_f$ be the stabilizer of $f$ in $P\Gamma$ as before, and take an $A_f$-equivariant open neighborhood $V_f \subset U_f$ such that $A_f \setminus V_f \subset P\Gamma \setminus Y$. By Proposition \ref{localmodel}\textcolor{blue}{.5}, we know that $A_f \setminus Y_f$ is isometric to $\Gamma_f \setminus \mathbb{R} H^n$ for some finite group of isometries of $\mathbb{R} H^n$. This implies that $A_f \setminus V_f$ is isometric to some open set $W'$ in $\Gamma_f \setminus \mathbb{R} H^n$. Take $W \subset \mathbb{R} H^n$ to be the preimage of $W'$.
But for any path metric space $X$ locally isometric to quotients of $\mathbb{R} H^n$ by finite groups of isometries, there is a unique real-hyperbolic orbifold structure on $X$ whose path metric is the given one,
for if $U$ and $U'$ are connected open subsets of $\mathbb{R} H^n$ and $\Gamma$ and $\Gamma'$ finite groups of isometries of $\mathbb{R} H^n$ preserving $U$ and $U'$ respectively, then any isometry $\bar \phi: \Gamma\setminus U \to \Gamma' \setminus U'$ extends to an isometry $\phi: \mathbb{R} H^n \to \mathbb{R} H^n$ such that $\phi(U) = U'$ and $\phi \Gamma{\phi}^{-1} = \Gamma' \subset \textnormal{Isom}(\mathbb{R} H^n)$.
We conclude that $P\Gamma \setminus Y$ is naturally a real hyperbolic orbifold.
\\
\\
2. Let us show that $O = \coprod_{\alpha \in C\mr A} [P\Gamma_\alpha \setminus \left(\mathbb{R} H^n_\alpha - \mr H \right)] \subset P\Gamma \setminus Y$ as a hyperbolic suborbifold. It suffices to show that for those $f = [x,\alpha] \in Y$ that have no nodes, the stabilizer $A_f = P\Gamma_f \subset P\Gamma$ of $f \in Y$ and the stabilizer $P\Gamma_{\alpha,x} \subset P\Gamma_{\alpha}$ of $x$ in $\mathbb{R} H^n_\alpha$ agree as subgroups of $P\Gamma$. Indeed, we have $Y_f = \mathbb{R} H^n_\alpha$, hence $A_f \setminus \mathbb{R} H^n_\alpha= A_f \setminus Y_f = \Gamma_f \setminus \mathbb{R} H^n$ with $\Gamma_f = A_f \setminus B_f = A_f$; by construction the orbifold chart will then be $W \to A_f \setminus W \subset P\Gamma_\alpha \setminus \mathbb{R} H^n_\alpha \subset Y$ for an invariant open subset $W$ of $\mathbb{R} H^n_\alpha$ containing $x$, which, if $A_f = P\Gamma_{\alpha,x}$, is also an orbifold chart for $O$ at the point $(x, \alpha)$.
\\
\\
To prove that $A_f = P\Gamma_{\alpha,x}$, we first observe that $p: \widetilde Y \to Y$ induces an isomorphism between $P\Gamma_{(x, \alpha)}$, the stabilizer of $(x, \alpha) \in \widetilde Y$ and $P\Gamma_f$, the stabilizer of $f = [x, \alpha] \in Y$. So it suffices to show that $P\Gamma_{(x,\alpha)} = P\Gamma_{\alpha,x}$. For this we use that the normalizer $N_{P\Gamma}(\alpha)$ and the stabilizer $P\Gamma_\alpha \subset P\Gamma$ of $\alpha$ in $P\Gamma$ are equal, which implies that $ P\Gamma_{\alpha,x} = P\Gamma_{(x, \alpha)}$ because
\begin{align*}
\left\{ g \in P\Gamma_\alpha : gx = x \right\} = \left\{ g \in N_{P\Gamma}(\alpha) : gx = x \right\}
= \left\{ g \in P\Gamma: g\cdot (x, \alpha) = (g(x), g\alpha g^{-1}) = (x, \alpha) \right\} .
\end{align*}
3. The real-hyperbolic orbifold $P\Gamma \setminus Y$ is complete by part \textcolor{blue}{1}, so the uniformization of the connected components of $P\Gamma \setminus Y$ follows from the uniformization theorem for $(G,X)$-orbifolds, see \cite[Proposition 13.3.2]{Thurston80}. This concludes the proof of Theorem \ref{theorem1}.
\end{proof}
\end{document}
\subsection{Real uniformization}
Write $\ca F_0(\mathbb{R}) = \ca F_0(\mathbb{R})$ for the preimage of $X_0(\mathbb{R})$ in the space $\ca F_0$. Then $\ca F_0 = \cup_{\alpha \in P\mr A} \ca F_0^\alpha$, where $\mr A$ is the set of anti-unitary involutions $\alpha : \Lambda \to \Lambda$ and $P\mr A = \mu_K \setminus \mr A$ as in Section \ref{realshimura}, $\kappa: X_0(\mathbb{C}) \to X_0(\mathbb{C})$ the anti-holomorphic involution that sends a binary quintic $F(x,y) = \sum_{i+j= 5}a_{ij}x^iy^j \in \mathbb{C}[x,y]$ to the binary quintic $\overline{F(x,y)}= \sum_{i+j =5}\overline{a_{ij}}x^iy^j \in \mathbb{C}[x,y]$, and $\ca F_0^\alpha$ the fixed point set of the natural anti-holomorphic involution $\alpha: \ca F_0 \to \ca F_0$ lying over $\kappa$ which is induced by $\alpha$ as follows. Consider a framing $(F, [f]) \in \ca F_0$, $f: \mathbb{V}_F \to \Lambda$, and let $[\alpha] \in P\mr A$. Let $C_F \to \mathbb{P}^1_\mathbb{C}$ be the quintic cover defined by a smooth binary quintic $F \in X_0(\mathbb{C})$. Complex conjugation $\sigma: \{F = 0\} = Z \to \kappa Z = \{\kappa\cdot F = 0 \}$ extends to an anti-holomorphic map $\sigma: C_F \to C_{\kappa \cdot F}$ which in turn extends to an anti-holomorphic map $\sigma: J(C_F) \to J(C_{\kappa \cdot F})$. Define $F_{\infty, F} = \sigma^\ast: \bb V_{\kappa \cdot F} \to \bb V_F$; then the composite
$
\mathbb{V}_{\kappa \cdot F} \xrightarrow{F_{\infty,F}} \mathbb{V}_F \xrightarrow{f} \Lambda \xrightarrow{\alpha} \Lambda
$ defines a framing of $\kappa \cdot F \in X_0(\mathbb{C})$. For the set of fixed points $ \ca F_0^\alpha$, we have $
\ca F_0^\alpha = \left\{ (F, [f]) \in \ca F_0 : \kappa \cdot F = F \textnormal{ and } [f \circ F_{\infty, \kappa \cdot F} \circ f^{-1}] = [\alpha]. \right\}$. If $[\alpha] \neq [\beta] \in P\mr A$, then $\ca F_0^\alpha \cap \ca F_0^\beta = \emptyset$, for if $(F, [f])$ is in the intersection, then $[\alpha] = [f \circ F_{\infty, \kappa \cdot F} \circ f^{-1}] = [\beta]$.
\begin{lemma}
The anti-holomorphic involution $[\alpha]: \ca F_0 \to \ca F_0$ defined by $[\alpha] \in P \mr A$ makes the period map $\ca P: \ca F_0 \to \mathbb{C} H^2$ Galois-equivariant.
\end{lemma}
\begin{proof}
Indeed, if $\sigma: \mathbb{C} \to \mathbb{C}$ is complex conjugation, then
the induced map $F_{\infty,t} \otimes \sigma: \bb V_{\kappa t}\otimes_\mathbb{Z} \mathbb{C} \to \bb V_t \otimes_\mathbb{Z} \mathbb{C}$ is anti-linear and preserves the Hodge decomposition as well as the eigenspaces.
\end{proof}
Consequently, we obtain a \textit{real period map}
$ \ca P_\mathbb{R}: \coprod_{\alpha \in P\mr A} \ca F_0^\alpha = \ca F_0(\mathbb{R}) \to \widetilde Y = \coprod_{\alpha \in P\mr A} \mathbb{R} H^2_\alpha$.
It is constant on $G(\mathbb{R})$-orbits since the same is true for $\ca P: \ca F_0 \to \mathbb{C} H^2$.
\begin{proposition} \label{prop:realsmoothperiods}
The real period map $\ca P_\mathbb{R}$ descends to a $P\Gamma$-equivariant diffeomorphism $ G(\mathbb{R}) \setminus \ca F_0(\mathbb{R}) \cong \coprod_{\alpha \in P\mr A} \mathbb{R} H^2_\alpha - \mr H$. Thus there is an isomorphism of real-analytic orbifolds
\begin{equation}
\ca P_{\mathbb{R}}: \mr M_\mathbb{R} = G(\mathbb{R}) \setminus X_0(\mathbb{R}) \cong \coprod_{\alpha \in C \mr A}\left[P\Gamma_\alpha \setminus \left( \mathbb{R} H^2_\alpha - \mr H
\right)\right].
\end{equation}
\end{proposition}
\begin{proof}
This follows from \cite[\textit{proof of Theorem 3.3}]{realACTsurfaces}. It is crucial that the actions of $G$ and $P\Gamma$ on $\ca F_0$ commute and are free, which is the case, see Corollary \ref{cor:freeaction}.
\end{proof}
The goal is to prove the real analogue of the isomorphism $\overline{\mr M}_\mathbb{C} = G(\mathbb{C}) \setminus X_s(\mathbb{C}) \cong P\Gamma \setminus \mathbb{C} H^2$ in Theorem \ref{th:delignemostow}. We need two more lemma's.
\begin{lemma}
\label{lemma:stabilizergroups}
The period map $\overline{\ca P}: \ca F_s \to \mathbb{C} H^2$ sends the points $f \in \widetilde \Delta : = \ca F_s - \ca F_0$ lying above binary quintics with $k$ nodes
to the locus of $\mathbb{C} H^2$ where exactly $k$ of the hyperplanes of $\mr H$ meet. If $f \in \widetilde \Delta$ is such a point, and $\textbf r = (r_1, \dotsc, r_k)$ a vector of short roots such that $\overline{\ca P}(f) = x \in \cap_i H_{r_i}$,
then $\overline{\ca P}: \ca F_s \to \mathbb{C} H^2$ induces a group isomorphism $P\Gamma_f \cong G(\textbf r)$.
$\hfill \qed$
\end{lemma}
The naturality of the Fox completion implies that each anti-unitary involution $\alpha: \ca F_0 \to \ca F_0$ extends to an anti-unitary involution $\alpha: \ca F_s \to \ca F_s$.
\begin{lemma} \label{lemma:alphaperiod}
For every $\alpha \in P\mr A$, the restriction of $\overline{\ca P}: \ca F_s \to \mathbb{C} H^2$ to $\ca F_s^\alpha$ defines a diffeomorphism $G(\mathbb{R}) \setminus \ca F_s^\alpha \cong \mathbb{R} H^2_\alpha$.
\end{lemma}
\begin{proof}
See \cite[Lemma 11.3]{realACTsurfaces}. It is crucial that $G$ acts freely on $\ca F_s$ (Corollary \ref{cor:freeaction}).
\end{proof}
We arrive at the main theorem of this section. Define $\ca F_s(\mathbb{R}) = \cup_{\alpha \in P\mr A} \ca F_s^\alpha = \pi^{-1}\left(X_s(\mathbb{R})\right)$. This is not a manifold because of the ramification of $\pi: \ca F_s \to X_s(\mathbb{C})$, but a union of embedded submanifolds.
\begin{theorem} \label{th:realstableperiod}
The real stable period map $\overline{\ca P}_{\mathbb{R}}: \coprod_{\alpha \in P\mr A} \ca F_s^\alpha \to \widetilde Y = \coprod_{\alpha \in P\mr A}\mathbb{R} H^2_\alpha$ extends $\ca P_\mathbb{R}$ and induces the following commutative diagram:
\begin{equation*}
\xymatrixcolsep{5pc}
\xymatrix{
\coprod_{\alpha \in P\mr A} \ca F_s^\alpha \ar[r]^{\overline{\ca P}_{\mathbb{R}}:}\ar[d] & \widetilde Y = \coprod_{\alpha \in P\mr A} \mathbb{R} H^2_\alpha \ar[d]\\
\ca F_s(\mathbb{R}) \ar[r]^{\overline{\ca P}_\mathbb{R}}\ar[d] & Y\ar@{=}[d] \\
G(\mathbb{R}) \setminus \ca F_s(\mathbb{R}) \ar[r]^{\mr P_\mathbb{R}}\ar[d] & Y\ar[d] \\
G(\mathbb{R}) \setminus X_s(\mathbb{R}) \ar[r]^{{\mr P}_\mathbb{R}} & P\Gamma \setminus Y,
}
\end{equation*}
and $\mr P_\mathbb{R}: G(\mathbb{R}) \setminus \ca F_s(\mathbb{R}) \to Y$ and ${\mr P}_\mathbb{R}: G(\mathbb{R}) \setminus \ca F_s(\mathbb{R}) / P\Gamma \to P\Gamma \setminus Y$ are homeomorphisms.
\end{theorem}
\begin{proof}
We first show that $\overline{\ca P}_{\mathbb{R}}: \coprod_{\alpha \in P\mr A} \ca F_s^\alpha \to \widetilde Y \to Y$ factors through $\ca F_s(\mathbb{R})$. Now $(f, \alpha)$ and $(g, \beta) \in \coprod_{\alpha \in P\mr A} \ca F_s^\alpha$ have the same image in $\ca F_s(\mathbb{R})$ if and only if $f = g \in \ca F_s^\alpha \cap \ca F_s^\beta$, in which case $x:=\ca P_{s}(f) = \ca P_{s}(g) \in \mathbb{R} H^2_\alpha \cap H^2_\beta$, so all we need to show is that $(x, \alpha) \sim (x, \beta) \in \widetilde Y$. But note that $\alpha\beta \in P\Gamma_f \cong (\mathbb{Z}/10)^k$,
and $\overline{\ca P}$ induces an isomorphism $P\Gamma_f \cong G(\textbf r)$ by Lemma \ref{lemma:stabilizergroups}. Hence $\alpha \beta \in G(\textbf r)$ so that indeed, $(x, \alpha) \sim (x, \beta)$. Let us prove the $G(\mathbb{R})$-equivariance of $\overline{\ca P}_{\mathbb{R}}$: if $f \in \ca F_s^\alpha, g \in \ca F_s^\beta$ with $a \cdot f = g \in \ca F_s(\mathbb{R})$ for some $a \in G(\mathbb{R})$, then $x:=\overline{\ca P}(f) = \overline{\ca P}(g) \in \mathbb{C} H^2$, so we need to show that $\alpha \beta \in G(\textbf r)$. The actions of $G(\mathbb{C})$ and $P\Gamma$ on $\mathbb{C} H^2$ commute, and the same holds for the actions of $G(\mathbb{R})$ and $P\Gamma'$ on $\ca F_s^\mathbb{R}$, where $P\Gamma' = \Gamma'/\mu_K$, $\Gamma'$ being the group of all unitary and anti-unitary automorphisms of $\Lambda$. It follows that $\alpha(g) = \alpha (a \cdot f) = a \cdot \alpha(f) = a \cdot f = g$, hence $g \in \ca F_s^\alpha \cap \ca F_s^\beta$. This implies in turn that $\alpha \beta (g) = g$, hence $\alpha \beta \in P\Gamma_g \cong G(\textbf r)$, so that indeed, $(x, \alpha) \sim (x, \beta)$. To prove that $\mr P_\mathbb{R}$ is injective, let again $(f, \alpha)$, $(g, \beta) \in \coprod_{\alpha \in P\mr A} \ca F_s^\alpha$ and suppose that they have the same image in $Y$. This implies that $x:=\overline{\ca P}(f) = \overline{\ca P}(g) \in \mathbb{R} H^2_\alpha \cap H^2_\beta$ and $\beta = \phi \circ \alpha$ for some $\phi \in G(\textbf r)$.
Now $\phi \in G(\textbf r) \cong P\Gamma_f$ (by Lemma \ref{lemma:stabilizergroups}) hence
\begin{equation}
\beta(f) = \phi \left(\alpha (f)\right) = \phi(f) = f.
\end{equation}
Therefore $f,g \in \ca F_s^\beta$; since $\overline{\ca P}(f) = \overline{\ca P}(g)$, it follows from Lemma \ref{lemma:alphaperiod} that there exists $a \in G(\mathbb{R})$ such that $a \cdot f = g$, which proves injectivity of $\mr P_\mathbb{R}$.
Surjectivity of $\mr P_\mathbb{R}: G(\mathbb{R}) \setminus \ca F_s(\mathbb{R}) \to Y$ is straightforward, using surjectivity of $\overline{\ca P}_{\mathbb{R}}$ (Lemma \ref{lemma:alphaperiod}). Finally, we claim that $\mr P_\mathbb{R}$ is open. Let $U \subset G(\mathbb{R}) \setminus \ca F_s^\mathbb{R}$; we can write $U = \mr P_\mathbb{R}^{-1} \mr P_\mathbb{R} \left(U\right)$. Let $V$ be the preimage of $U$ in $\coprod_{\alpha \in P\mr A}\ca F_s^\alpha$. Then $V = \overline{\ca P}_{\mathbb{R}}^{-1}\left( p^{-1}\left(\mr P_\mathbb{R}(U)\right)\right)$ and hence
$
\overline{\ca P}_{\mathbb{R}}\left( V \right) = p^{-1}\left(\mr P_\mathbb{R}(U)\right).
$
But the map $\overline{\ca P}_{\mathbb{R}}$ is open, being the coproduct of the maps $\ca F_s^\alpha \to \mathbb{R} H^2_\alpha$, each of which is open since its differential is surjective at each point. Thus $\mr P_\mathbb{R}(U)$ is open in $Y$.
\end{proof}
\begin{corollary} \label{cor:theorem2}
There is a lattice $P\Gamma_\mathbb{R} \subset \textnormal{PO}(2,1)$ and a homeomorphism
\begin{equation}
\mr M_\mathbb{R} = G(\mathbb{R}) \setminus X_s(\mathbb{R}) \cong P\Gamma_\mathbb{R} \setminus \mathbb{R} H^2.
\end{equation}
\end{corollary}
\begin{proof}
$G(\mathbb{R}) \setminus X_s(\mathbb{R}) \cong P\Gamma \setminus Y$ (Theorem \ref{th:realstableperiod}) and $\Gamma \setminus Y \cong P\Gamma_\mathbb{R} \setminus \mathbb{R} H^2$ (Theorem \ref{theorem1}).
\end{proof}
\begin{remark}
The proof of Theorem \ref{th:realstableperiod} also shows that $\ca M_s(\mathbb{R}) \cong P\Gamma \setminus Y$ if $\ca M_s$ is the moduli stack of either cubic surfaces or binary sextics. Notice that this strategy to uniformize the real moduli space is somewhat different than the one in \cite{realACTsurfaces}, \cite{realACTnonarithmetic}, \cite{realACTbinarysextics}, since we first glue together the different components of the real locus of the Shimura variety and only afterwards prove that our moduli space is homeomorphic to the resulting hyperbolic gluing.
\end{remark}
\end{document}
\subsection{Alternating and hermitian forms on $\Lambda$}
The goal of this subsection is to prove two lemma's that will be used in Section \ref{subsec:moduliabelianvarieties} to show that $P\Gamma \setminus \mathbb{C} H^n$ is a moduli space of abelian varieties
and interprete the divisor $P\Gamma \setminus \mr H \subset P\Gamma \setminus \mathbb{C} H^n$.
\begin{lemma} \label{lemma:equivalentforms}
The assignment $T \mapsto \textnormal{Tr}_{K/\mathbb{Q}} \circ T$ defines a bijection between:
\begin{enumerate}
\item \label{item:hermitian} The set of skew-hermitian forms $T: \Lambda_{\mathbb{Q}} \times \Lambda_{\mathbb{Q}} \to K$.
\item \label{item:alternating} The set of alternating forms $E: \Lambda_{\mathbb{Q}} \times \Lambda_{\mathbb{Q}} \to \mathbb{Q}$ such that $E(a \cdot x,y) = E( x, a^\sigma \cdot y)$.
\end{enumerate}
Under this correspondence, $T(\Lambda, \Lambda) \subset \mf D_K^{-1}$ if and only if $E(\Lambda, \Lambda) \subset \mathbb{Z}$.
\end{lemma}
\begin{proof}
Let $T: \Lambda_{\mathbb{Q}} \times \Lambda_{\mathbb{Q}} \to K$ be as in \ref{item:hermitian}. Define $E_T = \text{Tr}_{K/\mathbb{Q}} \circ T$.
Since $T$ is skew-hermitian, $\text{Tr}_{K/\mathbb{Q}}T(x,y) = - \text{Tr}_{K/\mathbb{Q}}\overline{T(y,x)}$. Since $K/\mathbb{Q}$ is separable, for any $x \in K$, we have
$
\text{Tr}_{K/\mathbb{Q}}(x) = \sum_{1 \leq i \leq g} \left( \tau_i(x) + \tau_i\sigma(x) \right)
$
hence $\text{Tr}_{K/\mathbb{Q}}(\sigma(x)) = \text{Tr}_{K/\mathbb{Q}}(x)$, so that $E_T(x,y) = - E_T(y,x)$ for any $x,y \in \Lambda_\mathbb{Q}$. The property in \ref{item:alternating} is easily checked.
Conversely, let $E: \Lambda_\mathbb{Q} \times \Lambda_\mathbb{Q} \to \mathbb{Q}$ be as in \ref{item:alternating}. Choose a basis $\{b_1, \dotsc, b_{n+1} \} \subset \Lambda$ for $\Lambda$ over $\mathcal{O}_K$.
Define $Q$ to be the induced map $K^{n+1} \times K^{n+1} \to \mathbb{Q}$ and consider the map
$K \to \mathbb{Q}$, $a \mapsto Q(a \cdot e_i, e_j)$. Since the trace pairing $K \times K \to \mathbb{Q}$, $(x,y) \mapsto \text{Tr}_{K/\mathbb{Q}}(xy)$ is non-degenerate \cite[\href{https://stacks.math.columbia.edu/tag/0BIE}{Tag 0BIE}]{stacks-project},
there exists a unique $t_{ij} \in K$ such that $Q(a \cdot e_i, e_j) = \text{Tr}_{K/\mathbb{Q}}(a \cdot t_{ij})$ for every $a \in K$. This gives a matrix $(t_{ij})_{ij} \in M_{n+1}(K)$ such that $\sigma(t_{ij}) = - t_{ji}$, and the basis $\{b_i\}$ induces a skew-hermitian form $T_E: \Lambda_\mathbb{Q} \times \Lambda_\mathbb{Q} \to K$.
The last claim follows by definition of $\mf D_K^{-1}$ as the trace dual of $\mathcal{O}_K$.
\end{proof}
\begin{examples} \label{examplesunitary}
\begin{enumerate}
\item \label{ex:unitone} Suppose $K = \mathbb{Q}(\sqrt \Delta)$ is imaginary quadratic with discriminant $\Delta$ and non-trivial Galois automorphism $a \mapsto a^\sigma$. Let $E: \Lambda \times \Lambda \to \mathbb{Z}$ be an alternating form with $E(a \cdot x,y) = E(x, a^\sigma \cdot y)$. The form $T: \Lambda \times \Lambda \to \mf D_{K}^{-1} = (\sqrt{\Delta})^{-1}$ is defined as
\begin{equation*}
T(x,y) = \frac{E( \sqrt{\Delta}\cdot x,y) + E(x,y)\sqrt{\Delta} }{2\sqrt \Delta}.
\end{equation*}
\item \label{ex:unittwo} Let $K = \mathbb{Q}(\zeta)$ where $\zeta = \zeta_p = e^{2 \pi i/p} \in \mathbb{C}$ for some prime number $p > 2$. Let $E: \Lambda \times \Lambda \to \mathbb{Z}$ be an alternating form with $E(a \cdot x,y) = E(x, a^\sigma \cdot y)$. Then $\mf D_K = \left(p/(\zeta - \zeta^{-1})\right)$ and
\begin{equation*}
T: \Lambda \times \Lambda \to \mf D_K^{-1}, \;\;\; T(x,y) = \frac{1}{p}\sum_{j = 0}^{p-1}\zeta^jE\left( x, \zeta^j \cdot y \right).
\end{equation*}
\end{enumerate}
\end{examples}
\begin{comment}
\begin{proof}
Since $E = \textnormal{Tr}_{K/\mathbb{Q}} \circ T$ and $Q = \textnormal{Tr}_{K/\mathbb{Q}} \circ R$, the `if' statement is clear. So suppose $E(\phi(x), \phi(y)) = Q(x,y)$ for $x,y \in L$. Consider the form $\phi^\ast T: L_\mathbb{Q} \times L_\mathbb{Q} \to K$ defined by $\phi^\ast T(x,y) = T(\phi(x),\phi(y))$. Then $\phi^\ast T$ is skew-hermitian, $\phi^\ast T(L, L) \subset \mf D_K^{-1}$, and
$$
\textnormal{Tr}_{K/\mathbb{Q}} \left( \phi^\ast T(x,y) \right) = \textnormal{Tr}_{K/\mathbb{Q}} \left( T(\phi(x), \phi(y)) \right) = E(\phi(x), \phi(y)) = Q(x,y) = \textnormal{Tr}_{K/\mathbb{Q}} \left( R(x,y) \right)
$$
for every $x,y \in L$. By Lemma \ref{lemma:equivalentforms}, we must have $\phi^\ast T = R$, proving what we want.
\end{proof}
\end{comment}
Now consider $E: \Lambda_\mathbb{Q} \times \Lambda_\mathbb{Q} \to \mathbb{Q}$ and $T: \Lambda_{\mathbb{Q}} \times \Lambda_{\mathbb{Q}} \to K $ as in Lemma \ref{lemma:equivalentforms}, and suppose that $E$ is non-degenerate. Let $\varphi: K \to \mathbb{C}$ be an embedding. Define a skew-hermitian form $T^\varphi$ as
\begin{equation}
T^\varphi: \Lambda \otimes_{\mathcal{O}_K, \varphi} \mathbb{C} \times \Lambda \otimes_{\mathcal{O}_K, \varphi} \mathbb{C}\to \mathbb{C}, \;\;\;
T^\varphi( \sum_i x_i \otimes \lambda_i, \sum_j y_j \otimes \mu_j) =
\sum_{ij} \lambda_i \overline{\mu_j} \cdot \varphi\left( T(x_i, y_j) \right).
\end{equation}
On $\Lambda_\mathbb{C}$, we also have the skew-hermitian form $A(x,y) = E_\mathbb{C}(x,\bar y)$. The composition $\left(\Lambda \otimes_\mathbb{Z} \mathbb{C} \right)_{\varphi} \to \Lambda \otimes_\mathbb{Z} \mathbb{C} \to \Lambda \otimes_{\mathcal{O}_K, \varphi} \mathbb{C}$ is an isomorphism. Define $A^\varphi$ to be the restriction of $A$ to the subspace $\left(\Lambda \otimes_\mathbb{Z} \mathbb{C} \right)_{\varphi} = \Lambda \otimes_{\mathcal{O}_K, \varphi} \mathbb{C} \subset \Lambda_\mathbb{C}$. Note that $\Lambda \otimes_\mathbb{Z} \mathbb{C} \cong \oplus_{\phi: K \to \mathbb{C}}\left(\Lambda \otimes_\mathbb{Z} \mathbb{C} \right)_{\phi}$; for $x \in \Lambda \otimes_\mathbb{Z} \mathbb{C}$, denote by $x^\phi$ the image of $x$ under the projection $\Lambda \otimes_\mathbb{Z} \mathbb{C} \to \left(\Lambda \otimes_\mathbb{Z} \mathbb{C} \right)_{\phi}$.
\begin{lemma} \label{lemma:agree}
We have $T^\varphi = A^\varphi$ as skew-hermitian forms $\left(\Lambda \otimes_\mathbb{Z} \mathbb{C} \right)_{\varphi} \times \left(\Lambda \otimes_\mathbb{Z} \mathbb{C} \right)_{\varphi} \to \mathbb{C}$. More precisely, we have $A(x,y) = \sum_{\phi: K \to \mathbb{C}} T^\phi(x^\phi, y^\phi)$ for every $x,y \in \Lambda \otimes_\mathbb{Z} \mathbb{C}$.
\end{lemma}
\begin{proof}
Write $V = \Lambda_\mathbb{Q}$.
The lemma follows from the fact that the following diagram commutes:
$$
\xymatrixcolsep{5pc}
\xymatrix{
V \times V\ar@{^{(}->}[d] \ar[d] \ar[r]^T& K\ar@{^{(}->}[d] \ar[r]^{\textnormal{Tr}_{K/\mathbb{Q}}} & \mathbb{Q} \ar@{^{(}->}[dd] \\
V \otimes_\mathbb{Q} \mathbb{C} \times V \otimes_\mathbb{Q} \mathbb{C} \ar[r]^{T_\mathbb{C}} \ar[drr]^{A(x, y)} \ar@{=}[d] & K \otimes_\mathbb{Q} \mathbb{C}\ar@{=}[d] & \\
\oplus_\phi \left(V \otimes_\mathbb{Q} \mathbb{C}\right)_\phi \times \left(V \otimes_\mathbb{Q} \mathbb{C}\right)_\phi \ar[r]^{\;\;\; \;\;\;\white\;\;\;\white \oplus T^\phi} & \oplus_\phi \mathbb{C}_\phi \ar[r]_{\sum} & \mathbb{C},
}
$$
where $\phi$ ranges over the set of embeddings $K \to \mathbb{C}$, $\mathbb{C}_\phi$ is the $K$-module $\mathbb{C}$ where $K$ acts via $\phi$, and $T_\mathbb{C}: V \otimes_\mathbb{Q} \mathbb{C} \times V \otimes_\mathbb{Q} \mathbb{C} \to K \otimes_\mathbb{Q}\mathbb{C}$ is the map that sends $(v \otimes \lambda, x \otimes \mu)$ to $\lambda \bar \mu T(v,w)$.
\end{proof}
\subsection{Moduli of abelian varieties with $\mathcal{O}_K$-action} \label{subsec:moduliabelianvarieties}
\begin{notation} \label{not:fixhermitian}
In the rest of Section \ref{unitaryshimura}, we fix a non-degenerate hermitian form $\mf h: \Lambda \times \Lambda \to \mf D_K^{-1}$. We also fix an element $\xi \in \mf D_K^{-1}$ such that $\sigma(\xi) = - \xi$ and $\Im \left(\tau_i(\xi) \right) < 0$ for $1 \leq i \leq g$ and write $\eta = \xi^{-1}$. This defines a skew-hermitian form $T: \Lambda \times \Lambda \to \mf D_K^{-1}$ where $ T = \xi \cdot \mf h$, in turn attached to a symplectic form $E: \Lambda \times \Lambda \to \mathbb{Z}$ such that $E(a x, y) = E(x, a^\sigma y)$ for all $a \in \mathcal{O}_K$, $x,y \in \Lambda$, see Lemma \ref{lemma:equivalentforms}.
Write $V_i = \Lambda_\mathbb{Q} \otimes_{K, \tau_i} \mathbb{C}$ and define $\mf h^{\tau_i}: V_i\times V_i \to \mathbb{C}$ to be the hermitian form restricting to $\tau_i \circ \mf h $ on $\Lambda$.
Let $(r_i,s_i)$ be the signature of the hermitian form $\mf h^{\tau_i}$.
\end{notation}
Consider a complex abelian variety $A$ equipped with a homomorphism $\iota: \ca O_K \to \text{End}(A)$ and a polarization $\lambda: A \to A^\vee$, satisfying the following conditions:
\begin{enumerate}
\item $\iota(a)^\dagger = i(a^\sigma)$ for the corresponding Rosati involution $\dagger: \textnormal{End}(A)_{\mathbb{Q}} \to\textnormal{End}(A)_{\mathbb{Q}} $,
\item
$
\text{char}(t, \iota(a) | \text{Lie}(A)) =
\prod_{\nu = 1}^g (t-a^{\tau_i})^{r_i}
\cdot
(t-a^{\tau_i\sigma})^{s_i} \;\;\; \in \;\;\; \mathbb{C}[t].
$
\end{enumerate}
Note that $\dim A = g(n+1)$.
Define $E_A : H_1(A, \mathbb{Z}) \times H_1(A, \mathbb{Z}) \to \mathbb{Z}$ to be the alternating form corresponding to $\lambda$. The condition on the Rosati involution implies that $E_A(\iota(a)x, y) = E_A(x, \iota(a^\sigma)y)$ for $x,y \in H_1(A, \mathbb{Q})$. Let $\mf h_A$ be the hermitian form $\mf h_A = \eta T_A$ on $H_1(A, \mathbb{Z})$, where $T_A: H_1(A, \mathbb{Z}) \times H_1(A, \mathbb{Z}) \to \mf D_K^{-1}$ is the skew-hermitian form attached to $E_A$ via Lemma \ref{lemma:equivalentforms}.
\\
\\
Let $\widetilde{\textnormal{Sh}}_{K}(\mf h)$ be the set of isomorphism classes of four-tuples $(A, i, \lambda, j)$ where $(A,i, \lambda)$ is as above and where $j: H_1(A, \mathbb{Z}) \to \Lambda$ is a symplectic isomorphism of $\mathcal{O}_K$-modules.
Finally, define $\bb D(V_i)$ to be the space of negative $s_i$-planes in the hermitian space $(V_{i}, \mf h^{\tau_i})$. We then have the following proposition which is due to Shimura, see \cite[Theorem 2]{Shimura1963ONAF} or \cite[\S 1]{shimuratranscendental}. We give a different proof since it will imply Proposition \ref{prop:HcorrespondsNonSimple} below, whereas we did not know how to deduce Proposition \ref{prop:HcorrespondsNonSimple} from \textit{loc. cit.} We remark that Shimura assumes $\Lambda$ to be an $R$-module for any order $R \subset \mathcal{O}_K$; our proof carries over but we don't need this generalization.
\begin{proposition} \label{prop:canonicalbijection}
There is a canonical bijection
$
\widetilde{\textnormal{Sh}}_K(\mf h) \cong \bb D(V_1) \times \dotsc \times \bb D(V_g).
$
\end{proposition}
\begin{proof}
Let $(A, i, \lambda, j)$ be a representative of an isomorphism class in $\widetilde{\textnormal{Sh}}_{K}(\mf h)$.
Let $H_1(A, \mathbb{C}) = H^{-1,0} \oplus H^{0,-1}$ be the Hodge decomposition of $A$. For $1 \leq i \leq g$ there is a decomposition
\begin{equation} \label{eq:posneg}
H_1(A, \mathbb{C})_{\tau_i} = H^{-1,0}_{\tau_i} \oplus H^{0,-1}_{\tau_i}
\end{equation}
where $\dim H^{-1,0}_{\tau_i} = r_i$ and $\dim H^{0,-1}_{\tau_i} = s_i$, where the latter holds because $\overline{ H^{-1,0}_{\tau_i\sigma}} = H^{0,-1}_{\tau_i}$. By Lemma \ref{lemma:agree}, $\tau_i(\eta)E_{A, \mathbb{C}}(x,\bar y)$ and ${\mf h}_{A, \mathbb{C}}^{\tau_i}(x,y)$ agree as hermitian forms on $H_1(A, \mathbb{Z}) \otimes_{\mathcal{O}_K, \tau_i} \mathbb{C}$. Since $\Im \tau_i(\eta) > 0$ for every $i$,
the decomposition of $H_1(A, \mathbb{C})_{\tau_i}$ in (\ref{eq:posneg}) is a decomposition into a positive definite $r_i$-dimensional subspace and a negative definite $s_i$-dimensional subspace.
The isomorphism $j: H_1(A, \mathbb{Q}) \to \Lambda_\mathbb{Q}$ induces an isometry $j_i: H_1(A, \mathbb{C})_{\tau_i} \to V_i$ for every $i$, and so we obtain a negative $s_i$-plane $j ( H^{0,-1}_{\tau_i})$ in the hermitian space $V_i$ for all $i$. Reversing the argument shows that given a negative $s_i$-plane $X_i \subset V_i$ for every $i$, there is a canonical polarized abelian variety $A = H^{-1,0}/\Lambda$, acted upon by $\mathcal{O}_K$ and inducing the planes $X_i \subset V_i$.
\end{proof}
Let $\textnormal{Sh}_{K}(\mf h)$ be the set of isomorphism classes of polarized $\mathcal{O}_K$-linear abelian varieties $(A, i, \lambda)$ such that $H_1(A, \mathbb{Z})$ is isometric to $\Lambda$ as hermitian $\mathcal{O}_K$-modules. Let $\Gamma(\mf h) = \textnormal{Aut}_{\mathcal{O}_K}(\Lambda, \mf h)$: this is the group of $\mathcal{O}_K$-linear automorphisms of $\Lambda$ preserving our form $\mf h: \Lambda \times \Lambda \to \mf D_K^{-1}$. The bijection in Proposition \ref{prop:canonicalbijection} being $\Gamma (\mf h)$-equivariant, we obtain the following:
\begin{corollary}
There is a canonical bijection $\textnormal{Sh}_{K}(\mf h) \cong \Gamma(\mf h) \setminus \prod_{i = 1}^g \bb D(V_i)$. $\hfill \qed$
\end{corollary}
\begin{comment}
\begin{corollary}
Suppose $n = 0$, i.e. $\Lambda$ is a free $\mathcal{O}_K$-module of rank $1$. Let $E : \Lambda \times \Lambda \to \mathbb{Z}$ be a symplectic (i.e. non-degenerate alternating) form such that $|\Lambda^\vee/\Lambda| = d^2$ and $E(ax,y) = E(x,a^\sigma y)$. Up to isomorphism there is one and only one polarized abelian variety $A$ with $\mathcal{O}_K$-multiplication of signature $\Psi$ having $(\Lambda, E)$ as underlying symplectic lattice. $\hfill \qed$
\end{corollary}
\end{comment}
\subsection{Abelian varieties in the hyperplane arrangement}
\begin{comment}
\begin{notation} \label{not:embeddingselementagain}
In the rest of Section \ref{unitaryshimura}, we fix the following. Let $\eta \in K$ such that \textcolor{blue}{(1)} $K = F(\eta)$, \textcolor{blue}{(2)} $\sigma(\eta) = - \eta$, $\eta^2 \in F$, and \textcolor{blue}{(3)} $\eta^{-1} \in \mf D_K^{-1}$. Moreover, let $\Psi$ be a set of embeddings $\{\tau_i: K \to \mathbb{C}\}_{1 \leq i \leq g}$ such that \textcolor{blue}{(4)} $\Psi \cup \sigma\Psi = \Hom(K, \mathbb{C})$ and \textcolor{blue}{(5)} $\Im \left(\tau_i(\eta) \right) > 0$ for $1 \leq i \leq g$.
Note that such $\eta$ and $\Psi$ exist. We also fix a skew-hermitian form $T: \Lambda \times \Lambda \to \mf D_K^{-1}$ and denote by $E$ the alternating form $\Lambda \times \Lambda \to \mathbb{Z}$ that corresponds to $T$ via Lemma \ref{lemma:equivalentforms}.
We define $\mf h: \Lambda_\mathbb{Q} \times \Lambda_\mathbb{Q} \to K$ to be the hermitian form $\mf h(x,y) = \eta T(x,y)$.
Let $V_i = \Lambda_\mathbb{Q} \otimes_{K, \tau_i} \mathbb{C}$ and assume that $\mf h^{\tau_i}$ is non-degenerate for each embedding $\tau_i: K \to \mathbb{C}$, of signature $(r_i,s_i)$.
\end{notation}
\end{comment}
\begin{comment}
\begin{condition} \label{condition:one}
The set of embeddings $\Psi = \{\tau_i: K \to \mathbb{C}\}_{1 \leq i \leq g}$ and the element $\eta \in K$ can be chosen such that $(r_1,s_1) = (n,1)$, $(r_i,s_i) = (n+1,0)$ for $2 \leq i \leq g$, as well as to satisfy the requirements of Notation \ref{notation:embeddingselementagain}. In other words, $(\Psi, \eta)$ satisfies the following:
\begin{enumerate}
\item $\Psi \cup \sigma\Psi = \Hom(K, \mathbb{C})$,
\item $\Im(\varphi(\eta)) > 0$ for every $\varphi \in \Psi$,
\item $K = F(\eta)$, $\eta \in \mathcal{O}_K$, $\sigma(\eta) = -\eta$, $\eta^2 \in F$,
\item the signature of $\tau_i(\eta) T^{\tau_i} $ equals $(r_1,s_1) = (n,1)$, $(r_i,s_i) = (n+1,0)$ for $2 \leq i \leq g$.
\end{enumerate}
\end{condition}
\end{comment}
\noindent
The set of embeddings $\Psi$ defines a map $\Psi: \mathcal{O}_K \to \mathbb{C}^g$
giving a complex torus $\mathbb{C}^g/\Psi(\mathcal{O}_K)$.
The map $Q: K \times K \to \mathbb{Q}$, $Q(x,y) = \text{Tr}_{K/\mathbb{Q}}(\xi x \bar y)$ is a non-degenerate $\mathbb{Q}$-bilinear form such that $Q(ax,y) = { Q}(x, a^\sigma y)$ for every $a,x,y \in K$, and $Q(\mathcal{O}_K, \mathcal{O}_K) \subset \mathbb{Z}$ because $\xi\in \mf D_K^{-1}$.
By \cite[Example 2.9 \& Footnote 16]{milneCM},
$ Q$ defines a Riemann form on the complex torus $\mathbb{C}^g/\Psi(\mathcal{O}_K)$.
\\
\\
Suppose that $(r_1,s_1) = (n,1)$ and $(r_i,s_i) = (n+1,0)$ for $2 \leq i \leq g$. As in Section \ref{realshimura}, let $\mathbb{C} H^n$ be the set of negative lines in $\Lambda \otimes_{\mathcal{O}_K, \tau_1} \mathbb{C}$, and define $\mr H = \cup_{\mf h(r,r) = 1} r_\mathbb{C}^\perp \subset \mathbb{C} H^n$.
\begin{proposition} \label{prop:HcorrespondsNonSimple}
Under the bijection $ \widetilde{\textnormal{Sh}}_{K}(\mf h) \cong \mathbb{C} H^n$ of Proposition \ref{prop:canonicalbijection}, the subset $\mr H \subset \bb C H^n$ corresponds to the isomorphism classes of those polarized marked $\mathcal{O}_K$-linear abelian varieties $A$
that admit a $\mathcal{O}_K$-linear homomorphism $\mathbb{C}^g/\Psi(\mathcal{O}_K) \to A$
of polarized abelian varieties.
\end{proposition}
\begin{proof}
Let $[A,i,\lambda,y] \in \widetilde{\textnormal{Sh}}_{K}(\mf h) $ correspond to $[x] \in \bb C H^n$. We may assume that $A = H^{-1,0} / \Lambda$ with $\Lambda \otimes_\mathbb{Z} \mathbb{C} = H^{-1,0} \oplus H^{0,-1}$, and that $T_A = T$. Let $\phi: \mathbb{C}^g/\Psi(\mathcal{O}_K) \to A$ be such a homomorphism. We obtain a homomorphism $\mathcal{O}_K \to \Psi(\mathcal{O}_K) \to H_1(A, \mathbb{Z}) = \Lambda$ which, for simplicity, we also denote by $\phi: \mathcal{O}_K \to \Lambda$. Let $r \in \Lambda$ be the image of $1 \in \mathcal{O}_K$.
The fact that $Q = \phi^\ast E_A$ implies that $T_Q = \phi^\ast T_A = \phi^\ast T$.
Therefore,
we have
$$\eta^{-1} = T_Q(1,1)= T_A(\phi(1), \phi(1)) = T(\phi(1), \phi(1)) = T(r,r),$$ so that $\mf h(r,r) = \eta \cdot T(r,r) = 1$. Next, we claim that $\mf h(x,r_\tau) = 0$, where $r_\tau \in \left(\Lambda \otimes_\mathbb{Z} \mathbb{C} \right)_\tau = H_\tau$ is the image of $r \in \Lambda$. Write $\Psi(\mathcal{O}_K) = L$, $L \otimes \mathbb{C} = W^{-1,0} \oplus W^{0,-1}$, and let $\alpha \in L$ correspond to $1 \in \mathcal{O}_K$. Notice that $\left(L \otimes_\mathbb{Z} \mathbb{C} \right)_\tau = W^{-1,0}_\tau$. Consequently, since the composition
$$
W^{-1,0}_\tau = \left(L \otimes_\mathbb{Z} \mathbb{C} \right)_\tau \to \left(\Lambda \otimes_\mathbb{Z} \mathbb{C} \right)_\tau = H^{-1,0}_\tau \oplus H^{0,-1}_\tau
$$
factors through the inclusion of $H^{-1,0}_\tau$ into $\left(L \otimes_\mathbb{Z} \mathbb{C} \right)_\tau$, we see that
$r_\tau = r^{-1,0}_\tau \in H^{-1,0}_\tau = \left( H^{0,-1}_\tau \right)^\perp = \langle x \rangle ^\perp$, and the claim follows.
Conversely, let $[x] \in r_\mathbb{C}^\perp \subset \mr H$ with $r \in \Lambda$ such that $\mf h(r,r) = 1$ and consider the marked abelian variety $A = H^{-1,0}/\Lambda$ corresponding to $[x]$. Define a homomorphism $\phi: \mathcal{O}_K \to \Lambda$ by $\phi(1) = r$. Then $\phi$ can easily be shown to be is a morphism of Hodge structures using the fact that its $\mathbb{C}$-linear extension preserves the eigenspace decompositions.
We obtain an $\mathcal{O}_K$-linear homorphism $\phi: \mathbb{C}^g/\Psi(\mathcal{O}_K) \to A$ it is easily shown that the fact that $\mf h(r,r) = 1$ implies that $\phi$ preserves the polarizations on both sides.
\end{proof}
Observe that if the different $\mf D_K\subset \mathcal{O}_K$ is a principal ideal $ (\eta) \subset \mathcal{O}_K$, then we have
$$
\{x \in K : \textnormal{Tr}_{K/\mathbb{Q}} \left(x \eta^{-1} \mathcal{O}_K\right) \subset \mathbb{Z} \} = \{x \in K: x \cdot \eta^{-1} \mathcal{O}_K \subset \eta^{-1} \mathcal{O}_K \} = \{x \in K: x \mathcal{O}_K \subset \mathcal{O}_K \} = \mathcal{O}_K,
$$
hence $Q: \Psi(\mathcal{O}_K) \times \Psi(\mathcal{O}_K) \to \mathbb{Z}$ defines a principal polarization on $\mathbb{C}^g/\Psi(\mathcal{O}_K)$. In fact, for $\beta \in K$, the rational Riemann form $\Psi(K) \times \Psi(K) \to \mathbb{Q}$, $(\Psi(x), \Psi(y)) \mapsto \textnormal{Tr}_{K/\mathbb{Q}} (\beta^{-1} x \bar y)$ defines a principal polarization on $\mathbb{C}^g/\Psi(\mathcal{O}_K)$ if and only if $\beta$ generates the different ideal $\mf D_K$, $\sigma(\beta) = -\beta$ and $\Im(\varphi(\beta)) > 0$ for every $\varphi \in \Psi$. This follows from the above; see also \cite{Wamelen99examplesof}.
\begin{comment}
\begin{lemma}
Consider the complex torus $B = \mathbb{C}^g/\Psi(\mathcal{O}_K)$. Let $\xi \in K$ and consider the corresponding rational Riemann form $\Psi(K) \times \Psi(K) \to \mathbb{Q}$, $(\Psi(x), \Psi(y)) \mapsto \textnormal{Tr}_{K/\mathbb{Q}} (\xi^{-1} x \bar y)$. Then $\xi$ defines a principal polarization on $B$ if and only if
\begin{enumerate}
\item $\xi$ generates the different ideal $\mf D_K$,
\item $\sigma(\xi) = - \xi$, and
\item $\Im(\varphi(\xi)) > 0$ for every $\varphi \in \Psi$.
\end{enumerate}
\end{lemma}
\begin{proof}
This follows from the above; see also \cite{Wamelen99examplesof}.
\end{proof}
\end{comment}
\begin{comment}
\begin{lemma} \label{lemma:selfdual}
Assume Condition \ref{condition:two}. Consider the skew-hermitian form $T : \Lambda \times \Lambda \to \mf D_K^{-1}$ and the hermitian form $h = \eta T: \Lambda_\mathbb{Q} \times \Lambda_\mathbb{Q} \to K$. Then $\mf h(\Lambda, \Lambda) \subset \mathcal{O}_K$, and
$$
\{x \in \Lambda_\mathbb{Q} : E(x, \Lambda) \subset \mathbb{Z} \}
= \{x \in \Lambda_\mathbb{Q} : T(x, \Lambda) \subset \mf D_K^{-1} \} = \{x \in \Lambda_\mathbb{Q} : \mf h(x, \Lambda) \subset \mathcal{O}_K \}.
$$
\end{lemma}
\begin{proof}
Indeed, $E(x, \Lambda) = \text{Tr}_{K/\mathbb{Q}} \left( T(x, \Lambda) \right) \subset \mathbb{Z}$ if and only if $T(x, \Lambda) \subset \mf D_K^{-1} = (\eta^{-1})$ if and only if $\mf h(x, \Lambda) = \eta T(x, \Lambda) \subset \mathcal{O}_K$.
\end{proof}
\end{comment}
\begin{comment}
\begin{examples} \label{examples:someexx}
\begin{enumerate}
\item Suppose that $K/\mathbb{Q}$ is an imaginary quadratic field $K = \mathbb{Q}(\sqrt \Delta)$ with discriminant $\Delta$. We view $K$ as a subfield of $\mathbb{C}$ via an embedding $\tau$ and require $\tau(\sqrt \Delta)$ to have positive imaginary part. The primes that ramify are the primes dividing $\Delta$; suppose that $-\Delta = \prod_{i = 1}^kp_i$, then for all $i$ there is a prime ideal $\mf p_i \subset \mathcal{O}_K$ such that $p_i\mathcal{O}_K = \mf p_i^2$. It follows that $\mf D_K = \prod_i \mf p_i = \left(\sqrt \Delta \right)$. Let $E : \Lambda \times \Lambda \to \mathbb{Z}$ be an alternating form such that $E(ax,y) = E(x, a^\sigma y)$ and let $T$ be the corresponding skew-hermitian form, see also Example \ref{examplesunitary}.\ref{ex:unitone}. Let $\eta = \sqrt \Delta$. Then $h = \sqrt \Delta T: \Lambda \times \Lambda \to \mathcal{O}_K$ is the corresponding hermitian form, and $\Lambda$ is self-dual for $E$ if and only if it is self-dual for $\mf h$. Moreover, $\Im \left( \tau(\sqrt \Delta) \right) > 0$ by construction.
\item \label{ex:unittwoex2} Let $K = \mathbb{Q}(\zeta)$ where $\zeta = \zeta_p = e^{2 \pi i/p} \in \mathbb{C}$ for some prime number $p > 2$. Consider $E: \Lambda \times \Lambda \to \mathbb{Z}$, $T_0 : \Lambda \times \Lambda \to \mathcal{O}_K$ and $T: \Lambda_\mathbb{Q} \times \Lambda_\mathbb{Q} \to K$ as in Example \ref{examplesunitary}.\ref{ex:unittwo}. Now observe that $\mathfrak p = (\zeta - \zeta^{-1}) \subset \mathcal{O}_K$ is prime because its norm $N_{K/\mathbb{Q}}(\mf p)$ is $(p) \subset \mathbb{Z}$. Since $(p) = N_{K/\mathbb{Q}}(\zeta - \zeta^{-1})$ and every conjugate of $\zeta - \zeta^{-1}$ is the product of $\zeta - \zeta^{-1}$ with a unit in $\mathcal{O}_K$, the ramification index $e(\mf p|p)$ equals $p-1$.
It follows that $\mf D_K = \mf p^{p-2} = (\zeta - \zeta^{-1})^{p-2} = (p) \mf p^{-1}$. Therefore, if we define $\eta = p / (\zeta - \zeta^{-1})$, then
$$
h: \Lambda \times \Lambda \to \mathcal{O}_K, \;\;\; \mf h(x,y) = \eta T(x,y) = \frac{p \cdot T(x,y)}{\zeta - \zeta^{-1}} = \frac{T_0(x,y)}{\zeta - \zeta^{-1}}
$$ is our required hermitian form. Moreover, define $\tau_i : K \to \mathbb{C}$ by $\tau_i(\zeta) = \zeta^{-i}$. Then
$\Im \left( \tau_i(\frac{1}{\zeta - \zeta^{-1}}) \right) > 0$ for $1 \leq i \leq (p-1)/2$.
\item More generally, suppose that $K = \mathbb{Q}(\zeta_m)$ for some integer $m > 2$. Then the element $\eta^{-1}_0$ below generates $\mf D_K$ with $\sigma(\eta^{-1}_0) = - \eta^{-1}_0$ (see \cite{li2021data}):
\begin{enumerate}
\item If $m$ is a prime number, then
$
\eta^{-1}_0 = m/ \left(\zeta_m^{(m+1)/2} - \zeta_m^{(m-1)/2} \right)$ by the above.
\item If $m = 2^k$, then $\eta^{-1}_0 = -2^{k-1}i$.
\item If $m = 3^k$, then $\eta^{-1}_0 = -3^{k-1} \sqrt 3 i$.
\item If $m = pq$ with $p,q$ distinct odd primes, then
$$
\eta^{-1}_0 = m\frac{\zeta_m^{(m+1)/2} - \zeta_m^{(m-1)/2}}{
\left(\zeta_m^{(q(p+1))/2} - \zeta_m^{q(p+1)/2}\right) \left(\zeta_m^{p(q+1)/2} - \zeta_m^{(p(q-1))/2}\right)}.
$$
\end{enumerate}
\end{enumerate}
\end{examples}
\end{comment}
\begin{comment}
\begin{corollary} \label{cor:abeliansubvariety}
Suppose that the different $\mf D_K$ is generated by an element $\eta \in \mathcal{O}_K$ such that $(\Psi, \eta)$ is a canonical tuple (Definition \ref{def:embeddingselement}). Suppose that the signature of $\mf h = \eta T$ equals $(r_1,s_1) = (n,1)$, $(r_i,s_i) = (n+1,0)$ for $2 \leq i \leq g$.
Under the bijection $\widetilde{\textnormal{Sh}}_{K}(\mf h) \cong \mathbb{C} H^n$ of Proposition \ref{prop:canonicalbijection}, the set $\mr H \subset \mathbb{C} H^n$ corresponds to isomorphism classes of polarized marked $\mathcal{O}_K$-linear abelian varieties $(A,i,\lambda,j)$ that admit an $\mathcal{O}_K$-linear embedding $\mathbb{C}^g/\Psi(\mathcal{O}_K)=B \hookrightarrow A$
of polarized abelian varieties. $\hfill \qed$
\end{corollary}
\end{comment}
\begin{theorem} \label{th:conditionsimplyhypothesis}
Suppose that the CM type $(K, \Psi)$ is primitive,
$\mf D_K = (\eta)$ for $\eta \in \mathcal{O}_K$ such that $\sigma(\eta) = - \eta$,
and
the signature of $\mf h^{\tau_i}$ is $(n,1)$ for $i = 1$ and $(n+1,0)$ for $i \neq 1$.
Let $r_1,r_2 \in \Lambda$ satisfy $H_{r_1} \cap H_{r_2} \neq \emptyset$ and $H_{r_1} \neq H_{r_2} \subset \bb C H^n$ for $H_{r_1} = (r_i)_\mathbb{C}^\perp \subset \mathbb{C} H^n$. Then $\mf h(r_1,r_2) = 0$.
\end{theorem}
\begin{proof}
Let $[x] \in H_r \cap H_t \subset \bb C H^n(V)$ and let $A$ be the abelian variety attached to $[x]$. Define $B$ to be the principally polarized abelian variety $\mathbb{C}^g/\Psi(\mathcal{O}_K)$. By Proposition \ref{prop:HcorrespondsNonSimple}, the roots $r$ and $t$ induce $\mathcal{O}_K$-linear embeddings $\phi_1: B \hookrightarrow A$ and $\phi_2 : B \hookrightarrow A$ of polarized abelian varieties.
By Lemma \ref{lemma:abeliansplit} below, each $\phi$ induces a decomposition
$
A \cong B \times C_i$ as polarized abelian varieties.
Note that $B$ is non-decomposable as an abelian variety because $\textnormal{End}(B) \otimes_\mathbb{Z} \mathbb{Q} = K$ is a field (here we use that the CM type $(K, \Psi)$ is primitive).
By \cite{debarrepolarisations}, the decomposition of $(A, \lambda)$ into non-decomposable polarized abelian subvarieties is unique, in the strong sense that if $(A_i, \lambda_i)$, $i\in \{1, \dotsc, r\}$ and $(B_j, \mu_j)$, $j \in \{1, \dotsc, m\}$ are polarized abelian subvarieties such that the natural homomorphisms $\prod_i(A_i, \lambda_i) \to (A, \lambda)$ and $\prod_j(B_j, \lambda_j) \to (A, \lambda)$ are isomorphisms, then $r = m$ and there exists a permutation $\sigma$ on $\{1, \dotsc, r\}$ such that $B_j$ and $A_{\sigma(j)}$ are \textit{equal} as polarized abelian subvarieties of $(A, \lambda)$, for every $j \in \{1, \dotsc, r\}$. Consequently, if we write $B_i = \phi_i(B) \subset A$, then either $B_1 = B_2 \subset A$ or $B_1 \cap B_2 = \{0\}$. Suppose first that $B_1 = B_2$. Then $\mathcal{O}_K\cdot r = \phi_1(\mathcal{O}_K) = \phi_2(\mathcal{O}_K) = \mathcal{O}_K \cdot t \subset \Lambda$. Consequently, $r = \lambda t$ for some $\lambda \in \mathcal{O}_K^\ast$; but then $H_r = H_t$ which is absurd. Therefore, we must have
$A \cong B_1 \times B_2 \times C$ as polarized abelian varieties for some polarized abelian subvariety $C$ of $A$. This implies that $H^{-1,0} = \text{Lie}(A) \cong \textnormal{Lie}(B_1) \times \textnormal{Lie}(B_2) \times \textnormal{Lie}(C)$, which is orthogonal for the positive definite hermitian form $iE_\mathbb{C}(x,\bar y)$ on $H^{-1,0}$. On the other hand, we have that $r_\tau = r^{-1,0}_\tau \in H^{-1,0}_\tau$ and $t_\tau = t^{-1,0}_\tau \in H^{-1,0}_\tau$: see the proof of Proposition \ref{prop:HcorrespondsNonSimple}. Moreover, by Lemma \ref{lemma:agree} we have
$$
\mf h(r,t) = \mf h^\tau(r_\tau, t_\tau) = \tau(\eta)\cdot T^\tau_\mathbb{C}(r_\tau, t_\tau) = \tau(\eta) \cdot E_{\mathbb{C}}(r_\tau, \bar{t_\tau}) = \tau(\eta) \cdot E_\mathbb{C} (r_\tau^{-1,0}, \overline{t_\tau^{-1,0}})
$$
so it suffices to show that $i E_\mathbb{C}(r_\tau^{-1,0}, \overline{t_\tau^{-1,0}}) = 0$. But $r_\tau^{-1,0} \in \textnormal{Lie}(B_1)$ and $t_\tau^{-1,0} \in \textnormal{Lie}(B_2)$.
\end{proof}
\begin{lemma} \label{lemma:abeliansplit}
Let $A$ be an abelian variety over a field $k$, with polarization $\lambda: A \to \widehat A$. Let $B \subset A$ be an abelian subvariety such that the polarization $\mu = \lambda|_B$ is principal. There exists a polarized abelian subvariety $Z \subset A$ such that $A \cong B \times Z$ as polarized abelian varieties.
\end{lemma}
\begin{proof}
Let $W = \textnormal{Ker}(A \xrightarrow{\lambda} \widehat A \to \widehat B )$. Let $Z = W^0_{\text{red}}$. Then $Z$ is an abelian subvariety of $A$, and has dimension $\dim(A) - \dim(B)$. The kernel of the natural homomorphism $B \times Z \to A$ is contained in $(B \cap Z) \times (B \cap Z)$; but $B \cap Z \subset B \cap W = \{0\}$ because $\mu: B \to \widehat B$ is an isomorphism. Therefore the natural homomorphism $B \times Z \to A$ is an isomorphism.
\end{proof}
Finally, we remark that the condition on the different $\mf D_K$ in Theorem \ref{th:conditionsimplyhypothesis} is quite often satisfied:
\begin{lemma} \label{lemma:discr}
Suppose that $K/\mathbb{Q}$ is imaginary quadratic or that $K = \mathbb{Q}(\zeta_n)$ for some integer $n \geq 3$.
Then $\mf D_K = (\eta) \subset \mathcal{O}_K$ for some element $\eta \in \mathcal{O}_K$ such that $F = K(\eta)$ and $\sigma(\eta) = -\eta$.
\end{lemma}
\begin{proof}
If $K/\mathbb{Q}$ is imaginary quadratic with discriminant $\Delta$,
then $\mf D_K
= (\sqrt \Delta)$ and the assertion is immediate.
Let $n \geq 3$ be an integer and consider $K = \mathbb{Q}(\zeta_n)$ and $F = \mathbb{Q}(\alpha)$ with $\alpha = \zeta_n + \zeta_n^{-1}$.
Since $\mathcal{O}_K = \mathbb{Z}[\zeta_n]$ by \cite[I, Proposition 10.2]{Neukirch}, we have $ \mathcal{O}_K = \mathcal{O}_F[\zeta_n]$. Notice that $f(x) = x^2 - \alpha x + 1 \in \mathcal{O}_F[x]$ is the minimal polynomial of $\zeta_n$ over $F$. We have $f'(\zeta_n) = 2 \zeta_n - \alpha \zeta_n = \zeta_n - \zeta_n^{-1}$. Therefore,
$\mf D_{K/F} = \left( f'(\zeta_n) \right) = \left(\zeta_n - \zeta_n^{-1}\right)$ \cite[III, Proposition 2.4]{Neukirch}. By \cite{Liang1976}, we know that $\mathcal{O}_F = \mathbb{Z}[\alpha]$. If $g(x) \in \mathbb{Z}[x]$ is the minimal polynomial of $\alpha$ over $\mathbb{Q}$, then $\mf D_{F/\mathbb{Q}} = (g'(\alpha))$. Since $\mf D_{K/\mathbb{Q}} = \mf D_{K/F} \mf D_{F/\mathbb{Q}}$ \cite[III, Proposition 2.2]{Neukirch}, we obtain:
$$
\mf D_{K/\mathbb{Q}} = \mf D_{K/F} \mf D_{F/\mathbb{Q}} = \left(\zeta_n - \zeta_n^{-1} \right) \cdot \left( g'(\alpha) \right) = \left( (\zeta_n - \zeta_n^{-1}) g'(\alpha) \right).
$$
\end{proof}
\begin{remark} \label{remark:avoidcondition}100
It would be more natural to attach an orthogonal hyperplane arrangement $\ca H \subset \mathbb{C} H^n$ to every primitive CM field $K$ and integral hermitian form $\mf h$ of hyperbolic signature
in such a way that $\ca H = \mr H = \cup_{\mf h(r,r) = 1} r_\mathbb{C}^\perp$ in case $\mf D_K = (\eta)$ for some $\eta \in \mathcal{O}_K$ such that $\sigma(\eta) = - \eta$. It appears that this can be done. The idea is as follows. Consider our CM field $K$. Choose $\beta \in \mathcal{O}_K - \mathcal{O}_F$ such that $\beta^2 \in \mathcal{O}_F$; then choose a CM type $\Psi = \{\tau_i: K \to \mathbb{C}\}_{1 \leq i \leq g}$ such that $\Im(\tau_i(\beta)) > 0$ for all $i$. Let $\mf h$ be a non-degenerate hermitian form $\Lambda \times \Lambda \to \mf D_K^{-1}$ such that $\textnormal{sign}(\mf h^{\tau}) = (n,1)$ and $\textnormal{sign}(\mf h^{\tau_i}) = (n+1,0)$ for $i \neq 1$.
Let $\mr S$ be the set of fractional ideals $\mf a \subset K$ for which there exist an element $b \in \mathcal{O}_F$ such that $\mf D_K \mf a \overline{\mf a} = (b\beta)$. By \cite[Theorem 4]{Wamelen99examplesof}, $\mr S$ is not empty.
For $\mf a \in \mr S$, define $\eta = b \beta \in \mathcal{O}_K$ and consider the complex torus $B = \mathbb{C}^g/\Psi(\mf a)$. It is equipped with the Riemann form $Q: \Psi(\mf a) \times \Psi(\mf a) \to \mathbb{Z}$, $(x,y) \mapsto \textnormal{Tr}_{K/\mathbb{Q}}( \eta^{-1} x \bar y)$,
and $Q$ defines a principal polarization on $B$ \cite[Theorem 3]{Wamelen99examplesof}. Let $\mr R$ be the set of embeddings $\phi: \mf a \to \Lambda$, $\mf a \in \mr S$, such that $\mf h(\phi(x), \phi(y)) = x \bar y$ for all $x,y \in \mf a$. For $\phi \in \mr R$, one obtains a hyperplane $H_\phi = \{x \in \mathbb{C} H^n: \mf h^{\tau}(x, \phi(\mf a)) = 0 \} \subset \mathbb{C} H^n$. The sought-for hyperplane arrangement $\ca H \subset \mathbb{C} H^n$ is defined as $\ca H = \cup_{\phi \in \mr R} H_\phi$. Indeed, if the CM type $(K, \Psi)$ is primitive, then $\ca H$ is an orthogonal arrangement by arguments similar to those used to prove Proposition \ref{prop:HcorrespondsNonSimple} and Theorem \ref{th:conditionsimplyhypothesis}.
\end{remark}
\end{document} |
\section{Introduction}
Businesses are increasingly turning to crowdsourcing as a means of obtaining knowledge, labour, and creativity from groups of distributed workers. Workers in crowdsourcing are rarely perfect, and the literature is abundant with papers that put forth various vote aggregation schemes to ensure quality control for the crowdsourced labels (e.g.,~\cite{chilton2013cascade,dai2013pomdp,mortensen2013crowdsourcing}). In practice, though, simple aggregation schemes, such as majority voting or its variations that require a stronger majority, are widely used, mainly due to their simplicity.
A commonly encountered variation of majority voting asks for multiple votes, until the positive votes are $\delta$ votes more than the negatives (or vice versa). We will refer to this process as \textit{$\delta$-margin majority voting}.\footnote{We use the term \emph{margin} in its colloquial meaning -- an amount beyond the minimum necessary (as opposed to its usage in statistics, where it typically connotes an idea of proportion and not difference).} While this consensus schemes of ``$\delta$-margin voting'' is salient in practice, the process tends to be perceived as heuristic and adhoc, lacking a solid theoretical foundation that describes its properties.
In this work, we show that a $\delta$-margin voting process can be easily modeled with well-established mathematical tools (namely, Markov chains with absorbing states). Our research contribution is simple: We provide a thorough theoretical description of key properties of $\delta$-margin voting, allowing decision-makers to give ex-ante answers to some of the more pervasive questions in crowdsourced experiment design:
\textit{How to structure a crowdsourced voting process to achieve a given label accuracy? What is the cost of running such a process? Can the same accuracy be achieved by a pool of lower accuracy workers?} We provide closed-form theoretical results that quantify the quality and cost of the results.
The organization of this paper is as follows. Section~\ref{sec:model} introduces the Markov chain formalism for the $\delta$-margin voting process. Section~\ref{sec:litrev} overviews existing work on consensus aggregation design and quality assurance, highlighting both the novelty and relevance of our theoretical presentations. Section~\ref{sec:characteristics} derives the theoretical equations for key characteristics of the $\delta$-margin voting process: quality of the results, expected value and variance of the number of votes required to reach consensus, and the overall distribution for time to completion. Section~\ref{sec:equiv} expands these results to demonstrate that $\delta$-margin processes with workers of different quality can have equivalent quality of results, and we further elaborate on this equivalence to propose a quality-sensitive payment scheme that connects result quality with worker accuracy. Section~\ref{sec:theory_vs_exp} presents experimental validation with real data comparing the theoretical predictions against results from real processes. Finally, Section~\ref{sec:future} discusses both the immediate and the more involved research developments that can be built upon this work.
\section{Markov chain model of $\delta$-margin majority voting}
\label{sec:model}
Consider the problem of labeling items on a crowdsourcing platform using \emph{$\delta$-margin voting}. For every item that needs to be assigned a binary label. We solicit binary votes from workers who assign the correct label with a probability $p$ and the incorrect label with probability $1-p$. We keep asking workers to assign labels to an item, until the absolute difference between the number of votes for the two classes exceeds a predefined \textit{consensus threshold} $\delta$.
\textbf{Example.} Set $\delta=2$. Then, the process will stop when the vote-count tuples $\langle n_{\textit{correct}}, n_{\textit{incorrect}}\rangle$ reach one of the following stages:
$\langle 2, 0\rangle$,
$\langle 0, 2\rangle$,
$\langle 3, 1\rangle$,
$\langle 1, 3\rangle$,
$\langle 4, 2\rangle$,
$\langle 2, 4\rangle$, and so on. A consensus vote obtained in one of the states $\{
\langle 2, 0\rangle, \langle 3, 1\rangle,
\langle 4, 2\rangle, ...\}$ is correct, while a consensus vote obtained in one of the states $\{\langle 0, 2\rangle, \langle 1, 3\rangle, \langle 2, 4\rangle, ...\}$ is incorrect.
Since the only desideratum for consensus is the difference between the two types of votes, the process can be modeled as a Markov random walk. Define the current state as the difference between the numbers of correct and incorrect votes. If the difference is $\delta$ or $-\delta$, the process terminates, the former resulting in a correct consensus label, and the latter in an incorrect one. In all other states, we procure an additional vote, which will be correct with probability $p$, which corresponds to a transition from state $i$ to state $i+1$; or it can be an incorrect vote with probability $1-p$, which makes for a transition from state $i$ to state $i-1$.
\begin{figure}[t]
\centering
\includegraphics[width=\columnwidth]{figs/diagram.pdf}
\caption{\rm{A Markov chain diagram illustrating state transitions for $\delta$-margin majority voting on a single item. Node labels indicate the difference between the numbers of correct ($n_1$) and incorrect ($n_0$) votes, $n_1-n_0$. Consensus is reached in absorbing states $\delta$ and $-\delta$, the former resulting in a correct
resolution, and the latter resulting in an incorrect resolution.}}
\label{fig:markov}
\end{figure}
This process has a state diagram illustrated in Figure~\ref{fig:markov} and is also known as \emph{Gambler's Ruin} model. The model is a common introductory example for random walks and describes the probability of a gambler winning a certain amount in a game of chance, vs.\ the probability of losing her entire gaming budget.\footnote{See~\cite[page~344]{FellerTextbook} for details.} Surprisingly, this model has not been used to describe the process for label aggregation in crowdsourcing, and we are not aware of anyone providing results about label aggregation quality and number of votes required, for this process in crowdsourcing. Before providing, in Section~\ref{sec:characteristics}, the analytical results that describe in closed form several key characteristics of $\delta$-margin voting, we provide next, in Section~\ref{sec:litrev} an overview of the related literature.
\section{Literature review}
\label{sec:litrev}
In this section, we provide a brief overview of related work in label aggregation in crowdsourcing (Section~\ref{sec:relwork-quality}) and then provide pointers to related work regarding $\delta$-margin voting, in particular (Section~\ref{sec:relwork-delta-margin}). To the best of our knowledge, there does not yet appear to be work that focuses on presenting theoretical results for ex-ante cost and quality estimation for $\delta$-margin voting.
\subsection{Literature on quality control in label aggregation}
\label{sec:relwork-quality}
A sizable proportion of literature on crowdsourcing proposes various quality controls for aggregating worker expertise. The majority of such work, however, is experimental \citep{kazai2011crowdsourcing, hansen2013quality, yin2014monetary}, while the proposed quality control mechanisms are often ad-hoc. Although informed by empirical intuition, such quality control interventions lack theoretical guarantees \citep{kucherbaev2016relauncher, de2017efficiently, dai2013pomdp}. Moreover, a significant fraction of this literature relies on availability of accurate priors on some of the key process or workforce parameters \citep{abassi2017adaptive, tao2018domain, heer2010crowdsourcing, dalvi2013aggregating, laureti2006information, jung2011improving, rutchick2020does}, which are costly to obtain \citep{de2017crowd, bonald2016minimax}.
Within the larger domain of literature concerned with the quality of the outcomes and process costs in crowdsourcing, a small but notable subspace of work offers theoretical guarantees.
The work of \cite{berend2014consistency} presents accuracy bounds for the weighted majority voting scheme.
\cite{khetan2016achieving} formulate a theoretical bound that describes the trade-off between budget
and accuracy in voting processes with \emph{adaptive task assignment} (i.e., tasks are assigned based on the data collected up until the moment of assignment, \cite{barowy2012automan}). The authors introduce an adaptive assignment scheme that attains the formulated fundamental limit.
Similarly, \cite{manino2018efficiency} consider adaptive assignment of the next worker to an item. The authors' theoretical formulation of the accuracy gap between the uniform allocation of workers \citep{karger2014budget}, an adaptive one, and an assignment that maximizes information gain \citep{simpson2015bayesian}. They derive tight yet not exact bounds on the accuracy of these assignment policies. \cite{livshits2014saving} is an example of a theoretical work that considers costs of voting completion: the authors use power analysis to obtain ex-ante estimates for the number of votes required to resolve each item with a certain level of statistical significance.
In contrast with these papers, our work focuses on the $\delta$-margin consensus rule and provides an exact expression of probability of error, expected time until consensus as well as its variance, enabling a more nuanced consideration of the requesters' utility -- one that considers worker \emph{wages}.
While the latter is seldom addressed in the literature, a notable counterexample is the work of \cite{singer2013pricing}. The authors present
incentives-compatible mechanisms for maximizing the number of tasks subject to budget constraints, as well as for minimizing worker payments
given a fixed number of tasks.
\subsection{Literature on $\delta$-margin voting}
\label{sec:relwork-delta-margin}
Literature on group decision-making and voting mechanisms recognizes two types of majority consensus vote: \emph{simple} majority and \emph{absolute} majority. The former is attained when the number of votes for option $A$ is greater than the number of votes for the alternative option $B$ ($s_A>s_B$), while the latter requires that the number of votes for option $A$ is greater than half of all votes ($s_A>\frac{n}{2}$)~\citep{laruelle2011majorities}.\footnote{Alternatively, the strength of majority condition is defined as a multiplier $k$ on the required inequality requirements that are modified to $s_A>k \cdot s_B$ for $k>1$, and $s_A>k \cdot n$ for $\frac{1}{2} \leq k \leq 1$. .} In the case of binary (or \emph{dichotomic}) voting, the distinction between simple and absolute majorities only exists when some voters abstain or vote neutrally.
While the $\delta$-margin voting aggregation method is a salient choice among practitioners, crowdsourcing literature typically involves $\delta$-margin voting only in empirical or experimental settings, missing a chance for ex-ante theoretical evaluation of its benefits.\footnote{Symptomatically, even outside of the crowdsourcing literature field the more notable papers that overview some of the theoretical properties of this consensus scheme refer to $\delta$-margin majority voting as `the forgotten decision rule'~\citep{llamazares2006forgotten, garcia2001majority}. Some of the earliest -- albeit brief -- acknowledgements of the $\delta$-margin voting scheme can be found in~\cite{fishburn2015theory} and~\cite{saari1990consistency}.}
In experimental comparisons of prototypical crowdsourcing tasks,~\cite{de2017crowd} remarks that while the $\delta$-margin method (called `Beat-By-$K$' following~\cite{goschin2014stochastic}) yields very accurate results for relatively high values of $\delta$, it is expensive to run in settings that prioritize utility without any budget constraints.
Several consensus aggregation methods in the literature capture a different but similar idea to $\delta$-margin voting -- namely, requiring a stronger, more confident agreement among the workers. For example, in the `Automan' scheme \citep{barowy2012automan} the requester continues sampling votes until a given statistical confidence value is attained by the voting process.
\cite{dietrich2007judgment} generalize majority voting to a larger family of `quota rules' (also known as `$k$-unanimity' or `$k$-majority' in cybernetics and discrete mathematics \citep{alon2006dominating, scheidler2015k}), in which an item is assigned a certain label if and only if the number of workers voting for that label exceeds a given threshold $k$. In a situation of simultaneous vote by all workers within a fixed size workforce, the quota scheme becomes similar to $\delta$-margin voting, with two major differences: (a)~there is no guarantee that at the end of the voting the margin will be satisfied, and (b)~the process may accumulate more votes than necessary, a disadvantage when every vote has an associated cost (as in the case of paid crowdsourcing).
\section{Theoretical characteristics of $\delta$-margin voting processes}
\label{sec:characteristics}
In this section, we provide closed-form theoretical formulations of key characteristics of the $\delta$-margin voting process. In particular, we employ well-known results from theory of Markov Chains to provide:
\begin{itemize}
\item a characterization of \emph{resulting label quality} for $\delta$-margin majority crowdsourcing processes;
\item the \emph{expected number of votes} to reach consensus is attained (i.e., \emph{expected time} until consensus);
\item the \emph{variance of number votes} to reach consensus;
\item an expression for the \emph{pdf} of time until consensus.
\end{itemize}
\subsection{Quality of the consensus vote}
\label{sec:quality}
The first quantity of interest is the quality of the final consensus votes. The process always starts at state $0$ (no votes). Let's denote as $\mathbbm{1}_C$ a random variable indicator of whether the consensus vote (in the event that it was reached) corresponds to the ground truth correct label of an item. We want to calculate the probability $Q:=P(\mathbbm{1}_C=1)$ of reaching terminal state $\delta$ (i.e., voters reach consensus, and the correct label is assigned) vs.\ reaching terminal state $-\delta$ (voters reach consensus, and the incorrect label is assigned, i.e. $\mathbbm{1}_C=0$). In all other states, the voting continues.
\vspace{0.25in}
\begin{theo}
\label{th:Q_nonrand}
For a $\delta$-margin voting process, with worker accuracy $p$ and a consensus threshold $\delta$, the probability $Q(\varphi,\delta)$ that the consensus vote $c$ is correct, is:
\begin{equation}
\label{eq:Q_nonrand}
Q(\varphi,\delta) :=
P(\mathbbm{1}_C=1)=
\frac{\varphi^{\delta}}{1+\varphi^{\delta}},
\end{equation}
\noindent
where $\varphi=\frac{p}{1-p}$ are the \emph{odds} of a single worker classifying the item correctly.
\end{theo}
\vspace{0.15in}
The proof is readily available in many introductory texts on Markov Chains (e.g., in~\cite[page~344]{FellerTextbook}) and we also provide it below, for completeness, and to also introduce notation that we use later in this paper.
\textbf{Proof.}
Consider an asymmetric Gambler's Ruin model, where a gambler begins with the amount $k$ of money and proceeds to gamble until she is either bankrupt, or wins the amount $N>k$. Let us formalize this model in the following way. Consider a series of random variables $X_t, t\in\mathbb{N}_0, X_t=k+\xi_1+...+\xi_t$, where
\begin{equation}\nonumber
\xi_t=\begin{cases}
1 & \text{with probability } p.\\
0 & \text{with probability } 1-p.\end{cases}
\end{equation}
We shall denote the \textit{time of termination} of the game of a gambler who began with $k$ amount of money (or $k$ vote difference) as $T_k:=\inf t, t\ge0: X_t\in\{0,N\}$.
Consider the following identity:
$$\mathbb{E}[S^{X_t}|X_{t-1},X_{t-2},...,X_1]=pS^{X_{t-1}+1}+(1-p)S^{X_{t-1}-1}.$$
A \textit{martingale} is a process such that:
$$\mathbb{E}[Z_t|Z_{t-1},...,Z_1]=Z_{t-1}.$$
Let us choose $S$ such that $\{S^{X_t}\}_{t\ge0}$ is a martingale. Namely, take $S=\frac{1-p}{p}$ (note that the solution $S=1$ is trivial). Using the fact that any martingale satisfies $\mathbb{E}[Z_t]=\mathbb{E}[Z_0],$ we obtain the equality
$$\mathbb{E}[S^{X_t}]=\mathbb{E}[S^{X_0}]=S^k.$$
In particular, $\mathbb{E}[S^{X_{T_k}}]=S^k$. At the same time,
\begin{multline}\mathbb{E}[S^{X_{T_k}}]=S^NP(X_{T_k}=N)+S^0(1-P(X_{T_k}=N))=S^NP(X_{T_k}=N)+(1-P(X_{T_k}=N))\end{multline}
Therefore,
\begin{equation}
\label{equ:prob_x_tk}
P(X_{T_k}=N)=\frac{S^k-1}{S^N-1} \text{\hspace{0.2cm} when } p\ne 1/2.
\end{equation}
In the context of a symmetric voting process, the initial value $k$ is set equidistantly apart from the ruin value of zero and from the winning value $N$. The number of votes needed for either a consensus for the incorrect answer (ruin) or a consensus for the correct answer (winning) is $\delta$. Thus, substituting $\delta$ for $k$ and $2\delta$ for $N$ while recalling our choice of $S=\frac{1-p}{p}=\frac{1}{\varphi}$, we obtain the statement of the theorem. $\square$
Figure~\ref{fig:Q_of_p} plots the dependence of consensus vote quality $Q$ on worker accuracy $p$ for $\delta \in \{2, 3, 4, 5\}$.
\begin{figure}[t]
\centering
\includegraphics[width=0.8\columnwidth]{figs/Q_of_p.png}
\caption{\rm{Theoretical values of quality $Q$ of consensus vote (Theorem~\ref{th:Q_nonrand}) as a function of the probability of a correct answer $p$, for a fixed consensus threshold $\delta$.}}
\label{fig:Q_of_p}
\end{figure}
\paragraph{\textbf{Discussion:}} Note that the parameter $\delta$ plays a significant role in the quality of the consensus vote: Based on Equation~\ref{eq:Q_nonrand}, the \emph{odds} of the \emph{consensus} vote being correct, are $\varphi^{\delta}$. Therefore by increasing $\delta$, we increase exponentially the odds that the consensus vote is correct.
For example, consider a pool of workers with accuracy $p=0.75$ (i.e., $\varphi=\frac{0.75}{1-0.75}=3$).
If we set $\delta=2$, the expected quality of the overall classification will be $Q(3,2)=0.9$ (i.e., odds 9 to 1 being correct). If we increase $\delta$ to $\delta=3$, $Q(3,3)=0.964$ (ie., odds 27 to 1), and if we increase to $\delta=4$, then $Q(3,4)=0.9878$ (i.e., odds 81 to 1). Next, we discuss how the cost of the process changes when we change $\delta$, and we show that we achieve exponential improvements in quality with a mostly linear increase in cost.
\subsection{Expected time until consensus}
\label{sec:votes}
The next question is whether the process is guaranteed to terminate, and how many votes do we expect to collect until reaching a consensus (in other words, what is the \emph{expected time} until consensus is reached).
We can estimate the number of votes it takes to reach state $\delta$ or state $-\delta$ in terms of transitions in the Markov chain as follows. The number of remaining steps $E_0$ from state $0$ is one step it takes to reach the next state (i.e., state 1 with probability $p$ or state -1 with probability $1-p$) plus the number of remaining steps from this resulting node. By solving, we get:
\vspace{0.25in}
\begin{theo}
\label{th:ET}
The expected number of votes $n_{votes}$ it takes to reach a (correct or incorrect) consensus when classifying an item using a $\delta$-margin voting scheme with worker accuracy $p$ and consensus threshold $\delta$ is:
\begin{equation}\label{eq:exp_votes}
\mathbb{E}[n_{votes}|\varphi,\delta]=\delta\cdot\frac{\varphi+1}{\varphi-1} \cdot \frac{\varphi^{\delta}-1}{\varphi^{\delta}+1},
\end{equation}
\noindent
where $\varphi=\frac{p}{1-p}$ is the \emph{odds} of a single worker classifying the item correctly.
\end{theo}
\vspace{0.15in}
\textbf{Proof.} Note that the expected number of votes needed to reach consensus when beginning at state $k$ of a Markov chain is nothing but $\mathbb{E}[T_k]$ in our earlier notation. Consider the equation
\begin{equation}\label{eq:exp_X}
\mathbb{E}[X_{T_k}]=P(X_{T_k}=N)\cdot N+(1-P(X_{T_k}=N))\cdot0=N\cdot P(X_{T_k}=N).
\end{equation}
Using Wald's equation for the expectation of a sum of random number of random variables, we obtain:
\begin{equation}\label{eq:wald}
\mathbb{E}[X_{T_k}]=k+\mathbb{E}\left[\sum_{i=1}^{T_k}\xi_i\right]=k+\mathbb{E}[T_k]\mathbb{E}[\xi_i]=
k+\left(\mathbb{E}[T_k]\right)(p-(1-p)).
\end{equation}
Combining (\ref{eq:exp_X}) with (\ref{eq:wald}) we get:
\begin{equation}\nonumber
\mathbb{E}[T_k]=\frac{N}{2p-1}P(X_{T_k}=N)-\frac{k}{2p-1}.\end{equation}
The statement of the theorem follows from making use of Equation~\ref{equ:prob_x_tk} and the identity $2p-1=\frac{\varphi-1}{\varphi+1}$, as well as performing the familiar substitutions: $N=2\delta$ and $k=\delta$. $\square$
\paragraph{\textbf{Discussion}:} Note that the expected cost increases \emph{mostly} linear with $\delta$. When $\varphi$ gets close to 1 (i.e., mostly random workers), then the expected time to termination peaks at $\delta^2$. Figure~\ref{fig:ET(d)} illustrates the dependence (\ref{eq:exp_votes}) of time until voting termination $\mathbb{E}(n_{votes})$ on worker accuracy $p$, for a given consensus threshold $\delta$.
\begin{figure}[t]
\centering
\includegraphics[width=0.8\columnwidth]{figs/E_T.png}
\caption{\rm{Expected time to voting termination as a function of probability of correct answer $p$, for a fixed consensus threshold $\delta$.}}
\label{fig:ET(d)}
\end{figure}
\subsection{Variance of time until consensus
\label{sec:var}
While the expected number of votes to completion is a useful metric for characterizing the voting process, it describes just the average across runs. In addition to the expectation, in reality we also want to know the robustness of a process and how reliably it will finish within the expected time frame. For this reason, it is important to know the \emph{variance} of the process in terms of number of votes required to completion and, more generally, the overall \emph{pdf} of the distribution. To derive these values, we rely on results from the theory of Markov Chains with absorbing states.
We showed that a $\delta$-margin voting process can be modeled as a Markov random walk. In particular, a $\delta$-margin voting process with a consensus threshold $\delta$ corresponds to a Markov chain process with two absorbing states and $2(\delta-1)+1$ transient states. If one enumerates the states such that the two absorbing states come last, then the transition matrix of this process has the canonical form~\citep{grinstead1997ch}:
\begin{displaymath}
\Large
\mathbf{M} = \left(\begin{smallmatrix}
\mathbf{T}_{tr} & \mathbf{T}_A \\
\textbf{0} & I_2
\end{smallmatrix}\right)
\end{displaymath}
Here $\mathbf{T}_{tr}$ is a matrix of size $(2\delta-1)\times(2\delta-1)$ containing the probabilities of transitioning from a transient state $i$ into a transient state $j$:
\begin{displaymath}
\Large
\mathbf{T}_{tr} = \left(
\begin{smallmatrix}
0 & p & 0 & 0 & \cdots & 0\\
q & 0 & p & 0 & \cdots & 0\\
0 & q & 0 & p & \cdots & 0\\
0 & 0 & q & 0 & \cdots & 0\\
\vphantom{\int\limits^x}\smash{\vdots} & \vdots & \vdots &\vdots & \ddots & \vdots\\
0 & 0 & 0 & 0 & q & 0\\
\end{smallmatrix}\right);
\end{displaymath}
$\mathbf{T}_{A}$ is a $(2\delta-1)\times2$ matrix of transition probabilities from transient into absorbing states:
\begin{displaymath}
\Large
\mathbf{T}_{A} = \left(
\begin{smallmatrix}
q & 0 & 0 & \cdots & 0 & 0\\
0 & 0 & 0 & \cdots & 0 & p\\
\end{smallmatrix}\right)^{\top};
\end{displaymath}
$\mathbf{I}_2$ is a $2\times 2$ identity matrix, and \textbf{0} is a $2\times((2\delta-1)+1)$ null matrix.
Using the above, we define the \emph{fundamental matrix} of this absorbing matrix chain as:
\begin{displaymath}
\Large \mathbf{N}:=\sum_{k=0}^{\infty}\left(\mathbf{T}_{tr}\right)^k
\end{displaymath}
It can be proven~\citep{grinstead1997ch} that:
\begin{displaymath}
\Large \mathbf{N}=\left(\mathbf{I}_{2\delta-1}-\mathbf{T}_{tr} \right)^{-1}
\end{displaymath}
The elements of $\mathbf{N}$ correspond to the expected number of visits to the state $j$, given that the initial state was $i$.
The expected number of steps before consensus, when starting in transient state $i$, is the $i$-th element of the vector
\begin{displaymath}
\mathbf{t} = \mathbf{N} \cdot \mathbf{1}
\end{displaymath}
where $\mathbf{1}$ is vector of ones of length $2\delta-1$. Finally, we have:
\vspace{0.25in}
\begin{propo}
The variance on the number of steps required to reach consensus when starting in a state $i$ is the $i$-th entry of vector
$$\bar{v} = (2\mathbf{N}-\mathbf{I}_{t})\mathbf {t} -\mathbf {t} _{\operatorname {sq} },$$
where $\mathbf {t}_{\operatorname{sq} }$ is the vector of squared elements of $\mathbf {t}$.
\end{propo}
\vspace{0.15in}
Applying this known result to our model for $\delta$-margin voting, we offer a closed-form expression for the variance of time until consensus:
\vspace{0.25in}
\begin{theo}
Variance of the number of votes it takes to reach consensus using $\delta$-margin voting process is:
\begin{equation}
\label{fla:var_nonrand}
Var(n_{votes}) = 4\delta\varphi\left(\frac{\varphi+1}{\varphi^\delta+1}\right)^2%
\left[
h(\delta)\cdot\varphi^{\delta-2} + %
\sum_{i=1}^{\delta-2}%
\Big(
h(\delta-i)(\varphi^{\delta+i-2}+\varphi^{\delta-i-2})
\Big)
\right],
\end{equation}
where coefficients $h$ are defined using the floor function:
$$h(z):=\left \lfloor{\frac{z^2}{4}}\right \rfloor. $$
\label{th:var_nonrand
\end{theo}
\vspace{0.15in}
Stated differently,
$$h(z) =\begin{cases}
\frac{z^2}{4}, & \text{if $z$ is even}\\
\frac{z-1}{2}\cdot\frac{z+1}{2}, & \text{if $z$ is odd.}
\end{cases}$$
The sequence of coefficients defined by $h(z)$ is also known as the \emph{quarter squares} sequence and can be defined as the interleaving of square numbers and pronic numbers \citep{losanitsch1897isomerie, sloane2019line} (the latter defined as the product of two consecutive integers).
Equation~\ref{fla:var_nonrand} offers a simplification of the formula proposed by \cite{andvel2012variance}, while agreeing with it numerically. Moreover, once the polynomial in square brackets is formulated for a given value of $\delta$, one can observe a simple and intuitive `pyramidal' pattern that governs coefficient generation.
In order to provide the reader with intuition on the look of the polynomial in Equation~\ref{fla:var_nonrand}, we list explicit formulas for $Var(n_{votes})$ for the first few values of $\delta$ in Table~\ref{tab:var_examples}.
\begin{table}[t]
\begin{center}
\bgroup
\def2.2{2.2}
\begin{tabular}{c|c}
$\bm{\delta}$ & $\mathbf{Var(n_{votes})}$ \\\hline\hline
2 & $8\varphi\left(\frac{\varphi+1}{\varphi^2+1}\right)^2 $ \\
\hline
3 & $12\varphi\left(\frac{\varphi+1}{\varphi^3+1}\right)^2 \left( \varphi^2 + 2\varphi + 1 \right) $ \\\hline
4 & $16\varphi\left(\frac{\varphi+1}{\varphi^4+1}\right)^2 \left( \varphi^4 + 2\varphi^3 + 4\varphi^2 + 2\varphi+ 1 \right) $ \\\hline
5 & $20\varphi\left(\frac{\varphi+1}{\varphi^5+1}\right)^2 \left( \varphi^6 + 2\varphi^5 + 4\varphi^4 + 6\varphi^3 + 4\varphi^2 + 2\varphi+ 1 \right) $ \\\hline
6 & $24\varphi\left(\frac{\varphi+1}{\varphi^6+1}\right)^2 \left( \varphi^8 + 2\varphi^7 + 4\varphi^6 + 6\varphi^5 + 9\varphi^4 + 6\varphi^3 + 4\varphi^2 + 2\varphi+ 1 \right) $
\\\hline
7 & $28\varphi\left(\frac{\varphi+1}{\varphi^6+1}\right)^2 \left( \varphi^{10} + 2\varphi^9 + 4\varphi^8 + 6\varphi^7 + 9\varphi^6 + 12\varphi^5 + 9\varphi^4 + 6\varphi^3 + 4\varphi^2 + 2\varphi+ 1 \right) $
\\\hline
\end{tabular}
\caption{Particular values of variance on the number of votes until consensus}
\label{tab:var_examples}
\egroup
\end{center}
\end{table}
\vspace{0.15in}
\begin{figure}[t]
\centering
\includegraphics[height=0.33\columnwidth]{figs/Var_th_croat.png}
\includegraphics[height=0.33\columnwidth]{figs/Var_T_over_E_T.png}
\caption{\rm{Left: $Var(n_{votes})$ (left) and $Var(n_{votes})/\mathbb{E}(n_{votes})$ (right) as a function of probability of correct answer $p$, for a fixed consensus threshold $\delta$ ($d$).}}
\label{fig:Var_by_d}
\end{figure}
Formulation in Equation~\ref{fla:var_nonrand} allows us to elaborate upon the plots in Figure~\ref{fig:ET(d)} by adding bands of the size $2\sqrt{Var(n_{votes})}$ to each trajectory of $\mathbb{E}[n_{votes}]$. Detailed plots for selected values of $\delta$ are shown in Figure~\ref{fig:ET_with_std}.
\begin{figure}[t]
\centering
\includegraphics[width=0.32\columnwidth]{figs/ET_std_d2.png}
\includegraphics[width=0.32\columnwidth]{figs/ET_std_d4.png}
\includegraphics[width=0.32\columnwidth]{figs/ET_std_d7.png}
\caption{\rm{Expected time until reaching consensus, with standard deviation bounds, for selected values of $\delta$.}}
\label{fig:ET_with_std}
\end{figure}
\subsection{Distribution of the time to consensus}
\label{sec:pdf}
Using the matrix notation used in the previous subsection, as well as known results about discrete phase-type distributions from queuing theory~\cite{neuts1994matrix, latouche1999introduction}, we can fully describe the probability density function $f(k)$ of time until consensus in $\delta$-margin voting.
\vspace{0.25in}
\begin{remo}
The event that a $\delta$-margin majority voting process will terminate after casting exactly $m$ votes when starting from zero votes is characterized by the following probability density function:
\begin{equation}
f(T_\delta = m) = \mathbf{z} \cdot (\mathbf{T}_{tr})^{m-1} \cdot (\mathbf{T}_A)
\label{eq:pdf_T}
\end{equation}
\noindent where $\mathbf{z}$ is a vector of length $2\delta-1$
that encodes the initial state of the process. Since we mainly care for the case where we start with no votes, we set
$\mathbf{z} :=(z_1, z_2, ..., z_{2\delta-1})$ with $z_{\delta} = 1$ and $z_i = 0$ for $i\ne \delta$.
\end{remo}
\vspace{0.15in}
The expression specifies separately the probability densities of the voting process terminating in each one of the two absorbing states -- i.e., reaching a consensus with a correct label, as well as with an incorrect one. Using our derived expression for the probability of label correctness (Equation~\ref{eq:Q_nonrand}), one may sum the values in the matrix expression of $f(T_\delta = m)$ in Equation~\ref{eq:pdf_T}, weighted by the respective probabilities of the consensus vote being correct/incorrect. Figure~\ref{fig:pdf_Tk} visualizes a result of this computation for a fixed selected threshold level $\delta = 4$ and for various values of $\varphi$.
\begin{figure}[t]
\centering
\includegraphics[width=0.7\columnwidth]{figs/pdf_Tk.png}
\caption{\rm{Probability density of reaching consensus in exactly $m$ steps (i.e., any consensus, correct or incorrect) $f(T_\delta = m)$ for $\delta=4$, as a function of the number of votes $m$, for varying worker odds $\varphi$.}}
\label{fig:pdf_Tk}
\end{figure}
\section{Equivalence conditions for two $\delta$-margin voting processes}
\label{sec:equiv}
In this section, we build in the results presented above, to explore the following questions:
\begin{itemize}
\item If we have two sets of workers, one with accuracy $p_1$ and the other with accuracy $p_2$, how should we configure the voting process to generate results of similar aggregate quality? (Section~\ref{sec:quality_equivalence})
\item If we have two sets of workers, one with accuracy $p_1$ and the other with accuracy $p_2$, how should we vary the worker payment, so that we get the same quality while keeping the overall cost the same? (Section~\ref{sec:pay_equiv})
\item If workers can vary their effort, and therefore the time invested and the resulting quality, what type of effort-quality curves result in unique payment-maximizing points, and what type of effort-quality curves make workers indifferent to the quality-sensitive payment scheme? (Section~\ref{sec:iso})
\end{itemize}
\subsection{Quality equivalence for processes with different values of $\delta$}
\label{sec:quality_equivalence}
Suppose that we have two sets of workers, one with accuracy $p_1$ and another with accuracy $p_2$. The first set of workers operate a $\delta$-margin majority voting scheme with threshold parameter $\delta_1$. Assuming, for now, that the employer only cares about the quality of the resulting work, what is the value of $\delta_2$, so that the set of workers with accuracy $p_2$ generate the same quality of the results?
\vspace{0.25in}
\begin{coro}
If an item is classified by workers with accuracy $p_1$ and threshold $\delta_1$, we can achieve the
same quality of results by a set of workers with accuracy $p_2$ if we set the threshold $\delta_2$ to be:
\begin{equation}
\label{fla:equiv}
\delta_2 = \delta_1 \cdot \frac{\ln{\varphi_1}}{\ln{\varphi_2}},
\end{equation}
where $\varphi_1=\frac{p_1}{1-p_1}$ and $\varphi_2=\frac{p_2}{1-p_2}$ are the \emph{odds} of a single worker classifying the item correctly. The result is obtained by setting $Q(\varphi_1,\delta_1) =Q(\varphi_2,\delta_2)$ (from Equation~\ref{eq:Q_nonrand}) and solving for $\delta_2$.
\end{coro}
\vspace{0.15in}
\subsection{Worker Pay Equivalence}
\label{sec:pay_equiv}
Assume that workers with accuracy $\varphi$ are paid $\textit{pay}(\varphi)$ per vote. Then a reasonable notion of cost associated with labeling a single example is simply:
$$ \textit{Cost}(\varphi,\delta) = \textit{pay}(\varphi) \cdot n_{votes}(\varphi,\delta),$$
where $pay(\varphi)$ is constrained to take only non-negative values. Then, using Equation~\ref{th:ET}, the expected cost of classifying an item is:
\begin{equation}
\mathbb{E}[\textit{Cost}|\varphi,\delta] = \textit{pay}(\varphi) \cdot \delta\cdot\frac{\varphi+1}{\varphi-1} \cdot \frac{\varphi^{\delta}-1}{\varphi^{\delta}+1}
\end{equation}
For now, we assume that an employer is \emph{risk-neutral} and cares only about the expected quality of the result and the expected cost.\footnote{It is reasonably easy to extend the results to the case of risk-averse requesters, by using our results for the variance of time until consensus (which is a proxy for the cost of the process).} A risk-neutral employer would like to pay teams of workers with different accuracies $p_1$ and $p_2$ in a way that is fair: as long as the teams can generate results of equal utility to the requester, they should be paid the same total amount.
From Section~\ref{sec:quality_equivalence}, we can set $\delta_2$ to adjust for the different worker accuracy $p_2$, thereby assuring that the quality of results is the same: $Q(\varphi_1,\delta_1)=Q(\varphi_2,\delta_2)$. Of course, as shown in Section~\ref{sec:votes}, a different consensus threshold $\delta_2$ also changes the expected number of votes required to reach consensus. When
$\delta_2 = \delta_1 \cdot \frac{\ln{\varphi_1}}{\ln{\varphi_2}}$, we have results of equal quality, we can ensure equal costs by setting:
\begin{eqnarray}
\mathbb{E}[\textit{Cost}|\varphi_1,\delta_1]
& = &
\mathbb{E}[\textit{Cost}|\varphi_2,\delta_2] \nonumber \\
\mathbb{E}[\textit{Cost}|\varphi_1,\delta_1]
& = &
\mathbb{E}[\textit{Cost}|\varphi_2,
\delta_1 \cdot \frac{\ln{\varphi_1}}{\ln{\varphi_2}}
] \nonumber \\
\textit{pay}(\varphi_1) \cdot \delta_1 \cdot
\frac{\varphi_1+1}{\varphi_1-1}
\cdot
\frac{\varphi_1^{\delta_1}-1}{\varphi_1^{\delta_1}+1}
& = &
\textit{pay}(\varphi_2) \cdot
\delta_1 \cdot \frac{\ln{\varphi_1}}{\ln{\varphi_2}}
\cdot
\frac{\varphi_2+1}{\varphi_2-1}
\cdot
\frac{\varphi_2^{\delta_1 \cdot \frac{\ln{\varphi_1}}{\ln{\varphi_2}}} -1 }
{\varphi_2^{\delta_1 \cdot \frac{\ln{\varphi_1}}{\ln{\varphi_2}}}+1}
\end{eqnarray}
\noindent Given that
$\varphi_2^{\delta_1 \cdot \frac{\ln{\varphi_1}}{\ln{\varphi_2}}} = \varphi_1^{\delta_1}$, we get the following corollary:
\vspace{0.25in}
\begin{coro}
If workers with response accuracy $\varphi_1$ are paid $\textit{pay}(\varphi_1)$ per vote, then workers with accuracy $\varphi_2$ will generate results of the same quality and cost if the ratio of the payments is:
\begin{equation}
\frac{\textit{pay}(\varphi_1)}{\textit{pay}(\varphi_2)}=\frac{\ln\varphi_1}{\ln\varphi_2}\cdot\frac{\varphi_2+1}{\varphi_1+1}\cdot\frac{\varphi_1-1}{\varphi_2-1},
\label{fla:pay_eq}
\end{equation}
where $\varphi_i=\frac{p_i}{1-p_i}$ are the odds that a worker belonging to pool $i$ classifies an item correctly.
\end{coro}
\vspace{0.15in}
\subsection{Iso-payment effort function}
\label{sec:iso}
Using the results of corollary 3, we can further investigate the relationship between payment per item and the per-item \emph{effort} made by a worker. Consider that for a given worker, an unknown function $p = g(\psi)$ governs the relationship between the effort $\psi$ expended by the worker on an item and the accuracy $p$ of the response given. (One possible measurable proxy for the level of effort is the amount of time spent working on an item.)
Given our understanding of the relative amounts of payment, given in (\ref{fla:pay_eq}), we can connect payment to effort for a given function $g(\psi)$.
For example, assume that the relationship between a worker's effort expended and accuracy of the answers is $p = e^\psi - 1$. This particular accuracy-effort function is represented with the x-marked line on the left panel of Figure~\ref{fig:iso-payment-effort}. Since we are interested in relative adjustments to payment rate, set $pay(\varphi_2)$ in equation~(\ref{fla:pay_eq}) to be equal to 1. We then get $pay(\varphi) = c_0\cdot ln(\varphi)\cdot\frac{\varphi-1}{\varphi+1}$, where $c_0$ is a constant, and $\varphi$ is equal to $\frac{p}{1-p}$ as before. We can now connect worker effort to the level of payment per unit of effort $pay(\varphi)/\psi$ (if we continue our example of effort measured as time spent per item, this last quantity would represent \emph{hourly wage}). The x-marked line on the right panel of Figure~\ref{fig:iso-payment-effort} represents the relationship for our particular chosen function $p=g(\psi)=e^\psi - 1$.
We now ask: what relationship between a worker's effort and the accuracy of her responses would make her payment invariant to effort expended? In other words, we are looking for such $g(\psi)$ that the `hourly wage' $pay(\varphi)/\psi$ is independent of $\psi$. It is immediate that all such functions
can be defined by
\begin{equation}
\psi = c\cdot ln\varphi\cdot\frac{\varphi-1}{\varphi+1} =: g^{-1}_{iso}(\varphi),
\label{fla:iso-payment-effort}
\end{equation}
where $c$ is a constant.
Figure~\ref{fig:iso-payment-effort} illustrates a sample of several functions $p = g_i(\psi)$, different only in the chosen values of the constant factor $c$.
\begin{figure}[t]
\centering
\includegraphics[height=0.31\columnwidth]{figs/p_of_effort.png}
\includegraphics[height=0.31\columnwidth]{figs/hourlywage_vs_effort.png}
\caption{\rm{Properties of various accuracy-effort curves, plotted for a range of accuracy values $0.51\le p \le 0.98$. Left: sample iso-payment accuracy curves, contrasted with two accuracy functions chosen without iso-payment considerations: $p = e^\psi - 1$ and $p \sim tanh(\psi)$. Right: visual demonstration of the iso-payment property.}}
\label{fig:iso-payment-effort}
\end{figure}
A more realistic hypothesis about a relationship between worker accuracy $\varphi$ and her effort $\psi$ would incorporate an assumption of diminishing returns. Figure~\ref{fig:iso-payment-effort} also includes a plot of an accuracy function $p = \frac{1}{2}+\frac{2}{5}\tanh(\frac{\psi}{4})$, which conforms to this assumption\footnote{The shape of a hyperbolic tangent is convenient for describing the natural assumption on the diminishing returns to effort. The constants are chosen to scale $\tanh$ with accordance to our context.}. Since a worker's natural objective is to maximize pay, the best strategy for a worker with this kind of a relationship between effort and accuracy would be to choose the global maximum of their hourly wage curve, if the global maximum is unique, and otherwise choose a global maximum corresponding to the lowest time expenditure.
\subsection{Utility-equalizing payments}
\label{sec:utility-eq-payments}
We offered a way to set the pay rate for workers in each pool relative to the other pool, subject to \emph{risk-neutral utility constraints}. A risk-neutral approach would entail a utility function that would only involve the expected cost of getting a consensus vote. However, a more nuanced notion of utility should consider not only the expected quality and expected number of votes required to reach consensus, but also the variance of each. Thus, a \emph{risk-averse} procedure for establishing an appropriate equivalence for payment for a single item is as follows:
$$
Q_1(\varphi_1, \delta_1)=Q_2(\varphi_2, x) \implies \delta_2:=x
$$
$$
U_i:= \mathbb{E}[Q(\varphi_i, \delta_i)] - \text{pay}_i(\varphi_i)\mathbb{E}[n_{votes, i}|\varphi_i, \delta_i] - \lambda\cdot\left( \sqrt{Var(Q(\varphi_i, \delta_i))} + \text{pay}_i(\varphi_i)\sqrt{Var(n_{votes, i})}\right)
$$
\begin{equation}\label{fla:utile_payments}
\begin{cases}U_1=U_2 \\ Q_1=Q_2 \end{cases} \iff \text{pay}_2:=\text{pay}_1\frac{\mathbb{E}[n_{votes, 1}|\varphi_1, \delta_1] + \lambda\sqrt{Var(n_{votes, 1})}}{\mathbb{E}[n_{votes, 2}|\varphi_2, \delta_2] + \lambda\sqrt{Var(n_{votes, 2})}}
\end{equation}
\\
A remark on notation: strictly, $Q$ is not a random variable but a first moment of $\mathbbm{1}_C$. We use the notation $Var(Q)$ to mean, more strictly, $Var(\mathbbm{1}_C)$. Although computing the term $Var(Q(\varphi_i, \delta_i))$ is not necessary for using the payment equivalence condition in Equation~\ref{fla:utile_payments}, it is worth mentioning that the variance of the indicator of reaching a correct consensus is the variance of a Bernoulli random variable: $Var(\mathbbm{1}_C) = \mathbb{E}[\mathbbm{1}^2_C]-(\mathbb{E}[\mathbbm{1}_C])^2 = P(\mathbbm{1}_C=1)(1-P(\mathbbm{1}_C=1))=Q(1-Q).$
The plots in Figure~\ref{fig:utile_payments} below visualize the following process: for every fixed value of utility parameter $\lambda$ and every triplet $(p_1, \delta_1, p_2)$ (where $0.5 < p_1 < 1, 0.5 < p_2 <1$), we compute $\delta_2$ that guarantees $Q_1=Q_2$. We then derive the ratio of payments $\text{pay}_2/\text{pay}_1$ for a given triplet, and we visually organize the results by the values of corresponding ratios of expected times to completion, $\mathbb{E}[n_{votes}(p_1,\delta_1 )] / \mathbb{E}[n_{votes}(p_2,\delta_2)]$. The plots display the results for selected values of the utility parameter: $\lambda=1, \lambda=2$, and $\lambda=10$.
\begin{figure}[t]
\makebox[\textwidth][c]{
\includegraphics[height=0.25\columnwidth]{figs/payment_ratio_lambda1.png}
\includegraphics[height=0.25\columnwidth]{figs/payment_ratio_lambda2.png}}
\makebox[\textwidth][c]{
\includegraphics[height=0.25\columnwidth]{figs/payment_ratio_lambda10.png}}
\caption{\rm{Utility-equalizing ratio of payment rates $pay(\varphi_2)/pay(\varphi_1)$ for workers on two voting processes (\ref{fla:utile_payments}), where $\delta_2$ is derived in a way that satisfies the equal quality constraint. Top left: $\lambda = 1$. Top right: $\lambda = 2$. Bottom: $\lambda = 10$.}}
\label{fig:utile_payments}
\end{figure}
\section{Experimental validation}
\label{sec:theory_vs_exp}
The results provided above are theoretical and asymptotic. How good are they in approximating the results of a real crowdsourced voting process with constrained resources? Obviously, in reality, some of the assumptions of our model will be violated: workers will not have identical qualities and items will have different degrees of difficulty. As the old adage states, \emph{``all models are wrong''}, hence we want to see if our model is actually useful in providing guidance on how to set up a crowdsourcing process to achieve certain levels of quality for certain levels of cost.
We compare the theoretical results with those obtained from experiments. In particular, we compare the theoretical values of quality, completion time, and variance thereof with two kinds of simulations: $\delta$-margin voting simulated as \emph{random walk}; and $\delta$-margin voting simulated as random draws from real-life \emph{crowdsourced data}.
\subsection{Dataset description}
\label{sec:bluebirds}
We simulate a voting process using a dataset of real votes collected on a major crowdsourcing platform. We used the \texttt{Bluebirds} dataset~\citep{bluebirds}, which contains binary labels for 108 unique images solicited from 39 workers, as well as ground truth labels for each item (with a 44:56 size ratio of the \texttt{True} and \texttt{False} classes). The histogram in Figure~\ref{fig:worker_acc_hist} shows the empirical distribution of individual accuracy among workers, computed as the frequency of labeling an item correctly. The mean worker accuracy stands at 0.636. The distribution of computed per-item rates of correct response suggests that a third of items are difficult (i.e., frequency of correct responses across all workers for an item is below 0.5).
\begin{figure}[t]
\centering
\includegraphics[width=0.45\columnwidth]{figs/wkr_acc_hist.png}
\includegraphics[width=0.46\columnwidth]{figs/item_acc_hist.png}
\caption{\rm{Descriptive statistics of the \texttt{Bluebirds} dataset. Left: a histogram of the response accuracy (as proportion of correct answers) for all workers. Right: a histogram of the per-item response accuracy.}}
\label{fig:worker_acc_hist}
\end{figure}
In the rest of this section, we simulate the $\delta$-margin voting process using simulated random walk as well as real life crowdsourced labels. We do so in order to both examine the asymptotic properties of experimental process quality and time to completion, as well as (and chiefly) to compare the experimental values with the theoretical results from Section~\ref{sec:characteristics}. Indeed, a key assumption underpinning the theoretical results presented in this paper is that individual worker accuracies are distributed uniformly, with each worker having accuracy $p$. In real life crowdsourced data --as in the \texttt{Bluebirds} dataset-- response accuracy varies across workers as well as across items (some being more difficult to classify). This is a deviation from our theoretical model, and the main purpose of the experiments in this section is to examine the extent to which empirical quantities obtained under such violations disagree with theoretical quantities.
\subsection{Theoretical quality vs. experimental quality of label aggregation}
\label{sec:q_vs_exp}
The first question of interest is how closely the theoretical quality in Equation~\ref{eq:Q_nonrand} captures a real or realistic process of $\delta$-margin voting in a crowdsourcing labeling task.
We begin by comparing the theoretical value of quality in Equation~\ref{eq:Q_nonrand} with the quality resulting from \emph{random walk simulations} of $\delta$-margin voting with the same accuracy $p$. For each value of delta $\delta \in [2,3, .., 7]$ and for each value of $p$, we simulate 100,000 random walk trials and record the average correctness of the consensus vote. Figure~\ref{fig:q_th_vs_q_sim} shows the relative error between the computed average quality of the vote and the quality estimated from simulations, $err(p):=Q_{sim}^{-1}\cdot \left(Q_{sim}-Q_{th})\right).$
\begin{figure}[t]
\centering
\includegraphics[width=0.8\columnwidth]{figs/Q_sim-Q_th.png}
\caption{\rm{Relative error $err(p):=Q_{sim}^{-1}\cdot \left(Q_{sim}-Q_{th})\right)$ of the average (expected) consensus vote quality, as a function of the probability of a correct answer $p$, for a range of fixed values of consensus threshold $\delta$.}}
\label{fig:q_th_vs_q_sim}
\end{figure}
Next, we transition to comparing the theoretical values to estimations with \emph{observational data} experiments. As discussed, the \texttt{Bluebirds} dataset violates several of our modeling assumptions (i.e., worker accuracy levels as well as item difficulty are not uniform). For this reason, estimating quantities of interest using random draws from this data is an informative benchmark for assessing real-life usefulness of presented theoretical formulations.
For a given $\delta \in [1, 2, .., 11]$, the general procedure that we used for simulating $\delta$-margin voting is as follows. For a given item, we iteratively draw votes
from labels given by a worker pool on this item, until the agreement threshold $\delta$ is attained. We run $r$
such simulations of the $\delta$-margin voting process for each item.
When examining quality and time to completion of experimental voting results in isolation, we performed sampling \emph{without replacement}, and we set the number of voting process repetitions per item to $r=20$. We additionally simulated a range of \emph{conditions} on worker pool accuracy and item difficulty. In other words, we ran $\delta$-margin voting simulations as described above for a range of pre-filtered versions of the \texttt{Bluebirds} dataset -- subsets that included only items of certain maximum levels of difficulty and workers with a certain minimum level of accuracy. In particular, we required that the worker accuracy be above $\tau_w$ and the item accuracy be above $\tau_w$ (where pairs $(\tau_w, \tau_i)$ take all values on a lattice $[0, 0.1, ..., 0.9]^2$). Thus, for each of the values of the consensus threshold $\delta$ we run 20 experiments for each of the items that passed a given worker-item accuracy filter. Our experimental process is described by the pseudocode procedure in Appendix~\ref{app:algs} (Algorithm~\ref{alg:simAsy}).
We first examine some of the asymptotic properties of quality estimates from our crowdsourced data experiments.
Figure~\ref{fig:mean_correctness_all_filters} illustrates the resulting quality of the labels aggregated in our experiments, summarized for all settings and grouped by levels of the filtering on the worker accuracy. The reason that the processes do not attain the $Q=1$ limit --even for the highest filtering thresholds of worker pool accuracy-- is likely that the quality is aggregated across experiments with various settings of \textit{item} accuracy, including settings where no minimum threshold for item `ease' is set. Thus, a non-negligible proportion of experiments (those without any filter on item accuracy) included difficult items which get incorrect consensus votes even when solicited from top worker pools.
\begin{figure}[t]
\centering
\includegraphics[width=0.8\columnwidth]{figs/correctness_all_filters.png}
\caption{\rm{Correctness of resulting consensus vote, averaged across all experiments and all settings for a fixed consensus threshold $\delta$ and fixed minimum worker accuracy threshold $\tau_w$.}}
\label{fig:mean_correctness_all_filters}
\end{figure}
Next, we will benchmark the results from crowdsourced data experiments against the theoretical values obtained from Equation~\ref{eq:Q_nonrand}.
To obtain the experimental values for the comparison, we performed sampling \emph{with replacement}\footnote{This prevents termination of the voting process due to exhausting the total set of available votes for high values of $\delta$. Repeated exposure to items is also a realistic scenario in some crowdsourced tasks, although our voting simulation necessitates that a worker always responds the same way to the same item.} and used the entire \texttt{Bluebirds} dataset. For a given $\delta$, we simulate $\delta$-margin voting and compare the resulting consensus labels with the ground truth. We record the average correctness (frequency of matching the ground truth) of $\delta$-margin consensus labels across all items. This process of procuring consensus votes and aggregating their correctness across all items is repeated 100 times for each value of $\delta$, collecting 100 statistical quantities, each of which summarizes consensus quality for the full set of items. (We also record an average number of votes it took to reach consensus for the full set of items -- we discuss the results in subsequent sections.)
For each value of $\delta$, we also compute the theoretical value of consensus vote quality from Equation~\ref{eq:Q_nonrand}. To do this, we take the worker pool accuracy level $p$ to be equal to the average of accuracies among all workers recruited for the task (regardless of whether a vote was solicited from them). For example, when recruiting the entire set of 39 workers to the task of labeling all 108 items using $\delta$-margin voting with threshold $\delta=3$, the empirically sourced pool-level odds of correct response in the \texttt{Bluebirds} dataset is $\varphi \approx \frac{0.6356}{1-0.6356} \approx 1.744$, and hence the theoretical quality of the voting process is $Q_{th}(\varphi, \delta) \approx \frac{1.744^3}{1+1.744^3} \approx 0.841$.
Finally, we compute the difference between the experimental and theoretical values of result quality. The simulation process is described in Algorithm~\ref{alg:simCmp} in Appendix~\ref{app:algs}.
Figure~\ref{fig:q_th_vs_q_exp} demonstrates the results for the total, unfiltered pool of workers and items. Theoretical quality makes for a reasonably small overestimation of the empirical results quality (with the exception of $\delta=2$, where part of the experiments demonstrated better-than-theoretical quality). The trend demonstrates that the gap between the theoretical quality derived earlier and the quality of (quasi-)empirical results is widening as the consensus majority requirement becomes stronger, however the variance of the difference becomes smaller. Figure~\ref{fig:q_th_vs_q_exp} also includes an illustration of the same quantities under an identical voting with replacement simulation, but with pre-set quality filters on the pool of workers.
\begin{figure}[t]
\centering
\includegraphics[width=0.46\columnwidth]{figs/q_th-q_exp.png}
\includegraphics[width=0.45\columnwidth]{figs/q_th_and_q_exp_sidebyside.png}
\includegraphics[width=0.28\columnwidth]{figs/q_th-q_ex_wkr_acc05.png}
\includegraphics[width=0.28\columnwidth]{figs/q_th-q_ex_wkr_acc07.png}
\includegraphics[width=0.28\columnwidth]{figs/q_th-q_ex_wkr_acc09.png}
\caption{\rm{Comparison between theoretical and empirical values of overall task result quality, for various values of consensus threshold $\delta$ and a fixed (empirically derived) worker pool accuracy $p$. Each data point is an average over 100 experiments. Top left: violin plot of distributions of the difference $Q_{th}-Q_{exp}$. Top right: side-by-side comparison and difference $Q_{th}-Q_{exp}$. Bottom row: combined violin and point difference plots, for pre-filtered datasets with worker quality $\tau_w>0.5$, $\tau_w>0.7$, and $\tau_w>0.9$ respectively.}}
\label{fig:q_th_vs_q_exp}
\end{figure}
\subsection{Theoretical vs. experimental time until completion}
As before, we begin by comparing the theoretical value of time until consensus in Equation~\ref{eq:exp_votes} with the average time (i.e., number of votes) cast in 100,000 \emph{random walk simulations}, for each pair $(p,\delta)$. Figure~\ref{fig:ET_th_vs_ET_sim} demonstrates that the relative error $err_{ET}(p):=\mathbb{E}^{-1}_{sim}[n_{votes}]\cdot \left(\mathbb{E}_{sim}[n_{votes}]-\mathbb{E}_{th}[n_{votes}])\right)$ between the computed theoretical values and the quality estimated from random walk simulations is small.
\begin{figure}[t]
\centering
\includegraphics[width=0.8\columnwidth]{figs/ET_sim-ET_th.png}
\caption{\rm{Relative error $err_{ET}(p):=\mathbb{E}^{-1}_{sim}[n_{votes}]\cdot \left(\mathbb{E}_{sim}[n_{votes}]-\mathbb{E}_{th}[n_{votes}])\right)$ of the average (expected) time until consensus, as a function of the probability of a correct answer $p$, for a range of fixed values of consensus threshold $\delta$.}}
\label{fig:ET_th_vs_ET_sim}
\end{figure}
Again, we follow up by a comparison between theoretical and \emph{empirical} data (as opposed to theoretical vs. \emph{simulated} data, above). We use the same procedure as described in the previous section to compute the empirical values of the number of votes it took to reach consensus. We estimate the empirical value of accuracy $p$ the same way as above in order to compute the theoretical value of expected number of votes using Equation~\ref{eq:exp_votes}.
\begin{figure}[t]
\centering
\includegraphics[height=0.3\columnwidth]{figs/nvotes_all_filters.png}\hspace{0.3cm}
\includegraphics[height=0.3\columnwidth]{figs/log_Nvotes.png}
\caption{\rm{Number of votes it took to reach consensus. Left: averaged across all experiments and all $\tau_i$ settings for a fixed consensus threshold $\delta$ and fixed minimum worker accuracy threshold $\tau_w$. Right: `violin plot' of $n_{votes}$ distributions for a fixed $\delta$ across all settings of $\tau_w$ and $\tau_i$, on a logarithmic scale.}}
\label{fig:mean_nvotes_all_filters}
\end{figure}
The results of this comparison are illustrated in Figure~\ref{fig:mean_nvotes_all_filters}. As one would intuitively expect, increasing the value of $\tau_w$ (a lower bound on the accuracy of recruited workers) pushes the number of votes required for consensus to be closer to the threshold $\delta$. In other words, any sample of workers is likely to agree, so a series of $\delta$ votes is likely to be enough to reach a $\delta$ majority. Also unsurprisingly, the adherence to the $N_{votes}=\delta$ line is tighter for smaller levels of $\delta$, and the graph begins to curve upwards for larger values of $\delta$. Indeed, the violin plot on the right panel of Figure~\ref{fig:mean_nvotes_all_filters} illuminates the fact (which we confirmed directly by examining individual items in our computation) that, for higher threshold values $\delta$, only a few items within a given setting may occasionally take an enormous number of votes to reach consensus, and such `outliers' drive up the mean value of $N_{votes}$ across all settings in the left panel for higher values of $\delta$. We confirmed this fact explicitly by examining the way time to completion was distributed across items in our computations.
\subsection{Theoretical variance of time until completion vs. random walk simulations}
\label{sec:var_vs_exp}
First, we compare the variance of time to termination in a simulated random walk with two absorbing states with the theoretical value of variance obtained as in Proposition~\ref{th:var_nonrand
.
To obtain the `experimental' variance, the first step is to simulate $n$ trials of a Markov random walk with given parameters $p$ and $\delta$ defining the initial value in a chain with absorbing states at $0$ and $2\delta$. We used $n=100,000$ trials for each pair of values $(p, \delta)$, recording the variance of the number of votes until consensus for each pair.
We are interested in establishing for which values of $p$ does the theoretical variance of time to termination start diverging from the value obtained via random walk simulations.
For a range of values of the threshold $\delta$, the plot in Figure~\ref{fig:var_th_vs_randomwalk} shows the relative error in the estimation of variance of a simulated random walk versus the theoretical values in Formula~\ref{fla:var_nonrand}
, as a function of $p\in[0.51, 0.93]$:\hspace{0.2cm}
$\ell(p):=Var_{sim}^{-1}(n_{votes})\cdot \left(Var_{sim}(n_{votes})-Var_{th}(n_{votes})\right).$
\begin{figure}[t]
\centering
\includegraphics[width=0.8\columnwidth]{figs/percentage_change.png}
\caption{\rm{Relative error $\ell(p):=Var_{sim}^{-1}(n_{votes})\cdot \left(Var_{sim}(n_{votes})-Var_{th}(n_{votes})\right)$ of the variance values between the theoretical and the simulated random walk experiments, as a function of the probability of a correct answer $p$, for a fixed consensus threshold $\delta$.}}
\label{fig:var_th_vs_randomwalk}
\end{figure}
\section{Discussion and future work}
\label{sec:future}
Delta-margin majority voting is a frequent method of choice when aggregating worker votes on crowdsourcing platforms. Adaptively requiring a strong enough majority ---instead of requesting a specific number of votes for consensus---, this method is favored in practice for promising higher accuracy of results without casting too large a number of votes when it is not necessary to do so. Yet wherever the preference for this method emerged ad hoc, a theoretical study of this method remained lacking.
In this work, we propose a novel way to model the $\delta$-margin voting process as a Markov process. We use well-established probability theory and statistics results to put forth closed-form expressions that quantify the expected quality of the consensus vote. We also formulate the PDF, expectation, and variance of the number of votes one needs to cast in order to reach consensus.
Furthermore, we formulate a series of useful equivalences. The first of these is an equation that defines how one ought to set the threshold parameter of a voting process in order to attain the same quality as that of a process with different (given) accuracy and consensus threshold. The second equivalence defines a risk-neutral utility-equalizing ratio of payments for two workflows which differ in worker response accuracies. In other words, it suggests hot to set the wage for the second process, relative to the wage for the first one, in such a way that the two processes attain the same utility, defined here as combining result quality and expected processing time (a proxy for process cost). We further refine this result to a case of risk-averse cost function -- one that considers both the expectation and the variance of time until consensus.
We probed the reliability of our theoretical results in real-life scenarios. Indeed, one limitation of our results is in making the assumption of uniform response accuracy among workers -- an assumption not likely to hold on crowdsourced platforms. However, comparing each of our theoretical quantifications with corresponding simulations (random walk as well as, separately, draws from real crowdsourced labels) suggested that our theoretical results are robust to violations of the worker accuracy uniformity assumption. This is an encouraging finding for practitioners seeking to design cost-optimal crowdsourced workflows: theoretical quantifications can be used to guide process design ex ante, without having to tune the parameters of the voting process after a series of costly trial-and-error runs.
While our experimental evaluations bode well for the applicability of the theoretical model as it stands (i.e., with the assumption of a uniform accuracy of all workers for all items), one future work direction would be to extend our model to formally handle worker and item heterogeneity. What are the equivalent results to Theorems~\ref{th:Q_nonrand}, ~\ref{th:ET} and the variance results in \ref{sec:var} that hold for voting tasks where worker accuracies are not constant and equal to $p$, but follow a distribution with mean $p$?
A further challenge to applicability that stems from our assumption about worker accuracy $p$ is that the accuracy parameter is seldom fully observed in practice. The question of interest to practitioners is that of an inverse problem: can our theoretical results be used to infer the accuracy distribution of a pool of workers, given limited or no access to ground truth labels? In future work, powerful statistical packages such as TensorFlow may be employed to solve an inferential version of the workflow problem with real-life crowdsourced task data and unknown worker accuracy parameters.
Another limitation of our work is its focus on binary classification.
A natural direction of future work is to derive corresponding results that hold for non-binary classification tasks\footnote{This question appears to be easily addressed by representing multi-class tasks as binary ones by pooling all but the correct class labels together.
} A further extension of this work could attempt to provide theoretical quality and cost estimates for continuous (regression) tasks.
Finally, in a broader outlook on future work, we plan on extending the
analysis herein onto other types of processes, with a focus on understanding the interplay between quality, time to completion, and payment costs. Ultimately, our goal is to have a set of `modular' analyses that allow practitioners to take as input the structure of a crowdsourcing task or process and provide an ex ante prediction of its behavior,
without having to execute the process to see the results.
Despite the fact that the voting process modeled using Gambler's Ruin random walk is guaranteed to terminate, the number of votes necessary for reaching consensus on a given item may be prohibitively large. This is one of the reasons for developing heuristics and, eventually, theoretical results for \textit{escalation workflows}. In this type of a crowdsourced labeling process, an item will either get resolved by a pool of workers recruited from a given pool, or remain unresolved after certain maximum amount of iterations (an \emph{escalation criterion}), after which the requester will transfer the item to the next `stage' where higher accuracy labourers will vote to conclude the item's consensus label\footnote{Another motivation for establishing escalation criteria is driven by cost efficiency: there may exist such combinations of accuracies $(p_1,p_2$ and wage rates $(pay_1,pay_2)$ that make it strictly more cost-efficient for a requester to escalate items sooner, regardless of the reasonable expectation of those items being resolved by the first, less accurate worker pool.}.
In relation to workflows, it is interesting to establish how many stages will resolve most of the items, given worker accuracy at each stage. Further, how many workers do we expect to recruit to resolve all items? What is the most cost-efficient stopping criterion? In this direction of future work, as with the current work, we aim to enable practitioners to infer cost-optimal parameters of multistage workflows ex ante.
\section{Procedures for simulating $\delta$-margin voting using observational data.}
\label{app:algs}
In this appendix, we present pseudo-code that describes various procedures we used to simulate $\delta$-margin voting for various analyses described in Section~\ref{sec:theory_vs_exp}. In particular, we describe the procedures we used for:
\begin{itemize}
\item examining asymptotic properties of \emph{consensus vote quality} as a function of accuracy $p$ for various values of the threshold $\delta$
\item examining asymptotic properties of \emph{time until consensus} as a function of accuracy $p$ for various values of the threshold $\delta$
\item comparing the experimental consensus vote quality (i.e., the average correctness of the consensus votes resulting from our simulations that use observational data) with the theoretical value of quality in Equation~\ref{eq:Q_nonrand}.
\item comparing the experimental time (i.e., number of votes it took) until consensus with the theoretical value of expected time until consensus in Equation~\ref{eq:exp_votes}.
\end{itemize}
\begin{algorithm}[H]
\begin{algorithmic}[1]
\Procedure{QualityAndTime\_VariousWorkerPools}{data = Bluebirds}
\For {$\tau_w \in [0, 0.1, .., 0.9]$}
\For {$\tau_i \in [0, 0.1, .., 0.9]$}
\State data\_subset $\gets filter($data,$ \tau_w, \tau_i)$
\For {$\delta \in [0, 1, .., 11]$}
\For {item $\in \{$data\_subset$\}$}
\For {$i \in [0,..,19]$}
\State this\_trial\_stats $\gets \textit{SimulateDMarginVoting}(\textit{item}, \delta, replace = \texttt{False})$
\State correctness\_and\_time $\gets get($this\_trial\_stats, [`is\_correct', `N\_votes'])
\State all\_trials\_single\_item $\gets append($correctness\_and\_time)
\EndFor
\State all\_trials\_all\_items $\gets append($all\_trials\_single\_item)
\EndFor
\State all\_deltas\_all\_trials\_all\_items $\gets append($all\_trials\_all\_items)
\EndFor
\State all\_itemfilters\_all\_deltas\_all\_trials\_all\_items $\gets append($all\_deltas\_all\_trials\_all\_items)
\EndFor
\State all\_filters\_correctness\_and\_time $\gets append($all\_itemfilters\_all\_deltas\_all\_trials\_all\_items)
\EndFor
\State $return$(all\_filters\_correctness\_and\_time)
\EndProcedure
\end{algorithmic}
\caption{\rm{The procedure simulating $\delta$-margin voting using various subsets of observational data, for computing experimental quality and time until consensus.}}
\label{alg:simAsy}
\end{algorithm}
\begin{algorithm}[H]
\begin{algorithmic}[1]
\Procedure{QualityExperimentsVsTheory}{data = Bluebirds}
\State p = $WorkerAccuracy$(data)
\For {$\delta \in [0, .. , 11]$}
\State quality\_theoretical = $ComputeTheoreticalQuality(\delta$, p)
\For {experiment\_id $\in [0, .. , 99]$}
\For {item $\in$ Bluebirds}
\State consensus\_vote\_stats $\gets SimulateDMarginVoting(\delta,$ item, replace = \texttt{True})
\State consensus\_vote\_correctness $\gets get($consensus\_vote\_stats, `is\_correct')
\State all\_items\_correctness $\gets append($consensus\_vote\_correctness)
\EndFor
\State quality\_experimental\_singleTrial $\gets average($all\_items\_correctness)
\State quality\_experimental\_multi $\gets append($quality\_experimental\_singleTrial)
\EndFor
\State theoretical\_vs\_experimental $\gets combine$(quality\_experimental\_multi, quality\_theoretical)
\State
theory\_vs\_100Experiments\_AllDeltas $\gets append($theoretical\_vs\_experimental)
\EndFor
\State $return$(theory\_vs\_100Experiments\_AllDeltas)
\EndProcedure
\end{algorithmic}
\caption{\rm{The procedure of simulating $\delta$-margin voting using observational data, for comparing theoretical and experimental values of quality and time to consensus.
}}
\label{alg:simCmp}
\end{algorithm}
\section{Croatian Proof of Theorem~\ref{fla:var_nonrand}}
\section{Iso-payment accuracy curves}
\label{app:iso}
Let us further examine the equation~\ref{fla:iso-payment-effort}. Figure~\ref{fig:iso_closeup} below displays the relationship between the odds $\varphi$ of classifying an item correctly, and effort $\psi$ expended per item. While the left panel may at first suggest that the relationship for all iso-curves is exponential in character, the change of scale on the right panel of Figure~\ref{fig:iso_closeup} reveals the existence of two function branches adjoined at $\psi=0$ for each iso-curve.
\begin{figure}[t]
\centering
\includegraphics[height=0.28\columnwidth]{figs/iso_curve.png}
\includegraphics[height=0.28\columnwidth]{figs/iso_curve_zoom.png}
\caption{\rm{A close-up plot of several iso-payment accuracy curves, revealing multiple branches jointly defining $p = g(\psi)$, where $p = \frac{\varphi}{\varphi+1}$ by definition of $\varphi$.}}
\label{fig:iso_closeup}
\end{figure}
Note that the setting relevant for crowdsourcing is the one where $p>0.5$, and hence $\varphi>1$. In this region, iso-payment curves are well-defined, and their shape is asymptotically exponential.
Let us further note the close match between the first iso-payment plot and our chosen non-iso accuracy curve on the left panel of Figure~\ref{fig:iso_closeup}. It is easy to attest to both visually and analytically that the iso-payment accuracy curve with $c=1$ converges to our initially chosen exponential curve: $\lim\limits_{\varphi\to\infty}\ln(1+\varphi) - \ln\varphi\frac{\varphi-1}{\varphi+1} = 0$. This means that as her accuracy pushes closer to being perfect, a worker whose accuracy and effort are connected as $p = e^\psi - 1$ tends to become indifferent to time spent per unit under this payment scheme.
\section{Symmetric walk (random voting) case}
\label{app:symm_case}
Let us extend the result of Theorem~\ref{th:Q_nonrand} to a case where $p=q$.
\vspace{0.25in}
\begin{theo}
For a random voting process ($p=1/2$), the probability $Q(\varphi,\delta)$ that the consensus vote $c$ is correct, for an item classified using a $\delta$-margin voting scheme
with worker accuracy $p$ and a consensus threshold $\delta$ is:
\begin{equation}
\label{eq:Q_rand}
Q(\varphi,\delta) = P(C=1|\varphi,\delta)=\delta^2.
\end{equation}
\end{theo}
\vspace{0.15in}
\textbf{Proof.}
First, derive the probability that the voting process terminates with a correct consensus label: $P(X_{T_k}=N)$.
Consider a random variable $X_{n+1}$. One can verify that it is a martingale: $\mathbb{E}[X_{n+1}|X_{n}, X_{n-1}, ... , X_0] = P(\xi_{n+1}=1)(X_n + 1) + P(\xi_{n+1}=-1)(X_n - 1) = \frac{1}{2} X_n + \frac{1}{2} + \frac{1}{2} X_n - \frac{1}{2} = X_n.$ Therefore, $\mathbb{E}[X_n]=\mathbb{E}[X_0]=k$, and in particular $\mathbb{E}[X_{T_k}]=k$.
At the same time, $\mathbb{E}[X_{T_k}]=0\cdot P(X_{T_K}=0)+N\cdot P(X_{T_k}=N).$ Hence,
\begin{equation}\label{deriv:5}
P(X_{T_k}=N) = \frac{k}{N}\text{\hspace{0.2cm} when } p=q=1/2.
\end{equation}
Now consider a random variable $X^2_{n+1}-(n+1).$ Let us verify that it is a martingale:
\begin{multline*}\mathbb{E}[X^2_{n+1}-(n+1) | X_n, X_{n-1}, ..., X_0] = \mathbb{E}[(\xi+X_n)^2 | X_n, X_{n-1}, ..., X_0] - (n+1)] = \\
= \frac{1}{2} (X_n+1)^2 + \frac{1}{2} (X_n - 1)^2 - (n+1) = X_n^2-n.
\end{multline*}
Therefore, $\mathbb{E}[X_{n+1}^2-(n+1)]=\mathbb{E}[X_0^2]=k^2,$ and in particular,
\begin{equation}\label{deriv:6}\mathbb{E}[X_{T_k}^2-T_k]=k^2.\end{equation}
At the same time,
\begin{equation}\label{deriv:7}\mathbb{E}[X_{T_k}^2-T_k]=\mathbb{E}[X_{T_k}^2]-\mathbb{E}[T_k]=0\cdot P(X_{T_k}=0)+N^2 P(X_{T_k}=N).\end{equation}
From Equations~\ref{deriv:5},~\ref{deriv:6}, and~\ref{deriv:7} we obtain:
\begin{equation}\mathbb{E}[T_k]= k(N-k).
\label{deriv:8}
\end{equation}
The statement of the theorem follows by substituting $2\delta$ for $N$ and $\delta$ for $k$. $\square$
\vspace{0.25in}
\begin{theo}
Variance of the number of votes it takes to reach consensus using $\delta$-margin voting with threshold $\delta$ for a case of random votes ($p=q=1/2$) is equal to:
\begin{equation}
Var(n_{votes}) = \frac{2}{3}\delta^2(\delta^2-1).
\label{fla:var_rand}
\end{equation}
\label{th:var_rand}
\end{theo}
\vspace{0.15in}
\textbf{Proof.}
First, consider a random variable $Z_t = X_n^3 - 3nX_n.$ It is a martingale, since
\begin{multline}\mathbb{E}[Z_t|Z_{t-1},...,Z_0] = \mathbb{E}[X_n^3|X_{n-1},...,X_0]-3n\mathbb{E}[X_n|X_{n-1},..,X_0] = \\=\frac{1}{2}(X_{n-1}+1)^3+\frac{1}{2}(X_{n-1}-1)^3-\frac{3}{2}n(X_{n-1}+1+X_{n-1}-1)=\\
=\frac{1}{2}(X^3_{n-1}-3X^2_{n-1}+3X_{n-1}-1)+\frac{1}{2}(X^3_{n-1}+3X^2_{n-1}+3X_{n-1}+1)-\frac{3}{2}2nX_{n-1} =\\
=X_{n-1}^3+3X_{n-1}-3(n-1)X_{n-1}-3X_{n-1}=X^3_{n-1}-3(n-1)X_{n-1} = Z_{n-1}.
\end{multline}
Therefore, $\mathbb{E}[Z_n]=\mathbb{E}[Z_0]=\mathbb{E}[X_0^3]=k^3$ and hence, by optional stopping theorem, \begin{equation}\mathbb{E}[Z_{T_k}]=k^3.
\label{proof:var_rand_1}
\end{equation}
At the same time,
$$ \mathbb{E}[Z_{T_k}]=\mathbb{E}[X_{T_k}^2-3T_kX_{T_k}]=\mathbb{E}[X^3_{T_k}]-3\mathbb{E}[T_kX_{T_k}]
= N^3P(T_k = N)-3\mathbb{E}[T_kX_{T_k}].
$$
Using (\ref{deriv:5}), we get:
\begin{equation}
\mathbb{E}[T_kX_{T_k}] = \frac{k(N^2-k^2)}{3}.\label{proof:var_rand_2}
\end{equation}
Now let us derive the conditional expectations of consensus time, given the correctness of the label:
$$\mathbb{E}[T_kX_{T_k}] = P(X_{T_k}=0)\mathbb{E}[T_kX_{T_k}|X_{T_k}=0] + P(X_{T_k}=N)\mathbb{E}[T_kX_{T_k}|X_{T_k}=N]
$$
Using (\ref{proof:var_rand_2}) and (\ref{deriv:5}), we get\footnote{It follows further from (\ref{proof:var_rand_3}) that $\mathbb{E}[T_k|X_{T_k}=0]=\frac{\mathbb{E}[T_k]-\mathbb{E}[T_k|X_{T_k}=N]P(X_{T_k}=N)}{P(X_{T_k}=0)}=\frac{k(2N-k)}{3}
$.}:
\begin{equation}
\mathbb{E}[T_k|X_{T_k}=N]=\frac{N^2-k^2}{3}.
\label{proof:var_rand_3}
\end{equation}
This allows us to derive the following quantity:
\begin{equation}
\mathbb{E}[T_kX_{T_k}^2]=P(X_{T_k}=N)\mathbb{E}[T_kX_{T_k}^2 | X_{T+k}=N]=Nk\mathbb{E}[T_k|X_{T_k}=N]=Nk\frac{N^2-k^2}{3}
\label{proof:var_rand_4}
\end{equation}
Finally, let us extract the expression for $\mathbb{E}[T_k^2]$ from the following martingale\footnote{Proof that $Z_n$ is a martingale is straightforward, and we leave it to the reader as an exercise.}: $Z_n = X_n^2-6nX^2_n+n(3n+2)$. Applying the martingale property and optional stopping theorem, we see that:
\begin{equation}
\mathbb{E}[Z_{T_k}]=\mathbb{E}[Z_0]=k^4.
\end{equation}
At the same time,
\begin{equation}
\mathbb{E}[Z_{T_k}]=\mathbb{E}[X^4_{T_k}]-6\mathbb{E}[T_kX_{T_k}^2]+3\mathbb{E}[T_k^2]+2\mathbb{E}[T_k].
\end{equation}
Applying the results in (\ref{deriv:5}), (\ref{proof:var_rand_4}), and (\ref{deriv:7}) to the two statements above, we get:
\begin{equation}
\mathbb{E}[T_k^2] = \frac{1}{3}\left(k^4-kN^3-2k(N-k)+2Nk(N^2-k^2)\right).
\end{equation}
Finally,
$$\mathbb{E}[T_k^2]-(\mathbb{E}[T_k])^2 = \frac{1}{3}\left(k^4-kN^3-2k(N-k)+2Nk(N^2-k^2)\right)-k^2(N-k)^2,$$
and hence \begin{equation}
Var(T_k) = \frac{k(N-k)}{3}(2k^2-2kN+N^2-2) \text{\hspace{0.2cm} when }p=q=1/2.
\end{equation}
The statement of the theorem follows from here by the usual substitution $k=\delta$, $N=2\delta$.
\section{Comparison with Majority Voting}
\label{sec:vs_MV}
Theoretical analysis further enables new insight into the performance of the $\delta$-margin majority voting scheme relative to other salient consensus aggregation schemes. In this chapter, we will compare $\delta$-margin voting to the most common consensus rule --- majority voting.
Figures~\ref{fig:Q_MV} and \ref{fig:MV_cost_ratio} display the expected quality of majority vote, as well as the estimated number of required workers for a majority voting aggregation scheme supplemented by an \textit{early stopping} mechanism that is used when a formal consensus number is not yet reached, but a particular voting outcome becomes imminent
\begin{figure}[t]
\centering
\includegraphics[height=0.43\columnwidth]{figs/Q_MV.png}
\caption{\rm{The ordinate axis represents the expected quality of a majority voting process with early stopping
with $n$ workers and worker accuracy $p$.}}
\label{fig:Q_MV}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[height=0.42\columnwidth]{figs/ratio_vanilla.png}
\includegraphics[height=0.42\columnwidth]{figs/ratio_early.png}
\caption{\rm{Left: ratio of expected cost (numbers of workers required for a consensus) between a $\delta$-margin majority voting scheme and a simple majority voting scheme. Right: ratio of expected cost between a $\delta$-margin voting scheme and a majority voting scheme with early stopping.}}
\label{fig:MV_cost_ratio}
\end{figure}
These graphs suggest strict dominance of the $\delta$-margin voting scheme compared with majority voting, in terms of the quality of the consensus vote.
|
\section*{Introduction: the structure-function problem in microbial communities.}
\label{S:1}
The evolutionary history of the biosphere is inextricably linked to the metabolic activities of microbes. Since life arose on this planet, microbes have lived in consortia
that saturated nearly every biochemical niche on the planet, driving global transformations in the chemical composition of the biosphere via metabolic processes from fermentation to photosynthesis to respiration~\cite{falkowski_global_2000,Canfield:2010ib,nelson_global_2016,sunagawa_structure_2015,zakem_redox-informed_2020}. As such, microbes and the communities in which they reside are the result of an ongoing eco-evolutionary process that couples the transformation of metabolites to the complex dynamics of interacting ecological systems across many spatial and temporal scales.
Given the importance of the metabolic activity of microbial communities we argue that a major goal for the field should be to predict, design, and control the metabolism of microbial communities in complex, natural, and engineered settings. Accomplishing this goal requires understanding how the structure of a community, in terms of the taxa present and its genomic composition, determines its metabolic activity in a given environmental context. The sequencing revolution has revealed the structure of microbial communities at the level of the taxa present, the genes they posses and the dynamics of gene expression.
This means that we now have a detailed and dynamic ``parts list'' for microbial communities in terms of taxa and genomic composition across a range of environments, from anaerobic digesters~\cite{Bocher:2015im,TOERIEN1969385,vanwonterghem_deterministic_2014} to the human gut~\cite{blanton_gut_2016,raman_sparse_2019}, soils~\cite{bahram_structure_2018}, and the ocean~\cite{sunagawa_structure_2015}. For some metabolic processes we can interpret gene content and taxa in terms of the specific metabolic processes that they are capable of. For example, we know the dominant taxa that perform processes such as nitrification~\cite{Bock2013} or polysaccharide degredation~\cite{sanchez-gorostiaga_high-order_2019}. Further, by annotating metagenomic data we can assign specific functional roles for many (but not all) of the genes present in a given community. As a result, we can measure the prevalence of enzymes that perform the reactions necessary for specific metabolic processes. Despite the remarkable scale and breadth of these sequencing data, we still do not have a predictive, quantitative framework for using these data to understand, predict, and design the metabolism of the communities in complex environments.
In this perspective we explore what makes this problem both challenging and important. We propose a specific approach to begin to address this question, and examine what types of communities and associated metabolic processes might be amenable to this approach. We review the techniques that are relevant to implementing the approach with a focus on methods for quantifying metabolites.
\subsection*{The significance of finding a solution}
Microbial communities play an outsized role in driving fluxes of nutrients through the biosphere. Photosynthetic microbes are responsible for nearly half of the carbon fixation on the planet~\cite{falkowski_role_1994}. These phototrophs work in concert with heterotrophic bacteria that enable primary productivity in terrestrial, marine and fresh water environments~\cite{kirchman_processes_2012,madigan_brock_2018}. We are only beginning to glimpse the role of the collective in this nearly 100 gigaton annual carbon flux. Bacteria and archaea in anaerobic environments degrade complex carbon sources to methane, playing an important role in carbon recycling and climate change~\cite{madigan_brock_2018}.
In the nitrogen cycle, microbes play a key role in nitrogen fixation (dinitrogen gas to ammonia), nitrification (ammonia to nitrate), and denitrification (nitrate to dinitrogen gas)~\cite{stein_nitrogen_2016}. These processes are key for wastewater treatment~\cite{cydzik-kwiatkowska_bacterial_2016} and human health~\cite{turnbaugh_human_2007}. A critical challenge is to form a quantitative and predictive understanding of how microbial communities drive these fluxes. To give a concrete example, the process of denitrification, performed by bacterial communities in soils, reduces nitrate to dinitrogen gas. An intermediate in the conversion of nitrate to dinitrogen is the potent greenhouse gas and ozone depleting compound nitrous oxide~\cite{tian_comprehensive_2020}. Denitrifying communities in some cases (especially in agricultural soils) can leak nitrous oxide, but in other cases fully convert nitrous oxide to harmless dinitrogen gas. The question then becomes: what controls the production of nitrous oxide from denitrifying communities in soils? Can we manipulate these microbial communities to limit nitrous oxide production? To address this we need to understand how the structure of the community and the environmental context determine the flux of metabolites through the system.
Similarly, the essential importance of resident microbiota in host health is now clear~\cite{turnbaugh_human_2007}, but as yet it is unclear how to rationally manipulate these communities to benefit the host. There exists tantalizing evidence that this can be done, for example by altering metabolic phenotypes~\cite{turnbaugh_obesity-associated_2006} or treating persistent infections~\cite{lawley_targeted_2012}, but we lack general approaches for developing such strategies. Here we focus on environmental microbiomes, but we emphasize that the strategy proposed here could, and in a few cases has been~\cite{raman_sparse_2019}, applied to host associated communities.
\subsection*{Defining structure and function}
Before moving forward we take a moment to define community structure and function. We define the structure of the community as the taxa present, as well as the genomic structure of each taxon, which may include everything from the detailed knowledge of the regulatory architecture of each gene, to the syntenic organization of the genome~\cite{junier_universal_2018}, to even the presence of phage. The structure of the community may, if necessary, include transcriptional or proteomic information at the metagenomic or single-taxon level as well.
We define the function of a community as the collective metabolic activity of all constituent organisms, which therefore operates in the space of metabolites. The dynamic or steady-state flux of metabolites through the consortium define its metabolic function. Depending on the context, the most important metabolic fluxes may include electron donors (e.g., organic carbon), acceptors (e.g., oxygen, nitrate), secondary metabolites, biomass, overall catabolic activity, or byproducts. A note about usage: some readers may find the term ``function" teleological, implying some sort of purpose on microbial communities. We use the term function to mean the \textit{activity or action} of a community without any implication of purpose. Despite this potential confusion, we find that using the term function is a useful shorthand. In particular, we would like to invoke a certain symmetry between the ideas presented here and the problem of sequence, structure and function at the level of proteins. The structure-function problem for microbial communities is therefore to deduce the mapping from the space of genes, transcripts, proteins and taxonomic organization to metabolite fluxes, and to understand the environmental dependence and context in which this mapping is relevant.
\subsection*{How should we approach the problem?} Understanding how the metabolic activity of a microbial community emerges from the taxa present and their metabolic capabilities is a problem of connecting heirarchical scales of biological organization, from genes to phenotypes and interactions in specific environmental contexts. What makes this challenging is the fact that processes at different scales feed back on one another. For example, compounds that mediate interactions between strains can do so by modulating gene expression~\cite{beliaev_inference_2014}. Similarly, phenotypic variation in one organism can modify interactions~\cite{mickalide_higher-order_2019}, and organisms modify their environment with widespread impacts on other members of the consortium~\cite{ratzke_ecological_2018}.
One way to proceed is via the reductionist mode that has motivated biology over the last century~\cite{woese2004}. In the context of communities, this would mean dissecting the mechanistic and physiological metabolic properties of each member of the community and understanding how metabolite dynamics emerge at the level of the collective. The challenge is the complexity of these systems, which makes a detailed, mechanistic understanding of the collective metabolism a massive undertaking. For synthetic communities comprising model organisms, where detailed information are available, this type of approach has had some success ~\cite{orth_what_2010,harcombe_metabolic_2014}. For comparatively simple communities of a few strains, such as bacteria cross-feeding amino acids~\cite{wintermute_emergent_2010} or ciliates consuming bacteria~\cite{mickalide_higher-order_2019}, detailed models have been constructed and validated. However, in environmental or host-associated contexts, where the number of strains present is enormous and many organisms present are poorly studied or challenging to work with in the laboratory, this approach faces huge challenges.
In this scenario what are we to do? On the one hand, building detailed models as described above is ill-advised. Even when such detailed models can be built, the challenge of distilling simple principles from these models increases with their complexity (see Borges ``On the exactitude of science''~\cite{borges_collected_1998}). On the other hand, we know from many examples that a huge number of processes from antibiotic warfare~\cite{vetsigian_structure_2011} to competition~\cite{Friedman:2017dk}, mutualism~\cite{hom_niche_2014}, and stress responses~\cite{amarnath_stress-induced_2021} influence interactions and metabolism. So how can we justify not building models that include these details?
In the spirit of Philip Anderson's influential essay~\cite{anderson_more_1972}, it may be that understanding communities requires explaining entirely new, and potentially simpler properties that are emergent at the level of the collective. In this case, due to the scale and complexity of the community, new and distinct phenomena emerge from the individual parts, and discovering the organizing principles requires an approach that goes beyond dissecting the detailed metabolic phenotypes of each individual member of the community. To be clear, we are not advocating the idea that reductionist approaches are not useful. Their utility is clear from many examples~\cite{jacob_genetic_1961,zumft_cell_1997,alon_robustness_1999,basan_universal_2020,amarnath_stress-induced_2021}. Instead, we are asking whether making headway on the ``structure-function problem'' as we have defined it might require a complementary approach that focuses not on the detailed mechanisms and physiology of each organism, but on patterns that are evident at the level of the collective. Here we discuss such an approach.
\subsection*{Learning the right variables: the power of statistics across ensembles}
We find it useful to cast the structure-function question in terms of a prediction problem. In this framing, the goal is to predict metabolite dynamics or fluxes from community structure for a given environment or set of environmental conditions. The key question then becomes, what structural elements must we quantify to predict function? Equivalently, we could ask, how predictive of metabolic function is community gene content, taxonomy, transcriptional or proteomic data?
One approach to this problem, which has found success in both physics and biology, is to look for statistical regularities across replicate systems and allow these patterns to naturally define variables that can be used to make predictions. In many cases, this approach can reveal the salient, emergent, features of a system and provide deep insight into their function.
Specifically, from proteins~\cite{halabi_protein_2009} to multicellular organisms~\cite{alba_global_2021}, examining statistical variation across many replicates of a system, or an ensemble, has proven powerful. For example, covariation across ensembles of homologous proteins have been used to reveal which amino acids are in contact in the folded structure~\cite{russ_natural-like_2005}, and co-evolving groups of residues that correlate with enzyme function~\cite{halabi_protein_2009}. Careful analysis of behavioral variation in large numbers of microbes~\cite{jordan_behavioral_2013} and flies~\cite{berman_mapping_2014} suggest that apparently very complex behaviors can in fact be described by a relatively small number of elementary behavioral features~\cite{berman_mapping_2014,katsov_dynamic_2017}. Statistical analysis of morphological variation across higher organisms suggests that morphologies adhere to constraints that some believe are associated with specific functional capabilities~\cite{raup_theoretical_1965,shoval_evolutionary_2012}. A recent study of variation in patterning in the fly wing shows that a single mode of variation describes the response of the wing developmental program to genetic and environmental perturbations~\cite{alba_global_2021}. For a recent piece discussing why low-dimensionality might be an inherent property of evolved systems see Eckmann and Tlusty~\cite{eckmann_dimensional_2021}.
The common feature of all of these examples is that by judiciously studying the variation across a carefully chosen ensemble of systems, one can often discover simple, relatively low-dimensional features that enable the prediction of the functional properties of the system. Here we advocate a similar approach to communities.
\subsection*{An ensemble approach to the structure-function problem}
In light of these considerations, we propose an ensemble approach to the structure-function problem in communities. We motivate this approach by analogy to a similar approach taken at the level of proteins (Fig.~\ref{Fig1}). For proteins the structure function problem is to predict the fold and function of a protein from its amino acid sequence. One way to approach this problem is by performing detailed physico-chemical simulations of a polypeptide chain via molecular dynamics (Fig~\ref{Fig1}A). While much progress has been made in this approach, it has proven a huge technical challenge. However, a statistical approach, ignoring the mechanistic details and instead considering only statistics of a multiple sequence alignment does a remarkably good job of predicting protein folds~\cite{morcos_direct-coupling_2011}. Similar approaches reveal low-dimensional structure in proteins which is predictive of function~\cite{halabi_protein_2009,russ_evolution-based_2020}. These studies suggest that much can be learned by carefully considering variation across a suitably chosen ensemble of systems.
In the context of microbial communities flux balance models of community metabolism are analogous to molecular dynamics simulations in proteins because they attempt a detailed mechanistic accounting of all of the phenomena within a community (Fig.~\ref{Fig1}B). However, in communities there is comparatively little work pursuing a statistical approach analogous to the one taken at the level of proteins.
Taking such an approach is precisely what we are advocating here. We propose quantifying the structure of a collection, or ensemble, of communities using sequencing while simultaneously measuring metabolite dynamics. Given such data, one can then approach the structure-function problem by asking whether variation in community structure (e.g., across metagenomes or metatransciptomes) permit quantitative insights into the functional properties of these communities. The proposal is then to leverage these insights to design, predict and control community function. Several studies in the past few years have begun to explore statistical approaches~\cite{Gowda:2021ve,raman_sparse_2019}. However, we suggest that this approach is under explored and that that there is a pressing need to collect new data sets that are explicitly designed to pursue a statistical approach to the structure-function problem in communities.
\begin{figure}
\centering\includegraphics[width=1\textwidth]{Fig1.png}
\caption{\textbf{Sequence, structure and function in proteins and microbial communities.} We propose that there exist analogous solutions to the sequence-structure problem in protein folding and the structure-function problem in microbial communities. (A) The mapping from amino acid sequence to 3D protein structure can be accomplished either by a simulation approach (e.g., molecular dynamics) or by a statistical approach (e.g., direct coupling analysis). The former is a computationally-intensive strategy to simulate 3D protein structure based on first-principles modeling of atomic interactions. The latter leverages information about residue coevolution from an ensemble of amino acid sequences to infer which residues are in contact, allowing for an elegant and interpretable statistical inference of 3D structure. (B) The mapping from genomic and metagenomic sequences to community metabolic activity can be achieved through community flux balance modeling, or, as we propose, a statistical ensembles approach. The former requires genome-level metabolic models of each organism to be built, a labor-intensive iterative process that so far has been successful primarily in a handful of model organisms. The latter leverages the diversity and variation in an ensemble of communities to learn an an effective mapping between community sequence content metabolic activity.}
\label{Fig1}
\end{figure}
\subsection*{Overview of the paper}
We will focus on three main challenges that must be surmounted to apply the ensemble approach to the structure-function problem: (1) choosing an ensemble across which one should make comparisons and look for patterns, (2) measuring metabolite dynamics, which requires analytical chemistry techniques that are often not standard practice in microbial ecology labs, and (3) using these data to distill the mapping from structure to function.
Our intention is to provide a roadmap for how such an approach might be applied across communities of interest. We recognize that this roadmap is far from complete, and that many pitfalls exist that may render this approach challenging in various circumstances.
We will not review sequencing technologies, which have been widely and capably recapped elsewhere~\cite{knight_best_2018}. We will focus largely on microbial communities in environmental contexts rather than health-related (human microbiome) contexts, in part because environmental microbiology is where our expertise lies, but also because of the abundance of existing literature on the latter topic. Finally, we will neglect the many recent advances in theoretical ecology, in particular the renaissance in consumer-resource models applied to communities, in service of focusing on questions that can be settled empirically.
\section*{Choosing communities and ensembles}
\label{S:2}
The first challenge is choosing a community and an associated metabolic process to interrogate. The choice is far from trivial and there is no simple prescription. Instead we appeal to the intuition of microbial ecologists, experts in physiology and the quantitative considerations of applied mathematics and physics. The goal should be to circumscribe a well-defined community and, if possible, an associated metabolic process where quantitative measurements in many replicates can be made.
Choosing a metabolic process, and therefore specific metabolite measurements to be made, is a significant challenge, and compromises are inevitable. Here we reiterate the point that microbial communities are not ``functional'' in the sense that they are designed with some purpose. Despite this fact, as we will discuss below, the importance of metabolic activity of communities in specific niches is clear, as is the relevance of these processes for the biosphere more broadly. To make this point clear, we begin by discussing specific communities and their associated environments and metabolic processes, with an eye towards how one might apply the ensemble approach to dissecting their function.
\section*{Model microbial communities}
\label{S:3}
\subsection*{Structure-function in the wild}
\begin{figure}
\centering\includegraphics[scale=0.25]{Fig2.png}
\caption{\textbf{Community structure and function in the wild.} (A) Algal blooms are a microbial successional processes that follow from the input of exogenous nutrients to aquatic environments. Reduced carbon fixed from CO$_2$ by algae are consumed, along with other nutrients, by heterotrohic bacteria in reproducible successional dynamics. (B) Marine snow particles are aggregates of organic carbon that are formed near the ocean's surface and subsequently sink to the ocean floor. Microbial communities can degrade these particles, and the amount of carbon that is mineralized to CO$_2$ versus the amount that is sequestered on the ocean floor depends strongly on the structure of the microbial community. (C) Microbial mats are layered communities that occur at air-water interfaces, often in extreme thermal environments such as hot springs. The spatial structure of these communities follows from exchanges of nutrients governed by redox gradients. (D) Pink berries are microbial aggregates that cryptically (internally) cycle sulfur between photosynthetic purple sulfur bacteria and anaerobic sulfate reducing bacteria.} \label{fig:inthewild}
\end{figure}
\subsubsection*{Soils}
Perhaps no microbial community on Earth is more important than that which inhabits soils. The soil microbiome plays a key role in plant growth and physiology ~\cite{saleem_more_2019}, in particular through nitrogen fixing bacteria that provide reduced nitrogen for plant hosts. The storage of carbon, and production of CO$_2$ via respiration of reduced organic compounds in soils are key components of the global carbon cycle~\cite{lal_soil_2004,lal_soil_2004}. As the climate warms, so the rate of microbial respiration of CO$_2$ from soils increases~\cite{kirschbaum_temperature_1995,allison_soil-carbon_2010}, potentially driving a positive feedback loop with dire consequences for the global climate. Similarly, denitrification in agricultural soils is responsible for roughly \SI{80}{\percent} of the anthropogenic release of nitrous oxide (N$_2$O) (see \url{https://www.epa.gov/ghgemissions/overview-greenhouse-gases} and~\cite{tian_comprehensive_2020}). Nitrous oxide is 300-fold more potent than CO$_2$ as a greenhouse gas and responsible for approximately \SI{10}{\percent} of the global warming potential from human activity. For these reasons, there is keen interest in associating soil microbiome structure to process rates such as CO$_2$ or N$_2$O production and N$_2$ fixation. However, most attempts to find a relationship between soil microbiome structure and the rates of key metabolic processes in soils have found only marginal success~\cite{graham_microbes_2016,rillig_role_2019,rocca_relationships_2015,fierer_embracing_2017}.
Despite these difficulties we see reason for optimism. It is known that a relatively small number of environmental factors are the dominant drivers of variation in soil community structure: pH, moisture, carbon and nitrogen availability, temperature, and redox potential~\cite{fierer_embracing_2017}. Moreover, while soils are routinely cited as very complex microbial communities, much like the human gut they are typically dominated by a handful of taxa (e.g., \textit{Acidobacteria}, \textit{Verrucomicrobia}~\cite{fierer_embracing_2017,crits-christoph_novel_2018}), with most other strains present in relatively low abundances. Moreover, there are clear patterns in the abundances of bacteria and fungi in soils, with high biomass turnover environments like grasslands dominated by bacteria and low biomass turnover forests dominated by fungi~\cite{fierer_embracing_2017}.
As acknowledged in recent meta-studies~\cite{graham_microbes_2016}, one challenge in associating soil community structure to metabolic function is a lack of high quality data sets where process rates (e.g., CO$_2$ production) are measured in a large ensemble of soil communities. Exceptions to this include a recent survey of global topsoil microbiomes~\cite{bahram_structure_2018}, and microcosm studies documenting the role of multiple environmental perturbations applied to soils~\cite{rillig_role_2019}. Despite these advances, the consensus remains that predicting metabolic processes in soils from microbial community structure is challenging~\cite{fierer_embracing_2017}. However, we note that in some cases this conclusion is derived from meta-studies that aggregate data from different experiments or labs. Such comparisons can be challenging given systematic errors in sequencing measurements between protocols~\cite{mclaren_consistent_2019}, and the strong dependence of measurements like soil pH on the technique employed~\cite{miller_comparison_2010}.
Given these considerations we propose that one route forward is the judicious collection of data from large ensembles of soil communities followed by careful quantification of process rates and community structure in a consistent manner. We acknowledge that even in the presence of such data, relating community composition at the taxonomic, genomic or transcriptomic levels to metabolite fluxes may remain a challenge. It may be that soil taxa are not ``good variables" for understanding function, and that chemical or abiotic properties such as the redox state of available organic carbon are necessary~\cite{keiluweit_anaerobic_2017}. Regardless, the importance of understanding the metabolic activities of soil communities cannot be overstated.
\subsubsection*{Algal blooms}
In aquatic systems, exogenous inputs of nitrogen and phosphorous, often driven by human activity, result in the dramatic successional process of the algal bloom~\cite{Teeling2012} (Fig.~\ref{fig:inthewild}A), wherein photosynthetic microbes explode in abundance. The rise of phototrophic microbes brings with it a complex community of non-photosynthetic (heterotrophic) bacteria that form tight symbioses with the phototrophs. These phototroph-heterotroph communities cycle large quantities of carbon, with phototrophs fixing CO$_2$ to reduced organic carbon, which is in turn consumed by the associated heterotrophic bacteria.
In this system, a natural framing of the structure-function problem would be to ask how the taxonomic composition of the community impacts the fixation of carbon and eventual bacterial biomass production. Teeling \emph{et al.}~\cite{Teeling2012} have found that specific classes of bacteria grow at different phases of the bloom, e.g., Alphaproteobacteria dominated the pre-bloom, as blooming commenced, Bacteroidetes increased more rapidly than others, and then Gammaproteobacteria grew much later. Concurrent metagenomic studies showed that the abundances of enzymes capable of degrading carbohydrates and sulfatases required to degrade sulfated algal polysaccharides increased in abundance during the bacterial succession. The degradation of the larger polysaccharides leads to the production of shorter organic compounds, which was revealed by the expression of the relevant transporters. These results, and those of other studies~\cite{kimbrel_host_2019,louati_structural_2015,mcfeters_growth_1978,parulekar_characterization_2017,ramanan_phycosphere_2015}, indicate a link between taxonomy of bacteria associated with certain photosynthetic strains during blooming events.
These results suggest that these associations are at least in part determined by the carbon catabolic activity of the bacteria and the fixed carbon that phototrophs excrete (see also \cite{Buchan2014}). Taking an ensemble approach to this problem could be accomplished by recapitulating bloom dynamics in a laboratory context~\cite{riemann_dynamics_2000}, where the identity of the phototroph and the composition of the bacterial community could be manipulated and the resulting total carbon flux quantified~\cite{astacio_closed_2020}. Such studies might shed insights on how to control blooms in natural settings.
\subsubsection*{Marine snow}
Marine snow is a term used to describe micrometer to millimeter-scale aggregates in the ocean, which are typically made of detritus containing carbon and nitrogen in the form of polysaccharides, microbes, and inorganic substances. These particles form at the upper levels of the ocean and sink to the ocean floor, transporting carbon in a ``pump'' that removes carbon from the atmosphere over geologic timescales~\cite{Alldredge1988,Gralka2020}. Communities of bacteria and other microbes consume these aggregates as they sink (Fig.~\ref{fig:inthewild}B). Understanding how the structure of marine snow communities impacts the degradation of carbon and the production of metabolic byproducts (e.g., CO$_2$) is a critical question for understanding global carbon fluxes.
Studying these particles \emph{in situ} is challenging~\cite{Kiorboe2007}. To circumvent this difficulty, some studies use synthetic particles to study community assembly and function. One study used agar beads \cite{Kiorboe2003} to study the colonization by microbes in raw seawater, revealing successional dynamics underlying particle degradation. The particles were initially colonized by bacteria, and later by flagellates. Cordero and colleagues have taken a similar approach to understand colonization dynamics~\cite{datta_microbial_2016,ebrahimi_cooperation_2019,Enke2019,pontrelli_hierarchical_2021}. These studies shed light on the metabolic roles played by different players during particle colonization dynamics: primary degraders excrete enzymes to break down polysaccharide chains, which creates a niche for secondary degraders capable of consuming monomers and oligomers of the polysaccharide, which in turn makes way for scavengers that take up metabolic byproducts of the primary and secondary degraders (ammonia, amino acids).
In terms of the approach proposed here for mapping structure to function, particle degradation offers a powerful model system because so much is known about how the communities collectively degrade the particle. For example, it would be interesting to see whether one could take a statistical approach to inferring the key metabolic traits of primary and secondary degraders and scavengers from sequencing data alone. In this case the function of the community to predict would be the fraction of carbon degraded or the total CO$_2$ respired. The power of this system is the ability to manipulate structure, but the particles make it challenging to measure carbon degradation directly (see the next section for further discussion).
\subsubsection*{Plastisphere}
Closely related to the degradation of polysaccharide particles in the ocean is the recent rise of plastic debris in freshwater and marine environments, and the microbial communities associated with its degradation \cite{Amaral-Zettler2020}. Given the fact that a few million tons of plastic enter the ocean per year \cite{Jambeck2015}, this is an important ecosystem to study, not only to understand how these pollutants affect the ecosystem, but also to find potentially efficient plastic degrading communities. Further, plastic is a comparatively new environment on evolutionary timescales, making it interesting to study from the perspective of evolution. It has been shown that the community composition on plastic is different from the composition on other substances in the same conditions, and that certain taxa are commonly found on plastics \cite{Dussud2018,Kirstein2019}. Diatoms have been observed in high numbers, although a strong succession dynamics was observed. Other photoheterotrophs, heterotrophs, ciliates, fungi, and pathogens have also been observed. Although the functional role of these microbes are unclear, it is speculated that chemotaxis, interactions with metals and degradation of low molecular weight polymers are important factors that determine community composition \cite{Amaral-Zettler2020}. In this system the structure-function problem is similar to that outlined for marine snow - how does community structure determine the rate of plastic degradation?
\subsubsection*{Microbial mats}
Microbial mats are stratified communities that often form at air-water interfaces in extreme environments such as hot springs~\cite{Klatt2013}. Mat communities harbor a top layer of photoautotrophic bacterium (typically \textit{Synechococcus} sp.) that use light to fix carbon during the day and often fix nitrogen at night. Below the top layer are strata of various heterotrophs and anaerobic autotrophs~\cite{ward_natural_1998} (Fig.~\ref{fig:inthewild}C). These communities have been studied for decades, and much is known about the metabolic roles each strain plays in the community~\cite{bateson_photoexcretion_1988,anderson_formation_1987} and metagenomic datasets are available~\cite{lee_metagenomics_2018}. It is remarkable that similar mat structures form in many different contexts across the globe. In these mats much of the nutrients are fixed from CO$_2$ and N$_2$~\cite{steunou_situ_2006}, heterotrophic community members then consume reduced organic carbon excreted by the autotrophs. In this context the question of relating structure to function falls to mapping the flux of C, N and other metabolites through the mat to the taxonomic and metagenomic structure of the system. For example, what is the simplest community that will stably form a mat? What pathways, for example in the primary autotroph, are essential to mat formation and which are dispensable? Further, given that mats support remarkable allelic diversity driven by extensive recombination~\cite{rosen_fine-scale_2015}, how is the functional genetic repertoire of these communities maintained? Preliminary work on mats in California \cite{Lee2018} showed the presence of a core genome across samples, and various other genes thought to be useful for specialized functions, but further metagenomic analysis is likely to be useful in addressing the structure-function question in mats. The main two challenges in these communities are obtaining axenic isolates from the mats and making high quality metabolite dynamics measurements \emph{in situ}. Petroff and colleagues have recently made exquisite quantitative measurements of oxygen dynamics in mats~\cite{petroff_subsurface_2017,tejera_oxygen_2018}, addressing the latter challenge, which opens the door to making quantitative links between community composition and metabolic activity.
\subsubsection*{Cryptic sulfur cycling in microbial aggregates}
Microbial metabolism drives the global cycling of sulfur through energy-generating oxidation and reduction reactions. It has become increasingly evident that much of this cycling occurs cryptically~\cite{Canfield:2010ib, Callbeck:2018iw} (i.e., with low steady state metabolites levels but substantial fluxes), often in the context of cellular aggregates where oxidation and reduction reactions occur in physical proximity~\cite{Wilbanks:2014kx, Callbeck:2018iw}. Inferring the presence and rate of cryptic sulfur cycling in microbial aggregates has important implications for our understanding of other elemental cycles, since sulfur cycling is often tightly coupled with the carbon and iron cycles.
Remarkable examples of cryptic sulfur cycling phenomenon are the so-called ``pink berry’’ consortia~\cite{Seitz:ht,Wilbanks:2014kx}, which are discrete macroscopic ($\sim$1 cm) aggregates that occur on the surface of submerged sediments in the intertidal pools of Sippewissett Salt Marsh (Massachusetts, USA) (Fig.~\ref{fig:inthewild}D). The bright pink coloration of these aggregates is attributable to the purple sulfur bacteria (PSB) that make up the majority of cellular biomass. These PSB oxidize sulfide to sulfate via the process of anoxygenic photosynthesis. Accompanying these PSB are sulfate reducing bacteria (SRB), which derive energy from anaerobic respiration by catalyzing the reverse process, reducing sulfate to sulfide. Together these PSB and SRB are capable of locally and cryptically cycling sulfur via the syntrophic exchange of oxidized and reduced sulfur compounds~\cite{Wilbanks:2014kx}.
While culture-based approaches to characterizing and reconstituting the pink berry consortia in the lab remain a challenge, the discrete nature of the pink berries presents an opportunity to statistically characterize the structure-function relationship at the level of individual aggregates harvested from the wild. Bulk metagenomic sequencing of pink berry consortia indicates that typically 2-3 phylotypes make up the majority of biomass in the aggregates~\cite{Wilbanks:2014kx}. Laboratory measurements of sulfide oxidation or sulfate reduction rates using spatially-explicit microprobe measurements followed by 16S metagenomic sequencing on harvested pink berries could provide insight into how the abundance of these phylotypes quantitatively relates to sulfur cycling. More generally, such an approach could be applied to characterizing cryptic sulfur cycling in other aggregated contexts, such as in marine particles~\cite{Callbeck:2018iw}. Predicting sulfur cycling is important to understanding other critical elemental cycles: sulfide oxidation by PSB can contribute substantially to carbon fixation~\cite{Dyksma:2016kh}, and sulfate reduction by SRB can be an important sink for reduced carbon~\cite{Liamleam:2007gc} and iron~\cite{Enning:2014ew} in environments where electron acceptors are otherwise scarce.
\subsection*{Community structure and function under domestication}
\begin{figure}
\centering\includegraphics[scale=0.4]{Fig3.png}
\caption{\textbf{Community structure and function under domestication.} (A) Kefir grains extracellular-polymeric aggregates host to microbial communities that inoculate milk for the production of kefir. These communities undergo a reproducible successional process that involves the production and consumption of fermentation byproducts, which ultimately give kefir its desired flavor. (B) Anaerobic bioreactors often use granulated microbial communities to remove waste products such as reduced carbon, nitrogen, and phosphorous from water. Improving the performance and efficiency of these systems through the ensembles-informed design of communities would increase their viability as alternatives to traditional wastewater treatment approaches that expend significant energy on aeration.}\label{fig:domestication}
\end{figure}
Many key insights in the Origin of Species came from studying trait variation under domestication. In the same way, we propose that learning the rules for mapping structure to function in communities should leverage the many instances in which communities have been domesticated. Here we explore some of these opportunities.
\subsubsection*{Microbial communities in the dairy industry}
The production of yogurt, cheese and other dairy products relies heavily on microbes. The work of Dutton and colleagues on cheese rinds \cite{Wolfe2014} is an important example in this context. This study compared and characterized more than 100 cheese rinds from across 10 countries, and found that less than 15 bacterial taxa and 10 fungal taxa were present in abundances of more than $1\%$. Further, most of these taxa are not present in the starting cultures, and their function is unknown. They found that the community composition is strongly correlated to the aging process and moisture rather than geography or milk source. The functional profile of the communities, found using shotgun metagenomics, was correlated to pH as well, and pathways were shown to correlate as expected with the cheese types. These results point to the idea that the chemical environment is perhaps the strongest determinant of community structure.
Remarkably, when the dominant taxa were reconstituted \emph{in vitro} and cultured as a community in media to mimic treatments in different types of cheese rinds, divergent communities were formed depending on the treatments, which showed some properties similar to the original cheese rinds including their abundance dynamics. This remarkable result shows that these communities can be reconstituted in the laboratory to recapitulate some of the basic functional features of the domesticated cheese communities. It would be interesting to extend these studies further by looking quantitatively at the metabolite dynamics. The fact that these communities can be so readily manipulated means that learning the relationship between composition and metabolic activity is now accessible. What remains unclear, to us, is what the salient metabolic features of these communities are that should be explained. One way to approach this question would be to chemically characterize the transformations that occur in specific rinds and to then ask whether subsets of the full community can or cannot recapitulate these processes.
Another promising microbiome in the dairy sector is the kefir community (Fig.~\ref{fig:domestication}A). Kefir grains are used to make a fermented drink like yogurt. They have about 50 bacterial and yeast taxa, which are resilient to stresses, and most of which perform lactic and acetic acid fermentation. A recent study \cite{Blasche2021} found that kefir grains, which are polysaccharide matrices synthesized by the bacterial consortium, collected from diverse locations had a very similar core community and differed only in rare species. Kefir grains are sustained much like sourdough starters and added to milk to initiate a fermentation process. When added to milk the composition of the grain community is stable while the community present in the milk exhibits a succession. Metabolite changes during the colonization showed similar succession dynamics. The study dissects the interactions between the community on the grains and that in the milk, and demonstrates reproducible metabolite dynamics in this system. As a result, kefir constitutes another powerful platform for manipulating community structure (e.g., composition of the community on the grains) and learning the impact of those changes on the metabolite dynamics. Blasche \emph{et al.} have already made significant progress in this regard, but it remains to investigate in a high-throughput statistical fashion, how the composition of the grain community confers the remarkable robustness they observe.
\subsubsection*{Anaerobic bioreactors}
The treatment of wastewater for reuse and release into the environment requires the removal of large quantities of organic matter, much of which is insoluble or otherwise slow to degrade. Anaerobic bioreactors serve an important role in this industrial process, harnessing microbial metabolism to degrade such organic matter into CO$_2$ and CH$_4$ gases (Fig.~\ref{fig:domestication}B). Because the bioreactors are fed a range of inputs and are operated under a variety of conditions, the microbial communities that populate bioreactors are highly functionally and taxonomically diverse~\cite{Werner:2011hm}. Organisms that excrete extracellular enzymes degrade insoluble polymers into soluble monomers, while fermenters consume these compounds and excrete products including acetate and H$_2$. Methanogenic archaea can then ferment these products to produce CH$_4$~\cite{TOERIEN1969385}. Many functional attributes are used in practice to characterize the performance of anaerobic bioreactors, including removal of chemical oxygen demand (COD, a proxy for aerobically metabolizable matter in a water sample) and methanogenic activity. The resilience of a bioreactor community to input fluctuations has also been of interest~\cite{Fernandez:2000cq, Hashsham:2000fw, Werner:2011hm}.
Several studies have explored the statistical relationship between community taxonomic structure and methanogenic bioreactor performance. In an important early study, Tiedje and colleagues found that, under constant conditions, COD removal in a laboratory bioreactor was stable while community composition varied substantially over a two year period~\cite{Fernandez:1999ew}. This work suggested the role of functional redundancy~\cite{Louca:2018cv} in maintaining community function, and implied a degeneracy in the relationship between community taxonomy and function. However a more recent study of several industrial-scale bioreactors observed relative stability in both community taxonomy and reactor performance over a year-long period~\cite{Werner:2011hm}. Notably, variations in reactor performance were found to be related to community composition. This indicated that taxonomy is predictive of community function, though the authors argued that taxonomy is simply a proxy for functional genomic content due a close correspondence between phylogeny and metabolic function for organisms in anaerobic bioreactor systems. The conflict between these two results indicates that much is still unknown about the structure-function relationship in anaerobic bioreactors.
Recent work has advanced our understanding of structure and function in methanogenic bioreactors, leveraging a statistical ensembles approach to discover a predictive relationship between gene content and methanogenic activity~\cite{Bocher:2015im}. The authors generated 49 diverse enrichment cultures by seeding laboratory bioreactors with inocula taken from a large and eclectic collection of industrial-scale bioreactors. This ensemble enabled a linear regression approach to mapping methanogen activity (as measured by methane production rate per unit biomass) to variation in genomic content, specifically the abundance of sequence variants of a gene important to methanogenesis (\textit{mcrA}). Remarkably their approach produced a predictive model with relatively few variables, suggesting that only a few key \textit{mcrA} variants, or strains possessing these variants, are important for methanogenic activity. A powerful consequence of this approach is a prediction for which gene variants would improve the performance of an underperforming bioreactor.
Given the importance and widespread use of bioreactors to process organic matter, these constitute important model systems. The field faces two important challenges. First, cultivating many of the slow growing taxa in these communities is difficult and this means that only equipped and experienced labs can readily work with these organisms. Second, the complexity of these communities makes carefully controlled and reproducible experiments a challenge, and as a result comparisons from one study to the next can be difficult. Standardizing conditions and starting inocula would therefore be a major advance.
\begin{figure}
\centering\includegraphics[scale=0.2]{fig4.png}
\caption{\textbf{Bringing wild communities into the lab.} (A) Winogradsky columns are laboratory-assembled are lab-assembled communities with distinctive and reproducible spatially-stratified metabolite fluxes. These fluxes, and consequently community structure, arise from emergent redox gradients. (B) Materially closed ecosystems are communities grown in sealed vessels whose only energy input is light. Nutrient cycling in closed ecosystems arises from phototrophic organisms generating reduced carbon by fixing CO$_2$, which can then be consumed by heterotrophic organisms. Predators such as ciliates can consume whole cells, facilitating the recycling of macromolecular biomass. (C) Serially-passaged communities enrich a complex environmentally-derived community on laboratory-controlled nutrient conditions (e.g., a fixed carbon source). The resulting communities are typically low-complexity, and demonstrate reproducible trophic roles and patterns of nutrient exchange.}
\label{Fig4}
\end{figure}
\subsection*{Bringing wild communities into the lab: enrichment cultures}
Another route to studying communities that is similar to domestication is to bring complex communities from nature into the laboratory and culture them in defined conditions. While these experiments allow the experimentalist to control the growth and incubation conditions, they often result in a drastic loss of diversity. As a result, this approach is likely poorly suited to understanding the structure of natural systems, but none the less should enable important insights into engineering and controlling communities.
\subsubsection*{Winogradsky columns}
One of the most well-known methods for enrichment culture is a Winogradsky column (Fig.~\ref{Fig4}A). The method was developed by Sergei Nikolaievich Winogradsky to study (and discover) chemosynthesis, a process where energy is derived from inorganic compounds \cite{Zavarzin2006}. Winogradsky columns are glass cylinders (or bottles) loaded with a sediment, supplemented an organic carbon source (typically paper) and a sulfur source, that is then sealed off and illuminated. With time, different microbes occupy different levels of the column based on their metabolic capabilities. Each layer is characterized by distinct redox reactions (electron donor/acceptor pairs) that support metabolism locally (much like the Mats discussed above), and via diffusion to other strata impact the metabolism of the entire column. For example, the top layer supports phototrophs that produce carbon and oxygen, which drives a second layer that aerobically respires carbon and oxygen, resulting in a third layer that is anaerobic and typically uses alternate electron acceptors such as nitrate or sulfate. Recent work has begun to show that these complex communities are amenable to quantitative interrogation in the laboratory \cite{pagaling_community_2014,pagaling_assembly_2017}. Of particular note is a study that used centimeter long glass capillaries to assemble stratified communities cultured in the presence of dyes to report pH and redox~\cite{quinn_winogradsky-based_2015}. A simple imaging setup then permitted the acquisition of quantitative spatiotemporal data on community assembly in many replicates. These systems were used to simulate communities in the lung of a cystic fibrosis patient, but an opportunity to extend this work remains and the platform is well-suited to the ensemble approach outlined here.
Given that metabolic niches are spatially separated in Winogradsky culture systems they are ideal systems for understanding how this stratification self-organizes, their energetics and how this organization depends on the boundary and initial conditions in the system. For example, if the capillary system of Quinn \emph{et al.}~\cite{quinn_winogradsky-based_2015} could be combined with defined nutrient conditions, controlled illumination and quantitative imaging one could ask how the layers (and therefore the metabolite fluxes) depend on community composition. Given the timescale of these experiments (months) working with many replicates in parallel will be key.
\subsubsection*{Materially closed ecosystems}
Closely related to Winogradsky columns are materially closed microbial ecosystems (CES, Fig.~\ref{Fig4}B). Closed ecosystems are hermetically sealed, typically aquatic, microbial communities that have been shown to sustain life with only light as an input for decades in some cases. These `ecospheres' are available commercially (\url{https://eco-sphere.com/}) and have been studied in an academic setting since the 1960s. CES contain photosynthetic microbes, typically algae or bacteria, and either simple or complex consortia of heterotrophic bacteria and predators. When provided with only light these communities self-organize to sustain nutrient cycles and therefore the community itself. CES act as model biospheres since they require nutrient cycling to persist~\cite{rillig_microbial_2019}. In the context of the structure-function problem the salient metabolic property of CES is nutrient cycling, and one question is how the composition of the community determines nutrient cycling rates and persistence.
Work from a number of groups~\cite{obenhuber_carbon_1988,obenhuber_eucaryoteprocaryote_1984,kearns_measurement_1981,kawabata_synthesis_1995,taub_closed_1974,taub_community_2009} showed that CES containing primarily microbes were tractable model systems and methods for quantifying nutrient cycling were developed~\cite{obenhuber_carbon_1988,taub_community_2009}. More recently, Leibler and Hekstra~\cite{hekstra_contingency_2012} and later Leibler, Frentz and Kuehn~\cite{frentz_strongly_2015} studied population dynamics in a synthetic closed ecosystem comprised of three species using sophisticated microscopy-based methods. These studies revealed remarkably deterministic population dynamics in CES. However, few studies have quantitatively characterized the nutrient cycling capabilities of CES with the notable exceptions of Obenhuber~\cite{obenhuber_carbon_1988} and later Taub~\cite{taub_community_2009}. The early work of Obenhuber showed that complex bacterial communities mixed with photoautotrophic algae or bacteria could sustain a carbon cycle for many months~\cite{obenhuber_carbon_1988}. Inspired by these studies we recently developed a higher throughput method for quantifying carbon cycling that uses low-cost microelectromechanical sensors (MEMS) made for mobile devices to measure small changes in pressure inside a CES~\cite{astacio_closed_2020}. As appreciated by Obenhuber and Folsome, changes in pressure reflect carbon cycling because oxygen is lower solubility in water than CO$_2$. When photosynthetic microbes fix CO$_2$ they produce oxygen and the pressure rises. When bacteria respire carbon the opposite happens and one can quantify carbon cycling by measuring pressure oscillations during light-dark cycles. We constructed CES using bacterial communities derived from soil samples combined with a domesticated strain of the alga \textit{Chlamydomonas reinhardtii}. We found that taxonomically diverse CES stably cycled carbon for as long as six months. Metabolic profiling of these communities showed that despite taxonomic variability across replicate CES each community exhibited similar metabolic capabilities in terms of the carbon compounds they could utilize. It remains to be understood how such taxonomically distinct communities can retain this diversity while exhibiting similar metabolic capabilities. To address this question using the approach outlined here would require studying many synthetic communities with varying composition while measuring carbon cycling.
We propose that CES constitute powerful model communities for understanding nutrient cycling. In the context of the approach outlined here, CES could be used to understand how initial nutrient supply (C, N, P, S etc) controls community structure and cycling. Further, since light is the only source of energy CES can be used to explore how energy availability impacts the structure-function mapping in terms of cycling.
\subsubsection*{Enrichment in defined media}
Recently Sanchez and collaborators performed serial enrichment cultures on complex natural communities in simple defined media containing a single carbon source~\cite{Goldford:2018jf,Estrela:hb} (Fig.~\ref{Fig4}C). In this case, the function of the community is the conversion of glucose to CO$_2$ and biomass. Over a few tens of generations these communities assembled into relatively simple, and predictable consortia comprised of a few cross feeding strains. This convergent structure was typified by characteristic ratio of strains from the Enterobacteriaceae and Pseudomonadaceae families. To explain this conserved structure, strains were isolated from the endpoint of the enrichment experiment. They were then assayed for growth on glucose, and Enterobacteriaceae strains were found to grow faster on glucose than the Pseudomonadaceae. Metabolomic measurements (mass spectrometry) revealed that Enterobacteriaceae strains excreted intermediates such as acetate, which Pseudomonadaceae were observed to consume rapidly. Therefore, similar to the polysaccharide particle experiments discussed above~\cite{datta_microbial_2016}, these experiments revealed a primary carbon degrader that rapidly consumed the supplied carbon source but released secondary metabolites (in this case acetate) that supported the growth of other strains. Very recent work suggests that this type of cross-feeding may arise from stress induced release of nutrients that arises due to serial dilution~\cite{amarnath_stress-induced_2021}. In these communities the salient metabolic property of the system is carbon degradation. These studies have already revealed much about how the structure of these communities impacts the carbon degradation (e.g., the role of secondary consumers). It would be interesting to more fully elucidate the mechanistic basis of this cross-feeding in the service of understanding how the genotypes of each strain present determine their uptake and release of carbon compounds. Hopefully these insights would allow us to understand how the cross-feeding depends on available carbon compounds and environmental parameters such as pH. These generalizations could prove insightful in natural contexts such as soils where carbon degradation is a critical phenomenon for climate change.
\subsubsection*{Bridging structure and function with synthetic communities}
In recent years a number of studies have attempted to bridge structure and function of wild communities by performing experiments on synthetically assembled communities of natural isolates. This approach offers an opportunity to capture much of the substantial taxonomic and genomic diversity of natural communities within a setting where environment and composition can be controlled, and function can be measured accurately.
A handful of studies have explored how interactions between strains in a community affect carbon utilization. In part, these studies attempt to determine how prevalent different types of interaction between strains are (e.g., mutualism, parasitism, etc.), and how these interactions vary based on community composition and the identity of the carbon substrate provided. Using strains isolated from tree-associated environments, Foster and Bell~\cite{Foster:2012kl} assembled synthetic communities and measured CO$_2$ evolution during growth on a complex medium as a metric for community productivity. The measured values of productivity were compared to a “non-interacting” null prediction obtained by summing the productivities of the constituent strains grown in monoculture. What was observed in the vast majority of pair cultures and higher-order communities was consistent with resource competition rather than mutualism, suggesting that competitive interactions for carbon utilization dominate in natural communities. This implies that the carbon utilization of a community should saturate with the diversity of a community, which is precisely what is shown in a more recent study by Yu et al.~\cite{Yu:2019dw}. In this study a diverse ensemble of communities with varying strain composition and diversity was generated from a seawater inoculum via serial dilution and dilute-to-extinction approaches. Cell density, protein concentration, and CO$_2$ evolution were measured along with community diversity (via 16S metagenomic sequencing). These community function measurements show saturating behavior as a function of taxonomic richness. Abundant strains were isolated to disentangle the effects driven by individuals and effects driven by interactions. Interactions increase and saturate with diversity, suggesting both competition and complementation increase simultaneously with diversity.
Additional recent studies have added nuance to the picture of what interactions are prevalent in carbon-degrading communities~\cite{Kehe:2019in, Kehe:2020ju}, finding a high prevalence of parasitic and mutualistic interactions within communities of isolates that are consistent with cross-feeding. These studies leverage a high-throughput droplet microfluidics platform to perform combinatorial community assembly and culture in multiple different carbon source conditions, and the growth of community constituents was measured using fluorescent labelling. In the first study~\cite{Kehe:2019in}, a high prevalence of competitive interactions was observed, particularly for communities and carbon sources where the constituent strains all showed strong growth on the carbon source in monoculture. However, positive interactions were frequently observed in cases where one strain grew poorly on a carbon source in monoculture, consistent with that strain growing in a community due to cross-feeding of metabolic intermediates. The second study~\cite{Kehe:2020ju} broadened this observation by focusing on communities comprising strains from two taxonomic orders, Enterobacterales and Pseudomonadales. Again it was observed that positive interactions were common in communities where one of the constituent strains grew poorly by itself on a given carbon source. It is likely that the positive interactions were generated by cross-feeding of overflow metabolism intermediates~\cite{amarnath_stress-induced_2021}. These positive interactions likely arise from the same mechanism discussed above in the enrichment culture experiments of Sanchez and colleagues~\cite{Goldford:2018jf,Estrela:hb}. Altogether, these results indicate that community carbon degradation can be a relatively simple function of the taxonomic structure of a community. Linking these relationships to genomic structure remains to be accomplished.
Another study by our group~\cite{Gowda:2021ve} set out to explicitly identify the genomic attributes of a community that are predictive of community function. Using a statistical ensemble of isolates that perform denitrification, a process of anaerobic respiration involving the reduction of oxidized nitrogen compounds, we first mapped the genotypes of individual isolates to the precisely quantified kinetics of nitrate and nitrite reduction, which were parameterized using a consumer-resource modeling approach. We used a regularized linear regression approach to predict nitrate and nitrite reduction kinetics from the presence and absence of denitrification-pathway genes possessed by each isolate. We then assembled communities of these isolates and determined that resource-competitive interactions are prevalent and predictable from single-strain kinetics via the consumer-resource model. Thus we inferred that the conserved properties of metabolic genes allow the prediction of community-level function. This study shows that synthetic communities comprised of natural isolates, combined with statistical approaches, can yield insights into the mapping from gene content to metabolite dynamics. It remains to be seen if this approach can be applied to more complex communities.
Our work in Gowda \emph{et al.}~\cite{Gowda:2021ve} points to the idea that focusing on the genomics and ecology of specific metabolic processes can be a powerful approach. In particular, this study leads us to believe that denitrification offers a remarkable system for the quantitative interrogation structure and function both using the statistical approach outlined here, but also for detailed physiological studies of specific metabolic processes. The advantages of denitrification include the fact that the taxa that perform the process are easily isolated and grow well in the laboratory~\cite{lycus_phenotypic_2017}, the metabolites can be quantified in high-throughput~\cite{Gowda:2021ve}, the molecular genetics of denitrification are well-understood~\cite{zumft_cell_1997} and the process has been characterized in the wild to some extent~\cite{tiedje_denitrification_1983}. These opportunities have spawned several compelling recent studies~\cite{lilja_segregating_2016,lilja_substrate_2019,goldschmidt_metabolite_2018}, including a study of the role of carbon source identity in driving denitrification in wild communities~\cite{carlson_selective_2020}.
\section*{Experimental advances and opportunities to enable the ensemble approach}
\label{S:4}
Having reviewed a variety of model communities for undertaking an ensemle approach to the structure function problem we now turn our attention to experimental methods to study these systems. We focus here on culturing and isolation methods as well as analytical techniques for measuring metabolites. We do not review sequencing approaches that have been discussed in detail elsewhere~\cite{hugerth_analysing_2017}.
\subsection*{Higher throughput/targeted isolation techniques}
A vast majority of microbes that are known to exist in nature remain uncultured in the laboratory. The staggering difference between the number of cells counted from microscopy and those obtained on agar plates was discovered in the early 19th century \cite{amannDirekteZahlungWasserbakterien1911} and was dubbed as the ‘the great plate count anomaly’ in 1985 by Staley and
Konopka \cite{staleyMeasurementSituActivities1985}. Developments in sequencing techniques have further widened the gap between the number of cultured and uncultured bacteria. The causes for microbial uncultivability include requirement of growth factors present in the natural environment, slow growth, need for interspecies interactions and transitions to dormancy. Development of isolation techniques that overcome these drawbacks is important to capture the high microbial diversity that exists in the wild. Such techniques will aid in the construction of synthetic communities with high genotypic and phenotypic diversity and hence benefit the study of structure-function problem
Recently developed isolation techniques that offer some advantages over the conventional approach of plating on agar include culturomics, microdroplets and diffusion chambers. In culturomics, communities are tested for growth in a multitude of media conditions using high-throughput techniques, followed by subjecting the communities to mass spectrometry and sequencing~\cite{lagierMicrobialCulturomicsParadigm2012,seng_ongoing_2009}.
By performing MALDI-TOF (see below) mass spectrometry directly on colonies, bacterial taxa can be identified with high fidelity~\cite{seng_ongoing_2009}. In case MALDI-TOF fails to identify taxa, 16S sequencing is used. The use of mass spectroscopy combined with sequencing facilitates accurate, rapid, and comprehensive strain identification. In a recent study~\cite{lagierCulturePreviouslyUncultured2016}, the culturomics approach was shown to be very successful in increasing the number of species isolated from the human gut (by \char`\~two fold).
A higher-throughput method involves encapsulating cells from natural communities in gel microdroplets (GMDs) made of agar. The GMDs are then incubated in growth chambers flushed with low nutrient media. Following this, GMDs with microcolonies are sorted using flow cytometry and individual GMDs are subsequently transferred into microtiter plate wells containing rich organic medium for biomass enrichment and isolation.\cite{zenglerCultivatingUncultured2002,zenglerHighThroughputCultivation2005} In this technique, the porous nature of the GMDs facilitates exchange of metabolites between droplets during the incubation. As a result, strains that require metabolites produced from resource-mediated interspecies interactions can be isolated using this technique.
Diffusion chambers work by culturing communities in chambers exposed to their native environments through porous membranes~\cite{kaeberleinIsolatingUncultivableMicroorganisms2002, bollmannIncubationEnvironmentalSamples2007,chaudharyDevelopmentNovelCultivation2019}. Thus, the setup allows for the growth of microbes that require growth factors present in their natural environments and/or produced from native community interactions. Significant developments improving the throughput of this isolation method include the isolation chip~\cite{berdySituCultivationPreviously2017a} and the Hollow-Fiber Membrane Chamber (HMFC)~\cite{aoiHollowFiberMembraneChamber2009}.
In addition to these non-targeted isolation techniques, targeted isolation techniques have been recently attempted. These involve designing the isolation methods to target desired phenotypes (e.g., antibiotic resistance or sporulation~\cite{browneCulturingUnculturableHuman2016}). One recent study successfully isolated desired cell types using fluorescently labeled antibodies against predicted cell surface proteins combined with flow cytometry for cell sorting~\cite{crossTargetedIsolationCultivation2019}.
Overall, both the available targeted and stochastic isolation techniques have proven to be useful for isolating previously unculturable bacteria. Hence, these techniques may prove valuable for generating ensembles of bottom-up assembled microbial communities.
\subsection*{High throughput experimental platforms}
Our proposed ensemble approach for studying structure-function relationship in microbial communities requires creation of many replicate communities. Hence, high throughput culturing platforms are critical for its implementation.
A majority of high throughput experimental platforms so far have been droplet-based or microfluidic-based devices. One such recently developed device is 'kChip', a microfluidic platform that facilitates combinatorial construction of microbial communities~\cite{blaineyMassivelyParallelOnchip2018}. A study involving synthetically constructed microbial communities on kChips successfully identified sets of strains among 19 soil isolates that promotes growth of model plant symbiont \emph{Herbaspirillum frisingense}, by screening $\sim$\num{100000} multispecies communities~\cite{keheMassivelyParallelScreening2019}. Though kChip is a high throughput platform, it only enables bottom-up construction of microbial communities that requires isolation of microbes prior to the experiments. Additionally, only metabolic functions with optical readouts can be assayed, as physical access to the microdroplets at this scale is not feasible.
Another similar microfluidic platform that enables parallel co-cultivation of microbial communities was developed by Park \emph{et al.}~\cite{parkMicrodropletEnabledHighlyParallel2011}. Their platform was able to successfully detect pairwise symbiotic interactions in communities when the symbionts were in as low an abundance as 3 percent of the total population. Here again, only optically detectable metabolic properties can be measured, but the device enables top-down construction of microbial consortia through random compartmentalisation of community members. Though this was not a structure-function study per se, inclusion of automated droplet sorting and characterization of communities in the retrieved droplets can easily enable structure-function studies. In fact, this was achieved in a more recent study by Terekhov \emph{et al.}, where microbes conferring antibiotic resistance in the oral microbiota of Siberian bears were identified~\cite{terekhovUltrahighthroughputFunctionalProfiling2018}. This was done by functional profiling of the encapsulated communities from the oral microbiome that suppressed the growth of the pathogen \emph{Staphylococcus aureus}.
From the aforementioned studies, it can be said that the choice of the experimental platform largely depends on the nature of the study. Some existing platforms support a top-down approach whereas others support a bottom-up approach. Further, the methods of determining structure and function can differ across platforms. There is room to improve these methods to incorporate other analytical techniques for measuring metabolites. For example, if large scale culturing platforms could be combined with spectroscopic or automated mass spectrometry methods, this would enable the rapid construction of large quantitative datasets.
\subsection*{Measuring metabolite dynamics}
Metabolites, unlike nucleic acids, require distinct analytical techniques depending on the metabolite of interest. Here we review the available methods, their applicability and opportunities for improving these methods for microbial consortia. See Table~\ref{MethodsTable} for the specific strengths and limitations of each technique discussed here.
\begin{table}[h!]
\centering
\begin{tabular}{||c c c c c||}
\hline
Method & Sensitivity & Specificity & \shortstack{Range of \\ applicability} & Throughput \\ [0.5ex]
\hline\hline
NMR & low & high & high & low \\
Mass Spectrometry & high & high & high & low/moderate \\
Infrared/Raman & moderate & high & high & high \\
UV/Vis & moderate & low & low & high \\
Targeted assays & moderate & high & low & high \\
\hline
\end{tabular}
\caption{Comparison of methods for measuring metabolites in microbial communities. Sensitivity refers to the minimum detectable concentration. Specificity refers to the ability of the assay to detect a specific metabolite. Range of applicability refers to the diversity of metabolites that can be detected with the technique. Throughout is the number of measurements that can be made in parallel.\label{MethodsTable}}
\end{table}
\subsection*{Nuclear magnetic resonance spectroscopy}
A number of high quality textbooks describe the fundamental physics~\cite{slichter_principles_1990} and chemistry~\cite{levitt_spin_2008} of nuclear magnetic resonsance (NMR) spectroscopy. Here we give an intuitive explanation of the basis of this technique and go on to the practical applications to measuring metabolites in microbial communities.
NMR spectroscopy exploits the spin magnetic moment of atomic nuclei such as hydrogen, carbon, and nitrogen to characterize chemical structure. In an applied magnetic field, nuclei behave as weak magnets, collectively aligning with the field. The collective alignment of the nuclear magnetic moments can then be manipulated with applied electromagnetic fields in the radio frequency (MHz) and detected as emitted fields in the same spectral region. The small magnetic moments of nuclei cause them to emit very weak radiation meaning that relatively high concentrations of metabolites of interest are necessary for detection. Nuclei experience minuscule changes in the local magnetic field due to their local chemical context, resulting in what are termed ``chemical shifts.'' For example, a proton in hydrogen on an alkane (e.g., methane) will emit a distinct radio frequency (its resonance frequency) from one in an aromatic hydrocarbon (e.g., benzene). These small changes in emitted radio frequency fields are of the order of parts per million (ppm). Typical resolution of modern instruments are a fraction of a ppm and depend on the field strength of the spectrometer and technical details of the detection scheme. State-of-the-art spectrometers (operating at \num{600}MHz and above) are widely available at core facilities.
For metabolomics the two most common types of NMR are proton ($^1$H) and carbon NMR, each of which has both advantages and disadvantages. First, the advantages of proton NMR are (1) rapid acquisition due to the relatively high signal-to-noise ratio in proton spectra, (2) the fact that no isotopic labeling is necessary, and (3) the ability of the technique to detect a broad range of relevant organic compounds. One disadvantage is the fact that spectra from mixtures of unknown metabolites are complex, often containing hundreds of peaks corresponding to the many different compounds present. As a result, it can be challenging to detect the presence or absence of specific metabolites via proton NMR. Further, water contributes a broad and strong solvent signal in the middle of the relevant range of chemical shifts. There are two main routes to removing this signal: (1) drying the sample and replacing the water with D$_2$O and (2) using clever pulse sequences to decouple the water signal from the metabolite signals~\cite{mckay_how_2011}. The former requires specialized equipment and increases the cost and reduces throughput. Therefore, it is recommended to use decoupling. The fundamental physics of how this decoupling works is beyond the scope of this review, but it is recommended to use 1-D NOESY (Nuclear Overhauser Effect Spectroscopy) to isolate metabolite signals from water~\cite{emwas_nmr_2019}. The approach is robust, widely applied, and requires no sample processing to be done. It should be noted that because of variation in technical specifications between instruments, performing measurements on a single spectrometer across samples is key to maintaining reproducibility of measurements~\cite{mckay_how_2011}.
Carbon NMR in contrast, detects signals from the magnetic moments of carbon nuclei. As with protons, the chemical context of the carbon nucleus gives rise to chemical shifts in the resonance and this permits the disambiguation of carbon nuclei in different compounds. One advantage of Carbon NMR is that it does not require suppression of water signals. The main drawback to carbon NMR is sensitivity. The dominant isotope of carbon ($^{12}$C, \SI{99}{\percent} prevalence) is not NMR active, while $^{13}$C is NMR active, but present at about \SI{1}{\percent} natural abundance. This means that most nuclei do not contribute to the observed signal. Second, $^{13}$C has a magnetic moment that is roughly \num{4}-fold lower than $^1$H reducing the signal-to-noise ratio. These two considerations imply that acquiring carbon spectra requires extensive averaging and can take hours for a single sample. However, the low isotopic abundance of $^{13}$C can be overcome by using $^{13}$C labeled compounds as nutrients, with order of magnitude increases in signal-to-noise. Unfortunately, these compounds are expensive (hundreds of dollars per gram) increasing costs.
Carbon and proton NMR are the two most commonly applied metabolomic profiling techniques for microbial communities. Typical studies range from targeted detection of a single metabolite in a well-defined community~\cite{andrade-dominguez_eco-evolutionary_2014} to untargeted profiling in very complex consortia such as anaerobic digesters\cite{gonzalez-gil_nmr_2015}. Here we present a few examples of NMR based measurements of metabolite dynamics in communities as case-studies that might be more broadly applicable.
One compelling approach taken by Date \emph{et al.}~\cite{date_new_2010} and Nakanishi \emph{et al.}~\cite{nakanishi_dynamic_2011} is to combine NMR based metabolite measurements in time with quantification of abundance dynamics. Date \emph{et al.} use $^{13}$C labeled glucose to initiate growth in fecal microbiota~\cite{date_new_2010} (note that $^{13}$C is a stable isotope). The authors then performed time series of abundance dynamics and carbon NMR measurements. Given labeled glucose as the sole carbon source, the authors could track the dynamic production and consumption of $^{13}$C labeled compounds as the glucose was converted to organic acids in time. The authors could then (crudely, given the electrophoresis methods at the time of the study) classify the community into primary and secondary degraders. The compelling aspect of this study is the potential to statistically correlate large scale variation in the community structure with metabolite dynamics. One could imagine a similar experiment with amplicon sequencing-based abundance dynamics measurements. This approach would be especially powerful for looking at the community level response to carbon fixation by autotrophs in systems like mats or CES. In these situations initiating a community with $^{13}$C labeled bicarbonate as the sole carbon source would allow the direct measurement of carbon flux from autotrophs to associated heterotroph communities. Since NMR relies on magnetic fields and emitted RF signals, it is non-invasive and could be applied to ongoing experiments (e.g. CES) without invasive sampling.
We conclude with a brief note on throughput. NMR spectrometers are large, expensive machines that rely on superconducting magnets to apply large magnetic fields. Running parallel experiments on NMR machines is therefore prohibitive. Throughput is achieved by using robotics or fluidic systems to automatically load samples into the spectrometer. Such experiments typically achieve a throughput of order \num{100} samples per day~\cite{macnaughtan_high-throughput_2003,soininen_high-throughput_2009}. Increasing NMR throughput by a factor of 10-100 would constitute a major advance.
Finally, despite the current limitations, there is a revolution underway in quantum sensing systems from superconducting quantum interference devices (SQUIDs)~\cite{mcdermott_liquid-state_2002} to spin-based magentic field sensors, impurities in diamonds (nitrogen-vacancy (NV) centers)~\cite{cujia_tracking_2019} or force measurements\cite{kuehn_advances_2008}. SQUIDs enable ultra-low field NMR, obviating the need for large expensive magnets, and NV-centers enable high sensitivity magnetic resonance detection at the single-molecule level. The applications of these technologies to metabolic function in microbial communities await future discovery, but one can imagine massively parallel NMR measurements or \emph{in situ} detection of metabolites in complex settings.
\subsection*{Mass spectrometry}
Mass spectrometry is the most widely used platform for metabolomics and several good reviews of the methodology are available~\cite{beale_review_2018,mastrangelo_sample_2015,dettmer_mass_2007,raftery_mass_2014,alseekh_mass_2021,Jemal2000}. Therefore, our discussion of mass spectrometry will be limited, but we include it as an important point of comparison with the other techniques discussed in this section.
Mass spectrometry ionizes the molecules in a sample, using a variety of different methods, and then accelerates the charged molecules using an electric field. The beam of ions is then passed through a magnetic field that (via the Lorentz force law) results in a force on each ion that depends on its mass to charge ratio. The result is a physical separation of ions in space in proportion to the mass-to-charge ratio. A huge number of variations exist on this basic theme, including measurements of time of flight (TOF) and quadrupole mass filters that apply oscillating fields to the ion beam. The reviews cited above contain detailed discussions of the type of ionization and detection methods that are best suited for metabolomic applications.
In the context of metabolomics, where samples are often of considerable chemical complexity, mass spectrometry is almost always preceded by either gas or liquid chromatography to separate compounds and therefore increase the specificity and sensitivity of downstream mass spectrometry. As a result, gas chromatography-mass spectrometry (GC-MS) and liquid chromatography-mass spectrometry (LC-MS) are the two most widely used forms of mass spectrometry for metabolomics. GC-MS is restricted to volatile compounds typically of molecular weight less than 600 Da~\cite{beale_review_2018}, while LC-MS applies to a broader spectrum of metabolites.
Mass spectrometry has significantly higher sensitivity that NMR, and can achieve single molecule sensitivity~\cite{robertson_single-molecule_2007}, although this is not routine. This advantage is especially crucial for measuring metabolite fluxes in cells and communities where concentrations are often low (micromolar and below)~\cite{bennett_absolute_2009,basan_universal_2020}. This combination of sensitivity and the ability to distinguish broad classes of compounds has contributed to the widespread usage of MS-based metabolomics methods~\cite{aiyar_antagonistic_2017}.
The throughput of these techniques is significantly lower than the optical methods discussed below and comparable, at present, to NMR. So it is routine to run $\sim$\num{100} samples over the course of a day or two (see~\cite{Jemal2000} or a review). Some robotic systems have been developed to automate the sampling and analysis process~\cite{molstad_robotized_2007}. However, making mass spectrometry measurements on thousands or tens of thousands of samples, while feasible, is costly and slow. Given the remarkable sensitivity, specificity and broad applicability of the technique especially for untargeted measurements of metabolite pools, it would represent a major advance if mass spectrometry could be routinely applied to thousands of samples in parallel. Here, we mention a handful of notable studies that leverage mass spectrometry to quantify metabolite dynamics in microbial communities. The goal is not to present an exhaustive list but simply to point the reader towards some representative studies.
Amarnath \emph{et al.}~\cite{amarnath_stress-induced_2021} used untargeted metabolomics to study the metabolites that are exchanged between two strains of bacteria in a serial dilution experiment. The authors revealed a broad spectrum of metabolites excreted by one strain in response to stress. These excreted compounds facilitated cross-feeding between the two strains. Shi \emph{et al.}~\cite{ShiLCMS} use LC-MS to study metabolite exchange in a fungal-bacterial community. A statistical analysis of the LC-MS data shows that the metabolites excreted are distinct for co-cultures and mono-cultures.
Mass spectrometry is widely used to study the degradation of compounds from pharmaceuticals to soil contaminants~\cite{PIEPER20104559,THELUSMOND20161241}. For example, a common environmental contaminant are polycyclic aromatic hydrocarbons (PAHs), which are routinely degraded by bacterial consortia. The breakdown of these compounds in time is typically interrogated by GC- or LC-MS~\cite{luan_study_2006}. %
Mass spectrometry is widely applied to food-related microbial communities. In these cases the untargeted nature of mass-spectrometry is important as the compounds of interest (e.g., for flavor) are typically unknown. For example, GC-MS has been used to identify starting components in traditional Cambodian rice wine \cite{ricewineGCMS}. Similarly, it was used in fermentation of red peppers to investigate changes in bacterial and fungal communities and volatile flavor compounds \cite{XUredpepper}, and high throughput GCMS has been used to correlate metabolites with taxonomic structure in kimchi fermentation \cite{PARK2019558}, glutinous rice wine \cite{HUANG2019593}, the liquor \textit{Daqu} \cite{JIN2019422} and pickled radishes \cite{RAO2020108804}.
\subsection*{Infrared spectroscopy}
Infrared spectroscopy detects absorption and emission of photons in the infrared region of the electromagnetic spectrum and characterizes molecular vibrations. As with NMR, the resonance frequency of a molecular vibration depends strongly on molecular structure. This dependence affords IR spectroscopy its chemical specificity. Infrared spectroscopy offers perhaps the best combination of sensitivity, specificity, high-dimensional characterization of complex metabolite pools and throughput. There are two commonly used methods for measuring infrared spectra that differ in their fundamental physical mechanisms: (1) IR absorption and (2) Raman spectroscopy.
Infrared absorption involves passing light in the infrared range ($\sim$\SI{2.5}{\micro\meter} to \SI{10}{\micro\meter} wavelengths or \SI{1000}{\per\centi\meter} to \SI{4000}{\per\centi\meter} wavenumbers) through an aqueous or gas phase sample and measuring the absorption. Compounds containing different chemical bonds absorb light at different frequencies and the resulting spectrum can provide extensive information on the chemical composition of the sample. As with proton NMR, a major downside of absorption spectroscopy is the broad absorption of water in the informative region of the spectrum (around \SI{3200}{\per\centi\meter} and \SI{1600}{\per\centi\meter}), which can limit the information for aqueous samples without cumbersome drying. Simple dispersive spectrometers that shine a narrow band of wavelengths through a sample have limits on sensitivity, spectral resolution and the duration of acquisition. These limitations can be overcome using Fourier Transform infrared spectroscopy (FTIR), which uses broadband excitation and an interferometer to rapidly acquire spectra in specific spectral bands, and this is the most commonly applied technique. Plate readers that perform FTIR measurements en masse on microtiter plates are available and can acquire data from both liquid and solid phase samples.
In contrast to FTIR or dispersive IR spectroscopy, Raman spectroscopy measures molecular vibrations using photons in the visible portion of the spectrum. When visible photons interact with a sample, most are scattered with the same energy as the incident photon (Reyleigh scattering). However, with low probability, incident light undergoes inelastic scattering and in the process photons are emitted from the sample with either slightly lower (Stokes) or higher (anti-Stokes) energy than the incident radiation. These small changes in the emitted photon wavelength correspond to the molecular vibrations in the sample. For reasons beyond the scope of this review, Raman spectroscopy does not suffer from broad band absorption from water, making it especially attractive for microbial communities in the aqueous phase. However, due to the inefficiency of inelastic photon scattering, Raman spectroscopy requires high laser power and the resulting heating can be a problem for biological samples, a limitation that can be overcome by techniques such as resonance Raman spectroscopy or surface enhanced Raman scattering. However, these methods are not yet routine for metabolomic profiling. Raman spectroscopy can be performed on bulk samples using plate readers, or integrated with a microscope for localized measurements. More recently, Raman spectra can be acquired via flow cytometry~\cite{suzuki_label-free_2019,nitta_raman_2020}. These platforms enable much higher throughput than is now standard by NMR or mass spectrometry.
FTIR and Raman spectroscopy have proven to be powerful methods for interrogating cellular physiology at the single-cell level when combined with microscopy (see ~\cite{hatzenpichler_next-generation_2020} for a recent review). Remarkably, Raman spectroscopy signals can be used as fingerprints to demarcate cells of one species in different growth states~\cite{escoriza_studying_2006}, or different taxa at the strain, species and genus levels~\cite{rosch_chemotaxonomic_2005,harz_micro-raman_2005}. A recent study showed that the global transcriptional profile of yeast and bacteria could be predicted via a linear model from single cell Raman spectra\cite{kobayashi-kirschvink_linear_2018}. This success owes to the high-dimensional nature of Raman spectra, which are often challenging to interpret in terms of individual peaks but are rich in information that can be decoded statistically.
Despite the power of infrared spectroscopies for chemical characterization, they have seen comparatively little use in the context of communities of microbes. We regard this as a missed opportunity, and suggest that these methods could and should be used more broadly. One of the limitations of the technique is the challenge of assigning specific peaks to specific compounds. As Kobayashi and co-workers have shown, this limitation can be overcome by using simple statistical methods to map infrared spectra to other cellular properties \cite{kobayashi-kirschvink_linear_2018}. The approach is to measure Raman or IR spectra on a set of samples and then use a lower throughput technique such as LC-MS to measure the absolute concentration of a metabolite of interest. A combination of dimensionality reduction and regression can then be used to map LC-MS data to infrared spectra. This approach has been used to track substrate concentrations in time in monocultures~\cite{paul_towards_2016} and phenol degredation in complex communities~\cite{wharfe_fourier_2010}.
The advantages of Raman spectroscopy, and to a lesser extent FTIR, over mass spectrometry and NMR are the potential for the rapid acquisition of high-dimensional characterization of metabolite pools. The complexity of the resulting spectra is similar to proton NMR, and therefore is perhaps most useful for statistically characterizing differences between community metabolite profiles. Such complex spectra can then be used either to measure specific metabolites via a calibration approach discussed in the previous paragraph, or to demarcate global metabolic states of consortia without concern for specific metabolite levels.
\subsection*{UV-visible spectroscopy}
We briefly note that simple ultraviolet and visible spectroscopy (UV-Vis) can be used to characterize electronic transitions in compounds of interest for metabolic characterization. These methods can be performed with widely available plate readers, particularly those that are equipped with monochromators rather than filters, which permit excitation and emission to be arbitrarily selected by the user. The main limitation of this technique is the fact that electronic transitions in the visible are restricted primarily to chemical species with delocalized electron density (e.g., conjugated rings such as benzene, tryptophan). As a result, these spectra are low specificity and cannot be used for targeted metabolomics. However, the high throughput of common plate readers facilitates rapid measurements, and the spectra can give coarse characterization of excreted compounds from autotrophs, for example~\cite{tenorio_impact_2017}. Moreover, targeted UV-Vis measurements can be integrated with common fluorescence microscopes and therefore offer the possibility of detecting metabolites in massively parallelized platforms such as droplet microfluidics~\cite{Kehe:2019in,Kehe:2020ju} or in spatially structured communities.
\subsection*{Targeted assays}
In situations where the metabolic function of the community under study is known apriori and restricted to a specific chemical compound a targeted metabolite assay can be powerful. Such assays typically utilize chemical reactions to create an optically active compound in proportion to the concentration of a metabolite of interest. For example, starch can be degraded to glucose enzymatically and then glucose concentration can be assayed via standard methods~\cite{holm_rapid_1986} or iodine can be used to stain starch directly and detect degredation~\cite{fuwa_new_1954}. Similarly, nitrate and nitrite can be detected via the Griess assay~\cite{miranda_rapid_2001}, which utilizes a colorometric reporter (dye) generated via a reaction with nitrite. Such assays can easily be performed in 96-well plates facilitating high throughput \cite{Gowda:2021ve}. These measurements can be powerful for studying specific metabolic processes in communities. However, typically the chemistry involved is not easily automated nor are the conditions of the reaction biocompatible. So, such measurements are made offline after sampling and are challenging to automate \emph{in situ}.
\section*{Quantifying structure}
We briefly outline the main ways in which community structure can be quantified. As mentioned above a suite of next generation sequencing technologies are capably of quantifying community structure on multiple levels. For example, amplicon sequencing of the 16S rRNA gene uses PCR to amplify this universally conserved ribosomal subunit and then uses the number of reads mapping to sequence variants (amplicon sequence variants, ASVs) as a proxy for the relative abundance of each taxon. This widely used method has many well documented downsides including variation in the copy number of the 16S gene across taxa, PCR bias and the challenge of associating taxonomy with metabolic capabilities of each strain \cite{Callahan2016}. Despite these shortcomings, amplicon sequencing does permit rapid and high throughput characterization of the community composition, and methods exist for inferring metabolic capabilites of strains from 16S gene sequence alone~\cite{douglas_picrust2_2020}.
In contrast, shotgun metagenomic sequencing amplifies all genomic DNA in a sample. These data enable the characterization of the gene content of an entire community by annotating reads. Metagenomics gives a much more complete picture of the genomic structure of a community but several technical hurdles limit this approach. First, annotating reads mapping to genes remains a challenge and roughly \num{30}-\num{50}\% of the open reading frames are annotated, leaving much of the genomic content unclassified in terms of function. Second, assigning annotated genes to specific taxa within the community and inferring their relative abundances remains a hard problem. In particular, assembling reads into genomes (metagenome-assembled genomes, MAGs) has been performed but the quality of these assemblies remains hard to assess. Applying cutting edge machine learning methods is likely to improve this process~\cite{nissen_improved_2021}. Another method to analyze shotgun metagenomics data is to use a database of reference genomes as templates to recruit reads from a metagenome. The upside of this approach is the ability to reliably detect variation at the level of single nucleotides~\cite{garud_evolutionary_2019}, and reliably assemble genomes from metagenomes. The cost of course is that the method misses any diversity that is not present in the reference genome database. Despite these challenges, metagenomics perhaps gives clearest picture of the genomic structure of a community as a whole.
Transcriptional profiling of entire communities is also feasible via RNA-sequencing based methods~\cite{antunes_microbial_2016,zhang_metatranscriptomics_2021}. Despite the potential predictive power of knowing which genes in a community are transcriptionally active, these measurements have been applied much less widely than taxonomic amplicon or shotgun metagenomic methods. However, as the costs of sequencing continue to fall, it remains a compelling proposition to use metagenomics and transcriptional profiling on the same samples. We propose that such measurements could very well lead to deeper insights into the community structure and function by potentially simplifying the picture. For example, transcriptional profiling could reveal which collective components of the metagenome are inactive and therefore could potentially be left out of a predictive framework.
\section*{Learning from data: function from structure}
\label{S:5}
Equipped with measurements of metabolites (either dynamically or at a fixed point in time) and some characterization of the community structure, we are then left to ask what to do with these data. In reality, the answer to this question is an empirical one that depends on the structure the data, the model system under study and the precise question being posed. We offer no pipeline or prescription for how best to proceed, but instead offer a few suggestions and examples and point out some important technical pitfalls. Our intention is to suggest some approaches to learning the structure-function mapping from these data and to leverage the results for predicting community metabolism. As with most data analysis tasks, simple is better. Using the latest methods in machine learning or dimension reduction may be tempting but it is almost always better to explore the data with methods that are simple to interpret and straightforward to implement.
Typically any sequencing based characterization of community structure will be high dimensional. For example, 16S amplicon sequencing will often yield \num{10} to \num{1000} of taxa per sample, similarly metagenomic data can contain \num{10000} or more annotated open reading frames depending on the complexity of the community. In contrast, assembling an ensemble of more than \num{100} or \num{1000} communities is a huge challenge even for the highest throughput methodologies. As a result, we are almost always in the limit of a small number of data points (communities) and large number of variables (taxa, genes, transcripts etc). In this regard, predicting functional properties from these structural data requires reducing the dimensionality of the data describing community structure. For the purposes of the discussion below we define the number of features (genes, taxa, transcripts) as $p$ and the number of communities in a given ensemble $n$ (Fig.~\ref{FigData}A). A sequencing data set can then be described by a matrix $X$ that is $n\times p$ with $n<<p$ in most cases. The rows of this matrix, $\vec{x}_i$ correspond to sequencing data for the $i$th community in the dataset. The entries of this vector are then the number of reads mapping to a specific ASV in a 16S data set or gene in the metagenome of that community. For each of these communities we assume that the data set includes some functional measurement, $y_i$, which may be a dynamic quantity ($y_i(t)$). The goal then is to learn a representation of this functional measurement of the community in terms of the columns of $X$.
\subsection*{Compositional data and zero counts}
All of the standard sequencing methods for quantifying community structure result in compositional data - that is they do not report the absolute abundances of taxa, genes or transcripts in the sample, but only the relative contribution (e.g., the rows of $X$ are defined only up to an unknown constant). Recently, methods using qPCR or the addition of oligos at known concentrations have been developed to measure absolute abundances via sequencing. However, as yet these methods are not widespread. Therefore, in any analysis we must contend with the compositional nature of sequencing data. Much has been written about this problem~\cite{gloor_microbiome_2017}, and members of the field are now generally aware of the issues that can arise when the compositional nature of the data are ignored.
Briefly, compositional data can, and should, be log-ratio transformed using log ratios of counts. Log-ratio transformations take compositional data from a simplex and map them to real numbers with the properties of a vector space. This transformation therefore permits the application of conventional statistical approaches to compositional data. Typically, this is done via the center-log transform (CLR) or an additive log-transform (ALR)~\cite{Aitchison2000,gloor_microbiome_2017}. Computing log-ratio is not compatible with zeros (e.g., zero abundances of an ASV or gene transcript), a problem that has received a significant amount of attention. A host of methods from adding psuedocounts uniformly to all zeros or using Bayesian approaches to replacing zeros ~\cite{Aitchison1982,love_moderated_2014,mcmurdie_waste_2014}. We urge caution here, as many methods are both ad hoc and can qualitatively impact the results of downstream analyses. We will not review the technical details, but readers should to engage carefully with their data rather than blindly applying existing pipelines for log-transforms and handling zeros. For example, variance decompositions applies to log-transformed data can be dominated by large numbers of taxa or transcripts with zero counts. In this scenario the details of how zeros are handled (e.g., the magnitude of the psuedocounts added to all taxa) can have huge impacts on the variance decomposition.
\begin{figure}
\centering\includegraphics[scale=0.8]{Fig5_Data.png}
\caption{\textbf{Learning the structure-function map from data.} (A) Hypothetical structure-function data from an ensemble of $n$ communities. $y$ denotes a metabolite measurement, either level or rate that could also be dynamic. Colored dots correspond to data points in (B,C). $X$ denotes a matrix of $n$ rows each denoting a single community in an ensemble. The columns denote the relative abundances (colored bars) of taxa, genes in the metagenome or transcripts in the metatranscriptome. (B) An unsupervised approach where dimensionality reduction is applied to $X$ yielding a lower dimensional representation of community structure that is then associated with communities of differing function. (C) A supervised approach where the function $f(\vec{x})$ is learned for mapping structural variation to functional variation. Regressors denote independent variables in a lower dimensional representation of $X$ that provides good predictive power of $y$.}
\label{FigData}
\end{figure}
\subsection*{Dimension reduction: with or without supervision}
The goal is to associate structural components or sets of components with the metabolic function of a community. As discussed above we are almost always in the limit of small numbers of data points. We are therefore forced to consider reducing the dimensionality of the data from $p$ by some form of dimension reduction. Above we suggested that low-dimensionality is a common feature of biological systems from proteins to higher organisms and behavior. Despite the fact that this observation seems to hold quite broadly it is important that we not take low-dimensionality as given in any analysis of a microbial community. In this sense, we must make a principled search for a simpler description of community structure while judiciously considering the possibility that no such description exists or that we are considering the system at the wrong level of organization. For example, in an ensemble of communities with strong functional redundancy, where very distinct taxa perform similar metabolic functions, it may not be possible to find a low-dimensional description of the ensemble in terms of taxa present across replicates in the ensemble.
There are two ways to go about approaching this problem: supervised and unsupervised dimension reduction (Fig.~\ref{FigData}B-C). We note that for most of the techniques described below the statistical learning textbook from Hastie, Tibshirani and Friedman~\cite{hastie_elements_2016} provides an excellent and readable reference. For a more recent review article we recommend the paper of Mehta \emph{et al.}~\cite{mehta_high-bias_2019} that covers both supervised and unsupervised methods.
\textbf{Unsupervised learning:} In unsupervised dimension reduction we seek a lower dimensional representation of the matrix $X$ without any explicit consideration given to $y$ in selecting low-dimensional features (Fig.~\ref{FigData}B). The idea is to reduce the dimensionality of $X$ and then examine the relationship between this lower dimensional representation of community structure and function, $\vec{y}$. The hope is that some lower dimensional representation of the community captures all of the useful information in the matrix $X$. All of these methods change the basis of $X$ in order to recast community structure in terms of groups of genes or taxa. The dimension reduction arises when a small number of such groups of taxa provide a good description of all communities in the data set. If this is case each community $\vec{x}$ can be represented not in terms of the abundances of $p$ taxa, but instead as a combination of $m<<p$ groups or combinations of taxa. Formally, these methods are based on a matrix decomposition ($X = AB$, with $A$ an $n \times m$ matrix and $B$ an $m \times p$ matrix) subjected to constraints. Dimension reduction comes about if $X$ can be well approximated for $m<<p$. In this case, each community in an ensemble can be represented in terms of just $m$ features. These low-dimensional representations, when they exist, can sometimes dramatically simplify our understanding of complex systems.
The canonical unsupervised methods are variance-based decompositions like principle components analysis (PCA) or the more general version, singular-values decomposition (SVD). These methods find the set of orthogonal directions in the $p$-dimensional feature space that maximize the variance in the data along each direction (eigenvectors of the covariance matrix of $X$). The advantage here is the clean interpretability of these decompositions. Each direction (principle component) is a simple $p$-dimensional vector representing a direction of high data variance. Each data point (community) in the original data set can be represented in the basis of principle components $\vec{x}_i \cdot \vec{p}_1,...,\vec{x}_i \cdot \vec{p}_j,... $ (with $\vec{p}_j$ the principle components). If a small number of principle components capture a substantial fraction of the data variance then $m$ is small and each community can be represented as projections on a handful of $\vec{p}_j$ components.
Many modifications to this basic idea exist. Perhaps the two most notable are independent components analysis (ICA)and non-negative matrix factorization (NMF). ICA finds a lower dimensional representation of $X$ in terms of components that are statistically independent, rather than simply uncorrelated, through an iterative process. The approach is to represent each community $\vec{x}$ as a sum of statistically independent components~\cite{hyvarinen_fast_1997}. In this case $\vec{x}_i = \sum_j a_{i,j} \vec{s}_j$ where the $a_{i,j}$ are the weights that `mix' the independent components $s$ to form the observed data $\vec{x}$. ICA is most often applied to signal separation problems where multiple independent inputs are combined. One can speculate that this perspective may be useful for communities if they possess a modular organization where independent modules within the community are responsible for functionally distinct processes. NMF is another matrix decomposition method that is applicable whenever the entries of $X$ are constrained to be positive (as is the case in sequencing data). NMF decomposes the data as $X \approx WH$ with all entries of $W$ and $H$ positive~\cite{lee_learning_1999}. In this case $\vec{x}_i = \sum_j w_{i,j}\vec{h}_j$ where the $w$ are weights and the $\vec{h}$ are vectors of length $p$. The advantage of this approach is that the columns of $H$, that act as `eigen-communities' contain all positive entries and are therefore interpretable. In contrast, the eigenvectors of a PCA decomposition can contain negative entries, which is not interpretable in the context of $X$ that contains only positive values (abundances or relative abundances). NMF has seen limited application in the microbiome context~\cite{cai_learning_2017}. We have focused here on well-established methods with simple interpretations. We are aware that over the past two decades many new methods have been developed, especially those that can learn low-dimensional representations of highly nonlinear data (e.g., autoencoders~\cite{kramer_nonlinear_1991} or stochastic neighbor embedding methods~\cite{hinton_stochastic_2002}). These methods may be useful in the context discussed here, but we advocate starting with the simpler approaches discussed above before moving on to these methods.
Regardless of the method of unsupervised learning applied the result is a new representation of community structure ($X$) in a new basis (e.g., principle components, $\vec{p}$, $\vec{s}$, $\vec{h}$). Ideally, $m<<p$ and communities with many hundreds or thousands of genes or taxa can be represented in a much lower dimensional space ((Fig.~\ref{FigData}B). The task then remains to associate this lower dimensional representation with metabolic function ($y_i$). A common approach to this problem is simply to ask which basis vectors correlate with specific metabolic properties of the community. One common approach is to treat the question as a regression problem to predict $y_i$ from the decomposed $\vec{x}_i$ for example by using the projections of each $\vec{x}_i$ along each principle component as independent variables.
\textbf{Supervised learning:} One major shortcoming of the unsupervised approach is that the low-dimensional representation of $X$ that we learn by unsupervised dimension reduction may not be the best representation of the data in terms of predicting $y_i$. In essence, PCA or NMF may find a low dimensional representation of the data, but there is no reason or guarantee that this representation will be informative of the community metabolic function. Indeed, the unsupervised approach artificially separates the process of finding low-dimensional descriptions of the community and predicting the response variable (metabolic function). A supervised approach overcomes this limitations by performing both dimension reduction and prediction at the same time.
In the supervised approach we seek some prediction of the metabolic function in terms of the structure (Fig.~\ref{FigData}C). Concretely, we would like to estimate $y_i = f(\vec{x}_i)$ for our entire data set. This can be posed as a regression problem where we either posit a functional form for $f(\vec{x}_i)$ or, using more flexible but less interpretable methods like neural networks, learn a mapping from $\vec{x}_i$ to $y_i$ without positing a specific functional form.
Before discussing how one can approach this problem we would like to clarify the meaning of $f(\vec{x}_i)$. We are proposing learning a statistical map from $\vec{x}_i$ to $y_i$. We are not proposing fitting an explicit ecological model such as a consumer-resource or Lotka-Volterra model to the data. We regard fitting such a model as a much harder proposition than learning a statistical mapping from structure to function. Indeed, a statistical approach to learning $f$ necessarily abstracts away these dynamics that relate $\vec{x}_i$ to $y_i$. We note that two of us recently took a hybrid approach to the problem with synthetic communities that did explicitly model the ecological dynamics~\cite{Gowda:2021ve}. In this case, we used a regression to map gene content to consumer-resource model parameters and then used the consumer-resource model to predict metabolite dynamics in consortia. Remarkably, the approach worked, but the downside is that it requires isolating individual taxa and constructing synthetic communities - a more laborious task than studying communities directly.
There are many statistical approaches to learning the function $f$. The simplest approaches are linear regression methods that simply posit a model of the form $f(\vec{x}_i) = \beta_0 + \sum_{k=1}^p \beta_k x_{i,k}$, where the $\beta$ are regression coefficients. There are two major problems with this approach. First, if $n<<p$ we have many fewer data points than independent variables, which means that an ordinary least squares regression will almost certainly overfit and yield poor out of sample predictions (as determined by cross validation). The second related issue is that this approach gives equal weight to each entry (gene, taxon) in $\vec{x}$ and does not provide any dimensional reduction. One way to solve this problem is via regularization~\cite{hastie_elements_2016} where the model is optimized with an additional penalty term seeks to reduce the number of non-zero $\beta$ coefficients (see LASSO and Ridge Regression). Regularization provides a solution to the problem of selecting which regressors (entries of $\vec{x}_i$) provide the most predictive power while also avoiding overfitting. In situations where the levels of noise are not too high and a sparse solution (small number of non-zero $\beta_k$) do allow for good predictions, these regularization methods typically succeed~\cite{fraebel_evolution_2020}. However, if the noise levels are high or the underlying process is not sparse, then even these methods will fail. Care must be taken in diagnosing when such a regression works and when it does not, see~\cite{fraebel_evolution_2020,Gowda:2021ve}.
A major shortcoming of the simple formulation outlined above it that it lacks any interaction terms (e.g., $x_{k,i} x_{l,i}$). Adding these terms to the regression above increases the number of independent variables from $p$ to $\sim p + p^2/2$. Given limited data $n$, it is typically not advised to take this approach. However, including such interaction terms is desireable given the utility of considering pairwise interactions in complex systems~\cite{bialek_rediscovering_2007,schneidman_weak_2006}. One way to proceed is to use linear regression approaches that use groups of independent variables as regressors. For example, principle component regression~\cite{hastie_elements_2016} uses principle components as independent variables in a linear regression. This is qualitatively similar to the unsupervised approach outlined above.
The models above are linear, and this aids in their interpretation. However, explicitly non-linear methods for estimating $f$ are also possible. When and why such approaches are more or less appropriate in the microbial context is not yet clear. However, decision tree based methods such as random forests have proven useful for relating taxonomic variation to host phenotypes in the microbiome~\cite{blanton_gut_2016}. Random forests can model complex non-linear relationships between regressors and response variables~\cite{hastie_elements_2016}, while retaining some interpretability by assigning `importance scores' to each independent variable. As a result these methods can be used to assess the impact of a given taxon or gene on the community function (in a statistical, not causal, sense). Random forest regressions fit many decision trees to bootstrapped replicates of the data and average the result. This averaging procedure, often called `bagging', reduces the variance of the prediction and as a result, high variance/low bias regression approaches can provide lower variance predictions.
Finally, as the amount of high quality data on microbial communities increases, the applicability of more recently developed neural network supervised prediction methods~\cite{mehta_high-bias_2019} will become more appropriate. These methods typically contain many millions or even billions of parameters and therefore require reasonably sized data sets to train. Advances in methods like transfer learning~\cite{weiss_survey_2016}, where existing trained networks are trained to solve a new problem given some data, mean that the user need not start from scratch. The challenge will be what we can learn from these networks once they are used to approximate $f$. In many cases neural networks do not generalize well and are susceptible to small amounts of noise on the input variables. We regard the application and interpretation of these network approaches to communities a problem at the forefront. It may be that we need to reconsider how networks are trained to properly learn the salient features of the structure function problem in microbial communities~\cite{blazek_explainable_2021}.
\subsection*{What we do and do not learn from a statistical approach}
What can the ensemble approach coupled with a statistical analysis like the one described above teach us about communities? We contend that by looking at an ensemble of well-chosen communities, and learning the main statistical features of community structure that determine function we can begin to learn what general properties of communities must be present to admit their functional properties. A handful of recent studies have begun to show the power of this approach~\cite{Goldford:2018jf,blanton_gut_2016,raman_sparse_2019,Gowda:2021ve}. What we recover from these studies is what reproducible features of communities are retained across replicate consortia. These reproducible features can be regarded as `good variables' for predicting community function from structure. In some cases, understanding these variables lets us control or predict the functional properties of consortia in synthetic systems~\cite{Gowda:2021ve} and in hosts~\cite{blanton_gut_2016}. Ultimately, we hope this approach can be used to design, predict and control microbial consortia in engineered and wild contexts to address the existential threat of climate change.
However, even when these statistical approaches succeed in predicting community structure from function we often do not understand why, at a mechanistic level, the prediction succeeds. For example, in the case of Gowda \emph{et al.}~\cite{Gowda:2021ve} the reason for the success of the regression from gene content to phenotypes is not entirely clear. Nor do we believe that the regression can predict the impact of gene gain or loss mutations. Similarly, in the case of Blanton \emph{et al.}~\cite{blanton_gut_2016} the precise metabolic role of each bacterial taxon in the stunting of the host is unclear. In essence, the statistical approach lets us find good variables for design and control of communities, but it does not, by itself, tell us why these are good variables.
\section*{Future directions: evolutionary rules of the structure-function mapping}
\label{S:6}
So, while the ensemble approach can help us solve the structure-function problem the deeper question of why nature constructs communities the way it does remains. We argue that the answer to this question will require considering the eco-evolutionary basis of the observed structure function mapping. Addressing this question is subject enough for a separate manuscript. However, we hope that through the careful application of quantitative methods, like those discussed here, to some of the model systems discussed above, we can open the door to understanding how nature constructs dynamic functional consortia.
\bibliographystyle{model1-num-names}
|
\section{Introduction}
\label{sec:introduction}
\IEEEPARstart{M}{ulti-modal} magnetic resonance imaging (MRI), including T1-weighted (T1w), T2-weighted (T2w) and fluid-attenuated inversion recovery (FLAIR) MRI, has been commonly adopted in clinical practice. During scanning, MRI protocols require separate acquisition for different modalities. However, the prolonged scan time could affect the efficiency of clinical diagnosis and research work, and may also lead to motion artifacts in acquired images and discomfort for patients. Accelerated MRI is necessary to shorten the scan duration. Most of the approaches acquire data at sub-Nyquist rates and reconstruct images from under-sampled \textit{k}-space data by exploiting redundancies within the data. Parallel imaging exploits redundancies of multiple receiver coils \cite{pruessmann1999sense,griswold2002generalized}. Compressed sensing MRI (CS-MRI) is based on the compressibility of images and utilizes sparse transformations for regularized reconstruction \cite{lustig2007sparse}. Although parallel imaging and CS-MRI have been widely used, their performance at highly sparse under-sampling factors is sub-optimal. For example, parallel imaging may result in noise amplification, and CS may lead to residual artifacts or blurring \cite{sung2013high,yang2018admm}. Moreover, CS-MRI requires lengthy iterative optimization and empirical fine-tuning of regularization parameters.
With recent development of deep learning, deep convolutional neural networks (CNNs) have been proved to be promising for accelerated MRI, including standalone denoising networks and unrolled deep learning networks \cite{wang2016accelerating, schlemper2017deep, yang2017dagan, lee2017deep, hyun2018deep, seitzer2018adversarial, qin2018convolutional}, both of which outperformed conventional MRI reconstruction techniques regarding both reconstruction speed and quality. Besides optimizing network architectures for improving performance of deep learning based MRI acquisition, efforts have also been made to better exploit additional information for MRI acceleration. Especially, considering that multi-modal MRI is often performed for diagnosis, it would be beneficial to utilize the fully-sampled or under-sampled images acquired of one modality to reconstruct under-sampled images of another modality. For example, some multi-modal reconstruction networks have been proposed to mitigate the under-sampling artifacts for T1w and T2w MR image pairs \cite{dar2020prior,ehrhardt2020multi,ehrhardt2016multicontrast,kim2018improving,weizman2016reference,xiang2018deep,xiang2018ultra, xuan2021multi}, by exploiting the highly-related anatomical information shared between the two modalities. In such methods, network architectures are designed to exploit the complementary cross-modal information, to achieve better reconstruction performance than the single-modal reconstruction. Furthermore, the information of the fully-sampled T1w MR image can be used to reconstruct the under-sampled T2w MR image more accurately. To maximize the assistance power of T1w MR image, the acquired T2w MR image should be complementary to the information that the network can learn from the fully-sampled T1w MR image. However, handcrafted under-sampling patterns are usually adopted in the T2w MR acquisition, and how to under-sample the data to achieve better multi-modal reconstruction performance remains an open question.
Inspired by the fact that the optimized under-sampling patterns tend to outperform typical fixed patterns \cite{zijlstra2016evaluation}, various under-sampling pattern optimization methods have been proposed to improve the performance of a reconstruction model \cite{bahadir2020deep,aggarwal2020jmodl,zhang2019reducing}. Based on the success of deep learning-based multi-modal reconstruction, the under-sampling pattern optimization algorithms for multi-modal MRI acquisition have gained great interest. For example, a fully-connected neural network was designed to accelerate the T2w MRI acquisition through simultaneous optimization of reconstruction and the under-sampling pattern with the assistance of the fully-sampled T1w MR image \cite{liu2021deep}, but the generated patterns are limited to 2D imaging. Due to the complexity of the model, particularly the fully-connected layers, extending this method to optimize the under-sampling pattern for 3D acquisition is not straightforward.
A straightforward extension of the pattern optimization method \cite{bahadir2019learning} from the single-modal image to the multi-modal image would be to stack the fully-sampled image of the reference modality (RM) along with the under-sampled image of the target modality (TM) as inputs of the reconstruction network. However, according to our experimental results, no significant improvement, compared to the single-modal reconstruction, was observed with this simple adaptation, indicating the necessity of developing new strategies to better exploit information of reference modality.
In this study, we propose to improve the target modality MRI reconstruction by optimizing the \textit{k}-space under-sampling pattern of the target modality MR image that can complement the fully-sampled reference modality MR image based on their statistical difference in \textit{k}-space. Particularly, we use a cross-modality translation network to map an MR image from the reference modality to the target modality, and the residual difference between the mapped scan and the actual scan in \textit{k}-space is used to guide the optimization of the target modality under-sampling pattern. Then, the residual can be adjusted and normalized in the form of probability distribution as part of the end-to-end reconstruction network, where the adjustment allows fine-tuning of the pattern to accommodate dataset-specific information. In this way, measurements in \textit{k}-space of the target modality MR image is complementary to existing information in the reference modality MR image, potentially allowing better target modality MRI reconstruction.
We evaluate our proposed framework on a public brain MRI dataset to allow the joint optimization of the under-sampling pattern for T2w or FLAIR MRI with the assistance of the fully-sampled T1w MR image. In this study, we considered under-sampling patterns in two phase-encoding directions, which can be applied to 3D turbo spin echo imaging where acceleration is essential in reducing the scan duration. Our results indicate that the under-sampling pattern generated by our proposed framework outperform those commonly used ones as well as those generated by a multi-modal pattern learning algorithm extended from a state-of-the-art single-modal pattern learning method. In addition, we also demonstrate that our proposed method is robust to small misalignment between reference modality and target modality MR images caused by inter-scan motion.
In summary, our contributions are as three-fold: 1) We propose to optimize the under-sampling patterns of target modality data in multi-modal MRI reconstruction; 2) The cross-modality translation network is applied to exploit the statistical difference in \textit{k}-space between the translated and true target modality MR images; 3) Robustness of our proposed framework to mis-registration between reference modality and target modality is also investigated.
The rest of the paper is organized as follows. Section \ref{sec:related} summarizes recent studies on single- or multi-modal MRI reconstruction, followed by methods of optimizing the under-sampling patterns to improve the reconstruction performance. Section \ref{sec:method} introduces the mathematical formulation and details of our proposed method. Section \ref{sec:experiments} presents the implementation details as well as the results based on the under-sampling patterns learned by our proposed method. Section \ref{sec:discussion} discusses and concludes our proposed method.
\begin{figure*}[th!]
\centering
\includegraphics[width=1\textwidth]{imgs/framework.pdf}
\caption{Illustration of our proposed methods. During training, given the fully-sampled T1w MR image $\mathbf{x}_{\rm{T1}}$ and the fully-sampled target modality \textit{k}-space signals $\mathbf{y}_{\rm{TM}}$, under-sampled target modality measurements $\mathbf{y}_{\rm{TMu}}$ can be produced through the pattern initialization (PI) and the adjustive under-sampling (AU) module, in which the learned under-sampling patterns aim to maximize the assistance power of the T1w MR image, as illustrated at the top-half of the figure. In PI, the pre-trained translation network firstly synthesizes the target modality MR image $\tilde{\mathbf{x}}_{\rm{TM}}$ from the corresponding T1w MR image $\mathbf{x}_{\rm{T1}}$, which is then compared against its ground-truth signals in \textit{k}-space to obtain the residual map $\mathbf{r}$ based on their difference. Values in $\mathbf{r}$ represent the initial sampling probability for each pixel, and can be further adjusted in AU to produce the refined under-sampling pattern $\mathbf{M}$. In AU, a learnable parameter $\mathbf{w}_{\rm{TM}}$ is applied to adjust the residual map $\mathbf{r}$ during the optimization of reconstruction network on an end-to-end basis, from which refined sampling probability matrix $\mathbf{m}$ can be acquired. Given the desired under-sampling factor, the binary mask $\mathbf{M}$ can be derived through binarization to under-sample the target modality signals $\mathbf{y}_{\rm{TMu}}$. With the fully-sampled T1w MR image $\mathbf{x}_{\rm{T1}}$ and the under-sampled target modality image $\mathcal{F}^{-1}(\mathbf{y}_{\rm{TMu}})$, multi-modal reconstruction network can be applied to recover the target modality MR image $\bar{\mathbf{x}}_\text{\rm{TM}}$. Ultimately, the framework will produce the optimized under-sampling pattern $\mathbf{M}$ and the trained reconstruction network as demonstrated at the bottom-half of the figure.}
\label{fig:framework}
\end{figure*}
\section{Related Works}
\label{sec:related}
\subsection{Multi-Modal MRI Reconstruction}
\label{sec:multimrirecon}
Considering the common underlying information among MR modalities, the fusion-based reconstruction methods have been proposed to take advantage of auxiliary information of multiple modalities. For example, a fusion network was developed to learn the latent representation of multi-modal MR images on the synthesis task \cite{zhou2020hi}. MR images scanned from previous visits for the same patient can be used as reference to accelerate the acquisition in the follow-up visits \cite{souza2020enhanced}. For accelerated multi-modal MRI, methods have been proposed to use MR images of one modality as reference to assist the reconstruction of under-sampled MR images of another modality. Accelerated T2w MRI reconstruction was achieved by concatenating the fully-sampled T1w MR image with the under-sampled T2w MR image as the input of reconstruction network \cite{xiang2018ultra, kim2018improving, dar2020prior, zhou2020dudornet}. Also, information pertaining to the under-sampled MR images can be mutually beneficial when reconstructing all the under-sampled multi-modal MR images \cite{sun2019deep}.
\subsection{Under-sampling Pattern Learning}
\label{sec:masklearning}
The under-sampling pattern used in \textit{k}-space plays an important role in MRI reconstruction. Given a fixed under-sampling factor, numerous physically plausible under-sampling patterns can be adopted, such as commonly used random uniform and variable density under-sampling patterns \cite{gamper2008compressed, wang2009variable, lustig2007sparse}. Compared to the fixed patterns, several studies have demonstrated that the reconstruction quality can be further improved by using under-sampling patterns that are data- or application-specific. However, the high computation cost is associated with the traditional methods for sampling-pattern optimization due to the fact that the under-sampled MR image has to be reconstructed before evaluation \cite{seeger2010optimization, gozcu2018learning, senel2019statistically}. Alternatively, deep learning can be applied to search for the optimal under-sampling pattern with the best reconstruction performance, and different strategies have been proposed to tackle both tasks of MRI reconstruction and pattern optimization. Bahadir \textit{et al.}~proposed a U-Net-like network architecture to encode the under-sampling pattern as a set of learnable parameters to allow simultaneous optimization of MR reconstruction and under-sampling patterns \cite{bahadir2020deep}, although extra hyper-parameters were introduced to make the learning of binary under-sampling pattern differentiable. For the same task, another method was proposed to optimize the under-sampling and reconstruction for parallel MRI, where under-sampling was performed in horizontal and vertical directions \cite{aggarwal2020jmodl}. However, the values in binary under-sampling patterns were replaced with continuous variables to allow for back-propagation, which is not in line with real MRI under-sampling. Furthermore, the learned patterns often sample data with the same probability over low and high frequency regions, not considering the different importance across frequency regions in \textit{k}-space. Instead of directly acquiring the patterns, an adaptive strategy known as active acquisition was proposed to allow a slice-by-slice Cartesian acquisition \cite{zhang2019reducing}.
\section{Method}
\label{sec:method}
Our approach is illustrated in Fig.~\ref{fig:framework}, consisting of two steps. 1) Weight initialization step for the under-sampling pattern, to produce the initial weight of the under-sampling pattern for target modality MR images, in which cross-modality translation is applied to synthesize the target modality MR image from the T1w MR image. The synthesized image is then transformed back to the \textit{k}-space domain and further compared with the \textit{k}-space of the fully-sampled target modality MR image to determine the regions with low synthesis accuracy, for which the under-sampling pattern of the target modality MR image can be designed to assign high priority. 2) Reconstruction-based under-sampling pattern refinement step, to further refine the under-sampling pattern while optimizing the multi-modal reconstruction network.
\subsection{Background and Problem Definition}
\label{subsec:overview_acc_mri}
Let $\textbf{y}_{\rm{T1}} \in \mathbb{C}^{M \times N}$ and $\textbf{y}_{\rm{TM}} \in \mathbb{C}^{M \times N}$ be two complex matrices that represent the acquired fully-sampled \textit{k}-space data of the T1w and the target modality MR images, where $M$ and $N$ are the height and the width of the 2D MR image, respectively. Then, the corresponding image can be obtained from the \textit{k}-space data by applying inverse Fourier transform $\mathcal{F}^{-1}$ as $\textbf{x}_{\rm{T1}} = \mathcal{F}^{-1}(\textbf{y}_{\rm{T1}})$ and $\textbf{x}_{\rm{TM}} = \mathcal{F}^{-1}(\textbf{y}_{\rm{TM}})$, where $\textbf{x}_{\rm{T1}} \in \mathbb{C}^{M \times N}$ and $\textbf{x}_{\rm{TM}} \in \mathbb{C}^{M \times N}$ are the T1w and the target modality MR image, respectively. To accelerate the target modality MRI acquisition, a binary mask $\textbf{M} \in \{0, 1\}^{M \times N}$ is applied as the under-sampling pattern, in which each value of $1$ corresponds to a position to be sampled in \textit{k}-space. The retrospectively under-sampled target modality \textit{k}-space data $\textbf{y}_{\rm{TMu}}$ can be obtained by masking out unwanted positions in fully-sampled target modality \textit{k}-space data as $\textbf{y}_{\rm{TMu}} = \textbf{M} \odot \textbf{y}_{\rm{TM}}$, where $\odot$ denotes element-wise multiplication. The basic zero-filled reconstructed target modality MR image can be obtained as $\textbf{x}_{\rm{TMu}} = \mathcal{F}^{-1}(\textbf{y}_{\rm{TMu}})$. In deep learning-based multi-modal reconstruction methods, a neural network $f_{\bm{\theta}}$ parameterized by $\bm{\theta}$ can be applied to approximate the fully-sampled target modality MR image from $\textbf{x}_{\rm{TMu}}$ and ${\textbf{x}}_{\rm{T1}}$ as $\bar{\bf{x}}_\text{TM} = f_{\bm{\theta}} (\textbf{x}_{\rm{TMu}}, \textbf{x}_{\rm{T1}})$. In this study, we aim to optimize the under-sampling pattern $\textbf{M}^{*}$ to improve the reconstruction performance of accelerated MRI of the target modality.
\subsection{Multi-modal Under-sampling Pattern Initialization}
\label{sec:mul_modal_und}
The cross-modality translation is the first step in the Pattern Initialization (PI) module, aiming to translate the MR image from the reference modality to the target modality, and generate the initial weights of the under-sampling pattern for further optimization. The method is designed in a simple but effective way to guide the learning of the target modality under-sampling pattern, allowing the fully-sampled T1w image to better complement and assist the reconstruction of the under-sampled target modality MR image, compared to other commonly used patterns.
For the cross-modality translation, a neural network $g_{\bm{\phi}}$ is deployed to map the $i$-th fully-sampled T1w MR image to the $i$-th target modality MR image in the dataset as $\tilde{\bf{x}}_\text{TM}^{(i)} = g_{\bm{\phi}} ({\bf{x}}_\text{T1}^{(i-1)}, {\bf{x}}_\text{T1}^{(i)}, {\bf{x}}_\text{T1}^{(i+1)})$. Two neighboring slices of ${\bf{x}}_\text{T1}^{(i)}$ are also included as input of $g_{\bm{\phi}}$, to allow tiny misalignment caused by motion to be captured by the network. Given the predicted target modality MR image, comparison with the ground-truth target modality MR image can be made, and the residual difference is calculated between the predicted and fully-sampled target modality MR \textit{k}-space signals as $\textbf{r} = \vert \mathcal{F}({{\bf{\tilde x}}_{\text{TM}}}) - {\textbf{y}}_{\text{TM}} \vert$. The residual difference map reflects how well the information from the T1w MR image can be captured by the network.
To reduce the variability of prediction, the residual difference $\bf{r}$ computed on all slices in the validation set are averaged to produce weights that can reflect the assistance power of T1w MRI data for the current data distribution, and thus allow the network to be more robust to noisy prediction.
In order to allow the learning of the under-sampling pattern, a learnable weighting map ${\bf{w}}_\text{TM}$ with the same size as $\mathbf{r}$ is randomly initialized with mean value of $0$ to adjust $\textbf{r}$ through their sum as $\bf{m} = \text{ReLU}(\text{clip}(\bf{w}_\text{TM}) + \text{norm}(\bf{r}))$, where three processing steps are applied: 1) $\bf{r}$ is normalized between $0$ to $1$ to represent the a probability distribution; 2) ${\bf{w}}_{\text{TM}}$ is clipped between $-1$ and $1$ to ensure that the adjustment power on $\bf{r}$ is significant enough but not dominant; and 3) The rectified linear unit operation is performed on their sum to only keep non-negative values.
To further derive the initial weights for the given under-sampling factor $R \in [0, 1]$, the final probabilistic mask $\mathbf{P}$ can be scaled as $\mathbf{P} = R\ \mathbf{m} / \overline{m}$, where $\overline{m}$ is the mean value of the mask $\mathbf{m}$, therefore the number of pixels to be sampled can be roughly equal to the number as required by the under-sampling factor. Lastly, the binary mask $\mathbf{M}$ is generated from $\mathbf{P}$, where the value of each pixel represents the probability of success in a Bernoulli distribution. During training, the Monte Carlo method as used in {\cite{bahadir2020deep}} is applied to allow back-propagation on ${\bf{w}}_{\text{TM}}$, achieving joint optimization of under-sampling pattern and the reconstruction network.
More specifically, a threshold matrix $\mathbf{\mathbf{w}_{\rm{th}}}$ is randomly generated from a uniform distribution ranged from $0$ to $1$, and the binary mask can be obtained through binarization from the sigmoid function as $\mathbf{M} = \sigma(\sigma_p(\mathbf{P} - \mathbf{\mathbf{w}_{\rm{th}}}))$, where $\sigma$ is the sigmoid operation, and $\sigma_p$ is its slope. During inference, with the learned weights $\bf{w}_\text{TM}$ fixed, the positions in \textit{k}-space that correspond to those with the top $R \times M \times N$ highest values in $\bf{P}$ are used to under-sample target modality MR images during scanning, to ensure the number of measurements to be consistent with the under-sampling factor.
Based on the obtained under-sampling pattern, the under-sampled target modality MR image can be acquired and provided to the reconstruction network along with $\mathbf{x}_{\rm{T1}}$ as input for multi-modal reconstruction. It should be noted that different from conventional simultaneous optimization of under-sampling pattern and the reconstruction, the translation network needs to be trained first to provide reasonable initialization for pattern optimization.
\subsection{Reconstruction-based Pattern Refinement}
\label{sec:rec_pat}
In the step of pattern refinement, the network $f_{\bm{\theta}}$ is iteratively optimized for both the under-sampling pattern and reconstruction. During training, the network takes fully-sampled T1w and under-sampled target modality MR images as input to reconstruct the target modality MR image. For the under-sampling pattern of the target modality, given the fixed initial weights $\textbf{r}$ from the pattern initialization step, the learnable parameters ${\bf{w}}_\text{TM}$ can be optimized along with the reconstruction network on an end-to-end basis.
With the under-sampled target modality MR image, the T1w MR image can be stacked with the target modality MR image as the input to the network, consisting of 4 channels (the real and the imaginary parts for each modality), to produce the reconstructed target modality MR image. Such a simple way of combining information from images of multiple modalities has been applied in multi-modal reconstruction tasks \cite{xiang2018ultra}, and was proved that information across modalities can be well captured by the network.
After training, the learned under-sampling pattern $\bf{M}$ is refined, and the network $f_{\bm{\theta}}$ is optimized for reconstruction. During inference, after having obtained the fully-sampled T1w MR image, the target modality MRI acquisition can be prospectively under-sampled based on the learned under-sampling pattern, and reconstructed using the trained network with the assistance of T1w MR images
\begin{figure*}[t!]
\centering
\includegraphics[width=0.7\textwidth]{imgs/baseline.pdf}
\caption{Examples of under-sampling patterns used in our experiments. The 1D Gaussian, center, and Poisson-disc variable density patterns are illustrated in the first three columns. Examples of learned under-sampling patterns based on the validation set using the LOUPE algorithm and our proposed method are on the last two columns. The under-sampling factors R of 1/4 and 1/8 are investigated.}
\label{fig:baseline}
\end{figure*}
\subsection{Model Details}
\label{sec:mod_de}
Two networks are used in our framework to perform two different tasks, namely the cross-modality translation and the reconstruction-based pattern refinement. In this study, we adopted the architecture similar to the U-Net {\cite{ronneberger2015u}} as used in the LOUPE algorithm {\cite{bahadir2020deep}} for both tasks.
Given a set of T1w and the target modality MR image pairs $S = {(\bf{x}_\text{T1}, \bf{x}_\text{TM})}^{n}_{i=0}$ the translation network is trained simply based on the similarity between synthesized and the true target modality MR images, which is supervised using a L2 loss as:
\begin{equation}
\mathcal{L}_\text{T}({\bm{\phi}}) = \frac{1}{2n} \sum_{i=1}^{n} ( g_{\bm{\phi}}({\bf{x}}_{\text{T1}}^{(i-1)}, {\bf{x}}_\text{T1}^{(i)}, {\bf{x}}_\text{T1}^{(i+1)}) - {\bf{x}}_\text{TM}^{(i)})^2.
\end{equation}
As for the reconstruction task, we calculate the mean squared error between reconstructed and the original images to train the network as:
\begin{equation}
\mathcal{L}_\text{rec}({\bm{\theta}}) = \frac{1}{2n} \sum_{i=1}^{n} ( f_{{\bm{\theta}}} (\mathcal{F}^{-1}({\bf{M}} \odot \mathcal{F}({\bf{x}}_\text{TM}^{(i)}))) - {\bf{x}}_\text{TM}^{(i)} )^2,
\end{equation}
subject to $\frac{1}{MN} \sum_{i=1}^{M} \sum_{j=1}^{N} {\bf{P}}(i, j) = R$. During optimization, we only used the magnitudes of the complex-value images to reduce the impact of confounding variables when comparing with the LOUPE algorithm, and also based on the experiments in \cite{bahadir2020deep} where training on magnitude-only data produced superior performance for the LOUPE algorithm.
\section{Experiments}
\label{sec:experiments}
\subsection{Dataset}
A dataset from the MICCAI Multi-modal Brain Tumor Segmentation (BraTS) challenge 2019 \cite{menze2014multimodal} was used to evaluate our proposed method. The dataset is a large-scale and public brain MRI dataset consisting of 335 preprocessed volumes for each modality. The data is acquired from multiple institutions. For each subject, we used the axial slices of T1w, T2w, and FLAIR modalities. Furthermore, each slice was cropped to $192 \times 192$ to reduce the proportion of background, followed by the intensity normalization to the range between 0 and 1. The corresponding \textit{k}-space data were retrospectively obtained by performing Fourier transform on the image data. In our experiments, T2w and FLAIR are considered as the target modality due to their longer acquisition time. We extracted 1005 images and randomly split them for training, validation and testing with the ratio of 3:1:1. It is noted that images for training, validation and testing are from different subjects.
\begin{table*}[t!]
\caption{Quantitative results of the LOUPE algorithm with two different inputs for 4-fold and 8-fold accelerated T2w and FLAIR reconstruction.}
\centering
\begin{tabular}{lllll}
\hline
\multirow{2}{*}{Task} & \multicolumn{2}{c}{R = 1/4} & \multicolumn{2}{c}{R = 1/8} \\ \cline{2-5}
& PSNR & SSIM & PSNR & SSIM \\ \hline
T2w (U) $\rightarrow$ T2w & 38.18 (0.98) & 0.981 (0.0026) & 36.30 (1.05) & 0.969 (0.0019) \\
T1w + T2w (U) $\rightarrow$ T2w & 37.51 (0.74) & 0.979 (0.0018) & 35.41 (0.73) & 0.968 (0.0022) \\ \hline
FLAIR (U) $\rightarrow$ FLAIR & 38.65 (0.98) & 0.979 (0.0008) & 37.06 (0.52) & 0.965 (0.0024) \\
T1w + FLAIR (U) $\rightarrow$ FLAIR & 38.26 (1.61) & 0.978 (0.0025) & 36.38 (0.72) & 0.961 (0.0027) \\ \hline
\end{tabular}
\label{tab:loupe_only}
\end{table*}
\begin{table*}[t!]
\caption{Quantitative results of T1-assisted reconstruction method with different under-sampling patterns and under-sampling factors of 1/4 and 1/8.}
\centering
\resizebox{2\columnwidth}{!}{
\begin{tabular}{lllllllll}
\hline
& \multicolumn{2}{c}{T1w + 1/4 T2w} & \multicolumn{2}{c}{T1w + 1/8 T2w} & \multicolumn{2}{c}{T1w + 1/4 FLAIR} & \multicolumn{2}{c}{T1w + 1/8 FLAIR} \\ \cline{2-9}
& PSNR & SSIM & PSNR & SSIM & PSNR & SSIM & PSNR & SSIM \\ \hline
1D Gaussian & 36.51 (0.75) & 0.955 (0.0039) & 33.10 (0.57) & 0.926 (0.0058) & 36.51 (0.19) & 0.939 (0.0048) & 33.20 (0.31) & 0.900 (0.0082) \\
Square & 39.49 (1.00) & 0.979 (0.0021) & 36.54 (0.84) & 0.961 (0.0037) & 39.38 (0.68) & 0.966 (0.0060) & 36.53 (0.59) & 0.939 (0.0098) \\
Variable Density & 33.83 (0.39) & 0.937 (0.0018) & 32.75 (0.41) & 0.924 (0.0050) & 34.57 (0.38) & 0.925 (0.0017) & 33.07 (0.37) & 0.903 (0.0033) \\
LOUPE & 37.51 (0.74) & 0.979 (0.0018) & 35.41 (0.73) & 0.968 (0.0022) & 38.26 (1.61) & 0.978 (0.0025) & 36.38 (0.72) & 0.961 (0.0027) \\
Ours & 44.68 (0.72) & 0.991 (0.0010) & 43.08 (0.70) & 0.986 (0.0013) & 45.00 (0.53) & 0.989 (0.0013) & 43.38 (1.01) & 0.985 (0.0031) \\ \hline
\end{tabular}}
\label{tab:result}
\end{table*}
\subsection{Implementation Details}
We extended the implementation of the pattern optimization algorithm of LOUPE in \cite{bahadir2019learning}, allowing the PI and AU modules to be directly integrated with LOUPE. It was mainly implemented using the TensorFlow library \cite{tensorflow2015-whitepaper}, and neural network models were trained on NVIDIA Titan X graphic cards.
In the PI module, the complex-valued T1w MR image was firstly used to train the translation network to predict the corresponding target modality MR image. We used the Adam \cite{kingma2014adam} optimizer with $\beta_1$ and $\beta_2$ being 0.5 and 0.999, respectively. A batch size of 16 and a learning rate of $2 \times 10^{-4}$ were applied. We trained the model for at least 50 epochs until the validation loss converges, after which the residual map $\mathbf{r}$ was computed based on \textit{k}-space signals of images from the validation set.
During the pattern refinement step, given the produced residual map $\mathbf{r}$, the under-sampling pattern used for the target modality MR image is generated through the binarization trick, in which slope of the sigmoid function $\sigma_p$ was set to 5 after a grid search. Though the slope is desired to be larger to achieve better binarization, but use of a large value is more likely to end up with an unstable model due to exploding gradients. We adopted the same configuration as used for training the translation network. After the joint optimization of the under-sampling pattern and the reconstruction network, model with the best performance on the validation set was kept for evaluation. During testing, the under-sampling patterns are directly extracted from $\mathbf{P}$ and are then fixed to under-sample the target modality MR image for further reconstruction.
\subsection{Evaluation}
Foe quantitative evaluation, the peak signal-to-noise ratio (PSNR) and structural similarity index measure (SSIM) are calculated to investigate the performance of generated under-sampling patterns based on reconstructed images. PSNR is commonly used to measure the reconstruction quality, and is defined as:
\begin{equation}
{\rm{PSNR}} ({\textbf{x}_{\rm{TM}}}, {\overline{\textbf{x}}_{\rm{TM}}}) = 10\ {\rm{log}_{10}} \frac{MN\ \rm{max}(\textbf{x}_{\rm{TM}})}{\sum_{i=1}^{MN}(\textbf{x}_{\rm{TM}}(i) - {\overline{\textbf{x}}_{\rm{TM}}}(i))^2},
\end{equation}
where $\rm{max}(\textbf{x}_{\rm{TM}})$ denotes the maximum value of pixels in $\textbf{x}_{\rm{TM}}$ ($1$ in our case). On the other hand, we followed the standard implementation of SSIM as defined in \cite{wang2004image} to assess the image quality, which is defined as:
\begin{equation}
{\rm{SSIM}} ({\textbf{x}_{\rm{TM}}}, {\overline{\textbf{x}}_{\rm{TM}}}) = \frac{(2 \mu_{\textbf{x}_{\rm{TM}}} \mu_{\overline{\textbf{x}}_{\rm{TM}}} + C_1)(2 \sigma_{{\textbf{x}_{\rm{TM}}} \overline{\textbf{x}}_{\rm{TM}}} + C_2)}{(\mu_{\textbf{x}_{\rm{TM}}}^2 + \mu_{\overline{\textbf{x}}_{\rm{TM}}}^2 + C_1)(\sigma_{\textbf{x}_{\rm{TM}}}^2 + \sigma_{\overline{\textbf{x}}_{\rm{TM}}}^2 + C_2)} ,
\end{equation}
where $\mu_{\textbf{x}_{\rm{TM}}}$ and $\mu_{\overline{\textbf{x}}_{\rm{TM}}}$ are local average values of the original and reconstructed TM images. Likewise, $\sigma_{\textbf{x}_{\rm{TM}}}^2$ and $\sigma_{\overline{\textbf{x}}_{\rm{TM}}}^2$ are local variances, whereas $\sigma_{{\textbf{x}_{\rm{TM}}} \overline{\textbf{x}}_{\rm{TM}}}$ is the local covariance. $C_1$ and $C_2$ are constant relaxation terms.
The evaluation was performed based on the reconstructed T2w and FLAIR MR images from the original and the transformed testing set separately. Using the under-sampling factor of $1/4$ and $1/8$, we retrospectively under-sampled the \textit{k}-space data with different patterns. Afterwards, the under-sampled \textit{k}-space data were transformed into the image domain for further evaluation, and a 5-fold cross-validation was performed on the dataset for thorough validation.
As a baseline, the performance of the translation network is also reported at the same under-sampling factors. Note that since the T1w MR image is always required to obtain the under-sampling pattern through cross-modality translation, we did not consider the evaluation using the under-sampled target modality MR image alone for reconstruction.
\subsection{Baseline Methods}
Several commonly used under-sampling patterns in compressed sensing MRI were tested on the reconstruction task. Examples of used patterns are illustrated in Fig.~\ref{fig:baseline}, including the one-dimensional Gaussian, center, and variable density pattern. In the one-dimensional Gaussian pattern, the unit Gaussian distribution was used to assign the probability of each row in the \textit{k}-space being measured, and certain numbers of rows proportional to the under-sampling factor were sampled from the associated probability as the pattern. In a center pattern, only the rectangular-shaped central part is measured, the proportion of which corresponds to the under-sampling factor. As for the variable density under-sampling pattern, the Poisson-disc sampling algorithm was applied to generate the pattern {\cite{bridson2007fast}}, and a binary search was performed to find the slope of density that allows generating a pattern of specific under-sampling factor.
In addition to fixed patterns, the state-of-the-art under-sampling pattern learning algorithm LOUPE {\cite{bahadir2020deep}} was also trained on the same dataset for comparison. Since it was originally designed for single-modal reconstruction, we followed the approach in \cite{xiang2018ultra} to modify the architecture to enable multi-modal MRI reconstruction with optimized under-sampling patterns, by adding the fully-sampled T1w MR image to the reconstruction input. In addition, we also investigated whether the LOUPE algorithm can be modified in such a simple way to improve the reconstruction performance by incorporating information from an extra modality.
\begin{table*}[t!]
\caption{Quantitative results of our method evaluated on the dataset under different under-sampling factors and different combinations of MR images as input.}
\centering
\begin{tabular}{lllll}
\hline
\multirow{2}{*}{Task} & \multicolumn{2}{c}{R = 1/4} & \multicolumn{2}{c}{R = 1/8} \\ \cline{2-5}
& PSNR & SSIM & PSNR & SSIM \\ \hline
T1w $\rightarrow$ T2w & 30.15 (1.14) & 0.945 (0.0115) & 30.15 (1.14) & 0.945 (0.0115) \\
T1w + T2w (U) $\rightarrow$ T2w & 44.68 (0.72) & 0.991 (0.0010) & 43.08 (0.70) & 0.986 (0.0013) \\
T1w (T) + T2w (U) $\rightarrow$ T2w & 43.89 (0.59) & 0.990 (0.0010) & 42.45 (0.65) & 0.985 (0.0015) \\ \hline
T1w $\rightarrow$ FLAIR & 32.09 (1.47) & 0.945 (0.0097) & 32.09 (1.47) & 0.945 (0.0097) \\
T1w + FLAIR (U) $\rightarrow$ FLAIR & 45.00 (0.53) & 0.989 (0.0013) & 43.38 (1.01) & 0.985 (0.0031) \\
T1w (T) + FLAIR (U) $\rightarrow$ FLAIR & 45.04 (0.53) & 0.989 (0.0010) & 42.30 (1.30) & 0.980 (0.0048) \\ \hline
\end{tabular}
\label{tab:ours_only}
\end{table*}
\section{Results}
\subsection{Investigation of Current Pattern Learning Methods}
As there is currently no pattern optimization method specifically designed for multi-modal MRI reconstruction, we modified the existing single-modal method, LOUPE \cite{bahadir2019learning}, to achieve pattern optimization for under-sampled T2w MRI with the assistance of the fully-sampled T1w MR image.
The reconstruction results are summarized in Table \ref{tab:loupe_only}. It can be seen that the introduction of the fully-sampled T1w MR image cannot improve T2w/FLAIR MRI reconstruction for neither low or high under-sampling factors. Some examples of the learned under-sampling patterns for reconstruction from MR images of under-sampled target modality only and fully-sampled T1w with under-sampled target modality are shown in Fig.~\ref{fig:loupe_only}, and no significant difference was found between these two forms of input data.
\begin{figure*}[t!]
\centering
\includegraphics[width=0.68\textwidth]{imgs/loupe_only.pdf}
\caption{Examples of under-sampling patterns learned by the LOUPE algorithm with under-sampling factors of 1/4 and 1/8 for T2w and FLAIR reconstruction. For each target modality (T2 or FLAIR), two types of input are tested, including target modality only and target modality + T1w.}
\label{fig:loupe_only}
\end{figure*}
Based on these results, it can be concluded that the LOUPE algorithm cannot be simply modified to allow learning of under-sampling patterns for T1-assisted T2w or FLAIR MRI reconstruction. Therefore, it motivates us to develop the two-step method of under-sampling pattern optimization for multi-modal MRI reconstruction.
\subsection{Comparisons with Baseline Patterns}
The reconstruction results are summarized in Table \ref{tab:result}. The PSNR and SSIM were calculated in the whole brain for each subject. It can be seen that the proposed method outperformed other fixed and learned patterns on T2w and FLAIR MRI reconstruction under $1/4$ and $1/8$ under-sampling factors. Selected reconstructions from different patterns are visualized in Fig.~\ref{fig:qualitative} along with the corresponding error maps compared with fully-sampled images. The learned patterns using our proposed method yield superior reconstruction performance, achieving better preservation of fine anatomical structures than other patterns.
Examples of under-sampling pattern optimized by our methods are shown in the fifth column in Fig.~\ref{fig:baseline}, the boundaries of which appear to be ellipse-shaped and include important information of low-frequency regions, and the under-sampling patterns extend evenly at the higher under-sampling factor. Meanwhile, sparse measurements are also taken from high-frequency regions that benefit the recovery of image details. Overall, the learned under-sampling pattern is in accordance with the common principle of under-sampling pattern design.
\subsection{Our Method with Different Inputs}
In order to understand the effect of data used as input for reconstructing TM images, we used various forms of MR images for reconstruction: T1w only (as applied in cross-modality translation); T1w and under-sampled target modality; and lastly transformed T1w and under-sampled target modality.
The results are listed in Table \ref{tab:ours_only}. For the translation network that takes the T1w MR image as input to synthesize the corresponding target modality MR image, its performance affects the distribution of residual maps. Example slices of the translated results can be seen in the bottom of Fig.~\ref{fig:ours_only}, where the translation can generally produce the corresponding target modality MR image but lack anatomical details, resulting in low PSNR and SSIM. Considering that the same network architecture was adopted for both translation and reconstruction, such results demonstrated the limitation of cross-modality synthesis. Therefore, the translated target modality MR image was not used for the final reconstruction, but used to generate the initial weight for pattern optimization.
\begin{figure*}[t!]
\centering
\includegraphics[width=0.7\textwidth]{imgs/ours_only.pdf}
\caption{Top: Examples of an original T1w image on the left, and the three motion augmented images. Bottom: Synthesized T2w and FLAIR images from the original T1w image.}
\label{fig:ours_only}
\end{figure*}
\subsection{Influence of Inter-scan Motion}
The dataset has been pre-processed by the data provider to make sure that multi-modal MR images are registered. However, in practice, the inter-scan motion is inevitable due to patient movement. Therefore, in addition to evaluation on the original dataset, we also simulated patient movement by randomly applying rigid transformations to the testing set of T1w MR images. Based on examples of tracked brain motion on healthy subjects during a scan in \cite{stucht2015highest}, we simulated from a similar range of motion. More specifically, before cropping and normalization, we performed random translation and rotation to all slices of T1w MR images, where images were randomly shifted between $-5$ to $5$ mm in both horizontal and vertical directions, and then randomly rotated between $-5$ to $5$ degrees.
To investigate the effect of inter-scan motion, evaluation was made on the testing set with simulated movement, and the reconstruction results are shown in the last row of Table \ref{tab:ours_only}. It can be seen that the PSNR drop is within $1$ dB for all cases, and the SSIM decrease by at most 0.004, which demonstrates the robustness of the proposed method to inter-scan misalignments. Some examples of transformed T1w MR images with simulated motion are visualized in Fig.~\ref{fig:ours_only}, which demonstrate the degree of simulated movement.
\begin{figure*}[t!]
\centering
\includegraphics[width=0.7\textwidth]{imgs/qualitative.pdf}
\caption{The reconstructed T2w (top) and FLAIR (bottom) images with five different under-sampling patterns and the corresponding error maps compared to the ground-truth T2w/FLAIR images for under-sampling factors of 1/4 and 1/8. Darker color in the error map indicates higher error.}
\label{fig:qualitative}
\end{figure*}
\section{Discussion and Conclusion}
\label{sec:discussion}
In this paper, we focus on optimizing the under-sampling patterns of target modalities being T2w or FLAIR, given the fully-sampled T1w MR image during multi-modal MRI. We proposed to utilize the target modality MR image synthesized from the corresponding T1w MR image to guide the learning of the under-sampling pattern of the target modality. The residual map reflects positions in \textit{k}-space where the T1w MR image cannot synthesize well during cross-modality translation, and the under-sampling patterns can be designed based on the residual map, so that the reconstruction model can best exploit the information from the fully-sampled T1w MR image. The basic principle is to sample data in the target modality MRI acquisition complementary to the information from the T1w MR image.
We conducted extensive experiments on a public MRI dataset, and demonstrated that the learned under-sampling patterns with the proposed method outperformed commonly used under-sampling patterns and the patterns learned by the state-of-the-art pattern optimization algorithm up to 8-fold acceleration. The initial weight generated based on the residual difference from the translation network is integrated with the state-of-the-art single-modal MRI pattern optimization algorithm for T1-assisted multi-modal reconstruction, which improves the multi-modal reconstruction performance by a large margin, compared to directly modifying the single-modal MRI pattern optimization algorithm by taking the fully-sampled T1w MR image as additional input.
In this study, we adopted a simple strategy to design the reconstruction task, as the main focus is to investigate the effectiveness of learned under-sampling patterns on multi-modal reconstruction. Therefore, our proposed pattern generation methods can also be potentially integrated with state-of-the-art reconstruction models to improve the performance of multi-modal reconstruction.
However, limitations also exist for this method. We have not evaluated the proposed method using prospectively under-sampled data, and the effectiveness of our method in real acquired under-sampled \textit{k}-space data are yet to be demonstrated. In addition, motion robustness of the proposed framework is only demonstrated for small motion, and it may fail when bulk motion occurs between T1w and target modality MRI acquisitions. Therefore, integration with a registration task could be a future research direction to further improve the robustness of the multi-modal pattern optimization and reconstruction framework to inter-scan motion.
\bibliographystyle{IEEEtran}
|
\section{Introduction}
Low-rank approximations and other related spectral methods are among the most fundamental and ubiquitous tools in data analysis. Their computational and statistical aspects have been studied extensively and are among the central themes in numerical analysis and multivariate statistics respectively. The classical approaches however face new challenges with massive data sets are being generated every day across diverse fields: gene expression analysis \citep[see, e.g.,][]{kluger2003spectral}, protein-to-protein interaction \citep[see, e.g.,][]{stelzl2005human}, MRI image analysis \citep[see, e.g.,][]{smith2004advances}, and the analysis of large graphs and social networks \citep[see, e.g.,][]{clauset2004finding, scott2017social}, to name a few. Scalable computation and valid statistical inferences in high dimensions for low-rank approximations are the subjects of fervent research interests in recent years.
Consider a data matrix $A\in \mathbbm{R}^{d_1\times d_2}$. As Eckart-Young Theorem indicates, its best rank-$r$ approximation, denoted by $A_r$ hereafter, can be uniquely identified by its top $r$ singular values and their corresponding vectors, assuming that its $r$th and $(r+1)$th singular values are different for simplicity. From a statistical perspective, we are interested in how well properties of a data-generating model can be inferred from $A_r$, be they principal components, statistical factors, empirical orthogonal functions, or else. Suppose that we are interested in estimating a parameter that can be represented by a matrix $\Theta\in {\mathbb R}^{d_1\times d_2}$ in the Frobenius norm $\|\cdot\|_{\rm F}$. These analyses often tell us about the stochastic behavior of the ``estimation error'' $\|A_r-\Theta\|_{\rm F}$. For example, if $A=\Theta+E$ where $E$'s entries are independent $N(0,\sigma^2)$ variables, then $\|A_r-\Theta\|_{\rm F}^2=O_p(r(d_1+d_2)\sigma^2)$ so that $A_r$ is a consistent estimate of $\Theta$ under the Frobenius norm whenever $\sigma^2\ll [r(d_1+d_2)]^{-1}$. Classical multivariate analysis has primarily focused on ``thin'' data matrices where $d_1$ (sample size) is large and $d_2$ (dimensionality or number of features) is small \citep[see, e.g.,][]{anderson2003introduction}. On the other hand, the emphasis of recent effort is on the so-called high dimensional paradigm where both $d_1$ and $d_2$ are large \citep[see, e.g.,][]{bai2010spectral, wainwright2019high}.
Similarly, from a computational perspective, tremendous progress has been made in the past couple of decades towards fast computation of low-rank approximations to $A$ when both $d_1$ and $d_2$ are large, and randomized algorithms have become increasingly popular for this purpose \citep[see, e.g.,][]{mahoney2011randomized, woodruff2014sketching}. The goal is to find an $\widehat{A}$ which may or may not be of rank $r$ that can approximate $A$ nearly as well as $A_r$ under suitable matrix norms. Typical guarantees, in terms of Frobenius norm, for instance, for the output $\widehat{A}$ from these algorithms take the form:
\begin{equation}
\label{eq:multerr}
\|A-\widehat{A}\|_{\rm F}\le (1+\epsilon)\|A-A_r\|_{\rm F},
\end{equation}
for a sufficiently small factor $\epsilon>0$.
It is, however, worth noting a subtle distinction between these two perspectives. While from a computational point of view, we are interested in finding a ``good'' approximation to $A$; from a statistical viewpoint, the focus is often on $A_r$ itself rather than the original data matrix $A$ because $A_r$ captures the more stable and oftentimes more reproducible properties of an otherwise complicated data-generating model, even if the data matrix $A$ itself may not necessarily be of low rank. These two goals are closely connected yet could also be rather different. In particular, \eqref{eq:multerr} cannot ensure that $\widehat{A}$ would inherit the nice statistical properties one may be able to establish for $A_r$: unless $A$ can be well approximated by $A_r$, \eqref{eq:multerr} does not imply that $\|A_r-\widehat{A}\|_{\rm F}$ is small so we may not be able to infer from \eqref{eq:multerr} that $\widehat{A}$ is necessarily a ``good'' estimate of $\Theta$ even if $\|A_r-\Theta\|_{\rm F}$ is known to be small. Consider, for example, the signal-plus-noise model $A=\Theta+E$ discussed before. Recall that $\|A-A_r\|_{\rm F}^2\asymp_p d_1d_2\sigma^2$ so that $\widehat{A}$ may be inconsistent even though $A_r$ is a consistent estimate of $\Theta$ when $(d_1d_2)^{-1}\ll \sigma^2\ll (d_1+d_2)^{-1}$ at least for small $r$s. Moreover, $\widehat{A}$ oftentimes has a rank much greater than $r$ and therefore unsuitable for situations where an exact rank-$r$ estimate is sought. Our work aims to fill in this gap between the two strands of fruitful research by investigating how well we can reconstruct $A_r$ from random sparsification of a general matrix $A$, and therefore contributes to growing literature to unify both statistical and computational perspectives \citep[see, e.g.,][]{raskutti2016statistical}.
To facilitate the storage, communication, or manipulation of a large data matrix, one often approximates the original data matrices with a more manageable amount of sketches. See \cite{mahoney2011randomized, woodruff2014sketching} for recent reviews. A popular idea behind many of these approaches is sparsification---creating a sparse matrix by zeroing out some entries of the original data matrix. Sparse sketching of a large data matrix not only reduces space complexity but also allows for efficient computations. See, e.g., \cite{frieze2004fast, arora2006fast, achlioptas2007fast, drineas2008relative, boutsidis2009improved, mahoney2009cur, drineas2011note, achlioptas2013near, krishnamurthy2013low}, among many others. In particular, we shall focus on matrix sparsification schemes that sample each entry of $A$ independently with a prescribed probability.
Denote by $\omega_{ij}$ the indicator that the $(i,j)$ entry of $A$ is sampled and $\Omega=(\omega_{ij})_{1\le i\le d_1,1\le j\le d_2}$. Write $\mathcal P_\Omega(A)=(a_{ij}\omega_{ij}/p_{ij})_{1\le i\le d_1,1\le j\le d_2}$ where $p_{ij}={\mathbb P}(\omega_{ij}=1)$ and we shall adopt the convention that $0/0=0$ in the rest of the paper. Many different sampling schemes have been developed in recent years so that the spectral error $\|\mathcal P_\Omega(A)-A\|$ can be made small even with a tiny fraction of the entries sampled, e.g., $\|\Omega\|_{\ell_1}=\sum_{i,j}|\omega_{ij}|\ll d_1d_2$. See, e.g., \cite{arora2006fast,achlioptas2007fast,drineas2011note} among numerous others. Given the closeness of $\mathcal P_\Omega(A)$ to $A$, it is natural to consider estimating $A_r$ by the best rank-$r$ approximation $[\mathcal P_\Omega(A)]_r$ of $\mathcal P_\Omega(A)$. We show that this is indeed a reasonable approach in light of the tight bounds controlled by $\|\mathcal P_\Omega(A) - A\|$ under the spectral error. But more interestingly, we argue that for an arbitrary $A$, by choosing $p_{ij}$s appropriately, we can derive a much better estimate of $A_r$ for all $r$s. To this end, we introduce a sampling scheme and a companion rank-$r$ estimator $\widehat{A}_r$ of $A_r$ from $\mathcal P_\Omega(A)$. We show that $\|\widehat{A}_r-A_r\|_{\rm F}$ can be bounded via the sampling error of $\|\mathcal P_\Omega(N_r)-N_r\|$ instead of $\|\mathcal P_\Omega(A)-A\|$ where $N_r=A-A_r$. This leads to a significantly improved estimate of $A_r$. In particular, if $A$ is of rank up to $r$, then $N_r=0$ so that our estimator $\widehat{A}_r$ recovers $A$ exactly with high probability. This makes an immediate connection with the popular matrix completion problem where one seeks the exact recovery of a large matrix from partial observations of its entries \citep[see, e.g.,][]{candes2009exact, gross2011recovering, recht2011simpler}
Despite such a similarity, there are also significant distinctions between our results and those typical for matrix completion. For exact matrix completion, it is usually assumed that the rank of $A$ is small and known apriori, and that its singular vectors are incoherent so that each entry carries a similar amount of information. On the other hand, our approach does not make such stipulations and is generally applicable. As we shall show, $\widehat{A}_r$ remains a good estimate of $A_r$ even if $A$ is of full rank, and the incoherence condition can be done away with through carefully designed sampling schemes.
The rest of the paper is organized as follows. We first discuss the na\"ive approach to approximate $A_r$ by $[\mathcal P_\Omega(A)]_r$ and show how it connects with existing results for matrix sparsification and completion in the next section. Section 3 introduces our sampling and estimation scheme for the improved approximation of $A_r$. We complement the theoretical results with numerical experiments in Section 4 to further demonstrate the merits of our approach. Proofs of the main results are presented in Section 5 with more technical details relegated to the Appendix
\section{Na\"ive Estimate of $A_r$
In the rest of the paper, we shall denote by $A_{i.}$, $A_{.j}$ and $a_{ij}$ the $i$-th row, $j$-th column, and $(i,j)$ entry of a matrix $A$, respectively. Similarly, $v_i$ is the $i$-th element of a vector $v$. Following the convention, $\|\cdot\|_\textup{F}$ denotes the Frobenius norm, $\|\cdot\|$ the spectral norm for a matrix and the $\ell_2$ norm for a vector, $\|\cdot\|_\infty$ the element-wise infinity norm, $\|\cdot\|_{\ell_1}$ the (vectorized) $\ell_1$ norm.
We shall also write the inner product between two matrices as $\langle A,B\rangle=\mathrm{trace}(A^\top B)$.
Most existing studies of the efficacy of sparsification algorithms focus on bounding the spectral error. For concreteness, consider a scheme that observes $a_{ij}$ independently with probability
\begin{align}\label{p:sparse}
p_{ij}=\min\left\{{n\over 2}\left({a_{ij}^2\over \|A\|_{\rm F}^2}+{|a_{ij}|\over \|A\|_{\ell_1}}\right),1\right\}.
\end{align}
It is not hard to see that, from Chernoff's bound, the total number $\|\Omega\|_{\ell_1}$ of observed entries is within $[n/2, 2n]$ with high probability. Also, there exists a numerical constant $C>4$ such that
$$
\left\|\mathcal P_\Omega(A)-A\right\|\le 4\|A\|_{\rm F}\sqrt{d\over n}+C{\|A\|_{\ell_1}\over n}\sqrt{\log (d)+t},
$$
with probability at least $1-e^{-t}$, where $d=\max\{d_1,d_2\}$ \citep[see, e.g.,][]{o2018random}. In what follows, we shall use $C$, and similarly $C_1$, $C_2$, etc., as generic numerical constants that may take different values at each appearance. Because
$$
\|A\|_{\ell_1}\le d\|A\|_{\rm F},
$$
we have, under the above event,
$$
\left\|\mathcal P_\Omega(A)-A\right\|\le C\|A\|_{\rm F}\left(\sqrt{d\over n}+{d\sqrt{\log (d)+t}\over n}\right).
$$
In particular, whenever $n\ge (\alpha+1)d\log d$, we get
\begin{equation}\label{specerr}
\left\|\mathcal P_\Omega(A)-A\right\|\le C\|A\|_{\rm F}\sqrt{d\over n},
\end{equation}
with probability at least $1- d^{-\alpha}$.
In light of such a bound on the spectral norm, it is natural to consider estimating $A_r$ by the respective best low-rank approximation of $\mathcal P_\Omega(A)$. The connection is made more precise by the following observation that links the closeness between $A_r$ and $[\mathcal P_\Omega(A)]_r$ with the spectral error $\|\mathcal P_\Omega(A)-A\|$.
More specifically, for a matrix $A\in \mathbb R^{d_1\times d_2}$,
its singular values are denoted by
$$
\sigma_1(A)\ge \sigma_2(A)\ge\cdots\ge \sigma_d(A).
$$
Then we have
\begin{lemma}
\label{le:lowrankappr}
Let $B=A+E$. If $\sigma_r(A)>\sigma_{r+1}(A)$ and $\|E\| \leq (\sigma_r(A) - \sigma_{r+1}(A))/2$, then there exists a numerical constant $C_0>0$ such that for any $1\le r<d$,
$$
\|B_r-A_r\|_{\rm F}\le C_0\sqrt{r}\|E\| \left(1+{\sqrt{\sigma_{r+1}(A)\sigma_1(A)}\over \sigma_r(A)-\sigma_{r+1}(A)}\right).
$$
\end{lemma}
In light of Lemma \ref{le:lowrankappr} and \eqref{specerr}, we can bound the difference between $A_r$ and $[\mathcal P_\Omega(A)]_r$ as follows:
\begin{theorem}
\label{th:old}
Assume that each entry of $\Omega$ is independently sampled from binomial trails with probability given by
\begin{equation*}
p_{ij}=\min\left\{{n\over 2}\left({a_{ij}^2\over \|A\|_{\rm F}^2}+{|a_{ij}|\over \|A\|_{\ell_1}}\right),1\right\}.
\end{equation*}
If $\sigma_r(A)>\sigma_{r+1}(A)$, then there exist numerical constants $C_1,\ C_2>0$ such that for any $\alpha>0$,
$$
\|[\mathcal P_\Omega(A)]_r-A_r\|_{\rm F}\le C_1\|A\|_{\rm F}\sqrt{rd\over n}\left(1+{\sqrt{\sigma_{r+1}(A)\sigma_1(A)}\over \sigma_r(A)-\sigma_{r+1}(A)}\right),
$$
with probability at least $1-d^{-\alpha}$ provided that
$$
n\ge C_2d\max\left\{(1+\alpha)\log d,{\|A\|_{\rm F}^2\over (\sigma_{r}(A) - \sigma_{r+1}(A))^2}\right\}.
$$
\end{theorem}
Theorem \ref{th:old} justifies $[\mathcal P_\Omega(A)]_r$ as a valid estimator of $A_r$. It is worth noting that both the sample size requirement and error bound depend on the eigengap $\sigma_r(A)-\sigma_{r+1}(A)$. This is inevitable since the eigengap characterizes the stability of best rank-$r$ approximation and in the extreme case where eigengap is 0, $A_r$ is not uniquely defined. We hereafter assume that $\sigma_r(A) > \sigma_{r+1}(A)$.
It is instructive to revisit the signal-plus-noise model we discussed in the Introduction: $A=\Theta+E\in \mathbb R^{d\times d}$ where, to fix ideas, we assume that the signal $\Theta={\mathbf u}{\mathbf v}^\top$ is a rank-one square matrix with both ${\mathbf u}$ and ${\mathbf v}$ of unit length, and that the noise $E$ has independent $N(0,\sigma^2)$ entries. As noted before, it is of special interest to focus on the case when $d^{-2}\ll\sigma^2\ll d^{-1}$ and there is a mismatch between the existing results from the statistical and computational sides. It is not hard to see that in this case $\|A\|_{\rm F}=O_p(d\sigma)$ and $\|A_1-\Theta\|_{\rm F}=O_p(\sigma\sqrt{d})$. In light of Theorem \ref{th:old}, this suggests that
$$
\|[\mathcal P_\Omega(A)]_1-\Theta\|_{\rm F}\le \|[\mathcal P_\Omega(A)]_1-A_1\|_{\rm F}+\|A_1-\Theta\|_{\rm F}=O_p\left(\sigma d^{3/2} n^{-1/2}\right).
$$
Hence $[\mathcal P_\Omega(A)]_1$ is consistent under the sample size requirement that $n\ge C\sigma^2d^3\log d$. In particular, if $\sigma^2\sim d^{-\alpha}$ for some $\alpha\in (1,2)$, then the sample size requirement can be expressed as $n\gtrsim d^{3-\alpha}\log d$ which means that a consistent estimate of $\Theta$ can be obtained from a vanishing proportion ($n/d^2\to 0$) of the entries of $A$.
Another interesting test case here is when $A$ is of rank up to $r$ so that $A=A_r$. If, in addition, all entries of $A$ are of the same order so that $p_{ij}\sim n/(d_1d_2)$, then it is not hard to see that
$$
\mathbb E\|\mathcal P_\Omega(A)-A\|_{\rm F}^2\le {C\|A\|_{\rm F}^2d^2\over n},
$$
so that we have the following performance bound for $\mathcal P_\Omega(A)$ as an approximation to $A$ under the Frobenius norm:
$$
\|\mathcal P_\Omega(A)-A\|_{\rm F}=O_p\left(\|A\|_{\rm F}\sqrt{d^2\over n}\right).
$$
On the other hand, Theorem \ref{th:old} indicates that
$$
\|[\mathcal P_\Omega(A)]_r-A\|_{\rm F} = O_p\left(\|A\|_{\rm F}\sqrt{rd\over n}\right).
$$
This immediately suggests that $[\mathcal P_\Omega(A)]_r$ is a better estimate of $A$ by leveraging the fact that $\mathrm{rank}(A)$ is small. Although this example shows the efficacy of $[\mathcal P_\Omega(A)]_r$ as an estimate of $A_r$, it also points to room for further improvement at least when $A$ is of low rank. In fact, with additional conditions, it may even be possible to recover a low-rank matrix $A$ exactly from $\mathcal P_\Omega(A)$!
For this to be possible, one usually assumes that $\mathrm{rank}(A)$ is small and known apriori. Another essential concept is the so-called incoherence condition. Denote by $A = U\Sigma V^\top$ its singular value decomposition. We say a rank-$r$ matrix $A$ is $\mu$-incoherent if
$$\max\left\{{d_1\over r} \|U_{i.}\|^2, {d_2\over r} \|V_{j.}\|^2\right\}\le \mu.$$
Intuitively, the incoherence condition ensures that each entry of $A$ is of similar importance so that missing any one of them will not prevent us from being able to recover $A$. Numerous tractable algorithms have been developed to reconstruct $A$ assuming that it is $\mu$-incoherent and each of its entries is sampled independently with probability $n/(d_1d_2)$. See, e.g., \cite{candes2009exact,candes2010power,keshavan2010matrix,sun2016guaranteed} among many others. Interested readers are also referred to \cite{davenport2016overview, chen2018harnessing} for a couple of recent surveys. For example, a natural approach is to reconstruct $A$ by the solution to
$$\min_{\mathrm{rank}(B)\le r} \left\langle \mathcal P_\Omega\left(A-B\right),A-B \right\rangle.$$
In particular, results from the recent work of \cite{chen2020noisy} indicate that we can recover $A$ exactly this way, with high probability, if
$$n \geq C\kappa^4\mu^2r^2(d_1+d_2)\log^3 (d_1+d_2),$$
where $\kappa=\sigma_1(A)/\sigma_r(A)$ is the condition number of $A$.
These results immediately suggest that we can do even better than $[\mathcal P_\Omega(A)]_r$, albeit under additional assumptions. But can we do better than $[\mathcal P_\Omega(A)]_r$ in general? Especially, can we do away with the incoherence assumption and the need to know $\mathrm{rank}(A)$ apriori? We shall now argue that the answer is indeed affirmative.
\section{Improved Estimate of $A_r$}
There are two main ingredients to our approach: sparsification with carefully chosen sampling probabilities to remove the need for incoherence; and an agnostic procedure based on projected gradient descent to estimate $A_r$ for any $r$.
The fact that incoherence is closely related to the uniform sampling was noted first by \citet{chen2015completing}. They observed that a rank-$r$ matrix $A$ could be recovered exactly without the incoherence condition by taking
$$p_{ij} \propto {d_1\over r} \|U_{i.}\|^2+ {d_2\over r} \|V_{j.}\|^2.$$
The difficulty of course is that it requires that $A$ be of rank $r$. Nonetheless, motivated by the observation that
$${\|A_{i.}\|\over\sigma_1(A)} \le \|U_{i.}\| \le {\|A_{i.}\|\over\sigma_r(A)},$$
and
$${\|A_{.j}\|\over\sigma_1(A)} \le \|V_{j.}\| \le {\|A_{j.}\|\over\sigma_r(A)},$$
we shall consider sampling $a_{ij}$ with probability
\begin{equation}
\label{eq:defp}
p_{ij}=\min\left\{{n\over 3}\left({\|A_{i\cdot}\|^2\over d_2\|A\|_{\rm F}^2}+{\|A_{\cdot j }\|^2\over d_1\|A\|_{\rm F}^2}+{|a_{ij}|\over \|A\|_{\ell_1}}\right), 1\right\},
\end{equation}
Compared with \eqref{p:sparse}, we essentially replaced $a_{ij}^2$ with $\|A_{i.}\|^2+\|A_{.j}\|^2$ which plays a similar role as $(d_1\|U_{i.}\|^2+d_2\|V_{j.}\|^2)/r$ when $A$ is indeed of rank $r$. Unlike the factor $(d_1\|U_{i.}\|^2+d_2\|V_{j.}\|^2)/r$, however, our choice of $p_{ij}$ does not depend on $r$. This is critical in allowing us to estimate $A_r$ for any $r$ from the sampled entries.
We shall now describe how to reconstruct $A_r$ from $\mathcal P_\Omega (A)$. Our approach is similar in spirit to \cite{ge2017no} and \cite{chen2020noisy}. See also \cite{jain2017non}. Recall that $A_r$ is the solution to
$$
\argmin_{\mathrm{rank}(B)\le r}\|A-B\|_{\rm F}^2.
$$
It is clear by construction
$$\mathbb E \left\langle \mathcal P_\Omega\left(A-B\right),A-B \right\rangle=A-B,$$
so that we can consider minimizing $\left\langle \mathcal P_\Omega\left(A-B\right),A-B \right\rangle$. Any matrix $B$ with rank up to $r$ can be written as $B=XY^\top$ for some $X\in \mathbb R^{d_1\times r}$ and $Y\in \mathbb R^{d_2\times r}$ so that this is equivalent to minimizing $\left\langle \mathcal P_\Omega\left(A-XY^\top\right),A-XY^\top \right\rangle$ over the couple $(X,Y)$. However, decomposition $B=XY^\top$ is not unique. To overcome such an identifiable issue, we shall consider estimating $A_r$ by the solution to
\begin{equation}\label{defopt}
\min_{\mathcal F_\beta} f(X,Y) =\frac{1}{2} \left\langle \mathcal P_\Omega\left(A-XY^\top\right),A-XY^\top \right\rangle + \frac{1}{8} \left\|X^\top X - Y^\top Y\right\|_{\rm F}^2 ,
\end{equation}
where
$$\mathcal F_\beta=\left\{ (X,Y)\in \mathbb R^{d_1 \times r}\times \mathbb R^{d_2\times r}: \|X_{i.}\|\leq \|A_{i.}\|/\beta,\ \|Y_{j.}\|\leq \|A_{.j}\|/\beta \right\},$$
and $\beta$ is a tuning parameter to be specified later.
The second term on the right-hand side of \eqref{defopt} forces $X^\top X=Y^\top Y$ so that if $XY^\top =\tilde{U}\tilde{\Sigma}\tilde{V}^\top$ then $X=\tilde{U}\tilde{\Sigma}^{1/2}$ and $Y= \tilde{V}\tilde{\Sigma}^{1/2}$. Note that
$$\|(A_r)_{i.}\|^2 = \sum_{j=1}^r \sigma_j(A)^2u_{ij}^2 \geq \sigma_r(A) \sum_{j=1}^r \sigma_j(A) u_{ij}^2.$$
When $X$ to be close to $U_r\Sigma^{1/2}_r$, then
$$\|X_{i.}\|^2 \approx \sum_{j=1}^r\sigma_j(A) u_{ij}^2 \leq \|(A_r)_{i.}\|^2/\sigma_r(A) \le \|A_{i.}\|^2/\sigma_r(A),$$
so that the constraint on $X_{i\cdot}$, and similarly that on $Y_{j\cdot}$, becomes inactive at least for a sufficiently small $\beta$.
The objective function $f$ in \eqref{defopt} is non-convex jointly over $(X,Y)$. Nonetheless, it is natural to solve it by projected gradient descent with a good initialization. To this end, let $\mathcal P_{\mathcal F_\beta}$ be the projection on the convex set $\mathcal F_\beta$, that is
$$
\mathcal P_{\mathcal F_\beta}(X)_{i.} = \frac{X_{i.}}{\|X_{i.}\|} \cdot \min\left\{\|X_{i.}\|,\frac{\|A_{i.}\|}{\beta} \right\},
$$
and
$$
\mathcal P_{\mathcal F_\beta}(Y)_{j.} = \frac{Y_{j.}}{\|Y_{j.}\|} \cdot \min\left\{\|Y_{j.}\|,\frac{\|A_{.j}\|}{\beta} \right\}.
$$
\begin{algorithm}[H]
\begin{algorithmic}[1]
\caption{Projected Gradient Descent}\label{alg1}
\REQUIRE Step size $\eta$, number of iterations $T$, tuning parameter $\beta$
\STATE \textbf{Initialization}: $X_0=U_0\Sigma_0^{1/2}$, $Y_0=V_0\Sigma_0^{1/2}$, where $[\mathcal P_\Omega(A)]_r =U_0\Sigma_0V_0^\top$ is its SVD
\STATE \textbf{for}: $t=0,1,2,\dots,T-1$
\STATE $\qquad X_{t+1} = X_t-\eta\nabla_X f(X_t,Y_t)$, $Y_{t+1} = Y_t-\eta\nabla_Y f(X_t,Y_t)$
\STATE $\qquad X_{t+1} = \mathcal P_{\mathcal F_\beta}(X_{t+1})$, $Y_{t+1} = \mathcal P_{\mathcal F_\beta}(Y_{t+1})$
\STATE \textbf{end for}
\ENSURE $(X_T,Y_T)$
\end{algorithmic}
\end{algorithm}
The intuition behind our approach is as follows. Write $A=A_r+N_r$. In light of the discussion from the previous subsection, we may be able to exactly recover $A_r$ from $\mathcal P_\Omega(A_r)$. However, what we have is $\mathcal P_\Omega(A) = \mathcal P_\Omega(A_r) + \mathcal P_\Omega(N_r)$, with an extra term $\mathcal P_\Omega(N_r)$, and we now hope to be able to control the error bound for $\widehat{A}_r:=X_TY_T^{\top}$ by the spectral error $\left\|\mathcal P_\Omega (N_r)-N_r\right\|$. The main advantage of our scheme is that we have tighter control of the perturbation $\left\|\mathcal P_\Omega(N_r)-N_r\right\|$ than $\left\|\mathcal P_\Omega(A)-A\right\|$ used in Theorem \ref{th:old}. Denote by
$$
\mu_r(A) = \max_{\substack{1\le i \le d_1\\1\le j \le d_2}}\left\{{\|(A_r)_{i\cdot}\|\over \|A_{i\cdot}\|},{\|(A_r)_{\cdot j}\|\over \|A_{\cdot j}\|}\right\},
$$
$$
\nu_r(A)^2=1-\min_{\substack{1\le i \le d_1\\1\le j \le d_2}}\left\{{\|(A_r)_{i\cdot}\|^2\over \|A_{i\cdot}\|^2},{\|(A_r)_{\cdot j}\|^2\over \|A_{\cdot j}\|^2}\right\},
$$
and
$$
\nu_r(A)_\infty = \max_{\substack{1\le i \le d_1\\1\le j \le d_2}}\left\{{|a_{ij}-(a_r)_{ij}|\over |a_{ij}| + (\|A_{i \cdot}\|^2 + \|A_{\cdot j}\|^2)/\|A\|_{\rm F}}\right\}.$$
We have
\begin{lemma}
\label{le:remainder}
Assume that each entry of $\Omega$ is independently sampled from binomial trails with probability given by \eqref{eq:defp}. If $\sigma_r(A)>\sigma_{r+1}(A)$, then there exist numerical constants $C_0,C_1>0$, with probability at least $1-d^{-\alpha}$, such that
\begin{equation}
\label{remainder:bound1}
\left\|\mathcal P_\Omega(N_r)-N_r\right\|\le C_0\left(\nu_r(A)+ \nu_r(A)_\infty\sqrt{d\log d\over n}\right)\|A\|_{\rm F}\sqrt{d\over n}.
\end{equation}
Further, an upper bound of $\nu_r(A)_\infty$ gives us
\begin{equation}
\label{remainder:bound2}
\left\|\mathcal P_\Omega(N_r)-N_r\right\|\le C_1\max \left\{ \nu_r(A),\left(1+{\mu_r(A)^2\|A\|_{\rm F}\over\sigma_r(A)}\right)\sqrt{d\log d\over n}\right\}\|A\|_{\rm F}\sqrt{d\over n}.
\end{equation}
\end{lemma}
It is worth pointing out that the sampling probability, and consequently $\Omega$ is determined by $A$ instead of $N_r$. As a result, $\mathcal P_\Omega(N_r)-N_r$ is not determined by $N_r$ alone and the error bound above depends also on properties of $A$. Compared with \eqref{specerr}, for sufficiently large $n$, $\nu_r(A)$ describes how much tighter control we can have for $\left\|\mathcal P_\Omega(N_r)-N_r\right\|$ than $\left\|\mathcal P_\Omega(A)-A\right\|$. By definition, $ \nu_r(A) \leq 1$, this means $\mathcal P_\Omega(N_r)$ concentrates around its mean more tightly than $\mathcal P_\Omega(A)$. We are now in the position to present our main theoretical guarantee which shows that the quality of recovery indeed rests upon the perturbation of $\mathcal P_\Omega(N_r)$, not $\mathcal P_\Omega(A)$.
To this end, write $X^* = U_r\Sigma_r^{1/2}, Y^* = V_r\Sigma^{1/2}_r$ and denote by
$$F_t = \begin{bmatrix}
X_t \\Y_t
\end{bmatrix},\quad F^* = \begin{bmatrix}
X^* \\Y^*
\end{bmatrix},$$
where $A_r= U_r\Sigma_r V_r^\top$ is its SVD. Due to rotation symmetry, the difference between $F_t$ ad $F^\ast$ can be measured by
$$\delta_t = \min_{RR^\top = R^\top R = I} \left\|F_t - F^*R \right\|_\textup{F}.$$
We have
\begin{theorem}
\label{th:main}
Assume $\sigma_r(A)>\sigma_{r+1}(A)$ and each entry of $\Omega$ is independently sampled with probability given by \eqref{eq:defp}, denote by $(X_t,Y_t)$ the iteration sequence from Algorithm \ref{alg1} with $\eta \leq (\sigma_r(A)-\sigma_{r+1}(A))\beta^4(1-\nu_r(A)^2)^2/(500\|A_r\|_{\rm F}^4)$. Then there exist numerical constants $C_0, C_1>0$ such that
$$\delta_{t} \leq \left[ 1-\frac{1}{5}\eta\left(\sigma_r(A)-\sigma_{r+1}(A)\right) \right]^{t/2}\cdot \delta_{0} +
C_0 \sqrt{r}\left\|\mathcal P_\Omega(N_r)-N_r\right\| \cdot \frac{\sqrt{\sigma_r(A)}}{\sigma_r(A)-\sigma_{r+1}(A)},$$
with probability at least $1-d^{-\alpha}$ provided that $\beta \leq \sqrt{\sigma_r(A)/(1-\nu_r(A)^2)}$ and
\begin{eqnarray*}
n&\ge& C_1(1+\alpha)d\log d \cdot {\|A\|_{\rm F}^2\over(\sigma_r(A)-\sigma_{r+1}(A))^4}\\
&&\cdot \max\left\{ \nu_r(A)^2r\sigma_1(A)\sigma_r(A),{r\sigma_1(A)^2\sigma_{r+1}(A)\over\sigma_r(A)},{\sigma_r^2(A)\over \beta^4(1-\nu_r(A))^2} \mu_r^2(A) \|A_r\|_{\rm F}^2\right\},
\end{eqnarray*}
where $d=\max\{d_1,d_2\}$. Moreover,
\begin{eqnarray*}
\|X_tY_t^\top-A_r\|_{\rm F}\le \sqrt{3\sigma_1(A)}\left[ 1-\frac{1}{5}\eta\left(\sigma_r(A)-\sigma_{r+1}(A)\right) \right]^{t/2}\cdot \delta_{0}\\
+C_0 \sqrt{r}\left\|\mathcal P_\Omega(N_r)-N_r\right\|\cdot \frac{\sigma_r(A)}{\sigma_r(A)-\sigma_{r+1}(A)}.
\end{eqnarray*}
\end{theorem}
A few remarks follow immediately. A benchmark case is when $A$ is of low rank. More specifically if $\mathrm{rank}(A) \le r$, then $A_r = A$ and $\nu_r(A) =\nu_r(A)_\infty =0$. From Theorem \ref{th:main} and \eqref{remainder:bound1},
$$\delta_{t} \leq \left[ 1-\frac{1}{5}\eta\sigma_r(A) \right]^{t/2}\cdot \delta_{0},$$
which implies that $F_t\to F^\ast$ and consequently $X_t{Y_t}^\top \to A$ as $t\to \infty$. It is worth pointing out that if $\mathrm{rank}(A) < r$, the $\sigma_r(A)$ in the above inequality can be replaced by $\sigma_{\mathrm{rank}(A)}(A)$. In other words, the exact recovery of $A$ can be achieved without either the incoherence condition or knowing the precise value of $\mathrm{rank}(A)$ apriori. Furthermore, if the incoherence condition indeed holds, the sample size requirement from Theorem \ref{th:main} is comparable to those typical in the matrix completion literature.
In general, Theorem \ref{th:main} and \eqref{remainder:bound2} suggests that whenever $\sigma_r(A)>\sigma_{r+1}(A)$,
\begin{eqnarray}
\label{newerr}
\|X_tY_t^\top-A_r\|_{\rm F}&\le& C_0\|A\|_{\rm F}\sqrt{rd\over n}\cdot \frac{\sigma_r(A)}{\sigma_r(A)-\sigma_{r+1}(A)} \nonumber\\
&& \cdot\max \left\{ \nu_r(A),\left(1+{\mu_r(A)^2\|A\|_{\rm F}\over\sigma_r(A)}\right)\sqrt{d\log d\over n}\right\},
\end{eqnarray}
for large enough $t$. This is to be compared with the bound given by Theorem \ref{th:old}:
$$
\|[\mathcal P_\Omega(A)]_r-A_r\|_{\rm F}\le C_0\|A\|_{\rm F}\sqrt{rd\over n}\left(1+{\sqrt{\sigma_{r+1}(A)\sigma_1(A)}\over \sigma_r(A)-\sigma_{r+1}(A)}\right).
$$
It is clear that the bound in \eqref{newerr} is smaller because
$$\nu_r(A)\le1,\quad \mu_r(A)\le1,\quad n \gtrsim d\log d\|A\|_{\rm F}^2/\sigma_r(A)^2,$$
and
\begin{eqnarray*}
1+{\sqrt{\sigma_{r+1}(A)\sigma_1(A)}\over \sigma_r(A)-\sigma_{r+1}(A)}&=&\frac{\sigma_r(A)}{\sigma_r(A)-\sigma_{r+1}(A)}+{\sqrt{\sigma_{r+1}(A)\sigma_1(A)}-\sigma_{r+1}(A)\over \sigma_r(A)-\sigma_{r+1}(A)}\\
&>&\frac{\sigma_r(A)}{\sigma_r(A)-\sigma_{r+1}(A)}.
\end{eqnarray*}
To gain further insights of the difference between the two estimates of $A_r$, we shall now take a closer look at these factors.
For brevity, we shall assume $d_1=d_2=d$ for the discussion. Recall that $A=UDV^\top$. It is not hard to see that
$$\nu_r(A)^2 = \max_{1\le j\le d} \left\{ \frac{\sum_{i=r+1}^d\sigma_i^2 u_{ij}^2}{\sum_{i=1}^d\sigma_i^2 u_{ij}^2}, \frac{\sum_{i=r+1}^d\sigma_i^2 v_{ij}^2}{ \sum_{i=1}^d\sigma_i^2 v_{ij}^2} \right\}.$$
If the entries of $U$ and $V$ are of similar order, then
$$\nu_r(A)^2\approx \sum_{i=r+1}^d\sigma_i^2/\sum_{i=1}^d\sigma_i^2=\|A-A_r\|_{\rm F}^2/\|A\|^2_{\rm F}.$$
Similarly, $\mu_r(A) \approx d\|A-A_r\|_\infty/\|A\|_{\rm F}$ so that, from \eqref{newerr}, there is more improvement over the na\"ive estimate $[\mathcal P_\Omega(A)]_r$ when $A$ is close to being of rank $r$.
On the other hand, it is instructive to consider the case when $\sigma_i(A)= i^{-\alpha}$. In this case,
$$
\frac{\sigma_r(A)}{\sigma_r(A)-\sigma_{r+1}(A)}\sim r,
$$
and on the other hand,
$$
1+{\sqrt{\sigma_{r+1}(A)\sigma_1(A)}\over \sigma_r(A)-\sigma_{r+1}(A)}\sim r^{\alpha/2+1}.
$$
This implies that much better bound can be achieved by our approach as $r$ and $\alpha$ increase.
\section{Numerical Experiments}
To complement the theoretical developments from the previous sections and further demonstrate the merits of our approach, we also conducted several sets of numerical experiments. In particular, we focused on several key operating characteristics of our method including the ability to recover low-rank matrices exactly; the impact of the target rank $r$; and the role of the eigengap $\sigma_r(A)-\sigma_{r+1}(A)$. We consider estimating $A_r$ by the na\"ive estimate $[\mathcal P_\Omega(A)]_r$ with $\Omega$ generated according to \eqref{p:sparse}, and by $\widehat{A}_r:=X_TY^\top_T$ as the output from Algorithm 1 with $\Omega$ generated according to \eqref{eq:defp}. For a fair comparison, we adjusted $n$ of the sampling probability \eqref{p:sparse} (for na\"ive method) and \eqref{eq:defp} (for the improved estimate) to ensure that both methods sample the same number of entries on average. In our implementation of the projected gradient descent, $\beta$ was set to be $\sqrt{\sigma_r(\mathcal P_\Omega(A))/2}$. To accelerate the optimization, we used line search (the line\_search function of scipy.optimize in Python) to determine the optimal stepsize. The stopping criterion was set to be $\|\nabla f(X,Y)\|_\infty \leq 10^{-6}$. In what follows, we shall refer to the ``relative error'' as $\|\widehat{A}_r-A_r\|_\textup{F}/\|A_r\|_\textup{F}$ or $\|[\mathcal P_\Omega(A)]_r-A_r\|_\textup{F}/\|A_r\|_\textup{F}$, and the ``spectral error'' as $\|\widehat{A}_r-A_r\|/\|A_r\|$ or $\|[\mathcal P_\Omega(A)]_r-A_r\|/\|A_r\|$. All the results presented are based on 100 independent simulation runs.
In our first set of experiments, we set $d_1 = d_2 =1000$ and generated $A$ as a low-rank matrix $A^*$ with additive noise $\sigma E$. More specifically,
$$A^* = U^*\Sigma^* {U^*}^\top/\|U^*\|_\textup{F}^2,\qquad \Sigma^* = \text{diag}\{1,0.9,0.8,0.7,0.6\},$$
where $U^\ast\in \mathbb R^{1000\times 5}$ is a Gaussian ensemble with entries independently sampled from the standard normal distribution, and $E$'s entries are independently from $N(0,1/4000)$. The variance of $E_{ij}$ is chosen such that $\|E\| \approx 1$.
We began with the case where $A$ is of low rank. More specifically, we set $A = A^*$ and $r=5$. In Figure \ref{fig:low rank}, we plotted the relative error of $[\mathcal P_\Omega(A)]_r$ and $\widehat{A}_r$ with the max number of iterations set to be $T=5$ or $10$. It is clear that the relative error of both estimates decreases quickly with the increased sample size, as predicted by our theoretical results. But the relative error of $\widehat{A}_r$ decreases quickly to 0 while the error of the na\"ive estimate levels off around 0.15 even with as many as 50\% of the entries observed. This highlights the ability of exact recovery for the improved approach as discussed in the previous section. Empirically, the projected gradient descent algorithm converges fairly quickly: there is little difference between setting the max number of iterations to 5 and 10, and 10 seems to suffice as a rule of thumb. In the rest of the experiments, we shall fix the max number of iterations at ten for consideration of computational speed.
\begin{figure}
\centering
\includegraphics[width=0.6\textwidth]{pre_no_noise_err_pct_obs.png}
\caption{``Relative error'' of different methods under low-rank scenario, averaged over 100 simulation runs. }
\label{fig:low rank}
\end{figure}
Next, we consider the case where $A$ is full rank yet we are primarily interested in its best rank approximations. To this end, we set $A = A^* + 0.05E$. We adjusted $n$ of the sampling probability so that on average 10\% of the entries were sampled. Figure \ref{fig: target rank} reports the mean and two-standard-deviance bands of both relative spectral and Frobenius errors from 100 simulation runs. A few interesting observations can be made. In particular, it is evident that $\widehat{A}_r$ is superior to the na\"ive estimator, in either error metric. Moreover, the na\"ive estimate is much more vulnerable to overshooting the ``effective'' rank of $A$. Note that $A$ is ``close'' to being of rank $5$. When consider estimating $A_r$ for $r>5$, the performance of $\widehat{A}_r$ only deteriorates mildly with an increasing $r$ yet on the other hand, for the na\"ive estimate the impact is much more significant
\captionsetup[subfigure]{labelfont=rm}
\begin{figure}
\centering
\begin{subfigure}{0.45\textwidth}
\includegraphics[width=\textwidth]{pre_0_05_noise_spectral_rank_maxiter_10.png}
\end{subfigure}
\begin{subfigure}{0.45\textwidth}
\includegraphics[width=\textwidth]{pre_0_05_noise_err_rank_maxiter_10.png}
\end{subfigure}
\caption{Effect of the targeted rank $r$ on different methods: averaged errors and $\pm$ two-standard-error bands were based on 100 simulation runs.}
\label{fig: target rank}
\end{figure}
To further investigate the impact of the eigengap $\sigma_r(A)-\sigma_{r+1}(A)$, we considered $A = A^*+ \sigma E$ with $\sigma$ varying from 0.02 to 0.4. Here $\sigma$ serves as a proxy of the relative eigengap $1-\sigma_6(A)/\sigma_5(A)$ as $\|E\| \approx 1$. The results, again based upon 100 simulation runs, were summarized in Figure \ref{fig:eigengap}. It is interesting to note that the error increases as the relative eigengap decreases for both methods, but the impact on $\widehat{A}_r$ is minimal when compared with the na\"ive method, especially with increased sampling proportions.
\captionsetup[subfigure]{labelfont=rm}
\begin{figure}
\centering
\begin{subfigure}{0.45\textwidth}
\includegraphics[width=\textwidth]{pre_noise_10_pct_obs_spectral_sigma_maxiter_10.png}
\caption{10\% observations}
\end{subfigure}
\begin{subfigure}{0.45\textwidth}
\includegraphics[width=\textwidth]{pre_noise_10_pct_obs_err_sigma_maxiter_10.png}
\caption{10\% observations}
\end{subfigure}
\vfill
\begin{subfigure}{0.45\textwidth}
\includegraphics[width=\textwidth]{pre_noise_40_pct_obs_spectral_sigma_maxiter_10.png}
\caption{40\% observations}
\end{subfigure}
\begin{subfigure}{0.45\textwidth}
\includegraphics[width=\textwidth]{pre_noise_40_pct_obs_err_sigma_maxiter_10.png}
\caption{40\% observations}
\end{subfigure}
\caption{Effect of eigengap and sampling proportion on different methods: averaged errors and $\pm$ two-standard-error bands were based on 100 simulation runs.}
\label{fig:eigengap}
\end{figure}
In our final example, we computed both estimators to the waterfall image \footnote{The original color version of image can be downloaded at \url{https://media.cntraveler.com/photos/571945e380cf3e034f974b7d/master/pass/waterfalls-Seljalandsfoss-GettyImages-457381095.jpg}.}. The original image was converted into greyscale.
The leading singular value of the $1536 \times2056$ matrix consisting of the pixel intensities accounts for 57.5\% of the total variation. The best rank-10, rank-20, and rank-30 approximations explain 88.1\%, 91.6\%, and 93.0\% of the variation respectively. We considered estimating the best rank-$r$ approximation to the image with $r=5, 10$, $20$, and $30$. For each $r$, we set the max number of iteration to be 10 and sampled 10\% pixels. Table \ref{table:waterfall} reports the mean and standard deviation of the ``relative error'' from 100 simulation runs. They again confirm that $\widehat{A}_r$ is a far more accurate estimate of $A_r$.
\begin{table}
\centering
\begin{tabular}{|c|c|c|c|c|}
\hline
Target rank $r$& Na\"ive Estimate &Improved Estimate \\
\hline
5 & 0.290 (0.0017)& 0.118 (0.0036) \\
10 & 0.432 (0.0012)& 0.154 (0.0018)\\
20& 0.609 (0.0009)& 0.174 (0.0012)\\
30&0.738 (0.0009)& 0.184 (0.0010)\\
\hline
\end{tabular}
\caption{Comparison between the na\"ive and the improved estimates of the best rank-$r$ approximations of the greyscale waterfall image for different targeted ranks. Reported are averaged ``relative error'' over 100 runs and numbers in parentheses are standard deviations.}
\label{table:waterfall}
\end{table}
To facilitate visual comparison between the improved and na\"ive methods, we now focus on the best rank-$30$ approximation of the image as shown in panel (a) of Figure \ref{fig:waterfall}. We fixed $n$ in the sampling probabilities so that the expected sampling proportion is 40\%. One typical realization of both estimates is given in Panels (b) and (c) of Figure \ref{fig:waterfall}. For this specific realization, $\widehat{A}_r$ has a relative error 0.11. This is to be compared with the na\"ive method which has a relative error 0.37, which again is in agreement with our theoretical findings.
\captionsetup[subfigure]{labelfont=rm}
\begin{figure}
\centering
\begin{subfigure}{0.3\textwidth}
\includegraphics[width=\textwidth]{water_30.png}
\caption{Exact approximation}
\end{subfigure}
\begin{subfigure}{0.3\textwidth}
\includegraphics[width=\textwidth]{water_benchmark_40pct_r30.png}
\caption{Na\"ive Method}
\end{subfigure}
\begin{subfigure}{0.3\textwidth}
\includegraphics[width=\textwidth]{water_our_40pct_r30.png}
\caption{Improved Method}
\end{subfigure}
\caption{The exact best rank-30 approximation of the original greyscale image is given in panel (a). Panels (b) and (c) are the output by the na\"ive method and the improved method.}
\label{fig:waterfall}
\end{figure}
\section{Proofs}
\begin{proof}
[Proof of Lemma \ref{le:lowrankappr}]
By definition,
$$
\|B_r-B\|_{\rm F}^2\le \inf_{\mathrm{rank}(M)\le r}\|M-B\|_{\rm F}^2.
$$
Observe that
$$
\|M-B\|_{\rm F}^2=\|M-A\|_{\rm F}^2+\|E\|_{\rm F}^2-2\langle E, M-A\rangle.
$$
We get
$$
\|B_r-A\|_{\rm F}^2\le \|M-A\|_{\rm F}^2+2\langle E, B_r-M\rangle\le \|M-A\|_{\rm F}^2+2\|E\|\|B_r-M\|_\ast,
$$
where we replaced $A$ with $B-E$ in the first inequality. Recall that
$$
\mathrm{rank}(B_r-M)\le \mathrm{rank}(B_r)+\mathrm{rank}(M)\le 2r.
$$
By Cauchy-Schwartz inequality,
$$
\|B_r-M\|_\ast\le \sqrt{2r}\|B_r-M\|_{\rm F}.
$$
Therefore,
$$
\|B_r-A\|_{\rm F}^2\le \|M-A\|_{\rm F}^2+2\sqrt{2r}\|E\|\|B_r-M\|_{\rm F}.
$$
In the case when $\mathrm{rank}(A)\le r$, we can take $M=A$ to yield
$$
\|B_r-A\|_{\rm F}^2\le 8r\|E\|^2.
$$
Now consider the case when $\mathrm{rank}(A)>r$. Taking $M=A_r$ yields
$$
\|B_r-A\|_{\rm F}^2\le \|A_r-A\|_{\rm F}^2+2\sqrt{2r}\|E\|\|B_r-A_r\|_{\rm F}.
$$
Observe that
$$
\|B_r-A\|_{\rm F}^2=\|B_r-A_r\|_{\rm F}^2+\|A_r-A\|_{\rm F}^2+2\langle B_r, A_r-A\rangle.
$$
We have
$$
\|B_r-A_r\|_{\rm F}^2\le 2\sqrt{2r}\|E\|\|B_r-A_r\|_{\rm F}+2\langle B_r, A-A_r\rangle.
$$
Let $A_r=U_r\Sigma_r V_r^\top$ and $B_r=\tilde{U}_r\tilde{\Sigma}_r \tilde{V}_r^\top$ be their respective singular value decomposition. Then
\begin{eqnarray*}
\langle B_r, A-A_r\rangle&=&\langle P_{U_r}^\perp B_rP_{V_r}^\perp, A-A_r\rangle\\
&\le& \sigma_{r+1}(A)\|P_{U_r}^\perp B_rP_{V_r}^\perp\|_\ast\\
&\le& r\sigma_{r+1}(A)\|P_{U_r}^\perp P_{\tilde{U}_r} B_rP_{\tilde{V}_r}P_{V_r}^\perp\|\\
&\le& r\sigma_{r+1}(A)\sigma_1(B)\|P_{U_r}^\perp P_{\tilde{U}_r}\|\|P_{\tilde{V}_r}P_{V_r}^\perp\|.
\end{eqnarray*}
By Davis-Kahan-Wedin's Theorems,
$$
\max\{\|P_{U_r}^\perp P_{\tilde{U}_r}\|,\|P_{\tilde{V}_r}P_{V_r}^\perp\|\}\le {4\|E\|\over \sigma_r(A)-\sigma_{r+1}(A)},
$$
$$|\sigma_1(A)-\sigma_1(B)| \le \|E\| < \sigma_1(A), $$
so that
$$
\langle B_r, A-A_r\rangle\le {16r\sigma_{r+1}(A)\sigma_1(B)\|E\|^2\over [\sigma_r(A)-\sigma_{r+1}(A)]^2}\le {32r\sigma_{r+1}(A)\sigma_1(A)\|E\|^2\over [\sigma_r(A)-\sigma_{r+1}(A)]^2}.
$$
This implies that
$$
\|B_r-A_r\|_{\rm F}\le 8\sqrt{r}\|E\| \left({1\over \sqrt{2}}+{\sqrt{\sigma_{r+1}(A)\sigma_1(A)}\over \sigma_r(A)-\sigma_{r+1}(A)}\right).
$$
\end{proof}
\begin{proof}[Proof of Lemma \ref{le:remainder}]
The proof relies on the following concentration bound.
\begin{lemma}
\label{le:concentration}
Let $B\in \mathbb R^{d_1\times d_2}$ be a fixed matrix and each entry of $\Omega$ is independently sampled from binomial trails with probability given by \eqref{eq:defp}. Then there exists a numerical constant $C_0>0$ such that
$$
\left\|\mathcal P_\Omega(B)-B\right\|\le 4\|A\|_{\rm F}\sqrt{d\over n}\max_{\substack{1\le i\le d_1\\1\le j \le d_2}}\left\{{\| B_{i\cdot}\|\over \|A_{i\cdot}\|},{\|B_{\cdot j}\|\over \|A_{\cdot j}\|}\right\}+ C_0\max_{\substack{i,j:\\ 0<p_{ij}<1}}\left\{ |b_{ij}|\over p_{ij}\right\}\sqrt{\log (d)+t},
$$
with probability at least $1-e^{-t}$, where $d=\max\{d_1,d_2\}$.
\end{lemma}
In particular, consider applying Lemma \ref{le:concentration} to $B=A-A_r$. It is not hard to see that
$$
{\|B_{i\cdot}\|^2\over \|A_{i\cdot}\|^2}=1-{\|(A_r)_{i\cdot}\|^2\over \|A_{i\cdot}\|^2} \le \nu_r(A)^2,
$$
and
$$
{\|B_{\cdot j}\|^2\over \|A_{\cdot j}\|^2}=1-{\|(A_r)_{\cdot j}\|^2\over \|A_{\cdot j}\|^2} \le \nu_r(A)^2.
$$
By definition of $\nu_r(A)_\infty$, we get
\begin{eqnarray*}
&&\max_{(i,j): 1> p_{ij}>0}\left\{|b_{ij}|\over p_{ij}\right\}\\
&\le& \nu_r(A)_\infty\cdot\max_{(i,j): 1> p_{ij}>0}\left\{|a_{ij}| + (\|A_{i \cdot}\|^2+\|A_{\cdot j}\|^2)/\|A\|_{\rm F}\over p_{ij}\right\}\\
&\leq &\nu_r(A)_\infty \cdot \max\left\{{3\|A\|_{\ell_1}\over n},{3d\|A\|_{\rm F}\over n} \right\}\\
&\leq &\nu_r(A)_\infty \cdot {3d\|A\|_{\rm F}\over n} ,
\end{eqnarray*}
where we used the fact that $\|A\|_{\ell_1} \le d\|A\|_{\rm F}$ and thus the first claim follows.
For the second claim, let $A = U\Sigma V^\top$ be its singular value decomposition,
\begin{eqnarray*}
| b_{ij}|&=&\left|\sum_{k>r}\sigma_k(A)u_{ik}v_{jk}\right|\\
&\le& |a_{ij}|+\left|\sum_{k\le r}\sigma_k(A)u_{ik}v_{jk}\right|\\
&\le& |a_{ij}|+\left(\sum_{k\le r}\sigma_k(A)u_{ik}^2\right)^{1/2}\left(\sum_{k\le r}\sigma_k(A)v_{jk}^2\right)^{1/2}\\
&\le& |a_{ij}|+{1\over \sigma_r(A)}\left(\sum_{k\le r}\sigma_k(A)^2u_{ik}^2\right)^{1/2}\left(\sum_{k\le r}\sigma_k(A)^2v_{jk}^2\right)^{1/2}\\
&=& |a_{ij}|+ {1\over \sigma_r(A)}\|(A_r)_{ i\cdot}\|\|(A_r)_{ \cdot j}\|.
\end{eqnarray*}
Hence,
$$
\nu_r(A)_\infty \le 1 + {\|A\|_{\rm F}\over \sigma_r(A)}\max\left\{{\|(A_r)_{i\cdot}\|^2\over \|A_{i\cdot}\|^2},{\|(A_r)_{\cdot j}\|^2\over \|A_{\cdot j}\|^2}\right\}.
$$
This implies that
\begin{eqnarray*}
&&\left\|\mathcal P_\Omega(A-A_r)-(A-A_r)\right\| \\
&\le&C_0\left(1+{\mu_r(A)^2\|A\|_{\rm F}\over \sigma_r(A)}\right){d\|A\|_{\rm F}\sqrt{\log d+t}\over n}\\
&&+4\nu_r(A)\|A\|_{\rm F}\sqrt{d\over n},
\end{eqnarray*}
with probability at least $1-e^{-t}$.
\end{proof}
\begin{proof}[Proof of Theorem \ref{th:main}]
In the rest of the proof, we shall omit $A$ in $\sigma_r(A)$ for brevity. Let
$$R_t = \argmin_{RR^\top = R^\top R = I} \left\|F_t - F^*R \right\|_\textup{F}.$$
We shall then write $S_t = X^*R_t$, and $T_t = Y^*R_t$. Moreover, denote by
$$\Delta_{X_t} =X_t- S_t,\qquad \Delta_{Y_t} =Y_t- T_t,\qquad \Delta_{F_t} =F_t- F^*R_t.$$
Our analysis relies on the follow two technical lemmas.
\begin{lemma}\label{b1}
Assume that $\beta \leq \sqrt{\sigma_r/(1-\nu_r(A)^2)}$ and $(X_t,Y_t)\in \mathcal F_\beta$ such that
$$\|\Delta_{F_t}\|_{\rm F} \leq \frac{\sigma_r-\sigma_{r+1}}{20\sqrt{\sigma_1}}.$$
There exists a numerical constant $C_1>0$ such that, with probability at least $1-d^{-\alpha}$, if
\begin{equation}
\label{eq:b1cond}
\max\left\{\|X_tY_t^\top-A_r\|_{\rm F},\sqrt{\sigma_r}\| \Delta_{F_t}\|_{\rm F}\right\}>\frac{31\sqrt{r}}{10(1-\sigma_{r+1}/\sigma_r)}\left\|\mathcal P_\Omega(N_r)-N_r\right\|,
\end{equation}
then
\begin{align*}
\langle \nabla_{X} f(X_t,Y_t),\Delta_{X_t}\rangle + \langle \nabla_Y f(X_t,Y_t),\Delta_{Y_t}\rangle
\geq {19(\sigma_r-\sigma_{r+1})\over 100}\|\Delta_{F_t}\|_{\rm F}^2;
\end{align*}
and otherwise,
\begin{align*}
\langle \nabla_{X} f(X_t,Y_t),\Delta_{X_t}\rangle + \langle \nabla_Y f(X_t,Y_t),\Delta_{Y_t}\rangle
\geq \frac{69(\sigma_r-\sigma_{r+1})}{100}\|\Delta_{F_t}\|_{\rm F}^2 - {5r\| \mathcal P_\Omega(N_r)-N_r\|^2\over (1-\sigma_{r+1}/\sigma_r)},
\end{align*}
provide that $n \geq C_1(1+\alpha)d(\log d)\sigma_r^2/\left[\beta^4(1-\nu_r(A)^2)^2\right] \mu_r^2(A)\|A_r\|_{\rm F}^2 \|A\|_{\rm F}^2/(\sigma_r-\sigma_{r+1})^4$.
\end{lemma}
Lemma \ref{b1} verifies the so-called local descent condition from \cite{chen2015fast}. It states that, if $\|\Delta_{F_t}\|_\textup{F} \geq O(\| \mathcal P_\Omega(N_r)-N_r\|)$, $\nabla f(X_t,Y_t)$ will have an acute angle with $\Delta_{F_t}$ and $f(X_t,Y_t)$ shows similar behavior to a convex function. It implies that if $\nabla f \approx 0$, then it is necessarily true that
$$\|\Delta_{F_t}\|_{\rm F}=O\left(\left\|\mathcal P_\Omega(N_r)-N_r\right\|\right),$$
so that, to bound $\|\Delta_{F_t}\|_\textup{F}$, it suffices to do so for $\| \mathcal P_\Omega(N_r)-N_r\|$.
\begin{lemma}\label{b2}
Under the conditions in Lemma \ref{b1}, there exists a numerical constant $C_1>0$ such that
\begin{align*}
\| \nabla_X f(X_t,Y_t)\|_{\rm F}^2 + \| \nabla_Y f(X_t,Y_t)\|_{\rm F}^2\leq \frac{81\|A_r\|_{\rm F}^4}{\beta^4(1-\nu_r(A)^2)^2}\|\Delta_{F_t}\|_\textup{F}^2 +18r\sigma_1\| \mathcal P_\Omega(N_r)-N_r\|^2,
\end{align*}
with probability at least $1-d^{-\alpha}$ provide $n \geq C_1\alpha d\log d\|A\|_{\rm F}^2/\|A_r\|_{\rm F}^2$.
\end{lemma}
We shall now use these lemmas to prove that for any $t$
\begin{equation}
\label{inducta}
\|\Delta_{F_t}\|_{\rm F} \leq \frac{\sigma_r-\sigma_{r+1}}{20\sqrt{\sigma_1}},
\end{equation}
and
\begin{equation}
\label{inductc}
\delta_t
\leq
\left[ 1-\frac{1}{5}\eta\left(\sigma_r-\sigma_{r+1}\right) \right]^{1/2}\cdot \delta_{t-1}+\frac{4\sqrt{r /\sigma_r}}{1-\sigma_{r+1}/\sigma_r}\left\|\mathcal P_\Omega(N_r)-N_r\right\|.
\end{equation}
We shall do so by induction. In fact, when $t=0$, it suffices to verify \eqref{inducta}. In light of Lemma \ref{le:remainder} and Lemma \ref{le:lowrankappr}, we have
$$\left\| X_0Y_0^\top -A_r \right\|_\textup{F} \leq O\left(\frac{\sigma_r-\sigma_{r+1}}{\sqrt{\sigma_1/\sigma_r}}\right),$$
with probability at least $1-d^{-\alpha}$ provided that
$$n\ge C_1(1+\alpha)rd(\log d)\cdot{(\sigma_1/\sigma_r)\|A\|_{\rm F}^2\over (\sigma_r-\sigma_{r+1})^2}\cdot\left(1+{\sigma_{r+1}\sigma_1\over (\sigma_r-\sigma_{r+1})^2}\right).$$
Then, \eqref{inducta} follows immediately from Lemma 6 and Lemma 42 of \cite{ge2017no}.
Now assume that \eqref{inducta} and \eqref{inductc} hold for $t$, we show that the same is true for $t+1$. We shall first verify \eqref{inductc}. Because the projection $\mathcal P_{\mathcal F_\beta}$ is contractive, we have
\begin{eqnarray*}
\|F_{t+1}-F^*R_t\|_\textup{F}^2 & \le& \|X_t-\eta\nabla_X f(X_t,Y_t)-S_t\|_\textup{F}^2 + \|Y_t-\eta\nabla_Y f(X_t,Y_t)-T_t\|_\textup{F}^2\\
&=& \|\Delta_{F_t}\|_\textup{F}^2 +\eta^2(\| \nabla_X f\|_{\rm F}^2 + \| \nabla_Y f\|_{\rm F}^2) - 2\eta \langle \nabla_{X} f(X_t,Y_t),\Delta_{X_t}\rangle \\
&&\hskip50pt -2\eta \langle \nabla_Y f(X_t,Y_t),\Delta_{Y_t}\rangle \\
&\leq& \|\Delta_{F_t}\|_\textup{F}^2 +{ 81\eta^2\|A_r\|_{\rm F}^4 \over \beta^4(1-\nu_r(A)^2)^2}\|\Delta_{F_t}\|_\textup{F}^2 +18\eta^2r\sigma_1\| \mathcal P_\Omega(N_r)-N_r\|^2\\
&&\hskip50pt - 2\eta \langle \nabla_{X} f(X_t,Y_t),\Delta_{X_t}\rangle -2\eta \langle \nabla_Y f(X_t,Y_t),\Delta_{Y_t}\rangle ,\label{b.4}
\end{eqnarray*}
where we used Lemma \ref{b2}.
By Lemma \ref{b1}, if \eqref{eq:b1cond} holds, then
\begin{eqnarray}
\delta_{t+1}^2 &\leq &\|F_{t+1}-F^*R_t\|_\textup{F}^2 \nonumber\\
&\leq& \|\Delta_{F_t}\|_\textup{F}^2 +{ 81\eta^2\|A_r\|_{\rm F}^4 \over \beta^4(1-\nu_r(A)^2)^2}\|\Delta_{F_t}\|_\textup{F}^2 \nonumber\\
&&+18\eta^2r\sigma_1\| \mathcal P_\Omega(N_r)-N_r\|^2 - \eta {19(\sigma_r-\sigma_{r+1})\over 50}\|\Delta_{F_t}\|_{\rm F}^2 \nonumber\\
&\leq& \left[ 1-\frac{1}{5}\eta(\sigma_r -\sigma_{r+1}) \right]\delta_t^2,\label{eq:main1}
\end{eqnarray}
where we used the condition that $\eta \leq (\sigma_r-\sigma_{r+1})\beta^4(1-\nu_r(A)^2)^2/(500\|A_r\|_{\rm F}^4)$.
On the other hand, if \eqref{eq:b1cond} does not hold, then by Lemma \ref{b1},
\begin{eqnarray*}
\delta_{t+1}^2 &\leq &\|F_{t+1}-F^*R_t\|_\textup{F}^2 \\
&\leq& \frac{31^2 \cdot r/\sigma_r \cdot \|\mathcal P_\Omega(N_r)-N_r\|^2}{10^2(1-\sigma_{r+1}/\sigma_r)^2} + \eta{5r\| \mathcal P_\Omega(N_r)-N_r\|^2\over (1-\sigma_{r+1}/\sigma_r)}\\
&&+{ 81\eta^2\|A_r\|_{\rm F}^4 \over \beta^4(1-\nu_r(A)^2)^2}\|\Delta_{F_t}\|_\textup{F}^2 +18\eta^2r\sigma_1\| \mathcal P_\Omega(N_r)-N_r\|^2\\
&&- \eta {69(\sigma_r-\sigma_{r+1})\over 50}\|\Delta_{F_t}\|_{\rm F}^2\\
&\leq& \frac{4^2 \cdot r/\sigma_r \cdot \|\mathcal P_\Omega(N_r)-N_r\|^2}{(1-\sigma_{r+1}/\sigma_r)^2},
\end{eqnarray*}
where we again used $\eta \leq (\sigma_r-\sigma_{r+1})\beta^4(1-\nu_r(A)^2)^2/(500\|A_r\|_{\rm F}^4)$.
In light of above two inequalities, we have,
$$\delta_{t+1} \le \max\left\{\delta_{t}, \frac{4\sqrt{r/\sigma_r}}{1-\sigma_{r+1}/\sigma_r}\cdot\|\mathcal P_\Omega(N_r)-N_r\|\right\}.$$
It is clear that \eqref{inducta} also continues to hold for $t+1$ in light of the inequality above and the Lemma \ref{le:remainder}. The first claim of Theorem \ref{th:main} follows immediately.
The second claim also follows, in light of the following bound:
\begin{lemma}
\label{le:apprest}
Under the assumptions of Theorem \ref{th:main},
$$
\|X_tY_t^\top-A_r\|_{\rm F}^2\le 3\sigma_1(A)\delta_t^2.
$$
\end{lemma}
If the second claim doesn't hold, then \eqref{eq:b1cond} holds and we have \eqref{eq:main1}. Combined with Lemma \ref{le:apprest}, there is a contradiction.
\end{proof}
\begin{proof}[Proof of Lemma \ref{b1}]
For brevity, we shall omit the subscript $t$ in what follows. Denote by $\kappa_r = \sigma_{r+1}/\sigma_r$. Observe that
$$
\nabla_X f = \mathcal P_{\Omega}\left(XY^\top - A\right)Y +{1\over 2}X(X^\top X - Y^\top Y),
$$
and
$$
\nabla_Y f= \left[\mathcal P_{\Omega}\left(XY^\top - A\right)\right]^\top X -{1\over 2}Y(X^\top X - Y^\top Y) .
$$
Therefore,
\begin{eqnarray}
&&\langle \nabla_X f,\Delta_X\rangle + \langle \nabla_Y f,\Delta_Y\rangle \nonumber\\
&=& \left\langle \mathcal P_\Omega\left(XY^\top-A\right)Y,\Delta_X\right\rangle + \left\langle \left[\mathcal P_\Omega\left(X Y^\top-A\right)\right]^\top X,\Delta_Y\right\rangle\nonumber\\
&& \hskip 75pt+ {1\over 2} \left\langle X^\top X - Y^\top Y ,X^\top(\Delta_X) - (\Delta_Y)^\top Y\right\rangle\nonumber\\
&=& \left\langle \mathcal P_\Omega \left(XY^\top-ST^\top\right),(\Delta_X)Y^\top+X(\Delta_Y)^\top\right\rangle \label{c.1} \\
&&\hskip 75pt + {1\over 2} \left\langle X^\top X - Y^\top Y ,X^\top(\Delta_X) - (\Delta_Y)^\top Y\right\rangle\label{c.3}\\
&& \hskip 75pt- \left\langle \mathcal P_\Omega(N_r),(\Delta_X)Y^\top+X(\Delta_Y)^\top\right\rangle \label{c.2}.
\end{eqnarray}
We now bound each of the term on the rightmost hand side to show that it can be lower bounded by $\|\Delta_{F_t}\|_{\rm F}^2$.
\paragraph{\indent Bounding \eqref{c.1}:} The term \eqref{c.1} can be bounded in a similar way as \cite{ge2016matrix}. However, since we are dealing with different $p_{ij}$ instead of uniform sampling in \cite{ge2016matrix}, new concentration inequalities are needed.
Denote by $D_1=S\Delta_Y^\top+\Delta_XT^\top$, $D_2=\Delta_X \Delta_Y^\top$, and
$$D_3 = \Delta_X^\top S + S^\top\Delta_X - \Delta_Y^\top T-T^\top \Delta_Y.$$
It is clear that
$$XY^\top-ST^\top=D_1+D_2,$$
and
$$(\Delta_X)Y^\top+X(\Delta_Y)^\top = D_1+2D_2.$$
Thus, by Cauchy-Schwartz inequality
\begin{eqnarray*}
&&\langle \mathcal P_\Omega(XY^\top-ST^\top),(\Delta_X)Y^\top+X(\Delta_Y)^\top\rangle \\
& =& \langle \mathcal P_\Omega(D_1+D_2),D_1+2D_2\rangle \\
& =& \langle \mathcal P_\Omega(D_1),D_1\rangle + 2 \langle \mathcal P_\Omega(D_2),D_2\rangle- 3\langle \mathcal P_\Omega(D_1), D_2 \rangle \\
& \geq& \langle \mathcal P_\Omega(D_1),D_1\rangle - 3\langle \mathcal P_\Omega(D_1), D_2 \rangle \\
& \geq&\langle \mathcal P_\Omega(D_1),D_1\rangle -3\sqrt{\langle \mathcal P_\Omega(D_1),D_1\rangle\langle \mathcal P_\Omega(D_2),D_2\rangle},
\end{eqnarray*}
We shall make use of the following concentration inequalities:
\begin{lemma}
\label{c1}
Under the assumptions of Theorem \ref{th:main}, with probability at least $1-d^{-\alpha}$,
$$\left| \langle \mathcal P_\Omega(D_1),D_1\rangle - \|D_1\|^2_\textup{F} \right| \leq \frac{1-\kappa_{r}}{20}\cdot \|D_1\|_\textup{F}^2,$$
and
$$\langle \mathcal P_\Omega(D_2),D_2\rangle \leq \left(\frac{1-\kappa_{r}}{20}\right)^2\cdot \sigma_r\|\Delta_F\|_\textup{F}^2.$$
\end{lemma}
In light of Lemma \ref{c1}, we have
\begin{align}
\label{c1bound}
& \langle \mathcal P_\Omega(XY^\top-ST^\top),(\Delta_X)Y^\top+X(\Delta_Y)^\top\rangle \nonumber \\
& \geq \left(1- \frac{1-\kappa_r}{20}\right)\|D_1\|_\textup{F}^2-3 \sqrt{\sigma_r}\|\Delta_F\|_\textup{F} \cdot \frac{1-\kappa_r}{20}\cdot \sqrt{1+\frac{1-\kappa_r}{20}}\|D_1\|_\textup{F}.
\end{align}
\paragraph{\indent Bounding \eqref{c.3}:} It is not hard to see that $S^\top S = T^\top T$. Thus, by rewriting $X = S + \Delta_X$ and $Y = T + \Delta_Y$, we have
\begin{align*}
&{1\over 2} \left\langle X^\top X - Y^\top Y ,X^\top(\Delta_X) - (\Delta_Y)^\top Y\right\rangle \\
=& {1\over 2} \left\langle X^\top X - Y^\top Y,\Delta_X^\top\Delta_X -\Delta_Y^\top \Delta_Y\right\rangle +{1\over 2} \left\langle\Delta_X^\top\Delta_X -\Delta_Y^\top \Delta_Y, S^\top \Delta_X - \Delta_Y^\top T \right\rangle \\
&\hskip 50pt+ {1\over 2} \left\langle \Delta_X^\top S + S^\top\Delta_X - \Delta_Y^\top T-T^\top \Delta_Y,S^\top \Delta_X - \Delta_Y^\top T \right\rangle\\
=& {1\over 2} \left\langle \Delta_X^\top S + 2S^\top\Delta_X - 2\Delta_Y^\top T-T^\top \Delta_Y+ \Delta_X^\top\Delta_X -\Delta_Y^\top \Delta_Y,\Delta_X^\top\Delta_X -\Delta_Y^\top \Delta_Y\right\rangle \\
&\hskip 50pt+ {1\over 4} \left\| \Delta_X^\top S + S^\top\Delta_X - \Delta_Y^\top T-T^\top \Delta_Y\right\|_{\rm F}^2,
\end{align*}
where we used the result that $\Delta_X^\top S + \Delta_Y^\top T$ is symmetric from Lemma 6 of \cite{ge2017no}. The first term on the righthand side can be bounded by
\begin{eqnarray*}
&&\left| {1\over 2}\left\langle \Delta_X^\top S + 2S^\top\Delta_X - 2\Delta_Y^\top T-T^\top \Delta_Y+ \Delta_X^\top\Delta_X -\Delta_Y^\top \Delta_Y,\Delta_X^\top\Delta_X -\Delta_Y^\top \Delta_Y\right\rangle \right| \\
&\leq&{1\over 2}(3\|S\|\|\Delta_X\|_{\rm F}+3\|T\|\|\Delta_Y\|_{\rm F}+ \|\Delta_X^\top\Delta_X -\Delta_Y^\top \Delta_Y\|_{\rm F})\|\Delta_X^\top\Delta_X -\Delta_Y^\top \Delta_Y\|_{\rm F} \\
&\leq&2\sqrt{\sigma_1}(\|\Delta_X\|_{\rm F}+\|\Delta_Y\|_{\rm F})(\|\Delta_X\|_{\rm F}^2+\|\Delta_Y\|_{\rm F}^2)\\
& \leq& 3\sqrt{\sigma_1}\|\Delta_F\|_{\rm F}^3,
\end{eqnarray*}
where the last inequality follow from the fact that $\max\{ \|\Delta_X\|, \|\Delta_Y\| \} \leq \|\Delta_F\|_{\rm F} \le \sqrt{\sigma_1}$.
\paragraph{\indent Bounding \eqref{c.2}:}
Observe that
\begin{align}
\left\langle \mathcal P_\Omega(N_r),(\Delta_X)Y^\top+X(\Delta_Y)^\top \right\rangle
= &\left\langle \mathcal P_\Omega(N_r)-N_r,(\Delta_X)Y^\top+X(\Delta_Y)^\top \right\rangle \nonumber\\
&+ \left\langle N_r,(\Delta_X)Y^\top+X(\Delta_Y)^\top \right\rangle.\label{c.4}
\end{align}
The first term on the righthand side can be bounded by
\begin{eqnarray*}
&& \left\langle \mathcal P_\Omega(N_r)-N_r,(\Delta_X)Y^\top+X(\Delta_Y)^\top \right\rangle \nonumber\\
&\leq & \left\| \mathcal P_\Omega(N_r)-N_r \right\| \left\| (\Delta_X)Y^\top+X(\Delta_Y)^\top \right\|_* \nonumber\\
&= & \left\| \mathcal P_\Omega(N_r)-N_r \right\| \sqrt{2r}\left\| (\Delta_X)Y^\top+X(\Delta_Y)^\top \right\|_\textup{F} \nonumber\\
&\leq & \sqrt{2r}\left\| \mathcal P_\Omega(N_r)-N_r \right\| \left(\left\|D_1 \right\|_\textup{F} + 2\|D_2\|_{\rm F}\right) \nonumber\\
&\leq & \sqrt{2r} \left\| \mathcal P_\Omega(N_r)-N_r \right\| (\|D_1\|_\textup{F} + \|\Delta_F\|_{\rm F}^2)
\end{eqnarray*}
where we made use of the facts that, for two arbitrary matrices $A$ and $B$,
$$\langle A,B \rangle \leq \|A\| \|B\|_*,\qquad \|A\|_*^2 \leq \mathrm{rank}(A)\|A\|_\textup{F}^2.$$
Note that by definition $S^\top N_r = 0$ and $N_r T =0$. This implies that
\begin{eqnarray*}
&&\left\langle N_r,(\Delta_X)Y^\top+X(\Delta_Y)^\top \right\rangle \nonumber\\
&=& \mathrm{trace}( (\Delta_X)^\top N_rY) + \mathrm{trace}( X^\top N_r (\Delta_Y)) \nonumber \\
&=& 2 \cdot \mathrm{trace}(\Delta_X^\top N_r \Delta_Y) \\
&\leq& 2\sigma_{r+1}\|\Delta_X\|_\textup{F}\|\Delta_Y\|_\textup{F}\nonumber \\
&\leq &\sigma_{r+1}\|\Delta_F\|_\textup{F}^2
\end{eqnarray*}
In summary, we have
\begin{eqnarray*}
&&\left\langle \mathcal P_\Omega(N_r)-N_r,(\Delta_X)Y^\top+X(\Delta_Y)^\top \right\rangle \\
&\leq& \sqrt{2r} \left\| \mathcal P_\Omega(N_r)-N_r \right\| (\|D_1\|_\textup{F} + \|\Delta_F\|_{\rm F}^2) + \sigma_{r+1}\|\Delta_F\|_\textup{F}^2.
\end{eqnarray*}
Together, the bounds for \eqref{c.1}, \eqref{c.3} and \eqref{c.2} imply that
\begin{eqnarray*}
&& \langle \nabla_X f,\Delta_X\rangle + \langle \nabla_Y f,\Delta_Y\rangle \\
&\geq & \|D_1\|_\textup{F}^2 + {1\over 4}\|D_3\|_{\rm F}^2- \sigma_{r+1}\|\Delta_F\|_\textup{F}^2 -\sqrt{2r} \left\| \mathcal P_\Omega(N_r)-N_r \right\|(\|D_1\|_\textup{F} + \|\Delta_F\|_{\rm F}^2)\\
&& - \frac{1-\kappa_r}{20}\|D_1\|_\textup{F}^2-3\sqrt{\sigma_r} \|\Delta_F\|_\textup{F} \cdot \frac{1-\kappa_r}{20}\cdot \sqrt{1+\frac{1-\kappa_r}{20}}\|D_1\|_\textup{F} -3\sqrt{\sigma_1}\|\Delta_F\|_{\rm F}^3 .
\end{eqnarray*}
Denote by
\begin{eqnarray*}
\epsilon^2 &= &\|\Delta_XT^\top +S\Delta_Y^\top\|_{\rm F}^2 + {1\over 4}\| \Delta_X^\top S + S^\top\Delta_X - \Delta_Y^\top T-T^\top \Delta_Y\|_{\rm F}^2\\
&= &\|\Delta_X T^\top\|_{\rm F}^2 +\|S \Delta_Y^\top\|_{\rm F}^2 + \langle \Delta_X T^\top, S\Delta_Y^\top \rangle + {1\over 2}\| \Delta_X^\top S\|_{\rm F}^2 + {1\over 2}\| \Delta_Y^\top T\|_{\rm F}^2\\
&&\hskip50pt + {1\over 2}\langle\Delta_X^\top S, S^\top\Delta_X \rangle + {1\over 2}\langle\Delta_Y^\top T, T^\top \Delta_Y \rangle - \langle \Delta_X^\top S ,\Delta_Y^\top T \rangle \\
&= & \|\Delta_X T^\top\|_{\rm F}^2 +\|S \Delta_Y^\top\|_{\rm F}^2 + {1\over 2}\| \Delta_X^\top S-\Delta_Y^\top T \|_{\rm F}^2 \\
&&\hskip50pt + {1\over 2}\langle \Delta_X^\top S+\Delta_Y^\top T , S^\top \Delta_X+T^\top \Delta_Y \rangle \\
&= & \|\Delta_X T^\top\|_{\rm F}^2 +\|S \Delta_Y^\top\|_{\rm F}^2 + {1\over 2}\| \Delta_X^\top S-\Delta_Y^\top T \|_{\rm F}^2+ {1\over 2}\| \Delta_X^\top S+\Delta_Y^\top T \|_{\rm F}^2,
\end{eqnarray*}
where we again used the result that $\Delta_X^\top S + \Delta_Y^\top T$ is symmetric from Lemma 6 of \cite{ge2017no} and the fact that $ \langle \Delta_X T^\top, S\Delta_Y^\top \rangle = \langle\Delta_X^\top S, T^\top\Delta_Y \rangle$.
This implies
\begin{align}
\label{epsilon:low}
\epsilon^2 \geq \sigma_r\|\Delta_F\|_{\rm F}^2.
\end{align}
Note that
\begin{align}
\label{deltaf}
\|\Delta_F\|_{\rm F}^2= \|\Delta_F\|_{\rm F}\cdot\|\Delta_F\|_{\rm F} \leq \frac{1-\kappa_{r}}{20\sqrt{\sigma_1}}\sigma_r\|\Delta_F\|_{\rm F}
\leq \frac{1-\kappa_{r}}{20}\sqrt{\sigma_r}\|\Delta_F\|_{\rm F} \leq \frac{1-\kappa_{r}}{20}\epsilon.
\end{align}
We get
\begin{eqnarray}
\label{gradient:lowbound}
&& \langle \nabla_X f,\Delta_X\rangle + \langle \nabla_Y f,\Delta_Y\rangle \nonumber\\
&\geq & \epsilon^2 (1 -\kappa_r)-\sqrt{2} \left\| \mathcal P_\Omega(N_r)-N_r \right\|(\|D_1\|_\textup{F} + \|\Delta_F\|_{\rm F}^2)- {31\over 5}\cdot\frac{1-\kappa_r}{20}\epsilon^2 \\
&\geq & \frac{69(1-\kappa_{r})}{100}\epsilon^2- 1.5\sqrt{r} \left\| \mathcal P_\Omega(N_r)-N_r \right\|\epsilon. \nonumber
\end{eqnarray}
In light of \eqref{deltaf}, if
\begin{equation}
\label{condition}
\max\left\{\|D_1+D_2\|_{\rm F},\sqrt{\sigma_r}\| \Delta_{F}\|_{\rm F}\right\}>\frac{31\sqrt{r}}{10(1-\kappa_{r})}\left\|\mathcal P_\Omega(N_r)-N_r\right\|,
\end{equation}
then
\begin{eqnarray*}
\|D_1+D_2\|_{\rm F}^2 &\leq& (\|D_1\|_{\rm F} + 0.5\|\Delta_F\|_{\rm F}^2)^2 \\
&\leq& (\|D_1\|_{\rm F} +\epsilon/40)^2\\
&\leq& \|D_1\|_{\rm F}^2 + \|D_1\|_{\rm F}\epsilon/20 + \epsilon^2/400\\
&<& 1.06\epsilon^2.
\end{eqnarray*}
Together with \eqref{condition}, we get
$${(1-\kappa_{r})\epsilon\over 3} > \sqrt{r}\left\|\mathcal P_\Omega(N_r)-N_r\right\|.$$
In light of \eqref{gradient:lowbound}, we have
\begin{eqnarray*}
&&\langle \nabla_X f,\Delta_X\rangle + \langle \nabla_Y f,\Delta_Y\rangle \\
&\geq& \frac{69(1-\kappa_{r})}{100}\epsilon^2-1.5\sqrt{r}\left\| \mathcal P_\Omega(N_r)-N_r \right\|\epsilon \\
&\geq& \left(\frac{69}{100}-{1\over 2}\right)(1-\kappa_{r})\epsilon^2\\
&\geq& {19\over 100}(1-\kappa_{r})\sigma_r\|\Delta_F\|_{\rm F}^2 .
\end{eqnarray*}
The case when \eqref{condition} does not hold can be handled in a similar fashion: since
\begin{align*}
\|D_1\|_{\rm F}&\leq \|D_1+D_2\|_{\rm F} + \|D_2\|_{\rm F} \leq \|D_1+D_2\|_{\rm F} + \|\Delta_F\|_{\rm F}^2 \\
& \leq \|D_1+D_2\|_{\rm F} + \frac{1-\kappa_{r}}{20}\sqrt{\sigma_r}\|\Delta_F\|_{\rm F} < \frac{63\sqrt{r}}{20(1-\kappa_{r})}\left\|\mathcal P_\Omega(N_r)-N_r\right\|,
\end{align*}
tegother with \eqref{gradient:lowbound}, we have
$$
\langle \nabla_X f,\Delta_X\rangle + \langle \nabla_Y f,\Delta_Y\rangle
\geq \frac{69(1-\kappa_{r})}{100}\sigma_r\|\Delta_F\|_{\rm F}^2 - {5r\| \mathcal P_\Omega(N_r)-N_r\|^2\over (1-\kappa_{r})}.
$$
\end{proof}
\begin{proof}[Proof of Lemma \ref{b2}]
By Cauchy-Schwartz inequality, for any $V \in \mathbb R^{d_1\times r}$ such that $\|V\|_\textup{F}=1$,
\begin{eqnarray}
\left|\langle \nabla_X f,V\rangle \right|^2
&= &\left| \langle \mathcal P_{\Omega}\left(XY^\top - A\right) , VY^\top \rangle +{1\over 2} \langle (X^\top X - Y^\top Y) , X^\top V \rangle\right|^2 \nonumber\\
&= &2\left| \langle \mathcal P_\Omega(XY^\top-ST^\top)-\mathcal P_\Omega(N_r),VY^\top \rangle\right|^2 \nonumber\\
&&\hskip 50pt+ {1\over 2}\left| \langle (X^\top X - Y^\top Y) , X^\top V \rangle\right|^2 \nonumber\\
&\leq &4 \langle \mathcal P_\Omega(XY^\top-ST^\top), VY^\top \rangle^2 + 4 \langle \mathcal P_\Omega(N_r),VY^\top \rangle^2 \nonumber\\
&&\hskip 50pt +{1\over 2}\left\| X^\top X - Y^\top Y\right\|_{\rm F}^2 \|X\|^2 \|V\|_{\rm F}^2 .\label{d.1}
\end{eqnarray}
The first term of \eqref{d.1} can be bounded by
\begin{eqnarray*}
&&\left| \langle \mathcal P_\Omega(XY^\top-ST^\top),VY^\top\rangle \right|^2 \\
&\leq & \left\langle \mathcal P_\Omega(XY^\top-ST^\top) ,XY^\top-ST^\top \right\rangle \left\langle \mathcal P_\Omega(VY^\top),VY^\top \right\rangle \\
&\leq& \left[\left\langle \mathcal P_\Omega(S(\Delta_Y)^\top), S(\Delta_Y)^\top \right\rangle^{1/2}+ \left\langle \mathcal P_\Omega((\Delta_X)Y^\top), (\Delta_X)Y^\top \right\rangle^{1/2} \right]^2\cdot\left\langle \mathcal P_\Omega(VY^\top),VY^\top \right\rangle.
\end{eqnarray*}
In the last line, we used the fact that $XY^\top-ST^\top = S(\Delta_Y)^\top + (\Delta_X)Y^\top$ and the Cauchy-Schwartz inequality. We shall need the following concentration inequality.
\begin{lemma}\label{f1}
Assume that each entry of $\Omega$ is independently sampled from binomial trails with probability given by \eqref{eq:defp}. There exists a numerical constant $C_1 >0$ such that, for any matrices $X,Y$,
\begin{align*}
\langle \mathcal P_{\Omega}(XY^\top),XY^\top\rangle \leq 2 \|A_r\|_{\rm F}^2 \min \left\{ \|X\|_\textup{F}^2\cdot \max_j \left(\frac{\|Y_{j.}\|^2}{\|(A_r)_{.j}\|^2}\right),\|Y\|_\textup{F}^2\cdot \max_i \left(\frac{\|X_{i.}\|^2}{\|(A_r)_{i.}\|^2}\right) \right\},
\end{align*}
with probability at least $1-d^{-\alpha}$ provided that $n \geq C_1\alpha d\log d \cdot \mu_r(A)\|A\|_{\rm F}^2/\|A_r\|_{\rm F}^2$.
\end{lemma}
By Lemma \ref{f1}
\begin{eqnarray*}
&& \left| \langle \mathcal P_\Omega(XY^\top-ST^\top),VY^\top \rangle \right|^2 \\
&\leq& \frac{4}{\beta^2(1-\nu_r(A)^2)} \|A_r\|_{\rm F}^2\|\Delta_F\|_\textup{F}^2\cdot \frac{2}{\beta^2(1-\nu_r(A)^2)} \|A_r\|_{\rm F}^2\|V\|_{\rm F}^2 \\
&=& \frac{8}{\beta^4(1-\nu_r(A)^2)^2} \|A_r\|_{\rm F}^4\|\Delta_F\|_\textup{F}^2,
\end{eqnarray*}
where we used the facts that
$$
\|S_{i.}\| \le \|(A_r)_{i.}\|/\sqrt{\sigma_{r}}\le \|(A_r)_{i.}\|/ (\beta\sqrt{1-\nu_r(A)^2}),$$
and
$$\|Y_{j.}\| \le {\|A_{.j}\|\over\beta} \le \max_j \left\{{\|A_{.j}\| \over \|(A_r)_{.j}\|} \right\}\cdot {\|(A_r)_{.j}\|\over\beta} = {\|(A_r)_{.j}\|\over \beta\sqrt{1-\nu_r(A)^2}}.$$
Here, we used the definition of $\nu_r(A)$,
$$1-\nu_r(A)^2 = \min_{\substack{1\le i \le d_1\\1\le j \le d_2}}\left\{{\|(A_r)_{i\cdot}\|^2\over \|A_{i\cdot}\|^2},{\|(A_r)_{\cdot j}\|^2\over \|A_{\cdot j}\|^2}\right\} = \left[\max_{\substack{1\le i \le d_1\\1\le j \le d_2}}\left\{{ \|A_{i\cdot}\|^2\over \|(A_r)_{i\cdot}\|^2},{ \|A_{\cdot j}\|^2\over\|(A_r)_{\cdot j}\|^2}\right\}\right]^{-1}.$$
Similarly, we can bound the second term on the righthand side of \eqref{d.1} by
\begin{eqnarray*}
&&\langle \mathcal P_\Omega(N_r), VY^\top\rangle^2 \\
&\leq& 2\langle \mathcal P_\Omega(N_r)-N_r, VY^\top\rangle^2 + 2\langle N_r, VY^\top\rangle^2 \\
&= & 2\langle \mathcal P_\Omega(N_r)-N_r, VY^\top\rangle^2 + 2 \cdot \mathrm{trace}\left( (\Delta_Y)^\top N_r^\top V \right)^2\\
&\leq & 2\| \mathcal P_\Omega(N_r)-N_r\|^2 \cdot \|VY^\top\|_*^2 + 2 \sigma_{r+1}^2 \cdot \|V\|_\textup{F}^2\|\Delta_Y\|_\textup{F}^2\\
&\leq & 2\| \mathcal P_\Omega(N_r)-N_r\|^2 \cdot r\|VY^\top\|_\textup{F}^2 + 2 \sigma_{r+1}^2 \|\Delta_Y\|_\textup{F}^2\\
&\leq & 2r\| \mathcal P_\Omega(N_r)-N_r\|^2 \|Y\|^2 + 2 \sigma_{r+1}^2\|\Delta_Y\|_\textup{F}^2\\
&\leq & 2r\| \mathcal P_\Omega(N_r)-N_r\|^2 \|Y\|^2 + {2 \|A_r\|_{\rm F}^4\over\beta^4(1-\nu_r(A)^2)^2}\|\Delta_Y\|_\textup{F}^2,
\end{eqnarray*}
where we used the fact that
$${\|A_r\|_{\rm F}^2 \over\beta^2(1-\nu_r(A)^2)} \geq {\|A_r\|_{\rm F}^2\over \sigma_r} \geq \sigma_r,$$
in the last inequality.
The last term of \eqref{d.1} can be bounded by
\begin{eqnarray*}
&&{1\over 2}\left\| X^\top X - Y^\top Y\right\|_{\rm F}^2 \|X\|^2 \|V\|_{\rm F}^2\\
&= &{1\over 2}\left\| X^\top \Delta_X +\Delta_X^\top S - Y^\top \Delta_Y -\Delta_Y ^\top T \right\|_{\rm F}^2 \|X\|^2\\
&\leq & {1\over 2}\left[ (\| X\|+\|S\|) \|\Delta_X\|_{\rm F} + (\| Y\|+\|T\|) \|\Delta_Y\|_{\rm F} \right]^2 \|X\|^2\\
&\leq & {9\over 2} \sigma_1^2 \|\Delta_F\|_{\rm F}^2\\
&\leq& {9 \|A_r\|_{\rm F}^4\over 2\beta^4(1-\nu_r(A)^2)^2}\|\Delta_F\|_{\rm F}^2,
\end{eqnarray*}
where we used the facts that
$$X = S+\Delta_X, \quad Y = T + \Delta_Y,\quad S^\top S = T^\top T,$$
$$\|A_r\|_{\rm F}^2 /(\beta^2(1-\nu_r(A)^2)) \geq \sigma_1,$$
\begin{equation}
\label{X:spectral_bound}
\|X\| \le \|S\| + \|\Delta_X\| \le \|S\| + \|\Delta_F\|_{\rm F} \le \sqrt{\sigma_1} + \frac{1-\kappa_{r}}{20\sqrt{\sigma_1}}\sigma_r \le 1.05\sqrt{\sigma_1},
\end{equation}
and same bound for $\|Y\|$ due to symmetry.
In summary, we have
$$
\| \nabla_X f\|_{\rm F}^2\leq { \|A_r\|_{\rm F}^4\cdot (73\|\Delta_F\|_\textup{F}^2 + 16 \|\Delta_Y\|_\textup{F}^2)\over 2\beta^4(1-\nu_r(A)^2)^2} +9r\sigma_1\| \mathcal P_\Omega(N_r)-N_r\|^2.
$$
The claim then follows from a similar bound for $\| \nabla_Y f\|_{\rm F}$.
\end{proof}
\newpage
|
\section{Introduction}
Kinetic inductance travelling-wave parametric amplifiers (KI-TWPAs) promise high gain, large bandwidth and potential for quantum limited noise performance \cite{Esposito2021}. Currently, most KI-TWPAs \cite{Eom2012, Vissers2016, Malnou3WM} have been fabricated for sub-20 GHz and \si{\milli\kelvin} temperature operation. However, there is interest in extending this to higher frequency applications where quantum limited amplification is needed such as dark matter detection \cite{Beurthey2020madmax} and high frequency quantum computing \cite{Fermarzi2021}. At mm-wave energy scales these superconducting circuits are less sensitive to thermal background noise due to the higher photon energies involved. Hence, these paramps may achieve their low noise figures at temperatures of \SI{1}{\kelvin}, which are accessible via higher heat lift fridges such as adsorption coolers. Additionally, even at temperatures of \SI{4}{\kelvin} KI-TWPAs are competitive with their transistor-based counterparts while producing an order of magnitude smaller heat load \cite{Malnou2021_4K}. Recently, narrowband parametric amplification has been achieved at \SI{30}{\giga\hertz} \cite{Banys2020} and at \SI{95}{\giga\hertz} \cite{Anferov2020}, while the highest reported frequency in a travelling wave design \cite{Shu2021} is approaching \SI{30}{\giga\hertz}.
At mm-wave frequencies KI-TWPAs based on conventional transmission lines such as co-planar waveguides and microstrips face challenges in connecting their \si{\micro\meter} size traces to external feeding structures without suffering from large dielectric losses. A potential method to alleviate dielectric losses at frequencies over 100 GHz is to use more exotic transmission lines that incorporate a hollow gap between conducting surfaces instead of one filled with dielectric and can be coupled to waveguide feeds via low loss transitions. One such structure is ridge gap waveguide (RGWG)\cite{Kildal2011} which is readily scalable to frequencies of a few hundred GHz \cite{Rahiminejad2012} and can support quasi-TEM mode of transmission over an octave of bandwidth. Coating these RGWGs with a thin superconducting (SC) film such as NbTiN and maintaining a conductor separation in the micrometer to sub-micrometer range can help achieve kinetic inductance fractions in excess of 90\%, which makes these transmission line structures potential candidates for mm-wave KI-TWPAs.
\section{High Kinetic Inductance Ridge Gap Waveguide}
RGWG as shown in Fig.~\ref{fig:ridgegapWG} consists of two parallel conducting plates where one of these is textured with a wave-guiding ridge that is surrounded by a periodic pin surface. This surface, often referred to as a bed of nails, on its own is able to stop the propagation of parallel plate modes that may exist in the hollow enclosed cavity if the nails were absent. This protecting characteristic is exhibited over a certain bandwidth that depends on the length of the nails and the gap to the top surface \cite{Kildal2011}. A wave guiding channel may be created between these nails, where the nails act as soft magnetic boundaries at the edges of the channel, creating a type of groove gap waveguide (GGWG). Finally, adding a central ridge to this groove allows the propagation of a TEM mode with fields constrained to the volume between the top of the ridge and the opposing lid.
\begin{figure}
\centering
\begin{subfigure}[b]{0.49\textwidth}
\centering
\includegraphics[width=\textwidth]{Fig1a.pdf}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.49\textwidth}
\centering
\includegraphics[width=\textwidth]{Fig1b.pdf}
\end{subfigure}
\caption{Left: A cross section view of RGWG model showing the central ridge of width $w=$ \SI{0.5}{\milli\meter} in a groove of width $a=$ \SI{2.54}{\milli\meter} and height $b=$ \SI{1.27}{\milli\meter}, with 3 rows of pins with width and separation of \SI{0.5}{\milli\meter}. The dimensions of $a$ and $b$ were chosen to match that of W band rectangular waveguide. Right: The dispersion of the first 8 modes from the eigenmode simulation of RGWG in HFSS. Modes 1-6, and 8 are the spurious modes that propagate in the whole internal volume and signify the lower and upper cut-off frequencies at which the bed of nails protects the propagation of the RGWG mode. As shown by the overlaid "light" line ($\beta = 2\pi f/c$) mode 7 is the quasi-TEM RGWG mode which is the sole propagating mode between 20 to 120 GHz (Color figure online).}
\label{fig:ridgegapWG}
\end{figure}
RGWG can be fabricated using conventional metal milling or silicon micro-machining techniques \cite{Rahiminejad2012}. In both cases a SC coating is necessary to provide a nonlinear kinetic inductance, which is required for parametric processes to take place. The extent to which this nonlinearity dominates over the total structure inductance is dictated by the kinetic inductance fraction, $\alpha$, which is dependent on the waveguide geometry and film properties. It is possible for $\alpha\rightarrow1$ to be achieved when the SC coating is thin enough such that $t\ll\lambda_L$, where $\lambda_L$ is the London penetration depth, and the hollow gap is small such that $h<1$ \si{\micro\meter}, which can be achieved via a deposition of a thin spacer along the outer perimeter of one of the involved bodies.
RGWG design begins with the engineering of the pin surface which creates the stopband for any global parallel plate or cavity modes that may exist. It was found that the pin surface protecting bandwidth was maximised when the nail width and separation were equal. A nail width of 0.5 mm was chosen with consideration to the mechanical fragility of the nails. Following this, the GGWG and RGWG are considered, which are dictated by the groove width, $a$, and ridge width $w$, respectively. The groove width was set to match the long wall of W band rectangular waveguide and three rows were chosen to provide sufficient isolation to the device edges in the GGWG region that would be needed in the transition between rectangular waveguide and RGWG via a taper. The GGWG mode fields decay by $\sim$ 30 dB per row of nails. The upper cut-off of the RGWG mode can be affected by $w$ especially when $w\rightarrow a$, since the added presence of a high KI coating can cause higher order RGWG modes to propagate. The main tool for designing and simulating these structures was Ansys HFSS\footnote{https://www.ansys.com/products/electronics/ansys-hfss}. The superconducting coating was included in S-parameter simulations via frequency dependent surface impedance boundaries \cite{Belitsky2006} that are imported from a calculation of the Zimmermann conductivity \cite{Zimmermann1991}. Specifically, the eigenmode solver of HFSS was used to calculate the dispersion properties of pin surface and RGWG, which is shown in Fig. \ref{fig:ridgegapWG}.
A thin SC coating enhances the linear kinetic inductance and sets the critical current of the line which needs to be reached for maximum non-linearity. Since the top of the ridge is easily acessible, the thin SC coating can be applied to this surface. A thin coating ($<\lambda_L$) on the textured surface deposited by a technique such as magnetron sputtering may cause current restrictions or discontinuities on sharp edges of the pins and ridges. These coating irregularities may give rise to undesired dissipative behaviour in regions of high current flow where the critical current may be surpassed locally. This may be alleviated by depositing a much thicker coating ($\gg\lambda_L$) on the textured surface or using a deposition technique such as atomic layer deposition, which can produce a thin conformal coating on all edges.
For this work, NbTiN coatings with a critical temperature $T_c \sim $ \SI{12.5}{\kelvin} and resistivity in the range of 200 \si{\micro\ohm} \si{\centi\meter} are considered. This coating would allow paramp operation at $\sim 1-4$ K with gain and noise figures comparable to equivalent transistor-based amplifiers \cite{Malnou2021_4K}. At these temperatures dissipation should be low due to the $\exp{(-T_c/T)}$ scaling of the SC surface resistance \cite{Gurevich2017}.
The linear kinetic inductance, $L_{k0}$, introduced via the SC coating scales with the dimensions and physical properties of the film. The kinetic inductance scales non-linearly with current \cite{Semenov2020} and when expanded to second order can be expressed as $L_k(I) = L_{k0}\left[1 + (I/I_*)^2\right]$, where $I_*$ is the scaling current which is dependent on the SC material properties and the geometry of the trace.
For parametric processes to be explored it is first necessary to estimate the values of the low current KI and scaling current. The value of $L_{k0}$ per square can be estimated \cite{Belitsky2006} using $L_{k0}= \mu_0\lambda_L\tanh{(t/\lambda_L)}$. To determine the extent to which the KI non-linearity dominates over the total inductance the KI fraction can be calculated using, $\alpha = L_k/L_g+L_k$, where $L_g=\mu_0h$ is the geometric inductance for a RGWG assuming a parallel plate (separated by $h$) approximation.
The scaling current is sensitive to a number of factors \cite{Zolochevskii2014} including the substrate roughness and composition, film geometry, deposition conditions and operating temperature which means it can vary over orders of magnitude for a given SC material. Analytic solutions of the scaling current usually only set the upper bound for this value. The scaling current, which is of the order of the critical current, $I_c$, takes on different solutions based on the current distribution on the SC trace. For a conductor of thickness $t \ll \lambda_L$ and width $w \ll \lambda_\perp$, where $\lambda_\perp =2\lambda_L^2/t$ is the perpendicular penetration depth, the current density can be assumed to be constant across the film \cite{Pearl1964}. This low temperature scaling current can be estimated \cite{Zmuidzinas2012, Shu2021} using $I_*=wt\kappa_*\sqrt{N_0\Delta^2/\mu_0\lambda_L^2}$, where $\kappa_*$ is constant of order 1 \cite{Zhao2020} and $N_0$ is the number density at the Fermi level. However, if the trace is wider than the perpendicular penetration depth $w \gg \lambda_\perp$, the current distribution becomes non-uniform and accumulates at the edges of the film \cite{Zolochevskii2014}. In this work, the SC traces are in the 100s \si{\micro\meter} which means solutions of the wide film critical current should be explored. In the case of a pristine film with perfect edges the critical current \cite{Zolochevskii2014} in SI units is given by
\begin{equation}
I_c(T) = \frac{2\Phi_0}{3\sqrt{3\pi}\xi(0)\mu_0}\left(\frac{w}{\lambda_\perp(0)}\right)^{1/2}(1-T/T_c),
\label{eq:IcWide}
\end{equation}
\noindent where $\mu_0$ is the permeability of free space, $\Phi_0$ is the magnetic flux quantum and $\xi(0)$ is the zero temperature Ginzburg-Landau coherence length. The perfect edge stipulation is important since $I_c$ in a wide film is dependent on the pinning properties that arise at the edges of the trace \cite{Zolochevskii2014}. Therefore, Eq. \ref{eq:IcWide} provides only the upper limit for the wide film critical current. The scaling current should be measured experimentally by measuring the kinetic inductance induced frequency shift of a resonant cavity with ac or dc current \cite{Eom2012}.
\section{Travelling-Wave Paramp using Ridge Gap Waveguide}
In the presence of a strong pump tone and weak signal a four photon mixing process can take place \cite{Erickson2017}, referred to as four wave mixing (4WM). These tones are related by $2\omega_p = \omega_s + \omega_i$, where $\omega_{p,s,i}$ are the frequencies of the pump, signal and idler tones, respectively.
For sufficient parametric gain the RGWG needs to be extended into an electrically long structure spanning hundreds of wavelengths. In addition, the phase relationship between the pump, signal and idler tones should be considered such that the dispersion difference is $\Delta\beta = \beta_s+\beta_i-2\beta_p=0$, where $\beta_{p,s,i}$ are the propagation constants of the involved tones. The nonlinear SC line introduces a power dependent phase shift, which must be compensated for by introducing anomalous dispersion \cite{Eom2012, Bockstiegel2014, Vissers2016, Shan2016} in order to satisfy the condition for wideband exponentially scaling gain. If this condition is not met, the gain will scale quadratically with propagation length.
The simplest dispersion engineering method involves the perturbation of the transmission line every $\lambda_{\text{per}}/2$ such that stopbands are created at integer values of frequency $f_{\text{per}}$ near which anomalous dispersion is introduced. By positioning the pump tone near these dispersion features, the nonlinear phase slippage may be partially or even fully cancelled out.
\begin{figure}
\centering
\begin{subfigure}[b]{0.49\textwidth}
\centering
\includegraphics[width=\textwidth]{Fig2a.pdf}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.49\textwidth}
\centering
\includegraphics[width=\textwidth]{Fig2b.pdf}
\end{subfigure}
\caption{Left: A schematic of a section of the ridge in a dispersion engineered RGWG using design principle from \cite{Vissers2016}. The physical width of the main line is \SI{500}{\micro\meter}, while the repeating perturbations are \SI{650}{\micro\meter} wide, and $\lambda/20=$ \SI{25}{\micro\meter} long. The separation of loading midpoints is $\lambda/2=$ \SI{250}{\micro\meter}. Right: An equivalent circuit model of an unperturbed and perturbed section of the RGWG. The impedances of the main line and perturbed sections were chosen to be $Z_0 \sim$ \SI{6}{\ohm} and $Z_1 \sim $ \SI{4.75}{\ohm}, respectively, which were confirmed in HFSS simulations and included the kinetic inductance contribution of two surfaces coated with 30 nm of NbTiN ($\lambda_L \sim$ \SI{360}{\nano\meter}). Using such a low impedance increased the induced AC current for a given input power, which had the effect of reducing the necessary drive power to stimulate nonlinear behaviour. In addition, the chosen ridge to lid spacing of $h = $ \SI{5}{\micro\meter} resulted in $\alpha \sim$ \SI{60}{\percent}. The corresponding equivalent circuit parameters were $L_{g,0} \sim$ \SI{12.5}{\nano\henry\per\meter}, $L_{g,1} \sim$ \SI{9.5}{\nano\henry\per\meter}, $L_{k,0} \sim$ \SI{22}{\nano\henry\per\meter}, $L_{k,1} \sim$ \SI{17}{\nano\henry\per\meter}, $C_{0} \sim$ \SI{0.9}{\nano\farad\per\meter}, $C_{1} \sim$ \SI{1.1}{\nano\farad\per\meter}. The scaling current was chosen to be \SI{50}{\milli\ampere}, which could be expected from a NbTiN coating deposited on an insulating layer with a rough metallic base (Color figure online).}
\label{fig:repeatingUnit}
\end{figure}
Dispersion engineering is achieved in RGWG by employing periodic widening of the ridge as shown in Fig. \ref{fig:repeatingUnit}. The impedance of the loaded and unloaded sections is estimated using $Z=\sqrt{(L_g+gL_{k0})/C}$, where $C$ is the capacitance and $g$ is a geometric factor accounting for the SC coating of both parallel surfaces of the RGWG. More accurate impedance values were extracted from HFSS simulations, these were within $\sim 5\%$ of the analytic solutions. The impedance of the perturbations was $\sim 20\%$ lower than the main line, which introduced a weak dispersion feature near the $f_{per}$ that becomes stronger as the number of repeating units is increased. As a result, a travelling wave structure spanning $450\lambda_{per}$ was investigated. Such an electrically long structure can be simulated directly in HFSS to show the stopband structure, however, HFSS is unable to take into account the current dependency of the KI nor does it support harmonic balance simulations to evaluate mixing. Keysight's ADS\footnote{https://www.keysight.com/us/en/products/software/pathwave-design-software/pathwave-advanced-design-system.html} is an industry standard amplifier design tool that is able to address these aspects of the paramp design process.
An equivalent circuit model of the nonlinear RGWG was created, which is described in Fig. \ref{fig:repeatingUnit}. This circuit takes into account the nonlinear inductance via a general symbolically defined device (SDD) component, which requires a voltage definition in the form $V = -\frac{d}{dt}L_k(I)I$. These repeating unit cells were nested within subcircuits to make up a total of $N=900$ cells. A large signal S-parameter (LSSP) simulation within ADS was used to generate the nonlinear S-parameters and dispersion properties as shown in Fig. \ref{fig:TW-KIPresults}. The dispersion of the TWPA was calculated via $\beta = -\text{unwrap}[\text{arg}(S_{21})]/L$, where $L$ is the device length.
\begin{figure}
\centering
\begin{subfigure}[b]{0.49\textwidth}
\centering
\includegraphics[width=\textwidth]{Fig3a.pdf}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.49\textwidth}
\centering
\includegraphics[width=\textwidth]{Fig3b.pdf}
\end{subfigure}
\caption{Left: Transmission coefficient and effective dispersion from a simulation of 900 $\lambda/2$ cells in ADS, corresponding to a physical propagation length of \SI{25}{\centi\meter}. The first 2 stopbands at $f_{per} \sim $ \SI{95}{\giga\hertz} and $f_{per} \sim$ \SI{190}{\giga\hertz} are shown. The unperturbed dispersion $\beta$ was found by fitting a second order polynomial to the perturbed data. Right: The simulated signal gain from the HB simulator of ADS. The -9 dBm pump tone is positioned at \SI{93}{\giga\hertz}, while the signal is at -80 dBm. The CME solution \cite{Chaudhuri2015} uses the dispersion properties extracted from the 900 cell LSSP simulation to calculate the phase mismatch between the 3 involved tones. The calculated CME gain matched the simulated ADS gain at $I/I_*\sim0.19$. (Color figure online).}
\label{fig:TW-KIPresults}
\end{figure}
The parametric gain of the dispersion engineered RGWG can be estimated using the simulated dispersion data and the analytic solutions from the coupled mode equations (CME) \cite{Chaudhuri2015} that describe the amplitudes of the propagating waves. The CME solutions used only take into account the amplitude of the first pump harmonic and the associated mixing products but do not take into account the transmission of the travelling-wave structure. These CMEs can be extended to include higher order pump harmonics \cite{Chaudhuri2015, Dixon2020}, which provide a more accurate gain description. Alternatively, the parametric gain can be simulated \cite{Sweetnam2022} using the harmonic balance (HB) simulator of ADS with the repeating circuit shown in Fig. \ref{fig:repeatingUnit}, which accurately describes the gain of such a paramp to a chosen number of input tone harmonics. The simulated gain of the \SI{25}{\centi\meter} RGWG paramp and is shown in Fig. \ref{fig:TW-KIPresults}, which shows a \SI{10}{\decibel} signal gain over the whole of W-band. This was achieved by sweeping the pump tone frequency and input power close to the first stop band generated by the periodic structure such that the phase mismatch $\Delta\beta$ is minimised and maximum gain is achieved.
The simulated gain shows promise that the fabricated RGWG KI-TWPA could provide sufficient gain as a front end amplifier. A logical prototype could be a series resonator of similar dimensions which would confirm the feasibility of the high KI RGWG at W-band and would allow the characterisation of the scaling current. The designed RGWG had a conservative conductor spacing of $h = $ \SI{5}{\micro\meter} and a \SI{30}{\nano\meter} NbTiN film coating, which resulted in a KI fraction of $\alpha \sim$ \SI{60}{\percent}. By reducing $h$ to \SI{1}{\micro\meter}, via a shorter deposition of the spacer layer, the value of $\alpha$ may reach $>90\%$, which would significantly reduce the required TWPA length.
The main practical challenges of the RGWG TWPA include the fabrication of the long structure, and the matching of the \SI{6}{\ohm} RGWG impedance to the chosen feedlines. The structure could be fabricated on a 4 inch wafer by reducing the groove width $a$ to \SI{700}{\micro\meter} and meandering the RGWG with a single row of nails between meanders to provide isolation between parallel lines. Transitioning to W band rectangular waveguide could be achieved via a slow tapering of the height of the ridge leading to the waveguide flange. In addition, the ridge could be reduced in width to $<$ \SI{50}{\micro\meter} which would increase the impedance to $>$ \SI{50}{\ohm} and ease the impedance mismatch. The use of an impedance taper would unavoidably cause significant gain ripples. However, such a tapered structure should still be able to demonstrate wideband parametric gain, confirming its' feasibility and paving the way for further development.
\section{Conclusion}
This work proposed a low loss structure for high frequency KI-TWPAs in the form of RGWG. This structure is scaleable to frequencies of a few \SI{100}{\giga\hertz} with feature sizes in the hundreds of \si{\micro\meter} range, meaning that it can be fabricated with high speed metal milling techniques or silicon micromachining and coated in a superconducting film. It was shown via superconducting HFSS simulations that this structure can have over an octave of bandwidth for single mode quasi-TEM transmission, which would be ultimately limited by external factors such as the transition to the chosen waveguide feeds. It was shown via ADS HB simulations that a NbTiN coated RGWG can be the basis of a TWPA and provide over \SI{10}{\decibel} of parametric gain over the whole of W-band (75--110 GHz) with a physical length of \SI{25}{\centi\meter}. This paper sets out the design methods and simulation pipeline for such KI RGWGs and their TWPA counterparts while including an accessible method to simulate the expected parametric gain.
\begin{acknowledgements}
This work was supported by STFC grant ST/R504956/1. D. Banys is supported by a STFC PhD studentship.
\end{acknowledgements}
\pagebreak
\bibliographystyle{aipnum4-1}
|
\section{Base network architecture} \label{base network}
The full base network is a two-layer real-valued CNN with a \texttt{ReLU} activation \cite{Glorot_AISTATS11_ReLU} in the first layer and a \textit{configuration product} activation in the second layer, as shown in Fig.\,\ref{fig:base_network}. In the configuration product activation function, every channel of the convolutional output is transformed into a number given by the product of all elements in the channel, and the sum over these numbers gives the network output $\psi_{\mathrm{b}}$. The name ``configuration product'' comes from the pair product states~\cite{Tahara_JPSJ08_mVMC} which sum over the products of pair entanglement factors to give the wave function. The full base network expresses $\psi_{\mathrm{b}}$ instead of $\log\psi_{\mathrm{b}}$, which permits the expression of sign structures without complex output values. The similar architectures are also used in some other NQS studies~\cite{Liang_PRB21_CNNPEPS, Li_arxiv21_LargeScaleCNN} and greatly improve the accuracy. In our numerical practice, the training of the base network is also faster compared with the networks expressing $\log\psi_{\mathrm{b}}$. In the case of the pyrochlore lattice, the channel number of both layers are set to 20 in the base network so that the total number of parameters in the base network and the sign network is similar to Ref.\,\cite{Astrakhantsev_arxiv21_pyrochlore}.
In this work, we would need an amplitude base network with strictly positive outputs. To this end, the \texttt{ShiftedReLU} activation, defined as
\begin{equation} \label{eq:ShiftedReLU}
\mathrm{ShiftedReLU}(x) = \mathrm{ReLU}(x) + 1 = \left\{
\begin{array}{cc}
x+1 & (x>0) \\
1 & (x<0)
\end{array} \right. ,
\end{equation}
is added before the configuration product activation function.
\section{Symmetry} \label{symmetries}
Symmetry plays an important role in the practice of NQS \cite{Nomura_JPCM2021_RBMsymm, Choo_PRL18_SymNet}. In the numerical experiments we apply suitable symmetry projectors to increase the accuracy of variational states. With the $q = 0$ translational symmetry already enforced by the CNN structure, we only discuss the remaining point group symmetries. Assuming the system permits a symmetry group of order $\nu$ represented by operators ${T_i}$ with characters ${\omega_i}$, the symmetrized wave function is then defined as
\begin{equation} \label{eq:wave_function_symm}
\begin{aligned}
\psi^{\mathrm{symm}}(\sigma) &= \frac{1}{\nu} \sum_i \omega_i^{-1} \psi(\hat T_i \sigma) \\
&= \frac{1}{\nu} \sum_i \omega_i^{-1} \psi_{\mathrm{b}}(\hat T_i \sigma) \times \psi_{\mathrm{h,sign}}(\hat T_i \sigma),
\end{aligned}
\end{equation}
with the original wave function $\psi(\sigma)$ given by Eq.\,\eqref{eq:wave_function}.
\begin{figure}[t]
\centering
\includegraphics[width=0.48\textwidth]{plots/pyrochlore_symm.pdf}
\caption{Symmetries of the pyrochlore lattice. $C_3$ is the three-fold rotation symmetry around the ``easy axis''. $I$ is the inversion symmetry with respect to the origin. $M$ is the mirror symmetry with the mirror plane passing the bond 2--3 and the middle point of 0 and 1.}
\label{fig:pyrochlore_symm}
\end{figure}
At the pyrochlore lattice, we apply the $C_3$ rotation symmetry around the ``easy axis'', $I$ inverse symmetry with respect to the origin and $M$ mirror symmetry shown in Fig.\,\ref{fig:pyrochlore_symm}. Those generators cover the full point symmetry group of the pyrochlore cluster. The ground state belongs to the trivial representation of the resulting symmetry group.
\section{Reweighting Monte Carlo} \label{reweighting}
Similarly to QMC, reweighting Monte Carlo has been applied to the study of statistical systems \cite{Munger_PRB91_ReweightingMC,Nakamura_JPSJ92_ReweightingQMC}. In this section we use the evaluation of energy as an example to illustrate the reweighting method.
To begin with, we define the two norms of the non-normalized quantum state $\ket{\Psi}$
\begin{equation}
||\Psi|| = \sqrt{\braket{\Psi|\Psi}} = \sqrt{\sum_\sigma|\psi_\sigma|^2},
\end{equation}
\begin{equation}
|\Psi| = \sum_\sigma|\psi_\sigma|.
\end{equation}
In the variational Monte Carlo approach, the energy is given by
\begin{equation} \label{eq:variational_energy2}
\begin{aligned}
E &= \sum_{\sigma,\sigma'}\frac{\psi_\sigma^*}{||\Psi||}\frac{\psi_{\sigma'}}{||\Psi||}H_{\sigma \sigma'} \\
&= \sum_\sigma \frac{|\psi_\sigma|^2}{||\Psi||^2} \sum_{\sigma'} \frac{\psi_{\sigma'}}{\psi_\sigma}H_{\sigma \sigma'}\\
&= \left\langle \sum_{\sigma'}\frac{\psi_\sigma'}{\psi_\sigma}H_{\sigma\sigma'} \right\rangle_2,
\end{aligned}
\end{equation}
where $H_{\sigma \sigma'} = \braket{\sigma|H|\sigma'}$, and $\left\langle \ldots \right\rangle_2$ denotes sampling with the weight $|\psi_{\sigma}|^2$.
The reweighting Monte Carlo estimates the energy as
\begin{equation}
\begin{aligned}
E &= \frac{|\Psi|}{||\Psi||^2} \sum_\sigma \frac{|\psi_\sigma|}{|\Psi|} \frac{|\psi_\sigma|}{\psi_\sigma} \sum_{\sigma'}\psi_{\sigma'} H_{\sigma\sigma'} \\
&= \frac{|\Psi|}{||\Psi||^2} \left\langle \frac{|\psi_\sigma|}{\psi_\sigma} \sum_{\sigma'}\psi_{\sigma'} H_{\sigma\sigma'} \right\rangle_1,
\end{aligned}
\end{equation}
where $\left\langle \ldots \right\rangle_1$ denotes sampling with the weight $|\psi_\sigma|$. The factor $|\Psi|/||\Psi||^2$ is given by
\begin{equation}
\frac{|\Psi|}{||\Psi||^2} = \frac{|\Psi|}{\sum_\sigma |\psi_\sigma|^2}
= \frac{1}{\sum_\sigma \frac{|\psi_\sigma|}{|\Psi|}|\psi_\sigma|}
= \frac{1}{\left\langle |\psi_\sigma| \right\rangle_1}.
\end{equation}
Finally, the energy in the reweighting VMC reads
\begin{equation} \label{eq:reweighting_energy}
E = \frac{1}{\left\langle |\psi_\sigma| \right\rangle_1} \left\langle \frac{|\psi_\sigma|}{\psi_\sigma} \sum_{\sigma'} \psi_{\sigma'} H_{\sigma \sigma'} \right\rangle_1.
\end{equation}
More generally,
\begin{equation}
\left\langle \ldots \right\rangle_2
= \frac{\left\langle |\psi_\sigma| \ldots \right\rangle_1}{\left\langle |\psi_\sigma| \right\rangle_1},
\end{equation}
which can help to write the SR optimization in a reweighting form.
When $|\psi_{\sigma'}|$ is much greater than $|\psi_\sigma|$, in VMC without reweighting the $\psi_{\sigma'}/\psi_\sigma$ ratio in Eq.\,\eqref{eq:variational_energy2} leads to large variance in the energy estimation, and the evolution strategy updates will focus on a few largest terms in the sampling, leading to unstable optimization. The reweighting method is introduced mainly for ameliorating this problem.
The reweighting Monte Carlo also makes it easier to update wave function entries with small amplitudes since the reweighting sampling probability is proportional to $|\psi|$ instead of $|\psi|^2$. Consequently, the reweighting Monte Carlo is adopted across all numerical experiments in this paper except calculating cumulative distribution functions to improve the accuracy of variational wave function.
\section{Covariance matrix adaptation evolution strategy (CMA-ES) details} \label{CMA-ES}
\textit{Hyperparameters.--\,}
The CMA-ES contains many hyperparameters. In numerical experiments, we set the ES sample number $N=100$, initial weights $\bm W = 0$ and the initial standard deviation of sampling $\sigma=\pi/4$. The default values in the \texttt{pycma} package~\cite{Hansen_19_pycma} are adopted for remaining hyperparameters.
The update weights $w_n$ are also chosen as the default values in the \texttt{pycma} package. In a descending order, they are given by
\begin{equation}
w_n' = \log \frac{N+1}{2} - \log n,
\end{equation}
\begin{equation}
w_n = \left\{
\begin{array}{rc}
w_n' / \sum |w_n'|^+, & w_n' \geq 0 \\
\alpha^- w_n' / \sum |w_n'|^-, & w_n' < 0
\end{array}
\right.,
\end{equation}
where $\sum |w_n'|^+$ and $\sum |w_n'|^-$ are the sum over all positive and negative $w_n'$ values, respectively. Here $\alpha^-$ is a multiplier introduced in Ref.\,\cite{Hansen_arxiv16_CMAES}, helping the information from previous iterations accelerate the optimization in a manner similar to the momentum term in gradient-based methods~\cite{Nesterov_03, Kingma_17_Adam}.
\textit{Estimation of energy in ES.--\,} An essential step in ES is the estimation of the variational energy $E_n$ for different sample weights $X_n$ through Eq.\,\eqref{eq:reweighting_energy}. To simplify the estimation, we use the same reweighting Monte Carlo samples for all sample networks because the Monte Carlo sampling probability is not influenced by the sign of the wave function if symmetries in Eq.\,\eqref{eq:wave_function_symm} are not applied. The base network outputs $\psi_{\mathrm{b}}(\sigma)$ and $\psi_{\mathrm{b}}(\sigma')$ are prepared for calculating the variational energy. These prepared values can also be used in the following SR step. After applying symmetries, the sign network is no longer trained. This simplification does not cause a significant loss of accuracy according to our numerical experiments.
Given $N$ sample networks in every iteration, an efficient calculation requires the parallel forward pass of all these networks. In practice we only define one sample network with its output channel number $N$ times of the original sign network so that it produces $N$ sign values corresponding to different sample weights in a single forward pass.
\textit{Comparison with simple ES.--\,} In the simple ES, the correlation of weights is not considered in sample generation and the weights with lowest variational energy are directly chosen as the new weights in the sign network. In our numerical experiments, the CMA-ES, as a popular variant of ES method, produces sign structures with lower variational energies compared with simple ES.
\begin{figure}[t]
\centering
\subfigure{ \label{fig:supplementary_comparison_energy}
\includegraphics[width=0.23\textwidth]{plots/energy_error.pdf}
}
\hspace{-0.6cm}
\subfigure{ \label{fig:supplementary_comparison_overlap}
\includegraphics[width=0.24\textwidth]{plots/overlap.pdf}
}
\caption{Comparison between different base and helper networks on the $4\times2^3$ pyrochlore Heisenberg model.}
\label{fig:pyrochlore_supplementary}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=0.45\textwidth]{plots/cumulative_pyrochlore.pdf}
\caption{Cumulative distribution function (CDF) of phase relative to ED on $4\times2^3$ pyrochlore lattice. The full network is combined with the phase or sign helper network. The phase is adjusted to lie within the range $(-\pi/2, 3\pi/2]$.}
\label{fig:pyrochlore_CDF}
\end{figure}
\section{Supplementary data for pyrochlore lattice} \label{supplementary_pyrochlore}
In Fig.\,\ref{fig:pyrochlore_supplementary}, we show the comparison of the energy error and the wave function overlap between different base networks and helper networks. The sign helper network outperforms the phase helper network, and the full base network also outperforms the amplitude base network.
Similar to Fig.\,\ref{fig:square_CDF}, in Fig.\,\ref{fig:pyrochlore_CDF} we depict the CDF of the variational phase obtained within the combination of the full base network and different helper networks relative to the phase $\phi_{\mathrm{ED}}$ expected within the ED. The CDF shows that the sign network trained by ES learns a better distribution of phases due to elimination of a redundant complex phase.
\section{Training details}
\textit{Training procedure.--\,} The training procedure is designed to ensure that the sign network expresses the primary sign structure such as MSR and the base network expresses the remaining sign structure. The whole training procedure contains the following steps:
\begin{enumerate}[itemsep=0pt, topsep=0pt]
\item No base network is used and all amplitudes are set to 1. The sign network is trained by the ES until the variational energy is converged.
\item The base network is put into the combined network structure and trained by the SR until the variational energy stabilizes.
\item The sign network is again trained by the ES.
\item Both networks are trained simultaneously for several thousands of iterations to convergence.
\item Point symmetry group defined in Appendix~\ref{symmetries} are applied and the base network is trained by the SR.
\end{enumerate}
\textit{Adjustment of base network weights.--\,}
Two possible problems may occur in the base network. Firstly, the output of the base network may overflow since the network expresses $\psi_{\mathrm{b}}$ directly instead of $\log\psi_{\mathrm{b}}$. Secondly, the scale of parameters may be significantly different in different layers, which leads to vanishing or exploding gradients.
Fortunately, one can adjust the weights of the full base network to alleviate the two problems. The architecture of the full base network in Fig.\,\ref{fig:base_network} contains two redundant degrees of freedom. One is free to multiply a number to all variational parameters in the first or second layer. This transformation essentially multiplies an overall constant to the wave function without changing the physical state.
Assuming one multiplies a constant $A$ to the weights in the first layer ${\bm W}^{(1)}$ and another constant $B$ to the weights in the second layer ${\bm W}^{(2)}$, then the transformation can be written in short as
\begin{equation} \label{eq:transform_weight}
{\bm W}^{(1)} \rightarrow A {\bm W}^{(1)}, \quad {\bm W}^{(2)} \rightarrow B {\bm W}^{(2)}.
\end{equation}
This transformation also modifies the neurons in the first layer ${\bm v}^{(1)}$, the neurons in the second layer ${\bm v}^{(2)}$ and the network output $\psi$ by
\begin{equation} \label{eq:transform_neuron}
{\bm v}^{(1)} \rightarrow A {\bm v}^{(1)}, \quad {\bm v}^{(2)} \rightarrow AB {\bm v}^{(2)}, \quad \psi \rightarrow (AB)^n \psi,
\end{equation}
where $n$ is the number of elements in each channel. After each SR step, we hope the maximal wave function entry $M$ in Monte Carlo samples can be normalized to 1, so
\begin{equation} \label{eq:normalize_M}
(AB)^n M = 1.
\end{equation}
On the other hand, the standard deviation of weights also undergoes similar transformation
\begin{equation}
\sigma^{(1)} \rightarrow A\sigma^{(1)}, \quad
\sigma^{(2)} \rightarrow B\sigma^{(2)},
\end{equation}
where $(1)$ and $(2)$ still represent different layers. In our numerical experiments, the network has best performance when $\sigma^{(1)}$ and $\sigma^{(2)}$ are kept equal, so we define their ratio
\begin{equation}
R = \frac{\sigma^{(1)}}{\sigma^{(2)}}
\rightarrow \frac{A}{B} R,
\end{equation}
which is measured in every iteration and modified to 1, giving
\begin{equation} \label{eq:normalize_R}
\frac{A}{B} R = 1
\end{equation}
Combining Eq.\,\eqref{eq:normalize_M} and Eq.\,\eqref{eq:normalize_R}, one obtains
\begin{equation}
A = \sqrt{\frac{1}{R M^{1/n}}}, \quad B = \sqrt{\frac{R}{M^{1/n}}}.
\end{equation}
We apply the transformation Eq.\,\eqref{eq:transform_weight} with the above $A$ and $B$ values to ensure the stability and efficiency of the training process.
\end{document}
|
\section{Introduction}
\label{sec:introduction}
\emph{Quantum information decoupling}~\cite{HOW2005partial,HOW2007quantum,
ADHW2009mother} is the procedure of removing the information of a reference
system from the system under control, via physically permitted operations.
It is a fundamental quantum information processing task, which has found
broad applications, ranging from quantum Shannon theory~\cite{HOW2005partial,
HOW2007quantum,ADHW2009mother,DevetakYard2008exact,YardDevetak2009optimal,
BDHSW2014quantum,BCR2011quantum,BBMW2018conditional,BFW2013quantum} to quantum thermodynamics~\cite{DARDV2011thermodynamic,BrandaoHorodecki2013area,
BrandaoHorodecki2015exponential,Aberg2013truly,DHRW2016relative} to black-hole
physics~\cite{HP2007black,BraunsteinPati2007quantum,BSZ2013better,
KamilChristoph2016one}. Since being introduced in~\cite{HOW2005partial,
HOW2007quantum,ADHW2009mother}, the problem of quantum information decoupling
has attracted continued interest of study from the community. This includes
the study of decoupling in the one-shot setting~\cite{BCR2011quantum,DBWR2014one,
ADJ2017quantum,MBDRC2017catalytic,WakakuwaNakata2021one,Dupuis2021privacy},
the search for more specific and more efficient decoupling operations~\cite{SDTR2013decoupling,BrownFawzi2015decoupling,NHMW2017decoupling},
and the investigation of the speed of asymptotic convergence of the decoupling performance~\cite{Sharma2015random,ABJT2020partially,Dupuis2021privacy}.
In particular, by introducing an independent system as catalyst, tight one-shot
characterization has been derived in~\cite{MBDRC2017catalytic,ADJ2017quantum},
which is able to provide the exact second-order asymptotics.
The \emph{reliability function} was introduced by Shannon in information theory~\cite{Shannon1959probability}. Defined as the rate of exponential decay of
the error with the increasing of blocklength, the reliability function provides
the desired precise characterization of how rapidly an information processing task
approaches the perfect in the asymptotic setting~\cite{Gallager1968information}.
Study of reliability functions in quantum information dates back to the work of
Holevo and Winter~\cite{BurnasherHolevo1998on,Winter1999coding,Holevo2000reliability}
more than two decades ago. In recent years, there has been a growing body of research
in this topic from the quantum community~\cite{Dalai2013lower,Hayashi2015precise,
DalaiWinter2017constant,CHT2019quantum,CHDH2020non,Dupuis2021privacy}. However,
complete characterization of the reliability functions in the quantum regime is
not known, even for classical-quantum channels. Nevertheless, see References~\cite{KoenigWehner2009strong,SharmaWarsi2013fundamental,WWY2014strong,
GuptaWilde2015multiplicativity,CMW2016strong,MosonyiOgawa2017strong,CHDH2020non}
for a partial list of the fruitful results on the strong converse exponent in
the quantum setting, which characterizes how fast a quantum information task
becomes the useless.
In this paper, we investigate the reliability function for the task of
quantum information decoupling in the catalytic setting. We have obtained
the exact formula when the decoupling cost is below a critical value.
Specifically, for a bipartite quantum state $\rho_{RA}$, we consider three
different types of decoupling operations on the $A$ system:
(a) decoupling via removing a subsystem,
(b) decoupling via projective measurement,
(c) decoupling via random unitary operation.
We show that under any of these three types of decoupling operations,
the reliability function is given by the Legendre transformation of the
\emph{sandwiched R\'enyi mutual information}
\[
I_{\alpha}(R:A)_\rho :=
\min_{\sigma_A\in\mc{S}(A)} D_{\alpha}(\rho_{RA} \| \rho_R \ox \sigma_A)
\]
of order $\alpha\in (1,2]$. Here $\mc{S}(A)$ is the set of all quantum
states on system $A$, and
\[D_{\alpha}(M\|N) :=\frac{1}{\alpha-1} \log \tr\big(N^{\frac{1-\alpha}{2\alpha}}MN^{\frac{1-\alpha}{2\alpha}}\big)^\alpha
\]
is the \emph{sandwiched R\'enyi divergence}~\cite{MDSFT2013on,WWY2014strong}.
This result is obtained by deriving respective upper and lower bounds, and
we show that the two bounds coincide when the rate of the decoupling cost is
below the critical value. On the one hand, we analyse the convex-split
lemma of~\cite{ADJ2017quantum} and derive for it a novel bound in terms of the
sandwiched R\'enyi divergence of order $\alpha\in(1,2]$, and this constitutes
the main technical tool for proving the lower bound. On the other hand, the
upper bound is obtained based on an asymptotic analysis of the smoothing
quantity of the max-information, for which we show that the exact exponent
is given by a formula in terms of the sandwiched R\'enyi mutual information
of order $\alpha\in(1,\infty)$. Furthermore, as application, we provide
similar characterization for the reliability function of quantum state
merging by exploiting the inherent connection between quantum state merging
and decoupling.
Our results, along with the concurrent work of~\cite{LYH2021tight} which
addresses different problems, have provided the sandwiched R\'enyi divergence~\cite{MDSFT2013on,WWY2014strong} with a new type of operational
interpretation by showing that it characterizes the exact exponents under which
certain quantum information tasks approach the perfect. This is in stark
contrast to what was previously known that the sandwiched R\'enyi divergence
characterizes the strong converse exponents---the optimal exponential rates
under which the underlying errors go to $1$~\cite{MosonyiOgawa2015quantum,
MosonyiOgawa2015two,CMW2016strong,HayashiTomamichel2016correlation,
MosonyiOgawa2017strong,CHDH2020non}. Therefore, we conclude that the meaning
of this fundamental entropic quantity can be more fruitful than what was
previously understood.
\emph{Relation to previous works.}
In References~\cite{Sharma2015random} and~\cite{Dupuis2021privacy}, exponential
achievability bounds for the decoupling error were given, which are in terms
of the sandwiched R\'enyi divergence of order $\alpha\in(1,2]$, too. However,
these bounds do not seem to be able to yield the optimal exponent in the
asymptotic setting, and thus do not provide much information on the reliability
function. Besides, there is no discuss on the converse bound for the exponent
of the decoupling error there. Another difference between these two works and
the present one is that, in~\cite{Sharma2015random} and~\cite{Dupuis2021privacy}
the decoupling error is measured using the trace distance, while in the present
paper we employ the purified distance, or equivalently, the fidelity function.
The reminder of this paper is organized as follows. In Section~\ref{sec:preliminaries}
we introduce the necessary notation, definitions and some basic properties. In
Section~\ref{sec:problem-results} we present the problem formulation, the main
results, and the application to quantum state merging. The proofs are given in
Section~\ref{sec:proof-main} and Section~\ref{sec:relations}, where in
Section~\ref{sec:proof-main} we prove the characterization of the reliability
functions, and in Section~\ref{sec:relations} we prove the relation between
different types of decoupling as well as the relation between decoupling and
quantum state merging. At last, in Section~\ref{sec:discussion} we conclude the
paper with some discussion and open questions.
\section{Preliminaries}
\label{sec:preliminaries}
\subsection{Notation and basic properties}
Let $\mc{H}$ be a Hilbert space, and $\mc{H}_A$ be the Hilbert space associated
with system $A$. $\mc{H}_{AB}$, denoting the Hilbert space of the composite
system $AB$, is the tensor product of $\mc{H}_A$ and $\mc{H}_B$. We restrict
ourselves to finite-dimensional Hilbert spaces throughout this paper. The
notation $|A|$ stands for the dimension of $\mathcal{H}_A$. We use $\1_A$
to denote the identity operator on $\mc{H}_A$. the notation $\supp(X)$ for
an operator $X$ is used for the support of $X$. The set of unitary operators
on $\mathcal{H}$ is denoted as $\mathcal{U}(\mathcal{H})$, and the set of
positive semidefinite operators on $\mc{H}$ is denoted as $\mathcal{P}(\mc{H})$.
The set of normalized quantum states and subnormalized quantum states on
$\mathcal{H}$ are denoted as $\mathcal{S}(\mathcal{H})$ and $\mathcal{S}_
{\leq}(\mathcal{H})$, respectively. That is,
\begin{align*}
\mc{S}(\mc{H}) &=\{\rho \in \mc{P}(\mc{H}) | \tr\rho=1 \},\\
\mc{S}_{\leq}(\mc{H}) &=\{\rho \in \mc{P}(\mc{H}) | \tr\rho \leq 1 \}.
\end{align*}
If the Hilbert space $\mc{H}$ is associated with system $A$, then the above
notations $\mc{U}(\mc{H})$, $\mc{P}(\mc{H})$, $\mc{S}(\mc{H})$ and
$\mc{S}_{\leq}(\mc{H})$ are also written as $\mc{U}(A)$, $\mc{P}(A)$,
$\mc{S}(A)$ and $\mc{S}_{\leq}(A)$, respectively. The discrete Weyl operators
on a $d$-dimensional Hilbert space $\mc{H}$ with an orthonormal basis
$\{\ket{a}\}_{a=0}^{d-1}$ are a collection of unitary operators
\[
W_{a,b}
=\sum_{c=0}^{d-1}\mathrm{e}^{\frac{2\pi\mathrm{i}bc}{d}}
\ket{(a+c)\!\!\!\mod d}\bra{c},
\]
where $a,b\in \{0,1,\ldots,d-1\}$.
For $X, Y \in \mc{P}(\mc{H})$, we write $X \geq Y$ if $X-Y \in \mc{P}(\mc{H})$
and $X \leq Y$ if $Y-X \in \mc{P}(\mc{H})$. $\{X \geq Y\}$ is the spectral
projection of $X-Y$ corresponding to all non-negative eigenvalues. $\{X>Y\}$,
$\{X \leq Y\}$ and $\{X <Y \}$ are similarly defined.
We use the purified distance~\cite{GLN2005distance, TCR2009fully} to measure
the closeness of a pair of states $\rho, \sigma \in \mathcal{S}_{\leq}(\mathcal{H})$.
The purified distance is defined as $P(\rho,\sigma):=\sqrt{1-F^2(\rho,\sigma)}$,
where
\[
F(\rho,\sigma):=\tr\sqrt{\sqrt{\sigma}\rho\sqrt{\sigma}}
+ \sqrt{(1-\tr\rho)(1-\tr \sigma)}
\]
is the fidelity. The Uhlmann's theorem~\cite{Uhlmann1976transition}, stated as
follows, will play a key role in later proofs. Let $\rho_{AB}\in \mc{S}_{\leq}
(\mc{H_{AB}})$ be a bipartite state, and let $\sigma_{A} \in \mc{S}_{\leq}
(\mc{H_{A}})$. Then there exists an extension $\sigma_{AB}$ of $\sigma_{A}$ such
that $P(\rho_{AB},\sigma_{AB}) = P(\rho_{A},\sigma_{A})$.
A quantum operation or quantum channel $\Phi$ is a linear, completely positive,
and trace-preserving~(CPTP) map acting on quantum states. We denote by $\Phi_
{A \rar B}$ a quantum operation from system $A$ to system $B$. The Stinespring
representation theorem~\cite{Stinespring1955positive} states that there is an
ancillary system $C$ in a pure state $\proj{0}_C$, a system $E$ and a unitary
$U_{AC \rar BE}$ such that $\Phi_{A \rar B}(\rho_A)=\tr_{E} \big(U(\rho_A \ox
\proj{0}_C)U^*\big)$. A quantum measurement is described by a set of positive
semidefinite operators $\{M_x\}_x$ such that $\sum_xM_x=\1$. It outputs $x$ with
probability $\tr(\rho M_x)$ when the underlying state is $\rho$. If a measurement
$\{Q_x\}_x$ is such that all the $Q_x$ are projections onto mutually orthogonal
subspaces, then it is called a projective measurement. We associate each
quantum measurement $\mc{M}=\{M_x\}_x$ with a measurement channel
$\Phi_\mc{M}:\rho\mapsto \sum_x(\tr\rho M_x)\proj{x}$, where
$\{\ket{x}\}$ is an orthonormal basis.
Let $\sigma$ be a self-adjoint operator on $\mc{H}$ with spectral projections
$Q_1, \ldots, Q_{v(\sigma)}$, where $v(\sigma)$ is the number of different
eigenvalues of $\sigma$. The associated pinching map $\mathcal{E}_\sigma$ is
defined as
\[
\mathcal{E}_\sigma : X \rar \sum_i Q_i X Q_i.
\]
The pinching inequality~\cite{Hayashi2002optimal} states that for any $X \in
\mc{P}(\mc{H})$,
\begin{equation}
\label{equ:pinchingineq}
X \leq v(\sigma)\mc{E}_\sigma(X).
\end{equation}
For $n \in \mathbb{N}$, let $S_n$ be the symmetric group of the permutations of
$n$ elements. The set of symmetric states and subnormalized symmetric states on
$\mc{H}_{A^n}$ are defined, respectively, as
\begin{align*}
\mathcal{S}^{\rm{sym}}(A^n) & :=
\big\{\rho_{A^n} | \rho_{A^n} \in \mathcal{S}(A^n),~W_{\pi} \rho_{A^n}
W_{\pi}^*=\rho_{A^n},~\forall \pi \in S_n\big\}, \\
\mathcal{S}^{\rm{sym}}_{\leq}(A^n) & :=
\big\{\rho_{A^n} | \rho_{A^n} \in \mathcal{S}_{\leq}(A^n),~
W_{\pi} \rho_{A^n} W_{\pi}^*=\rho_{A^n},~\forall \pi \in S_n\big\},
\end{align*}
where $W_{\pi} : \ket{\psi_1} \ox \cdots \ox \ket{\psi_n} \mapsto
\ket{\psi_{\pi^{-1}(1)}} \ox \cdots \ox \ket{\psi_{\pi^{-1}(n)}}$ is the natural
representation of $\pi \in S_n$. The set of the symmetric states can be dominated
by a single symmetric state, in the sense of the following Lemma~\ref{lemma:unisym},
and two different constructions are given in~\cite{CKR2009postselection}
and~\cite{Hayashi2009universal}, respectively.
See~\cite[Appendix A]{MosonyiOgawa2017strong} for a detailed proof.
\begin{lemma}
\label{lemma:unisym}
For every Hilbert space $\mathcal{H}_A$ and $n\in\mathbb{N}$, there exists a
universal symmetric state $\omega^{(n)}_{A^n} \in \mc{S}^{\rm{sym}}(A^n)$, such
that for any $\rho_{A^n} \in \mc{S}^{\rm{sym}}(A^n)$ we have
\begin{align*}
\rho_{A^n} &\leq g_{n,|A|} \omega^{(n)}_{A^n}, \\
v(\omega^{(n)}_{A^n}) &\leq (n+1)^{|A|-1},
\end{align*}
where $g_{n,|A|} \leq (n+1)^{\frac{(|A|+2)(|A|-1)}{2}}$ and $v(\omega^{(n)}_{A^n})$
denotes the number of different eigenvalues of $\omega^{(n)}_{A^n}$.
\end{lemma}
Throughout this paper, $\log$ is with base $2$ and $\ln$ is with base $e$.
\subsection{Quantum entropies and information divergences}
The sandwiched R\'enyi divergence has been introduced
in~\cite{MDSFT2013on,WWY2014strong} and is a quantum generalization of the
classical R\'enyi information divergence. For quantum states $\rho, \sigma
\in \mc{S}(\mc{H})$ and a parameter $\alpha\in (0,1)\cup (1,\infty)$, it is
defined as
\[
D_{\alpha}(\rho\|\sigma) :=\frac{1}{\alpha-1} \log \tr\big
(\sigma^{\frac{1-\alpha}{2\alpha}}\rho\sigma^{\frac{1-\alpha}{2\alpha}}\big)^\alpha
\]
if either $\alpha>1$ and $\supp(\rho)\subseteq\supp(\sigma)$ or $\alpha<1$ and $\supp(\rho)\not\perp\supp(\sigma)$, otherwise we set $D_{\alpha}(\rho\|\sigma)
=+\infty$.
For a bipartite quantum state $\rho_{AB} \in \mc{S}(AB)$ and $\alpha\in
(0,1)\cup (1,\infty)$, the sandwiched R\'enyi mutual information of order $\alpha$
is defined as~\cite{WWY2014strong, Beigi2013sandwiched}
\[
I_{\alpha}(A:B)_\rho :=
\min_{\sigma_B\in\mc{S}(B)} D_{\alpha}(\rho_{AB} \| \rho_A \ox \sigma_B),
\]
and we consider a version of the sandwiched R{\'e}nyi conditional
entropy~\cite{MDSFT2013on}
\[
H_{\alpha}(A|B)_\rho:=
-\min_{\sigma_B\in\mc{S}(B)}D_{\alpha}(\rho_{AB} \| \1_A \ox \sigma_B).
\]
when the system $B$ is of dimension $1$, we recover from the sandwiched
R{\'e}nyi conditional entropy the R{\'e}nyi entropy $H_{\alpha}(A)_\rho
:=-D_{\alpha}(\rho_{A} \| \1_A )=\frac{1}{1-\alpha}\log\tr\rho_A^\alpha$.
The quantum relative entropy~\cite{Umegaki1954conditional}
\[
D(\rho \|\sigma) :=
\begin{cases}
\tr(\rho(\log\rho-\log\sigma)) & \text{ if }\supp(\rho)\subseteq\supp(\sigma), \\
+\infty & \text{ otherwise}
\end{cases}
\]
of states $\rho$ and $\sigma$ is the limit of the sandwiched R\'enyi divergence
when $\alpha\rar 1$. In the case $\alpha\rar \infty$, we get the max-relative
entropy~\cite{Datta2009min}
\[
D_{\rm{max}}(\rho \| \sigma) := \inf\{ \lambda~|~\rho \leq 2^\lambda \sigma \}.
\]
The limits of $I_{\alpha}(A:B)_\rho$ and $H_{\alpha}(A|B)_\rho$ when
$\alpha\rar 1$ are the quantum mutual information and quantum conditional entropy,
respectively: for $\rho_{AB} \in \mc{S}(AB)$,
\begin{align*}
I(A:B)_\rho:&=\min_{\sigma_B\in\mc{S}(B)} D(\rho_{AB} \| \rho_A \ox \sigma_B)
=D(\rho_{AB} \| \rho_A \ox \rho_B) ,\\
H(A|B)_\rho:&=-\min_{\sigma_B\in\mc{S}(B)} D(\rho_{AB} \| \1_A \ox \sigma_B)
=-D(\rho_{AB} \| \1_A \ox \rho_B).
\end{align*}
The max-information~\cite{CBR2013smooth} for $\rho_{AB} \in \mc{S}(AB)$,
\[
I_{\rm{max}}(A:B)_\rho
:=\max_{\sigma_B\in\mc{S}(B)}D_{\rm{max}}(\rho_{AB} \| \rho_A \ox \sigma_B),
\]
emerges as the limit of $I_{\alpha}(A:B)_\rho$ when $\alpha\rar\infty$.
\section{Problem Statement and Main Results}
\label{sec:problem-results}
\subsection{Catalytic quantum information decoupling}
Let a bipartite quantum state $\rho_{RA}\in\mc{S}(RA)$ be given. Quantum
information decoupling is the procedure of removing the information about
system $R$ from the $A$ system, by performing a quantum operation on the
$A$ system. In catalytic quantum information decoupling, an auxiliary
system $A'$ in a sate that is independent of the state $\rho_{RA}$ can be
added as a catalyst during the decoupling operation. Readers are referred
to~\cite{MBDRC2017catalytic} for a detailed description of standard and
catalytic decoupling. We consider three different types of decoupling
operations: (a) decoupling via removing a subsystem~\cite{ADHW2009mother},
(b) decoupling via projective measurement~\cite{HOW2007quantum},
(c) decoupling via random unitary operation~\cite{GPW2005quantum}.
\emph{(a) decoupling via removing a subsystem.} For quantum state $\rho_{RA}$,
a catalytic decoupling scheme via removing a subsystem
consists of a catalytic system $A'$ in the state $\sigma_{A'}$ and an
isometry operation $U:\mc{H}_{AA'} \rar \mc{H}_{A_1A_2}$, where $A_2$ is
the system to be removed and $A_1$ is the remaining system. Without loss
of generality, we require that $|AA'|=|A_1A_2|$ and hence $U_{AA'\rar A_1A_2}$
becomes unitary. The cost of the decoupling is given by the number of qubits
that is removed, $\log |A_2|$. The performance is measured by the
purified distance between the remaining state $\tr_{A_2}U(\rho_{RA}\ox
\sigma_{A'})U^*$ and the nearest product state of the form $\rho_R \ox
\omega_{A_1}$. We are interested in the optimal performance when the cost,
namely, the number of removed qubits, is bounded.
\begin{definition}
\label{def:perform-dec}
Let $\rho_{RA}\in \mc{S}(RA)$ be a bipartite quantum state. For a given
size of removed system $k\geq 0$ (in qubits), the optimal performance
of catalytic decoupling via removing a subsystem is given by
\begin{equation}
\label{eq:perform-dec}
P^{\rm dec}_{R:A}(\rho_{RA},k):=
\min P\big(\tr_{A_2}U(\rho_{RA}\ox \sigma_{A'})U^*,\rho_R \ox
\omega_{A_1}\big),
\end{equation}
where the minimization is over all system dimensions $|A'|$, $|A_1|$, $|A_2|$
such that $|AA'|=|A_1A_2|$ and $\log |A_2|\leq k$, all states $\sigma_{A'}\in
\mc{S}(A')$, $\omega_{A_1}\in\mc{S}(A_1)$, and all unitary operations
$U:\mc{H}_{AA'} \rar \mc{H}_{A_1A_2}$.
\end{definition}
\emph{(b) decoupling via projective measurement.} For quantum state $\rho_{RA}$,
a catalytic decoupling scheme via projective measurement consists of a catalytic
system $A'$ in the state $\sigma_{A'}$ and an projective measurement
$\mc{Q}=\{Q^x\}_{x=1}^m$ on the composite system $AA'$. The cost of the decoupling
is accounted by the number (in bits) of measurement outcomes, $\log m$. The
performance is measured by the purified distance between the post-measurement
state $\sum_{x=1}^m Q^x_{AA'}(\rho_{RA}\ox\sigma_{A'})Q^x_{AA'}$ and the nearest
product state of the form $\rho_R \ox\omega_{AA'}$. We are interested in the
optimal performance when the cost, namely, the number of measurement outcomes,
is bounded.
\begin{definition}
\label{def:perform-dec-m}
Let $\rho_{RA}\in \mc{S}(RA)$ be a bipartite quantum state. For a given
number of measurement outcomes $k\geq 0$ (in bits), the optimal performance
of catalytic decoupling via projective measurement is given by
\begin{equation}
\label{eq:perform-dec-m}
P^{\rm{dec\text{-}m}}_{R:A}(\rho_{RA},k):=
\min P\Big(\sum_{x=1}^m Q^x_{AA'}(\rho_{RA}\ox\sigma_{A'})Q^x_{AA'},
\rho_R \ox\omega_{AA'}\Big),
\end{equation}
where the minimization is over all system dimensions $|A'|$, all states
$\sigma_{A'}\in\mc{S}(A')$, $\omega_{AA'}\in\mc{S}(AA')$, and all projective
measurements $\{Q^x_{AA'}\}_{x=1}^m$ such that $\log m \leq k$.
\end{definition}
\emph{(c) decoupling via random unitary operation.} For quantum state $\rho_{RA}$,
a catalytic decoupling scheme via random unitary operation consists of a
catalytic system $A'$ in the state $\sigma_{A'}$ and a random unitary operation
$\Lambda_{AA'}: X\mapsto \frac{1}{m}\sum_{i=1}^{m} U_i X U_i^*$ acting on the
composite system $AA'$. The cost of the decoupling is accounted by the number
(in bits) of unitary operators in $\Lambda_{AA'}$, $\log m$. The performance is
measured by the purified distance between the resulting state $\Lambda_{AA'}
(\rho_{RA}\ox\sigma_{A'})$ and the nearest product state of the form $\rho_R
\ox\omega_{AA'}$. We are interested in the optimal performance when the cost,
namely, the number of unitary operators, is bounded.
\begin{definition}
\label{def:perform-dec-u}
Let $\rho_{RA}\in \mc{S}(RA)$ be a bipartite quantum state. For a given
number of unitary operators $k\geq 0$ (in bits), the optimal performance
of catalytic decoupling via random unitary operation is given by
\begin{equation}
\label{eq:perform-dec-u}
P^{\rm{dec\text{-}u}}_{R:A}(\rho_{RA},k):=
\min P\Big(\Lambda_{AA'}(\rho_{RA} \ox \sigma_{A'}),\rho_R \ox\omega_{AA'}\Big),
\end{equation}
where the minimization is over all system dimensions $|A'|$, all states
$\sigma_{A'}\in \mc{S}(A')$, $\omega_{AA'} \in \mc{S}(AA')$ and all random
unitary operations $\Lambda_{AA'}(\cdot)=\frac{1}{m}\sum_{i=1}^{m} U_i (\cdot)
U_i^*$ with $U_i \in\mc{U}(AA')$ such that $\log m \leq k$.
\end{definition}
The reliability function of quantum information decoupling characterizes the
the speed at which perfect decoupling can be approached in the asymptotic
setting, in which the underlying bipartite quantum state is in the form of
tensor product of $n$ identical copies. Specifically, it is the rate of
exponential decreasing of the optimal performance, as a function of the
cost.
\begin{definition}
\label{def:reliability-dec}
Let $\rho_{RA}\in \mc{S}(RA)$ be a bipartite quantum state, and $r\geq 0$.
The reliability functions of catalytic quantum information
decoupling for the state $\rho_{RA}$, via the three different types of decoupling
operations described above, are defined respectively as
\begin{align}
E^{\rm dec}_{R:A}(\rho_{RA},r) &:=\limsup_{n\rar\infty}\frac{-1}{n} \log
P^{\rm dec}_{R^n:A^n}\big(\rho_{RA}^{\ox n},nr\big),
\label{eq:reliability-dec} \\
E^{\rm dec\text{-}m}_{R:A}(\rho_{RA},r)&:=\limsup_{n\rar\infty}\frac{-1}{n} \log
P^{\rm{dec\text{-}m}}_{R^n:A^n}\big(\rho_{RA}^{\ox n},nr\big),
\label{eq:reliability-dec-m} \\
E^{\rm dec\text{-}u}_{R:A}(\rho_{RA},r)&:=\limsup_{n\rar\infty}\frac{-1}{n} \log
P^{\rm{dec\text{-}u}}_{R^n:A^n}\big(\rho_{RA}^{\ox n},nr\big).
\label{eq:reliability-dec-u}
\end{align}
\end{definition}
\subsection{Main results}
At first, we show in the following Proposition~\ref{prop:dec-relation} equalities
that relate the optimal performances or the reliability functions for the three
different decoupling operations. With this, we are able to deal with them in a
unified way.
\begin{proposition}
\label{prop:dec-relation}
For $\rho_{RA} \in \mathcal{S}(RA)$ and $k,r \geq 0$, we have
\begin{align}
&P_{R:A}^{\emph{dec}}(\rho_{RA}, k)
=P_{R:A}^{\rm dec\text{-}m}(\rho_{RA}, 2k)
=P_{R:A}^{\rm dec\text{-}u}(\rho_{RA}, 2k), \label{eq:dec-relation-P} \\
&E_{R:A}^{\emph{dec}}(\rho_{RA}, r)
=E_{R:A}^{\rm dec\text{-}m}(\rho_{RA}, 2r)
=E_{R:A}^{\rm dec\text{-}u}(\rho_{RA}, 2r). \label{eq:dec-relation-E}
\end{align}
\end{proposition}
Then, we derive a one-shot achievability bound for the performance of
catalytic quantum information decoupling. The bound is given in terms
of the sandwiched R\'enyi information divergence. In doing so, we have
employed the convex-split lemma~\cite{ADJ2017quantum} as a key technical
tool. While we present the bound only for the case of decoupling via
removing a subsystem in the following Theorem~\ref{thm:dec-p-up},
similar results hold for the other two cases in light of
Proposition~\ref{prop:dec-relation}.
\begin{theorem}
\label{thm:dec-p-up}
Let $\rho_{RA}\in \mc{S}(RA)$. For any $m\in\mathbb{N}$, $0<s\leq 1$ and
$\sigma_A\in\mc{S}(A)$, the optimal performance of decoupling $A$ from $R$
is bounded as
\[
P^{\rm dec}_{R:A}(\rho_{RA},\log m) \leq
\sqrt{\frac{v^s}{s}}\exp\Big\{-(\ln2)\ s\big(\log m-\frac{1}{2}D_{1+s}
(\rho_{RA}\| \rho_R\ox\sigma_A)\big)\Big\},
\]
where $v$ is the number of distinct eigenvalues of $\rho_R\ox \sigma_A$.
\end{theorem}
Our main result is the characterization of reliability functions.
This is given in Theorem~\ref{thm:reliability-dec} for the case of decoupling
via removing a subsystem. Thanks to Proposition~\ref{prop:dec-relation},
similar results follow directly for the other two cases. We mention that
we have completely determined the reliability functions when the respective
cost is below a critical value.
\begin{theorem}
\label{thm:reliability-dec}
Let $\rho_{RA}\in \mc{S}(RA)$ be a bipartite quantum state, and consider
the problem of decoupling quantum information in $A^n$ from the
reference system $R^n$ for the quantum state $\rho_{RA}^{\ox n}$. When
$r\leq R_{\rm critical}:=\frac{1}{2}\frac{\mathrm{d}}
{\mathrm{d}s} sI_{1+s}(R:A)_{\rho}\big|_{s=1}$, we have
\beq\label{eq:exp-dec}
E^{\rm dec}_{R:A}(\rho_{RA},r) = \max_{0\leq s\leq 1}
\Big\{s\big(r-\frac{1}{2}I_{1+s}(R:A)_\rho\big)\Big\}.
\eeq
In general, we have
\begin{align}
E^{\rm dec}_{R:A}(\rho_{RA},r) &\geq \max_{0\leq s\leq 1}
\Big\{s\big(r-\frac{1}{2}I_{1+s}(R:A)_\rho\big)\Big\}, \label{eq:exp-dec-low} \\
E^{\rm dec}_{R:A}(\rho_{RA},r) &\leq \ \sup_{s \geq 0}
\ \, \Big\{s\big(r-\frac{1}{2}I_{1+s}(R:A)_\rho\big)\Big\}.\label{eq:exp-dec-up}
\end{align}
\end{theorem}
The lower bound of Eq.~\eqref{eq:exp-dec-low} is a consequence of
Theorem~\ref{thm:dec-p-up}. For the upper bound of Eq.~\eqref{eq:exp-dec-up},
we first bound the optimal decoupling performance using the smoothing quantity
associated with the max-information, and then we derive the exact exponent in
smoothing the max-information. Eq.~\eqref{eq:exp-dec} is derived from the
combination of Eqs.~\eqref{eq:exp-dec-low} and \eqref{eq:exp-dec-up}.
The results presented in Theorem~\ref{thm:reliability-dec} are depicted
in Figure~\ref{fig:reliability}. Above the critical value, we are unable
to determine the formula for the reliability function. This is indeed a
common hard problem in the topic of reliability functions (see, e.g., References~\cite{Gallager1968information,Dalai2013lower,
HayashiTan2016equivocations,LYH2021tight}). More comments on
Theorem~\ref{thm:reliability-dec} can be found in Remark~\ref{rem:mainresult}.
\begin{figure}[ht]
\includegraphics[width=9.5cm]{reliability.eps}
\caption{Reliability function of quantum information decoupling.
$E_u(r):=\sup_{s \geq 0}
\big\{s\big(r-\frac{1}{2}I_{1+s}(R:A)_\rho\big)\big\}$
is the upper bound of Eq.~(\ref{eq:exp-dec-up}).
$E_l(r):=\sup_{0\leq s\leq1}
\big\{s\big(r-\frac{1}{2}I_{1+s}(R:A)_\rho\big)\big\}$
is the lower bound of Eq.~(\ref{eq:exp-dec-low}). The
two bounds are equal in the interval $[0,R_\text{critical}]$,
giving the exact reliability function. The reliability function
equals $0$ when $r<\frac{1}{2}I(R:A)_\rho$ and it is strictly
positive when $r>\frac{1}{2}I(R:A)_\rho$. Above the critical value
$R_\text{critical}$, the upper bound $E_u(r)$ becomes larger than
the lower bound and it is $\infty$ when $r>\frac{1}{2}I_{\rm{max}}
(R:A)_\rho$, while the lower bound $E_l(r)$ becomes linear and
reaches $\log |A|-\frac{1}{2}I_2(R:A)_\rho$ at $r=\log |A|$.}
\label{fig:reliability}
\end{figure}
\subsection{Applications to quantum state merging}
\label{subsec:merging}
The inherent connection between quantum information decoupling and quantum
state merging has been established since the invention of these two tasks~\cite{HOW2005partial,HOW2007quantum,ADHW2009mother}. It was further
explored later on, e.g., in~\cite{DBWR2014one,ADJ2017quantum,MBDRC2017catalytic,
ABJT2020partially}. Exploiting this connection, we are able to extend the
results on the reliability function of quantum information decoupling to
the scenario of quantum state merging.
Let $\rho_{RAB}$ be a tripartite pure state. Alice, Bob and a referee hold
system $A$, $B$ and $R$ respectively. Quantum state merging is the task of
transmitting the quantum information stored in the $A$ system from Alice to
Bob. There are two different ways to achieve this. One is by classical
communication and is introduced in~\cite{HOW2005partial}. The other one is
by quantum communication, firstly considered in~\cite{ADHW2009mother}.
In both cases, free pre-shared entanglement between Alice and Bob is allowed.
A formal description is as follows.
A quantum state merging protocol via quantum communication, $\mathcal{M}_1$,
consists of using a shared bipartite entangled pure state $\phi_{A'B'}$, Alice
applying local unitary $U_{AA' \rightarrow A_1A_2}$ and sending the system $A_2$
to Bob, Bob applying local unitary $V_{A_2BB' \rightarrow ABB_1}$ and they
discarding the systems $A_1$ and $B_1$. A CPTP map $\mathcal{M}_2$ is a quantum
state merging protocol via classical communication if it consists of using a
shared bipartite entangled pure state $\psi_{A'B'}$, applying local operation at
Alice's side, sending $k$ classical bits from Alice to Bob and applying local
operation to reproduce systems $A$ and $B$ at Bob's side.
The performances of both protocols are given by the purified distance between
$\rho_{RAB}$ and the final state on the referee's system $R$ and Bob's system
$A$ and $B$. The cost of state merging that we are concerned with, is the number
of qubits ($\log |A_2|$ in $\mathcal{M}_1$) or classical bits ($k$ in $\mc{M}_2$)
that Alice sends to Bob.
\begin{definition}
Let $\rho_{RAB} \in \mathcal{S}({RAB})$ be a tripartite pure state and
$r \geq 0$. Let $P_{A \Rightarrow B}^{\rm{merg}}(\rho_{RAB}, k)$ denote
the optimal performance of quantum state merging via quantum communication of at
most $k$ qubits, and let $P_{A \rightarrow B}^{\rm{merg}}(\rho_{RAB}, k)$ denote
the optimal performance of quantum state merging via classical communication of
at most $k$ bits. They are defined respectively as
\begin{align}
\label{eq:p-merg-q}
&P_{A \Rightarrow B}^{\rm{merg}}(\rho_{RAB}, k)
:= \min_{\mathcal{M}_1} P(\mathcal{M}_1(\rho_{RAB}), \rho_{RAB} ),\\
\label{eq:p-merg-c}
&P_{A \rightarrow B}^{\rm{merg}}(\rho_{RAB}, k)
:= \min_{\mathcal{M}_2} P(\mathcal{M}_2(\rho_{RAB}), \rho_{RAB}),
\end{align}
where $\mathcal{M}_1$ is the protocol via quantum communication and the
minimization in Eq.~\eqref{eq:p-merg-q} is over all such protocols
whose communication cost is bounded by $k$ qubits, and $\mathcal{M}_2$
is the protocol via classical communication and the minimization in
Eq.~\eqref{eq:p-merg-c} is over all possible $\mathcal{M}_2$ whose
communication cost is bound by $k$ bits.
\end{definition}
The reliability function of quantum state merging characterizes the rate of
exponential decreasing of the optimal performance in the asymptotic limit.
\begin{definition}
Let $\rho_{RAB}\in\mathcal{S}({RAB})$ be a tripartite pure state and $r\geq 0$.
The reliability functions $E_{A \Rightarrow B}^{\rm{merg}}(\rho_{RAB}, r)$
and $E_{A \rightarrow B}^{\rm{merg}}(\rho_{RAB}, r)$ of quantum state merging
via quantum communication and classical communication respectively, are defined
as
\begin{align}
&E_{A \Rightarrow B}^{\rm{merg}}(\rho_{RAB}, r)
:=\limsup_{n\rightarrow\infty} \frac{-1}{n}
\log P_{A^n \Rightarrow B^n}^{\rm{merg}}(\rho_{RAB}^{\ox n}, nr),\\
&E_{A \rightarrow B}^{\rm{merg}}(\rho_{RAB}, r)
:=\limsup_{n\rightarrow\infty} \frac{-1}{n}
\log P_{A ^n\rightarrow B^n}^{\rm{merg}}(\rho_{RAB}^{\ox n}, nr).
\end{align}
\end{definition}
To apply our results on decoupling to the problem of quantum state merging,
we show exact equalities relating the optimal performances (or reliability
functions) of catalytic decoupling and quantum state merging.
Eq.~\eqref{perf-merg-dec} in Proposition~\ref{prop:merging-dec} is
essentially due to Uhlmann's theorem and is implicitly used in previous
works~\cite{HOW2005partial,HOW2007quantum,ADHW2009mother,DBWR2014one,
ADJ2017quantum,MBDRC2017catalytic,ABJT2020partially}. However, our
definition of the optimal decoupling performance
(cf. Definition~\ref{def:perform-dec}) is subtle to enable such an
equality relation.
\begin{proposition}
\label{prop:merging-dec}
For a tripartite pure state $\rho_{RAB} \in \mathcal{S}({RAB})$ and
$k\geq 0$, $r \geq 0$, we have
\begin{align}
&P^{\rm{merg}}_{A \Rightarrow B}(\rho_{RAB}, k)
=P^{\rm{merg}}_{A \rightarrow B}(\rho_{RAB}, 2k)
=P^{\rm{dec}}_{R:A}(\rho_{RA}, k), \label{perf-merg-dec} \\
&E^{\rm{merg}}_{A \Rightarrow B}(\rho_{RAB}, r)
=E^{\rm{merg}}_{A \rightarrow B}(\rho_{RAB}, 2r)
=E^{\rm{dec}}_{R:A}(\rho_{RA}, r). \label{reliab-merg-dec}
\end{align}
\end{proposition}
With Proposition~\ref{prop:merging-dec}, we immediately obtain results
in analogy to Theorem~\ref{thm:dec-p-up} and Theorem~\ref{thm:reliability-dec}
for quantum state merging. We do not lay them out entirely but only exhibit
the following.
\begin{corollary}
\label{cor:reliability-merging}
Let $\rho_{RAB}\in\mathcal{S}({RAB})$ be a tripartite pure state. When the
rate of qubits transmission $r$ is such that $0\leq r \leq R_{\rm critical}
\equiv\frac{1}{2}\frac{\mathrm{d}}{\mathrm{d}s} sI_{1+s}(R:A)_{\rho}\big|_{s=1}$,
the reliability function of quantum state merging via quantum communication
is given by
\beq\label{eq:exp-merging-q}
E^{\rm{merg}}_{A \Rightarrow B}(\rho_{RAB}, r)
=\max_{0\leq s\leq 1}\Big\{s\big(r-\frac{1}{2}I_{1+s}(R:A)_\rho\big)\Big\};
\eeq
when the rate of classical bits transmission $r$ is such that $0\leq
\frac{r}{2}\leq R_{\rm critical}\equiv\frac{1}{2}\frac{\mathrm{d}}{\mathrm{d}s}
sI_{1+s}(R:A)_{\rho}\big|_{s=1}$, the reliability function of quantum state
merging via classical communication is given by
\beq\label{eq:exp-merging-c}
E^{\rm{merg}}_{A \rightarrow B}(\rho_{RAB}, r)
=\max_{0\leq s\leq 1}\Big\{\frac{s}{2}\big(r-I_{1+s}(R:A)_\rho\big)\Big\}.
\eeq
\end{corollary}
A special case where Bob holds no side information is worth looking at. It
can be understood as entanglement-assisted quantum source coding. In this case,
Alice and a referee share a pure state $\rho_{RA}$ and Alice wants to send
the information in the $A$ system to Bob with the assistance of unlimited
entanglement and using noiseless quantum or classical communication. It has
been shown in~\cite[Lemma 8]{WWW2019quantifying}
(see also~\cite[Proposition 4.86]{KhatriWilde2020principles}) that
\[
I_{1+s}(R:A)_\rho=\frac{2s+1}{s}\log\tr \rho_A^{\frac{1}{2s+1}}
=2H_{\frac{1}{2s+1}}(\rho_A),
\]
which can also be verified by employing the dual relation
of~\cite[Lemma 6]{HayashiTomamichel2016correlation}.
So the reliability function in this case has a simpler
formula involving only the R\'enyi entropy of one single system.
\section{Proof of the Characterization of Reliability Functions}
\label{sec:proof-main}
In this section, we prove Theorem~\ref{thm:reliability-dec} on the
characterization of the reliability functions. As an intermediate step,
we also derive the one-shot achievability bound of Theorem~\ref{thm:dec-p-up}.
The proof is organized as follows. In Section~\ref{subsec:convex-split}, we
analyse the convex-split lemma~\cite{ADJ2017quantum}, obtaining a new bound
that employs the sandwiched R{\'e}nyi divergence. This will be a crucial tool
for proving Theorem~\ref{thm:reliability-dec}, regarding the achievability
bound of Eq.~\eqref{eq:exp-dec-low}. Then in Section~\ref{subsec:MI-CE}, we
derive the exact exponent for the asymptotic decreasing of the smoothing
quantity for the max-information. This will serve as another key technical
tool for proving Theorem~\ref{thm:reliability-dec}, for the converse bound of
Eq.~\eqref{eq:exp-dec-up}. At last, in Section~\ref{subsec:thm-proof}, we
accomplish the proof of Theorem~\ref{thm:reliability-dec} as well as
Theorem~\ref{thm:dec-p-up}, by employing the established tools mentioned above.
\subsection{A convex-split lemma}
\label{subsec:convex-split}
The convex-split lemma was introduced in~\cite{ADJ2017quantum} and has broad
applications in topics such as one-shot quantum Shannon theory~\cite{ADJ2017quantum,
MBDRC2017catalytic,AJW2017generalized,AJW2018building,AJW2019convex}, entanglement
and general resource theories~\cite{AHJ2018quantifying,BertaMajenz2018disentanglement},
and quantum thermodynamics~\cite{FBB2021thermodynamic,LipkaSkrzypczyk2021all}.
Roughly speaking, it quantifies how well the information of a distinct object
located among many other identical ones, can be erased by randomly mixing all of
them. Originally in~\cite{ADJ2017quantum} and in all of the previous applications,
the effect of this erasure is bounded using the max-relative entropy, which, after
being smoothed, is sufficiently tight for those purposes. However, for our purpose
of deriving the reliability function, this bound does not work any more. Instead,
we prove a version of the convex-split lemma with a new bound, employing directly
the sandwiched R{\'e}nyi divergence.
\begin{lemma}
\label{lem:convex-split}
Let $\rho_{RA}\in \mc{S}(RA)$ and $\sigma_A\in \mc{S}(\mc{H}_A)$
be quantum states such that $\supp (\rho_A)\subseteq\supp (\sigma_A)$.
Consider the following state
\[
\tau_{RA_1A_2\cdots A_m}:=\frac{1}{m}\sum_{i=1}^m
\rho_{RA_i}\ox \big[\sigma^{\ox (m-1)}\big]_{A^m/A_i},
\]
where $A^m/A_i$ denotes the composite system consisting of $A_1,A_2,\cdots,
A_{i-1},A_{i+1},\cdots,A_m$ and $\big[\sigma^{\ox (m-1)}\big]_{A^m/A_i}$
is the product state $\sigma^{\ox (m-1)}$ on these $m-1$ systems. Let
$v=v(\rho_R \ox \sigma_A)$ denote the number of distinct eigenvalues of
$\rho_R \ox \sigma_A$. Then for any $0<s\leq 1$,
\[
D\left(\tau_{RA_1A_2\cdots A_m}\big\| \rho_R \ox (\sigma^{\ox m})_{A^m}\right)
\leq \frac{v^s}{(\ln2)s}\exp\Big\{-(\ln2)\ s\big(\log m-D_{1+s}
(\rho_{RA}\| \rho_R\ox\sigma_A)\big)\Big\}.
\]
\end{lemma}
\begin{proof}
We use the shorthand $\xi_i \equiv \rho_{RA_i}\ox \big[\sigma^{\ox (m-1)}
\big]_{A^m/A_i}$ for simplicity and start with
\begin{align*}
& D\Big(\tau_{RA_1A_2\cdots A_m}\big\| \rho_R \ox (\sigma^{\ox m})_{A^m}\Big) \\
= &\tr\Big[\Big(\frac{1}{m}\sum_{i}\xi_i\Big) \Big(\log\big(\frac{1}{m}
\sum_{i}\xi_i\big)-\log\big(\rho_R \ox (\sigma^{\ox m})_{A^m}\big)\Big)\Big] \\
= &\tr\Big[\xi_1\Big(\log\big(\frac{1}{m}\sum_{i}\xi_i\big)-
\log\big(\rho_R \ox (\sigma^{\ox m})_{A^m}\big)\Big)\Big] \\
= &\tr\Big[\xi_1\Big(\log\xi_1-\log\big(\rho_R\ox (\sigma^{\ox m})_{A^m}\big)\Big)\Big]
-\tr\Big[\xi_1\Big(\log\xi_1-\log\big(\frac{1}{m}\sum_{i}\xi_i\big)\Big)\Big] \\
= &D\Big(\rho_{RA_1}\ox (\sigma^{\ox (m-1)})_{A_2\cdots A_m}\big\|\rho_R\ox (\sigma^{\ox m})_{A^m}\Big)
- D\Big(\rho_{RA_1}\ox (\sigma^{\ox (m-1)})_{A_2\cdots A_m}\big\| \frac{1}{m}\sum_{i}\xi_i\Big) \\
\leq & D\big(\rho_{RA_1} \| \rho_R \ox \sigma_{A_1}\big) - D\Big(\rho_{RA_1}\big\|
\frac{1}{m}\rho_{RA_1}+\frac{m-1}{m}\rho_{R}\ox\sigma_{A_1}\Big) \\
=&\tr\Big[\rho_{RA}\Big(\log\big(\frac{1}{m}\rho_{RA}+\frac{m-1}{m}
\rho_{R}\ox\sigma_{A}\big)-\log(\rho_{R}\ox\sigma_{A})\Big)\Big],
\end{align*}
where the third line is due to the symmetry of the states
$\frac{1}{m}\sum_{i}\xi_i$ and $\rho_R \ox (\sigma^{\ox m})_{A^m}$ over systems
$A_1,A_2,\cdots,A_m$, and for the inequality we have used the data
processing inequality for relative entropy under partial trace. Now
employ the pinching map $\mc{E}_{\rho_{R}\ox\sigma_{A}}$ and write $\bar{\rho}_{RA}:=\mc{E}_{\rho_{R}\ox\sigma_{A}}(\rho_{RA})$.
The pinching inequality together with the operator monotonicity of the
logarithm gives
\[
\log\Big(\frac{1}{m}\rho_{RA}+\frac{m-1}{m}\rho_{R}\ox\sigma_{A}\Big)\leq
\log\Big(\frac{v}{m}\bar{\rho}_{RA}+\rho_{R}\ox\sigma_{A}\Big).
\]
Making use of this, we proceed as follows.
\[\begin{split}
& D\Big(\tau_{RA_1A_2\cdots A_m}\big\| \rho_R \ox (\sigma^{\ox m})_{A^m}\Big) \\
\leq & \tr\left[\rho_{RA}\left(\log\big(\frac{v}{m}\bar{\rho}_{RA}+
\rho_{R}\ox\sigma_{A}\big)-\log(\rho_{R}\ox\sigma_{A})\right)\right] \\
= & \tr\left[\bar{\rho}_{RA}\log\left(\frac{v}{m}\bar{\rho}_{RA}
(\rho_{R}\ox\sigma_{A})^{-1}+\1_{RA}\right)\right] \\
\leq & \frac{1}{(\ln2)s}\tr\Big[\bar{\rho}_{RA} \Big( \frac{v^s}{m^s}
(\bar{\rho}_{RA})^s (\rho_{R}\ox\sigma_{A})^{-s}\Big)\Big] \\
= & \frac{v^s}{(\ln2)s}\exp\Big\{-(\ln2)\ s\big(\log m-D_{1+s}
(\bar{\rho}_{RA}\| \rho_R\ox\sigma_A)\big)\Big\} \\
\leq & \frac{v^s}{(\ln2)s}\exp\Big\{-(\ln2)\ s\big(\log m-D_{1+s}
(\rho_{RA}\| \rho_R\ox\sigma_A)\big)\Big\},
\end{split}\]
where for the third line note that the density matrices $\bar{\rho}_{RA}$
and $\rho_{R}\ox\sigma_{A}$ commute, for the fourth line we have used the
inequality $\ln(1+x)\leq \frac{1}{s}x^s$ for $x\geq 0$ and $0<s\leq 1$,
and the last line is by the data processing inequality for the sandwiched
R{\'e}nyi divergence~\cite{WWY2014strong, MDSFT2013on, FrankLieb2013monotonicity,
Beigi2013sandwiched}.
\end{proof}
\subsection{Smoothing of max-information and conditional min-entropy}
\label{subsec:MI-CE}
Recall that the smooth max-relative entropy is defined for $\rho \in
\mc{S}(\mc{H})$, $\sigma \in \mc{P}(\mc{H})$, and $0\leq\delta<1$
as~\cite{Datta2009min}
\[
D^\delta_{\rm max}(\rho \| \sigma):=\min\left\{\lambda:\exists\,\tilde{\rho}\in
\mc{S}_{\leq}(\mc{H}) {\rm \ \ s.t.\ \ } P(\tilde{\rho}, \rho) \leq \delta,
\,\tilde{\rho} \leq 2^\lambda\sigma\right\}.
\]
In~\cite{LYH2021tight} we have introduced the smoothing quantity for
the max-relative entropy
\[
\delta(\rho \| \sigma,\lambda):=\min\left\{ P(\tilde{\rho}, \rho): \tilde{\rho}\in
\mc{S}_{\leq}(\mc{H}),\,\tilde{\rho} \leq 2^\lambda\sigma\right\}
\]
and obtained
\begin{equation}
\label{eq:exp-mre}
\lim_{n\rightarrow\infty} \frac{-1}{n}\log \delta\big(\rho^{\ox n}\big\|
\sigma^{\ox n}, nr\big) = \frac{1}{2}\sup_{s \geq 0}\left\{ s\big(r-D_{1+s}
(\rho \| \sigma)\big)\right\}.
\end{equation}
In this section, we are interested in the smooth max-information and the
smooth conditional min-entropy defined for $\rho_{AB}\in\mc{S}(AB)$ and
$0\leq\delta<1$, respectively as~\cite{Renner2005security,BCR2011quantum,
ABJT2020partially}
\begin{align*}
I^\delta_{\rm max}(A:B)_\rho
&:=\ \min_{\sigma_B\in\mc{S}(B)}D^\delta_{\rm max}(\rho_{AB}\|\rho_A\ox\sigma_B),\\
H^\delta_{\rm min}(A|B)_\rho
&:=-\min_{\sigma_B\in\mc{S}(B)} D^\delta_{\rm max}(\rho_{AB}\|\1_A\ox\sigma_B).
\end{align*}
Our purpose is to derive the asymptotic exponents for the smoothing of these
two one-shot entropies. Note that in the classical case where $\rho_{AB}$
is a probability distribution, this kind of exponential analysis was done
for the conditional min-entropy based on the trace distance~\cite{Hayashi2016security}.
In doing so, we define the corresponding smoothing quantities.
\begin{definition}
\label{def:smoothing-q}
Let $\rho_{AB}\in\mc{S}(AB)$ and $\lambda\in\mathbb{R}$. The smoothing quantity
for the max-information and for the conditional min-entropy is defined,
respectively, as
\begin{align}
\label{eq:def-smi}
\delta_{A:B}(\rho_{AB},\lambda)&:=
\min_{\sigma_B\in\mc{S}(B)} \delta(\rho_{AB}\|\rho_A\ox\sigma_B,\lambda),\\
\label{eq:def-sce}
\delta_{A|B}(\rho_{AB},\lambda)&:=
\min_{\sigma_B\in\mc{S}(B)} \delta(\rho_{AB}\|\1_A\ox\sigma_B,-\lambda).
\end{align}
\end{definition}
Note that alternative expressions for the smoothing quantities are
\begin{align}
\label{eq:alt-smi}
\delta_{A:B}(\rho_{AB},\lambda)&=
\min\left\{\delta\in\mathbb{R}:I^\delta_{\rm max}(A:B)_\rho\leq\lambda\right\},\\
\label{eq:alt-sce}
\delta_{A|B}(\rho_{AB},\lambda)&=
\min\left\{\delta\in\mathbb{R}:H^\delta_{\rm min}(A|B)_\rho\geq\lambda\right\},
\end{align}
from which the relation to the smooth max-information and smooth conditional
min-entropy is easily seen. The main result is stated in the following.
\begin{theorem}
\label{theorem:exp-mice}
For $\rho_{AB}\in\mc{S}(AB)$ and $r\in\mathbb{R}$, we have
\begin{align}
\label{eq:exp-mi}
\lim_{n\rar\infty}\frac{-1}{n}\log\,\delta_{A^n:B^n}\left(\rho_{AB}^{\ox n},nr\right)
&=\frac{1}{2}\sup_{s \geq 0}\left\{ s\big(r-I_{1+s}(A:B)_\rho\big)\right\},\\
\label{eq:exp-ce}
\lim_{n\rar\infty}\frac{-1}{n}\log\,\delta_{A^n|B^n}\left(\rho_{AB}^{\ox n},nr\right)
&=\frac{1}{2}\sup_{s \geq 0}\left\{ s\big(H_{1+s}(A|B)_\rho-r\big)\right\}.
\end{align}
\end{theorem}
\begin{proof}
Since the r.h.s. of Eqs.~(\ref{eq:def-smi}) and~(\ref{eq:def-sce}) are
similar, it suffices to prove, for any $M\in\mc{P}(A)$ and $r\in\mathbb{R}$,
\beq\label{eq:exp-mices}
\lim_{n\rar\infty}\frac{-1}{n}\log\!\min_{\sigma^{(n)}_{B^n}\in\mc{S}(B^n)}\!\!
\delta\left(\rho_{AB}^{\ox n}\big\|M_A^{\ox n}\ox\sigma^{(n)}_{B^n},nr\right)
\!=\!\frac{1}{2}\sup_{s \geq 0}\left\{s\big(r-\!\!\!\min_{\sigma_B\in\mc{S}(B)}
\!\!\!D_{1+s}(\rho_{AB}\|M_A\ox\sigma_B)\big)\right\}.
\eeq
Substituting $\rho_A$ for $M_A$ results in Eq.~(\ref{eq:exp-mi}), and the
substitution of $M_A\leftarrow\1$ and $r\leftarrow -r$ recovers
Eq.~(\ref{eq:exp-ce}).
Observing that
\[
\min_{\sigma^{(n)}_{B^n}\in\mc{S}(B^n)}
\delta\left(\rho_{AB}^{\ox n}\big\|M_A^{\ox n}\ox\sigma^{(n)}_{B^n},nr\right)
\leq \min_{\sigma_B^{\ox n}\in\mc{S}(B^n)}
\delta\left(\rho_{AB}^{\ox n}\big\|M_A^{\ox n}\ox\sigma_B^{\ox n},nr\right),
\]
we obtain the "$\geq$" part of Eq.~(\ref{eq:exp-mices}) by invoking
Eq.~(\ref{eq:exp-mre}). That is,
\beq\label{eq:exp-mices-low}
\liminf_{n\rar\infty}\frac{-1}{n}\log\!\min_{\sigma^{(n)}_{B^n}\in\mc{S}(B^n)}\!\!
\delta\left(\rho_{AB}^{\ox n}\big\|M_A^{\ox n}\ox\sigma^{(n)}_{B^n},nr\right)
\!\geq\!\frac{1}{2}\sup_{s \geq 0}\left\{s\big(r-\!\!\!\min_{\sigma_B\in\mc{S}(B)}
\!\!\!D_{1+s}(\rho_{AB}\|M_A\ox\sigma_B)\big)\right\}.
\eeq
Now we turn to the proof of the opposite direction. At first, we have
\beq\begin{split}\label{eq:mices-1}
&\min_{\sigma^{(n)}_{B^n}\in\mc{S}(B^n)}\delta\left(\rho_{AB}^{\ox n}
\big\|M_A^{\ox n}\ox\sigma^{(n)}_{B^n},nr\right) \\
=&\min\left\{P\big(\rho_{AB}^{\ox n}, \gamma^{(n)}_{A^nB^n}\big):
\gamma^{(n)}_{A^nB^n}\in\mc{S}_\leq(A^nB^n), \big(\exists\ \sigma^{(n)}_{B^n}\in\mc{S}(B^n)\big)\
\gamma^{(n)}_{A^nB^n}\leq 2^{nr}M_A^{\ox n}\ox\sigma^{(n)}_{B^n}\right\} \\
=&\min\left\{P\big(\rho_{AB}^{\ox n}, \gamma^{(n)}_{A^nB^n}\big):
\gamma^{(n)}_{A^nB^n}\in\mc{S}^{\rm sym}_\leq(A^nB^n), \big(\exists\ \sigma^{(n)}_{B^n}
\in\mc{S}^{\rm sym}(B^n)\big)\
\gamma^{(n)}_{A^nB^n}\leq 2^{nr}M_A^{\ox n}\ox\sigma^{(n)}_{B^n}\right\} \\
\geq &\min\left\{P\big(\rho_{AB}^{\ox n}, \gamma^{(n)}_{A^nB^n}\big):
\gamma^{(n)}_{A^nB^n}\in\mc{S}_\leq(A^nB^n),
\gamma^{(n)}_{A^nB^n}\leq 2^{nr}g_{n, |B|}M_A^{\ox n}\ox\omega^{(n)}
_{B^n}\right\}, \\
\end{split}\eeq
where the second line is by definition, in the third line we restrict the
minimization to over the symmetric states by making a random permutation
and this makes no difference because random permutation operation as a
CPTP map keeps operator inequality and does not increase the purified
distance, in the fourth line we have employed the universal symmetric
state $\omega^{(n)}_{B^n}$ and made use of Lemma~\ref{lemma:unisym}.
Let $\gamma^{(n)*}_{A^nB^n}$ be the optimal state that makes the last
line of Eq.~(\ref{eq:mices-1}) achieves the minimum. Let $\mc{E}^n
\equiv \mc{E}_{M_A^{\ox n}\ox\omega^{(n)}_{B^n}}$ be the pinching map
associated with $M_A^{\ox n}\ox\omega^{(n)}_{B^n}$. Then we have
\beq\begin{split}\label{eq:mices-2}
\min_{\sigma^{(n)}_{B^n}\in\mc{S}(B^n)}\delta\left(\rho_{AB}^{\ox n}
\big\|M_A^{\ox n}\ox\sigma^{(n)}_{B^n},nr\right)
\geq &P\left(\rho_{AB}^{\ox n}, \gamma^{(n)*}_{A^nB^n}\right)\\
\geq &P\left(\mc{E}^n\big(\rho_{AB}^{\ox n}\big), \mc{E}^n
\big(\gamma^{(n)*}_{A^nB^n}\big)\right),
\end{split}\eeq
and in addition, $\gamma^{(n)*}_{A^nB^n}\leq 2^{nr} g_{n, |B|}M_A^{\ox n}
\ox\omega^{(n)}_{B^n}$, which after the CPTP map $\mc{E}^n$ being applied
to both sides yields
\beq\label{eq:mices-3}
\mc{E}^n\big(\gamma^{(n)*}_{A^nB^n}\big)\leq 2^{nr} g_{n, |B|}M_A^{\ox n}
\ox\omega^{(n)}_{B^n}.
\eeq
To proceed, we construct a projective measurement $\{\Pi_n,\1-\Pi_n\}$
with
\[
\Pi_n:=\Big\{\mc{E}^n\big(\rho_{AB}^{\ox n}\big)\geq 9\cdot2^{nr}
g_{n, |B|}M_A^{\ox n}\ox\omega^{(n)}_{B^n}\Big\}
\]
and set
\begin{align}
\label{eq:p_n}
p_n=&\tr\big(\mc{E}^n(\rho_{AB}^{\ox n})\Pi_n\big), \\
\label{eq:q_n}
q_n=&\tr\big(\mc{E}^n(\gamma^{(n)*}_{A^nB^n})\Pi_n\big).
\end{align}
By Eq.~(\ref{eq:mices-3}) and the construction of $\Pi_n$, it can be verified
that
\[q_n \leq \frac{1}{9} p_n.\]
So letting $\mc{M}^n:X\mapsto\tr(X\Pi_n)\proj{0}+\tr(X(\1-\Pi_n))\proj{1}$
be the measurement map associated with $\{\Pi_n,\1-\Pi_n\}$, we are able
to obtain
\beq\begin{split}\label{eq:mices-4}
&\min_{\sigma^{(n)}_{B^n}\in\mc{S}(B^n)}\delta\left(\rho_{AB}^{\ox n}
\big\|M_A^{\ox n}\ox\sigma^{(n)}_{B^n},nr\right) \\
\geq &P\left(\mc{M}^n\circ\mc{E}^n\big(\rho_{AB}^{\ox n}\big),
\mc{M}^n\circ\mc{E}^n\big(\gamma^{(n)*}_{A^nB^n}\big)\right) \\
= &\sqrt{1-\left(\sqrt{p_n}\sqrt{q_n}+\sqrt{1-p_n}
\sqrt{\tr(\gamma^{(n)*}_{A^nB^n})-q_n}\right)^2} \\
\geq &\sqrt{p_n\big(\frac{1}{3}-\frac{p_n}{9}\big)} \\
\geq &\frac{\sqrt{2p_n}}{3},
\end{split}\eeq
where the second line follows from Eq.~(\ref{eq:mices-2}) and the data
processing inequality for purified distance, for the fourth line we
have used $q_n \leq \frac{1}{9} p_n$ and $\tr(\gamma^{(n)*}_{A^nB^n})
-q_n\leq1$, and for the last line note that $p_n\leq 1$.
We prove later in Proposition~\ref{prop:usym-Imax} the asymptotics for
$p_n$ (cf. Eq.~(\ref{eq:p_n}) for its expression), that is,
\beq\label{eq:asym-p}
\lim_{n\rar\infty}\frac{-1}{n}\log p_n
=\sup_{s \geq 0}\left\{s\big(r-\min_{\sigma_B\in\mc{S}(B)}
D_{1+s}(\rho_{AB}\|M_A\ox\sigma_B)\big)\right\}.
\eeq
Then it follows from Eq.~(\ref{eq:asym-p}) and Eq.~(\ref{eq:mices-4})
that
\beq\label{eq:exp-mices-up}
\limsup_{n\rar\infty}\frac{-1}{n}\log\!\min_{\sigma^{(n)}_{B^n}\in
\mc{S}(B^n)}\!\!\delta\left(\rho_{AB}^{\ox n}\big\|M_A^{\ox n}\ox
\sigma^{(n)}_{B^n},nr\right)\!\leq\!\frac{1}{2}\sup_{s \geq 0}\left\{
s\big(r-\!\!\!\min_{\sigma_B\in\mc{S}(B)}\!\!\!D_{1+s}(\rho_{AB}\|M_A
\ox\sigma_B)\big)\right\}.
\eeq
Eventually, Eq.~(\ref{eq:exp-mices-up}) and Eq.~(\ref{eq:exp-mices-low})
together lead to Eq.~(\ref{eq:exp-mices}), and we are done.
\end{proof}
\begin{proposition}
\label{prop:usym-Imax}
Let $\rho_{AB}\in\mc{S}(AB)$ and $M_A\in\mc{P}(A)$ be such that $\supp(\rho_A)
\subseteq\supp(M_A)$. Let $\omega^{(n)}_{B^n}$ be the universal symmetric
state. Let $\mc{E}^n\equiv \mc{E}_{M_A^{\ox n}\ox\omega^{(n)}_{B^n}}$ be
the pinching map, and let $f(n)\geq 1$ be any sub-exponential function of $n$.
For given $r\in\mathbb{R}$, consider the sequence
\[
p_n:=\tr\left[\mc{E}^n\big(\rho_{AB}^{\ox n}\big)
\Big\{\mc{E}^n\big(\rho_{AB}^{\ox n}\big)\geq f(n)2^{nr}
M_A^{\ox n}\ox\omega^{(n)}_{B^n}\Big\} \right]
\]
for $n\in\mathbb{N}$. We have
\beq\label{eq:usym-Imax}
\lim_{n\rar\infty}\frac{-1}{n}\log p_n
=\sup_{s \geq 0}\left\{s\big(r-\min_{\sigma_B\in\mc{S}(B)}
D_{1+s}(\rho_{AB}\|M_A\ox\sigma_B)\big)\right\}.
\eeq
\end{proposition}
\begin{proof}
For simplicity, we use the shorthands
\[
P_n:=\mc{E}^n\big(\rho_{AB}^{\ox n}\big),\quad
Q_n:=M_A^{\ox n}\ox\omega^{(n)}_{B^n}.
\]
We will make frequent use of the fact that $P_n$ commutes with $Q_n$ as
well as that, by Lemma~\ref{lem:additive-Is},
\beq\label{eq:Ds-asympt}
\frac{1}{n}D_{1+s}(P_n\|Q_n)\longrightarrow \min_{\sigma_B\in\mc{S}(B)}
D_{1+s}(\rho_{AB}\|M_A\ox\sigma_B), \quad \text{as}\ \ n\rar\infty.
\eeq
In what follows we will first show that the r.h.s. of Eq.~(\ref{eq:usym-Imax})
is an achievable rate for the exponential decreasing of $p_n$, and then
we prove that this is indeed the optimal rate.
Now we prove the achievability part. For any $s>0$,
\[\begin{split}
\frac{-1}{n}\log p_n
&= \frac{-1}{n}\log\tr\big[P_n\left\{P_n\geq f(n) 2^{nr}Q_n\right\}\big]\\
&\geq\frac{-1}{n}\log\tr\Big[P_n\Big(\frac{P_n}{f(n) 2^{nr}Q_n}\Big)^s\Big]\\
& = sr-\frac{1}{n}\log\tr\big[P_n^{1+s}Q_n^{-s}\big]+\frac{s\log f(n)}{n}\\
& = sr-\frac{s}{n}D_{1+s}\big(P_n\|Q_n\big)+\frac{s\log f(n)}{n}\\
&\overset{n\to\infty}{\longrightarrow} s\big(r-\min_{\sigma_B\in\mc{S}(B)}D_{1+s}(\rho_{AB}\|M_A\ox\sigma_B)\big).
\end{split}\]
Noticing that $s>0$ is arbitrary, we conclude from the above estimation
that
\beq\label{eq:usym-Imax-low}
\liminf_{n\rar\infty}\frac{-1}{n}\log p_n
\geq\sup_{s \geq 0}\left\{s\big(r-\min_{\sigma_B\in\mc{S}(B)}
D_{1+s}(\rho_{AB}\|M_A\ox\sigma_B)\big)\right\}.
\eeq
For the other direction, we follow the method
of~\cite{HayashiTomamichel2016correlation} and employ the G\"{a}rtner-Ellis
theorem of large deviation theory. Let
\[
Z_n:=\frac{1}{n}\big(\log P_n-\log Q_n -nr -\log f(n) \big).
\]
We have
\beq\begin{split}\label{eq:usym-Imax-1}
p_n=&\tr\big[P_n\left\{P_n\geq f(n) 2^{nr}Q_n\right\}\big] \\
=&\tr\big[P_n\left\{Z_n\geq 0\right\}\big].
\end{split}\eeq
We see that the observable $Z_n$ commutes with the state $P_n$. So the
above Eq.~(\ref{eq:usym-Imax-1}) has a classical-probability-theoretic
explanation in which $Z_n$ is regarded as a random variable with $P_n$
being its distribution. Specifically, let the set of orthonormal vectors
$\{\ket{a_x^n}\}_x$ be the common eigenvectors of $Z_n$ and $P_n$. Then
$Z_n$ takes value $\bra{a^n_x}Z_n\ket{a^n_x}$ with probability
$\bra{a^n_x}P_n\ket{a^n_x}$. Now, Eq.~\eqref{eq:usym-Imax-1} translates
to
\[
p_n=\operatorname{Pr}\big\{Z_n\geq 0\big\},
\]
with respect to the probability distribution $\{\bra{a^n_x}P_n\ket{a^n_x}
\}_x$. To apply the G\"{a}rtner-Ellis theorem (in the form of
Lemma~\ref{lem:Gartner-Ellis}), we calculate the asymptotic cumulant
generating function of the sequence $\mc{Z}=\{Z_n\}_{n\in\mathbb{N}}$:
\beq\begin{split}\label{eq:cgf}
\Lambda_{\mc{Z}}(s)
=&\lim_{n\rar\infty}\frac{1}{n}\log\mathbb{E}\big[2^{nsZ_n}\big]\\
=&\lim_{n\rar\infty}\frac{1}{n}\log\mathbb{E}
\Big[\big(\frac{P_n}{Q_n}\big)^s(f(n))^{-s}2^{-nsr}\Big]\\
=&\lim_{n\rar\infty}\frac{1}{n}\log\tr
\Big[P_n\big(\frac{P_n}{Q_n}\big)^s(f(n))^{-s}2^{-nsr}\Big]\\
=&\lim_{n\rar\infty}\Big(\frac{s}{n}D_{1+s}(P_n\|Q_n)-sr-
\frac{s\log f(n)}{n}\Big) \\
=&\ s\big(\min_{\sigma_B\in\mc{S}(B)}
D_{1+s}(\rho_{AB}\|M_A\ox\sigma_B)-r\big).
\end{split}\eeq
Set $F(s):=\min\limits_{\sigma_B\in\mc{S}(B)}D_{1+s}(\rho_{AB}\|M_A\ox
\sigma_B)$ for later convenience. To proceed, we restrict our attention
to the case where $F(0)<r<F(+\infty)$ at the moment. It has been proven
in~\cite{HayashiTomamichel2016correlation} that the function $s\mapsto F(s)$
is monotonically increasing, continuously
differentiable, and $s\mapsto sF(s)$ is convex. As a result, we have that
$s\mapsto \Lambda_{\mc{Z}}(s)$ is differentialbe in $(0,+\infty)$, and
\begin{align*}
\lim_{s\rar 0} \Lambda_{\mc{Z}}'(s)
& = D(\rho_{AB}\|M_A\ox\rho_B)-r=F(0)-r<0, \\
(\exists s_0>0)\ \lim_{s\rar s_0} \Lambda_{\mc{Z}}'(s)
&\geq \min_{\sigma_B\in\mc{S}(B)}D_{1+s_0}(\rho_{AB}\|M_A\ox\sigma_B)-r
=F(s_0)-r>0.
\end{align*}
Hence, Lemma~\ref{lem:Gartner-Ellis} applies, yielding for $r\in(F(0),
F(\infty))$,
\beq\begin{split}\label{eq:usym-Imax-up}
\limsup_{n\rar\infty}\frac{-1}{n}\log p_n
=&\limsup_{n\rar\infty}\frac{-1}{n}\log{\rm Pr}\big\{Z_n\geq 0\big\} \\
\leq&\!\sup_{0<s<s_0}\!\left\{s\big(r-\min_{\sigma_B\in\mc{S}(B)}
D_{1+s}(\rho_{AB}\|M_A\ox\sigma_B)\big)\right\}\\
\leq&\ \ \sup_{s>0}\ \ \left\{s\big(r-\min_{\sigma_B\in\mc{S}(B)}
D_{1+s}(\rho_{AB}\|M_A\ox\sigma_B)\big)\right\}.
\end{split}\eeq
Now combining Eq.~(\ref{eq:usym-Imax-low}) and Eq.~(\ref{eq:usym-Imax-up})
together lets us arrive at
\beq\label{eq:usym-Imax-done}
\lim_{n\rar\infty}\frac{-1}{n}\log p_n
=\sup_{s \geq 0}\left\{s\big(r-\min_{\sigma_B\in\mc{S}(B)}
D_{1+s}(\rho_{AB}\|M_A\ox\sigma_B)\big)\right\}
\eeq
for $r\in(F(0),F(\infty))$. To complete the proof, we show that the
equality of Eq.~(\ref{eq:usym-Imax-done}) can be extended to the whole
range $r\in\mathbb{R}$. At first, we observe by definition that, the l.h.s.
of Eq.~(\ref{eq:usym-Imax-done}) is nonnegative and monotonically
increasing with $r$. Moreover, the equality of Eq.~(\ref{eq:usym-Imax-done})
established for $r\in(F(0),F(\infty))$ shows that the l.h.s. goes
to $0$ when $r\searrow F(0)=D(\rho_{AB}\|M_A\ox\rho_B)$. So we conclude
that the l.h.s. of Eq.~(\ref{eq:usym-Imax-done}) equals $0$ for $r\leq F(0)$,
coinciding with the right hand side. Next, we consider the case $r>F(\infty)
=\min_{\sigma_B\in\mc{S}(B)}D_{\rm max}(\rho_{AB}\|M_A\ox\sigma_B)$.
Letting $s\rar\infty$ in Lemma~\ref{lem:additive-Is} we get
\[
D_{\rm max}\Big(\mc{E}^n(\rho_{AB}^{\ox n})\big\|M_A^{\ox n}\ox
\omega^{(n)}_{B^n}\Big) =n\left( F(\infty)+O\Big(\frac{\log n}{n}\Big)\right).
\]
This implies that, for $r>F(\infty)$ and $n$ big enough,
\[
\Big\{\mc{E}^n\big(\rho_{AB}^{\ox n}\big)\geq
f(n)2^{nr}M_A^{\ox n}\ox\omega^{(n)}_{B^n}\Big\}=0
\]
and hence $p_n=0$. As a result, we conclude that the l.h.s. of
Eq.~(\ref{eq:usym-Imax-done}) is $+\infty$ for $r>F(\infty)$,
coinciding with the right hand side, too.
\end{proof}
\bigskip
The following lemma is due to Hayashi and
Tomamichel~\cite{HayashiTomamichel2016correlation}.
\begin{lemma}
\label{lem:additive-Is}
Let $\rho_{AB}\in\mc{S}(AB)$ and $M_A\in\mc{P}(A)$ such that $\supp(\rho_A)
\subseteq\supp(M_A)$. Let $\omega^{(n)}_{B^n}$ be the universal symmetric
state. Let $\mc{E}^n\equiv \mc{E}_{M_A^{\ox n}\ox\omega^{(n)}_{B^n}}$ be
the pinching map. For any $s\geq \frac{1}{2}$ we have
\[
\frac{1}{n}D_s\Big(\mc{E}^n(\rho_{AB}^{\ox n})\big\|M_A^{\ox n}\ox
\omega^{(n)}_{B^n}\Big) = \min_{\sigma_B\in\mc{S}(B)}D_s(\rho_{AB}\|M_A
\ox\sigma_B)+O\Big(\frac{\log n}{n}\Big).
\]
The implicit constants of the last term are independent of $s$.
\end{lemma}
\bigskip
The following lemma is due to Chen~\cite{Chen2000generalization} and
reformulated in~\cite{HayashiTomamichel2016correlation}.
\begin{lemma}
\label{lem:Gartner-Ellis}
Let $\{Z_n\}_{n\in\mathbb{N}}=:\mc{Z}$ be a sequence of random variables,
with the asymptotic cumulant generating function
\[
\Lambda_{\mc{Z}}(t)
:=\lim_{n\rar\infty}\frac{1}{n}\log\mathbb{E}\big[2^{ntZ_n}\big].
\]
If the above limit exists and the function $t\mapsto\Lambda_{\mc{Z}}(t)$ is
differentiable in some interval $(a,b)$, then for any
$z\in\big(\lim\limits_{t\searrow a}\Lambda_{\mc{Z}}'(t),\lim\limits_
{t\nearrow b}\Lambda_{\mc{Z}}'(t)\big)$, it holds that
\[
\limsup_{n\rar\infty}\frac{-1}{n}\log\,{\rm Pr}\{Z_n\geq z\}
\leq \sup_{t\in(a,b)}\big\{zt-\Lambda_{\mc{Z}}(t)\big\}.
\]
\end{lemma}
\subsection{Proof of Theorem~\ref{thm:reliability-dec}}
\label{subsec:thm-proof}
Based on the results obtained in the above two subsections, we are now ready
to complete the proof of Theorem~\ref{thm:reliability-dec}. At first, we
prove the one-shot achievability bound of Theorem~\ref{thm:dec-p-up}.
\begingroup
\def\thetheorem{\ref{thm:dec-p-up}}
\begin{theorem}[restatement]
Let $\rho_{RA}\in \mc{S}(RA)$. For any $m\in\mathbb{N}$, $0<s\leq 1$ and
$\sigma_A\in\mc{S}(A)$, the optimal performance of decoupling $A$ from $R$
is bounded as
\[
P^{\rm dec}_{R:A}(\rho_{RA},\log m) \leq
\sqrt{\frac{v^s}{s}}\exp\Big\{-(\ln2)\ s\big(\log m-\frac{1}{2}D_{1+s}
(\rho_{RA}\| \rho_R\ox\sigma_A)\big)\Big\},
\]
where $v$ is the number of distinct eigenvalues of $\rho_R\ox \sigma_A$.
\end{theorem}
\addtocounter{theorem}{-1}
\endgroup
\begin{proof}
We will employ the convex-split lemma. The convex-split lemma provides
directly a catalytic decoupling strategy via random unitary operation,
which in turn can be converted into a way for catalytic decoupling via
removing a subsystem (cf. Section~\ref{subsec:relation-dec}). Identify
$A$ with $A_1$, and let $A'=A_2A_3\cdots A_m$, where all the $A_i$ systems
have equal dimension. Let $A'$ in the state $\bar{\sigma}_{A'}=\sigma_{A_2}
\ox\sigma_{A_3}\ox\cdots\ox\sigma_{A_m}$ be the catalytic system. We
construct a random unitary channel
\[
\Lambda_{AA'}: X \mapsto \frac{1}{m}\sum_{i=1}^mU_iXU_i^*,
\]
where $U_i=W_{(1,i)}$ is the swapping between $A_1$ and $A_i$ (we set
$W_{(1,1)}=\1$). Then
\[
\Lambda_{AA'}\big(\rho_{RA}\ox \bar{\sigma}_{A'}\big)
=\frac{1}{m}\sum_{i=1}^m \rho_{RA_i}\ox\big[\sigma^{\ox m-1}\big]_{A^m/A_i},
\]
which is in the form of the state in Lemma~\ref{lem:convex-split}. So we
have
\[\begin{split}
P_{R:A}^{\rm dec\text{-}u}(\rho_{RA}, \log m)
\leq &P\Big(\Lambda_{AA'}\big(\rho_{RA}\ox \bar{\sigma}_{A'}\big),
\rho_{R}\ox\big[\sigma^{\ox m}\big]_{A^m}\Big) \\
\leq &\Big[(\ln2)\ D\Big(\Lambda_{AA'}\big(\rho_{RA}\ox \bar{\sigma}_{A'}\big)
\big\|\rho_R\ox\big[\sigma^{\ox m}\big]_{A^m} \Big)\Big]^{\frac{1}{2}}\\
\leq &\sqrt{\frac{v^s}{s}}\exp\Big\{-(\ln2)\ s\big(\frac{1}{2}\log m-
\frac{1}{2}D_{1+s}(\rho_{RA}\| \rho_R\ox\sigma_A)\big)\Big\},
\end{split}\]
where for the second line we have used the relation $P(\rho,\sigma)\leq
\sqrt{(\ln 2)D(\rho\|\sigma)}$, and for the third line we have used
Lemma~\ref{lem:convex-split}. At last, invoking
Proposition~\ref{prop:dec-relation} lets us confirm the claim.
\end{proof}
\medskip
The following one-shot converse bound is proved using standard techniques
(see, e.g.~\cite{DBWR2014one} and~\cite{MBDRC2017catalytic}).
\begin{proposition}
\label{prop:dec-p-low}
Let $\rho_{RA}\in \mc{S}(RA)$. For $k\geq 0$, the optimal performance of
decoupling $A$ from $R$ is bounded by the smoothing quantity for the
max-information (cf. Eq.~(\ref{eq:def-smi}) in Definition~\ref{def:smoothing-q}):
\[
P^{\rm dec}_{R:A}(\rho_{RA},k) \geq \delta_{R:A}(\rho_{RA},2k).
\]
\end{proposition}
\begin{proof}
Consider an arbitrary decoupling scheme that discards not more than than $k$
qubits. Let the catalytic state be $\sigma_{A'}$ and the unitary decoupling
operation be $U:\mc{H}_{AA'}\rar\mc{H}_{A_1A_2}$. Then $\log |A_2|\leq k$
and the performance for this scheme is
\[
\epsilon:=\min_{\omega_{A_1}\in\mc{S}(A_1)} P\big(\tr_{A_2}U(\rho_{RA}\ox
\sigma_{A'})U^*,\rho_R \ox\omega_{A_1}\big).
\]
Let $\omega_{A_1}^*$ be the optimal state in the above minimization. By
Uhlmann's theorem~\cite{Uhlmann1976transition}, there is a state
$\tilde{\rho}_{RA_1A_2}$ such that
\begin{align}
\label{eq:decpl-1}
&P\big(U(\rho_{RA}\ox\sigma_{A'})U^*,\tilde{\rho}_{RA_1A_2}\big)=\epsilon,\\
\label{eq:decpl-2}
&\tr_{A_2} \tilde{\rho}_{RA_1A_2} = \rho_R \ox\omega_{A_1}^*.
\end{align}
From Eq.~(\ref{eq:decpl-1}) we get
\beq\label{eq:decpl-3}
P\big(\rho_{RA},\tr_{A'}U^*\tilde{\rho}_{RA_1A_2}U\big)\leq\epsilon.
\eeq
Eq.~(\ref{eq:decpl-2}) implies that
\[
\tilde{\rho}_{RA_1A_2}
\leq|A_2|^2\rho_R\ox\omega_{A_1}^*\ox \frac{\1_{A_2}}{|A_2|}
\leq 2^{2k}\rho_R\ox\omega_{A_1}^*\ox\frac{\1_{A_2}}{|A_2|},
\]
which further yields
\beq\label{eq:decpl-4}
\tr_{A'}U^*\tilde{\rho}_{RA_1A_2}U\leq2^{2k}\rho_R\ox
\tr_{A'}U^*\big(\omega_{A_1}^*\ox\frac{\1_{A_2}}{|A_2|}\big)U.
\eeq
Inspecting the definition in Eq.~(\ref{eq:def-smi}), we obtain from
Eq.~(\ref{eq:decpl-3}) and Eq.~(\ref{eq:decpl-4}) that
\[\delta_{R:A}(\rho_{RA},2k)\leq\epsilon.\]
Since by assumption $\epsilon$ is the performance of an arbitrary decoupling
scheme with cost not larger than $k$, the optimal performance satisfies the
above relation as well. That is,
\[\delta_{R:A}(\rho_{RA},2k)\leq P^{\rm dec}_{R:A}(\rho_{RA},k).\]
\end{proof}
\begin{proofof}[of Theorem~\ref{thm:reliability-dec}]
By applying Theorem~\ref{thm:dec-p-up} and making the substitution
$\rho_{RA}\leftarrow\rho_{RA}^{\ox n}$, $\log m\leftarrow nr$, $\sigma_A
\leftarrow\sigma_A^{\ox n}$, we get
\[
\frac{-1}{n}\log P^{\rm dec}_{R^n:A^n}\big(\rho_{RA}^{\ox n},nr\big)\geq
\frac{-1}{2n}\log\frac{v_n^s}{s}+s\big(r-\frac{1}{2}D_{1+s}(\rho_{RA}\|
\rho_R\ox\sigma_A)\big),
\]
where $v_n\leq(n+1)^{|R| + |A|}$ is the number of distinct eigenvalues
of $\rho_R^{\ox n}\ox\sigma_A^{\ox n}$. Letting $n\rar\infty$, and then
maximizing the r.h.s. over $s\in (0,1]$ and $\sigma_A\in\mc{S}(A)$, we
arrive at
\[
E^{\rm dec}_{R:A}(\rho_{RA},r)
\geq\liminf_{n\rar\infty}\frac{-1}{n}\log
P^{\rm dec}_{R^n:A^n}\big(\rho_{RA}^{\ox n},nr\big)
\geq\sup_{0<s\leq 1}\Big\{s\big(r-\frac{1}{2}I_{1+s}(R:A)_\rho\big)\Big\},
\]
which is Eq.~(\ref{eq:exp-dec-low}).
For Eq.~(\ref{eq:exp-dec-up}), we apply Proposition~\ref{prop:dec-p-low}
and Theorem~\ref{theorem:exp-mice} to obtain
\[\begin{split}
E^{\rm dec}_{R:A}(\rho_{RA},r)
= & \limsup_{n\rar\infty}\frac{-1}{n}\log
P^{\rm dec}_{R^n:A^n}\big(\rho_{RA}^{\ox n},nr\big) \\
\leq& \lim_{n\rar\infty}\frac{-1}{n}\log
\delta_{R^n:A^n}\big(\rho_{RA}^{\ox n},2nr\big) \\
= & \sup_{s\geq0}\Big\{s\big(r-\frac{1}{2}I_{1+s}(R:A)_\rho\big)\Big\}.
\end{split}\]
At last, Eq.~(\ref{eq:exp-dec}) follows from Eq.~(\ref{eq:exp-dec-low}) and
Eq.~(\ref{eq:exp-dec-up}). To see this, we consider the function
\beq
f(s)=s\big(r-\frac{1}{2}I_{1+s}(R:A)_\rho\big)
\eeq
for $s\in[0,\infty)$. $f(s)$ is concave because $s\mapsto sI_{1+s}(R:A)_\rho$
is convex. So, we have that $f(s)$ reaches the maximum in $[0,1]$ and
therefore
\beq
\max_{s\geq 0}f(s) = \max_{0\leq s\leq 1} f(s),
\eeq
if $f'(1)\leq 0$. This condition is equivalent to $r\leq R_{\rm critical}=
\frac{1}{2}\frac{\mathrm{d}}{\mathrm{d}s} sI_{1+s}(R:A)_{\rho}\big|_{s=1}$.
\end{proofof}
\begin{remark} \label{rem:mainresult}
Some comments on the main results are as follows.
\begin{enumerate}[(a)]
\item Although the reliability function $E^{\rm dec}_{R:A}(\rho_{RA},r)$ is
defined as a limit superior, we see in the above proof that when we are
able to determine it for $r\leq R_{\rm critical}$, the limit of $\frac{-1}
{n}\log P^{\rm dec}_{R^n:A^n}\big(\rho_{RA}^{\ox n},nr\big)$ actually
exists.
\item When we employ the convex-split lemma (Lemma~\ref{lem:convex-split})
to derive the achievability results of Theorem~\ref{thm:dec-p-up} and
Eq.~(\ref{eq:exp-dec-low}), we can keep using the relative entropy as the
measure of closeness between states. Meanwhile, Eq.~(\ref{eq:exp-dec-up})
can be converted into an upper bound for the rate of exponential decay
of the performance measured by relative entropy, making use of the
relation $P(\rho,\sigma)\leq\sqrt{(\ln 2)D(\rho\|\sigma)}$. This actually
enables us to derive results similar to Theorem~\ref{thm:reliability-dec},
with the purified distance being replaced by the relative entropy in the
definition of the reliability function.
\item The upper bound of Eq.~\eqref{eq:exp-dec-up} diverges when $r>\frac{1}{2}
I_{\rm max}(R:A)_\rho$, which can be easily seen. One may guess
that it goes to $\infty$ when $r\nearrow\frac{1}{2}I_{\rm max}(R:A)_\rho$.
However, we show in Proposition~\ref{prop:bounded} in the Appendix
that it is actually bounded when $r<\frac{1}{2}I_{\rm max}(R:A)_\rho$.
\end{enumerate}
\end{remark}
\section{Proof of the relations between quantum tasks}
\label{sec:relations}
In this section, we prove Proposition~\ref{prop:dec-relation} on
the relation between the three types of catalytic quantum information
decoupling, and Proposition~\ref{prop:merging-dec} on the relation
between catalytic quantum information decoupling and quantum state
merging.
\subsection{Proof of the relation between decouplings}
\label{subsec:relation-dec}
\begingroup
\def\thetheorem{\ref{prop:dec-relation}}
\begin{proposition}[restatement]
For $\rho_{RA} \in \mathcal{S}(RA)$ and $k\geq 0$, $r \geq 0$, we have
\begin{align}
&P_{R:A}^{\rm dec}(\rho_{RA}, k)
=P_{R:A}^{\rm dec\text{-}m}(\rho_{RA}, 2k)
=P_{R:A}^{\rm dec\text{-}u}(\rho_{RA}, 2k), \label{eqr:dec-P} \\
&E_{R:A}^{\rm dec}(\rho_{RA}, r)
=E_{R:A}^{\rm dec\text{-}m}(\rho_{RA}, 2r)
=E_{R:A}^{\rm dec\text{-}u}(\rho_{RA}, 2r). \label{eqr:dec-E}
\end{align}
\end{proposition}
\addtocounter{theorem}{-1}
\endgroup
\begin{proof}
Eq.~\eqref{eqr:dec-E} follows from Eq.~\eqref{eqr:dec-P} directly. So it
suffices to prove Eq.~\eqref{eqr:dec-P}. In the following, we show the
identities $P_{R:A}^{\rm dec}(\rho_{RA}, k)=P_{R:A}^{\rm dec\text{-}m}
(\rho_{RA}, 2k)$ and $P_{R:A}^{\rm dec}(\rho_{RA}, k)
=P_{R:A}^{\rm dec\text{-}u}(\rho_{RA}, 2k)$ separately.
We first prove the former identity. Consider an arbitrary decoupling scheme
via removing a subsystem of size not more than $k$ qubits. Let $\sigma_{A'}$
be the catalytic state and $U:\mc{H}_{AA'}\rar \mc{H}_{A_1A_2}$ be the
unitary operation, where $A_2$ is the system to be removed and $\log |A_2|
\leq k$. Introduce a system $A''$ in the maximally mixed state
$\frac{\1_{A''}}{|A''|}$ such that $|A''|=|A_2|$. We construct a decoupling
scheme via projective measurement: the catalytic state is chosen to be
$\sigma_{A'}\ox \frac{\1_{A''}}{|A''|}$, and the projective measurement
is given by $\{Q^x_{AA'A''}\}_{x=1}^m$, with $m=|A_2|^2$ and
\[
Q^x_{AA'A''}=(U\ox\1_{A''})^*\big((W_{A_2}^x\ox\1_{A''})\Psi_{A_2A''}
(W_{A_2}^x\ox\1_{A''})^*\ox\1_{A_1}\big)(U\ox\1_{A''}),
\]
where $\{W_{A_2}^x\}_{x=1}^{m}$ is the set of discrete Weyl operators on
$\mc{H}_{A_2}$ and $\Psi_{A_2A''}$ is the maximally entangled state. So,
the cost is $\log m=\log |A_2|^2 \leq 2k$. By the definition of
$P_{R:A}^{\rm{dec\text{-}m}}$, for arbitrary state $\omega_{A_1}$, we have
\begin{equation}
\label{eq:decmleqdec}
\begin{split}
&P_{R:A}^{\rm{dec\text{-}m}}(\rho_{RA}, 2k) \\
\leq &P\Big(\sum_{x=1}^m Q^x_{AA'A''}\big(\rho_{RA} \ox \sigma_{A'} \ox
\frac{\1_{A''}}{|A''|}\big)Q^x_{AA'A''},\,\rho_R\ox U^*\big(\omega_{A_1}
\ox \frac{\1_{A_2}}{|A_2|}\big)U \ox \frac{\1_{A''}}{|A''|}\Big) \\
= &P\Big(U^*\big(\tr_{A_2} \big[U(\rho_{RA} \ox \sigma_{A'})U^*\big]
\ox \frac{\1_{A_2}}{|A_2|}\big)U\ox \frac{\1_{A''}}{|A''|},\,
\rho_R\ox U^*\big(\omega_{A_1}\ox \frac{\1_{A_2}}{|A_2|}\big)U
\ox \frac{\1_{A''}}{|A''|}\Big) \\
= &P\big(\tr_{A_2}U(\rho_{RA}\ox\sigma_{A'})U^*,\rho_R\ox\omega_{A_1}\big),
\end{split}
\end{equation}
where the third line can be verified by direct calculation. Minimizing
the last line of Eq.~(\ref{eq:decmleqdec}) over $\sigma_{A'}$,
$\omega_{A_1}$ and $U:\mc{H}_{AA'}\rar \mc{H}_{A_1A_2}$, and by
Definition~\ref{def:perform-dec}, we get from Eq.~(\ref{eq:decmleqdec}) that
\begin{equation}
\label{eq:decmleqdec1}
P_{R:A}^{\rm{dec\text{-}m}}(\rho_{RA},2k)\leq P^{\rm{dec}}_{R:A}(\rho_{RA},k).
\end{equation}
Next, we deal with the other direction. Let $\sigma_{A'}$ and $\{Q^x_{AA'}\}
_{x=1}^m$ be the catalytic state and the projectors of a decoupling scheme
via projective measurement with cost $\log m\leq 2k$. Introduce systems $C$
and $C'$ such that $|C|=|C'|=\sqrt{m}$ and denote by $\Psi_{CC'}$ the maximally
entangled state on $C$ and $C'$. We construct a decoupling scheme via removing
a subsystem: the catalytic state is given by $\sigma_{A'}\ox \frac{\1_{CC'}}
{|C|\cdot|C'|}$, and the unitary operation is given by
\[
U_{AA'CC'}
=\sum_{x=1}^m \Big(\sum_{y=1}^m \mathrm{e}^{\frac{2\pi\mathrm{i}xy}{m}}
Q_{AA'}^y\Big)\ox \big(W^x_C\ox\1_{C'}\big)\Psi_{CC'}
\big(W^x_C\ox\1_{C'}\big)^*,
\]
where $\{W^x_C\}_{x=1}^m$ is the set of discrete Weyl operators on
$\mc{H}_C$. Set $A_1\equiv AA'C$ be the remaining system, and $A_2\equiv C'$
be the removed system. So, the cost is $\log|A_2|=\log|C'|\leq k$.
By the definition of $P_{R:A}^{\rm{dec}}$, for arbitrary state
$\omega_{AA'}$, we have
\begin{equation}
\label{eq:decleqdecm}
\begin{split}
& P_{R:A}^{\rm{dec}}(\rho_{RA}, k)\\
\leq & P\Big(\tr_{C'}\big[U_{AA'CC'}\big(\rho_{RA}\ox\sigma_{A'}\ox
\frac{\1_{CC'}}{|C|\cdot|C'|}\big)U_{AA'CC'}^*\big],\,
\rho_R\ox\omega_{AA'}\ox\frac{\1_C}{|C|} \Big) \\
= & P\Big(\sum_xQ^x_{AA'}\big(\rho_{RA}\ox \sigma_{A'}\big)Q^x_{AA'}\ox
\frac{\1_C}{|C|},\,\rho_R \ox \omega_{AA'} \ox \frac{\1_C}{|C|} \Big)\\
= & P\Big(\sum_xQ^x_{AA'}\big(\rho_{RA}\ox \sigma_{A'}\big)Q^x_{AA'},\,
\rho_R\ox\omega_{AA'}\Big),
\end{split}
\end{equation}
where the third line can be verified by direct calculation. Minimizing the
last line of Eq.~(\ref{eq:decleqdecm}) over $\sigma_{A'}$, $\omega_{AA'}$
and $\{Q^x_{AA'}\}_{x=1}^m$, and by Definition~\ref{def:perform-dec-m},
we get from Eq.~(\ref{eq:decleqdecm}) that
\begin{equation}
\label{eq:decleqdecm1}
P_{R:A}^{\rm{dec}}(\rho_{RA},k) \leq P_{R:A}^{\rm{dec\text{-}m}}(\rho_{RA},2k).
\end{equation}
The combination of Eq.~\eqref{eq:decmleqdec1} and Eq.~\eqref{eq:decleqdecm1}
leads to the identity $P_{R:A}^{\rm dec}(\rho_{RA}, k)=P_{R:A}^{\rm dec\text{-}m}
(\rho_{RA}, 2k)$.
Now we prove the latter identity. Consider an arbitrary decoupling scheme
via removing a subsystem of size not more than $k$ qubits. Let $\sigma_{A'}$
be the catalytic state and $U:\mc{H}_{AA'}\rar \mc{H}_{A_1A_2}$ be the
unitary operation, where $A_2$ is the system to be removed and $\log |A_2|
\leq k$. We construct a decoupling scheme via random unitary operation:
the catalytic state is chosen to be $\sigma_{A'}$ and the random unitary
operation is given by
\begin{equation}
\Lambda_{AA'}(\cdot)=\frac{1}{|A_2|^2} \sum_{i=1}^{|A_2|^2}
(\1_{A_1}\ox W_{A_2}^i) U (\cdot) U^* (\1_{A_1}\ox W_{A_2}^i)^*,
\end{equation}
where $\{W_{A_2}^i\}_{i=1}^{|A_2|^2}$ are all the discrete Weyl operators
on $\mc{H}_{A_2}$. By the definition of $P_{R:A}^{\rm{dec\text{-}u}}$, for
arbitrary state $\omega_{A_1}$, we have
\begin{equation}
\label{eq:eraslardec}
\begin{split}
&P_{R:A}^{\rm{dec\text{-}u}}(\rho_{RA}, 2k) \\
\leq &P\Big(\Lambda_{AA'}(\rho_{RA} \ox \sigma_{A'}), \rho_R
\ox \omega_{A_1}\ox \frac{\1_{A_2}}{|A_2|}\Big) \\
= &P\Big(\tr_{A_2} \big[U(\rho_{RA} \ox \sigma_{A'})U^*\big] \ox
\frac{\1_{A_2}}{|A_2|}, \rho_R \ox \omega_{A_1}
\ox \frac{\1_{A_2}}{|A_2|}\Big) \\
= &P\Big(\tr_{A_2} \big[U(\rho_{RA} \ox \sigma_{A'})U^*\big],
\rho_R \ox \omega_{A_1}\Big).
\end{split}
\end{equation}
Minimizing the last line of Eq.~(\ref{eq:eraslardec}) over $\sigma_{A'}$,
$\omega_{A_1}$ and $U:\mc{H}_{AA'}\rar \mc{H}_{A_1A_2}$, and by
Definition~\ref{def:perform-dec}, we get from Eq.~(\ref{eq:eraslardec}) that
\begin{equation}
\label{eq:eraslardec1}
P_{R:A}^{\rm{dec\text{-}u}}(\rho_{RA},2k)\leq P^{\rm{dec}}_{R:A}(\rho_{RA},k).
\end{equation}
Next, we deal with the other direction. Let $\sigma_{A'}$ and $\Lambda_{AA'}$
be the catalytic state and the random unitary channel of a decoupling scheme
via random unitary operation with cost bounded by $2k$ bits. $\Lambda_{AA'}$
can be written as
\begin{equation}
\Lambda_{AA'}(\cdot)=\frac{1}{m} \sum_{i=1}^m V_{AA'}^{i} (\cdot) (V_{AA'}^{i})^*,
\end{equation}
where $V_{AA'}^i\in\mc{U}(AA')$ and $\log m \leq 2k$. Introduce systems $C$
and $C'$ such that $|C|=|C'|=\sqrt{m}$ and denote by $\Psi_{CC'}$ the maximally
entangled state on $C$ and $C'$. We construct a decoupling scheme via removing
a subsystem: the catalytic state is given by
\[
\sigma_{A'CC'}=\frac{1}{m}\sum_{i=1}^{m}
\sigma_{A'}\ox(W_C^i\ox\1_{C'})\Psi_{CC'}(W_C^i\ox\1_{C'})^*
\]
with $\{W_C^i\}_{i=1}^{m}$ be all the discrete Weyl operators on $\mc{H}_C$,
and the unitary operation is given by
\[
U_{AA'CC'}
=\sum_{i=1}^m V_{AA'}^i \ox (W_C^i\ox\1_{C'})\Psi_{CC'}(W_C^i\ox\1_{C'})^*.
\]
Set $A_1\equiv AA'C$ be the remaining system, and $A_2\equiv C'$ be the
removed system. So, the cost is $\log|A_2|=\log|C'|\leq k$. By the
definition of $P_{R:A}^{\rm{dec}}$, for arbitrary state $\omega_{AA'}$, we
have
\begin{equation}
\label{eq:declareras}
\begin{split}
& P_{R:A}^{\rm{dec}}(\rho_{RA}, k)\\
\leq & P\Big(\tr_{C'} \big[U_{AA'CC'}(\rho_{RA}\ox\sigma_{A'CC'})
U_{AA'CC'}^*\big],\rho_R\ox\omega_{AA'}\ox\frac{\1_C}{|C|} \Big) \\
= & P\Big(\Lambda_{AA'}(\rho_{RA}\ox \sigma_{A'})\ox \frac{\1_C}{|C|},
\rho_R \ox \omega_{AA'} \ox \frac{\1_C}{|C|} \Big)\\
= & P\big(\Lambda_{AA'}(\rho_{RA} \ox \sigma_{A'}),
\rho_R\ox\omega_{AA'}\big).
\end{split}
\end{equation}
Minimizing the last line of Eq.~(\ref{eq:declareras}) over $\sigma_{A'}$,
$\Lambda_{AA'}$ and $\omega_{AA'}$, and by Definition~\ref{def:perform-dec-u},
we get from Eq.~(\ref{eq:declareras}) that
\begin{equation}
\label{eq:declareras1}
P_{R:A}^{\rm{dec}}(\rho_{RA},k) \leq P_{R:A}^{\rm{dec\text{-}u}}(\rho_{RA},2k).
\end{equation}
The combination of Eq.~\eqref{eq:eraslardec1} and Eq.~\eqref{eq:declareras1}
completes the proof.
\end{proof}
\subsection{Proof of the relation between decoupling and quantum state merging}
\label{subsec:relation-dec-merg}
\begingroup
\def\thetheorem{\ref{prop:merging-dec}}
\begin{proposition}[restatement]
For a tripartite pure state $\rho_{RAB} \in \mathcal{S}({RAB})$ and
$k\geq 0$, $r \geq 0$, we have
\begin{align}
&P^{\rm{merg}}_{A \Rightarrow B}(\rho_{RAB}, k)
=P^{\rm{merg}}_{A \rightarrow B}(\rho_{RAB}, 2k)
=P^{\rm{dec}}_{R:A}(\rho_{RA}, k), \label{eqr:merg-dec-P} \\
&E^{\rm{merg}}_{A \Rightarrow B}(\rho_{RAB}, r)
=E^{\rm{merg}}_{A \rightarrow B}(\rho_{RAB}, 2r)
=E^{\rm{dec}}_{R:A}(\rho_{RA}, r). \label{eqr:merg-dec-E}
\end{align}
\end{proposition}
\addtocounter{theorem}{-1}
\endgroup
\begin{proof}
Eq.~\eqref{eqr:merg-dec-E} follows from Eq.~\eqref{eqr:merg-dec-P}
directly. So it suffices to prove Eq.~\eqref{eqr:merg-dec-P}. Assisted
by free entanglement, a merging protocol via quantum communication of
$k$ qubits, can simulate a merging protocol via classical communication
of $2k$ bits, by using teleportation. Conversely, a merging protocol
via classical communication of $2k$ bits, can simulate a merging protocol
via quantum communication of $k$ qubits, by using dense coding. So, we
easily obtain the relation $P_{A \Rightarrow B}^{\rm{merg}}(\rho_{RAB},k)
=P_{A \rar B}^{\rm{merg}}(\rho_{RAB}, 2k)$. In the following, we prove
$P_{A \Rightarrow B}^{\text{merg}}(\rho_{RAB},k)= P_{R:A}^{\text{dec}}
(\rho_{RA}, k)$.
Consider a decoupling scheme with catalytic state $\sigma_{A'}$, in which
the system $A_2$ of size $\log|A_2|\leq k$ is removed after a unitary
operator $U_{AA'\rar A_1A_2}$ being applied. Let $\omega_{A_1}$ be an
arbitrary state on $A_1$. Let $\sigma_{A'B'}$ and $\omega_{A_1B_1}$ be
the purification of $\sigma_{A'}$ and $\omega_{A_1}$, respectively. By
Uhlmann's theorem~\cite{Uhlmann1976transition}, there exists an isometry
$V_{A_2BB' \rar ABB_1}$ such that
\begin{equation}
\label{equ:isometry}
\begin{split}
&P\big(\tr_{A_2} \big[U_{AA'\rar A_1A_2}(\rho_{RA} \ox \sigma_{A'})
U_{AA'\rar A_1A_2}^*\big], \rho_R \ox \omega_{A_1} \big) \\
=&P\big(V_{A_2BB' \rightarrow ABB_1}U_{AA'\rar A_1A_2}(\rho_{RAB} \ox
\sigma_{A'B'})U_{AA'\rar A_1A_2}^*V_{A_2BB' \rar ABB_1}^*, \rho_{RAB}
\ox \omega_{A_1B_1}\big).
\end{split}
\end{equation}
We construct a merging protocol $\mc{M}_1$ which consists of using
shared entanglement $\sigma_{A'B'}$, Alice applying $U_{AA'\rar A_1A_2}$
and sending $A_2$ to Bob, and Bob applying isometry $V_{A_2BB' \rar ABB_1}$.
The cost of quantum communication is $\log |A_2| \leq k$. We have
\begin{equation}
\label{eq:M1}
\mc{M}_1(\rho_{RAB})=\tr_{A_1B_1}\big[V_{A_2BB' \rar ABB_1}
U_{AA'\rar A_1A_2}(\rho_{RAB} \ox \sigma_{A'B'})U_{AA'\rar A_1A_2}^*
V_{A_2BB' \rar ABB_1}^*\big].
\end{equation}
By the monotonicity of purified distance under partial trace, we get
from Eq.~\eqref{equ:isometry} and Eq.~\eqref{eq:M1} that
\begin{equation}
\label{eq:mleqdec}
P\big(\mc{M}_1(\rho_{RAB}), \rho_{RAB}\big)
\leq P\big(\tr_{A_2}\big[U_{AA'\rar A_1A_2}(\rho_{RA} \ox \sigma_{A'})
U_{AA'\rar A_1A_2}^*\big], \rho_R \ox \omega_{A_1} \big)
\end{equation}
Eq.~\eqref{eq:mleqdec} together with the definitions of
$P_{A \Rightarrow B}^{\text{merg}}$ and $P_{R:A}^{\text{dec}}$ implies
\begin{equation}
\label{equ:mergeleqdec}
P_{A \Rightarrow B}^{\text{merg}}(\rho_{RAB}, k)
\leq P_{R:A}^{\text{dec}}(\rho_{RA}, k).
\end{equation}
Now, we turn to the proof of the other direction. Let $\mc{M}_1$ be a
merging protocol via quantum communication of not more than $k$ qubits.
We write $\mc{M}_1$ as
\[
\mc{M}_1(\cdot)=
\tr_{A_1B_1} \big[V_{A_2BB' \rar ABB_1}U_{AA' \rar A_1A_2}\big((\cdot)
\ox \phi_{A'B'}\big)U_{AA' \rar A_1A_2}^*V_{A_2BB' \rar ABB_1}^*\big],
\]
where $\phi_{A'B'}$ is the shared entanglement, $U_{AA' \rar A_1A_2}$
is performed by Alice, $V_{A_2BB' \rar ABB_1}$ is performed by Bob, and
$A_2$ with $\log |A_2| \leq k$ is sent from Alice to Bob. Using Uhlmann's
theorem again, we know that there exists a pure state $\varphi_{A_1B_1}$
such that
\begin{equation}
\begin{split}
&P\big(\mc{M}_1(\rho_{RAB}), \rho_{RAB}\big) \\
=&P\big(V_{A_2BB' \rar ABB_1}U_{AA' \rar A_1A_2}\big(\rho_{RAB} \ox
\phi_{A'B'}\big)U_{AA' \rar A_1A_2}^*V_{A_2BB' \rar ABB_1}^*,
\rho_{RAB} \ox \varphi_{A_1B_1}\big).
\end{split}
\end{equation}
Because the purified distance decreases under the action of partial trace,
we get
\begin{equation}
\label{equ:mergeuhm}
P(\mc{M}_1(\rho_{RAB}), \rho_{RAB})
\geq P\big(\tr_{A_2}\big[U_{AA' \rar A_1A_2}\big(\rho_{RA} \ox \phi_{A'}\big)
U_{AA' \rar A_1A_2}^*\big], \rho_R \ox \varphi_{A_1}\big).
\end{equation}
Since $\tr_{A_2}\big[U_{AA' \rar A_1A_2}\big(\rho_{RA} \ox \phi_{A'}\big)
U_{AA' \rar A_1A_2}^*\big]$ in the right hand side of Eq.~\eqref{equ:mergeuhm}
describes a decoupling scheme via removing $\log |A_2| \leq k$ qubits,
by the definitions of $P_{A \Rightarrow B}^{\text{merg}}$ and
$P_{R:A}^{\text{dec}}$, we conclude that
\begin{equation}
\label{equ:decleqmerge}
P_{A \Rightarrow B}^{\text{merg}}(\rho_{RAB}, k)
\geq P_{R:A}^{\text{dec}}(\rho_{RA},k).
\end{equation}
\end{proof}
\section{Discussion}
\label{sec:discussion}
In this work, we have characterized the reliability function of catalytic quantum
information decoupling using the sandwiched R{\'e}nyi divergence, and have applied
it to quantum state merging. Our results add new operational meanings to the
sandwiched R{\'e}nyi information quantities. Before ending the paper, we list a
few open problems.
\begin{enumerate}
\item Our results are obtained for the catalytic setting. To characterize the
reliability function of quantum information decoupling without involving any
catalytic system is an open problem. It is known that when only an asymptotically
vanishing error is concerned, the best rate of cost is the same no matter a
catalytic system is allowed or not~\cite{HOW2005partial,HOW2007quantum,
ADHW2009mother,MBDRC2017catalytic,ADJ2017quantum}. However, for the reliability
function we do not know whether catalysis makes a difference.
\item Another open problem is to derive the reliability function when the rate
of cost is above the critical value. The upper and lower bounds obtained in
the present work do not match in the high-rate case. However, note that the
existence of critical points in the study of reliability functions is a common
phenomenon, where at the unsolved side the problem becomes more of a
combinatorial feature and is hard to tackle~\cite{Gallager1968information,
Dalai2013lower,HayashiTan2016equivocations,LYH2021tight}.
\item Moreover, it is interesting to consider the reliability function when
the performance is measured by the trace distance. Our method works for the
purified distance, and equivalently, for the fidelity. But for trace distance,
quite different techniques may be needed. It is worth mentioning that, as
was pointed out in~\cite{Dupuis2021privacy}, for trace distance there is no
obvious way to transfer the reliability function from quantum information
decoupling to other quantum tasks (such as quantum state merging). This is
in contrast to what we have done in Proposition~\ref{prop:merging-dec} and Corollary~\ref{cor:reliability-merging} which make the transference by
employing the Uhlmann's theorem.
\item Finally, we hope that our work will be stimulating and find more
applications, in deriving the reliability functions for more quantum
information tasks. Indeed, in a later work~\cite{LiYao2021reliable} we
have applied the present results to quantum channel simulation. Can we
extend this line of research to entanglement-assisted communication of noisy
channels~\cite{BSST2002entanglement}, and to the treatment of the amount
of comsumed/generated entanglement in quantum tasks? These problems are
closely related to the size of the remaining system in quantum information
decoupling.
\end{enumerate}
\bigskip
|
\section{\label{intro}INTRODUCTION\protect\\}
Zigzag graphene nanoribbons (ZGNRs) have attracted a great deal of interest in the last decades as ones of the most promising graphene derivatives for technological applications, in particular for their potential use in spintronic devices. Indeed, density functional theory (DFT) calculations
showed that these systems are spin gapped, with the occurrence of spin-polarized zigzag edges\cite{nakada_edge_1996} that become antiferromagnetically (AFM) ordered\cite{son_energy_2006,wu_electronic_2015,correa_braiding_2018,jiang_unique_2007}. It has been suggested that an in-plane electric field perpendicular to the graphene nanoribbon could lead to a half-metal\cite{son_half-metallic_2006}, where only one spin carries electric current. One can make use of such desirable properties in several technological applications such as ribbon-based spintronics, sensors, and storage devices\cite{wang_2021_graphene}, as long as magnetism in these materials is strong enough to survive at operating temperatures.
However,
ZGNRs
are quasi-one-dimensional (quasi-1D) systems where
electron-electron interaction
together with quantum fluctuations could play a major role in stabilizing different phases, competing each other in energy. This is particularly true for the narrowest nanoribbons, made of an infinite chain of fused benzene rings, such as to form a zigzag edge. The number of chains will determine the width of the ribbon. According to previous conventions\cite{fujita1996peculiar}, we will refer to the ribbon made of $n$ zigzag chains as an $n$-ZGNR.
In the narrowest ribbon, i.e. the 2-ZGNR (see Fig.~\ref{fig:unit-cell}), the effects of quantum fluctuations are enhanced by low transverse dimensionality. A previous quantum Monte Carlo (QMC) study\cite{dupuy_fate_2018} on the polyacene molecules, the finite-length counterpart of 2-ZGNRs,
revealed how low dimensionality and
electron correlation favored a resonating-valence-bond (RVB) -like ground state, by penalizing solutions with spin polarized edges. Given these premises, a thorough verification of the DFT predictions needs to be carried out. Indeed, strong correlation and large quantum fluctuations, typical of 1D systems, are critical for DFT, which
could fail to predict the correct ground-state (GS) properties in these situations.
\begin{figure}[ht]
\centering
\includegraphics[scale=0.2]{unit_cell.pdf}
\caption{\label{fig:unit-cell} Schematic view of the unit cell of 2-ZGNR, represented in square brackets. This shows also the numbering of atoms, as used throughout the paper. The unit cell contains $N_\textrm{C}=4$ carbon atoms, and $N_\textrm{H}=2$ hydrogen atoms. The dashed lines are periodic replicas, extending to infinity on both sides.
}
\end{figure}
A route to verify the DFT predictions
is
experimental. Unfortunately, ZGNRs
are
difficult to synthesize in a stable form because of their extremely reactive edges. Thus, a direct experimental demonstration of their
GS
as an AFM-ordered phase is still to come, although significant progress has been made recently towards a controlled fabrication of these systems\cite{han2007energy,chen2007graphene,magda2014room,ruffieux_-surface_2016,fedotov2020excitonic,houtsma_atomically_2021,chen2021sub}.
A gap opening in these samples has been experimentally detected by resistivity measurements\cite{han2007energy,chen2007graphene} and scanning tunneling spectroscopy up to room temperature\cite{magda2014room}, and some degree of electron localization has been reported\cite{ruffieux_-surface_2016}. Still, the link to a possible AFM order could
be made only
on the basis of DFT or model Hubbard Hamiltonians gap-size predictions.
Another possible way of verification is to perform first-principles calculations by using high-level theories,
more accurate than
standard
DFT but also computationally much more expensive.
Previous studies on the acene molecular series
revealed that the paramagnetic solution for their
GS
is challenged by the competition against the formation of localized edge states, AFM coupled in an open-shell singlet configuration\cite{bendikov2004oligoacenes,hachmann2007radical,hajgato2009benchmark,chai2012density,rivero2013entanglement,yang2016nature,dupuy_fate_2018,schriber2018combined,trinquier2018qualitative}. Several correlated post-Hartree-Fock methods\cite{hachmann2007radical,hajgato2009benchmark,yang2016nature,schriber2018combined} and various types of DFT functionals\cite{bendikov2004oligoacenes,chai2012density} have been employed to
unveil
the true
GS,
with results strongly dependent on the level of theory\cite{tonshoff2021pushing}. Moreover, in this class of systems, numerical DFT predictions quantitatively depend on the chosen functional.
On the other hand, QMC in its diffusion Monte Carlo implementation is one of the most accurate \emph{ab initio} theories, and it can be utilized to provide benchmark results in such
correlated systems\cite{foulkes_quantum_2001,Wagner_2016,motta2020ground}.
This is the way we will purse in this paper in order to study the
GS
properties of the 2-ZGNR.
Various other attempts have been made to model strong correlations in graphene and graphene derivatives using Hubbard Hamiltonians\cite{raghu_structural_2002,srinivasan_structural_1998} on a honeycomb lattice. However, a relevant question still debated
in this framework is about the correct value of Hubbard U repulsion one has to use to reproduce the physics taking place in a more realistic setting\cite{jung2011nonlocal,wehling2011strength,schuler2013optimal,csacsiouglu2017strength,PhysRevB.98.205123}.
At variance with model Hamiltonian approaches\cite{golor2013quantum}, which suffer from this issue, our QMC framework
includes electron correlations entirely from first principles. By carrying out extensive variational and diffusion QMC calculations, we show that the AFM phase is the prime candidate for the $n$-ZGNR
GS, for $n=2$ and presumably for larger $n$.
We then compare the results obtained at the QMC level with
the
widely used DFT
to validate the use of certain DFT functionals, such as the hybrid PBE0, HSE and GauPBE
schemes,
and determine the
optimal
value of U to be used in the DFT+U framework.
The paper is organized as follows. Secs.~\ref{QMC-methods} and \ref{DFT-methods} introduce the QMC and DFT methods used in this work, respectively. Sec.~\ref{QMC-results} reports the ground-state properties obtained by QMC, while in Sec.~\ref{DFT-results} various DFT functionals are compared against QMC in order to find the ones that best reproduce the physics of 2-ZGNRs. Finally, Sec.~\ref{conclusions} draws the conclusions.
\section{METHODS}
\label{methods}
\subsection{QMC}
\label{QMC-methods}
All the QMC calculations were done using the TurboRVB package\cite{nakano_turborvb_2020}, a complete suite of \emph{ab initio} codes based on real-space Jastrow-correlated wave functions $\Psi$.
With the aim of unveiling the ground-state (GS) nature of the narrowest zigzag graphene nanoribbon, we compared the variational Monte Carlo (VMC) energies ($E_\textrm{VMC}$) of two types of wave functions:
\begin{enumerate}
\item A Jastrow-correlated single-determinant wave function, describing a paramagnetic (PM) state;
\item A wave function with spin-broken symmetry in both the Jastrow and the determinantal part with $z$-projected total spin $\hat{S}_z=0$, developing an AFM long-range order.
\end{enumerate}
To reach the closest solution to the true GS, by means of the lattice regularized diffusion Monte Carlo (LRDMC)\cite{casula_diffusion_2005} method
we
projected these wave functions to the lowest energy state $\Psi_\textrm{FN}$ compatible with the fixed node (FN) approximation\cite{reynolds_fixednode_1982}. The closer the nodes of the variational wave function to the GS ones, the lower the corresponding FN energy is. Thus, by comparing the FN energies ($E_\textrm{FN}$) of the two wave functions, one can determine what is the symmetry of the state that best represents the GS.
The two different PM and AFM many-body wave functions used in this work can be written in a unified form as:
\begin{equation}
\Psi^\textbf{k}(\textbf{R})=\exp\{-U(\textbf{R},\Sigma)\} \det\{\phi^{\textbf{k},\uparrow}_j(\textbf{r}^\uparrow_i)\}\det\{\phi^{\textbf{k},\downarrow}_j(\textbf{r}^\downarrow_i)\},
\label{wf}
\end{equation}
for $i,j \in \{1,\ldots,N/2\}$, where $N$ is the number of electrons in the simulation supercell, $\textbf{k}$ is the twist belonging to the supercell Brillouin zone,
$U$ is the Jastrow function, $\phi^{\textbf{k},\sigma}_i$ are one-body orbitals,
$\textbf{R}=\{\textbf{r}^\uparrow_1,\ldots,\textbf{r}^\uparrow_{N/2},\textbf{r}^\downarrow_1,\ldots,\textbf{r}^\downarrow_{N/2}\}$ is the $N$-electron spacial coordinate, and $\Sigma$ is the $N$-electron spin configuration. For the paramagnetic case, both Jastrow function and one-body orbitals do not depend on $\Sigma$, i.e. $U(\textbf{R},\Sigma)=U(\textbf{R})$ and $\phi^{\textbf{k},\uparrow}_i=\phi^{\textbf{k},\downarrow}_i=\phi^\textbf{k}_i$. Otherwise, a spin-broken solution can be stabilized.
The Jastrow function $U$ is split into electron-nucleus, electron-electron, and electron-electron-nucleus parts:
\begin{equation}
U=U_{en}+U_{ee}+U_{een}.
\label{JasU}
\end{equation}
The electron-nucleus function has an exponential form and it reads as
\begin{equation}
U_{en}=\sum_{iI} J_{1b}(r_{iI}),
\label{JasUen}
\end{equation}
where the index $i$($I$) runs over electrons (nucleus), $r_{iI}$ is the electron-nucleus distance, and
\begin{equation}
J_{1b}(r)=\alpha \left( 1-e^{-r/\alpha} \right),
\label{J1b}
\end{equation}
with $\alpha$ a variational parameter. $J_{1b}$ cures the nuclear cusp conditions, and allows the use of the bare Coulomb potential for the hydrogen atom in our QMC framework. For the carbon atom, we used instead the Burkatzki-Filippi-Dolg (BFD) pseudopotential\cite{burkatzki_energy-consistent_2007}, to replace the $1s$ core electrons. The electron-electron function has a Pad\'e form and it reads as
\begin{equation}
U_{ee}=-\sum_{ij} J_{2b}(r_{ij},\sigma_i,\sigma_j),
\label{JasUee}
\end{equation}
where the indices $i$ and $j$ run over electrons, $r_{ij}$ is the electron-electron distance, $\sigma_i$ is the spin of the $i$-th electron, and
\begin{eqnarray}
J_{2b}(r,\sigma,\bar{\sigma}) & = & 0.5~ r/(1+ \beta r),
\label{J2b_anti}
\\ J_{2b}(r,\sigma,\sigma) & = & 0.25 ~r/(1+ \beta r),
\label{J2b_para}
\end{eqnarray}
with $\beta$ a variational parameter. This two-body Jastrow term fulfills the cusp conditions for both parallel $(\sigma_i=\sigma_j=\sigma)$ and antiparallel $(\sigma_i=\sigma, \sigma_i=\bar{\sigma})$ spins. The last term in the Jastrow factor is the electron-electron-nucleus function:
\begin{equation}
U_{een}=\sum_{ijI} \sum_{\gamma\delta} M_{\gamma \delta I}(\sigma_i,\sigma_j)\chi_{\gamma I}(r_{iI}) \chi_{\delta I}(r_{jI}),
\label{JasUeen}
\end{equation}
with $M_{\gamma \delta I}(\sigma,\sigma^\prime)$ a matrix of variational parameters symmetric in both orbital $(\gamma,\delta)$ and spin $(\sigma,\sigma^\prime)$ sectors, and $\chi_{\gamma I}(r)$ a $(2s,2p,1d)$ Gaussian basis set, with orbital index $\gamma$, centered on the nucleus $I$. Note that the sum in Eq.~\ref{JasUeen} runs over a single nuclear index $I$. Thus, in the variational wave function we neglect inhomogeneous correlations between electrons belonging to different sites, in order to reduce the number of relevant parameters. The $J_{1b}$ and $J_{2b}$ Jastrow functions have been periodized using a $\textbf{r} \rightarrow \textbf{r}^\prime$ mapping that makes the distances diverge at the border of the supercell, as explained in Ref~\onlinecite{nakano_turborvb_2020}. For the inhomogeneous $U_{een}$ part, the Gaussian basis set $\chi$ has been made periodic by summing over replicas translated by lattice vectors. The Gaussian basis set of the Jastrow factor is constructed such that the radial part of the second-$s$ and second-$p$ Gaussian functions have a $r^2 e^{-Z r^2}$ form, to enforce orthogonality in the basis set. For the paramagnetic system, $M_{\gamma \delta I}(\sigma_i,\sigma_j)$ is taken spin independent, with the same entries for all spin sectors. Similarly, $J_{2b}(r,\sigma,\sigma^\prime)=0.5 r/(1+ \beta r) ~~ \forall \sigma, \forall \sigma^\prime$, namely $J_{2b}$ will fulfill only the electron cusps conditions for antiparallel spins in the spin-independent case.
The one-body orbitals $\phi$ are expanded onto a primitive $(8s,8p,2d,1f)$ Gaussian basis set for carbon (C), $(4s,2p,1d)$ for hydrogen (H), which we contracted into an optimal set of 9 hybrid orbitals for C and 2 hybrid orbitals for H (9C\&2H), by using the geminal embedding orbitals (GEO) contraction scheme\cite{sorella_geminal_2015} at the $\Gamma$ point. $\phi$s' are made periodic by using the same scheme as for the $\chi$s'.
The best orbital contractions are determined at the DFT level within the local density approximation (LDA), for the system in the paramagnetic phase.
The TurboRVB package allows one to perform LDA-DFT calculations in a Gaussian basis set, and with a mean-field wave function as in Eq.~\ref{wf} supplemented with the one-body Jastrow factor only, namely with $U=U_{en}$ in Eq.~\ref{JasU}.
Note that thanks to the one-body Jastrow factor included in the DFT wave function, it is possible to use the \emph{ab initio} Hamiltonian with bare Coulomb potential for the H atom. Not only the fulfillment of electron-ion cusp conditions\cite{pack_cusp_1966} via the Jastrow factor allows the use of divergent electron-ion potentials in the Hamiltonian, but it also significantly speeds up the basis set convergence already at the DFT level. Thus, it avoids the use of large Gaussian exponents, which are then cut off from the regular basis set. By employing a primitive Gaussian basis set of cc-pVTZ\cite{kendall_electron_1992,davidson_comment_1996} quality, optimized for the C pseudopotential\cite{burkatzki_energy-consistent_2007}, we determined the optimal GEO-contracted orbitals by requiring a convergence within 0.5 mH from the cc-pVTZ energy. It turns out that the 9C\&2H GEO contractions (see Tab.~\ref{tab:geo-contractions}) yield the desired accuracy at a less computational expense and with less variational parameters in the subsequent QMC calculations.
\begin{table}[ht]
\caption{\label{tab:geo-contractions}
2-ZGNR energies from LDA-DFT $\Gamma$-point calculations for the primitive cc-pVTZ and GEO contracted basis sets. Sub-mHa convergence is reached with a 9C\&2H GEO contraction.
}
\begin{tabular}{ccc}
\hline
\hline
Basis set & Energy per C (Ha) & Overlap with cc-pVTZ \\
\hline
\hline
cc-pVTZ & -5.95858 & 1.00000 \\
4C\&2H & -5.75062 & 0.89992 \\
7C\&2H & -5.95597 & 0.99927 \\
9C\&2H & -5.95802 & 0.99983 \\
\hline
\hline
\end{tabular}
\end{table}
Beside the basis set error, which is usually very mild in LRDMC calculations, another source of bias in QMC
energies
of extended systems comes from finite-size (FS) errors. In this work, the 2-ZGNR is put in a 3-dimensional periodic box, aligned along the $x$-axis and with vacuum separation of 7.2 \AA\ and 7 \AA\ in $y$ and $z$ directions, respectively, to avoid interaction between periodic images (see App.~\ref{DFT_appendix} and Fig.~\ref{fig:vac-z}). We used twisted boundary conditions for box-periodicity\cite{lin2001twist}.
Finite-size scaling analysis has been done by extending the unit cell in the $x$ direction (see Fig.~\ref{fig:unit-cell}).
We controlled FS errors by using the exact special twist method\cite{dagrada_exact_2016}. The special twist is the $\textbf{k}_s$-point in the first Brillouin zone (IBZ) where the energy is equal (or numerically very close) to the energy averaged over the full IBZ. Thus, it is very similar in spirit to the Baldereschi $\textbf{k}$-point\cite{baldereschi_mean-value_1973}, which is instead determined by symmetry arguments only. This procedure is ``exact'' in DFT or in other one-body theories, while it is an approximation in QMC, since the latter is affected by both one-body and two-body FS errors\cite{kent_finite-size_1999,kwee_finite-size_2008}. However, it has been proven that QMC calculations performed at the special twist show a smooth energy dependence on the supercell size\cite{dagrada_exact_2016}.
We carried out LDA-DFT calculations in the Gaussian GEO-contracted basis set to determine the special twist at a given supercell size in the PM phase. The quasi-1D nature of the 2-ZGNRs restricts this search to $\textbf{k}_s=(k_x,0,0)$, for a ribbon oriented along the $x$ direction. The $k_x$ found for different supercell sizes is reported in Tab.~\ref{tab:pm-QMCenergies}. We used the same $\textbf{k}_s$ in the AFM phase (Tab.~\ref{tab:afm-QMCenergies}).
Final LDA-DFT calculations at the special $\textbf{k}_s$-point are performed to find the self-consistent Kohn-Sham orbitals, plugged in the variational wave function of Eq.~\ref{wf} as starting $\phi^{\textbf{k}_s,\sigma}_i$ one-body orbitals. These calculations are performed at both LDA and local spin density approximation (LSDA) levels. The LDA-DFT does not
break the spin symmetry and initializes the one-body orbitals of the PM wave function. For the LSDA-DFT, we impose an initial magnetic field which stabilizes an AFM pattern, and switched off after the first iteration, in order to initialize the orbitals of the AFM wave function.
To correlate the Slater determinant generated by DFT, we applied the Jastrow factor leading to the QMC wave function of Eq.~\ref{wf}. In order to obtain the best trial wave function in the variational sense (i.e. as close as possible to the true ground-state), we optimize\cite{sorella_wave_2005,umrigar2007alleviation} the variational parameters by minimizing the QMC energy of the variational wave function $\Psi$. At the first QMC step, the $\alpha$, $\beta$ and $M_{\gamma \delta I}(\sigma,\sigma^\prime)$ linear coefficients of the Jastrow part are optimized. Then, the Gaussian exponents in the Jastrow basis set $\chi$ are relaxed as well. Afterwards, the linear coefficients in $\phi^{\textbf{k}_s,\sigma}_i$ and the Gaussian exponents of the determinantal part are also optimized, simultaneously with the Jastrow factor. As a last step, the 2-ZGNR geometry is relaxed together with the wave function, thanks to the latest progress in the QMC calculations of ionic forces\cite{doi:10.1063/1.3516208}.
After fully optimizing the QMC trial wavefunctions, we performed
VMC simulations to evaluate the variational energy $E_\textrm{VMC}$ for both paramagnetic and AFM states. Finally, we carried out
LRDMC calculations based on the optimized variational wave functions. The LRDMC is a lattice regularized version of the standard diffusion Monte Carlo (DMC) algorithm\cite{casula_diffusion_2005}, obtained by discretizing the Laplacian on an adaptive mesh\cite{nakano2020speeding}. We used a discretization step of 0.25 Bohr, which guarantees unbiased energy differences. Like DMC\cite{reynolds_fixednode_1982}, the LRDMC is a stochastic projection method to solve the imaginary-time Schr\"odinger equation, within the constraints imposed by the FN approximation. Moreover, LRDMC is fully compatible with the use of non-local pseudopotentials, such as the one for the carbon atom.
The wave function projection is done through a multi-walker branching algorithm with fixed number $n_w$ of walkers\cite{buonaura1998numerical}. We used $n_w=9216$ for the largest system, with 50 hops per walker between two branching steps, and less than 0.5\% of the walkers population killed per step. The residual finite population bias is then corrected by the correcting factors\cite{buonaura1998numerical}.
At convergence, the LRDMC algorithm samples the mixed distribution $\Psi_\textrm{FN}\Psi$, being $\Psi_\textrm{FN}$ the projected wave function. Thus, one evaluates the FN energy as $E_\textrm{FN}=\langle \Psi_\textrm{FN} | H_\textrm{FN} | \Psi \rangle / \langle \Psi_\textrm{FN} | \Psi \rangle $. $H_\textrm{FN}$ is the effective FN Hamiltonian implementing the FN constraints,
whose GS is exactly $\Psi_\textrm{FN}$.
The variational character of $E_\textrm{FN}$ implies that $E_0 \le E_\textrm{FN} < E_\textrm{VMC} $, where $E_0$ is the unknown true GS energy of the system, and $E_\textrm{VMC}$ is the energy of the wave function $\Psi$ before projection. The difference between $E_0$ and $E_\textrm{FN}$ is the FN error.
The above steps (wave function initialization, optimization and projection) are repeated for different supercell sizes, in order to reduce many-body FS errors. For each phase, we employed 3 supercells, made of 12 C (3-ring periodic), 24 C (6-ring periodic) and 36 C (9-ring periodic), respectively. The VMC and LRDMC energies of these 3 supercells are then corrected by applying the Kwee-Zhang-Krakauer (KZK)\cite{kwee_finite-size_2008} energy functional. The KZK-corrected energies, reported in Tabs.~\ref{tab:pm-QMCenergies} and \ref{tab:afm-QMCenergies}, are finally extrapolated to the thermodynamic limit by performing a linear fit in $1/N$ (see Fig.~\ref{fig:FS-total-energy}). The extrapolated energies are thus FS error free. Not only the energies but also other physical observables, such as the absolute magnetization, can be extrapolated to their thermodynamic values by following their appropriate FS-scaling (Fig.~\ref{fig:mag-mom}).
\subsection{DFT}
\label{DFT-methods}
While DFT calculations that are preliminary for our QMC simulations were done by using the Gaussian basis set code as implemented in the TurboRVB package (see Sec.~\ref{QMC-methods}),
all the other
DFT\cite{hohenberg_inhomogeneous_1964} calculations were performed using the plane-wave (PW) based Quantum ESPRESSO\cite{giannozzi_quantum_2009,giannozzi_advanced_2017} package.
The 2-ZGNR unit cell was defined as in the QMC calculations (see Fig.~\ref{fig:unit-cell}). At variance with QMC, no supercells are used in DFT. The vacuum separation in $y$ and $z$ directions was kept to 7.2 \AA~ and 7 \AA, respectively (see Fig.~\ref{fig:vac-z} of App.~\ref{DFT_appendix}).
One of the aims of this work is to study how DFT performs in ZGNRs as compared with benchmark QMC results. Various functionals have been used for this comparison:
LDA\cite{ceperley_ground_1980,perdew_self-interaction_1981}, LDA+U\cite{PhysRevB.52.R5467,PhysRevB.44.943}, the generalized gradient approximation in the Perdew–Burke–Ernzerhof implementation (GGA-PBE)\cite{perdew_generalized_1996}, GGA-PBE+U\cite{PhysRevB.57.1505}, the DFT-DF2 (van der Waals corrected) functional\cite{PhysRevB.82.081101}, the PBE functional revised for solids (PBEsol)\cite{PhysRevLett.100.136406}, the Becke-Lee-Yang-Parr (BLYP) functional\cite{PhysRevB.37.785}, the Gaussian-PBE (GauPBE)\cite{song_communication_2011}, Heyd–Scuseria–Ernzerhof (HSE)\cite{doi:10.1063/1.1564060,doi:10.1063/1.2204597}, and PBE0\cite{doi:10.1063/1.472933,adamo_1999_toward} hybrid fuctionals.
In our PW-DFT calculations, the electron-ion interactions are described by the optimized norm-conserving Vanderbilt (ONCV) pseudopotentials\cite{PhysRevB.88.085117} for PBE and PBE-based hybrid functionals, while we used projector-augmented wave (PAW) pseudopotentials\cite{paw_pseudos} for LDA, LDA+U, PBE+U, PBEsol, and BLYP functionals.
A kinetic energy (charge density) cut-off of 80 Ry (320 Ry) was used in all PW calculations. This guarantees a sub-meV accuracy for both ONCV and PAW pseudopotentials. For the $\textbf{k}$-points integration, the Brillouin zone was sampled with 32 $\textbf{k}$-points in the ribbon direction, corresponding to a
$32 \times 1 \times 1$ Monkhorst-Pack grid\cite{monkhorst_special_1976}. We used a Marzari-Vanderbilt\cite{marzari_ensemble_1997} smearing of 0.006 Ry. This setup yields an accuracy below 1 meV in total energies (see App.~\ref{DFT_appendix} and Fig.~\ref{fig:smearing-convergence}).
In the case of hybrid functionals, a downsampled $\textbf{q}$-grid of $16 \times 1 \times 1$ has been employed for the calculation of the exact-exchange operator, which again fulfills a target accuracy of 1 meV in total energies.
The DFT+U calculations were performed within the
Cococcioni's scheme to evaluate the Hubbard U operator \cite{cococcioni2005linear},
acting on the carbon $p$ states. This approach turns out to be equivalent to the Liechtenstein's formulation\cite{harmon_calculation_1995},
where we set the Hund's parameter $J=0$. Indeed, $\pi$-electron localization involves a single $p_z$ orbital per C site. In this configuration, the on-site intra-orbital Hubbard $U$ parameter is enough to fully characterize the Hubbard repulsion, and the two approaches coincide.
To check the computational accuracy of the aforementioned exchange-correlation functionals for
the 2-ZGNR,
we performed different set of calculations. We computed magnetic and non-magnetic solutions of 2-ZGNR
with:
\begin{itemize}
\item
fully relaxed DFT geometries within LDA, LDA+U, PBE, PBE+U, PBEsol and BLYP functionals. For the calculations with hybrid functionals, we used PBE-relaxed geometries;
\item
fully relaxed QMC geometries.
\end{itemize}
Geometry relaxation was converged with ionic forces below the $10^{-4}$ Ha/\AA\ threshold, pressure below $10^{-1}$ kbar.
The final geometries are reported in Tabs.~\ref{tab:table-pm-geo} and \ref{tab:table-afm-geo}. The energy stability of the AFM phase and its absolute magnetization are reported in Sec.~\ref{DFT-results}.
\section{RESULTS AND DISCUSSION}
\label{results}
\subsection{Ground-state properties from QMC}
\label{QMC-results}
We performed QMC calculations for two different possible solutions of the 2-ZGNR ground state: the unpolarized
PM phase and
the
AFM state. If for the GS of the system it is energetically more convenient to break the spin symmetry and become spin-polarized, the Lieb's theorem\cite{lieb1989two}, applied to a bipartite honeycomb lattice with on-site interactions, states that the solution must be antiferromagnetic. Thus, the PM energy needs first to be compared with the AFM one, in order to know whether the 2-ZGNR GS develops a long-range spin order, as predicted by previous DFT calculations. This question is particularly relevant for quasi-1D systems, where quantum fluctuations are strong and can destroy long-range charge and spin orders. Electron correlations on top of enhanced quantum fluctuations make the GS determination a hard task. We will see that electron correlations in 2-ZGNR are indeed not negligible, and this adds another layer of complexity to the problem.
To tackle the study of 2-ZGNR, we used a fully-fledged quantum Monte Carlo approach, as described in Sec.~\ref{QMC-methods}. By first optimizing and then projecting the variational wave function in Eq.~\ref{wf}, we obtained the fixed-node GS energy, namely the best fixed-node DMC energy compatible with the nodes of the wave function before projection. We would like to highlight here that the starting LDA nodes, as defined by the LDA Kohn-Sham orbitals of the preparatory Gaussian DFT calculations, have been relaxed and optimized in the energy minimization step, in the presence of the correlated Jastrow factor. Thus, in the DMC step we used improved nodes which have a milder FN error. The combination of a full wave function optimization performed at the VMC level, followed by its further projection through the LRDMC
diffusion Monte Carlo algorithm, makes the QMC a very accurate framework in solid state physics\cite{Wagner_2016,saritas2017investigation,benali2020toward,nakano2020speeding,motta2020ground}.
During the wave function optimization step, not only we improved the electronic wave function parameters but also we relaxed the geometry of the system. Indeed, both wave function optimization and structural relaxation are based on the variational energy minimization. Thus, they can be performed simultaneously. By starting from the PBE geometry, we relaxed the symmetry-independent structural parameters till convergence.
Our
QMC statistics
is such that, at convergence, the bond lengths fluctuate around $\approx 0.02$\AA. The optimal geometries found at the VMC level for the PM and AFM phases are reported in Tabs.~\ref{tab:table-pm-geo} and \ref{tab:table-afm-geo}, respectively. It is worth noting that the variation between the two relaxed geometries mainly involves the C3-C6 vertical distance, and to a minor extent the C1-C6 and C-H bond lengths (see Fig.~\ref{fig:unit-cell}). The AFM phase has looser bonds linked to the outermost carbon sites of the zigzag chain. This affects the values of both C1-C6 and C6-H1 independent parameters. As we will see later, this is due to a quite strong electron localization, evidenced by the formation of
sizable
local magnetic moments at the edges of the 2-ZGNR. As a consequence, the electron localization in the AFM state weakens the covalent chemical bonds around the terminal C sites, causing their elongation. Another concomitant effect is the widening of the $\widehat{\textrm{C2-C1-C6}}$ angle, leading to a larger C3-C6 distance in the AFM phase. This angle goes from 118.56(11) degrees in the PM state to 119.09(18) degrees in the AFM phase.
\begin{table}[t!]
\begin{ruledtabular}
\caption{\label{tab:table-pm-geo} Fully relaxed PM geometries (lengths in \AA) at different level of theory. We report only the values of independent parameters, from which the whole geometry can be fully reconstructed. By symmetry, C1-C6 $=$ C5-C6 $=$ C2-C3 $=$ C4-C3, C1-C2 $=$ C4-C5, and C6-H1 $=$ C3-H2.
}
\begin{tabular}{cccccc}
Bond length (\AA) & \multicolumn{1}{l}{QMC} & \multicolumn{1}{l}{PBE} & \multicolumn{1}{l}{BLYP} & \multicolumn{1}{l}{LDA+U} & \multicolumn{1}{l}{PBE+U} \\
\hline
\hline
C1-C6 & 1.391 & 1.402 & 1.408 & 1.353 & 1.373 \\
C1-C2 & 1.439 & 1.461 & 1.469 & 1.411 & 1.432 \\
C3-C6 & 2.769 & 2.811 & 2.821 & 2.701 & 2.745 \\
C6-H1 & 1.078 & 1.091 & 1.092 & 1.081 & 1.081
\end{tabular}
\end{ruledtabular}
\end{table}
\begin{table}[t!]
\begin{ruledtabular}
\caption{\label{tab:table-afm-geo} As in Tab.~\ref{tab:table-pm-geo} but for fully relaxed AFM geometries (lengths in \AA).}
\begin{tabular}{cccccc}
Bond length (\AA) & \multicolumn{1}{l}{QMC} & \multicolumn{1}{l}{PBE} & \multicolumn{1}{l}{BLYP} & \multicolumn{1}{l}{LDA+U} & \multicolumn{1}{l}{PBE+U} \\
\hline
\hline
C1-C6 & 1.398 & 1.404 & 1.408 & 1.362 & 1.381 \\
C1-C2 & 1.439 & 1.459 & 1.469 & 1.403 & 1.424 \\
C3-C6 & 2.799 & 2.811 & 2.822 & 2.703 & 2.746 \\
C6-H1 & 1.083 & 1.091 & 1.092 & 1.080 & 1.081
\end{tabular}
\end{ruledtabular}
\end{table}
The FN-GS energies yielded by the PM and AFM wave functions computed at their respective relaxed geometries are plotted in Fig.~\ref{fig:FN_energies}, as a function of $1/N$.
\begin{figure}[ht!]
\includegraphics[width=0.9\columnwidth]{KZK_scaling.pdf}
\caption{\label{fig:FS-total-energy} Finite-size scaling of AFM (green) and PM (violet) fixed-node LRDMC energies per C atom.}
\label{fig:FN_energies}
\end{figure}
One can see that,
although
the PM and AFM energies per C atom are nearly degenerate for a 3-ring supercell, the different slope of their finite-size scaling let the AFM energy be the lowest in the thermodynamic limit. Therefore, our QMC results confirm the early DFT finding of an antiferromagnetic ground state. Remarkably, the difference between the PM and AFM FN energies (AFM energy gain), once extrapolated to the thermodynamic limit, is large. It amounts to 36$\pm$3 meV per C atom (see Tab.~\ref{tab:afm-gain-qmc-geo}), significantly larger than previous predictions.
To give a more exhaustive explanation of this finding, let us analyze the two-dimensional (2D) spin magnetization density $\sigma_\textrm{2D}(x,y)=\langle \hat{\sigma}_\textrm{2D}(x,y) \rangle$, defined as the expectation value of the following operator:
\begin{equation}
\hat{\sigma}_\textrm{2D}(x,y)= - g \mu_B \sum_{i=1}^N \hat{\sigma}_z(i) \delta(\hat{x}_i - x)\delta(\hat{y}_i -y),
\label{spin_density}
\end{equation}
where $\hat{\sigma}_z(i)$ is the $z$-component of the spin vector operator acting on the $i$-th particle spinor, $\hat{x}$ ($\hat{y}$) is the $x$ ($y$) component of the position operator, $\mu_B$ is the Bohr magneton, and the $g$-factor is taken equal to 2. The LRDMC expectation value of the spin density operator in Eq.~\ref{spin_density} is computed over the mixed distribution, such that $\sigma^\textrm{LRDMC}_\textrm{2D}(x,y)=\langle \Psi_\textrm{FN} | \hat{\sigma}_\textrm{2D}(x,y) | \Psi \rangle/ \langle \Psi_\textrm{FN}| \Psi \rangle$. As usual for operators that do not commute with the Hamiltonian, this value does not coincide with the ``pure'' estimator of the FN ground state: $\sigma_\textrm{2D}(x,y)=\langle \Psi_\textrm{FN} | \hat{\sigma}_\textrm{2D}(x,y) | \Psi_\textrm{FN} \rangle$. We corrected for the mixed distribution bias by approximating the pure estimators through the linear-order correction in $O(|\Psi_{FN}-\Psi|)$, i.e.
$\sigma_\textrm{2D} \approx 2 \sigma^\textrm{LRDMC}_\textrm{2D} - \sigma^\textrm{VMC}_\textrm{2D}$, where $\sigma^\textrm{VMC}_\textrm{2D}(x,y)=\langle \Psi | \hat{\sigma}_\textrm{2D}(x,y) | \Psi \rangle$, the regular VMC expectation value.
\begin{figure}[ht]
\includegraphics[width=\columnwidth]{2d_spin_new2.pdf}
\caption{\label{fig:contour-plot} Spin magnetization density
for AFM fixed-node GS wave function in a 9-ring supercell. The quantum expectation values over the FN GS are corrected for the LRDMC mixed distribution (see text). Units are in $\mu_B$/\AA$^2$.}
\end{figure}
In Fig.~\ref{fig:contour-plot}, we plot $\sigma_\textrm{2D}(x,y)$ for the AFM fixed-node GS of a 9-ring supercell, which yields values close to the thermodynamic limit. It is apparent that the spins are arranged in an AFM pattern, as expected from the LSDA initialization of the AFM wave function (see Sec.~\ref{QMC-methods}). However, the outermost C sites host magnetic moments as large as 0.5 $\mu_B$, much larger than the initial LSDA values (see also Sec.~\ref{DFT-results}). The spread of the peaks in the $\sigma_\textrm{2D}(x,y)$ distribution show a high degree of localization of the magnetic moments. For the inner C sites, the spin moments are
smaller than the terminal ones. The change of intensity between the outermost and the inner C atoms has also been reported in other ZGNR studies\cite{son_half-metallic_2006,PhysRevLett.99.186801,jiang_unique_2007,magda2014room}. Finally, in Fig.~\ref{fig:contour-plot} it can be seen that, in the AFM phase, the magnetic moments at the opposite edges are antiferromagnetically aligned, while across the ribbon direction the moments are ferromagnetically coupled. Instead, the PM phase does not show any pattern in $\sigma^\textrm{LRDMC}_\textrm{2D}(x,y)$, because the FN constraint
hampers the spin-symmetry breaking during the projection,
and the corresponding spin density vanishes within the error bars.
Fig.~\ref{fig:contour-plot} shows that the AFM pattern is a robust feature of the spin-broken LRDMC wave function. The intensity of this pattern is much more pronounced than in the LSDA calculations used to generate the initial wave function. Thus, electron correlations must enhance electron localization and local spin moment formation. To rigorously quantify this enhancement, we compute the absolute magnetization per unit cell $M$, defined as
\begin{equation}
M = \int_V \!\textrm{d}x \textrm{d}y ~ |\sigma_\textrm{2D}(x,y)|,
\end{equation}
where the integral is performed over the unit cell volume $V$.
$M/N_\textrm{C}$ gives the average magnetic moment per C atom in the cell. As we did for the spin magnetization density, we computed the value $M$ of the unbiased estimator for the LRDMC fixed-node wave function by means the first-order expression: $M=2M_\textrm{LRDMC}- M_\textrm{VMC}$. The finite-size scaling of $M$ is shown in Fig.~\ref{fig:mag-mom}. It linearly extrapolates to the value of 1.13 $\mu_\textrm{B}$, as also reported in Tab.~\ref{tab:afm-gain-qmc-geo}.
\begin{figure}[ht]
\includegraphics[width=0.9\columnwidth]{magnetic_moment_new.pdf}
\caption{\label{fig:mag-mom} Finite-size extrapolation of absolute magnetization per unit cell $M$, obtained for
the AFM FN ground state, corrected for the LRDMC mixed distribution.
}
\end{figure}
In Sec.~\ref{DFT-results}, we will compare our QMC results with the outcome of several DFT exchange-correlation functionals. We will see that the value of $M$ strongly correlates with the AFM energy gain. The larger $M$, the stronger the AFM pattern, and the more energetically stable the AFM GS is. This can easily be explained by assuming that the underlying model Hamiltonian governing the energetics and magnetism of the 2-ZGNR is the $t-J$ model restricted to the $p_z$ orbitals, where $t$ is the hopping, or hybridization strength, between two neighboring C sites, and $J$ is the spin-exchange coupling between nearest neighbors. Already at the mean-field level\cite{inui1988coexistence,jayaprakash1989mean,kyung2000new}, both energy and staggered magnetization of the $t-J$ model scale with $J$. Thus, they are correlated in the same way, and this correlation is stronger and stronger as the electrons get more and more localized. Indeed, strictly speaking, the $t-J$ representation holds in the strong electron localization limit. The positive linear relationship between the AFM energy gain and $M$ is also verified by their finite-size scaling behavior. Indeed, they both increase linearly as $N\rightarrow \infty$, as shown in Figs.~\ref{fig:FN_energies} and \ref{fig:mag-mom}.
\subsection{Ground-state properties from DFT and comparison with QMC}
\label{DFT-results}
As explained in Sec.~\ref{DFT-methods},
we fully relaxed the geometries with LDA, LDA+U, PBE, PBE+U, PBEsol and BLYP functionals for both PM and AFM phases, whenever the latter exists as stable phase. For further calculations with hybrid functionals, we used the PBE relaxed geometries. The relaxed PM and AFM geometries are reported in Tabs.~\ref{tab:table-pm-geo} and \ref{tab:table-afm-geo}, respectively, for the most relevant functionals, while the corresponding PM-AFM energy difference (AFM energy gain) per C atom and AFM absolute magnetization per unit cell $M$ are reported in Tab.~\ref{tab:afm-gain-pbe-geo}, and compared to our QMC benchmark calculations. Moreover, for the sake of comparison, in Tab.~\ref{tab:afm-gain-qmc-geo} we report the AFM energy gain and $M$ values computed for different functionals at the same PM and AFM geometries taken from QMC and kept then frozen.
\begin{table}[ht]
\caption{\label{tab:afm-gain-pbe-geo} AFM energy gain per C atom with respect to fully relaxed PM phase, and AFM absolute magnetization per unit cell $M$. All the calculations were done at the DFT optimized geometries, except for reference QMC values. These data are plotted in Figs.~\ref{fig:afm-gain-mag-mom}(a) and \ref{fig:afm-gain-mag-mom}(b).
}
\begin{ruledtabular}
\begin{tabular}{lcc}
Level of theory & \multicolumn{1}{l}{AFM gain (meV)} & \multicolumn{1}{l}{ $M$ ($\mu_B$)} \\
\hline
\hline
BLYP & 1.3 & 0.35 \\
PBE & 3.4 & 0.53 \\
DFT-DF2 & 5.9 & 0.63 \\
HSE & 24.2 & 0.92 \\
PBE0 & 34.0 & 1.02 \\
GauPBE & 28.6 & 1.03 \\
LDA+U=7.6 eV & 39.1 & 1.12 \\
PBE+U=5.0 eV & 33.9 & 1.13 \\
\hline
QMC (reference) & 36$\pm$3 & 1.13$\pm$0.01
\end{tabular}
\end{ruledtabular}
\end{table}
Figs.~\ref{fig:afm-gain-mag-mom}(a) and \ref{fig:afm-gain-mag-mom}(b) plot the data in Tab.~\ref{tab:afm-gain-pbe-geo}. By analyzing them, it is apparent that in 2-ZGNR there is a strong correlation between the size of absolute magnetization and the AFM energy gain, as already pointed out. The larger the magnetization, the more stable the AFM phase is. In fact,
in ZGNRs
$M$ is a proxy for the electron localization strength, as the local moment formation is directly related to the charge localization of a single $p_z$ electron per site, carrying a $1/2$ spin moment. The AFM exchange interaction $J$ between localized spins is then responsible for the lowering of the total energy in the AFM phase, and eventually for its gain with respect to the PM one. This is a strong coupling picture, where the Hubbard $U$ repulsion makes
the particles more localized. Therefore, only \emph{ab initio} schemes able to deal with strong correlation are also capable of correctly describing the 2-ZGNR ground state.
It turns out that LDA and PBEsol fail to yield the AFM as stable phase, as they are, among the tested functionals, those that tend to delocalize the electrons the most. As a consequence, there is no stable AFM solution within these two functionals. It is worth noting that this result crucially depends on the correct $\textbf{k}$-point sampling of the IBZ. Indeed, if the $\textbf{k}$-mesh is not dense enough, i.e. it contains less than 30 $\textbf{k}$-points along the ribbon direction (see the $\textbf{k}$-point convergence plot of Fig.~\ref{fig:smearing-convergence}), one can still obtain a stable AFM solution in LDA. However, this results into a very fragile phase, i.e. with very small $M$ and low AFM energy gain\footnote{This is why we have been able to initialize an AFM wave function in the Gaussian LSDA framework for further QMC calculations.}. This could explain the previous LDA outcome published in Ref.~\onlinecite{correa_braiding_2018}, which reports an AFM stabilization energy of a few meV, while in our case the AFM gain is null. The sensitiveness to $\textbf{k}$-points sampling is clearly due to a Dirac cone formation in the PM band structure, arising from edge states\cite{correa_braiding_2018}. This is particulary evident in the 2-ZGNR, while the Dirac velocities flatten out for larger $n$-ZGNR, yielding a braided band structure.
At variance with LDA and PBEsol,
PBE and BLYP functionals
predict a stable AFM order, although the resulting phase has a too weak electron localization, and correspondingly a too low AFM energy gain, if compared to QMC.
Even the inclusion of dispersion interactions
through
the DFT-DF2 functional does not improve the AFM energy gain and absolute magnetization. This shows that dispersion interactions are not so significant in curing the correlation problem in ZGNR, as expected.
\begin{table}[ht]
\caption{\label{tab:afm-gain-qmc-geo} As in Tab.~\ref{tab:afm-gain-pbe-geo} but for QMC optimized geometries. These data are plotted in Figs.~\ref{fig:afm-gain-mag-mom}(c) and \ref{fig:afm-gain-mag-mom}(d).
}
\begin{ruledtabular}
\begin{tabular}{lcc}
Level of theory & \multicolumn{1}{l}{AFM gain (meV)} & \multicolumn{1}{l}{$M$ ($\mu_\textrm{B}$)} \\
\hline
\hline
BLYP & 6.4 & 0.35 \\
PBE & 7.2 & 0.53 \\
DFT-DF2 & 9.5 & 0.63 \\
HSE & 25.1 & 0.91 \\
PBE0 & 30.1 & 0.99 \\
GauPBE & 28.9 & 1.01 \\
LDA+U=7.6 eV & 40.0 & 1.32 \\
PBE+U=5.0 eV & 32.3 & 1.21 \\
\hline
QMC (reference) & 36$\pm$3 & 1.13$\pm$0.01
\end{tabular}
\end{ruledtabular}
\end{table}
\begin{figure*}
\includegraphics[width=0.49\textwidth]{afm-gain-pbe-geo.pdf}
\includegraphics[width=0.49\textwidth]{mag-mom-pbe-geo.pdf}
\includegraphics[width=0.49\textwidth]{afm-gain-qmc-geo.pdf}
\includegraphics[width=0.49\textwidth]{mag-mom-qmc-geo.pdf}
\caption{\label{fig:afm-gain-mag-mom}
(a) AFM energy gain defined as energy difference between PM and AFM phases, computed by different exchange-correlation functionals.
The value of the Hubbard repulsion U used in DFT+U calculations is reported at the bottom.
The reference line corresponds to the value (36 meV) obtained at the QMC level. All the calculations are done at the corresponding DFT optimized geometries. (b) Absolute magnetization per unit cell
as a function of the same exchange-correlation functionals as used in panel (a).
The reference line corresponds to the value (1.13 $\mu_B$) obtained at the QMC level. All the calculations are done at corresponding DFT optimized geometries. (c)
The same as in (a) but for the QMC optimized geometries, taken same for all functionals.
(d)
The same as in (b) but for the QMC optimized geometries, taken the same for all functionals.
}
\end{figure*}
Since ``weakly correlated'' functionals, such as LDA, PBE, BLYP, PBEsol and DFT-DF2 functionals, fail to give an accurate description of the true GS, then use of
hybrid functionals
becomes
a
natural
choice.
Indeed,
hybrid functionals
can treat electron correlation effects in a better way,
because of explicit incorporation of a portion of exact exchange from Hartree-Fock theory\cite{FISCHER1986274}.
As a matter of fact,
the results obtained with HSE, PBE0 and GauPBE show a significant improvement over the LDA and GGA-based functionals.
PBE0 turns out to be the best hybrid functional for 2-ZGNRs, as it yields results close to QMC,
both in terms of AFM energy gain and $M$ values.
The reason of its apparent success
is the incorporation of fully non-local Hartree-Fock exchange\cite{adamo_1999_toward},
which favors the stabilization of
broken-symmetry phases, such as the
AFM long-range order, as verified in ZGNR by model Hamiltonians\cite{jung2011nonlocal}.
In HSE and GauPBE, the Hartree-Fock exchange is included only at short range, by using the error function and a Gaussian envelope as attenuation schemes\cite{song_communication_2011}, respectively, with the aim of screening the bare Coulomb interaction and making hybrid calculations more efficient.
They provide qualitatively similar results, with small variations arising from the different attenuation scheme employed. Nevertheless, on average, GauPBE performs slightly better than HSE in 2-ZGNR.
Despite their clear improvement with respect to weakly correlated functionals, hybrid functionals
are not able to fully
recover the QMC results. Therefore, in this work we explored
another popular way to include strong correlation in DFT, i.e.
the explicit incorporation of the on-site Coulomb interaction U in the DFT+U approach.
In order to find the optimal value of U, we required that the AFM solution provided by DFT+U yields the \emph{same} $M$ value as QMC \emph{at relaxed geometries}. Indeed, electron localization probed by $M$ is a key quantity to assess the correlation level reached in the system. Requiring the same level of localization is physically more sound than choosing the AFM energy gain as target quantity, which could depend on how the density functional is built and defined. We noticed that geometry affects the $p_z$ electron localization. Thus, we relaxed the geometry within DFT+U at a given U value and we then computed the final corresponding $M$. In this way,
we
found U=5.0 eV (U=7.6 eV) as optimal value in PBE+U (LDA+U).
Interestingly enough, these values
are fully in range with those predicted by
restricted random phase approximation (cRPA) calculations\cite{PhysRevB.98.205123} of ZGNRs, further strengthening our procedure. Indeed, site-specific cRPA estimates of the local U repulsion, based on the same PBE functional, show a reduction from the ``bulk'' value of graphene (U=9.3 eV)\cite{wehling2011strength,csacsiouglu2017strength} to the edge value of $\approx$ 5 eV, in nice agreement with our findings. In the narrowest 2-ZGNR, the edges play a dominant role in determining the site-averaged U repulsion, as the one we provide. It is also interesting to note that LDA+U requires larger U values than PBE+U. This agrees with a recent Bayesian calibration of Hubbard parameters in strongly correlated materials\cite{tavadze2021exploring}.
The DFT+U framework with optimal U values provides a further step forward with respect to the hybrid functionals for the 2-ZGNR description. LDA+U and PBE+U yield AFM energy gains that lie within one error bar from the best QMC estimates (Fig.~\ref{fig:afm-gain-mag-mom}(a)). At the same time, the $M$ values equal the QMC ones, thanks to the optimal-U construction (Fig.~\ref{fig:afm-gain-mag-mom}(b)).
The corresponding LDA+U and PBE+U geometries are reported in Tabs.~\ref{tab:table-pm-geo} and \ref{tab:table-afm-geo}. The agreement between DFT+U and QMC is less good for the geometries than for the energies.
From Tabs.~\ref{tab:table-pm-geo} and \ref{tab:table-afm-geo}, it is clear that a highly correlated method such as QMC gives shorter equilibrium bond lengths than the PBE and BLYP functionals. However, DFT+U overestimates the bond length contraction. LDA+U severely overshoots the bond shortening, while this effect is much milder in PBE+U, which provides at the end much better geometries than LDA+U.
Another consequence of electron correlation
is to enhance the structural differences between PM and AFM phases. While in PBE and BLYP the PM and AFM geometries are nearly the same, there is a significant variation in the QMC geometries across the phase change.
Upon inclusion of the on-site Coulomb interaction U, the difference between AFM and PM geometries becomes noticeable also in DFT+U. The C1-C6 and C3-C6 bond lengths in the AFM phase are longer than their counterparts in PM phase. This is true in both LDA+U and PBE+U, and it is in agreement with the QMC results.
Overall, supplementing the LDA and PBE functionals with local Hubbard U leads to equilibrium geometries that mimic more closely the structural behavior seen in QMC, with PBE+U outperforming LDA+U.
To test the robustness of the results shown in Figs.~\ref{fig:afm-gain-mag-mom}(a) and \ref{fig:afm-gain-mag-mom}(b) against structural variations, we computed both AFM gain and absolute magnetization $M$ at the PM and AFM geometries borrowed from QMC and kept the same for all functionals. This makes the comparison with QMC more direct because it avoids a possible source of disagreement. The corresponding results are reported in Tab.~\ref{tab:afm-gain-qmc-geo} and plotted in Figs.~\ref{fig:afm-gain-mag-mom}(c) and \ref{fig:afm-gain-mag-mom}(d).
The qualitative picture does not depend on the actual geometries chosen. The local and semi-local functionals drastically fail, while hybrid and DFT+U functionals provide much more reliable results. Even quantitatively, the picture stays almost the same, with variations of a few meV between the AFM gains computed by using relaxed geometries and QMC geometries. The most relevant difference is the change of absolute magnetization $M$ in the DFT+U framework upon geometry variation. At the QMC geometry, both LDA+U and PBE+U yield larger $M$ than the QMC values. Nevertheless, the PBE+U $M$ value is still very good, as it lies within 0.1 $\mu_B$ from the QMC reference. At the same time,
the PBE+U AFM gain is the closest to QMC among all tested functionals. The superior performances of PBE+U (U=5.0 eV) with respect to LDA+U (U=7.6 eV) and to the other functionals is thus verified for various independent properties: equilibrium geometries, AFM gain per C atom, AFM absolute magnetization per unit cell, robustness against geometry variation.
We finally looked for other possible instabilities arising from magnetism or geometry. Breaking the
crystal
symmetry by introducing \emph{cis} and \emph{trans} type of distortions\cite{kivelson_polyacene_1983,dos_santos_electronic_2006} in the fused benzene rings at both PBE and
GauPBE levels of theory does not lead to any structural instability in the AFM phase, for both functionals. Therefore, our results discard the existence of multiferroic ground states in neutral 2-ZGNR\cite{fernandez2008prediction,jung2009carrier}.
Instead, we did find structural instabilities in the PM phase. Nevertheless, the energy gained by introducing such distortions is very tiny ($< 0.1$ meV) as compared to the energy gain obtained by spin symmetry breaking. Thus, structural \emph{cis}/\emph{trans}- instabilities are irrelevant for the 2-ZGNR in its GS.
As far as the spin sector is concerned,
the ferromagnetic phase melts at the PBE level, and it is certainly not the GS
in GauPBE,
being always bound from below by the AFM phase. This is in agreement with Lieb's theorem\cite{lieb1989two}.
The AFM symmetry breaking is by far the most robust among possible instabilities. Our findings, based on QMC and ``correlated'' functionals, support the picture of a 2-ZGNR GS with localized $\pi$ electrons and long-ranged AFM correlations.
\section{CONCLUSIONS}
\label{conclusions}
In this work, we studied the ground state properties of the 2-ZGNR by means of very accurate QMC calculations. We found that the best candidate for the 2-ZGNR ground state is a wave function developing an AFM long-range order at zero temperature. Despite the low dimensionality of the system, the AFM phase is very robust against spin quantum fluctuations, and the AFM magnetic pattern is stable in our LRDMC simulations for all ribbons with supercell size equal to or more than 3 fused benzene rings. This is at variance with the result obtained for the acene series\cite{dupuy_fate_2018}, the molecular analogues of the 2-ZGNR, where a paramagnetic wave function is the primary candidate for their ground state. The consequences of making the ribbon length \emph{finite}, by chopping a 2-ZGNR into an acene molecule, deserve further studies.
This work also provides QMC benchmark results which have been used to validate different DFT functionals, both in terms of AFM energy gain and AFM absolute magnetization. We
tested
different DFT functionals such as LDA, LDA+U, PBE, PBE+U, DFT-DF2, BLYP, HSE, PBE0 and GauPBE functionals. For DFT+U frameworks, we determined the optimal value of U, thanks to the comparison with our QMC results.
We can conclude that the PBE+U functional with U=5.0 eV is the best among all the DFT functionals reported in this work, in terms of geometry, electron localization, magnetic moment, and AFM stabilization energy. The optimal U repulsion strength in PBE+U is in a very good agreement with its cRPA-PBE determination\cite{PhysRevB.98.205123}, and with a previous estimate based on the measured magnitude of ZGNR gaps and on the semiconductor–metal transition ZGNR width, found experimentally\cite{magda2014room}.
Hence, it would be rather safe to extend the use of PBE+U
from 2-ZGNR to computing materials properties of analogous C-based systems, with a graphene pattern: nanotubes, nanoribbons,
graphene impurities with doping and magnetism.
To conclude,
both
QMC and ``correlated'' functionals support the picture of a significant stabilization energy
of the AFM long-range order at zero temperature. This energy gain is much stronger than what predicted by previous DFT calculations using local or semi-local functionals. The quantitative failure of ``weakly correlated'' schemes is spectacular here. The AFM stabilization energy predicted by LRDMC, hybrids functionals and DFT+U schemes is from 5 to 7 times larger than the PBE one, and the associated local moments are from 2 to 3 times larger.
GW calculations suggested that many-body correlation effects could make the ZGNR band gaps 2-3 times larger than in PBE\cite{yang2007quasiparticle}. In this work we show how a non-perturbative many-body approach such as QMC further enhances the tendency to AFM ordering in 2-ZGNR, entirely from first principles. Previous hints based on non-perturbative many-body calculations came mainly from the solution of the Hubbard model on a honeycomb lattice, where the choice of the effective on-site repulsion remains critical for quantitative estimates\cite{wehling2011strength,schuler2013optimal}.
The remarkably strong stability of the AFM phase found in the 2-ZGNRs at zero temperature by \emph{ab initio} QMC techniques points towards the possibility of having stable antiferromagnetism above room temperature in this class of $\pi$-conjugated materials.
\begin{acknowledgments}
We are indebted to Prof. Prasenjit Ghosh, who took part in the early stage of the project, when RM was enrolled in the BS-MS Dual Degree Program involving the Department of Physics, Indian Institute of Science Education and Research (IISER), Pune, India, and the Institut de Min\'eralogie, de Physique des Mat\'eriaux et de Cosmochimie (IMPMC), Sorbonne Université, Paris, France. Prof. Prasenjit Ghosh acted as IISER supervisor.
We thank the European union for providing the Erasmus+ International Credit Mobility Grant for carrying out this collaborative project. RM would like to thank IISER Pune for providing a local cluster, Prithvi, which was used for preliminary DFT calculations. RM would also like to thank Wageningen University and Dutch National Supercomputing agency SURFsara for providing access to the Cartesius machine under the project EINF-750.
MC thanks GENCI for providing computational resources under the grant number 0906493, the Grands Challenge DARI for allowing calculations on the Joliot-Curie Rome HPC cluster under the project number gch0420, and RIKEN for the access to the Hokusai Greatwave supercomputer with the account number G19030.
This work was partially supported by the European Centre of Excellence in Exascale Computing TREX—Targeting Real Chemical Accuracy at the Exascale. This project has received funding from the European Union’s Horizon 2020 Research and Innovation program under Grant Agreement No. 952165.
\end{acknowledgments}
|
\section{Introduction}\label{Sec: introduction}
Stochastic partial differential equations (SPDEs) play a prominent r\^{o}le in modern analysis, probability theory and mathematical physics due to their effectiveness in modeling different phenomena, ranging from turbulence to interface dynamics. At a structural level, several progresses in the study of non linear problems have been made in the past few years, thanks to Hairer's work on the theory of regularity structures \cite{Hairer14, Hairer15} and to the theory of paracontrolled distributions \cite{Gubinelli} based on Bony paradifferential calculus \cite{Bony}.
Without entering into the details of these approaches, we stress that a common hurdle in all of them is the necessity of applying a renormalization scheme to cope with ill-defined product of distributions. In all these instances the common approach consists of introducing a suitable $\epsilon$-regularization scheme which makes manifest the pathological divergences in the limit $\epsilon\to 0^+$, see {\it e.g.} \cite{Hairer14}. This strategy is very much inspired by the standard approach to a similar class of problems which appears in theoretical physics and, more precisely, in quantum field theory on Minkowski spacetime, studied in momentum space.
Despite relying on a specific renormalization scheme, all these approaches have been tremendously effective in developing the solution theory of non linear stochastic partial differential equations in presence of an additive white noise. At the same time not much attention has been devoted to computing explicitly the expectation value and the correlations of the underlying solutions, features which are of paramount relevance in the applications, especially when inspired by physics.
In view of these comments, in a recent paper, \cite{Dappiaggi:2020gge}, it has been developed a novel framework to analyze scalar, non linear SPDEs in presence of an additive white noise. The inspiration as well as the starting point for such work comes from the algebraic approach to quantum field theory, see \cite{BFDY15, Rejzner:2016hdj} for reviews. In a few words this is a specific setup which separates on one side observables, collecting them in a suitable unital $*$-algebra encoding specific structural properties ranging from dynamics, to causality and the canonical commutation relations. On the other side, one finds states, that is normalized, positive linear functionals on the underlying algebra, which allow to recover via the GNS theorem the standard probabilistic interpretation of quantum theories.
Without entering into more details, far from the scope of this work, we stress that this approach, developed to be effective both in coordinate and in momentum space, has the key advantage of allowing an analysis of interacting theories within the realm of perturbation theory \cite{Brunetti:2009qc}.
Most notably, renormalization plays an ubiquitous r\^{o}le and, following an approach \`a la Epstein-Glaser, this is codified intrinsically within this framework without resorting to any specific \textit{ad hoc} regularization scheme \cite{Brunetti-Fredenhagen-00}.
From a technical viewpoint the main ingredients in this successful approach are a combination of the algebraic structures at the heart of the perturbative series together with the microlocal properties both of the propagators ruling the linear part of the underlying equations of motion and of the two-point correlation function of the chosen state.
In \cite{Dappiaggi:2020gge} it has been observed that the algebraic approach could as well be adapted also to analyze stochastic, scalar semi-linear SPDEs such as
\begin{align}\label{Eq: Generic equation}
E\widehat{\Phi}=\xi+F[\widehat\Phi].
\end{align}
Here $\widehat{\Phi}$ must be interpreted as a random distribution on the underlying manifold $M$, $\xi$ denotes the standard Gaussian, real white noise centered at $0$ whose covariance is $\mathbb{E}[\xi(x)\xi(y)]=\delta(x-y)$.
Furthermore $F:\mathbb{R}\to\mathbb{R}$ is a non-linear potential which can be considered for simplicity of polynomial type, while $E$ is a linear operator either of elliptic or of parabolic type -- see \cite{Dappiaggi:2020gge} for more details and comments.
Following \cite{Brunetti:2009qc} and inspired by the so-called functional formalism \cite{CDDR20, DDR20, Fredenhagen:2014lda}, we consider a specific class of distributions with values in polynomial functionals over $C^\infty(M)$.
While we refer a reader interested in more details to \cite{Dappiaggi:2020gge}, we sketch briefly the key aspects of this approach.
The main ingredients are two distinguished elements
\begin{align*}
\Phi(f;\varphi)=\int_M \varphi(x) f(x)\mu(x),\quad\boldsymbol{1}(f;\varphi)=\int_M f(x)\mu(x)
\end{align*}
where $\mu$ is a strictly positive density over $M$, $\varphi\in C^\infty(M)$ while $f\in C^\infty_0(M)$. These two functionals are employed as generators of a commutative algebra $\mathcal{A}$ whose composition is the pointwise product. The main rationale at the heart of \cite{Dappiaggi:2020gge} and inspired by the algebraic approach \cite{Brunetti:2009qc} is the following: The stochastic behavior codified by the white noise $\xi$ can be encoded in $\mathcal{A}$ by deforming its product setting for all $\tau_1,\tau_2\in\mathcal{A}$
\begin{align}\label{Eq: formal delta-product}
(\tau_1\cdot_Q \tau_2)(f;\varphi)
&=\sum_{k\geq 0}\frac{1}{k!}[(\delta_{\mathrm{Diag}_2}\otimes Q^{\otimes k})\cdot(\tau_1^{(k)}\otimes \tau_2^{(k)})](f\otimes1_{1+2k};\varphi)
=\sum_{k\geq 0}t_k(f\otimes1_{1+2k};\varphi)\,,
\end{align}
where $\tau_i^{(k)}$, $i=1,2$ indicate the $k$-th functional derivatives, while $Q=G\circ G^*$. Here $G$ ({\em resp.} $G^*$) is a fundamental solution associated to $E$ ({\em resp.} $E^*$, the formal adjoint of $E$), while $\circ$ indicates the composition of distributions. By a careful analysis of the singular structure of $G$ and in turn of $Q$ one can infer that the distributions $t_k$ are well-defined on $M^{2k+2}$ \clacomment{up to the total diagonal $\mathrm{Diag}_{2k+2}$ of $M^{2k+2}$, that is, $t_k(\cdot\,;\varphi)\in\mathcal{D}'(M^{2k+2}\setminus\mathrm{Diag}_{2k+2})$}.
Yet, adapting to the case in hand the results of \cite{Brunetti-Fredenhagen-00}, in \cite{Dappiaggi:2020gge} it has been proven that it is possible to extend $t_k$ to $\hat{t}_k\in\mathcal{D}^\prime(M^{2k+2})$. This renormalization procedure, when existent, might not be unique, but the ambiguities have been classified, giving ultimately a mathematical precise meaning to Equation \eqref{Eq: formal delta-product}.
For a further applications of these techniques to the analysis of \textit{a priori} ill-defined product of distributions see \cite{DRS21}.
As a consequence one constructs a deformed algebra $\mathcal{A}_{\cdot Q}$ whose elements encompasses at the algebraic level the information brought by the white noise.
Without entering into many details, which are left to \cite{Dappiaggi:2020gge} for an interested reader, we limit ourselves at focusing our attention on a specific, yet instructive example. More precisely we highlight that, if one evaluates at the configuration $\varphi=0$ the product of two generators $\Phi$ of $\mathcal{A}_{\cdot Q}$, one obtains
$$\left(\Phi\cdot_Q\Phi\right)(f;0)=\widehat{Q\delta_{\mathrm{Diag}_2}}(f),$$
where $f\in C^\infty_0(M)$ and where $=\widehat{Q\delta_{\mathrm{Diag}_2}}$ indicates a renormalized version of the otherwise ill-defined composition between the operator $Q$ and $\delta_{\mathrm{Diag}_2}$. The latter is the standard bi-distribution lying $\mathcal{D}^\prime(M\times M)$ such that $\delta_{Diag_2}(h)=\int_M\mu(x)\,h(x,x)$ for every $h\in C^\infty_0(M\times M)$.
A direct inspection shows that we have obtained the expectation value
$\mathbb{E}(\widehat{\varphi}^2(x))$ of the random field $\widehat{\varphi}^2(x)$, where $\widehat{\varphi}\vcentcolon=G\ast\xi$ is the so-called \emph{stochastic convolution}, which is a solution of Equation \ref{Eq: Generic equation} when $F=0$ and for vanishing initial conditions. With a similar procedure one can realize that $\mathcal{A}_{\cdot Q}$ encompasses the renormalized expectation values of all finite products of the underlying random field.
Without entering into further unnecessary details, we stress that an additional extension of the deformation procedure outlined above allows also to identify another algebra whose elements encompass at the algebraic level the information on the correlations between the underlying random fields. All these data have been applied in \cite{Dappiaggi:2020gge} to the $\Phi^3_d$ model. This has been considered as a prototypical case of a non linear SPDE and it has been analyzed at a perturbative level constructing both the solutions and their two-point correlations. Most notably renormalization and its associated freedoms have been intrinsically encoded yielding order by order in perturbation theory a renormalized equations which takes them automatically into account.
In view of the novelty of \cite{Dappiaggi:2020gge}, several questions are left open and goal of this paper is to address and solve a specific one. Most notably the class of SPDEs considered in the above reference contains only scalar distributions and consistently a real Gaussian white noise. Yet, there exists several models not falling within this class and a notable one, at the heart of this paper, goes under the name of stochastic nonlinear Schr\"odinger equation.
The reasons to consider this model are manifold. From a physical viewpoint, such specific equation is used in the analysis of several relevant physical phenomena ranging from Bose-Einstein condensates to type II superconductivity when coupled to an external magnetic field \cite{De Dominicis, Stoof, Sasik}. From a mathematical perspective, such class of equations is particularly relevant for its distinguished structural properties and it has been studied by several authors, see {\it e.g.} \cite{Bourgain, Hoshino-pre, Hoshino} for a list of those references which have been of inspiration to this work.
It is worth stressing that in many instances the attention has been given to the existence and uniqueness of the solutions rather than in their explicit construction or in the characterization of the mean and of the correlation functions. It is therefore natural to wonder whether the algebraic approach introduced in \cite{Dappiaggi:2020gge} can be adapted also to this scenario. A close investigation of the system in hand, see Section \ref{Sec: SNLS} for the relevant definitions, unveils that this is not a straightforward transition and a close investigation is necessary. The main reason can be ascribed to the presence of an additive {\em complex} Gaussian white noise, see Equations \eqref{Eq: complex white noise} and \eqref{Eq: complex white noise - 2} which entails that the only non vanishing correlations are those between the white noise and its complex conjugate.
This property has severe consequences, most notably the necessity of modifying significantly the algebraic structure at the heart of \cite{Dappiaggi:2020gge}.
\clacomment{A further remarkable difference with respect to the cases considered in \cite{Dappiaggi:2020gge} has to be ascribed to the singular structure of the fundamental solution of the Schr\"odinger operator. Without entering here into the technical details, we observe that especially the Epstein-Glaser renormalization procedure bears the consequences of this feature, since this calls for a thorough study of the scaling degree of all relevant integral kernels with respect to the submanifold $\Lambda^{2k+2}_t=\{(\widehat{t}_{2k+2},\widehat{x}_{2k+2})\,\vert\,t_1=\ldots=t_{2k+2}\}\subset M^{2k+2}$ rather than with respect to the total diagonal $\mathrm{Diag}_{2k+2}$ of $M^{2k+2}$, as in \cite{Dappiaggi:2020gge}.
As we shall discuss in Section \ref{Sec: graph}, the distinguished r\^{o}le of this richer singular structure emerges in the analysis of the subcritical regime which occurs only if the space dimension is $d=1$, contrary to what occurs in many parabolic models as highlighted in \cite{Dappiaggi:2020gge}}.
The goal of this work will be to reformulate the algebraic approach to SPDEs for the case on a non-linear Schr\"odinger equation and for simplicity of the exposition we shall focus our attention only to the case of $\mathbb{R}^{d+1}$ as an underlying manifold. Although the extension to more general curved backgrounds \cite{Dappiaggi:2020gge, RS21} is possible with a few minor modifications, we feel that this might lead us astray from the main goal of showing the versatility of the algebraic approach and thus we consider only the scenario which is more of interest in the concrete applications.
In the next subsection we discuss in detail the specific model that we study in this paper, but, prior to that, we conclude the introduction with a short synopsis of this work. In Chapter \ref{Sec: Preliminaries} we introduce the main algebraic and analytic ingredients necessary in this paper. In particular we discuss the notion of functional-valued distributions, adapting it to the case of an underlying complex valued partial differential equation. In addition we show how to construct a commutative algebra of functionals using the pointwise product and we discuss the microlocal properties of its elements. We stress that we refer to \cite{Hormander-I-03} for the basic notions concerning the wavefront set and the connected operations between distributions, while we rely on \cite[App. B]{Dappiaggi:2020gge} for a concise summary of the scaling degree of a distribution and of its main properties. In Section \ref{Sec: expectations} we prove the existence of $\mathcal{A}^{\mathbb{C}}_{\cdot Q}$ a deformation of the algebra identified in the previous analysis and we highlight how it can codify both the information of a complex white noise. Renormalization is a necessary tool in this construction and it plays a distinguished r\^{o}le in Theorem \ref{Thm: deformation map cdot} which is one of the main results of this work. In Section \ref{Sec: correlations} we extend the analysis first identifying a non-local algebra constructed out of $\mathcal{A}^{\mathbb{C}}_{\cdot Q}$ and then deforming its product so to be able to compute multi-local correlation functions of the underlying random field. At last, in Section \ref{Sec: perturbative analysis} we focus our attention on the stochastic non-linear Schr\"odinger equation. First we show how to construct at a perturbative level the solutions using the functional formalism. In particular we show that, at each order in perturbation theory, the expectation value of the solution vanishes and we compute up to first order the two-point correlation function. As last step, we employ a diagrammatic argument to discuss under which constraint on the dimension of the underlying spacetime, the perturbative analysis to all orders has to cope with a finite number of divergences, to be tamed by means of renormalization. In this respect we greatly improve a similar procedure proposed in \cite[Sec. 6.3]{Dappiaggi:2020gge}. To conclude, in Appendix \ref{App: microlocal} we give a proof of Theorem \ref{Thm: WF(G)}.
\subsection{The Stochastic Nonlinear Schr\"odinger Equation}\label{Sec: SNLS}
In this short subsection, we introduce the main object of our investigation, namely the {\em stochastic nonlinear Schr\"odinger equation} on $\mathbb{R}^{d+1}$, $d\geq 1$. In addition, throughout this paper, we assume that the reader is familiar with the basic concepts of microlocal analysis, see {\it e.g.} \cite[Ch. 8]{Hormander-I-03}, as well as with the notion of scaling degree, see in particular \cite{Brunetti-Fredenhagen-00} and \cite[App. B]{Dappiaggi:2020gge}.
More precisely, with a slight abuse of notation, by $\mathcal{D}^\prime(\mathbb{R}^d)$ we denote the collection of all random distributions and, inspired by \cite{Hoshino} we are interested in $\psi\in\mathcal{D}^\prime(\mathbb{R}^d)$ such that
\begin{equation}\label{Eq: Stochastic NLS}
i\partial_t\psi = \Delta\psi +\lambda|\psi|^{2\kappa}\psi+\xi,
\end{equation}
where $\lambda\in\mathbb{R}$, $\Delta$ is the Laplace operator on the Euclidean space $\mathbb{R}^d$, while $t$ plays the r\^{o}le of the time coordinate along $\mathbb{R}$. The parameter $\kappa\in\mathbb{N}$ controls the nonlinear behaviour of the equation and it is here left arbitrary since, in our approach, it is not necessary to fix a specific value, although a reader should bear in mind that, in almost all concrete models, {\it e.g.} Bose-Einstein condensates or the Ginzburg-Landau theory of type II superconductors, $\kappa=1$. At the same time we remark that the framework that we develop allows to consider a more general class of potentials in Equation \eqref{Eq: Stochastic NLS}, but we refrain from moving in this direction so to keep a closer contact with models concretely used in physical applications. For this reason, unless stated otherwise, henceforth we set $\kappa=1$ in Equation \eqref{Eq: Stochastic NLS}. The stochastic character of this equation is codified in $\xi$, which is a complex, additive Gaussian random distribution, fully characterized by its mean and covariance:
\begin{gather}\label{Eq: complex white noise}
\mathbb{E}[\xi(f)]=\mathbb{E}[\overline{\xi}(f)]=0,\\
\mathbb{E}(\xi(f)\xi(h))=\mathbb{E}(\overline{\xi}(f)\overline{\xi}(h))=0,\quad\mathbb{E}(\overline{\xi}(f)\xi(h))=(\overline{f},h)_{L^2},\label{Eq: complex white noise - 2}
\end{gather}
where $f,h\in\mathcal{D}(\mathbb{R}^{d+1})$, while $(,)_{L^2}$ represents the standard inner product in $L^2(\mathbb{R}^{d+1})$. The symbol $\mathbb{E}$ stands for the expectation value.
\begin{remark}
We stress that we have chosen to work with a Gaussian white noise centered at $0$ only for convenience and without loss of generality. If necessary and mutatis mutandis we can consider a shifted white noise, i.e. $\xi$ is such that the covariance is left unchanged from Equation \eqref{Eq: complex white noise} while
$$\mathbb{E}[\xi(f)]=\varphi(f)\quad\textrm{and}\quad\mathbb{E}[\overline{\xi}(f)]=\overline{\varphi}(f),$$
where $\varphi\in\mathcal{E}(\mathbb{R}^{d+1})$, while $\varphi(f)\doteq\int\limits_{\mathbb{R}^{d+1}}dx\,\varphi(x)f(x)$.
\end{remark}
To conclude the section, we observe that the linear part of Equation \eqref{Eq: Stochastic NLS} is ruled by the Schr\"odinger operator $L\doteq i\partial_t+\Delta$, which is a formally self-adjoint operator. For later convenience we introduce the fundamental solution $G\in\mathcal{D}^\prime(\mathbb{R}^{d+1}\times\mathbb{R}^{d+1})$ whose integral kernel reads
\begin{equation}\label{Eq: Integral Kernel}
G(x,y)=\frac{\Theta(t-t^\prime)}{(4\pi i (t-t^\prime))^{\frac{d}{2}}}e^{-\frac{|\underline{x}-\underline{y}|^2}{4i(t-t^\prime)}},
\end{equation}
where $x=(t,\underline{x})$, $y=(t^\prime,\underline{y})$ while $\Theta$ is the Heaviside function.
In comparison to the cases considered in \cite{Dappiaggi:2020gge}, $L$ is not a microhypoelliptic operator.
This translates into a distinguished singular structure of $G$ which is codified in its wavefront set.
In particular, we have the following theorem, whose proof is given in Appendix \ref{App: microlocal}.
\clacomment{
\begin{theorem}\label{Thm: WF(G)}
Let $G\in\mathcal{D}^\prime(\mathbb{R}^{d+1}\times\mathbb{R}^{d+1})$ be the fundamental solution of the Schr\"odinger operator as per Equation \eqref{Eq: Integral Kernel}.
Then
\begin{align}\label{Eq: Wavefront Set of G}
\mathrm{WF}(G)=\mathrm{WF}(\delta_{Diag_2})\cup\{(t,x,t,y,\omega,0,-\omega,0)\in T^*(\mathbb{R}^{1+d})^2\,,\,\omega\neq0\}.
\end{align}
\end{theorem}}
\begin{remark}[{\bf Notation}]\label{Rem: Cut-Off on G}
In our analysis, we shall be forced to introduce a cut-off to avoid infrared divergences, namely we consider a real valued function $\chi\in\mathcal{D}(\mathbb{R}^{d+1})$ and we define
$$G_\chi\doteq G\cdot(1\otimes \chi).$$
We stress that this modification does not change the microlocal structure of the underlying distribution, {\it i.e.} $\mathrm{WF}(G_\chi)=\mathrm{WF}(G)$. Therefore, since the cut-off is ubiquitous in our work and in order to avoid the continuous use of the subscript $\chi$, with a slight abuse of notation we will employ only the symbol $G$, leaving $\chi$ understood.
\end{remark}
\section{Analytic and Algebraic Preliminaries}\label{Sec: Preliminaries}
In this section we introduce the key analytic and algebraic tools which are used in this work, also fixing notation and conventions. As mentioned in the introduction, our main goal is to extend and to adapt the algebraic and microlocal approach to a perturbative analysis of stochastic partial differential equations (SPDEs), so to be able to discuss specific models encoding complex random distributions. In particular our main target is the stochastic non-linear Schr\"odinger equation as in Equation \eqref{Eq: Stochastic NLS} and, for this reason we shall adapt all the following definitions and structures to this case, commenting when necessary on the extension to more general scenarios.
\noindent The whole algebraic and microlocal programme is based on the concept of functional-valued distribution which is here spelt out. We recall that with $\mathcal{E}(\mathbb{R}^d)$ ({\em resp.} $\mathcal{D}(\mathbb{R}^d)$) we indicate the space of smooth ({\em resp.} smooth and compactly supported) complex valued functions on $\mathbb{R}^d$ endowed with their standard locally convex topology.
\begin{remark}\label{Rem: Key difference}
As mentioned in the introduction, due to the presence of a complex white noise in Equation \eqref{Eq: Stochastic NLS}, we cannot apply slavishly the framework devised in \cite{Dappiaggi:2020gge} since it does not allow to account for the defining properties listed in Equation \eqref{Eq: complex white noise}. Heuristically, one way to bypass this hurdle consists of adopting a different viewpoint which is inspired by the analysis of complex valued quantum fields such as the Dirac spinors, see {\it e.g.} \cite{Araki:1971id}. More precisely, using the notation and nomenclature of Section \ref{Sec: SNLS}, we shall consider $\psi$ and $\bar{\psi}$ as two a priori independent fields, imposing only at the end the constraint that they are related by complex conjugation. As it will become clear in the following, this shift of perspective has the net advantage of allowing a simpler construction of the algebra deformations, see in particular Theorem \ref{Thm: deformation map cdot} avoiding any potential ordering problem between the field and its complex conjugate. This might arise if we do not keep them as independent.
\end{remark}
\begin{definition}\label{Def: Functionals}
Let $d,m\in\mathbb{N}$. We call {\bf functional-valued distribution} $u\in\mathcal{D}^\prime(\mathbb{R}^{d+1};\mathsf{Fun}_{\mathbb{C}})$ a map
\begin{equation}\label{Eq: Functional valued distribution}
u:\mathcal{D}(\mathbb{R}^{d+1})\times\mathcal{E}(\mathbb{R}^{d+1})\times\mathcal{E}(\mathbb{R}^{d+1})\to\mathbb{C}\,,\quad (f,\eta,\eta^\prime)\mapsto u(f;\eta,\eta^\prime)\,,
\end{equation}
which is linear in the first entry and continuous in the locally convex topology of $\mathcal{D}(\mathbb{R}^{d+1})\times\mathcal{E}(\mathbb{R}^{d+1})\times\mathcal{E}(\mathbb{R}^{d+1})$.
In addition, we indicate the $(k,k^\prime)$-th order functional derivative as the distribution $u^{(k,k^\prime)}\in \mathcal{D}^\prime(\underbrace{\mathbb{R}^{d+1}\times\dots\times\mathbb{R}^{d+1}}_{k+k^\prime+1};\mathrm{Fun}_\mathbb{C})$ such that
\begin{gather}\label{Eq: Functional derivative}
u^{(k,k^\prime)}(f\otimes\eta_1\otimes\dots\otimes\eta_k\otimes\eta^\prime_1\otimes\dots\otimes\eta^\prime_{k^\prime};\eta,\eta^\prime)\vcentcolon=\notag\\
\vcentcolon=\left.\frac{\partial^{k+k^\prime}}{\partial s_1\dots\partial s_k\partial s^\prime_1\dots\partial s^\prime_{k^\prime}}u(f;s_1\eta_1+\dots+s_k\eta_k+\eta,s^\prime_1\eta^\prime_1\dots+s^\prime_{k^\prime}\eta_{k^\prime}+\eta^\prime))\right|_{s_1=\dots=s_k=s^\prime_1=\dots=s^\prime_{k^\prime}=0}\,.
\end{gather}
We say that a functional-valued distribution is {\bf polynomial} if $\exists\,(\bar{k},\bar{k}^\prime)\in\mathbb{N}_0\times\mathbb{N}_0$ with $\mathbb{N}_0=\mathbb{N}\cup\{0\}$ such that $u^{(k,k^\prime)}=0$ whenever at least one of these conditions holds true: $k\geq\bar{k}$ or $k^\prime\geq\bar{k}^\prime$.
The collection of all these functionals is denoted by $\mathcal{D}^\prime(\mathbb{R}^{d+1};\mathsf{Pol}_{\mathbb{C}})$.
\end{definition}
Observe that the subscript $\mathbb{C}$ in $\mathcal{D}^\prime(\mathbb{R}^{d+1};\mathsf{Fun}_{\mathbb{C}})$ and in $\mathcal{D}^\prime(\mathbb{R}^{d+1};\mathsf{Pol}_{\mathbb{C}})$ is here introduced in contrast to the notation of \cite{Dappiaggi:2020gge} to highlight that, since we consider an SPDE with an additive complex white noise, we need to work with a different notion of functional-valued distribution. More precisely Equation \eqref{Eq: Functional valued distribution} codifies that, in addition to the test-function $f$, we need two independent configurations, $\eta,\eta^\prime$ in order to build a functional-valued distribution. The goal is to encode in this framework the information that we want to consider a priori as independent the random distribution $\psi$ and $\bar{\psi}$ in Equation \eqref{Eq: Stochastic NLS}. Their mutual relation via complex conjugation is codified only at the end of our analysis.
\noindent An immediate structural consequence of Definition \ref{Def: Functionals} can be encoded in the following corollary, whose proof is immediate and, therefore we omit it.
\begin{corollary}\label{Cor: Pointwise Algebra Structure}
The collection of polynomial functional-valued distributions can be endowed with the structure of a commutative $\mathbb{C}$-algebra such that for all $u,v\in\mathcal{D}^\prime(\mathbb{R}^{d+1};\mathsf{Pol}_{\mathbb{C}})$, and for all $f\in\mathcal{D}(\mathbb{R}^{d+1})$ and $\eta,\eta^\prime\in\mathcal{E}(\mathbb{R}^{d+1})$,
\begin{equation}\label{Eq: Pointwise Product}
(uv)(f;\eta,\eta^\prime)=u(f;\eta,\eta^\prime)v(f;\eta,\eta^\prime)\,.
\end{equation}
\end{corollary}
A close inspection of Equation \eqref{Eq: Functional derivative} suggests the possibility of introducing a related notion of {\em directional derivative} of a functional $u\in\mathcal{D}^\prime(\mathbb{R}^{d+1};\mathsf{Fun}_{\mathbb{C}})$ by taking an arbitrary but fixed $\zeta\in\mathcal{E}(\mathbb{R}^{d+1})$ and setting for all $(f,\eta)\in\mathcal{D}(\mathbb{R}^{d+1})$,
\begin{equation}\label{Eq: Directional derivative}
\delta_\zeta u(f;\eta,\eta^\prime)\doteq u^{(1,0)}(f\otimes\zeta;\eta,\eta^\prime)\quad\textrm{and}\quad\overline{\delta}_\zeta u(f;\eta,\eta^\prime)\doteq u^{(0,1)}(f\otimes\zeta;\eta,\eta^\prime)\,.
\end{equation}
In order to make Definition \ref{Def: Functionals} more concrete, we list a few basic examples of polynomial functional-valued distributions, which shall play a key r\^{o}le in our investigation, particularly in the construction of the algebraic structures at the heart of our approach.
\begin{example}\label{Ex: Basic Functionals}
For any $f\in\mathcal{D}(\mathbb{R}^{d+1})$ and $\eta,\eta^\prime\in\mathcal{E}(\mathbb{R}^{d+1})$, we call
\begin{align*}
\mathbf{1}(f;\eta,\eta^\prime)\doteq\int_{\mathbb{R}^{d+1}} dx\,f(x),\quad\Phi(f;\eta,\eta^\prime)=\int_{\mathbb{R}^{d+1}} dx\,f(x)\eta(x)\quad\overline{\Phi}(f;\eta,\eta^\prime)=\int_{\mathbb{R}^{d+1}} dx\,f(x)\eta^\prime(x)\,,
\end{align*}
where $dx$ is the standard Lesbegue measure on $\mathbb{R}^{d+1}$. Notice that $\Phi$ and $\overline{\Phi}$ are two independently defined functionals and a priori they are not related by complex conjugation as the symbols might suggest. This difference originates from our desire to follow an approach similar to the one often used in quantum field theory, see {\it e.g.} \cite{Araki:1971id}, when analyzing the quantization of spinors. In this case it is convenient to consider a field and its complex conjugate as a priori independent building blocks, since this makes easier the implementation of the canonical anticommutation relations. Here we wish to follow a similar rationale and we decided to keep the symbols $\Phi$ and $\overline{\Phi}$ as a memento that, at the end of the analysis, one has to restore their mutual relation codified by complex conjugation.
In addition, we can construct composite functionals, namely for any $p,q\in\mathbb{N}$, $f\in\mathcal{D}(\mathbb{R}^{d+1})$ and $\eta,\eta^\prime\in\mathcal{E}(\mathbb{R}^{d+1})$, we call
\begin{align}\label{Eq: generic functional product}
\overline{\Phi}^p\Phi^q(f,\eta,\eta^\prime)=(\Phi^q\overline{\Phi}^p)(f,\eta,\eta^\prime)=\int_{\mathbb{R}^{d+1}} dx\,f(x)\eta^p(x)(\eta^\prime)^q(x)\,.
\end{align}
Observe that, for convenience, we adopt the notation $|\Phi|^{2k}\equiv\Phi^k\overline{\Phi}^k=\overline{\Phi}^k\Phi^k$.
One can readily infer that, for all $p,q\in\mathbb{N}$, $\mathbf{1},\Phi,\overline{\Phi},\overline{\Phi}^p\Phi^q\in\mathcal{D}^\prime(\mathbb{R}^{d+1};\mathsf{Pol}_{\mathbb{C}})$.
On the one hand all functional derivatives of $\mathbf{1}$ vanish, while, on the other hand
\begin{align*}
\Phi^{(1,0)}(f\otimes\eta_1;\eta,\eta^\prime)=\overline{\Phi}^{(0,1)}(f\otimes\eta_1;\eta,\eta^\prime)=\int_{\mathbb{R}^{d+1}}dx\,f(x)\eta_1(x)\,,
\end{align*}
whereas $\Phi^{(k,k^\prime)}=0$ for all $k^\prime\neq 0$ or if $k\geq 1$.
A similar conclusion can be drawn for $\overline{\Phi}$ and for $\overline{\Phi}^p\Phi^q$.
\end{example}
In the spirit of a perturbative analysis of Equation \eqref{Eq: Stochastic NLS} the next step consists of encoding in the polynomial functionals the information that the linear part of the dynamics is ruled by the Schr\"odinger operator $L=i\partial_t-\Delta$.
To this end, let $u\in\mathcal{D}^\prime(\mathbb{R}^{d+1};\mathsf{Fun}_{\mathbb{C}})$ and let $G$ be the fundamental solution of $L$ whose integral kernel is as per Equation \eqref{Eq: Integral Kernel}.
Then, for all $\eta,\eta^\prime\in\mathcal{E}(\mathbb{R}^{d+1})$ and for all $f\in\mathcal{D}(\mathbb{R}^{d+1})$,
\begin{equation}\label{Eq: Action of G}
(G\circledast u)(f;\eta,\eta^\prime)\doteq u(G\circledast f;\eta,\eta^\prime),
\end{equation}
where, in view of Equation \eqref{Eq: Wavefront Set of G} and of \cite[Th. 8.2.12]{Hormander-I-03} $G\circledast f\in C^\infty(\mathbb{R}^{d+1})$ is such that, for all $h\in\mathcal{D}(\mathbb{R}^{d+1})$, $(G\circledast f)(h)\doteq G(f\otimes h)$.
We can now collect all the ingredients introduced, building a distinguished commutative algebra.
We proceed in steps, adapting to the case in hand the procedure outlined in \cite{Dappiaggi:2020gge}.
As a starting point, we introduce
\begin{equation}\label{Eq: Pointwise algebra - A_0}
\mathcal{A}^{\mathbb{C}}_0\doteq\mathcal{E}[\mathbf{1},\Phi,\overline{\Phi}],
\end{equation}
that is the polynomial ring on $\mathcal{E}(\mathbb{R}^{d+1})$ whose generators are the functionals $\mathbf{1},\Phi,\overline{\Phi}$ defined in Example \ref{Ex: Basic Functionals}.
The algebra product is the pointwise one introduced in Corollary \ref{Cor: Pointwise Algebra Structure}.
Subsequently we encode the action of the fundamental solution of the Schr\"odinger operator as
\begin{subequations}
\begin{equation}\label{Eq: Pointwise algebra - GA_0}
G\circledast\mathcal{A}^{\mathbb{C}}_0\doteq\{u\in\mathcal{D}^\prime(\mathbb{R}^{d+1};\mathsf{Fun}_{\mathbb{C}})\;|\;u=G\circledast v,\;\textrm{with}\; v\in\mathcal{A}^{\mathbb{C}}_0\},
\end{equation}
\begin{equation}\label{Eq: Pointwise algebra - ccGA_0}
\overline{G}\circledast\mathcal{A}^{\mathbb{C}}_0\doteq\{u\in\mathcal{D}^\prime(\mathbb{R}^{d+1};\mathsf{Fun}_{\mathbb{C}})\;|\;u=\overline{G}\circledast v,\;\textrm{with}\; v\in\mathcal{A}^{\mathbb{C}}_0\},
\end{equation}
\end{subequations}
where the action of $G$ is defined in Equation \eqref{Eq: Action of G}.
Here $\overline{G}$ stands for the complex conjugate of the fundamental solution $G$ as in Equation \eqref{Eq: Integral Kernel}.
Its action on a function is defined in complete analogy with Equation \eqref{Eq: Action of G}.
In order to account for the possibility of applying to our functionals more than once the fundamental solution $G$, as well as $\overline{G}$, we proceed inductively defining, for every $j\geq 1$
\begin{equation}\label{Eq: Pointwise algebra - A_j}
\mathcal{A}^{\mathbb{C}}_j=\mathcal{E}[\mathcal{A}^{\mathbb{C}}_{j-1}\cup G\circledast\mathcal{A}^{\mathbb{C}}_{j-1}\cup \overline{G}\circledast\mathcal{A}^{\mathbb{C}}_{j-1}].
\end{equation}
Since $\mathcal{A}^{\mathbb{C}}_j\subset\mathcal{A}^{\mathbb{C}}_{j-1}$ for all $j\geq 1$, the following definition is natural.
\begin{definition}\label{Def: Pointwise Algebra}
Let $\mathcal{A}^{\mathbb{C}}_j$, $j\geq 0$, be the rings as per Equation \eqref{Eq: Pointwise algebra - A_0} and \eqref{Eq: Pointwise algebra - A_j}.
Then we call $\mathcal{A}^{\mathbb{C}}$ the unital, commutative $\mathbb{C}$-algebra obtained as the direct limit
\begin{align*}
\mathcal{A}^{\mathbb{C}}\doteq\varinjlim\mathcal{A}^{\mathbb{C}}_j\,.
\end{align*}
\end{definition}
\begin{remark}\label{moduli2}
For later convenience, it is important to realize that $\mathcal{A}^{\mathbb{C}}$ ({\em resp.} $\mathcal{A}^{\mathbb{C}}_j$, $j\geq 0$) can be regarded as a graded algebra over $\mathcal{E}(\mathbb{R}^{d+1})$, {\it i.e.}
\begin{equation}
\mathcal{A}^{\mathbb{C}}=\bigoplus_{m,l,m',l'\in\mathbb{N}_0}\mathcal{M}_{m,m',l,l'}, \hspace{2cm} \mathcal{A}^{\mathbb{C}}_j=\bigoplus_{m,l,m',l'\in\mathbb{N}_0}\mathcal{M}_{m,m',l,l'}^j\,.
\end{equation}
Here $\mathcal{M}_{m,m',l,l'}$ is the $\mathcal{E}(\mathbb{R}^{d+1})$-module generated by those elements in which the fundamental solutions $G$ and $\overline{G}$ act $l$ and $l'$ times, respectively while the overall polynomial degree in $\Phi$ is $m$ and the one in $\overline{\Phi}$ is $m'$.
The components of the decomposition of $\mathcal{A}^{\mathbb{C}}_{j}$ are defined as $\mathcal{M}_{m,m',l,l'}^j=\mathcal{M}_{m,m',l,l'}\cap\mathcal{A}^{\mathbb{C}}_j$.
In the following a relevant r\^{o}le is played by the polynomial degree of an element lying in $\mathcal{A}^{\mathbb{C}}$, ignoring the occurrence of $G$ and $\overline{G}$.
Therefore we introduce
\begin{equation}
\mathcal{M}_{m,m'}\doteq\bigoplus\limits_{\substack{l,l'\in\mathbb{N}_0 \\ p\leq m\\q\leq m'}}\mathcal{M}_{p,q,l,l'},\hspace{2cm} \mathcal{M}^j_{m,m'}\doteq\bigoplus\limits_{\substack{l,l'\in\mathbb{N}_0 \\ p\leq m\\q\leq m'}}\mathcal{M}_{p,q,l,l'}^j\,.
\end{equation}
Since it holds that $\mathcal{M}_k\subseteq\mathcal{M}_{k+1}$ for all $k\geq0$, the direct limit is well defined and it holds:
\begin{equation}
\mathcal{A}^{\mathbb{C}}=\lim_{m,m'\rightarrow\infty}\mathcal{M}_{m,m'}\,.
\end{equation}
\end{remark}
\noindent To conclude the section, we establish an estimate on the wavefront set of the derivatives of the functionals lying in $\mathcal{A}^{\mathbb{C}}$, see Definition \ref{Def: Pointwise Algebra}.
To this end, we fix the necessary preliminary notation, namely, for any $k\in\mathbb{N}$ we set
\begin{equation}\label{Eq: Delta Diag}
\delta_{Diag_k}\in\mathcal{D}^\prime(\mathbb{R}^{(d+1)k})\quad|\quad \mathcal{D}(\mathbb{R}^{(d+1)k})\ni f(x_1,\cdots,x_k)\mapsto \delta_{Diag_k}(f)\doteq\int_{\mathbb{R}^{d+1}}dx_1 f(x_1,\cdots,x_1)\,,
\end{equation}
where $x_1,\cdots,x_k\in\mathbb{R}^{d+1}$.
\begin{definition}\label{Def: Cm set}
For any but fixed $m\in\mathbb{N}$, let us consider any arbitrary partition of the set $\{1,\ldots,m\}$ into the disjoint union of $p$ non-empty subsets $I_1\uplus\ldots\uplus I_p$, $p$ being arbitrary.
Employing the notations $\widehat{x}_m=(x_1,\ldots,x_m)\in \mathbb{R}^{(d+1)m}$ and $\widehat{t}_{m}=(t_1,\ldots,t_m)\in\mathbb{R}^m$, as well as their counterpart at the level of covectors, we set
\clacomment{
\begin{equation}\label{C_m}
\begin{split}
C_m:=&\{(\widehat{t}_m, \widehat{x}_m,\widehat{\omega}_m,\widehat{k}_m)\in T^\ast\mathbb{R}^{(d+1)m}\setminus\{0\}\,\vert\,\exists\,l\in\{1,\ldots,m-1\},\\
&\{1,\ldots,m\}=I_1\uplus\ldots\uplus I_l\text{ such that }\forall i\neq j,\\
&\forall (a,b)\in I_i\times I_j,\text{ then }t_a\neq t_b\text{ and }\forall j\in\{1,\ldots, l\},\\
&t_n=t_m\,\forall n,m\in I_j\text{ and }\sum_{m\in I_j}\omega_m=0,\sum_{m\in I_j}k_m=0\}\,,
\end{split}
\end{equation}}
where $\vert I_j\vert$ denotes the cardinality of the set $I_j$.
Accordingly we call
\begin{equation}\label{Eq: Condition_C}
\mathcal{D}'_C(\mathbb{R}^{(d+1)};\mathsf{Pol}_{\mathbb{C}})\doteq\{u\in\mathcal{D}'(\mathbb{R}^{(d+1)};\mathsf{Pol}_{\mathbb{C}})\;\vert\;\mathrm{WF}(u^{(k,k^\prime)})\subseteq C_{k+k^\prime+1}\,,\;\forall\, k,k^\prime\geq 0\}\,.
\end{equation}
\end{definition}
\begin{remark}\label{Rem: smooth distributions}
We observe that the elements of the space $\mathcal{D}'_C(\mathbb{R}^{(d+1)};\mathsf{Pol}_{\mathbb{C}})$ are distributions generated by smooth functions, as one can deduce from Equations \eqref{C_m} and \eqref{Eq: Condition_C} setting $k=k^\prime=0$.
\end{remark}
\begin{remark}\label{Rem: stability wrt G}
The space $\mathcal{D}'_C(\mathbb{R}^{(d+1)};\mathsf{Pol}_{\mathbb{C}})$ is stable with respect to the action of the fundamental solution $G$ and of its complex conjugate, more precisely
\begin{align*}
G\circledast\tau,\,\overline{G}\circledast\tau\in\mathcal{D}'_C(\mathbb{R}^{(d+1)};\mathsf{Pol}_{\mathbb{C}})\,,\qquad\forall\,\tau\in\mathcal{D}'_C(\mathbb{R}^{(d+1)};\mathsf{Pol}_{\mathbb{C}})\,.
\end{align*}
The proof of this property follows the same lines of \cite[Lemma 2.14]{Dappiaggi:2020gge} and thus we omit it. \clacomment{We underline that the only difference concerns the different form of $\mathrm{WF}(G)$, which is nonetheless accounted for by the definition of the sets in Equation \eqref{C_m}.
Another important result pertains the estimate of the scaling degree of $G\circledast \tau$. Before dwelling into its calculation, we recall the definition of weighted scaling degree at a point: given $u\in\mathcal{D}'(\mathbb{R}^{d+1})$ and $f\in\mathcal{D}(\mathbb{R}^{d+1})$, consider the scaled function $f_\lambda(t,x)=\lambda^{-(d+2)}f(\lambda^{-2}t,\lambda^{-1}x)$ and $u_\lambda(f)=u(f_\lambda)$.
\begin{align*}
\mathrm{sd}_{t_0,x_0}(u)=\inf_{\omega^\prime}\{\omega^\prime\in\mathbb{R}\,\vert\,\lim_{\lambda\rightarrow0}\lambda^{\omega^\prime}u_\lambda(t_0,x_0)=0\}.
\end{align*}
In this work we are interested in the scaling degree with respect to the hypersurface $\Lambda_t$, see \cite{Brunetti-Fredenhagen-00}. Under the aforementioned parabolic scaling, the fundamental solution $G$ behaves homogeneously and a direct computation shows that $\mathrm{wsd}_{\Lambda_t}(G)=\mathrm{wsd}_{\Lambda_t}(\overline{G})=d$.}
\clacomment{Having in mind the preceding properties and referring to \cite[App. B]{Dappiaggi:2020gge} and \cite[Lemma 2.14 ]{Dappiaggi:2020gge} for the technical details, it holds
\begin{align*}
\mathrm{wsd}_{\Lambda^{1+k_1+k_2}_t}(G\circledast\tau)^{(k_1,k_2)}\,,\mathrm{wsd}_{\Lambda^{1+k_1+k_2}_t}(\overline{G}\circledast\tau)^{(k_1,k_2)}<\infty\,,
\end{align*}
whenever $\mathrm{wsd}_{\Lambda^{1+k_1+k_2}_t}(\tau)^{(k_1,k_2)}<\infty$. Here $\mathrm{wsd}_{\Lambda^{1+k_1+k_2}_t}$ denotes the scaling degree with respect to the subset
\begin{align*}
\Lambda^{1+k_1+k_2}_t=\{(t_1,\ldots,t_{1+k_1+k_2};\widehat{x}_1,\ldots,\widehat{x}_{1+k_1+k_2})\,\vert\,t_1=\ldots=t_{1+k_1+k_2}\}\subset\mathbb{R}^{(d+1)(1+k_1+k_2)}.
\end{align*}}
\end{remark}
\begin{remark}
\clacomment{We underline that the choice of a parabolic scaling, weighting the time variable twice the spatial ones, is natural since this is the scaling transformation under which the linear part of the Schr\"odinger equation as well as its fundamental solution are scaling invariant.}
\end{remark}
\noindent The following proposition is the main result of this section.
\begin{proposition}\label{Prop: WF of the Algebra}
Let $\mathcal{A}^{\mathbb{C}}$ be the algebra introduced in Definition \ref{Def: Pointwise Algebra}.
In view of Definition \ref{Def: Cm set}, it holds that
\begin{align*}
\mathcal{A}^{\mathbb{C}}\subset\mathcal{D}^\prime_C(\mathbb{R}^{d+1}; \mathsf{Pol}_{\mathbb{C}})\,.
\end{align*}
\end{proposition}
\begin{proof}
In view of the characterization of $\mathcal{A}^{\mathbb{C}}$ as an inductive limit, see Definition \ref{Def: Pointwise Algebra}, it suffices to prove the statement for each $\mathcal{A}^{\mathbb{C}}_j$ as in Equation \eqref{Eq: Pointwise algebra - A_0} and \eqref{Eq: Pointwise algebra - A_j}.
To this end we proceed inductively on the index $j$.
\vskip .2cm
\noindent{\em Step 1 -- $j=0$:} In view of Equation \eqref{Eq: Pointwise algebra - A_0}, it suffices to focus the attention on the collection of generators $u_{p,q}\equiv\overline{\Phi}^p\Phi^q$, $p,q\geq 0$, see Example \ref{Ex: Basic Functionals}.
If $p=q=0$, there is nothing to prove since we are considering the identity functional $\mathbf{1}$ whose derivatives are all vanishing.
Instead, in all other cases, we observe that a direct calculation shows that the derivatives yield either smooth functionals or suitable products between smooth functions and Dirac delta distributions.
Per comparison with Definition \ref{Def: Cm set}, we can conclude
$$\mathrm{WF}(u_{p,q}^{(k,k^\prime)})\subseteq C_{k+k^\prime+1}\,.$$
\vskip .2cm
\noindent{\em Step 2 -- $j\geq 1$:} As first step, we observe that if $u\in\mathcal{A}^{\mathbb{C}}_j\cap\mathcal{D}^\prime_C(\mathbb{R}^{d+1}; \mathsf{Pol}_{\mathbb{C}})$, then $G\circledast u,\overline{G}\circledast u\in\mathcal{D}^\prime_C(\mathbb{R}^{d+1};\mathsf{Pol}_{\mathbb{C}})$.
As a matter of fact, Equation \eqref{Eq: Action of G} entails that, for every $k,k^\prime\geq 0$, $(G\circledast u)^{(k,k^\prime)}=u^{(k,k^\prime)}\cdot (G\otimes 1_{k+k^\prime})$, the dot standing for the pointwise product of distributions, while $1$ stands here for the identity operator.
Since $\mathrm{WF}(u^{(k,k^\prime)})\subseteq C_{k+k^\prime+1}$ per hypothesis, it suffices to apply \cite[Lemma 2.14]{Dappiaggi:2020gge}, \clacomment{see Remark \ref{Rem: stability wrt G}}, in combination with Equation \eqref{Eq: Wavefront Set of G} to conclude that $\mathrm{WF}((G\circledast u)^{(k,k^\prime)})\subseteq C_{k+k^\prime+1}$.
The same line of reasoning entails that an identical conclusion can be drawn for $\overline{G}$.
We can now focus on the inductive step.
Therefore, let us assume that the statement of the proposition holds true for $\mathcal{A}^{\mathbb{C}}_j$. In view of Equation \eqref{Eq: Pointwise algebra - A_j}, $\mathcal{A}^{\mathbb{C}}_{j+1}$ is generated by $\mathcal{A}^{\mathbb{C}}_j$, $G\circledast\mathcal{A}^{\mathbb{C}}_j$ and $\overline{G}\circledast\mathcal{A}^{\mathbb{C}}_j$.
On account of the inductive step, it suffices to focus on the pointwise product of two functionals, say $u$ and $v$ lying in one of the generating algebras. \\
For definiteness we focus on $u\in\mathcal{A}^{\mathbb{C}}_j$ and $v\in\mathcal{A}^{\mathbb{C}}_j$, the other cases following suit. Using the Leibniz rule, for every $k,k^\prime\geq 0$, it turns out that $(uv)^{(k,k^\prime)}$ is a linear combination with smooth coefficient of products of distributions of the form $u^{(p,p^\prime)}\otimes v^{(q,q^\prime)}$ with $0\leq p\leq k$, $0\leq q\leq k^\prime$ and $p+q=k$ while $p^\prime+q^\prime=k^\prime$.
The inductive hypothesis entails
\begin{align*}
\mathrm{WF}(u^{(p,p^\prime)})\subseteq C_{p+p^\prime+1},\quad\textrm{and}\quad\mathrm{WF}(v^{(q,q^\prime)})\subseteq C_{q+q^\prime+1}\,.
\end{align*}
\clacomment{
A direct inspection of Equation \eqref{C_m} entails that if $(t,x,\omega_1,\kappa_1,s_{p+p'},y_{p+p^\prime},\omega_{s}; \kappa_y)\in C_{p+p^\prime+1}$ while $(t,x,\omega_2,\kappa_2,r_{q+q'},z_{q+q^\prime},\omega_{r}; \kappa_z)\in C_{q+q^\prime+1}$ then
\begin{align*}
(t,x,\omega_1+\omega_2,\kappa_1+\kappa_2,s_{p+p'},y_{p+p^\prime},\omega_s,\kappa_y,, r_{q+q'},y_{q+q^\prime},\omega_r,\kappa_z)\in C_{k+k^\prime+1}\,,
\end{align*}
from which it descends that $\mathrm{WF}(u^{(p,p^\prime)}\otimes v^{(q,q^\prime)})\subseteq C_{k+k^\prime+1}$.
}
\end{proof}
\section{The Algebra $\mathcal{A}^{\mathbb{C}}_{\cdot_Q}$}\label{Sec: expectations}
The algebra $\mathcal{A}^{\mathbb{C}}$ introduced in Definition \ref{Def: Pointwise Algebra} does not codify the information associated to the stochastic nature of the underlying white noise.
In order to encode such datum in the functional-valued distributions and in the same spirit of \cite{Dappiaggi:2020gge}, we
introduce a new algebra which is obtained as a \emph{deformation} of the pointwise product of $\mathcal{A}^{\mathbb{C}}$. As it will be manifest in the following discussion, this construction is a priori purely formal unless a suitable renormalization procedure is implemented
\begin{remark}\label{Rem: Q}
As a premise, we introduce the following bi-distribution, constructed out of the fundamental solution $G\in\mathcal{D}^\prime(\mathbb{R}^{d+1}\times\mathbb{R}^{d+1})$ as per Equation \eqref{Eq: Integral Kernel}:
\begin{equation}\label{Eq: definition of Q}
Q\vcentcolon=G\circ G\in\mathcal{D}^\prime(\mathbb{R}^{d+1}\times\mathbb{R}^{d+1})\,,
\end{equation}
where $\circ$ denotes the composition of bi-distributions, namely, for any $f_1,f_2\in\mathcal{D}(\mathbb{R}^{d+1}\times\mathbb{R}^{d+1})$,
\begin{align*}
Q(f_1,f_2)\vcentcolon=(G\circ G)(f_1\otimes f_2)\vcentcolon=[(G\otimes G)\cdot(1_{d+1}\otimes\delta_{\mathrm{Diag}_2}\otimes1_{d+1})](f_1\otimes1_{d+1}\otimes1_{d+1}\otimes f_2)\,.
\end{align*}
Here, with a slight abuse of notation, $1_{d+1}\in\mathcal{D}^\prime(\mathbb{R}^{d+1})$ stands for the distribution generated by the constant smooth function $1$ on $\mathbb{R}^{d+1}$ while $\cdot$ indicates the pointwise product of distributions, see \cite[Thm. 8.2.10]{Hormander-I-03}. Further properties of the composition $\circ$ are discussed in \cite[App. A]{Dappiaggi:2020gge}.
From the perspective of the stochastic process at the heart of Equation \eqref{Eq: Stochastic NLS}, we observe that the bi-distribution $Q$ codifies the covariance of the complex Gaussian random field $\widehat{\varphi}=G\ast\xi$.
More explicitly, it holds
\begin{equation}\label{Eq: covariance complex Gaussian random field}
\mathbb{E}[\widehat{\varphi}(f_1)\overline{\widehat{\varphi}}(f_2)]=Q(f_1,f_2), \quad\forall f_1,f_2\in\mathcal{D}(\mathbb{R}^{d+1}\times\mathbb{R}^{d+1}).
\end{equation}
Observe, in addition that $\widehat{\varphi}$ can be read also as a solution of the stochastic linear Schr\"odinger equation, namely Equation \eqref{Eq: Stochastic NLS} setting $\lambda=0$.
To conclude our excursus on the bi-distribution $Q$, we highlight two notable properties of its singular structure:
\begin{enumerate}
\item in view of \cite[Thm. 8.2.14]{Hormander-I-03},
\begin{align*}
\mathrm{WF}(Q)\subseteq\mathrm{WF}(G)\,.
\end{align*}
\item as a consequence of \cite[Lemma B.12]{Dappiaggi:2020gge},
$$\mathrm{wsd}_{\clacomment{\Lambda^2_t}}(Q)<\infty,$$
where $\mathrm{wsd}_{\clacomment{\Lambda^2_t}}(Q)$ denotes the scaling degree of the distribution $Q\in\mathcal{D}^\prime(\mathbb{R}^{d+1}\times\mathbb{R}^{d+1})$ with respect to $\Lambda_t^2:=\{(t_1,x_1,t_2,x_2)\in\mathbb{R}^{d+1}\times\mathbb{R}^{d+1}\,\vert\,t_1=t_2\}$.
\end{enumerate}
\end{remark}
We are now in a position to introduce the sought deformation which encodes the stochastic properties due to the complex white noise present in Equation \eqref{Eq: Stochastic NLS}. Inspired by \cite{Dappiaggi:2020gge}, as a tentative starting point we set, for any $\tau_1,\tau_2\in\mathcal{A}^\mathbb{C}$, $f\in\mathcal{D}(\mathbb{R}^{d+1})$ and $\eta,\eta^\prime\in\mathcal{E}(\mathbb{R}^{d+1})$,
\begin{equation}\label{Eq: prodotto deformato2}
\begin{split}
[\tau_1\cdot_Q \tau_2](f;\eta, \eta^\prime):=\sum_{k\geq0}\sum\limits_{\substack{k_1,k_2 \\ k_1+k_2=k}}\frac{1}{k_1!k_2!}\big(\delta_{Diag_2}\otimes Q^{\otimes k_1}\otimes\overline{Q}^{\otimes k_2}\big)
\cdot\big[\tau_1^{(k_1,k_2)}\Tilde{\otimes}\tau_2^{(k_1,k_
2)}\big](f\otimes 1_{1+2k};\eta, \eta^\prime)\,.
\end{split}
\end{equation}
Here with $\tilde{\otimes}$ we denote the tensor product though modified in terms of a permutation of its arguments, which, at the level of integral kernel, reads
\begin{gather}\label{Eq: Tilde otimes}
(\delta_{Diag_2}\otimes Q^{\otimes k_1}\otimes\overline{Q}^{\otimes k_2}\big)
\cdot\big[\tau_1^{(k_1,k_2)}\Tilde{\otimes}\tau_2^{(k_1,k_
2)}\big]\notag\\
=\delta(x_1,x_2)\prod_{j=1}^{k_1}\prod_{\ell=1}^{k_2}Q(z_j,z^\prime_{j})\overline{Q}(y_{\ell},y^\prime_\ell)\tau_1^{(k_1,k_2)}(x_1,z_1,\ldots,z_{k_1},y_{1},\ldots,y_{k_2})\tau_2^{(k_1,k_2)}(x_2,z^\prime_{1},\ldots,z^\prime_{k_1},y^\prime_{1},\ldots,y^\prime_{k_2})\,.
\end{gather}
\begin{remark}\label{Rem: deformed product}
We observe that only a finite number of terms contributes to the sum on the right hand side of Equation \eqref{Eq: prodotto deformato2} on account of the polynomial nature of the functional-valued distributions $\tau_1$ and $\tau_2$.
Nonetheless, at this stage, the right hand side of Equation \eqref{Eq: prodotto deformato2} is only a formal expression, since it can include a priori ill-defined structures such as the coinciding point limit of $Q$ and $\overline{Q}$. In the following theorem we shall bypass this hurdle by means of a renormalization procedure so to give meaning to Equation \eqref{Eq: prodotto deformato2} for any $\tau_1,\tau_2\in\mathcal{A}^\mathbb{C}$.
\end{remark}
The main motivation for the introduction of a deformation of the algebraic structure is to build an explicit algorithm for computing expectation values and correlation functions of polynomial expressions in the random fields $\widehat{\varphi}=G\ast\xi$ and $\overline{\widehat{\varphi}}=\overline{G}\ast\overline{\xi}$. For this reason. let us illustrate the stochastic interpretation of the deformed product $\cdot_Q$ and its link to the expectation values via the following example.
\begin{example}\label{Ex: Example interpretation dot}
Formally, at the level of integral kernels and referring to the defining properties of the complex white noise in Equation \eqref{Eq: complex white noise}, it holds that
\begin{equation*}
\begin{split}
\mathbb{E}[\widehat{\varphi}^2(f)]&=\mathbb{E}\Big[\int_{(\mathbb{R}^{d+1})^2} dx'dy'\,G(x,x')G(x,y')\xi(x')\xi(y')f(x)\Big]\\
&=\int_{(\mathbb{R}^{d+1})^2}dx'dy'\,G(x,x')G(x,y')\underbrace{\mathbb{E}\bigl[\xi(x')\xi(y')\bigr]}_{=0}f(x)=0,
\end{split}
\end{equation*}
for every $f\in\mathcal{D}(\mathbb{R}^{d+1})$. Now let us consider the expectation value of $\vert\widehat{\varphi}\vert^2=\widehat{\varphi}\overline{\widehat{\varphi}}$:
\begin{equation*}
\begin{split}
\mathbb{E}[\widehat{\varphi}\overline{\widehat{\varphi}}(f)]&=\mathbb{E}\Big[\int_{(\mathbb{R}^{d+1})^2} dx'dy'\,G(x,x')\overline{G}(x,y')\xi(x')\overline{\xi}(y')f(x)\Big]\\
&=\int_{(\mathbb{R}^{d+1})^2}dx'dy'\,G(x,x')\overline{G}(x,y')\underbrace{\mathbb{E}\bigl[\xi(x')\xi(y')\bigr]}_{=\delta(x'-y')}f(x)\\
&=\int_{\mathbb{R}^{d+1}}dx'\,G(x,x')\overline{G}(x,x')f(x)=(G\circ G^\ast)(f\delta_{\text{Diag}_2})=Q(f\delta_{\text{Diag}_2}),
\end{split}
\end{equation*}
where we used Equation \eqref{Eq: complex white noise - 2} together with the relation between the fundamental solution of $L$ and its formal adjoint $G^\ast$, $G^\ast(x,x')=\overline{G(x',x)}$. On account of Equations \eqref{Eq: prodotto deformato2} and \eqref{Eq: Gammadot on products} we can compute
\begin{equation*}
(\Phi\cdot_Q\Phi)(f;\eta,\overline{\eta})=\Phi^2(f;\eta,\overline{\eta}),
\end{equation*}
\begin{equation*}
(\Phi\cdot_Q\overline{\Phi})(f;\eta,\overline{\eta})=\Phi\overline{\Phi}(f;\eta,\overline{\eta})+Q(f\delta_{{Diag}_2}).
\end{equation*}
Observe that, fixing $\eta'=\overline{\eta}$, the two configurations are no longer independent, allowing us to make contact with the stochastic nature of the equation. Evaluating these expressions for $\eta=0$, we obtain
\begin{equation*}
(\Phi\cdot_Q\Phi)(f;0,0)=0\equiv\mathbb{E}[\widehat{\varphi}^2(f)],\hspace{1cm} (\Phi\cdot_Q\overline{\Phi})(f;0,0)=Q(f\delta_{{Diag}_2})\equiv\mathbb{E}[\widehat{\varphi}\overline{\widehat{\varphi}}(f)].
\end{equation*}
As stated in Remark \ref{Rem: deformed product}, the expressions above are a priori ill-defined, accounting for the singular contribution $Q\delta_{Diag_2}$. This problem is tackled in Theorem \ref{Thm: deformation map cdot}.
This example can be readily extended to arbitrary polynomial expressions of $\varphi$ and $\overline{\varphi}$, highlighting how our guess for the deformed product $\cdot_Q$ codifies properly the expectation values.
\end{example}
\begin{theorem}\label{Thm: deformation map cdot}
Let $\mathcal{A}^\mathbb{C}$ be the algebra introduced in Definition \ref{Def: Pointwise Algebra} and let $\mathcal{M}_{m,m^\prime}$ be the moduli as per Remark \ref{moduli2}.
There exists a linear map $\Gamma_{\cdot_Q}^\mathbb{C}:\mathcal{A}^\mathbb{C}\to\mathcal{D}^\prime_C(\mathbb{R}^{d+1};\mathrm{Pol}_\mathbb{C})$ such that:
\begin{enumerate}
\item for any $\tau\in\mathcal{M}_{1,0},\,\mathcal{M}_{0,1}$, it holds
\begin{align}\label{Eq: Gammadot on M_1}
\Gamma_{\cdot_Q}^\mathbb{C}(\tau)=\tau\,;
\end{align}
\item for any $\tau\in\mathcal{A}^\mathbb{C}$, it holds
\begin{align}\label{Eq: Gammadot Gtau}
\nonumber
\Gamma_{\cdot_Q}^\mathbb{C}(G\circledast\tau)=G\circledast\Gamma_{\cdot_Q}^\mathbb{C}(\tau)\,,\\
\Gamma_{\cdot_Q}^\mathbb{C}(\overline{G}\circledast\tau)=\overline{G}\circledast\Gamma_{\cdot_Q}^\mathbb{C}(\tau)\,
\end{align}
\item for any $\tau\in\mathcal{A}^\mathbb{C}$ and $\psi\in\mathcal{E}(\mathbb{R}^{d+1})$, it holds
\begin{align}\label{Eq: Gammadot and derivatives}
\nonumber
\Gamma_{\cdot_Q}^\mathbb{C}\cdot\delta_\psi&=\delta_{\psi}\circ\Gamma_{\cdot_Q}^\mathbb{C}\,,\qquad
&&\Gamma_{\cdot_Q}^\mathbb{C}\cdot\delta_{\overline{\psi}}=\delta_{\overline{\psi}}\circ\Gamma_{\cdot_Q}^\mathbb{C}\,,\\
\Gamma_{\cdot_Q}^\mathbb{C}(\psi\tau)&=\psi\Gamma_{\cdot_Q}^\mathbb{C}(\tau)\,,\qquad
&&\Gamma_{\cdot_Q}^\mathbb{C}(\overline{\psi}\tau)=\overline{\psi}\Gamma_{\cdot_Q}^\mathbb{C}(\tau)\,,
\end{align}
where $\delta_\psi$ denotes the directional functional derivative along $\psi$ as per Equation \eqref{Eq: Directional derivative};
\item denoting by $\sigma_{(p,q)}(\tau)=\mathrm{wsd}_{\mathrm{Diag}_{p+q+1}}(\tau^{(p,q)})$ the weighted scaling degree of $\tau^{(p,q)}$ with respect to the total diagonal of $(\mathbb{R}^{d+1})^{p+q+1}$, see \cite[Rmk. B.9]{Dappiaggi:2020gge}, it holds
\begin{align}\label{Eq: Gammadot and sd}
\sigma_{(p,q)}(\Gamma_{\cdot_Q}^\mathbb{C}(\tau))<\infty\,,\qquad\forall\,\tau\in\mathcal{A}^\mathbb{C}\,.
\end{align}
\end{enumerate}
\end{theorem}
\begin{proof}
\noindent\textbf{Strategy:} The construction of a map $\Gamma_{\cdot_Q}^\mathbb{C}$ satisfying the conditions in the statement of the theorem goes by induction with respect to the indices $m$ and $m^\prime$ in the decomposition
$\mathcal{A}^\mathbb{C}=\bigoplus_{m,m^\prime\in\mathbb{N}}\mathcal{M}_{m,m^\prime}$, discussed in Remark \ref{moduli2}. Since the proof shares many similarities with the counterpart in \cite{Dappiaggi:2020gge} for the case of a stochastic, scalar, partial differential equation, we shall focus mainly on the different aspects. We start from Equation \eqref{Eq: Gammadot on M_1} and, for $\tau_1,\ldots,\tau_n\in\mathcal{A}^\mathbb{C}$, we set
\begin{align}\label{Eq: Gammadot on products}
\Gamma_{\cdot_Q}^\mathbb{C}(\tau_1\ldots\tau_n)\vcentcolon=\Gamma_{\cdot_Q}^\mathbb{C}(\tau_1)\cdot_Q\ldots\cdot_Q\Gamma_{\cdot_Q}^\mathbb{C}(\tau_n)\,,
\end{align}
where the product $\cdot_Q$ is given by Equation \eqref{Eq: prodotto deformato2}.
As we anticipated above, the right hand side of Equation \eqref{Eq: Gammadot on products} might be only formal since the product $\cdot_Q$ is not \textit{a priori} well defined on the whole $\mathcal{D}^\prime_C(\mathbb{R}^{d+1};\mathrm{Pol}_\mathbb{C})$.\\
\noindent\textbf{The Case $\mathbf{d=1}$:} In this scenario the divergence of $Q\delta_{\text{Diag}_2}$ needs no taming. Working at the level of integral kernels, see Equation \eqref{Eq: Integral Kernel}, the composition $Q=G\circ G$ yields a contribution proportional to $\int_{\mathbb{R}}dt/t\sim \log(t)$, which is locally integrable.
As a consequence, in such scenario the proof does not need renormalization and it is straightforward starting from the condition in the statement of the theorem as well as from Equation \eqref{Eq: Gammadot on products}.\\
\noindent\textbf{The Case $\mathbf{d\geq2}$: Step 1.} We start by showing that, whenever $\Gamma_{\cdot_Q}^\mathbb{C}$ is well defined for $\tau\in\mathcal{A}^\mathbb{C}$ in such a way in particular that Equations \eqref{Eq: Gammadot and derivatives} and \eqref{Eq: Gammadot and sd} hold true, then the same applies for $G\circledast\tau$ and $\overline{G}\circledast\tau$. We shall only discuss the case of $G\circledast\tau$, the other following suit.
First of all we notice that $\Gamma_{\cdot_Q}^\mathbb{C}(G\circledast\tau)$ is completely defined by Equation \eqref{Eq: Gammadot Gtau}.
In addition, Equation \eqref{Eq: Gammadot and sd} for $\Gamma_{\cdot_Q}^\mathbb{C}(G\circledast\tau)$ is a direct consequence of Remark \ref{Rem: stability wrt G}.
Finally, Equation \eqref{Eq: Gammadot and derivatives} descends from an iteration of the following argument, namely, for any $\eta,\eta^\prime,\zeta\in\mathcal{E}(\mathbb{R}^{d+1})$ and $f\in\mathcal{D}(\mathbb{R}^{d+1})$,
\begin{align*}
\delta_\zeta\circ\Gamma_{\cdot_Q}^\mathbb{C}(G\circledast\tau)(f;\eta,\eta^\prime)&=
\delta_\zeta\circ\Gamma_{\cdot_Q}^\mathbb{C}(\tau)(G\circledast f;\eta,\eta^\prime)=
\Gamma_{\cdot_Q}^\mathbb{C}(\delta_\zeta\tau)(G\circledast f;\eta,\eta^\prime)\\
&=\Gamma_{\cdot_Q}^\mathbb{C}(G\circledast\delta_\zeta\tau)(f;\eta,\eta^\prime)=
\Gamma_{\cdot_Q}^\mathbb{C}(\delta_\zeta G\circledast\tau)(f;\eta,\eta^\prime)\\
&=\Gamma_{\cdot_Q}^\mathbb{C}\circ\delta_\zeta(G\circledast\tau)(f;\eta,\eta^\prime)\,.
\end{align*}
With these data in our hand, we can start the inductive procedure. \\
\noindent\textbf{Step 2: $\mathbf{(m,m^\prime)=(1,1)}$.} As a starting point, we observe that Equation \eqref{Eq: Gammadot on M_1} completely determines the map $\Gamma_{\cdot_Q}^\mathbb{C}$ restricted to the moduli $\mathcal{M}_{(1,0)}$ and $\mathcal{M}_{(0,1)}$. In addition all other required properties hold true automatically. Next, we focus our attention on $\mathcal{M}_{(1,1)}$. In order to consistently construct $\Gamma_{\cdot_Q}^\mathbb{C}$ on $\mathcal{M}_{(1,1)}$ we rely on an inductive argument with respect to the index $j$ subordinated to the decomposition -- see Remark \ref{moduli2}:
\begin{align*}
\mathcal{M}_{(1,1)}=\bigoplus_{j\in\mathbb{N}}\mathcal{M}_{(1,1)}^j\,,\qquad\mathcal{M}_{(1,1)}^j\vcentcolon=\mathcal{M}_{(1,1)}\cap\mathcal{A}_j^\mathbb{C}\,.
\end{align*}
\noindent Setting $j=0$, we observe that
\begin{align*}
\mathcal{M}_{(1,1)}^0=\mathrm{span}_{\mathcal{E}(\mathbb{R}^{d+1};\mathbb{C})}\{\mathbf{1},\Phi,\overline{\Phi},\Phi\overline{\Phi}\}\,.
\end{align*}
On account of the previous discussion, we are left with the task of defining the action of $\Gamma_{\cdot_Q}^\mathbb{C}$ on $\Phi\overline{\Phi}$. This suffices since $\Gamma_{\cdot_Q}^\mathbb{C}$ can be consequently extended to the whole $\mathcal{M}_{(1,1)}^0$ by linearity.
Recalling that $\Phi\overline{\Phi}$ is defined as in Equation \eqref{Eq: generic functional product}, we set, formally, for any $\eta,\eta^\prime\in\mathcal{E}(\mathbb{R}^{d+1})$ and $f\in\mathcal{D}(\mathbb{R}^{d+1})$,
\begin{align*}
\Gamma_{\cdot_Q}^\mathbb{C}(\Phi\overline{\Phi})(f;\eta,\eta^\prime)\vcentcolon=[\Gamma_{\cdot_Q}^\mathbb{C}(\Phi)\cdot_Q\Gamma_{\cdot_Q}^\mathbb{C}(\overline{\Phi})](f;\eta,\eta^\prime)
=\Phi\overline{\Phi}(f;\eta,\eta^\prime)+(G\cdot\overline{G})(f\otimes1_{d+1})\,,
\end{align*}
where in the last equality we exploited Equation \eqref{Eq: prodotto deformato2} and where $G\cdot\overline{G}$ denotes the pointwise product between $G$ and $\overline{G}$.
We observe that the above formula is purely formal due to the presence of the product $G\cdot\overline{G}$ which is \textit{a priori} ill-defined. As a matter of fact, due to the microlocal behaviour of $G$ codified in Equation \eqref{Eq: Wavefront Set of G} and using H\"ormander criterion for the multiplication of distributions \cite[Th. 8.2.10]{Hormander-I-03}, we can only conclude that
\clacomment{$G\cdot\overline{G}\in\mathcal{D}^\prime(\mathbb{R}^{d+1}\times\mathbb{R}^{d+1}\setminus\Lambda^2_t)$.}
Since $\mathrm{wsd}_{\Lambda^2_t}(G\cdot\overline{G})=2d$, \cite[Thm. B.8]{Dappiaggi:2020gge} and \cite[Rmk. B.9]{Dappiaggi:2020gge} entail the existence of an extension $\widehat{G\cdot\overline{G}}\in\mathcal{D}^\prime(\mathbb{R}^{d+1}\times\mathbb{R}^{d+1})$ of $G\cdot\overline{G}$ which preserves both the scaling degree and the wave-front set. Having chosen once and for all any such extension, we set for all $\eta,\eta^\prime\in\mathcal{E}(\mathbb{R}^{d+1})$ and $f\in\mathcal{D}(\mathbb{R}^{d+1})$,
\begin{align}\label{Eq: renormalized squared modulus}
\Gamma_{\cdot_Q}^\mathbb{C}(\Psi\overline{\Psi})(f;\eta,\eta^\prime)\vcentcolon=\Psi\overline{\Psi}(f;\eta,\eta^\prime)+(\widehat{G\cdot\overline{G}})(f\otimes 1)\,.
\end{align}
To conclude, we observe that Equation \eqref{Eq: renormalized squared modulus} implies Equations \eqref{Eq: Gammadot and derivatives} and \eqref{Eq: Gammadot and sd} by direct inspection, as well as that $\Gamma_{\cdot_Q}^\mathbb{C}(\Psi\overline{\Psi})\in\mathcal{D}^\prime_C(\mathbb{R}^{d+1};\mathrm{Pol}_\mathbb{C})$ on account of Equation \eqref{Eq: Wavefront Set of G}.
Having consistently defined $\Gamma_{\cdot_Q}^\mathbb{C}$ on $\mathcal{M}_{(1,1)}^0$, we assume that the map $\Gamma_{\cdot_Q}^\mathbb{C}$ has been coherently assigned on $\mathcal{M}_{(1,1)}^j$ and we prove the inductive step by constructing it on $\mathcal{M}_{(1,1)}^{j+1}$.
We remark that, given $\tau\in\mathcal{M}_{(1,1)}^{j+1}$, either $\tau=G\circledast\tau^\prime$ with $\tau^\prime\in\mathcal{M}_{(1,1)}^j$ (the case with the complex conjugate $\overline{G}$ is analogous), or $\tau=\tau_1\tau_2$ with $\tau_1\in\mathcal{M}_{(1,0)}^j\cup G\circledast\mathcal{M}_{(1,0)}^j\cup\overline{G}\circledast\mathcal{M}_{(1,0)}^j$ while $\tau_2\in\mathcal{M}_{(0,1)}^j\cup G\circledast\mathcal{M}_{(0,1)}^j\cup\overline{G}\circledast\mathcal{M}_{(0,1)}^j$.
In the first case, $\Gamma_{\cdot_Q}^\mathbb{C}(\tau)$ is defined as per Step \textit{1}. of this proof, using in addition the inductive hypothesis. In the second one, we start from the formal expression which descends from Equation \eqref{Eq: Gammadot on products}. It yields for all $\eta,\eta^\prime\in\mathcal{E}(\mathbb{R}^{d+1})$ and $f\in\mathcal{D}(\mathbb{R}^{d+1})$,
\begin{align*}
\Gamma_{\cdot_Q}^\mathbb{C}(\tau)(f;\eta,\eta^\prime)=
[\Gamma_{\cdot_Q}^\mathbb{C}(\tau_1)\cdot_Q\Gamma_{\cdot_Q}^\mathbb{C}(\tau_2)](f;\eta,\eta^\prime)=
[\Gamma_{\cdot_Q}^\mathbb{C}(\tau_1)\Gamma_{\cdot_Q}^\mathbb{C}(\tau_2)](f;\eta,\eta^\prime)+U(f\otimes1_3)\,,
\end{align*}
where
\begin{align}\label{Eq: ill-defined U}
U(f\otimes1_3)\vcentcolon=[(\delta_{\mathrm{Diag}_2}\otimes Q)\cdot(\Gamma_{\cdot_Q}^\mathbb{C}(\tau_1)^{(1,0)}\Tilde{\otimes}\Gamma_{\cdot_Q}^\mathbb{C}(\tau_2)^{(0,1)})](f\otimes1_3)\,,
\end{align}
and where $\Gamma_{\cdot_Q}^\mathbb{C}(\tau_1)\Gamma_{\cdot_Q}^\mathbb{C}(\tau_2)$ denotes the pointwise product between the functional-valued distributions, which is well-defined on account of Remark \ref{Rem: smooth distributions}.
Yet, Equation \eqref{Eq: ill-defined U} involves the product of singular distributions and it calls for a renormalization procedure.
First of all we observe that, due to the microlocal behaviour of $Q$ -- \textit{cf.} Remark \ref{Rem: Q} -- and due to the inductive hypothesis entailing
\begin{align*}
\mathrm{WF}(\Gamma_{\cdot_Q}^\mathbb{C}(\tau_1)^{(1,0)}),\mathrm{WF}(\Gamma_{\cdot_Q}^\mathbb{C}(\tau_2)^{(0,1)})\subset C_2\,,
\end{align*}
it follows that
\clacomment{
\begin{align*}
U\in\mathcal{D}'((\mathbb{R}^{d+1})^4\setminus\Lambda_t^{4,\mathrm{Big}}),
\end{align*}
where $\Lambda_t^{4,\mathrm{Big}}:=\{(\widehat{t},\widehat{x})\in(\mathbb{R}^{d+1})^4\,\vert\,\exists a,b\in\{1,2,3,4\}, a\neq b, t_a=t_b\}$.\\
This estimate can be improved observing that, whenever $(\widehat{t},\widehat{x})\in\Lambda_t^{4,\mathrm{Big}}\setminus\Lambda_t^4$, one of the factors among $(\delta_{\mathrm{Diag}_2}\otimes Q)$, $\Gamma_{\cdot_Q}^\mathbb{C}(\tau_1)^{(1,0)}$ and $\Gamma_{\cdot_Q}^\mathbb{C}(\tau_2)^{(0,1)}$ is smooth and the product of the remaining two is well-defined.
}
\clacomment{As a consequence, $U\in\mathcal{D}^\prime((\mathbb{R}^{d+1})^4\setminus\Lambda_t^4)$. Furthermore \cite[Rmk. B.7]{Dappiaggi:2020gge} here applied to the weighted scaling degree guarantees that
\begin{align*}
\mathrm{wsd}_{\Lambda_t^4}(U)
\leq&\mathrm{wsd}_{\Lambda_t^4}(\delta_{\mathrm{Diag}_2}\otimes Q)+\mathrm{wsd}_{\Lambda_t^4}(\Gamma_{\cdot_Q}^\mathbb{C}(\tau_1)^{(1,0)}\Tilde{\otimes}\Gamma_{\cdot_Q}^\mathbb{C}(\tau_2)^{(0,1)})\\
\leq&\mathrm{wsd}_{\Lambda_t^2}(\delta_{\mathrm{Diag}_2})+\mathrm{wsd}_{\Lambda_t^2}(Q)+\mathrm{wsd}_{\Lambda_t^2}(\Gamma_{\cdot_Q}^\mathbb{C}(\tau_1)^{(1,0)})+\mathrm{wsd}_{\Lambda_t^2}(\Gamma_{\cdot_Q}^\mathbb{C}(\tau_2)^{(0,1)})<\infty\,.
\end{align*}}
Once more, \cite[Thm. B.8]{Dappiaggi:2020gge} and \cite[Rem. B.9]{Dappiaggi:2020gge} entail the existence of an extension $\widehat{U}\in\mathcal{D}^\prime((\mathbb{R}^{d+1})^4)$ of $U\in\mathcal{D}^\prime((\mathbb{R}^{d+1})^4\setminus\Lambda_t^4)$ which preserves both the scaling degree and the wave-front set. Eventually, we set
\begin{align}\label{Eq: renormalized U}
\Gamma_{\cdot_Q}^\mathbb{C}(\tau)(f;\eta,\eta^\prime)=
[\Gamma_{\cdot_Q}^\mathbb{C}(\tau_1)\Gamma_{\cdot_Q}^\mathbb{C}(\tau_2)](f;\eta,\eta^\prime)+\widehat{U}(f\otimes1_3)\,.
\end{align}
By direct inspection, one can realize that Equation \eqref{Eq: renormalized U} satisfies the conditions codified in Equations \eqref{Eq: Gammadot and derivatives} and \eqref{Eq: Gammadot and sd}. This concludes the proof for the case $(m,m^\prime)=(1,1)$.\\
\noindent\textbf{Step 3: Generic $\mathbf{(m,m^\prime)}$.}
We can now prove the inductive step with respect to the indices $(m,m^\prime)$. In particular, we assume that $\Gamma_{\cdot_Q}^\mathbb{C}$ has been consistently assigned on $\mathcal{M}_{(m,m^\prime)}$ and we prove that the same holds true for $\mathcal{M}_{(m+1,m^\prime)}$. We stress that one should prove the same statement also for $\mathcal{M}_{(m,m^\prime+1)}$, but since the analysis is mutatis mutandis the same as for $\mathcal{M}_{(m+1,m^\prime)}$, we omit it.
Once more we employ an inductive procedure exploiting that $\mathcal{M}_{(m+1,m^\prime)}=\bigoplus_{j\in\mathbb{N}}\mathcal{M}_{(m+1,m^\prime)}^j$, with $\mathcal{M}_{(m+1,m^\prime)}^j:=\mathcal{M}_{(m+1,m^\prime)}\cap\mathcal{A}_j^\mathbb{C}$. First of all, we focus on the case $j=0$, observing that
\begin{align*}
\mathcal{M}_{(m+1,m^\prime)}^0=\mathrm{span}_{\mathcal{E}(\mathbb{R}^{d+1};\mathbb{C})}\{\mathbf{1},\Phi,\overline{\Phi},\ldots,\Phi^{m+1}\overline{\Phi}^{m^\prime}\}\,.
\end{align*}
On account of the inductive hypothesis, the map $\Gamma_{\cdot_Q}^\mathbb{C}$ has been already assigned on all the generators of $\mathcal{M}_{(m+1,m^\prime)}^0$ but $\Phi^{m+1}\overline{\Phi}^{m^\prime}$.
Hence, to determine $\Gamma_{\cdot_Q}^\mathbb{C}$ on the whole $\mathcal{M}_{(m+1,m^\prime)}^0$, it suffices to establish its action on $\Phi^{m+1}\overline{\Phi}^{m^\prime}$ extending it by linearity.
To this end, we exploit again Equation \eqref{Eq: Gammadot on products} which, on account of the explicit form of Equation \eqref{Eq: prodotto deformato2}, yields
\begin{align*}
[\Gamma_{\cdot_Q}^\mathbb{C}(\Phi^{m+1}\overline{\Phi}^{m^\prime})](f;\eta,\eta^\prime)=\sum_{k=0}^{m+1\wedge m^\prime}k!\binom{m+1}{k}\binom{m^\prime}{k}[Q_{2k}\cdot(\Gamma_{\cdot_Q}^\mathbb{C}(\Phi)^{m+1-k})\Gamma_{\cdot_Q}^\mathbb{C}(\overline{\Phi})^{m^\prime-k}](f;\eta,\eta^\prime)\,,
\end{align*}
where $m+1\wedge m^\prime=\min\{m+1,m^\prime\}$ and where the symbol $Q_{2k}$ is a shortcut notation for
\begin{align*}
Q_{2k}(f)\vcentcolon=(G\overline{G})^{\otimes k}\cdot(\delta_{\mathrm{Diag}_k}\otimes 1_k)(f\otimes 1_{2k-1})\,.\quad\forall f\in\mathcal{D}(\mathbb{R}^{d+1}).
\end{align*}
We observe that, although the distribution $Q_{2k}$ is a priori ill-defined, as discussed in Step 2 of this proof, \clacomment{$G\overline{G}\in\mathcal{D}^\prime((\mathbb{R}^{d+1})^2\setminus\Lambda_t^2)$} admits an extension $\widehat{G\overline{G}}$. In turn this yields a renormalized version of $Q_{2k}$, \textit{i.e.},
\begin{align*}
\widehat{Q}_{2k}(f)\vcentcolon=(\widehat{G\overline{G}})^{\otimes k}\cdot(\delta_{\mathrm{Diag}_k}\otimes 1_k)(f\otimes 1_{2k-1})\,.
\end{align*}
Bearing in mind these premises, we set
\begin{align*}
[\Gamma_{\cdot_Q}^\mathbb{C}(\Phi^{m+1}\overline{\Phi}^{m^\prime})](f;\eta,\eta^\prime)=\sum_{k=0}^{m+1\wedge m^\prime}k!\binom{m+1}{k}\binom{m^\prime}{k}[\widehat{Q}_{2k}\cdot(\Gamma_{\cdot_Q}^\mathbb{C}(\Phi)^{m+1-k})\Gamma_{\cdot_Q}^\mathbb{C}(\overline{\Phi})^{m^\prime-k}](f;\eta,\eta^\prime)\,,
\end{align*}
which is well-defined since it involves only products of distributions generated by smooth functions. All required properties of $\Gamma_{\cdot_Q}^\mathbb{C}$ are satisfied by direct inspection.
Finally, we discuss the last inductive step, namely we assume that $\Gamma_{\cdot_Q}^\mathbb{C}$ has been assigned on $\mathcal{M}_{(m+1,m^\prime)}^j$ and we construct it consistently on $\mathcal{M}_{(m+1,m^\prime)}^{j+1}$. To this end, let us consider $\tau\in\mathcal{M}_{(m+1,m^\prime)}^{j+1}$. Similarly to the inductive procedure in Step {\em 2.}, on account of the definition of $\mathcal{M}_{(m+1,m^\prime)}^{j+1}$ and of the linearity of $\Gamma_{\cdot_Q}^\mathbb{C}$, it suffices to consider elements $\tau$ either of the form $\tau=G\circledast\tau^\prime$ with $\tau^\prime\in\mathcal{M}_{(m+1,m^\prime)}^{j}$ or of the form
\begin{align*}
\tau=\tau_1\ldots\tau_\ell\,,\qquad\tau_i\in\mathcal{M}_{(m_i,m_i^\prime)}^{j}\cup G\circledast\mathcal{M}_{(m_i,m_i^\prime)}^{j}\cup \overline{G}\circledast\mathcal{M}_{(m_i,m_i^\prime)}^{j}\,,\qquad i\in\{1,\ldots,\ell\}\,,\qquad\ell\in\mathbb{N}\,,
\end{align*}
where for any $i\in\{1,\ldots,\ell\}$ it holds $m_i,m_i^\prime\in\mathbb{N}\setminus\{0\}$ and $\sum_{i=1}^\ell m_i=m+1$ and $\sum_{i=1}^\ell m^\prime_i=m^\prime$.
We observe that in the first case the sought result descends as a direct consequence of Step {\em 1.} and of the inductive hypothesis. Hence we focus only on the second one. In view of the inductive hypothesis, $\Gamma_{\cdot_Q}^\mathbb{C}$ has already been assigned on $\tau_i$ for all $i\in\{1,\ldots,\ell\}$.
As before, Equation \eqref{Eq: Gammadot on products} yields a formal expression, namely
\begin{align}
\Gamma_{\cdot_Q}^\mathbb{C}(\tau)(f;\eta,\eta^\prime)=
\sum_{N,M\geq0}\sum_{\substack{N_1+\ldots+N_\ell=N+M\\ M_1+\ldots+M_\ell=N+M}}\mathcal{F}(N_1,\ldots&,N_\ell;M_1,\ldots,M_\ell)\cdot\bigl[(\delta_{Diag_l}\otimes Q^{\otimes N}\otimes\overline{Q}^{\otimes M})\cdot\notag\\
&\Bigl(U_{1}^{(N_1,M_1)}\Tilde{\otimes}\ldots\Tilde{\otimes} U_{\ell}^{(N_\ell,M_\ell)}\Bigr)\bigr](f\otimes\mathbf{1}_{\ell-1+2N+2M};\eta,\eta^\prime)\,,\label{Eq: Gamma long}
\end{align}
where we have introduced the notation $\Gamma_{\cdot_Q}^\mathbb{C}(\tau_i)=U_i$ and where the symbols $\mathcal{F}(N_1,\ldots,N_\ell;M_1,\ldots,M_\ell)$ are $\mathbb{C}$-numbers stemming from the underlying combinatorics. Their explicit form plays no r\^{o}le and hence we omit them.
First of all, since we consider only polynomial functionals, only a finite number of terms in the above sum is non-vanishing. Nonetheless, the expression in Equation \eqref{Eq: Gamma long} is still purely formal and renormalization needs to be accounted for. To be more precise, let us introduce
\begin{align*}
U_{N,M}\vcentcolon=\bigl[(\delta_{Diag_\ell}\otimes Q^{\otimes N}\otimes\overline{Q}^{\otimes M})\cdot\Bigl(U_{1}^{(N_1,M_1)}\Tilde{\otimes}\ldots\Tilde{\otimes} U_{\ell}^{(N_\ell,M_\ell)}\Bigr)\bigr]\,.
\end{align*}
On the one hand, by \cite[Thm. 8.2.9]{Hormander-I-03},
\begin{align*}
\mathrm{WF}(\delta_{Diag_\ell}\otimes Q^{\otimes N}\otimes\overline{Q}^{\otimes M})\subseteq\{(\widehat{x}_\ell,\widehat{z}_{2N+2M}&;\widehat{k}_\ell,\widehat{q}_{2N,2M})\in T^\ast (\mathbb{R}^{d+1})^{\ell+2N+2M}\setminus\{0\}\,\vert\, \\
&(\widehat{x}_\ell,\widehat{k}_\ell)\in\mathrm{WF}(\delta_{\text{Diag}_\ell})\,, (\widehat{z}_{2N+2M};\widehat{q}_{2N,2M})\in\mathrm{WF}(Q^{\otimes N}\otimes\overline{Q}^{\otimes M})\}\,,
\end{align*}
while, on the other hand, by the inductive hypothesis, $\mathrm{WF}(U_{i}^{(N_i,M_i)})\subseteq C_{N_i+M_i+1}$ for any $i\in\{1,\ldots,\ell\}$.
These two data together with \cite[Thm. 8.2.10]{Hormander-I-03} imply \clacomment{$U_{N,M}\in\mathcal{D}'(M^{\ell+2N+2M}\setminus\Lambda_t^{\ell+2N+2M,\mathrm{Big}})$}, where
\clacomment{
\begin{align*}
\Lambda_t^{\ell+2N+2M,\mathrm{Big}}\vcentcolon=\{(\widehat{t}_{\ell+2N+2M},\widehat{x}_{\ell+2N+2M})\in\mathbb{R}^{(d+1)(\ell+2N+2M)}\,\vert\,\exists i,j\in\{1,\ldots,\ell+2N+2M\}, t_i=t_j\}\,.
\end{align*}}
In addition, again by \cite[Thm. 8.2.10]{Hormander-I-03}, it holds
\begin{equation*}
\begin{split}
\mathrm{WF}(U_{N,M}&)=\{(\widehat{x}_\ell,\widehat{z}_{2N+2M},\widehat{k}_\ell+\widehat{k'}_{\ell},\widehat{q}_{N_1,M_1}+\widehat{q'}_{N_1,M_1},\ldots,\widehat{q}_{N_\ell,M_\ell}+\widehat{q'}_{N_\ell,M_\ell})\in\\
&\hspace{1cm}\in T^\ast (\mathbb{R}^{d+1})^{\ell+2N+2M}\setminus\{0\}\,\vert\, (\widehat{x}_\ell,\widehat{k}_\ell)\in\mathrm{WF}(\delta_{\text{Diag}_\ell}),\\
&\hspace{2cm}\,(\widehat{z}_{2N+2M},\widehat{q}_{2N+2M})\in\mathrm{WF}(Q^{\otimes N}\otimes\overline{Q}^{\otimes M}),\\
&\hspace{3cm}(x_i,\widehat{z}_{N_i,M_i};k_i,\widehat{q'}_{N_i,M_i})\in C_{1+N_i+M_i} \forall i\in\{1,\ldots\ell\}\}\,.
\end{split}
\end{equation*}
This estimate can be improved through the following argument, which is a generalization of the one used in Step {\em 2.}
Let $\{A,B\}$ be a partition of the index set $\{1,\ldots,\ell+2N+2M\}$, in two disjoint sets such that if
\clacomment{$\{t_1,x_1,\ldots,t_{\ell+2N+2M},x_{\ell+2N+2M}\}=\{\widehat{t}_A,\widehat{x}_A,\widehat{t}_B,\widehat{x}_B\}$, then $t_a\neq t_b$ for any $t_a\in\widehat{t}_A$ and $t_b\in\widehat{t}_B$.}
In such scenario, the integral kernel of $U_{N,M}$ can be decomposed as\clacomment{
\begin{align*}
U_{N,M}(t_1,x_1,\ldots,t_{\ell+2N+2M},x_{\ell+2N+2M})=K^A_{N,M}(\widehat{t}_A,\widehat{x}_A)S_{N,M}(\widehat{t}_A,\widehat{x}_A,\widehat{t}_B,\widehat{x}_B)K^B_{N,M}(\widehat{t}_B,\widehat{x}_B)\,,
\end{align*}
where the kernel $S_{N,M}(\widehat{t}_A,\widehat{x}_A,\widehat{t}_B,\widehat{x}_B)$ is smooth on such a partition and where $K^A_{N,M}(\widehat{t}_A,\widehat{x}_A)$ and $K^B_{N,M}(\widehat{t}_B,\widehat{x}_B)$ are kernels of distributions appearing in the definition of the map $\Gamma_{\cdot_Q}^\mathbb{C}$ on $\mathcal{M}_{(k,k^\prime)}^p$ with $k<m+1$, $k^\prime<m^\prime$ and $p<j+1$.
This, together with the inductive hypothesis, implies that the product $(K^A_{N,M}\otimes K^B_{N,M})\cdot S_{N,M}$ is well defined.
Since this argument is independent from the partition $\{A,B\}$, it follows that $U_{N,M}\in\mathcal{D}'(M^{\ell+2N+2M}\setminus\Lambda_t^{\ell+2N+2M})$.
To conclude, observing that
\begin{equation*}
\begin{split}
\text{wsd}_{\Lambda_t^{\ell+2N+2M}}(U_{N,M})\leq \text{wsd}_{\Lambda_t^{\ell+2N+2M}}(\delta_{Diag_\ell}\otimes Q^{\otimes N}\otimes\overline{Q}^{\otimes M})+\sum_{i=1}^\ell\text{wsd}_{\Lambda_t^{1+N_i+M_i}}(U_{i}^{(N_i,M_i)})<\infty\,,
\end{split}
\end{equation*}
once more, \cite[Thm. B.8]{Dappiaggi:2020gge} and \cite[Rem. B.9]{Dappiaggi:2020gge} grant the existence of an extension $\widehat{U}_{N,M}\in\mathcal{D}^\prime((\mathbb{R}^{d+1})^{\ell+2N+2M})$ of $U\in\mathcal{D}^\prime((\mathbb{R}^{d+1})^{\ell+2N+2M}\setminus\Lambda_t^{\ell+2N+2M})$ which preserves the scaling degree and the wave-front set.
This allows us to set
\begin{align}\label{Eq: extension of the general case}
\Gamma_{\cdot_Q}^\mathbb{C}(\tau)(f;\eta,\eta^\prime)=
\sum_{N,M\geq0}\sum_{\substack{N_1+\ldots+N_\ell=N+M\\ M_1+\ldots+M_\ell=N+M}}\mathcal{F}(N_1,\ldots&,N_\ell;M_1,\ldots,M_\ell)\cdot\widehat{U}_{N,M}\,.
\end{align}}
Following the same strategy as in the proof of \cite[Thm. 3.1]{Dappiaggi:2020gge}, \textit{mutatis mutandis}, one can show that Equation \eqref{Eq: extension of the general case} satisfies the conditions required in the statement of the theorem.
\end{proof}
A map $\Gamma_{\cdot_Q}^\mathbb{C}$ satisfying the properties listed in Theorem \ref{Thm: deformation map cdot} is the main ingredient for introducing a deformation of the pointwise product allowing the class of functional-valued distributions to encode the stochastic behaviour induced by the white noise.
This is codified in the algebra introduced in the following theorem, whose proof we omit since, being completely algebraic, it is analogous to the one of \cite[Cor. 3.3]{Dappiaggi:2020gge}.
\begin{theorem}\label{thm_struttura_deformata}
Let $\Gamma_{\cdot_Q}^\mathbb{C}:\mathcal{A}^\mathbb{C}\rightarrow\mathcal{D}^\prime_C(\mathbb{R}^{d+1};\mathrm{Pol}_\mathbb{C})$ be the linear map introduced via Theorem \ref{Thm: deformation map cdot}.
Moreover, let us $\mathcal{A}^\mathbb{C}_{\cdot_Q}\vcentcolon=\Gamma_{\cdot_Q}^\mathbb{C}(\mathcal{A})$.
In addition, let
\begin{equation}
u_1\cdot_{\Gamma_{\cdot_Q}^\mathbb{C}}u_2\vcentcolon=\Gamma_{\cdot_Q}^\mathbb{C}\bigl[{\Gamma_{\cdot_Q}^\mathbb{C}}^{-1}(u_1){\Gamma_{\cdot_Q}^\mathbb{C}}^{-1}(u_2)\bigr],\hspace{1cm}\forall\,u_1,u_2\in\mathcal{A}^\mathbb{C}_{\cdot_Q}\,.
\end{equation}
Then $(\mathcal{A}^\mathbb{C}_{\cdot_Q},\cdot_{\Gamma_{\cdot_Q}^\mathbb{C}})$ is a unital, commutative and associative $\mathbb{C}$-algebra.
\end{theorem}
\paragraph{Uniqueness}
The main result of the previous section consists of the existence of a map $\Gamma_{\cdot_Q}^\mathbb{C}$ satisfying the condition as per Theorem \ref{Thm: deformation map cdot} and which, through a renormalization procedure, codifies at an algebraic level the stochastic properties of $\Psi$ induced by the complex white noise $\xi$.
In the following we shall argue that the map $\Gamma_{\cdot_Q}^\mathbb{C}$ is not unique and, to this end, it is convenient to focus on Equation \eqref{Eq: renormalized squared modulus} and in particular on $\widehat{G\cdot\overline{G}}\in\mathcal{D}^\prime(\mathbb{R}^{d+1}\times\mathbb{R}^{d+1})$.
\clacomment{As discussed above, this is an extension of $G\cdot\overline{G}\in\mathcal{D}^\prime(\mathbb{R}^{d+1}\times\mathbb{R}^{d+1}\setminus\Lambda_t^2)$.}
Due to \cite[Thm. B.8]{Dappiaggi:2020gge} and \cite[Rem. B.9]{Dappiaggi:2020gge}, such an extension might not be unique depending on the dimension $d\in\mathbb{N}$ of the underlying space.
In spite of this, we can draw some conclusions on the relation between different prescriptions for the deformation map $\Gamma_{\cdot_Q}^\mathbb{C}$ and in turn for the associated algebras $\mathcal{A}_{\cdot_Q}^\mathbb{C}$.
Results in this direction are analogous to those stated in \cite[Thm. 5.2]{Dappiaggi:2020gge}. Hence mutatis mutandis we adapt them to the complex scenario. Once more we do not give detailed proofs, since these can be readily obtained from the counterparts in \cite{Dappiaggi:2020gge}.
The following two results characterize completely the arbitrariness in choosing the linear map $\Gamma_{\cdot_Q}^\mathbb{C}$, as well as the ensuing link between the deformed algebras.
\begin{theorem}\label{thm_unicc}
Let $\Gamma_{\cdot_Q}^\mathbb{C},\,{\Gamma_{\cdot_Q}^\mathbb{C}}':\mathcal{A}^\mathbb{C}\rightarrow\mathcal{D}^\prime_C(\mathbb{R}^{d+1};\mathrm{Pol}_\mathbb{C})$ be two linear maps satisfying the requirements listed in Theorem \ref{Thm: deformation map cdot}.
There exists a family $\{C_{\ell,\ell^\prime}\}_{\ell,\ell^\prime\in\mathbb{N}_0}$ of linear maps $C_{\ell,\ell'}:\mathcal{A}^\mathbb{C}\rightarrow\mathcal{M}_{\ell,\ell'}$ satisfying the following properties
\begin{itemize}
\item For all $m, m'\in\mathbb{N}$ such that either $m\leq j+1$ or $m'\leq j'+1$, it holds
\begin{equation*}
C_{j,j'}[\mathcal{M}_{m,m'}]=0\,,
\end{equation*}
\item For all $\ell,\ell'\in\mathbb{N}_0$ and for all $\tau\in\mathcal{A}^\mathbb{C}$, it holds
\begin{equation*}
C_{\ell,\ell'}[G\circledast\tau]=G\circledast C_{\ell,\ell'}[\tau],\hspace{1cm}C_{\ell,\ell'}[\overline{G}\circledast\tau]=\overline{G}\circledast C_{\ell,\ell'}[\tau]\,,
\end{equation*}
\item For all $\ell,\ell'\in\mathbb{N}$ and for all $\zeta\in\mathcal{E}(\mathbb{R}^{d+1};\mathbb{C})$, it holds
\begin{equation*}
\delta_\zeta\circ C_{\ell,\ell'}=C_{\ell-1,\ell'}\circ\delta_\zeta,\hspace{1cm}\delta_{\overline{\zeta}}\circ C_{\ell,\ell'}=C_{\ell,\ell'-1}\circ\delta_{\overline{\zeta}}\,.
\end{equation*}
\item For all $\tau\in\mathcal{M}_{m,m'}$, it holds
\begin{equation*}
{\Gamma_{\cdot_Q}^\mathbb{C}}'(\tau)=\Gamma_{\cdot_Q}^\mathbb{C}\bigl(\tau+C_{m-1,m'-1}(\tau)\bigr)\,.
\end{equation*}
\end{itemize}
\end{theorem}
\begin{corollary}
Under the hypothesis of Theorem \ref{thm_unicc}, the algebras $\mathcal{A}^\mathbb{C}_{\cdot_Q}=\Gamma_{\cdot_Q}^\mathbb{C}(\mathcal{A}^\mathbb{C})$ and ${\mathcal{A}^\mathbb{C}}'_{\cdot_Q}={\Gamma_{\cdot_Q}^\mathbb{C}}'(\mathcal{A}^\mathbb{C})$, defined as per Theorem \ref{thm_struttura_deformata}, are isomorphic.
\end{corollary}
\section{The Algebra $\mathcal{A}^{\mathbb{C}}_{\bullet_Q}$}\label{Sec: correlations}
The construction of Section \ref{Sec: expectations} allows to compute the expectation values of the stochastic distributions appearing in the \emph{perturbative expansion} of the stochastic non-linear Schr\"odinger equation, as we shall discuss in Section \ref{Sec: perturbative analysis}.
Alas, this falls short from characterizing, again at a perturbative level, the stochastic behaviour of the perturbative solution on Equation \eqref{Eq: Stochastic NLS}, since it does not allow the computation of multi-local \emph{correlation functions}, which are instead of great relevance for applications.
Similarly to the case of expectation values, also correlation functions can be obtained through a deformation procedure of a suitable non-local algebra constructed out of $\mathcal{A}^\mathbb{C}_{\cdot_Q}$. Before moving to the main result of this section, we introduce the necessary notation.
As a starting point, we need a multi-local counterpart for the space of polynomial functional-valued distributions, namely
\begin{equation}\label{eq_T'}
\mathcal{T}_{C}^{'\scriptscriptstyle \mathbb{C}}(\mathbb{R}^{d+1};\text{Pol})\vcentcolon=\mathbb{C}\oplus\bigoplus_{\ell\geq1}\mathcal{D}^\prime_C((\mathbb{R}^{d+1})^\ell;\mathsf{Pol}_{\mathbb{C}})\,.
\end{equation}
We are also interested in
\begin{equation}\label{eq_modulo2}
\mathcal{T}^{\scriptscriptstyle \mathbb{C}}(\mathcal{A}^\mathbb{C}_{\cdot_Q})\vcentcolon=\mathcal{E}(\mathbb{R}^{d+1};\mathbb{C})\oplus\bigoplus_{\ell>0}{\mathcal{A}^\mathbb{C}}_{\cdot_Q}^{\,\otimes\ell}\,.
\end{equation}
As anticipated, we encode the information on correlation functions through a deformed algebra structure induced over $ \mathcal{T}^{\scriptscriptstyle \mathbb{C}}(\mathcal{A}^\mathbb{C}_{\cdot_Q})$.
The main ingredient is the bi-distribution $Q\in\mathcal{D}^\prime(\mathbb{R}^{d+1}\times\mathbb{R}^{d+1})$ we introduced in Section \ref{Sec: expectations}, which is the $2$-point correlation function of the stochastic convolution $G\circledast\xi$.
Starting from $Q$ we construct the following product $\bullet_Q$: for any $\tau_1\in\mathcal{D}^\prime_C((\mathbb{R}^{d+1})^{\ell_1};\mathsf{Pol}_{\mathbb{C}})$ and $\tau_2\in\mathcal{D}^\prime_C((\mathbb{R}^{d+1})^{\ell_2};\mathsf{Pol}_{\mathbb{C}})$, for any $f_1\in\mathcal{D}((\mathbb{R}^{d+1})^{\ell_1})$, $f_2\in\mathcal{D}((\mathbb{R}^{d+1})^{\ell_2})$ and for all $\eta,\eta^\prime\in\mathcal{E}(\mathbb{R}^{d+1})$
\begin{gather}
(\tau_1\bullet_Q\tau_2)(f_1\otimes f_2;\eta,\eta^\prime)=\notag\\
\sum\limits_{\substack{k\geq0\\ k_1+k_2=k}}\frac{1}{k_1!k_2!}\big[(\mathbf{1}_{\ell_1+\ell_2}\otimes Q^{\otimes k_1}\otimes\overline{Q}^{\otimes k_2})\cdot(u_1^{(k_1,k_2)}\Tilde{\otimes}u_2^{(k_1,k_2)})\bigr](f_1\otimes f_2\otimes1_{2k};\eta,\eta^\prime)\,.\label{eq_bullet2}
\end{gather}
Here $\Tilde{\otimes}$ codifies a particular permutation of the arguments of the tensor product as in Equation \eqref{Eq: Tilde otimes}.
\begin{theorem}\label{thm_bullet2}
Let $\mathcal{A}^\mathbb{C}_{\cdot_Q}=\Gamma_{\cdot_Q}^\mathbb{C}(\mathcal{A}^\mathbb{C})$ be the deformed algebra defined in Theorem \ref{thm_struttura_deformata} and let us consider the space $ \mathcal{T}_{C}^{'\scriptscriptstyle \mathbb{C}}(\mathbb{R}^{d+1};\mathsf{Pol}_{\mathbb{C}})$ introduced in Equation \eqref{eq_T'}, as well as the universal tensor module $\mathcal{T}^{\scriptscriptstyle \mathbb{C}}(\mathcal{A}^\mathbb{C}_{\cdot_Q})$ defined via Equation \eqref{eq_modulo2}. Then there exists a linear map $\Gamma_{\bullet_Q}^\mathbb{C}:\mathcal{T}^{\scriptscriptstyle\mathbb{C}}(\mathcal{A}^\mathbb{C}_{\cdot_Q})\rightarrow\mathcal{T}'_C(\mathbb{R}^{d+1};\mathsf{Pol}_{\mathbb{C}})$ satisfying the following properties:
\begin{enumerate}
\item For all $\tau_1\ldots,\tau_\ell\in\mathcal{A}^\mathbb{C}_{\cdot_Q}$ with $\tau_1\in\Gamma_{\cdot_Q}^\mathbb{C}(\mathcal{M}_{1,0})$ or $\tau_1\in\Gamma_{\cdot_Q}^\mathbb{C}(\mathcal{M}_{0,1})$, see Remark \ref{moduli2}, it holds
\begin{equation}
\Gamma_{\bullet_Q}^\mathbb{C}(\tau_1\otimes\ldots\otimes\tau_\ell)\vcentcolon=\tau_1\bullet_Q\Gamma_{\bullet_Q}^\mathbb{C}(\tau_2\otimes\ldots\otimes\tau_n)\,,
\end{equation}
where $\bullet_Q$ is the product defined via Equation \ref{eq_bullet2}.
\item For any $\tau_1,\ldots,\tau_\ell\in\mathcal{A}^\mathbb{C}_{\cdot_Q}$, $\eta,\eta^\prime\in\mathcal{E}(M;\mathbb{C})$ and $f_1,\ldots,f_n\in\mathcal{D}(\mathbb{R}^{d+1};\mathbb{C})$ such that there exists $I \subsetneq\{1,\ldots,n\}$ for which
\begin{equation*}
\bigcup_{i\in I}\text{supp}(f_i)\cap\bigcup_{j\notin I}\text{supp}(f_j)=\emptyset\,,
\end{equation*}
it holds
\begin{equation*}
\Gamma_{\bullet_Q}^\mathbb{C}(\tau_1\otimes\ldots\otimes\tau_n)(f_1\otimes\ldots\otimes f_n;\eta,\eta^\prime)
=\Bigl[\Gamma_{\bullet_Q}^\mathbb{C}\bigl(\bigotimes_{i\in I}\tau_i\bigr)\bullet_Q \Gamma_{\bullet_Q}^\mathbb{C}\bigl(\bigotimes_{j\notin I}\tau_j\bigr)\Bigr](f_1\otimes\ldots\otimes f_n;\eta,\eta^\prime)\,.
\end{equation*}
\item Denoting by $\delta_\zeta,\delta_{\overline{\zeta}}$ the functional derivatives in the direction of $\zeta,\overline{\zeta}\in\mathcal{E}(\mathbb{R}^{d+1};\mathbb{C})$, $\Gamma_{\bullet_Q}^\mathbb{C}$ satisfies the following identities:
\begin{align}
&\Gamma_{\bullet_Q}^\mathbb{C}(\tau)=\tau\,,\hspace{1cm}\forall\tau\in\mathcal{A}^\mathbb{C}_{\cdot_Q},\label{identita_A}\\
&\Gamma_{\bullet_Q}^\mathbb{C}\circ\delta_{\zeta}=\delta_\zeta\circ\Gamma_{\bullet_Q}^\mathbb{C}\,,\hspace{0.6cm}\Gamma_{\bullet_Q}^\mathbb{C}\circ\delta_{\overline{\zeta}}=\delta_{\overline{\zeta}}\circ\Gamma_{\bullet_Q}^\mathbb{C}\,,\hspace{0.6cm}\forall\zeta\in\mathcal{E}(\mathbb{R}^{d+1};\mathbb{C})\,,
\end{align}
\begin{equation}
\begin{split}
\Gamma_{\bullet_Q}^\mathbb{C}(\tau_1&\otimes\ldots\otimes G\circledast\tau_i\otimes\ldots\otimes\tau_n)\\
&=(\delta_{\text{Diag}_2}^{\otimes i-1}\otimes G\otimes \delta_{\text{Diag}_2}^{\otimes n-i})\circledast\Gamma_{\bullet_Q}^\mathbb{C}(\tau_1\otimes\ldots\otimes\tau_i\otimes\ldots\otimes\tau_n)\,.
\end{split}
\end{equation}
\begin{equation}
\begin{split}
\Gamma_{\bullet_Q}^\mathbb{C}(\tau_1&\otimes\ldots\otimes \overline{G}\circledast\tau_i\otimes\ldots\otimes\tau_n)\\
&=(\delta_{\text{Diag}_2}^{\otimes i-1}\otimes\overline{G}\otimes\delta_{\text{Diag}_2}^{\otimes n-i})\circledast\Gamma_{\bullet_Q}^\mathbb{C}(\tau_1\otimes\ldots\otimes \tau_i\otimes\ldots\otimes\tau_n)\,.
\end{split}
\end{equation}
for all $\tau_1,\ldots,\tau_n\in\mathcal{A}^\mathbb{C}_{\cdot_Q}$, $n\in\mathbb{N}_0$.
\end{enumerate}
\end{theorem}
\noindent Similarly to Section \ref{Sec: expectations}, through a map $\Gamma_{\bullet_Q}^\mathbb{C}$ we can induce an algebra structure.
\begin{theorem}\label{thm_bulletalgebra}
Let $\Gamma_{\bullet_Q}^\mathbb{C}:\mathcal{T}^{\scriptscriptstyle \mathbb{C}}(\mathcal{A}^\mathbb{C}_{\cdot_Q})\rightarrow\mathcal{T}_C^{'\scriptscriptstyle \mathbb{C}}(\mathbb{R}^{d+1};\mathsf{Pol}_{\mathbb{C}})$ be a linear map satisfying the constraints of Theorem \ref{thm_bullet2} and let us define
\begin{equation}\label{defoQ}
\mathcal{A}^\mathbb{C}_{\bullet_Q}\vcentcolon=\Gamma_{\bullet_Q}^\mathbb{C}(\mathcal{A}^\mathbb{C}_{\cdot_Q})\subseteq\mathcal{T}_C^{'\scriptscriptstyle \mathbb{C}}(\mathbb{R}^{d+1};\mathsf{Pol}_{\mathbb{C}})\,.
\end{equation}
Furthermore, let us consider the bilinear map $\bullet_{\Gamma_{\bullet_Q}^\mathbb{C}}:\mathcal{A}^\mathbb{C}_{\bullet_Q}\times\mathcal{A}^\mathbb{C}_{\bullet_Q}\rightarrow\mathcal{A}^\mathbb{C}_{\bullet_Q}$
\begin{equation}\label{prod_bullet}
\tau_1\bullet_{\Gamma_{\bullet_Q}^\mathbb{C}}\tau_2\vcentcolon=\Gamma_{\bullet_Q}^\mathbb{C}\bigl({\Gamma_{\bullet_Q}^\mathbb{C}}^{-1}(u_1)\otimes{\Gamma_{\bullet_Q}^\mathbb{C}}^{-1}(u_1)\bigr)\hspace{1cm}\forall\,\tau_1,\tau_2\in\mathcal{A}^\mathbb{C}_{\bullet_Q}\,.
\end{equation}
Then $(\mathcal{A}^\mathbb{C}_{\bullet_Q},\bullet_{\Gamma_{\bullet_Q}^\mathbb{C}})$ identifies a unital, commutative and associative algebra.
\end{theorem}
\begin{remark}
For the sake of brevity, we omit the proofs of Theorems \ref{thm_bullet2} and \ref{thm_bulletalgebra} since, barring minor modifications, they follow the same lines of those outlined in \cite[Sec. 4]{Dappiaggi:2020gge}.
\end{remark}
\begin{example}\label{Ex: Example interpretation bullet}
To better grasp the r\^ole played by this deformation of the algebraic structure, in total analogy with Example \ref{Ex: Example interpretation dot} we compare the two point correlation function of the random field $\varphi$, defined via Equation \eqref{Eq: covariance complex Gaussian random field}, with the functional counterpart
\begin{align*}
\Gamma_{\bullet_Q}^{\mathbb{C}}(\Phi\overline{\Phi})(f_1\otimes f_2;\eta,\overline{\eta})&=(\Phi\bullet_Q\overline{\Phi})(f_1\otimes f_2;\eta, \overline{\eta})=\Phi\otimes\overline{\Phi}(f_1\otimes f_2;\eta,\overline{\eta})\\
&+[(\mathbf{1}_2\otimes Q)\cdot(\delta_{Diag_2}\Tilde{\otimes}\delta_{Diag_2})](f_1\otimes f_2\otimes 1_2;\eta, \overline{\eta})\\
&=\Phi\otimes\overline{\Phi}(f_1\otimes f_2;\eta,\overline{\eta})+Q(f_1\otimes f_2).
\end{align*}
which, evaluated at the configurations $\eta=0$, yields $(\Phi\bullet_Q\overline{\Phi})(f_1\otimes f_2;0,0)=Q(f_1\otimes f_2)\equiv\omega_2(f_1\otimes f_2)$.
Note that we could have also computed the correlation function of the random field $\varphi$ with itself, obtaining zero. More generally, the two-point correlation function of a polynomial expression of $\varphi$ and $\overline{\varphi}$ with itself turns out to be trivial, while the one of the expression and its complex conjugate contains relevant stochastic information, as shown in Section \ref{Sec: perturbative analysis}.
\end{example}
Similarly to the case of the deformation map $\Gamma_{\cdot_Q}^\mathbb{C}$ and the algebra $\mathcal{A}^\mathbb{C}_{\cdot_Q}$, also for $\Gamma_{\bullet_Q}^\mathbb{C}$ and the algebra $\mathcal{A}^\mathbb{C}_{\bullet_Q}$ one can discuss the issue of uniqueness due to the renormalization procedure exploited in the construction of $\Gamma_{\bullet_Q}^\mathbb{C}$. The following theorem deals with this hurdle. Once more we do not give a detailed proof, since it can be readily obtained from the counterpart in \cite{Dappiaggi:2020gge}.
\begin{theorem}\label{Thm: Uniqueness bullet algebra}
Let $\Gamma_{\bullet_Q}^{\mathbb{C}}\,{\Gamma_{\bullet_Q}^{\mathbb{C}}}':\mathcal{A}^\mathbb{C}_{\cdot_Q}\rightarrow\mathcal{T}^{'\scriptscriptstyle \mathbb{C}}_C(\mathbb{R}^{d+1};\mathsf{Pol}_{\mathbb{C}})$ be two linear maps satisfying the requirements listed in Theorem \ref{thm_bullet2}. Then there exists a family $\{C_{\underline{m},\underline{m}'}\}_{\underline{m},\underline{m}'\in\mathbb{N}_0^{\mathbb{N}_0}}$ of linear maps $C_{\underline{m},\underline{m}'}:\mathcal{T}^{\scriptscriptstyle \mathbb{C}}(\mathcal{A}^\mathbb{C}_{\cdot_Q})\rightarrow\mathcal{T}^{\scriptscriptstyle \mathbb{C}}(\mathcal{A}^\mathbb{C}_{\cdot_Q})$, the space $\mathcal{T}^{\scriptscriptstyle \mathbb{C}}(\mathcal{A}^\mathbb{C}_{\cdot_Q})$ being defined via Equation \eqref{eq_modulo2}, satisfying the following properties:
\begin{enumerate}
\item For all $j\in\mathbb{N}_0$, it holds
\begin{equation*}
C_{\underline{m},\,\underline{m}'}[(\mathcal{\mathcal{A}^\mathbb{C}_{\cdot_Q}})^{\otimes j}]\subseteq\mathcal{M}_{m_1,m_1'}\otimes\ldots\otimes\mathcal{M}_{m_j,m_j'},
\end{equation*}
while, if either $m_i\leq l_i-1$ or $m'_i\leq l'_i-1$ for some $i\in\{1,\ldots, j\}$, then
\begin{equation*}
C_{\underline{l},\,\underline{l}'}[\mathcal{M}_{m_1,m_1'}\otimes\ldots\otimes\mathcal{M}_{m_j,m'_j}]=0.
\end{equation*}
\item For all $j\in\mathbb{N}\cup\{0\}$ and $u_1\ldots, u_j\in\mathcal{A}^\mathbb{C}_{\cdot_Q}$, the following identities hold true:
\begin{equation*}
\begin{split}
C_{\underline{m},\,\underline{m}'}[u_1\otimes\ldots&\otimes G\circledast u_k\otimes\ldots\otimes u_j]\\
&=(\delta_{\text{Diag}_2}^{\otimes(k-1)}\otimes G\otimes\delta_{\text{Diag}_2}^{\otimes(j-k)})\circledast C_{\underline{m},\,\underline{m}'}[u_1\otimes\ldots\otimes u_j],
\end{split}
\end{equation*}
\begin{equation*}
\begin{split}
C_{\underline{m},\,\underline{m}'}[u_1\otimes\ldots&\otimes \overline{G}\circledast u_k\otimes\ldots\otimes u_j]\\
&=(\delta_{\text{Diag}_2}^{\otimes(k-1)}\otimes \overline{G}\otimes\delta_{\text{Diag}_2}^{\otimes(j-k)})\circledast C_{\underline{m},\,\underline{m}'}[u_1\otimes\ldots\otimes u_j],
\end{split}
\end{equation*}
\begin{equation*}
\delta_\psi C_{\underline{m},\,\underline{m}'}[u_1\otimes\ldots\otimes u_j]=\sum_{a=1}^jC_{\underline{m}(a),\,\underline{m}'}[u_1\otimes\ldots\otimes\delta_\psi u_a\otimes\ldots\otimes u_j],
\end{equation*}
\begin{equation*}
\delta_{\overline{\psi}} C_{\underline{m},\,\underline{m}'}[u_1\otimes\ldots\otimes u_j]=\sum_{a=1}^jC_{\underline{m},\,\underline{m}'(a)}[u_1\otimes\ldots\otimes\delta_{\overline{\psi}} u_a\otimes\ldots\otimes u_j],
\end{equation*}
where $\underline{m}(a)_i=m_i$ if $i\neq a$ and $\underline{m}(a)_a=m_a-1$, while $\delta_\psi$, $\delta_{\overline{\psi}}$ are the directional derivatives with respect to $\psi,\overline{\psi}\in\mathcal{E}(\mathbb{R}^{d+1};\mathbb{C})$, see Definition \ref{Def: Functionals}. Here $G\in\mathcal{D}^\prime(\mathbb{R}^{d+1}\times \mathbb{R}^{d+1})$ is a fundamental solution of the parabolic operator $L$.
\item Let us consider two linear maps $\Gamma_{\cdot_Q}^\mathbb{C},\,{\Gamma_{\cdot_Q}^\mathbb{C}}':\mathcal{A}^\mathbb{C}\rightarrow\mathcal{D}'_\mathbb{C}(\mathbb{R}^{d+1};\mathsf{Pol}_{\mathbb{C}})$ compatible with the constraints of Theorem \ref{Thm: deformation map cdot}. For all $u_{m_1,m'_1},\ldots, u_{m_j,m'_j}\in\mathcal{A}^\mathbb{C}$ with $u_{m_i,m'_i}\in\mathcal{M}_{m_i,m'_i}$ for all $i\in\{1,\ldots,j\}$ and for $f_1,\ldots,f_j\in\mathcal{D}(\mathbb{R}^{d+1})$ it holds
\begin{equation*}
\begin{split}
{\Gamma_{\bullet_Q}^\mathbb{C}}'(&{\Gamma_{\cdot_Q}^\mathbb{C}}^{'\,\otimes j}(u_{m_1,m'_1}\otimes\ldots\otimes u_{m_j,m'_j}))(f_1\otimes\ldots\otimes f_j)\\
&=\Gamma_{\bullet_Q}^\mathbb{C}({\Gamma_{\cdot_Q}^\mathbb{C}}^{\otimes j}(u_{m_1,m_1'}\otimes\ldots\otimes u_{m_j,m'_j}))(f_1\otimes\ldots\otimes f_j)\\
&+\sum_{\mathcal{G}\in \mathcal{P}(1,\ldots,j)}\Gamma_{\bullet_Q}^\mathbb{C}\Bigl[{\Gamma_{\cdot_Q}^\mathbb{C}}^{\otimes\vert\mathcal{G}\vert}C_{\underline{m}_{\mathcal{G}},\,\underline{m}'_{\mathcal{G}}}\Bigl(\bigotimes_{I\in\mathcal{G}}\prod_{i\in I}u_{m_i,\,m'_i}\Bigr)\Bigr]\Bigl(\bigotimes_{I\in\mathcal{G}}\prod_{i\in I}f_i\Bigr).
\end{split}
\end{equation*}
Here $\mathcal{P}(1,\ldots,j)$ denotes the set of all possible partitions of $\{1,\ldots,j\}$ into non-empty disjoint subsets while $\underline{m}_{\mathcal{G}}=(m_I)_{I\in\mathcal{G}}$, where $m_I:=\sum_{i\in I}m_i$.
\end{enumerate}
\end{theorem}
\section{Perturbative Analysis of the Nonlinear Schr\"odinger Dynamics}\label{Sec: perturbative analysis}
In the following we apply the framework devised in Sections \ref{Sec: expectations} and \ref{Sec: correlations} to the study of the stochastic nonlinear Schr\"odinger equation, that is Equation \eqref{Eq: Stochastic NLS}, where we set for simplicity $\kappa=1$.
The first step consists of translating the equation of interest to a functional formalism, replacing the unknown random distribution $\psi$ with a polynomial functional-valued distribution $\Psi\in\mathcal{D}^\prime(\mathbb{R}^{d+1};\mathsf{Pol}_{\mathbb{C}})$ as per Definition \ref{Def: Functionals}. Recalling that we denote by $G\in\mathcal{D}'(\mathbb{R}^{d+1}\times \mathbb{R}^{d+1})$ the fundamental solution of the Schr\"odinger operator, see Equation \eqref{Eq: Integral Kernel}, the integral form of Equation \eqref{Eq: Stochastic NLS} reads
\begin{equation}\label{Eq: Functional equation}
\Psi= \Phi+\lambda G\circledast\overline{\Psi}\Psi^2,\hspace{1cm}\lambda\in\mathbb{R}_+,
\end{equation}
where $\Phi\in\mathcal{D}^\prime(\mathbb{R}^{d+1};\mathsf{Pol}_{\mathbb{C}})$ is the functional defined in Example \ref{Ex: Basic Functionals}.
Equation \eqref{Eq: Functional equation} cannot be solved exactly and therefore we rely on a perturbative analysis. Hence we expand the solution as a formal power series in the coupling constant $\lambda\in\mathbb{R}_{+}$ with coefficients lying in $\mathcal{A}^{\mathbb{C}}\subset\mathcal{D}^\prime_C(\mathbb{R}^{d+1};\mathsf{Pol}_{\mathbb{C}})$, see Definition \ref{Def: Pointwise Algebra}:
\begin{equation}\label{Eq: Perturbative expansion}
\Psi\llbracket\lambda\rrbracket=\sum_{k\geq0}\lambda^{k}F_{k},\hspace{1cm}F_k\in\mathcal{A}^{\mathbb{C}}.
\end{equation}
The coefficients of Equation \eqref{Eq: Perturbative expansion} are determined through an iteration procedure yielding, at first orders
\begin{equation}\label{Eq: Espressione coefficienti}
\begin{split}
&F_0=\Phi\,,\\
&F_1=G\circledast\overline{\Phi}\Phi^2\,,\\
&F_2=G\circledast(\overline{F_1}F_0^2+2\overline{F_0}F_0F_1)\,,
\end{split}
\end{equation}
while the $k$-th order contribution reads
\begin{equation}\label{Eq: Generic coefficient}
F_k=\sum\limits_{\substack{k_1,k_2,k_3\in\mathbb{N}_0 \\ k_1+k_2+k_3=k-1}} G\circledast(\overline{F}_{k_1}F_{k_2}F_{k_3}),\hspace{1cm}k\geq3\,.
\end{equation}
We also underline that the $k$-th coefficient is constructed out of lower order coefficients.
\begin{remark}
Working on $\mathbb{R}^{d+1}$ we are forced to introduce a cut-off function $\chi\in\mathcal{D}(\mathbb{R}^{d+1})$ in order to implement the necessary support conditions yielding a well-defined convolution $G\circledast u$, see \cite[Chap. 4]{Hormander-I-03}. As per Remark \ref{Rem: Cut-Off on G}, henceforth the cut-off is left implied.
\end{remark}
The perturbative expansion of Equation \eqref{Eq: Perturbative expansion} is particularly useful for computing statistical quantities at every order in perturbation theory by means of the deformation maps, as outlined in the preceding sections. In what follows we explicitly calculate the expectation value of the solution at first order in $\lambda$, as well as its two point correlation function.
\vskip .3cm
\noindent\textbf{\normalsize Expectation value.}
The strategy is thus the following: we first construct the formal perturbative solution $\Psi\llbracket\lambda\rrbracket$ in the algebra $\mathcal{A}\llbracket\lambda\rrbracket$ and then, on account of Theorem \ref{thm_struttura_deformata} and of the stochastic interpretation of the map $\Gamma_{\cdot_Q}^{\mathbb{C}}$ introduced in Theorem \ref{Thm: deformation map cdot}, we consider $\Gamma_{\cdot_Q}^{\mathbb{C}}(\Psi\llbracket\lambda\rrbracket)$.
Bearing in mind this comment we consider
\begin{equation}\label{Eq: perturbative deformed solution}
\Psi_{\cdot_Q}\llbracket\lambda\rrbracket\vcentcolon=\Gamma_{\cdot_Q}^{\mathbb{C}}(\Psi\llbracket\lambda\rrbracket)=\sum_{k}\lambda^k\Gamma_{\cdot_Q}^{\mathbb{C}}(F_k)\in\mathcal{A}_{\cdot_Q}^{\mathbb{C}}\llbracket\lambda\rrbracket\,,
\end{equation}
where we exploited Equation \eqref{Eq: Perturbative expansion} as well as the linearity of $\Gamma_{\cdot_Q}^{\mathbb{C}}$.
As outlined in Example \ref{Ex: Example interpretation dot}, the action of $\Gamma_{\cdot_Q}^{\mathbb{C}}$ on elements lying in $\mathcal{A}^{\mathbb{C}}$ coincides with the expectation value of the equivalent expression in terms of the random field $\varphi$. In view of Equation \eqref{Eq: perturbative deformed solution} up to the first order in perturbation theory, it holds
\begin{equation*}
\Psi\llbracket\lambda\rrbracket=\Phi+\lambda G\circledast\overline{\Phi}\Phi^2+ \mathcal{O}(\lambda^2)\,.
\end{equation*}
Hence, taking into account the defining properties of $\Gamma_{\cdot_Q}^{\mathbb{C}}$ encoded in Equations \eqref{Eq: Gammadot on M_1} and \eqref{Eq: Gammadot Gtau}, the $\eta$-shifted expectation value of the solution reads
\begin{equation*}
\begin{split}
\mathbb{E}[\psi_\eta\llbracket\lambda\rrbracket(f)]&=\Gamma_{\cdot_Q}^{\mathbb{C}}(\Psi\llbracket\lambda\rrbracket)(f;\eta,\overline{\eta})=\Gamma_{\cdot_Q}^{\mathbb{C}}(\Phi)(f;\eta,\overline{\eta})+\lambda\Gamma_{\cdot_Q}^{\mathbb{C}}(G\circledast\overline{\Phi}\Phi^2)(f;\eta,\overline{\eta})\\
&=\Phi(f;\eta,\overline{\eta})+\lambda G\circledast\Gamma_{\cdot_Q}^{\mathbb{C}}(\overline{\Phi}\Phi^2)(f;\eta,\overline{\eta})\,,
\end{split}
\end{equation*}
for all $f\in\mathcal{D}(\mathbb{R}^{d+1}), \eta\in\mathcal{E}(\mathbb{R}^{d+1})$. Notice that, at this stage, we have restored the information that $\Phi$ and $\overline{\Phi}$ are related by complex conjugation by choosing $\eta$ and $\bar{\eta}$ as the underlining configurations. The explicit expression of $\cdot_Q$ in Equation \eqref{Eq: prodotto deformato2} entails
\begin{equation*}
\begin{split}
\Gamma_{\cdot_Q}^{\mathbb{C}}(\overline{\Phi}\Phi^2)(f;\eta,\overline{\eta})&=(\overline{\Phi}\cdot_Q\Gamma_{\cdot_Q}^{\mathbb{C}}(\Phi^2))(f;\eta,\overline{\eta})=\overline{\Phi}\Phi^2(f;\eta,\overline{\eta})\\
&+2(\delta_{Diag_2}\otimes\overline{Q})(\delta_{Diag_2}\otimes\Phi\delta_{Diag_2})(f\otimes\delta_{Diag_3};\eta,\overline{\eta})\\
&=\overline{\Phi}\Phi^2(f;\eta,\overline{\eta})+2\overline{C}\Phi(f;\eta,\overline{\eta})\,,
\end{split}
\end{equation*}
where $\overline{C}\in\mathcal{E}(\mathbb{R}^{d+1})$ is a smooth function whose integral kernel reads $\overline{C}(x)=\chi(x)(\widehat{G\cdot\overline{G}})(\delta_x\otimes\chi)$. Here $\widehat{G\cdot\overline{G}}$ is any but fixed extension of \clacomment{$G\cdot\overline{G}\in\mathcal{D}'(\mathbb{R}^{d+1}\times\mathbb{R}^{d+1}\setminus\Lambda_t^2)$} to the whole $\mathbb{R}^{(d+1)}\times\mathbb{R}^{(d+1)}$, see the proof of Theorem \ref{Thm: deformation map cdot}. We adopted the notation $\psi_\eta$ to underline that its expectation value is in terms of the shifted random distribution $\varphi_\eta=G\circledast\xi+\eta$, which is still Gaussian, but with expectation value $\mathbb{E}[\varphi_\eta]=\eta$. To wit, the centered Gaussian noise properties can be recollected setting $\eta=0$. Thus the expectation value of the solution up to order $\mathcal{O}(\lambda^2)$ reads
\begin{equation}
\mathbb{E}[\psi_0\llbracket\lambda\rrbracket]=\mathcal{O}(\lambda^2)\,.
\end{equation}
This result is a direct consequence of the cubic nonlinear term and it can be extended to any order in perturbation theory.
\begin{theorem}\label{Thm: valor medio nullo}
Let $\Psi\llbracket\lambda\rrbracket\in\mathcal{A}^\mathbb{C}$ be the perturbative solution of Equation \eqref{Eq: Functional equation}. Then
\begin{equation*}
\mathbb{E}[\psi_0\llbracket\lambda\rrbracket(f)]=\Gamma_{\cdot_Q}^\mathbb{C}(\Psi\llbracket\lambda\rrbracket)(f;0,0)=0,\hspace{1cm}\forall f\in\mathcal{D}(\mathbb{R}^{d+1})\,.
\end{equation*}
\end{theorem}
\begin{proof}
By direct inspection of Equations \eqref{Eq: prodotto deformato2} and \eqref{Eq: Gammadot on products} one can infer that the action of $\Gamma_{\cdot_Q}^\mathbb{C}$ on $u\in\mathcal{A}^{\mathbb{C}}$ yields a sum of terms whose polynomial degree in $\Phi$ and $\overline{\Phi}$ is decreased by an even number due to contractions between pairs of fields $\Phi$ and $\overline{\Phi}$. Hence it is sufficient to show that all perturbative coefficients $F_k$ of Equation \eqref{Eq: Perturbative expansion} lie in $\mathcal{M}_{m,m'}$ with $m+m'=2l+1$ for $l\in\mathbb{N}\setminus\{0\}$. If this holds true, all terms resulting from the action of $\Gamma_{\cdot_Q}^\mathbb{C}(u)$ would contain at least one field $\Phi$ or $\overline{\Phi}$ and their evaluation at $\eta=0$ would vanish.
Let us define the vector space $\mathcal{O}\subset\mathcal{A}^\mathbb{C}$ of polynomial functional-valued distributions of odd polynomial degree, namely $u^{(k_1,k_2)}(\cdot,0)=0$ for all $u\in\mathcal{O}$ with
$k_1,k_2\in\mathbb{N}_0$ such that $k_1+k_2=2l$, $l\in\mathbb{N}$, see Definition \ref{Def: Functionals}. Observe that $\mathcal{O}$ is closed under the action of $\Gamma_{\cdot_Q}^\mathbb{C}$. In addition it holds that $u_1u_2u_3\in\mathcal{O}$ for any $u_1,u_2,u_3\in\mathcal{O}$ and if $u\in\mathcal{O}$, then $\overline{u}\in\mathcal{O}$.
Focusing on the problem under scrutiny, the thesis translates into requiring that $\Psi\llbracket\lambda\rrbracket\in\mathcal{O}$, which via Equation \eqref{Eq: Perturbative expansion} is equivalent to $F_k\in\mathcal{O}$ for all $k\in\mathbb{N}$.
Proceeding inductively, it holds true that, for $k=0$, $F_{0}=\Phi\in\mathcal{O}$.
Assuming that $F_j\in\mathcal{O}$ for any $j\leq k\in\mathbb{N}$, our goal is to prove that the same holds true for $F_{k+1}$.
Invoking Equation \eqref{Eq: Generic coefficient}, it holds
\begin{equation*}
F_{k+1}=\sum\limits_{\substack{k_1,k_2k_3\in\mathbb{N} \\ k_1+k_2+k_3=k}}G\circledast(\overline{F}_{k_1}F_{k_2}F_{k_3})\,.
\end{equation*}
Since $k_i\leq k$ for all $i\in\{1,2,3\}$, the inductive hypothesis entails that $F_{k_i}\in\mathcal{O}$.
From the aforementioned properties of $\mathcal{O}$ we conclude that $F_{k+1}\in\mathcal{O}$.
\end{proof}
\begin{remark}
Theorem \ref{Thm: valor medio nullo} holds true for any nonlinear term of the form $\vert\psi\vert^{2\kappa}\psi=\overline{\psi}^\kappa\psi^{\kappa+1}$, $\kappa\in\mathbb{N}$ in Equation \eqref{Eq: Stochastic NLS} since the proof relies on the property that the perturbative coefficients contain an odd number of fields.
\end{remark}
\vskip .3cm
\noindent\textbf{\normalsize Two-point correlation function.} Another crucial information encoded in the solution of an SPDE resides in the correlation functions which describe the non-local behaviour of the system. Being interested in the calculation of the two-point correlation function, we focus on the deformed algebra $\mathcal{A}^\mathbb{C}_{\bullet_Q}=\Gamma^\mathbb{C}_{\bullet_Q}(\mathcal{A}_{\cdot_Q}^\mathbb{C})$, see Theorems \ref{thm_bullet2} and \ref{thm_bulletalgebra}. In view of Example \ref{Ex: Example interpretation bullet}, the two-point correlation function of the perturbative solution $\psi_\eta\llbracket\lambda\rrbracket$, $\eta\in\mathcal{E}(\mathbb{R}^{d+1})$, localized via $f_1,f_2\in\mathcal{D}(\mathbb{R}^{d+1})$, reads
\begin{equation}\label{Eq: due punti}
\begin{split}
\omega_2(\psi_\eta\llbracket\lambda\rrbracket;f_1&\otimes f_2)=\bigl(\Psi_{\cdot_Q}\llbracket\lambda\rrbracket\bullet_{\Gamma_{\bullet_Q}^\mathbb{C}}\overline{\Psi}_{\cdot_Q}\llbracket\lambda\rrbracket\bigr)(f_1\otimes f_2;\eta,\overline{\eta})\\
&=\Gamma_{\bullet_Q}^\mathbb{C}\bigl[\Gamma_{\cdot_Q}^\mathbb{C}(\Psi\llbracket\lambda\rrbracket)\otimes\Gamma_{\cdot_Q}^{\mathbb{C}}(\overline{\Psi}\llbracket\lambda\rrbracket)\bigr](f_1\otimes f_2;\eta;\overline{\eta})\\
&=\sum_{k\geq0}\lambda^k\sum\limits_{\substack{k_1,k_2 \\ k_1+k_2=k}}\Gamma_{\bullet_Q}\big(\Gamma_{\cdot_Q}(F_{k_1})\otimes\Gamma_{\cdot_Q}(\overline{F}_{k_2})\big)(f_1\otimes f_2;\eta,\overline{\eta})\,,
\end{split}
\end{equation}
where we used once more the linearity of $\Gamma_{\cdot_Q}^\mathbb{C}$ and $\Gamma_{\bullet_Q}^\mathbb{C}$. With reference to the perturbative coefficients as per Equation \eqref{Eq: Espressione coefficienti}, Equation \eqref{Eq: due punti} up to order $\mathcal{O}(\lambda^2)$ reads
\begin{equation*}
\begin{split}
\omega_{2}(\psi_\eta\llbracket\lambda\rrbracket;f_1\otimes f_2)&=\Gamma_{\bullet_Q}^\mathbb{C}(\Gamma_{\cdot_Q}^\mathbb{C}(\Phi)\otimes\Gamma_{\cdot_Q}^\mathbb{C}(\overline{\Phi}))(f_1\otimes f_2;\eta,\overline{\eta})\\
&+\lambda\Gamma_{\bullet_Q}^\mathbb{C}(\Gamma_{\cdot_Q}^\mathbb{C}(G\circledast\overline{\Phi}\Phi^2)\otimes\Gamma_{\cdot_Q}^\mathbb{C}\overline{\Phi})(f_1\otimes f_2;\eta,\overline{\eta})\\
&+\lambda\Gamma_{\bullet_Q}^\mathbb{C}(\Gamma_{\cdot_Q}^\mathbb{C}(\Phi)\otimes\Gamma_{\cdot_Q}^\mathbb{C}(\overline{G}\circledast\Phi\overline{\Phi}^2))(f_1\otimes f_2;\eta,\overline{\eta})\,.
\end{split}
\end{equation*}
Taking into account that Equation \eqref{Eq: Gammadot on M_1} entails that $\Gamma_{\cdot_Q}^\mathbb{C}(\Phi)=\Phi$, $\Gamma_{\cdot_Q}^\mathbb{C}(\overline{\Phi})=\overline{\Phi}$, Equation \eqref{eq_bullet2} yields that each separate term in the last expression reads
\begin{equation*}
\Gamma_{\bullet_Q}^\mathbb{C}(\Phi\otimes\overline{\Phi})(f_1\otimes f_2;\eta,\overline{\eta})=(\Phi\bullet_Q\overline{\Phi})(f_1\otimes f_2;\eta,\overline{\eta})=(\Phi\otimes\overline{\Phi})(f_1\otimes f_2;\eta,\overline{\eta})+Q(f_1\otimes f_2)\,,
\end{equation*}
\begin{equation*}
\begin{split}
\Gamma_{\bullet_Q}^\mathbb{C}(\Gamma_{\cdot_Q}^\mathbb{C}(G\circledast\overline{\Phi}\Phi^2)&\otimes\overline{\Phi})(f_1\otimes f_2;\eta,\overline{\eta})=(G\circledast\overline{\Phi}\Phi^2+2G\circledast\overline{C}\,\Phi)\bullet_Q\overline{\Phi})(f_1\otimes f_2;\eta,\overline{\eta})\\
&=(G\circledast(\overline{\Phi}\Phi^2+2\overline{C}\Phi)\otimes\overline{\Phi})(f_1\otimes f_2;\eta,\overline{\eta})+Q\cdot(G\circledast(2\overline{\Phi}\Phi)\otimes \mathbf{1})(f_1\otimes f_2;\eta,\overline{\eta})\\
&+2\overline{Q}\cdot(G\circledast\overline{C}\mathbf{1}\otimes \mathbf{1})(f_1\otimes f_2;\eta,\overline{\eta})\,,
\end{split}
\end{equation*}
\begin{equation*}
\begin{split}
\Gamma_{\bullet_Q}^\mathbb{C}(\Phi&\otimes\Gamma_{\cdot_Q}^\mathbb{C}(\overline{G}\circledast\Phi\overline{\Phi}^2))(f_1\otimes f_2;\eta,\overline{\eta})=(\Phi\bullet_Q(\overline{G}\circledast\Phi\overline{\Phi}^2+2\overline{G}\circledast C\,\overline{\Phi}))(f_1\otimes f_2;\eta,\overline{\eta})\\
&=(\Phi\otimes \overline{G}\circledast(\Phi\overline{\Phi}^2+2C\overline{\Phi}))(f_1\otimes f_2;\eta,\overline{\eta})+\overline{Q}\cdot(\mathbf{1}\otimes \overline{G}\circledast(2\overline{\Phi}\Phi))(f_1\otimes f_2;\eta,\overline{\eta})\\
&+2Q\cdot(\overline{G}\circledast C\mathbf{1}\otimes \mathbf{1})(f_1\otimes f_2;\eta,\overline{\eta})\,,
\end{split}
\end{equation*}
where $\mathbf{1}\in\mathcal{D}^\prime(\mathbb{R}^{d+1};\mathsf{Pol}_{\mathbb{C}})$ is the identity functional defined in Example \eqref{Ex: Basic Functionals}.
Gathering all contributions together and setting $\eta=0$, at first order in $\lambda$ the two-point correlation function of the perturbative solution reads
\begin{equation}
\omega_2(\psi_0\llbracket\lambda\rrbracket;f_1\otimes f_2)=Q(f_1\otimes f_2)+2\lambda\overline{Q}(G\circledast\overline{C}\mathbf{1}\otimes\mathbf{1})(f_1\otimes f_2)+2\lambda Q\cdot(\overline{G}\circledast C\mathbf{1}\otimes \mathbf{1})(f_1\otimes f_2)+\mathcal{O}(\lambda^2)\,.
\end{equation}
\begin{remark}
In this section we limited ourselves to constructing the two-point correlation function of the solution.
However it is important to highlight that, using the algebraic approach, we could have studied the $m$-point correlation function at any order in perturbation theory, for any $m\in\mathbb{N}$.
Yet, being the underlying random distribution Gaussian, all odd correlation functions vanish, while the even ones can be computed as
\begin{equation*}
\begin{split}
&\omega_{m}(\psi\llbracket\lambda\rrbracket;f_1\otimes\ldots\otimes f_m)\\
&=\bigl(\underbrace{\Psi_{\cdot_Q}\llbracket\lambda\rrbracket\bullet_{\Gamma_{\bullet_Q}^\mathbb{C}}\overline{\Psi}_{\cdot_Q}\llbracket\lambda\rrbracket\bullet_{\Gamma_{\bullet_Q}^\mathbb{C}}\ldots\bullet_{\Gamma_{\bullet_Q}^\mathbb{C}}\Psi_{\cdot_Q}\llbracket\lambda\rrbracket\bullet_{\Gamma_{\bullet_Q}^\mathbb{C}}\overline{\Psi}_{\cdot_Q}\llbracket\lambda\rrbracket}_{m\text{ pairs}}\bigr)(f_1\otimes\ldots\otimes f_m;0,0)\,,
\end{split}
\end{equation*}
for all $f_1,\ldots, f_m\in\mathcal{D}(\mathbb{R}^{d+1})$.
The curly bracket encompasses $m$ pairs of terms of the form $\Psi_{\cdot_Q}\llbracket\lambda\rrbracket\bullet_{\Gamma_{\bullet_Q}^\mathbb{C}}\overline{\Psi}_{\cdot_Q}\llbracket\lambda\rrbracket$, with $m=2m'$ for any $m'\in\mathbb{N}$.
\end{remark}
\subsection{Renormalized equation}
As previously underlined, in order for the functional Equation \eqref{Eq: Functional equation} to account for the stochastic character codified by the complex white noise in Equation \eqref{Eq: Stochastic NLS}, we must deform the underlying algebra. In other words we act on Equation \eqref{Eq: Functional equation} with the deformation map $\Gamma_{\cdot_Q}^\mathbb{C}$, see Theorem \ref{Thm: deformation map cdot}, obtaining
\begin{equation}\label{Eq: Equazione deformata}
\begin{split}
\Psi_{\cdot_Q}\llbracket\lambda\rrbracket&=\Gamma_{\cdot_Q}^\mathbb{C}(\Psi\llbracket\lambda\rrbracket)=\Gamma_{\cdot_Q}^\mathbb{C}(\Phi+\lambda G\circledast \overline{\Psi}\Psi^2)\\
&=\Phi +\lambda G\circledast(\overline{\Psi}_{\cdot_Q}\cdot_Q\Psi_{\cdot_Q}\cdot_Q\Psi_{\cdot_Q})\,,
\end{split}
\end{equation}
where $\Psi_{\cdot_Q}=\Gamma_{\cdot_Q}^\mathbb{C}(\Psi)$ and $\overline{\Psi}_{\cdot_Q}=\Gamma_{\cdot_Q}^\mathbb{C}(\overline{\Psi})$, while $\cdot_Q$ is defined in Equation \eqref{Eq: prodotto deformato2}.
Yet, the presence of this deformed product makes quite cumbersome working with Equation \eqref{Eq: Equazione deformata}. Hence it is desirable to address whether one can rewrite such equation in an equivalent form in which only the pointwise product enters the game, namely we seek for the so called \textit{renormalized equation} \cite{Dappiaggi:2020gge}. The price to pay is the occurrence of new contributions to the equation known in the physics literature as \textit{counter-terms}.
These are a by product both of the deformation procedure (and hence of the stochastic properties of the solution) and of the renormalization one.
We now present the fundamental result of this section, which is an extension of \cite[Prop. 6.6]{Dappiaggi:2020gge}.
\begin{theorem}\label{Thm: Equazione Rinormalizzata}
Let us denote with $\Psi\llbracket\lambda\rrbracket\in\mathcal{A}^\mathbb{C}$ a perturbative solution of Equation \eqref{Eq: Functional equation} and with $\Psi_{\cdot_Q}=\Gamma_{\cdot_Q}^\mathbb{C}(\Psi)$ its counterpart lying in $\mathcal{A}^\mathbb{C}_{\cdot_Q}$. Then there exists a sequence of linear operator-valued functionals $\{M_k\}_{k\geq1}$ such that
\begin{enumerate}
\item \label{punto1} for all $k\geq1$ and for all $\eta,\eta^\prime\in\mathcal{E}(\mathbb{R}^{d+1})$, it holds:
\begin{equation}\label{eq_rinormalizzatac}
M_k(\eta,\eta^\prime):\mathcal{E}(\mathbb{R}^{d+1})\times\mathcal{E}(\mathbb{R}^{d+1})\rightarrow\mathcal{E}(\mathbb{R}^{d+1})\times\mathcal{E}(\mathbb{R}^{d+1}).
\end{equation}
\item \label{punto2} Every element of the sequence has a polynomial dependence on $\eta,\eta^\prime$ and, for all $i,j,k\in\mathbb{N}$ such that $i+j=2k+1$, $M_k^{(i,j)}(0)=0$.
\end{enumerate}
In addition, calling $M:=\sum_{k\geq1}\lambda^kM_k$, $\Psi_{\cdot_Q}$ is a solution of
\begin{equation}\label{Eq: Equazione rinormalizzata}
\Psi_{\cdot_Q}=\Phi+\lambda G\circledast\overline{\Psi}_{\cdot_Q}\Psi_{\cdot_Q}^2+G\circledast (M\Psi_{\cdot_Q}).
\end{equation}
\end{theorem}
\begin{proof} Referring to the formal power series expansion
\begin{equation}
M:=\sum_{k\geq1}\lambda^kM_k,
\end{equation}
we proceed inductively with respect to the perturbative order $k$,
providing at every inductive step a suitable candidate $M_k$ satisfying the desired properties. On account of Equation \eqref{Eq: Gammadot on M_1}, at order $k=0$ the solution reads $\Psi_{\cdot_Q}=\Phi+\mathcal{O}(\lambda)$ and no correction is required. We can use the expression of $F_1$ in Equation \eqref{Eq: Espressione coefficienti} to compute $M$ at first order, namely
\begin{equation*}
\begin{split}
\mathrm{LHS}&=\Gamma_{\cdot_Q}^\mathbb{C}(\Phi+\lambda G\circledast\overline{\Phi}{\Phi}^2)+\mathcal{O}(\lambda^2)\\
&=\Phi+\lambda G\circledast\overline{\Phi}\Phi^2+2\lambda G\circledast\overline{C}\Phi+\mathcal{O}(\lambda^2)\,,
\end{split}
\end{equation*}
\begin{equation*}
\begin{split}
\mathrm{RHS}&=\Phi+\lambda G\circledast\overline{\Phi}\Phi^2+\lambda G\circledast M_1\Phi+\mathcal{O}(\lambda^2)\,,
\end{split}
\end{equation*}
where $\overline{C}\in\mathcal{E}(\mathbb{R}^{d+1})$ has been defined in the proof of Theorem \ref{Thm: deformation map cdot} and where with $\mathrm{LHS}$ and $\mathrm{RHS}$ we denote the left hand side and the right hand side, respectively, of Equation \eqref{Eq: Equazione rinormalizzata}.
Comparing these expressions, we conclude $M_1=2\overline{C}\mathbb{I}$ where $\mathbb{I}$ is the identity operator.
All the properties listed in the statement of the theorem, as well as Equation \eqref{Eq: Equazione rinormalizzata} up to order $\mathcal{O}(\lambda^2)$ are fulfilled as one can infer by direct inspection.
To complete the induction procedure we assume that $M_{k'}$ has been defined for all $k'\leq k-1$, $k\in\mathbb{N}$ so to satisfy all properties listed in the statement and so that Equation \eqref{Eq: Equazione rinormalizzata} holds true up to order $\mathcal{O}(\lambda^{k})$.
We show how one can individuate $M_k$ sharing the same properties up to order $\mathcal{O}(\lambda^{k+1})$.
Mimicking the procedure adopted for $k=1$ it descends
\begin{equation*}
\begin{split}
\mathrm{LHS}=T_{k-1}&+\lambda^k\sum_{k_1+k_2+k_3=k-1}G\circledast\Gamma_{\cdot_Q}^\mathbb{C}(\overline{F}_{k_1}F_{k_2}F_{k_3})+\mathcal{O}(\lambda^{k+1})\,,
\end{split}
\end{equation*}
\begin{equation*}
\begin{split}
\mathrm{RHS}=T_{k-1}&+\lambda^k\sum_{k_1+k_2+k_3=k-1}G\circledast\bigl[\Gamma_{\cdot_Q}^\mathbb{C}(\overline{F}_{k_1})\Gamma_{\cdot_Q}(F_{k_2})\Gamma_{\cdot_Q}^\mathbb{C}(F_{k_3})\bigr]\\
&+\lambda^k\sum\limits_{\substack{k_1+k_2=k\\k_2\neq0}}G\circledast(M_{k_1}\Gamma_{\cdot_Q}^\mathbb{C}(F_{k_2}))+\lambda^kG\circledast(M_k\Phi)+\mathcal{O}(\lambda^{k+1})\,.
\end{split}
\end{equation*}
Here $T_{k-1}$ summarizes all contributions up to order $\mathcal{O}(\lambda^k)$ and thanks to the inductive hypothesis it satisfies Equation \eqref{Eq: Equazione rinormalizzata}. Thus it does not contribute to the construction of $M_k$.
In the proof of Theorem \ref{Thm: valor medio nullo} we introduced the vector space $\mathcal{O}\subset\mathcal{A}^\mathbb{C}$ of distributions $u\in\mathcal{A}^\mathbb{C}$ having odd polynomial degree in $\Phi,\overline{\Phi}\in\mathcal{A}^\mathbb{C}$.
Since the inductive hypothesis entails that $M_l$ has an even polynomial degree in both $\Phi$ and $\overline{\Phi}$ for all $l\leq k-1$, it follows that, since $\Gamma_{\cdot_Q}^\mathbb{C}(F_{k_2})\in\mathcal{O}$, then $M_{k_1}\Gamma_{\cdot_Q}^\mathbb{C}(F_{k_2})\in\mathcal{O}$.
We can further characterize the perturbative coefficients as particular elements of the vector space $\mathcal{O}$: The specific form of the nonlinear potential $V(\psi,\overline{\psi})=\overline{\psi}\psi^2$ implies that every $F_l$, $l\in\mathbb{N}$ depends precisely on $m$ fields $\overline{\Phi}$ and on $m+1$ fields $\Phi$, with $m\in\mathbb{N}$ such that $m\leq l$. This characterization entails that there is always an unpaired field $\Phi$. Note also that this property is preserved by the action of $\Gamma_{\cdot_Q}^\mathbb{C}$ since contractions collapse pairs of $\Phi$ and $\overline{\Phi}$. We conclude that, besides the term involving $M_k$, all other contributions in both sides of Equation \eqref{Eq: Equazione rinormalizzata} are of the form $P\circledast u$, $u\in\mathcal{O}$. Moreover, $u$ can be written as $u=K\Phi$ with $K$ a linear operator abiding to all the properties listed in the statement of the theorem for the renormalization counterterms. Thus we can write
\begin{equation*}
\begin{split}
\Psi_{\cdot_Q}-\Phi-&\lambda G\circledast\Psi_{\cdot_Q}+\lambda G\circledast\overline{\Psi}_{\cdot_Q}\Psi_{\cdot_Q}^2-G\circledast (M\Psi_{\cdot_Q})\\
&=\lambda^kG\circledast\bigl[K-M_k\bigr]\Phi+\mathcal{O}(\lambda^{k+1}).
\end{split}
\end{equation*}
If we set $M_k=K$, it suffices to observe that Equation \eqref{Eq: Equazione rinormalizzata} holds true up to order $\mathcal{O}(\lambda^{k})$. This concludes the induction procedure and the proof.
\end{proof}
\begin{remark}
We refer to \cite[Rem. 6.7]{Dappiaggi:2020gge} for the explicit computation of the coefficient $M_2$ in the real scenario.
The complex case can be discussed similarly.
\end{remark}
\begin{remark}
One may wonder why Equation \eqref{Eq: Equazione rinormalizzata} does not encompass a correction involving $\overline{\Psi}_{\cdot_Q}$, namely a term of the form $G\circledast(N\overline{\Psi}_{\cdot_Q})$ with $N$ the same defining properties of $M$. Although, a priori such kind of contribution should be taken into account, Theorem \ref{Thm: Equazione Rinormalizzata} entails that such additional term is not necessary. Furthermore, our renormalization procedure makes a direct contact with analogous results within other non-perturbative frameworks, such as the theory of regularity structures and of paracontrolled calculus. More precisely, in full analogy with \cite{Hoshino}, only correction terms linear in $\Psi$ should be expected.
\end{remark}
\subsection{A graphical approach to renormalizability}\label{Sec: graph}
As we have seen in the previous sections, the construction of the algebra $\mathcal{A}^\mathbb{C}_{\cdot_Q}$, which is a necessary step in order to compute at any order in perturbation theory the expectation value of the perturbative solution, requires renormalization in order to be meaningful.
As a matter of fact, in the construction of the perturbative solution only some elements of the algebra $\mathcal{A}^\mathbb{C}_{\cdot_Q}$ are involved.
The question we address in this section is whether we can find a condition under which the number of ill-defined contributions to be renormalized in order to compute the expectation value of the solution remains finite to all orders in $\lambda$.
As a consequence, such a condition would tell whether one needs a finite or infinite number of independent renormalization steps in order to compute the expectation value of the solution at any order in perturbation theory.
This is tantamount to classifying a model as \emph{super-renormalizable} or, equivalently, \emph{sub-critical}.
It is useful to tackle such problem using a graphical representation, taking advantage of an analogous formulation for the stochastic $\varphi^3_d$ model analyzed in \cite{Dappiaggi:2020gge}. We can associate to every perturbative contribution a graph via the following prescription:
\begin{itemize}
\item the occurrence of the field $\Phi\in\mathcal{A}^\mathbb{C}$ is represented by the symbol $\fiammifero$, while the complex conjugate field $\overline{\Phi}$ corresponds to $\fiammiferoCC$.
\item The convolution with $G$ and $\overline{G}$ is denoted by segments $\propagatore$ and $\propagatoreCC$ , respectively.
\item The graphical analogue of the pointwise product between elements of $\mathcal{A}^\mathbb{C}$ consists of joining the roots of the respective graphs into a single vertex.
\end{itemize}
\begin{example}
In order to better grasp the notation, consider the following two examples:
\begin{equation*}
\Phi\overline{\Phi}=
\begin{tikzpicture}[thick,scale=1.2]
\draw[red] (0,0) -- (0.2,0.3);
\filldraw[red] (0.2,0.3)circle (1pt);
\draw (0,0) -- (-0.2,0.3);
\filldraw (-0.2,0.3)circle (1pt);
\end{tikzpicture},
\hspace{3cm}
G\circledast\Phi\overline{\Phi}=
\begin{tikzpicture}[thick,scale=1.2]
\draw[red] (0,0) -- (0.2,0.3);
\filldraw[red] (0.2,0.3)circle (1pt);
\draw (0,0) -- (-0.2,0.3);
\filldraw (-0.2,0.3)circle (1pt);
\draw (0,0) -- (0,-0.3);
\end{tikzpicture}.
\end{equation*}
\end{example}
The rules outlined above encompass all the algebraic structures required to represent the perturbative coefficients, but the need of renormalization occurs when acting with the deformation map $\Gamma_{\cdot_Q}^\mathbb{C}$. Its evaluation against elements lying in $\mathcal{A}^\mathbb{C}$ has been thoroughly studied in Section \ref{Sec: expectations}. The net effect amounts to a sequence of contractions between $\Phi$ and $\overline{\Phi}$, related to the occurrence of powers of $Q$ and $\overline{Q}$. Hence at the graphical level it translates into progressively collapsing pairs of leaves into a single loop.
Observe how, in view of the stochastic properties of the complex white noise, see Equation \eqref{Eq: complex white noise}, only contractions between leaves of different colours are allowed.
\begin{example}
The graphical counterpart of $\Gamma_{\cdot_Q}(\Phi\overline{\Phi})(f;\eta,\overline{\eta})=\Phi\overline{\Phi}(f;\eta,\overline{\eta})+C(f)$, $f\in\mathcal{D}(\mathbb{R}^{d+1}),\eta\in\mathcal{E}(\mathbb{R}^{d+1})$, is
\begin{equation*}
\Gamma_{\cdot_Q}(\Phi\overline{\Phi})=
\begin{tikzpicture}[thick,scale=1.2]
\draw[red] (0,0) -- (0.2,0.3);
\filldraw[red] (0.2,0.3)circle (1pt);
\draw (0,0) -- (-0.2,0.3);
\filldraw (-0.2,0.3)circle (1pt);
\end{tikzpicture}
+\fish.
\end{equation*}
\end{example}
The next step consists of identifying all possible graphs appearing in the perturbative decomposition of a solution. From a direct inspection of the coefficients $F_k$, $k\in\mathbb{N}$ one can prove inductively that all such graphs have a tree structure with branches of the form
\begin{equation*}
\begin{tikzpicture}[thick,scale=1.2]
\draw (0,0) -- (0,0.3);
\draw[red] (0,0.3) -- (0.3,0.6);
\filldraw[red] (0.3,0.6)circle (1pt);
\draw (0,0.3) -- (0,0.7);
\filldraw (0,0.7)circle (1pt);
\draw (0,0.3) -- (-0.3,0.6);
\filldraw (-0.3,0.6)circle (1pt);
\end{tikzpicture}
\hspace{1cm}\text{or}\hspace{1cm}
\begin{tikzpicture}[thick,scale=1.2]
\draw[red] (0,0) -- (0,0.3);
\draw[red] (0,0.3) -- (0.3,0.6);
\filldraw[red] (0.3,0.6)circle (1pt);
\draw[red] (0,0.3) -- (0,0.7);
\filldraw[red] (0,0.7)circle (1pt);
\draw (0,0.3) -- (-0.3,0.6);
\filldraw (-0.3,0.6)circle (1pt);
\end{tikzpicture}.
\end{equation*}
Furthermore, increasing by one the perturbative order in $\lambda$ translates into adding a vertex with this topology. Such result entails a constraint on the class of graphs involved in the construction of a solution and it allows us to perform an analysis with respect to the perturbative order. This differs from the approach adopted in \cite{Dappiaggi:2020gge}, where the authors consider a larger class, including also graphs which do not enter the construction. As a results we improve the bound on the renormalizability condition.
\begin{remark}
Non contracted leaves correspond, at the level of distributions, to the multiplication by a smooth function $\eta\in\mathcal{E}(\mathbb{R}^{d+1})$, which does not contribute to the divergence of the graph. Hence we restrict ourselves to maximally contracted graphs, being aware that adding uncontracted leaves does not affect the degree of divergence.
\end{remark}
\begin{definition}
A graph is said to be admissible if it derives from a graph associated to any perturbative coefficient $F_k$ via a maximal contraction of its leaves.
\end{definition}
Following the prescription presented above, we construct a graph $\mathcal{G}$ and an associated distribution $u_{\mathcal{G}}$ completely characterized by the following features:
\begin{itemize}
\item $\mathcal{G}$ has $L$ edges and $N$ vertices of valence $1$ or $4$. Here the valence of a vertex is the number of edges connected to it.
\item Each edge $e$ of the graph is the pictorial representation of a propagator $G(x_{s(e)},x_{t(e)})$, where $s(e)$ stands for the origin of $e$, while with $t(e)$ we denote the target,
\item the integral kernel of $u_{\mathcal{G}}$ is a product of propagators:
\begin{equation}\label{eq_propagatori}
u_{\mathcal{G}}(x_1,\dots,x_M):=\prod_{e\in E_{\mathcal{G}}}G(x_{s(e)},x_{t(e)}),
\end{equation}
where $E_{\mathcal{G}}=\{e_i, i=1,...,N\}$ represents the set of edges of $\mathcal{G}$.
\end{itemize}
When acting with $\Gamma_{\cdot_Q}^\mathbb{C}$ on elements of the perturbative expansion, we are working with distributions $u\in\mathcal{D}^\prime(U)$, $U\subseteq\mathbb{R}^{N(d+1)}$. As outlined in the proof of Theorem \ref{Thm: deformation map cdot}, the singular support of $u$ corresponds to \clacomment{$\Lambda_t^{N}=\{(t_1,x_1,\ldots,t_N,x_N)\in(\mathbb{R}^{d+1})^N\,\vert\,t_1=\ldots=t_N\}$} and renormalization can be dealt with via microlocal analytical techniques based on the calculation of the scaling degree of $u$ with respect to the submanifold $\Lambda_t^N$ of $\mathbb{R}^{N(d+1)}$. A direct computation yields\\
\clacomment{
\begin{equation}\label{Eq: Degree of divergence G}
\rho_{\text{Diag}_{N(d+1)}}(u_\mathcal{G}):=Ld-2(N-1)\,,
\end{equation}
where $\rho_{\Lambda_t^N}(u_\mathcal{G})$ is the weighted degree of divergence of $u_{\mathcal{G}}$ with respect to $\Lambda_t^N$, while $2(N-1)$ is the weighted co-dimension of $\Lambda_t^N$, see \cite[App. B]{Dappiaggi:2020gge}.}
Our goal is to find an expression of Equation \eqref{Eq: Degree of divergence G} in terms of the perturbative order $k$. We divide the analysis in several steps.\\
\noindent \textbf{Step 1: L(k)}.\hspace{1mm} Since the number of edges is not affected by contractions into a single vertex, at this level we can ignore the action of $\Gamma_{\cdot_Q}$. We state that increasing by one the perturbative order corresponds to adding three edges arranged in one of the following admissible branches:
\begin{equation*}
\begin{tikzpicture}[thick,scale=1.2]
\draw[red] (0,0) -- (0.3,0.3);
\filldraw[red] (0.3,0.3)circle (1pt);
\draw (0,0) -- (0,0.4);
\filldraw (0,0.4)circle (1pt);
\draw (0,0) -- (-0.3,0.3);
\filldraw (-0.3,0.3)circle (1pt);
\end{tikzpicture},
\hspace{1cm}
\begin{tikzpicture}[thick,scale=1.2]
\draw[red] (0,0) -- (0.3,0.3);
\filldraw[red] (0.3,0.3)circle (1pt);
\draw[red] (0,0) -- (0,0.4);
\filldraw[red] (0,0.4)circle (1pt);
\draw (0,0) -- (-0.3,0.3);
\filldraw (-0.3,0.3)circle (1pt);
\end{tikzpicture}\,,
\end{equation*}
that is, the number of vertices of valence $4$ coincides with the perturbative order. This implies that
\begin{equation}\label{Eq: L(k)}
L(k)=3k+1.
\end{equation}
\noindent \textbf{Step 2: N(k)}.\hspace{0.1cm} The value $N(k)$ does not depend on how the branches are arranged. It is manifest that switching a branch from a vertex to another leaves $N(k)$ unchanged. For this reason in what follows we refer to tree-like graphs where all vertices of a layer are saturated before moving to the next one. As an example, we do not consider a graph with the following shape:
\begin{equation*}
\begin{tikzpicture}[thick,scale=1.2]
\draw (0,0) -- (0,0.3);
\draw[red] (0,0.3) -- (0.5,0.8);
\draw[red] (0.5,0.8) -- (0.2,1.2);
\filldraw[red] (0.2,1.2)circle (1pt);
\draw[red] (0.5,0.8) -- (0.5,1.3);
\filldraw[red] (0.5,1.3)circle (1pt);
\draw (0.5,0.8) -- (1,1.4);
\draw (1,1.4) -- (0.7,1.8);
\filldraw (0.7,1.8)circle (1pt);
\draw (1,1.4) -- (1,1.9);
\filldraw (1,1.9)circle (1pt);
\draw[red] (1,1.4) -- (1.3,1.8);
\filldraw[red] (1.3,1.8)circle (1pt);
\draw (0,0.3) -- (0,0.7);
\filldraw (0,0.7)circle (1pt);
\draw (0,0.3) -- (-0.3,0.6);
\filldraw (-0.3,0.6)circle (1pt);
\end{tikzpicture}\,,
\end{equation*}
but rather the simpler
\begin{equation*}
\begin{tikzpicture}[thick,scale=1.2]
\draw (0,0) -- (0,0.3);
\draw[red] (0,0.3) -- (0.5,0.8);
\draw[red] (0.5,0.8) -- (0.2,1.2);
\filldraw[red] (0.2,1.2)circle (1pt);
\draw[red] (0.5,0.8) -- (0.5,1.3);
\filldraw[red] (0.5,1.3)circle (1pt);
\draw (0.5,0.8) -- (0.8,1.2);
\filldraw (0.8,1.2)circle (1pt);
\draw (0,0.3) -- (0,0.7);
\filldraw (0,0.7)circle (1pt);
\draw (0,0.3) -- (-0.5,0.8);
\draw (-0.5,0.8) -- (-0.8,1.2);
\filldraw (-0.8,1.2)circle (1pt);
\draw (-0.5,0.8) -- (-0.5,1.3);
\filldraw (-0.5,1.3)circle (1pt);
\draw[red] (-0.5,0.8) -- (-0.2,1.2);
\filldraw[red] (-0.2,1.2)circle (1pt);
\end{tikzpicture}\,.
\end{equation*}
Thanks to an argument similar to that of the derivation of $L(k)$ and reasoning on these simplified configurations, one can infer that
\begin{equation*}
N(k)=3k+1.
\end{equation*}
Yet, this holds true for non-contracted graphs, while we are looking for configurations showing the highest number of divergences. Since the number of vertices of valence 1 at fixed perturbative order $k$ is $2k+1$ and, in view of the observation that there remains always a single non-contracted leaf, performing all permissible contractions amounts to removing $\frac{1}{2}(2k+1-1)=k$ vertices, namely
\begin{equation}\label{Eq: N(k)}
N(k)=3k+1-k=2k+1.
\end{equation}
\noindent \textbf{Step 3}.\hspace{0.1cm} We can express the perturbative order $k$ in terms of $N$ by inverting Equation \eqref{Eq: N(k)}, namely
\begin{equation}
k=\frac{N-1}{2}\,.
\end{equation}
Since every admissible configuration has an odd number of vertices, $k(N)$ is an integer. Inserting this expression of the perturbative order in Equation \eqref{Eq: L(k)} we obtain
\begin{equation}\label{L(N)}
L(N)=\frac{3}{2}N-\frac{1}{2}\,.
\end{equation}
We are interested in the scenarios when only a finite number of graphs shows a singular behaviour. In the present setting this condition translates into requiring that, for $N$ large enough, the degree of divergence of admissible graphs becomes negative.
\begin{theorem}
If \clacomment{$d=1$}, only a finite number of graphs in the perturbative solution of Equation \eqref{Eq: Functional equation} needs to be renormalized.
\end{theorem}
\begin{proof}
\clacomment{
In view of Equation \eqref{Eq: Degree of divergence G} and of the relations derived in the preceding steps, it descends
\begin{equation}\label{Eq: rho}
\begin{split}
\rho(u_{\mathcal{G}})&=Ld-2(N-1)=\Bigl(\frac{3}{2}N-\frac{1}{2}\Bigr)d-2(N-1)\\
&=N\Bigl(\frac{3}{2}d-2\Bigr)-\frac{d}{2}+2\,.
\end{split}
\end{equation}
For the condition $\rho(u_{\mathcal{G}})<0$ to occur, the coefficient in front of $N$ must be negative, so that for sufficiently large $N$ the inequality holds true.} \clacomment{This translates to the inequality
\begin{equation*}\label{Eq: Subcritical regime}
\frac{3}{2}d-2<0\;\Rightarrow\;d<\frac{4}{3}\,.
\end{equation*}
}
\end{proof}
\begin{remark}
Once the spatial dimension $d$ has been fixed, Equation \eqref{Eq: rho} allows us to find the minimal number of vertices needing renormalization and for which further divergences do not occur.
Subsequently, via Equation \eqref{Eq: N(k)}, we can also infer the maximum number of graphs requiring renormalization.
\end{remark}
\begin{remark}
Even if in our analysis we focused on the stochastic Schr\"odinger equation with a cubic non-linearity, this approach to the study of the subcritical regime can be extended to a more general polynomial potential of the form $\vert\psi\vert^{2\kappa}\psi$ with $\kappa>1$, as in Equation \eqref{Eq: Stochastic NLS}. At a graphical level the net effect is that admissible graphs present vertices of valence at most $2\kappa+2$. Hence the derivation of Equation \eqref{Eq: Subcritical regime} follows slavishly the preceding steps, yielding a condition on the dimension of the underlying space depending on $\kappa$.
\end{remark}
\paragraph{Acknowledgements.}
We are thankful to N. Drago and V. Moretti for helpful discussions. A.B. is supported by a PhD fellowship of the University of Pavia, while P.R by a postdoc fellowship of the Institute for Applied Mathematics of the University of Bonn.
|
\section*{#1
\else
\vspace{.05in}\footnotesize
\parindent .2in
{\upshape\bfseries #1. }\ignorespaces
\fi}
{\if@twocolumn\else\par\vspace{.1in}\fi}
\def\supess{\mathop{\rm ess\: sup }}
\begin{document}
\title{\sc Relaxation for an optimal design problem in $BD(\O)$}
\author{{\sc Ana Cristina Barroso}\\
Departamento de Matem\'atica and CMAFcIO \\
Faculdade de Ci\^encias da Universidade de Lisboa\\
Campo Grande, Edif\' \i cio C6, Piso 1\\
1749-016 Lisboa, Portugal\\
[email protected]\\
\\
{\sc Jos\'e Matias}\\
Departamento de Matem\'{a}tica and CAMGSD\\
Instituto Suprior T\'{e}cnico\\
Av. Rovisco Pais 1\\
1049-001 Lisboa, Portugal\\
[email protected]\\
\\
and
\\
\\
{\sc Elvira Zappale} \\
Dipartimento di Scienze di Base e Applicate per l'Ingegneria\\
Sapienza - Universit\`a di Roma\\
Via Antonio Scarpa, 16\\
00161 Roma (RM), Italy\\
<EMAIL>}
\maketitle
\begin{abstract}
We obtain a measure representation for a functional arising in the context of optimal design problems under linear growth conditions. The functional in question corresponds to the relaxation with respect to a pair $(\chi,u)$, where $\chi$ is the characteristic function of a set of finite perimeter and $u$ is a function of bounded deformation, of an energy with a bulk term depending on the symmetrised gradient as well as a perimeter term.
\end{abstract}
\smallskip
{\bf MSC (2020)}: 49J45, 49Q10
{\bf Keywords}: (special) fields of bounded deformation, optimal design, sets of finite perimeter, symmetric quasiconvexity
\section{Introduction}
In optimal design one aims to find an optimal shape which minimises a cost functional. The optimal shape is a subset $E$ of a bounded, open set
$\Omega \subset \mathbb R^N$ which is described by its characteristic function
$\chi: \Omega \to \{0,1\}$, $E = \{\chi= 1\}$, and, in the linear elasticity framework, the cost functional is usually a quadratic energy, so we are lead to the problem
\begin{equation}\label{orpb}
\displaystyle \min_{(\chi,u)} \int_{\Omega}\chi(x) W_1({\cal E} u(x))
+ (1 - \chi(x))W_0(\E u(x)) \, dx,
\end{equation}
where $W_0$ and $W_1$ are two {\em elastic} densities, with $W_0 \geq W_1$, and $\E u$ denotes the symmetrised gradient of the displacement $u$. We refer to the seminal papers \cite{AL, KS1, KS2, KS3, MT}, among a wide literature (see, for instance, the recent contributions \cite{BIR} and \cite{BIR2}).
However, as soon as plasticity comes into play, the
observed stress-strain relation is no longer linear and, due to the linear growth of the stored elastic energy and to the lack of reflexivity of the space $L^1$, a suitable functional space is necessary to account for fields $u$ whose strains are measures. The space of special fields with bounded deformation, $BD(\Omega)$, was first proposed in \cite{S1}-\cite{S4}
and starting from these pioneering papers a vast literature developed.
Indeed, already in the case where $\chi\equiv \chi_\Omega$,
the search for equilibria in the context of perfect plasticity leads naturally to the study of lower semicontinuity properties, and eventually relaxation, for energies of the type
\begin{equation}\label{calE}
\int_{\Omega}f(\E u(x)) \, dx
\end{equation}
where
$f$ is the volume energy density. As mentioned above, $u$ belongs to the space $BD(\Omega)$ of functions of bounded deformation composed of integrable vector-valued functions for which all components $E_{ij}$, $i,j = 1, \ldots,N$, of the deformation tensor $Eu := \frac{Du + Du^T}{2}$ are bounded Radon measures and
$\E u$ stands for the absolutely continuous part, with respect to the Lebesgue measure, of the symmetrised distributional derivative $Eu$.
Lower semicontinuity for \eqref{calE} was established in \cite{BCDM} under convexity assumptions on $f$ and in \cite{E} for symmetric quasiconvex integrands, under linear growth conditions and for $u \in LD(\Omega)$, the subspace of $BD(\Omega)$ comprised of functions for which the singular part $E^su$ of the measure $Eu$ vanishes. For a symmetric quasiconvex density $f$ with an explicit dependence on the position in the body and satisfying superlinear growth assumptions, lower semicontinuity properties were established in \cite{E2} for $u \in SBD(\Omega)$.
In the case where the energy density takes the form $\|\E u\|^2$ or
$\|\E^D u\|^2 + (\text{div } u)^2$ (where $A^D$ stands for the deviator of the
$N \times N$ matrix $A$ given by $A^D := A - \frac{1}{N}\text{tr} (A)I$), and the total energy also includes a surface term, a first relaxation result was proved in \cite{BDV}. We also refer to \cite{KR} for the relaxation in the case where there is no surface energy and to \cite{Mo}, \cite{JS}, and \cite{MMS} for related models concerning evolutions and homogenization, among a wider list of contributions.
For general energy densities $f$, Barroso, Fonseca $\&$ Toader \cite{BFT} studied the relaxation of \eqref{calE} for $u \in SBD(\Omega)$ under linear growth conditions but placing no convexity assumptions on $f$. They showed that the relaxed functional admits an integral representation where a surface energy term arises naturally. The global method for relaxation due to Bouchitté, Fonseca $\&$ Mascarenhas \cite{BFM} was used to characterise the density of this term, whereas the identification of the relaxed bulk energy term relied on the blow-up method \cite{FM1} together with a Poincaré-type inequality.
Ebobisse $\&$ Toader \cite{ET} obtained an integral representation result for general local functionals defined in $SBD(\Omega)$ which are lower semicontinous with respect to the $L^1$ topology and satisfy linear growth and coercivity conditions. The functionals under consideration are restrictions of Radon measures and are assumed to be invariant with respect to rigid motions. Their work was extended to the space $SBD^p(\Omega)$, $p > 1$, which arises in connection with the study of fracture and damage models, by Conti, Focardi $\&$ Iurlano \cite{CFI} in the 2-dimensional setting. A crucial and novel ingredient of their proof is the construction of a $W^{1,p}$ approximation of an $SBD^p$ function $u$ using finite-elements on a countable mesh which is chosen according to $u$ (recall that $SBD^p$ denotes the space of fields with bounded deformation such that the symmetrised gradient is the sum of an $L^p$ field and a measure supported on a set of finite $\mathcal H^{N-1}$ measure).
The analysis of an integral representation for a variational functional satisfying lower semicontinuity, linear growth conditions and the usual measure theoretical properties, was extended to the full space $BD(\Omega)$ by Caroccia, Focardi $\&$ Van Goethem \cite{CFVG}. In this work, the invariance of the studied functional with respect to rigid motions, required in \cite{ET}, is replaced by a weaker condition stating continuity with respect to infinitesimal rigid motions. Their result relies, as in papers mentioned above, on the global method for relaxation, as well as on the characterisation of the Cantor part of the measure $Eu$, due to De Philippis $\&$ Rindler \cite{DPR0}, which extends to the $BD$ case the result of Alberti's rank-one theorem in $BV$.
In the study of the minimisation problem \eqref{orpb} one usually prescribes the volume fraction of the optimal shape, leading to a constraint of the form
$\displaystyle \frac{1}{{\mathcal L}^N(\Omega)}\int_\Omega \chi(x) dx= \theta, \;\theta \in (0,1)$. It is sometimes convenient to replace this constraint
by inserting, instead, a Lagrange multiplier in the modelling functional
which, in the optimal design context, becomes
\begin{equation}\label{functnoper}F(\chi,u;\Omega) : =
\displaystyle \int_{\Omega}\chi(x) W_1({\cal E} u(x))
+ (1 - \chi(x))W_0(\E u(x)) + \int_\Omega k \chi(x) dx.
\end{equation}
Despite the fact that we have compactness for $u$ in $BD(\Omega)$ for functionals of the form \eqref{functnoper}, it is well known that the problem
of minimising \eqref{functnoper} with respect to $(\chi,u)$, adding suitable forces and/or boundary conditions, is ill-posed, in the sense that
minimising sequences $\chi_n \in L^\infty(\Omega;\{0,1\})$ tend to highly oscillate and develop microstructure, so that in the limit we may no longer obtain a characteristic function. To avoid this phenomenon, as in \cite{ABper} and \cite{KL}, we add a perimeter penalisation along the interface between the two zones
$\{\chi = 0\}$ and $\{\chi = 1\}$ (see \cite{CZ} for the analogous analysis performed in $BV$, and \cite{CZ0, BZ, BZ2} for the Sobolev settings, also in the presence of a gap in the growth exponents).
Thus, with an abuse of notation (i.e. denoting $W_1+ k$, in \eqref{functnoper}, still by $W_1$), our aim in this paper is to study the energy functional given by
\begin{equation}\label{Fint}
F(\chi,u;\Omega) : =
\displaystyle \int_{\Omega}\chi(x) W_1({\cal E} u(x))
+ (1 - \chi(x))W_0(\E u(x)) \, dx + |D\chi|(\Omega),
\end{equation}
where $u \in BD(\Omega)$, $\chi \in BV(\Omega;\{0,1\})$ and the densities $W_i$, $i = 0,1$, are continuous functions satisfying the following linear growth conditions from above and below,
\begin{equation}\label{growthint}
\exists \, \alpha, \beta > 0 \text{ such that }
\alpha |\xi| \leq W_i(\xi) \leq \beta (1 + |\xi|), \; \; \forall
\xi \in \Rb^{N \times N}_s.
\end{equation}
We point out that no convexity assumptions are placed on $W_i$, $i = 0,1$.
To simplify the notation, in the sequel, we let
$f: \{0,1\} \times \mathbb R^{N \times N}_s \to [0,+\infty)$
be defined as
\begin{equation}\label{densityint}
f\left( q,\xi\right) :=q W_1(\xi)+ (1-q)W_0(\xi),
\end{equation}
and for a fixed $q \in\{0,1\}$, we recall that the recession function of $f$, in its second argument, is given by
\begin{equation}\label{recS}
f^{\infty}(q,\xi) := \limsup_{t\rightarrow+\infty}\frac{f(q,t\xi)}{t}.
\end{equation}
Since we place no convexity assumptions on $W_i$, we consider
the relaxed localised functionals arising from the energy \eqref{Fint},
defined, for an open subset $A \subset \Omega$, by
\begin{align}\label{calFint}
\mathcal{F}\left(\chi,u;A\right) &
:=\inf\bigg\{ \liminf_{n\rightarrow +\infty} F(\chi_n,u_n;A):
u_{n} \in W^{1,1}(A;\mathbb{R}^{N}),
\chi_{n} \in BV(A;\{0,1\}), \\
& \hspace{4cm}
u_{n}\to u\text{ in }
L^{1}(A;\mathbb{R}^{N}),
\chi_{n}\to\chi\text{ in }
L^1(A;\{0,1\}) \bigg\},\nonumber
\end{align}
and
\begin{align}\label{calFLDint}
\mathcal{F}_{LD}\left(\chi,u;A\right) &
:=\inf\bigg\{ \liminf_{n\rightarrow +\infty} F(\chi_n,u_n;A):
u_{n} \in LD(A), \chi_{n} \in BV(A;\{0,1\}), \\
& \hspace{4cm}
u_{n}\to u\text{ in }
L^{1}(A;\mathbb{R}^{N}),
\chi_{n}\to\chi\text{ in }
L^1(A;\{0,1\}) \bigg\},\nonumber
\end{align}
where $LD(\Omega):=\{u \in BD(\Omega): E^s u=0\}$.
Due to the expression of \eqref{Fint}, and to the fact that
$\chi_n \overset{\ast}{\rightharpoonup}\chi$ in $BV$ if and only if $\{\chi_n\}$ is uniformly bounded in $BV$ and $\chi_n \to \chi$ in $L^1$, it is equivalent to take
$\chi_n \overset{\ast}{\rightharpoonup}\chi$ in $BV$ or $\chi_n \to \chi$ in $L^1$
in the definitions of the functionals \eqref{calFint} and \eqref{calFLDint}, obtaining for each of them the same infimum regardless of the considered convergence.
As a simple consequence of the density of smooth functions in $LD(\Omega)$ we show in Remark~\ref{remkW11LD} that, under the above growth conditions on $W_0, W_1$,
$$\mathcal{F}\left(\chi,u;A\right) = \mathcal{F}_{LD}\left(\chi,u;A\right), \;
\mbox{for every} \; \chi \in BV(A;\left\{0,1\right\}), u \in BD(\Omega),
A \in {\cal O}(\Omega).$$
We prove in Proposition~\ref{traceRm} that $\mathcal{F}\left(\chi,u;\cdot\right)$ is the restriction to the open subsets of $\Omega$ of a Radon measure, the main result of our paper concerns the characterisation of this measure.
\begin{theorem}\label{main}
Let $f:\{0,1\} \times \mathbb R^{N \times N}_s\to [0, + \infty)$ be a continuous function as in \eqref{densityint}, where $W_0$ and $W_1$ satisfy \eqref{growthint}, and consider
$F:BV(\Omega;\{0,1\})\times BD(\Omega)\times \mathcal O(\Omega)$
defined in \eqref{Fint}. Then
\begin{align}\label{F11}
\mathcal{F}\left( \chi,u;A\right)&=\int_A SQf(\chi(x), \E u(x)) \, dx +
\int_{A \cap (J_\chi \cup J_u)}
g(x,\chi^+(x),\chi^-(x),u^+(x),u^-(x),\nu(x))\, d\cH^{N-1}(x) \nonumber \\
& + \int_A(SQf)^{\infty}(\chi(x), \frac{d E^c u}{d |E^c u|}(x)) \, d|E^c u|(x),
\end{align}
where $SQf$ is the symmetric quasiconvex envelope of $f$ and
$(SQf)^{\infty}$ is its recession function (cf. Subsection~\ref{defqcx} and \eqref{recS}, respectively). The relaxed surface energy density is given by
$$
g(x_0,a,b,c,d,\nu) := \limsup_{\varepsilon \to 0^+}
\frac{m(\chi_{a,b,\nu}(\cdot - x_0),u_{c,d,\nu}(\cdot - x_0);Q_\nu(x_0,\e))}
{\e^{N-1}}
$$
where $Q_{\nu}(x_0,\e)$ stands for an open cube with centre $x_0$, sidelength $\e$ and two of its faces parallel to the unit vector $\nu$,
$$m(\chi,u;V) := \inf\left\{\mathcal F (\theta,v;V) :
\theta \in BV(\O;\{0,1\}), v \in BD(\O), \theta = \chi \mbox{ on } \partial V,
v = u \mbox{ on } \partial V\right\},$$
for any $V$ open subset of $\Omega$ with Lipschitz boundary,
and, for $(a,b,c,d,\nu) \in \{0,1\} \times \{0,1\} \times \mathbb{R}^{N} \times\mathbb{R}^{N} \times S^{N-1},$ the functions $\chi_{a,b,\nu}$ and
$u_{c,d,\nu}$ are defined as
\begin{equation*}
\chi_{a,b,\nu}(y) := \begin{cases}
a, & {\rm if } \; y \cdot \nu > 0\\
b, & {\rm if } \; y \cdot \nu < 0
\end{cases}
\; \; \; {\rm and } \; \; \;
u_{c,d,\nu}(y) := \begin{cases}
c, & {\rm if } \; y \cdot \nu > 0\\
d, & {\rm if } \; y \cdot \nu < 0.
\end{cases}
\end{equation*}
\end{theorem}
For the notation regarding the jump sets $J_\chi$, $J_u$ and the corresponding vectors
$\chi^+(x)$, $\chi^-(x)$, $\nu_\chi(x)$, $u^+(x)$, $u^-(x)$ and $\nu_u(x)$ we refer to Subsections~\ref{BV}, \ref{BD} and \ref{surface}.
The above expression for the relaxed surface energy density arises as an application of the global method for relaxation \cite{BFM}. However, as we will see in Subsection~\ref{surface}, in the case where $f$ satisfies the additional hypothesis \eqref{finfty}, this density can be described more explicitly, leading to an integral representation for \eqref{calFint}, in the $BD$ setting, entirely similar to the one in $BV$, obtained in \cite{CZ}, when $W_0$ and $W_1$ depend on the whole gradient $\nabla u$. Indeed, under this assumption, we show that
$$g(x_0,a,b,c,d,\nu) = K(a,b,c,d,\nu)$$
where
\begin{equation}\label{KSQ}
K(a,b,c,d,\nu):=\inf\left\{
\displaystyle\int_{Q_{\nu}}
(SQf)^{\infty}(\chi(x),\E u(x)) \, dx+|D\chi|(Q_{\nu}):\left(\chi,u\right)
\in\mathcal{A}(a,b,c,d,\nu)\right\},
\end{equation}
and, for $(a,b,c,d,\nu) \in \{0,1\} \times \{0,1\} \times \mathbb{R}^{N} \times\mathbb{R}^{N} \times S^{N-1},$ the set of admissible functions is
\begin{align*}
\mathcal{A}(a,b,c,d,\nu) & :=\bigg\{ \left(\chi,u\right)
\in BV_{\rm loc}\left(S_{\nu};\{0,1\}\right) \times
W^{1,1}_{\rm loc}\left(S_{\nu};\mathbb{R}^{N}\right) : \nonumber \\
& (\chi(y),u(y)) = (a,c) \text{ if } y\cdot\nu=\frac{1}{2},
~(\chi(y),u(y)) = (b,d) \text{ if } y\cdot\nu=-\frac{1}{2},\nonumber\\
& (\chi, u)\text{ are 1-periodic in the directions of }\nu_{1},\dots,\nu_{N-1}\bigg\},
\end{align*}
$\left\{\nu_{1},\nu_{2},\dots,\nu_{N-1},\nu\right\}$ is an orthonormal
basis of $\mathbb{R}^{N}$ and $S_\nu$ is the strip given by
$$S_\nu = \left\{x \in \Rb^N : |x \cdot \nu| < \frac{1}{2}\right\}.$$
As an application of the result of Caroccia, Focardi $\&$ Van Goethem, obtained in the abstract variational functional setting in \cite{CFVG}, the authors proved an integral representation for the relaxed functional, defined in $BD(\Omega) \times \mathcal O(\Omega)$,
$$
\mathcal F_{0}(u; A) := \inf \Big \{\liminf_{n \to + \infty}
F_0(u_n;A) : u_n \in W^{1, 1}(A; \Rb^N), u_n \to u \; {\rm in} \; L^1(A;\Rb^N)
\Big \},$$
where
$$F_0(u;A) :=\begin{cases}
\displaystyle \int_A f_{0} (x,u(x),\E u_n(x)) \, dx, & {\rm if} \; u \in W^{1,1}(\Omega;\Rb^N)\\
+\infty, & {\rm otherwise}
\end{cases}$$
and the density $f_0$ satisfies linear growth conditions from above and below
$$\frac{1}{C}|A| \leq f_0(x,u,A) \leq C\left(1 + |A|\right), \; \forall (x,u,A) \in
\Omega \times \Rb^N \times \Rb^{N \times N}_s$$
as well as a continuity condition with respect to $(x,u)$. This generalises to the full space $BD(\Omega)$, and to the case of densities $f_0$ depending explicitly on $(x,u)$, the results obtained in \cite{BFT}. We will make use of their work in Subsection~\ref{Cantor} to prove both lower and upper bounds for the density of the Cantor part of the measure
$\mathcal F(\chi,u;\cdot)$, by means of an argument based on Chacon's Biting Lemma which allows us to fix $\chi$ at an appropriately chosen point $x_0$, as in \cite{MMZ}.
The contents of this paper are organised as follows. In Section~\ref{prelim}
we fix our notation and provide some results pertaining to $BV$ and $BD$ functions and notions of quasiconvexity
which will be used in the sequel.
Section~\ref{auxres} contains some auxiliary results which are needed to prove our main theorem. In particular, in Proposition~\ref{traceRm} we show that $\mathcal F(\chi,u;\cdot)$ is the restriction to the open subsets of $\Omega$ of a Radon measure $\mu$. Section~\ref{mainthm} is dedicated to the proof of our main theorem, which characterises this measure. In each of Subsections~\ref{bulk}, \ref{Cantor} and \ref{surface} we prove lower and upper bounds of the densities of $\mu$ with respect to the bulk and Cantor parts of $Eu$, as well as with respect to a surface measure which is concentrated on the union of the jump sets of $\chi$ and $u$.
The fact that our functionals have an explicit dependence on the $\chi$ field
prevented us from applying existing results (such as \cite{ARDPR} and \cite{BDG}) directly and required us to obtain direct proofs.
\section{Preliminaries}\label{prelim}
In this section we fix notations and quote some definitions and
results that will be used in the sequel.
Throughout the text $\Omega \subset \mathbb R^{N}$ will denote an open, bounded set
with Lipschitz boundary.
We will use the following notations:
\begin{itemize}
\item ${\mathcal B}(\Omega)$, ${\mathcal O}(\Omega)$ and ${\mathcal O}_{\infty}(\Omega)$ represent the families of all Borel, open and open subsets of $\Omega$ with Lipschitz boundary, respectively;
\item $\mathcal M (\Omega)$ is the set of finite Radon
measures on $\Omega$;
\item $\left |\mu \right |$ stands for the total variation of a measure $\mu\in \mathcal M (\Omega)$;
\item $\mathcal L^{N}$ and $\mathcal H^{N-1}$ stand for the $N$-dimensional Lebesgue measure
and the $\left( N-1\right)$-dimensional Hausdorff measure in $\mathbb R^N$, respectively;
\item the symbol $d x$ will also be used to denote integration with respect to $\mathcal L^{N}$;
\item the set of symmetric $N \times N$ matrices is denoted by $\Rb_s^{N\times N}$;
\item given two vectors $a, b \in \Rb^N$, $a \odot b$ is the symmetric $N \times N$ matrix defined by $a \odot b : = \dfrac{a \otimes b + b \otimes a}{2}$, where $\otimes$ indicates tensor product;
\item $B(x, \e)$ is the open ball in $\Rb^N$ with centre $x$ and radius $\e$,
$Q(x,\e)$ is the open cube in $\Rb^N$ with two of its faces parallel to the unit vector $e_N$, centre $x$ and sidelength $\e$, whereas
$Q_{\nu}(x,\e)$ stands for a cube with two of its faces parallel to the unit vector $\nu$; when $x=0$ and $\e = 1$, $\nu=e_N$ we simply write $B$ and $Q$;
\item $S^{N-1} := \partial B$ is the unit sphere in $\Rb^N$;
\item $C_c^{\infty}(\Omega; \Rb^N)$ and $C_{\rm per}^{\infty}(\Omega; \Rb^N)$ are the spaces of $\Rb^N$-valued smooth functions with compact support in $\Omega$ and
smooth and $Q$-periodic functions from $Q$ to $\Rb^N$, respectively;
\item by
$\displaystyle \lim_{\delta, n}$ we mean
$\displaystyle \lim_{\delta \to 0^+} \lim_{n\to +\infty}$,
$\displaystyle \lim_{k, n}$ means
$\displaystyle \lim_{k \to +\infty} \lim_{n\to +\infty}$;
\item $C$ represents a generic positive constant that may change from line to line.
\end{itemize}
\subsection{BV Functions and Sets of Finite Perimeter}\label{BV}
In the following we give some preliminary notions regarding functions of bounded variation and sets of finite perimeter. For a detailed treatment we refer to \cite{AFP}.
\smallskip
Given $u \in L^1(\Omega; \Rb^d)$ we let $\O _u$ be the set of Lebesgue points of $u$,
i.e., $x\in \O _u$ if there exists $\widetilde u(x)\in {\mathbb{R}}^d$ such
that
\begin{equation*}
\lim_{\varepsilon\to 0^+} \frac{1}{\e^N}\int_{B(x,\varepsilon)}
|u(y)-\widetilde u(x)|\,dy=0,
\end{equation*}
$\widetilde u(x)$ is called the approximate limit of $u$ at $x$.
The Lebesgue discontinuity set $S_u$ of $u$ is defined as
$S_u := \Omega \setminus \O _u$. It is known that ${\mathcal{L}}^{N}(S_u) = 0$ and the function
$x \in \Omega \mapsto \widetilde u(x)$, which coincides with $u$ $\L ^N$- a.e.
in $\O _u$, is called the Lebesgue representative of $u$.
The jump set of the function $u$, denoted by $J_u$, is the set of
points $x\in \O \setminus \O _u$ for which there exist
$a, \,b\in {\mathbb{R}}^d$ and a unit vector $\nu \in S^{N-1}$, normal to $J_u$ at $x$, such that $a\neq b$ and
\begin{equation*}
\lim_{\varepsilon \to 0^+} \frac {1}{\varepsilon^N} \int_{\{ y \in
B(x,\varepsilon) : (y-x)\cdot\nu > 0 \}} | u(y) - a| \, dy = 0,
\qquad
\lim_{\varepsilon \to 0^+} \frac {1}{\varepsilon^N} \int_{\{ y \in
B(x,\varepsilon) : (y-x)\cdot\nu < 0 \}} | u(y) - b| \, dy = 0.
\end{equation*}
The triple $(a,b,\nu)$ is uniquely determined by the conditions above
up to a permutation of $(a,b)$ and a change of sign of $\nu$
and is denoted by $(u^+ (x),u^- (x),\nu_u (x)).$ The jump of $u$ at $x$ is defined by
$[u](x) : = u^+(x) - u^-(x).$
\smallskip
We recall that a function $u\in L^{1}(\Omega;{\mathbb{R}}^{d})$ is said to be of bounded variation, and we write $u\in BV(\Omega;{\mathbb{R}}^{d})$, if
all its first order distributional derivatives $D_{j}u_{i}$ belong to
$\mathcal{M}(\Omega)$ for $1\leq i\leq d$ and $1\leq j\leq N$.
The matrix-valued measure whose entries are $D_{j}u_{i}$ is denoted by $Du$
and $|Du|$ stands for its total variation.
The space $BV(\O ; {\mathbb{R}}^d)$ is a Banach space when endowed with the norm
\begin{equation*}
\|u\|_{BV(\O ; {\mathbb{R}}^d)} = \|u\|_{L^1(\O ; {\mathbb{R}}^d)} + |Du|(\O )
\end{equation*}
and we observe that if $u\in BV(\Omega;\mathbb{R}^{d})$ then $u\mapsto|Du|(\Omega)$ is lower semicontinuous in $BV(\Omega;\mathbb{R}^{d})$ with respect to the
$L_{\mathrm{loc}}^{1}(\Omega;\mathbb{R}^{d})$ topology.
By the Lebesgue Decomposition Theorem, $Du$ can be split into the sum of two
mutually singular measures $D^{a}u$ and $D^{s}u$, the absolutely continuous
part and the singular part, respectively, of $Du$ with respect to the
Lebesgue measure $\mathcal{L}^N$. By $\nabla u$ we denote the
Radon-Nikod\'{y}m derivative of $D^{a}u$ with respect to $\mathcal{L}^N$, so that we
can write
\begin{equation*}
Du= \nabla u \mathcal{L}^N \lfloor \O + D^{s}u.
\end{equation*}
If $u \in BV(\O )$ it is well known that $S_u$ is countably $(N-1)$-rectifiable, see \cite{AFP},
and the following decomposition holds
\begin{equation*}
Du= \nabla u \mathcal{L}^N \lfloor \O + [u] \otimes \nu_u {\mathcal{H}}^{N-1}\lfloor S_u + D^cu,
\end{equation*}
\noindent where $D^cu$ is the Cantor part of the
measure $Du$.
If $\Omega$ is an open and bounded set with Lipschitz boundary then the
outer unit normal to $\partial \Omega$ (denoted by $\nu$) exists ${\mathcal{H}}^{N-1}$-a.e. and the trace for functions in $BV(\Omega;{\mathbb{R}}^d)$ is
defined.
\begin{theorem}
(Approximate Differentiability)\label{approxdiff}
If $u \in BV(\Omega; {\mathbb{R}}^d),$ then for $\mathcal{L}^N$-a.e. $x \in\Omega$
\begin{equation}\label{apdif}
\lim_{\varepsilon \rightarrow 0^+} \frac{1}{\varepsilon^{N+1} }
\int_{Q(x, \varepsilon)} |u(y) - u(x) - \nabla u(x).(y-x)|\, dy = 0.
\end{equation}
\end{theorem}
\begin{definition}
\label{Setsoffiniteperimeter} Let $E$ be an $\mathcal{L}^{N}$- measurable
subset of $\mathbb{R}^{N}$. For any open set $\Omega\subset\mathbb{R}^{N}$ the
{\em perimeter} of $E$ in $\Omega$, denoted by $P(E;\Omega)$, is
given by
\begin{equation}
\label{perimeter}P(E;\Omega):=\sup\left\{ \int_{E} \mathrm{div}\varphi(x) \,dx:
\varphi\in C^{1}_{c}(\Omega;\mathbb{R}^{d}), \|\varphi\|_{L^{\infty}
\leq1\right\} .
\end{equation}
We say that $E$ is a {\em set of finite perimeter} in $\Omega$ if $P(E;\Omega) <+
\infty.$
\end{definition}
Recalling that if $\mathcal{L}^{N}(E \cap\Omega)$ is finite, then $\chi_{E}
\in L^{1}(\Omega)$, by \cite[Proposition 3.6]{AFP}, it follows
that $E$ has finite perimeter in $\Omega$ if and only if $\chi_{E} \in
BV(\Omega)$ and $P(E;\Omega)$ coincides with $|D\chi_{E}|(\Omega)$, the total
variation in $\Omega$ of the distributional derivative of $\chi_{E}$.
Moreover, a
generalised Gauss-Green formula holds:
\begin{equation}\nonumber
{\int_{E}\mathrm{div}\varphi(x) \, dx
=\int_{\Omega}\left\langle\nu_{E}(x),\varphi(x)\right\rangle \, d|D\chi_{E}|,
\;\;\forall\,\varphi\in C_{c}^{1}(\Omega;\mathbb{R}^{d})},
\end{equation}
where $D\chi_{E}=\nu_{E}|D\chi_{E}|$ is the polar decomposition of $D\chi_{E}$.
The following approximation result can be found in \cite{BA}.
\begin{lemma}\label{polyhedra}
Let $E$ be a set of finite perimeter in $\Omega$. Then, there exists a sequence of polyhedra $E_n$, with characteristic functions $\chi_n$, such that $\chi_n\to \chi$ in $L^1(\Omega;\{0,1\})$ and $P (E_n;\Omega)\to P(E;\Omega)$.
\end{lemma}
\subsection{BD and LD Functions}\label{BD}
We now recall some facts about functions of bounded deformation. More details can be found in \cite{ACDM, BFT, BCDM, T, TS}.
\smallskip
A function $u\in L^{1}(\Omega;{\mathbb{R}}^{N})$ is said to be of bounded deformation, and we write $u\in BD(\Omega)$, if the symmetric part of its distributional derivative $Du$, $Eu : = \dfrac{Du + Du^T}{2},$ is a matrix-valued bounded Radon measure.
The space $BD(\O)$ is a Banach space when endowed with the norm
\begin{equation*}
\|u\|_{BD(\Omega)} = \|u\|_{L^1(\Omega; \Rb^N)} + |Eu|(\O ).
\end{equation*}
We denote by $LD(\O)$ the subspace of $BD(\O)$ comprised of functions $u$ such that
$Eu \in L^1(\O;\Rb^{N\times N}_s)$, a counterexample due to Ornstein \cite{O} shows that $W^{1,1}(\O;\mathbb R^N) \subsetneq LD(\O)$.
The intermediate topology in the space $BD(\O)$ is the one determined by the distance
$$d(u,v) := \|u-v\|_{L^1(\O;\Rb^N)} + \big| |Eu|(\O) - |Ev |(\O)\big|,
\;u, v \in BD(\O).$$
Hence, a sequence $\{u_n\} \subset BD(\O)$ converges to a function $u \in BD(\O)$ with respect to this topology, written $u_n \stackrel{i}{\to}u$, if and only if,
$u_n \to u$ in $L^1(\O;\Rb^N)$, $Eu_n \wsto Eu$ in the sense of measures and
$|Eu_n|(\O) \to |Eu|(\O)$.
Recall that if $u_n \to u$ in $L^1(\Omega;\Rb^N)$ and there exists $C > 0$ such that $|Eu_n|(\Omega) \leq C, \forall n \in \Nb$, then $u \in BD(\Omega)$ and
\begin{equation}\label{Elsc}
|Eu|(\Omega) \leq \liminf_{n\to +\infty}|Eu_n|(\Omega).
\end{equation}
By the Lebesgue Decomposition Theorem, $Eu$ can be split into the sum of two
mutually singular measures $E^{a}u$ and $E^{s}u$, the absolutely continuous
part and the singular part, respectively, of $Eu$ with respect to the
Lebesgue measure $\mathcal{L}^N$. The
Radon-Nikod\'{y}m derivative of $E^{a}u$ with respect to $\mathcal{L}^N$, is denoted by $\E u$ so we have
\begin{equation*}
Eu= \E u \, \mathcal{L}^N \lfloor \O + E^{s}u.
\end{equation*}
With these notations we may write
$$LD(\Omega):=\{u \in BD(\Omega): E^s u=0\}$$
and (cf. \cite{T}) $LD(\Omega)$ is a Banach space when endowed with the norm
$$\|u\|_{LD(\Omega)} : = \|u\|_{L^1(\Omega;\mathbb R^N)} +
\|{\cal E} u\|_{L^1(\Omega;\mathbb R^N)}.$$
If $\O$ is a bounded, open subset of $\Rb^N$ with Lipschitz boundary $\Gamma$, then
there exists a linear, surjective and continuous, both with respect to the norm and to the intermediate topologies, trace operator
$$\text{tr }: BD(\O) \to L^1(\O;\Rb^N)$$
such that tr $u = u$ if $u \in BD(\O) \cap C(\overline{\O};\Rb^N)$. Furthermore, the following Gauss-Green formula holds
\begin{equation}\label{GGBD}
\int_{\O}(u \odot D \varphi)(x) \, dx + \int_{\O}\varphi(x) \, dEu(x) =
\int_{\Gamma}\varphi (x)(\text{tr}u \odot \nu)(x) \, d \mathcal H ^{N-1}(x),
\end{equation}
for every $\varphi \in C^1(\overline{\O})$ (cf. \cite{ACDM, T}).
\smallskip
The following lemma is proved in \cite{BFT}.
\begin{lemma}\label{lemma2.2BFT}
Let $u \in BD(\O)$ and let $\rho \in C_0^{\infty}(\Rb^N)$ be a non-negative function such that ${\rm supp}(\rho) \subset\subset B(0,1)$, $\rho(-x) = \rho(x)$ for every $x \in \Rb^N$ and $\displaystyle \int_{\Rb^N}\rho(x) \, dx = 1$. For any $n \in \Nb$ set $\rho_n(x) : = n^N\rho(nx)$ and
$$u_n(x) := (u *\rho_n)(x) = \int_{\O}u(y)\rho_n(x-y) \, dy, \; \;
\mbox{for} \; x \in \left\{y \in \O : {\rm dist}(y,\partial \O) > \frac{1}{n}\right\}.$$
Then $u_n \in C^{\infty}\left(\left\{y \in \O : {\rm dist}(y,\partial \O) > \frac{1}{n}\right\};\Rb^N\right)$ and
\begin{itemize}
\item[i)] for any non-negative Borel function $h : \O \to \Rb$
$$\int_{B(x_0,\varepsilon)}h(x) |\E u_n(x)| \, dx \leq
\int_{B(x_0,\varepsilon+ \frac{1}{n})}(h*\rho_n)(x) \, d|E u|(x),$$
whenever $\e + \frac{1}{n} < {\rm dist}(x_0,\partial \O)$;
\item[ii)] for any positively homogeneous of degree one, convex function
$\theta : \Rb^{N\times N}_{\rm sym} \to [0,+\infty[$ and any $\e \in \, ]0,{\rm dist}(x_0,\partial \O)[$ such that $|E u|(\partial B(x_0,\e)) = 0$,
$$\lim_{n\rightarrow +\infty}\int_{B(x_0,\varepsilon)} \theta(\E u_n(x)) \, dx =
\int_{B(x_0,\varepsilon)}\theta\left(\frac{d Eu}{d |Eu|}\right) \, d |Eu|,$$
\item[iii)] $\displaystyle \lim_{n \to + \infty}u_n(x) = \widetilde u(x)$ and
$\displaystyle \lim_{n \to + \infty}(|u_n -u| * \rho_n)(x) = 0$ for every
$x \in \O \setminus S_u$, whenever $u \in L^{\infty}(\O;\Rb^N)$.
\end{itemize}
\end{lemma}
The following result, proved in \cite{T}, see also \cite[Theorem 2.6]{BFT}, shows that it is possible to approximate any $BD(\O)$ function $u$ by a sequence of smooth functions which preserve the trace of $u$.
\begin{theorem}\label{densitysmooth}
Let $\O$ be a bounded, connected, open set with Lipschitz boundary. For every $u \in BD(\O)$, there exists a sequence of smooth functions $\{u_n\} \subset C^{\infty}(\O;\Rb^N) \cap W^{1,1}(\O;\Rb^N)$ such that $u_n \stackrel{i}{\to}u$ and
tr $u_n =$ tr $u$. If, in addition, $u \in LD(\O)$, then $\E u_n \to \E u$ in
$L^1(\O;\Rb_s^{N\times N}).$
\end{theorem}
It is also shown in \cite{T} that if $\O$ is an open, bounded subset of $\Rb^N$, with Lipschitz boundary, then $BD(\O)$ is compactly embedded in $L^q(\O;\Rb^N)$, for every $1 \leq q < \frac{N}{N-1}.$
In particular,
the following result holds.
\begin{theorem}\label{THM2.10BFT}
Let $\Omega$ be an open, bounded subset of $\mathbb R^N$, with Lipschitz boundary
and let $1\leq q <\frac{N}{N-1}$. If $\{u_n\}$ is bounded in $BD(\Omega)$, then there exist $u \in BD(\Omega)$ and a subsequence $\{ u_{n_k}\}$ of $\{u_n\}$ such that $u_{n_k}\to u$ in $L^q (\Omega;\mathbb R^N)$.
\end{theorem}
\smallskip
If $u \in BD(\O )$ then $J_u$ is countably $(N-1)$-rectifiable, see \cite{ACDM},
and the following decomposition holds
\begin{equation*}
Eu= \E u \mathcal{L}^N \lfloor \O + [u] \odot \nu_u {\mathcal{H}}^{N-1}\lfloor J_u + E^cu,
\end{equation*}
\noindent where $[u] = u^+ - u^-$, $u^{\pm}$ are the traces of $u$ on the sides of $J_u$ determined by the unit normal $\nu_u$ to $J_u$ and $E^cu$ is the Cantor part of the measure $Eu$ which vanishes on Borel sets $B$ with
$\mathcal H^{N-1}(B) < + \infty.$
We end this subsection by pointing out that the equivalent of \eqref{apdif}, with $\E u(x)$ replacing $\nabla u(x)$, is false (see \cite{ACDM}).
However the following result holds (cf. \cite[Theorem 4.3]{ACDM} and \cite[Theorem 2.5]{E2}).
\begin{theorem}
(Approximate Symmetric Differentiability)\label{approxsymdiff}
If $u \in BD(\Omega),$ then, for $\mathcal{L}^N$-a.e. $x \in\Omega$ ,
there exists an $N\times N$ matrix $\nabla u(x)$ such that
\begin{equation}\label{apdifBD}
\lim_{\varepsilon \rightarrow 0^+} \frac{1}{\varepsilon^{N+1} }
\int_{B_\varepsilon(x)} |u(y) - u(x) - \nabla u(x).(y-x)|\, dy = 0,
\end{equation}
\begin{equation}\label{apsymdif}
\lim_{\varepsilon \rightarrow 0^+} \frac{1}{\varepsilon^{N} }
\int_{B_\varepsilon(x)} \frac{|\langle u(y) - u(x) - \E u(x).(y-x), y-x\rangle|}{|y-x|^2}\, dy = 0,
\end{equation}
for $\mathcal L^N$- a.e. $x \in \Omega$.
Furthermore
$${\mathcal L}^N(\{x\in \Omega:|\nabla u(x)|>t \})\leq
\frac {C(N,\Omega)}{t}\|u\|_{BD(\Omega)}, \;\;\forall t>0,$$
with $C(N,\Omega)>0$ depending only on $N$ and $\Omega$.
\end{theorem}
From \eqref{apdifBD} and \eqref{apsymdif} it follows that
$\displaystyle \E u=\frac{\nabla u+\nabla u^T}{2}$.
\medskip
We denote by $\mathcal R$ the kernel of the linear operator $E$ consisting of the class of rigid motions in $\mathbb R^N$, i.e., affine maps of the form $Mx + b$ where $M$ is a skew-symmetric $N \times N$ matrix and $b\in \mathbb R^N$.
$\mathcal R$ is therefore closed and finite-dimensional so
it is possible to define the orthogonal projection $P : BD(\Omega)\to \mathcal R$. This operator belongs to the class considered in the following Poincar\'e-Friedrichs
type inequality for $BD$ functions (see \cite{ACDM}, \cite{K} and \cite{T}).
\begin{theorem}\label{Thm2.8BFT}
Let $\Omega$ be a bounded, connected, open subset of $\mathbb R^N$, with Lipschitz boundary, and let $R : BD(\Omega)\to \mathcal R$ be a continuous linear map which leaves the elements of $\mathcal R$ fixed. Then there exists a constant
$C(\Omega, R)$ such that
\begin{align*}
\int_\Omega |u(x) - R(u)(x)| \, dx \leq C(\Omega,R) \, |E u|(\O), \; \mbox{ for every } u\in BD(\Omega).
\end{align*}
\end{theorem}
\subsection{Notions of Quasiconvexity}\label{defqcx}
\begin{definition}
[\cite{BFT}, Definition 3.1]\label{BFTsqcx}
A Borel measurable
function $f:\mathbb R^{N\times N}_s\to \mathbb R$ is said to be
{\em symmetric quasiconvex} if
\begin{equation}\label{sqcx}
f(\xi)\leq\int_Q
f(\xi+\E \varphi(x)) \, dx,
\end{equation}
for every $\xi \in \mathbb R^{N\times N}_s$ and for
every $\varphi \in C^{\infty}_{\rm per}(Q;\mathbb R^N)$.
\end{definition}
\begin{remark}\label{LDper}
{\rm The above property \eqref{sqcx} is independent of the size, orientation and centre of the cube over which the integration is performed. Also, if $f$ is upper semicontinuous and locally bounded from above, using Fatou's Lemma and the density of smooth functions in $LD(Q)$, it follows that in \eqref{sqcx} $C^{\infty}_{\rm per}(Q;\mathbb R^N)$ may be replaced by $LD_{\rm per}(Q).$}
\end{remark}
Given $f:\mathbb R^{N\times N}_s\to \mathbb R$, the symmetric quasiconvex envelope of $f$, $SQf$, is defined by
\begin{equation}\label{sqcxenv}
SQf(\xi) : = \inf \bigg\{\int_Q f(\xi+\E \varphi(x)) \, dx : \varphi \in C^{\infty}_{\rm per}(Q;\mathbb R^N)\bigg\}.
\end{equation}
It is possible to show that $SQf$ is the greatest symmetric quasiconvex function that is less than or equal to $f$. Moreover, definition \eqref{sqcxenv} is independent of the domain, i.e.
\begin{align}\label{sqcx0}SQf(\xi) : = \inf \bigg\{\frac{1}{\L ^N(D)}\int_{D} f(\xi+\E \varphi(x)) \, dx : \varphi \in C^{\infty}_{0}(D;\mathbb R^N)\bigg\}
\end{align}
whenever $D \subset \Rb^N$ is an open, bounded set with $\L ^N(\partial D) = 0$.
In \cite{E}, a Borel measurable function $f:\mathbb R^{N\times N}_s\to \mathbb R$ is said to be symmetric quasiconvex if and only if
\begin{align}\label{SQfE}
f(\xi)\leq \frac{1}{\mathcal L^N(D)}\int_D f(\xi +\E\varphi(x)) \, dx
\hbox{ for all }\varphi \in W^{1,\infty}_0(D;\mathbb R^N),
\end{align}
and it is stated that $f$ is symmetric quasiconvex if and only if $f\circ \pi$ is quasiconvex in the sense of Morrey,
where $\pi$ is the projection of $\mathbb R^{N \times N}$ onto
$\mathbb R^{N\times N}_{s}$.
Let us show that these two notions coincide. Observe first that, for any
$\varphi \in C^\infty_0(D;\mathbb R^N)$,
\begin{align}\label{SQfestW11}
SQf(\xi)\leq \frac{1}{\mathcal L^N(D)}\int_D SQf(\xi+ \E \varphi(x))dx=
\frac{1}{\mathcal L^N(D)}\int_D (SQf \circ \pi)(\xi+ \nabla\varphi(x))dx.
\end{align}
If $f$ is upper semicontinuous and satisfies a growth condition from above as in \eqref{growthint}, then $SQf$ in \eqref{sqcx0} is symmetric quasiconvex also in the sense of \cite{E}. Indeed, $SQf$ satisfies the same growth condition \eqref{growthint} and a density argument as in \cite{BM} shows that
$SQf \circ \pi$ is $W^{1,1}$-quasiconvex, hence $W^{1,\infty}$-quasiconvex, i.e., $\varphi$ can be chosen in $W^{1,\infty}_0(D;\mathbb R^N)$. Thus,
\begin{align}\label{SQfisE}
SQf(\xi)\leq \frac{1}{\mathcal L^N(D)}\int_D SQf(\xi+ \E \varphi(x)) \, dx
\leq \frac{1}{\mathcal L^N(D)}\int_D f(\xi+ \E \varphi(x)) \, dx,
\end{align}
for every $\varphi \in W^{1,\infty}_0(D;\mathbb R^N)$.
Therefore, denoting by $SQf_E$ the symmetric quasiconvexification
\begin{align}\label{symqcxE}
SQf_E(\xi) : = \inf \bigg\{\frac{1}{\L ^N(D)}\int_{D} f(\xi+\E \varphi(x)) \, dx : \varphi \in W^{1,\infty}_{0}(D;\mathbb R^N)\bigg\},
\end{align}
and by $SQf$ the symmetric quasiconvexification defined through \eqref{sqcx0},
trivially $SQf_E\leq SQf$ and by \eqref{SQfisE} we have equality.
Actually, under linear growth conditions and upper semicontinuity of $f$, we may
also conclude that
\begin{align*}
SQf_E(\xi) : = \inf \bigg\{\frac{1}{\L^N(D)}\int_{D} f(\xi+\E \varphi(x)) \, dx : \varphi \in W^{1,1}_{0}(D;\mathbb R^N)\bigg\}.
\end{align*}
\section{Auxiliary Results}\label{auxres}
We recall that for $u \in BD(\Omega)$ and $\chi \in BV(\Omega;\{0,1\})$ the energy under consideration is
\begin{equation}\label{F}
F(\chi,u;\Omega) : =
\displaystyle \int_{\Omega}\chi(x) W_1({\cal E} u(x))
+ (1 - \chi(x))W_0(\E u(x)) \, dx + |D\chi|(\Omega),
\end{equation}
and our aim is to obtain an integral representation for the localised relaxed functionals, defined for $A \in \mathcal O(\Omega)$, by
\begin{align}\label{calF}
\mathcal{F}\left(\chi,u;A\right) &
:=\inf\bigg\{ \liminf_{n\rightarrow +\infty} F(\chi_n,u_n;A):
u_{n} \in W^{1,1}(A;\mathbb{R}^{N}),
\chi_{n} \in BV(A;\{0,1\}), \\
& \hspace{4cm}
u_{n}\to u\text{ in }
L^{1}(A;\mathbb{R}^{N}),
\chi_{n}\to\chi\text{ in }
L^1(A;\{0,1\}) \bigg\},\nonumber
\end{align}
\begin{align}\label{calFLD}
\mathcal{F}_{LD}\left(\chi,u;A\right) &
:=\inf\bigg\{ \liminf_{n\rightarrow +\infty} F(\chi_n,u_n;A):
u_{n} \in LD(A), \chi_{n} \in BV(A;\{0,1\}), \\
& \hspace{4cm}
u_{n}\to u\text{ in }
L^{1}(A;\mathbb{R}^{N}),
\chi_{n}\to\chi\text{ in }
L^1(A;\{0,1\}) \bigg\},\nonumber
\end{align}
where the densities $W_i$, $i = 0,1$, are continuous functions such
that
\begin{equation}\label{growth}
\exists \, \alpha, \beta > 0 \text{ such that }
\alpha |\xi| \leq W_i(\xi) \leq \beta (1 + |\xi|), \; \; \forall
\xi \in \Rb^{N \times N}_s
\end{equation}
and where, for purposes of notation, we let
$f: \{0,1\} \times \mathbb R^{N \times N}_s \to [0,+\infty)$
be defined as
\begin{equation}\label{density}
f\left( q,\xi\right) :=q W_1(\xi)+ (1-q)W_0(\xi).
\end{equation}
It follows from the definition of the recession function \eqref{recS} and from the growth conditions \eqref{growth} that for every $q \in \{0,1\}$ and every
$\xi \in \Rb^{N \times N}_s$
\begin{equation}\label{finftygr}
\alpha |\xi| \leq f^{\infty}(q,\xi) \leq \beta |\xi|.
\end{equation}
It is an immediate consequence of \eqref{growth} that
\begin{equation}\label{G}
|f(q_1,\xi) - f(q_2,\xi)| \leq \beta \, |q_1 - q_2|(1 + |\xi|), \; \;
\forall q_1, q_2 \in \{0,1\},
\forall \xi \in \Rb^{N \times N}_s,
\end{equation}
from which it follows that
\begin{equation}\label{Gfinfty}
|f^{\infty}(q_1,\xi) - f^{\infty}(q_2,\xi)| \leq \beta \, |q_1 - q_2| \,|\xi|, \; \;
\forall q_1, q_2 \in \{0,1\},
\forall \xi \in \Rb^{N \times N}_s.
\end{equation}
The following additional hypothesis will be used to write the density of the jump term in the form given in \eqref{KSQ}
\begin{equation}\label{finfty}
\exists \, 0 < \gamma \leq 1, \exists \, C, L > 0 : \; t\,|\xi| > L \Rightarrow
\left| f^{\infty}(q,\xi) - \frac{f(q,t\xi)}{t}\right|
\leq C \frac{|\xi|^{1-\gamma}}{t^\gamma},
\end{equation}
for every $q \in \{0,1\}$ and every $\xi \in \Rb^{N \times N}_s$.
As pointed out in \cite{FM}, this can be stated equivalently as
\begin{equation}\label{finfty2}
\exists \, 0 < \gamma \leq 1, \exists \, C > 0 \mbox{ such that }
\left| f^{\infty}(q,\xi) - f(q,\xi)\right|
\leq C \left(1 + |\xi|^{1-\gamma}\right),
\end{equation}
for every $q \in \{0,1\}$ and every $\xi \in \Rb^{N \times N}_s$.
Under our assumed growth conditions \eqref{growth}, we observe that if $f$ satisfies \eqref{finfty}, or equivalently \eqref{finfty2}, then the same holds for its symmetric quasiconvex envelope $SQf$.
To this end, we recall that, under the hypothesis \eqref{growth}, the
recession function of a symmetric quasiconvex function is still symmetric quasiconvex (see \cite[Remarks 8 and 9]{R}) and we begin by stating the following results (cf. \cite[(iv) and (v) in Remark 3.2]{CZ} and \cite[Propositions 2.6, 2.7]{RZ} for the quasiconvex counterpart).
\begin{proposition}\label{SQfinfty=}
Let $f:\{0,1\} \times \mathbb R^{N \times N}_s\to [0, + \infty)$ be a continuous function as in \eqref{density} and satisfying \eqref{growth} and \eqref{finfty}. Let $f^\infty$ and $SQf$ be its recession function and its symmetric quasiconvex envelope, defined by \eqref{recS} and \eqref{sqcxenv}, respectively.
Then
\begin{equation}\label{QWinf=}
SQ(f^\infty)(q,\xi)= (SQf)^\infty(q,\xi) \;\;\;\hbox{ for every }(q,\xi) \in \{0,1\} \times \mathbb R^{N \times N}_s.
\end{equation}
\end{proposition}
\begin{proposition}\label{propperH5}
Let $f:\{0,1\} \times \mathbb R^{N \times N}_s\to [0, + \infty)$ be a continuous function as in \eqref{density}, satisfying \eqref{growth} and \eqref{finfty}.
Then, there exist $\gamma\in [0,1)$ and $C>0$ such that
$$ \displaystyle\left| (SQf)^\infty(q,\xi)- SQf(q, \xi)\right|
\leq C\big( 1+|\xi|^{1-\gamma}\big), \;\;
\forall\ (q,\xi)\in \{0,1\} \times \mathbb R^{N \times N}_s.$$
\end{proposition}
The growth conditions \eqref{growth}, as well as standard diagonalisation arguments, allow us to prove the following properties of the functional
$\mathcal F (\chi,u;A)$ defined in \eqref{calF}.
\begin{proposition}\label{firstprop}
Let $A \in \mathcal O(\Omega)$, $u \in BD(A)$, $\chi \in BV(A;\{0,1\})$ and
$F(\chi,u;A)$ be given by \eqref{F}. If $W_i$, $i = 0,1$, satisfy \eqref{growth}, then
\begin{itemize}
\item[i)] there exists $C > 0$ such that
$$C\left(|Eu|(A) + |D\chi|(A)\right) \leq \mathcal F (\chi,u;A) \leq
C\left(\L^N(A) + |Eu|(A) + |D\chi|(A)\right);$$
\item[ii)] $\mathcal F (\chi,u;A)$ is always attained, that is, there exist sequences
$\{u_n\} \subset W^{1,1}(A;\mathbb{R}^{N})$ and
$\{\chi_{n}\} \subset BV(A;\{0,1\})$ such that $u_{n}\to u$ in $L^{1}(A;\mathbb{R}^{N})$,
$\chi_{n}\to\chi$ in $L^1(A;\{0,1\})$ and
$$\mathcal F(\chi,u;A) = \lim_{n\rightarrow\infty}F(\chi_n,u_n;A);$$
\item[iii)] if $\{u_n\} \subset W^{1,1}(A;\mathbb{R}^{N})$ and
$\{\chi_{n}\} \subset BV(A;\{0,1\})$ are such that $u_{n}\to u$ in $L^{1}(A;\mathbb{R}^{N})$ and
$\chi_{n}\to\chi$ in $L^1(A;\{0,1\})$, then
$$\mathcal F(\chi,u;A) \leq \liminf_{n\to +\infty}\mathcal F(\chi_n, u_n;A).$$
\end{itemize}
\end{proposition}
\begin{proof}
$i)$ The upper bound follows from the growth condition from above of $W_i$,
$i =0,1$ and by fixing $\chi_n = \chi$ as a test sequence for
$\mathcal F (\chi,u;A)$, whereas the lower bound is a consequence of the inequality from below in \eqref{growth}, \eqref{Elsc} and the lower semicontinuity of the total variation of Radon measures.
The conclusions in $ii)$ and $iii)$ follow by standard diagonalisation arguments.
\end{proof}
\begin{remark}{\rm Analogous conclusions also hold for the functional
$\mathcal F_{LD} (\chi,u;A)$.}
\end{remark}
\begin{remark}\label{remkW11LD}{\rm
Assuming that the continuous functions $W_0$ and $W_1$ satisfy the growth hypothesis \eqref{growth}, it follows from the density of smooth functions in $LD(\Omega)$ and a diagonalisation argument that
$$\mathcal{F}\left(\chi,u;A\right) = \mathcal{F}_{LD}\left(\chi,u;A\right), \;
\mbox{for every} \; \chi \in BV(A;\left\{0,1\right\}), u \in BD(\Omega),
A \in {\cal O}(\Omega).$$}
\end{remark}
\begin{proof}
As $W^{1,1}(A;\mathbb R^N) \subset LD(A)$, one inequality is trivial. In order to show the reverse one, let $\{u_n\} \subset LD(A)$,
$\{\chi_n\} \subset BV(A;\left\{0,1\right\})$ be such that
$u_{n}\to u$ in $L^{1}(A;\mathbb{R}^{N})$,
$\chi_{n}\to\chi$ in $L^1(A;\{0,1\})$ and
$$\mathcal{F}_{LD}\left(\chi,u;A\right) = \lim_n \left[
\int_{A}\chi_n(x) W_1({\cal E} u_n(x)) + (1 - \chi_n(x))W_0(\E u_n(x)) \, dx + |D\chi_n|(A) \right].$$
By Theorem \ref{densitysmooth}, for each $n \in \Nb$, let $v_{n,k} \in W^{1,1}(A;\mathbb R^N)$ be such that
$v_{n,k} \to u_n$ in $L^{1}(A;\mathbb{R}^{N})$, as $k \to + \infty$, and
$\E v_{n,k} \to \E u_n$ in $L^{1}(A;\mathbb{R}^{N\times N}_s)$, as $k \to + \infty$.
By passing to a subsequence, if necessary, assume also that $\displaystyle \lim_{k\rightarrow\infty}\E v_{n,k}(x) = \E u_n(x)$, for a.e. $x \in A$.
By \eqref{growth} and Fatou's Lemma we obtain
$$
\int_A\chi_n(x)\left[C(1 + |\E u_n(x)|) - W_1(\E u_n(x))\right] dx \leq
\liminf_{k\to +\infty}\int_A\chi_n(x)
\left[C(1 + |\E v_{n,k}(x)|) - W_1(\E v_{n,k}(x))\right] dx $$
so that
$$
\int_A\chi_n(x)W_1(\E u_n(x)) \, dx \geq
\limsup_{k\to +\infty}\int_A\chi_n(x)W_1(\E v_{n,k}(x)) \, dx, $$
and likewise for the term involving $(1 - \chi_n)W_0$. From the previous inequalities
we conclude that
$$F(\chi_n,u_n;A) \geq \limsup_{k\to +\infty}F(\chi_n,v_{n,k};A).$$
Since $v_{n,k} \to u_n$ in $L^{1}(A;\mathbb{R}^{N})$, as $k \to + \infty$, and
$u_{n}\to u$ in $L^{1}(A;\mathbb{R}^{N})$, by a diagonalisation argument there exists
a sequence $k_n \to + \infty$ such that $v_{n,k_n} \to u$ in $L^{1}(A;\mathbb{R}^{N})$ and
$$F(\chi_n,v_{n,k_n};A) \leq F(\chi_n,u_n;A) + \frac{1}{k_n}.$$
As $\{\chi_n\}$, $\{v_{n,k_n}\}$ are admissible for $\mathcal{F}\left(\chi,u;A\right)$ it follows that
\begin{align*}
\mathcal{F}\left(\chi,u;A\right) \leq \liminf_{n\to +\infty}F(\chi_n,v_{n,k_n};A)
\leq \limsup_{n\to +\infty} \left(F(\chi_n,u_n;A) + \frac{1}{k_n}\right) =
\mathcal{F}_{LD}\left(\chi,u;A\right).
\end{align*}
\end{proof}
A straightforward adaptation of the proof of
\cite[Proposition 3.7]{BFT} yields the following result which enables us to prove the nested subadditivity property of the functional $\mathcal{F}\left(\chi,u;\cdot\right)$.
\begin{proposition}\label{slicing}
Let $A \in \mathcal O (\Omega)$ and assume that $W_0, W_1$ satisfy the growth condition \eqref{growth}. Let $\{\chi_n\} \subset BV(A;\{0,1\})$ and
$\{u_n\}, \{v_n\} \subset BD(A;\mathbb R^N)$ be sequences
satisfying $u_{n} - v_n \to 0$ in $L^{1}(A;\mathbb{R}^{N})$,
$\sup_n |Eu_n|(A) < + \infty$, $|Ev_n| \overset{\ast}{\rightharpoonup} \mu$ and
$|Ev_n| \to \mu(A)$. Then there exist subsequences $\{v_{n_k}\}$ of $\{v_n\}$,
$\{\chi_{n_k}\}$ of $\{\chi_n\}$ and there exists a sequence $\{w_k\} \subset BD(A)$
such that $w_k = v_{n_k}$ near $\partial A$,
$w_k - v_{n_k} \to 0$ in $L^{1}(A;\mathbb{R}^{N})$ and
$$\limsup_{k\to +\infty}F(\chi_{n_k},w_k;A) \leq
\liminf_{n\to +\infty}F(\chi_n,u_n;A).$$
\end{proposition}
It is clear from the proof that if the original sequences $\{u_n\}, \{v_n\}$ belong
to $W^{1,1}(A;\mathbb{R}^{N})$ then the sequence $\{w_k\}$ will also be in this space.
\begin{proposition}\label{nestedsa}
Assume that $W_0$ and $W_1$ are continuous functions satisfying \eqref{growth}.
Let $u \in BD(\Omega)$, $\chi \in BV(\Omega;\{0,1\})$ and $S, U, V \in \mathcal O(\Omega)$ be such that $S \subset \subset V \subset U.$ Then
$$\mathcal{F}\left(\chi,u;U\right) \leq \mathcal{F}\left(\chi,u;V\right)
+ \mathcal{F}\left(\chi,u;U\setminus \overline S\right).$$
\end{proposition}
\begin{proof}
By Proposition \ref{firstprop}, $ii)$, let
$\{v_n\} \subset W^{1,1}(V;\mathbb{R}^{N})$,
$\{w_n\} \subset W^{1,1}(U\setminus \overline S;\mathbb{R}^{N})$,
$\{\chi_n\} \subset BV(V;\{0,1\})$ and
$\{\theta_n\} \subset BV(U\setminus \overline S;\{0,1\})$ be such that
$v_n \to u$ in $L^{1}(V;\mathbb{R}^{N})$,
$w_n \to u$ in $L^{1}(U\setminus \overline S;\mathbb{R}^{N})$,
$\chi_{n}\to\chi$ in $L^1(V;\{0,1\})$
$\theta_{n}\to\chi$ in
$L^1(U\setminus \overline S;\{0,1\})$ and
\begin{align}\label{VUS1}
\mathcal{F}\left(\chi,u;V\right) & = \lim_{n\rightarrow\infty}F(\chi_n,v_n;V) \\
\mathcal{F}\left(\chi,u;U\setminus \overline S\right) & = \lim_{n\rightarrow\infty}F(\theta_n,w_n;U\setminus \overline S). \label{VUS2}
\end{align}
Let $V_0 \in \mathcal O_{\infty}(\O)$ satisfy
$S \subset \subset V_0 \subset \subset V$ and $|E u| (\partial V_0) = 0$,
$|D\chi|(\partial V_0) = 0$. Applying Proposition \ref{slicing} to $\{v_n\}$ and $u$ in $V_0$, we obtain a subsequence $\{\overline \chi_n\}$ of $\{\chi_n\}$ and a sequence $\{\overline v_n\} \subset W^{1,1}(V_0;\mathbb{R}^{N})$
such that $\overline v_n = u$ near $\partial V_0$, $\overline v_n \to u$ in
$L^{1}(V_0;\mathbb{R}^{N})$ and
\begin{equation}\label{sl1}
\limsup_{n\to +\infty}F(\overline \chi_n,\overline v_n;V_0) \leq
\liminf_{n\to +\infty}F(\chi_n,v_n;V_0).
\end{equation}
A further application of Proposition \ref{slicing}, this time to $\{w_n\}$ and $u$ in $U \setminus \overline V_0$, yields a subsequence $\{\overline \theta_n\}$ of $\{\theta_n\}$ and a sequence
$\{\overline w_n\} \subset W^{1,1}(U \setminus \overline V_0;\mathbb{R}^{N})$
such that $\overline w_n = u$ near $\partial V_0$, $\overline w_n \to u$ in
$L^{1}(U \setminus \overline V_0;\mathbb{R}^d)$ and
\begin{equation}\label{sl2}
\limsup_{n\to +\infty}F(\overline \theta_n,\overline w_n;U \setminus \overline V_0) \leq \liminf_{n\to +\infty}F(\theta_n,w_n;U \setminus \overline V_0).
\end{equation}
Define
$$z_n : = \begin{cases}
\overline v_n,& {\rm in } \; V_0\\
\overline w_n,& {\rm in } \; U \setminus V_0,
\end{cases}$$
notice that, by the properties of $\{\overline v_n\}$ and $\{\overline w_n\}$,
$\{z_n\} \subset W^{1,1}(U;\mathbb{R}^{N})$ and
$z_n \to u$ in $L^{1}(U;\mathbb{R}^{N})$.
We must now build a transition sequence $\{\eta_n\}$ between $\{\overline\chi_n\}$ and $\{\overline\theta_n\}$, in such a way that an upper bound
for the total variation of $\eta_n$ is obtained.
In order to connect these functions without adding more interfaces, we argue
as in \cite{BMMO} (see also \cite{BZ}).
For $\delta > 0$ consider
$$
V_\delta:=\{x \in V: {\rm dist}(x,V_0) < \delta\},
$$
where $\delta$ is small enough so that $\overline w_n = u$ in
$V_{\delta} \setminus \overline V_0$ and
\begin{equation}\label{Odelta}
\int_{V_{\delta} \setminus \overline V_0}C(1 + |u(x)|) \, dx = O(\delta).
\end{equation}
Given $x \in V$, let $d(x) := {\rm dist}(x;V_0)$. Since the distance function
to a fixed set is Lipschitz continuous,
applying the change of variables formula (see Theorem 2, Section 3.4.3,
in \cite{EG}) yields
$$
\int_{V_\delta \setminus \overline{V_0}}|\overline\chi_n(x)-\overline\theta_n(x)|
|{\rm det}\nabla d(x)| \, dx =
\int_0^\delta \left[ \int_{d^{-1}(y)} |\overline\chi_n(x)-\overline\theta_n(x)| \,
d {\mathcal H}^{N-1}(x) \right] dy$$
and, as $|{\rm det}\nabla d(x)|$
is bounded and $\overline\chi_n - \overline\theta_n \to 0$ in
$L^1(V \cap (U \setminus \overline S);\{0,1\})$, it follows that, for almost every $\rho \in [0; \delta]$, we have
\begin{equation}\label{rho}
\lim_{n\to +\infty}\int_{d^{-1}(\rho)}|\overline\chi_n(x) - \overline\theta_n(x)| \,
d{\mathcal H}^{N-1}(x) =
\lim_{n\to +\infty}\int_{\partial V_\rho}|\overline\chi_n(x)-\overline\theta_n(x)| \,
d {\mathcal H}^{N-1}(x) = 0.
\end{equation}
Fix $\rho_0\in [0; \delta]$ such that $|D\chi|(\partial V_{\rho_0})= 0$ and \eqref{rho} holds. We
observe that $V_{\rho_0}$ is a set with locally Lipschitz boundary since
it is a level set of a Lipschitz function (see, for example, \cite{EG}). Hence, for every $n$, we can consider $\overline\chi_n, \overline\theta_n$ on $\partial V_{\rho_0}$ in the sense of traces and define
\begin{equation*}
\eta_n :=\begin{cases}
\overline\chi_n, &\hbox{ in } V_{\rho_0}\\
\overline\theta_n, &\hbox{ in } U\setminus V_{\rho_0}.
\end{cases}
\end{equation*}
Then $\{\eta_n\} \subset BV(U;\{0,1\})$,
$\eta_{n}\to\chi$ in $L^1(U;\{0,1\})$ and so $\{\eta_n\}$
and $\{z_n\}$ are admissible for $\mathcal{F}\left(\chi,u;U\right)$. Therefore, by
\eqref{rho}, \eqref{growth}, \eqref{sl1}, \eqref{sl2}, \eqref{Odelta}, \eqref{VUS1} and \eqref{VUS2},
\begin{align*}
\mathcal{F}\left(\chi,u;U\right) &\leq \liminf_{n\to +\infty}F(\eta_n,z_n;U) \\
& = \liminf_{n\to +\infty}\left[F(\overline\chi_n,\overline v_n; V_0) +
\int_{V_{\rho_0}\setminus \overline V_0}\overline\chi_n(x)W_1(\E u(x)) +
(1 - \overline \chi_n(x))W_0(\E u(x)) \, dx \right.\\
& \left. \hspace{2cm} + |D\overline \chi_n|(V_{\rho_0}\setminus V_0)
+ F(\overline\theta_n,\overline w_n; U \setminus V_{\rho_0}) + \int_{\partial V_{\rho_0} }|\overline \chi_n(x) - \overline \theta_n(x)|
\, d {\mathcal H}^{N-1}(x)\right] \\
& \leq \limsup_{n\to +\infty}F(\overline\chi_n,\overline v_n; V_0) +
\limsup_{n\to +\infty}F(\overline\theta_n,\overline w_n; U \setminus \overline V_{0})
+ \int_{V_{\rho_0}\setminus \overline V_0}C \big(1 + |\E u(x)|\big) \, dx \\
& \hspace{7.9cm} + \limsup_{n\to +\infty} |D\chi_n|(V_{\rho_0}\setminus V_0) \\
& \leq \liminf_{n\to +\infty}F(\chi_n,v_n; V_0) +
\liminf_{n\to +\infty}F(\theta_n,w_n; U \setminus \overline V_{0}) + O(\delta) +
\limsup_{n\to +\infty} |D\chi_n|(V_{\rho_0}\setminus V_0) \\
& \leq \limsup_{n\to +\infty}F(\chi_n,v_n; V) +
\limsup_{n\to +\infty}F(\theta_n,w_n; U \setminus \overline S) + O(\delta) \\
& = \mathcal{F}\left(\chi,u;V\right) +
\mathcal{F}\left(\chi,u;U\setminus \overline S\right) + O(\delta)
\end{align*}
so the result follows by letting $\delta \to 0^+$.
\end{proof}
\begin{proposition}\label{traceRm}
Let $W_0$ and $W_1$ be continuous functions satisfying \eqref{growth}.
For every $u \in BD(\Omega)$, $\chi \in BV(\Omega;\{0,1\})$, $\mathcal{F}\left(\chi,u;\cdot\right)$ is the restriction to $\mathcal O(\O)$ of a Radon measure.
\end{proposition}
\begin{proof}
By Proposition \ref{firstprop}, $ii)$, let
$\{u_n\} \subset W^{1,1}(\O;\mathbb{R}^{N})$,
$\{\chi_n\} \subset BV(\O;\{0,1\})$, be such that
$u_n \to u$ in $L^{1}(\O;\mathbb{R}^{N})$,
$\chi_{n}\to\chi$ in $L^1(\O;\{0,1\})$ and
$$\mathcal{F}\left(\chi,u;\O\right) = \lim_{n\rightarrow\infty}F(\chi_n,u_n;\O).$$
Let $\mu_n = f(\chi_n(\cdot), \E u_n (\cdot)) {\cal L}^N\lfloor{\Omega}
+ |D \chi_n|$ and extend this sequence of measures outside of
$\Omega$ by setting, for any Borel set $E \subset \mathbb R^N$,
$$\lambda_n(E) = \mu_n(E \cap \Omega).$$
Passing, if necessary, to a subsequence, we can assume that there exists a
non-negative Radon measure $\mu$ (depending on $\chi$ and $u$) on
$\overline{\Omega}$ such that $\lambda_n \wsto \mu$ in the sense of measures
in $\overline {\Omega}$.
Let $\varphi_k \in C_0(\overline{\Omega})$ be an increasing sequence of functions
such that $0 \leq \varphi_k \leq 1$ and $\varphi_k(x) \to 1$ a.e. in
$\overline{\Omega}$. Then, by Fatou's Lemma and by the choice of $\{u_n\}$, $\{\chi_n\}$, we have
\begin{align*}
\mu(\overline{\Omega}) & = \int_{\overline{\Omega}}
\liminf_{k \to + \infty} \varphi_k(x) \, d\mu
\leq \liminf_{k \to +\infty}\int_{\overline{\Omega}}\varphi_k(x) \, d\mu \\
& = \liminf_{k \to +\infty}\lim_{n \to + \infty}\left ( \int_{\Omega}
\varphi_k(x) f(\chi_n(x),\E u_n (x)) \, dx +
\int_{\Omega}\varphi_k(x) \, d |D\chi_n| \right )\\
& \leq \lim_{n \to + \infty} \left (\int_{\Omega}
f(\chi_n(x),\E u_n (x)) \, dx + |D\chi_n|(\Omega) \right )
= {\cal F}(\chi,u;\Omega),
\end{align*}
so that
\begin{equation}\label{mu1}
\mu(\overline{\Omega})\leq {\cal F}(\chi, u; \Omega).
\end{equation}
On the other hand, by the upper semicontinuity of weak $\ast$ convergence of
measures on compact sets,
for every open set $V \subset \Omega$, it follows that
\begin{equation}\label{mu2}
{\cal F}(\chi, u; V) \leq \liminf_{n \to +\infty} F(\chi_n, u_n; V)
= \liminf_{n \to +\infty} \mu_n(V)
\leq \limsup_{n \to +\infty} \mu_n(\overline{V})
\leq \mu(\overline{V}).
\end{equation}
Now let $V \in \mathcal O(\O)$ and $\e>0$ be fixed and consider an open set
$S \subset \subset V$ such that $\mu(V \setminus S) < \e$. Then
\begin{equation}\label{mu3}
\mu(V) \leq \mu(S) + \e = \mu(\overline \O) - \mu(\overline \O \setminus S) + \e,
\end{equation}
and so, by \eqref{mu3}, \eqref{mu1}, \eqref{mu2} and Proposition \ref{nestedsa}
we have
\begin{equation*}
\mu(V)\leq \mu({\overline \Omega})- \mu( {\overline\Omega} \setminus S)
+ \varepsilon
\leq {\cal F}(\chi, u; \Omega)-
{\cal F}(\chi, u; \Omega \setminus {\overline S}) + \varepsilon
\leq {\cal F}(\chi, u; V)+\varepsilon.
\end{equation*}
Letting $\e \to 0^+$, we obtain
$$
\mu(V)\leq {\cal F}(\chi, u; V),
$$
whenever $V$ is an open set such that $V \subset \subset \Omega$.
For a general open subset $V \subset \Omega$ we have
$$
\mu(V) = \sup \{\mu(O) : O \subset \subset V \}
\leq \sup \{{\cal F}(\chi, u; O) : O \subset \subset V \}
\leq {\cal F}(\chi, u; V).
$$
It remains to show that ${\cal F}(\chi, u; U) \leq \mu(U)$,
$\forall \, U \in \mathcal O(\O)$. Fix $\e > 0$ and choose $V, S \in \mathcal O(\O)$
such that $S \subset \subset V \subset \subset U$ and
$\mathcal L^N(U \setminus \overline S) + |Eu|(U \setminus \overline S)
+ |D\chi|(U \setminus \overline S) < \e$. By Proposition \ref{firstprop} $i)$,
\eqref{mu2} and the nested subadditivity result, it follows that
\begin{align*}
{\cal F}(\chi, u; U) &\leq {\cal F}(\chi, u; V) +
{\cal F}(\chi, u; U \setminus {\overline S}) \\
&\leq \mu(\overline V) +
C\left(\mathcal L^N(U \setminus \overline S) + |Eu|(U \setminus \overline S)
+ |D\chi|(U \setminus \overline S)\right) \leq \mu(U) + C\e,
\end{align*}
so it suffices to let $\e \to 0^+$ to conclude the proof.
\end{proof}
Combining the arguments given in the proofs of Propositions \ref{slicing} and
\ref{nestedsa} it is possible to obtain the following refined version of Proposition
\ref{slicing}.
\begin{proposition}\label{newslicing}
Let $A \in \mathcal O (\Omega)$ and assume that $W_0, W_1$ satisfy the growth condition \eqref{growth}. Let
$\{u_n\}, \{v_n\} \subset BD(A;\mathbb R^N)$ and $\{\chi_n\}, \{\theta_n\} \subset BV(A;\{0,1\})$ be sequences
satisfying $u_{n} - v_n \to 0$ in $L^{1}(A;\mathbb{R}^{N})$,
$\chi_{n} - \theta_n \to 0$ in $L^1(A;\{0,1\})$,
$\sup_n |E^su_n|(A) < + \infty$, $|Ev_n| \overset{\ast}{\rightharpoonup} \mu$,
$|Ev_n| \to \mu(A)$, $\sup_n |D\chi_n|(A) < + \infty$ and
$\sup_n |D\theta_n|(A) < + \infty$. Then there exist subsequences $\{v_{n_k}\}$ of $\{v_n\}$,
$\{\theta_{n_k}\}$ of $\{\theta_n\}$ and there exist sequences
$\{w_k\} \subset BD(A)$, $\{\eta_k\} \subset BV(A;\{0,1\})$
such that $w_k = v_{n_k}$ near $\partial A$,
$\eta_k = \theta_{n_k}$ near $\partial A$,
$w_k - v_{n_k} \to 0$ in $L^{1}(A;\mathbb{R}^{N})$,
$\eta_{k} - \theta_{n_k} \to 0$ in $L^1(A;\{0,1\})$ and
$$\limsup_{k\to +\infty}F(\eta_{k},w_k;A) \leq
\liminf_{n\to +\infty}F(\chi_n,u_n;A).$$
\end{proposition}
As in Proposition \ref{slicing}, the new sequence $\{w_k\}$ has the same regularity as the original sequences $\{u_n\}, \{v_n\}$ as it is obtained through a convex combination of these ones using smooth cut-off functions.
\smallskip
The following proposition, whose proof is standard (cf. for instance \cite[Lemma 3.1]{RZ} or \cite[Proposition 2.14]{CZasy}), allows us to assume without loss of generality that $f$ is symmetric quasiconvex.
\begin{proposition}\label{Frel=Frel**}
Let $W_0$ and $W_1$ be continuous functions satisfying \eqref{growth} and consider the functional
$F:BV(\Omega;\{0,1\})\times BD(\Omega)\times \mathcal O(\Omega)$
defined in \eqref{F}.
Consider furthermore the relaxed functionals given in \eqref{calF} and
\begin{align}\label{Fsqxrelax}
\mathcal F_{SQf}(\chi,u;A) & :=\inf\Big\{\liminf_{n\to +\infty}
\int_A SQf(\chi_n(x),\E u_n(x)) \, dx + |D \chi_n|(A) : \\
& (\chi_n,u_n) \in BV(A;\{0,1\})\times LD(A),
u_{n}\to u\text{ in }
L^{1}(A;\mathbb{R}^{N}), \chi_{n} \to \chi \text{ in } L^1(A;\{0,1\})\Big\}.\nonumber
\end{align}
Then, ${\mathcal F}(\cdot,\cdot;\cdot)$ coincides with
${\mathcal F}_{SQf}(\cdot, \cdot; \cdot)$ in
$BV(\Omega;\{0,1\})\times BD(\Omega)\times \mathcal O(\Omega)$.
\end{proposition}
In the sequel we rely on the result of Proposition \ref{Frel=Frel**} and assume
that $f$ is symmetric quasiconvex. Together with \eqref{growth}, this entails the Lipschitz continuity of $f$ with respect to the second variable (see \cite{DPR2}). Under this quasiconvexity hypothesis, assuming in addition that \eqref{finfty} holds and taking also into account Proposition \ref{propperH5}, we
recall (cf. \eqref{KSQ}) that our relaxed surface energy density is given by
\begin{equation}\label{K}
K(a,b,c,d,\nu):=\inf\left\{
\displaystyle\int_{Q_{\nu}}
f^{\infty}(\chi(x),\E u(x)) \, dx+|D\chi|(Q_{\nu}):\left(\chi,u\right)
\in\mathcal{A}(a,b,c,d,\nu)\right\},
\end{equation}
where, for $(a,b,c,d,\nu) \in \{0,1\} \times \{0,1\} \times \mathbb{R}^{N} \times\mathbb{R}^{N} \times S^{N-1},$ the set of admissible functions is
\begin{align}\label{adm}
\mathcal{A}(a,b,c,d,\nu) & :=\bigg\{ \left(\chi,u\right)
\in BV_{\rm loc}\left(S_{\nu};\{0,1\}\right) \times
W^{1,1}_{\rm loc}\left(S_{\nu};\mathbb{R}^{N}\right) : \\
& (\chi(y),u(y)) = (a,c) \text{ if } y\cdot\nu=\frac{1}{2},
~(\chi(y),u(y)) = (b,d) \text{ if } y\cdot\nu=-\frac{1}{2},\nonumber\\
& (\chi, u)\text{ are 1-periodic in the directions of }\nu_{1},\dots,\nu_{N-1}\bigg\} ,\nonumber
\end{align}
$\left\{\nu_{1},\nu_{2},\dots,\nu_{N-1},\nu\right\}$ is an orthonormal
basis of $\mathbb{R}^{N}$ and $S_\nu$ is the strip given by
$$S_\nu = \left\{x \in \Rb^N : |x \cdot \nu| < \frac{1}{2}\right\}.$$
The following result provides an alternative characterisation of $K(a,b,c,d,\nu)$
which will be useful to obtain the surface term of the relaxed energy, under hypothesis \eqref{finfty}.
To this end, given $(a,b,c,d,\nu) \in \{0,1\} \times \{0,1\} \times \mathbb{R}^{N} \times\mathbb{R}^{N} \times S^{N-1},$ we consider the functions
\begin{equation}\label{targets}
\chi_{a,b,\nu}(y) := \begin{cases}
a, & {\rm if } \; y \cdot \nu > 0\\
b, & {\rm if } \; y \cdot \nu < 0
\end{cases}
\; \; \; {\rm and } \; \; \;
u_{c,d,\nu}(y) := \begin{cases}
c, & {\rm if } \; y \cdot \nu > 0\\
d, & {\rm if } \; y \cdot \nu < 0.
\end{cases}
\end{equation}
\begin{proposition}\label{Ktilde}
For every $(a,b,c,d,\nu) \in \{0,1\} \times \{0,1\} \times \mathbb{R}^{N} \times\mathbb{R}^{N} \times S^{N-1}$ we have
$$K(a,b,c,d,\nu) = \widetilde{K}(a,b,c,d,\nu)$$
where
\begin{align}\label{tilK}
\widetilde{K}(a,b,c,d,\nu) & := \inf\bigg\{\liminf_{n \to + \infty}
\left[\displaystyle\int_{Q_{\nu}}f^{\infty}(\chi_n(x),\E u_n(x)) \, dx
+ |D\chi_n|(Q_{\nu})\right] : \chi_n \in BV\left(Q_{\nu};\{0,1\}\right), \nonumber\\
& \hspace{1,3cm} u_n \in W^{1,1}\left(Q_{\nu};\mathbb{R}^{N}\right), \chi_n \to \chi_{a,b,\nu}
\; {\rm in } \; L^1(Q_{\nu};\{0,1\}), u_n \to u_{c,d,\nu} \; {\rm in } \; L^1(Q_{\nu};\Rb^N) \bigg\}.
\end{align}
\end{proposition}
\begin{proof}
The conclusion follows as in \cite[Proposition 3.5]{BBBF}, by proving a double inequality.
To show that $K(a,b,c,d,\nu) \leq \widetilde{K}(a,b,c,d,\nu)$ we take sequences $\{\chi_n\}$, $\{u_n\}$ as in the definition of $\widetilde{K}(a,b,c,d,\nu)$ and use
Proposition \ref{newslicing}, applied to $\{\chi_n\}$, $\{\chi_{a,b,\nu}\}$,
$\{u_n\}$ and $\{v_n\}$, where $v_n$ is a regularization of $u_{c,d,\nu}$ which preserves its boundary values (cf. Theorem \ref{densitysmooth}).
The reverse inequality is based on the periodicity of the admissible functions for $K(a,b,c,d,\nu)$, together with the Riemann-Lebesgue Lemma.
\end{proof}
\section{Proof of the Main Theorem}\label{mainthm}
Given $\chi \in BV(\Omega;\{0,1\})$ and $u \in BD(\Omega)$, by Proposition \ref{traceRm} we know that $\mathcal F(\chi,u,;\cdot)$ is the restriction to $\mathcal O(\Omega)$ of a Radon measure $\mu$. By Proposition~\ref{firstprop} $i)$ we may decompose $\mu$ as
$$\mu = \mu^a \mathcal L ^N + \mu^j
+ \mu^c, \;\; \hbox{ with } \mu^j \ll |E^j u|+ |D \chi|.$$
Our aim in this section is to characterise the density $\mu^a$ and the measures
$\mu^j$ and $\mu^c$.
We point out that the measure $\mu^j$ is given by
$\sigma^j \mathcal H^{N-1} \lfloor (J_\chi \cup J_u)$, for a certain density $\sigma^j$. Indeed, due to the fact that, for $BV$ functions,
$\mathcal H^{N-1}(S_u \setminus J_u) = 0$, the measure $|D\chi|$ is concentrated on
$J_\chi$ apart from an $\mathcal H^{N-1}$-negligible set, whereas, by
\cite[Remark 4.2 and Proposition 4.4]{ACDM},
$|E^ju|$ is concentrated on $J_u$ and it is the only part of the measure $Eu$ that is concentrated on $(n-1)$-dimensional sets.
\subsection{The Bulk Term}\label{bulk}
\begin{proposition}\label{lbbulk}
Let $u \in BD(\Omega)$, $\chi \in BV(\Omega;\{0,1\})$ and let $W_0$ and $W_1$ be continuous functions satisfying \eqref{growth}.
Assume that $f$ given by \eqref{density} is symmetric quasiconvex.
Then, for $\LL ^N$ a.e. $x_0 \in \O$,
$$ \mu^a(x_0) = \dfrac{d\mathcal F(\chi, u; \cdot)}{d {\mathcal L}^N}(x_0) \geq f(\chi(x_0), \mathcal E u(x_0)).$$
\end{proposition}
\begin{proof}
Let $x_0 \in \O$ be a point satisfying
\begin{equation}\label{RNdev}
\mu^a(x_0) = \frac{d \mu}{d \mathcal L^N}(x_0)=
\lim_{\e \to 0^+}\frac{\mu(Q(x_0,\e))}{\e^ N}
\; \; \mbox{exists and is finite}
\end{equation}
and
\begin{equation}\label{440bis}
\frac{d |E^su|}{d\mathcal{L}^N}(x_0)= 0, \; \;
\frac{d |D\chi|}{d\mathcal{L}^N}(x_0)=0.
\end{equation}
Furthermore, we choose $x_0$ to be a point of approximate continuity for $u$, for $\E u$ and for $\chi$, namely we assume that
\begin{equation}\label{uac}
\lim_{\e\rightarrow 0^+}
\frac{1}{\e^{N}}\int_{Q\left(x_0,\e\right)}
\left\vert u(x)-u(x_0)\right\vert \,dx = 0,
\end{equation}
\begin{equation}\label{364}
\lim_{\e \to 0^+} \frac{1}{\e^N} \int_{Q(x_0, \e)}
|\mathcal E u(x) - \mathcal Eu(x_0)|\, dx = 0
\end{equation}
and
\begin{equation}\label{chiad}
\lim_{\e\rightarrow 0^+}
\frac{1}{\e^{N}}\int_{Q\left(x_0,\e\right)}
\left\vert \chi(x)-\chi(x_0)\right\vert \,dx = 0.
\end{equation}
We observe that the above properties hold for ${\mathcal L}^N$ a.e. $x_0 \in \Omega$ (applying, for instance, \cite[eq. (2.5)]{ACDM} to $u$,
$\mathcal E u$ and $\chi$).
Assuming that the sequence $\e_k \to 0^+$ is chosen in such a way that
$\mu(\partial Q(x_0,\e_k)) = 0$, we have
\begin{equation*}
\begin{split}
\mu^a(x_0) = \lim_{\e_k \to 0^+}\frac{\mu (Q(x_0,\e_k))}{\e_k^N} &= \lim_{\e_k,n}\left[\frac{1}{\e_k^N}\int_{Q(x_0,\e_k)}
f(\chi_n(x),\mathcal E u_n(x))\,dx + |D \chi_n|(Q(x_0,\e_k))\right]\\
&\geq \lim_{\e_k,n}\int_Q f(\chi_n(x_0+\e_k y), \mathcal E u_n(x_0+\e_k y))\,dy,
\end{split}
\end{equation*}
where $\chi_n \in BV(Q(x_0,\e_k);\{0,1\})$, $\chi_n \to \chi$ in $L^1(Q(x_0,\e_k);\{0,1\})$ and $u_n \in W^{1,1}(Q(x_0,\e_k);\Rb^N)$, $u_n \to u$ in $L^1(Q(x_0,\e_k);\Rb^N)$.
Defining
$$\chi_{n,\e_k}(y):=\chi_n(x_0+\e_k y)- \chi(x_0),$$
it follows by \eqref{chiad} that
\begin{equation}\label{chindeltalim}
\begin{split}
\lim_{\e_k,n}\|\chi_{n,\e_k}\|_{L^1(Q)} = &\lim_{\e_k,n}
\int_Q|\chi_n(x_0+\e_k y)-\chi(x_0)|\,dy \\
=&\lim_{\e_k,n}\frac{1}{\e_k^{N}}\int_{Q(x_0,\e_k)}|\chi_n(x)-\chi(x_0)|\,dx\\
=&\lim_{\e_k \to 0^+}\frac{1}{\e_k^{N}}\int_{Q(x_0,\e_k)}|\chi(x)-\chi(x_0)|\, dx =0.
\end{split}
\end{equation}
Analogously, letting
$$u_{n,\e_k}(y):=\frac{u_n(x_0+\e_k y)- u(x_0)}{\e_k},$$
then
$\mathcal E u_{n,\e_k}(y)= \mathcal E u_n(x_0+\e_k y)$
and, since $u_{n,\e_k}\in W^{1,1}(\Omega;\Rb^N)$,
$E u_{n,\e_k}= \mathcal E u_{n,\e_k} \LL^N$.
Moreover, arguing as in the proof of \cite[Proposition 4.1]{BFT}, exploiting the coercivity of $f$ in the second variable and Theorems \ref{Thm2.8BFT} and \ref{THM2.10BFT}, we conclude that there exists a function $v \in BD(\Omega)$, such that
$$
\lim_{\e_k,n}\|u_{n,\e_k}- P(u_{n,\e_k})-v\|_{L^1(Q;\mathbb R^N)} = 0,
$$
where $P$ is the projection of $BD(\O)$ onto the kernel of the operator $E$.
Furthermore, given that the point $x_0$ was chosen to satisfy \eqref{440bis} and \eqref{364}, it was shown in \cite[Proposition 4.1, (4.8)]{BFT} that
\begin{equation}\label{measw*}
Ev = \mathcal E u(x_0)\mathcal L^N.
\end{equation}
Therefore, a diagonalisation argument allows to extract subsequences
$u_k := u_{n_k,\e_k} - P(u_{n_k,\e_k})$ and
$\chi_k := \chi_{n_k,\e_k}$, such that
\begin{equation}\label{radii}
\begin{split}
& \lim_{k \to +\infty}\|\chi_{k}\|_{L^1(Q)} = 0,\\
& \lim_{k \to +\infty} \|u_{k}- v\|_{L^1(Q;\mathbb R^N)} = 0
\end{split}
\end{equation}
and
\begin{align}\label{8}
\mu^a(x_0) = \frac{d\mu}{d\mathcal{L}^N}(x_0)\geq \lim_{k \to + \infty}
\int_Q f(\chi(x_0)+ \chi_{k}(y),\mathcal E u_{k}(y)) \, dy.
\end{align}
Our next step is to fix $\chi(x_0)$ in the first argument of $f$ in the previous integral. To this end we make use of Chacon's Biting Lemma
(see \cite[Lemma 5.32]{AFP}). Indeed, by the coercivity hypothesis \eqref{growth} and
\eqref{8}, the sequence $\{\E u_k\}$ is bounded in $L^1(Q;\Rb^{N\times N}_s)$ so the Biting Lemma guarantees the existence of a (not relabelled) subsequence of $\{u_k\}$ and of a decreasing sequence of Borel sets $D_r$, such that
$\displaystyle \lim_{r \to + \infty}{\mathcal L}^N(D_r) = 0$ and the sequence
$\{\mathcal Eu_k\}$ is equiintegrable in $Q\setminus D_r$,
for any $r \in \mathbb N$.
Since $f \geq 0$, by \eqref{G} and \eqref{8}, we have
\begin{align}\label{9}
\mu^a(x_0) & \geq \lim_{k \to + \infty}\int_{Q\setminus D_r}
f(\chi(x_0) + \chi_k(y), \mathcal E u_{k}(y))\, dy \nonumber\\
& \geq \lim_{k \to + \infty}
\left\{\int_{Q\setminus D_r} f(\chi(x_0), \mathcal E u_{k}(y)) \, dy - \int_{Q\setminus D_r} C |\chi_{k}(y)|\cdot (1 + |\mathcal E u_k(y)|) \, dy \right\}\nonumber \\
& \geq \lim_{k \to + \infty} \int_{Q \setminus D_r}
f(\chi(x_0), \mathcal E u_{k}(y)) \, dy - \limsup_{k \to + \infty}
\int_{Q\setminus D_r} C |\chi_{k}(y)| \cdot |\mathcal E u_k(y)| \, dy,
\end{align}
where we used \eqref{chindeltalim}.
We claim that for each $j \in \mathbb N$, there exist $k=k(j)$ and
$r_j \in \mathbb N$, such that
\begin{equation} \label{489}
\int_{Q\setminus D_{r_j}} f(\chi(x_0), \E u_{k(j)}(y)) \, dy \geq
\int_Q f(\chi(x_0), \E u_{k(j)}(y)) \, dy - \frac{C}{j}.
\end{equation}
In light of \eqref{growth}, in order to guarantee that \eqref{489} holds, it suffices to show that for each $j \in \Nb$, there exist $k=k(j)$ and $r_j \in \Nb$, such that
\begin{equation}\label{77}
\int_{D_{r_j}} 1 + |\mathcal E u_{k(j)}(y)| \, dy \leq \frac{1}{j}.
\end{equation}
Suppose not. Then, there exists $j_0 \in \Nb$ such that, for all $r, k \in \Nb$,
\begin{equation} \label{497}
\int_{D_r}1 + |\mathcal E u_{k}(y)| \, dy > \frac{1}{j_0}
\end{equation}
which contradicts the equiintegrability of the constant sequence $\{1 + |\E u_k|\}$,
for $k$ fixed, and the fact that
$\displaystyle \lim_{r \to + \infty}{\mathcal L}^N(D_r) = 0$.
For this choice of $k(j)$ and $r_j$, we now estimate the last term in \eqref{9}. Since
$|\chi_{k(j)}| \to 0$, as $j \to + \infty$, in $L^1(Q)$, this sequence also converges to zero in measure. Thus, denoting by
$$\displaystyle A_{k(j)} : = \left\{x \in Q \setminus D_{r_j} : |\chi_{k(j)}(x)| = 1 \right\},$$
it follows that for every $\delta > 0$, there
exists $j_0 \in \Nb$ such that $\LL^N(A_{k(j)}) < \delta$, for all $j > j_0$.
On the other hand, because the sequence
$\{\mathcal Eu_{k(j)}\}$ is equiintegrable in $Q\setminus D_{r_j}$, we know that
for every $\e > 0$, there exists $\delta = \delta(\e) > 0$ such that for any measurable set $A \subset Q\setminus D_{r_j}$ with $\LL^N(A) < \delta(\e)$ we have
$\displaystyle \int_{A}|\mathcal Eu_{k(j)}(y)| \, dy < \e$.
Choosing $j$ large enough so that $\LL^N(A_{k(j)}) < \delta(\e)$ we obtain
$\displaystyle \int_{A_{k(j)}}|\mathcal Eu_{k(j)}(y)| \, dy < \e$ and hence
\begin{equation}\label{11}
\int_{Q\setminus D_{r_j}}
|\chi_{k(j)}(y) \cdot |\mathcal Eu_{k(j)}(y)| \, dy < \e,
\end{equation}
for every sufficiently large $j$.
Therefore, up to the extraction of a further subsequence, and denoting in what follows $\chi_j := \chi_{k(j)}$, $v_j : = u_{k(j)}$ and
$ D_j:=D_{r_j}$, \eqref{9}, \eqref{489} and \eqref{11} yield
\begin{align*}
\mu^a(x_0) = \frac{d\mu}{d\mathcal{L}^N}(x_0) & \geq \liminf_{j \to + \infty}
\left(\int_Q f(\chi(x_0),\mathcal E v_j(y)) \, dy - \frac{C}{j}\right)
-\limsup_{j \to + \infty}\int_{Q\setminus D_{j}}
C |\chi_j(y)| \cdot |\mathcal E v_j(y)|\, dy \\
& \geq \liminf_{j \to + \infty} \int_Q f(\chi(x_0),\mathcal E v_j(y)) \, dy - \e.
\end{align*}
Since $v_j \to v$ in $L^1(Q;\Rb^N)$, Proposition \ref{slicing} allows us to assume, without loss of generality, that $v_j = v$ on $\partial Q$. Hence, using the symmetric quasiconvexity of $f$ in the second variable, which also holds for test functions in $LD_{\rm per}(Q)$ (cf. Remark \ref{LDper}), and \eqref{measw*}, we obtain
\begin{align*}
\mu ^a(x_0) &\geq
\liminf_{j \to + \infty} \int_Q f(\chi(x_0),\mathcal E v_j(y)) \, dy - \e \\
&\geq \liminf_{j \to + \infty}
\int_Q f(\chi(x_0),\E u(x_0) + \mathcal E (v_j - v)(y)) \, dy - \e \\
& \geq f(\chi(x_0),\mathcal E u(x_0)) - \e,
\end{align*}
so to conclude it suffices to let $\e \to 0^+$.
\end{proof}
\begin{proposition}\label{upbbulk}
Let $u \in BD(\Omega)$, $\chi \in BV(\Omega;\{0,1\})$ and let $W_0$ and $W_1$ be continuous functions satisfying \eqref{growth}.
Let $f$ be given by \eqref{density} and assume that $f$ is symmetric quasiconvex.
Then, for $\LL^N$ a.e. $x_0 \in \O$,
$$ \mu^a(x_0) = \dfrac{d\mathcal F(\chi, u; \cdot)}{d {\mathcal L}^N}(x_0) \leq f(\chi(x_0), \mathcal E u(x_0)).$$
\end{proposition}
\begin{proof}
Choose a point $x_0 \in \O$ such that \eqref{uac},\eqref{364}, \eqref{chiad} hold,
\begin{equation}\label{x02}
\lim_{\varepsilon \to 0^+}
\frac{1}{\e ^N}|E^s u|(Q(x_0, \e)) = 0,
\end{equation}
\begin{equation}\label{x03}
\lim_{\varepsilon \to 0^+}
\frac{1}{\e ^N}|D \chi|(Q(x_0, \e)) = 0,
\end{equation}
and, furthermore, such that
\begin{equation}\label{x05}
\mu^a(x_0) = \lim_{\varepsilon \to 0^+}
\frac{\mathcal F(\chi, u;Q(x_0,\e))}{\e ^N} \; \; \; \mbox{exists and is finite},
\end{equation}
where the sequence of $\e \to 0^+$ is chosen so that $|E u|(\partial Q(x_0,\e)) = 0$.
Notice that $\LL^N$ almost every point $x_0 \in \O$ satisfies the above properties.
For the purposes of this proof we assume that $\chi(x_0) = 1$, the case
$\chi(x_0) = 0$ is treated in a similar fashion. Thus, it follows from \eqref{chiad} that
\begin{equation}\label{x07}
\lim_{\varepsilon \to 0^+}
\frac{1}{\e ^N}\LL^N\big(Q(x_0, \e) \cap \{\chi = 0\}\big) = 0.
\end{equation}
Using the symmetric quasiconvexity of $f$, fix $\delta > 0$ and let
$\phi \in C^{\infty}_{\rm per}(Q;\Rb^N)$ be such that
\begin{equation}\label{phi}
\int_Qf(\chi(x_0),\E u(x_0) + \E \phi(x)) \, dx \leq f(\chi(x_0),\E u(x_0)) + \delta.
\end{equation}
We extend $\phi$ to $\Rb^N$ by periodicity, define
$\phi_n(x) := \dfrac{1}{n}\phi(nx)$ and consider the sequence of functions in $W^{1,1}(Q(x_0,\e);\Rb^N)$ given by
$$u_{n,\e}(x) : = (\rho_n * u)(x) + \e \phi_n\left(\frac{x - x_0}{\e}\right).$$
The periodicity of $\phi$ ensures that, as $ n \to + \infty$, $u_{n,\e} \to u$ in
$L^1(Q(x_0,\e);\Rb^N)$ and so, letting $\chi_n = \chi$, $\forall n \in \Nb$, the
sequences $\{u_{n,\e}\}_n$ and $\{\chi_n\}_n$ are admissible for
$\mathcal F(\chi,u;Q(x_0,\e))$. Hence, by \eqref{x03}, we have
\begin{align*}
\mu^a(x_0) &= \lim_{\varepsilon \to 0^+}
\frac{\mathcal F(\chi, u;Q(x_0,\e))}{\e ^N} \leq
\liminf_{\varepsilon \to 0^+}\liminf_{n \to +\infty}\frac{1}{\e^N}
\Big(\int_{Q(x_0, \e)}f(\chi(x),\E u_{n,\e}(x)) \, dx + |D\chi|(Q(x_0,\e))\Big)\\
&=\liminf_{\varepsilon \to 0^+}\liminf_{n \to +\infty}\frac{1}{\e^N}
\int_{Q(x_0, \e)}f\left(\chi(x),\E u_{n,\e}(x)\right) \, dx \\
&\leq \limsup_{\varepsilon \to 0^+}\limsup_{n \to +\infty}\frac{1}{\e^N}
\int_{Q(x_0, \e)}f\left(\chi(x_0),\E u(x_0) + \E \phi_n\left(\frac{x-x_0}{\e}\right)\right) \, dx \\
& + \limsup_{\varepsilon \to 0^+}\limsup_{n \to +\infty}\frac{1}{\e^N}
\int_{Q(x_0, \e)} f\left(\chi(x),\E u_{n,\e}(x)\right) -
f\left(\chi(x_0),\E u_{n,\e}(x))\right)\, dx \\
& + \limsup_{\varepsilon \to 0^+}\limsup_{n \to +\infty}\frac{1}{\e^N}
\int_{Q(x_0, \e)} f\left(\chi(x_0),\E u_{n,\e}(x)\right) -
f\left(\chi(x_0),\E u(x_0) + \E \phi_n\left(\frac{x-x_0}{\e}\right)\right)\, dx \\
& =: I_1 + I_2 + I_3.
\end{align*}
By changing variables, using the periodicity of $\phi$ and \eqref{phi}, it follows that
\begin{align*}
I_1 &= \limsup_{n \to +\infty}\int_Q f(\chi(x_0),\E u(x_0) + \E \phi_n(y)) \, dy
= \limsup_{n \to +\infty}\int_Q f(\chi(x_0),\E u(x_0) + \E \phi(ny)) \, dy\\
&= \limsup_{n \to +\infty}\int_Q f(\chi(x_0),\E u(x_0) + \E \phi(x)) \, dx
\leq f(\chi(x_0),\E u(x_0)) + \delta.
\end{align*}
Consequently, to complete the proof it remains to show that $I_2 = I_3 = 0$ and finally to let $\delta \to 0^+$.
To conclude that $I_3 = 0$ we reason exactly as in \cite[Proposition 4.2]{BFT} since
$\chi(x_0)$ is fixed in both terms of the integrand. As for $I_2$, since
$\chi(x_0) = 1$, we have by \eqref{G},
\begin{align*}
I_2 &= \limsup_{\varepsilon \to 0^+}\limsup_{n \to +\infty}\frac{1}{\e^N}
\int_{Q(x_0, \e) \cap \{\chi = 0\}} f\left(0,\E u_{n,\e}(x)\right) -
f\left(1,\E u_{n,\e}(x))\right)\, dx \\
&\leq \limsup_{\varepsilon \to 0^+}\limsup_{n \to +\infty}\frac{C}{\e^N}
\int_{Q(x_0, \e) \cap \{\chi = 0\}} 1 +
\left |\E (u*\rho_n)(x) + \E \phi_n\left(\frac{x-x_0}{\e}\right)\right| \, dx,
\end{align*}
where, by periodicity and the Riemann-Lebesgue Lemma,
\begin{align*}
\limsup_{\varepsilon \to 0^+}\limsup_{n \to +\infty}\frac{C}{\e^N}
\int_{Q(x_0, \e) \cap \{\chi = 0\}}
&\left |\E \phi_n\left(\frac{x-x_0}{\e}\right)\right| \, dx
= \limsup_{\varepsilon \to 0^+}\limsup_{n \to +\infty}
C \int_{Q \cap \{y : \chi(x_0 + \e y) = 0\}}\left |\E\phi(ny)\right| \, dy\\
&\hspace{2cm} = \limsup_{\varepsilon \to 0^+}C \int_{Q \cap \{y : \chi(x_0 + \e y) = 0\}}
\left(\int_Q|\E \phi(x)| \, dx\right) \, dy\\
&\hspace{2cm} = \limsup_{\varepsilon \to 0^+}\frac{C}{\e^N}
\LL^N\left(Q(x_0,\e) \cap \{\chi = 0\}\right)\int_Q|\E \phi(x)| \, dx = 0
\end{align*}
by \eqref{x07}.
On the other hand, since $|E u|$ does not charge the boundary of $Q(x_0,\e)$, using Lemma \ref{lemma2.2BFT}, \eqref{x02}, \eqref{364} and \eqref{x07},
it follows that
\begin{align*}
& \limsup_{\varepsilon \to 0^+}\limsup_{n \to +\infty}\frac{C}{\e^N}
\int_{Q(x_0, \e) \cap \{\chi = 0\}}
\left |\E (u*\rho_n)(x)\right| \, dx \\
& \leq \limsup_{\varepsilon \to 0^+}\limsup_{n \to +\infty}\frac{C}{\e^N}
\int_{Q(x_0, \e + \frac{1}{n}) \cap \{\chi = 0\}}d|E u|(x) \\
&= \limsup_{\varepsilon \to 0^+}\frac{C}{\e^N}
\int_{Q(x_0, \e) \cap \{\chi = 0\}}|\E u(x)| \, dx \\
& \leq \limsup_{\varepsilon \to 0^+}\frac{C}{\e^N}
\int_{Q(x_0, \e)}|\E u(x) - \E u(x_0)| \, dx +
\limsup_{\varepsilon \to 0^+}\frac{C |\E u(x_0)|}{\e^N}
\LL^N\left(Q(x_0, \e) \cap \{\chi = 0\}\right) = 0.
\end{align*}
Therefore, a final application of \eqref{x07} allows us to conclude that $I_2 = 0$.
\end{proof}
\begin{remark}\label{norestsqcx}
{\rm We stress that the symmetric quasiconvexity hypothesis on $f$ in Proposition \ref{upbbulk} is not a restriction for the proof of Theorem \ref{main}, in view of Proposition \ref{Frel=Frel**}.}
\end{remark}
\subsection{The Cantor Term}\label{Cantor}
This section is devoted to the identification of the density of $\mathcal F$ in \eqref{calFint} with respect to $|E^c u|$. To this end, we start by observing that, by virtue of Proposition \ref{Frel=Frel**}, there is no loss of generality in assuming that $f$ is symmetric quasiconvex. If this symmetric quasiconvexity hypothesis on $f$ is omitted, the result of the next proposition holds provided we replace $f^\infty$ by $(SQf)^\infty$, whereas, due to the inequality
$(SQf)^{\infty} \leq f^{\infty}$, \eqref{ubCantor} holds as stated.
\begin{proposition}\label{lbCantor}
Let $u \in BD(\Omega)$, $\chi \in BV(\Omega;\{0,1\})$ and let $W_0$ and $W_1$ be continuous functions satisfying \eqref{growth}.
Assume that $f$ given by \eqref{density} is symmetric quasiconvex.
Then, for $|E^cu|$ a.e. $x_0 \in \O$,
$$ \mu^c(x_0) = \dfrac{d\mathcal F(\chi, u; \cdot)}{d |E^cu|}(x_0) \geq f^{\infty}\left(\chi(x_0), \frac{d E^c u}{d |E^c u|}(x_0)\right).$$
\end{proposition}
\begin{proof}
Let $x_0 \in \O$ be a point satisfying \eqref{uac}, \eqref{chiad} and
\begin{equation}\label{Cdens}
\mu^c(x_0) = \dfrac{d\mathcal F(\chi, u; \cdot)}{d |E^cu|}(x_0)
= \frac{d \mu}{d |E^cu|}(x_0) =
\lim_{\e \to 0^+}\frac{\mu(Q(x_0,\e))}{|E^cu|(Q(x_0,\e))}
\; \; \mbox{exists and is finite,}
\end{equation}
these properties hold for $|E^c u|$ a.e. $x_0 \in \Omega$. Indeed, by \cite[Theorem 6.1]{ACDM},
$|E u|(S_u\setminus J_u)=0$, thus $|E^c u|(S_u\setminus J_u)=0$. Hence, by \cite[Propositions 3.5 and 4.4]{ACDM}, we have
$$|E^c u|(S_u)= |E^c u|(J_u)+ |E^c u|(S_u\setminus J_u)=0,$$ which justifies the validity of \eqref{uac}. As for \eqref{chiad}, this is a well known property of $BV$ functions (cf. \cite{AFP}).
We define
$$f_0(\xi) = f(0, \xi) \mbox{ and } f_1(\xi) = f(1, \xi), \forall \xi \in \Rb^{N \times N}_s$$
and we consider the auxiliary functionals
\begin{align}\label{auxf}
\mathcal F_i(u; A) &:= \inf \Big \{\liminf_{n \to + \infty}
\int_A f_i (\E u_n(x)) \, dx : u_n \in W^{1, 1}(A; \Rb^N), u_n \to u \; {\rm in} \; L^1(A;\Rb^N)\Big \}, \, i=0,1.
\end{align}
Referring to Theorem 6.1, Remark 6.4 and Corollary 6.8 in \cite{CFVG},
$\mathcal F_i(u;\cdot)$, $i=0,1$,
are the restriction to $\mathcal O(\Omega)$ of Radon measures whose densities with respect to $|E^c u|$ are given by
\begin{equation}\label{CdensFi}
\frac{d \mathcal F_i(u;\cdot)}{d|E^c u|}(x_0)
= f_{i}^\infty \Big(\frac{d E^c u}{d|E^c u|}(x_0)\Big) =
f^\infty \Big(i,\frac{d E^c u}{d|E^c u|}(x_0)\Big)
\end{equation}
for $|E^c u|$ a.e. $x_0 \in \Omega$. Choose $x_0$ so that it also satisfies \eqref{CdensFi}, $i=0,1$.
In what follows we assume, without loss of generality, that $\chi(x_0)=1$, the case $\chi(x_0)=0$ can be treated similarly. Bearing this choice in mind we work with the
functional \eqref{auxf} and we will make use of \eqref{CdensFi}, with $i=1$.
\color{black}
Selecting the sequence $\e_k \to 0^+$ in such a way that
$\mu(\partial Q(x_0,\e_k)) = 0$ and $Q(x_0,\e_k) \subset \Omega$, we have
\begin{align*}
\mu^c(x_0) &= \lim_{k \to +\infty}\frac{\mu (Q(x_0,\e_k))}{|E^cu|(Q(x_0,\e_k))} \\
&= \lim_{k,n}\left[\frac{1}{|E^cu|(Q(x_0,\e_k))}\int_{Q(x_0,\e_k)}
f(\chi_n(x),\mathcal E u_n(x))\,dx + |D \chi_n|(Q(x_0,\e_k))\right]
\end{align*}
where $\chi_n \in BV(Q(x_0,\e_k);\{0,1\})$, $\chi_n \to \chi$ in $L^1(Q(x_0,\e_k);\{0,1\})$, $u_n \in W^{1,1}(Q(x_0,\e_k);\Rb^N)$, $u_n \to u$ in $L^1(Q(x_0,\e_k);\Rb^N)$.
Taking into account that we are searching for a lower bound for $\mu^c(x_0)$,
we neglect the perimeter term $|D \chi_n|(Q(x_0,\e_k))$ and obtain
\begin{align}\label{lbmuc}
\mu^c(x_0) & \geq
\liminf_{k,n}\frac{1}{|E^cu|(Q(x_0,\e_k))}
\int_{Q(x_0,\e_k)}f(\chi_n(x),\mathcal E u_n(x))\,dx \\
& \geq
\liminf_{k,n}\frac{1}{|E^cu|(Q(x_0,\e_k))}
\int_{Q(x_0,\e_k)}f_1(\mathcal E u_n(x))\,dx \nonumber\\
& \hspace{3cm} + \liminf_{k,n}\frac{1}{|E^cu|(Q(x_0,\e_k))}
\int_{Q(x_0,\e_k)}f(\chi_n(x),\mathcal E u_n(x)) - f(1,\mathcal E u_n(x))\,dx \nonumber\\
& \geq \liminf_k \frac{\mathcal F_1(u;Q(x_0,\e_k))}{|E^cu|(Q(x_0,\e_k))}+ \liminf_{k,n}I_{k,n} \nonumber\\
& \geq \frac{d \mathcal F_{1}(u;\cdot)}{d|E^c u|}(x_0) + \liminf_{k,n}I_{k,n}
\nonumber\\
& \geq f^\infty \Big(1,\frac{d E^c u}{d|E^c u|}(x_0)\Big) + \liminf_{k,n}I_{k,n}
\label{lbmuc2}
\end{align}
where
$$I_{k,n} = \frac{1}{|E^cu|(Q(x_0,\e_k))}
\int_{Q(x_0,\e_k)}f(\chi_n(x),\mathcal E u_n(x)) - f(1,\mathcal E u_n(x))\,dx.$$
It remains to estimate this term. Changing variables we get
\begin{align}\label{Ikn}
\left|I_{k,n}\right| &= \left|\frac{\e_k^N}{|E^cu|(Q(x_0,\e_k))}
\int_Q f(\chi_n(x_0+\e_k y),\mathcal E u_n(x_0+\e_k y)) -
f(1, \mathcal E u_n(x_0+\e_k y))\,dy\right| \nonumber \\
&= \left|\delta_k
\int_Q f(\chi_{n,k}(y) + 1,\mathcal E u_{n,k}(y)) -
f(1, \mathcal E u_{n,k}(y))\,dy\right|
\end{align}
where
$$\displaystyle \delta_k : = \frac{\e_k^N}{|E^cu|(Q(x_0,\e_k))}, \; \;
\chi_{n,k}(y) := \chi_n(x_0+\e_k y) - 1, \; \;
u_{n,k}(y) := \frac{u_n(x_0+\e_k y) - u(x_0)}{\e_k}.$$
By \eqref{chiad} it follows that
$\displaystyle \lim_{k,n}\|\chi_{n,k}\|_{L^1(Q)} = 0$ (see \eqref{chindeltalim})
and $\displaystyle \lim_k \delta_k = 0$.
Thus, using also \eqref{G}, we have from \eqref{Ikn}
\begin{align}\label{Ikn2}
\liminf_{k,n}\left|I_{k,n}\right| &\leq \limsup_{k,n} \delta_k
\int_Q\left | f(\chi_{n,k}(y) + 1,\mathcal E u_{n,k}(y)) -
f(1, \mathcal E u_{n,k}(y))\right|\,dy \nonumber \\
&\leq \limsup_{k,n} C \delta_k
\int_Q |\chi_{n,k}(y)|\big(1 + |\mathcal E u_{n,k}(y)|\big) \, dy \nonumber \\
& = \limsup_{k,n} C \delta_k
\int_Q |\chi_{n,k}(y)| \, |\mathcal E u_{n,k}(y)| \, dy.
\end{align}
From the growth condition from below on $f$, \eqref{lbmuc} and \eqref{Cdens} we conclude that
\begin{align}
\limsup_{k,n} C \delta_k
\int_Q |\chi_{n,k}(y)| \, |\mathcal E u_{n,k}(y)| \, dy &\leq
\limsup_{k,n} \frac{C}{|E^cu|(Q(x_0,\e_k))}
\int_{Q(x_0,\e_k)}|\mathcal E u_{n}(x)| \, dx \nonumber \\
&\leq \limsup_{k,n} \frac{C}{|E^cu|(Q(x_0,\e_k))}
\int_{Q(x_0,\e_k)}f(\chi_n(x),\mathcal E u_{n}(x)) \, dx \nonumber \\
&\leq C \mu^c(x_0) < + \infty. \nonumber
\end{align}
Using a diagonalisation argument, let $\chi_k := \chi_{n(k),k}$,
$u_k := u_{n(k),k}$ be such that $\chi_k \to 0$ in $L^1(Q)$ and
\begin{equation}\label{Ikn3}
\limsup_{k,n}C \delta_k
\int_Q |\chi_{n,k}(y)| \, |\mathcal E u_{n,k}(y)| \, dy =
\lim_{k}C \delta_k
\int_Q |\chi_{k}(y)| \, |\mathcal E u_{k}(y)| \, dy < + \infty.
\end{equation}
Therefore, the sequence $\{\delta_k \chi_k \, \mathcal E u_{k}\}$ is bounded in $L^1(Q;\Rb^{N \times N}_s)$
so, by the Biting Lemma, there exists a subsequence (not relabeled) and there exist
sets $D_r \subset Q$ such that
$\displaystyle \lim_{r \to + \infty}{\mathcal L}^N(D_r) = 0$ and the sequence
$\{\delta_k \chi_k \, \mathcal Eu_k\}$ is equiintegrable in $Q\setminus D_r$,
for any $r \in \mathbb N$. Following the reasoning in the proof of
Proposition~\ref{lbbulk} (see \eqref{77}), for any $j \in \Nb$ there exist $k(j), r(j) \in \Nb$ such that
\begin{equation}\label{Ikn4}
\delta_k(j)
\int_{D_{r(j)}} |\chi_{k(j)}(y)| \, |\mathcal E u_{k(j)}(y)| \, dy \leq \frac{1}{j}.
\end{equation}
The fact that $\chi_{k(j)} \to 0$, as $j \to + \infty$, in $L^1(Q)$ and the equiintegrability of $\{\delta_{k(j)} \chi_{k(j)} \, \mathcal Eu_{k(j)}\}$ in
$Q\setminus D_{r(j)}$ ensures that, for any $\e > 0$,
\begin{equation}\label{Ikn5}
\delta_k(j)
\int_{Q \setminus D_{r(j)}} |\chi_{k(j)}(y)| \, |\mathcal E u_{k(j)}(y)| \, dy < \e,
\end{equation}
provided $j$ is large enough (see the argument used to obtain \eqref{11}).
Hence, from \eqref{lbmuc2}, \eqref{Ikn2}, \eqref{Ikn3}, \eqref{Ikn4} and \eqref{Ikn5}
we conclude that
$$\mu^c(x_0) \geq f^\infty \Big(1,\frac{d E^c u}{d|E^c u|}(x_0)\Big),$$
which completes the proof.
\end{proof}
\begin{proposition}
Let $u \in BD(\Omega)$, $\chi \in BV(\Omega;\{0,1\})$ and let $W_0$ and $W_1$ be continuous functions satisfying \eqref{growth}.
Assume that $f$ given by \eqref{density} is symmetric quasiconvex.
Then, for $|E^cu|$ a.e. $x_0 \in \O$,
\begin{equation}\label{ubCantor}
\mu^c(x_0) = \dfrac{d\mathcal F(\chi, u; \cdot)}{d |E^cu|}(x_0) \leq f^{\infty}\left(\chi(x_0), \frac{d E^c u}{d |E^c u|}(x_0)\right).
\end{equation}
\end{proposition}
\begin{proof}
Let $x_0 \in \Omega$ be a point satisfying \eqref{Cdens}, \eqref{uac}, \eqref{chiad} (which hold for $|E^cu|$ a.e. $x \in \Omega$, as observed in the proof of Proposition \ref{lbCantor}) and, in addition,
\begin{equation}\label{Dchi}
\lim_{\e \to 0^+}\frac{|D\chi|(Q(x_0,\e))}{|E^cu|(Q(x_0,\e))} = 0,
\end{equation}
Assuming, once again, that $\chi(x_0) = 1$, we also require that $x_0$ satisfies \eqref{CdensFi}.
Choosing the sequence $\e_k \to 0^+$ in such a way that
$\mu(\partial Q(x_0,\e_k)) = 0$ and $Q(x_0,\e_k) \subset \Omega$,
let $u_n \in W^{1,1}(Q(x_0,\e_k);\Rb^N)$ be such that $u_n \to u$ in $L^1(Q(x_0,\e_k);\Rb^N)$ and
\begin{equation}\label{F1Cdens}
\dfrac{d\mathcal F_1(u; \cdot)}{d |E^cu|}(x_0) =
\lim_{k \to +\infty}\frac{\mathcal F_1(u;Q(x_0,\e_k))}{|E^cu|(Q(x_0,\e_k))}
= \lim_{k,n}\frac{1}{|E^cu|(Q(x_0,\e_k))}
\int_{Q(x_0,\e_k)}f_1(\mathcal E u_n(x)) \, dx.
\end{equation}
Then, as the constant sequence $\chi_n = \chi$ is admissible for
$\mathcal F(\chi,u;Q(x_0,\e_k))$, from \eqref{Dchi}, \eqref{F1Cdens} and \eqref{CdensFi} with $i=1$, it follows that
\begin{align*}
\mu^c(x_0) &= \lim_{k \to +\infty}
\frac{\mathcal F(\chi,u;Q(x_0,\e_k))}{|E^cu|(Q(x_0,\e_k))} \\
&\leq \liminf_{k,n}\left[\frac{1}{|E^cu|(Q(x_0,\e_k))}\int_{Q(x_0,\e_k)}
f(\chi(x),\mathcal E u_n(x))\,dx + |D \chi|(Q(x_0,\e_k))\right] \\
&\leq \lim_{k,n}\frac{1}{|E^cu|(Q(x_0,\e_k))}\int_{Q(x_0,\e_k)}
f(1,\mathcal E u_n(x))\,dx \\
&\hspace{3cm} + \limsup_{k,n}\frac{1}{|E^cu|(Q(x_0,\e_k))}\int_{Q(x_0,\e_k)}
f(\chi(x),\mathcal E u_n(x)) - f(1,\mathcal E u_n(x)) \,dx \\
& = f^{\infty}\left(\chi(x_0), \frac{d E^c u}{d |E^c u|}(x_0)\right) + \limsup_{k,n}\frac{1}{|E^cu|(Q(x_0,\e_k))}\int_{Q(x_0,\e_k)}
\hspace{-0,4cm}f(\chi(x),\mathcal E u_n(x)) - f(1,\mathcal E u_n(x)) \,dx.
\end{align*}
The same argument used in the proof of Proposition~\ref{lbCantor}, now applied to the sequences
$$\chi_{k}(y) = \chi(x_0 + \e_ky) - 1, \; \;
u_{n,k}(y) := \frac{u_n(x_0+\e_k y) - u(x_0)}{\e_k},$$ yields
$$\limsup_{k,n}\frac{1}{|E^cu|(Q(x_0,\e_k))}\int_{Q(x_0,\e_k)}
f(\chi(x),\mathcal E u_n(x)) - f(1,\mathcal E u_n(x)) \,dx = 0$$
from which the conclusion follows.
\end{proof}
\subsection{The Surface Term}\label{surface}
Given $x_0 \in J_\chi \cup J_u$ we denote by $\nu(x_0)$ the vector $\nu_u(x_0)$,
if $x_0 \in J_u \setminus J_\chi$, whereas $\nu(x_0) := \nu_{\chi}(x_0)$ if
$x_0 \in J_\chi \setminus J_u$, these vectors are well defined as Borel measurable functions for $\cH^{N-1}$ a.e. $x_0 \in J_\chi \cup J_u$. Due to the rectifiability of both $J_\chi$ and $J_u$ (cf. \cite[Theorems 3.77 and 3.78]{AFP} and \cite[Proposition 3.5 and Remark 3.6]{ACDM}),
for $\cH^{N-1}$ a.e. $x_0 \in J_\chi \cap J_u$ we may select
$\nu(x_0) := \nu_{\chi}(x_0) = \nu_u(x_0)$ where the orientation of $\nu_{\chi}(x_0)$ is
chosen so that $\chi^+(x_0) = 1$, $\chi^-(x_0) = 0$ and then $u^+(x_0)$ and $u^-(x_0)$ are selected according to this orientation.
Thus, in the sequel for $\mathcal H^{N-1}$ a.e. $x_0 \in J_{\chi} \cup J_u$, the vector $\nu(x_0)$ is defined according to the above considerations.
Given that $\mathcal H^{N-1}(S_\chi \setminus J_\chi) = 0$ and that all points in $\Omega \setminus S_\chi$ are Lebesgue points of $\chi$, in what follows we take
$\chi^+(x_0) = \chi^-(x_0) = \widetilde\chi(x_0)$ for
$\mathcal H^{N-1}$- a.e. $x_0 \in J_u \setminus J_\chi$, where $\tilde v$ denotes the precise representative of a field $v$ in $BV$, cf. Section \ref{BV}. On the other hand, for a $BD$ function $u$ it is not known whether
$\mathcal H^{N-1}(S_u \setminus J_u) = 0$. However, given that all points in
$\Omega \setminus S_u$ are Lebesgue points of $u$ and that, by \cite[Remark 6.3]{ACDM} and the $\mathcal H^{N-1}$ rectifiability of $J_\chi$,
$\mathcal H^{N-1}(S_u \setminus J_u)\cap J_\chi)=0$, we may consider
$u^+(x_0) = u^-(x_0) = \widetilde u(x_0)$ for $\mathcal H^{N-1}$ a.e.
$x_0 \in J_\chi \setminus J_u$, where $\tilde v$ denotes the Lebesgue representative of a field $v$ in $BD$ (cf. \cite[page 206]{ACDM}), see also \cite{B}.
In order to describe $\mu^j$ we will follow the ideas of the global method for relaxation introduced in \cite{BFM} (see also \cite{BFT} and \cite{CFVG}), the sequential characterisation of $K(a,b,c,d, \nu)$, obtained in Proposition \ref{Ktilde}, will also be used.
Given $u \in BD(\O)$, $\chi \in BV(\O;\{0,1\})$ and $V \in \mathcal O_{\infty}(\O)$
we define
\begin{equation}\label{m}
m(\chi,u;V) := \inf\left\{\mathcal F (\theta,v;V) :
\theta \in BV(\O;\{0,1\}), v \in BD(\O), \theta = \chi \mbox{ on } \partial V,
v = u \mbox{ on } \partial V\right\}.
\end{equation}
Our goal is to show the following result.
\begin{proposition}\label{mg}
Let $f$ be given by \eqref{density}, where $W_0$ and $W_1$ are continuous functions satisfying \eqref{growth}.
Given $u \in SBD(\Omega)$ and $\chi \in BV(\Omega;\{0,1\})$, we have
$$\mathcal{F}( \chi, u; V \cap (J_\chi \cup J_u)) =
\int_{V \cap (J_\chi \cup J_u)}
g(x,\chi^+(x),\chi^-(x),u^+(x),u^-(x),\nu(x))\, d\cH^{N-1}(x),$$
where
\begin{equation}\label{defg}
g(x_0,a,b,c,d,\nu) := \limsup_{\varepsilon \to 0^+}
\frac{m(\chi_{a,b,\nu}(\cdot - x_0),u_{c,d,\nu}(\cdot - x_0);Q_\nu(x_0,\e))}
{\e^{N-1}}
\end{equation}
and $\chi_{a,b,\nu}$, $u_{c,d,\nu}$ were defined in \eqref{targets}.
\end{proposition}
The proof of the above proposition relies on a series of auxiliary results, based on Lemmas 3.1, 3.3 and 3.5 in \cite{BFM} and which were adapted to the $BD$ case in \cite{BFT}[Lemmas 3.10, 3.11 and 3.12]. The properties of $\mathcal F(\chi,u;A)$ established in Proposition \ref{firstprop}, and the fact that
$\mathcal F(\chi,u;\cdot)$ is a Radon measure, ensure that we can apply the reasoning given in their respective proofs.
\begin{lemma}\label{BFT310}
Let $f$ be given by \eqref{density}, where $W_0$ and $W_1$ are continuous functions satisfying \eqref{growth}.
Then there exists a positive constant $C$ such that
$$|m(\chi_1,u_1;V) - m(\chi_2,u_2;V)| \leq
C\left[\int_{\partial V}|{\rm tr } \, \chi_1(x) - {\rm tr}\,\chi_2(x)| +
|{\rm tr}\,u_1(x) - {\rm tr}\,u_2(x)| \, d\cH^{N-1}(x) \right],$$
for every $\chi_1,\chi_2 \in BV(\O;\{0,1\})$, $u_1,u_2 \in BD(\O)$ and any
$V \in \mathcal O_{\infty}(\O).$
\end{lemma}
\begin{proof}
The proof follows that of Lemma 3.10 in \cite{BFT}. Given $\delta > 0$ let
$V_\delta := \{x \in V : {\rm dist}(x,\partial V) > \delta\}$ and select
$\theta \in BV(\O;\{0,1\})$ and $v \in BD(\O)$ such that $\theta = \chi_2$ and
$v = u_2$ on $\partial V$. Now define $\theta_\delta \in BV(\O;\{0,1\})$ and
$v_\delta \in BD(\O)$ by
$$\theta_\delta : = \begin{cases}
\theta, & {\rm in} \; V_\delta\\
\chi_1, & {\rm in} \; \O \setminus V_\delta
\end{cases} \; \; \; {\rm and} \; \; \;
v_\delta : = \begin{cases}
v, & {\rm in} \; V_\delta\\
u_1, & {\rm in} \; \O \setminus V_\delta.
\end{cases}
$$
The definition of $m(\cdot,\cdot;\cdot)$ and the additivity and locality of
$\mathcal F(\cdot,\cdot;\cdot)$, as well as the inequality from above in Proposition
\ref{firstprop} $i)$, lead to the conclusion.
\end{proof}
Fixing $\chi \in BV(\O;\{0,1\})$, $u \in BD(\O)$ and
$\nu \in S^{N-1}$, we define $\lambda := \LL^N + |E^s u| + |D\chi|$ and, following \cite{BFM}, we let
$$\mathcal O^*:= \left\{Q_\nu(x,\e) : x \in \O,\, \e > 0\right\}$$
and, for $\delta > 0$ and $V \in \mathcal O(\O)$, set
\begin{align*}
m^\delta(\chi,u;V) &:= \inf\Big\{\sum_{i = 1}^{+\infty}m(\chi,u;Q_i) :
Q_i \in \mathcal O^*, Q_i \cap Q_j = \emptyset \; {\rm if} \; i \neq j,\\
& \hspace{4cm} Q_i \subset V, {\rm diam} \, Q_i < \delta,
\lambda\left(V \setminus \displaystyle\bigcup_{i=1}^{+\infty}Q_i\right) = 0\Big\}.
\end{align*}
Clearly, $\delta \mapsto m^\delta(\chi,u;V)$ is a decreasing function, so we define
$$m^*(\chi,u;V) := \sup \left\{m^\delta(\chi,u;V) : \delta > 0\right\}
= \lim_{\delta \to 0^+}m^\delta(\chi,u;V).$$
\begin{lemma}\label{BFT311}
Let $f$ be given by \eqref{density}, where $W_0$ and $W_1$ are continuous functions satisfying \eqref{growth}.
Given $\chi \in BV(\O;\{0,1\})$, $u \in BD(\O)$, we have
$$\mathcal F(\chi,u;V) = m^*(\chi,u;V), \; \mbox{ for every } V \in \mathcal O(\O).$$
\end{lemma}
\begin{proof}
The inequality
$$m^*(\chi,u;V) \leq \mathcal F(\chi,u;V)$$ is an immediate consequence
of the fact that $m(\chi,u;Q_i) \leq \mathcal F(\chi,u;Q_i)$ and that
$\mathcal F(\chi,u;\cdot)$ is a Radon measure.
The proof of the reverse inequality relies on the lower semicontinuity of
$\mathcal F(\cdot,\cdot;V)$ obtained in
Proposition \ref{firstprop} $iv)$ and on the definitions of $m^\delta(\chi,u;V)$, $m(\chi,u;V)$ and $m^*(\chi,u;V)$. Indeed, fixing $\delta > 0$, we consider $(Q_i^\delta)$ an admissible family for $m^\delta(\chi,u;V)$ such that,
letting $\displaystyle N^\delta := V \setminus \displaystyle\cup_{i=1}^{+\infty}Q_i^\delta$,
$$\sum_{i =1}^{+\infty}m(\chi,u;Q_i^\delta) < m^\delta(\chi,u;V) + \delta
\; \mbox{ and } \; \lambda(N^\delta) = 0,$$
and we now let $\theta_i^\delta \in BV(\O;\{0,1\})$ and $v_i^\delta \in BD(\O)$
be such that $\theta_i^\delta = \chi$ on $\partial Q_i^\delta$, $v_i^\delta = u$ on
$\partial Q_i^\delta$ and
$$\mathcal F(\theta_i^\delta,v_i^\delta;Q_i^\delta) \leq
m(\chi,u;Q_i^\delta) + \delta \LL^N(Q_i^\delta).$$
Setting
$\displaystyle N_0^\delta := \O\setminus\displaystyle\cup_{i=1}^{+\infty}Q_i^\delta$,
we define
$$\theta^\delta := \sum_{i =1}^{+\infty}\theta_i^\delta \,\chi_{Q_i^\delta}
+ \chi \,\chi_{N_0^\delta} \;\;\; \mbox{ and } \; \;\;
v^\delta := \sum_{i =1}^{+\infty}v_i^\delta \,\chi_{Q_i^\delta}
+ u \,\chi_{N_0^\delta}.$$
Following the computations in the proof of \cite[Lemma 3.11]{BFT}, we may show that $\theta^\delta \in BV(\O;\{0,1\})$, $v^\delta \in BD(\O)$,
$\theta^\delta \to \chi$ in $L^1(V;\{0,1\})$ and
$v^\delta \to u$ in $L^1(V;\Rb^N)$, as $\delta \to 0^+$, and also
$$\mathcal F(\theta^\delta,v^\delta;N^\delta) \leq C \lambda(N^\delta) = 0.$$
Using the additivity of $\mathcal F(\theta^\delta,v^\delta;\cdot)$ we have
\begin{align*}
\mathcal F(\theta^\delta,v^\delta;V) &= \sum_{i =1}^{+\infty}
\mathcal F(\theta_i^\delta,v_i^\delta;Q_i^\delta) +
\mathcal F(\theta^\delta,v^\delta;N^\delta)\\
&\leq \sum_{i =1}^{+\infty}m(\chi,u;Q_i^\delta) + \delta \LL^N(V)
\leq m^\delta(\chi,u;V) + \delta + \delta \LL^N(V),
\end{align*}
so that the lower semicontinuity of $\mathcal F(\cdot,\cdot;V)$ yields
\begin{align*}
\mathcal F(\chi,u;V) &\leq \liminf_{\delta \to 0^+}
\mathcal F(\theta^\delta,v^\delta;V) \\
&\leq \liminf_{\delta \to 0^+}
\left( m^\delta(\chi,u;V) + \delta + \delta \LL^N(V)\right) = m^*(\chi,u;V)
\end{align*}
and this completes the proof.
\end{proof}
Finally, a straightforward adaptation of \cite[Lemma 3.5]{BFM} leads to the following
result.
\begin{lemma}\label{BFT312}
Let $f$ be given by \eqref{density}, where $W_0$ and $W_1$ are continuous functions satisfying \eqref{growth}.
Given $\chi \in BV(\O;\{0,1\})$, $u \in BD(\O)$, we have
$$\lim_{\e \to 0^+}\frac{\mathcal F(\chi,u;Q_\nu(x_0,\e))}{\lambda(Q_\nu(x_0,\e))} = \lim_{\e \to 0^+}\frac{m(\chi,u;Q_\nu(x_0,\e))}{\lambda(Q_\nu(x_0,\e))},$$
for $\lambda$ a.e. $x_0 \in \O$ and for every $\nu \in S^{N-1}$.
\end{lemma}
We now proceed with the proof of Proposition \ref{mg}.
\begin{proof}[Proof of Proposition \ref{mg}]
In the sequel, for simplicity of notation, we will write $\nu = \nu(x_0)$.
Let $x_0 \in \O \cap (J_\chi \cup J_u)$ be a point satisfying
\begin{equation}\label{chitilde}
\lim_{\varepsilon \rightarrow 0^+}\frac{1}{\e^N}
\int_{Q_{\nu}(x_0,\e)}|\chi(x) - \widetilde{\chi}(x_0)| \, dx = 0,
\; {\rm if} \;x_0 \in \Omega \setminus J_{\chi},
\end{equation}
\begin{equation}\label{chipm}
\lim_{\varepsilon \rightarrow 0^+}\frac{1}{\e^N}
\int_{Q^+_{\nu}(x_0,\e)}\hspace{-0,2cm}|\chi(x) - {\chi}^+(x_0)| \, dx =
\lim_{\varepsilon \rightarrow 0^+}\frac{1}{\e^N}
\int_{Q^-_{\nu}(x_0,\e)}\hspace{-0,2cm}|\chi(x) - {\chi}^-(x_0)| \, dx =0,
\; {\rm if} \;x_0 \in \Omega \cap J_{\chi},
\end{equation}
\begin{equation}\label{utilde}
\lim_{\varepsilon \rightarrow 0^+}\frac{1}{\e^N}
\int_{Q_{\nu}(x_0,\e)}|u(x) - \widetilde{u}(x_0)| \, dx = 0,
\; {\rm if} \;x_0 \in \Omega \setminus J_{u},
\end{equation}
\begin{equation}\label{upm}
\lim_{\varepsilon \rightarrow 0^+}\frac{1}{\e^N}
\int_{Q^+_{\nu}(x_0,\e)}\hspace{-0,2cm}|u(x) - {u}^+(x_0)| \, dx =
\lim_{\varepsilon \rightarrow 0^+}\frac{1}{\e^N}
\int_{Q^-_{\nu}(x_0,\e)}\hspace{-0,2cm}|u(x) - {u}^-(x_0)| \, dx =0,
\; {\rm if} \;x_0 \in \Omega \cap J_{u},
\end{equation}
where
$$Q^{\pm}_{\nu}(x_0,\e) =
\left\{x \in Q_{\nu}(x_0,\e) : (x - x_0) \cdot (\pm \nu) > 0\right\},$$
and
\begin{equation}\label{muj}
\mu^j(x_0) = \lim_{\varepsilon \rightarrow 0^+}
\frac{\mathcal F(\chi,u;Q_{\nu}(x_0,\e))}
{\mathcal H^{N-1} \lfloor (J_\chi \cup J_u)(Q_{\nu}(x_0,\e))} =
\lim_{\varepsilon \rightarrow 0^+}\frac{1}{\e^{N-1}}\int_{Q_{\nu}(x_0,\e)}d\mu(x) \;
\mbox{ exists and is finite}.
\end{equation}
In view of the considerations made at the beginning of this subsection, these properties hold for $\mathcal H^{N-1}$ a.e.
$x_0 \in \Omega\cap (J_\chi \cup J_u)$.
Furthermore, we require that $x_0$ also satisfies
\begin{equation}\label{symjump}
\lim_{\e \to 0^+}\frac{1}{\e^{N-1}}|Eu|(Q_{\nu}(x_0,\e)) =
|([u]\odot \nu)(x_0)| = |Eu_0|(Q_\nu)
\end{equation}
and
\begin{equation}\label{perchi0}
\lim_{\e \to 0^+}\frac{1}{\e^{N-1}}|D\chi|(Q_{\nu}(x_0,\e)) = 1 = |D\chi_0|(Q_\nu),
\end{equation}
where we are denoting by $\chi_0$ and $u_0$ the functions given by
\eqref{targets} with $\nu = \nu(x_0)$ and $a=\chi^+(x_0)$,
$b=\chi^-(x_0)$, $c=u^+(x_0)$ and $d=u^-(x_0)$.
Letting $\sigma := \cH^{N-1}\lfloor (J_\chi \cup J_u)$, by Lemma \ref{BFT312}
it follows that, for $\sigma$ a.e. $x_0 \in \O$,
\begin{equation}\label{msigma}
\frac{d\mathcal F(\chi,u;\cdot)}{d \sigma}(x_0) =
\lim_{\varepsilon \to 0^+}
\frac{\mathcal F(\chi,u;Q_\nu(x_0,\e))}{\sigma(Q_\nu(x_0,\e))} =
\lim_{\varepsilon \to 0^+}
\frac{m(\chi,u;Q_\nu(x_0,\e))}{\sigma(Q_\nu(x_0,\e))}.
\end{equation}
Let $\chi_{\e}: Q_\nu \to \{0,1\}$ and $u_\e: Q_\nu \to \Rb^N$ be defined by
$\chi_{\e}(y) := \chi(x_0 + \e y)$, $u_\e(y) :=u(x_0 + \e y)$. Properties \eqref{chitilde} or \eqref{chipm}, and \eqref{utilde} or \eqref{upm}, respectively,
guarantee that $\chi_{\e} \to \chi_0$ in $L^1(Q_\nu;\{0,1\})$ and
$u_\e \to u_0$ in $L^1(Q_\nu;\Rb^N)$. On the other hand, by \eqref{symjump} and
\eqref{perchi0} we have
$$\lim_{\e \to 0^+}|Eu_\e|(Q_\nu) =
\lim_{\e \to 0^+}\frac{1}{\e^{N-1}}|Eu|(Q_\nu(x_0,\e))
= |([u]\odot \nu)(x_0)| = |Eu_0|(Q_\nu)$$
and
$$\lim_{\e \to 0^+}|D\chi_\e|(Q_\nu) =
\lim_{\e \to 0^+}\frac{1}{\e^{N-1}}|D\chi|(Q_\nu(x_0,\e)) = |D\chi_0|(Q_\nu).$$
Due to the continuity of the trace operator with respect to the intermediate topology
we conclude that
\begin{align}\label{trcont}
& \lim_{\e \to 0^+}\frac{1}{\e^{N-1}}\int_{\partial Q_{\nu}(x_0,\e)}
|{\rm tr } \, \chi(x) - {\rm tr}\,\chi_0(x-x_0)| +
|{\rm tr}\,u(x) - {\rm tr}\,u_0(x-x_0)| \, d\cH^{N-1}(x) \nonumber \\
& = \lim_{\e \to 0^+}\int_{\partial Q_{\nu}}
|{\rm tr } \, \chi_\e(y) - {\rm tr}\,\chi_0(y)| +
|{\rm tr}\,u_\e(y) - {\rm tr}\,u_0(y)| \, d\cH^{N-1}(y) = 0.
\end{align}
Hence, from \eqref{msigma}, \eqref{muj}, Lemma \ref{BFT310} and \eqref{trcont}, we obtain
\begin{align*}
&\frac{d\mathcal F(\chi,u;\cdot)}{d \sigma}(x_0) =
\lim_{\varepsilon \to 0^+}
\frac{m(\chi,u;Q_\nu(x_0,\e))}{\sigma(Q_\nu(x_0,\e))} \\
& = \lim_{\varepsilon \to 0^+}\frac{m(\chi,u;Q_\nu(x_0,\e))-
m(\chi_0(\cdot - x_0),u_0(\cdot - x_0);Q_\nu(x_0,\e))
+ m(\chi_0(\cdot - x_0),u_0(\cdot - x_0);Q_\nu(x_0,\e))}{\e^{N-1}}\\
& = \lim_{\varepsilon \to 0^+}
\frac{m(\chi_0(\cdot - x_0),u_0(\cdot - x_0);Q_\nu(x_0,\e))}{\e^{N-1}}
\end{align*}
and, therefore,
\begin{align*}
\mathcal{F}( \chi, u; V \cap (J_\chi \cup J_u)) &=
\int_{V \cap (J_\chi \cup J_u)}
\frac{d\mathcal F(\chi,u;\cdot)}{d \sigma}(x) \, d\sigma(x) \\
&=\int_{V \cap (J_\chi \cup J_u)}
g(x,\chi^+(x),\chi^-(x),u^+(x),u^-(x),\nu(x))\, d\cH^{N-1}(x).
\end{align*}
\end{proof}
In the final two propositions we will show that, under assumption \eqref{finfty}, the surface energy density $g(x_0,a,b,c,d,\nu)$
may be more explicitly characterised. For this purpose we need an additional lemma
which states that more regular functions can be considered in the definition of the
Dirichlet functional
$m(\chi,u;V)$ in \eqref{m}. In what follows, for
$u \in BD(\O)$, $\chi \in BV(\O;\{0,1\})$ and $V \in \mathcal O_{\infty}(\O)$
we define
$$m_0(\chi,u;V) := \inf\left\{F (\theta,v;V) :
\theta \in BV(\O;\{0,1\}), v \in W^{1,1}(\O;\Rb^N), \theta = \chi \mbox{ on } \partial V,
v = u \mbox{ on } \partial V\right\}.$$
\begin{lemma}\label{BFT314}
Let $f$ be given by \eqref{density}, where $W_0$ and $W_1$ are continuous functions satisfying \eqref{growth}.
Given $\chi \in BV(\O;\{0,1\})$, $u \in BD(\O)$, we have
$$m(\chi,u;V) = m_0(\chi,u;V), \; \mbox{ for every } V \in \mathcal O_{\infty}(\O).$$
\end{lemma}
\begin{proof}
The inequality $m(\chi,u;V) \leq m_0(\chi,u;V)$ is clear since, given any
$\theta \in BV(\O;\{0,1\})$ such that $\theta = \chi \mbox{ on } \partial V$ and any $v \in W^{1,1}(\O;\Rb^N)$ such that
$v = u \mbox{ on } \partial V$, we have
$$m(\chi,u;V) \leq \mathcal F(\theta,v;V) \leq F(\theta,v;V).$$
To show the reverse inequality, we fix $\e > 0$ and let $\theta \in BV(\O;\{0,1\}),
v \in BD(\O)$ be such that
$\theta = \chi \mbox{ on } \partial V, v = u \mbox{ on } \partial V$ and
$$m(\chi,u;V) + \e \geq \mathcal F(\theta,v;V).$$
By Proposition~\ref{firstprop} $ii)$, let $\chi_n \in BV(\O;\{0,1\}), u_n \in W^{1,1}(\O;\Rb^N)$ satisfy $\chi_n \to \theta$ in $L^1(\O;\{0,1\})$,
$u_n \to v$ in $L^1(\O;\Rb^N)$ and
$$\mathcal F(\theta,v;V) = \lim_{n \to + \infty}F(\chi_n,u_n;V).$$
Theorem~\ref{densitysmooth} ensures the existence of a sequence
$v_n \in W^{1,1}(\O;\Rb^N)$ such that $v_n \to v$ in $L^1(\O;\Rb^N)$,
$v_n = v = u \mbox{ on } \partial V$ and $|Ev_n|(V) \to |Ev|(V)$.
We now apply Proposition~\ref{newslicing} to conclude that there exists a subsequence
$\{v_{n_k}\}$ of $\{v_n\}$ and there exist sequences $w_k \in W^{1,1}(\O;\Rb^N)$,
$\eta_k \in BV(\O;\{0,1\})$ verifying $w_k = v_{n_k} = u$ on $\partial V$,
$\eta_k = \theta = \chi$ on $\partial V$ and
$$\limsup_{k \to +\infty}F(\eta_k,w_k;V) \leq
\liminf_{n \to +\infty}F(\chi_n,u_n;V).$$
Therefore,
$$m_0(\chi,u;V) \leq \limsup_{k \to +\infty}F(\eta_k,w_k;V) \leq
\liminf_{n \to +\infty}F(\chi_n,u_n;V) = \mathcal F(\theta,v;V) \leq
m(\chi,u;V) + \e,$$
so the desired inequality follows by letting $\e \to 0^+.$
\end{proof}
\begin{proposition}
Let $f$ be given by \eqref{density}, where $W_0$ and $W_1$ are continuous functions satisfying \eqref{growth}.
Assume that $f$ is symmetric quasiconvex and that \eqref{finfty} holds.
Given $u \in BD(\Omega)$ and $\chi \in BV(\Omega;\{0,1\})$,
for $\mathcal H^{N-1}$ a.e. $x_0 \in \Omega \cap (J_\chi \cup J_u)$, we have
$$g(x_0,\chi^+(x_0),\chi^-(x_0),u^+(x_0),u^-(x_0),\nu(x_0)) \geq
K(\chi^+(x_0),\chi^-(x_0),u^+(x_0),u^-(x_0),\nu(x_0)),$$
where $\chi^+(x_0) = \chi^-(x_0) = \widetilde\chi(x_0)$ if
$x_0 \in J_u \setminus J_\chi$ and $u^+(x_0) = u^-(x_0) = \widetilde u(x_0)$ if
$x_0 \in J_\chi \setminus J_u$, and $K$ is given by \eqref{K}.
\end{proposition}
\begin{proof}
As before, for simplicity of notation, we write $\nu = \nu(x_0)$.
By Lemma~\ref{BFT314} we have
\begin{align*}
&g(x_0,\chi^+(x_0),\chi^-(x_0),u^+(x_0),u^-(x_0),\nu(x_0)) \\
& \qquad =\limsup_{\e \to 0^+}\frac{1}{\e^{N-1}}
\inf\Big\{F (\theta,v;Q_\nu(x_0,\e)) :
\theta \in BV(\O;\{0,1\}), v \in W^{1,1}(\O;\Rb^N), \\
& \hspace{3,5cm} \theta = \chi_0(\cdot - x_0) \mbox{ on } \partial Q_\nu(x_0,\e),
v = u_0(\cdot - x_0) \mbox{ on } \partial Q_\nu(x_0,\e)\Big\},
\end{align*}
where $\chi_0$ and $u_0$ are given by \eqref{targets} with $\nu = \nu(x_0)$ and $a=\chi^+(x_0)$,
$b=\chi^-(x_0)$, $c=u^+(x_0)$ and $d=u^-(x_0)$, respectively. Thus, for every
$n \in \Nb$, there exist $\theta_{n,\e} \in BV(\O;\{0,1\})$,
$v_{n,\e} \in W^{1,1}(\O;\Rb^N)$ such that $\theta_{n,\e} = \chi_0(\cdot - x_0)$ on $\partial Q_\nu(x_0,\e)$, $v_{n,\e} = u_0(\cdot - x_0)$ on $\partial Q_\nu(x_0,\e)$
and
\begin{align}\label{g}
& g(x_0,\chi^+(x_0),\chi^-(x_0),u^+(x_0),u^-(x_0),\nu(x_0)) + \frac{1}{n} \nonumber\\
&\hspace{2cm}\geq \limsup_{\e \to 0^+}\frac{1}{\e^{N-1}}
\left[\int_{Q_{\nu}(x_0,\e)}f(\theta_{n,\e}(x), \E v_{n,\e}(x)) \, dx + |D\theta_{n,\e}|(Q_{\nu}(x_0,\e))\right] \nonumber\\
& \hspace{2cm}= \limsup_{\varepsilon \rightarrow 0^+}
\left[\int_{Q_{\nu}}\e f(\theta_{n,\e}(x_0+\e y), \E v_{n,\e}(x_0+\e y)) \, dy +
\int_{Q_{\nu}\cap \frac{1}{\e}(J_{\theta_{n,\e}}-x_0)}d\mathcal H^{N-1}(y)\right] \nonumber\\
& \hspace{2cm}= \limsup_{\varepsilon \rightarrow 0^+}
\left[\int_{Q_{\nu}}\e f(\chi_{n,\e}(y), \frac{1}{\e}\E u_{n,\e}(y)) \, dy +
|D\chi_{n,\e}|(Q_{\nu})\right] \nonumber\\
& \hspace{2cm} \geq \liminf_{\varepsilon \rightarrow 0^+}
\left[\int_{Q_{\nu}}f^{\infty}(\chi_{n,\e}(y),\E u_{n,\e}(y)) \, dy +
|D\chi_{n,\e}|(Q_{\nu})\right] \nonumber\\
& \hspace{2cm} + \liminf_{\varepsilon \rightarrow 0^+}
\int_{Q_{\nu}}\left[\e f(\chi_{n,\e}(y), \frac{1}{\e}\E u_{n,\e}(y)) -
f^{\infty}(\chi_{n,\e}(y),\E u_{n,\e}(y))\right] dy,
\end{align}
where $\chi_{n,\e}(y) = \theta_{n,\e}(x_0 + \e y)$ and
$u_{n,\e}(y) = v_{n,\e}(x_0+ \e y)$.
We claim that
\begin{equation}\label{limit0}
\liminf_{\e \to 0^+}
\int_{Q_{\nu}}\left[\e f(\chi_{n,\e}(y), \frac{1}{\e}\E u_{n,\e}(y))- f^{\infty}(\chi_{n,\e}(y),\E u_{n,\e}(y))\right] dy = 0.
\end{equation}
If so, noticing that $(\chi_{n,\e},u_{n,\e}) \in \mathcal{A}(\chi^+(x_0),\chi^-(x_0),u^+(x_0),u^-(x_0),\nu(x_0))$, we have
from \eqref{g}, \eqref{limit0} and the definition of $K(a,b,c,d,\nu)$,
\begin{align*}
&g(x_0,\chi^+(x_0),\chi^-(x_0),u^+(x_0),u^-(x_0),\nu(x_0)) + \frac{1}{n} \\
&\hspace{3cm}
\geq \liminf_{\e \to 0^+}
\left[\int_{Q_{\nu}}f^{\infty}(\chi_{n,\e}(y),\E u_{n,\e}(y)) \, dy +
|D\chi_{n,\e}(Q_{\nu})\right] \\
&\hspace{3cm}\geq K(\chi^+(x_0),\chi^-(x_0),u^+(x_0),u^-(x_0),\nu(x_0)),
\end{align*}
hence the result follows by letting $n \to +\infty$.
It remains to prove \eqref{limit0}. We write
\begin{align*}
& \int_{Q_{\nu}}\e f(\chi_{n,\e}(y), \frac{1}{\e}\E u_{n,\e}(y))- f^{\infty}(\chi_{n,\e}(y),\E u_{n,\e}(y)) \, dy \\
& = \int_{Q_{\nu}\cap \{\frac{1}{\e}|\E u_{n,\e}(y)| \leq L\}}\e f(\chi_{n,\e}(y), \frac{1}{\e}\E u_{n,\e}(y))- f^{\infty}(\chi_{n,\e}(y),\E u_{n,\e}(y)) \, dy \\
& + \int_{Q_{\nu}\cap \{\frac{1}{\e}|\E u_{n,\e}(y)| > L\}}\e f(\chi_{n,\e}(y), \frac{1}{\e}\E u_{n,\e}(y))- f^{\infty}(\chi_{n,\e}(y),\E u_{n,\e}(y)) \, dy
= : I_1 + I_2.
\end{align*}
By the growth hypothesis \eqref{growth} and \eqref{finftygr} we have
\begin{align*}
|I_1| &\leq \int_{Q_{\nu}\cap \{|\E u_{n,\e}(y)| \leq \e L\}}
\e \, C \left(1 + \frac{1}{\e}|\E u_{n,\e}(y)|\right) + C \, |\E u_{n,\e}(y)| \, dy \\
& \leq \int_{Q_{\nu}}\e \, C \, dy = O(\e)
\end{align*}
and, by hypothesis \eqref{finfty} with $t = \frac{1}{\e}$, H\"older's inequality and
\eqref{growth},
\begin{align*}
|I_2| &\leq \int_{Q_{\nu}\cap \{\frac{1}{\e}|\E u_{n,\e}(y)| > L\}}
\left|\e f(\chi_{n,\e}(y), \frac{1}{\e}\E u_{n,\e}(y))-
f^{\infty}(\chi_{n,\e}(y),\E u_{n,\e}(y))\right| \, dy \\
& \leq \int_{Q_{\nu}\cap \{\frac{1}{\e}|\E u_{n,\e}(y)| > L\}}
C \, \e^{\gamma} \, |\E u_{n,\e}(y)|^{1-\gamma} \, dy \\
& \leq C \, \e^{\gamma}
\left( \int_{Q_{\nu}}|\E u_{n,\e}(y)| \, dy\right)^{1-\gamma}\\
& \leq C \, \e^{\gamma}
\left( \int_{Q_{\nu}}\e \,f(\chi_{n,\e}(y),\frac{1}{\e}\E u_{n,\e}(y)) \, dy\right)^{1-\gamma} = O(\e^{\gamma}),
\end{align*}
since the integral in the last expression is uniformly bounded by \eqref{g}. The above estimates yield \eqref{limit0} and complete the proof.
\end{proof}
\begin{proposition}
Let $f$ be given by \eqref{density}, where $W_0$ and $W_1$ are continuous functions satisfying \eqref{growth}, and assume that $f$ is symmetric quasiconvex
and that \eqref{finfty} holds.
Given $u \in BD(\Omega)$ and $\chi \in BV(\Omega;\{0,1\})$,
for $\mathcal H^{N-1}$ a.e. $x_0 \in \Omega \cap (J_\chi \cup J_u)$ we have
$$g(x_0,\chi^+(x_0),\chi^-(x_0),u^+(x_0),u^-(x_0),\nu(x_0)) \leq
K(\chi^+(x_0),\chi^-(x_0),u^+(x_0),u^-(x_0),\nu(x_0)).$$
\end{proposition}
\begin{proof}
Using the sequential characterisation of $K$ given in Proposition \ref{Ktilde}, let
$\chi_n \in BV(Q_\nu;\{0,1\})$, $u_n \in W^{1,1}(Q_\nu;\Rb^N)$ be such that
$\chi_n \to \chi_0$ in $L^1(Q_\nu;\{0,1\})$, $u_n \to u_0$ in $L^1(Q_\nu;\Rb^N)$
and
$$K(\chi^+(x_0),\chi^-(x_0),u^+(x_0),u^-(x_0),\nu(x_0)) =
\lim_{n\rightarrow+\infty}
\left[ \int_{Q_\nu} f^\infty(\chi_n(y),\E u_n(y)) \, dy + |D\chi_n|(Q_\nu)\right],$$
where $\chi_0$, $u_0$ are as in the proof of Proposition~\ref{mg}.
For $x \in Q_\nu(x_0,\e)$, set
$\displaystyle \theta_n(x) := \chi_n\left(\frac{x - x_0}{\e}\right)$ and
$\displaystyle v_n(x) := u_n\left(\frac{x - x_0}{\e}\right)$. Then, changing variables and using the positive homogeneity of $f^\infty(q,\cdot)$, we have
\begin{align}\label{Kg1}
&K(\chi^+(x_0),\chi^-(x_0),u^+(x_0),u^-(x_0),\nu(x_0)) =
\lim_{n\rightarrow+\infty}
\left[ \int_{Q_\nu} f^\infty(\chi_n(y),\E u_n(y)) \, dy + |D\chi_n|(Q_\nu)\right] \nonumber \\
& \hspace{2cm}= \frac{1}{\e^{N-1}}\lim_{n\rightarrow+\infty}
\left[ \int_{Q_\nu(x_0,\e)} f^\infty(\theta_n(x),\E v_n(x)) \, dx + |D\theta_n|(Q_\nu(x_0,\e))\right] \nonumber \\
& \hspace{2cm}\geq \frac{1}{\e^{N-1}}\liminf_{n\rightarrow+\infty}
\left[ \int_{Q_\nu(x_0,\e)} f(\theta_n(x),\E v_n(x)) \, dx + |D\theta_n|(Q_\nu(x_0,\e))\right] \nonumber \\
& \hspace{2cm} + \frac{1}{\e^{N-1}}\liminf_{n\rightarrow+\infty}
\int_{Q_\nu(x_0,\e)} \Big(f^\infty(\theta_n(x),\E v_n(x)) - f(\theta_n(x),\E v_n(x))\Big)\, dx =: I_1 + I_2.
\end{align}
Given that $\chi_n \to \chi_0$ in $L^1(Q_\nu;\{0,1\})$ and $u_n \to u_0$ in $L^1(Q_\nu;\Rb^N)$, it follows that $\theta_n \to \chi_0(\cdot - x_0)$ in $L^1(Q_\nu(x_0,\e);\{0,1\})$ and $v_n \to u_0(\cdot - x_0)$ in $L^1(Q_\nu(x_0,\e);\Rb^N)$. Thus,
\begin{equation}\label{Kg2}
I_1 \geq \frac{1}{\e^{N-1}}
\mathcal F(\chi_0(\cdot - x_0),u_0(\cdot - x_0);Q_\nu(x_0,\e))
\geq \frac{1}{\e^{N-1}}
m(\chi_0(\cdot - x_0),u_0(\cdot - x_0);Q_\nu(x_0,\e)).
\end{equation}
On the other hand, the same calculations that were used to prove \eqref{limit0} by means
of hypothesis \eqref{finfty} allow us to conclude that
\begin{equation}\label{Kg3}
\limsup_{\varepsilon \to 0^+}I_2 = 0.
\end{equation}
Hence, from \eqref{Kg1}, \eqref{Kg2} and \eqref{Kg3} we obtain
$$K(\chi^+(x_0),\chi^-(x_0),u^+(x_0),u^-(x_0),\nu(x_0)) \geq
g(x_0,\chi^+(x_0),\chi^-(x_0),u^+(x_0),u^-(x_0),\nu(x_0)).$$
\end{proof}
\medskip
\noindent\textbf{Acknowledgements}.
The research of ACB was partially supported by National Funding from FCT -
Funda\c c\~ao para a Ci\^encia e a Tecnologia through project UIDB/04561/2020.
JM acknowledges support from CAMGSD and from the Fundação para a Ciência e a Tecnologia through the grant UID/MAT/04459/2020.
EZ is a member of INdAM GNAMPA, whose support is gratefully acknowledged, also through the GNAMPA project 2020, coordinated by Prof. Marco Bonacini. JM and EZ thank the Isaac
Newton Institute for Mathematical Sciences (Cambridge) for its support and hospitality during the programme DNM (Design of New Materials), where this research project started,
supported by EPSRC grant no EP/R014604/1.
\color{black}
|
\section{Introduction}
\label{sec:intro}
Very long baseline interferometry (VLBI) provides the highest direct angular resolution in astronomy, recently culminating in the first image of a black hole \citep{PaperI}. Images of black holes are expected to have a rich interferometric response on long baselines, including small-scale power from both their irregular accretion flows \citep[e.g.,][]{Roelofs_2017,Medeiros_2018,Gelles_2021} and from their distinctive gravitationally lensed features such as the ``photon ring'' \citep[e.g.,][]{Johnson_Ring,Gralla_2020}. A crucial target for VLBI studies is the Milky Way's nuclear black hole, Sagittarius~A* (Sgr~A$^\ast$\xspace), which subtends the largest angular size of any known black hole.
The intrinsic and scattered structure of Sgr~A$^\ast$\xspace\ have been intensively studied using VLBI at millimeter and centimeter wavelengths \citep[e.g.,][]{Jauncey_1989,Alberdi_1993,Lo_1998,Krichbaum_1998,Bower_2004,Shen_2005,Doeleman_2008,Lu_2011,Johnson_2018,Issaoun_2019}. A major challenge for these studies is the unusually strong interstellar scattering of Sgr~A$^\ast$\xspace.
\citet{Kellermann_1977} reported early VLBI observations of Sgr~A$^\ast$\xspace\ at $\lambda = 3.8\,{\rm cm}$, which indicated that 25\% of the emission came from a compact component, $\mathrel{\raise.3ex\hbox{$<$\kern-.75em\lower1ex\hbox{$\sim$}}} 1\,{\rm mas}$ in size.
Such a bright, compact component would be detectable using ground-space VLBI with the orbiting telescope Spektr\nobreakdash-R of the \emph{RadioAstron}\xspace program \citep{2013ARep...57..153K}. More recent measurements have not found evidence for this component at centimeter wavelengths \citep[e.g.,][]{Lo_1998,Bower_2004,Lu_2011,Bower_2015,Johnson_2018}. However, \citet{Gwinn_2014} discovered compact image substructure in Sgr~A$^\ast$\xspace\ at $\lambda = 1.3\,{\rm cm}$, with approximately $1\%$ of the total flux density measured on baselines up to $3000\,{\rm km}$ (${\approx}\, 250 \times 10^6 \lambda$, or $0.8\,{\rm mas}$ resolution). This signal is consistent with expected ``refractive substructure'' produced by scattering in the ionized interstellar medium \citep{NG_1989,GN_1989,Johnson_Gwinn_2015}.
Refractive substructure was also detected in AGN using space-VLBI \citep{2016ApJ...820L..10J,2018MNRAS.474.3523P}.
In this Letter, we examine the interferometric properties of Sgr~A$^\ast$\xspace\ on the much longer baselines, using ground-space VLBI provided by the orbiting telescope Spektr\nobreakdash-R. Launched in 2011, Spektr\nobreakdash-R was located on a highly elliptic orbit, with perigee of $350{,}000\,{\rm km}$ and an orbital period of approximately 9~days. It successfully observed using receivers at four wavelengths, extending from 1.3\,cm to 92\,cm \citep[e.g.,][]{2017MNRAS.465..978P,Kovalev_2020}. Our observations of Sgr~A$^\ast$\xspace\ include projected baselines 4 times the Earth's diameter, providing the sharpest view of this source at centimeter wavelengths and the first using space-VLBI.
\floattable
\begin{deluxetable*}{lccccc}[ht!]
\tablecaption{Observing setup and array for 1.35\,cm observations of Sgr~A$^\ast$\xspace.}
\tablehead{
\colhead{Telescope} & \colhead{Telescope} & \colhead{Diameter} & \colhead{Bitrate} & \colhead{IFs per} & \colhead{Observing} \\[-6pt]
\colhead{name} & \colhead{abbreviation} & \colhead{[m]} & \colhead{[Mbps]} & \colhead{polarization [MHz]} & \colhead{time [UT]}
}
\startdata
Spektr\nobreakdash-R & RA & 10 & 128 & 2$\times$16 & 13.09.2015 01:30 -- 04:30\\
VLBA St.~Croix & SC & 25 & 2048 & 4$\times$64 & 12.09.2015 23:00 -- 13.09.2015 04:30 \\
VLBA Hancock & HN & 25 & 2048 & 4$\times$64 & 12.09.2015 23:00 -- 13.09.2015 04:30 \\
VLBA North Liberty & NL & 25 & 2048 & 4$\times$64 & 12.09.2015 23:00 -- 13.09.2015 04:30 \\
VLBA Fort Davis & FD & 25 & 2048 & 4$\times$64 & 12.09.2015 23:00 -- 13.09.2015 04:30 \\
VLBA Los Alamos & LA & 25 & 2048 & 4$\times$64 & 12.09.2015 23:00 -- 13.09.2015 04:30 \\
VLBA Pie Town & PT & 25 & 2048 & 4$\times$64 & 12.09.2015 23:00 -- 13.09.2015 04:30 \\
VLBA Kitt Peak & KP & 25 & 2048 & 4$\times$64 & 12.09.2015 23:00 -- 13.09.2015 04:30 \\
VLBA Owens Valley & OV & 25 & 2048 & 4$\times$64 & 12.09.2015 23:00 -- 13.09.2015 04:30 \\
VLBA Brewster & BR & 25 & 2048 & 4$\times$64 & 12.09.2015 23:00 -- 13.09.2015 04:30 \\
VLBA Mauna Kea & MK & 25 & 2048 & 4$\times$64 & 12.09.2015 23:00 -- 13.09.2015 04:30 \\
VLA & VLA & 27$\times$25 & 2048 & 4$\times$64 & 12.09.2015 23:00 -- 13.09.2015 04:30 \\
GBT & GBT & 100 & 2048 & 4$\times$64 & 12.09.2015 23:00 -- 13.09.2015 03:00 \\
LBA ATCA & ATCA & 6$\times$22 & 1024 & 2$\times$64 & 13.09.2015 02:00 -- 12:00 \\
LBA Mopra & MP & 22 & 1024 & 2$\times$64 & 13.09.2015 02:00 -- 12:00 \\
LBA Ceduna & CD & 30 & 512 & 1$\times$64 & 13.09.2015 02:00 -- 12:00 \\
LBA Hobart & HO & 26 & 512 & 1$\times$64 & 13.09.2015 02:00 -- 12:00 \\
KVN Yonsei & KY & 21 & 1024 & 2$\times$64 & 13.09.2015 07:00 -- 12:00 \\
KVN Ulsan & KU & 21 & 1024 & 2$\times$64 & 13.09.2015 07:00 -- 12:00 \\
KVN Tamna & KT & 21 & 1024 & 2$\times$64 & 13.09.2015 07:00 -- 12:00 \\
Tian Mai & T6 & 65 & 512 & 4$\times$32 & 13.09.2015 07:00 -- 12:00 \\
\enddata
\label{tab:array}
\tablecomments{The VLA and ATCA are linked radio interferometers of 27 and 6 antennas, respectively, which were each phased up and coherently summed for these observations.
All telescopes observed dual circular polarizations except Tian Mai, which observed only left-hand circular polarization (LCP).
All ground telescopes used 2-bit sampling; Spektr\nobreakdash-R used 1-bit sampling. Note that the recorded bandwidth and number of intermediate frequency (IF) bands varied significantly across the array, necessitating a multi-stage correlation and analysis procedure.
}
\end{deluxetable*}
\begin{figure*}[ht]
\begin{center}
\includegraphics[width=0.49\textwidth]{u-v_Coverage_1.pdf}
\includegraphics[width=0.49\textwidth]{u-v_Coverage_2.pdf}
\end{center}
\caption{Sgr~A$^\ast$\xspace\ baseline coverage of \emph{RadioAstron}\xspace 1.3~cm observations. The left panel shows the full range of baseline coverage, including only the three most sensitive baselines to Spektr\nobreakdash-R; the right panel shows a zoomed in region with only the ground baselines. Single-baseline detections/non-detections using PIMA are colored green/black.
We do not detect Sgr~A$^\ast$\xspace\ on any baselines to Spektr\nobreakdash-R. Detections on ground baselines all lie within the region where the scattering kernel is at least $1\%$, denoted by the yellow ellipse.}
\label{fig:uv_coverage}
\end{figure*}
\section{Observations, Correlation, and Processing}
\label{sec:observations}
We observed Sgr~A$^\ast$\xspace\ on September 13, 2015 using Spektr\nobreakdash-R together with a global ground network of 20 telescopes (observing codes RAGS01B and BK193A). To maximize the likelihood of detections on Sgr~A$^\ast$\xspace, which has angular broadening from scattering ${\propto}\,\lambda^2$, our observations used the highest Spektr\nobreakdash-R observing band (a central frequency of 22.2\,GHz). In addition, we selected an observing interval where the Earth-space baselines were predominantly North-South, which is the orientation in which the scattering is both weaker and more poorly constrained.
\autoref{tab:array} gives details of the array and observing setup.
The time interval of Spektr\nobreakdash-R observations was determined mainly by its cooling constrains. The Green Bank earth station \citep{2014SPIE.9145E..0BF} was used to collect science data from Spektr\nobreakdash-R.
Correlation was performed using a special version of the software correlator {\tt DiFX} \citep{deller_difx_2011, bruni_difx_2016} that is capable of dealing with Spektr\nobreakdash-R data.
Due to scheduling limitations, there was no common scan for all stations on the strong fringe calibrator 1730$-$130, however, separate scans for the US and LBA-KVN-T6 parts of the array produced fringes for all ground telescopes.
Because of the long slew time, Spektr\nobreakdash-R did not observe the calibrator. This decision was supported by an extensive experience with RadioAstron operations; onboard telemetry provided comprehensive information on the technical status of the space telescope.
Due to different recorded bandwidth at different antennas, the correlation was performed in three separate passes to recover as much information as possible:
(i) native 64~MHz channels for all ground stations excluding T6;
(ii) 32~MHz channels for all ground stations;
(iii) $2{\times}16$~MHz channels for baselines between Spektr\nobreakdash-R and GBT, ATCA, and VLA.
Using the VLBI software processing package \texttt{PIMA}\footnote{\url{http://astrogeo.org/pima/}}, we performed baseline-based fringe fitting, which included determination of the phase acceleration term to statistically separate observations with significant interferometric signal and noise. To do this, we applied the approach described in \cite{Petrov_2011} and \cite{Kovalev_2020} to estimate the probability of false detection (PFD) for each interferometric fringe obtained. Fringes with $\text{PFD} > 10^{-4}$ were classified as non-detections; for each non-detection, we computed the associated visibility amplitude upper limit.
The ground array dataset with 64-MHz channels (IFs) was also processed in \texttt{AIPS}\footnote{\url{http://www.aips.nrao.edu}} with global fringe fitting to determine single-IF antenna-based solutions. The amplitude was calibrated using \textit{a~priori} gain and system temperature information provided by the observing facilities.
We were able to perform and use the global antenna-based fringe fitting with multi-IF solutions for the US portion of the array due to availability and successful application of instrumental phase-calibration data. With 4 IFs, this procedure improved the SNR of our solutions by up to a factor of two.
For our subsequent analysis and discussion, we use results from both these software packages, relying on PFDs and upper limits from \texttt{PIMA} for baselines without detected fringes, and using the highest possible signal-to-noise ratio of ground antenna-based fringe fitting solutions in \texttt{AIPS}.
\autoref{fig:uv_coverage} shows the baseline coverage of our observations of Sgr~A$^\ast$\xspace. This figure also indicates which baselines gave statistically significant interferometric single-baseline fringes using PIMA. We did not detect fringes on any baselines to Spektr\nobreakdash-R.
\section{Analysis and Results}
\label{sec:Results}
Analysis of our observation was complicated by usually short coherence times and poor phasing of the VLA. To ensure accurate results that recover as much information as possible, we performed our analysis in several stages, using strong detections on short baselines to recover weak detections on long baselines. In all cases, we compare our measurements with predictions from a physically motivated model for the scattering of Sgr~A$^\ast$\xspace that was developed by \citet{Psaltis_2018}. We use source and scattering parameters from \citet{Johnson_2018} (hereafter, \citetalias{Johnson_2018}), who estimated them by analyzing archival VLBI measurements of Sgr~A$^\ast$\xspace\ extending from 1.3\,mm to 29\,cm wavelength. The predictions of this model have been successfully tested using VLBI including ALMA at 3.5\,mm \citep{Issaoun_2019,Issaoun_2021} and using monitoring observations with the East Asian VLBI Network at 1.35 and 0.7\,cm \citep{Cho_2021}. Specifically, the \citetalias{Johnson_2018} model has an intrinsic source that is a circular Gaussian with a wavelength-dependent full width at half maximum (FWHM) size $(0.4\,{\rm mas}) \times \lambda_{\rm cm}$, and a scattering screen located 2.7\,kpc from the Earth and 5.4\,kpc from the source (with a corresponding magnification $M \approx 0.53$). The scattering screen is stochastic, with a shallow scattering power-law index of $\alpha = 1.38$,\footnote{The power-law index $\alpha$ corresponds to that of the phase structure function in the inertial range. The 2D power spectrum of phase fluctuations has an index $\beta = \alpha+2$; the angular broadening scales as $\lambda^{1+2/\alpha}$ at short wavelengths and as $\lambda^2$ at long wavelengths. For a review of interstellar scattering, see \citet{Rickett_1990} or \citet{Narayan_1992}.} an inner scale of $r_{\rm in} = 800\,{\rm km}$, and anisotropic angular broadening at a position angle $\phi_{\rm PA} = 81.9^\circ$ with FWHMs along the principal axes $\theta_{\rm maj} = (1.380\,{\rm mas}) \times \lambda_{\rm cm}^2$ and $\theta_{\rm min} = (0.703\,{\rm mas}) \times \lambda_{\rm cm}^2$ (as $\lambda \rightarrow \infty$).
On baselines that do not significantly resolve Sgr~A$^\ast$\xspace, the visibility function is expected to be well approximated by that of the ensemble-averaged image -- the product of the visibility function of the intrinsic (unscattered) source and the scattering kernel \citep[e.g.,][]{Coles_1987,GN_1989}. For baselines with $b \mathrel{\raise.3ex\hbox{$<$\kern-.75em\lower1ex\hbox{$\sim$}}} (1+M) r_{\rm in} \approx 1200\,{\rm km}$, the scattering kernel is well-approximated by an elliptical Gaussian \citep{GN_1989,Psaltis_2018}; for baselines that only partially resolve the intrinsic source, the source visibility function is also well-approximated by an elliptical Gaussian. Thus, we first fit our short-baseline data using an elliptical Gaussian model, described in \autoref{sec::Gaussian_fits}.
Next, we used the Gaussian fit on short baselines to derived amplitude and phase self-calibration solutions that were applied to all baselines. After discarding data that had no self-calibration solution, we coherently averaged the data in two-minute intervals and across all four IFs to maximize the available sensitivity. We analyzed the averaged data on long ground baselines, which were then sufficiently sensitive to detect a $\mathrel{\raise.3ex\hbox{$<$\kern-.75em\lower1ex\hbox{$\sim$}}} 1\%$ signal, described in \autoref{sec:long_baseline_analysis}.
Finally, we assessed whether the long-baseline signal was consistent with the expected refractive noise using the \citetalias{Johnson_2018} model. Because refractive substructure is partially quenched by an extended source, this provides a test of both the intrinsic and scattering model parameters. However, the comparison is statistical in nature because the signal from refractive noise is stochastic. Our comparisons are described in \autoref{sec:consistency_check}.
\subsection{Short-Baseline Analysis: Gaussian Fits}
\label{sec::Gaussian_fits}
We first fit a Gaussian model to the short-baseline data. To avoid biasing this fit with non-Gaussian contributions from refractive-noise dominated measurements, we fit only the data on baselines with an expected ensemble-average flux density of at least $2\%$ of the total flux density in the \citetalias{Johnson_2018} model. The longest baseline meeting this criterion had $|\mathbf{u}| = 97.4$\,M$\lambda$ (BR-VLA). We began by coherently averaging the data in frequency (across each IF) and in time (in 2-second intervals). We then applied amplitude and phase self-calibration solutions before coherently averaging the data further, in 2-minute intervals. At this stage, the primary purpose of self calibration is to avoid phase decoherence in the coherent averaging.
Next, we fit an elliptical Gaussian model to the averaged visibilities. We fit the model using diagonalized log closure amplitudes \citep{Blackburn_2020}, with flat model priors on the major axis, minor axis, and position angle. To avoid non-Gaussian errors from low signal-to-noise ratio (S/N) measurements, we only included closure amplitudes with ${\rm S/N} > 5$. Because of the wavelength-dependent image size, we fit data from each IF independently. To perform the fit, we utilized {\tt eht-imaging} \citep{Chael_16} together with the nested sampling package {\tt dynesty} \citep{Speagle_2020}.
\autoref{tab:Gaussian_Fits} shows the results of this Gaussian model fitting. Our data tightly constrain all the model parameters, particularly the major axis and position angle. Note that the fitted uncertainties on these parameters only give the uncertainties from thermal noise; refractive scattering causes an additional ${\sim}1\%$ variation in the major and minor axis sizes, which is comparable to the error budget from thermal noise for the better-constrained major axis. All our fitted values are consistent with the predictions from the \citetalias{Johnson_2018} scattering model.
{
\begin{deluxetable}{l|ccc}[t]
\tablewidth{\textwidth}
\tablecaption{Summary of Gaussian fits to Sgr~A$^\ast$\xspace.}
\tablehead{
\colhead{$\lambda$ [cm]} & \colhead{$\theta_{\rm maj}$ [$\mu{\rm as}$]} & \colhead{$\theta_{\rm min}$ [$\mu{\rm as}$]} & \colhead{P.A. [deg]}
}
\startdata
1.352 & $2587 \pm 18$ & $1210 \pm 136$ & $81.8 \pm 0.9$ \\
& $2580$ & $1394$ & $81.9$\\
\hline
1.348 & $2572 \pm 17$ & $1313 \pm 115$ & $82.2 \pm 0.8$ \\
& $2566$ & $1387$ & $81.9$\\
\hline
1.344 & $2571 \pm 17$ & $1356 \pm 112$ & $82.2 \pm 0.9$ \\
& $2552$ & $1381$ & $81.9$\\
\hline
1.341 & $2543 \pm 17$ & $1302 \pm 116$ & $82.5 \pm 0.9$ \\
& $2538$ & $1374$ & $81.9$
\enddata
\label{tab:Gaussian_Fits}
\tablecomments{Fitted elliptical Gaussian parameters for Sgr~A$^\ast$\xspace\ for each sub-band (for details, see \autoref{sec::Gaussian_fits}). Below each fitted parameter is its corresponding expected value using the \citetalias{Johnson_2018} source and scattering parameters.}
\end{deluxetable}
}
\begin{figure*}[t]
\begin{center}
\includegraphics[width=\textwidth]{RadioAstron_Summary_Combine_IF0-crop.pdf}
\end{center}
\caption{Top: Visibility amplitude as a function of baseline length for Sgr~A$^\ast$\xspace. Green points show strong detections used to derive the Gaussian self-calibration solution. Gray points show the long-baseline visibilities after coherently averaging using the short-baseline self-calibration; for clarity, only visibilities with thermal noise less than $10\,{\rm mJy}$ and $|\mathbf{u}| > 80 \times 10^6$ are shown. Yellow dashed lines shown the expected major (lower) and minor (upper) axes of the ensemble-average image; orange curves show the expected root-mean-square (renormalized) refractive noise along the major (upper) and minor (lower) axes. Colored triangles show the PIMA fringe amplitudes for representative long baselines, including the three most sensitive ground-space baselines; the corresponding thermal noise for each indicated with a cross. None of these fringe amplitudes is a statistically significant detection. Bottom: Zoom-in on ground baselines. Colored points show a subset of the most sensitive long ground baselines; the incoherent average on each baseline is also shown and indicated with a diamond. Uncertainties shown are $\pm1\sigma$. Both NL-VLA and BR-VLA have significant flux expected in the ensemble-average image, but the remaining long baselines are expected to be dominated by refractive noise.}
\label{fig:radplot_combine}
\end{figure*}
\subsection{Long-Baseline Analysis: Refractive Noise}
\label{sec:long_baseline_analysis}
After performing the Gaussian fits on short baselines, we derived amplitude and phase self-calibration solutions and applied them to all baselines. For the amplitude self-calibration, we used a total flux density of $1\,{\rm Jy}$, which is comparable to the average of historical values at 1.3\,cm \citep[e.g.,][]{Bower_Spectrum}. This choice of normalization does not affect our remaining analysis, which only relies on the fractional visibility amplitudes relative to the total flux density. In addition, we averaged the data across all IFs, to maximize our sensitivity.
\autoref{fig:radplot_combine} shows the resulting visibility measurements, including upper limits on baselines to Spektr\nobreakdash-R (computed by {\tt PIMA}) and highlighting the amplitudes for our most sensitive ground baselines. These baselines are expected to have signals that are dominated by refractive noise; they have $S/N$ up to 8.4 after the final incoherent averaging in time, indicating a reliable detection of image substructure.
For each figure, we show the expected envelope of the ensemble-average image and the predicted ``renormalized refractive noise'' $\hat{\sigma}_{\rm ref}(\mathbf{u}) = \left\langle \left|\Delta \hat{V}(\mathbf{u}) \right|^2 \right \rangle^{1/2}$, where $\hat{V}$ is the complex visibility function of the source after centering the image and normalizing the total flux density \citepalias[for details, see Appendix~A of][]{Johnson_2018}.
\subsection{Consistency with Expected Properties}
\label{sec:consistency_check}
We now examine whether the long-baselines visibilities are consistent with the expected level of refractive noise for the \citetalias{Johnson_2018} model. For baselines that resolve the ensemble-average image, refractive noise is well approximated as a circular complex Gaussian random variable \citep{Johnson_Narayan_2016}. The visibility amplitude is then drawn from a Rayleigh distribution. The middle 95\% of samples drawn from a Rayleigh distribution fall between $0.18$ and $2.2$ times the mean, $\left \langle \left|\Delta \hat{V}(\mathbf{u}) \right| \right \rangle = \frac{\sqrt{\pi}}{2} \hat{\sigma}_{\rm ref}(\mathbf{u}) \approx 0.89 \hat{\sigma}_{\rm ref}(\mathbf{u})$. All the long-baseline measurements in \autoref{fig:radplot_combine} are, thus, consistent with the expected level of refractive noise.
We can make more precise comparisons by averaging measurements from different baselines, which sample partially independent elements of the refractive noise. As a simple comparison, we took an incoherent average of all (debiased) visibility amplitudes on baselines longer than $150\times 10^6$ and with thermal noise less than $10\,{\rm mJy}$. This procedure gave an average amplitude of $5.7\,{\rm mJy}$. When then generated a set of 1000 scattered images using the {\tt stochastic-optics} module of {\tt eht-imaging} \citep{Johnson_2016}; we created synthetic observations for each, with baseline coverage and thermal noise matching our observations. For each synthetic observation, we computed the incoherent average on long baselines exactly as was done for the real observation. This procedure gave a median amplitude of $5.4\,{\rm mJy}$, with the inner 95\% of samples falling between $4.3$ and $8.0\,{\rm mJy}$. Hence, the average of our long-baseline measurements is consistent with the predicted range of refractive noise.
Our non-detections to Spektr\nobreakdash-R are likewise consistent with theoretical expectations for a power-law theory of the scattering and substructure (see \autoref{fig:radplot_combine}). We do not find evidence for any long-baseline visibilities above the expected refractive noise floor, $\sigma_{\rm ref} \sim 1\,{\rm mJy} \times \left( |\mathbf{u}|/{10^9} \right)^{-0.69}$.
In particular, this refractive noise is ${\sim}2$ orders of magnitude below the most stringent upper limits of ${\sim}150\,{\rm mJy}$ on baselines to Spektr\nobreakdash-R.
An unexpected feature of our long ground baseline measurements is that they are quite close in amplitude to those of \citet{Gwinn_2014} \citepalias[see Figure~6 of][]{Johnson_2018}, which were taken on March 7, 2014. Namely, both observations measure visibility amplitudes of ${\sim}4\,{\rm mJy}$ on baselines from 150-200\,M$\lambda$, with a rise to ${\sim} 8\,{\rm mJy}$ on the ${\sim}250\,{\rm M}\lambda$ baselines GBT-BR and GBT-OV. These observations are separated by 554~days, while the expected correlation timescale of the refractive noise is a few months (for a characteristic scattering velocity of $50\,{\rm km/s}$). Thus, we expect that the similarity is simply a chance alignment; if future measurements show that this signal is persistent, it would either indicate a much slower scattering velocity or that the measurements are sensitive to the ensemble-average properties of the scattering rather than of refractive substructure.
\section{Summary}
\label{sec:Summary}
Space-VLBI is an exciting frontier for black hole astrophysics, with the potential to resolve the gravitationally lensed ``photon rings'' of nearby supermassive black holes \citep{Johnson_Ring}, to measure the masses of \emph{thousands} of supermassive black holes via their ``shadow'' diameters \citep{Pesce_2021}, and to track the orbits of many supermassive black hole binaries \citep{DOrazio_2018}.
Proposed mission concepts to observe Sgr~A$^\ast$\xspace\ have focused on exploring configurations that allow for rapid baseline sampling to reconstruct movies \citep[e.g.,][]{Palumbo_2019,Johnson_Astro2020} as well as enabling observations at submillimeter wavelengths for which interstellar scattering is sharply reduced \citep[e.g.,][]{Roelofs_2019,Kudriashov_2021}.
Space VLBI at longer wavelengths provides crucial input for these designs and gives firm system requirements. In this Letter, we have presented the first space-VLBI of Sgr~A$^\ast$\xspace, observing at $\lambda = 1.35\,{\rm cm}$ using Spektr\nobreakdash-R and 20 ground antennas. Our short ground baselines are well fit by an elliptical Gaussian image, with parameters matching those of historical measurements. Our long ground baselines confirm the presence of persistent small-scale structure in the scattered image of Sgr~A$^\ast$\xspace, originally discovered by \citet{Gwinn_2014}, at a level that is also consistent with predictions for refractive interstellar scattering. We do not detect interferometric fringes on any baselines to Spektr\nobreakdash-R.
Our work highlights the severe challenges for observing Sgr~A$^\ast$\xspace\ with space-VLBI. Because of diffractive interstellar scattering, improving the angular resolution achievable from the ground will likely require observations with $\lambda \mathrel{\raise.3ex\hbox{$<$\kern-.75em\lower1ex\hbox{$\sim$}}} 2\,{\rm mm}$, while detecting features such as the lensed photon ring will require $\lambda \mathrel{\raise.3ex\hbox{$<$\kern-.75em\lower1ex\hbox{$\sim$}}} 0.8\,{\rm mm}$.
\begin{acknowledgments}
We thank Eduardo Ros for useful comments on the manuscript, the anonymous ApJL referee for comments that clarified the manuscript, Amy Mioduszewski for consultations regarding VLA observations in the VLBI mode, and Rebecca Azulay for correlator support at MPIfR. The \emph{RadioAstron}\xspace project was led by the Astro Space Center of the Lebedev Physical Institute of the Russian Academy of Sciences and the Lavochkin Scientific and Production Association under a contract with the State Space Corporation ROSCOSMOS, in collaboration with partner organizations in Russia and other countries.
The National Radio Astronomy Observatory and the Green Bank Observatory
are facilities of the National Science Foundation operated under cooperative agreement by Associated Universities, Inc.
The Australian Long Baseline Array is part of the Australia Telescope National Facility which is funded by the Australian Government for operation as a National Facility managed by CSIRO.
This research is based on observations correlated at the Bonn Correlator, jointly operated by the Max Planck Institute for Radio Astronomy (MPIfR), and the Federal Agency for Cartography and Geodesy (BKG). MDJ thanks the National Science Foundation (AST-1716536, AST-1935980) and the Gordon and Betty Moore Foundation (GBMF-5278) for financial support of this work. YYK is supported in the framework of the State project ``Science'' by the Ministry of Science and Higher Education of the Russian Federation under the contract 075-15-2020-778.
\end{acknowledgments}
\facilities{RadioAstron Space Radio Telescope (Spektr\nobreakdash-R), VLBA, GBT, VLA, LBA, KVN}
\software{
AIPS \citep{Greisen_2003},
eht-imaging \citep{Chael_2016},
PIMA \citep{Petrov_2011}
}
|
\section{Introduction}
Traffic-related injuries and fatalities are major health risks in the United States. Mobile phone use while driving (MPUWD) quadruples the risk for a motor vehicle crash \cite{Redelmeier1997-fj}. In 2019 alone, 3,142 people died in crashes with distracted drivers \cite{nhtsa-dd}. Drivers using a smartphone have reduced ability to detect and respond to roadway stimuli and events \cite{Caird2008-tu, Fitch2013-vh, Handel2014-ku, Klauer2014-tf, Strayer2013-tp}. Teens in general have a stronger tendency to use their mobile phones \cite{Walsh2008-rd}, and studies have shown that nearly half of adolescent drivers engage in MPUWD \cite{Curry2011-cf, Neyens2008-sn}. Legal bans on phone use while driving have shown only modest results, likely due to the difficulties of enforcement \cite{Delgado2016-sr}. Modern mobile phones have built-in functionality to detect when a phone is in a moving vehicle; however, those libraries are unable to differentiate between drivers and passengers, and commercially available MPUWD software typically includes a simple option for users to manually disable it during drives where they are a passenger.
\paragraph{Related work.} To overcome this limitation on current phones researchers have utilized vehicle additions or modifications to differentiate between drivers and passengers. Previous efforts to determine the seating location of a phone user have relied on either Bluetooth signals \cite{Yuan2020-np, Yang2011-sh}, phone sensor data \cite{Park2018-py}, or car sensor data \cite{Kar2017-ew}. Many of these methods rely on either installing Bluetooth beacons, having the user's phone paired to the car's Bluetooth radio, or access the car's sensor bus, and the accuracy has not exceeded 93\%. Another method is to use the phone's camera to determine the user's location using computer vision \cite{Khurana2020-qc}. This work was able to reach 94\% accuracy, but in a a limited dataset. Camera-based methods require no hardware installation or Bluetooth pairing by using only the user's phone camera to determine the user location within the car. We are expanding on this work by demonstrating the feasibility of a computer vision method in a larger, more varied, dataset.
\section{Methodology}
\paragraph{Dataset.} The seat detection dataset was collected by at a number of used car dealerships. A custom mobile phone app was written to assist users in taking a series of pictures from each seat of the vehicle with prompts for phone orientation. Multiple images were acquired from each seat using slightly different orientations. Images were acquired from driver seats as well as front and rear passenger seats. The make, model, and year of each car was recorded along with the time of day. The image acquisition times varied between full sun, overcast, dawn, dusk and night. A total of 12,042 images were acquired from 120 cars and trucks. There were 3,721 images from the driver seat and 8,321 from the passenger seats. The model years ranged from 1996 - 2020. The dataset was randomly split into training, development, and testing sets using the car as the split key, so the same car did not appear in multiple sets.
\paragraph{CNN architecture.} The neural network was implemented using Keras and Tensorflow. The architecture was MobileNet-V2 \cite{Sandler2018-ix} feature extractor provided with Keras pretrained using the ImageNet dataset \cite{Deng2009-ov}. A custom head was attached to the MobileNet feature extractor that consisted of: 1) a $1 \times 1$ convolution followed by batch normalization \cite{Ioffe2015-il}, a ReLU activation, and dropout, 2) a $7 \times 7$ convolution followed by batch normalization, a ReLU activation, and dropout, and 3) a global max-pooling and a final fully connected layer.
\paragraph{Training procedure.} The training set consisted of 9,127 images, and the development set consisted of 1,229 images. The network was trained using a binary cross-entropy loss function with a stochastic gradient descent optimizer. The optimizer had a learning rate of 1E-4 and momentum of 0.9. The model was trained until the accuracy on the development set decreased for 8 consecutive epochs, then the weights were reset to those of the epoch with maximum accuracy on the development set. The images were converted to gray-scale then rescaled and randomly cropped down to $224 \times 224$ pixels. Image augmentation through rotations was attempted, but it reduced performance. We hypothesize that this was because subtle perspective rotation can be a differentiator between the driver and front passenger seats.
\begin{figure}[h]
\centering
\begin{subfigure}{0.33\textwidth}
\centering
\includegraphics[width=.8\linewidth]{correct-passenger-fit.jpg}
\caption{Truth: Passenger \newline Prediction: {\color{green}Passenger}}
\label{fig:figure1a}
\end{subfigure}
%
\begin{subfigure}{0.33\textwidth}
\centering
\includegraphics[width=.8\linewidth]{correct-passenger-pilot.jpg}
\caption{Truth: Passenger \newline Prediction: {\color{green}Passenger}}
\label{fig:figure1b}
\end{subfigure}
%
\begin{subfigure}{0.33\textwidth}
\centering
\includegraphics[width=.8\linewidth]{correct-driver-pilot.jpg}
\caption{Truth: Driver \newline Prediction: {\color{green}Driver}}
\label{fig:figure1c}
\end{subfigure}
\begin{subfigure}{0.33\textwidth}
\centering
\includegraphics[width=.8\linewidth]{incorrect-driver-outback.jpg}
\caption{Truth: Passenger \newline Prediction: {\color{red}Driver}}
\label{fig:figure1d}
\end{subfigure}
%
\begin{subfigure}{0.33\textwidth}
\centering
\includegraphics[width=.8\linewidth]{incorrect-driver-pilot.jpg}
\caption{Truth: Passenger \newline Prediction: {\color{red}Driver}}
\label{fig:figure1e}
\end{subfigure}
%
\begin{subfigure}{0.33\textwidth}
\centering
\includegraphics[width=.8\linewidth]{incorrect-passenger-rav4.jpg}
\caption{Truth: Driver \newline Prediction: {\color{red}Passenger}}
\label{fig:figure1f}
\end{subfigure}
\caption{Example images showing both correct and incorrect predictions in a variety of lighting conditions and image qualities.}
\label{figure1}
\end{figure}
\section{Results}
The final test set consisted of 1,686 images from a variety of cars, trucks, and lighting conditions. The network was able correctly identify a driver vs. a passenger with an accuracy of 94.9\% when using a 0.5 decision threshold. Examples of correct and incorrect predictions are shown in \textbf{Figure 1}.
\section{Discussion}
Our results confirm and expand the results of Khurana \textit{et al.} \cite{Khurana2020-qc} and demonstrate the feasibility of using a phone's camera to determine the user's location within a vehicle. Compared to the method in Khurana \textit{et al.}, our method uses only the rear camera, which is typically much better in low light conditions. Our method also uses a purely neural network model compared to a geometric model, so it should generalize better to different vehicles and modes of transportation. This seat detection model could be provided as a library for application developers to use to enable or disable features or to lock the phone completely.
\clearpage
\bibliographystyle{IEEEtran}
|
\section{Introduction}
Tensor categories have come to play an important role in noncommutative analysis, arising as categories of bimodules of $\rm C^*$ and von Neumann algebras and as representation categories of compact quantum groups.
In subfactor theory, the standard invariant of a finite index $\rm II_1$ subfactor \cite{MR1334479,math.QA/9909027} can be described by a unitary tensor category (a.k.a.~a semisimple rigid $\rm C^*$-tensor category), together with a chosen unitary Frobenius algebra object internal to this category \cite{MR1966524,MR1027496,MR2909758,MR3948170}.
In operator algebraic approaches to quantum field theories (AQFT) and topologically ordered spin systems, braided tensor categories arise in the DHR theory of superselection sectors of nets of von Neumann algebras \cite{MR0297259, MR1405610, MR1838752, MR3426207, arXiv:2106.15741}.
Here, the presence of a braiding yields an incredibly rich structure theory which does not have an obvious analog in the purely `noncommutative' world of ordinary subfactors.
There are, however, less widely recognized instances of braided tensor categories arising in the theory of $\rm{II}_{1}$ factors.
In \cite{MR377534}, Connes introduced an invariant $\chi(M)$ of a $\rm{II}_{1}$ factor $M$, which is the abelian subgroup of $\operatorname{Out}(M)$ consisting of the image of approximately inner and centrally trivial automorphisms.
In \cite{MR585235}, V.~Jones defines a quadratic form $\kappa$ on the group $\chi(M)$.
Eilenberg and MacLane have shown that an abelian group together with a quadratic form defines (uniquely up to braided equivalence) a braided 2-group \cite{MR65163}, which linearizes to a braided unitary tensor category.
Generalizations of these constructions in the relative context were studied by Kawahigashi \cite{MR1230287, MR1321700} and utilized in the study of orbit equivalence of group actions by Ioana \cite{MR2342972}.
In his groundbreaking work \cite{MR1317367, MR1339767}, Popa introduced notions of approximately inner and centrally free for finite index subfactors,
which played a key role in his remarkable classification result for subfactors in terms of their standard invariants.
In \cite[Def.~2.5]{MR1317367},
Popa also considers a definition of a centrally trivial subfactor (as an `opposite' to his notion of centrally free), and in \cite[Rem.~2.7]{MR1317367}, he discusses how this definition and his definition of approximately inner subfactor have a natural generalization to bimodules.
He introduces the unitary tensor category $\tilde{\chi}(M)$ of dualizable approximately inner and centrally trivial bimodules of a $\rm{II}_{1}$ factor, generalizing Connes' $\chi(M)$.
Popa asks whether this $\tilde{\chi}(M)$ is a `commutative' tensor category, i.e., does it admit a braiding?
In this article, we answer Popa's question positively.
\begin{thmalpha}
\label{thm:ExistsBraiding}
Let $M$ be a $\rm{II}_{1}$ factor.
Then $\tilde{\chi}(M)$ admits a unitary braiding (see Equation \eqref{eq:uXY}).
Furthermore, if
$N$ is another $\rm{II}_{1}$ factor stably isomorphic to $M$, then $\tilde{\chi}(M)\cong \tilde{\chi}(N)$ as braided unitary tensor categories.
\end{thmalpha}
Notably, $\tilde{\chi}(M)$ recovers Connes' $\chi(M)$ as the equivalence classes of invertible bimodules in $\tilde{\chi}(M)$ whose left and right von Neumann dimension are equal (to one), and our braiding on $\tilde{\chi}(M)$ recovers Jones' $\kappa$ (see Example \ref{ex:ConnesChi} for more details).
We may thus think of $\tilde{\chi}(M)$ as a unitary braided categorical extension of the braided 2-group $(\chi(M), \kappa)$.
The existence of a braiding on $\tilde{\chi}(M)$ is surprising from a categorical viewpoint.
Indeed, von Neumann algebras form a 2-category whose 1-morphisms are bimodules and whose 2-morphisms are intertwiners.
The unitary tensor category $\tilde{\chi}(M)$
is a full subcategory of $\operatorname{End}(M)\cong \mathsf{Bim}(M)$ in this 2-category.
Braidings arise formally in the context of 3-categories by looking at endomorphisms of some identity 1-morphism.
This is the algebraic structure underlying conformal nets, which produces unitary modular tensor categories in the rational setting \cite{MR1838752,MR3439097,MR3773743}.
The presence of a braiding, together with its behavior under local extensions in Theorem \ref{thmalpha:LocalExtension} below, suggests that von Neumann algebras may be objects in a yet to be discovered 3-category.
As expected from experience with Connes' $\chi$, using \cite{MR377534} and \cite{MR1317367,MR2661553},
it is straightforward to show that
$\tilde{\chi}(R)$, $\tilde{\chi}(N)$ and $\tilde{\chi}(R\otimes N)$ are trivial where $R$ denotes the hyperfinite $\rm{II}_{1}$ factor and $N$ is any non-Gamma $\rm{II}_{1}$ factor.
In order to leverage these facts to compute some non-trivial examples, we prove the following theorem, which is similar in spirit to Connes' short exact sequence \cite{MR377534} (c.f.~\cite[Prop.~6.4]{MR3424476} for a parallel result in the conformal net context).
\begin{thmalpha}
\label{thmalpha:LocalExtension}
Let $N\subseteq M$ be a finite index $\rm{II}_{1}$ subfactor such that $Q:={}_{N} L^{2}M_{N}\in\tilde{\chi}(N)$ is a commutative Q-system.
Then $\tilde{\chi}(M)\cong \tilde{\chi}(N)_Q^\operatorname{loc}$ as braided unitary tensor categories.
\end{thmalpha}
In \cite{MR2661553}, Popa studied Connes' $\chi$ in the context of inductive limits of Jones towers of finite depth finite index non-Gamma $\rm II_1$ subfactors $N\subseteq M$.
He shows $\chi(M_{\infty})=1$ for a large class of non-Gamma inclusions $N\subseteq M$ for which $M_{\infty}$ is McDuff but not isomorphic to $R\otimes N$ for $N$ non-Gamma, resolving a question of Connes. Building off Popa's techniques, in this paper we will directly compute $\tilde{\chi}(M_{\infty})$ as a unitary braided tensor category.
To state our results, recall the standard invariant of a finite depth finite index $\rm II_1$ subfactor $N\subseteq M$ consists of the indecomposable $2\times 2$ multifusion category $\cC(N\subseteq M)$
of $N-N$, $N-M$, $M-M$, and $M-N$ bimodules generated by $L^2M$, together with the choice of generating object ${}_NL^2M_M$.
We define Morita equivalence of two standard invariants $\cC(N_{1}\subseteq N_2)$ and $\cC(M_1\subseteq M_{2})$ as Morita equivalence of the underlying multifusion categories \cite[\S7.12]{MR3242743}.
The Drinfeld center $\cZ(\cC(N\subseteq M))$ is a braided unitary fusion category, and indecomposable multifusion categories are Morita equivalent if and only if their Drinfeld centers are equivalent as braided fusion categories \cite[\S8.5]{MR3242743}. We have the following theorem.
\begin{thmalpha}
\label{thmalpha:InductiveLimit}
Let $N\subseteq M$ be a finite depth finite index inclusion of non-Gamma $\rm{II}_{1}$ factors, and let $M_{\infty}$ denote the inductive limit factor of the Jones tower.
Then $\tilde{\chi}(M_{\infty})\cong \cZ(\cC(N\subseteq M))$ as braided unitary tensor categories.
\end{thmalpha}
We get the following immediate corollary.
\begin{coralpha}
If $N_{1}\subseteq N_{2}$ and $M_{1}\subseteq M_{2}$ are finite depth inclusions of non-Gamma $\rm{II}_{1}$ factors with $\cC(N_{1}\subseteq N_{2})$ not Morita equivalent to $\cC(M_{1}\subseteq M_{2})$, then the $\rm II_1$ factors $N_{\infty}$ and $M_{\infty}$ are not stably isomorphic.
\end{coralpha}
This corollary shows that remarkably, the stable isomorphism class of the inductive limit $\rm{II}_{1}$ factor $M_\infty$ remembers the standard invariant of the initial finite depth subfactor $N\subseteq M$ up to Morita equivalence.
In fact, our result complements the rigidity result of Popa, which states that the $\rm{II}_{1}$ factor $M_{\infty}$ remembers the inclusion $N\subseteq M$ up to weak equivalence \cite[Def.~3.1.4 and Cor.~3.6]{MR2888236}. As another consequence, our computation for $\tilde{\chi}(M_{\infty})$ also computes the ordinary $\chi(M_{\infty})$ and $\kappa$ invariants as the braided subcategory of invertible objects in $\cZ(\cC(N\subseteq M))$.
By Popa's result \cite[Thm.~4.2]{MR2661553}, this is isomorphic to the relative $\chi(M'\cap M_{\infty} \subseteq N'\cap M_{\infty} )$ studied by Kawahigashi \cite{MR1230287}.
The latter inclusion is a finite index hyperfinite $\rm II_1$ subfactor with standard invariant equivalent to $\cC(N\subseteq M)$.
There are many examples of finite depth non-Gamma inclusions. Popa and Shlyakhtenko showed that every subfactor standard invariant can be realized as an inclusion of $\rm II_1$ factors isomorphic to $L\bbF_{\infty}$ \cite{MR2051399}.
In \cite{MR2732052}, Guionnet-Jones-Shlyakhtenko provide an alternative realization of finite depth standard invariants as inclusions of interpolated free group factors \cite{MR2807103}
in their diagrammatic reproof of Popa's celebrated subfactor reconstruction theorem \cite{MR1278111}.
As every indecomposable unitary multifusion category is Morita equivalent to any of its unitary fusion category diagonal summands, we obtain the following corollary.
\begin{coralpha}
For $\cC$ a unitary fusion category, its Drinfeld center $\cZ(\cC)$ is realized as $\tilde{\chi}(M)$ for some McDuff $\rm{II}_{1}$ factor $M$.
\end{coralpha}
While we have stated our results above for $\tilde{\chi}(M)$, our analysis actually occurs in a much more general categorical setting.
We define the notions of approximately inner and centrally trivial for endofunctors on an arbitrary $\rm W^*$-category with separable preduals (see \S\ref{ApproxInnerCentTrival}). Functors which are both are said to be \textit{local}.
We construct a unitary braiding on this category (without any dualizability assumptions), and prove all the axioms are satisfied here.
Thus we get a new construction of a canonical braided $\rm W^*$-tensor category ${\End}_{\loc}(\cC)$ from an arbitrary $\rm W^*$-category $\cC$.
When $\cC=\mathsf{Mod_{fgp}}(M)$, the finitely generated projective modules of a separable finite von Neumann algebra, under the well known equivalence between $\operatorname{End}(\mathsf{Mod_{fgp}}(M))$ and $\mathsf{Bim_{fgp}}(M)$,
our definitions of approximately inner and centrally trivial agree with Popa's from \cite{MR1317367}.
We use this equivalence to express the braiding as a bimodule intertwiner in Equation \eqref{eq:uXY}.
Aside from the greater generality, one of the reasons we use the functor language for our constructions and proofs is that the category of endofunctors is strict, making some commuting diagrams significantly simpler.
In addition, in the categorically oriented functor approach, the concept of approximately commuting diagram significantly reduces the complexity of proofs and their verification.
\subsection*{Outline}
In \S\ref{sec:Preliminaries}, we recall the notion of $\rm W^*$-category, and we pay special attention to the $\rm W^*$ 2-category of von Neumann algebras, bimodules, and intertwiners.
In \S\ref{sec:CMC*Cat}, we discuss the canonical $\sigma$-strong* topology on the hom spaces of a $\rm W^*$-category, which is essential to our construction.
In \S\ref{sec:W*ApproxNatTrans}, we introduce the notion of an approximate natural transformation between endofunctors of a $\rm W^*$-category, which we use to define the notions of approximately inner and centrally trivial for endofunctors.
We show that the local endofunctors which are both approximately inner and centrally trivial admit a canonical unitary braiding in \S\ref{sec:RelativeBraiding}.
In \S\ref{sec:Translation}, we translate our construction into the language of dualizable bimodules over a $\rm II_1$ factor $M$, and we calculate many examples of $\tilde{\chi}(M)$ in \S\ref{sec:Examples}.
In \S\ref{sec:LocalExtensionChapter}, we prove Theorem \ref{thmalpha:LocalExtension}, and in \S\ref{sec:Calculation}, we prove Theorem \ref{thmalpha:InductiveLimit}.
To prove these theorems, we make heavy use of the Q-system realization machinery developed in \cite{MR3948170,2105.12010} and the graphical calculus for unitary tensor categories.
\subsection*{Acknowledgements}
This project began through conversations with Vaughan Jones.
This work would not have been possible without his generous support and encouragement.
We dedicate this article to his memory.
The authors would like to thank
Dietmar Bisch,
Andr\'e Henriques,
Yasuyuki Kawahigashi,
Sorin Popa, and
Stefaan Vaes
for helpful comments and conversations.
Quan Chen and David Penneys were supported by NSF grant DMS 1654159.
Corey Jones was supported by NSF grant DMS 1901082/2100531.
\section{Preliminaries}
\label{sec:Preliminaries}
We assume the reader is relatively familiar with the basics of von Neumann algebras, in particular $\rm II_1$ factors, where our main references include \cite{MR1873025,MR2188261,JonesVNA,ClaireSorinII_1}.
Most von Neumann algebras that appear in this article are assumed to be \emph{separable} (their preduals are separable), with the exception of ultraproducts in \S\ref{sec:Examples} below.
We also assume the reader is relatively familiar with the basics of tensor categories and 2-categories, where our main references include \cite{MR3242743,MR3971584,2002.06055}.
Of particular importance is the graphical string diagrammatic calculus for 2-categories and tensor categories \cite[\S1.1.1 and 8.1.2]{MR3971584}.
For a 2-category $\cC$, objects are represented by 2D shaded regions, 1-morphisms are represented by labelled 1D strands read from left-to-right, and 2-morphisms are represented by labelled 0D coupons which are read bottom-to-top.
1-composition is read left-to-right similar to the relative tensor product of bimodules, and 2-composition is read bottom-to-top.
These string diagrams are formally dual to pasting diagrams, and typically associators and unitors are suppressed whenever possible.
As a tensor category is equivalent to a 2-category with one object, the graphical calculus for tensor categories has no shadings for regions;
objects are represented by labelled 1D strings, and morphisms are represented by labelled 0D coupons read bottom-to-top.
Tensor product is read left-to-right, and composition of morphisms is read bottom-to-top.
Our 2-categories and tensor categories are $\rm C^*/W^*$ (see \S\ref{sec:W*Cat} and \ref{sec:Modules} below), and we represent the $\dag$-operation by vertical reflection of diagrams.
\subsection{\texorpdfstring{$\rm C^*/W^*$}{C*/W*}-categories}
\label{sec:W*Cat}
We begin with the basics of $\rm C^*$ and $\rm W^*$-categories.
The latter were first introduced in \cite{MR808930}.
\begin{defn}
A \textit{$\rm C^*$-category} is a $\bbC$-linear category $\cC$ such that:
\begin{itemize}
\item
for each pair of objects $a,b\in \cC$, there is a conjugate linear involution $\dag:\cC(a\to b)\rightarrow \cC(b\to a)$, satisfying $(f\cdot g)^\dag=g^\dag\cdot f^\dag$,
\item
for each pair of objects $a,b\in \cC$, there is a Banach norm on $\cC(a\to b)$ satisfying $\|f\|^{2}=\|f^\dag\cdot f\|=\|f\cdot f^\dag\|$ for all $f\in \cC(a\to b)$, and
\item
for all $f\in \cC(a\to b)$, $f^\dag\cdot f$ is a positive element in the $\rm C^*$-algebra $\cC(a\to a)$.
That is, there is a $g\in \cC(a\to a)$ such that $f^\dag \cdot f = g^\dag \cdot g$.
\end{itemize}
A \emph{$\rm W^*$-category} is a $\rm C^*$-category such that every hom space $\cC(a\to b)$ admits a predual Banach space.
We call a $\rm W^*$-category \emph{separable} if all such preduals are separable Banach spaces.
\end{defn}
\begin{assumption}
In this article, we assume all $\rm C^*/W^*$-categories are \emph{unitarily Cauchy complete}, meaning they admit all finite orthogonal direct sums and are orthogonal projection complete.
There is a formal construction to complete any $\rm C^*/W^*$-category which satisfies a universal property; we refer the reader to \cite[\S3.1.1]{1810.06076} for more details.
\end{assumption}
\begin{rem}
Every unitarily Cauchy complete $\rm C^*/W^*$-category admits a canonical left $\mathsf{Hilb_{fd}}$-module category structure.
That is, for each $c\in \cC$ and finite dimensional Hilbert space $H$, there is an object $H\rhd c\in \cC$, unique up to canonical unitary isomorphism.
Moreover, there is a canonical unitary associator $H\rhd (K\rhd a)\cong (H\otimes K)\rhd a$.
In the sequel, we will assume our $\mathsf{Hilb_{fd}}$-module category structure is \emph{strictly unital}, i.e., $\bbC\rhd c = c$ for all $c\in \cC$.
\end{rem}
\begin{defn}
A \emph{$\dag$-functor} between between $\rm C^*$-categories is a functor $F: \cC\to \cD$ such that $F(f^\dag) = F(f)^\dag$ for all morphisms $f$ in $\cC$.
Two $\rm C^*$-categories are \emph{(unitarily) equivalent} if there are $\dag$-functors each way whose appropriate composites are unitarily naturally isomorphic to the appropriate identity $\dag$-functors.
For $\rm W^*$-categories, one restricts to the \emph{normal} $\dag$-functors which are weak*-continuous on hom spaces.
Equivalence is defined similarly as before, but restricting to normal $\dag$-functors.
\end{defn}
\begin{ex}
\label{ex:RightFGPMod}
The most important example of a $\rm W^*$-category for our article is the finitely generated projective right modules for a $\rm II_1$ factor $M$.
There are two dagger equivalent such categories that one can work with:
\begin{itemize}
\item
Hilbert spaces $H$ equipped with a normal right $M$-action such that the von Neumann dimension $\dim(H_M)$ is finite, or
\item
finitely generated projective right Hilbert $\rm W^*$-modules (see \S\ref{sec:Modules} below for more details).
\end{itemize}
To see the equivalence, the map from the first to the second is taking bounded vectors (the $\xi \in H$ such that $\widehat{m} \mapsto \xi m$ extends to a bounded map $L^2M \to H$), and the map from the second to the first is $-\otimes_M L^2M$ (the inner product is given by $\langle \eta\otimes \widehat{m} , \xi\otimes \widehat{n} \rangle:=\langle \langle \xi | \eta\rangle_M \widehat{m}, \widehat{n}\rangle_{L^2M}$).
We will use the second definition above for the convenience that we may state many results for all $\xi \in X_M$ rather than for all bounded vectors.
However, one can work with the first definition provided that one restricts to bounded vectors when appropriate.
\end{ex}
\begin{ex}
For a separable $\rm C^*$-algebra $A$, $\mathsf{Rep}(A)$ is the $\rm W^*$-category of (non-degenerate) $*$-representations of $A$ on separable Hilbert spaces. This category is relevant in the operator algebraic study of quantum statistical mechanics.
\end{ex}
\begin{defn}
\label{defn:sigmaFH}
Given a $\mathsf{Hilb_{fd}}$-module $\rm C^*$ category $\cC$, a finite dimensional Hilbert space $H$, and any $\dag$-functor $F\in \operatorname{End}(\cC)$, we have a canonical \emph{braid-like} unitary natural isomorphism
$$
\sigma_{F,H}: F(H\rhd -)\rightarrow H\rhd F(-)
$$
defined as follows.
For an orthonormal basis $\{e_{i}\}$, we may identify its elements as bounded operators $e_i\in B(\bbC, H)$ defined by $1\mapsto e_{i}$.
Then $e^{\dag}_{i}\in B(H, \bbC)$ is given by $e^{\dag}_{j}e_{i}=\delta_{i,j}$.
We then define $\sigma_{F,H}$ in components by
$$
\sigma^{a}_{F,H}:=\sum_{i} (e_{i}\rhd F(1_{a}))\cdot F(e^{\dag}_{i}\rhd 1_{a}),
$$
which does not depend on the choice of orthonormal basis of $H$.
Unitarity is straightforward to verify.
To show naturality, let $f\in \cC(a\to b)$.
Then
\begin{align*}
(1_{H}\rhd F(f))\cdot \sigma^{a}_{F,H}
& = \sum_{i} (e_{i}\rhd F(f))\cdot F(e^{\dag}_{i}\rhd 1_{a}) = \sum_{i} (e_{i}\rhd 1_{b})\cdot F(f)\cdot F(e^{\dag}_{i}\rhd 1_{a})
\\
& = \sum_{i} (e_{i}\rhd 1_{b})\cdot F(e^{\dag}_{i}\rhd f)=\sigma^{b}_{F,H}\cdot F(1_{H}\rhd f). \qedhere
\end{align*}
\end{defn}
The family $\sigma$ also satisfies the following monoidality conditions (where we have suppressed the module category associator).
\begin{prop}
\label{prop:sigmabraid}
\mbox{}
\begin{enumerate}
\item For any $F\in \operatorname{End}(\cC)$, $\sigma^{a}_{F,H\otimes K}=(1_{H}\rhd \sigma^{a}_{F,K})\cdot \sigma^{K\rhd a}_{F,H}$.
\item For any $G\in \operatorname{End}(\cC)$ we have $\sigma^{a}_{F\circ G, H}=\sigma^{G(a)}_{F,H}\cdot F(\sigma^{a}_{G,H})$.
\end{enumerate}
\end{prop}
\begin{proof}
Let $\{f_j\}$ be an orthonormal basis for $K$, which we identify with bounded operators $f_j\in B(\bbC,K)$.
Then
\begin{align*}
(1_H\rhd\sigma^a_{F,K})\cdot \sigma^{K\rhd a}_{F,H}
& = \sum_{i,j} (1_H\rhd f_j\rhd F(1_a))\cdot (1_H\rhd F(f_j^\dag\rhd 1_a))\cdot (e_i\rhd F(1_{K\rhd a})\cdot F(e_i^\dag\rhd 1_{K\rhd a}) \\
& = \sum_{i,j} (1_H\rhd f_j\rhd F(1_a))\cdot (e_i\rhd F(1_a))\cdot F(f_j^\dag\rhd 1_a) \cdot F(e_i^\dag\rhd 1_{K\rhd a})
= \sigma^a_{F,H\otimes K} \\
\intertext{and}
\sigma^{G(a)}_{F,H}\cdot F(\sigma^{a}_{G,H})
& = \sum_i (e_i\rhd FG(1_a))\cdot F(e_i^\dag\rhd G(1_a))\cdot F(e_i\rhd G(1_a))\cdot FG(e_i^\dag\rhd 1_a) \\
& = \sum_i (e_i\rhd FG(1_a))\cdot FG(e_i^\dag\rhd 1_a)
= \sigma^{a}_{F\circ G, H}.
\qedhere
\end{align*}
\end{proof}
\subsection{Modules and correspondences of \texorpdfstring{$\rm W^*$}{W*}-algebras}
\label{sec:Modules}
We now recall the definition of the $\rm W^*$ 2-category $\mathsf{W^*Alg_{fgp}}$ of finitely generated projective right $\rm W^*$-correspondences, after which we formally define the finitely generated projective right modules $\mathsf{Mod_{fgp}}(M)$.
Our exposition follows \cite[\S2.2]{2105.12010}, which was adapted from \cite[\S8]{MR2111973}.
Other references include \cite{MR355613,MR0367670}.
\begin{defn}
The $\rm W^*$ 2-category $\mathsf{W^*Alg_{fgp}}$ is given as follows.
\begin{itemize}
\item
objects are von Neumann algebras
\item
1-morphisms are finitely generated projective right $\rm W^*$-corresponendences.
In more detail, given von Neumann algebras $A,B$, a 1-morphism ${}_AX_B$ is a Banach space equipped with a right $B$-action and a right $B$-valued inner product satisfying
\begin{itemize}
\item $\langle \eta| \xi_1 +\xi_2 b\rangle_B = \langle \eta | \xi_1\rangle_B + \langle \eta | \xi_2\rangle_B b$,
\item $\langle \eta_1 + \eta_2 b| \xi\rangle_B = \langle \eta_1 | \xi\rangle_B+ b^*\langle \eta_2 | \xi\rangle_B$,
\item $\langle \eta|\xi\rangle_B^* = \langle \xi | \eta\rangle_B$, and
\item $\langle \xi | \xi \rangle_B \geq 0$ with equality if and only if $\xi = 0$.
\end{itemize}
By the Cauchy-Schwarz inequality, $\|\langle \xi|\xi\rangle_B\|_B$ defines a norm on $X$, which is required to be complete.
Moreover, we require the left $A$-action to be by adjointable operators.
The finitely generated projective condition says that as a right $B$-module, $X_B$ is unitarily isomorphic to $pB^n$ for some (adjointable) orthogonal projection $p\in \operatorname{End}_{-B}(B^n)$.
The $\rm W^*$ condition amounts to requiring that:
\begin{itemize}
\item ${}_AX_B$ has a predual,
\item the $B$-valued inner product $\langle \,\cdot\,|\,\cdot\,\rangle_B$ is separately weak*-continuous, and
\item the left $A$-action $A\to \operatorname{End}(X_B)$ is normal.
\end{itemize}
Composition of 1-morphisms is the relative tensor product.
\item
2-morphisms ${}_AX_B\Rightarrow {}_AY_B$ are the adjointable right $B$-linear operators that commute with the left $A$-action.
\end{itemize}
\end{defn}
\begin{defn}
For a von Neumann algebra $M$, we define $\mathsf{Mod_{fgp}}(M):=\mathsf{W^*Alg_{fgp}}(\bbC \to M)$.
Given $X_M \in \mathsf{Mod_{fgp}}(M)$, a finite $X_M$-basis
is a finite subset $\{\beta\}\subseteq X$ such that $\sum_{\beta} \beta\langle \beta|\xi\rangle_M = \xi$ for all $\xi\in X$ \cite[\S1.1.3]{MR1278111},
\cite[\S3.1.1]{1111.1362}.
As we only work with finitely generated projective modules in this article, all $X_M$-bases will be finite, so we omit the word `finite' without confusion.
We call such a basis \emph{orthogonal} if $\langle \beta|\beta'\rangle_M$ is equal to $\delta_{\beta=\beta'}$ times an orthogonal projection in $M$.
Given an $X_M$-basis, one can always obtain an orthogonal $X_M$-basis using the Gram-Schmidt orthogonalization procedure \cite[Lem.~8.5.2]{ClaireSorinII_1}.
\end{defn}
\begin{rem}
\label{rem:OppositeIssue}
It is well known that the $\rm W^*$-tensor category $\operatorname{End}(\mathsf{Mod_{fgp}}(M))$ of normal $\dag$-endofunctors is dagger equivalent to $\mathsf{Bim_{fgp}}(M)^{\rm mp}$, the monoidal opposite of $\mathsf{Bim_{fgp}}(M)$.
That is, every normal $\dag$-endofunctor of $\mathsf{Mod_{fgp}}(M)$ is of the form $-\boxtimes_M X$ for some $X\in \mathsf{Mod_{fgp}}(M)$.
For example, this equivalence is explained in \cite[\S3.2]{2004.08271} for infinite von Neumann algebras using the fact that $\operatorname{End}(\mathsf{Mod_{fgp}}(M))$ is unitarily equivalent to the orthogonal projection completion of the $\rm W^*$-tensor category $\operatorname{End}(M)$.
\end{rem}
\begin{rem}
There is another way that the category of $\operatorname{End}(\mathsf{Mod_{fgp}}(N))$ is used in practice, particularly among the $\rm{II}_{1}$ factor community.
This stems from the fact that $\mathsf{Mod_{fgp}}(N)$ is equivalent to the unitary Cauchy completion of $N$ thought of as a $\rm W^*$-category with one object.
Objects in the completion are pairs $(n,p)$, where $n\in \bbN$ and $p\in M_{n}(N)$ is a projection.
Morphisms $(n,p)\to (m,q)$ are elements of $qM_{m\times n}(N)p$.
By the universal property of Cauchy completion, an endofunctor is determined by where it sends $(1,1_{N})$ together with its action on $\operatorname{End}((1, 1_{N}))\cong N$.
In other words, an endofunctor in $\operatorname{End}(\mathsf{Mod_{fgp}}(N))$ is completely determined up to unitary natural isomorphism by a (unital) homomorphism $N\to pM_{n}(N)p$ for some projection $p\in M_{n}(N)$, called a \emph{cofinite morphism} of $N$ in \cite{MR1049618}.
Furthermore, a natural transformation is uniquely determined by its $(1,1_{N})$-component.
\end{rem}
\subsection{The \texorpdfstring{$\sigma$}{sigma}-strong* topology on a \texorpdfstring{$\rm W^*$}{W*}-category}
\label{sec:CMC*Cat}
Let $\cC$ be a separable $\rm W^*$-category, which has a canonical weak* topology on each hom space.
\begin{defn}
For each $a,b\in \cC$,
the $\sigma$-\emph{strong* topology} $\tau$ on $\cC(a\to b)$
is defined as follows:
$f_i \to 0$ $\sigma$-strong* if and only if
$f_i^\dag f_i \to 0$
and
$f_if_i^\dag \to 0$
weak* ($\sigma$-weakly).
\end{defn}
\begin{facts}
The $\sigma$-strong* topology $\tau$ on the hom spaces of $\cC$ satisfies the following properties:
\begin{enumerate}[label=($\tau$\arabic*)]
\item
\label{tau:composition}
composition is jointly $\tau$-continuous on norm bounded subsets (if $\{a_{n}\}, \{b_n\}$ are uniformly norm bounded, and $a_{n}\rightarrow a$ and $b_{n}\rightarrow b$ then $a_{n}b_{n}\rightarrow ab$).
\item
\label{tau:dagcontinuity}
$\dag$ is $\tau$-continuous on norm bounded subsets of morphism spaces.
\item
\label{tau:metrizable}
$\tau$ restricted to the unit ball of any morphism space is completely metrizable.
\end{enumerate}
\end{facts}
The following proposition basically follows from \cite[III.2.2.2]{MR2188261}.
\begin{prop}
Suppose $M,N$ are von Neumann algebras and $\Phi: M \to N$ is a unital $*$-homomorphism.
Then $\Phi$ is normal if and only if it is $\sigma$-strong* continuous on the unit ball of $M$.
\end{prop}
\begin{proof}
If $\Phi$ is $\sigma$-strong* continuous on bounded sets, then for any increasing bounded net $(x_i)$ in $M$ with $x_i \nearrow x$, $x_i \to x$ in the $\sigma$-strong* topology.
Hence $\Phi(x_i)\nearrow \Phi(x)$, and $\Phi$ is normal.
The converse argument is similar to \cite[III.2.2.2]{MR2188261}.
If $x_i \to 0$ $\sigma$-strong*, then $x_ix_i^*\to 0$ and $x_i^*x_i \to 0$ $\sigma$-weakly.
Hence $\Phi(x_i)^*\Phi(x_i) \to 0$ and $\Phi(x_i)\Phi(x_i)^* \to 0$ $\sigma$-weakly, which implies $\Phi(x_i) \to 0$ $\sigma$-strong*.
\end{proof}
As all $\rm W^*$-categories were assumed to admit finite orthogonal direct sums, we have the following immediate corollary.
\begin{cor}
Suppose $\cC,\cD$ are $\rm W^*$-categories and $F: \cC\to \cD$ is a $\dag$-functor.
Then $F$ is normal if and only if $F$ is $\sigma$-strong* continuous on norm bounded subsets of hom spaces in $\cC$.
\end{cor}
\begin{ex}
Let $(M,\operatorname{tr}_M)$ be a finite von Neumann algebra equipped with a faithful tracial state $\tau$.
In this case, the $\sigma$-strong* topology on the unit ball of $M$ is exactly the $\|\cdot\|_2$-topology, where $\|x\|_2^2 := \operatorname{tr}_M(x^*x)$ \cite[Prop.~9.1.1]{JonesVNA}.
In fact, we may also describe the entire $\sigma$-strong* topology $\tau$ on $\mathsf{Mod_{fgp}}(M)$ on norm bounded sets as coming from a $\|\cdot\|_2$-norm induced by canonical commutant traces.
In more detail, for $X_M\in \mathsf{Mod_{fgp}}(M)$, the canonical \emph{commutant trace}
\cite[\S1.1.3(c)]{MR1278111},
\cite[Def.~3.1.4]{1111.1362}
on the finite von Neumann algebra $\operatorname{End}(X_M)$ is given by
$$
\operatorname{Tr}_X(f)
:=
\sum_{b} \operatorname{tr}_M(\langle b| xb\rangle^X_M)
$$
where $\{b\}$ is any finite $X_M$-basis.
Observe that $\operatorname{Tr}_X$ is independent of the choice of basis.
When $M$ is a $\rm II_1$ factor, $\operatorname{Tr}(\operatorname{id}_X)$ equals the right von Neumann dimension of $X\otimes_M L^2M$.
Observe that the maps $(\operatorname{Tr}_X)_{X\in \mathsf{Mod_{fgp}}(M)}$ endow $\mathsf{Mod_{fgp}}(M)$ with a \emph{unitary categorical trace} in the spirit of \cite[Def.~3.7]{MR3019263} (see also \cite[Def.~3.59]{1810.06076}).
Indeed, for $f\in \operatorname{Hom}(X_M \to Y_M)$,
we choose a finite $X_M$-basis $\{b\}$ and a finite $Y_M$-basis $\{c\}$,
and we calculate
\begin{align*}
\operatorname{Tr}_X(f^\dag f)
&=
\sum_{b} \operatorname{tr}_M(\langle b| f^\dag fb\rangle^X_M)
=
\sum_{b} \operatorname{tr}_M(\langle fb| fb\rangle^Y_M)
=
\sum_{b,c} \operatorname{tr}_M(\langle fb| c \langle c|fb\rangle^Y_M \rangle^Y_M )
\\&
=
\sum_{b,c} \operatorname{tr}_M(\langle fb| c\rangle^Y_M \langle c|fb\rangle^Y_M )
=
\sum_{b,c} \operatorname{tr}_M( \langle f^\dag c|b\rangle^X_M \langle b| f^\dag c\rangle^X_M )
=
\sum_{b,c} \operatorname{tr}_M( \langle b\langle b|f^\dag c\rangle^X_M| f^\dag c\rangle^X_M )
\\&=
\sum_{c} \operatorname{tr}_M( f^\dag c| f^\dag c\rangle^X_M )
=
\sum_{c} \operatorname{tr}_M( c| ff^\dag c\rangle^X_M )
=
\operatorname{Tr}_Y(ff^\dag).
\end{align*}
Using this categorical trace, for $f\in \operatorname{Hom}(X_M \to Y_M)$, we define $\|f\|_2 := \operatorname{Tr}_X(f^\dag\circ f)^{1/2}$.
Observe that $\|\cdot\|_2$ on $\operatorname{Hom}(X_M \to Y_M)$ is exactly
the restriction of
$\|\cdot\|_2$ on $\operatorname{End}(X_M \oplus Y_M)$, which is again defined using the canonical commutant trace $\operatorname{Tr}_{X\oplus Y}$.
Thus the $\sigma$-strong* topology $\tau$ on $\mathsf{Mod_{fgp}}(M)$ exactly corresponds to the $\|\cdot\|_2$-topology on norm-bounded sets.
\end{ex}
The following remark will be used later in \S\ref{sec:LocalExtension}.
\begin{rem}
Suppose $(M,\operatorname{tr}_M)$ is a finite von Neumann algebra equipped with a normal faithful trace $\operatorname{tr}_M$.
Suppose $X_M\in \mathsf{Mod_{fgp}}(M)$ and $N\subseteq (M,\operatorname{tr}_M)$ is strongly Markov inclusion \cite[Def.~2.8]{MR2812459}, i.e., there is a finite $M_N$-basis $\{c\}$ which satisfies $[M:N]:=\sum_c cc^* \in [1,\infty)$.
(This definition was based on \cite{MR945550} and \cite[\S1.1.3 and 1.1.4]{MR1278111}.)
Here, the right $N$-valued inner product on $M_N$ is given by $\langle a|b\rangle_N = E_N(a^*b)$ where $E_N: M\to N$ is the unique trace-preserving conditional expectation.
We now compare $\operatorname{Tr}_{X_M}$ and $\operatorname{Tr}_{X\boxtimes_M M_N}$.
If $\{b\}$ is a basis for $X_M$ and $\{c\}$ is a basis for $M_N$, then $\{b\boxtimes c\}$ is a basis for $X\boxtimes_M M_N$.
Thus for $f\in \operatorname{End}(X_M)$,
\begin{align*}
\operatorname{Tr}_{X\boxtimes_M M_N}(f\boxtimes \operatorname{id}_{M_N})
&=
\sum_{b,c} \operatorname{tr}_N(\langle b\boxtimes c | fb\boxtimes c\rangle^{X\boxtimes_M M_N}_N)
=
\sum_{b,c} \operatorname{tr}_N(E_N(c^* \langle b |fb\rangle^X_M c))
\\&=
\sum_{b,c} \operatorname{tr}_M(c^* \langle b |fb\rangle^X_M c)
=
\sum_{b,c} \operatorname{tr}_M(\langle b |fb\rangle^X_M cc^* )
\\&=
[M:N]
\sum_{b} \operatorname{tr}_M(\langle b |fb\rangle^X_M)
=
[M:N]
\operatorname{Tr}_X(f).
\end{align*}
This says the (faithful) restriction functor $-\boxtimes M_N : \mathsf{Mod_{fgp}}(M) \to \mathsf{Mod_{fgp}}(N)$ is a continuous embedding (a homeomorphism onto its image) of hom spaces with respect to the $\tau_M-\tau_N$ topologies.
This means that $f_n \to 0$ in $\tau_M$ if and only if $f_n \boxtimes \operatorname{id}_{M_N} \to 0$ in $\tau_N$.
\end{rem}
\section{Approximate natural transformations and local endofunctors}
\label{sec:W*ApproxNatTrans}
In this section, given a $\rm W^*$-category, we define its canonical braided $\rm W^*$-tensor category of local endofunctors which are both approximately inner and centrally trivial.
To define these notions, we first introduce the concept of an approximate natural transformation.
\subsection{Approximate natural transformations}
Suppose $\cC$ is a separable $\rm W^*$-category, and recall $\operatorname{End}(\cC)$ denotes the normal $\dag$-endofunctors of $\cC$.
We define $\ell^{\infty}(\bbN,\cC)$ as the $\rm W^*$-category with the same objects as $\cC$ and whose morphisms are uniformly norm-bounded sequences of morphisms in $\cC$.
The composition and $\dag$ in $\ell^{\infty}(\bbN,\cC)$ are defined pointwise.
\begin{defn}
For each $a,b\in \cC$, we define
\begin{align*}
\cI(a\to b)
&:=
\set{
f=(f_{n})\in \ell^{\infty}(\bbN,\cC(a\to b))
\,}{\,
f_{n}\rightarrow_{\tau} 0
}
\\
\cC^{\infty}(a\to b)
&:=
\set{f\in \ell^{\infty}(\bbN,\cC(a\to b))
\,}{\,
\parbox{6.2cm}{
$\forall\, g\in \cI(b\to c), g\cdot f\in \cI(a\to c)$
and
$\forall\, h\in \cI(d\to a), f\cdot h\in \cI(d\to b)$
}
}.
\end{align*}
We view $\cC^\infty$ as the \emph{idealizer} of $\cI$ in $\ell^{\infty}(\bbN,\cC)$.
We call $f\in \cC^{\infty}(a\to b)$ an \emph{approximate morphism}, and we say two morphisms $f,g\in \cC^{\infty}(a\to b)$ are \emph{equivalent} or \textit{approximately equal} if $f-g\in \cI(a\to b)$.
Observe that $\cC^\infty$ is a category under pointwise composition of approximate morphisms.
By \ref{tau:composition} and \ref{tau:dagcontinuity}, $\cI(a\to b)\subseteq \cC^\infty(a\to b)$ for all $a,b$, and $\cI$ defines a $\dag$-closed ideal in $\cC^\infty$.
We define a $\dag$-category $\widetilde{\cC}$
with the same objects as $\cC$, and hom spaces
$\widetilde{\cC}(a\to b):=\cC^\infty(a\to b)/\cI(a\to b)$.
For $f\in \cC^\infty(a\to b)$, we write $\widetilde{f}$ for its image in $\widetilde{\cC}(a\to b)$.
Observe we can view $\cC$ as a $\dag$-subcategory of $\widetilde{\cC}$ by mapping $f\in \cC(a\to b)$
to the image of the constant sequence $\widetilde{(f)} \in \widetilde{\cC}(a\to b)$.
In what follows,
we identify $f\in\cC(a\to b)$ with $(f)_{n\in\bbN}\in \cC^\infty(a\to b)$ and $\widetilde{(f)}_{n\in\bbN}\in \widetilde{\cC}(a\to b)$.
\end{defn}
\begin{defn}
A pasting diagram in $\cC^\infty$ is said to \emph{approximately commute} if the corresponding pasting diagram in the quotient $\widetilde{\cC}=\cC^\infty/\cI$ actually commutes.
That is, given $a,b,c,d\in\cC$
and morphisms $f\in \cC^\infty(a\to b)$, $g\in \cC^\infty(b\to d)$, $h\in \cC^\infty(a\to c)$, and $k\in \cC^\infty(c\to d)$,
the diagram
\[
\begin{tikzcd}
a \arrow{r}{f_n} \arrow[swap]{d}{h_n} & b \arrow{d}{g_n}
\\
c \arrow{r}{k_n} & d
\end{tikzcd}
\]
approximately commutes if
$g_n\cdot f_n-k_n\cdot h_n\to_\tau 0$.
\end{defn}
For a normal $\dag$-endofunctor $F\in\operatorname{End}(\cC)$, for any $f=(f_n)\in\cC^\infty(a\to b)$,
$(F(f_n))\in \cC^\infty(F(a)\to F(b))$,
so $F$ descends to a $\dag$-endofunctor on $\widetilde{\cC}$.
\begin{defn}
Given two functors $F,G\in \operatorname{End}(\cC)$ an \textit{approximate natural transformation} is a family $\{\eta^{a}\in \cC^{\infty}(F(a)\rightarrow G(a) \}_{a\in \cC}$
such that for every $f\in \cC(a\to b)$, $\eta^{b}\cdot f=f\cdot \eta^{a}$ in $\widetilde{\cC}$.
In other words, the following diagram approximately commutes:
$$
\begin{tikzcd}
F(a)
\arrow[d,"F(f)"']
\arrow[r,"\eta^a_n"]
&
G(a)
\arrow[d,"G(f)"]
\\
F(b)
\arrow[r,"\eta^b_n"]
&
G(b).
\end{tikzcd}
$$
Clearly every natural transformation $\eta: F\Rightarrow G$ gives an approximate natural transformation.
\end{defn}
\begin{warn}
The collection of normal $\dag$-endofunctors of $\cC$ and approximate natural transformations between them (up to $\cI$) clearly forms a $\dag$-category
containing $\operatorname{End}(\cC)$ as a (non-full) subcategory.
It is tempting to think that this should also form a tensor category,
with tensor product being
composition of endofunctors as usual.
However, the horizontal composition of two approximate natural transformations is not well-defined in general.
This is a fundamental point: the category of endofunctors and approximately natural transformations is \emph{not} a tensor category in general, as endomorphisms of the `unit' in this category (see Definition \ref{defn:CentralSequenceAsApproximateNT}) is not a commutative algebra.
\end{warn}
\begin{defn}
We call $v\in\cC^\infty(a\to b)$ an \textit{approximate isometry} if its image $\widetilde{v}\in\widetilde{\cC}(a\to b)$ is an isometry.
For $\dag$-endofunctors $F,G\in\operatorname{End}(\cC)$, an approximate natural transformation $v:F\Rightarrow G$ is called an \textit{approximately isometric natural transformation} if $v^a\in \cC^\infty(F(a)\to G(a))$ is an approximate isometry for each $a\in\cC$.
\end{defn}
\begin{rem}(Arrow flipping)\label{rem:ArrowFlipping}
Suppose
$f\in \cC^\infty(a\to c)$, $v\in \cC^\infty(a\to b)$, and $w\in \cC^\infty(b\to c)$ such that the diagram
$$
\begin{tikzcd}
a\arrow{rr}{f_n} \arrow[swap]{dr}{v_n} & & c\\
& b \arrow[swap]{ur}{w_n} &
\end{tikzcd}
$$
approximately commutes.
\begin{itemize}
\item
If $w$ is an approximate isometry, then
$
\begin{tikzcd}
a\arrow{rr}{f_n} \arrow[swap]{dr}{v_n} & & c \arrow{dl}{w^{\dag}_n}\\
& b &
\end{tikzcd}
$
approximately commutes.
\item
If $v$ is an approximate coisometry, then
$
\begin{tikzcd}
a\arrow{rr}{f_n} & & c \\
& b \arrow{ul}{v_n^{\dag}} \arrow[swap]{ur}{w_n} &
\end{tikzcd}
$
approximately commutes.
\end{itemize}
Indeed, these remarks follow since $\widetilde{v}$, respectively $\widetilde{w}$, is an actual isometry, respectively coisometry, in $\widetilde{\cC}$.
\end{rem}
\begin{defn}
\label{defn:CentralSequenceAsApproximateNT}
An approximately natural transformation from the identity functor to itself is called a \textit{central sequence}.
\end{defn}
\begin{lem}
\label{lem:CauchyCompletionPreservesCentralSequences}
Suppose $\cC$ is a separable $\rm W^*$-category which is not necessarily unitarily Cauchy complete.
Every central sequence of $\cC$ has a canonical extension to the unitary Cauchy completion of $\cC$.
Moreover, this extension gives a bijective correspondence between
the equivalence classes of central sequences of $\cC$ and
the equivalence classes of central sequences of the unitary Cauchy completion of $\cC$.
\end{lem}
\begin{proof}
We proceed in 2 steps;
first, we show the result for the orthogonal direct sum completion $\mathsf{Add}(\cC)$, and second, we show the result for the orthogonal projection completion $\mathsf{Proj}(\cC)$.
\medskip
\item[\underline{Step 1:}]
We claim a central sequence $\eta$ of $\cC$ gives a central sequence of $\mathsf{Add}(\cC)$ by $\mathsf{Add}(\eta)_{ii}:=\pi_i^\dag \eta^{c_i} \pi_i$ for $\bigoplus_i c_i\in \mathsf{Add}(\cC)$, where $\pi_j : \bigoplus c_i \to c_j$ denotes the canonical projections satisfying $\sum_i \pi_i^\dag\pi_i = \operatorname{id}_{\bigoplus_i c_i}$ and $\pi_i \circ \pi_j^\dag = \delta_{i=j}\operatorname{id}_{c_i}$.
Let $f^{n}=(f^{n}_{ji})\in \mathsf{Add}(\cC)^\infty(\bigoplus_{i} c_{i}\rightarrow \bigoplus_{j} d_{j})$, with each $(f^{n}_{ji})\in \cC^\infty(c_{i}\rightarrow d_{j})$ a bounded sequence of morphisms.
Then clearly $(f^{n})\in \cI(\bigoplus_i c_{i}\rightarrow \bigoplus_j d_{j}) $ if and only if each $(f^{n}_{ji})\in \cI(c_i \rightarrow d_{j})$.
Therefore since $\eta\in \cC^{\infty}$, we have $\mathsf{Add}(\eta)\in\mathsf{Add}(\cC)^\infty$.
To see that $\mathsf{Add}(\eta)$ defines a natural transformation of the identity functor, let $f=(f_{ji})\in \mathsf{Add}(\cC)(\bigoplus c_{i}\rightarrow \bigoplus d_{j})$ as above.
Then in $\widetilde{\cC}$, $\mathsf{Add}(\eta)f=(\eta^{d_j}_{n}f_{ji})= (f_{ji}\eta^{c_i}_{n})=f\mathsf{Add}(\eta)$.
Moreover, the assignment $\eta\mapsto \mathsf{Add}(\eta)$ clearly preserves equivalence of central sequences.
Conversely, given a central sequence $\eta$ of $\mathsf{Add}(\cC)$, we automatically get a central sequence of $\cC$ by considering the canonical embedding $\cC \hookrightarrow \mathsf{Add}(\cC)$.
Moreover, for $\bigoplus_i c_i \in \mathsf{Add}(\cC)$, the off-diagonal terms of $\eta^{\bigoplus_i c_i}$ go to zero $\sigma$-strong* as $\eta^{\bigoplus_i c_i}$ approximately commutes with the $\pi_j$, and the diagonal term corresponding to $c_i$ must be approximately equivalent to $\eta^{c_i}$.
\medskip
\item[\underline{Step 2:}]
A central sequence $\eta$ of $\cC$ gives a central sequence of $ \mathsf{Proj}(\cC)$ by defining
$\mathsf{Proj}(\eta)^{(c,p)}:= p\eta^c p$ for $(c,p)\in \mathsf{Proj}(\cC)$.
Given $(a,p),(b,q)\in \mathsf{Proj}(\cC)$,
$\cI((a,p) \to (b,q)) = q\cI(a\to b)p$, so $\mathsf{Proj}(\eta)\in \mathsf{Proj}(\cC)^\infty$. Moreover, this construction preserves equivalence of central sequences.
Conversely, given a central sequence $\eta$ of $\mathsf{Proj}(\cC)$, we automatically get a central sequence of $\cC$ by considering the canonical embedding $\cC \hookrightarrow \mathsf{Proj}(\cC)$.
Clearly starting with a central sequence in $\cC$, extending to $\mathsf{Proj}(\cC)$ as above, and restricting back to $\cC$ yields the same central sequence.
In the other direction, let $\eta$ be a central sequence in $\mathsf{Proj}(\cC)$.
We need to show that in $\widetilde{\mathsf{Proj}(\cC)}$, $p\eta^{(c,\operatorname{id}_{c})}p= \eta^{(c,p)}$.
But note we can view an orthogonal projection $p\in \operatorname{End}_\cC(c)$ as a morphism in $\mathsf{Proj}(\cC)((c,p)\to (c,\operatorname{id}_c))$.
Then in $\widetilde{\mathsf{Proj}(\cC)}$, $\eta^{(c,p)}=\eta^{(c,p)}p=(\eta^{(c,p)}p)p=p\eta^{(c,\operatorname{id}_c)}p$ as desired.
\end{proof}
\begin{ex}
\label{ex:MCentralSequencesAgree}
For a separable von Neumann algebra $M$, central sequences in the sense of Definition \ref{defn:CentralSequenceAsApproximateNT} for the $\rm W^*$-category $\mathsf{Mod_{fgp}}(M)$ exactly agree with the usual notion of central sequences of $M$.
Indeed, consider the category with one object $M_M$ whose endomorphisms is $M$ acting by left multiplication.
An approximate natural transformation of the identiy functor is exactly a sequence $(x_n)$ such that $x_n m - mx_n \to 0$ in the $\sigma$-strong* topology for all $m\in M$.
Now since $\mathsf{Mod_{fgp}}(M)$ is the unitary Cauchcy completion of this one object category, the claim follows by Lemma \ref{lem:CauchyCompletionPreservesCentralSequences}.
\end{ex}
The next lemma will be important in the next subsection.
\begin{lem}
\label{Lem:ApproNT&CentralSq}
Suppose $H,K$ are finite dimensional Hilbert spaces with orthonormal bases $\{e_i\},\{f_j\}$ respectively.
For a collection of maps $(\eta^a \in \cC^\infty (H\rhd a \to K\rhd a))_{a\in \cC}$,
we define $\eta^a_{i,j}:=(f_j^\dag\rhd 1_a)\eta^a(e_i\rhd 1_a)\in \cC^\infty(a\to a)$.
Then
$\eta:H\rhd- \Rightarrow K\rhd-$ defines an approximate natural transformation if and only if $\eta_{i,j}$ is a central sequence for each $i,j$.
\end{lem}
\begin{proof}
Note that $\eta^a = \sum_{i,j} (f_j\rhd 1_a)\eta^a_{i,j}(e_i^\dag\rhd 1_a)$.
Then for $g\in\cC(a\to b)$,
\begin{align*}
g\eta^a_{i,j} - \eta^b_{i,j} g
& = g (f_j^\dag\rhd 1_a) \eta^a (e_i\rhd 1_a) - (f_j^\dag\rhd 1_b) \eta^b (e_i\rhd 1_b) g \\
& = (f_j^\dag\rhd 1_a) (1_K\rhd g)\eta^a (e_i\rhd 1_a) - (f_j^\dag\rhd 1_b) \eta^b (1_H\rhd g)(e_i\rhd 1_b) \\
& = (f_j^\dag\rhd 1_b) \left((1_K\rhd g) \eta^a -\eta^b (1_H\rhd g)\right) (e_i\rhd 1_a) ,
\\
\intertext{and}
(1_K\rhd g) \eta^a - \eta^b (1_H\rhd g)
&=
\sum_{i,j} (f_j\rhd 1_b)\left(g \eta^a_{i,j} - \eta^b_{i,j} g \right) (e_i^\dag\rhd 1_a).
\end{align*}
Thus
$(1_K\rhd g) \eta^a - \eta^b (1_H\rhd g) =0$ in $\widetilde{\cC}$
if and only if
$g\eta^a_{i,j} - \eta^b_{i,j} g =0$ in $\widetilde{\cC}$ for all $i,j$.
\end{proof}
\subsection{Centrally trivial and approximately inner endofunctors}\label{ApproxInnerCentTrival}
For the rest of this section, $\cC$ is a fixed separable $\rm W^*$-category.
\begin{defn}
\label{Defn:CentrallyTrivial}
A functor $F\in\operatorname{End}(\cC)$ is called \textit{centrally trivial} if for all finite dimensional Hilbert spaces and all approximate natural transformations $\eta:H\rhd- \Rightarrow K\rhd-$, the following diagram approximately commutes.
$$
\begin{tikzcd}
F(H\rhd a)
\arrow{r}{F(\eta^{a}_n)}
\arrow[swap]{d}{\sigma^{a}_{F,H}}
& F(K\rhd a)
\arrow{d}{\sigma^{a}_{F,K}}
\\
H\rhd F(a)
\arrow{r}{\eta^{F(a)}_n}
& K\rhd F(a)
\end{tikzcd}
$$
\end{defn}
\begin{prop}
\label{Prop:CentralTrivial&CentralSq}
A functor $F$ is centrally trivial if and only if for all central sequences $\eta$,
$F(\eta^a)=\eta^{F(a)}$ in $\widetilde{\cC}$.
\end{prop}
\begin{proof}
It is clear that $F(\eta^a)=\eta^{F(a)}$ in $\widetilde{\cC}$ for all central sequences $\eta$ if $F$ is centrally trivial.
Conversely,
for each approximate natural transformation $\eta:H\rhd -\Rightarrow K\rhd -$,
$\eta_{i,j}$ defined in Lemma \ref{Lem:ApproNT&CentralSq} is a central sequence for each $i,j$.
Then by Definition \ref{defn:sigmaFH},
\begin{align*}
\sigma^a_{F,K} F(\eta^a) (\sigma^a_{F,H})^\dag
& = \sum_j (f_j\rhd F(1_a)) F(f_j^\dag\rhd 1_a) F(\eta^a) \sum_i F(e_i\rhd 1_a) (e_i^\dag\rhd F(1_a)) \\
& = \sum_{i,j} (f_j\rhd F(1_a)) F(\eta_{i,j}^a) (e_i^\dag\rhd F(1_a)) \\
& = \sum_{i,j} (f_j\rhd F(1_a)) \eta_{i,j}^{F(a)} (e_i^\dag\rhd F(1_a))
= \eta^{F(a)},
\end{align*}
which implies that $F$ is centrally trivial.
\end{proof}
\begin{defn}
We denote the full subcategory of $\operatorname{End}(\cC)$ of centrally trivial endofunctors by ${\End}_{\ct}(\cC)$.
\end{defn}
\begin{prop}
\label{prop:ctEnd(C)FullReplete}
${\End}_{\ct}(\cC)$ is a replete unitarily Cauchy complete $\rm W^*$-tensor subcategory of $\operatorname{End}(\cC)$.
\end{prop}
\begin{proof}
Suppose $G,G'$ are centrally trivial.
For each central sequence $\eta$,
by Proposition \ref{Prop:CentralTrivial&CentralSq}, we have
$$G(G'(\eta^a))=G(\eta^{G'(a)})=\eta^{G(G'(a))},$$
which implies $G\circ G'$ is centrally trivial.
Now suppose $G$ is centrally trivial and $v:F\Rightarrow G$ is an isometric natural transformation for some other endofunctor $F\in\operatorname{End}(\cC)$.
That $F(\eta^a)=\eta^{F(a)}$ for all central sequences $\eta$ follows from the following approximately commuting diagram
\begin{equation}
\label{eq:IsometryCT}
\begin{tikzcd}
F(a)
\arrow[rrr, bend left=40, "F(\eta_n^{a})"]
\arrow{r}{v^a}
\arrow[rrr, bend right=40, "\eta_n^{F(a)}"]
&
G(a)\arrow[r, bend left= 30, "G(\eta_n^{a})"]
\arrow[r, bend right=30, "\eta_n^{G(a)}"]
&
G(a)
\arrow{r}{(v^{a})^\dag}
&
F(a)
\end{tikzcd},
\end{equation}
where we have used Remark \ref{rem:ArrowFlipping} to flip the arrow on the right.
Considering the case when $v$ is unitary shows that ${\End}_{\ct}(\cC)$ is replete.
As $\cC$ admits orthogonal direct sums, so does $\operatorname{End}(\cC)$.
Suppose $G,G'$ are centrally trivial, and let $\eta$ be a central sequence.
Then
$$
(G\oplus G')(\eta^a) = G(\eta^a)\oplus G'(\eta^a) = \eta^{G(a)}\oplus \eta^{G'(a)}.
$$
By the proof of Step 1 of Lemma \ref{lem:CauchyCompletionPreservesCentralSequences}, $\eta^{G(a)}\oplus \eta^{G'(a)}=\eta^{(G\oplus G')(a)}$ in $\widetilde{\cC}$, and thus $G\oplus G'$ is centrally trivial.
As $\cC$ is orthogonal projection complete, so is $\operatorname{End}(\cC)$.
Suppose $G$ is centrally trivial and $\pi : G\Rightarrow G$ is an orthogonal projection natural transformation.
Then $\pi$ orthogonally splits in $\operatorname{End}(\cC)$, so there is an $F\in \operatorname{End}(\cC)$ and an isometry natural transformation $v: F\Rightarrow G$.
But then $F\in {\End}_{\ct}(\cC)$ by \eqref{eq:IsometryCT} above.
\end{proof}
\begin{ex}
For a finite dimensional Hilbert space $H$, the functor $H\rhd -$ is centrally trivial.
Note that the identity functor $\operatorname{id}_\cC$ is centrally trivial, and thus so is $\bigoplus_i \operatorname{id}_\cC$.
Since the functor $H\rhd -$ is equivalent to $\bigoplus_i \operatorname{id}_\cC$,
by Proposition \ref{prop:ctEnd(C)FullReplete}, $H\rhd -$ is centrally trivial.
\end{ex}
\begin{defn}
\label{Defn:ApproximatelyInner}
A functor $F\in\operatorname{End}(\cC)$ is called \textit{approximately inner} if there exists a finite dimensional Hilbert space $H$ and an approximate isometry natural transformation $v: F\Rightarrow H\rhd -$ in $\operatorname{End}(\widetilde{\cC})$, i.e.,
$$
\begin{tikzcd}
F(a) \arrow{r}{v^{a}_n} \arrow[swap]{d}{F(f)} & H\rhd a \arrow{d}{1_{H}\rhd f} \\
F(b) \arrow{r}{v^{b}_n} & H\rhd b
\end{tikzcd}
$$
approximately commutes.
The pair $(v,H)$ is called an \textit{approximating sequence} for $F$.
\end{defn}
\begin{defn}
We denote the full subcategory of $\operatorname{End}(\cC)$ of approximately inner endofunctors by ${\End}_{\ai}(\cC)$.
\end{defn}
\begin{prop}
\label{prop:aiRepleteComplete}
${\End}_{\ai}(\cC)$ is a replete unitarily Cauchy complete $\rm W^*$-tensor subcategory of $\operatorname{End}(\cC)$.
\end{prop}
\begin{proof}
Suppose $F$ and $F'$ are approximately inner, with approximating sequences $(v,H)$ and $(w,K)$ respectively.
We claim $(\alpha^{-1}_{H,K,-}\cdot v^{K\rhd -}\cdot F(w),H\otimes K)$ is an approximating sequence for $F\circ F'$,
where the unitary natural transformation $\alpha_{H,K,-}:(H\otimes K)\rhd -\Rightarrow H\rhd (K\rhd -)$ is the left module associator.
For $f\in \cC(a\to b)$, consider the diagram
$$
\begin{tikzcd}[column sep=3em]
{F(F'(a))}
\arrow["F(w^a_n)", rightarrow]{r}
\arrow["F(F'(f))", rightarrow]{d}
&
{F(K\rhd a)}
\arrow["v_n^{K\rhd a}", rightarrow]{r}
\arrow["F(1_K\rhd f)", rightarrow]{d}
&
{H\rhd (K\rhd a)}
\arrow["\alpha^\dag_{H,K,a}", rightarrow]{r}
\arrow["1_H\rhd(1_K\rhd f)", rightarrow]{d}
&
{(H\otimes K)\rhd a}
\arrow["1_{H\otimes K}\rhd f", rightarrow]{d}
\\
{F(F'(b))}
\arrow["F(w_n^b)"', rightarrow]{r}
&
{F(K\rhd b)}
\arrow["v_n^{K\rhd b}"', rightarrow]{r}
&
{H\rhd (K\rhd b)}
\arrow["\alpha^\dag_{H,K,b}"', rightarrow]{r}
&
{(H\otimes K)\rhd b}
\end{tikzcd}.
$$
The left square commutes because $F'$ is approximately inner and $F$ is $\tau$-continuous on bounded subsets.
The middle square commutes because $F$ is approximately inner.
The right square commutes because $\alpha^\dag_{H,K,-}$ is natural.
Now suppose $F$ is approximately inner with approximating sequence $(v,H)$
and $u:G\Rightarrow F$ is a isometric natural transformation for some other endofunctor $G\in \operatorname{End}(\cC)$.
It is easy to see $(v\cdot u,H)$ is an approximating sequence for $G$ from the fact that $v\cdot u$ is again approximately isometric.
The case when $u$ is unitary shows that ${\End}_{\ai}(\cC)$ is replete.
The case when $u$ is an isometry shows that since $\cC$ is orthogonal projection complete, then so is ${\End}_{\ai}(\cC)$.
Finally, as $\cC$ is orthogonal direct sum complete, the orthogonal direct sum of functors is defined.
If $F$ and $F'$ are approximately inner with approximating sequence $(v,H)$ and $(w,K)$ respectively, then $(v\oplus w, H\oplus K)$ is an approximating sequence for $F\oplus F'$, so $F\oplus F' \in {\End}_{\ai}(\cC)$.
\end{proof}
\subsection{Relative braiding between centrally trivial and approximately inner endofunctors}
\label{sec:RelativeBraiding}
The goal of this section is to show that the subcategories ${\End}_{\ct}(\cC)$ and ${\End}_{\ai}(\cC)$ of $\operatorname{End}(\cC)$ `commute' with each other in the sense of Definition \ref{Defn:centralizing} below.
\begin{prop}
\label{prop:vwdagCommuteG}
Suppose $F\in{\End}_{\ai}(\cC)$ with approximating sequence $(v,H)$ and $G\in {\End}_{\ct}(\cC)$.
Then $(v^{G(a)})^\dag\cdot\sigma^a_{G,H}\cdot G(v^a)\in\widetilde{\cC}(G(F(a))\to F(G(a)))$ is independent of the choice of approximating sequence $(v,H)$ for $F$.
\end{prop}
\begin{proof}
Suppose $(v,H)$ and $(w,K)$ are approximating sequences for the endofunctor $F\in{\End}_{\ai}(\cC)$.
Observe that $vw^\dag:K\rhd - \Rightarrow H\rhd -$ is approximately natural.
Now since $G$ is centrally trivial,
by Definition \ref{Defn:CentrallyTrivial},
$$
\begin{tikzcd}
& G(F(a))
\arrow{dr}{G(w_n^{a})} &
\\
G(H\rhd a)\arrow[swap]{d}{\sigma^{a}_{G,H}}
\arrow{ur}{G((v_n^{a})^{\dag})}
& & G(K\rhd a)
\arrow{d}{\sigma^{a}_{G,K}}
\\
H\rhd G(a)
\arrow{r}{(v_n^{G(a)})^{\dag}}
& F(G(a))
\arrow{r}{w_n^{G(a)}}
& K\rhd G(a)
\end{tikzcd}
$$
approximately commutes for each object $a\in\cC$.
Then by Remark \ref{rem:ArrowFlipping},
$$
\begin{tikzcd}
& G(F(a))
\arrow[swap]{dl}{G(v_n^{a})}
\arrow{dr}{G(w_n^{a})} &
\\
G(H\rhd a)
\arrow[swap]{d}{\sigma^{a}_{G,H}} & &
G(K\rhd a)
\arrow{d}{\sigma^{a}_{G,K}}
\\
H\rhd G(a)
\arrow{r}{(v_n^{G(a)})^{\dag}}
& F(G(a))
& K\rhd G(a)
\arrow[swap]{l}{(w_n^{G(a)})^{\dag}}
\end{tikzcd}
$$
approximately commutes.
Therefore, $(v^{G(a)})^\dag\cdot\sigma^a_{G,H}\cdot G(v^a)\in\widetilde{\cC}(G(F(a))\to F(G(a)))$ is independent of the choice of approximating sequences $(v,H)$ for $F$.
\end{proof}
\begin{lem}\label{lem:CauchySequenceBySubsequence}
Suppose $(x_n)$ is a sequence in a metric space $(X,d)$
which satisfies the following property:
\begin{itemize}
\item
For all functions $k: \bbN \to \bbN$ such that $n< k_n\leq k_{n+1}$,
$d(x_n,x_{k_n})\to 0$.
(Note here that $(x_{k_n})$ is not quite a subsequence as terms can repeat.)
\end{itemize}
Then $(x_n)$ is Cauchy.
\end{lem}
\begin{proof}
Suppose for contradiction that $(x_n)$ is not Cauchy.
Then there is an $\varepsilon>0$ such that for all $N\in \bbN$, there are $m,n>N$ such that $d(x_m,x_n)\geq \varepsilon$.
Pick a subsequence $(x_{\ell_n})$ with $\ell_1>1$ and $d(x_{\ell_{2n-1}},x_{\ell_{2n}})\geq \varepsilon$.
We now define a function $k:\bbN \to \bbN$ by
$$
k_n:=
\begin{cases}
\ell_1
&\text{if }1\le n < \ell_1
\\
\ell_{i+1}
&\text{if }\ell_i \leq n < \ell_{i+1}.
\end{cases}
$$
Since $k: \bbN \to \bbN$ satisfies $n< k_n \leq k_{n+1}$, we have $d(x_n , x_{k_n})\to 0$.
But observe that when $i$ is odd and $n=\ell_i$, we have
$$
d(x_{\ell_i} , x_{k_{\ell_i}})
=
d(x_{\ell_i} , x_{\ell_{i+1}})
\geq \varepsilon,
$$
a contradiction.
\end{proof}
\begin{thm}
\label{ThmDef:uGFa}
Suppose $F\in {\End}_{\ai}(\cC)$ and $G\in {\End}_{\ct}(\cC)$.
For each $a\in\cC$,
there exists a unique morphism in $\cC$, $u^a_{G,F}\in \cC(G(F(a))\to F(G(a)))$, such that for all approximate sequences $(v,H)$,
the following diagram approximately commutes.
$$
\begin{tikzcd}[column sep=3.4em]
{G(F(a))}
\arrow[" u^a_{G,F}", dashrightarrow]{r}
\arrow["G(v_n^a)"', rightarrow]{d}
&
{F(G(a))}
\\
{G(H\rhd a)}
\arrow["\sigma^a_{G,H}"', rightarrow]{r}
&
{H\rhd G(a)}
\arrow["(v_n^{G(a)})^\dag"', rightarrow]{u}
\end{tikzcd}
$$
\end{thm}
\begin{proof}
Suppose $(v,H)$ is an approximating sequence for $F$.
Then for any function $k:\bbN\to\bbN$ such that $n<k_n\le k_{n+1}$, $(v':=(v_{k_n})_n,H)$ is also an approximating sequence for $F$.
By Proposition \ref{prop:vwdagCommuteG},
$$(v^{G(a)})^\dag\cdot\sigma^a_{G,H}\cdot G(v^a) = ({v'}^{G(a)})^\dag\cdot\sigma^a_{G,H}\cdot G({v'}^a)$$
in $\widetilde{\cC}(G(F(a))\to F(G(a)))$.
By \ref{tau:metrizable},
the $\tau$ topology on
any bounded subspace of $\cC(G(F(a))\to F(G(a)))$ is completely metrizable.
Then by Lemma \ref{lem:CauchySequenceBySubsequence} and the definition of $\widetilde{\cC}$,
the bounded sequence
$$(v^{G(a)})^\dag\cdot\sigma^a_{G,H}\cdot G(v^a) = \left((v_n^{G(a)})^\dag\cdot\sigma^a_{G,H}\cdot G(v_n^a)\right)_n$$
is a Cauchy sequence,
and we denote $u^a_{G,F}$ to be its unique limit.
Again, by Proposition \ref{prop:vwdagCommuteG}, $u^a_{G,F}$ does not depend on the choice of the approximating sequence for $F$.
\end{proof}
\begin{prop}
\label{prop:uGFUnitary}
For $F\in {\End}_{\ai}(\cC)$ and $G\in {\End}_{\ct}(\cC)$,
$u^a_{G,F}$ is unitary.
\end{prop}
\begin{proof}
By Proposition \ref{prop:vwdagCommuteG},
for an approximating sequence $(v,H)$ for $F$,
the following diagram approximately commutes.
$$
\begin{tikzcd}[column sep=4.5em]
G(H\rhd a)
\arrow{r}{G(v_n^{a}\cdot (v_n^a)^\dag)}
\arrow[swap]{d}{\sigma^{a}_{G,H}}
& G(H\rhd a)
\arrow{d}{\sigma^{a}_{G,H}}
\\
H\rhd G(a)
\arrow{r}{v_n^{G(a)}\cdot (v_n^{G(a)})^\dag}
& H\rhd G(a)
\end{tikzcd}
$$
Since $\sigma^a_{G,H}$ is unitary, then in $\widetilde{\cC}$, we have
\begin{align*}
u^a_{G,F} \cdot (u^a_{G,F})^\dag
& = \left((v^{G(a)})^\dag\cdot \sigma^a_{G,H} \cdot G(v^a)\right)\cdot \left(G(v^a)^\dag \cdot (\sigma^a_{G,H})^\dag \cdot v^{G(a)}\right) \\
& = (v^{G(a)})^\dag\cdot \left(\sigma^a_{G,H} \cdot G(v^a) \cdot G(v^a)^\dag \cdot (\sigma^a_{G,H})^\dag \right) \cdot v^{G(a)}\\
& = (v^{G(a)})^\dag \cdot\left(v^{G(a)}\cdot(v^{G(a)})^\dag \right)\cdot v^{G(a)}\\
& = \operatorname{id}_{F(G(a))},\\
(u^a_{G,F})^\dag \cdot u^a_{G,F}
& = \left(G(v^a)^\dag \cdot (\sigma^a_{G,H})^\dag \cdot v^{G(a)}\right) \cdot \left((v^{G(a)})^\dag\cdot \sigma^a_{G,H} \cdot G(v^a)\right) \\
& = G(v^a)^\dag \cdot \left((\sigma^a_{G,H})^\dag \cdot v^{G(a)} \cdot (v^{G(a)})^\dag\cdot \sigma^a_{G,H}\right) \cdot G(v^a) \\
& = G(v^a)^\dag \cdot \left(G(v^a) \cdot G(v^a)^\dag\right) \cdot G(v^a) \\
& = \operatorname{id}_{G(F(a))}.
\end{align*}
Since the inclusion $\cC\hookrightarrow \widetilde{\cC}$ is faithful, we are finished.
\end{proof}
\begin{prop}
\label{prop:uGF-Natural}
For $F\in {\End}_{\ai}(\cC)$ and $G\in {\End}_{\ct}(\cC)$,
the family $u_{G,F}:\{u^{a}_{G,F}\}_{a\in \cC}$ is a natural transformation $G\circ F \Rightarrow F\circ G$.
\end{prop}
\begin{proof}
For $f\in\cC(a\to b)$, it suffices to prove the following diagram approximately commutes.
\[
\begin{tikzcd}
{G(F(a))}
\arrow["u^a_{G,F}", rightarrow]{rrr}
\arrow["G(F(f))", rightarrow, swap]{ddd}
\arrow["G(v_n^a)", rightarrow]{rd}
&
{}
&
{}
&
{F(G(a))}
\arrow["F(G(f))", rightarrow]{ddd}
\\
{}
&
{G(H\rhd a)}
\arrow["\sigma^a_{G,H}", rightarrow]{r}
\arrow["G(1_H\rhd f)", rightarrow, swap]{d}
&
{H\rhd G(a)}
\arrow["1_H\rhd G(f)", rightarrow]{d}
\arrow["(v_n^{G(a)})^\dag", rightarrow]{ru}
&
{}
\\
{}
&
{G(H\rhd b)}
\arrow["\sigma^b_{G,H}", rightarrow, swap]{r}
&
{H\rhd G(b)}
\arrow["(v_n^{G(b)})^\dag", rightarrow, swap]{rd}
&
{}
\\
{G(F(b))}
\arrow["u^b_{G,F}", rightarrow, swap]{rrr}
\arrow["G(v_n^b)", rightarrow, swap]{ru}
&
{}
&
{}
&
{F(G(b))}
\end{tikzcd}
\]
The top and bottom squares commute by definition of $u^a_{G,F}$ by Theorem \ref{ThmDef:uGFa}.
The left square commute because $F$ is approximately inner and $G$ is $\tau$-continuous on bounded sets.
The right square commutes by approximate naturality.
The middle square commutes by naturality of $\sigma$.
\end{proof}
In the next definition, we use strict monoidal categories simply because the case we care about is strict, but one obtains the general definition by inserting coheretors where appropriate.
\begin{defn}\label{Defn:centralizing} Let $\cC$ be a (strict) $\rm C^*$-tensor category, and $\cA, \cB$ full, replete $\dag$ tensor subcategories of $\cC$. A \textit{centralizing structure} on the pair $(\cA, \cB)$ is a family of unitary isomorphisms $u_{a,b}:a\otimes b\to b\otimes a$ for $a\in \cA$ and $b\in \cB$
satisfying the following conditions:
\begin{enumerate}
\item
(Natural) For $a,a'\in \cA$ and $f\in \cC(a\to a')$ and $b,b\in \cB$, $g\in \cC(b\to b')$, $(g\otimes f)\circ u_{a,b}=u_{a',b'}\circ (f\otimes g)$.
\item
(Braid relation 1) For $a\in \cA$, $b,b'\in \cB$, $(1_{a}\otimes u_{a,b'})\circ (u_{a,b}\otimes 1_{b'})=u_{a, b\otimes b'}$.
\item
(Braid relation 2) For $a,a'\in \cA$ and $b\in \cB$, $(u_{a,b}\otimes 1_{a'})\circ (1_{a}\otimes u_{a',b})=u_{a\otimes a', b}$.
\end{enumerate}
\end{defn}
The goal of this section is to construct a \textit{centralizing structure} for the pair of full, replete subcategories $( {\End}_{\ct}(\cC), {\End}_{\ai}(\cC))$ inside $ \operatorname{End}(\cC)$.
\begin{prop}
\label{prop:uGF-Centralizing-Natural}
For $F\in {\End}_{\ai}(\cC)$ and $G\in {\End}_{\ct}(\cC)$,
$u_{G,F}$ satisfies condition (1) in Definition \ref{Defn:centralizing}.
\end{prop}
\begin{proof}
Suppose $G,G'\in{\End}_{\ct}(\cC)$ and $F,F'\in{\End}_{\ai}(\cC)$ with approximating sequences $(v,H)$ and $(w,K)$ respectively.
Let $\eta\in\operatorname{End}(\cC)( F\Rightarrow F' )$ be a natural transformation.
We shall show $u^a_{G,F'}\cdot G(\eta^a) =\eta^{G(a)}\cdot u^a_{G,F}$.
It suffices to prove the following diagram approximately commutes.
\[
\begin{tikzcd}
{G(F(a))}
\arrow["u^a_{G,F}", rightarrow]{rrr}
\arrow["G(\eta^a)", rightarrow, swap]{ddd}
\arrow["G(v_n^a)", rightarrow]{rd}
&
{}
&
{}
&
{F(G(a))}
\arrow["\eta^{G(a)}", rightarrow]{ddd}
\\
{}
&
{G(H\rhd a)}
\arrow["\sigma^a_{G,H}", rightarrow]{r}
\arrow["G((w_n\cdot\eta\cdot v_n^\dag)^a)", rightarrow, swap]{d}
&
{H\rhd G(a)}
\arrow["(w_n\cdot\eta\cdot v_n^\dag)^{G(a)}", rightarrow]{d}
\arrow["(v_n^{G(a)})^\dag", rightarrow]{ru}
&
{}
\\
{}
&
{G(K\rhd a)}
\arrow["\sigma^a_{G,K}", rightarrow, swap]{r}
&
{K\rhd G(a)}
\arrow["(w_n^{G(a)})^\dag", rightarrow, swap]{rd}
&
{}
\\
{G(F'(a))}
\arrow["u^a_{G,F'}", rightarrow, swap]{rrr}
\arrow["G(w_n^a)", rightarrow, swap]{ru}
&
{}
&
{}
&
{F'(G(a))}
\end{tikzcd}
\]
The left/right squares commute since $v,w$ are approximately isometric.
The top/bottom squares commute by the definition of $u_{G,F}$.
The middle square commutes because $G$ is centrally trivial.
Let $\psi\in\operatorname{End}(\cC)(G\Rightarrow G')$.
We shall show $u^a_{G',F}\cdot \psi^{F(a)} = F(\psi^a)\cdot u^a_{G,F}$.
It suffices to prove the following diagram approximately commutes.
\[
\begin{tikzcd}
{G(F(a))}
\arrow["u^a_{G,F}", rightarrow]{rrr}
\arrow["\psi^{F(a)}", rightarrow, swap]{ddd}
\arrow["G(v_n^a)", rightarrow]{rd}
&
{}
&
{}
&
{F(G(a))}
\arrow["\psi^{G(a)}", rightarrow]{ddd}
\\
{}
&
{G(H\rhd a)}
\arrow["\sigma^a_{G,H}", rightarrow]{r}
\arrow["\psi^{H\rhd a}", rightarrow, swap]{d}
&
{H\rhd G(a)}
\arrow["1_H\rhd \psi^a", rightarrow]{d}
\arrow["(v_n^{G(a)})^\dag", rightarrow]{ru}
&
{}
\\
{}
&
{G'(H\rhd a)}
\arrow["\sigma^a_{G',H}", rightarrow, swap]{r}
&
{H\rhd G'(a)}
\arrow["(v_n^{G'(a)})^\dag", rightarrow, swap]{rd}
&
{}
\\
{G'(F(a))}
\arrow["u^a_{G',F}", rightarrow, swap]{rrr}
\arrow["G'(v_n^a)", rightarrow, swap]{ru}
&
{}
&
{}
&
{F(G'(a))}
\end{tikzcd}
\]
The left square commutes by the naturality of $\psi$.
The right square commutes because $F$ is approximately inner.
The top/bottom squares commute by the definition of $u_{G,F}$.
The middle square commutes by the naturality of $\sigma$.
\end{proof}
\begin{prop}
\label{braid4}
For $F\in {\End}_{\ai}(\cC)$ and $G,G'\in {\End}_{\ct}(\cC)$,
$u_{G,F}$ satisfies condition (2) of Definition \ref{Defn:centralizing}.
\end{prop}
\begin{proof}
It suffices to show that for an approximating sequence $(v,H)$ for $F$, the following diagram approximately commutes.
\[
\begin{tikzcd}[column sep=5em]
GG'F(a)
\arrow[dddddd, swap, bend right=55, "G(u^a_{G',F})"']
\arrow[rrrddd, bend left=25, "u^a_{G\circ G',F}"]
\arrow[dd, "G(G'(v_n^a))"]
\arrow[rdd, "G(G'(v_n^a))"]
& & & \\
& & & & \\
GG'(H\rhd a)
\arrow[r, swap, "G(G'(v_n^a(v_n^a)^\dag))"]
\arrow[dd, "G(\sigma^a_{G',H})"]
& GG'(H\rhd a)
\arrow[rd, "\sigma^a_{G\circ G',H}"]
\arrow[dd, swap, "G(\sigma^a_{G',H})"]
& & \\
& &
H\rhd GG'(a)
\arrow[r, "(v_n^{G(G'(a))})^\dag"]
& FGG'(a)
\\
G(H\rhd G'(a))
\arrow[dd, "G(v_n^{G'(a)})^\dag"]
\arrow[r, "G(v_n^{G'(a)}(v_n^{G'(a)})^\dag)"]
& G(H\rhd G'(a))
\arrow[ru, swap, "\sigma^{G'(a)}_{G,H}"]
& & \\
& & & & \\
GFG'(a)
\arrow[ruu, swap, "G(v_n^{G'(a)})"]
\arrow[rrruuu, bend right=25, swap, "u^{G'(a)}_{G,F}"]
& & &
\end{tikzcd}
\]
For any choice of approximating sequence, the outer three cells approximately commute by the definition of $u$ from Theorem \ref{ThmDef:uGFa}.
The upper triangle approximately commutes as $v$ is approximately isometric.
The middle triangle approximately commutes by Proposition \ref{prop:sigmabraid}(2).
The lower triangle is trivial.
Finally, the remaining square approximately commutes because $vv^\dag:H\rhd -\Rightarrow H\rhd -$ is approximately natural and $G'$ is centrally trivial (see also Proposition \ref{prop:uGFUnitary}).
\end{proof}
\begin{prop}
\label{braid3}
For $F,F'\in {\End}_{\ai}(\cC)$ and $G\in {\End}_{\ct}(\cC)$,
$u_{G,F}$ satisfies condition (3) of Definition \ref{Defn:centralizing}.
\end{prop}
\begin{proof}
We must prove for each fixed $a\in \cC$, $F(u^a_{G,F'})u_{G,F}^{F'(a)}=u^a_{G,F\circ F'}$.
To do so, we carefully choose approximating sequences $(v,H)$ and $(w,K)$ for $F$ and $F'$ respectively such that the following diagram approximately commutes.
\[
\begin{tikzcd}
GFF'(a)
\arrow{dd}{G(v_n^{F'(a)})}
\arrow[ddr, "G(v_n^{F'(a)})"]
\arrow[dddddddddddd, swap, bend right=50, "u^{a}_{G, F\circ F'}"']
\arrow[ddddddrrr, bend left=35, "u^{F'(a)}_{G,F}"]
& & & \\
& & & \\
G(H\rhd F'(a))
\arrow[dd, swap, "G(1_{H}\rhd w_n^{a})"]
& G(H\rhd F'(a))
\arrow{ddl}{G(1_{H}\rhd w_n^{a})}
\arrow[ddr, swap, "\sigma^{F'(a)}_{G,H}"]
& & \\
& & & \\
G(H\rhd(K\rhd a))
\arrow{dddd}{\sigma^{a}_{G, H\otimes K}}
\arrow{ddr}{\sigma^{K\rhd a}_{G,H}}
& & H\rhd GF'(a)
\arrow{ddl}[swap]{1_{H}\rhd G(w_n^{a})}
\arrow[swap]{ddr}{(v_n^{GF'(a)})^{\dag}}
& \\
& & & \\
& H\rhd G(K\rhd a)
\arrow{ddl}{1_{H}\rhd \sigma^{a}_{G,K}}
\arrow[swap]{ddr}{(v_n^{G(K\rhd a)})^{\dag}} & & FGF'(a)
\arrow[swap]{ddl}{FG(w_n^{a})}
\arrow[ddddddlll, bend left=35, "F(u^{a}_{G,F'})"]\\
& & & \\
H\rhd K\rhd G(a)
\arrow[swap]{dd}{1_{H}\rhd (w_n^{G(a)})^{\dag}}
\arrow{ddr}{(v_n^{K\rhd G(a)})^{\dag}}
& & FG(K\rhd a)
\arrow[swap]{ddl}{F(\sigma^{a}_{G,K})} & \\
& & & \\
H\rhd F'G(a)
\arrow{dd}{(v_n^{F'G(a)})^{\dag} }
& F(K\rhd G(a))
\arrow[ddl, "F(w_n^{G(a)})^{\dag}"]
& & \\
& & & \\
FF'G(a)
& & & \\
\end{tikzcd}
\]
For any choices of approximating sequences, the outer three cells approximately commute by the definition of $u$ from Theorem \ref{ThmDef:uGFa}.
The upper left square is trivial.
The adjacent square to its lower right approximately commutes by the naturality of $\sigma_{G,H}$.
The left middle triangle approximately commutes by Proposition \ref{prop:sigmabraid}(1),
and the square to the lower right of this triangle approximately commutes by approximate naturality of $v$.
This leaves us to consider the two remaining squares
$$
\begin{tikzcd}[column sep=3.4em]
{H\rhd K\rhd G(a)}
\arrow["(v_n^{K\rhd G(a)})^\dag", rightarrow]{r}
\arrow["1\rhd(w_n^{G(a)})^\dag"', rightarrow]{d}
&
{F(K\rhd G(a))}
\arrow["F(w_n^{G(a)})^\dag", rightarrow]{d}
\\
{H\rhd F'G(a)}
\arrow["(v_n^{F'G(a)})^\dag"', rightarrow]{r}
&
{FF'G(a)}
\end{tikzcd}
\quad\text{and}\quad
\begin{tikzcd}[column sep=3.4em]
{H\rhd GF'(a)}
\arrow["(v_n^{GF'(a)})^\dag", rightarrow]{r}
\arrow["1_H\rhd G(w_n^a)"', rightarrow]{d}
&
{FGF'(a)}
\arrow["FG(w_n^a)", rightarrow]{d}
\\
{H\rhd G(K\rhd a)}
\arrow["(v_n^{G(K\rhd a)})^\dag"', rightarrow]{r}
&
{FG(K\rhd a)}
\end{tikzcd}
\,.
$$
These squares may not approximately commute for an arbitrary choice of approximating sequences, but we can get around this by replacing $v$ with a subsequence, as any approximating sequence for $F$ can be used.
Indeed, for $b,c\in \cC$, let $d_{b\to c}$ denote a metric inducing the $\tau$-topology on bounded subsets of $\cC(b\to c)$.
Since $(w_n^{G(a)})^\dag$ and $G(w_n^a)$ are morphisms in $\cC$ for each fixed $n$, using approximate naturality of $v^\dag$,
we may inductively choose $1\leq k_{n-1} < k_n$ so that both
\begin{align*}
d_{H\rhd K\rhd G(a)\to FF'G(a)}\left(F(w_n^{G(a)})^\dag\cdot (v_{k_n}^{K\rhd G(a)^\dag}),\ (v_{k_n}^{F'G(a)^\dag})\cdot (1_H\rhd (w_n^{G(a)})^\dag) \right)
&< \frac{1}{n}
\qquad\qquad\text{and}
\\
d_{H\rhd GF'(a) \to FG(K\rhd a)}\left(FG(w_n^a)\cdot (v_{k_n}^{GF'(a)^\dag}),\ (v_{k_n}^{G(K\rhd a)^\dag})\cdot (1_H\rhd G(w_n^a) \right)
&< \frac{1}{n}
\end{align*}
simultaneously.
Thus replacing $(v_n^c)$ with $(v_{k_n}^c)$ for all $c\in\cC$,
the previous arguments still hold, as $((v_{k_n}^c)_c,H)$ is still an approximating sequence for $F$, and
these two squares approximately commute for our fixed $a\in \cC$.
Since we only need to verify condition (2) of Definition \ref{Defn:centralizing} for one $a\in\cC$ at a time, the result follows.
\end{proof}
\begin{defn}
We define the category ${\End}_{\loc}(\cC)$ of \textit{local endofunctors}, to be
the full
$\rm W^*$-monoidal subcategory of $\operatorname{End}(\cC)$ whose objects are normal $\dag$-endofunctors which are both approximately inner and centrally trivial.
By Propositions \ref{prop:ctEnd(C)FullReplete} and \ref{prop:aiRepleteComplete}, ${\End}_{\loc}(\cC)$ is replete and unitarily Cauchy complete.
The family of unitary natural transformations $u_{G,F}: G\circ F\Rightarrow F\circ G$ equips ${\End}_{\loc}(\cC)$ with the structure of a \textit{braided $\rm W^*$-tensor category}.
\end{defn}
\section{\texorpdfstring{$\tilde{\chi}(M)$}{chi(M)} for finite von Neumann algebras via bimodules}
\label{sec:Translation}
In this section, we give the main application of the construction of the last section to give a definition of $\tilde{\chi}(M)$ for a $\rm II_1$-factor $M$ in terms of bimodules.
Given a W*-tensor category $\cC$, the \textit{dualizable part}, denoted $\cC_{\rm dualizable}$, is the full tensor subcategory whose objects have two sided duals.
Observe that if $\operatorname{End}_\cC(1_\cC)$ is finite dimensional, then $\cC_{\rm dualizable}$ is a rigid $\rm C^*/W^*$ tensor category.
If moreover $\cC_{\rm dualizable}$ is semisimple (equivalently orthogonal projection complete), it is called a \emph{unitary multitensor category} \cite{MR4133163}; it is called a \emph{unitary tensor category} if $\operatorname{End}_\cC(1_\cC)$ is one dimensional.
\begin{defn}
Given a von Neumann algebra $M$, we denote the braided unitary tensor category $\tilde{\chi}(M):={\End}_{\loc}(\mathsf{Mod_{fgp}}(M))_{\rm dualizable}$.
\end{defn}
Identifying $\operatorname{End}(\mathsf{Mod_{fgp}}(M))\simeq\mathsf{Bim_{fgp}}(M)^{\rm mp}$ as in Remark \ref{rem:OppositeIssue}, we call a bimodule $X\in \mathsf{Bim_{fgp}}(M)$ approximately inner (respectively centrally trivial) if the functor $-\boxtimes_M X$ is approximately inner (respectively centrally trivial).
We see the underlying unitary tensor category of $\tilde{\chi}(M)$ agrees with the definition of $\tilde{\chi}(M)$ from \cite[Rem.~2.7]{MR1317367}.
We may thus think of $\tilde{\chi}(M)$ as the dualizable approximately inner and centrally trivial bimodules of $M$ whose conjugate bimodule is also approximately inner and centrally trivial.
Since ${\End}_{\loc}(\mathsf{Mod_{fgp}}(M))$ is braided, we get a monoidal equivalence from $\tilde{\chi}(M)$ to its monoidal opposite $\tilde{\chi}(M)^{\rm mp}$, which allows us to bypass this opposite issue.
We address this in detail in Remark \ref{rem:OppositeWithBraiding} below and the discussion thereafter.
Note the dualizable objects in $\mathsf{Bim_{fgp}}(M)$ are precisely the bifinite Hilbert bimodules of $M$, and the dual object is the conjugate bimodule.
It is easy to see that $H$ is centrally trivial if and only if $\overline{H}$ is centrally trivial, but for approximately inner, any such relationship is not obvious.
Thus it may be possible for a centrally trivial bifinite bimodule to be approximately inner, but its conjugate bimodule may not be approximately inner.
We do not have an example, but we cannot rule this out at this time.
\subsection{Module and bimodule realization}
\label{sec:ModuleRealization}
In order to translate the results of \S\ref{sec:W*ApproxNatTrans}, we use the graphical calculus for $\mathsf{Mod_{fgp}}(M)$ as a right $\mathsf{Bim_{fgp}}(M)$-module $\rm W^*$-category.
One way to do this is to use the realization graphical calculus from \cite{2105.12010} based on \cite[\S4.1]{MR3221289}.
We only introduce the part of the graphical calculus that we need in this section, and we introduce the rest of the graphical calculus for Q-system realization in \S\ref{sect:Q-systemRealization} below.
In the 2D graphical calculus for $\mathsf{W^*Alg}$, von Neumann algebras are denoted by shaded regions, bimodules are denoted by 1D strands, and intertwiners are denoted by 0D coupons.
For the rest of this section, let $M$ be a $\rm II_1$ factor.
Of particular importance is the right $M$-module $M_M$.
The missing label on the left hand side is inferred to be $\bbC$, which is always represented by the empty shading.
That is, we identify $\mathsf{Mod_{fgp}}(M)=\mathsf{W^*Alg}(\bbC \to M)$ in the 2D graphical calculus.
We denote ${}_\bbC M_M$ by a dashed line which is shaded by $M$ on the right hand side.
$$
\tikzmath{\filldraw[\rColor, rounded corners=5, very thin, baseline=1cm] (0,0) rectangle (.6,.6);}=M
\qquad
\qquad
\tikzmath{
\filldraw[white, rounded corners=5, very thin, baseline=1cm] (0,0) rectangle (.6,.6);
\draw[dotted, rounded corners=5pt] (0,0) rectangle (.6,.6);
}=\bbC
\qquad
\qquad
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (0,0) rectangle (.6,.6);
\filldraw[\rColor] (.3,0) rectangle (.6,.6);
\end{scope}
\draw[dashed] (.3,0) -- (.3,.6);
\draw[dotted, rounded corners=5pt] (.3,0) -- (0,0) -- (0,.6) -- (.3,.6);
}={}_\bbC M_M.
$$
A bounded, adjointable intertwiner $f: Y_M \to Z_M$ is denoted graphically by
$$
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.3,-.7) rectangle (.6,.7);
\filldraw[\rColor] (0,-.7) -- (0,.7) -- (.7,.7) -- (.7,-.7);
\end{scope}
\draw[blue, thick] (0,-.7) -- (0,0);
\draw[violet, thick] (0,0) -- (0,.7);
\roundNbox{unshaded}{(0,0)}{.3}{0}{0}{\scriptsize{$f$}};
}
:Y_M \to Z_M;
\qquad\qquad\qquad
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (0,0) rectangle (.6,.6);
\filldraw[\rColor] (.3,0) rectangle (.6,.6);
\end{scope}
\draw[blue, thick] (.3,0) -- (.3,.6);
\draw[dotted, rounded corners=5pt] (.3,0) -- (0,0) -- (0,.6) -- (.3,.6);
}=Y_M
\qquad\text{and}\qquad
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (0,0) rectangle (.6,.6);
\filldraw[\rColor] (.3,0) rectangle (.6,.6);
\end{scope}
\draw[violet, thick] (.3,0) -- (.3,.6);
\draw[dotted, rounded corners=5pt] (.3,0) -- (0,0) -- (0,.6) -- (.3,.6);
}=Z_M.
$$
\begin{construction}[{\cite[Const.~4.1]{2105.12010}}]
Given $Y_M\in \mathsf{Mod_{fgp}}(M)$, the map $x\mapsto L_y$ where $L_y(m) := ym$ gives a canonical isomorphism $Y_M \cong |Y|_M:= \operatorname{Hom}(M_M \to Y_M)$
such that $\langle x|y\rangle_M = L_x^\dag L_y$.
In \S\ref{sec:CTandAIbimods} and \S\ref{sec:LocalExtension} below, we make heavy use of this identification.
\[
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.3,-.7) rectangle (.6,.7);
\filldraw[\rColor] (0,-.7) -- (0,.7) -- (.7,.7) -- (.7,-.7);
\end{scope}
\draw[dashed] (0,-.7) -- (0,0);
\draw[blue,thick] (0,0) -- (0,.7);
\roundNbox{unshaded}{(0,0)}{.3}{0}{0}{\scriptsize{$y$}};
}
\in
|Y|:= \operatorname{Hom}(M_M \to Y_M).
\]
The right $M$-action is given by identifying $M=\operatorname{End}(M_M)$ and stacking coupons:
\[
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.3,-1.7) rectangle (.6,.7);
\filldraw[\rColor] (0,-1.7) -- (0,.7) -- (.7,.7) -- (.7,-1.7);
\end{scope}
\draw[dashed] (0,-1.7) -- (0,0);
\draw[blue,thick] (0,0) -- (0,.7);
\roundNbox{unshaded}{(0,0)}{.3}{0}{0}{\scriptsize{$y$}};
\roundNbox{unshaded}{(0,-1)}{.3}{0}{0}{\scriptsize{$a$}};
}\,
=
y\lhd a.
\]
The condition that $\{c_j\}$ is a $Y_M$-basis can be written graphically as
\[
\sum_j
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.15,-1.2) rectangle (.6,1.2);
\filldraw[\rColor] (0,-1.2) rectangle (.6,1.2);
\end{scope}
\draw[dashed] (0,-.5) -- (0,.5);
\draw[blue,thick] (0,.5) -- (0,1.2);
\draw[blue,thick] (0,-.5) -- (0,-1.2);
\roundNbox{unshaded}{(0,.5)}{.3}{0}{0}{\scriptsize{$c_j$}};
\roundNbox{unshaded}{(0,-.5)}{.3}{0}{0}{\scriptsize{$c_j^\dag$}};
}
=
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.15,-.5) rectangle (.45,.5);
\filldraw[\rColor] (0,-.5) rectangle (.45,.5);
\end{scope}
\draw[blue,thick] (0,-.5) -- (0,.5);
}\,.
\]
If $X\in \mathsf{Bim_{fgp}}(M)$, then we define the realization $|X|$ slightly differently:
$$
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.3,-.7) rectangle (.6,.7);
\filldraw[\rColor] (0,-.7) -- (0,0) -- (-.2,0) -- (-.2,.7) -- (.7,.7) -- (.7,-.7);
\end{scope}
\draw[dashed] (0,-.7) -- (0,0);
\draw[dashed] (-.2,0) -- (-.2,.7);
\draw[\XColor,thick] (.2,0) -- (.2,.7);
\roundNbox{unshaded}{(0,0)}{.3}{.05}{.05}{\scriptsize{$x$}};
}
\in
|X|:= \operatorname{Hom}(M_M \to M\boxtimes_M X_M).
$$
While this definition is canonically isomorphic to the previous definition via the canonical unitor $M\boxtimes_M X \cong X$, this definition offers the advantage of depicting both the left and the right $M$-actions graphically by
\[
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.4,-.7) rectangle (.75,1.7);
\filldraw[\rColor] (0,-.7) -- (0,0) -- (-.25,0) -- (-.25,1.7) -- (.75,1.7) -- (.75,-.7);
\end{scope}
\draw[dashed] (-.25,0) -- (-.25,1.7);
\draw[\XColor,thick] (.25,0) -- (.25,1.7);
\draw[dashed] (0,-.7) -- (0,0);
\roundNbox{unshaded}{(-.25,1)}{.3}{0}{0}{\scriptsize{$a$}};
\roundNbox{unshaded}{(0,0)}{.3}{.15}{.15}{\scriptsize{$x$}};
}
\,=
a\rhd x
\qquad\text{and}\qquad
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.4,-1.7) rectangle (.7,.7);
\filldraw[\rColor] (0,-1.7) -- (0,0) -- (-.2,0) -- (-.2,.7) -- (.75,.7) -- (.75,-1.7);
\end{scope}
\draw[dashed] (-.2,0) -- (-.2,.7);
\draw[\XColor,thick] (.2,0) -- (.2,.7);
\draw[dashed] (0,-1.7) -- (0,0);
\roundNbox{unshaded}{(0,0)}{.3}{.05}{.05}{\scriptsize{$x$}};
\roundNbox{unshaded}{(0,-1)}{.3}{0}{0}{\scriptsize{$b$}};
}
\,=
x\lhd b.
\]
\end{construction}
\subsection{Centrally trivial and approximately inner bimodules}
\label{sec:CTandAIbimods}
In this section we clarify the equivalence between our definitions of approximately inner and centrally trivial given in terms of endofunctors, and Popa's original definitions as translated to bimodules, which are much more natural from the point of view of a single von Neumann algebra \cite[Rem.~2.7]{MR1317367}.
For this section, let $M$ be a finite separable von Neumann algebra with faithful normal trace $\operatorname{tr}_M$.
\begin{nota}
For norm bounded sequences $(f_n)_n,(g_n)_n\subseteq \operatorname{Hom}(X_M\to Y_M)$,
we write $f_n\approx g_n$ if $\lim_n\|f_n-g_n\|_2 = 0$.
For $f\in \operatorname{Hom}(X_M\to Y_M)$, we write $f\approx f_n$ if $\lim_n \|f-f_n\|_2 = 0$.
As a consequence, $f\approx g$ if and only if $f=g$.
We remark that since composition is jointly $\tau$-continuous on norm-bounded subsets of hom spaces, we may use $f_n\approx g_n$ as a \emph{local relation} amongst morphisms in $\mathsf{Mod_{fgp}}(M)$.
If we precompose with an appropriate $h$, we still have $f_n \circ h \approx g_n\circ h$, and similarly for composing on the other side, or both sides simultaneously.
\end{nota}
\begin{prop}
\label{prop:BimodualCT&CS}
$X$ is centrally trivial over $M$ if and only if for all central sequences $(a_n)_n\subseteq M$ and for all $x\in X$, $\|a_nx-xa_n\|_2\to 0$, i.e.,
\[
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.4,-.7) rectangle (.75,1.7);
\filldraw[\rColor] (0,-.7) -- (0,0) -- (-.25,0) -- (-.25,1.7) -- (.75,1.7) -- (.75,-.7);
\end{scope}
\draw[dashed] (-.25,0) -- (-.25,1.7);
\draw[\XColor,thick] (.25,0) -- (.25,1.7);
\draw[dashed] (0,-.7) -- (0,0);
\roundNbox{unshaded}{(-.25,1)}{.3}{0}{0}{\scriptsize{$a_n$}};
\roundNbox{unshaded}{(0,0)}{.3}{.15}{.15}{\scriptsize{$x$}};
}
\approx
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.4,-1.7) rectangle (.7,.7);
\filldraw[\rColor] (0,-1.7) -- (0,0) -- (-.2,0) -- (-.2,.7) -- (.75,.7) -- (.75,-1.7);
\end{scope}
\draw[dashed] (-.2,0) -- (-.2,.7);
\draw[\XColor,thick] (.2,0) -- (.2,.7);
\draw[dashed] (0,-1.7) -- (0,0);
\roundNbox{unshaded}{(0,0)}{.3}{.05}{.05}{\scriptsize{$x$}};
\roundNbox{unshaded}{(0,-1)}{.3}{0}{0}{\scriptsize{$a_n$}};
}
\qquad\qquad
\forall\, x\in X,
\qquad
\forall\text{ central sequences }(a_n)\subseteq M.
\]
\end{prop}
\begin{proof}
Recall a central sequence of $\mathsf{Mod_{fgp}}(M)$ is a natural transformation of the identity functor.
By Example \ref{ex:MCentralSequencesAgree}, equivalence classes of these central sequences agree with the usual equivalence classes of central sequences of $M$.
The result now follows directly from Proposition \ref{Prop:CentralTrivial&CentralSq}, which translates into the displayed condition in the statement of the proposition.
\end{proof}
Now by definition, the functor $-\boxtimes_M X$ is approximately inner if there exists an approximately natural isometry $v_n^Y:Y\boxtimes_M X\to Y\otimes H$, i.e., we have
\[
v_n^Y=
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.15,-.7) rectangle (.9,.7);
\filldraw[\rColor] (0,-.7) -- (0,.7) -- (1,.7) -- (1,-.7);
\end{scope}
\draw[blue,thick] (0,-.7) node[below]{$\scriptstyle Y$} -- (0,.7);
\draw[\HColor,thick] (.4,0) -- (.4,.7) node[above,black]{$\scriptstyle H$};
\draw[\XColor,thick] (.4,0) -- (.4,-.7) node[below]{$\scriptstyle X$};
\roundNbox{unshaded}{(.2,0)}{.3}{.1}{.1}{\scriptsize{$v_n^{Y}$}};
}
\qquad\text{such that}\qquad
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.15,-1.7) rectangle (1,.7);
\filldraw[\rColor] (0,-1.7) -- (0,.7) -- (1,.7) -- (1,-1.7);
\end{scope}
\draw[blue,thick] (0,-.7) -- (0,.7);
\draw[\XColor,thick] (.5,-1.7) -- (.5,0);
\draw[violet,thick] (0,-1.7) -- (0,-.7);
\draw[\HColor,thick] (.5,0) -- (.5,.7);
\roundNbox{unshaded}{(.2,0)}{.3}{.2}{.2}{\scriptsize{$v_n^{Y}$}};
\roundNbox{unshaded}{(0,-1)}{.3}{0}{0}{\scriptsize{$f$}};
}
\approx
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.15,-.7) rectangle (1,1.7);
\filldraw[\rColor] (0,-.7) -- (0,1.7) -- (1,1.7) -- (1,-.7);
\end{scope}
\draw[blue,thick] (0,1) -- (0,1.7);
\draw[\XColor,thick] (.5,-.7) -- (.5,0);
\draw[violet,thick] (0,-.7) -- (0,1);
\draw[\HColor,thick] (.5,0) -- (.5,1.7);
\roundNbox{unshaded}{(.2,0)}{.3}{.2}{.2}{\scriptsize{$v_n^{Z}$}};
\roundNbox{unshaded}{(0,1)}{.3}{0}{0}{\scriptsize{$f$}};
}
\qquad\text{and}\qquad
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.15,-1.7) rectangle (.9,.7);
\filldraw[\rColor] (0,-1.7) -- (0,.7) -- (1,.7) -- (1,-1.7);
\end{scope}
\draw[blue,thick] (0,-1.7) -- (0,.7);
\draw[\XColor,thick] (.4,-1.7) -- (.4,-1);
\draw[\HColor,thick] (.4,-1) -- (.4,0);
\draw[\XColor,thick] (.4,0) -- (.4,.7);
\roundNbox{unshaded}{(.2,0)}{.3}{.1}{.1}{\scriptsize{$(v_n^{Y})^\dag$}};
\roundNbox{unshaded}{(.2,-1)}{.3}{.1}{.1}{\scriptsize{$v_n^{Y}$}};
}
\approx
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.15,-1.7) rectangle (.9,.7);
\filldraw[\rColor] (0,-1.7) -- (0,.7) -- (1,.7) -- (1,-1.7);
\end{scope}
\draw[blue,thick] (0,-1.7) -- (0,.7);
\draw[\XColor,thick] (.4,-1.7) -- (.4,.7);
}
\]
for all intertwiners $f\in \operatorname{Hom}(Z_M\to Y_M)$.
\begin{defn}
\label{defn:ApproximatePPBasis}
For $Y\in\mathsf{Mod_{fgp}}(M)$, an \emph{approximate $Y_M$-basis} is a sequence $\{b_i^{(n)}\}_{i=1}^{m(n)}\subseteq Y$ such that
$\sup_n m(n) <\infty$,
$\sup_{i,n} \|\langle b_i^{(n)}| b_i^{(n)}\rangle_M^Y\| <\infty$,
and
$$
\lim_{n\to\infty} \left\|x-\sum_{i=1}^m b_i^{(n)}\langle b_i^{(n)}|x\rangle_M^Y \right\|_2=0
\qquad\qquad
\forall\,x\in Y.
$$
For $X\in \mathsf{Bim_{fgp}}(X)$, an \emph{approximately inner} $X_M$-basis is an approximate $X_M$-basis such that
\begin{equation}
\label{eq:AICommutativity}
\left\|ab_i^{(n)}-b_i^{(n)}a\right\|_2\to 0
\qquad\qquad
\forall\,a\in M.
\end{equation}
\end{defn}
\begin{prop}
\label{prop:AI&APPB}
A bimodule $X$ is approximately inner over $M$ if and only if there exists an approximately inner $X_M$-basis.
\end{prop}
\begin{proof}
Suppose $X$ is approximately inner.
By Definition \ref{Defn:ApproximatelyInner}, there exists a finite dimensional Hilbert space $H$ and an approximate natural isometry $v=(v_n):-\boxtimes_M X\to H\rhd-$.
Let $\{e_i\}$ be an orthonormal basis of $H$, and define $b_i^{(n)}$ as follows.
\[
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.3,-.7) rectangle (.6,.7);
\filldraw[\rColor] (0,-.7) -- (0,0) -- (-.2,0) -- (-.2,.7) -- (.7,.7) -- (.7,-.7);
\end{scope}
\draw[dashed] (0,-.7) -- (0,0);
\draw[dashed] (-.2,0) -- (-.2,.7);
\draw[\XColor,thick] (.2,0) -- (.2,.7);
\roundNbox{unshaded}{(0,0)}{.3}{.05}{.05}{\scriptsize{$b_i^{(n)}$}};
}
: =
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.5,-.7) rectangle (.8,.7);
\filldraw[\rColor] (-.25,-.7) -- (-.25,.7) -- (1,.7) -- (1,-.7);
\end{scope}
\draw[dashed] (-.25,-.7) -- (-.25,.7);
\draw[\XColor,thick] (.25,0) -- (.25,.7);
\draw[\HColor,thick] (.25,0) -- (.25,-.55);
\filldraw[\HColor] (.25,-.55) circle (.05cm);
\roundNbox{unshaded}{(0,0)}{.3}{.15}{.15}{\scriptsize{$(v_n^{M})^\dag$}};
\node at (.5,-.55) {\scriptsize{$e_i$}};
}
\,.
\]
Observe that for all $x\in X$, we have
\[
\sum_{i} b_i^{(n)}\langle b_i^{(n)}|x\rangle_M^X
=
\sum_i
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.5,-2.2) rectangle (.8,1.2);
\filldraw[\rColor] (0,-2.2) -- (0,-1.5) -- (-.2,-1.5) -- (-.2,-.5) -- (0,-.5) -- (0,.5) -- (-.2,.5) -- (-.2,1.2) -- (.8,1.2) -- (.8,-2.2);
\end{scope}
\draw[dashed] (-.2,-1.5) -- (-.2,-.5);
\draw[dashed] (-.2,.5) -- (-.2,1.2);
\draw[dashed] (0,-.5) -- (0,.5);
\draw[dashed] (0,-1.5) -- (0,-2.2);
\draw[\XColor,thick] (.2,-1.5) -- (.2,-.5);
\draw[\XColor,thick] (.2,.5) -- (.2,1.2);
\roundNbox{unshaded}{(0,.5)}{.3}{.2}{.2}{\scriptsize{$b_i^{(n)}$}};
\roundNbox{unshaded}{(0,-.5)}{.3}{.2}{.2}{\scriptsize{$(b_i^{(n)})^\dag$}};
\roundNbox{unshaded}{(0,-1.5)}{.3}{.2}{.2}{\scriptsize{$x$}};
}
=
\sum_i
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.5,-2.4) rectangle (1,1.4);
\filldraw[\rColor] (0,-2.4) -- (0,-1.5) -- (-.3,-1.5) -- (-.3,1.4) -- (1,1.4) -- (1,-2.4);
\end{scope}
\draw[dashed] (-.3,-1.7) -- (-.3,1.4);
\draw[dashed] (0,-2.4) -- (0,-1.7);
\draw[\XColor,thick] (.3,-1.7) -- (.3,-.7);
\draw[\XColor,thick] (.3,.7) -- (.3,1.4);
\draw[\HColor,thick] (.3,.15) -- (.3,.4);
\draw[\HColor,thick] (.3,-.15) -- (.3,-.4);
\filldraw[\HColor] (.3,.15) circle (.05cm);
\filldraw[\HColor] (.3,-.15) circle (.05cm);
\roundNbox{unshaded}{(0,.7)}{.3}{.2}{.2}{\scriptsize{$(v_n^{M})^\dag$}};
\roundNbox{unshaded}{(0,-.7)}{.3}{.2}{.2}{\scriptsize{$v_n^{M}$}};
\roundNbox{unshaded}{(0,-1.7)}{.3}{.2}{.2}{\scriptsize{$x$}};
\node at (.55,.15) {\scriptsize{$e_i$}};
\node at (.55,-.15) {\scriptsize{$e_i^*$}};
}
=
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.5,-2.2) rectangle (1,1.2);
\filldraw[\rColor] (0,-2.2) -- (0,-1.5) -- (-.3,-1.5) -- (-.3,1.2) -- (1,1.2) -- (1,-2.2);
\end{scope}
\draw[dashed] (-.3,-1.5) -- (-.3,1.2);
\draw[dashed] (0,-2.2) -- (0,-1.5);
\draw[\XColor,thick] (.3,-1.5) -- (.3,-.5);
\draw[\XColor,thick] (.3,.5) -- (.3,1.2);
\draw[\HColor,thick] (.3,-.5) -- (.3,.5);
\roundNbox{unshaded}{(0,.5)}{.3}{.2}{.2}{\scriptsize{$(v_n^{M})^\dag$}};
\roundNbox{unshaded}{(0,-.5)}{.3}{.2}{.2}{\scriptsize{$v_n^{M}$}};
\roundNbox{unshaded}{(0,-1.5)}{.3}{.2}{.2}{\scriptsize{$x$}};
}
=
(v_n^{M})^\dag(v_n^{M}(x)).
\]
The condition \eqref{eq:AICommutativity} follows immediately by approximate naturality of $v$ on the $M$-component.
Conversely, starting with an approximately inner $X_M$-basis, we define
\[
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.5,-.7) rectangle (.8,.7);
\filldraw[\rColor] (-.25,-.7) -- (-.25,.7) -- (1,.7) -- (1,-.7);
\end{scope}
\draw[dashed] (-.25,-.7) -- (-.25,.7);
\draw[\HColor,thick] (.25,0) -- (.25,.7);
\draw[\XColor,thick] (.25,0) -- (.25,-.7);
\roundNbox{unshaded}{(0,0)}{.3}{.15}{.15}{\scriptsize{$v_n^{M}$}};
\node at (.25,.9) {\scriptsize{$H$}};
}
:=
\sum_i
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.7,-.7) rectangle (.9,1.1);
\filldraw[\rColor] (-.25,-.7) -- (-.25,1.1) -- (1,1.1) -- (1,-.7);
\end{scope}
\draw[dashed] (-.25,-.7) -- (-.25,1.1);
\draw[\XColor,thick] (.25,0) -- (.25,-.7);
\draw[\HColor,thick] (.25,0) -- (.25,.55);
\draw[\HColor,thick] (.25,.85) -- (.25,1.1);
\filldraw[\HColor] (.25,.55) circle (.05cm);
\filldraw[\HColor] (.25,.85) circle (.05cm);
\roundNbox{unshaded}{(0,0)}{.3}{.15}{.15}{\scriptsize{$v_n^{M}$}};
\node at (.5,.55) {\scriptsize{$e_i^*$}};
\node at (.5,.85) {\scriptsize{$e_i$}};
}
=
\sum_i
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.3,-.7) rectangle (.8,.8);
\filldraw[\rColor] (-.2,-.7) -- (-.2,.8) -- (.8,.8) -- (.8,-.7);
\end{scope}
\draw[dashed] (-.2,-.7) -- (-.2,.8);
\draw[\XColor,thick] (.2,0) -- (.2,-.7);
\draw[\HColor] (.2,.55) -- (.2,.8);
\filldraw[\HColor] (.2,.55) circle (.05cm);
\roundNbox{unshaded}{(0,0)}{.3}{.2}{.2}{\scriptsize{$(b_i^{(n)})^\dag$}};
\node at (.45,.55) {\scriptsize{$e_i$}};
}\,,
\]
and we define each $v_n^Y$ for $Y\in\mathsf{Mod_{fgp}}(M)$ in terms of $v_n^{M}$ and a $Y_M$-basis $\{c_j\}$:
\[
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.15,-.7) rectangle (.9,.7);
\filldraw[\rColor] (0,-.7) -- (0,.7) -- (1,.7) -- (1,-.7);
\end{scope}
\draw[blue,thick] (0,-.7) node[below]{$\scriptstyle Y$} -- (0,.7);
\draw[\HColor,thick] (.4,0) -- (.4,.7);
\draw[\XColor,thick] (.4,0) -- (.4,-.7) node[below]{$\scriptstyle X$};
\roundNbox{unshaded}{(.2,0)}{.3}{.1}{.1}{\scriptsize{$v_n^{Y}$}};
}
:=
\sum_j
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.45,-1.7) rectangle (.9,1.7);
\filldraw[\rColor] (-.3,-1.7) rectangle (.9,1.7);
\end{scope}
\draw[dashed] (-.3,-1) -- (-.3,1);
\draw[\HColor,thick] (.3,0) -- (.3,1.7);
\draw[\XColor,thick] (.3,0) -- (.3,-1.7);
\draw[blue,thick] (-.3,-1.7) -- (-.3,-1);
\draw[blue,thick] (-.3,1) -- (-.3,1.7);
\roundNbox{unshaded}{(0,0)}{.3}{.3}{.3}{\scriptsize{$v_n^{M}$}};
\roundNbox{unshaded}{(-.3,1)}{.3}{0}{0}{\scriptsize{$c_j$}};
\roundNbox{unshaded}{(-.3,-1)}{.3}{0}{0}{\scriptsize{$c_j^\dag$}};
}\,.
\]
We have $v_n^Y$ is norm-bounded as $\sum_j \|c_j\|_2^2<\infty$.
To see the approximate naturality of $\{v_n^Y\}_{Y\in\mathsf{Mod_{fgp}}(M)}$, for $f\in \operatorname{Hom}(Z_M\to Y_M)$,
\[
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.15,-1.7) rectangle (1,.7);
\filldraw[\rColor] (0,-1.7) -- (0,.7) -- (1,.7) -- (1,-1.7);
\end{scope}
\draw[blue,thick] (0,-.7) -- (0,.7) node[above]{$\scriptstyle Y$};
\draw[\XColor,thick] (.5,-1.7) node[below]{$\scriptstyle X$} -- (.5,0);
\draw[violet,thick] (0,-1.7) node[below]{$\scriptstyle Z$} -- (0,-.7);
\draw[\HColor,thick] (.5,0) -- (.5,.7);
\roundNbox{unshaded}{(.2,0)}{.3}{.2}{.2}{\scriptsize{$v_n^{Y}$}};
\roundNbox{unshaded}{(0,-1)}{.3}{0}{0}{\scriptsize{$f$}};
}
=
\sum_j
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.45,-2.7) rectangle (.9,1.7);
\filldraw[\rColor] (-.3,-2.7) rectangle (.9,1.7);
\end{scope}
\draw[dashed] (-.3,-1) -- (-.3,1);
\draw[\HColor,thick] (.3,0) -- (.3,1.7);
\draw[\XColor,thick] (.3,0) -- (.3,-2.7);
\draw[blue,thick] (-.3,-2) -- (-.3,-1);
\draw[blue,thick] (-.3,1) -- (-.3,1.7);
\draw[violet,thick] (-.3,-2.7) -- (-.3,-2);
\roundNbox{unshaded}{(0,0)}{.3}{.3}{.3}{\scriptsize{$v_n^{M}$}};
\roundNbox{unshaded}{(-.3,1)}{.3}{0}{0}{\scriptsize{$c_j$}};
\roundNbox{unshaded}{(-.3,-1)}{.3}{0}{0}{\scriptsize{$c_j^\dag$}};
\roundNbox{unshaded}{(-.3,-2)}{.3}{0}{0}{\scriptsize{$f$}};
}
=
\sum_{j,k}
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.45,-4.7) rectangle (.9,1.7);
\filldraw[\rColor] (-.3,-4.7) rectangle (.9,1.7);
\end{scope}
\draw[dashed] (-.3,-1) -- (-.3,1);
\draw[dashed] (-.3,-4) -- (-.3,-3);
\draw[\HColor,thick] (.3,0) -- (.3,1.7);
\draw[\XColor,thick] (.3,0) -- (.3,-4.7);
\draw[blue,thick] (-.3,-2) -- (-.3,-1);
\draw[blue,thick] (-.3,1) -- (-.3,1.7);
\draw[violet,thick] (-.3,-3) -- (-.3,-2);
\draw[violet,thick] (-.3,-4.7) -- (-.3,-4);
\roundNbox{unshaded}{(0,0)}{.3}{.3}{.3}{\scriptsize{$v_n^{M}$}};
\roundNbox{unshaded}{(-.3,1)}{.3}{0}{0}{\scriptsize{$c_j$}};
\roundNbox{unshaded}{(-.3,-1)}{.3}{0}{0}{\scriptsize{$c_j^\dag$}};
\roundNbox{unshaded}{(-.3,-2)}{.3}{0}{0}{\scriptsize{$f$}};
\roundNbox{unshaded}{(-.3,-3)}{.3}{0}{0}{\scriptsize{$d_k$}};
\roundNbox{unshaded}{(-.3,-4)}{.3}{0}{0}{\scriptsize{$d_k^\dag$}};
}
\approx
\sum_{j,k}
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.45,-1.7) rectangle (.9,4.7);
\filldraw[\rColor] (-.3,-1.7) rectangle (.9,4.7);
\end{scope}
\draw[dashed] (-.3,-1) -- (-.3,1);
\draw[dashed] (-.3,3) -- (-.3,4);
\draw[\HColor,thick] (.3,0) -- (.3,4.7);
\draw[\XColor,thick] (.3,0) -- (.3,-1.7);
\draw[violet,thick] (-.3,-1.7) -- (-.3,-1);
\draw[violet,thick] (-.3,1) -- (-.3,2);
\draw[blue,thick] (-.3,2) -- (-.3,3);
\draw[blue,thick] (-.3,4) -- (-.3,4.7);
\roundNbox{unshaded}{(0,0)}{.3}{.3}{.3}{\scriptsize{$v_n^{M}$}};
\roundNbox{unshaded}{(-.3,1)}{.3}{0}{0}{\scriptsize{$d_k$}};
\roundNbox{unshaded}{(-.3,-1)}{.3}{0}{0}{\scriptsize{$d_k^\dag$}};
\roundNbox{unshaded}{(-.3,2)}{.3}{0}{0}{\scriptsize{$f$}};
\roundNbox{unshaded}{(-.3,3)}{.3}{0}{0}{\scriptsize{$c_j^\dag$}};
\roundNbox{unshaded}{(-.3,4)}{.3}{0}{0}{\scriptsize{$c_j$}};
}
=
\sum_k
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.45,-1.7) rectangle (.9,2.7);
\filldraw[\rColor] (-.3,-1.7) rectangle (.9,2.7);
\end{scope}
\draw[dashed] (-.3,-1) -- (-.3,1);
\draw[\HColor,thick] (.3,0) -- (.3,2.7);
\draw[\XColor,thick] (.3,0) -- (.3,-1.7);
\draw[violet,thick] (-.3,-1.7) -- (-.3,-1);
\draw[violet,thick] (-.3,1) -- (-.3,2);
\draw[blue,thick] (-.3,2) -- (-.3,2.7);
\roundNbox{unshaded}{(0,0)}{.3}{.3}{.3}{\scriptsize{$v_n^{M}$}};
\roundNbox{unshaded}{(-.3,1)}{.3}{0}{0}{\scriptsize{$d_k$}};
\roundNbox{unshaded}{(-.3,-1)}{.3}{0}{0}{\scriptsize{$d_k^\dag$}};
\roundNbox{unshaded}{(-.3,2)}{.3}{0}{0}{\scriptsize{$f$}};
}
=
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.15,-.7) rectangle (1,1.7);
\filldraw[\rColor] (0,-.7) -- (0,1.7) -- (1,1.7) -- (1,-.7);
\end{scope}
\draw[blue,thick] (0,1) -- (0,1.7);
\draw[\XColor,thick] (.5,-.7) -- (.5,0);
\draw[violet,thick] (0,-.7) -- (0,1);
\draw[\HColor,thick] (.5,0) -- (.5,1.7);
\roundNbox{unshaded}{(.2,0)}{.3}{.2}{.2}{\scriptsize{$v_n^{Z}$}};
\roundNbox{unshaded}{(0,1)}{.3}{0}{0}{\scriptsize{$f$}};
}\,.
\]
We see $v_n^Y$ is independent of the choice of $Y_M$-basis by taking $f=\operatorname{id}_Y$ above.
\end{proof}
In \cite[Def.~1.1]{MR1317367}, Popa gave a definition of approximate innerness for a finite index $\rm II_1$ subfactor $N\subseteq M$.
We can also view ${}_NM_N \in \mathsf{Bim_{fgp}}(N)$ with $\langle x|y\rangle_N := E_N(x^*y)$.
We will show in Proposition \ref{PopaEq} below that
$N\subseteq M$ is approximately inner in the sense of \cite[Def.~1.1]{MR1317367} if and only if ${}_NM_N$ is approximately inner.
We quickly recall the notion of ultraproduct for $\rm II_1$ (sub)factors following \cite[\S5.4]{ClaireSorinII_1}.
Let $\omega$ be a non-principal unltrafilter on $\bbN$.
For a $\rm{II}_{1}$ factor $N$, define $N^{\omega}=\ell^{\infty}(\bbN, M)/\cI$, where $\cI$ is the ideal of sequences which converge to $0$ in $\|\cdot \|_{2}$ along $\omega$.
Then $N^{\omega}$ is a $\rm II_1$ factor, with trace given by taking the limit along $\omega$.
Now consider a finite index $\rm{II}_{1}$ subfactor $N\subseteq M$.
Then $N^{\omega}\subseteq M^{\omega}$ is another $\rm II_1$ subfactor with the same Jones index and trace preserving condition expectation extending $E:M\rightarrow N$.
Using this expectation, we can view $M^{\omega}$ as an $N^{\omega}$ bimodule.
We can also consider the inclusion $N_{\omega}:=N^{\prime}\cap N^{\omega}\subseteq N^{\prime}\cap M^{\omega}=M_{\omega}$, and $E$ restricts to the trace preserving conditional expectation $E: M_{\omega}\rightarrow N_{\omega}$.
We recall the following definition due to Popa.
\begin{defn}[{\cite[Def.~1.1 and Prop.~1.2]{MR1317367}}]
A finite index $\rm II_1$ subfactor $N\subseteq M$ is called \textit{approximately inner} if the inclusion $N_{\omega}\subseteq_{E} M_{\omega}$ is $[M:N]^{-1}$-Markov, i.e., there is a (finite) Pimsner-Popa basis $\{b\}$ for $M_\omega$ over $N_\omega$ which satisfies $\sum_b bb^* \in [1,\infty)$.
\end{defn}
\begin{prop}
\label{PopaEq}
A finite index $\rm II_1$ subfactor $N\subseteq M$ is approximately inner if and only if ${}_{N} M_{N} \in \mathsf{Bim_{fgp}}(N)$ is approximately inner.
\end{prop}
\begin{proof}
It is easy to see that for a finite index subfactor $N\subseteq M$, if ${}_NM_N\in \mathsf{Bim_{fgp}}(N)$ is approximately inner, then an approximate $M_N$-basis
gives an honest Pimsner-Popa basis for $M^\omega$ over $N^\omega$, and the extra commutativity condition \eqref{eq:AICommutativity} means this Pimsner-Popa basis lies in $M_\omega=N'\cap M^\omega$.
Thus we have a Pimsner-Popa basis for $M_\omega$ over $N_\omega$ which lifts to a Pimsner-Popa basis for $M^\omega$ over $N^\omega$, which implies $N_\omega \subseteq_E M_\omega$ is $[M:N]^{-1}$-Markov.
Conversely, suppose $N\subseteq M$ is approximately inner.
Let $\{b^{(n)}_{i}\}^{k}_{i=1}$ be a representative of the Pimsner-Popa basis for $N_{\omega}\subseteq_{E} M_{\omega}$ which lifts to a Pimsner-Popa basis for $N^{\omega}\subseteq_{E} M^{\omega}$, which exists by \cite[Proposition 2.2 (3)]{MR1339767}.
Note that for every $x\in M$ and $y\in N$,
$$
\left\|x-\sum b^{(n)}_{i}E((b^{(n)}_{i})^{*}x)\right\|_{2}\longrightarrow_{\omega}0
\qquad\text{and}\qquad
\left\|yb^{(n)}_{i}-b^{(n)}_{i}y\right\|_{2}\longrightarrow_{\omega} 0.
$$
Now let $F\subseteq N$ and $G\subseteq M$ be countable $\sigma$-strong* dense subsets, and write $F=\bigcup_{l} F_{l}$ and $G=\bigcup_{l} G_{l}$ where the $F_{l}$ and $G_{l}$ increasing sequences of finite subsets.
For each $l$, we have
$$
\sum_{x\in G_{l}}\left\|x-\sum b^{(n)}_{i}E((b^{(n)}_{i})^{*}x)\right\|_{2}+\sum_{y\in F_{l}} \left\|yb^{(n)}_{i}-b^{(n)}_{i}y\right\|_{2}\longrightarrow_{\omega} 0.
$$
Therefore there exists a subsequence of the $b^{(n_{j})}_{i}$ such that
$$
\sum_{x\in G_{l}}\left\|x-\sum b^{(n_{j})}_{i}E((b^{(n_{j})}_{i})^{*}x)\right\|_{2}+\sum_{y\in F_{l}} \left\|yb^{(n_{j})}_{i}-b^{(n_{j})}_{i}y\right\|_{2}\longrightarrow 0.
$$
In particular, we can choose $c^{(l)}_{i}:=b^{(n_{j})}_{i}$ where $n_{j}$ is sufficiently large so that the above sum is less than $2^{-l}$.
We claim is the sequence $\{c^{(l)}_{i}\}$ is an approximately inner $L^2M_N$-basis.
Clearly the collection $\{c^{(l)}_{i}\}$ satisfies the conditions from Definition \ref{defn:ApproximatePPBasis} for $x\in G$ while it satisfies \eqref{eq:AICommutativity} for $y\in F$.
The result follows since $G$ and $F$ are $\sigma$-strong* dense in $M$ and $N$ respectively.
\end{proof}
\begin{rem}[{cf.~\cite[Ex.~8.1.9]{MR3242743}}]
\label{rem:OppositeWithBraiding}
For a monoidal category $\cC$, the \emph{monoidal opposite} $\cC^{\rm mp}$ is the same category with the opposite monoidal product given by
$a \otimes_{\rm mp} b := b \otimes_\cC a$.
It is equipped with the inverse associator: $$
\alpha_{a,b,c}^{\rm mp}
:
(a \otimes_{\rm mp} b) \otimes_{\rm mp} c
:=
c \otimes (b \otimes a)
\xrightarrow{\alpha_{c,b,a}^{-1}}
(c \otimes b) \otimes a
=:
a \otimes_{\rm mp} (b \otimes_{\rm mp} c).
$$
Now suppose $\cC$ is a braided monoidal category.
Observe that the braiding endows the linear equivalence $\cC \to \cC^{\rm mp}$
with a monoidal structure
$$
\mu_{a,b}
:
\operatorname{id}(a\otimes b)
=
a\otimes b
\xrightarrow{\beta_{a,b}}
b\otimes a
=
\operatorname{id}(a)\otimes_{\rm mp} \operatorname{id}(b)
$$
giving an equivalence of monoidal categories $\cC\simeq \cC^{\rm mp}$.
Now transporting $\beta$ to $\cC^{\rm mp}$ along this monoidal equivalence endows $\cC^{\rm mp}$ with the braiding
$$
\beta^{\rm mp}_{a,b}
:=
\mu_{b,a} \circ \operatorname{id}(\beta_{a,b})\circ \mu^{-1}_{a,b}
:
a \otimes_{\rm mp} b
=
b \otimes a
\xrightarrow{
\beta_{b,a}\circ \beta_{a,b}\circ \beta_{a,b}^{-1}
=
\beta_{b,a}
}
a \otimes b
=
b \otimes_{\rm mp} a
$$
such that $(\operatorname{id},\mu): (\cC, \alpha, \beta) \to (\cC^{\rm mp}, \alpha^{\rm mp}, \beta^{\rm mp})$ is a braided monoidal equivalence.
(Observe that if we chose our monoidal structure for $\operatorname{id}$ to be $\mu_{a,b}:=\beta_{b,a}^{-1}$, we would still obtain $\beta^{\rm mp}_{a,b} = \beta_{b,a}$ as the transported braiding on $\cC^{\rm mp}$.)
\end{rem}
Using the above remark, we now translate the definition of the unitary braiding into the language of bimodules.
For $X,Y \in \mathsf{Bim_{fgp}}(M)$,
we write $u_{X,Y}=u^{L^2M}_{-\boxtimes_M Y, -\boxtimes_M X}$.
Choosing approximately inner $X_M,Y_M$-bases $\{b_i^{(n)}\}_i,\{c_j^{(n)}\}_j$
and ordinary $X_M,Y_M$-bases $\{b_i\}_i,\{c_j\}_j$ respectively,
unpacking Theorem \ref{ThmDef:uGFa} and Definition \ref{defn:sigmaFH} gives the following formulas.
\begin{align}
u_{X,Y} &=
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.55,-.5) rectangle (.7,.5);
\filldraw[\rColor] (-.4,-.5) -- (-.4,.5) -- (.7,.5) -- (.7,-.5);
\end{scope}
\draw[dashed] (-.4,-.5) -- (-.4,.5);
\draw[blue,thick] (.4,-.5) node[below]{$\scriptstyle Y$} .. controls ++(90:.45cm) and ++(270:.45cm) .. (0,.5);
\filldraw[\rColor] (.2,0) circle (.05cm);
\draw[\XColor,thick] (0,-.5) node[below]{$\scriptstyle X$} .. controls ++(90:.45cm) and ++(270:.45cm) .. (.4,.5);
}
\approx
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.7,-1.5) rectangle (.9,1.5);
\filldraw[\rColor] (-.4,-1.5) -- (-.4,1.5) -- (1,1.5) -- (1,-1.5);
\end{scope}
\draw[dashed] (-.4,-1.5) -- (-.4,1.5);
\draw[\XColor,thick] (0,-1.5) -- (0,-.8);
\draw[blue,thick] (.4,-1.5) -- (.4,-.5) .. controls ++(90:.45cm) and ++(270:.45cm) .. (0,.5) -- (0,1.5);
\filldraw[\rColor] (.2,0) circle (.05cm);
\draw[\XColor,thick] (.4,.8) -- (.4,1.5);
\draw[\HColor,thick] (0,-.5) .. controls ++(90:.45cm) and ++(270:.45cm) .. (.4,.5);
\roundNbox{unshaded}{(0,.8)}{.3}{.35}{.35}{\scriptsize{$(v_n^Y)^\dag$}};
\roundNbox{unshaded}{(-.2,-.8)}{.3}{.15}{.15}{\scriptsize{$v_n^{M}$}};
}
=
\sum_i
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.7,-1.5) rectangle (.9,1.5);
\filldraw[\rColor] (-.4,-1.5) -- (-.4,1.5) -- (1,1.5) -- (1,-1.5);
\end{scope}
\draw[dashed] (-.4,-1.5) -- (-.4,1.5);
\draw[\XColor,thick] (0,-1.5) -- (0,-.8);
\draw[\HColor,thick] (0,-.5) -- (0,-.25);
\draw[\HColor,thick] (.4,.5) -- (.4,.25);
\filldraw[\HColor] (0,-.25) circle (.05cm);
\filldraw[\HColor] (.4,.25) circle (.05cm);
\draw[blue,thick] (.4,-1.5) -- (.4,-.5) .. controls ++(90:.45cm) and ++(270:.45cm) .. (0,.5) -- (0,1.5);
\draw[\XColor,thick] (.4,.8) -- (.4,1.5);
\roundNbox{unshaded}{(0,.8)}{.3}{.35}{.35}{\scriptsize{$(v_n^Y)^\dag$}};
\roundNbox{unshaded}{(-.2,-.8)}{.3}{.15}{.15}{\scriptsize{$v_n^{M}$}};
\node at (.65,.25) {\scriptsize{$e_i$}};
}
=
\sum_{i,j}
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.7,-1.5) rectangle (.9,3.5);
\filldraw[\rColor] (-.4,-1.5) -- (-.4,.5) -- (-.2,.5) -- (-.2,2.5) -- (-.4,2.5) -- (-.4,3.5) -- (1,3.5) -- (1,-1.5);
\end{scope}
\draw[dashed] (-.4,-1.5) -- (-.4,.5);
\draw[\XColor,thick] (0,-1.5) -- (0,-.8);
\draw[\HColor,thick] (0,-.5) -- (0,-.25);
\draw[\HColor,thick] (.4,1.5) -- (.4,1.25);
\filldraw[\HColor] (0,-.25) circle (.05cm);
\filldraw[\HColor] (.4,1.25) circle (.05cm);
\draw[blue,thick] (.4,-1.5) -- (.4,-.5) .. controls ++(90:.45cm) and ++(270:.45cm) .. (0,.5);
\draw[dashed] (-.2,.5) -- (-.2,2.5);
\draw[\XColor,thick] (.4,1.8) -- (.4,3.5);
\draw[dashed] (-.4,2.5) -- (-.4,3.5);
\draw[blue,thick] (0,2.5) -- (0,3.5);
\roundNbox{unshaded}{(-.2,.8)}{.3}{.15}{.15}{\scriptsize{$c_j^\dag$}};
\roundNbox{unshaded}{(-.2,2.8)}{.3}{.15}{.15}{\scriptsize{$c_j$}};
\roundNbox{unshaded}{(0,1.8)}{.3}{.35}{.35}{\scriptsize{$(v_n^{M})^\dag$}};
\roundNbox{unshaded}{(-.2,-.8)}{.3}{.15}{.15}{\scriptsize{$v_n^{M}$}};
\node at (.65,1.25) {\scriptsize{$e_i$}};
}
=
\sum_{i,j}
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.35,-2.2) rectangle (1,2.2);
\filldraw[\rColor] (-.2,-2.2) -- (-.2,-1.5) -- (0,-1.5) -- (0,-.5) -- (.3,-.5) -- (.3,.5) -- (0,.5) -- (0,1.5) -- (-.2,1.5) -- (-.2,2.2) -- (1.05,2.2) -- (1.05,-2.2);
\end{scope}
\draw[dashed] (-.2,-2.2) -- (-.2,-1.5);
\draw[\XColor,thick] (.2,-2.2) -- (.2,-1.5);
\draw[dashed] (0,-1.5) -- (0,-.5);
\draw[blue,thick] (.6,-2.2) -- (.6,-.5);
\draw[dashed] (.3,-.5) -- (.3,.5);
\draw[dashed] (0,.5) -- (0,1.5);
\draw[\XColor,thick] (.6,.5) -- (.6,2.2);
\draw[dashed] (-.2,1.5) -- (-.2,2.2);
\draw[blue,thick] (.2,1.5) -- (.2,2.2);
\roundNbox{unshaded}{(0,1.5)}{.3}{.1}{.1}{\scriptsize{$c_j$}};
\roundNbox{unshaded}{(.3,.5)}{.3}{.15}{.15}{\scriptsize{$b_i^{(n)}$}};
\roundNbox{unshaded}{(.3,-.5)}{.3}{.15}{.15}{\scriptsize{$c_j^\dag$}};
\roundNbox{unshaded}{(0,-1.5)}{.3}{.1}{.1}{\scriptsize{$(b_i^{(\!n\!)}\!)^\dag$}};
}
\label{eq:OverBraiding}
\\
u_{Y,X}^\dag&=
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.55,-.5) rectangle (.7,.5);
\filldraw[\rColor] (-.4,-.5) -- (-.4,.5) -- (.7,.5) -- (.7,-.5);
\end{scope}
\draw[dashed] (-.4,-.5) -- (-.4,.5);
\draw[\XColor,thick] (0,-.5) .. controls ++(90:.45cm) and ++(270:.45cm) .. (.4,.5);
\filldraw[\rColor] (.2,0) circle (.05cm);
\draw[blue,thick] (.4,-.5) .. controls ++(90:.45cm) and ++(270:.45cm) .. (0,.5);
}
\approx
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.55,-1.5) rectangle (.9,1.5);
\filldraw[\rColor] (-.4,-1.5) -- (-.4,1.5) -- (1,1.5) -- (1,-1.5);
\end{scope}
\draw[dashed] (-.4,-1.5) -- (-.4,1.5);
\draw[\XColor,thick] (0,-1.5) -- (0,-.5) .. controls ++(90:.45cm) and ++(270:.45cm) .. (.4,.5) -- (.4,1.5);
\filldraw[\rColor] (.2,0) circle (.05cm);
\draw[\HColor,thick] (.4,-.5) .. controls ++(90:.45cm) and ++(270:.45cm) .. (0,.5);
\draw[blue,thick] (.4,-1.5) -- (.4,-.5);
\draw[blue,thick] (0,.5) -- (0,1.5);
\roundNbox{unshaded}{(-.2,.8)}{.3}{.15}{.15}{\scriptsize{$(w_n^{M})^\dag$}};
\roundNbox{unshaded}{(0,-.8)}{.3}{.35}{.35}{\scriptsize{$w_n^X$}};
}
=
\sum_j
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.55,-1.5) rectangle (.9,1.5);
\filldraw[\rColor] (-.4,-1.5) -- (-.4,1.5) -- (1,1.5) -- (1,-1.5);
\end{scope}
\draw[dashed] (-.4,-1.5) -- (-.4,1.5);
\draw[\HColor,thick] (0,.5) -- (0,.25);
\draw[\HColor,thick] (.4,-.25) -- (.4,-.5);
\filldraw[\HColor] (0,.25) circle (.05cm);
\filldraw[\HColor] (.4,-.25) circle (.05cm);
\draw[\XColor,thick] (0,-1.5) -- (0,-.5) .. controls ++(90:.45cm) and ++(270:.45cm) .. (.4,.5) -- (.4,1.5);
\draw[blue,thick] (.4,-1.5) -- (.4,-.5);
\draw[blue,thick] (0,.5) -- (0,1.5);
\roundNbox{unshaded}{(-.2,.8)}{.3}{.15}{.15}{\scriptsize{$(w_n^{M})^\dag$}};
\roundNbox{unshaded}{(0,-.8)}{.3}{.35}{.35}{\scriptsize{$w_n^{X}$}};
\node at (-.2,.25) {\scriptsize{$f_j$}};
}
=
\sum_{i.j}
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.55,-3.5) rectangle (.9,1.5);
\filldraw[\rColor] (-.4,-3.5) -- (-.4,-2.5) -- (-.2,-2.5) -- (-.2,-.5) -- (-.4,-.5) -- (-.4,1.5) -- (1,1.5) -- (1,-3.5);
\end{scope}
\draw[dashed] (-.4,-.5) -- (-.4,1.5);
\draw[\HColor,thick] (0,.5) -- (0,.25);
\draw[\HColor,thick] (.4,-1.25) -- (.4,-1.5);
\filldraw[\HColor] (0,.25) circle (.05cm);
\filldraw[\HColor] (.4,-1.25) circle (.05cm);
\draw[\XColor,thick] (0,-.5) .. controls ++(90:.45cm) and ++(270:.45cm) .. (.4,.5) -- (.4,1.5);
\draw[dashed] (-.2,-2.5) -- (-.2,-.5);
\draw[blue,thick] (.4,-3.5) -- (.4,-1.5);
\draw[blue,thick] (0,.5) -- (0,1.5);
\draw[dashed] (-.4,-3.5) -- (-.4,-2.5);
\draw[\XColor,thick] (0,-3.5) -- (0,-2.5);
\roundNbox{unshaded}{(-.2,.8)}{.3}{.15}{.15}{\scriptsize{$(w_n^{M})^\dag$}};
\roundNbox{unshaded}{(0,-1.8)}{.3}{.35}{.35}{\scriptsize{$w_n^{M}$}};
\roundNbox{unshaded}{(-.2,-.8)}{.3}{.15}{.15}{\scriptsize{$b_i$}};
\roundNbox{unshaded}{(-.2,-2.8)}{.3}{.15}{.15}{\scriptsize{$b_i^\dag$}};
\node at (-.2,.25) {\scriptsize{$f_j$}};
}
=
\sum_{i,j}
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.35,-2.2) rectangle (1,2.2);
\filldraw[\rColor] (-.2,-2.2) -- (-.2,-1.5) -- (0,-1.5) -- (0,-.5) -- (.3,-.5) -- (.3,.5) -- (0,.5) -- (0,1.5) -- (-.2,1.5) -- (-.2,2.2) -- (1.05,2.2) -- (1.05,-2.2);
\end{scope}
\draw[dashed] (-.2,-2.2) -- (-.2,-1.5);
\draw[\XColor,thick] (.2,-2.2) -- (.2,-1.5);
\draw[dashed] (0,-1.5) -- (0,-.5);
\draw[blue,thick] (.6,-2.2) -- (.6,-.5);
\draw[dashed] (.3,-.5) -- (.3,.5);
\draw[dashed] (0,.5) -- (0,1.5);
\draw[\XColor,thick] (.6,.5) -- (.6,2.2);
\draw[dashed] (-.2,1.5) -- (-.2,2.2);
\draw[blue,thick] (.2,1.5) -- (.2,2.2);
\roundNbox{unshaded}{(0,1.5)}{.3}{.1}{.1}{\scriptsize{$c_j^{(n)}$}};
\roundNbox{unshaded}{(.3,.5)}{.3}{.15}{.15}{\scriptsize{$b_i$}};
\roundNbox{unshaded}{(.3,-.5)}{.3}{.15}{.15}{\scriptsize{$(c_j^{(\!n\!)}\!)^\dag$}};
\roundNbox{unshaded}{(0,-1.5)}{.3}{.1}{.1}{\scriptsize{$b_i^\dag$}};
}
\label{eq:UnderBraiding}
\end{align}
That is, the braidings $u_{X,Y}, u_{Y,X}^\dag$ can be expressed as the following $\|\cdot\|_2$-limits of $M$-finite rank operators:
\begin{equation}
\label{eq:uXY}
u_{X,Y}
=
\lim_\tau\sum_{i,j} | c_j \boxtimes b_i^{(n)} \rangle\langle b_i^{(n)} \boxtimes c_j |
\qquad\text{and}\qquad
u_{Y,X}^\dag
=
\lim_\tau\sum_{i,j} | c_j^{(n)} \boxtimes b_i \rangle\langle b_i \boxtimes c_j^{(n)} |.
\end{equation}
The content of \S\ref{sec:W*ApproxNatTrans} then translates into the facts that:
\begin{itemize}
\item
The limit $u_{X,Y}$ exists and is independent of the choice of (approximately inner) $X_M,Y_M$-bases (Theorem \ref{ThmDef:uGFa} and Proposition \ref{prop:uGF-Natural}),
\item
$u_{X,Y}$ is unitary (Proposition \ref{prop:uGFUnitary}),
\item
$u_{X,Y}$ is natural in $X$ and $Y$ (Proposition \ref{prop:uGF-Centralizing-Natural}), and
\item
$u_{X,Y}$ satisfies the braid relations (Propositions \ref{braid3} and \ref{braid4}).
\end{itemize}
\subsection{Examples}
\label{sec:Examples}
We now compute many examples of $\tilde{\chi}(M)$ for various von Neumann algebras $M$.
In this section, in order to easily make contact with other results in the literature, we work with the Hilbert space version of $\mathsf{Mod_{fgp}}(M)$, i.e., Hilbert spaces with a normal right $M$-action such that $\dim(H_M)<\infty$ (see Example \ref{ex:RightFGPMod} for the equivalence with $\mathsf{W^*Alg_{fgp}}(\bbC \to M)$).
Now we recall two different notions of central sequence in the von Neumann algebra literature which agree for finite von Neumann algebras but are subtly distinct in general. The standard notion to use outside the finite setting is \textit{centralizing sequence}, introduced by Connes in \cite{MR0358374}.
One way to define this is a say a norm bounded sequence $\{x_{i}\}\subseteq M$ is centralizing if for all $\eta\in L^{2}M$, $\|x_{i}\eta-\eta x_{i}\|_2\rightarrow 0$.
This is in contrast to \textit{central sequences}, which satisfy $x_{i}m-mx_{i}\rightarrow 0$, in the strong*-topology (the latter notion is compatible with our use of the term central sequence). Note every centralizing sequence is central, but in general the converse is not true. For finite von Neumann algebras, however, these two notions agree.
In both settings, we say a sequence is \textit{trivial} if there exists a scalar $\lambda$ such that $x_{i}-\lambda 1_{M}\rightarrow 0$ in the strong*-topology.
For our formalism, the notion of central sequence is the correct one; however, we will occasionally need to make use of results that are stated in terms of centralizing sequences.
\begin{ex}
[$L^{\infty}(X, \mu)$]
For the abelian von Neumann algebra $L^{\infty}(X,\mu)$ over a finite measure space $(X, \mu)$, constant sequences are centrally trivial.
Thus any centrally trivial right-finite correspondence is simply a finitely generated projective module, made into a bimodule by defining the right action to be the left action.
All of these correspondences are inner, hence approximately inner.
This braided category is equivalent to the symmetric monoidal category of finitely generated projective modules, which is equivalent to the category of finite dimensional measurable Hilbert bundles over $(X, \mu)$ \cite{MR641217}.
\end{ex}
\begin{ex}
[Connes' $\chi(M)$ \cite{MR377534} and Jones' $\kappa$ \cite{MR585235}]
\label{ex:ConnesChi}
Let $M$ be a separable finite von Neumann algebra with faithful normal trace $\operatorname{tr}_M$.
We show that Connes' $\chi(M)$ embeds as a multiplicative subgroup of the monoid of equivalence classes of invertible objects in $\tilde{\chi}(M)$.
For $\alpha\in\operatorname{Aut}(M)$, we define the corresponding $M-M$ bimodule as ${}_\alpha L^2M$.
Denoting the image of $1_M$ in ${}_\alpha L^2M$ by $\Omega$,
the left action is given by $a\rhd x\Omega:=\alpha(a)x\Omega$,
the right action is given by $x\Omega\lhd b=xb\Omega$, and
the $M$-valued inner product is given by $\langle x\Omega|x'\Omega\rangle^{L^2M}_M: = x^*x'$.
Moreover, the map $\alpha \mapsto {}_\alpha L^2M$ descends to a group isomorphism from $\operatorname{Out}(M)$ onto the group of unitary equivalence classes of
invertible $M-M$ bimodules $H$ such that $\dim({}_MH) = \dim(H_M) = 1$.
Recall that an automorphism $\alpha$ is approximately inner if there exists a sequence of unitaries $u_{n}\in U(M)$ such that $\|\alpha(x)u_n - u_{n} x\|_2 \to 0$ for all $x\in M$.
According to Proposition \ref{prop:AI&APPB}, to show ${}_\alpha L^2M$ is approximately inner,
it suffices to show $\{u_n\Omega\}$ is an approximate Pimsner-Popa basis centralized by $M$.
It is clear that $u_n\rhd \langle u_n\Omega|x\Omega\rangle^{{}_\alpha L^2M}_M = x$ for all $x\in M$,
and for all $a\in M$,
$$
\|a\rhd u_n\Omega - u_n\Omega\lhd a\|_2
=
\|\alpha(a)u_n\Omega - u_na\Omega\|_2
\longrightarrow
0.
$$
Therefore, ${}_\alpha L^2M$ is approximately inner.
By Proposition \ref{prop:BimodualCT&CS},
the bimodule ${}_\alpha L^2M$ is centrally trivial if $\alpha$ is centrally trivial.
Now assume $M$ is a $\rm{II}_{1}$ factor.
The subcategory of $\tilde{\chi}(M)$ spanned by the image of $\chi(M)$ is a pointed braided unitary tensor category.
The entire braided tensor structure here is uniquely determined by the quadratic form $\kappa$ on $\chi(M)$ determined by $u:=u_{{}_\alpha L^2M,{}_\alpha L^2M}=\kappa(\alpha) 1_{{}_\alpha L^2M\boxtimes_{M} {}_\alpha L^2M}$.
Since
$$
u=
\|\cdot\|_2-\lim_n |\Omega\boxtimes u_n\Omega\rangle \langle u_n\Omega\boxtimes \Omega|,
$$
for an arbitrary bounded vector $\Omega\boxtimes m\Omega$,
we see
$$
\kappa(\alpha)(\Omega \boxtimes m\Omega)
=
u(\Omega\boxtimes m\Omega)
=
\Omega \boxtimes \|\cdot\|_2-\lim_n u^{*}_{n}\alpha(u_{n})m\Omega.
$$
Thus this $\kappa$ is precisely Jones' quadratic form $\kappa$ on $\chi(M)$ \cite[Def.~2.4]{MR585235}.
Further references for Connes' $\chi(M)$ and Jones' $\kappa$ include \cite{VaughanNotesOnChi,JianduanChenThesis,MR1642584}.
\end{ex}
\begin{ex}
[$\tilde{\chi}(R)$ is trivial]
The following proposition specialized to the case $S=\bbC$ shows that the only centrally trivial bimodules of $R$ are inner, and thus $\tilde{\chi}(R)$ is trivial.
This extends Connes' result that $\chi(R)$ is trivial \cite{MR377534}.
\begin{prop}
\label{CTRHYPER}
Let $M=R\otimes S$ where $R$ is the hyperfinite $\rm{II}_1$ factor and $S$ is any factor. Let $H$ be a separable $M-M$ Hilbert bimodule such that for all $\xi\in H$ and all central sequences $x=(x_{n})_{n\in \mathbb{N}}\subseteq R$, $\|(x_{n}\otimes 1_{S})\xi -\xi(x_{n}\otimes 1_{S})\|_{2}\rightarrow 0$.
Then $H\cong L^{2}R\otimes K$ for some $S-S$ bimodule $K$.
\end{prop}
\begin{proof}
Let $\xi\in H$ with $\|\xi\|=1$. Represent $R=\otimes^{\infty}_{i=1} M_{2}(\bbC)$, and denote $R_{n}=\otimes^{n}_{i=1} M_{2}(\bbC)\subseteq R$. We claim there is some $n_0$ so that for all unitaries $u\in R^{\prime}_{n_0}\cap R$, $\|(u\otimes 1_{S})\xi-\xi (u\otimes 1_{S})\|<\frac{1}{2}$. Otherwise, we could find a sequence of unitaries $u_{n}\in R^{\prime}_{n}\cap R$ with $\|(u_{n}\otimes 1_{S})\xi-\xi (u_{n}\otimes 1_{S})\|\ge \frac{1}{2}$. However by construction the sequence $(u_{n})_{n\in \bbN}\subseteq R$ is central, contradicting the hypothesis.
Choose such an $n_{0}$, and consider the weakly compact convex subset
$$
\overline{\operatorname{co}}^{w}\set{(u\otimes 1_{S})\xi(u^{*}\otimes 1_{S})}{ u\in R^{\prime}_{n_0}\cap R}\subseteq H.
$$
This has a unique element of minimal norm, $\xi_{0}$. By hypothesis
$$
2\|\xi\|^{2}-2\operatorname{Re}\langle (u\otimes 1_{S})\xi(u^{*}\otimes 1_{S}), \xi\rangle\
=
\|(u\otimes 1_{S})\xi(u^{*}\otimes 1_{S})-\xi \|^{2}
<
\frac{1}{4}
$$
and thus
$$
\frac{7}{8} <\ \operatorname{Re}\langle (u\otimes 1_{S})\xi(u^{*}\otimes 1_{S}), \xi\rangle.
$$
Therefore $\xi_{0}\ne 0$.
Since $\|(u\otimes 1_{S})\xi_{0}(u^{*}\otimes 1_{S})\|=\|\xi_{0}\|$,
uniqueness of $\xi_0$ implies
$(u\otimes 1_{S})\xi_0=\xi_0 (u\otimes 1_{S})$ for all unitaries (hence all elements) $u\in R^{\prime}_{n}\cap R$.
Note that $R_{n}\xi_{0} R_{n}$ is a cyclic bimodule over the finite dimensional full matrix algebra $R_{n}$, and thus contains a non-zero $R_{n}$ central vector $\xi_{1}$ (which is evidently $R^{\prime}_{n}\cap R$-central).
Thus since $R=\langle R_{n}, R^{\prime}_{n}\cap R\rangle$, we see that $\xi_{1}$ is R-central and thus $R$-bounded by \cite[Lem.~3.20]{MR3040370}.
This means the $R-R$ bimodule $H_{1}:=\overline{(R\otimes 1_{S})\xi_{1} (R\otimes 1_{S})}^{\|\cdot\|}$ is canonically isomorphic to $L^{2}R$ as an $R$-$R$ bimodule via the map defined by $1_{R}\rightarrow \xi_{1}$.
Choosing a vector $\xi^{\prime}\in H^{\perp}_{1}$, we can repeat this procedure, obtaining a decomposition as $R$-bimodules $H\cong L^{2}R\otimes K $ where $K$ is a separable multiplicity space. But note $K\cong \Omega_{R}\otimes K$ is the space of $R\otimes 1_{S}$ central vectors hence is closed under the left and right actions of $1_{R}\otimes S$.
\end{proof}
\end{ex}
The following proposition is well known to experts.
We record it here for completeness and convenience of the reader.
\begin{prop}\label{quasitrivialsubfactor}
Let $N\subseteq M$ be a finite index $\rm{II}_{1}$ subfactor.
Then $M=N\vee (N^{\prime}\cap M)$ if and only if $ L^{2}M\cong \bigoplus_{i}L^{2}N$ as $N-N$ bimodules.
\end{prop}
\begin{proof}
Suppose $M=N\vee (N^{\prime}\cap M)$.
Since $N^{\prime}\cap M$ is finite dimensional \cite{MR0696688}, pick an orthonormal basis $ \{e_{i}\}\subseteq N^{\prime}\cap M $ with respect to $\operatorname{tr}_{M}=E_N|_{N'\cap M}$.
Then each $e_iL^2N \cong L^2N$,
and
$L^{2}M\cong \bigoplus_{i} e_{i}L^{2}N$
as
$$
\langle e_in_i, e_j n_j \rangle_{L^2M}
=
\operatorname{tr}_M(n_j^*e_j^*e_i n_i)
=
\operatorname{tr}_N(n_j^* E_N(e_j^*e_i)n_i)
=
\delta_{i=j}\langle n_i, n_j\rangle_{L^2N}.
$$
The converse is obvious.
\end{proof}
\begin{ex}
[$\tilde{\chi}(N)$ is trivial for $N$ non-Gamma]
Let $N$ be a non-Gamma $\rm{II}_1$ factor, and $H\in \tilde{\chi}(N)$ irreducible.
Setting $X=L^{2}N\oplus H$ and $M:= |X\boxtimes_{N} \overline{X}|$, we get a finite index $\rm II_1$ subfactor $N\subseteq M$ by Example \ref{ex:RealizeSeparatedQSystem} which is approximately inner by Proposition \ref{PopaEq}.
By \cite[Proposition 2.6 (iv)]{MR1339767}, $M=N\vee (N^{\prime}\cap M)$,
so by Proposition \ref{quasitrivialsubfactor}, ${}_{N} L^{2}M_{N}\cong \bigoplus_{i} {}_{N} L^{2}N_{N} $.
On the other hand, $L^{2}M\cong X\boxtimes_{N} \overline{X}\cong L^{2}N\oplus H\oplus \overline{H}\oplus (H\boxtimes_{N} \overline{H})$ contains $H$ as an irreducible summand, and thus we have $H\cong L^{2}N$ as an $N$-$N$ bimodule.
\end{ex}
\begin{ex}
[$\tilde{\chi}(R\otimes N)$ is trivial for $N$ non-Gamma] \label{ex:Chi_fus(RboxtimesN)}
Let $H\in \tilde{\chi}(R\otimes N)$ be irreducible, and consider $X=L^{2}(R\otimes N)\oplus H$.
Then since $H$ is centrally trivial, so is $X$.
By Lemma \ref{CTRHYPER}, $X\cong L^{2}(R)\otimes K$ where $K$ is an $N$-$N$ bifinite bimodule.
In particular, setting $M:=|K\boxtimes_{N} \overline{K}|$, $N\subseteq M$ is a finite index $\rm II_1$ subfactor by Example \ref{ex:RealizeSeparatedQSystem}.
Furthermore, since $N$ is non-Gamma, $M$ is non-Gamma by \cite[Prop.~1.11]{MR860811}.
Now since $H$ is approximately inner, the $R\otimes N-R\otimes N$ bimodule $X\boxtimes_{R\otimes N} \overline{X}\cong L^{2}(R\otimes M)$ is approximately inner, hence $R\otimes N\subseteq R\otimes M$ is a finite index approximately inner subfactor.
But note this subfactor is simply $R\otimes (N\subseteq M)$.
Thus the inclusion $(R\otimes N)_{\omega}\subseteq (R\otimes M)_{\omega}$
is $[M:N]^{-1}$-Markov, and has a finite Pimsner-Popa basis lifting to a Pimsner-Popa basis of $(R\otimes N)^{\omega}\subseteq (R\otimes M)^{\omega}$ (where $\omega$ is a non-principal ultrafilter).
Furthermore, since
$R \otimes N= (R\otimes 1)\ \vee (1\otimes N)$,
we have the equality
\begin{equation}
\label{eq:RNcommutant}
(R\otimes N)^{\prime}\cap (R\otimes M)^{\omega}
=
\left((1\otimes N)^{\prime}\cap (R\otimes M)^{\omega}\right) \cap \left( (R\otimes 1)^{\prime} \cap (R\otimes M)^{\omega}\right).
\end{equation}
By \cite[Prop.~3.2(1)]{MR2661553}, the inclusion $1\otimes N\subseteq R\otimes M$ has spectral gap, and thus
\begin{equation}
\label{eq:1Ncommutant}
(1\otimes N)^{\prime}\cap (R\otimes M)^{\omega}=\left((1\otimes N)^{\prime}\cap (R\otimes M)\right)^{\omega}=(R\otimes (N^{\prime}\cap M))^{\omega}.
\end{equation}
Since $(R\otimes 1)^{\omega}\subseteq (R\otimes M)^{\omega}$,
combining \eqref{eq:RNcommutant} and \eqref{eq:1Ncommutant},
we have
$$
(R\otimes N)^{\prime}\cap (R\otimes M)^{\omega}= (R\otimes (N^{\prime} \cap M))^{\omega} \cap \left( (R\otimes 1)^{\prime} \cap (R\otimes M)^{\omega}\right)=(R\otimes 1)^{\prime} \cap (R \otimes (N^{\prime}\cap M))^{\omega}.
$$
Thus there is a finite Pimsner-Popa basis $\{\tilde{m}_{i}\}$ for $(R\otimes N)^{\omega}\subseteq (R\otimes M)^{\omega}$ with $\tilde{m}_{i}=\{b^{(k)}_{i}\}_{k\in \bbN}$ where each $b^{(k)}_{i}\in R\otimes (N^{\prime}\cap M)$.
In particular, for any $m\in M$ and any $\varepsilon>0$, there exists finitely many elements $r_{i}\in R$, $k_{i}\in N^{\prime}\cap M$, and $n_{i}\in N$ such that $\|(1_{R}\otimes m)-\sum_{i} r_{i}\otimes k_{i}n_{i}\|_{2}<\varepsilon.$
But then applying the trace preserving conditional expectation $E=\operatorname{tr}_R\otimes \operatorname{id}_M:R\otimes M\rightarrow M$, we have
$$
\|m-\sum_{i} \operatorname{tr}_{R}(r_{i})k_{i}n_{i}\|_{2}=\|E((1_{R}\otimes m)-\sum_{i} r_{i}\otimes k_{i}n_{i})\|_{2}\le \|(1_{R}\otimes m)-\sum_{i} r_{i}\otimes k_{i}n_{i}\|_{2}<\varepsilon.
$$
Therefore $N\vee (N^{\prime}\cap M)=M$. By Proposition \ref{quasitrivialsubfactor}, this implies that ${}_{N}L^{2}M_{N}\cong \bigoplus_{i} {}_{N}L^{2}N_{N}$.
But since $ |X\boxtimes_{R\otimes N} \overline{X}|$ is isomorphic to $L^{2}R\otimes |K\boxtimes_{R\otimes N} \overline{K}|$ as $N-N$ bimodules, we have that $X\boxtimes_{R\otimes N} \overline{X} \cong \bigoplus_{i} L^{2}(R\otimes N) $ as $R\otimes N-R\otimes N$ bimodules.
But recall $X\boxtimes_{R\otimes N} \overline{X}\cong L^{2}(R\otimes N)\oplus H\oplus \overline{H}\oplus (H\boxtimes_{R\otimes N} \overline{H})$, which contains $H$ as an irreducible summand. Thus $H$ is trivial.
\end{ex}
\section{Local extension}
\label{sec:LocalExtensionChapter}
In this section, we prove Theorem \ref{thmalpha:LocalExtension}, i.e., $\tilde{\chi}(|Q|) = \tilde{\chi}(M)_Q^{\operatorname{loc}}$ for a commutative Q-system $Q\in \tilde{\chi}(M)$, where $|Q|$ is the realization of $Q$ defined in \S\ref{sect:Q-systemRealization} below.
This result appears as Theorem \ref{thm:LocalExtension}.
\subsection{Q-system realization}
\label{sect:Q-systemRealization}
Q-systems are unitary versions of Frobenius algebra objects which were originally introduced by Longo in \cite{MR1257245} to describe the canonical endomorphism for type $\rm III$ subfactors \cite{MR1027496}.
In this section, we define the \emph{realization} procedure \cite{MR3948170,2105.12010}
(based on \cite{MR2097363} and \cite[\S4.1]{MR3221289})
which
given a Q-system $Q$ over a $\rm II_1$ factor $M$, recovers a von Neumann algebra $|Q|$ containing $M$ and a conditional expectation $E_M: |Q| \to M$ with finite Pimsner-Popa index.
This story works in much broader generality, but we restrict to $\rm II_1$ factors here for ease of exposition.
As in \S\ref{sec:ModuleRealization}, for this section, $M$ is a $\rm II_1$ factor, and we denote $\bbC,M, {}_\bbC M_M$ as before:
$$
\tikzmath{\filldraw[\rColor, rounded corners=5, very thin, baseline=1cm] (0,0) rectangle (.6,.6);}=M
\qquad
\qquad
\tikzmath{
\filldraw[white, rounded corners=5, very thin, baseline=1cm] (0,0) rectangle (.6,.6);
\draw[dotted, rounded corners=5pt] (0,0) rectangle (.6,.6);
}=\bbC
\qquad
\qquad
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (0,0) rectangle (.6,.6);
\filldraw[\rColor] (.3,0) rectangle (.6,.6);
\end{scope}
\draw[dashed] (.3,0) -- (.3,.6);
\draw[dotted, rounded corners=5pt] (.3,0) -- (0,0) -- (0,.6) -- (.3,.6);
}={}_\bbC M_M.
$$
\begin{defn}
Given a $\rm II_1$ factor $M$, a \emph{Q-system} in $\mathsf{Bim_{fgp}}(M)$ is a triple $(Q,m,i)$
where $Q\in \mathsf{Bim}(M)$ is bimodule, and $m: Q\boxtimes_M Q \to Q$ and $i: M \to Q$ are bounded maps that satisfy certain relations best described graphically.
Representing $M$ by a shaded region and $M$ by a strand, $m$ is a trivalent vertex, and $i$ is a univalent vertex; adjoints are represented by vertical reflection.
$$
\tikzmath{\filldraw[\rColor, rounded corners=5, very thin, baseline=1cm] (0,0) rectangle (.6,.6);}=M
\qquad
\tikzmath{
\fill[\rColor, rounded corners=5pt ] (0,0) rectangle (.6,.6);
\draw[black,thick] (.3,0) -- (.3,.6);
}={}_MQ_M.
\qquad
\tikzmath{
\fill[\rColor, rounded corners=5pt] (-.3,0) rectangle (.9,.6);
\draw[black,thick] (0,0) arc (180:0:.3cm);
\draw[black,thick] (.3,.3) -- (.3,.6);
\filldraw[black] (.3,.3) circle (.05cm);
}=m
\qquad
\tikzmath{
\fill[\rColor, rounded corners=5pt] (-.3,0) rectangle (.9,-.6);
\draw[black,thick] (0,0) arc (-180:0:.3cm);
\draw[black,thick] (.3,-.3) -- (.3,-.6);
\filldraw[black] (.3,-.3) circle (.05cm);
}=m^\dag
\qquad
\tikzmath{
\fill[\rColor, rounded corners=5pt] (0,0) rectangle (.6,.6);
\draw[black,thick] (.3,.3) -- (.3,.6);
\filldraw[black] (.3,.3) circle (.05cm);
}=i
\qquad
\tikzmath{
\fill[\rColor, rounded corners=5pt] (0,0) rectangle (.6,-.6);
\draw[black,thick] (.3,-.3) -- (.3,-.6);
\filldraw[black] (.3,-.3) circle (.05cm);
}=i^\dag.
$$
The axioms that $m,i$ must satisfy are associativity, unitality, the Frobenius relations, and unitary separability.
We refer the reader to \cite[\S3.1]{2105.12010} for a full discussion with many helpful diagrams.
We call a Q-system $Q\in \mathsf{Bim_{fgp}}(M)$ \emph{connected} if $\operatorname{Hom}_{M-M}(M\to Q) =\bbC i$.
\end{defn}
\begin{defn}[{\cite[\S4.1]{2105.12010}}]
For a Q-system $(Q,m,i)\in \mathsf{Bim}(M)$,
its \emph{realization}
$|Q|$ is the unital $*$-algebra
with underlying vector space is $\operatorname{Hom}_{\bbC - M}({}_{\bbC}M_M \to {}_{\bbC}M\boxtimes_M Q_M)$,
whose elements are denoted by
\[
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.3,-.7) rectangle (.6,.7);
\filldraw[\rColor] (0,-.7) -- (0,0) -- (-.15,0) -- (-.15,.7) -- (.7,.7) -- (.7,-.7);
\end{scope}
\draw[dashed] (0,-.7) -- (0,0);
\draw[dashed] (-.15,0) -- (-.15,.7);
\draw[black,thick] (.15,0) -- (.15,.7);
\roundNbox{unshaded}{(0,0)}{.3}{0}{0}{\scriptsize{$q$}};
}
\in |Q|
:=
\operatorname{Hom}_{\bbC - M}({}_{\bbC}M_M \to {}_{\bbC}M\boxtimes_M Q_M).
\]
The multiplication, unit, and adjoint, respectively, of $|Q|$ are given by
\[
q_1\cdot q_2:=
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-1,-.7) rectangle (.5,2);
\filldraw[\rColor] (-.2,-.7) -- (-.2,0) -- (-.4,.3) .. controls ++(90:.2cm) and ++(270:.2cm) .. (-.6,.7) -- (-.8,1) -- (-.8,2.8) -- (.9,2.8) -- (.9,-.7);
\end{scope}
\draw[dashed] (-.2,-.7) -- (-.2,0);
\draw[dashed] (-.4,.3) .. controls ++(90:.2cm) and ++(270:.2cm) .. (-.6,.7);
\draw[dashed] (-.8,1) -- (-.8,2);
\filldraw[\rColor] (-.2,1.8) circle (.05cm);
\draw[black,thick] (-.4,1.3) -- (-.4,1.5) arc (180:0:.2cm) -- (0,0);
\draw[black,thick] (-.2,1.7) -- (-.2,2);
\filldraw[black] (-.2,1.7) circle (.05cm);
\roundNbox{unshaded}{(-.6,1)}{.3}{.1}{.1}{\scriptsize{$q_1$}};
\roundNbox{unshaded}{(-.2,0)}{.3}{.1}{.1}{\scriptsize{$q_2$}};
}
\qquad\qquad
1_{|Q|}
:=
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.5,-.4) rectangle (.3,.4);
\filldraw[\rColor] (-.3,-.5) rectangle (.3,.5);
\end{scope}
\draw[dashed] (-.3,-.4) -- (-.3,.4);
\draw[black,thick] (0,0) -- (0,.4);
\filldraw[black] (0,0) circle (.05cm);
}
\qquad\qquad
q^*:=
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.5,-1.2) rectangle (.85,.7);
\filldraw[\rColor] (-.15,-1.2) -- (-.15,0) -- (0,0) -- (0,.7) -- (.85,.7) -- (.85,-1.2);
\end{scope}
\draw[dashed] (0,0) -- (0,.7);
\draw[dashed] (-.15,-1.2) -- (-.15,0);
\draw[black,thick] (.15,0) -- (.15,-.5) arc (-180:0:.2cm) -- (.55,.7);
\draw[black,thick] (.35,-1) -- (.35,-.7);
\filldraw[black] (.35,-.7) circle (.05cm);
\filldraw[black] (.35,-1) circle (.05cm);
\roundNbox{unshaded}{(0,0)}{.3}{0}{0}{\scriptsize{$q^\dag$}};
}
\]
Identifying $M=\operatorname{End}(M_M)$, the inclusion $M\hookrightarrow |Q|$ is given by
$$
\operatorname{End}({}_\bbC M_M)
\ni
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.3,-.7) rectangle (.6,.7);
\filldraw[\rColor] (0,-.7) rectangle (.6,.7);
\end{scope}
\draw[dashed] (0,-.7) -- (0,.7);
\roundNbox{unshaded}{(0,0)}{.3}{0}{0}{\scriptsize{$m$}};
}
\mapsto
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.3,-.7) rectangle (.6,.7);
\filldraw[\rColor] (0,-.7) rectangle (.6,.7);
\end{scope}
\draw[dashed] (0,-.7) -- (0,.7);
\draw[black, thick] (.2,.5) -- (.2,.7);
\filldraw[black] (.2,.5) circle (.05cm);
\roundNbox{unshaded}{(0,0)}{.3}{0}{0}{\scriptsize{$m$}};
}
\in |Q|
$$
By \cite[Rem.~4.4 and Prop.~4.6]{2105.12010}, $|Q|$ is a finite (Pimsner-Popa) index von Neumann algebra over $M$.
Moreover, $|Q|$ is a $\rm II_1$ factor if and only if $\operatorname{End}_{Q-Q}(Q)=\bbC\operatorname{id}_Q$.
In this case, the unique trace-preserving conditional expectation is given by
$$
E_M(q) =
\frac{1}{[|Q|:M]}
\cdot\,
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.3,-.7) rectangle (.6,.7);
\filldraw[\rColor] (0,-.7) -- (0,0) -- (-.15,0) -- (-.15,.7) -- (.7,.7) -- (.7,-.7);
\end{scope}
\draw[dashed] (0,-.7) -- (0,0);
\draw[dashed] (-.15,0) -- (-.15,.7);
\draw[black,thick] (.15,0) -- (.15,.5);
\filldraw[black] (.15,.5) circle (.05cm);
\roundNbox{unshaded}{(0,0)}{.3}{0}{0}{\scriptsize{$q$}};
}
$$
\end{defn}
\begin{ex}[Canonical Q-systems]
\label{ex:RealizeSeparatedQSystem}
Suppose ${}_NH_P \in \mathsf{W^*Alg_{fgp}}(N\to P)$.
Then $H\boxtimes_P \overline{H}$ is a Q-system in $\mathsf{Bim_{fgp}}(N)$.
By \cite[Prop.~3.1]{MR703809}, there is a canonical isomorphism $H\boxtimes_P \overline{H} \cong L^2(P^{\operatorname{op}})'\cap B(H)$, and thus $|H\boxtimes_P \overline{H}|\cong (P^{\operatorname{op}})'=:M$.
Observe that the relative commutant of $N \subseteq M$ is exactly given by
$N'\cap M = N'\cap (P^{\operatorname{op}})' = \operatorname{End}_{N-P}(H)$.
\end{ex}
\begin{rem}
\label{rem:AllQSystems}
Let $N$ be a $\rm II_1$ factor.
If $Q$ is a connected Q-system in $\mathsf{Bim_{fgp}}(N)$, then $N\subseteq |Q|$ is a finite index irreducible $\rm II_1$ subfactor.
By the $\rm W^*$ version of \cite[Prop.~4.16]{2105.12010}, $|Q|$ is also a connected Q-system in $\mathsf{Bim_{fgp}}(N)$, and $Q\cong |Q|$ as Q-systems.
Hence the connected Q-systems in $\mathsf{Bim_{fgp}}(N)$ are exactly $\rm II_1$ factors $M$ containing $N$ such that $N\subseteq M$ is finite index and irreducible.
\end{rem}
\begin{defn}[{\cite[\S4.2]{2105.12010}}]
\label{defn:RealizationOfBimodules}
Suppose $P,Q\in\mathsf{Bim_{fgp}}(M)$ are two Q-systems and $X$ is a $P-Q$ bimodule.
The \emph{realization} $|X|:=\operatorname{Hom}_{\bbC-M}({}_\bbC M_M \to {}_\bbC M \boxtimes_M X_M)$ of $X$ is a $|P|-|Q|$ bimodule whose elements are denoted by
$$
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.3,-.7) rectangle (.6,.7);
\filldraw[\rColor] (0,-.7) -- (0,0) -- (-.15,0) -- (-.15,.7) -- (.7,.7) -- (.7,-.7);
\end{scope}
\draw[dashed] (0,-.7) -- (0,0);
\draw[dashed] (-.15,0) -- (-.15,.7);
\draw[\XColor,thick] (.15,0) -- (.15,.7);
\roundNbox{unshaded}{(0,0)}{.3}{0}{0}{\scriptsize{$x$}};
}
\in |X|
:=
\operatorname{Hom}_{\bbC - M}({}_{\bbC}M_M \to {}_{\bbC}M\boxtimes_M X_M).
$$
with left $|P|$ and right $|Q|$-actions given respectively by
\[
p\rhd x
:=
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-1,-.7) rectangle (.5,2);
\filldraw[\rColor] (-.2,-.7) -- (-.2,0) -- (-.4,.3) .. controls ++(90:.2cm) and ++(270:.2cm) .. (-.6,.7) -- (-.8,1) -- (-.8,2.8) -- (.9,2.8) -- (.9,-.7);
\end{scope}
\draw[dashed] (-.2,-.7) -- (-.2,0);
\draw[dashed] (-.4,.3) .. controls ++(90:.2cm) and ++(270:.2cm) .. (-.6,.7);
\draw[dashed] (-.8,1) -- (-.8,2);
\filldraw[\rColor] (-.2,1.8) circle (.05cm);
\draw[black,thick] (-.4,1.3) arc (180:90:.4cm);
\draw[\XColor,thick] (0,0) -- (0,2);
\filldraw[\XColor] (0,1.7) circle (.05cm);
\roundNbox{unshaded}{(-.6,1)}{.3}{.1}{.1}{\scriptsize{$p$}};
\roundNbox{unshaded}{(-.2,0)}{.3}{.1}{.1}{\scriptsize{$x$}};
}
\qquad\qquad
x\lhd q:=
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-1,-.7) rectangle (.5,2);
\filldraw[\rColor] (-.2,-.7) -- (-.2,0) -- (-.4,.3) .. controls ++(90:.2cm) and ++(270:.2cm) .. (-.6,.7) -- (-.8,1) -- (-.8,2.8) -- (.9,2.8) -- (.9,-.7);
\end{scope}
\draw[dashed] (-.2,-.7) -- (-.2,0);
\draw[dashed] (-.4,.3) .. controls ++(90:.2cm) and ++(270:.2cm) .. (-.6,.7);
\draw[dashed] (-.8,1) -- (-.8,2);
\filldraw[\rColor] (-.2,1.8) circle (.05cm);
\draw[\XColor,thick] (-.4,1.3) -- (-.4,2);
\draw[black,thick] (0,0) -- (0,1.3) arc (0:90:.4cm);
\filldraw[\XColor] (-.4,1.7) circle (.05cm);
\roundNbox{unshaded}{(-.6,1)}{.3}{.1}{.1}{\scriptsize{$x$}};
\roundNbox{unshaded}{(-.2,0)}{.3}{.1}{.1}{\scriptsize{$q$}};
}
=x\lhd q
\qquad\qquad
\langle x_1|x_2\rangle_{|Q|}^X
:=
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.4,-1.4) rectangle (.9,1.4);
\filldraw[\rColor] (0,-1.4) -- (0,-.5) -- (-.2,-.5) -- (-.2,.5) -- (0,.5) -- (0,1.4) -- (1.1,1.4) -- (1.1,-1.4);
\end{scope}
\draw[dashed] (0,.5) -- (0,1.4);
\draw[dashed] (0,-1.4) -- (0,-.5);
\draw[dashed] (-.2,-.5) -- (-.2,.5);
\draw[\XColor,thick] (.2,-.5) -- (.2,.5);
\draw[black,thick] (.2,0) arc (-90:0:.4cm) -- (.6,1.4);
\filldraw[\XColor] (.2,0) circle (.05cm);
\roundNbox{unshaded}{(0,.7)}{.3}{.05}{.05}{\scriptsize{$x_1^\dag$}};
\roundNbox{unshaded}{(0,-.7)}{.3}{.05}{.05}{\scriptsize{$x_2$}};
}
\]
Clearly $|X|$ has a predual as it is a hom space in a $\rm W^*$-category.
By \cite[Lem.~2.3 and Prop.~2.4]{2105.12010}, the right $|Q|$-valued inner product is separately weak*-continuous, and the left action of $|P|$ on $X$ is normal.
Hence $|X|\in \mathsf{W^*Alg_{fgp}}(|P|\to |Q|)$ is a $\rm W^*$-correspondence.
\end{defn}
\begin{nota}
Given another Q-system $R\in \mathsf{Bim_{fgp}}(M)$ and a $Q-R$ bimodule $Y$,
there is a notion of the relative tensor product of $X$ with $Y$ over $Q$, denoted $X\otimes_Q Y$.
We refer the reader to \cite[\S3.2]{2105.12010} for the detailed definition.
We have two canonical projectors which we denote graphically as follows:
\begin{equation}
\label{eq:CanonicalProjectors}
\tikzmath{
\draw[thin, dotted, rounded corners = 5] (-.2,0) -- (-.5,0) -- (-.5,.8) -- (-.2,.8);
\draw[thin, dotted, rounded corners = 5] (.2,0) -- (.5,0) -- (.5,.8) -- (.2,.8);
\filldraw[\rColor] (-.2,0) rectangle (.2,.4);
\filldraw[\rColor] (-.2,.4) rectangle (.2,.8);
\draw[\XColor,thick] (-.2,0) -- (-.2,.8);
\draw[blue,thick] (.2,0) -- (.2,.8);
\draw[black, thick] (-.2,.4) -- (.2,.4);
}
:X\boxtimes_M Y \to X\otimes_Q Y
\qquad\text{and}\qquad
\tikzmath{
\draw[thin, dotted, rounded corners = 5] (-.2,0) -- (-.5,0) -- (-.5,.8) -- (-.2,.8);
\draw[thin, dotted, rounded corners = 5] (.2,0) -- (.5,0) -- (.5,.8) -- (.2,.8);
\filldraw[\rColor] (-.2,0) rectangle (.2,.4);
\filldraw[\QrColor] (-.2,.4) rectangle (.2,.8);
\draw[\XColor,thick] (-.2,0) -- (-.2,.8);
\draw[blue,thick] (.2,0) -- (.2,.8);
\draw[black, thick] (-.2,.4) -- (.2,.4);
}
:|X\boxtimes_M Y| \to |X|\boxtimes_{|Q|} |Y|;
\qquad
\tikzmath{
\fill[\QrColor, rounded corners=5] (0,0) rectangle (.6,.6);
}
=
|Q|.
\end{equation}
For the second diagram, we omit the external shadings, which may denote either a left/right $M$-action, or a left $|P|$ and right $|R|$-action depending on context.
\end{nota}
\begin{rem}
\label{rem:QBimodules}
By the $\rm W^*$ version of \cite[Thm.~A]{2105.12010},
realization $|\,\cdot\,|$ gives a dagger 2-equivalence from the Q-system completion of $\mathsf{W^*Alg}$ to $\mathsf{W^*Alg}$.
Thus $|\,\cdot\,|$ gives an equivalence from the unitary tensor category of
$Q-Q$ bimodules in $\mathsf{Bim_{fgp}}(M)$
with the tensor product $\otimes_Q$
to $\mathsf{Bim_{fgp}}(|Q|)$ with the Connes fusion relative tensor product $\boxtimes_Q$.
Moreover, the canonical tensorator $\mu_{X,Y}:|X|\boxtimes_{|Q|} |Y| \to |X\otimes_Q Y|$ fits into a commuting diagram with the canonical projectors \eqref{eq:CanonicalProjectors}:
\begin{equation}
\label{eq:CanonicalProjectorTriangle}
\tikzmath{
\node (a) at (0,0) {$|X\boxtimes_M Y|$};
\node (b) at (4,1) {$|X|\boxtimes_{|Q|}|Y|$};
\node (c) at (4,-1) {$|X\otimes_QY|$};
\draw[->] (a) -- (b);
\draw[->] (a) -- (c);
\draw[->] (b) -- node[right]{$\scriptstyle \mu_{X,Y}$} (c);
\coordinate (a) at (1.2,-1.2);
\draw ($ (a) + (-.05,0) $) -- ($ (a) + (-.05,.6) $) ;
\draw ($ (a) + (.85,0) $) -- ($ (a) + (.85,.6) $) ;
\draw[dotted, thin, rounded corners=5] ($ (a) + (.2,0) $) -- (a) -- ($ (a) + (0,.6) $) -- ($ (a) + (.2,.6) $);
\draw[dotted, thin, rounded corners=5] ($ (a) + (.6,0) $) -- ($ (a) + (.8,0) $) -- ($ (a) + (.8,.6) $) -- ($ (a) + (.6,.6) $);
\filldraw[\rColor] ($ (a) + (.2,0) $) rectangle ($ (a) + (.6,.6)$);
\draw[\XColor,thick] ($ (a) + (.2,0) $) -- ($ (a) + (.2,.6) $);
\draw[blue,thick] ($ (a) + (.6,0) $) -- ($ (a) + (.6,.6) $);
\draw[black, thick] ($ (a) + (.2,.3) $) -- ($ (a) + (.6,.3) $);
\coordinate (b) at (1.2,.6);
\draw[dotted, thin, rounded corners=5] ($ (b) + (.2,0) $) -- (b) -- ($ (b) + (0,.6) $) -- ($ (b) + (.2,.6) $);
\draw[dotted, thin, rounded corners=5] ($ (b) + (.6,0) $) -- ($ (b) + (.8,0) $) -- ($ (b) + (.8,.6) $) -- ($ (b) + (.6,.6) $);
\filldraw[\rColor] ($ (b) + (.2,.3) $) rectangle ($ (b) + (.6,0)$);
\filldraw[\QrColor] ($ (b) + (.2,.3) $) rectangle ($ (b) + (.6,.6)$);
\draw[\XColor,thick] ($ (b) + (.2,0) $) -- ($ (b) + (.2,.6) $);
\draw[blue,thick] ($ (b) + (.6,0) $) -- ($ (b) + (.6,.6) $);
\draw[black, thick] ($ (b) + (.2,.3) $) -- ($ (b) + (.6,.3) $);
}
\end{equation}
\end{rem}
\subsection{Local extension}
\label{sec:LocalExtension}
We now turn to the proof that $\tilde{\chi}(|Q|) = \tilde{\chi}(M)_Q^{\operatorname{loc}}$.
\begin{defn}
A Q-system $Q$ in a unitary braided tensor category $\cC$ is called \textit{commutative} if
$$
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.3,-.5) rectangle (.7,1);
\filldraw[\rColor] (-.4,-.5) -- (-.4,1) -- (.7,1) -- (.7,-.5);
\end{scope}
\draw[black,thick] (0,-.5) .. controls ++(90:.45cm) and ++(270:.45cm) .. (.4,.5);
\draw[black,thick] (0,.5) arc (180:0:.2cm);
\draw[black,thick] (.2,.7) -- (.2,1);
\filldraw[\rColor] (.2,0) circle (.05cm);
\filldraw[black] (.2,.7) circle (.05cm);
\draw[black,thick] (.4,-.5) .. controls ++(90:.45cm) and ++(270:.45cm) .. (0,.5);
}
=
\tikzmath{
\fill[\rColor, rounded corners=5pt] (-.3,0) rectangle (.9,.6);
\draw[black,thick] (0,0) arc (180:0:.3cm);
\draw[black,thick] (.3,.3) -- (.3,.6);
\filldraw[black] (.3,.3) circle (.05cm);
}
=
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.3,-.5) rectangle (.7,1);
\filldraw[\rColor] (-.4,-.5) -- (-.4,1) -- (.7,1) -- (.7,-.5);
\end{scope}
\draw[black,thick] (.4,-.5) .. controls ++(90:.45cm) and ++(270:.45cm) .. (0,.5);
\draw[black,thick] (0,.5) arc (180:0:.2cm);
\draw[black,thick] (.2,.7) -- (.2,1);
\filldraw[\rColor] (.2,0) circle (.05cm);
\filldraw[black] (.2,.7) circle (.05cm);
\draw[black,thick] (0,-.5) .. controls ++(90:.45cm) and ++(270:.45cm) .. (.4,.5);
}\,.
$$
Suppose $Q\in\cC$ is a commutative connected Q-system and let $X\in\cC$ be a right $Q$-module. We say $X$ is \textit{local} if
\[
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.3,-.5) rectangle (.7,1.3);
\filldraw[\rColor] (-.4,-.5) -- (-.4,1.3) -- (.7,1.3) -- (.7,-.5);
\end{scope}
\draw[black,thick] (0,-.5) .. controls ++(90:.45cm) and ++(270:.45cm) .. (.4,.5) arc (0:90:.4cm);
\filldraw[\rColor] (.2,0) circle (.05cm);
\draw[\XColor,thick] (.4,-.5) .. controls ++(90:.45cm) and ++(270:.45cm) .. (0,.5) -- (0,1.3);
\filldraw[\XColor] (0,.9) circle (.05cm);
}
=
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.3,-.5) rectangle (.7,1.3);
\filldraw[\rColor] (-.4,-.5) -- (-.4,1.3) -- (.7,1.3) -- (.7,-.5);
\end{scope}
\draw[\XColor,thick] (.4,-.5) .. controls ++(90:.45cm) and ++(270:.45cm) .. (0,.5) -- (0,1.3);
\filldraw[\rColor] (.2,0) circle (.05cm);
\draw[black,thick] (0,-.5) .. controls ++(90:.45cm) and ++(270:.45cm) .. (.4,.5) arc (0:90:.4cm);
\filldraw[\XColor] (0,.9) circle (.05cm);
}
\]
We can turn a local $Q$-module $X_Q$ into a $Q-Q$ bimodule by defining the left action map by
\[
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.7,0) rectangle (.3,1);
\filldraw[\rColor] (-.7,-.5) -- (-.7,1.3) -- (.7,1.3) -- (.7,-.5);
\end{scope}
\draw[\XColor,thick] (0,0) -- (0,1);
\draw[black,thick] (-.4,0) -- (-.4,.2) arc (180:90:.4cm);
\filldraw[\XColor] (0,.6) circle (.05cm);
}
: =
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.3,-.5) rectangle (.7,1.3);
\filldraw[\rColor] (-.4,-.5) -- (-.4,1.3) -- (.7,1.3) -- (.7,-.5);
\end{scope}
\draw[black,thick] (0,-.5) .. controls ++(90:.45cm) and ++(270:.45cm) .. (.4,.5) arc (0:90:.4cm);
\filldraw[\rColor] (.2,0) circle (.05cm);
\draw[\XColor,thick] (.4,-.5) .. controls ++(90:.45cm) and ++(270:.45cm) .. (0,.5) -- (0,1.3);
\filldraw[\XColor] (0,.9) circle (.05cm);
}
=
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.3,-.5) rectangle (.7,1.3);
\filldraw[\rColor] (-.4,-.5) -- (-.4,1.3) -- (.7,1.3) -- (.7,-.5);
\end{scope}
\draw[\XColor,thick] (.4,-.5) .. controls ++(90:.45cm) and ++(270:.45cm) .. (0,.5) -- (0,1.3);
\filldraw[\rColor] (.2,0) circle (.05cm);
\draw[black,thick] (0,-.5) .. controls ++(90:.45cm) and ++(270:.45cm) .. (.4,.5) arc (0:90:.4cm);
\filldraw[\XColor] (0,.9) circle (.05cm);
}\,.
\]
It is well known that under the bimodule relative tensor product $\otimes_Q$,
the collection of local $Q$-modules $\cC^{\operatorname{loc}}_{Q}$ is a unitary braided tensor category, where the braiding is inherited from $\cC$.
\end{defn}
Our goal now is to prove
$\tilde{\chi}(|Q|)\cong\tilde{\chi}(M)_Q^{\operatorname{loc}}$ as braided unitary tensor categories.
\begin{lem}
\label{lem:CSinMCSin|Q|}
If $(a_n)_n\subseteq M$ is a central sequence, then $(a_n)_n$ is also a central sequence in $|Q|$.
\end{lem}
\begin{proof}
Since $Q$ is centrally trivial, for all $q\in |Q|$, $\|a_n q-qa_n\|_2\to 0$ by Proposition \ref{prop:BimodualCT&CS}.
\end{proof}
\begin{prop}
\label{prop:APPBofQisCSinQ}
A Q-system $Q\in \tilde{\chi}(M)$ is commutative if and only if
for every approximate inner $Q_M$-basis $\{q_i^{(n)}\}$,
$(q_i^{(n)})_n$ is a central sequence in $|Q|$ for
each fixed $i$.
\end{prop}
\begin{proof}
Suppose $\{q_j\}_j$ is a $Q_M$-basis and $\{q_i^{(n)}\}_i$ is an approximate $Q_M$-basis.
By \eqref{eq:OverBraiding} and \eqref{eq:UnderBraiding},
\begin{equation}
\label{eq:ExpandBraidingsQ}
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.55,-.5) rectangle (.7,.5);
\filldraw[\rColor] (-.4,-.5) -- (-.4,.5) -- (.7,.5) -- (.7,-.5);
\end{scope}
\draw[dashed] (-.4,-.5) -- (-.4,.5);
\draw[black,thick] (.4,-.5) .. controls ++(90:.45cm) and ++(270:.45cm) .. (0,.5);
\filldraw[\rColor] (.2,0) circle (.05cm);
\draw[black,thick] (0,-.5) .. controls ++(90:.45cm) and ++(270:.45cm) .. (.4,.5);
}
\approx
\sum_{i,j}
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.35,-2.2) rectangle (1,2.2);
\filldraw[\rColor] (-.2,-2.2) -- (-.2,-1.5) -- (0,-1.5) -- (0,-.5) -- (.3,-.5) -- (.3,.5) -- (0,.5) -- (0,1.5) -- (-.2,1.5) -- (-.2,2.2) -- (1.05,2.2) -- (1.05,-2.2);
\end{scope}
\draw[dashed] (-.2,-2.2) -- (-.2,-1.5);
\draw[black,thick] (.2,-2.2) -- (.2,-1.5);
\draw[dashed] (0,-1.5) -- (0,-.5);
\draw[black,thick] (.6,-2.2) -- (.6,-.5);
\draw[dashed] (.3,-.5) -- (.3,.5);
\draw[dashed] (0,.5) -- (0,1.5);
\draw[black,thick] (.6,.5) -- (.6,2.2);
\draw[dashed] (-.2,1.5) -- (-.2,2.2);
\draw[black,thick] (.2,1.5) -- (.2,2.2);
\roundNbox{unshaded}{(0,1.5)}{.3}{.1}{.1}{\scriptsize{$q_j$}};
\roundNbox{unshaded}{(.3,.5)}{.3}{.15}{.15}{\scriptsize{$q_i^{(n)}$}};
\roundNbox{unshaded}{(.3,-.5)}{.3}{.15}{.15}{\scriptsize{$q_j^\dag$}};
\roundNbox{unshaded}{(0,-1.5)}{.3}{.1}{.1}{\scriptsize{$(q_i^{(\!n\!)}\!)^\dag$}};
}
\qquad\text{and}\qquad
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.55,-.5) rectangle (.7,.5);
\filldraw[\rColor] (-.4,-.5) -- (-.4,.5) -- (.7,.5) -- (.7,-.5);
\end{scope}
\draw[dashed] (-.4,-.5) -- (-.4,.5);
\draw[black,thick] (0,-.5) .. controls ++(90:.45cm) and ++(270:.45cm) .. (.4,.5);
\filldraw[\rColor] (.2,0) circle (.05cm);
\draw[black,thick] (.4,-.5) .. controls ++(90:.45cm) and ++(270:.45cm) .. (0,.5);
}
\approx
\sum_{i,j}
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.35,-2.2) rectangle (1,2.2);
\filldraw[\rColor] (-.2,-2.2) -- (-.2,-1.5) -- (0,-1.5) -- (0,-.5) -- (.3,-.5) -- (.3,.5) -- (0,.5) -- (0,1.5) -- (-.2,1.5) -- (-.2,2.2) -- (1.05,2.2) -- (1.05,-2.2);
\end{scope}
\draw[dashed] (-.2,-2.2) -- (-.2,-1.5);
\draw[black,thick] (.2,-2.2) -- (.2,-1.5);
\draw[dashed] (0,-1.5) -- (0,-.5);
\draw[black,thick] (.6,-2.2) -- (.6,-.5);
\draw[dashed] (.3,-.5) -- (.3,.5);
\draw[dashed] (0,.5) -- (0,1.5);
\draw[black,thick] (.6,.5) -- (.6,2.2);
\draw[dashed] (-.2,1.5) -- (-.2,2.2);
\draw[black,thick] (.2,1.5) -- (.2,2.2);
\roundNbox{unshaded}{(0,1.5)}{.3}{.1}{.1}{\scriptsize{$q_j^{(n)}$}};
\roundNbox{unshaded}{(.3,.5)}{.3}{.15}{.15}{\scriptsize{$q_i$}};
\roundNbox{unshaded}{(.3,-.5)}{.3}{.15}{.15}{\scriptsize{$(q_j^{(\!n\!)}\!)^\dag$}};
\roundNbox{unshaded}{(0,-1.5)}{.3}{.1}{.1}{\scriptsize{$q_i^\dag$}};
}\,.
\end{equation}
First, if each $q_i^{(n)}$ is a central sequence, it is straightforward to see that $Q$ is commutative using \eqref{eq:ExpandBraidingsQ}.
Conversely, if $Q$ is commutative, to show each $q_i^{(n)}$ is a central sequence,
since $\|q_i^{(n)}a-aq_i^{(n)}\|_2\to 0$ for each $a\in M$, it suffices to prove $\|q_i^{(n)}q_j-q_jq_i^{(n)}\|_2\to 0$ for each $j$.
For each $i,j$ and $n$, define
\[
x_n^{i,j}:=q_i^{(n)}q_j-q_jq_i^{(n)}=
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.45,-1.2) rectangle (1.05,1.6);
\filldraw[\rColor] (.3,-1.2) -- (.3,-.5) -- (0,-.5) -- (0,.5) -- (-.2,.5) -- (-.2,2) -- (1.05,2) -- (1.05,-1.2);
\end{scope}
\draw[dashed] (-.2,.5) -- (-.2,1.6);
\draw[dashed] (0,-.5) -- (0,.5);
\draw[dashed] (.3,-1.2) -- (.3,-.5);
\draw[black,thick] (.2,.8) -- (.2,1) arc (180:0:.2cm) -- (.6,-.5);
\draw[black,thick] (.4,1.2) -- (.4,1.6);
\filldraw[black] (.4,1.2) circle (.05cm);
\roundNbox{unshaded}{(0,.5)}{.3}{.1}{.1}{\scriptsize{$q_i^{(n)}$}};
\roundNbox{unshaded}{(.3,-.5)}{.3}{.15}{.15}{\scriptsize{$q_j$}};
}
-
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.45,-1.2) rectangle (1.05,1.6);
\filldraw[\rColor] (.3,-1.2) -- (.3,-.5) -- (0,-.5) -- (0,.5) -- (-.2,.5) -- (-.2,2) -- (1.05,2) -- (1.05,-1.2);
\end{scope}
\draw[dashed] (-.2,.5) -- (-.2,1.6);
\draw[dashed] (0,-.5) -- (0,.5);
\draw[dashed] (.3,-1.2) -- (.3,-.5);
\draw[black,thick] (.2,.8) -- (.2,1) arc (180:0:.2cm) -- (.6,-.5);
\draw[black,thick] (.4,1.2) -- (.4,1.6);
\filldraw[black] (.4,1.2) circle (.05cm);
\roundNbox{unshaded}{(0,.5)}{.3}{.1}{.1}{\scriptsize{$q_j$}};
\roundNbox{unshaded}{(.3,-.5)}{.3}{.15}{.15}{\scriptsize{$q_i^{(n)}$}};
}
\in\,
\operatorname{Hom}(M_M \to Q_M).
\]
Observe that in $\operatorname{End}(Q_M)$, then again by \eqref{eq:ExpandBraidingsQ} we have
\begin{align*}
\sum_{i,j} x_n^{i,j} (x_n^{i,j})^\dag
&=
\sum_{i,j}
\left(
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.35,-2.5) rectangle (1,2.5);
\filldraw[\rColor] (-.2,-2.5) -- (-.2,-1.5) -- (0,-1.5) -- (0,-.5) -- (.3,-.5) -- (.3,.5) -- (0,.5) -- (0,1.5) -- (-.2,1.5) -- (-.2,2.5) -- (1.05,2.5) -- (1.05,-2.5);
\end{scope}
\draw[dashed] (-.2,-2.5) -- (-.2,-1.5);
\draw[black,thick] (.2,-1.5) -- (.2,-2) arc (-180:0:.2cm) -- (.6,-.5);
\draw[dashed] (0,-1.5) -- (0,-.5);
\draw[black,thick] (.4,-2.5) -- (.4,-2.2);
\draw[dashed] (.3,-.5) -- (.3,.5);
\draw[dashed] (0,.5) -- (0,1.5);
\draw[black,thick] (.4,2.2) -- (.4,2.5);
\draw[dashed] (-.2,1.5) -- (-.2,2.5);
\draw[black,thick] (.2,1.5) -- (.2,2) arc (180:0:.2cm) -- (.6,.5);
\filldraw[black] (.4,-2.2) circle (.05cm);
\filldraw[black] (.4,2.2) circle (.05cm);
\roundNbox{unshaded}{(0,1.5)}{.3}{.1}{.1}{\scriptsize{$q_i^{(n)}$}};
\roundNbox{unshaded}{(.3,.5)}{.3}{.15}{.15}{\scriptsize{$q_j$}};
\roundNbox{unshaded}{(.3,-.5)}{.3}{.15}{.15}{\scriptsize{$q_j^\dag$}};
\roundNbox{unshaded}{(0,-1.5)}{.3}{.1}{.1}{\scriptsize{$(q_i^{(\!n\!)}\!)^\dag$}};
}
-
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.35,-2.5) rectangle (1,2.5);
\filldraw[\rColor] (-.2,-2.5) -- (-.2,-1.5) -- (0,-1.5) -- (0,-.5) -- (.3,-.5) -- (.3,.5) -- (0,.5) -- (0,1.5) -- (-.2,1.5) -- (-.2,2.5) -- (1.05,2.5) -- (1.05,-2.5);
\end{scope}
\draw[dashed] (-.2,-2.5) -- (-.2,-1.5);
\draw[black,thick] (.2,-1.5) -- (.2,-2) arc (-180:0:.2cm) -- (.6,-.5);
\draw[dashed] (0,-1.5) -- (0,-.5);
\draw[black,thick] (.4,-2.5) -- (.4,-2.2);
\draw[dashed] (.3,-.5) -- (.3,.5);
\draw[dashed] (0,.5) -- (0,1.5);
\draw[black,thick] (.4,2.2) -- (.4,2.5);
\draw[dashed] (-.2,1.5) -- (-.2,2.5);
\draw[black,thick] (.2,1.5) -- (.2,2) arc (180:0:.2cm) -- (.6,.5);
\filldraw[black] (.4,-2.2) circle (.05cm);
\filldraw[black] (.4,2.2) circle (.05cm);
\roundNbox{unshaded}{(0,1.5)}{.3}{.1}{.1}{\scriptsize{$q_i^{(n)}$}};
\roundNbox{unshaded}{(.3,.5)}{.3}{.15}{.15}{\scriptsize{$q_j$}};
\roundNbox{unshaded}{(.3,-.5)}{.3}{.15}{.15}{\scriptsize{$(q_i^{(\!n\!)})^\dag$}};
\roundNbox{unshaded}{(0,-1.5)}{.3}{.1}{.1}{\scriptsize{$q_j^\dag$}};
}
-
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.35,-2.5) rectangle (1,2.5);
\filldraw[\rColor] (-.2,-2.5) -- (-.2,-1.5) -- (0,-1.5) -- (0,-.5) -- (.3,-.5) -- (.3,.5) -- (0,.5) -- (0,1.5) -- (-.2,1.5) -- (-.2,2.5) -- (1.05,2.5) -- (1.05,-2.5);
\end{scope}
\draw[dashed] (-.2,-2.5) -- (-.2,-1.5);
\draw[black,thick] (.2,-1.5) -- (.2,-2) arc (-180:0:.2cm) -- (.6,-.5);
\draw[dashed] (0,-1.5) -- (0,-.5);
\draw[black,thick] (.4,-2.5) -- (.4,-2.2);
\draw[dashed] (.3,-.5) -- (.3,.5);
\draw[dashed] (0,.5) -- (0,1.5);
\draw[black,thick] (.4,2.2) -- (.4,2.5);
\draw[dashed] (-.2,1.5) -- (-.2,2.5);
\draw[black,thick] (.2,1.5) -- (.2,2) arc (180:0:.2cm) -- (.6,.5);
\filldraw[black] (.4,-2.2) circle (.05cm);
\filldraw[black] (.4,2.2) circle (.05cm);
\roundNbox{unshaded}{(0,1.5)}{.3}{.1}{.1}{\scriptsize{$q_j$}};
\roundNbox{unshaded}{(.3,.5)}{.3}{.15}{.15}{\scriptsize{$q_i^{(n)}$}};
\roundNbox{unshaded}{(.3,-.5)}{.3}{.15}{.15}{\scriptsize{$q_j^\dag$}};
\roundNbox{unshaded}{(0,-1.5)}{.3}{.1}{.1}{\scriptsize{$(q_i^{(\!n\!)}\!)^\dag$}};
}
+
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.35,-2.5) rectangle (1,2.5);
\filldraw[\rColor] (-.2,-2.5) -- (-.2,-1.5) -- (0,-1.5) -- (0,-.5) -- (.3,-.5) -- (.3,.5) -- (0,.5) -- (0,1.5) -- (-.2,1.5) -- (-.2,2.5) -- (1.05,2.5) -- (1.05,-2.5);
\end{scope}
\draw[dashed] (-.2,-2.5) -- (-.2,-1.5);
\draw[black,thick] (.2,-1.5) -- (.2,-2) arc (-180:0:.2cm) -- (.6,-.5);
\draw[dashed] (0,-1.5) -- (0,-.5);
\draw[black,thick] (.4,-2.5) -- (.4,-2.2);
\draw[dashed] (.3,-.5) -- (.3,.5);
\draw[dashed] (0,.5) -- (0,1.5);
\draw[black,thick] (.4,2.2) -- (.4,2.5);
\draw[dashed] (-.2,1.5) -- (-.2,2.5);
\draw[black,thick] (.2,1.5) -- (.2,2) arc (180:0:.2cm) -- (.6,.5);
\filldraw[black] (.4,-2.2) circle (.05cm);
\filldraw[black] (.4,2.2) circle (.05cm);
\roundNbox{unshaded}{(0,1.5)}{.3}{.1}{.1}{\scriptsize{$q_j$}};
\roundNbox{unshaded}{(.3,.5)}{.3}{.15}{.15}{\scriptsize{$q_i^{(n)}$}};
\roundNbox{unshaded}{(.3,-.5)}{.3}{.15}{.15}{\scriptsize{$(q_i^{(\!n\!)})^\dag$}};
\roundNbox{unshaded}{(0,-1.5)}{.3}{.1}{.1}{\scriptsize{$q_j^\dag$}};;
}
\right)
\displaybreak[1]\\
&\approx
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.8,-.6) rectangle (.6,.6);
\filldraw[\rColor] (-.6,-.6) rectangle (.6,.6);
\end{scope}
\draw[dashed] (-.6,-.6) -- (-.6,.6);
\draw[black,thick] (0,-.6) -- (0,-.3) arc (-90:270:.3cm);
\draw[black,thick] (0,.3) -- (0,.6);
\filldraw[black] (0,-.3) circle (.05cm);
\filldraw[black] (0,.3) circle (.05cm);
}
-
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.5,-1) rectangle (.7,1);
\filldraw[\rColor] (-.3,-1) rectangle (.7,1);
\end{scope}
\draw[dashed] (-.3,-1) -- (-.3,1);
\draw[black,thick] (0,-.5) .. controls ++(90:.45cm) and ++(270:.45cm) .. (.4,.5);
\draw[black,thick] (0,.5) arc (180:0:.2cm);
\draw[black,thick] (.2,.7) -- (.2,1);
\draw[black,thick] (0,-.5) arc (-180:0:.2cm);
\draw[black,thick] (.2,-.7) -- (.2,-1);
\filldraw[\rColor] (.2,0) circle (.05cm);
\draw[black,thick] (.4,-.5) .. controls ++(90:.45cm) and ++(270:.45cm) .. (0,.5);
\filldraw[black] (.2,.7) circle (.05cm);
\filldraw[black] (.2,-.7) circle (.05cm);
}
-
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.5,-1) rectangle (.7,1);
\filldraw[\rColor] (-.3,-1) rectangle (.7,1);
\end{scope}
\draw[dashed] (-.3,-1) -- (-.3,1);
\draw[black,thick] (.4,-.5) .. controls ++(90:.45cm) and ++(270:.45cm) .. (0,.5);
\draw[black,thick] (0,.5) arc (180:0:.2cm);
\draw[black,thick] (.2,.7) -- (.2,1);
\draw[black,thick] (0,-.5) arc (-180:0:.2cm);
\draw[black,thick] (.2,-.7) -- (.2,-1);
\filldraw[\rColor] (.2,0) circle (.05cm);
\draw[black,thick] (0,-.5) .. controls ++(90:.45cm) and ++(270:.45cm) .. (.4,.5);
\filldraw[black] (.2,.7) circle (.05cm);
\filldraw[black] (.2,-.7) circle (.05cm);
}
+
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.8,-.6) rectangle (.6,.6);
\filldraw[\rColor] (-.6,-.6) rectangle (.6,.6);
\end{scope}
\draw[dashed] (-.6,-.6) -- (-.6,.6);
\draw[black,thick] (0,-.6) -- (0,-.3) arc (-90:270:.3cm);
\draw[black,thick] (0,.3) -- (0,.6);
\filldraw[black] (0,-.3) circle (.05cm);
\filldraw[black] (0,.3) circle (.05cm);
}
= 0.
\end{align*}
But since $\sum_{i,j} x_n^{i,j} (x_n^{i,j})^\dag$ is positive, we must have
$\|\cdot\|_2-\lim_n x_n^{i,j} (x_n^{i,j})^\dag=0$ for each $i,j$.
Since $\mathsf{W^*Alg}$ is $\rm W^*$, we must have $\lim_n x_n^{i,j} = 0$, so each $q_i^{(n)}$ is a central sequence.
\end{proof}
\begin{lem}
\label{lem:APPAfromMto|Q|}
Suppose $X\in \mathsf{Bim_{fgp}}(M)$ is a right $Q$-module.
If $\{b_i^{(n)}\}$ is an approximate $X_M$-basis, then it is also an approximate $|X|_{|Q|}$-basis.
Similarly, if $\{b_i\}$ is an $X_M$-basis, then it is also an $|X|_{|Q|}$-basis.
\end{lem}
\begin{proof}
We prove the approximate version, and the ordinary version is similar, but easier.
Identifying $X_M\cong |X|_M$ as right $M$-modules, observe
\[
\sum_i b_i^{(n)}\langle b_i^{(n)}|x\rangle^X_M
=
\sum_i
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.35,-2.2) rectangle (.7,1.2);
\filldraw[\rColor] (0,-2.2) -- (0,-1.5) -- (-.2,-1.55) -- (-.2,-.5) -- (0,-.5) -- (0,.5) -- (-.2,.5) -- (-.2,1.7) -- (1.3,1.7) -- (1.3,-2.55);
\end{scope}
\draw[dashed] (0,-2.2) -- (0,-1.5);
\draw[dashed] (-.2,-1.55) -- (-.2,-.5);
\draw[dashed] (0,-.5) -- (0,.5);
\draw[dashed] (-.2,.5) -- (-.2,1.2);
\draw[\XColor,thick] (.2,-1.55) -- (.2,-.5);
\draw[\XColor,thick] (.2,.5) -- (.2,1.2);
\roundNbox{unshaded}{(0,.5)}{.3}{.1}{.1}{\scriptsize{$b_i^{(n)}$}};
\roundNbox{unshaded}{(0,-.5)}{.3}{.1}{.1}{\scriptsize{$(b_i^{(\!n\!)}\!)^\dag$}};
\roundNbox{unshaded}{(0,-1.5)}{.3}{.1}{.1}{\scriptsize{$x$}};
}
\approx x
\qquad\Longrightarrow\qquad
\sum_i b_i^{(n)}\langle b_i^{(n)}|x\rangle^{|X|}_{|Q|} =
\sum_i
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.35,-2.6) rectangle (1,1.6);
\filldraw[\rColor] (0,-2.6) -- (0,-1.9) -- (-.2,-1.9) -- (-.2,-.5) -- (0,-.5) -- (0,.5) -- (-.2,.5) -- (-.2,1.7) -- (1.3,1.7) -- (1.3,-2.6);
\end{scope}
\draw[dashed] (0,-2.6) -- (0,-1.9);
\draw[dashed] (-.2,-1.9) -- (-.2,-.5);
\draw[dashed] (0,-.5) -- (0,.5);
\draw[dashed] (-.2,.5) -- (-.2,1.6);
\draw[\XColor,thick] (.2,-1.9) -- (.2,-.5);
\draw[\XColor,thick] (.2,.5) -- (.2,1.6);
\draw[black,thick] (.2,-1.2) arc (-90:0:.4cm) -- (.6,.8) arc (0:90:.4cm);
\filldraw[\XColor] (.2,-1.2) circle (.05cm);
\filldraw[\XColor] (.2,1.2) circle (.05cm);
\roundNbox{unshaded}{(0,.5)}{.3}{.1}{.1}{\scriptsize{$b_i^{(n)}$}};
\roundNbox{unshaded}{(0,-.5)}{.3}{.1}{.1}{\scriptsize{$(b_i^{(\!n\!)}\!)^\dag$}};
\roundNbox{unshaded}{(0,-1.9)}{.3}{.1}{.1}{\scriptsize{$x$}};
}
\approx x.
\qedhere
\]
\end{proof}
\begin{prop}
\label{Chi(M)locQsubsetChi(Q)}
Realization $|\,\cdot\,|$ takes every bimodule in $\tilde{\chi}(M)_Q^\operatorname{loc}$ into $\tilde{\chi}(|Q|)$.
\end{prop}
\begin{proof}
Suppose $X\in \tilde{\chi}(M)_Q^{\operatorname{loc}}$.
We show ${}_{|Q|}|X|_{|Q|}$ is approximately inner and centrally trivial.
Since $X$ is approximately inner over $M$,
by Proposition \ref{prop:AI&APPB},
there is an approximately inner $X_M$-basis
$\{b_i^{(n)}\}\subseteq X$.
By Lemma \ref{lem:APPAfromMto|Q|}, $\{b_i^{(n)}\}$ is an approximate $|X|_{|Q|}$-basis.
Now we show \eqref{eq:AICommutativity}, i.e., $\|qb_i^{(n)}-b_i^{(n)}q\|_2\to 0$ for all $q\in |Q|$.
\[
q b_i^{(n)}
=
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.45,-1.2) rectangle (1.05,1.6);
\filldraw[\rColor] (.3,-1.2) -- (.3,-.5) -- (0,-.5) -- (0,.5) -- (-.2,.5) -- (-.2,2) -- (1.05,2) -- (1.05,-1.2);
\end{scope}
\draw[dashed] (-.2,.5) -- (-.2,1.6);
\draw[dashed] (0,-.5) -- (0,.5);
\draw[dashed] (.3,-1.2) -- (.3,-.5);
\draw[black,thick] (.2,.8) arc (180:90:.4cm);
\draw[\XColor,thick] (.6,-.5) -- (.6,1.6);
\filldraw[\XColor] (.6,1.2) circle (.05cm);
\roundNbox{unshaded}{(0,.5)}{.3}{.1}{.1}{\scriptsize{$q$}};
\roundNbox{unshaded}{(.3,-.5)}{.3}{.15}{.15}{\scriptsize{$b_i^{(n)}$}};
}
=
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-1,-.7) rectangle (.5,3.1);
\filldraw[\rColor] (-.2,-.7) -- (-.2,0) -- (-.4,.3) .. controls ++(90:.2cm) and ++(270:.2cm) .. (-.6,.7) -- (-.8,1) -- (-.8,3.1) -- (.9,3.1) -- (.9,-.7);
\end{scope}
\draw[dashed] (-.2,-.7) -- (-.2,0);
\draw[dashed] (-.4,.3) .. controls ++(90:.2cm) and ++(270:.2cm) .. (-.6,.7);
\draw[dashed] (-.8,1) -- (-.8,3.1);
\draw[black,thick] (-.4,1.3) .. controls ++(90:.45cm) and ++(270:.45cm) .. (0,2.3) arc (0:90:.4cm);
\filldraw[\rColor] (-.2,1.8) circle (.05cm);
\filldraw[\XColor] (-.4,2.7) circle (.05cm);
\draw[\XColor,thick] (0,0) -- (0,1.3) .. controls ++(90:.45cm) and ++(270:.45cm) .. (-.4,2.3) -- (-.4,3.1);
\roundNbox{unshaded}{(-.6,1)}{.3}{.1}{.1}{\scriptsize{$q$}};
\roundNbox{unshaded}{(-.2,0)}{.3}{.1}{.1}{\scriptsize{$b_i^{(n)}$}};
}
\approx
\sum_{j,k}
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.45,-4.2) rectangle (1,2.6);
\filldraw[\rColor] (.3,-4.2) -- (.3,-3.5) -- (0,-3.5) -- (0,-2.5) -- (-.2,-2.5) -- (-.2,-1.5) -- (0,-1.5) -- (0,-.5) -- (.3,-.5) -- (.3,.5) -- (0,.5) -- (0,1.5) -- (-.2,1.5) -- (-.2,2.6) -- (1,2.6) -- (1,-4.2);
\end{scope}
\draw[dashed] (.3,-4.2) -- (.3,-3.5);
\draw[dashed] (0,-3.5) -- (0,-2.5);
\draw[dashed] (-.2,-2.5) -- (-.2,-1.5);
\draw[black,thick] (.2,-2.5) -- (.2,-1.5);
\draw[dashed] (0,-1.5) -- (0,-.5);
\draw[\XColor,thick] (.6,-3.5) -- (.6,-.5);
\draw[dashed] (.3,-.5) -- (.3,.5);
\draw[dashed] (0,.5) -- (0,1.5);
\draw[black,thick] (.6,.5) -- (.6,1.8) arc (0:90:.4cm);
\draw[dashed] (-.2,1.5) -- (-.2,2.2);
\draw[\XColor,thick] (.2,1.5) -- (.2,2.6);
\filldraw[\XColor] (.2,2.2) circle (.05cm);
\roundNbox{unshaded}{(0,1.5)}{.3}{.1}{.1}{\scriptsize{$b_j^{(n)}$}};
\roundNbox{unshaded}{(.3,.5)}{.3}{.15}{.15}{\scriptsize{$q_k$}};
\roundNbox{unshaded}{(.3,-.5)}{.3}{.15}{.15}{\scriptsize{$(b_j^{(n)})^\dag$}};
\roundNbox{unshaded}{(0,-1.5)}{.3}{.1}{.1}{\scriptsize{$q_k^\dag$}};
\roundNbox{unshaded}{(0,-2.5)}{.3}{.1}{.1}{\scriptsize{$q$}};
\roundNbox{unshaded}{(.3,-3.5)}{.3}{.15}{.15}{\scriptsize{$b_i^{(n)}$}};
}
\approx
\sum_{j,k}
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.45,-4.2) rectangle (1,2.6);
\filldraw[\rColor] (.3,-4.2) -- (.3,-3.5) -- (0,-3.5) -- (0,-2.5) -- (.3,-2.5) -- (.3,-1.5) -- (0,-1.5) -- (0,-.5) -- (.3,-.5) -- (.3,.5) -- (0,.5) -- (0,1.5) -- (-.2,1.5) -- (-.2,2.6) -- (1,2.6) -- (1,-4.2);
\end{scope}
\draw[dashed] (.3,-4.2) -- (.3,-3.5);
\draw[dashed] (0,-3.5) -- (0,-2.5);
\draw[dashed] (.3,-2.5) -- (.3,-1.5);
\draw[black,thick] (.6,-3.5) -- (.6,-2.5);
\draw[dashed] (0,-1.5) -- (0,-.5);
\draw[\XColor,thick] (.6,-1.5) -- (.6,-.5);
\draw[dashed] (.3,-.5) -- (.3,.5);
\draw[dashed] (0,.5) -- (0,1.5);
\draw[black,thick] (.6,.5) -- (.6,1.8) arc (0:90:.4cm);
\draw[dashed] (-.2,1.5) -- (-.2,2.6);
\draw[\XColor,thick] (.2,1.5) -- (.2,2.6);
\filldraw[\XColor] (.2,2.2) circle (.05cm);
\roundNbox{unshaded}{(0,1.5)}{.3}{.1}{.1}{\scriptsize{$b_j^{(m)}$}};
\roundNbox{unshaded}{(.3,.5)}{.3}{.15}{.15}{\scriptsize{$q_k$}};
\roundNbox{unshaded}{(.3,-.5)}{.3}{.15}{.15}{\scriptsize{$(b_j^{(n)})^\dag$}};
\roundNbox{unshaded}{(.3,-1.5)}{.3}{.15}{.15}{\scriptsize{$b_i^{(n)}$}};
\roundNbox{unshaded}{(.3,-2.5)}{.3}{.15}{.15}{\scriptsize{$q_k^\dag$}};
\roundNbox{unshaded}{(.3,-3.5)}{.3}{.15}{.15}{\scriptsize{$q$}};
}
\approx
\sum_{j,k}
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.45,-4.2) rectangle (1,2.6);
\filldraw[\rColor] (.3,-4.2) -- (.3,-3.5) -- (0,-3.5) -- (0,-2.5) -- (.3,-2.5) -- (.3,-1.5) -- (0,-1.5) -- (0,-.5) -- (-.2,-.5) -- (-.2,.5) -- (0,.5) -- (0,1.5) -- (-.2,1.5) -- (-.2,2.6) -- (1,2.6) -- (1,-4.2);
\end{scope}
\draw[dashed] (.3,-4.2) -- (.3,-3.5);
\draw[dashed] (0,-3.5) -- (0,-2.5);
\draw[dashed] (.3,-2.5) -- (.3,-1.5);
\draw[black,thick] (.6,-3.5) -- (.6,-2.5);
\draw[dashed] (0,-1.5) -- (0,-.5);
\draw[\XColor,thick] (.2,-.5) -- (.2,.5);
\draw[dashed] (-.2,-.5) -- (-.2,.5);
\draw[dashed] (0,.5) -- (0,1.5);
\draw[black,thick] (.6,-1.5) -- (.6,1.8) arc (0:90:.4cm);
\draw[dashed] (-.2,1.5) -- (-.2,2.6);
\draw[\XColor,thick] (.2,1.5) -- (.2,2.6);
\filldraw[\XColor] (.2,2.2) circle (.05cm);
\roundNbox{unshaded}{(0,1.5)}{.3}{.1}{.1}{\scriptsize{$b_j^{(n)}$}};
\roundNbox{unshaded}{(0,.5)}{.3}{.1}{.1}{\scriptsize{$(b_j^{(\!n\!)})^\dag$}};
\roundNbox{unshaded}{(0,-.5)}{.3}{.1}{.1}{\scriptsize{$b_i^{(n)}$}};
\roundNbox{unshaded}{(.3,-1.5)}{.3}{.15}{.15}{\scriptsize{$q_k$}};
\roundNbox{unshaded}{(.3,-2.5)}{.3}{.15}{.15}{\scriptsize{$q_k^\dag$}};
\roundNbox{unshaded}{(.3,-3.5)}{.3}{.15}{.15}{\scriptsize{$q$}};
}
\approx
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.45,-1.2) rectangle (1.05,1.6);
\filldraw[\rColor] (.3,-1.2) -- (.3,-.5) -- (0,-.5) -- (0,.5) -- (-.2,.5) -- (-.2,1.6) -- (1.05,1.6) -- (1.05,-1.2);
\end{scope}
\draw[dashed] (-.2,.5) -- (-.2,1.6);
\draw[dashed] (0,-.5) -- (0,.5);
\draw[dashed] (.3,-1.2) -- (.3,-.5);
\draw[\XColor,thick] (.2,.5) -- (.2,1.6);
\draw[black,thick] (.6,-.5) -- (.6,.8) arc (0:90:.4cm);
\filldraw[\XColor] (.2,1.2) circle (.05cm);
\roundNbox{unshaded}{(0,.5)}{.3}{.1}{.1}{\scriptsize{$b_i^{(n)}$}};
\roundNbox{unshaded}{(.3,-.5)}{.3}{.15}{.15}{\scriptsize{$q$}};
}
=
b_i^{(n)} q
\]
The second equality uses the hypothesis that $X$ is a local bimodule.
The third $\approx$ uses \eqref{eq:UnderBraiding}.
The fourth $\approx$ holds because $\langle q_k|q\rangle^Q_M \in M$ and $\|[x,b_i^{(n)}]\|_2\to 0$ for $x\in M$.
The fifth $\approx$ holds because $\langle b_j^{(n)}|b_i^{(n)}\rangle^X_M$ is a central sequence in $M$ for each $i,j$
and $Q$ is centrally trivial over $M$.
The sixth $\approx$ holds from the definition of (approximate) basis for $X_M$ and $Q_M$.
We conclude that $|X|$ is approximately inner over $|Q|$.
Since $X$ is centrally trivial over $M$, by Proposition \ref{prop:BimodualCT&CS}, for all central sequences $(a_n)_n\subseteq M$ and all $x\in X$, $\|a_n x-xa_n\|_2 \to 0$.
If $(q_n)_n\subseteq |Q|$ is a central sequence, then
\[
q_n x
=
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.45,-1.2) rectangle (1.05,1.6);
\filldraw[\rColor] (.3,-1.2) -- (.3,-.5) -- (0,-.5) -- (0,.5) -- (-.2,.5) -- (-.2,2) -- (1.05,2) -- (1.05,-1.2);
\end{scope}
\draw[dashed] (-.2,.5) -- (-.2,1.6);
\draw[dashed] (0,-.5) -- (0,.5);
\draw[dashed] (.3,-1.2) -- (.3,-.5);
\draw[black,thick] (.2,.8) arc (180:90:.4cm);
\draw[\XColor,thick] (.6,-.5) -- (.6,1.6);
\filldraw[\XColor] (.6,1.2) circle (.05cm);
\roundNbox{unshaded}{(0,.5)}{.3}{.1}{.1}{\scriptsize{$q_n$}};
\roundNbox{unshaded}{(.3,-.5)}{.3}{.15}{.15}{\scriptsize{$x$}};
}
=
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-1,-.7) rectangle (.5,3);
\filldraw[\rColor] (-.2,-.7) -- (-.2,0) -- (-.4,.3) .. controls ++(90:.2cm) and ++(270:.2cm) .. (-.6,.7) -- (-.8,1) -- (-.8,3) -- (.9,3) -- (.9,-.7);
\end{scope}
\draw[dashed] (-.2,-.7) -- (-.2,0);
\draw[dashed] (-.4,.3) .. controls ++(90:.2cm) and ++(270:.2cm) .. (-.6,.7);
\draw[dashed] (-.8,1) -- (-.8,3);
\draw[\XColor,thick] (0,0) -- (0,1.3) .. controls ++(90:.45cm) and ++(270:.45cm) .. (-.4,2.3) -- (-.4,3);
\filldraw[\rColor] (-.2,1.8) circle (.05cm);
\draw[black,thick] (-.4,1.3) .. controls ++(90:.45cm) and ++(270:.45cm) .. (0,2.3) arc (0:90:.4cm);
\filldraw[\XColor] (-.4,2.7) circle (.05cm);
\roundNbox{unshaded}{(-.6,1)}{.3}{.1}{.1}{\scriptsize{$q_n$}};
\roundNbox{unshaded}{(-.2,0)}{.3}{.1}{.1}{\scriptsize{$x$}};
}
\approx
\sum_{j,k}
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.45,-4.2) rectangle (1,2.5);
\filldraw[\rColor] (.3,-4.2) -- (.3,-3.5) -- (0,-3.5) -- (0,-2.5) -- (-.2,-2.5) -- (-.2,-1.5) -- (0,-1.5) -- (0,-.5) -- (.3,-.5) -- (.3,.5) -- (0,.5) -- (0,1.5) -- (-.2,1.5) -- (-.2,2.5) -- (1,2.5) -- (1,-4.2);
\end{scope}
\draw[dashed] (.3,-4.2) -- (.3,-3.5);
\draw[dashed] (0,-3.5) -- (0,-2.5);
\draw[dashed] (-.2,-2.2) -- (-.2,-1.5);
\draw[black,thick] (.2,-2.2) -- (.2,-1.5);
\draw[dashed] (0,-1.5) -- (0,-.5);
\draw[\XColor,thick] (.6,-3.5) -- (.6,-.5);
\draw[dashed] (.3,-.5) -- (.3,.5);
\draw[dashed] (0,.5) -- (0,1.5);
\draw[black,thick] (.6,.5) -- (.6,1.8) arc (0:90:.4cm);
\draw[dashed] (-.2,1.5) -- (-.2,2.5);
\draw[\XColor,thick] (.2,1.5) -- (.2,2.5);
\filldraw[\XColor] (.2,2.2) circle (.05cm);
\roundNbox{unshaded}{(0,1.5)}{.3}{.1}{.1}{\scriptsize{$b_j$}};
\roundNbox{unshaded}{(.3,.5)}{.3}{.15}{.15}{\scriptsize{$q_k^{(n)}$}};
\roundNbox{unshaded}{(.3,-.5)}{.3}{.15}{.15}{\scriptsize{$b_j^\dag$}};
\roundNbox{unshaded}{(0,-1.5)}{.3}{.1}{.1}{\scriptsize{$(q_k^{(\!n\!)})^\dag$}};
\roundNbox{unshaded}{(0,-2.5)}{.3}{.1}{.1}{\scriptsize{$q_n$}};
\roundNbox{unshaded}{(.3,-3.5)}{.3}{.15}{.15}{\scriptsize{$x$}};
}
\approx
\sum_{j,k}
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.45,-4.2) rectangle (1,2.5);
\filldraw[\rColor] (.3,-4.2) -- (.3,-3.5) -- (0,-3.5) -- (0,-2.5) -- (.3,-2.5) -- (.3,-1.5) -- (0,-1.5) -- (0,-.5) -- (.3,-.5) -- (.3,.5) -- (0,.5) -- (0,1.5) -- (-.2,1.5) -- (-.2,2.5) -- (1,2.5) -- (1,-4.2);
\end{scope}
\draw[dashed] (.3,-4.2) -- (.3,-3.5);
\draw[dashed] (0,-3.5) -- (0,-2.5);
\draw[dashed] (.3,-2.5) -- (.3,-1.5);
\draw[black,thick] (.6,-3.5) -- (.6,-2.5);
\draw[dashed] (0,-1.5) -- (0,-.5);
\draw[\XColor,thick] (.6,-1.5) -- (.6,-.5);
\draw[dashed] (.3,-.5) -- (.3,.5);
\draw[dashed] (0,.5) -- (0,1.5);
\draw[black,thick] (.6,.5) -- (.6,1.8) arc (0:90:.4cm);
\draw[dashed] (-.2,1.5) -- (-.2,2.5);
\draw[\XColor,thick] (.2,1.5) -- (.2,2.5);
\filldraw[\XColor] (.2,2.2) circle (.05cm);
\roundNbox{unshaded}{(0,1.5)}{.3}{.1}{.1}{\scriptsize{$b_j$}};
\roundNbox{unshaded}{(.3,.5)}{.3}{.15}{.15}{\scriptsize{$q_k^{(n)}$}};
\roundNbox{unshaded}{(.3,-.5)}{.3}{.15}{.15}{\scriptsize{$b_j^\dag$}};
\roundNbox{unshaded}{(.3,-1.5)}{.3}{.15}{.15}{\scriptsize{$x$}};
\roundNbox{unshaded}{(.3,-2.5)}{.3}{.15}{.15}{\scriptsize{$(q_k^{(\!n\!)})^\dag$}};
\roundNbox{unshaded}{(.3,-3.5)}{.3}{.15}{.15}{\scriptsize{$q_n$}};
}
\approx
\sum_{j,k}
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.45,-4.2) rectangle (1,2.5);
\filldraw[\rColor] (.3,-4.2) -- (.3,-3.5) -- (0,-3.5) -- (0,-2.5) -- (.3,-2.5) -- (.3,-1.5) -- (0,-1.5) -- (0,-.5) -- (-.2,-.5) -- (-.2,.5) -- (0,.5) -- (0,1.5) -- (-.2,1.5) -- (-.2,2.5) -- (1,2.5) -- (1,-4.2);
\end{scope}
\draw[dashed] (.3,-4.2) -- (.3,-3.5);
\draw[dashed] (0,-3.5) -- (0,-2.5);
\draw[dashed] (.3,-2.5) -- (.3,-1.5);
\draw[black,thick] (.6,-3.5) -- (.6,-2.5);
\draw[dashed] (0,-1.5) -- (0,-.5);
\draw[\XColor,thick] (.2,-.5) -- (.2,.5);
\draw[dashed] (-.2,-.5) -- (-.2,.5);
\draw[dashed] (0,.5) -- (0,1.5);
\draw[black,thick] (.6,-1.5) -- (.6,1.8) arc (0:90:.4cm);
\draw[dashed] (-.2,1.5) -- (-.2,2.5);
\draw[\XColor,thick] (.2,1.5) -- (.2,2.5);
\filldraw[\XColor] (.2,2.2) circle (.05cm);
\roundNbox{unshaded}{(0,1.5)}{.3}{.1}{.1}{\scriptsize{$b_j$}};
\roundNbox{unshaded}{(0,.5)}{.3}{.1}{.1}{\scriptsize{$b_j^\dag$}};
\roundNbox{unshaded}{(0,-.5)}{.3}{.1}{.1}{\scriptsize{$x$}};
\roundNbox{unshaded}{(.3,-1.5)}{.3}{.15}{.15}{\scriptsize{$q_k^{(n)}$}};
\roundNbox{unshaded}{(.3,-2.5)}{.3}{.15}{.15}{\scriptsize{$(q_k^{(\!n\!)})^\dag$}};
\roundNbox{unshaded}{(.3,-3.5)}{.3}{.15}{.15}{\scriptsize{$q_n$}};
}
\approx
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.45,-1.2) rectangle (1.05,1.6);
\filldraw[\rColor] (.3,-1.2) -- (.3,-.5) -- (0,-.5) -- (0,.5) -- (-.2,.5) -- (-.2,1.6) -- (1.05,1.6) -- (1.05,-1.2);
\end{scope}
\draw[dashed] (-.2,.5) -- (-.2,1.6);
\draw[dashed] (0,-.5) -- (0,.5);
\draw[dashed] (.3,-1.2) -- (.3,-.5);
\draw[\XColor,thick] (.2,.5) -- (.2,1.6);
\draw[black,thick] (.6,-.5) -- (.6,.8) arc (0:90:.4cm);
\filldraw[\XColor] (.2,1.2) circle (.05cm);
\roundNbox{unshaded}{(0,.5)}{.3}{.1}{.1}{\scriptsize{$x$}};
\roundNbox{unshaded}{(.3,-.5)}{.3}{.15}{.15}{\scriptsize{$q_n$}};
}
=
x q_n
\]
The second equality uses the hypothesis that $X$ is local.
The third $\approx$ uses \eqref{eq:OverBraiding}.
The fourth $\approx$ holds because $\langle q_k^{(n)}| q_n\rangle^Q_M$ is a central sequence in $M$, and $X$ is centrally trivial.
The fifth $\approx$ holds because $\langle b_j|x\rangle^X_M\in M$ and $\|[a,q_k^{(n)}]\|_2 \to 0$ for $a\in M$.
The sixth $\approx$ holds from the definition of (approximate) basis for $X_M$ and $Q_M$.
We conclude that $|X|$ is centrally trivial over $|Q|$.
\end{proof}
The following proposition is straightforward; we omit its proof.
\begin{prop}
\label{prop:(A)PPBfromQtoM}
\mbox{}
\begin{enumerate}
\item
If $\{b_i\}_i$ is an $|X|_{|Q|}$-basis and $\{q_j\}_j$ is a $Q_M$-basis,
then
$$\left\{\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.3,-.7) rectangle (.6,.7);
\filldraw[\rColor] (0,-.7) -- (0,0) -- (-.2,0) -- (-.2,.7) -- (.7,.7) -- (.7,-.7);
\end{scope}
\draw[dashed] (0,-.7) -- (0,0);
\draw[dashed] (-.2,0) -- (-.2,.7);
\draw[\XColor,thick] (.2,0) -- (.2,.7);
\roundNbox{unshaded}{(0,0)}{.3}{.1}{.1}{\scriptsize{$c_{i,j}$}};
}
:=
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.45,-1.2) rectangle (1.05,1.6);
\filldraw[\rColor] (.3,-1.2) -- (.3,-.5) -- (0,-.5) -- (0,.5) -- (-.2,.5) -- (-.2,1.6) -- (1.05,1.6) -- (1.05,-1.2);
\end{scope}
\draw[dashed] (-.2,.5) -- (-.2,1.6);
\draw[dashed] (0,-.5) -- (0,.5);
\draw[dashed] (.3,-1.2) -- (.3,-.5);
\draw[\XColor,thick] (.2,.5) -- (.2,1.6);
\draw[black,thick] (.6,-.5) -- (.6,.8) arc (0:90:.4cm);
\filldraw[\XColor] (.2,1.2) circle (.05cm);
\roundNbox{unshaded}{(0,.5)}{.3}{.1}{.1}{\scriptsize{$b_i$}};
\roundNbox{unshaded}{(.3,-.5)}{.3}{.15}{.15}{\scriptsize{$q_j$}};
}\right\}_{i,j}$$
is an $|X|_M$-basis.
\item
If $\{b_i^{(n)}\}$ is an approximate $X_{|Q|}$-basis and $\{q_j^{(n)}\}$ is an approximately inner $Q_M$-basis,
then
$$
\left\{\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.3,-.7) rectangle (.6,.7);
\filldraw[\rColor] (0,-.7) -- (0,0) -- (-.2,0) -- (-.2,.7) -- (.7,.7) -- (.7,-.7);
\end{scope}
\draw[dashed] (0,-.7) -- (0,0);
\draw[dashed] (-.2,0) -- (-.2,.7);
\draw[\XColor,thick] (.2,0) -- (.2,.7);
\roundNbox{unshaded}{(0,0)}{.3}{.1}{.1}{\scriptsize{$c_{i,j}^{(n)}$}};
}
:=
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.45,-1.2) rectangle (1.05,1.6);
\filldraw[\rColor] (.3,-1.2) -- (.3,-.5) -- (0,-.5) -- (0,.5) -- (-.2,.5) -- (-.2,1.6) -- (1.05,1.6) -- (1.05,-1.2);
\end{scope}
\draw[dashed] (-.2,.5) -- (-.2,1.6);
\draw[dashed] (0,-.5) -- (0,.5);
\draw[dashed] (.3,-1.2) -- (.3,-.5);
\draw[\XColor,thick] (.2,.5) -- (.2,1.6);
\draw[black,thick] (.6,-.5) -- (.6,.8) arc (0:90:.4cm);
\filldraw[\XColor] (.2,1.2) circle (.05cm);
\roundNbox{unshaded}{(0,.5)}{.3}{.1}{.1}{\scriptsize{$b_i^{(n)}$}};
\roundNbox{unshaded}{(.3,-.5)}{.3}{.15}{.15}{\scriptsize{$q_j^{(n)}$}};
}\right\}_{i,j}
$$ is an approximate $|X|_M$-basis.
Moreover, if $\{b_i^{(n)}\}_i$ is approximately inner, so is $\{c_{i,j}^{(n)}\}$.
\end{enumerate}
Since $X_M$ is canonically isomorphic to $|X|_M$, we may view (1) as an $X_M$-basis and (2) as an approximate(ly inner) $X_M$-basis.
\end{prop}
\begin{prop}
\label{Chi(Q)subsetChi(M)locQ}
Every bimodule in $\tilde{\chi}(|Q|)$ is unitarily isomorphic to a realization of a bimodule in $\tilde{\chi}(M)_Q^\operatorname{loc}$.
\end{prop}
\begin{proof}
By Remark \ref{rem:QBimodules}, it suffices to consider a $Q-Q$ bimodule ${}_QX_Q$ in $\mathsf{Bim_{fgp}}(M)$ such that $|X| \in \tilde{\chi}(|Q|)$.
In order to show $X\in\tilde{\chi}(M)_Q^\operatorname{loc}$, must prove $X$ is centrally trivial and approximately inner over $M$, and $X$ is a local $Q-Q$ bimodule.
By Lemma \ref{lem:CSinMCSin|Q|}, any central sequence $(a_n)\subseteq M$ is also a central sequence in $|Q|$.
By Proposition \ref{prop:BimodualCT&CS}, $X$
is centrally trivial over $|Q|$, so $X$ is centrally trivial over $M$.
By Proposition \ref{prop:(A)PPBfromQtoM} and Proposition \ref{prop:AI&APPB},
we have $X$
is approximately inner over $M$.
Therefore,as an $M-M$ bimodule, $X\in \tilde{\chi}(M)$.
Since $X$ is already a $Q-Q$ bimodule, it remains to show $X$ is local.
Let $\{b_i\},\{q_j\}$ be $X_{|Q|},Q_M$-bases
and let
$\{b_i^{(n)}\},\{q_j^{(n)}\}$ be approximately inner $X_{|Q|},Q_M$-bases respectively.
Defining
$\{c_{i,j}\}$ and $\{c_{i,j}^{(n)}\}$ as in
Proposition \ref{prop:(A)PPBfromQtoM}
gives an $X_M$-basis and an approximately inner $X_M$-basis respectively.
For the over braiding,
\begin{align*}
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.55,-1.2) rectangle (.7,.5);
\filldraw[\rColor] (-.4,-1.2) rectangle (.7,.5);
\end{scope}
\draw[dashed] (-.4,-1.2) -- (-.4,.5);
\draw[\XColor,thick] (.4,-1.2) -- (.4,-.5) .. controls ++(90:.45cm) and ++(270:.45cm) .. (0,.5);
\filldraw[\rColor] (.2,0) circle (.05cm);
\draw[black,thick] (.4,-.9) arc (270:180:.4cm) -- (0,-.5) .. controls ++(90:.45cm) and ++(270:.45cm) .. (.4,.5);
\filldraw[\XColor] (.4,-.9) circle (.05cm);
}
&\approx
\sum_{i,j,k}
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.35,-2.5) rectangle (1,2.2);
\filldraw[\rColor] (-.2,-2.5) -- (-.2,-1.5) -- (0,-1.5) -- (0,-.5) -- (.3,-.5) -- (.3,.5) -- (0,.5) -- (0,1.5) -- (-.2,1.5) -- (-.2,2.2) -- (1.05,2.2) -- (1.05,-2.5);
\end{scope}
\draw[dashed] (-.2,-2.5) -- (-.2,-1.5);
\draw[black,thick] (.6,-2.2) arc (270:180:.4cm);
\draw[dashed] (0,-1.5) -- (0,-.5);
\draw[\XColor,thick] (.6,-2.5) -- (.6,-.5);
\draw[dashed] (.3,-.5) -- (.3,.5);
\draw[dashed] (0,.5) -- (0,1.5);
\draw[black,thick] (.6,.5) -- (.6,2.2);
\draw[dashed] (-.2,1.5) -- (-.2,2.2);
\draw[\XColor,thick] (.2,1.5) -- (.2,2.2);
\filldraw[\XColor] (.6,-2.2) circle (.05cm);
\roundNbox{unshaded}{(0,1.5)}{.3}{.1}{.1}{\scriptsize{$c_{i,j}$}};
\roundNbox{unshaded}{(.3,.5)}{.3}{.15}{.15}{\scriptsize{$q_k^{(n)}$}};
\roundNbox{unshaded}{(.3,-.5)}{.3}{.15}{.15}{\scriptsize{$c_{i,j}^\dag$}};
\roundNbox{unshaded}{(0,-1.5)}{.3}{.1}{.1}{\scriptsize{$(q_k^{(\!n\!)})^\dag$}};
}
=
\sum_{i,j,k}
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.75,-3.7) rectangle (1,3.5);
\filldraw[\rColor] (-.6,-3.7) -- (-.6,-2.5) -- (-.4,-2.5) -- (-.4,-1.5) -- (-.1,-1.5) -- (-.1,-.5) -- (.25,-.5) -- (.25,.5) -- (-.1,.5) -- (-.1,1.5) -- (-.4,1.5) -- (-.4,2.5) -- (-.6,2.5) -- (-.6,3.5) -- (1,3.5) -- (1,-3.7);
\end{scope}
\draw[dashed] (-.6,-3.7) -- (-.6,-2.5);
\draw[dashed] (-.4,-2.5) -- (-.4,-1.5);
\draw[black,thick] (.2,-3.4) arc (270:180:.4cm);
\draw[dashed] (-.1,-1.5) -- (-.1,-.5);
\draw[black,thick] (.6,-.5) -- (.6,-1.8) arc (0:-90:.4cm);
\draw[\XColor,thick] (.2,-3.7) -- (.2,-1.5);
\draw[dashed] (.25,-.5) -- (.25,.5);
\draw[dashed] (-.1,.5) -- (-.1,1.5);
\draw[black,thick] (.6,.5) -- (.6,3.5);
\draw[dashed] (-.4,1.5) -- (-.4,2.5);
\draw[black,thick] (.2,1.5) -- (.2,2.8) arc (0:90:.4cm);
\draw[dashed] (-.6,2.5) -- (-.6,3.5);
\draw[\XColor,thick] (-.2,2.5) -- (-.2,3.5);
\filldraw[\XColor] (.2,-3.4) circle (.05cm);
\filldraw[\XColor] (.2,-2.2) circle (.05cm);
\filldraw[\XColor] (-.2,3.2) circle (.05cm);
\roundNbox{unshaded}{(-.4,2.5)}{.3}{.1}{.1}{\scriptsize{$b_i$}};
\roundNbox{unshaded}{(-.1,1.5)}{.3}{.15}{.15}{\scriptsize{$q_j$}};
\roundNbox{unshaded}{(.25,.5)}{.3}{.2}{.2}{\scriptsize{$q_k^{(n)}$}};
\roundNbox{unshaded}{(.25,-.5)}{.3}{.2}{.2}{\scriptsize{$q_j^\dag$}};
\roundNbox{unshaded}{(-.1,-1.5)}{.3}{.15}{.15}{\scriptsize{$b_i^\dag$}};
\roundNbox{unshaded}{(-.4,-2.7)}{.3}{.1}{.1}{\scriptsize{$(q_k^{(\!n\!)})^\dag$}};
}
=
\sum_{i,j,k}
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.75,-3.7) rectangle (1,3.5);
\filldraw[\rColor] (-.6,-3.7) -- (-.6,-2.5) -- (-.4,-2.5) -- (-.4,-1.5) -- (-.1,-1.5) -- (-.1,-.5) -- (.25,-.5) -- (.25,.5) -- (-.1,.5) -- (-.1,1.5) -- (-.4,1.5) -- (-.4,2.5) -- (-.6,2.5) -- (-.6,3.5) -- (1,3.5) -- (1,-3.7);
\end{scope}
\draw[dashed] (-.6,-3.7) -- (-.6,-2.5);
\draw[dashed] (-.4,-2.5) -- (-.4,-1.5);
\draw[black,thick] (.2,-3.2) arc (270:180:.4cm);
\draw[dashed] (-.1,-1.5) -- (-.1,-.5);
\draw[black,thick] (.6,-.5) -- (.6,-3) arc (0:-90:.4cm);
\draw[\XColor,thick] (.2,-3.7) -- (.2,-1.5);
\draw[dashed] (.25,-.5) -- (.25,.5);
\draw[dashed] (-.1,.5) -- (-.1,1.5);
\draw[black,thick] (.6,.5) -- (.6,3.5);
\draw[dashed] (-.4,1.5) -- (-.4,2.5);
\draw[black,thick] (.2,1.5) -- (.2,2.8) arc (0:90:.4cm);
\draw[dashed] (-.6,2.5) -- (-.6,3.5);
\draw[\XColor,thick] (-.2,2.5) -- (-.2,3.5);
\filldraw[\XColor] (.2,-3.4) circle (.05cm);
\filldraw[\XColor] (.2,-3.2) circle (.05cm);
\filldraw[\XColor] (-.2,3.2) circle (.05cm);
\roundNbox{unshaded}{(-.4,2.5)}{.3}{.1}{.1}{\scriptsize{$b_i$}};
\roundNbox{unshaded}{(-.1,1.5)}{.3}{.15}{.15}{\scriptsize{$q_j$}};
\roundNbox{unshaded}{(.25,.5)}{.3}{.2}{.2}{\scriptsize{$q_k^{(n)}$}};
\roundNbox{unshaded}{(.25,-.5)}{.3}{.2}{.2}{\scriptsize{$q_j^\dag$}};
\roundNbox{unshaded}{(-.1,-1.5)}{.3}{.15}{.15}{\scriptsize{$b_i^\dag$}};
\roundNbox{unshaded}{(-.4,-2.5)}{.3}{.1}{.1}{\scriptsize{$(q_k^{(\!n\!)})^\dag$}};
}
\approx
\sum_{i,j,k}
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.75,-3.7) rectangle (1,3.5);
\filldraw[\rColor] (-.6,-3.7) -- (-.6,-2.5) -- (-.4,-2.5) -- (-.4,-1.5) -- (-.1,-1.5) -- (-.1,-.5) -- (.25,-.5) -- (.25,.5) -- (-.1,.5) -- (-.1,1.5) -- (-.4,1.5) -- (-.4,2.5) -- (-.6,2.5) -- (-.6,3.5) -- (1,3.5) -- (1,-3.7);
\end{scope}
\draw[dashed] (-.6,-3.7) -- (-.6,-2.5);
\draw[dashed] (-.4,-2.5) -- (-.4,-1.5);
\draw[black,thick] (-.2,-3.2) arc (-90:0:.4cm) -- (.2,-1.5);
\draw[dashed] (-.1,-1.5) -- (-.1,-.5);
\draw[black,thick] (.6,-.5) -- (.6,-2.6) arc (0:-90:.8cm);
\draw[\XColor,thick] (-.2,-3.7) -- (-.2,-2.5);
\draw[dashed] (.25,-.5) -- (.25,.5);
\draw[dashed] (-.1,.5) -- (-.1,1.5);
\draw[black,thick] (.6,.5) -- (.6,3.5);
\draw[dashed] (-.4,1.5) -- (-.4,2.5);
\draw[black,thick] (.2,1.5) -- (.2,2.8) arc (0:90:.4cm);
\draw[dashed] (-.6,2.5) -- (-.6,3.5);
\draw[\XColor,thick] (-.2,2.5) -- (-.2,3.5);
\filldraw[\XColor] (-.2,-3.4) circle (.05cm);
\filldraw[\XColor] (-.2,-3.2) circle (.05cm);
\filldraw[\XColor] (-.2,3.2) circle (.05cm);
\roundNbox{unshaded}{(-.4,2.5)}{.3}{.1}{.1}{\scriptsize{$b_i$}};
\roundNbox{unshaded}{(-.1,1.5)}{.3}{.15}{.15}{\scriptsize{$q_j$}};
\roundNbox{unshaded}{(.25,.5)}{.3}{.2}{.2}{\scriptsize{$q_k^{(n)}$}};
\roundNbox{unshaded}{(.25,-.5)}{.3}{.2}{.2}{\scriptsize{$q_j^\dag$}};
\roundNbox{unshaded}{(-.1,-1.5)}{.3}{.15}{.15}{\scriptsize{$(q_k^{(n)})^\dag$}};
\roundNbox{unshaded}{(-.4,-2.5)}{.3}{.1}{.1}{\scriptsize{$b_i^\dag$}};
}
\displaybreak[1]\\
&\approx
\sum_i
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.95,-2) rectangle (.7,1.8);
\filldraw[\rColor] (-.8,-2) -- (-.8,-.8) -- (-.6,-.8) -- (-.6,.8) -- (-.8,.8) -- (-.8,1.8) -- (1.3,1.8) -- (1.3,-2);
\end{scope}
\draw[dashed] (-.8,-2) -- (-.8,-.8);
\draw[\XColor,thick] (-.4,-2) -- (-.4,-.8);
\draw[dashed] (-.6,-.8) -- (-.6,.8);
\draw[black,thick] (-.4,-1.7) arc (-90:0:.8cm) -- (.4,-.5) .. controls ++(90:.45cm) and ++(270:.45cm) .. (0,.5) -- (0,1.1) arc (0:90:.4cm);
\filldraw[\rColor] (.2,0) circle (.05cm);
\draw[black,thick] (-.4,-1.5) arc (-90:0:.4cm) -- (0,-.5) .. controls ++(90:.45cm) and ++(270:.45cm) .. (.4,.5) -- (.4,1.8);
\draw[dashed] (-.8,.8) -- (-.8,1.8);
\draw[\XColor,thick] (-.4,.8) -- (-.4,1.8);
\filldraw[\XColor] (-.4,-1.7) circle (.05cm);
\filldraw[\XColor] (-.4,-1.5) circle (.05cm);
\filldraw[\XColor] (-.4,1.5) circle (.05cm);
\roundNbox{unshaded}{(-.6,-.8)}{.3}{.1}{.1}{\scriptsize{$b_i^\dag$}};
\roundNbox{unshaded}{(-.6,.8)}{.3}{.1}{.1}{\scriptsize{$b_i$}};
}
=
\sum_i
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.35,-1.7) rectangle (1.3,1.5);
\filldraw[\rColor] (-.2,-1.7) -- (-.2,-.5) -- (0,-.5) -- (0,.5) -- (-.2,.5) -- (-.2,1.5) -- (1.3,1.5) -- (1.3,-1.7);
\end{scope}
\draw[dashed] (-.2,-1.7) -- (-.2,-.5);
\draw[dashed] (0,-.5) -- (0,.5);
\draw[dashed] (-.2,.5) -- (-.2,1.5);
\draw[black,thick] (.2,1.2) arc (90:0:.4cm) -- (.6,-.8) arc (0:-90:.4cm);
\draw[black,thick] (.2,-1.4) arc (-90:0:.8cm) -- (1,1.5);
\draw[\XColor,thick] (.2,-1.7) -- (.2,-.5);
\draw[\XColor,thick] (.2,.5) -- (.2,1.5);
\filldraw[\XColor] (.2,-1.4) circle (.05cm);
\filldraw[\XColor] (.2,-1.2) circle (.05cm);
\filldraw[\XColor] (.2,1.2) circle (.05cm);
\roundNbox{unshaded}{(0,.5)}{.3}{.1}{.1}{\scriptsize{$b_i$}};
\roundNbox{unshaded}{(0,-.5)}{.3}{.1}{.1}{\scriptsize{$b_i^\dag$}};
}
=
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.55,-.4) rectangle (.7,.7);
\filldraw[\rColor] (-.4,-.4) rectangle (.7,.7);
\end{scope}
\draw[\XColor,thick] (0,-.4) -- (0,.7);
\draw[black,thick] (0,0) arc (-90:0:.4cm) -- (.4,.7);
\draw[dashed] (-.4,-.4) -- (-.4,.7);
\filldraw[\XColor] (0,0) circle (.05cm);
}
\end{align*}
The first $\approx$ uses \eqref{eq:OverBraiding}, the second equality is the defintion of $c_{i,j}$, the third equality uses associativty of the $Q-Q$ bimodule actions, the fourth $\approx$ uses that $X$ is centrally trivial over $|Q|$, the fifth $\approx$ uses \eqref{eq:ExpandBraidingsQ}, and the sixth equation
follows from
\begin{equation}
\label{eq:UseCommutativityForRichtAction}
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.7,-1.6) rectangle (.7,.5);
\filldraw[\rColor] (-.7,-1.6) rectangle (.7,.5);
\end{scope}
\draw[\XColor,thick] (-.4,-1.6) -- (-.4,.5);
\draw[black,thick] (-.4,-1.3) arc (-90:0:.8cm) -- (.4,-.5) .. controls ++(90:.45cm) and ++(270:.45cm) .. (0,.5);
\filldraw[\rColor] (.2,0) circle (.05cm);
\draw[black,thick] (-.4,-1.1) arc (-90:0:.4cm) -- (0,-.5) .. controls ++(90:.45cm) and ++(270:.45cm) .. (.4,.5);
\filldraw[\XColor] (-.4,-1.3) circle (.05cm);
\filldraw[\XColor] (-.4,-1.1) circle (.05cm);
}
=
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.7,-1.6) rectangle (.7,.5);
\filldraw[\rColor] (-.7,-1.6) rectangle (.7,.5);
\end{scope}
\draw[\XColor,thick] (-.4,-1.6) -- (-.4,.5);
\draw[black,thick] (0,-.5) arc (-180:0:.2cm) -- (.4,-.5) .. controls ++(90:.45cm) and ++(270:.45cm) .. (0,.5);
\filldraw[\rColor] (.2,0) circle (.05cm);
\draw[black,thick] (0,-.5) .. controls ++(90:.45cm) and ++(270:.45cm) .. (.4,.5);
\draw[black,thick] (.2,-.7) arc (0:-90:.6cm);
\filldraw[black] (.2,-.7) circle (.05cm);
\filldraw[\XColor] (-.4,-1.3) circle (.05cm);
}
=
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.7,-.9) rectangle (.7,.5);
\filldraw[\rColor] (-.7,-1.6) rectangle (.7,.5);
\end{scope}
\draw[\XColor,thick] (-.4,-.9) -- (-.4,.5);
\draw[black,thick] (0,.5) -- (0,.2) arc (-180:0:.2cm) -- (.4,.5);
\draw[black,thick] (.2,0) arc (0:-90:.6cm);
\filldraw[black] (.2,0) circle (.05cm);
\filldraw[\XColor] (-.4,-.6) circle (.05cm);
}
=
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.7,-.9) rectangle (.7,.5);
\filldraw[\rColor] (-.7,-1.6) rectangle (.7,.5);
\end{scope}
\draw[\XColor,thick] (-.4,-.9) -- (-.4,.5);
\draw[black,thick] (0,.5) -- (0,0) arc (0:-90:.4cm);
\draw[black,thick] (.4,.5) -- (.4,.2) arc (0:-90:.8cm);
\filldraw[\XColor] (-.4,-.4) circle (.05cm);
\filldraw[\XColor] (-.4,-.6) circle (.05cm);
}
=
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.7,-1.6) rectangle (.7,.5);
\filldraw[\rColor] (-.7,-1.6) rectangle (.7,.5);
\end{scope}
\draw[\XColor,thick] (-.4,-1.6) -- (-.4,.5);
\draw[black,thick] (-.4,-1.1) arc (-90:0:.4cm) -- (0,-.5) .. controls ++(90:.45cm) and ++(270:.45cm) .. (.4,.5);
\filldraw[\rColor] (.2,0) circle (.05cm);
\draw[black,thick] (-.4,-1.3) arc (-90:0:.8cm) -- (.4,-.5) .. controls ++(90:.45cm) and ++(270:.45cm) .. (0,.5);
\filldraw[\XColor] (-.4,-1.3) circle (.05cm);
\filldraw[\XColor] (-.4,-1.1) circle (.05cm);
}\,.
\end{equation}
For the under braiding,
\begin{align*}
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.55,-1.2) rectangle (.7,.5);
\filldraw[\rColor] (-.4,-1.2) rectangle (.7,.5);
\end{scope}
\draw[dashed] (-.4,-1.2) -- (-.4,.5);
\draw[black,thick] (.4,-.9) arc (270:180:.4cm) -- (0,-.5) .. controls ++(90:.45cm) and ++(270:.45cm) .. (.4,.5);
\filldraw[\rColor] (.2,0) circle (.05cm);
\draw[\XColor,thick] (.4,-1.2) -- (.4,-.5) .. controls ++(90:.45cm) and ++(270:.45cm) .. (0,.5);
\filldraw[\XColor] (.4,-.9) circle (.05cm);
}
&\approx
\sum_{i,j,k}
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.35,-2.5) rectangle (1,2.2);
\filldraw[\rColor] (-.2,-2.5) -- (-.2,-1.5) -- (0,-1.5) -- (0,-.5) -- (.3,-.5) -- (.3,.5) -- (0,.5) -- (0,1.5) -- (-.2,1.5) -- (-.2,2.2) -- (1.05,2.2) -- (1.05,-2.5);
\end{scope}
\draw[dashed] (-.2,-2.5) -- (-.2,-1.5);
\draw[black,thick] (.6,-2.2) arc (270:180:.4cm);
\draw[dashed] (0,-1.5) -- (0,-.5);
\draw[\XColor,thick] (.6,-2.5) -- (.6,-.5);
\draw[dashed] (.3,-.5) -- (.3,.5);
\draw[dashed] (0,.5) -- (0,1.5);
\draw[black,thick] (.6,.5) -- (.6,2.2);
\draw[dashed] (-.2,1.5) -- (-.2,2.2);
\draw[\XColor,thick] (.2,1.5) -- (.2,2.2);
\filldraw[\XColor] (.6,-2.2) circle (.05cm);
\roundNbox{unshaded}{(0,1.5)}{.3}{.1}{.1}{\scriptsize{$c_{i,j}^{(n)}$}};
\roundNbox{unshaded}{(.3,.5)}{.3}{.15}{.15}{\scriptsize{$q_k$}};
\roundNbox{unshaded}{(.3,-.5)}{.3}{.15}{.15}{\scriptsize{$(c_{i,j}^{(n)})^\dag$}};
\roundNbox{unshaded}{(0,-1.5)}{.3}{.1}{.1}{\scriptsize{$q_k^\dag$}};
}
=
\sum_{i,j,k}
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.75,-3.7) rectangle (1,3.5);
\filldraw[\rColor] (-.6,-3.7) -- (-.6,-2.5) -- (-.4,-2.5) -- (-.4,-1.5) -- (-.1,-1.5) -- (-.1,-.5) -- (.25,-.5) -- (.25,.5) -- (-.1,.5) -- (-.1,1.5) -- (-.4,1.5) -- (-.4,2.5) -- (-.6,2.5) -- (-.6,3.5) -- (1,3.5) -- (1,-3.7);
\end{scope}
\draw[dashed] (-.6,-3.7) -- (-.6,-2.5);
\draw[dashed] (-.4,-2.5) -- (-.4,-1.5);
\draw[black,thick] (.2,-3.4) arc (270:180:.4cm);
\draw[dashed] (-.1,-1.5) -- (-.1,-.5);
\draw[black,thick] (.6,-.5) -- (.6,-1.8) arc (0:-90:.4cm);
\draw[\XColor,thick] (.2,-3.7) -- (.2,-1.5);
\draw[dashed] (.25,-.5) -- (.25,.5);
\draw[dashed] (-.1,.5) -- (-.1,1.5);
\draw[black,thick] (.6,.5) -- (.6,3.5);
\draw[dashed] (-.4,1.5) -- (-.4,2.5);
\draw[black,thick] (.2,1.5) -- (.2,2.8) arc (0:90:.4cm);
\draw[dashed] (-.6,2.5) -- (-.6,3.5);
\draw[\XColor,thick] (-.2,2.5) -- (-.2,3.5);
\filldraw[\XColor] (.2,-3.4) circle (.05cm);
\filldraw[\XColor] (.2,-2.2) circle (.05cm);
\filldraw[\XColor] (-.2,3.2) circle (.05cm);
\roundNbox{unshaded}{(-.4,2.5)}{.3}{.1}{.1}{\scriptsize{$b_i^{(n)}$}};
\roundNbox{unshaded}{(-.1,1.5)}{.3}{.15}{.15}{\scriptsize{$q_j^{(n)}$}};
\roundNbox{unshaded}{(.25,.5)}{.3}{.2}{.2}{\scriptsize{$q_k$}};
\roundNbox{unshaded}{(.25,-.5)}{.3}{.2}{.2}{\scriptsize{$(q_j^{(n)})^\dag$}};
\roundNbox{unshaded}{(-.1,-1.5)}{.3}{.15}{.15}{\scriptsize{$(b_i^{(n)})^\dag$}};
\roundNbox{unshaded}{(-.4,-2.7)}{.3}{.1}{.1}{\scriptsize{$q_k^\dag$}};
}
=
\sum_{i,j,k}
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.75,-3.7) rectangle (1,3.5);
\filldraw[\rColor] (-.6,-3.7) -- (-.6,-2.5) -- (-.4,-2.5) -- (-.4,-1.5) -- (-.1,-1.5) -- (-.1,-.5) -- (.25,-.5) -- (.25,.5) -- (-.1,.5) -- (-.1,1.5) -- (-.4,1.5) -- (-.4,2.5) -- (-.6,2.5) -- (-.6,3.5) -- (1,3.5) -- (1,-3.7);
\end{scope}
\draw[dashed] (-.6,-3.7) -- (-.6,-2.5);
\draw[dashed] (-.4,-2.5) -- (-.4,-1.5);
\draw[black,thick] (.2,-3.2) arc (270:180:.4cm);
\draw[dashed] (-.1,-1.5) -- (-.1,-.5);
\draw[black,thick] (.6,-.5) -- (.6,-3) arc (0:-90:.4cm);
\draw[\XColor,thick] (.2,-3.7) -- (.2,-1.5);
\draw[dashed] (.25,-.5) -- (.25,.5);
\draw[dashed] (-.1,.5) -- (-.1,1.5);
\draw[black,thick] (.6,.5) -- (.6,3.5);
\draw[dashed] (-.4,1.5) -- (-.4,2.5);
\draw[black,thick] (.2,1.5) -- (.2,2.8) arc (0:90:.4cm);
\draw[dashed] (-.6,2.5) -- (-.6,3.5);
\draw[\XColor,thick] (-.2,2.5) -- (-.2,3.5);
\filldraw[\XColor] (.2,-3.4) circle (.05cm);
\filldraw[\XColor] (.2,-3.2) circle (.05cm);
\filldraw[\XColor] (-.2,3.2) circle (.05cm);
\roundNbox{unshaded}{(-.4,2.5)}{.3}{.1}{.1}{\scriptsize{$b_i^{(n)}$}};
\roundNbox{unshaded}{(-.1,1.5)}{.3}{.15}{.15}{\scriptsize{$q_j^{(n)}$}};
\roundNbox{unshaded}{(.25,.5)}{.3}{.2}{.2}{\scriptsize{$q_k$}};
\roundNbox{unshaded}{(.25,-.5)}{.3}{.2}{.2}{\scriptsize{$(q_j^{(n)})^\dag$}};
\roundNbox{unshaded}{(-.1,-1.5)}{.3}{.15}{.15}{\scriptsize{$(b_i^{(n)})^\dag$}};
\roundNbox{unshaded}{(-.4,-2.5)}{.3}{.1}{.1}{\scriptsize{$q_k^\dag$}};
}
\approx
\sum_{i,j,k}
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.75,-3.7) rectangle (1,3.5);
\filldraw[\rColor] (-.6,-3.7) -- (-.6,-2.5) -- (-.4,-2.5) -- (-.4,-1.5) -- (-.1,-1.5) -- (-.1,-.5) -- (.25,-.5) -- (.25,.5) -- (-.1,.5) -- (-.1,1.5) -- (-.4,1.5) -- (-.4,2.5) -- (-.6,2.5) -- (-.6,3.5) -- (1,3.5) -- (1,-3.7);
\end{scope}
\draw[dashed] (-.6,-3.7) -- (-.6,-2.5);
\draw[dashed] (-.4,-2.5) -- (-.4,-1.5);
\draw[black,thick] (-.2,-3.2) arc (-90:0:.4cm) -- (.2,-1.5);
\draw[dashed] (-.1,-1.5) -- (-.1,-.5);
\draw[black,thick] (.6,-.5) -- (.6,-2.6) arc (0:-90:.8cm);
\draw[\XColor,thick] (-.2,-3.7) -- (-.2,-2.5);
\draw[dashed] (.25,-.5) -- (.25,.5);
\draw[dashed] (-.1,.5) -- (-.1,1.5);
\draw[black,thick] (.6,.5) -- (.6,3.5);
\draw[dashed] (-.4,1.5) -- (-.4,2.5);
\draw[black,thick] (.2,1.5) -- (.2,2.8) arc (0:90:.4cm);
\draw[dashed] (-.6,2.5) -- (-.6,3.5);
\draw[\XColor,thick] (-.2,2.5) -- (-.2,3.5);
\filldraw[\XColor] (-.2,-3.4) circle (.05cm);
\filldraw[\XColor] (-.2,-3.2) circle (.05cm);
\filldraw[\XColor] (-.2,3.2) circle (.05cm);
\roundNbox{unshaded}{(-.4,2.5)}{.3}{.1}{.1}{\scriptsize{$b_i^{(n)}$}};
\roundNbox{unshaded}{(-.1,1.5)}{.3}{.15}{.15}{\scriptsize{$q_j^{(n)}$}};
\roundNbox{unshaded}{(.25,.5)}{.3}{.2}{.2}{\scriptsize{$q_k$}};
\roundNbox{unshaded}{(.25,-.5)}{.3}{.2}{.2}{\scriptsize{$(q_j^{(n)})^\dag$}};
\roundNbox{unshaded}{(-.1,-1.5)}{.3}{.15}{.15}{\scriptsize{$q_k^\dag$}};
\roundNbox{unshaded}{(-.4,-2.5)}{.3}{.1}{.1}{\scriptsize{$(b_i^{(\!n\!)})^\dag$}};
}
\\
&\approx
\sum_i
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.95,-2) rectangle (.7,1.8);
\filldraw[\rColor] (-.8,-2) -- (-.8,-.8) -- (-.6,-.8) -- (-.6,.8) -- (-.8,.8) -- (-.8,1.8) -- (1.3,1.8) -- (1.3,-2);
\end{scope}
\draw[dashed] (-.8,-2) -- (-.8,-.8);
\draw[\XColor,thick] (-.4,-2) -- (-.4,-.8);
\draw[dashed] (-.6,-.8) -- (-.6,.8);
\draw[black,thick] (-.4,-1.5) arc (-90:0:.4cm) -- (0,-.5) .. controls ++(90:.45cm) and ++(270:.45cm) .. (.4,.5) -- (.4,1.8);
\filldraw[\rColor] (.2,0) circle (.05cm);
\draw[black,thick] (-.4,-1.7) arc (-90:0:.8cm) -- (.4,-.5) .. controls ++(90:.45cm) and ++(270:.45cm) .. (0,.5) -- (0,1.1) arc (0:90:.4cm);
\draw[dashed] (-.8,.8) -- (-.8,1.8);
\draw[\XColor,thick] (-.4,.8) -- (-.4,1.8);
\filldraw[\XColor] (-.4,-1.7) circle (.05cm);
\filldraw[\XColor] (-.4,-1.5) circle (.05cm);
\filldraw[\XColor] (-.4,1.5) circle (.05cm);
\roundNbox{unshaded}{(-.6,-.8)}{.3}{.1}{.1}{\scriptsize{$(b_i^{(\!n\!)})^\dag$}};
\roundNbox{unshaded}{(-.6,.8)}{.3}{.1}{.1}{\scriptsize{$b_i^{(n)}$}};
}
=
\sum_i
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.35,-1.7) rectangle (1.3,1.5);
\filldraw[\rColor] (-.2,-1.7) -- (-.2,-.5) -- (0,-.5) -- (0,.5) -- (-.2,.5) -- (-.2,1.5) -- (1.3,1.5) -- (1.3,-1.7);
\end{scope}
\draw[dashed] (-.2,-1.7) -- (-.2,-.5);
\draw[dashed] (0,-.5) -- (0,.5);
\draw[dashed] (-.2,.5) -- (-.2,1.5);
\draw[black,thick] (.2,1.2) arc (90:0:.4cm) -- (.6,-.8) arc (0:-90:.4cm);
\draw[black,thick] (.2,-1.4) arc (-90:0:.8cm) -- (1,1.5);
\draw[\XColor,thick] (.2,-1.7) -- (.2,-.5);
\draw[\XColor,thick] (.2,.5) -- (.2,1.5);
\filldraw[\XColor] (.2,-1.4) circle (.05cm);
\filldraw[\XColor] (.2,-1.2) circle (.05cm);
\filldraw[\XColor] (.2,1.2) circle (.05cm);
\roundNbox{unshaded}{(0,.5)}{.3}{.1}{.1}{\scriptsize{$b_i^{(n)}$}};
\roundNbox{unshaded}{(0,-.5)}{.3}{.1}{.1}{\scriptsize{$(b_i^{(\!n\!)})^\dag$}};
}
\approx
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.55,-.4) rectangle (.7,.7);
\filldraw[\rColor] (-.4,-.4) rectangle (.7,.7);
\end{scope}
\draw[\XColor,thick] (0,-.4) -- (0,.7);
\draw[black,thick] (0,0) arc (-90:0:.4cm) -- (.4,.7);
\draw[dashed] (-.4,-.4) -- (-.4,.7);
\filldraw[\XColor] (0,0) circle (.05cm);
}\,.
\end{align*}
The first $\approx$ uses \eqref{eq:UnderBraiding}, the second equality is the definition of $c_{i,j}^{(n)}$, the third equality uses associativity of the $Q-Q$ bimodule actions, the fourth $\approx$ uses $X$ is centrally trivial over $|Q|$ so $\|[b_i^{(n)},q_k]\|_2\to 0$,
the fifth $\approx$ uses \eqref{eq:ExpandBraidingsQ}, and the sixth equality uses \eqref{eq:UseCommutativityForRichtAction} again.
\end{proof}
\begin{thm}
\label{thm:LocalExtension}
Realization gives a braided unitary equivalence
$\tilde{\chi}(M)_Q^{\operatorname{loc}}\to \tilde{\chi}(|Q|)$.
\end{thm}
\begin{proof}
By Remark
\ref{rem:QBimodules},
realization $|\,\cdot\,|$ gives a unitary tensor equivalence from $Q-Q$ bimodules in $\mathsf{Bim_{fgp}}(M)$ to $\mathsf{Bim_{fgp}}(|Q|)$.
By Proposition \ref{Chi(M)locQsubsetChi(Q)}, for $X\in \tilde{\chi}(M)_Q^{\operatorname{loc}}$, $|X| \in \tilde{\chi}(|Q|)$, and by Proposition \ref{Chi(Q)subsetChi(M)locQ}, every bimodule in $\tilde{\chi}(|Q|)$ arises in this way.
Since
$\tilde{\chi}(M)_Q^{\operatorname{loc}}$ is a full subcategory of the $Q-Q$ bimodules in $\mathsf{Bim_{fgp}}(M)$,
$\tilde{\chi}(|Q|)$ is a full subcategory of $\mathsf{Bim_{fgp}}(|Q|)$, and realization $|\,\cdot\,|$ is fully faithful,
it restricts to a unitary tensor equivalence $\tilde{\chi}(M)_Q^{\operatorname{loc}} \to \tilde{\chi}(|Q|)$.
It remains to verify that $|\,\cdot\,|:\tilde{\chi}(M)_Q^{\operatorname{loc}} \to \tilde{\chi}(|Q|)$ is braided, i.e., the following diagram commutes.
$$
\begin{tikzcd}
{|X| \boxtimes_{|Q|} |Y|}
\arrow[d,"\mu_{X,Y}"]
\arrow[rrr,"u_{X,Y}^{|Q|}"]
\arrow[dr, dashed]
&&&
{|Y| \boxtimes_{|Q|} |X|}
\arrow[d,"\mu_{Y,X}"]
\\
{|X\otimes_Q Y|}
\arrow[r,hookrightarrow]
&
{|X\boxtimes_M Y|}
\arrow[r,"|u_{X,Y}^M|"]
&
{|Y\boxtimes_M X|}
\arrow[ur, dashed]
\arrow[r, two heads]
&
{|Y\otimes_Q X|}
\end{tikzcd}
$$
The two triangles on either side commute by \eqref{eq:CanonicalProjectorTriangle}, so it remains to prove the inner square commutes.
Graphically denoting the $\rm II_1$ factor $|Q|$ and the canonical projector $|X\boxtimes_M Y| \to |X|\otimes_{|Q|} |Y|$ as in \eqref{eq:CanonicalProjectors},
as realization is fully faithful,
this is the condition that
\[
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-1.35,-.9) rectangle (1.1,.9);
\filldraw[\QrColor] (-.8,-.9) rectangle (1.1,.9);
\filldraw[\rColor] (-1.2,-.9) rectangle (-.8,.9);
\filldraw[\rColor] (0,-.5) -- (.4,-.5) arc (-90:0:.4cm) -- (.8,.1) arc (0:90:.4cm) -- (0,.5) arc (90:180:.4cm) -- (-.4,-.1) arc (180:270:.4cm);
\end{scope}
\draw[dashed] (-1.2,-.9) -- (-1.2,.9);
\draw[black,thick] (-.8,-.9) -- (-.8,.9);
\draw[black,thick] (0,-.5) -- (.4,-.5) arc (-90:0:.4cm) -- (.8,.1) arc (0:90:.4cm) -- (0,.5) arc (90:180:.4cm) -- (-.4,-.1) arc (180:270:.4cm);
\draw[blue,thick] (.4,-.9) -- (.4,-.5) .. controls ++(90:.45cm) and ++(270:.45cm) .. (0,.5) -- (0,.9);
\filldraw[\rColor] (.2,0) circle (.05cm);
\draw[\XColor,thick] (0,-.9) -- (0,-.5) .. controls ++(90:.45cm) and ++(270:.45cm) .. (.4,.5) -- (.4,.9);
}
\overset{?}{=}
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.95,-.5) rectangle (.7,.5);
\filldraw[\QrColor] (-.4,-.9) rectangle (.7,.9);
\filldraw[\rColor] (-.8,-.9) rectangle (-.4,.9);
\end{scope}
\draw[dashed] (-.8,-.5) -- (-.8,.5);
\draw[black,thick] (-.4,-.5) -- (-.4,.5);
\draw[blue,thick] (.4,-.5) .. controls ++(90:.45cm) and ++(270:.45cm) .. (0,.5);
\filldraw[\QrColor] (.2,0) circle (.05cm);
\draw[\XColor,thick] (0,-.5) .. controls ++(90:.45cm) and ++(270:.45cm) .. (.4,.5);
}\,.
\]
Let $\{c_k\}$ be a $Y_{|Q|}$-basis, and let $\{b_i^{(n)}\}$ be an approximately inner $X_{|Q|}$-basis.
Let $\{q_l\}$ be a $Q_M$-basis and $\{q_j^{(n)}\}$ be an approximately inner $Q_M$-basis.
According to Proposition \ref{prop:(A)PPBfromQtoM},
$\{c_kq_l\}$ is a $Y_M$ basis and $\{b_i^{(n)}q_j^{(n)}\}$ is an approximately inner $X_M$ basis.
Then
\begin{align*}
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.95,-.9) rectangle (1.1,.9);
\filldraw[\QrColor] (-.4,-.9) rectangle (1.1,.9);
\filldraw[\rColor] (-.8,-.9) -- (-.4,-.9) arc (180:90:.4cm) -- (.4,-.5) arc (-90:0:.4cm) -- (.8,.1) arc (0:90:.4cm) -- (0,.5) arc (270:180:.4cm) -- (-.8,.9);
\end{scope}
\draw[dashed] (-.8,-.9) -- (-.8,.9);
\draw[black,thick] (-.4,-.9) arc (180:90:.4cm) -- (.4,-.5) arc (-90:0:.4cm) -- (.8,.1) arc (0:90:.4cm) -- (0,.5) arc (270:180:.4cm);
\draw[blue,thick] (.4,-.9) -- (.4,-.5) .. controls ++(90:.45cm) and ++(270:.45cm) .. (0,.5) -- (0,.9);
\filldraw[\rColor] (.2,0) circle (.05cm);
\draw[\XColor,thick] (0,-.9) -- (0,-.5) .. controls ++(90:.45cm) and ++(270:.45cm) .. (.4,.5) -- (.4,.9);
}
&\approx
\sum_{i,j,k,l}
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-1.55,-5.1) rectangle (1.3,5.1);
\filldraw[\rColor] (1.3,-5.1) -- (-1.4,-5.1) -- (-1.4,-3.7) -- (-1,-3.4) .. controls ++(90:.2cm) and ++(270:.2cm) .. (-.6,-3) -- (-.4,-2.7) -- (-.4,-1.5) -- (-.1,-1.5) -- (-.1,-.5) -- (.25,-.5) -- (.25,.5) -- (-.1,.5) -- (-.1,1.5) -- (-.4,1.5) -- (-.4,2.5) -- (-.6,3) .. controls ++(90:.2cm) and ++(270:.2cm) .. (-1,3.4) -- (-1.4,3.7) -- (-1.4,5.1) -- (1.3,5.1);
\filldraw[\QrColor] (1.3,-5.1) -- (-1,-5.1) arc (180:90:.4cm) -- (.2,-4.7) arc (-90:0:.8cm) -- (1,3.9) arc (0:90:.8cm) -- (-.6,4.7) arc (270:180:.4cm) -- (1.3,5.1);
\end{scope}
\draw[dashed] (-1.4,-5.1) -- (-1.4,-3.7);
\draw[dashed] (-1,-3.4) .. controls ++(90:.2cm) and ++(270:.2cm) .. (-.6,-3);
\draw[dashed] (-.4,-2.7) -- (-.4,-1.5);
\draw[dashed] (-.1,-1.5) -- (-.1,-.5);
\draw[dashed] (.25,-.5) -- (.25,.5);
\draw[dashed] (-.1,.5) -- (-.1,1.5);
\draw[dashed] (-.4,1.5) -- (-.4,2.5);
\draw[dashed] (-.6,3) .. controls ++(90:.2cm) and ++(270:.2cm) .. (-1,3.4);
\draw[dashed] (-1.4,3.7) -- (-1.4,5.1);
\draw[\XColor,thick] (-.6,-5.1) -- (-.6,-3.7);
\draw[blue,thick] (.2,-5.1) -- (.2,-1.5);
\draw[\XColor,thick] (.2,1.5) -- (.2,5.1);
\draw[blue,thick] (-.6,3.7) -- (-.6,5.1);
\draw[black,thick] (-1,-5.1) arc (180:90:.4cm) -- (.2,-4.7) arc (-90:0:.8cm) -- (1,3.9) arc (0:90:.8cm) -- (-.6,4.7) arc (270:180:.4cm);
\draw[black,thick] (-.2,2.7) -- (-.2,4) arc (0:90:.4cm);
\draw[black,thick] (.6,.5) -- (.6,1.8) arc (0:90:.4cm);
\draw[black,thick] (-.2,-2.7) -- (-.2,-4) arc (0:-90:.4cm);
\draw[black,thick] (.6,-.5) -- (.6,-1.8) arc (0:-90:.4cm);
\filldraw[\XColor] (-.6,-4.4) circle (.05cm);
\filldraw[blue] (.2,-2.2) circle (.05cm);
\filldraw[\XColor] (.2,2.2) circle (.05cm);
\filldraw[blue] (-.6,4.4) circle (.05cm);
\roundNbox{unshaded}{(-1,3.7)}{.3}{.25}{.25}{\scriptsize{$c_k$}};
\roundNbox{unshaded}{(-.4,2.7)}{.3}{.1}{.1}{\scriptsize{$q_l$}};
\roundNbox{unshaded}{(-.1,1.5)}{.3}{.15}{.15}{\scriptsize{$b_i^{(n)}$}};
\roundNbox{unshaded}{(.25,.5)}{.3}{.2}{.2}{\scriptsize{$q_j^{(n)}$}};
\roundNbox{unshaded}{(.25,-.5)}{.3}{.2}{.2}{\scriptsize{$q_l^\dag$}};
\roundNbox{unshaded}{(-.1,-1.5)}{.3}{.15}{.15}{\scriptsize{$c_k^\dag$}};
\roundNbox{unshaded}{(-.4,-2.7)}{.3}{.1}{.1}{\scriptsize{$(q_j^{(\!n\!)})^\dag$}};
\roundNbox{unshaded}{(-1,-3.7)}{.3}{.25}{.25}{\scriptsize{$(b_i^{(n)})^\dag$}};
}
=
\sum_{i,j,k,l}
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-1.55,-5.1) rectangle (1.3,5.1);
\filldraw[\rColor] (1.3,-5.1) -- (-1.4,-5.1) -- (-1.4,-3.7) -- (-1,-3.4) .. controls ++(90:.2cm) and ++(270:.2cm) .. (-.6,-2.8) -- (-.4,-2.7) -- (-.4,-1.5) -- (-.1,-1.5) -- (-.1,-.5) -- (.25,-.5) -- (.25,.5) -- (-.1,.5) -- (-.1,1.5) -- (-.4,1.5) -- (-.4,2.5) -- (-.6,2.8) .. controls ++(90:.2cm) and ++(270:.2cm) .. (-1,3.4) -- (-1.4,3.7) -- (-1.4,5.1) -- (1.3,5.1);
\filldraw[\QrColor] (1.3,-5.1) -- (-1,-5.1) arc (180:90:.4cm) -- (.2,-4.7) arc (-90:0:.8cm) -- (1,3.9) arc (0:90:.8cm) -- (-.6,4.7) arc (270:180:.4cm) -- (1.3,5.1);
\end{scope}
\draw[dashed] (-1.4,-5.1) -- (-1.4,-3.7);
\draw[dashed] (-1,-3.4) .. controls ++(90:.2cm) and ++(270:.2cm) .. (-.6,-2.8);
\draw[dashed] (-.4,-2.7) -- (-.4,-1.5);
\draw[dashed] (-.1,-1.5) -- (-.1,-.5);
\draw[dashed] (.25,-.5) -- (.25,.5);
\draw[dashed] (-.1,.5) -- (-.1,1.5);
\draw[dashed] (-.4,1.5) -- (-.4,2.5);
\draw[dashed] (-.6,2.8) .. controls ++(90:.2cm) and ++(270:.2cm) .. (-1,3.4);
\draw[dashed] (-1.4,3.7) -- (-1.4,5.1);
\draw[\XColor,thick] (-.6,-5.1) -- (-.6,-3.7);
\draw[blue,thick] (.2,-5.1) -- (.2,-1.5);
\draw[\XColor,thick] (.2,1.5) -- (.2,5.1);
\draw[blue,thick] (-.6,3.7) -- (-.6,5.1);
\draw[black,thick] (-1,-5.1) arc (180:90:.4cm) -- (.2,-4.7) arc (-90:0:.8cm) -- (1,3.9) arc (0:90:.8cm) -- (-.6,4.7) arc (270:180:.4cm);
\draw[black,thick] (-.2,2.8) arc (180:90:.4cm);
\draw[black,thick] (.6,.5) -- (.6,2.8) arc (180:90:.4cm);
\draw[black,thick] (-.2,-2.8) arc (180:270:.4cm);
\draw[black,thick] (.6,-.5) -- (.6,-2.8) arc (180:270:.4cm);
\filldraw[blue] (.2,-3.2) circle (.05cm);
\filldraw[\XColor] (.2,3.2) circle (.05cm);
\filldraw[black] (1,-3.2) circle (.05cm);
\filldraw[black] (1,3.2) circle (.05cm);
\roundNbox{unshaded}{(-1,3.7)}{.3}{.25}{.25}{\scriptsize{$c_k$}};
\roundNbox{unshaded}{(-.4,2.5)}{.3}{.1}{.1}{\scriptsize{$q_l$}};
\roundNbox{unshaded}{(-.1,1.5)}{.3}{.15}{.15}{\scriptsize{$b_i^{(n)}$}};
\roundNbox{unshaded}{(.25,.5)}{.3}{.2}{.2}{\scriptsize{$q_j^{(n)}$}};
\roundNbox{unshaded}{(.25,-.5)}{.3}{.2}{.2}{\scriptsize{$q_l^\dag$}};
\roundNbox{unshaded}{(-.1,-1.5)}{.3}{.15}{.15}{\scriptsize{$c_k^\dag$}};
\roundNbox{unshaded}{(-.4,-2.5)}{.3}{.1}{.1}{\scriptsize{$(q_j^{(\!n\!)})^\dag$}};
\roundNbox{unshaded}{(-1,-3.7)}{.3}{.25}{.25}{\scriptsize{$(b_i^{(n)})^\dag$}};
}
\approx
\sum_{i,j,k,l}
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-1.55,-5.1) rectangle (1.3,5.1);
\filldraw[\rColor] (1.3,-5.1) -- (-1.4,-5.1) -- (-1.4,-3.7) -- (-1,-3.4) .. controls ++(90:.2cm) and ++(270:.2cm) .. (-.6,-2.8) -- (-.4,-2.7) -- (-.4,-1.5) -- (-.1,-1.5) -- (-.1,-.5) -- (.25,-.5) -- (.25,.5) -- (-.1,.5) -- (-.1,1.5) -- (-.4,1.5) -- (-.4,2.5) -- (-.6,2.8) .. controls ++(90:.2cm) and ++(270:.2cm) .. (-1,3.4) -- (-1.4,3.7) -- (-1.4,5.1) -- (1.3,5.1);
\filldraw[\QrColor] (1.3,-5.1) -- (-1,-5.1) arc (180:90:.4cm) -- (-.2,-4.7) arc (-90:0:1.2cm) -- (1,3.5) arc (0:90:1.2cm) -- (-.6,4.7) arc (270:180:.4cm) -- (1.3,5.1);
\end{scope}
\draw[dashed] (-1.4,-5.1) -- (-1.4,-3.7);
\draw[dashed] (-1,-3.4) .. controls ++(90:.2cm) and ++(270:.2cm) .. (-.6,-2.8);
\draw[dashed] (-.4,-2.7) -- (-.4,-1.5);
\draw[dashed] (-.1,-1.5) -- (-.1,-.5);
\draw[dashed] (.25,-.5) -- (.25,.5);
\draw[dashed] (-.1,.5) -- (-.1,1.5);
\draw[dashed] (-.4,1.5) -- (-.4,2.5);
\draw[dashed] (-.6,2.8) .. controls ++(90:.2cm) and ++(270:.2cm) .. (-1,3.4);
\draw[dashed] (-1.4,3.7) -- (-1.4,5.1);
\draw[\XColor,thick] (-.6,-5.1) -- (-.6,-3.7);
\draw[blue,thick] (-.2,-5.1) -- (-.2,-2.5);
\draw[\XColor,thick] (-.2,2.5) -- (-.2,5.1);
\draw[blue,thick] (-.6,3.7) -- (-.6,5.1);
\draw[black,thick] (-1,-5.1) arc (180:90:.4cm) -- (-.2,-4.7) arc (-90:0:1.2cm) -- (1,3.5) arc (0:90:1.2cm) -- (-.6,4.7) arc (270:180:.4cm);
\draw[black,thick] (.2,1.5) -- (.2,2.8) arc (0:90:.4cm);
\draw[black,thick] (.6,.5) -- (.6,2.8) arc (180:90:.4cm);
\draw[black,thick] (.2,-1.5) -- (.2,-2.8) arc (0:-90:.4cm);
\draw[black,thick] (.6,-.5) -- (.6,-2.8) arc (180:270:.4cm);
\filldraw[blue] (-.2,-3.2) circle (.05cm);
\filldraw[\XColor] (-.2,3.2) circle (.05cm);
\filldraw[black] (1,-3.2) circle (.05cm);
\filldraw[black] (1,3.2) circle (.05cm);
\roundNbox{unshaded}{(-1,3.7)}{.3}{.25}{.25}{\scriptsize{$c_k$}};
\roundNbox{unshaded}{(-.4,2.5)}{.3}{.1}{.1}{\scriptsize{$b_i^{(n)}$}};
\roundNbox{unshaded}{(-.1,1.5)}{.3}{.15}{.15}{\scriptsize{$q_l$}};
\roundNbox{unshaded}{(.25,.5)}{.3}{.2}{.2}{\scriptsize{$q_j^{(n)}$}};
\roundNbox{unshaded}{(.25,-.5)}{.3}{.2}{.2}{\scriptsize{$q_l^\dag$}};
\roundNbox{unshaded}{(-.1,-1.5)}{.3}{.15}{.15}{\scriptsize{$(q_j^{(\!n\!)})^\dag$}};
\roundNbox{unshaded}{(-.4,-2.5)}{.3}{.1}{.1}{\scriptsize{$c_k^\dag$}};
\roundNbox{unshaded}{(-1,-3.7)}{.3}{.25}{.25}{\scriptsize{$(b_i^{(n)})^\dag$}};
}
\\
&\approx
\sum_{i,k}
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-1.55,-3) rectangle (1.3,3);
\filldraw[\rColor] (1.3,-3) -- (-1.4,-3) -- (-1.4,-2) -- (-1,-1.7) .. controls ++(90:.2cm) and ++(270:.2cm) .. (-.6,-1.1) -- (-.4,-1) -- (-.4,1) -- (-.6,1.1) .. controls ++(90:.2cm) and ++(270:.2cm) .. (-1,1.7) -- (-1.4,2) -- (-1.4,3) -- (1.3,3);
\filldraw[\QrColor] (1.3,-3) -- (-1,-3) arc (180:90:.4cm) -- (-.2,-2.6) arc (-90:0:1.2cm) -- (1,1.4) arc (0:90:1.2cm) -- (-.6,2.6) arc (270:180:.4cm) -- (1.3,3);
\end{scope}
\draw[dashed] (-1.4,-3) -- (-1.4,-2);
\draw[dashed] (-1,-1.7) .. controls ++(90:.2cm) and ++(270:.2cm) .. (-.6,-1.1);
\draw[dashed] (-.4,-1) -- (-.4,1);
\draw[dashed] (-.6,1.1) .. controls ++(90:.2cm) and ++(270:.2cm) .. (-1,1.7);
\draw[dashed] (-1.4,2) -- (-1.4,3);
\draw[black,thick] (.6,-.5) .. controls ++(90:.45cm) and ++(270:.45cm) .. (.2,.5);
\filldraw[\rColor] (.4,0) circle (.05cm);
\draw[black,thick] (.2,-.5) .. controls ++(90:.45cm) and ++(270:.45cm) .. (.6,.5);
\draw[black,thick] (-1,-3) arc (180:90:.4cm) -- (-.2,-2.6) arc (-90:0:1.2cm) -- (1,1.4) arc (0:90:1.2cm) -- (-.6,2.6) arc (270:180:.4cm);
\draw[black,thick] (.2,.5) -- (.2,1.1) arc (0:90:.4cm);
\draw[black,thick] (.6,.5) -- (.6,1.1) arc (180:90:.4cm);
\draw[black,thick] (.2,-.5) -- (.2,-1.1) arc (0:-90:.4cm);
\draw[black,thick] (.6,-.5) -- (.6,-1.1) arc (180:270:.4cm);
\draw[\XColor,thick] (-.6,-3) -- (-.6,-2);
\draw[blue,thick] (-.2,-3) -- (-.2,-.8);
\draw[\XColor,thick] (-.2,.8) -- (-.2,3);
\draw[blue,thick] (-.6,2) -- (-.6,3);
\filldraw[blue] (-.2,-1.5) circle (.05cm);
\filldraw[\XColor] (-.2,1.5) circle (.05cm);
\filldraw[black] (1,-1.5) circle (.05cm);
\filldraw[black] (1,1.5) circle (.05cm);
\roundNbox{unshaded}{(-1,2)}{.3}{.25}{.25}{\scriptsize{$c_k$}};
\roundNbox{unshaded}{(-.4,.8)}{.3}{.1}{.1}{\scriptsize{$b_i^{(n)}$}};
\roundNbox{unshaded}{(-.4,-.8)}{.3}{.1}{.1}{\scriptsize{$c_k^\dag$}};
\roundNbox{unshaded}{(-1,-2)}{.3}{.25}{.25}{\scriptsize{$(b_i^{(n)})^\dag$}};
}
=
\sum_{i,k}
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-1.55,-2.5) rectangle (.5,2.5);
\filldraw[\rColor] (.5,-2.5) -- (-1.4,-2.5) -- (-1.4,-1.5) -- (-1,-1.2) .. controls ++(90:.2cm) and ++(270:.2cm) .. (-.6,-.8) -- (-.4,-.5) -- (-.4,.5) -- (-.6,.8) .. controls ++(90:.2cm) and ++(270:.2cm) .. (-1,1.2) -- (-1.4,1.5) -- (-1.4,2.5) -- (.5,2.5);
\filldraw[\QrColor] (.5,-2.5) -- (-1,-2.5) arc (180:90:.4cm) -- (-.2,-2.1) arc (-90:0:.4cm) -- (.2,1.7) arc (0:90:.4cm) -- (-.6,2.1) arc (270:180:.4cm) -- (.5,2.5);
\end{scope}
\draw[dashed] (-1.4,-2.5) -- (-1.4,-1.5);
\draw[dashed] (-1,-1.2) .. controls ++(90:.2cm) and ++(270:.2cm) .. (-.6,-.8);
\draw[dashed] (-.4,-.5) -- (-.4,.5);
\draw[dashed] (-.6,.8) .. controls ++(90:.2cm) and ++(270:.2cm) .. (-1,1.2);
\draw[dashed] (-1.4,1.5) -- (-1.4,2.5);
\draw[black,thick] (-1,-2.5) arc (180:90:.4cm) -- (-.2,-2.1) arc (-90:0:.4cm) -- (.2,1.7) arc (0:90:.4cm) -- (-.6,2.1) arc (270:180:.4cm);
\draw[\XColor,thick] (-.6,-2.5) -- (-.6,-1.5);
\draw[blue,thick] (-.2,-2.5) -- (-.2,-.5);
\draw[\XColor,thick] (-.2,.5) -- (-.2,2.5);
\draw[blue,thick] (-.6,1.5) -- (-.6,2.5);
\roundNbox{unshaded}{(-1,1.5)}{.3}{.25}{.25}{\scriptsize{$c_k$}};
\roundNbox{unshaded}{(-.4,.5)}{.3}{.1}{.1}{\scriptsize{$b_i^{(n)}$}};
\roundNbox{unshaded}{(-.4,-.5)}{.3}{.1}{.1}{\scriptsize{$c_k^\dag$}};
\roundNbox{unshaded}{(-1,-1.5)}{.3}{.25}{.25}{\scriptsize{$(b_i^{(n)})^\dag$}};
}
=
\sum_{i,k}
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-1.55,-3) rectangle (.9,3);
\filldraw[\rColor] (1.3,-3) -- (-1.4,-3) -- (-1.4,-1.9) -- (-1,-1.6) .. controls ++(90:.2cm) and ++(270:.2cm) .. (-.2,-.8) -- (0,-.5) -- (0,.5) -- (-.2,.8) .. controls ++(90:.2cm) and ++(270:.2cm) .. (-1,1.6) -- (-1.4,1.9) -- (-1.4,3) -- (1.3,3);
\filldraw[\QrColor] (1.3,-3) -- (-1,-3) arc (180:90:.4cm) arc (-90:0:.4cm) -- (-.2,-1.6) arc (180:90:.4cm) arc (-90:0:.4cm) -- (.6,.8) arc (0:90:.4cm) arc (270:180:.4cm) -- (-.2,2.2) arc (0:90:.4cm) arc (270:180:.4cm) -- (1.3,3);
\end{scope}
\draw[dashed] (-1.4,-3) -- (-1.4,-1.9);
\draw[dashed] (-1,-1.6) .. controls ++(90:.2cm) and ++(270:.2cm) .. (-.2,-.8);
\draw[dashed] (0,-.5) -- (0,.5);
\draw[dashed] (-.2,.8) .. controls ++(90:.2cm) and ++(270:.2cm) .. (-1,1.6);
\draw[dashed] (-1.4,1.9) -- (-1.4,3);
\draw[black,thick] (-1,-3) arc (180:90:.4cm) arc (-90:0:.4cm) -- (-.2,-1.6) arc (180:90:.4cm) arc (-90:0:.4cm) -- (.6,.8) arc (0:90:.4cm) arc (270:180:.4cm) -- (-.2,2.2) arc (0:90:.4cm) arc (270:180:.4cm);
\draw[\XColor,thick] (-.6,-3) -- (-.6,-1.9);
\draw[blue,thick] (.2,-3) -- (.2,-.5);
\draw[\XColor,thick] (.2,.5) -- (.2,3);
\draw[blue,thick] (-.6,1.9) -- (-.6,3);
\roundNbox{unshaded}{(-1,1.9)}{.3}{.25}{.25}{\scriptsize{$c_k$}};
\roundNbox{unshaded}{(0,.5)}{.3}{.1}{.1}{\scriptsize{$b_i^{(n)}$}};
\roundNbox{unshaded}{(0,-.5)}{.3}{.1}{.1}{\scriptsize{$c_k^\dag$}};
\roundNbox{unshaded}{(-1,-1.9)}{.3}{.25}{.25}{\scriptsize{$(b_i^{(n)})^\dag$}};
}
\approx
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.95,-.5) rectangle (.7,.5);
\filldraw[\QrColor] (-.4,-.9) rectangle (.7,.9);
\filldraw[\rColor] (-.8,-.9) rectangle (-.4,.9);
\end{scope}
\draw[dashed] (-.8,-.5) -- (-.8,.5);
\draw[black,thick] (-.4,-.5) -- (-.4,.5);
\draw[blue,thick] (.4,-.5) .. controls ++(90:.45cm) and ++(270:.45cm) .. (0,.5);
\filldraw[\QrColor] (.2,0) circle (.05cm);
\draw[\XColor,thick] (0,-.5) .. controls ++(90:.45cm) and ++(270:.45cm) .. (.4,.5);
}\,.
\end{align*}
The first $\approx$ uses \eqref{eq:OverBraiding} for $X\boxtimes_M Y$, and the second equality uses associativity of the bimodule actions.
The third $\approx$ uses $X$ is approximately inner over $|Q|$, $Y$ is centrally trivial over $|Q|$ and $(q^{(n)}_l)$ is a central sequence in $|Q|$, and the fourth $\approx$ uses \eqref{eq:ExpandBraidingsQ}.
The fifth equality uses an argument similar to \eqref{eq:UseCommutativityForRichtAction}.
The sixth equality is just isotopy,
and the final $\approx$ uses \eqref{eq:OverBraiding} for $X \boxtimes_{|Q|} Y$.
Finally, since
\[
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.55,-.6) rectangle (.7,.6);
\filldraw[\rColor] (-.4,-.9) rectangle (.7,.9);
\filldraw[\QrColor] (0,-.9) rectangle (.4,.9);
\end{scope}
\draw[dashed] (-.4,-.6) -- (-.4,.6);
\draw[black,thick] (0,-.6) -- (0,.6);
\draw[black,thick] (.4,-.6) -- (.4,.6);
}
=
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.55,-.6) rectangle (.7,.6);
\filldraw[\rColor] (-.4,-.9) rectangle (.7,.9);
\filldraw[\QrColor] (.4,-.6) arc (0:90:.4cm) -- (0,-.6);
\filldraw[\QrColor] (.4,.6) arc (0:-90:.4cm) -- (0,.6);
\end{scope}
\draw[dashed] (-.4,-.6) -- (-.4,.6);
\draw[black,thick] (0,-.6) -- (0,.6);
\draw[black,thick] (.4,-.6) arc (0:90:.4cm);
\draw[black,thick] (.4,.6) arc (0:-90:.4cm);
\filldraw[black] (0,-.2) circle (.05cm);
\filldraw[black] (0,.2) circle (.05cm);
}
=
\sum_j
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.55,-1.5) rectangle (.7,1.5);
\filldraw[\rColor] (.7,-1.5) -- (-.4,-1.5) -- (-.4,-.5) -- (-.2,-.5) -- (-.2,.5) -- (-.4,.5) -- (-.4,1.5) -- (.7,1.5);
\filldraw[\QrColor] (.4,-1.5) arc (0:90:.4cm) -- (0,-1.5);
\filldraw[\QrColor] (.4,1.5) arc (0:-90:.4cm) -- (0,1.5);
\end{scope}
\draw[dashed] (-.4,-1.5) -- (-.4,-.5);
\draw[dashed] (-.2,-.5) -- (-.2,.5);
\draw[dashed] (-.4,.5) -- (-.4,1.5);
\draw[black,thick] (0,-1.5) -- (0,-.5);
\draw[black,thick] (0,.5) -- (0,1.5);
\draw[black,thick] (.4,-1.5) arc (0:90:.4cm);
\draw[black,thick] (.4,1.5) arc (0:-90:.4cm);
\filldraw[black] (0,-1.1) circle (.05cm);
\filldraw[black] (0,1.1) circle (.05cm);
\roundNbox{unshaded}{(-.2,-.5)}{.3}{.1}{.1}{\scriptsize{$q_j^\dag$}};
\roundNbox{unshaded}{(-.2,.5)}{.3}{.1}{.1}{\scriptsize{$q_j$}};
}
\qquad\text{and}\qquad
\tikzmath{
\filldraw[\rColor] (0,-.8) rectangle (.8,.8);
\filldraw[\QrColor] (0,-.4) arc (-90:0:.4cm) arc (180:90:.4cm) -- (.8,.8) -- (0,.8);
\draw[black,thick] (0,-.8) -- (0,.8);
\draw[black,thick] (.8,-.8) -- (.8,.8);
\draw[black,thick] (0,-.4) arc (-90:0:.4cm) arc (180:90:.4cm);
\filldraw[black] (0,-.4) circle (.05cm);
\filldraw[black] (.8,.4) circle (.05cm);
}
=
\tikzmath{
\filldraw[\rColor] (0,-.8) rectangle (.8,.8);
\filldraw[\QrColor] (0,.4) arc (90:0:.4cm) arc (180:270:.4cm) -- (.8,.8) -- (0,.8);
\draw[black,thick] (0,-.8) -- (0,.8);
\draw[black,thick] (.8,-.8) -- (.8,.8);
\draw[black,thick] (0,.4) arc (90:0:.4cm) arc (180:270:.4cm);
\filldraw[black] (0,.4) circle (.05cm);
\filldraw[black] (.8,-.4) circle (.05cm);
}\,,
\]
we have
\[
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-1.35,-.9) rectangle (1.1,.9);
\filldraw[\QrColor] (-.8,-.9) rectangle (1.1,.9);
\filldraw[\rColor] (-1.2,-.9) rectangle (-.8,.9);
\filldraw[\rColor] (0,-.5) -- (.4,-.5) arc (-90:0:.4cm) -- (.8,.1) arc (0:90:.4cm) -- (0,.5) arc (90:180:.4cm) -- (-.4,-.1) arc (180:270:.4cm);
\end{scope}
\draw[dashed] (-1.2,-.9) -- (-1.2,.9);
\draw[black,thick] (-.8,-.9) -- (-.8,.9);
\draw[black,thick] (0,-.5) -- (.4,-.5) arc (-90:0:.4cm) -- (.8,.1) arc (0:90:.4cm) -- (0,.5) arc (90:180:.4cm) -- (-.4,-.1) arc (180:270:.4cm);
\draw[blue,thick] (.4,-.9) -- (.4,-.5) .. controls ++(90:.45cm) and ++(270:.45cm) .. (0,.5) -- (0,.9);
\filldraw[\rColor] (.2,0) circle (.05cm);
\draw[\XColor,thick] (0,-.9) -- (0,-.5) .. controls ++(90:.45cm) and ++(270:.45cm) .. (.4,.5) -- (.4,.9);
}
=
\sum_j
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-1.35,-2.2) rectangle (1.1,2.2);
\filldraw[\rColor] (1.1,-2.2) -- (-1.2,-2.2) -- (-1.2,-1.2) -- (-1,-.9) -- (-1,.9) -- (-1.2,1.2) -- (-1.2,2.2) -- (1.1,2.2);
\filldraw[\QrColor] (1.1,-2.2) -- (-.8,-2.2) -- (-.8,-1.9) arc (-90:0:.4cm) -- (-.4,-.9) arc (180:90:.4cm) -- (.4,-.5) arc (-90:0:.4cm) -- (.8,.1) arc (0:90:.4cm) -- (0,.5) arc (270:180:.4cm) -- (-.4,1.5) arc (0:90:.4cm) -- (-.8,2.2) -- (1.1,2.2);
\end{scope}
\draw[dashed] (-1.2,-2.2) -- (-1.2,-1.2);
\draw[dashed] (-1,-.9) -- (-1,.9);
\draw[dashed] (-1.2,1.2) -- (-1.2,2.2);
\draw[black,thick] (-.8,1.5) -- (-.8,2.2);
\draw[black,thick] (-.8,-2.2) -- (-.8,-1.5);
\draw[black,thick] (-.8,-1.9) arc (-90:0:.4cm) -- (-.4,-.9) arc (180:90:.4cm) -- (.4,-.5) arc (-90:0:.4cm) -- (.8,.1) arc (0:90:.4cm) -- (0,.5) arc (270:180:.4cm) -- (-.4,1.5) arc (0:90:.4cm);
\draw[blue,thick] (.4,-2.2) -- (.4,-.5) .. controls ++(90:.45cm) and ++(270:.45cm) .. (0,.5) -- (0,2.2);
\filldraw[\rColor] (.2,0) circle (.05cm);
\draw[\XColor,thick] (0,-2.2) -- (0,-.5) .. controls ++(90:.45cm) and ++(270:.45cm) .. (.4,.5) -- (.4,2.2);
\filldraw[black] (-.8,-1.9) circle (.05cm);
\filldraw[black] (-.8,1.9) circle (.05cm);
\roundNbox{unshaded}{(-1,-1.2)}{.3}{.1}{.1}{\scriptsize{$q_j^\dag$}};
\roundNbox{unshaded}{(-1,1.2)}{.3}{.1}{.1}{\scriptsize{$q_j$}};
}
=
\sum_j
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-1.45,-1.8) rectangle (.7,1.8);
\filldraw[\rColor] (.7,-1.8) -- (-1.2,-1.8) -- (-1.2,-.8) -- (-1,-.8) -- (-1,.8) -- (-1.2,.8) -- (-1.2,1.8) -- (.7,1.8);
\filldraw[\QrColor] (.7,-1.8) -- (-.8,-1.8) -- (-.8,-1.5) arc (-90:0:.4cm) -- (-.4,1.1) arc (0:90:.4cm) -- (-.8,1.8) -- (.7,1.8);
\end{scope}
\draw[dashed] (-1.2,-1.8) -- (-1.2,-.8);
\draw[dashed] (-1,-.8) -- (-1,.8);
\draw[dashed] (-1.2,.8) -- (-1.2,1.8);
\draw[black,thick] (-.8,-1.8) -- (-.8,-.8);
\draw[black,thick] (-.8,.8) -- (-.8,1.8);
\draw[black,thick] (-.8,-1.5) arc (-90:0:.4cm) -- (-.4,1.1) arc (0:90:.4cm);
\draw[blue,thick] (.4,-1.8) -- (.4,-.5) .. controls ++(90:.45cm) and ++(270:.45cm) .. (0,.5) -- (0,1.8);
\filldraw[\QrColor] (.2,0) circle (.05cm);
\draw[\XColor,thick] (0,-1.8) -- (0,-.5) .. controls ++(90:.45cm) and ++(270:.45cm) .. (.4,.5) -- (.4,1.8);
\filldraw[black] (-.8,-1.5) circle (.05cm);
\filldraw[black] (-.8,1.5) circle (.05cm);
\roundNbox{unshaded}{(-1,-.8)}{.3}{.1}{.1}{\scriptsize{$q_j^\dag$}};
\roundNbox{unshaded}{(-1,.8)}{.3}{.1}{.1}{\scriptsize{$q_j$}};
}
=
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.95,-.5) rectangle (.7,.5);
\filldraw[\QrColor] (-.4,-.9) rectangle (.7,.9);
\filldraw[\rColor] (-.8,-.9) rectangle (-.4,.9);
\end{scope}
\draw[dashed] (-.8,-.5) -- (-.8,.5);
\draw[black,thick] (-.4,-.5) -- (-.4,.5);
\draw[blue,thick] (.4,-.5) .. controls ++(90:.45cm) and ++(270:.45cm) .. (0,.5);
\filldraw[\QrColor] (.2,0) circle (.05cm);
\draw[\XColor,thick] (0,-.5) .. controls ++(90:.45cm) and ++(270:.45cm) .. (.4,.5);
}\,.
\qedhere
\]
\end{proof}
\section{Calculation of \texorpdfstring{$\tilde{\chi}(M_\infty)$}{Minfty} for a non-Gamma finite depth \texorpdfstring{$\rm II_1$}{II1} subfactor}
\label{sec:Calculation}
In this section, we calculate $\tilde{\chi}(M_\infty)$ for the inductive limit $\rm II_1$ factor obtained from iterating Jones' basic construction for a finite depth finite index non-Gamma $\rm II_1$ subfactor $N\subseteq M$.
These examples are motivated by \cite{MR2661553}.
Suppose $N\subseteq M$ a finite depth, finite index $\rm II_1$ subfactor, and let $\cC={}_N\cC_N$ denote the unitary fusion category of $N-N$ bimodules generated by ${}_NM_N$.
The results of \cite[\S3 and 4]{MR3801484} give a bijective correspondence between equivalence classes of (bifinite) bimodules of $M_{\infty}$ which restrict to $R\otimes N$-bimodules of the form $\cC^{\operatorname{op}}\boxtimes \cC$ and objects of the Drinfeld center $\cZ(\cC)$.
The main goal of this section is to extend this bijection to a fully faithful unitary tensor functor $\Phi:\cZ(\cC)\to \mathsf{Bim_{fgp}}(M_\infty)$ such that when $N$ is non-Gamma, $\Phi$ takes values in $\tilde{\chi}(M_\infty)$ and is a braided unitary equivalence.
To do so, we rely on the Q-system realization language from \cite{2105.12010} together with the coend realization viewpoint of \cite{MR3948170}.
We begin this section with some basics on unitary fusion categories and subfactor standard invariants.
\subsection{Basics on unitary fusion categories and subfactor standard invariants}
A \emph{unitary fusion category} is a unitary tensor category with only finitely many isomorphism classes of simple objects.
A unitary fusion category $\cC$ has three commuting involutions $\dag, \vee, \overline{\,\cdot\,}$, and the composite of any two is the third.
Here, $\vee$ is the unique unitary dual functor \cite{MR2091457,MR4133163} giving the canonical unitary spherical structure of $\cC$ \cite{MR1444286}, and we may define $\overline{\,\cdot\,}:=\vee\dag = \dag \vee$.
\begin{defn}
\label{def:Z(C)}
The \emph{Drinfeld center} of a unitary fusion category $\cC$ is $\cZ(\cC)=\operatorname{End}_{\cC-\cC}(\cC)$, the Morita dual of $\cC^{\rm mp}\boxtimes \cC$ acting on $\cC$ by $(a^{\rm mp}\boxtimes b)\rhd c := b\otimes c \otimes a$, where $\cC^{\rm mp}$ is the monoidal opposite of $\cC$ from Remark \ref{rem:OppositeWithBraiding}.
Note that the unitary dual functor $\vee$ gives a unitary tensor equivalence $\cC^{\rm mp}\to \cC^{\operatorname{op}}$, the opposite fusion category with the opposite arrows, but the same tensor product.
It is useful in the subsections below to identify $\cZ(\cC)$ with $\operatorname{End}_{\cC^{\operatorname{op}}\boxtimes \cC}(\cC)$
with the action $(a^{\operatorname{op}}\boxtimes b)\rhd c := b\otimes c \otimes \overline{a}$.
Now $\cZ(\cC)$ has another description in terms of pairs $(z,\sigma_z)$ of an object $z\in \cC$ equipped with a half-braiding $\sigma_z$, where $\cZ(\cC)$ acts on $\cC$ via the forgetful functor $(z,\sigma_z)\mapsto z$ \cite[\S7.13 and 8.5]{MR3242743}.
Our convention for the half-braiding $\sigma_z$ is that the strands for objects in $\cC$ pass over the $z$-strand:
$$
\sigma_{c,z}
:=
\tikzmath{
\draw[blue,thick] (.4,-.5) node[below]{$\scriptstyle z$} .. controls ++(90:.45cm) and ++(270:.45cm) .. (0,.5);
\draw[\XColor,thick, knot] (0,-.5) node[below]{$\scriptstyle c$} .. controls ++(90:.45cm) and ++(270:.45cm) .. (.4,.5);
}
$$
Thus the braiding $(z,\sigma_z)\otimes (w,\rho_w) \to (w, \rho_w)\otimes (z,\sigma_z)$ in $\cZ(\cC)$ is given by $\rho_{z,w}$.
\end{defn}
\begin{defn}
There are many equivalent notions of the standard invariant for a finite index $\rm II_1$ subfactor $N\subseteq M$.
For this article, the standard invariant will mean the $2\times 2$ unitary multitensor category $\cC(N\subseteq M)$ of $N-N$, $N-M$, $M-N$, and $M-M$ bimodules generated by $L^2M$ under $\boxtimes$, $\oplus$, and $\subseteq$, with generating object ${}_NL^2M_M$.
$$
\cC(N\subseteq M) =
\begin{pmatrix}
{}_N\cC_N & {}_N\cC_M
\\
{}_M\cC_N & {}_M \cC_M
\end{pmatrix}
$$
Observe that $\cC(N\subseteq M)$ is multifusion if and only if $N\subseteq M$ is finite depth.
In this case, the corners ${}_N\cC_N$ and ${}_M \cC_M$ of $N-N$ and $M-M$ bimodules generated by $L^2M$ respectively are unitary fusion categories which are Morita equivalent,
and thus share the same Drinfeld center $\cZ(\cC)$.
\end{defn}
\begin{rem}
\label{rem:MoritaEquivalenceRealization}
Suppose $\cC$ and $\cD$ are two unitary fusion categories and ${}_\cC \cM_\cD$ is an indecomposable unitary $\cC-\cD$ bimodule category witnessing a Morita equivalence.
Using the internal hom \cite{MR1976459} (see also \cite[Appendix~A]{MR3933035}), we can form a $2\times 2$ unitary multifusion category by
\begin{equation}
\label{eq:2x2UMFC}
\begin{pmatrix}
\cC & \cM
\\
\cM^{\operatorname{op}} & \cD
\end{pmatrix}.
\end{equation}
For a simple $X\in \cM$, we get two Q-systems $X\otimes \overline{X} =\underline{\operatorname{End}}_\cC(X)\in \cC$ and $\overline{X}\otimes X=\underline{\operatorname{End}}_\cD(X) \in \cD$.
The map $\operatorname{Ad}(X) : d\mapsto X\otimes d\otimes \overline{X}$ gives a unitary tensor equivalence between $\cD$ and $X\otimes \overline{X}-X\otimes \overline{X}$ bimodules in $\cC$.
A similar result holds on the other side.
Suppose now we have a fully faithful unitary tensor functor $F: \cC\to \mathsf{Bim_{fgp}}(N)$ for a $\rm II_1$ factor $N$.
Then the realization $M:= |X\otimes \overline{X}|$ is a $\rm II_1$ factor containing $N$, and the standard invariant of $N\subseteq M$ is unitarily equivalent to the $2\times 2$ unitary multifusion category \eqref{eq:2x2UMFC} with generator $|X|$ as an $N-M$ bimodule.
By Remark \ref{rem:QBimodules}, we get a fully faithful unitary tensor functor $G: \cD \to \mathsf{Bim_{fgp}}(M)$ from realization as
$$
\cD
\xrightarrow{\operatorname{Ad}(X)} \mathsf{Bim}_\cC(X\otimes \overline{X})
\xrightarrow{|\,\cdot\,|}
\mathsf{Bim_{fgp}}(M).
$$
\end{rem}
We now give an important example of Remark \ref{rem:MoritaEquivalenceRealization} which will be used in this section below.
\begin{ex}
\label{ex:MoritaEquivalenceBetweenCopCandZ(C)}
Let $\cC$ be a unitary fusion category, and consider the $\cC\boxtimes \cC^{\rm mp} - \cZ(\cC)$
Morita equivalence bimodule $\cC$.
One calculates that $\underline{\operatorname{End}}_{\cC^{\rm mp}\boxtimes \cC}(1_\cC) = \bigoplus_{c\in \operatorname{Irr}(\cC)} \overline{c}^{\rm mp}\boxtimes c$, which we call the \emph{symmetric enveloping Q-system} after \cite{MR1302385,MR1729488}. The infinite version of this algebra object plays a very important role for the study of analytic properties of infinite unitary tensor categories \cite{MR3406647}.
Identifying $\cC^{\rm mp}\cong \cC^{\operatorname{op}}$ via $\vee$, which will be useful in the sequel, the symmetric enveloping Q-system is given by $S:=\bigoplus_{c\in \operatorname{Irr}(\cC)} c^{\operatorname{op}}\boxtimes c$.
By Remark \ref{rem:MoritaEquivalenceRealization}, $\cZ(\cC)\cong \mathsf{Bim}_{\cC^{\operatorname{op}}\boxtimes \cC}(S)$.
On the other hand, one calculates that $\underline{\operatorname{End}}_{\cZ(\cC)}(1_\cC) = I(1_\cC)$, where $I: \cC\to \cZ(\cC)$ is the adjoint of the forgetful functor.
\end{ex}
\subsection{Q-system realization as a coend}
Suppose $\cC$ is a unitary fusion category and $G: \cC\to \mathsf{Bim_{fgp}}(N)$ is a unitary tensor functor, where $N$ is a $\rm II_1$ factor.
Given a Q-system $Q\in \cC$, the realization
$|G(Q)|$ is a $\rm II_1$ multifactor (finite direct sum of $\rm II_1$ factors) which is a factor if and only if $Q$ is simple as a $Q-Q$ bimodule in $\cC$.
By the Yoneda lemma, we have a canonical isomorphism of vector spaces
$$
|G(Q)|
:=
\operatorname{Hom}(N_N \to N \boxtimes_N G(Q)_N)
\cong
\bigoplus_{c\in \cC} \cC(c\to Q) \otimes_\bbC G(c).
$$
We graphically represent elements of this tensor product by
$$
\sum_{c\in\operatorname{Irr}(\cC)}
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.7,0) -- (-.7,-1.4) -- (.7,-1.4) -- (.7,0);
\filldraw[\rColor] (-.7,0) -- (-.7,-1.4) -- (.7,-1.4) -- (.7,0);
\end{scope}
\draw[dotted] (-.7,0) -- (.7,0);
\draw[orange,thick] (0,-1.4) -- (0,-.7);
\draw (0,-.7) -- (0,.7);
\draw[black, thick] (0,1) -- (0,1.4);
\roundNbox{unshaded}{(0,-.7)}{.3}{0}{0}{$\xi_c$};
\roundNbox{unshaded}{(0,.7)}{.3}{0}{0}{$f_c$};
\node at (.2,.2) {\scriptsize{$c$}};
\node at (.2,-.2) {\scriptsize{$c$}};
\node at (.2,1.2) {\scriptsize{$Q$}};
}
:=
\sum_{c\in\operatorname{Irr}(\cC)}
\tikzmath{
\draw[black, thick] (0,.7) -- (0,0);
\draw (0,0) -- (0,-.7);
\roundNbox{unshaded}{(0,0)}{.3}{0}{0}{$f_c$};
\node at (.2,.5) {\scriptsize{$Q$}};
\node at (.2,-.5) {\scriptsize{$c$}};
}
\otimes_{\bbC}
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.7,-.7) rectangle (.7,.7);
\filldraw[\rColor] (-.7,-.7) rectangle (.7,.7);
\end{scope}
\draw (0,0) -- (0,.7);
\draw[orange,thick] (0,-.7) -- (0,0);
\roundNbox{unshaded}{(0,0)}{.3}{0}{0}{$\xi_c$};
\node at (0,.9) {\scriptsize{$c$}};
}
=
\sum_{c\in \operatorname{Irr}(\cC)} f_c \otimes_\bbC \xi_c.
$$
Here, the orange line represents the
functor $G^\circ := \operatorname{Forget}\circ \,G$
viewed as a $\rm W^*$-algebra object in $\mathsf{Fun}(\cC \to {\mathsf{Vect}})$ \cite[Prop.~2.18]{MR3948170}, where $\operatorname{Forget}: \mathsf{Bim_{fgp}}(N) \to {\mathsf{Vect}}$ is the forgetful functor.
The shaded half of the diagram is read top to bottom, and the tensorator $G^2_{a,b}$ is denoted by appending an orange trivalent vertex below.
Under this isomorphism of vector spaces, the multiplication and $*$-structure are given by
$$
(f_a\otimes \xi_a)(g_b\otimes \eta_b)
=
\sum_{\substack{ a,b,c\in\operatorname{Irr}(\cC) \\ \alpha\in\operatorname{ONB}(a\otimes b\to c)}}
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-1.2,0) -- (-1.2,-2.1) -- (1.2,-2.1) -- (1.2,0);
\filldraw[\rColor] (-1.2,0) -- (-1.2,-2.1) -- (1.2,-2.1) -- (1.2,0);
\end{scope}
\draw[dotted] (-1.2,0) -- (1.2,0);
\draw (0,-.3) -- (0,.3);
\draw (.4,.7) arc (0:-180:.4cm);
\draw (.4,-.7) arc (0:180:.4cm);
\filldraw (0,-.3) circle (.05cm);
\filldraw (0,.3) circle (.05cm);
\draw[black, thick] (-.4,1.3) arc (180:0:.4cm);
\draw[black, thick] (0,1.7) -- (0,2.1);
\filldraw (0,1.7) circle (.05cm);
\draw[orange, thick] (-.4,-1.3) arc (-180:0:.4cm);
\draw[orange, thick] (0,-1.7) -- (0,-2.1);
\filldraw[orange] (0,-1.7) circle (.05cm);
\roundNbox{unshaded}{(.4,-1)}{.3}{0}{0}{$\eta_b$};
\roundNbox{unshaded}{(.4,1)}{.3}{0}{0}{$g_b$};
\roundNbox{unshaded}{(-.4,-1)}{.3}{0}{0}{$\xi_a$};
\roundNbox{unshaded}{(-.4,1)}{.3}{0}{0}{$f_a$};
\node at (.6,.5) {\scriptsize{$b$}};
\node at (.6,-.5) {\scriptsize{$b$}};
\node at (-.6,.5) {\scriptsize{$a$}};
\node at (-.6,-.5) {\scriptsize{$a$}};
\node at (.2,.2) {\scriptsize{$c$}};
\node at (.2,-.2) {\scriptsize{$c$}};
\node at (.6,1.5) {\scriptsize{$Q$}};
\node at (-.6,1.5) {\scriptsize{$Q$}};
\node at (0,.5) {\scriptsize{$\alpha^\dag$}};
\node at (0,-.5) {\scriptsize{$\alpha$}};
}
\qquad\qquad
(f_a\otimes \xi_a)^*
=
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.7,0) -- (-.7,-1.4) -- (.7,-1.4) -- (.7,0);
\filldraw[\rColor] (-.7,0) -- (-.7,-1.4) -- (.7,-1.4) -- (.7,0);
\end{scope}
\draw[dotted] (-.7,0) -- (.7,0);
\draw[orange,thick] (0,-1.4) -- (0,-.7);
\draw (0,-.7) -- (0,.7);
\draw[black, thick] (0,1) -- (0,1.4);
\roundNbox{unshaded}{(0,-.7)}{.3}{0}{0}{$\overline{\xi_a}$};
\roundNbox{unshaded}{(0,.7)}{.3}{0}{0}{$\overline{f_a}$};
\node at (.2,.2) {\scriptsize{$\overline{a}$}};
\node at (.2,-.2) {\scriptsize{$\overline{a}$}};
\node at (.2,1.2) {\scriptsize{$Q$}};
}\,,
$$
and the unit is given by
$$
1=
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.7,0) -- (-.7,-1.4) -- (.7,-1.4) -- (.7,0);
\filldraw[\rColor] (-.7,0) -- (-.7,-1.4) -- (.7,-1.4) -- (.7,0);
\end{scope}
\draw[dotted] (-.7,0) -- (.7,0);
\filldraw (0,.3) circle (.05cm);
\draw[black, thick] (0,.3) -- node[right]{$\scriptstyle Q$} (0,.7);
\draw[orange,thick] (0,-1.4) -- (0,-.7);
\roundNbox{unshaded}{(0,-.7)}{.3}{0}{0}{\scriptsize{$\Omega_N$}};
}\,;
\qquad\qquad
\Omega_N = 1\in {}_NN_N.
$$
\begin{ex}
\label{ex:RealizedJonesTower}
Suppose now $N\subseteq M$ is a finite depth, finite index $\rm II_1$ subfactor.
The algebra $M$ considered as an $N-N$ bimodule ${}_NM_N \in {}_N\cC_N$ is the canonical Q-system $X\boxtimes_M \overline{X}$ corresponding to the generator $X:={}_NM_M \in {}_N\cC_{M}$ as discussed in Example \ref{ex:RealizeSeparatedQSystem}.
The realization $|{}_NM_N|$ is canonically $*$-isomorphic to $M$:
$$
|{}_NM_N|
=
\operatorname{Hom}(N_N \to N \boxtimes_N M_N)
=
\operatorname{Hom}(N_N \to N \boxtimes_N M \boxtimes_M M_N)
\cong
\operatorname{Hom}(M_M \to M_M)
=
M.
$$
Now consider the Jones tower obtained by iterating Jones' basic construction defined inductively by $M_{n+1} := \operatorname{End}((M_n)_{M_{n-1}})$ \cite{MR0696688,MR999799}
$$
M_0=N \subseteq M = M_1 \subseteq M_2 \subseteq M_3\subseteq \cdots.
$$
The $\rm II_1$ factor $M_n$ is $*$-isomorphic to the realization of the Q-system $(X\boxtimes_M \overline{X})^{\boxtimes n}\cong X^{\operatorname{alt}\boxtimes n} \boxtimes \overline{X^{\operatorname{alt}\boxtimes n}}$ which has multiplication and unit given by
$$
\tikzmath{
\draw (-.3,0) node[below]{$\scriptstyle n$} arc (180:0:.3cm) node[below]{$\scriptstyle n$};
\draw (-.6,0) node[below]{$\scriptstyle n$} .. controls ++(90:.3cm) and ++(270:.3cm) .. (-.3,.8);
\draw (.6,0) node[below]{$\scriptstyle n$} .. controls ++(90:.3cm) and ++(270:.3cm) .. (.3,.8);
}
\qquad\qquad
\tikzmath{
\draw (-.3,0) node[above]{$\scriptstyle n$} -- (-.3,-.3) arc (-180:0:.3cm) -- (.3,0) node[above]{$\scriptstyle n$};
}
\qquad\qquad
\tikzmath{
\draw (0,0) --node[right]{$\scriptstyle n$} (0,1);
}
=
\operatorname{id}_{X^{\operatorname{alt}\boxtimes n}};
\qquad\qquad
X^{\operatorname{alt}\boxtimes n}:=\underbrace{X\boxtimes_M \overline{X} \boxtimes_N \cdots \boxtimes X^?}_{\text{$n$ tensorands}}.
$$
Indeed,
$$
\operatorname{Hom}(N_N \to N \boxtimes_N X^{\operatorname{alt}\boxtimes n} \boxtimes \overline{X^{\operatorname{alt}\boxtimes n}})
\cong
\operatorname{End}(N\boxtimes_N X^{\operatorname{alt}\boxtimes n}_{M \text{ or }N})
\cong
M_n
$$
by the multistep Jones basic construction \cite{MR965748,MR1424954}.
Another way to see this is to use Remark \ref{rem:MoritaEquivalenceRealization};
for example, the map $\operatorname{Ad}(X)$ takes the basic construction $\langle M, N\rangle=\overline{X}\boxtimes_N X$ to
$(X\boxtimes_M \overline{X})^{\boxtimes 2}$ with the multiplicaiton as claimed.
As a coend realization, we have a canonical Frobenius reciprocity isomorphism
\begin{equation}
\label{eq:CoendForMn}
M_n =
\bigoplus_{c\in \operatorname{Irr}(\cC)}
\cC(c\to X^{\operatorname{alt}\boxtimes n} \boxtimes \overline{X^{\operatorname{alt}\boxtimes n}})
\otimes_\bbC G(c)
\cong
\bigoplus_{c\in \operatorname{Irr}(\cC)}
\cC(c\boxtimes X^{\operatorname{alt}\boxtimes n} \to X^{\operatorname{alt}\boxtimes n})
\otimes_\bbC G(c).
\end{equation}
Under this isomorphism, in the coend realization diagrammatic calculus, we can draw the $X^{\operatorname{alt}\boxtimes n}$ horizontally, where the horizontal line should be viewed as slightly tilted going from the bottom right to the top left, as indicated by the cyan arrows below.
\begin{equation}
\label{eq:BottomRightTopLeft}
\tikzmath{
\draw (-.7,0) -- (.7,0);
\draw (0,0) -- (0,-.7);
\draw[cyan, dashed] (-.7,-.7) -- (.7,.7);
\draw[cyan, ->] (-.6,-.7) -- (-.8,-.5);
\draw[cyan, ->] (.7,.6) -- (.5,.8);
\roundNbox{unshaded}{(0,0)}{.3}{0}{0}{$f$};
\node at (-.9,0) {\scriptsize{$k$}};
\node at (.9,0) {\scriptsize{$k$}};
\node at (.2,-.5) {\scriptsize{$c$}};
}
\in
\cC(c\boxtimes X^{\operatorname{alt}\boxtimes n} \to X^{\operatorname{alt}\boxtimes n})
\end{equation}
The multiplication, $*$, and unit in the realization $|X^{\operatorname{alt}\boxtimes n} \boxtimes \overline{X^{\operatorname{alt}\boxtimes n}}|\cong M_n$
are now represented respectively by
$$
\sum_{\substack{ a,b,c\in\operatorname{Irr}(\cC) \\ \alpha\in\operatorname{ONB}(a\otimes b\to c)}}
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-1.2,0) -- (-1.2,-2.2) -- (1.2,-2.2) -- (1.2,0);
\filldraw[\rColor] (-1.2,0) -- (-1.2,-2.2) -- (1.2,-2.2) -- (1.2,0);
\end{scope}
\draw[dotted] (-1.2,0) -- (1.2,0);
\draw[orange,thick] (-.5,-1.4) arc (-180:0:.5cm);
\draw[orange, thick] (0,-1.9) -- (0,-2.2);
\filldraw[orange] (0,-1.9) circle (.05cm);
\draw (0,-.3) -- (0,.3);
\draw (.5,.8) arc (0:-180:.5cm);
\draw (.5,-.8) arc (0:180:.5cm);
\draw (-1.2,1.1) -- (1.2,1.1);
\filldraw (0,-.3) circle (.05cm);
\filldraw (0,.3) circle (.05cm);
\roundNbox{unshaded}{(.5,-1.1)}{.3}{0}{0}{$\eta_b$};
\roundNbox{unshaded}{(.5,1.1)}{.3}{0}{0}{$g_b$};
\roundNbox{unshaded}{(-.5,-1.1)}{.3}{0}{0}{$\xi_a$};
\roundNbox{unshaded}{(-.5,1.1)}{.3}{0}{0}{$f_a$};
\node at (.7,.6) {\scriptsize{$b$}};
\node at (.7,-.6) {\scriptsize{$b$}};
\node at (-.7,.6) {\scriptsize{$a$}};
\node at (-.7,-.6) {\scriptsize{$a$}};
\node at (.2,.2) {\scriptsize{$c$}};
\node at (.2,-.2) {\scriptsize{$c$}};
\node at (1,1.3) {\scriptsize{$n$}};
\node at (-1,1.3) {\scriptsize{$n$}};
\node at (0,1.3) {\scriptsize{$n$}};
\node at (0,.5) {\scriptsize{$\alpha^\dag$}};
\node at (0,-.5) {\scriptsize{$\alpha$}};
}\,,
\qquad\qquad
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.7,0) -- (-.7,-1.4) -- (.7,-1.4) -- (.7,0);
\filldraw[\rColor] (-.7,0) -- (-.7,-1.4) -- (.7,-1.4) -- (.7,0);
\end{scope}
\draw[dotted] (-.7,0) -- (.7,0);
\draw[orange,thick] (0,-1.4) -- (0,-.7);
\draw (0,-.7) -- (0,.7);
\draw (-.7,.7) -- (.7,.7);
\roundNbox{unshaded}{(0,-.7)}{.3}{0}{0}{$\overline{\xi_a}$};
\roundNbox{unshaded}{(0,.7)}{.3}{0}{0}{$\overline{f_a}$};
\node at (.2,.2) {\scriptsize{$\overline{a}$}};
\node at (.2,-.2) {\scriptsize{$\overline{a}$}};
\node at (.5,.9) {\scriptsize{$n$}};
\node at (-.5,.9) {\scriptsize{$n$}};
}\,,
\qquad\qquad
\text{and}
\qquad\qquad
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.7,0) -- (-.7,-1.4) -- (.7,-1.4) -- (.7,0);
\filldraw[\rColor] (-.7,0) -- (-.7,-1.4) -- (.7,-1.4) -- (.7,0);
\end{scope}
\draw[dotted] (-.7,0) -- (.7,0);
\draw (-.7,.7) -- (.7,.7);
\node at (0,.9) {\scriptsize{$n$}};
}
=
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.7,0) -- (-.7,-1.4) -- (.7,-1.4) -- (.7,0);
\filldraw[\rColor] (-.7,0) -- (-.7,-1.4) -- (.7,-1.4) -- (.7,0);
\end{scope}
\draw[dotted] (-.7,0) -- (.7,0);
\draw (-.7,.7) -- (.7,.7);
\draw[orange,thick] (0,-1.4) -- (0,-.7);
\roundNbox{unshaded}{(0,-.7)}{.3}{0}{0}{\scriptsize{$\Omega_N$}};
\node at (0,.9) {\scriptsize{$n$}};
}\,.
$$
Here, the $X^{\operatorname{alt}\boxtimes n}$ horizontal strand is read bottom to top, i.e.,
$$
\tikzmath{
\draw (-.7,0) -- node[above]{$\scriptstyle n$} (.7,0);
}
\qquad
=
\tikzmath{
\draw (-.7,0) -- node[above]{$\scriptstyle X^{\operatorname{alt}\boxtimes n}$} (.7,0);
}
\qquad
=
\qquad
\tikzmath{
\draw (-.7,-1) -- node[above]{$\scriptstyle X$} (.7,-1);
\draw (-.7,-.5) -- node[above]{$\scriptstyle \overline{X}$} (.7,-.5);
\draw (-.7,0) -- node[above]{$\scriptstyle X$} (.7,0);
\node at (0,.7) {$\scriptstyle \vdots$};
\draw (-.7,1) -- node[above]{$\scriptstyle X^?$} (.7,1);
}\,.
$$
\end{ex}
\subsection{The inductive limit factor as a realization}
\label{sec:InductiveLimit}
We now give a graphical representation of the inductive limit $\rm II_1$ factor $M_\infty$ from the realized Jones tower from Example \ref{ex:RealizedJonesTower}.
We begin with a short remark about inductive limits in the category of tracial von Neumann algebras and trace-preserving unital $*$-homomorphisms, followed by a brief review of construction of the hyperfinite $\rm II_1$ subfactor $R\subseteq P$ with the opposite standard invariant as our finite depth, finite index $\rm II_1$ subfactor $N\subseteq M$.
\begin{rem}
Consider the category whose objects are pairs $(M,\operatorname{tr})$ where $M$ is a separable von Neumann algebra and $\operatorname{tr}$ is a faithful normal tracial state, and whose morphisms are trace-preserving unital $*$-homomorphisms (which are automatically normal by \cite[III.2.2.2]{MR2188261} and \cite[Prop.~9.1.1]{JonesVNA}).
It is well known that this category admits inductive limits.
We briefly recall the construction for completeness and convenience of the reader.
For an increasing sequence of tracial von Neumann algebras $(M_n,\operatorname{tr}_n)$, we get a tracial von Neumann algebra $\varinjlim M_n$ by taking the bicommutant of $\bigcup_n M_n$ on its GNS Hilbert space with respect to the limit trace, which is equipped with the faithful tracial state $\varinjlim \operatorname{tr} := \langle \,\cdot\, \Omega, \Omega\rangle_{L^2\bigcup_n M_n}$.
For any tracial von Neumann algebra $(N,\operatorname{tr})$ and trace-preserving maps $\varphi_n :M_n \to N$ compatible with the inclusions, we get a unique trace-preserving map $\varphi:\bigcup_n M_n \to N$,
which extends uniquely to a trace-preserving map $\varinjlim M_n\to N$.
Indeed, for a fixed $x\in \varinjlim M_n$, let $\xi \in L^2N$ be the image of $x\Omega$ under the induced map of GNS spaces $L^2(\bigcup_n M_n) \to L^2N$ given by the extension of $m\Omega \mapsto \varphi(m)\Omega$.
Since there is a bounded sequence $(x_k)\subseteq \bigcup_n M_n$ with $x_k \to x$ in $\|\cdot\|_2$, we see
\begin{align*}
\|L_\xi n\Omega\|_2
&=
\|\xi n \|_2
=
\| Jn^*J \xi\|_2
=
\lim \|Jn^*J \varphi(x_k)\Omega\|_2
=
\lim \|\varphi(x_k) n\Omega\|_2
\\&\leq
\limsup \|\varphi(x_k)\|\cdot \|n\Omega\|_2.
=
\limsup \|x_k\|\cdot \|n\Omega\|_2.
\end{align*}
Thus $\xi$ is $N$-bounded, and necessarily of the form $\varphi(x)\Omega$ for some $\varphi(x)\in N$.
Since multiplication is jointly SOT-continuous on bounded subsets
and $*$ is SOT-continuous on bounded subsets
of $\varinjlim M_n$,
it is easily verified that the above definition of $\varphi(x)$ extends $\varphi$ to a unital $*$-homomorphism $\varinjlim M_n \to N$.
\end{rem}
For the rest of this section, we fix a finite depth finite index $\rm II_1$ subfactor $N\subseteq M$.
Recall that its standard invariant can also be described as the subfactor planar algebra $\cP_\bullet$ whose box spaces are given by
$$
\cP_{k,+}:= \operatorname{End}(X^{\operatorname{alt}\boxtimes k})
\qquad\text{and}\qquad
\cP_{k,-}:= \operatorname{End}(\overline{X}^{\operatorname{alt}\boxtimes k}).
$$
These finite dimensional von Neumann algebras are equipped with the canonical traces which agree with the categorical traces on $\cC(N\subseteq M)$.
We now rapidly recall how to construct a hyperfinite $\rm II_1$ subfactor with the opposite standard invariant \cite{MR996454,MR1278111}.
For a detailed diagrammatic exposition (in the multifactor setting), see \cite[\S5.1]{2010.01067}.
Since $N\subseteq M$ is finite depth, the inductive limit tracial von Neumann algebras
$\varinjlim \cP_{k,\pm}$ are hyperfinite $\rm II_1$ factor.
We have a trace-preserving injection $\cP_{n,+}\hookrightarrow \cP_{n+1,-}$ by adding an $\overline{X}$ strand to the left, giving a $\rm II_1$ subfactor
$$
R := \varinjlim \cP_{n,+} \subseteq \varinjlim \cP_{n,-} =:P.
$$
It is well known that by the Ocneanu Compactness Theorem \cite[Thm.~5.7.6]{MR1473221}, the inclusion $R^{\operatorname{op}}\subseteq P^{\operatorname{op}}$ has the same standard invariant as $N\subseteq M$ \cite{MR996454,MR1278111}, and thus $R\subseteq P$ has the opposite standard invariant.
Letting $\cC:={}_N\cC_N$ be the unitary fusion category generated by ${}_NM_N$, the above construction gives a fully faithful unitary tensor functor $F:\cC^{\operatorname{op}}\to \mathsf{Bim_{fgp}}(R)$.
By \cite[Prop.~3.2]{MR1424954}, simple objects $c\in \operatorname{Irr}(\cC)$ correspond to minimal projections $p\in M_0'\cap M_{2n}$ under the correspondence
$$
M_0'\cap M_{2n} \ni p \longmapsto pM_n \in \mathsf{Bim_{fgp}}(N).
$$
Now consider the Jones tower
$$
R_0 = R \subseteq P = R_1 \subseteq R_2 \subseteq R_3\subseteq \cdots
$$
of our hyperfinite $\rm II_1$ subfactor $R\subseteq P$.
Simple objects $c^{\operatorname{op}}\in \operatorname{Irr}(\cC^{\operatorname{op}})$ correspond to the opposite projections $p^{\operatorname{op}}\in R_0'\cap R_{2n} \cong (M_0'\cap M_{2n})^{\operatorname{op}}$, which corresponds to the bimodule $p^{\operatorname{op}}R_n$.
As $R_n$ is also isomorphic to $\varinjlim \cP_{k,\pm}$ (here, $\pm$ depends on the parity of $n$), we can realize the bimodule $p^{\operatorname{op}}R_n$ graphically as an inductive limit:
\begin{equation}
\label{eq:HyperfiniteAction}
F(c^{\operatorname{op}})
=
p^{\operatorname{op}}R_n
\cong
\varinjlim
\cC(
c\boxtimes X^{\operatorname{alt}\boxtimes k}\to X^{\operatorname{alt}\boxtimes k}
)
\end{equation}
under the isometric right $R_n$-inner product preserving inclusions
$$
\tikzmath{
\draw (-.7,0) -- (.7,0);
\draw (0,0) -- (0,-.7);
\roundNbox{unshaded}{(0,0)}{.3}{0}{0}{$f$};
\node at (-.9,0) {\scriptsize{$k$}};
\node at (.9,0) {\scriptsize{$k$}};
\node at (.2,-.5) {\scriptsize{$c$}};
}
\longmapsto
\tikzmath{
\draw (-.7,0) -- (.7,0);
\draw (-.7,.5) -- (.7,.5);
\draw (0,0) -- (0,-.7);
\roundNbox{unshaded}{(0,0)}{.3}{0}{0}{$f$};
\node at (-.9,0) {\scriptsize{$k$}};
\node at (.9,0) {\scriptsize{$k$}};
\node at (.2,-.5) {\scriptsize{$c$}};
}\,.
$$
Indeed, $p^{\operatorname{op}}R_n\cong p^{\operatorname{op}}R_{n}f_n\subseteq p^{\operatorname{op}}R_{2n}f_n$, where $f_n$ is the multistep Jones projection \cite{MR965748,MR1424954}.
In diagrams, for $\xi \in \cC(c\boxtimes X^{\operatorname{alt}\boxtimes k+n}\to X^{\operatorname{alt}\boxtimes k+n}) \subseteq R_n \hookrightarrow R_{2n}$, we have
$$
f_n =
\frac{1}{[M:N]^{n/2}}
\tikzmath{
\draw (-.5,.3) node[above]{$\scriptstyle n$} arc (90:-90:.3cm);
\draw (.5,.3) node[above]{$\scriptstyle n$} arc (90:270:.3cm);
}
\qquad\qquad
\Longrightarrow
\qquad\qquad
p^{\operatorname{op}} \xi f_n
=
\frac{1}{[M:N]^{n/2}}
\tikzmath{
\draw[rounded corners=5, dashed] (-1.5,-1) rectangle (.7,.6);
\draw (-1.7,.2) node[left]{$\scriptstyle k$} -- (1.1,.2) node[right]{$\scriptstyle k$};
\draw (-.7,-.2) -- node[above]{$\scriptstyle n$} (-.3,-.2);
\draw (.3,-.2) arc (90:-90:.2cm) -- node[below]{$\scriptstyle n$} (-.7,-.6);
\draw (1.1,-.2) node[right]{$\scriptstyle n$} arc (90:270:.2cm) node[right]{$\scriptstyle n$};
\draw (-1.3,-.4) -- (-1.7,-.4) node[left]{$\scriptstyle 2n$};
\roundNbox{unshaded}{(0,0)}{.4}{-.1}{-.1}{$\xi$};
\roundNbox{unshaded}{(-1,-.4)}{.4}{-.1}{-.1}{$p$};
}\,.
$$
It is now visibly evident how to implement the isomorphism \eqref{eq:HyperfiniteAction}.
Now, since we have two subfactors $N\subseteq M$ and $R\subseteq P$ with opposite standard invariants, we get two fully faithful unitary tensor functors $F: \cC^{\operatorname{op}}\to \mathsf{Bim_{fgp}}(R)$ and $G: \cC\to \mathsf{Bim_{fgp}}(N)$.
Consider the \emph{symmetric enveloping Q-system}
$S:= \bigoplus_{c\in \operatorname{Irr}(\cC)} c^{\operatorname{op}}\boxtimes c \in \cC^{\operatorname{op}}\boxtimes \cC$, which is simple as an $S-S$ bimodule, giving the realized
$\rm II_1$ factor $|(F\boxtimes G)(S)|$.
\begin{prop}
\label{prop:SymmetricEvelopeInductiveLimit}
The realization $|(F\boxtimes G)(S)|$ is $*$-isomorphic to the inductive limit $\rm II_1$ factor $M_\infty = \varinjlim M_n$.
\end{prop}
\begin{proof}
Observe that by the Yoneda Lemma, we have canonical isomorphisms
\begin{align*}
|(F\boxtimes G)(S)|
&\cong
\bigoplus\limits_{a,b\in\operatorname{Irr}(\cC)} (\cC^\operatorname{op}\boxtimes\cC)(a^\operatorname{op}\boxtimes b\to S)\otimes_{\bbC}(F\boxtimes G)(a^\operatorname{op}\boxtimes b)
\\&=
\bigoplus\limits_{c\in\operatorname{Irr}(\cC)} (\cC^\operatorname{op}\boxtimes\cC)(c^\operatorname{op}\boxtimes c\to S)\otimes_{\bbC}(F\boxtimes G)(c^\operatorname{op}\boxtimes c)
\\&\cong
\bigoplus\limits_{c\in\operatorname{Irr}(\cC)} (F\boxtimes G)(c^\operatorname{op}\boxtimes c)
\\&=
\bigoplus\limits_{c\in\operatorname{Irr}(\cC)} F(c^{\operatorname{op}}) \otimes_{\bbC} G(c)
\end{align*}
Thus $|(F\boxtimes G)(S)|$ is $*$-isomorphic to the more general coend realization of $F^\circ$ and $G^\circ$ from \cite[\S4]{MR3948170} (see also \cite[Ex.~5.33]{MR3687214}), where as above, $\circ$ denotes taking the underlying vector space.
Using this identification,
\[
\bigoplus\limits_{c\in\operatorname{Irr}(\cC)} F(c^{\operatorname{op}}) \otimes_{\bbC} G(c)
\cong
\varinjlim
\bigoplus\limits_{c\in\operatorname{Irr}(\cC)} \cC(c\boxtimes X^{\operatorname{alt}\boxtimes k} \to X^{\operatorname{alt}\boxtimes k})
\otimes_{\bbC} G(c)
\underset{\text{\eqref{eq:CoendForMn}}}{\cong}
\varinjlim M_n.
\qedhere
\]
\end{proof}
\subsection{Embedding \texorpdfstring{$\cZ(\cC)$}{Z(C)} into \texorpdfstring{$\tilde{\chi}(M_\infty)$}{chi(Minfty)} when \texorpdfstring{$N\subseteq M$}{N in M} is finite depth}
As in \S\ref{sec:InductiveLimit} above, for this section, we fix a finite depth, finite index $\rm II_1$ subfactor $N\subseteq M$, and let $\cC={}_N\cC_N$ denote the unitary fusion category of $N-N$ bimodules generated by ${}_NM_N$.
In this section, we extend this bijection to a unitary tensor functor $\Phi:\cZ(\cC)\to \mathsf{Bim_{fgp}}(M_\infty)$ such that when $N$ is non-Gamma, $\Phi$ takes values in $\tilde{\chi}(M_\infty)$ as is braided.
To do so, we use the description of the inductive limit $\rm II_1$ factor $M_\infty=|(F\boxtimes G)(S)|$ obtained from iterating the Jones basic construction afforded by Proposition \ref{prop:SymmetricEvelopeInductiveLimit}
for the fully faithful unitary tensor functors
$F: \cC^{\operatorname{op}}\to \mathsf{Bim_{fgp}}(R)$ from \eqref{eq:HyperfiniteAction} and $G: \cC\to \mathsf{Bim_{fgp}}(N)$ associated to the subfactor $N\subset M$.
First, applying \cite[Cor.~C]{2105.12010} in the $\rm W^*$ setting to the fully faithful unitary tensor functor $F\boxtimes G : \cC^{\operatorname{op}}\boxtimes \cC\to \mathsf{Bim_{fgp}}(R\otimes N)$, bimodule realization gives a fully faithful tensor functor from $\mathsf{Bim}_{\cC^{\operatorname{op}}\boxtimes \cC}(S) \to \mathsf{Bim_{fgp}}(M_\infty)$.
(See also Remark \ref{rem:QBimodules} above.)
Explicitly, on an $S-S$ bimodule
$X
=
\bigoplus_{a,b\in \operatorname{Irr}(\cC)} X_{ab} \otimes (a^{\operatorname{op}} \boxtimes b)$,
we have
$$
|X|
=
\operatorname{Hom}\left(
(R\otimes N)_{R\otimes N}
\to
\bigoplus_{a,b\in \operatorname{Irr}(\cC)} X_{ab} \otimes \big(F(a^{\operatorname{op}}) \otimes G(b)\big)_{R\otimes N}
\right)
\cong
\bigoplus_{a,b\in \operatorname{Irr}(\cC)}
X_{ab}\otimes F(a^{\operatorname{op}}) \otimes G(b).
$$
Using the well-known equivalence
$\mathsf{Bim}_{\cC^{\operatorname{op}}\boxtimes \cC}(S)\cong \cZ(\cC)$
from Example \ref{ex:MoritaEquivalenceBetweenCopCandZ(C)},
we get the following proposition.
\begin{prop}
Bimodule realization gives a fully faithful unitary tensor functor $\Phi:\cZ(\cC) \to \mathsf{Bim_{fgp}}(M_\infty)$.
\end{prop}
We now want an explicit model for $\Phi(z,\sigma_z)$ for each object $(z,\sigma_z)\in \cZ(\cC)$.
To do so, we give an explicit description of the $S-S$ bimodule $X_z \in \mathsf{Bim}_{\cC^{\operatorname{op}}\boxtimes \cC}(S)$ under the unitary tensor equivalence.
\begin{defn}
Given $(z,\sigma_z)\in \cZ(\cC)$, we define $X_z:=\bigoplus_{a,b\in\operatorname{Irr}(\cC)} \cC(b\to z\otimes a) \otimes (a^\operatorname{op}\boxtimes b)$.
The left $S$-module structure of $X_z$ is given as follows.
First, we observe
\begin{align*}
S\otimes_{\cC^{\operatorname{op}}\boxtimes \cC} X_z
&=
\bigoplus_{a,b,c\in\operatorname{Irr}(\cC)}
\cC(b\to z\otimes a) \otimes
( (c^\operatorname{op}\otimes a^\operatorname{op})\boxtimes (c\otimes b) )
\\
&=
\bigoplus_{a,b,c,d,e\in\operatorname{Irr}(\cC)}
\cC(b\to z\otimes a) \otimes \cC^\operatorname{op}(d^\operatorname{op}\to c^\operatorname{op}\otimes a^\operatorname{op}) \otimes \cC(e\to c\otimes b) \otimes
(d^\operatorname{op}\boxtimes e)
\\
&=
\bigoplus_{a,b,c,d,e\in\operatorname{Irr}(\cC)}
\cC(b\to z\otimes a) \otimes \cC(c\otimes a\to d) \otimes \cC(e\to c\otimes b) \otimes
(d^\operatorname{op}\boxtimes e)
\end{align*}
The left action map is given component-wise by
$$
\tikzmath{
\draw[thick, \zColor] (-.15,.3) -- node[left]{$\scriptstyle z$} (-.15,.7);
\draw (.15,.3) -- node[right]{$\scriptstyle a$} (.15,.7);
\draw (0,-.7) -- node[right]{$\scriptstyle b$} (0,-.3);
\roundNbox{fill=white}{(0,0)}{.3}{0}{0}{$f$}
}
\otimes
\tikzmath{
\draw (-.15,-.3) -- node[left]{$\scriptstyle c$} (-.15,-.7);
\draw (.15,-.3) -- node[right]{$\scriptstyle a$} (.15,-.7);
\draw (0,.7) -- node[right]{$\scriptstyle d$} (0,.3);
\roundNbox{fill=white}{(0,0)}{.3}{0}{0}{$g$}
}
\otimes
\tikzmath{
\draw (-.15,.3) -- node[left]{$\scriptstyle c$} (-.15,.7);
\draw (.15,.3) -- node[right]{$\scriptstyle b$} (.15,.7);
\draw (0,-.7) -- node[right]{$\scriptstyle e$} (0,-.3);
\roundNbox{fill=white}{(0,0)}{.3}{0}{0}{$h$}
}
\longmapsto
\tikzmath{
\draw (0,1) -- (0,1.9);
\draw (-.075,-1.7) -- (-.075,-1);
\draw (.3,0) -- (.3,.9);
\draw (.15,-.7) -- (.15,0);
\draw[\zColor,thick] (0,.3) .. controls ++(90:.3cm) and ++(270:.3cm) .. (-.6,.9) -- (-.6,1.9);
\draw[knot] (-.3,-.7) -- (-.3,.9);
\roundNbox{fill=white}{(0,1.2)}{.3}{.15}{.15}{$g$};
\roundNbox{fill=white}{(-.075,-1)}{.3}{.075}{.075}{$h$};
\roundNbox{fill=white}{(.15,0)}{.3}{0}{0}{$f$};
\node at (.125,-1.5) {\scriptsize{$e$}};
\node at (-.5,0) {\scriptsize{$c$}};
\node at (.35,-.5) {\scriptsize{$b$}};
\node at (.5,.5) {\scriptsize{$a$}};
\node[\zColor] at (-.8,1.7) {\scriptsize{$z$}};
\node at (.2,1.7) {\scriptsize{$d$}};
}
$$
The right $S$-module structure is defined similarly.
Observe
\begin{align*}
X_z\otimes_{\cC^{\operatorname{op}}\boxtimes \cC} S
&=
\bigoplus_{a,b,c\in\operatorname{Irr}(\cC)}
\cC(b\to z\otimes a)
\otimes
(a^\operatorname{op}\otimes c^\operatorname{op})\boxtimes (b\otimes c)
\\
&=
\bigoplus_{a,b,c,d,e\in\operatorname{Irr}(\cC)}
\cC(b\to z\otimes a)\otimes \cC(a\otimes c\to d)\otimes \cC(e\to b\otimes c) \otimes
(d^\operatorname{op}\boxtimes e),
\end{align*}
and the right action map is given component-wise by
$$
\tikzmath{
\draw[thick, \zColor] (-.15,.3) -- node[left]{$\scriptstyle z$} (-.15,.7);
\draw (.15,.3) -- node[right]{$\scriptstyle a$} (.15,.7);
\draw (0,-.7) -- node[right]{$\scriptstyle b$} (0,-.3);
\roundNbox{fill=white}{(0,0)}{.3}{0}{0}{$f$};
}
\otimes
\tikzmath{
\draw (-.15,-.3) -- node[left]{$\scriptstyle a$} (-.15,-.7);
\draw (.15,-.3) -- node[right]{$\scriptstyle c$} (.15,-.7);
\draw (0,.7) -- node[right]{$\scriptstyle d$} (0,.3);
\roundNbox{fill=white}{(0,0)}{.3}{0}{0}{$g$};
}
\otimes
\tikzmath{
\draw (-.15,.3) -- node[left]{$\scriptstyle b$} (-.15,.7);
\draw (.15,.3) -- node[right]{$\scriptstyle c$} (.15,.7);
\draw (0,-.7) -- node[right]{$\scriptstyle e$} (0,-.3);
\roundNbox{fill=white}{(0,0)}{.3}{0}{0}{$h$};
}
\longmapsto
\tikzmath{
\draw (0,0) -- (0,.7);
\draw (-.2,-.7) -- (-.2,0);
\draw (.4,-.7) -- (.4,.7);
\draw (.2,1) -- (.2,1.7);
\draw (.1,-1.7) -- (.1,-1);
\draw[thick, \zColor] (-.4,0) -- (-.4,1.7);
\roundNbox{fill=white}{(.2,1)}{.3}{.1}{.1}{$g$};
\roundNbox{fill=white}{(-.2,0)}{.3}{.1}{.1}{$f$};
\roundNbox{fill=white}{(.1,-1)}{.3}{.2}{.2}{$h$};
\node at (.3,-1.5) {\scriptsize{$e$}};
\node at (-.4,-.5) {\scriptsize{$b$}};
\node at (.6,0) {\scriptsize{$c$}};
\node at (-.2,.5) {\scriptsize{$a$}};
\node[\zColor] at (-.6,1.5) {\scriptsize{$z$}};
\node at (.4,1.5) {\scriptsize{$d$}};
}
$$
\end{defn}
We now describe the realization
\begin{equation}
\label{eq:RealizationXz}
|X_z|
=
\varinjlim
\bigoplus_{a,b\in \operatorname{Irr}(\cC)}
\cC(b\to z\otimes a)
\otimes_\bbC
\cC(a\boxtimes X^{\operatorname{alt} \boxtimes n}\to X^{\operatorname{alt} \boxtimes n})
\otimes_\bbC
G(b).
\end{equation}
First, note that by semisimplicity and Frobenius reciprocity, we can alternatively describe the first two tensorands in the direct sum for $|X_z|$ as
\begin{equation}
\label{eq:SemisimpleIso}
\left.
\begin{aligned}
\bigoplus_{a\in \operatorname{Irr}(\cC)}
\cC(b\to z\otimes a)
&\otimes_\bbC
\cC(a\boxtimes X^{\operatorname{alt} \boxtimes n} \to X^{\operatorname{alt} \boxtimes n})
\\&\cong
\bigoplus_{a\in \operatorname{Irr}(\cC)}
\cC( \overline{z}\boxtimes b\to a)
\otimes_\bbC
\cC(a \to X^{\operatorname{alt} \boxtimes n}\boxtimes \overline{X^{\operatorname{alt} \boxtimes n}})
\\&\cong
\cC(\overline{z}\boxtimes b
\to
X^{\operatorname{alt} \boxtimes n}\boxtimes \overline{X^{\operatorname{alt} \boxtimes n}})
\\&\cong
\cC(
b\boxtimes X^{\operatorname{alt} \boxtimes n}
\to
z\boxtimes
X^{\operatorname{alt} \boxtimes n})
\end{aligned}
\right\}
\end{equation}
We may thus graphically represent elements in a dense subspace of $|X_z|$ as
$$
\bigoplus_{b\in\operatorname{Irr}(\cC)}
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.7,0) -- (-.7,-1.4) -- (.7,-1.4) -- (.7,0);
\filldraw[\rColor] (-.7,0) -- (-.7,-1.4) -- (.7,-1.4) -- (.7,0);
\end{scope}
\draw[dotted] (-.7,0) -- (.7,0);
\draw[orange,thick] (0,-1.4) -- (0,-.7);
\draw (0,-.7) -- (0,.7);
\draw (-.7,.95) node[left]{$\scriptstyle n$} -- (.7,.95) node[right]{$\scriptstyle n$};
\draw[thick, \zColor] (-.7,.65) node[left]{$\scriptstyle z$} -- (0,.65);
\draw[cyan, dashed] (-.7,.1) -- (.7,1.5);
\draw[cyan, ->] (-.6,.1) -- (-.8,.3);
\draw[cyan, ->] (.7,1.4) -- (.5,1.6);
\roundNbox{unshaded}{(0,-.7)}{.3}{0}{0}{$\xi_b$};
\roundNbox{unshaded}{(0,.8)}{.35}{-.05}{-.05}{$f_b$};
\node at (.2,.2) {\scriptsize{$b$}};
\node at (.2,-.2) {\scriptsize{$b$}};
}
\in
\varinjlim
\bigoplus_{b\in \operatorname{Irr}(\cC)}
\cC(b\boxtimes X^{\operatorname{alt} \boxtimes n} \to z\boxtimes X^{\operatorname{alt} \boxtimes n})
\otimes_\bbC
G(b)
$$
where as in \eqref{eq:BottomRightTopLeft}, the horizontal line in the top half of the diagram should be viewed as slightly tilted going from the bottom right to the top left, as indicated by the cyan arrows above.
The right $M_\infty$-action on $|X_z|$ is the obvious diagrammatic one, and the left one is similar, but uses the half-braiding for $z$:
\[
(f_a\otimes\xi_a)\rhd (g_b\otimes \eta_b)
:=
\sum_{\substack{
c\in \operatorname{Irr}(\cC)
\\
\alpha\in\operatorname{ONB}(a\otimes b\to c)
}}
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-1.2,0) -- (-1.2,-2.2) -- (1.2,-2.2) -- (1.2,0);
\filldraw[\rColor] (-1.2,0) -- (-1.2,-2.2) -- (1.2,-2.2) -- (1.2,0);
\end{scope}
\draw[dotted] (-1.2,0) -- (1.2,0);
\draw[orange,thick] (-.5,-1.4) arc (-180:0:.5cm);
\draw[orange, thick] (0,-1.9) -- (0,-2.2);
\filldraw[orange] (0,-1.9) circle (.05cm);
\draw[thick, \zColor] (.2,1.1) .. controls ++(180:.3cm) and ++(0:.3cm) .. (-.4,.7) -- (-1.2,.7) node[above, xshift=.2cm]{$\scriptstyle z$};
\draw (-1.2,1.3) node[above, xshift=.2cm]{$\scriptstyle n$} -- node[above]{$\scriptstyle n$} (1.2,1.3) node[above, xshift=-.2cm]{$\scriptstyle n$};
\draw[knot] (.5,.8) arc (0:-180:.5cm) -- (-.5,1);
\draw (.5,-.8) arc (0:180:.5cm);
\draw (0,-.3) -- (0,.3);
\filldraw (0,-.3) circle (.05cm);
\filldraw (0,.3) circle (.05cm);
\roundNbox{unshaded}{(.5,-1.1)}{.3}{0}{0}{$\eta_b$};
\roundNbox{unshaded}{(.5,1.2)}{.4}{-.1}{-.1}{$g_b$};
\roundNbox{unshaded}{(-.5,-1.1)}{.3}{0}{0}{$\xi_a$};
\roundNbox{unshaded}{(-.5,1.3)}{.3}{0}{0}{$f_a$};
\node at (.7,.6) {\scriptsize{$b$}};
\node at (.7,-.6) {\scriptsize{$b$}};
\node at (-.7,.5) {\scriptsize{$a$}};
\node at (-.7,-.6) {\scriptsize{$a$}};
\node at (.2,.2) {\scriptsize{$c$}};
\node at (.2,-.2) {\scriptsize{$c$}};
\node at (0,.5) {\scriptsize{$\alpha^\dag$}};
\node at (0,-.5) {\scriptsize{$\alpha$}};
}
\,.
\]
The $M_\infty$-valued inner product of $|X_z|$ is given by
\[
\langle f_a\otimes \xi_a|g_b\otimes \eta_b\rangle_{M_\infty}^{|X_z|}
:=
\sum_{\substack{
c\in \operatorname{Irr}(\cC)
\\
\alpha\in\operatorname{ONB}(a\otimes b\to c)
}}
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-1.2,0) -- (-1.2,-2.2) -- (1.2,-2.2) -- (1.2,0);
\filldraw[\rColor] (-1.2,0) -- (-1.2,-2.2) -- (1.2,-2.2) -- (1.2,0);
\end{scope}
\draw[dotted] (-1.2,0) -- (1.2,0);
\draw[orange,thick] (-.5,-1.4) arc (-180:0:.5cm);
\draw[orange, thick] (0,-1.9) -- (0,-2.2);
\filldraw[orange] (0,-1.9) circle (.05cm);
\draw[thick, \zColor] (.2,1.1) -- node[below]{$\scriptstyle z$} (-.2,1.1);
\draw (-1.2,1.3) node[above, xshift=.2cm]{$\scriptstyle n$} -- node[above]{$\scriptstyle n$} (1.2,1.3) node[above, xshift=-.2cm]{$\scriptstyle n$};
\draw (.5,.8) arc (0:-180:.5cm);
\draw (.5,-.8) arc (0:180:.5cm);
\draw (0,-.3) -- (0,.3);
\filldraw (0,-.3) circle (.05cm);
\filldraw (0,.3) circle (.05cm);
\roundNbox{unshaded}{(.5,-1.1)}{.3}{0}{0}{$\eta_b$};
\roundNbox{unshaded}{(.5,1.2)}{.4}{-.1}{-.1}{$g_b$};
\roundNbox{unshaded}{(-.5,-1.1)}{.3}{0}{0}{$\overline{\xi_a}$};
\roundNbox{unshaded}{(-.5,1.2)}{.4}{-.1}{-.1}{$\overline{f_a}$};
\node at (.7,.6) {\scriptsize{$b$}};
\node at (.7,-.6) {\scriptsize{$b$}};
\node at (-.7,.6) {\scriptsize{$a$}};
\node at (-.7,-.6) {\scriptsize{$a$}};
\node at (.2,.2) {\scriptsize{$c$}};
\node at (.2,-.2) {\scriptsize{$c$}};
\node at (0,.5) {\scriptsize{$\alpha^\dag$}};
\node at (0,-.5) {\scriptsize{$\alpha$}};
}
\,.
\]
By the definition of the realization \eqref{eq:RealizationXz} for $|X_z|$ and $|X_w|$
and the right and left $S$-action on $X_z$ and $X_w$ respectively, the tensorator $\Phi^2_{z,w}$ is given on $|X_z|\boxtimes_{M_\infty} |X_w|$ by
$$
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.7,0) -- (-.7,-1.4) -- (.7,-1.4) -- (.7,0);
\filldraw[\rColor] (-.7,0) -- (-.7,-1.4) -- (.7,-1.4) -- (.7,0);
\end{scope}
\draw[dotted] (-.7,0) -- (.7,0);
\draw[orange,thick] (0,-1.4) -- (0,-.7);
\draw (0,-.7) -- (0,.8);
\draw (-.7,1.1) node[above, xshift=.2cm]{$\scriptstyle n$} -- (.7,1.1) node[above, xshift=-.2cm]{$\scriptstyle n$};
\draw[\zColor,thick] (0,.4) -- (-.7,.4) node[above, xshift=.2cm]{$\scriptstyle z$};
\filldraw[\zColor] (0,.4) circle (.05cm);
\roundNbox{unshaded}{(0,1.1)}{.3}{0}{0}{$f_a$};
\roundNbox{unshaded}{(0,-.7)}{.3}{0}{0}{$\xi_b$};
\node at (.2,.6) {$\scriptstyle a$};
\node at (.2,.2) {$\scriptstyle b$};
\node at (.2,-.2) {$\scriptstyle b$};
}
\underset{M_\infty}{\boxtimes}
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.7,0) -- (-.7,-1.4) -- (.7,-1.4) -- (.7,0);
\filldraw[\rColor] (-.7,0) -- (-.7,-1.4) -- (.7,-1.4) -- (.7,0);
\end{scope}
\draw[dotted] (-.7,0) -- (.7,0);
\draw[orange,thick] (0,-1.4) -- (0,-.7);
\draw (0,-.7) -- (0,.8);
\draw (-.7,1.1) node[above, xshift=.2cm]{$\scriptstyle n$} -- (.7,1.1) node[above, xshift=-.2cm]{$\scriptstyle n$};
\draw[red,thick] (0,.4) -- (-.7,.4) node[above, xshift=.2cm]{$\scriptstyle w$};
\filldraw[red] (0,.4) circle (.05cm);
\roundNbox{unshaded}{(0,1.1)}{.3}{0}{0}{$g_c$};
\roundNbox{unshaded}{(0,-.7)}{.3}{0}{0}{$\eta_d$};
\node at (.2,.6) {$\scriptstyle c$};
\node at (.2,.2) {$\scriptstyle d$};
\node at (.2,-.2) {$\scriptstyle d$};
}
\quad
\overset{\Phi^2_{z,w}}{\longmapsto}
\quad
\sum_{\substack{
e,f\in \operatorname{Irr}(\cC)
\\
\alpha\in\operatorname{ONB}(a\otimes c\to e)
\\
\beta\in\operatorname{ONB}(b\otimes d\to f)
}}
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-1.2,0) -- (-1.2,-2.2) -- (1.2,-2.2) -- (1.2,0);
\filldraw[\rColor] (-1.2,0) -- (-1.2,-2.2) -- (1.2,-2.2) -- (1.2,0);
\end{scope}
\draw[dotted] (-1.2,0) -- (1.2,0);
\draw[orange,thick] (-.5,-1.4) arc (-180:0:.5cm);
\draw[orange, thick] (0,-1.9) -- (0,-2.2);
\filldraw[orange] (0,-1.9) circle (.05cm);
\draw (0,-.3) -- (0,.3);
\draw (.5,.8) arc (0:-180:.5cm);
\draw (.5,-.8) arc (0:180:.5cm);
\filldraw (0,-.3) circle (.05cm);
\filldraw (0,.3) circle (.05cm);
\draw[\zColor] (-.5,.8) -- (-1.2,.8) node[above, xshift=.2cm]{$\scriptstyle z$};
\draw[red] (.5,1.2) -- (-1.2,1.2) node[above, xshift=.2cm]{$\scriptstyle w$};
\draw[knot] (-.5,.8) -- (-.5,1.2) arc (180:0:.5cm) -- (.5,.8);
\filldraw[\zColor] (-.5,.8) circle (.05cm);
\filldraw[red] (.5,1.2) circle (.05cm);
\draw (0,1.7) -- (0,2.3);
\draw (-.5,2.8) arc (-180:0:.5cm);
\filldraw (0,1.7) circle (.05cm);
\filldraw (0,2.3) circle (.05cm);
\draw (-1.2,3.1) node[above, xshift=.2cm]{$\scriptstyle n$} -- node[above]{$\scriptstyle n$} (1.2,3.1) node[above, xshift=-.2cm]{$\scriptstyle n$};
\roundNbox{unshaded}{(.5,-1.1)}{.3}{0}{0}{$\eta_d$};
\roundNbox{unshaded}{(.5,3.1)}{.3}{0}{0}{$g_c$};
\roundNbox{unshaded}{(-.5,-1.1)}{.3}{0}{0}{$\xi_b$};
\roundNbox{unshaded}{(-.5,3.1)}{.3}{0}{0}{$f_a$};
\node at (.7,2.6) {\scriptsize{$c$}};
\node at (.7,1.4) {\scriptsize{$c$}};
\node at (-.7,2.6) {\scriptsize{$a$}};
\node at (-.7,1.4) {\scriptsize{$a$}};
\node at (.2,2) {\scriptsize{$e$}};
\node at (.7,.6) {\scriptsize{$d$}};
\node at (.7,-.6) {\scriptsize{$d$}};
\node at (-.7,.6) {\scriptsize{$b$}};
\node at (-.7,-.6) {\scriptsize{$b$}};
\node at (.2,.2) {\scriptsize{$f$}};
\node at (.2,-.2) {\scriptsize{$f$}};
\node at (0,.5) {\scriptsize{$\beta^\dag$}};
\node at (0,-.5) {\scriptsize{$\beta$}};
\node at (0,2.5) {\scriptsize{$\alpha^\dag$}};
\node at (0,1.5) {\scriptsize{$\alpha$}};
}
\quad
\underset{\text{\eqref{eq:SemisimpleIso}}}{\longleftrightarrow}
\quad
\sum_{f,\beta}
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-1.6,0) -- (-1.6,-2.2) -- (1.2,-2.2) -- (1.2,0);
\filldraw[\rColor] (-1.6,0) -- (-1.6,-2.2) -- (1.2,-2.2) -- (1.2,0);
\end{scope}
\draw[dotted] (-1.6,0) -- (1.2,0);
\draw[orange,thick] (-.5,-1.4) arc (-180:0:.5cm);
\draw[orange, thick] (0,-1.9) -- (0,-2.2);
\filldraw[orange] (0,-1.9) circle (.05cm);
\draw[thick, red] (.2,1) .. controls ++(180:.3cm) and ++(0:.3cm) .. (-.4,.6) -- (-.8,.6) .. controls ++(180:.45cm) and ++(0:.45cm) .. (-1.6,1) node[left]{$\scriptstyle w$};
\draw[thick, \zColor, knot] (-.8,1) .. controls ++(180:.45cm) and ++(0:.45cm) .. (-1.6,.6) node[left]{$\scriptstyle z$};
\draw (-1.6,1.3) node[above, xshift=.2cm]{$\scriptstyle n$} -- (1.2,1.3) node[above, xshift=-.2cm]{$\scriptstyle n$};
\draw[knot] (.5,.8) arc (0:-180:.5cm) -- (-.5,1);
\draw (.5,-.8) arc (0:180:.5cm);
\draw (0,-.3) -- (0,.3);
\filldraw (0,-.3) circle (.05cm);
\filldraw (0,.3) circle (.05cm);
\roundNbox{unshaded}{(.5,-1.1)}{.3}{0}{0}{$\eta_d$};
\roundNbox{unshaded}{(.5,1.15)}{.35}{-.05}{-.05}{$g_b'$};
\roundNbox{unshaded}{(-.5,-1.1)}{.3}{0}{0}{$\xi_b$};
\roundNbox{unshaded}{(-.5,1.15)}{.35}{-.05}{-.05}{$f_b'$};
\node at (0,1.5) {\scriptsize{$n$}};
\node at (.7,.6) {\scriptsize{$d$}};
\node at (.7,-.6) {\scriptsize{$d$}};
\node at (-.6,.4) {\scriptsize{$b$}};
\node at (-.7,-.6) {\scriptsize{$b$}};
\node at (.2,.2) {\scriptsize{$f$}};
\node at (.2,-.2) {\scriptsize{$f$}};
\node at (0,.5) {\scriptsize{$\beta^\dag$}};
\node at (0,-.5) {\scriptsize{$\beta$}};
$$
under the semisimplicity isomorphism \eqref{eq:SemisimpleIso}.
We now show that the image of the unitary tensor functor $\Phi: \cZ(\cC) \to \mathsf{Bim_{fgp}}(M_\infty)$ lies in $\tilde{\chi}(M_\infty)$ when $N$ is non-Gamma.
\begin{lem}
Since $N\subseteq M$ is finite depth, there is a $k>0$ such that every $c\in \operatorname{Irr}(\cC)$ is isomorphic to a summand of $(X\boxtimes \overline{X}))^{\boxtimes k}=X^{\operatorname{alt}\boxtimes 2k}$.
There is a finite subset $\{e_i\}_{i=1}^m\subseteq \cC^\operatorname{op}(X^{\operatorname{alt} \boxtimes 2k} \to z\boxtimes X^{\operatorname{alt} \boxtimes 2k})$
such that
\begin{equation}
\label{eq:zX2k-PPbasis}
\operatorname{id}_{z\boxtimes X^{\operatorname{alt}\boxtimes 2k}}
=
\tikzmath{
\draw (-.7,.15) -- node[above]{$\scriptstyle 2k$} (.7,.15);
\draw[thick, \zColor] (-.7,-.15) -- node[below]{$\scriptstyle z$} (.7,-.15);
}
=
\sum_{i=1}^m
\tikzmath{
\draw (-1.2,.15) node[left]{$\scriptstyle 2k$} -- node[above]{$\scriptstyle 2k$} (1.2,.15) node[right]{$\scriptstyle 2k$};
\draw[thick, \zColor] (-1.2,-.15) node[left]{$\scriptstyle z$} -- (-.5,-.15);
\draw[thick, \zColor] (1.2,-.15) node[right]{$\scriptstyle z$} -- (.5,-.15);
\roundNbox{unshaded}{(-.5,0)}{.35}{-.05}{-.05}{$e_i$};
\roundNbox{unshaded}{(.5,0)}{.35}{-.05}{-.05}{$e_i^\dag$};
}\,.
\end{equation}
\end{lem}
\begin{proof}
For each $c\in \operatorname{Irr}(\cC)$, we pick
\begin{itemize}
\item
a finite family of isometries $\{\iota_c^i:c\to X^{\operatorname{alt}\boxtimes 2k}\}_{i=1}^{m_c}$ such that $(\iota_c^i)^\dag\cdot \iota_c^i=1_c$ and $\sum_{c}\sum_{i=1}^{m_c} \iota_c^i\cdot (\iota_c^i)^\dag =1_{X^{\operatorname{alt}\boxtimes 2k}}$, and
\item
an orthonormal basis $\{\alpha_c\} \subseteq \cC(c\to z\boxtimes X^{\operatorname{alt}\boxtimes 2k})$ under the isometry inner product, i.e., $\alpha_c^\dag \cdot \alpha_c' = \delta_{\alpha_c=\alpha'_c} \operatorname{id}_c$.
\end{itemize}
Then we have (reading diagrams right to left)
\[
\tikzmath{
\draw (-.7,.15) -- node[above]{$\scriptstyle 2k$} (.7,.15);
\draw[thick, \zColor] (-.7,-.15) -- node[below]{$\scriptstyle z$} (.7,-.15);
}
=
\sum_{c\in\operatorname{Irr}(\cC)}
\sum_{\alpha_c}
\tikzmath{
\draw (-.3,0) -- (.3,0);
\draw (-.3,0) arc (0:90:.3cm) -- (-.8,.3) node[left]{$\scriptstyle 2k$};
\draw[thick, \zColor] (-.3,0) arc (0:-90:.3cm) -- (-.8,-.3) node[left]{$\scriptstyle z$};
\draw (.3,0) arc (180:90:.3cm) -- (.8,.3) node[right]{$\scriptstyle 2k$};
\draw[thick, \zColor] (.3,0) arc (180:270:.3cm) -- (.8,-.3) node[right]{$\scriptstyle z$};
\filldraw (-.3,0) node[left]{$\scriptstyle \alpha_c$} circle (.05cm);
\filldraw (.3,0) node[right]{$\scriptstyle \alpha_c^\dag$} circle (.05cm);
\node at (0,-.2) {\scriptsize{$c$}};
}
=
\sum_{c\in\operatorname{Irr}(\cC)}
\sum_{\alpha_c}
\sum_{i=1}^{m_c}
\tikzmath{
\draw[dashed,rounded corners = 5] (-1.75,.5) rectangle (.2,-.5);
\draw (-1.2,0) -- (1.2,0);
\draw (-1.2,0) arc (0:90:.3cm) -- (-1.9,.3) node[left]{$\scriptstyle 2k$};
\draw[\zColor, thick] (-1.2,0) arc (0:-90:.3cm) -- (-1.9,-.3) node[left]{$\scriptstyle z$};
\draw (1.2,0) arc (180:90:.3cm) -- (1.7,.3) node[right]{$\scriptstyle 2k$};
\draw[\zColor] (1.2,0) arc (180:270:.3cm) -- (1.7,-.3) node[right]{$\scriptstyle z$};
\filldraw (1.2,0) node[right]{$\scriptstyle \alpha_c^\dag$} circle (.05cm);
\filldraw (-1.2,0) node[left]{$\scriptstyle \alpha_c$} circle (.05cm);
\roundNbox{unshaded}{(-.5,0)}{.3}{.05}{.05}{$\scriptstyle (\iota_c^i)^\dag$};
\roundNbox{unshaded}{(.65,0)}{.3}{0}{0}{$\scriptstyle \iota_c^i$};
\node at (-1.05,-.2) {\scriptsize{$c$}};
\node at (.05,-.2) {\scriptsize{$2k$}};
}\,.
\]
So we define our set $\{e_i\}$ to be $\bigcup_{c\in \operatorname{Irr}(\cC)}\{\alpha_c\cdot (\iota^i_c)^\dag\}_{i=1}^{m_c}$.
\end{proof}
\begin{prop}
\label{prop:-otimesXAI}
Let $\{e_i\}$ be as in \eqref{eq:zX2k-PPbasis} above.
For $n\geq 0$, define subsets $\{b_i\}$ and $\{b_i^{(n)}\}$ of $|X_z|$ by
\begin{equation}
\label{eq:b_i^nOverS}
b_i
:=
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-.7,0) -- (-.7,-1.4) -- (.7,-1.4) -- (.7,0);
\filldraw[\rColor] (-.7,0) -- (-.7,-1.4) -- (.7,-1.4) -- (.7,0);
\end{scope}
\draw[dotted] (-.7,0) -- (.7,0);
\draw[dashed] (0,-.7) -- (0,.7);
\draw (.7,.95) node[right]{$\scriptstyle 2k$} -- (-.7,.95) node[left]{$\scriptstyle 2k$};
\draw[orange,thick] (0,-1.4) -- (0,-.7);
\draw[\zColor, thick] (-.3,.65) -- (-.7,.65) node[left]{$\scriptstyle z$};
\roundNbox{unshaded}{(0,-.7)}{.3}{0}{0}{\scriptsize{$\Omega_N$}};
\roundNbox{unshaded}{(0,.8)}{.35}{-.05}{-.05}{$e_i$};
\node at (.2,.2) {\scriptsize{$1_\cC$}};
\node at (.2,-.2) {\scriptsize{$1_\cC$}};
}
\qquad\qquad
b_i^{(n)}
:=
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-1.3,0) -- (-1.3,-1.4) -- (.7,-1.4) -- (.7,0);
\filldraw[\rColor] (-1.7,0) -- (-1.7,-1.4) -- (.7,-1.4) -- (.7,0);
\end{scope}
\draw[dotted] (-1.3,0) -- (.7,0);
\draw (.7,1.5) node[right]{$\scriptstyle 2k$} -- (-1.3,1.5) node[left]{$\scriptstyle 2k$};
\draw[orange,thick] (0,-1.4) -- (0,-.7);
\draw[\zColor, thick] (-.3,1.2) .. controls ++(180:.45cm) and ++(0:.45cm) .. (-1.1,.4) -- (-1.3,.4) node[left]{$\scriptstyle z$};
\draw[knot] (.7,.8) node[right]{$\scriptstyle 2n$} -- (-1.3,.8) node[left]{$\scriptstyle 2n$};
\draw[dashed] (0,-.7) -- (0,1.35);
\roundNbox{unshaded}{(0,-.7)}{.3}{0}{0}{\scriptsize{$\Omega_N$}};
\roundNbox{unshaded}{(0,1.35)}{.35}{-.05}{-.05}{$e_i$};
\node at (.2,.2) {\scriptsize{$1_\cC$}};
\node at (.2,-.2) {\scriptsize{$1_\cC$}};
}.
\end{equation}
Then $\{b_i\}$ is an $|X_z|$-basis and $\{b_i^{(n)}\}$ is an approximately inner $|X_z|_{M_\infty}$-basis.
This implies $|X_z|$ is approximately inner.
\end{prop}
\begin{proof}
The first claim is immediate from \eqref{eq:zX2k-PPbasis}.
Similarly, $\{b_i^{(n)}\}$ is an $|X_z|_{M_\infty}$-basis for every fixed $n$, and moreover, $[b_i^{(n)}, a]=0$ for all $a\in M_{2n}\subseteq M_\infty$.
Since $M_\infty = \varinjlim M_n$, for $a\in M_\infty$, $\|a-E_{M_n}(a)\|_2\to 0$
(e.g., see \cite[Lem.~B.7]{2010.01067}).
Then for all $a\in M_\infty$,
\begin{align*}
\|ab_i^{(n)}-b_i^{(n)}a\|_2
&\le \|(a-E_{M_n}(a))b_i^{(n)}\|_2+\|E_{M_n}(a)b_i^{(n)}-b_i^{(n)}E_{M_n}(a)\|_2+\|b_i^{(n)}(E_{M_n}(a)-a)\|_2 \\
& \le 2\|b_i^{(n)}\|_2\|E_{M_n}(a)-a\|_2 \longrightarrow 0.
\qedhere
\end{align*}
\end{proof}
\begin{assumption}
\label{assume:nonGamma}
For the remainder of this section, we now assume the $\rm II_1$ factor $N$ in our finite index finite depth subfactor $N\subseteq M$ is non-Gamma.
This implies $M$ is also non-Gamma by \cite[Prop.~1.11]{MR860811}.
\end{assumption}
\begin{lem}
\label{lem:CentralSequenceCommPPbasis}
Let $\{b_i\}$ be as in \eqref{eq:b_i^nOverS}.
For each central sequence $(a_n)\subseteq N_\infty$, $\|a_nb_i-b_ia_n\|_2\to 0$.
\end{lem}
\begin{proof}
Since $N\subseteq M$ is finite depth
and $N$ is non-Gamma,
by \cite[Prop.~3.2(3)]{MR2661553}, every Jones basic construction $M_n$ has spectral gap in $M_\infty$ for every $n$.
This implies that $\|a_n-E_{M_{k+1}'\cap M_\infty}(a_n)\|_2\to 0$.
Since $[E_{M_{k+1}'\cap M_\infty}(a_n),b_i]=0$, we have
\begin{align*}
\|a_nb_i-b_ia_n\|_2
&\le \|(a_n-E_{M_{k+1}'\cap M_\infty}(a_n))b_i\|_2
+
\|E_{M_{k+1}'\cap M_\infty}(a_n)b_i-b_iE_{M_{k+1}'\cap M_\infty}(a_n)\|_2
\\
&\quad
+\|b_i(E_{M_{k+1}'\cap M_\infty}(a_n)-a_n)\|_2
\\
& \le \|b_i\|_2\cdot \|a_n-E_{M_{k+1}'\cap M_\infty}(a_n)\|_2
+
\|a_n-E_{N_{k+1}'\cap M_\infty}(a_n)\|_2\cdot \|b_i\|_2
\longrightarrow
0. \qedhere
\end{align*}
\end{proof}
\begin{prop}
\label{prop:-otimesCT}
$|X_z|$ is centrally trivial.
\end{prop}
\begin{proof}
By Proposition \ref{prop:BimodualCT&CS},
we must show that for each central sequence $(a_n)\subseteq M_\infty$,
$\|a_nx-xa_n\|_2\to 0$, for every $x\in |X_z|$.
Suppose $(a_n)$ is such a central sequence.
Let $\{b_i\}$ be the $|X_z|_{M_\infty}$-basis as in \eqref{eq:b_i^nOverS}, and let $K>0$ such that $\|b_i\|_2\le K$ for all $i$.
By Lemma \ref{lem:CentralSequenceCommPPbasis},
\begin{align*}
\|a_nx-xa_n\|_2
&=
\left\|a_n \sum_i b_i\langle b_i|x\rangle^{|X_z|}_{M_\infty} \right. - \left.\sum_i b_i\langle b_i|x\rangle^{|X_z|}_{M_\infty} a_n\right\|_2 \\& \le
\left\|\sum_i (a_n b_i- b_ia_n)\langle b_i|x\rangle^{|X_z|}_{M_\infty}\right\|_2
+
\left\|\sum b_i\left(a_n\langle b_i|x\rangle^{|X_z|}_{M_\infty}-\langle b_i|x\rangle^{|X_z|}_{M_\infty} a_n\right)\right\|_2
\\& \le
K\|x\|_2 \cdot \sum_i\|a_n b_i- b_i a_n\|_2
+
K\sum_i\left\|a_n\langle b_i|x\rangle^{|X_z|}_{M_\infty} -\langle b_i|x\rangle^{|X_z|}_{M_\infty} a_n\right\|_2 \\
& \longrightarrow 0. \qedhere
\end{align*}
\end{proof}
Combining the statements of Propositions \ref{prop:-otimesXAI} and \ref{prop:-otimesCT},
our unitary tensor functor $\Phi: \cZ(\cC) \to \mathsf{Bim_{fgp}}(M_\infty)$ lands in $\tilde{\chi}(M_\infty)$.
\begin{prop}
The unitary tensor functor $\Phi:\cZ(\cC)\to \tilde{\chi}(M_\infty)$ is braided, i.e., for $z,w\in\cZ(\cC)$, the following diagram commutes
\begin{equation}
\label{eq:BraidedFunctor}
\begin{tikzcd}[column sep=4em]
{|X_z|\boxtimes_{M_\infty} |X_w|}
\arrow["u_{|X_z|,|X_w|}", rightarrow]{r}
\arrow["\Phi^2_{z,w}"', rightarrow]{d}
&
{|X_w|\boxtimes_{M_\infty} |X_z|}
\arrow["\Phi^2_{w,z}", rightarrow]{d}
\\
{|X_{z\otimes w}|}
\arrow["\Phi(\sigma_{z,w})"', rightarrow]{r}
&
{|X_{w\otimes z}|}
\end{tikzcd}
\end{equation}
\end{prop}
\begin{proof}
Let $\{b_i^{(n)}\}\subseteq |X_z|$ and $\{c_j\}\subseteq |X_w|$ be approximately inner $|X_z|_{M_\infty}$-basis and $|X_w|_{M_\infty}$-basis respectively as in \eqref{eq:b_i^nOverS}.
By \eqref{eq:uXY},
$u_{|X_z|,|X_w|} = \lim_n \sum_{i,j} |c_j\boxtimes b_i^{(n)}\rangle \langle b_i^{(n)}\boxtimes c_j|$.
For $x=(f_a\otimes \xi_a)\in |X_z|$ and $y=(g_b\otimes \eta_b)\in |X_w|$ with $f_a\in \cC(X^{\operatorname{alt}\boxtimes 2t}\to a\boxtimes X^{\operatorname{alt}\boxtimes 2t})$
and
$g_b\in \cC(X^{\operatorname{alt}\boxtimes 2t}\to b\boxtimes X^{\operatorname{alt}\boxtimes 2t})$ for $t$ sufficiently large,
we have
\begin{align*}
\sum_{i,j}
\Phi^2_{w,z}(|c_j\boxtimes b_i^{(n)}\rangle
&
\langle b_i^{(n)}\boxtimes c_j|x\boxtimes y\rangle_{M_\infty}^{|X_z|\boxtimes_{M_\infty}|X_w|}
)
\\&=
\sum_{i,j}
\sum_{\substack{
c\in \operatorname{Irr}(\cC)
\\
\alpha\in\operatorname{ONB}(a\otimes b\to c)}}
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-2.3,-2.7) rectangle (6,.1);
\filldraw[\rColor] (-2.3,-2.7) rectangle (6,-.4);
\end{scope}
\draw[dotted] (-2.3,-.4) -- (6,-.4);
\draw[orange,thick] (4,-1.8) arc (-180:0:.65cm and .5cm);
\draw[orange, thick] (4.65,-2.3) -- (4.65,-2.7);
\filldraw[orange] (4.65,-2.3) circle (.05cm);
\draw[\zColor, thick] (.2,1.7) .. controls ++(180:.45cm) and ++(0:.45cm) .. (-.8,.2) -- (-1.3,.2) .. controls ++(180:.45cm) and ++(0:.45cm) .. (-2.3,.6) node[left]{$\scriptstyle z$};
\draw[red, thick, knot] (-1.3,.6) .. controls ++(180:.45cm) and ++(0:.45cm) .. (-2.3,.2) node[left]{$\scriptstyle w$};
\draw[\zColor, thick, knot] (2.8,1.7) .. controls ++(0:.45cm) and ++(180:.45cm) .. (3.7,.6);
\draw[red, thick, knot] (1.8,.6) .. controls ++(0:.45cm) and ++(180:.45cm) .. (2.8,.2) -- (4,.2) .. controls ++(0:.45cm) and ++(180:.45cm) .. (5,.6);
\draw[knot] (-2.3,2.4) -- (6,2.4);
\draw[knot] (-2.3,2) node[left]{$\scriptstyle 2k$} -- (6,2);
\draw[knot] (-2.3,1.3) node[left]{$\scriptstyle 2n$} -- (6,1.3);
\draw[knot] (-2.3,.9) node[left]{$\scriptstyle 2k$} -- (6,.9);
\draw[knot] (4,.4) arc (-180:0:.65cm and .5cm);
\draw (4,-1.2) arc (180:0:.65cm and .5cm);
\draw (4.65,-.7) -- (4.65,-.1);
\filldraw[black] (4.65,-.1) circle (.05cm);
\filldraw[black] (4.65,-.7) circle (.05cm);
\roundNbox{unshaded}{(-1,.75)}{.35}{-.05}{-.05}{$c_j$};
\roundNbox{unshaded}{(.5,1.85)}{.35}{-.05}{-.05}{$b_i$};
\roundNbox{unshaded}{(1.5,.75)}{.35}{-.05}{-.05}{$c_j^\dag$};
\roundNbox{unshaded}{(2.5,1.85)}{.35}{-.05}{-.05}{$b_i^\dag$};
\roundNbox{unshaded}{(4,1.5)}{1.1}{-.8}{-.8}{$f_a$};
\roundNbox{unshaded}{(5.3,1.5)}{1.1}{-.8}{-.8}{$g_b$};
\roundNbox{unshaded}{(4,-1.5)}{.3}{0}{0}{$\xi_a$};
\roundNbox{unshaded}{(5.3,-1.5)}{.3}{0}{0}{$\eta_b$};
\node at (3.9,.05) {\scriptsize{$a$}};
\node at (3.9,-.85) {\scriptsize{$a$}};
\node at (5.4,.1) {\scriptsize{$b$}};
\node at (5.4,-.9) {\scriptsize{$b$}};
\node at (4.8,-.25) {\scriptsize{$c$}};
\node at (4.8,-.55) {\scriptsize{$c$}};
\node at (4.65,.1) {\scriptsize{$\alpha^\dag$}};
\node at (4.65,-.9) {\scriptsize{$\alpha$}};
}
\displaybreak[1]\\&=
\sum_{i,j}
\sum_{\substack{
c\in \operatorname{Irr}(\cC)
\\
\alpha\in\operatorname{ONB}(a\otimes b\to c)}}
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (-1.9,-2.7) rectangle (6,.1);
\filldraw[\rColor] (-1.9,-2.7) rectangle (6,-.4);
\end{scope}
\draw[dotted] (-1.9,-.4) -- (6,-.4);
\draw[orange,thick] (4,-1.8) arc (-180:0:.65cm and .5cm);
\draw[orange, thick] (4.65,-2.3) -- (4.65,-2.7);
\filldraw[orange] (4.65,-2.3) circle (.05cm);
\draw[\zColor, thick, knot] (-.8,1.7) .. controls ++(180:.45cm) and ++(0:.45cm) .. (-1.8,.6) -- (-1.9,.6) node[left]{$\scriptstyle z$};
\draw[\zColor, thick, knot] (2.8,1.7) .. controls ++(0:.45cm) and ++(180:.45cm) .. (3.7,.6);
\draw[red, thick, knot] (.2,.6) .. controls ++(180:.45cm) and ++(0:.45cm) .. (-.8,.2) -- (-1.9,.2) node[left]{$\scriptstyle w$};
\draw[red, thick, knot] (1.8,.6) .. controls ++(0:.45cm) and ++(180:.45cm) .. (2.8,.2) -- (4,.2) .. controls ++(0:.45cm) and ++(180:.45cm) .. (5,.6);
\draw[knot] (-1.9,2.4) -- (6,2.4);
\draw[knot] (-1.9,2) node[left]{$\scriptstyle 2k$} -- (6,2);
\draw[knot] (-1.9,1.3) node[left]{$\scriptstyle 2n$} -- (6,1.3);
\draw[knot] (-1.9,.9) node[left]{$\scriptstyle 2k$} -- (6,.9);
\draw[knot] (4,.4) arc (-180:0:.65cm and .5cm);
\draw (4,-1.2) arc (180:0:.65cm and .5cm);
\draw (4.65,-.7) -- (4.65,-.1);
\filldraw[black] (4.65,-.1) circle (.05cm);
\filldraw[black] (4.65,-.7) circle (.05cm);
\roundNbox{unshaded}{(-.5,1.85)}{.35}{-.05}{-.05}{$b_i$};
\roundNbox{unshaded}{(.5,.75)}{.35}{-.05}{-.05}{$c_j$};
\roundNbox{unshaded}{(1.5,.75)}{.35}{-.05}{-.05}{$c_j^\dag$};
\roundNbox{unshaded}{(2.5,1.85)}{.35}{-.05}{-.05}{$b_i^\dag$};
\roundNbox{unshaded}{(4,1.5)}{1.1}{-.8}{-.8}{$f_a$};
\roundNbox{unshaded}{(5.3,1.5)}{1.1}{-.8}{-.8}{$g_b$};
\roundNbox{unshaded}{(4,-1.5)}{.3}{0}{0}{$\xi_a$};
\roundNbox{unshaded}{(5.3,-1.5)}{.3}{0}{0}{$\eta_b$};
\node at (3.9,.05) {\scriptsize{$a$}};
\node at (3.9,-.85) {\scriptsize{$a$}};
\node at (5.4,.1) {\scriptsize{$b$}};
\node at (5.4,-.9) {\scriptsize{$b$}};
\node at (4.8,-.25) {\scriptsize{$c$}};
\node at (4.8,-.55) {\scriptsize{$c$}};
\node at (4.65,.1) {\scriptsize{$\alpha^\dag$}};
\node at (4.65,-.9) {\scriptsize{$\alpha$}};
}
\displaybreak[1]\\&=
\sum_{\substack{
c\in \operatorname{Irr}(\cC)
\\
\alpha\in\operatorname{ONB}(a\otimes b\to c)}}
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (3.3,-2.7) rectangle (6,.1);
\filldraw[\rColor] (2.7,-2.7) rectangle (6,-.4);
\end{scope}
\draw[dotted] (3.3,-.4) -- (6,-.4);
\draw[orange,thick] (4,-1.8) arc (-180:0:.65cm and .5cm);
\draw[orange, thick] (4.65,-2.3) -- (4.65,-2.7);
\filldraw[orange] (4.65,-2.3) circle (.05cm);
\draw (3.3,.9) node[left]{$\scriptstyle 2t$} -- (6,.9);
\draw[\zColor, thick] (3.7,.6) -- (3.3,.6) node[left]{$\scriptstyle z$};
\draw[red, thick, knot] (5,.6) .. controls ++(180:.45cm) and ++(0:.45cm) .. (4,.2) -- (3.7,.2) -- (3.3,.2) node[left]{$\scriptstyle w$};
\draw[knot] (4,.4) arc (-180:0:.65cm and .5cm);
\draw (4,-1.2) arc (180:0:.65cm and .5cm);
\draw (4.65,-.7) -- (4.65,-.1);
\filldraw[black] (4.65,-.1) circle (.05cm);
\filldraw[black] (4.65,-.7) circle (.05cm);
\roundNbox{unshaded}{(4,.75)}{.35}{-.05}{-.05}{$f_a$};
\roundNbox{unshaded}{(5.3,.75)}{.35}{-.05}{-.05}{$g_b$};
\roundNbox{unshaded}{(4,-1.5)}{.3}{0}{0}{$\xi_a$};
\roundNbox{unshaded}{(5.3,-1.5)}{.3}{0}{0}{$\eta_b$};
\node at (3.9,.05) {\scriptsize{$a$}};
\node at (3.9,-.85) {\scriptsize{$a$}};
\node at (5.4,.1) {\scriptsize{$b$}};
\node at (5.4,-.9) {\scriptsize{$b$}};
\node at (4.8,-.25) {\scriptsize{$c$}};
\node at (4.8,-.55) {\scriptsize{$c$}};
\node at (4.65,.1) {\scriptsize{$\alpha^\dag$}};
\node at (4.65,-.9) {\scriptsize{$\alpha$}};
}
\\&=
\sum_{\substack{
c\in \operatorname{Irr}(\cC)
\\
\alpha\in\operatorname{ONB}(a\otimes b\to c)}}
\tikzmath{
\begin{scope}
\clip[rounded corners = 5] (1.7,-2.7) rectangle (6,.1);
\filldraw[\rColor] (1.7,-2.7) rectangle (6,-.4);
\end{scope}
\draw[dotted] (1.7,-.4) -- (6,-.4);
\draw[orange,thick] (4,-1.8) arc (-180:0:.65cm and .5cm);
\draw[orange, thick] (4.65,-2.3) -- (4.65,-2.7);
\filldraw[orange] (4.65,-2.3) circle (.05cm);
\draw (1.7,.9) node[left]{$\scriptstyle 2t$} -- (6,.9);
\draw[red, thick] (5,.6) .. controls ++(180:.45cm) and ++(0:.45cm) .. (4,.2) -- (3.7,.2) .. controls ++(180:.45cm) and ++(0:.45cm) .. (2.7,.6) .. controls ++(180:.45cm) and ++(0:.45cm) .. (1.7,.2) node[left]{$\scriptstyle w$};
\draw[\zColor, thick, knot] (3.7,.6) .. controls ++(180:.45cm) and ++(0:.45cm) .. (2.7,.2) .. controls ++(180:.45cm) and ++(0:.45cm) .. (1.7,.6) node[left]{$\scriptstyle z$};
\draw[knot] (4,.4) arc (-180:0:.65cm and .5cm);
\draw (4,-1.2) arc (180:0:.65cm and .5cm);
\draw (4.65,-.7) -- (4.65,-.1);
\filldraw[black] (4.65,-.1) circle (.05cm);
\filldraw[black] (4.65,-.7) circle (.05cm);
\roundNbox{unshaded}{(4,.75)}{.35}{-.05}{-.05}{$f_a$};
\roundNbox{unshaded}{(5.3,.75)}{.35}{-.05}{-.05}{$g_b$};
\roundNbox{unshaded}{(4,-1.5)}{.3}{0}{0}{$\xi_a$};
\roundNbox{unshaded}{(5.3,-1.5)}{.3}{0}{0}{$\eta_b$};
\node at (3.9,.05) {\scriptsize{$a$}};
\node at (3.9,-.85) {\scriptsize{$a$}};
\node at (5.4,.1) {\scriptsize{$b$}};
\node at (5.4,-.9) {\scriptsize{$b$}};
\node at (4.8,-.25) {\scriptsize{$c$}};
\node at (4.8,-.55) {\scriptsize{$c$}};
\node at (4.65,.1) {\scriptsize{$\alpha^\dag$}};
\node at (4.65,-.9) {\scriptsize{$\alpha$}};
\draw[dashed,rounded corners = 5] (1.7,.05) rectangle (2.7,.75);
}
\\&=
\Phi(\sigma_{z,w})\cdot \Phi^2_{z,w} (x\boxtimes y).
\end{align*}
Thus \eqref{eq:BraidedFunctor} commutes on a dense subspace of $|X_z|\boxtimes_{M_\infty}|X_w|$, and the result follows.
\end{proof}
\begin{cor}
Let $\cC$ be a braided fusion category,
then there exists a $\mathrm{II}_1$ factor $M$ with a braided fully faithful monoidal functor $\cC\hookrightarrow \cZ(\cC)\to \tilde{\chi}(M)$.
\end{cor}
\subsection{Calculation of \texorpdfstring{$\tilde{\chi}(M_\infty)$}{chi(Minfty)} when \texorpdfstring{$N\subseteq M$}{N in M} is non-Gamma and finite depth}
As in \S\ref{sec:InductiveLimit} above we suppose $N\subseteq M$ is a fixed finite depth finite index $\rm II_1$ subfactor.
We also continue Assumption \ref{assume:nonGamma} that $N$ is non-Gamma.
We now prove our main result, which uses a technical result on centralizers in braided tensor categories in \S\ref{sec:TechnicalBTC} below.
\begin{thm}
\label{thm:nonGamma}
Let $N\subseteq M$ be a finite depth finite index inclusion of non-Gamma $\rm II_1$ factors.
Let $M_\infty$ denote the inductive limit $\rm II_1$ factor from the Jones tower, and let $\cC={}_N\cC_N$ be the even part of the standard invariant $\cC(N\subseteq M)$.
Then $\tilde{\chi}(M_\infty)\cong \cZ(\cC)$.
\end{thm}
\begin{proof}
Consider our construction of $\Phi:\cZ(\cC)\to\tilde{\chi}(M_\infty)$.
Let $L:=I(1_\cC)\in \cZ(\cC)$ be the canonical Lagrangian algebra, where $I: \cC\to \cZ(\cC)$ is adjoint to the forgetful functor.
By Example \ref{ex:MoritaEquivalenceBetweenCopCandZ(C)}, the Q-systems $L\in \cZ(\cC)$ and $S\in \cC^{\operatorname{op}}\boxtimes \cC$ are related as in Remark \ref{rem:MoritaEquivalenceRealization}, i.e., $S = X\otimes \overline{X}$ and $L=\overline{X}\otimes X$ for $X=1_\cC$ in the Morita equivalence $\cC^{\operatorname{op}}\boxtimes \cC - \cZ(\cC)$ bimodule category $\cC$.
Thus we can identify $|\Phi(L)|$ with the basic construction of $R\otimes N\subseteq M_{\infty}$ by the discussion in Example \ref{ex:RealizedJonesTower}.
But this implies that $|\Phi(L)|$ is Morita equivalent to $R\otimes N$, and in particular, by Example \ref{ex:Chi_fus(RboxtimesN)},
$\tilde{\chi}(|\Phi(L)|)=\tilde{\chi}(R\otimes N)$ is trivial.
By Proposition \ref{thm:LocalExtension}, $\tilde{\chi}(M_\infty)_{\Phi(L)}^{\operatorname{loc}}\cong \tilde{\chi}(|\Phi(L)|)$.
This implies no non-trivial simple object in $\tilde{\chi}(M_\infty)$ centralizes $\Phi(\cZ(\cC))$,
since the free module functor $x\mapsto x\otimes L$ for $x\in\Phi(\cZ(\cC))'\subseteq \tilde{\chi}(M_\infty)$ is fully faithful.
Thus $\Phi(\cZ(\cC))'\subseteq \tilde{\chi}(M_\infty)$ is trivial.
Since $\cZ(\cC)$ is non-degenerately braided by \cite{MR1966525},
by Proposition \ref{EXISTCENTRAL} in the next subsection,
$\Phi(\cZ(\cC))=\tilde{\chi}(M_\infty)$.
\end{proof}
\begin{rem}
Kawahigashi studied a relative version of Connes' $\chi(M)$ and Jones $\kappa$ invariant for finite index $\rm II_1$ subfactors $N\subseteq M$ \cite{MR1230287, MR1321700}.
In particular, Kawahigashi provides bounds and computations for relative $\chi$ for finite depth finite index subfactors of the hyperfinite $\rm{II}_{1}$ factor.
For a given finite depth \textit{hyperfinite} subfactor $N\subseteq M$, there exists a non-Gamma inclusion $A\subseteq B$ with the same standard invariant \cite{MR2051399}.
By \cite[Thm.~4.2]{MR2661553}, $\chi(A_{\infty})\cong \chi(N\subseteq M)$.
By Theorem \ref{thm:nonGamma} and Example \ref{ex:ConnesChi},
$\chi(N\subseteq M)\cong \operatorname{Inv}(\cZ(\cC(N\subseteq M)))$, the group of isomorphism classes of invertible objects of the Drinfeld center of $\cC(N\subseteq M)$.
\end{rem}
\begin{rem}
Suppose $N\subseteq M$ is a finite index inclusion of non-Gamma $\rm II_1$ factors with $A_{2n}$ Jones-Temperley-Lieb standard invariant.
Then $\cZ(\cC(N\subseteq M))$ is a unitary modular tensor category with no invertible objects.
This distinguishes the corresponding $M_{\infty}$ factors pairwise, but they all have the same trivial Connes' $\chi$ invariant.
Popa considers these examples of $\rm{II}_{1}$ factors with trivial $\chi$ which are not s-McDuff in \cite{MR2661553}, answering a question of Connes in the negative.
(Recall a $\rm{II}_{1}$ factor is \emph{s-McDuff} if it is of the form $R\otimes N$ for $N$ non-Gamma).
This leads us to ask the natural extension of Connes question.
\end{rem}
\begin{quest}
If $M$ is McDuff and $\tilde{\chi}(M)$ is trivial, is $M$ s-McDuff?
\end{quest}
\subsection{A technical result on centralizers in braided tensor categories}
\label{sec:TechnicalBTC}
The goal of this section is to prove the following technical result for braided unitary tensor categories.
We expect this result holds in the greater generality of semisimple ribbon tensor categories; see the paragraph before \cite[Prop.~2.5]{MR1990929} for more details in this direction.
\begin{prop}\label{EXISTCENTRAL}
Suppose $\cC$ is an arbitrary braided unitary tensor category and $\cD\subsetneq \cC$ is a non-degenerately braided proper fusion subcategory.
There exists $a\in \operatorname{Irr}(\cC)\backslash \operatorname{Irr}(\cD)$ which centralizes $\cD$.
\end{prop}
Let $\cC$ be a braided unitary tensor category.
Let $\operatorname{tr}_\cC$ denote the (unnormalized) categorical trace corresponding to the unique unitary spherical structure from minimal solutions to the conjugate equations \cite{MR1444286}.
(The normalization is
$\operatorname{tr}_\cC(\operatorname{id}_c) = \dim(c)$ for each $c\in \cC$.)
For each $a\in \operatorname{Irr}(\cC)$ define a function $\gamma_{a}: \operatorname{Irr}(\cC)\rightarrow \bbC$ by
$$
\gamma_{a}(b)
:=\frac{1}{d_{b}}\operatorname{tr}_\cC(\sigma_{b,a}\sigma_{a,b}).
$$
Then $\gamma_{a}(b)$ extends linearly to a \textit{character} on the fusion ring, i.e.
$$
\gamma_{a}(b)\gamma_{a}(c)=\sum_{d} N^{d}_{bc}\gamma_{a}(d).
$$
Furthermore, characters are related by the following equation:
\begin{equation}
\label{eq:CharacterProductExpansion}
\frac{\gamma_{a}(c)\gamma_{b}(c)}{d_{c}}=\sum_{e} \frac{d_{e}}{d_{a}d_{b}}N^{c}_{ab} \gamma_{e}(c).
\end{equation}
Suppose $\cD$ is a non-degenerately braided full fusion subcategory of $\cC$ (which is thus modular by unitarity).
Then by non-degeneracy, $\{\gamma_{a}\}_{a\in \operatorname{Irr}(\cD)}$ forms a complete set of characters of $K_0(\cD)$.
But any $b\in \operatorname{Irr}(\cC)$ also defines a character $\gamma_{b}$, and thus $\gamma_{b}|_{\operatorname{Irr}(\cD)}=\gamma_{f(b)}$ for some uniquely defined $f(b)\in \operatorname{Irr}(\cD)$.
Thus we have a function $f:\operatorname{Irr}(\cC)\rightarrow \operatorname{Irr}(\cD)$.
Now we define the \emph{fusion hypergroup} of a semisimple unitary tensor category $\cC$ as the fusion algebra $K_0(\cC)$ with the distinguished basis $\lambda_{a}=\frac{[a]}{d_{a}}$.
We then have
$$
\lambda_{a}\lambda_{b}=\sum_{c} M^{c}_{ab} \lambda_{c}
\qquad\qquad
\text{where}
\qquad\qquad
M^{c}_{ab}:=\frac{d_{c}}{d_{a}d_{b}} N^{c}_{ab}
.
$$
\begin{lem}
The assignment $f(\lambda_{a}):= \lambda_{f(a)}$ extends to a homomorphism of fusion algebras $K_0(\cC)\rightarrow K_0(\cD)$.
\end{lem}
\begin{proof}
For $x\in \operatorname{Irr}(\cD)$ and $a,b\in \operatorname{Irr}(\cC)$,
we compute $d_x^{-1}\gamma_{a}(x)\gamma_{b}(x)$ in two ways.
First, we can apply \eqref{eq:CharacterProductExpansion} and then swap $\gamma_{c}$ with $\gamma_{f(c)}$, or
we can swap
$\gamma_{a}, \gamma_{b}$ with $\gamma_{f(a)},\gamma_{f(b)}$ respectively and then apply \eqref{eq:CharacterProductExpansion}.
Equating these two computations gives the equality
$$
\sum_{c\in \operatorname{Irr}(\cC)} \frac{d_{c}}{d_{a}d_{b}}N^{c}_{ab} \gamma_{f(c)}(x)=\sum_{y\in \operatorname{Irr}(\cD)} \frac{d_{y}}{d_{f(a)}d_{f(b)}} N^{y}_{f(a)f(b)}\gamma_{y}(x),
$$
which implies
\begin{equation}
\label{eq:HypergroupIdentity}
\sum_{y\in \operatorname{Irr}(\cD)} \left( \sum_{c\in f^{-1}(y)} M^{c}_{ab} -M^{y}_{f(a)f(b)}\right) \gamma_{y}(x) =0.
\end{equation}
Since \eqref{eq:HypergroupIdentity} holds for all $x\in \operatorname{Irr}(\cD)$, we have
$$
\sum_{y\in \operatorname{Irr}(\cD)} \left( \sum_{c\in f^{-1}(y)} M^{c}_{ab} -M^{y}_{f(a)f(b)}\right) \gamma_{y}=0,
$$
which is an equation in the space of functions on $\operatorname{Irr}(\cD)$.
But since $\{\gamma_{y}\}_{y\in \operatorname{Irr}(\cD)}$ is a complete set of characters for the fusion algebra $K_0(\cD)$, it forms a basis for the space $\mathsf{Fun}(\operatorname{Irr}(\cD)\to \cC)$ (where we idenitfy $\mathsf{Fun}(\operatorname{Irr}(\cD)\to \cC)$ with the dual space $K_0(\cD)^{\vee}$), and is thus linearly independent.
This immediately implies
$$
\sum_{c\in f^{-1}(y)} M^{c}_{ab} =M^{y}_{f(a)f(b)}
\qquad\qquad
\forall\,y\in \operatorname{Irr}(\cD).
$$
We then see that
$$
f(\lambda_{a})f(\lambda_{b})=\lambda_{f(a)} \lambda_{f(b)}=\sum_{y} M^{y}_{f(a)f(b)}\lambda_{y}
=
\sum_{y} \left(\sum_{c\in f^{-1}(y)} M^{c}_{ab} \right) \lambda_{y}
=
\sum_{c} M^{c}_{ab} \lambda_{f(c)}
=
f(\lambda_{a}\lambda_{b})
$$
as claimed.
\end{proof}
\begin{proof}[Proof of Proposition~\ref{EXISTCENTRAL}]
We prove the contrapositive.
That is, we will show that if $\cC$ is a braided unitary tensor category and
$\cD\subseteq \cC$ is a non-degenerately braided fusion subcategory, then the absence of a nontrivial centralizing simple object for $\cD$ in $\operatorname{Irr}(\cC)$ implies $\cC=\cD$.
First, by \cite[Prop.~2.5]{MR1990929}, $c\in \operatorname{Irr}(\cC)$ centralizes $\cD$ if and only if $\gamma_{c}|_{\cD}=\gamma_{f(c)}=\gamma_{1_\cD}$.
Suppose that the only $c\in \operatorname{Irr}(\cC)$ for which $f(c)=1_\cD$ is $c=1_\cC$.
Let $\tau_\cC$ be the functional on $K_0(\cC)$ which picks off the coefficient of the identity object, and similarly define $\tau_\cD$ on $K_0(\cD)$.
Note that $\tau_\cC,\tau_\cD$ are positive definite on the $*$-algebras $K_0(\cC),K_0(\cD)$ respectively.
For any $\eta\in K_0(\cC)$, our hypothesis implies $\tau(\eta)=\tau(f(\eta))$, where we have
extended $f$ linearly.
But then $f:K_0(\cC)\rightarrow K_0(\cD)$ is injective, since if
$f(\eta)=0$, then $0=f(\eta^{*})f(\eta)=f(\eta^{*}\eta)$, and thus $0=\tau_\cD(f(\eta^{*}\eta))=\tau_\cC(\eta^{*}\eta)$, which implies $\eta=0$.
In particular,
$$
\operatorname{rank}(\cC)=\dim(K_0(\cC))\le \dim(K_0(\cD))=\operatorname{rank}(\cD).
$$
But as $\cD\subseteq \cC$, we must have $\cD=\cC$.
\end{proof}
\bibliographystyle{alpha}
{\footnotesize{
|
\section*{Plain Language Summary}
The solar wind represents a continuous outflow of charged particles from the Sun’s upper atmosphere into the solar system. Upon reaching Earth's magnetosphere, the solar wind's dynamic pressure is balanced by the magnetic pressure of Earth's magnetic field in a boundary layer known as the magnetopause. This boundary layer represents the entry point of the solar wind's energy into Earth's magnetosphere and upper atmosphere, playing a crucial role in energy transport throughout the interconnected system. Plasma density and temperature differences across the boundary layer generate an electric current that supports the magnetopause. In this paper, we clarify the physical mechanism of the magnetopause current by using high-resolution data from NASA's MMS mission. We found a significant ion temperature contribution to the magnetopause current not identified in previous studies. Our results also indicated that the plasma electrons' contribution to the magnetopause current was significantly smaller than the ion contribution.
\section{Introduction}
The magnetopause is a magnetosphere boundary layer created through the dynamic pressure balance between the solar wind's kinetic pressure and Earth's magnetic field. The solar wind causes distortions in the magnetosphere’s magnetic field topology supported by a current sheet first proposed by Chapman and Ferraro in 1931 \cite{CF1931}, often termed the Chapman-Ferraro (CF) current, which runs in a dawn-to-dusk direction around the magnetopause \cite{Ganushkina2017}. This current structure is believed to be generated through pressure gradients at the magnetopause boundary layer where, as explained in \citeA{Hasegawa2012}, the magnetosheath plasma has a higher plasma density, while the magnetosphere will have a higher ion temperature. The resulting changes in plasma density and temperature across the magnetopause leads to gradients that generate ion and electron diamagnetic currents running perpendicular to the magnetic field \cite{Ganushkina2017}.
Because of the magnetopause's important role in magnetic reconnection and the resulting transfer of plasma and energy into the magnetosphere it has been the focus of numerous studies [\citeA{Cahill1963}; \citeA{Le1994}; \citeA{PhanLarson1996}; \citeA{PhanPaschmann1996}; \citeA{Haaland2014}; \citeA{Paschmann2018}; \citeA{Haaland2019}; \citeA{Shuster2019}; \citeA{Haaland2020}; etc.] and missions [MMS, THEMIS, and Cluster] which have delved deeper into the current sheet’s structure and creation. From \citeA{Paschmann2018} and their MMS magnetopause crossing database, the total current density across the dayside magnetopause was studied in detail. The flank magnetopause total current density was then surveyed in \citeA{Haaland2019} and \citeA{Haaland2020} where the flanks were found to have a weaker current density and a correspondingly thicker boundary layer than the dayside.
While the literature generally recognizes the importance of both the density and temperature in generating diamagnetic currents, a large scale systematic analysis of {their contribution to the} magnetopause current system has not yet been accomplished. To help fill this gap in the literature, we considered four years of burst mode MMS mission data over the magnetopause crossing intervals provided by \citeA{Paschmann2018}'s MMS database. We measured both the density and temperature diamagnetic current components and created current accumulations of their
contributions. In doing so, we found that the temperature diamagnetic current component is a statistically significant factor to both the dayside and flank magnetopause current sheet by acting against the density component and thus reducing the total diamagnetic current density.
\section{Data and Analysis}
\subsection{MMS Database}
For this study, we used four years of data from \citeA{Paschmann2018,Paschmann2020}'s and \citeA{Haaland2020}'s database of MMS current sheet crossings. MMS is a mission comprised of four separate spacecraft traveling in a tetrahedron pattern through the magnetopause \cite{MMS_paper}. This database catalogues MMS current sheet crossings based on Fast Plasma Investigation (FPI) \cite{FPI}, Fluxgate Magnetometer \cite{FGM}, and Hot Plasma Composition Analyzer (HPCA) measurements \cite{HPCA}. The magnetopause transit times are captured through an automated minimum variance analysis {(MVA)} of the magnetic field in boundary normal or LMN coordinates \cite{Paschmann2018,Paschmann2020}. The start and end times of the current sheet crossing were then assigned as covering 76\% of the primary magnetic field component's {($B_L$'s)} transition across the magnetopause boundary, {which represents the thickness of a tanh fit of the $B_L$ profile}.
{The magnetopause velocities for each crossing are obtained through a de Hoffman-Teller analysis and stored by the database as the de Hoffman-Teller velocity ($V_{HT}$). We then took these velocities and dotted them with the normal direction in the LMN coordinate system to get the normal magnetopause boundary velocity ($V_n$) \cite{Sonnerup1998,Khrabrov1998a,Paschmann2018}. The magnetopause thicknesses are then determined by taking this normal velocity and multiplying it by the time span of the current sheet as determined by the MVA analysis.}
The database places identifiers on the individually identified magnetopause crossings, classifying their characteristics and structure. A full description of this process and the current sheet identifiers can be found in \citeA{Paschmann2018} and \citeA{Haaland2020}. An example crossing, with the database defined magnetopause current sheet crossing indicated by the dashed orange lines, is given in Figure \ref{fig:example_crossing}.
\begin{figure}[htbp]
\centering
\noindent\includegraphics[width=0.90\textwidth]{Jdia_LMN_2987.jpg}
\caption{Example crossing during a monotonic (constant magnetopause velocity), complete 2015 MMS transit from the magnetosheath into the magnetosphere {at the following Cartesian GSE position along the dayside magnetopause, indicated in Earth Radii (X,Y,Z): (9.4 Re, -6.5 Re, -1.0 Re)}. The orange lines represent the magnetopause current sheet as identified by {the MMS database. (a) Magnitude of the magnetic field, (b) magnetic field in LMN coordinates, (c,d) ion and electron omnidirectional spectrograms, (e) ion number density, (f) ion perpendicular and parallel temperature, and (g,h,i,j) curlometer, total ion diamagnetic, density component, and temperature component current densities respectively in LMN coordinates with magnitudes indicated in black.}}
\label{fig:example_crossing}
\end{figure}
\subsection{Magnetopause Current Sheet Selection}
We chose database defined crossings from 2015 to 2018 using event identifiers to select for complete and monotonic magnetopause crossings, where monotonic indicated events that had a constant magnetopause velocity so their thicknesses and durations could be computed. Additionally, we included Harris sheet-like events, or simple clear magnetopause crossings that were also complete, monotonic events in our data set. Events with unusually high (above $2,000$ $nA/m^2$) current {density} spikes during the magnetopause crossing time, such as would occur during a reconnection event, or when data flags for the various instruments were non-nominal, were manually removed from our data set.
{Alongside the database defined event criteria, we imposed two additional conditions on our events in order to ensure high signal-to-noise ratios and typical magnetopause plasma number densities as described below:}
{First, we considered the current measured during the magnetopause crossing and selected for events that reported significant peaks in their current densities. Where “significant" in our case was considered to be a current crossing where at least 50\% of the crossing duration was within 15\% of the maximum current peak during that crossing. This condition enabled us to select events with a strong current and high signal-to-noise ratios, thus ensuring that the temperature and density gradients and their resulting diamagnetic current components were not artificially diminished by the higher noise floor of low current density magnetopause crossings.}
{Second, we used HPCA measurements to compare the number densities of $H^{+}$ with $O^{+}$ for each magnetopause crossing. If $O^{+}$ exceeded $0.2$ $cm^{- 3}$ during the transit and $H^{+}$ remained below $1.5$ $cm^{- 3}$, then we considered $O^{+}$ to dominate the magnetospheric ion mass density by more than a factor of 2 as described by \citeA{Fuselier2019}. Events fitting this classification were also removed from our data set as they represented densities not typically found in the magnetopause current sheet.}
{From the application of the database identifiers as well as our conditions, we were able to identify 561 events. The locations and corresponding years of the selected crossings are denoted in Figure \ref{fig:crossings}.}
\begin{figure}[htbp]
\centering
\noindent\includegraphics[width=0.60\textwidth]{Event_Locs.jpg}
\caption{Diagram of our 561 MMS magnetopause crossings from 2015 (red), 2016 (green), 2017 (orange), and 2018 (blue). We define a local spherical coordinate system with $\phi$ in the $X_{GSE}$ - $Y_{GSE}$ plane, positively defined from the $+X_{GSE}$ axis, R defined as radially outward, and $\theta$ as the {polar} angle into the $Z_{GSE}$ direction. Note, every $15^{\circ}$ in $\phi$ is equal to 1 hour of MLT with 12 MLT corresponding to $0^{\circ}$ in $\phi$, or along the +$X_{GSE}$ axis. {The Dawn flank is defined as $-50^{\circ}$ to $-90^{\circ}$ in $\phi$, the Dusk flank as $+50^{\circ}$ to $+90^{\circ}$, and the Dayside as $+50^{\circ}$ to $-50^{\circ}$}.
Note, MMS first launched in 2015 with an orbit focusing on the dayside magnetopause, but after 2017, this orbit was extended to a wider orbit focusing on the flank magnetopause \cite{Haaland2020}. Because of the varying solar cycle, this has the possibility of creating an asymmetry between the dawn and dusk flank plasma measurements.}
\label{fig:crossings}
\end{figure}
\subsection{Current Calculations}
MMS's four separate spacecraft allows the total current to be calculated by the curlometer method \cite{Dunlop1988} that uses all four spacecraft to perform the curl of the observed magnetic field in order to approximate Ampere's law in the MHD approximation {(e.g. \citeA{Ganushkina2017})}. This current we call the curlometer current:
\begin{equation}
\textbf{J}_{curl} = \frac{\nabla \times \textbf{B}}{\mu_0}.
\end{equation}
Using the curlometer method, we also calculated the gradient of the ion density, and the divergence of the temperature tensor to get the total ion diamagnetic current and its current components ($\textbf{J}_{dia \ \nabla N_i}$ and $\textbf{J}_{dia \ \nabla \cdot \overleftrightarrow{T_i}}$ ). We found the electron diamagnetic current to be at least one order of magnitude smaller, on average, than the ion diamagnetic current. This is in agreement with the results of \citeA{Dong2018} in their case study where they found that the ion diamagnetic current was the primary source of perpendicular current in the magnetopause. Thus we are presenting results for the ion components and will drop the signifier “i” going forward. The components {of the total diamagnetic current are found by imputing the ideal gas law, $\overleftrightarrow{P} = Nk_b \overleftrightarrow{T}$, into the expression for the total diamagnetic current $J_{\perp} = \frac{\vec{B} \times \nabla \cdot \overleftrightarrow{P}}{|B|^2}$, see e.g. \citeA{Ganushkina2017} and the references therein, resulting in the following expressions}:
\begin{equation}
\textbf{J}_{dia \ \nabla N} = \frac{\textbf{B}\times(K_b\overleftrightarrow{T} \cdot \nabla N)}{|\textbf{B}|^2} \ \ , \ \
\textbf{J}_{dia \ \nabla \cdot \overleftrightarrow{T}} = \frac{\textbf{B}\times(K_bN\nabla \cdot \overleftrightarrow{T})}{|\textbf{B}|^2}.
\end{equation}
Where, by definition, $\textbf{J}_{dia \ Total} = \textbf{J}_{ dia \ \nabla N} + \textbf{J}_{dia \ \nabla \cdot \overleftrightarrow{T}}$.
{These currents were then averaged over our selected magnetopause crossings for each event, leading to single averaged current values for each of our 561 crossings. See Figure \ref{fig:example_crossing} for an example crossing on the dayside magnetopause. Note the two vertical orange lines represent the database defined current sheet crossing where the resulting currents would be averaged over.}
All of the data taken from MMS, as well as the current calculations, was interpolated to the 30 ms FPI electron time resolution. As our main results involve ion diamagnetic currents and the total current as computed from the curlometer method, any sub 150 ms variations in the ion measurements will not impact our results.
For non-curlometer calculations, we averaged over all four spacecraft to create a single data stream where, on average, the MMS separation during 2015 - 2018 was 10 - 60 km while the magnetopause current crossings analyzed typically had thicknesses greater than several hundred km, sufficiently larger than the max 60 km tetrahedron separation. These calculations were completed in GSE coordinates and then converted to a local spherical coordinate system built off of the Cartesian GSE coordinates. See Figure \ref{fig:crossings} for a depiction of our spherical coordinate system and the definition of the dusk and dawn sectors.
{Note, one limitation with applying the curlometer method to a magnetopause current sheet crossing is that the curlometer method requires simultaneous measurements from all four spacecraft to calculate a gradient, curl, or divergence of a quantity. Thus, errors occur when one or more of the MMS constellation is outside of the magnetopause boundary as the spacecraft are no longer all measuring the same medium. For our study, this caveat is mitigated by the fact that the magnetopause boundary is, on average, at least one order of magnitude larger than the average MMS separation, with median magnetopause thicknesses often reported as approximating 1,000 km (e.g. \citeA{Haaland2019,Haaland2020}). This makes times where the curlometer method results in erroneous measurements brief and limited to the outskirts of a current sheet crossing.}
\subsection{Current Accumulations}
The averaged currents from the 561 magnetopause crossings were put into bins corresponding to MMS’s physical location in our local spherical coordinate system. The angle $\phi$ was used to create 1-dimensional bins from $90^{\circ}$ to $-90^{\circ}$ in $20^{\circ}$ increments. This was done for each component in the spherical coordinate system as described in Figure \ref{fig:crossings} and shown in Figure \ref{fig:accumulations}. Error bars for each figure were computed using the standard error or $\sigma/\sqrt{N}$, with $\sigma$ the standard deviation of the values in each bin and N the number of events that fell inside that bin, {with the bins visually depicted in Figure 3 (a-d) by the grey-dashed lines.}
Using Figure \ref{fig:accumulations} we can make several observations. The first is that the $\textbf{J}_{curl}$, $\textbf{J}_{dia \ Total}$, and $\textbf{J}_{dia \ \nabla N}$ $\phi$-component currents are all in the $+ \phi$ direction across the magnetopause, or in the classical CF, dawn-to-dusk direction. However, the $\textbf{J}_{dia \ \nabla \cdot \overleftrightarrow{T}}$ $\phi$ component is consistently in the $-\phi$ direction, or from dusk-to-dawn across the magnetopause. Therefore the two components of the ion diamagnetic current, $\textbf{J}_{dia \ \nabla N}$ and $\textbf{J}_{dia \ \nabla \cdot \overleftrightarrow{T}}$, are oppositely directed across the magnetopause as can be seen in Figure \ref{fig:accumulations}b. The directions of these components are as expected when using the magnetospheric quantities evaluated by \citeA{Hasegawa2012} where the density component should run in the traditional CF current direction as the plasma density is higher in the magnetosheath and lower in the magnetosphere. At the same time, it is expected that the ion temperature is lower in the magnetosheath and higher in the magnetosphere, leading to the ion temperature component typically running counter to the CF current direction. On average, however, $\textbf{J}_{dia \ \nabla N}$ is stronger than $\textbf{J}_{dia \ \nabla \cdot \overleftrightarrow{T}}$, which allows the total diamagnetic current, $\textbf{J}_{dia \ Total}$, to still flow in the classical CF direction.
\begin{figure}[htbp]
\centering
\noindent\includegraphics[width=\textwidth]{Figure_3.jpg}
\caption{ a) through d) depict current density accumulations for $\textbf{J}_{curl}$, $\textbf{J}_{dia \ Total}$, $\textbf{J}_{dia \ \nabla N}$, and $\textbf{J}_{dia \ \nabla \cdot \overleftrightarrow{T}}$ across the dayside and flank magnetopause sectors, represented by the {orange} dashed lines. {The grey dashed lines represent the edges of the $20^{\circ}$ accumulation bins, centered at $80^{\circ}$, $60^{\circ}$, $40^{\circ}$, etc.} Moving from top to bottom: a). represents the R-component of the current in our local spherical coordinate system (described in Figure \ref{fig:crossings}). b). $\phi$ - component, c). $\theta$ - component, d). magnitude of the current components. e) diagram of the contributions and directions of $\textbf{J}_{dia \ Total}$, $\textbf{J}_{dia \ \nabla N}$, and $\textbf{J}_{dia \ \nabla \cdot \overleftrightarrow{T}}$ across the dusk, dayside, and dawn magnetopause. Note the size of the arrows in each sector indicates the relative magnitude of their current densities and the direction indicates the current component's flow around the magnetopause.}
\label{fig:accumulations}
\end{figure}
\subsection{Current Measurement Results}
We then used our data to create a table of results over the dusk, dayside, and dawn magnetopause including the mean, median, and standard errors for our 561 magnetopause crossings as seen in Table \ref{tab:table1}.
{From this table, $\textbf{J}_{curl}$ is strongest on the dayside, with a dusk-dawn asymmetry as the dawn curlometer current is stronger than the dusk. Both $\textbf{J}_{dia \ Total}$ and $\textbf{J}_{dia \ \nabla N}$ show similar distributions with the dayside again being the strongest sector, but the dusk and dawn results are now in agreement within their standard errors. The $\textbf{J}_{dia \ \nabla \cdot \overleftrightarrow{T}}$ component shows a dusk-dawn asymmetry with a significantly stronger dusk current density than the dawn. The total electron diamagnetic current, $\textbf{J}_{e \ dia \ Total}$, is the weakest current component studied and shows a dusk-dawn asymmetry with the dusk being significantly weaker than both the dusk and the dayside.}
{Comparing these results with past studies, our magnetopause thickness are, in general, higher than those found by \citeA{Haaland2020}; however, this is to be expected as we are considering a specific subset of database events as described in the previous sections. For similar reasons, the current densities found by \citeA{Haaland2020} show differences, with our dawn current densities matching closely, but the dusk and dayside values showing deviations.}
\newpage
\begin{table}[h]
\caption{Magnetopause parameters, magnitudes, {and $\phi$ components} of current densities across Dusk: {($+50^{\circ}$ to $+90^{\circ}$), Dawn: ($-50^{\circ}$ to $-90^{\circ}$), and Dayside: ( $>$ $-50^{\circ}$ and $<$ $+50^{\circ}$)} crossings with the following format: mean (median) $\pm$ standard error of the values measured over each event's magnetopause crossing. Note, $V_n$ values are unsigned averages and medians of the normal magnetopause velocity.}
\centering
\begin{tabular}{c c c c}
\hline
Parameter & Dusk & Dayside & Dawn \\
\hline
Number of Crossings & 215 & 225 & 121 \\
Thickness (km) & 1961 (1429) $\pm$ 132 & 2032 (1379) $\pm$ 135 & 1950 (1393) $\pm$ 180 \\
Thickness ($d_i$) & 21.4 (14.3) $\pm$ 1.5 & 26.7 (16.4) $\pm$ 1.8 & 28.0 (15.9) $\pm$ 3.0 \\
Thickness ($R_{gi}$) & 35.6 (18.7) $\pm$ 4.3 & 85.9 (33.2) $\pm$ 12 & 33.2 (17.9) $\pm$ 4.0 \\
$V_n$ (km/s) & 131.5 (109.8) $\pm$ 7.0 & 99.8 (71.1) $\pm$ 7.5 & 94.3 (86.0) $\pm$ 6.3 \\
Duration (s) & 16.0 (11.2) $\pm$ 0.8 & 25.0 (18.7) $\pm$ 1.3 & 20.6 (14.7) $\pm$ 1.3 \\
\ \\
$|\textbf{J}_{curl}|$ $(nA/m^2)$ & 30.0 (17.6) $\pm$ 3.2 & 83.2 (67.7) $\pm$ 5.6 & 52.1 (35.0) $\pm$ 6.3 \\
$|\textbf{J}_{i \ dia \ Total}|$ $(nA/m^2)$ & 19.4 (8.4) $\pm$ 5.8 & 59.0 (47.5) $\pm$ 6.9 & 25.1 (17.5) $\pm$ 5.0 \\
$|\textbf{J}_{dia \ \nabla N_i}|$ $(nA/m^2)$ & 50.0 (28.3) $\pm$ 8.5 & 78.2 (64.7) $\pm$ 7.2 & 43.6 (32.7) $\pm$ 6.6 \\
$|\textbf{J}_{dia \ \nabla \cdot \overleftrightarrow{T}_i}|$ $(nA/m^2)$ & 31.5 (16.6) $\pm$ 7.0 & 20.0 (13.5) $\pm$ 5.8 & 19.6 (10.8) $\pm$ 2.9 \\
\ \\
$\textbf{J}_{curl} \ ({\phi})$ $(nA/m^2)$ & 21.1 (10.9) $\pm$ 2.7 & 83.0 (67.4) $\pm$ 5.3 & 45.3 (30.3) $\pm$ 5.9 \\
$\textbf{J}_{i \ dia \ Total} \ ({\phi})$ $(nA/m^2)$ & 18.2 (8.3) $\pm$ 5.2 & 58.7 (47.3) $\pm$ 5.2 & 22.6 (16.6) $\pm$ 5.0 \\
$\textbf{J}_{dia \ \nabla N_i} \ ({\phi})$ $(nA/m^2)$ & 43.2 (26.1) $\pm$ 6.7 & 78.0 (64.7) $\pm$ 6.7 & 39.4 (30.4) $\pm$ 6.6 \\
$\textbf{J}_{dia \ \nabla \cdot \overleftrightarrow{T}_i} \ ({\phi})$ $(nA/m^2)$ & -25.0 (-16.1) $\pm$ 6.2 & -19.3 (-13.3) $\pm$ 4.1 & -16.8 (-10.7) $\pm$ 2.7 \\
\ \\
$|\textbf{J}_{e \ dia \ Total}|$ $(nA/m^2)$ & 1.7 (0.5) $\pm$ 0.7 & 5.9 (4.4) $\pm$ 1.0 & 5.4 (3.6) $\pm$ 0.7 \\
\hline
\end{tabular}
\label{tab:table1}
\end{table}
\section{Temperature Gradient's Impact on the Magnetopause Current System}
Using our results from Figure \ref{fig:accumulations} and Table \ref{tab:table1}, we can posit three primary ways the ion temperature gradient impacts the magnetopause current {density} and, in doing so, create a 2D diagram to summarize our findings as shown in Figure \ref{fig:accumulations}e.
\begin{enumerate}
\item The divergence of the ion temperature tensor generates up to one third of the total ion diamagnetic current {density} in the $\phi$ direction.
\
Specifically, in the $\phi$ direction, {the ratio of the temperature-driven component to the density-driven component's contribution to the total ion diamagnetic current gives the following break down with} $\textbf{J}_{dia \ \nabla \cdot \overleftrightarrow{T}}$ making up {$37\%$} of the diamagnetic current {density} on the dusk, {$20\%$} on the dayside, and {$30\%$} on the dawn.
\
\item $\textbf{J}_{dia \ \nabla \cdot \overleftrightarrow{T}}$ goes in the opposite direction of the classical Chapman-Ferraro Current.
\
$\textbf{J}_{dia \ \nabla \cdot \overleftrightarrow{T}}$ is clearly in the $-\phi$ direction across the magnetopause when considering Figure \ref{fig:accumulations}b and Table 1. This results in $\textbf{J}_{dia \ \nabla \cdot \overleftrightarrow{T}}$ lowering the contribution of $\textbf{J}_{dia \ \nabla N}$, making the $\textbf{J}_{dia \ Total}$ less than $\textbf{J}_{curl}$ on average as seen in Table \ref{tab:table1}.
\
\item $\textbf{J}_{dia \ \nabla \cdot \overleftrightarrow{T}}$’s contribution to the magnetopause current {density} becomes more important toward the flank magnetopause.
\
$\textbf{J}_{dia \ Total}$ and $\textbf{J}_{dia \ \nabla N}$ are strongest on the dayside and grow steadily weaker on the dusk and the dawn flanks, with both flanks showing similar results for the current densities. This is in contrast to $\textbf{J}_{dia \ \nabla \cdot \overleftrightarrow{T}}$ which {increases in strength, particularly on the dusk flank}, resulting in the total diamagnetic current being decreased even further by $\textbf{J}_{dia \ \nabla \cdot \overleftrightarrow{T}}$'s impact on the flank mangetopause.
\end{enumerate}
{From Table \ref{tab:table1} and Figure \ref{fig:accumulations} we may also notice the difference between the curlometer and the total diamagnetic current densities across the magnetopause. This difference is generally expected as the curlometer current represents the total current density as measured by MMS across the magnetopause current layer, which includes contributions from both ion and electron currents and their perpendicular and parallel components. The total ion diamagnetic current is thus one component of the curlometer current. This being said, the total diamagnetic current density still represents the main contributor to the curlometer current, accounting for almost 70\% of the current density on the dayside.}
\section{Discussion}
\subsection{Gradient Formation}
These observations are generally consistent with previously literature regarding ion populations in the magnetosphere. From \citeA{Chappell2008}, the warm plasma cloak is defined as a population of 10 eV to 3 keV ions energized in the polar cap and magnetotail, which circulates in a dawn-to-dusk circulation pattern throughout the inner magnetosphere out to the magnetopause. As \citeA{Chappell2008} notes, the warm plasma cloak ions can occupy the same space as the much warmer and more energetic ring current ions, which circulate in the opposite direction across the magnetosphere, from dusk-to-dawn. Thus it is possible, on a simplified level of magnetospheric circulation, that the colder warm plasma cloak ions provide generating pressure for the density gradient component across the magnetopause in its dawn-to-dusk, CF current like direction, while the warmer ring current ions provide generating pressure for the temperature gradient component in its dusk-to-dawn direction. The density component's dayside-flank asymmetry could also be explained by additional density gradients generated by the plasmasphere drainage plume \cite{Borovsky2008}, which exhausts through the dayside magnetopause during storm conditions, enhancing the dayside with more cold ions, thus leading to an enhanced dayside $\textbf{J}_{dia \ \nabla N}$ while leaving the dusk and dawn components reliant solely on the warm plasma cloak ion population. {The presence of magnetosphere particle populations in the mangetopause and their effects on the magnetopause current sheet, specifically regarding magnetic reconnection, is explored in the following studies: e.g. \citeA{Borovsky2008}, \citeA{Fuselier2017}, and \citeA{Walsh2013}. Additional study regarding the effects of the warm plasma cloak and plasmasphere particle populations on the diamagnetic current and its} component generation is needed however.
\subsection{Large Scale Current System Closure}
{From Figure \ref{fig:accumulations} and Table \ref{tab:table1}, we can see the total diamagnetic current density is lower on the flanks, helped by an increasingly prominent temperature component. This indicates current closure with the larger 3D current system where the magnetopause current steadily curves toward the parallel as it goes around the flanks. As the total current through the magnetopause current system must remain constant, as charge is conserved, the current density of the total current's components must then fluctuate appropriately along the dayside and flanks to “transfer" the current density from the perpendicular dominated dayside to the increasingly parallel dominated flanks. This indicates that the total diamagnetic (perpendicular) current decreases in order for more of the total current density to be diverted toward parallel currents such as the field aligned currents. Additionally, the magnetopause is generally thinner on the dayside and thicker on the flanks while the magnetopause current must remain the same as it flows through the magnetopause (e.g. \citeA{Haaland2019,Haaland2020}). Thus the total perpendicular current density must change to compensate for either a thinner or thicker magnetopause. From Figure \ref{fig:accumulations}, we can see this is, indeed, the case as the total diamagnetic current density decreases on the flanks and is strongest toward the dayside.}
\subsection{Ion vs. Electron Diamagnetic Current Densities}
Even though the electron diamagnetic current was found to be significant on electron scale current sheets by \citeA{Shuster2019,Shuster2021}, we found the electron current {density} to be less significant over our ion scale magnetopause current sheets. Specifically the electron diamagnetic current density is {$7\%$} of the ion diamagnetic current density in the $\phi$ direction on the dusk, {$10\%$} on the dayside, and {$20\%$} on the dawn. This presents an interesting asymmetry for the electron current as it is noticeably weaker on average on the dusk than it is on the dawn; however, in both cases, the electron diamagnetic current is also weaker than the contribution made by the ion current. The weaker electron current density may be explained based on the fact that we are averaging over many electron scale current sheets when considering our ion scale magnetopause crossing, thus lowering the resulting current density.
\section{Conclusions}
{In this paper, we have quantified the diamagnetic current's composite nature. Based on our systematic analysis of four years of MMS magnetopause crossings, the diamagnetic current is composed of two competing components: one generated by density gradients and one by temperature gradients.}
{We have found that the temperature generated component acts against the density component, weakening the total diamagnetic current's net strength, particularly on the flanks where the temperature component's contribution can reach up to {37\%} of the diamagnetic current density along the magnetopause. We also found that ions contribute the majority of the current density to the diamagnetic current, with electrons accounting for only {7\%} to {20\%} of the ion's contribution.}
{Taking these findings into account, we can posit that enhancements of the temperature gradient along the magnetopause boundary may lead to a corresponding weakening of both the diamagnetic current and, by extension, the total current in the magnetopause. This implied weakening of the magnetopause current by the temperature gradient leads to a more complicated picture of the interaction between the solar wind and Earth's magnetosphere, especially on small scales where situations can arise where the two components of the diamagnetic current become equal yet opposite, leading to the net cancellation of the diamagnetic current in that region. Studying the small-scale consequences of this interaction is the basis of our future work.}
\acknowledgments
The MMS current sheet database was created by Goetz Paschmann and Stein Haaland, and further developed by the International Space Science Institute Team 442, “Study of the physical processes in magnetopause and magnetosheath current sheets using a large MMS database". We thank them as well as the entire MMS team and instrument leads for the data access and support. We also thank the pySPEDAS team for their support and data analysis tools. This research was supported by the NASA Magnetospheric Multiscale Mission in association with NASA contract NNG04EB99C. J. M. H. B. and V. M. U. were supported through the cooperative agreements NNG11PL10A and 80NSSC21M0180.
\section*{Open Research}
\noindent The MMS data used in this study is publicly available at https://lasp.colorado.edu/mms/sdc/public/datasets/ from the FPI,
FIELDS, and HPCA datasets. The averaged MMS data over the 561 MP crossings used for this study are available through a Harvard Dataverse public database: https://doi.org/10.7910/DVN/SRBJCR.
\section{} command to identify level 1 heads;
Math coded inside display math mode \[ ...\]
will not be numbered, e.g.,:
\[ x^2=y^2 + z^2\]
Math coded inside \begin{equation} and \end{equation} will
be automatically numbered, e.g.,:
\begin{equation}
x^2=y^2 + z^2
\end{equation}
\begin{eqnarray}
x_{1} & = & (x - x_{0}) \cos \Theta \nonumber \\
&& + (y - y_{0}) \sin \Theta \nonumber \\
y_{1} & = & -(x - x_{0}) \sin \Theta \nonumber \\
&& + (y - y_{0}) \cos \Theta.
\end{eqnarray}
|
\section{Optical tweezers}
Mechanical effects of light were first argued by Kepler to explain comet tails
\cite{kepler1619cometis}. The advent of the laser technology in the 60s enabled to reproduce these
effects on our planet starting a real scientific revolution
\cite{ashkin1970acceleration,ashkin1970atomic}. Optical tweezers
\cite{ashkin1986observation,jones2015optical,marago2013optical,Polimeno2018} are tools based on
tightly focused laser beams capable to trap, manipulate, and characterize a wide range of
microscopic and nanoscopic particles, in liquids, air, and vacuum
\cite{marago2013optical,Polimeno2018}. Since the pioneering work by Ashkin
\cite{ashkin1970acceleration,ashkin1970atomic,ashkin1986observation}, that led him to the Nobel
prize in Physics 2018, key applications of this contactless manipulation technique have been
developed in a wide range of fields: from biology, soft matter, and ultra-sensitive spectroscopy to
atomic physics, nanoscience, photonics, spectroscopy, and aerosols science \cite{jones2015optical}.
A crucial advancement has been the realization of Raman tweezers, \textit{i.e.}, the coupling of
optical tweezers with a Raman spectrometer \cite{petrov2007raman}. This allows the chemical and
physical analysis of a trapped particle through its vibrational fingerprints
\cite{gillibert2019raman,donato2018optical}.
Despite this tremendous experimental progress, the accurate light scattering modeling of optical
tweezers, that takes into account particle size, shape, and composition, has been developed only
recently \cite{Polimeno2018,borghese2007optical,borghese2008radiation}. In the limiting cases of
particles much smaller or much larger than the laser wavelength, optical forces in optical tweezers
can be divided in two components \cite{jones2015optical}: a gradient force, proportional to the
intensity gradient of the laser spot, responsible for trapping; and a scattering force,
proportional to the light intensity that tends to push particles away from the laser focus
destabilizing single-beam trapping of particles with large extinction. Such detrimental effects can
be suppressed through the use of two counter-propagating beams to null the opposite scattering
forces \cite{zemanek2003theoretical}. These dual-beam traps are based on the use of low numerical
aperture (NA) lenses and allow the trapping of particles with reduced incident power in a focal
region that is wider than for standard optical tweezers \cite{donato2018optical}. Thus they are
well suited for operation in air or vacuum where detrimental effects by radiation pressure are
enhanced by the reduced viscous damping \cite{gong2018optical,svak2018transverse,alali2020laser}.
Notwithstanding the remarkable improvement in optical trapping techniques, their application to
planetary exploration (see the pictorial representation in Fig. \ref{figure1a}) is still to be
developed, even though already conceived by, \textit{e.g.}, NASA \cite{NASA2014}. The development
of the optical trapping technique to collect and analyze \textit{in situ} or return to Earth a
variety of extraterrestrial particles will open doors to information on space materials that is
currently unreachable: 1) as the dust volatile component, not measurable in situ by dust
instruments, \textit{e.g.}, those onboard Rosetta/ESA, and not retrievable by samples return
missions, \textit{e.g.}, Stardust/NASA, Hayabusa; 2) have biases due to collection media
contamination, \textit{e.g.}, aerogel used by the Stardust/NASA space probe
\cite{rotundi2008combined,rotundi2014two}).
Here, we first review the role of dust in the universe: from interstellar, to interplanetary,
cometary and planetary dust particles. Thus, we describe the models and methods we use to calculate
light pressure and optical trapping properties on a variety of realistic dust particle models.
Then, we show results on solar radiation pressure calculations that can help to a greater
understanding of micro-particle dynamics as well as to estimate its detrimental role in optical
trapping in space. Finally, we consider and compare results of calculations for optical trapping of
dust particles in standard optical tweezers in water (typical laboratory conditions) with those
calculated for OT in air or in space.
\begin{figure}
\centering
\resizebox{12 cm}{!}{%
\includegraphics{Figure1.pdf}
}
\caption{Pictorial representation of space tweezers, space applications of optical tweezers. Interplanetary or planetary dust can be collected and investigated directly \textit{in situ} (open space or extraterrestrial surfaces). The inset represents a closeup up of a grain of interplanetary dust trapped by a single-beam optical tweezers.}\label{figure1a}
\end{figure}
\section{A dusty universe}
We live in a dusty galaxy, one of the billions of galaxies in the universe. Almost all of them are
dusty. Up to the mid-20th century, all of this dust was considered as an unfortunate impediment to
making precise observations of stars and galaxies, regarded as the most important items in the
universe. The modern view of dust as a cosmic component is almost entirely the reverse of that
earlier view. We now know that almost every aspect of the formation of planets, stars, and galaxies
is influenced in some way by interstellar dust \cite{CCP12}. The involvement of dust grains in
providing molecules important for the origin of life, and in the safe transmission of those species
to newly-forming planets orbiting Sun-like stars \cite{McGuire}, has been one of the greatest
surprises of all. It is no wonder that considerable efforts are currently placed in trying to
understand nature, composition, and evolution of dust in the interstellar medium of galaxies.
There is now a robust evidence that dust grains condense in circumstellar environments, mainly AGB
(Asymptotic Giant Branch) stars \cite{gail2014} and supernovae \cite{sarangi15,gall14}, from which
they are ejected in the general interstellar medium. There, dust is found to be tightly mixed with
the gas, with the dust representing only a minor fraction of the total mass. The gas-to-dust mass
ratio locally assumes the fairly constant value of 100 \cite{giannetti17}. The dependence of this
ratio on the metal content among galaxies and within a galaxy is an important issue from a cosmic
perspective for a number of reasons \cite{zafar13}, including the appearance of the first solids in
the early universe \cite{gall11}. In late-type galaxies, like our own, this ratio scales with the
metallicity, decreasing with the galactocentric radius, a clear indication that dust growth in the
interstellar medium dominates over destruction \cite{mattsson14}. As a consequence, stardust,
literally dust formed around stars, is not the same as interstellar dust, and must be regarded as
the raw material from which true interstellar grains are formed. Such a material is modified and
destructured by violent interstellar processes \cite{Bocchio14}, before being reformed and
reassembled in denser interstellar regions \cite{Draine09}.
Modifying processes continue to act throughout the lifetime of a dust grain, some hundreds of
millions of years in the Milky Way galaxy. In some cases, such modifications may be catastrophic,
as during the formation of a planetary system, a filter that severely modifies the composition of
interstellar dust. Those materials passing through the filter are the most robust to destruction
and erosion. In the solar system, the patterns of isotopic abundances that are found in survived
presolar grains clearly identify their origins in the cool envelopes of evolved low mass stars and
in supernovae ejecta. These particles have passed through many episodes of possible destruction,
including their ejection from the stellar envelopes, their passage through the interstellar medium,
where they have been subjected to intense radiation fields and dynamical shocks, their
incorporation into the molecular cloud that formed the solar system, and through all the varied and
violent processes involved in the formation of the Sun and its planets. As near-stellar dust is
modified in the interstellar medium and becomes interstellar dust, so interstellar dust may be
severely modified, when not obliterated, when it is incorporated in the gas that forms newly-born
planetary system.
\begin{figure}
\centering
\resizebox{\textwidth}{!}{%
\includegraphics{figure2.pdf}
}
\caption{Scattering models with shape and composition inspired by interstellar, interplanetary
(DUSTER mission \cite{rietmeijer2016laboratory}), and planetary \cite{oliva2019database} particles.
On the top, the models emulate hypothetical interstellar dust grains whose constituents are olivine and aliphatic carbon.
In (a) and (b), the constituents refractive indexes are mixed according to the Bruggeman criterion. Instead, in (c) and (d),
the olivine is considered covered by a carbon layer. In (e), the model of a silica particle arranged in quenched melt spheres shown in the
FESEM image \cite{rietmeijer2016laboratory} (f). In (g), the model of condensed Ca[O] nanograins that are accreted onto a larger melted aggregate
of tiny carbonate grains shown in the FESEM image \cite{rietmeijer2016laboratory} (h). The larger sphere is calcite and the other spheres
are CaO. In (i), a spherical model of the particle Fe, Mg-rich 'TP2' \cite{rotundi2014two}, in which we consider an effective refractive
index constructed mixing iron (67\%) and magnesium (33\%), according to the Bruggeman criterion. In (j), a spherical model of
Martian hematite \cite{oliva2019database}. In (k), a spherical model of Lunar regolith \cite{egan1973optical}. In (l), a model of an
ellipsoidal fassaite shown in the FESEM image (m). Here, we consider an effective refractive index constructed mixing
silica (53\%), CaO (27\%), FeO (10\%), Al$_2$O$_3$ (10\%), according to the Bruggeman criterion \cite{bohren2008absorption}.}\label{figure1}
\end{figure}
\subsection{Interstellar dust}
Most of the information that we have about interstellar dust is obtained remotely, by the influence
of dust on various kinds of astronomical observations. These observations may be carried out at a
very wide range of wavelengths, from X-rays to radio waves, but traditionally the most important
and defining data have come from the infrared, visible, and ultraviolet parts of the spectrum,
summarized through a wavelength-dependent extinction curve along the lines of sight to individual
stars.
There are detailed extinction measurements along hundreds of lines of sight in the Milky Way Galaxy
\cite{Fitzpatrick19}, and less accurate extinction data for the interstellar media of external
galaxies \cite{Naveen20}. Such observed extinction profiles are different, but are clearly members
of the same family showing a basic similarity in their shapes: the extinction typically increases
from low values in the infrared to high values in the far ultraviolet, a near-linear portion in the
optical region, (in most cases) a pronounced and broad "bump" near 217.5 nm, and a final rise of
varying slope into the far ultraviolet. Both the general aspect of the extinction and the details
of specific curves along particular lines of sight provide useful information, and generally
indicate that grains of a wide range of sizes (roughly nanometers to micrometers) are required
\cite{CCP15}.
The dominant feature in the extinction curve is the prominent bump in the near ultraviolet. Its
central position is fixed, although the width of the feature can vary significantly from one line
of sight to another. It is widely attributed to $\pi^\star \leftarrow \pi$ transitions in aromatic
carbon solids \cite{gadallah11} or polycyclic aromatic hydrocarbons \cite{tielensbook}. The
far-ultraviolet rise may be decomposed in a linear contribution, due to nano-sized particles, and a
non-linear component belonging to the partially invisible (because located beyond the Lyman
continuum) $\sigma^\star \leftarrow \sigma$ resonance in aromatic carbon. In the near-infrared,
there is a weak absorption feature at a wavelength of 3.4~$\mu$m, detectable on long paths through
diffuse gas. It is characteristic of absorption in the $sp^3$ (aliphatic) C-H stretching bond
\cite{Sandford91}. Further into the infrared are two stronger absorption features, at 9.7 and
18~$\mu$m, ascribable to silicate materials, from Si-O stretching and O-Si-O bending modes,
respectively. Taken together, these features strongly support the presence of some form of
carbon/hydrocarbon and silicate in the dust. X-ray scattering and absorption edges provide
constraints on grain size and composition, specifically O, Mg, Si, Fe, and C atoms \cite{Draine03}.
There is also a set of detected infrared emission features occurring at 3.3, 6.2, 7.7, 8.6, and
11.3~$\mu$m, indicative of aromatic CH groups. The mechanism responsible for the excitation of such
infrared emission may involve non-equilibrium emission from polycyclic aromatic hydrocarbons
stochastically heated to high temperatures by the absorption of individual photons from the
interstellar radiation field \cite{Leger84}. However, the requirement on microscopic sizes can be
relaxed if the emitters of the 3.3~$\mu$m and other infrared bands are heated by the chemical
energy released from reactions within larger carbon interstellar grains of mixed $sp^2/sp^3$ carbon
composition \cite{DW11}. Such structures have been in fact observed in some extragalactic objects
\cite{dartois07}. Exploiting a ternary phase diagram where the hydrogen content and the two main
bonding types ($sp^2$ and $sp^3$) for carbon constitute the poles, Dartois et al. (2007)
\cite{dartois07} were able to identify the carrier of the spectral features as an interstellar
hydrocarbon belonging to the class of polymeric-like hydrogenated amorphous carbon (a-C:H),
dominated by an aliphatic/olefinic backbone structure. The change from aliphatic to aromatic
structures may occur in environments that selectively dehydrogenate the a-C:H, providing an
opportunity for aromatic molecules to form. These observations, together with observations of very
evolved stars (protoplanetary and planetary nebulae), suggest an evolution in which aliphatics are
converted into aromatic structures \cite{antonia2008stratified,CCP10}.
The interpretation of dust observations must take also account of the available abundances along
the line of sight. The hydrogen abundance is often well-determined; abundances of other elements
relative to hydrogen are assumed, using solar or other relative values as a standard. It is unclear
which of these standards is the appropriate one to use \cite{Nieva08}, but whichever one is
adopted, the inventory of some materials is demonstrably incomplete. In fact, silicates alone are
unable to account for the entirety of either the oxygen solid phase or Fe abundances
\cite{Jenkins09}. The remainder of the Fe could be in iron oxides or in metallic form
\cite{Draine13}. The unaccounted oxygen is less readily explained \cite{Whittet10}. There is an
evident problem of oxygen budget in the dense interstellar medium, as the combined contribution of
gaseous CO and silicate/oxide dust are by far less of the lowest reference abundances
\cite{Asplund09}. Solid H$_2$O is an obvious O-bearing candidate. Astronomers have known for
decades that water is fairly common in the universe. Interstellar water ice, as distinct from
gaseous water molecules, was first identified toward the embedded protostellar Becklin-Neugebauer
object by Gillette \& Forrest (1973) \cite{Gillett73}. It is now widely detected in interstellar
dark clouds. These ices, composed mainly of water, also contain carbon monoxide and dioxide, simple
hydrides such as methane and ammonia, and a few other species \cite{boogert15}. Still, ice
contribution is not enough. According to Whittet (2010) \cite{Whittet10}, the most plausible
candidate for the depleted oxygen appears to be a form of O-bearing carbonaceous matter similar to
the organics found in cometary particles returned by the Stardust mission. Such materials may share
some similarities with refractory organic residues arising in the irradiation of icy mixtures by
ions \cite{baratta14} or X-rays \cite{ciaravella19}.
For its very nature, astronomical dust is likely to be an amalgam of a number of different
materials, very chaotic in composition and structure, with different individual substances
dominating at different wavelengths. These materials are thus fundamentally different from
terrestrial materials. Nuth \& Hecht (1990) \cite{Nuth90} introduced the concept of chaotic
silicates in which the level of disorder is even greater than for glasses, that are characterized
by the absence of long-range order in the atomic arrangement beyond nearest neighbours. Since
materials may be assembled in the agglomerate, an astronomical silicate cannot be considered a
solid with a definite stoichiometric composition. They may also occur in groups that recall solid
solutions, in which one or more types of atoms or molecules of the solid may be partly substituted
for the original atoms and molecules without changing the structure. Olivine and enstatite are
excellent examples of solid solutions. Forsterite, Mg$_2$SiO$_4$, and fayalite, Fe$_2$SiO$_4$, have
identical structures because the ions Mg$^{2+}$ and Fe$^{2+}$ are very nearly the same size and are
chemically similar. Very frequently amorphous silicates in space are misleadingly described in
terms of the optical properties of these materials. Indeed, as pointed out by Rietmeijer and Nuth
(2013) \cite{Rietmeijer13}, there are no amorphous silicates, as the word "silicate" already
implies that the material is crystalline and could be a mineral. Moreover, astronomical solids may
be porous and therefore of much lower density than a glass. Ultimately, the nature of an
astronomical silicate is rather loosely constrained, to same extent just limited to a material
whose infrared spectrum is dominated by Si-O stretching and bending vibrations. Thermal annealing
(e.g. in shocks) or intense X-ray irradiation \cite{Ciaravella16} of precursor materials, that were
probably amorphous, may explain the presence of crystalline silicates (see however Ritmejer and
Nuth 2013 \cite{Rietmeijer13}, for petrologic constraints) in circumstellar regions and
protoplanetary disks \cite{juhasz10}.
Even carbonaceous materials in space are difficult to constrain. A striking example is given by the
nature of the carrier of the interstellar ultraviolet extinction bump at 217.5 nm, that was
originally attributed to small crystalline graphite particles \cite{Stecher65}, followed by a
plethora of proposals including mixture of spheres composed of graphite, amorphous carbon, and
silicate \cite{Mathis89M}, irregular or fractal arrangement of graphite and amorphous carbon
\cite{Wright87}, polycyclic aromatic hydrocarbons \cite{Joblin92}, natural coal \cite{Papoular96},
and even electronic transitions of OH$^{-}$ ions in sites of low coordination in silicates
\cite{Steel87}. In general, carbonaceous materials contain greater or lesser hydrogen fractions,
varying proportions of different chemical bonding, and different degrees of long-range order. All
these forms of carbon can, under suitable conditions, be readily converted from one to another. The
manifold of possible bonding arrangements produces several allotropes of carbon of which the best
known are graphite, diamond and amorphous carbon. The physical properties of carbon vary widely
with the allotropic form. Amorphous carbonaceous materials cover a wide range of compositions, from
wide band gap, H-rich, aliphatic-rich a-C:H to narrow band gap, H-poor, aromatic-rich a-C
materials. The properties of a-C:H materials are determined by the $sp^3/sp^2$ ratio for the carbon
atoms and the hydrogen concentration. A C-H bond contributes to the formation of $sp^3$ bonding and
the reduction of the defects in the amorphous carbon network. In general, it is found that the
optical energy gap increases with hydrogen concentration \cite{Higa06}.
Remote observations such as extinction profiles were at one time the only sources of information
about interstellar dust. However, even if challenging it is now also possible to study in situ
interstellar grains entering the heliosphere by spacecrafts, and to collect them using
stratospheric aircrafts and balloons, and satellite probes. Rare interstellar dust grains can now
be examined in the laboratory being returned by the Stardust/NASA mission. During the Cassini
mission around Saturn, millions of dust particles have been analyzed and now, thanks to the Cosmic
Dust Analyzer on board the probe, 36 grains of dust supposed coming from outside our solar system
have been detected \cite{altobelli16}. Results from different missions generally deviate one from
another, and typically they are not very consistent with the dust picture obtained by remote
observations \cite{Draine09}. Obviously, this is a field in which much work is required.
\begin{figure}
\centering
\resizebox{\textwidth}{!}{%
\includegraphics{figure3.pdf}
}
\caption{Optical forces distributions exerted by the Sun on dust particles modeled as shown in Fig.~\ref{figure1}.
The optical force parallel, $f_\parallel(\lambda)$ (blue line), and perpendicular, $f_\perp(\lambda)$ (red line),
are intended respect to the light propagation direction $\hat{\mathbf{k}}$. In (a-d) we show calculations on olivine-aliphatic
carbon structures. In (a) and (b), the constituents refractive indexes are mixed according to the Bruggeman
criterion \cite{bohren2008absorption}. While in (c) and (d) the olivine and the carbon are distributed in a core-shell structure.
In (e) we show results for the quenched melt silica particle, in (f) for the bunch-of-grape carbon Ca-rich, in (g) for the 'TP2' sphere, in (h) for the martian hematite sphere, and in (i) for the fassaite ellipsoid.}\label{figure2}
\end{figure}
\subsection{The interplanetary dust complex}
The interplanetary space of the solar system is very dusty, appreciable to the naked eye through
the faint solar colour cone of light above the western horizon after sunset, or above the eastern
horizon just before sunrise: the Zodiacal Light, a huge amount of fine dust particles that scatter
solar radiation in the interplanetary space. The brightness of the Zodiacal Light provides
information on the overall space density of the dust known as the interplanetary dust complex.
The interplanetary dust complex consists of microscopic (mainly rocky) particles, typically less
than a few millimetres in size, called micrometeoroids, moving in the interplanetary space of the
solar system \cite{rubin2010meteorite}. Dust is produced by collisions among solid bodies, by
disruptions of icy bodies \cite{nesvorny2006physical,flynn2009dust} and by cometary activity
\cite{fulle2020comets,guttler2019synthesis}. As such the interplanetary dust complex is an
inventory of the constituent materials of a large variety of solid bodies in the solar system.
Their cosmochemical study thus allows investigation of great diversity of astrogeobiological
processes occurred on their parent bodies. Remarkably, these include the most primitive bodies not
further processed by planetesimal accretion after aggregation in the protoplanetary disk
\cite{blum2017evidence}. In addition, since the bulk of extraterrestrial matter accreted by the
Earth is in the form of cosmic dust from the interplanetary dust complex, understanding its
composition provides also clues on outstanding issues like the contribution of extraterrestrial
matter to the Earth's geochemical budget \cite{peucker1996accretion}, including its bearing on the
emergence of life \cite{sandford2016organic}. Based on particle-impact detection at the Long
Duration Exposure Facility/NASA, the main mass fraction of the submillimetre dust particles in the
Zodiacal cloud in the vicinity of the Earth have characteristic diameters of $\sim$100-200 $\mu$m,
and the flux of micrometeoroids entering the Earth's atmosphere is 40000$\pm$20000 tonnes per year
\cite{love1993direct}. Gr\"un et al. \cite{grun1985collisional} reached a similar conclusion and
quantified the spatial mass of cosmic debris at 1 AU as $10^{-16}$ g m$^{-3}$, with the largest
fraction in the $10^{-6}$ to $10^{-4}$ g mass range, based on the lunar microcrater record and
spacecraft micrometeoroid detectors.
How much asteroid collisions versus cometary activity contribute to the interplanetary dust complex
is an outstanding issue. Matter is complicated since such contributions are expected to vary
through the solar system's history; this is due to the stochastic fluctuations in the number of
mutual collisions in the asteroid belt and the number and level of activity of comets entering the
inner solar system. If astronomical observations carried out by the Infrared Astronomical
Satellite (IRAS) and Cosmic Background Explorer (COBE) satellites in the 1980's indicated that
collisions in the Main Asteroid Belt (MAB) are the dominant source of dust in the near-Earth space
\cite{kortenkamp2001sources}, dynamical simulations predict in turn that the bulk of the Zodiacal
cloud can be best produced by debris derived from Jupiter Family Comets (JFCs) through spontaneous
disruption \cite{nesvorny2010cometary}. Oxygen isotopic data from relatively large micrometeorites
in the 100-1000 $\mu$m size range collected at the Earth's surface
\cite{cordier2014oxygen,suttle2020extraterrestrial,suttle2020isotopic} indicates that the
interplanetary dust complex is dominated by dust produced by cometary activity and by collisions
between primitive hydrous asteroids of carbonaceous chondrite compositions, with a subordinate
contribution from more evolved anhydrous asteroids (mainly ordinary chondritic and Vesta-like). An
additional small fraction of micrometeorites, with heavy oxygen isotope composition, may sample an
unknown body in the solar system, namely a body not yet sampled by macroscopic meteorites. The
isotopic statistics shows that primitive, volatile-rich material dominates the small size fraction,
whereas evolved anhydrous material dominates the large size fraction of the dust complex.
\paragraph{Physical and compositional properties of the interplanetary dust complex.}
Most of the compositional properties of the interplanetary dust complex derives from the
cosmochemical analyses of samples recovered from the Earth's: 1) surface, e.g., micrometeorites
collected in Antarctica \cite{cordier2014oxygen,dionnet2020x,suttle2020extraterrestrial}; 2)
stratosphere, by balloon born instruments as, e.g., DUSTER, designed for non-destructive and
uncontaminated collection of solid particles from tens of microns down to 200 nm in size
\cite{della2012situ} and by stratospheric NASA/aircraft passive sticking on silicon oil coated
plates \cite{rietmeijer2002earliest}. A critical contribution is also given by laboratory analyses
of samples, i.e., collected, and brought back to Earth, from asteroid surfaces
\cite{dionnet2020combining} and in a cometary coma \cite{rotundi2008combined,rotundi2014two}. In
addition, cometary dust was studied in situ from the onset of cometary activity to its cessation
after perihelion by the Rosetta/ESA space mission
\cite{della201667p,fulle2015density,rotundi2015dust}.
The pristine mineralogy of micrometeorites is best unravelled by the study of a specific class: the
unmelted, i.e., particles that did not experienced alteration by frictional heating during
atmospheric passage \cite{luigi15micrometeorites}. Based on petrographic and geochemical data, most
of the micrometeorite flux reaching Earth today is dominated (>50\% of unmelted micrometeorites
across all size fractions) by fine-grained and hydrated carbonaceous chondrite material affine to
the CM, CR matrices and CI chondrites, with some specimens similar to the ungrouped meteorite
Tagish Lake. They are therefore a major component of the near-Earth dust complex in agreement with
isotopic statistics \cite{cordier2014oxygen}. Coarse-grained micrometeorites, which are primarily
fragments of chondrule, and CAI (Ca-, Al-rich inclusions) represent a smaller component (<25\%) of
the micrometeorite flux \cite{van2012chondritic,suttle2020extraterrestrial}. Thus, the major
constituent minerals are olivine, low-Ca pyroxene, magnetite, sulphides, metal and hydrous Fe-Mg
silicates like serpentine and saponite. The density of unmelted micrometeorites varies greatly
according to their mineral composition. Mean density for unmelted fine-grained micrometeorites is
$\sim$1.4 g cm$^{-3}$ (average data from Kohout et al. \cite{kohout2014density}), but can be more
than twice for coarse-grained micrometeorites. Micrometeorites range in size from 10 $\mu$m to 2000
$\mu$m \cite{rubin2010meteorite}, however particles up to 3000 $\mu$m in size have been reported in
the Transantarctic Mountains micrometeorite collection
\cite{rochette2008micrometeorites,suavet2009statistical}, and their mass varies within the nanogram
to the milligram range. A statistically significant number of micrometeorites in the 200-700 $\mu$m
size range shows that the micrometeorite size distribution is bimodal, with peaks centred at
$\sim$145 and $\sim$250 $\mu$m. This suggests that the micrometeorite flux is composed of multiple
dust sources with distinct size distributions: fine-grained material from primitive objects and
coarse-grained material from evolved bodies \cite{suttle2020extraterrestrial}.
DUSTER particles, collected with a strict contamination protocol during different stratospheric
balloon flight campaigns at 30-40 km altitudes, are the smallest (down to 200 nm) meteoritic
objects available for laboratory investigation (see Fig. \ref{figure1}f,h,l). These are the
residues of porous, structurally weak bolides crossing the Earth's atmosphere. They contribute to
the daily input of stratospheric extra-terrestrial material with: 1)tens of microns down to
sub-micrometre aggregates of nanometer-sized spherical grains, produced by vaporization and
quenching; 2) partially annealed minerals; and 3) unprocessed minerals
\cite{corte2013meteoric,rietmeijer2016laboratory}. DUSTER particles composition includes alumina,
aluminosilica, plagioclase, fassaite, silica, CaCO$_3$, CaO, extreme F-rich C-O-Ca particles, and
oxocarbon. These are particle linked to the friable CI and CM carbonaceous chondrite, and
unequilibrated ordinary chondrite meteors that are the most common source of bolides and fireballs
\cite{rietmeijer2016laboratory}.
Similar particles are collected at lower quotes in the stratosphere by NASA/aircrafts:
Interplanetary Dust Particles (IDPs), numerous and fine-grained, are available for laboratory
analyses \cite{rietmeijer2002earliest,rotundi2007combined}. They have a chondritic overall
composition, masses in the order of pico- to nano-grams and sizes of few microns
\cite{rubin2010meteorite}, but they can reach hundreds of microns for the Giant Cluster IDPs class.
They mainly consist of extremely fragile aggregate of crystalline and amorphous materials with
grain-size typically in the order of 100s of nm or smaller. Hydrous IDPs are mainly massive objects
with fibrous or platy surface textures. They consist mainly of hydrous mineral assemblages, mostly
smectite and lesser serpentine and silicate glass, with minor amounts of diopside, forsterite,
chromite, Fe- and Ni-sulfides, schreibersite (Ni-phosphide), magnetite, glasses of silicatic
composition and disordered carbonaceous material. Among IDPs are the aggregate anhydrous IDPs,
which are extremely fine-grained, highly porous (up to 70\%) and low-density (0.3-0.6 g cm$^{-3}$).
Following Rietmeijer's classification\cite{rietmeijer2002earliest}, they consist of a matrix
aggregates of generally spherical entities, $\sim$0.1 to $\sim$1 micron in size, with embedded
variable amounts of $\sim$5 micron-sized Mg, Fe- and Ca, Mg, Fe-silicates, Ni-free and low-Ni
pyrrhotite, iron oxides, and amorphous materials. Aggregate IDPs are supposed to have a cometary
origin \cite{brunetto2011mid} testified also by the comparison with cometary dust particles
returned by the Stardust/NASA space probe \cite{rotundi2008combined,rotundi2014two}. In fact, dust
particles after ejection from the nucleus form a dust flux \cite{della2019giada}, which undergoes
to solar radiation pressure effects, as observed in situ by GIADA onboard the ESA Rosetta space
probe \cite{della2016giada}, to gas drag and to gravity force \cite{zakharov2018asymptotics}. These
dust particles either remain tied to the nucleus or they are liberated into space feeding the
interplanetary dust complex, where their orbital evolution is mainly controlled by radiation
pressure and the Poynting-Robertson light drag \cite{dermott2001orbital}. This causes particles
from 1 to 1000s of $\mu$m in size to spiral slowly into the Sun, i.e., to gradually reduce the
orbit's eccentricity and semi-major axis, in a $10^4$ to $10^6$ year time scales, intercepting in
some cases Earth's atmosphere, where they are collected as IDPs.
\subsection{Planetary dust}
Many terrestrial planets and satellites of both terrestrial and giant planets in our solar system
show dusty environmental conditions. In particular, the mostly explored dusty bodies are the Moon
and Mars.
\paragraph{The Moon.}
The Moon is the only Earth's natural satellite. Since the Moon has neither a magnetic field nor a
significant atmosphere, the lunar regolith and the near-surface environment are mainly affected by
space weathering processes such as meteoroid impacts, solar ultraviolet radiation, solar wind,
galactic cosmic rays and plasma processes in the tail of the Earth's magnetosphere.
On the Moon, all the locations explored so far have cratered surfaces covered with loose regolith
of several meters. We can therefore assume that the entire surface is covered by regolith although
the thickness may vary (about 5 m in the maria, up to 20 m in the highlands). Such regolith layer
is a cohesive, dark grey to light grey, very-fine-grained, clastic material consisting of a mixture
of a wide variety of materials including fragments of highland anorthosites, some forms of KREEP
volcanic rocks, mare basalts and volcanic glass, plus a small meteoritic component
\cite{lucey2006understanding}. The influence of the solar wind and high-energy particles of solar
and cosmic origins induces the implantation of H, He, and many rare elements into the regolith.
Continued reworking by micrometeoroids of the hydrogen-enriched regolith particles causes the
material melting and the reactions (particularly, H with FeO), producing water vapor and submicron
grains of metallic iron sintered into glass \cite{mckay1991lunar}.
Considering the samples returned by the Apollo and Luna missions, the fine fraction of the lunar
regolith (45-100 $\mu$m in mean size) comprises about a half of the lunar regolith by weight; the
superfine fraction (particles smaller than 10 $\mu$m) about 10\%, while the particles smaller than
2 $\mu$m, the finest regolith fraction, make up 1-2\% of the mass. The shape of dust particles is
extremely irregular and highly variable, ranging from angular with sharp edges to spheroidal. The
density of individual particles is usually assumed in the range from 2.7 to 3.0 g cm$^{-3}$
\cite{carrier1991physical}. Individual particles may be glass-bounded aggregates
\cite{mckay1991lunar} called agglutinates. Agglutinate particles may have lower density (higher
porosity), they are usually <1 mm and contain minute droplets of Fe metal (much of which is very
fine-grained, single domain FeO), and troilite (FeS). They formed by the melting and mixing
produced by micrometeoritic bombardment of the lunar regolith.
The optical parameters of the dust particles (\textit{i.e.}, the real and imaginary parts of the
complex refractive index) depend on the composition and can vary significantly, not only over a
wide region, but even locally in a microscopic scale. For the glassy component, the real part
values range from 1.570 to 1.749 and they vary directly with the total Fe and Ti contents and
inversely with the Al content \cite{chao1970lunar}. The imaginary part of the complex refractive
index is more variable and can span from 0.0005 up to 0.15 according to the composition and
wavelength of interest.
Electrical properties of the dust particles are characterized by the extremely low electrical
conductivity (approximately $10^{-14}$ ohm/m in shadowed areas \cite{carrier1991physical}, but
about a hundred times higher under solar-light exposure) permitting accumulation of electrostatic
charge under ultraviolet irradiation.
\paragraph{Mars.}
Mars, the fourth planet of our solar system, has many features in common with the Earth. Much of
the Martian surface is covered by unconsolidated soils (dust) derived from impact, aeolian and
other sedimentary processes. Such dust is likely produced by impacts early in Mars' history and
subsequently recycled at its surface. Sometimes the wind in Mars' thin atmosphere blows the dust on
the surface into dust storms, carrying dust particles up to altitudes of about 50 km. The
composition of the Martian dust grains was obtained from several space missions, starting from the
mission Mariner 9 in 1971 \cite{toon1977physical}.
It is possible to build a mineralogical model of the dust using all the data acquired from the
Mariner 9 and Viking Landers \cite{clark1982chemical} space missions. Best fit results are those
obtained by mixing different kinds of clay, with montmorillonite Si$_2$Al$_4$O$_{10}$(OH)$_2$
$\cdot$nH$_2$O as a base, mixed with nontronite (where part of Al is replaced by Fe$^{3+}$) and
saponite (where Al is replaced by Mg). Indeed, data by the X spectrometer aboard the Pathfinder
showed higher abundances of magnesium and aluminum, and lower abundances of iron, chlorine and
sulphur \cite{foley2003final} respect to what expected. The rocks analyzed were similar to
andesites, terrestrial materials derived from the fractionation of basalt in intrusive conditions.
Data from recent rovers (e.g. MSL Curiosity at Rocknest, Oct 2012) and orbital spacecraft show that
Martian surface is dominated by a soil (dust grain dimension <150 $\mu$m) of basaltic composition
with primarily pyroxene, plagioclase feldspar, and olivine, as well as minor amounts of Fe and Ti
oxides (e.g., magnetite, ilmenite, and hematite) and alteration minerals (e.g., sulfates,
phyllosilicates, and carbonates)
\cite{morris2006,yen2005integrated,mcsween2010determining,bish2013x}. Dust particles properties
changes during dust storms \cite{chen2020dust}. In general, dust particle sizes range from 1.2 to
about 4.1 $\mu$m with a mean dust radius of about 1.6-1.8 $\mu$m
\cite{oliva2018properties,wolff2009wavelength} and a positive correlation between dust opacity and
particle sizes.
Phase function results show asymmetry parameter values of $g = 0.601$ $\pm$ 0.108 for high
atmospheric dust loading scenarios and $g = 0.710$ $\pm$ 0.065 for non-dusty periods. Regarding the
shape of the particles, considering a modified log-normal aspect ratio distribution for a mixture
of spheroids, data suggest more elongated particles are present during dust storms, with aspect
ratios of 2.8 $\pm$ 0.9 for high-opacity days, in contrast to values of 1.8 measured during
post-storm period. The particles single scattering albedo is found ranging in the solar band
between 0.89-0.90 (dark regions) and 0.92-0.94 (bright regions) \cite{wolff2010ultraviolet}.
Moreover, there are a very strong and a fast feedback between sand/dust emission and electric field
enhancement and a linear relation between the concentration of lifted dust and the generated
electric field \cite{esposito2016role}.
\begin{figure}
\centering
\resizebox{0.93\textwidth}{!}{%
\includegraphics{figure4.pdf}
}
\caption{Extinction cross-sections, $\langle \sigma_\mathrm{ext} \rangle$, and optical trapping force, $F$, components along $x$ (blue line), $y$ (red line), and $z$ (yellow dots) when the particle is trapped in water (center column) or in air (right column). Results in (a-c) are related to the olivine-aliphatic carbon core-shell sphere. (d-f) show calculations for the olivine-aliphatic carbon core-shell cluster. (g-i) concern the quenched melt silica. (j-l) show the results for the bunch-of-grape carbon Ca-rich. (m-o) are the results for the 'TP2' sphere. For the optical trapping calculations the laser power is fixed at 50 mW and the wavelength at $0.83$ $\mu$m.}\label{figure3}
\end{figure}
\section{Methods and models for space tweezers calculations}\label{sect:Model and methods}
Optical forces and optical trapping are the consequence of the electromagnetic momentum
conservation during a light scattering process \cite{jones2015optical,borghese2007scattering}. To
understand theoretically how light interacts mechanically with matter, Maxwell's equations and the
integration of the averaged Maxwell stress tensor have to be performed
\cite{mishchenko2002scattering,borghese2007scattering}. However, such calculations can be
computationally complex, $e.g.$ when we deal with non-spherical or non-homogeneous particles, and
the use of approximations can be also advantageous \cite{jones2015optical,polimeno2019optical}. The
range of validity of the different approximations is determined by the size parameter $x =
k_\mathrm{m} r$, where $k_\mathrm{m} = 2 \pi n_\mathrm{m} /\lambda_0$ is the radiation wavenumber
in the medium surrounding the particle, $r$ is the particle size, $\lambda_0$ is the wavelength in
vacuum, and $n_\mathrm{m}$ is the medium refractive index \cite{jones2015optical}. When $x \ll 1$,
the particle size is much smaller than the wavelength, and the dipole approximation can be adopted,
treating the particle as a dipole \cite{chaumet2000time,arias2003optical,albaladejo2009scattering}.
In the opposite case, when $x \gg 1$, the ray optics regime appears computationally the most
suitable, and the optical fields are simply represented as a collection of $N$ light rays according
to the geometrical optics \cite{ashkin1992forces,jones2015optical}. Each ray has a $N$-th portion
of the total incident power $P_\mathrm{i}$, and a linear momentum per second $n_\mathrm{m}
P_{\mathrm{i},N}/c$ \cite{born1999principles}. According to the Snell's law, a single ray which
impinges on a surface with an certain incident angle is partly reflected and partly transmitted.
Therefore, the ray changes its direction and, consequently, its momentum causing a reaction force
on the center of mass of the particle \cite{callegari2015computational}.
In the intermediate regime, that is when the particle size is comparable with the light wavelength
($x \simeq 1$) or for non-spherical and non-homogeneous particles like cosmic dust, we need to
resort to full electromagnetic theory and the T-Matrix formalism has proved to be computationally
advantageous \cite{waterman1971symmetry,borghese1980addition,borghese2007scattering}. Any particle
can be modeled through a sphere, clusters or aggregates of spheres, spheres with spherical
(eccentric) inclusions, and multilayered spheres \cite{borghese2007scattering}. This formalism
consists in the expansion of the electromagnetic fields into a basis of vector spherical harmonics
applying the boundary conditions across the particle surface
\cite{mishchenko2002scattering,borghese2007scattering}. The incident, and the internal fields
inside the particle have to be regular at the origin while the scattered field is such as to
satisfy the radiation condition at infinity
\cite{borghese2007scattering,fucile1997optical,ishimaru1991wave}. The transition matrix,
$\mathbb{T}$, connects the amplitudes of the scattered fields to the amplitudes of the incident
ones upon a multipole expansion of the fields. Its order is, in principle, infinite and thus it
must be truncated to a multipole index value $L_M$ \cite{borghese2007scattering}. Such value is
chosen to ensure the required accuracy of the transition matrix elements. However, for a cluster of
$N$ spheres this implies the solution of a system of order $D_M = 2 N L_M (L_M + 2)$, which may
become too large \cite{borghese2007scattering}. Actually, the inversion of the scattering matrix is
responsible for most of the computation time required for the calculation that scales as $D_M^3$.
Thus, on account of the definition of $D_M$, the computation time scales as $L_M^6$ whereas the
storage requirements scale as $L_M^4$, so that it pays, in terms of both CPU time and storage
requirements, to keep $L_M$ as low as practicable. Thus, the choice of an appropriate value of
$L_M$, in order to satisfy computation time and storage requirements, is of crucial importance
\cite{iati2004optical,saija2003efficient}.
The diversity of dust particles in an astrophysical context implies a richness of models that we
need to build in order to calculate realistic optical forces for space tweezers applications. Here,
we consider several models of extraterrestrial dusts with shape and composition inspired by
interstellar particles, DUSTER samples \cite{corte2013meteoric,rietmeijer2016laboratory}, and Moon
or Mars dust analogs. We show these models in Fig.~\ref{figure1}. On the top row, the homogeneous
and stratified single/aggregated spheres emulate hypothetical interstellar dust grains whose
constituents are olivine and aliphatic carbon \cite{antonia2008stratified}. The olivine refractive
index is provided by Draine \& Li while the carbon one by Ashok et al.
\cite{draine1984optical,palik1998handbook}. In Figs.~\ref{figure1}a and \ref{figure1}b, the
constituents refractive indexes are mixed in such a way as to treat the particle homogeneously with
a single effective refractive index according to the Bruggeman criterion
\cite{bohren2008absorption}. On the other hand, in Figs.~\ref{figure1}c and \ref{figure1}d, an
olivine core is considered covered by a carbon layer \cite{antonia2008stratified}. The spheres
radius of Figs.~\ref{figure1}a and \ref{figure1}c is $r=0.15$ $\mu$m. On the other hand, the
clusters of Figs.~\ref{figure1}b and \ref{figure1}d are composed by $9$ spheres of different sizes
with the major semi-axis $r=0.16$ $\mu$m. In Fig.~\ref{figure1}e, we present a model according to a
Field Emission Scanning Electron Microscope (FESEM) image of a silica particle clustering arranged
in quenched melt spheres, shown in Fig.~\ref{figure1}f, and collected by DUSTER
\cite{rietmeijer2016laboratory}. The refractive index is provided by Malitson
\cite{malitson1965interspecimen}. The model is composed by 4 spheres of different radius with the
major semi-axis $r=0.23$ $\mu$m. Fig.~\ref{figure1}g represents the model of condensed Ca[O]
nanograins that are accreted onto a larger melted aggregate of tiny carbonate grains, shown in the
FESEM image of Fig.~\ref{figure1}h, and collected by DUSTER
\cite{corte2013meteoric,rietmeijer2016laboratory}. The larger sphere is calcite and the other
spheres are CaO. The cluster model is composed by 30 spheres with the calcite refractive index
provided by Ghosh while the calcium oxide one is provided by Liu \& Sieckmann
\cite{ghosh1999dispersion,liu1966refractive}. Moreover, its major semi-axis has $r=0.25$ $\mu$m. In
Fig.~\ref{figure1}i, a spherical model ($r=1.25$ $\mu$m) of the particle Fe, Mg-rich TP2, collected
by DUSTER, is shown \cite{rotundi2014two}. We consider an effective refractive index obtained
mixing iron (67 \%) and magnesium (33 \%), according to the Bruggeman criterion whose the two
refractive indexes are respectively provided by Johnson \& Christy and Hagemann
\cite{johnson1972optical,hagemann1974desy,bohren2008absorption}. In Fig.~\ref{figure1}j, a
spherical model of Martian hematite ($r = 1$ $\mu$m) \cite{oliva2019database}. In
Fig.~\ref{figure1}k, a spherical model of Lunar regolith ($r = 1$ $\mu$m) \cite{egan1973optical}.
Fig.~\ref{figure1}l shows the model of a microscale fassaite ellipsoidal, collected by DUSTER, and
shown in a FESEM image of Fig.~\ref{figure1}m \cite{rietmeijer2016laboratory}. We consider an
effective refractive index constructed mixing silica (65 \%), and CaO (35 \%) according to the
Bruggeman criterion \cite{bohren2008absorption}, the major semi-axis is 2 $\mu$m. All non-spherical
models are oriented in such a way that their major axis is aligned with the incident light
propagation direction.
\begin{table}
\centering
\resizebox{0.9\textwidth}{!}{%
\includegraphics{fig_table.pdf}
}
\caption{Summary table of radiation pressure and optical trapping properties. For each model particle we show: the radius, $r$, or the radius of the smallest sphere enclosing the cluster; the geometric cross-section, $\sigma_\mathrm{geom}$; the integrated optical force on the solar intensity spectrum along the parallel, $\bar{F}_\parallel$, and perpendicular direction, $\bar{F}_\perp$, with respect to the light propagation; the extinction cross-section $\sigma_\mathrm{ext}$ at the trapping wavelength $\lambda=0.83$ $\mu$m; the real, $n$, and the imaginary part, $k$, of the refractive index at the trapping wavelength $\lambda=0.83$ $\mu$m; and the optical trap stiffnesses along the optical axis, $\kappa_z$, and perpendicular to it, $\tilde{\kappa}$, in water and in air.}\label{figure4}
\end{table}
The calculation of the solar radiation pressure and optical trapping forces on model particles of
Figs.~\ref{figure1}a - \ref{figure1}k is carried out with the T-matrix formalism because their size
parameters fall within the range $x \approx [0.1 - 6]$. On the other hand, when the particle size
parameter is too high (e.g. $x\geq 10$) such as in the fassaite micro ellipsoid of
Fig.~\ref{figure1}l, the calculation is carried out in the ray optics approximation. For this case
we exploited the optimized computational MATLAB codes for dielectric particles provided by
Callegari et al. \cite{callegari2015computational}
(http://opticaltweezers.org/software/otgo-optical-tweezers-geometrical-optics/).
\section{Results on solar radiation pressure and optical trapping of dust particles}\label{sect:Results and discussion}
Solar radiation pressure calculations are important to understand its relevance in optical trapping
applications in space. In general external forces, such as solar radiation pressure, can have
detrimental effects on optical trapping of particles in space or the high atmosphere. Here we aim
to show that T-matrix methods can be used to give accurate estimates of these effects on individual
dust particles. The total radiation force, $\mathbf{\bar{F}}$, and force spectrum, $f_s (\lambda)$,
that the Sun exerts on particles are calculated as:
\begin{equation}\label{eq:solar_force_integration}
\bar{F}_s = \int_{\lambda} \mathrm{d}\lambda \ f_s (\lambda) \ ,
\end{equation}
\begin{equation}\label{eq:solar_force_distribution}
f_s (\lambda) = \frac{i_\odot(\lambda)}{I_\odot} \ F_{\mathrm{rad},s} (\lambda) \ ,
\end{equation}
\noindent where the index $s=(\parallel,\perp)$ specifies respectively the parallel and the
orthogonal component of $F_{\mathrm{rad},s} (\lambda)$ respect to the radiation incident direction,
$\hat{\mathbf{k}}$. The expressions of $F_{\mathrm{rad},s}(\lambda)$ are the wavelength dependent
optical forces calculated in the ray optics approximation \cite{jones2015optical} or in the
T-matrix approach \cite{borghese2007scattering} according to the different models under
investigation. The term $f_s (\lambda)$ specifies the spectral force distribution obtained scaling
the computation outputs $F_{\mathrm{rad},s} (\lambda)$ upon the solar radiation intensity
distribution $i_\odot(\lambda)$ normalized at the solar irradiance $I_\odot$, $i.e.$, $I_\odot =
\int_{\lambda} \mathrm{d}\lambda \ i_\odot(\lambda) = 1.34$ kW$/$m$^2$ \cite{standard2012g173}.
Therefore, the considered spectrum is chosen in the range $\lambda = [0.28 - 2]$ $\mu$m that is
when the light emission by the Sun has the maximum intensity.
In the Fig.~\ref{figure2}, we plot the solar radiation force distributions $f_s (\lambda)$
(Eq.~\ref{eq:solar_force_distribution}) for all the models presented in Fig.~\ref{figure1}. In
particular, we show the force, $f_\parallel(\lambda)$, parallel to the propagation direction as a
red line, and the perpendicular one, $f_\perp(\lambda)$, as a blue line. Figures.~\ref{figure2}a-d
concern the calculation on different models for the olivine-aliphatic carbon structures.
Figure~\ref{figure2}e is referred to the quenched melt silica, Fig.~\ref{figure2}f the
bunch-of-grape carbon Ca-rich, Fig.~\ref{figure2}g the 'TP2' sphere, and Fig.~\ref{figure2}h the
martian hematite sphere. Finally, Fig.~\ref{figure2}i shows the fassaite ellipsoid for which the
radiation force is calculated with ray optics. We note how the parallel component, $f_\parallel$,
is much larger than the perpendicular one, $f_\perp$. In fact, the parallel component is
proportional to the particle extinction cross section, while the transverse one is related to the
asymmetry parameters, $g_i$, related to the non-sphericity of the scatterer
\cite{mishchenko2001radiation,borghese2007scattering,saija2005transverse}. Thus, for cylindrically
symmetric particles the component $f_\perp$ is zero (Figs.~\ref{figure2}a, \ref{figure2}c,
\ref{figure2}g, \ref{figure2}h, \ref{figure2}i). Figures.~\ref{figure2}b, \ref{figure2}d, and
\ref{figure2}e show results for elongated nanoscale clusters that generally align with the incident
electric field direction \cite{borghese2008radiation} yielding a low value of $f_\perp$. In
Fig.~\ref{figure2}f we report the highly non-symmetrical model of the bunch-of-grape carbon Ca-rich
for which $f_\perp$ is quite strong and $\bar{F}_\perp \simeq 1.46$ pN, a value comparable to
$\bar{F}_\parallel \simeq 18.2$ pN (Tab.~\ref{figure4}). In summary, the parallel component of the
solar radiation pressure, $f_\parallel$, describes an optomechanical interaction of the solar
radiation pressure with the extraterrestrial dust models in the tens of piconewton range. On the
other hand, for non-spherical particles the perpendicular component of the solar radiation pressure
can drive more complex transverse or rotational dynamics \cite{klavcka2001motion}.
We now focus on the systematic characterization of optical trapping forces in optical tweezers,
$i.e$, a single Gaussian beam focused by a high-NA objective. In our calculations we fix the laser
wavelength at $0.83$ $\mu$m, that is a typical wavelength for optical tweezers experiments in the
near-infrared, and the power $P = 50$ mW. In Fig.~\ref{figure3}, we show the three Cartesian
components of the trapping force in the neighborhood of the optical tweezers paraxial nominal focus
placed at the origin of the coordinate system ($x=y=z=0$). The trapping position of the particle in
the axial $z$-direction does not typically coincide with the origin because of the 'pushing' effect
of the optical scattering force. To calculate the force on the particle at the equilibrium position
$C_\mathrm{eq} = (x_\mathrm{eq}, y_\mathrm{eq}, z_\mathrm{eq})$, the $z$ axial coordinate at which
the axial force vanishes must firstly be found. Hence, the force plots in the transverse directions
$(x, y)$ can then be calculated. In the left column of the Fig.~\ref{figure3}, we present the
extinction cross-sections $\sigma_\mathrm{ext}$ for the visible and near-infrared wavelength
spectrum. The extiction cross section, $\sigma_\mathrm{ext}$, takes into account the rate at which
the energy is removed from the electromagnetic wave through scattering and the absorption, allowing
us to understand how effectively trapping takes place \cite{jones2015optical}. The trapping arises
when the focused incoming field generates a restoring force proportional to the particle's
displacement from an equilibrium point, and that, for small displacements, behaves harmonically
\cite{Polimeno2018}. Therefore, trap stiffnesses are defined as:
\begin{equation}
\kappa_i \equiv \frac{\mathrm{d}F_i}{\mathrm{d} x_i}\bigg|_{x_{\mathrm{eq},i}} \ .
\end{equation}
\noindent We calculate optical trapping in water ($n_{\rm m}=1.33$) with an objective NA $= 1.3$
(middle column of Fig.~\ref{figure3}) and in air or vacuum ($n_{\rm m}=1$) with NA $= 0.9$ (right
column of Fig.~\ref{figure3}). We note that here we focus only on electromagnetic calculations,
neglecting thermal fluctuations and hydrodynamics effects due to the surrounding medium
\cite{volpe2013simulation}. Generally, for a given particle, optical trapping in water is
stabilized by the overdamped viscous dynamics in the fluid, while in air or vacuum the underdamped
dynamics might be more critical for stable optical trapping \cite{svak2018transverse}. Moreover,
the higher NA in water causes the equilibrium point $C_\mathrm{eq}$ to be closer to the nominal
focus than in air. This is confirmed by comparing the graphs in Fig.~\ref{figure3}, central column
(in water), with those of the right column (in air). The dielectric particles, such as the quenched
melt silica (Figs.~\ref{figure3}h, \ref{figure3}i), and carbonate cluster (Figs.~\ref{figure3}k,
\ref{figure3}l) exhibit fairly stable trapping. Even the interstellar dust model of a sphere
aggregate can be trapped thanks to the relatively low extinction (Figs.~\ref{figure3}e,
\ref{figure3}f). Indeed, for these model particles we are able to extrapolate the trapping
constants along the axial direction, $\kappa_z$, and along the perpendicular direction,
$\tilde{\kappa} \equiv (\kappa_x + \kappa_y)/2$ (Tab.~\ref{figure4}). On the other hand, model
particles, like the interstellar dust sphere (Figs.~\ref{figure3}b, \ref{figure3}c), or the Fe-Mg
sphere (Figs.~\ref{figure3}n, \ref{figure3}o), can not be trapped neither in water nor in air. A
similar behavior is also exhibited by the hematite sphere, regolith sphere, and the ellipsoidal
fassaite as shown in Tab.~\ref{figure4} by observing the corresponding trapping constants. The
reason is to be found in their large size (Fe-Mg sphere, hematite, regolith, and fassaite) and on
their strong absorption (interstellar dust, Fe-Mg sphere, hematite, regolith and fassaite). To
further confirm, we have calculated the trapping properties of the Fe-Mg, hematite spheres by
reducing their size such that $r_\mathrm{Fe-Mg} = 0.05$ $\mu$m and $r_\mathrm{hem} = 0.1$ $\mu$m.
In this way, the trap stiffnesses can be extrapolated (Tab.~\ref{figure4}) and a quantitative
indication of $\sigma_\mathrm{ext}$ is provided in Tab.~\ref{figure4}. This behaviour is similar to
the optical trapping of metal nanoparticles that can be efficiently trapped at small size where
absorption and extinction cross sections are still small in the near-infrared
\cite{amendola2017surface}.
\section{Conclusions}
After a review on the role of interstellar, interplanetary, and planetary dust in the universe, we
studied computationally the solar radiation force and optical trapping properties for different
cosmic dust particles. We showed that the radiation force exerted by the Sun is not trivially
negligible: it can actively influence the dynamics of the modeled dust particles at the nano- and
microscale when compared to, \textit{e.g.}, optical trapping forces. Furthermore, we calculated
single-beam optical trapping properties for cosmic dust modeled particles both in water and in air
with parameters close to typical optical tweezers experiments in the near-infrared. We found that
dielectric and weakly absorbing particles can be captured, while microparticles with stronger
absorption show a scattering component of the optical force that prevents optical trapping in a
standard single-beam optical tweezers. This detrimental effects can be overcome through dual-beam
traps made by stationary counter-propagating laser beams
\cite{zemanek2002simplified,donato2018optical}.
This study opens perspectives for the application of optical tweezers techniques to solar system
study, $e.g.$, cometary particles analyses including the volatile component, dust particles in the
Martian atmosphere and/or on the Martian and Lunar surfaces. Such applications will also be
strategic within curator facilities designed for the uncontaminated handling and preliminary
characterization of extraterrestrial samples returned by space probes. The opportunity to apply
optical tweezers to planetary particulate matter in the frame of curation facilities can pave the
way for space applications for \textit{in situ} physical chemical analyses. The accurate knowledge
of the (optical, microchemical and electrical) properties of extraterrestrial dust allows the
design of an appropriate optical tweezers system for space applications, \textit{e.g.}, to be
mounted on stratospheric platforms, onboard future cometary probes, or landers/rovers to Mars or to
the Moon.
\section*{Acknowledgements}
We acknowledge financial contribution from the agreement ASI-INAF n.2018-16-HH.0, project ``SPACE
Tweezers''. M.A.I., D.B.C., M.G.D., A.F., P.G.G., and O.M.M. acknowledge support from the MSCA ITN
(ETN) project ``Active Matter'', project number 812780.
\bibliographystyle{unsrt}
|
\section{Introduction}
\textit{This paper is a sequel to the second author's paper \cite{Saw6}, and
continues to be dedicated to the memory of Professor Elias M. Stein.}
We prove the following $T1$ and cancellation theorem for general
Calder\'{o}n-Zygmund operators and doubling measures in $\mathbb{R}^{n}$. The
case of Lebesgue measures was obtained for testing conditions by David and
Journ\'{e} \cite{DaJo} (with less than smooth kernels), and for cancellation
conditions by Stein \cite[Theorem 4, page 306]{Ste2}. A weaker version of this
theorem was first proved in \cite{Saw6}, but with restrictive extra conditions
and assumptions. Note that the theorem below is a complete characterization,
for all smooth Calder\'{o}n-Zygmund kernels, including iterated Riesz
transforms, of the two weight norm inequality in terms of classical $A_{2}$
and testing conditions, albeit restricted to doubling measures. All previous
results to date have considered only specific operators such as Hilbert,
Beurling, Cauchy and (noniterated) Riesz transforms, or have assumed strong
energy conditions on the pairs of weights.
\begin{theorem}
\label{main}Suppose $0\leq\alpha<n$, and let $T^{\alpha}$ be an $\alpha
$-fractional Calder\'{o}n-Zygmund singular integral operator on $\mathbb{R
^{n}$ with a smooth $\alpha$-fractional kernel $K^{\alpha}$. Assume that
$\sigma$ and $\omega$ are doubling measures on $\mathbb{R}^{n}$. Set
$T_{\sigma}^{\alpha}f=T^{\alpha}\left( f\sigma\right) $ for any smooth
truncation of $T^{\alpha}$.\newline
\begin{enumerate}
\item Then the best constant $\mathfrak{N}_{T^{\alpha}}$ in the weighted norm
inequalit
\[
\left\Vert T_{\sigma}^{\alpha}f\right\Vert _{L^{2}\left( \omega\right) }\leq
C\left\Vert f\right\Vert _{L^{2}\left( \sigma\right) }\ ,
\]
satisfie
\begin{equation}
\mathfrak{N}_{T^{\alpha}}\leq C_{\alpha,n}\left( \sqrt{A_{2}^{\alpha
}+\mathfrak{T}_{T^{\alpha}}+\mathfrak{T}_{\left( T^{\alpha}\right) ^{\ast
}\right) , \label{bound
\end{equation}
where the constant $C_{\alpha,n}$ depends on the Calder\'{o}n-Zygmund constant
$C_{CZ}$ of the kernel, $A_{2}^{\alpha}$ is the classical Muckenhoupt
constant, and $\mathfrak{T}_{T^{\alpha}},\mathfrak{T}_{\left( T^{\alpha
}\right) ^{\ast}}$ are the dual testing constants.
\item If the kernel $K^{\alpha}$ is elliptic in the sense of Stein, see
(\ref{steinelliptic}) for the definition, then the inequality can be reversed
\[
\mathfrak{N}_{T^{\alpha}}\approx\sqrt{A_{2}^{\alpha}}+\mathfrak{T}_{T^{\alpha
}}+\mathfrak{T}_{\left( T^{\alpha}\right) ^{\ast}}.
\]
\item In the presence of finiteness of the classical Muckenhoupt constant
$\sqrt{A_{2}^{\alpha}}$, then the following four conditions are
equivalent\footnote{Conditions (a) and (c) represent two of the classical ways
to make sense of boundedness of singular integrals on weighted spaces.}:
\begin{enumerate}
\item The operator $T_{\sigma}^{\alpha}$ is bounded from $L^{2}\left(
\sigma\right) $ to $L^{2}\left( \omega\right) ,$ in the sense that
admissible truncations are uniformly bounded.
\item The testing constants $\mathfrak{T}_{T^{\alpha}}$ and $\mathfrak{T
_{\left( T^{\alpha}\right) ^{\ast}}$ are finite.
\item There exists a bounded operator $T^{\alpha}:L^{2}\left( \sigma\right)
\rightarrow L^{2}\left( \omega\right) $, that is associated with the kernel
$K^{\alpha}$ in the sense that whenever $f\in\mathcal{S}$ has compact support,
the tempered distribution $T^{\alpha}f$ can be identified, in the complement
of the support, with the function obtained by integration with respect to the
kernel, i.e.\footnote{In \cite[Remark 12 on page 12]{Saw6}, it was assumed, in
order to include condition (c) in the equivalence when $\alpha=0$, that there
was an operator $T^{0}$ associated with $K^{0}$ that is bounded on unweighted
$L^{2}\left( \mathbb{R}^{n}\right) $. This is no longer needed here due to
our use of the NTV reach for Alpert wavelets.
\begin{equation}
T^{\alpha}f\left( x\right) \equiv\int K^{\alpha}\left( x,y\right) f\left(
y\right) d\sigma\left( y\right) ,\ \ \ \ \ \text{for }x\in\mathbb{R
^{n}\setminus\operatorname{Supp}f. \label{identify
\end{equation}
\item The cancellation constants $\mathfrak{A}_{K^{\alpha}}$ and
$\mathfrak{A}_{K^{\alpha,\ast}}$ are finite.
\end{enumerate}
\end{enumerate}
\end{theorem}
We remark that the Beurling, Cauchy,\ and Riesz transform kernels, as well as
the kernels of each iterated Riesz transform, are all elliptic in the sense of Stein.
As mentioned earlier, Theorem \ref{main} is a two weight extension of the
David-Journ\'{e} and Stein theorems to pairs of \emph{doubling} measures
$\left( \sigma,\omega\right) $, and \textbf{all} smooth Calder\'{o}n-Zygmund
operators. Up until now, such characterizations for doubling measures have
been restricted to Calder\'{o}n-Zygmund operators with very special
properties, such as Hilbert and Riesz transforms, and have excluded in
particular imaginary powers of the Laplacian (arising e.g. in unique
continuation \cite{JeKe}), double layer potentials (arising e.g. in the
solution to the Dirichlet problem \cite{Ver}), and iterated Riesz transforms
(arising e.g. in Helmholtz projections and in recovering second order partial
derivatives from the Laplacian \cite{Ste}).
Since the only restrictions imposed here to obtain such a $T1$
characterization of boundedness are that the operator be smooth and that the
measures be doubling, this gives support to the conjecture that the two weight
$T1$ theorem holds for general measure pairs and general Calderon-Zygmund
operators. However, it seems that the techniques in this paper have been
pushed to their limit, so probably new ideas are required, even in order to
relax the smoothness assumption on the kernel. The reader should note that for
arbitrary pairs of measures, the $T1$ theorem has only been attained to date
for the Hilbert transform, \cite{LaSaShUr3}\&\cite{Lac} and \cite{Hyt2}, after
groundbreaking work in \cite{NTV}-\cite{NTV4}. For doubling measures the case
of Riesz transforms was proved earlier in \cite{LaWi}, see also
\cite{SaShUr10}.
At the same time, from the viewpoint of extending the David-Journ\'{e} and
Stein theorems to characterize $L^{2}$ boundedness in terms of a
\emph{cancellation} condition, Theorem \ref{main} may turn out to be optimal,
inasmuch as the proof of Stein's cancellation conditions being equivalent to
the testing conditions seems to require that the measures are doubling.
Here the classical fractional Muckenhoupt constant $A_{2}^{\alpha
=A_{2}^{\alpha}\left( \sigma,\omega\right) $ and the testing constant
$\mathfrak{T}_{T^{\alpha}}=\mathfrak{T}_{T^{\alpha}}\left( \sigma
,\omega\right) $ are defined b
\begin{equation}
A_{2}^{\alpha}\left( \sigma,\omega\right) \equiv\sup_{Q\in\mathcal{P}^{n
}\frac{\left\vert Q\right\vert _{\sigma}}{\left\vert Q\right\vert
^{1-\frac{\alpha}{n}}}\frac{\left\vert Q\right\vert _{\omega}}{\left\vert
Q\right\vert ^{1-\frac{\alpha}{n}}};\ \ \ \ \ \left( \mathfrak{T}_{T^{\alpha
}}\left( \sigma,\omega\right) \right) ^{2}\equiv\sup_{Q\in\mathcal{P}^{n
}\frac{1}{\left\vert Q\right\vert _{\sigma}}\int_{Q}\left\vert T_{\sigma
}^{\alpha}\mathbf{1}_{Q}\right\vert ^{2}\omega, \label{Muck and test
\end{equation}
and the cancellation constant $\mathfrak{A}_{K^{\alpha}}=\mathfrak{A
_{K^{\alpha}}\left( \sigma,\omega\right) $ is the best constant $C$ in the
inequalit
\begin{align}
& \int_{\left\Vert x-x_{0}\right\Vert <N}\left\vert \int_{\varepsilon
<\left\Vert x-y\right\Vert <N}K^{\alpha}\left( x,y\right) d\sigma\left(
y\right) \right\vert ^{2}d\omega\left( x\right) \leq C\left( \sigma
,\omega\right) \ \int_{\left\Vert x_{0}-y\right\Vert <N}d\sigma\left(
y\right) ,\label{can cond}\\
& \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \text{for all
}0<\varepsilon<N\text{ and }x_{0}\in\mathbb{R}^{n},\nonumber
\end{align}
where $\left\Vert y\right\Vert \equiv\max_{1\leq k\leq n}\left\vert
y_{k}\right\vert $ (which we later replace by the Euclidean norm $\left\vert
y\right\vert $: it follows from the $T_{\operatorname*{translate}}$ theorem in
Section \ref{T translate section} that we can use ball-testing instead of
cube-testing, and circular annuli in (\ref{can cond}) as originally done in
\cite[Theorem 4 page 306]{Ste2}). Similar suprema and inequalities define the
dual constants $\mathfrak{T}_{T^{\alpha,\ast}}$ and $\mathfrak{A
_{K^{\alpha,\ast}}$, in which the measures $\sigma$ and $\omega$ are
interchanged and $K^{\alpha}\left( x,y\right) $ is replaced by
$K^{\alpha,\ast}\left( x,y\right) =K^{\alpha}\left( y,x\right) $. It
should also be noted that (\ref{can cond}) is not simply the testing condition
for a truncation of $T$ over subsets of a ball, but instead has the historical
form of bounding in an average sense, integrals of the kernel over annuli,
which are here cubical rather than spherical.
\subsection{Application}
An application of this theorem will appear in an upcoming paper, where the
authors show that for pairs of $A_{\infty}$ weights, the two weight norm
inequality for individual iterated Riesz transforms is stable under rotations
of the weights, but for pairs of doubling measures, it is \emph{unstable}. In
particular, it is no longer true that boundedness of one iterated Riesz
transform implies boundedness of any other even when the measures are
doubling, thus showing the same instability for controlling second order
partial derivatives by the Laplacian. Finally, this application of our theorem
provides a connection between operator theory and the old problem of
distinguishing $A_{\infty}$ from doubling that was solved in 1974 by Fefferman
and Muckenhoupt \cite{FeMu}.
\subsection{A brief history of cancellation and testing}
Given a Calder\'{o}n-Zygmund kernel $K\left( x,y\right) $ in Euclidean space
$\mathbb{R}^{n}$, a classical problem for many decades was to identify optimal
cancellation conditions on $K$ so that there would exist an associated
singular integral operator $Tf\left( x\right) \sim\int K\left( x,y\right)
f\left( y\right) dy$ bounded on $L^{2}\left( \mathbb{R}^{n}\right) $. The
first such cancellation condition was that $K$ be an odd smooth homogeneous
convolution kernel, and an early more general example (\cite[see Theorem 2 on
page 35]{Ste}) was that a smooth convolution kernel $K\left( x\right) $
satisfy derivative bounds $\left\vert \nabla^{\ell}K\left( x\right)
\right\vert \leq C_{n,\ell}\left\vert x\right\vert ^{-n-\ell}$, and the
cancellation conditio
\[
\int_{\varepsilon<\left\vert x\right\vert <N}K\left( x\right)
dx=0,\ \ \ \ \ \text{for }0<\varepsilon<N<\infty.
\]
This cancellation condition was later generalized by many authors to
non-convolution kernels in various versions of the form
\[
\sup_{x\in\mathbb{R}^{n}}\left\vert \int_{\varepsilon<\left\vert
x-y\right\vert <N}K\left( x,y\right) dy\right\vert \leq
C,\ \ \ \ \ \text{for }0<\varepsilon<N<\infty.
\]
See e.g. \cite[page 53]{Ste} for references to the earlier work in this direction.
These efforts culminated in the decisive $T1$ theorem of David and Journ\'{e}
\cite{DaJo}, in which boundedness of an operator $T$ on $L^{2}\left(
\mathbb{R}^{n}\right) $ associated to $K$ was characterized by two testing
conditions taken over indicators of cubes
\[
\int_{Q}\left\vert T\mathbf{1}_{Q}\left( x\right) \right\vert ^{2
dx\lesssim\left\vert Q\right\vert \text{ and }\int_{Q}\left\vert T^{\ast
}\mathbf{1}_{Q}\left( x\right) \right\vert ^{2}dx\lesssim\left\vert
Q\right\vert ,\ \ \ \ \ \text{all cubes }Q\subset\mathbb{R}^{n}.
\]
This formulation is easily seen to be equivalent to the original one that used
$T1,T^{\ast}1\in BMO$ and the weak boundedness property. The corresponding
optimal cancellation conditions, involving $L^{2}$ norms of integrals of the
kernel over shells, were given in \cite[Theorem 4, page 306]{Ste2}
\begin{align*}
& \frac{1}{\int_{\left\vert x_{0}-y\right\vert <N}dy}\int_{\left\vert
x-x_{0}\right\vert <N}\left\vert \int_{\varepsilon<\left\vert x-y\right\vert
<N}K\left( x,y\right) dy\right\vert ^{2}dx\leq\mathfrak{A}_{K}\ ,\\
& \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \text{for all
}0<\varepsilon<N\text{ and }x_{0}\in\mathbb{R}^{n},
\end{align*}
and its dual. In the words of Stein this cancellation characterization was `a
rather direct consequence of' the $T1$ theorem.
An obvious next step is to replace Lebsegue measure with a fixed $A_{2}$
weight $w$
\[
\left( \frac{1}{\left\vert Q\right\vert }\int_{Q}w\left( x\right)
dx\right) \left( \frac{1}{\left\vert Q\right\vert }\int_{Q}\frac{1}{w\left(
x\right) }dx\right) \lesssim1,\ \ \ \ \ \text{all cubes }Q\subset
\mathbb{R}^{n}\ ,
\]
and ask when $T$ is bounded on $L^{2}\left( w\right) $. For elliptic
Calder\'{o}n-Zygmund operators $T$, this question is easily reduced to the
David-Journ\'{e} theorem using the Stein-Weiss interpolation with change of
measures theorem \cite{StWe}, and the Coifman-Fefferman extension \cite{CoFe}
of the one weight Hilbert transform inequality of Hunt, Muckenhoupt and
Wheeden \cite{HuMuWh} to a general class of Calder\'{o}n-Zygmund
operators\footnote{Indeed, if $T$ is bounded on $L^{2}\left( w\right) $,
then by duality it is also bounded on $L^{2}\left( \frac{1}{w}\right) $, and
the Stein-Weiss interpolation theorem with change of measure \cite{StWe} shows
that $T$ is bounded on unweighted $L^{2}\left( \mathbb{R}^{n}\right) $.
Conversely, if $T$ is bounded on unweighted $L^{2}\left( \mathbb{R
^{n}\right) $, the proof in \cite{CoFe} shows that $T$ is bounded on
$L^{2}\left( w\right) $ using $w\in A_{2}$.}.
We follow the general scheme of proof initiated by Nazarov, Treil and Volberg
for the two weight Hilbert transform inequality, and developed further in work
of Lacey, Sawyer, Shen and Uriarte-Tuero, but with several changes exploiting
the use of doubling conditions in the absence of energy conditions. For
example, in the first paper in this series \cite{Saw6}, it was observed that
doubling implies a weaker form of the NTV\ pivotal condition involving `tails'
$\frac{1}{\left( 1+\left\vert y\right\vert \right) ^{n+\kappa-\alpha}}$ with
a large $\kappa$ depending on the doubling constant, whose use requires both
$\kappa^{th}$-order Taylor expansions of the kernel, and weighted Alpert
wavelets with $\kappa$ vanishing moments, in place of weighted Haar wavelets.
The subsequent use of quasiorthogonality methods in Calder\'{o}n-Zygmund
coronas required supremum estimates for weighted Alpert wavelets, that turned
out to hold \emph{if and only if} the measures were doubling. It was also
observed there that the kernels of the operators need only satisfy
Calder\'{o}n-Zygmund smoothness conditions up to order $\kappa$, but we will
not track this dependence in this paper as the parameter $\kappa$ goes to
infinity as the doubling constant does. The reader should note that we use
neither energy conditions, which are essentially only necessary for
boundedness of the Hilbert transform and its perturbations on the real line
(see \cite{SaShUr11} and \cite{Saw}), nor the $k$\emph{-energy dispersed
}conditions introduced in \cite{SaShUr9}, which essentially only hold for
perturbations of Riesz transforms in higher dimensions.
But most importantly, the NTV paraproduct - neighbour - stopping decomposition
developed in this paper now requires bounding a new commutator term arising
from the fact that the polynomials in the Alpert wavelets do not commute with
singular integrals, and this innovation turns out to be considerably
challenging, involving new quasiorthogonality arguments, organization of
terms, and covering arguments. The paraproduct, neighbour and stopping forms
also require new arguments due to the fact that products of Alpert wavelets
and polynomials are no longer orthogonal to each other. We also use doubling
of the measures to reduce our bounds to just the classical Muckenhoupt
condition and the usual cube testing conditions; either the one-tailed
Muckenhoupt conditions or the full testing conditions are typically used
elsewhere in the literature.
The problem, beyond doubling measures, of bounding a general
Calder\'{o}n-Zygmund operator $T$ from one $L^{2}\left( \sigma\right) $
space to another $L^{2}\left( \omega\right) $ space, is wide open in
general. For the Hilbert transform $H$, the two weight inequality was
completely solved in the two part paper \cite{LaSaShUr3};\cite{Lac}, where it
was shown that $H$ is bounded from $L^{2}\left( \sigma\right) $ to
$L^{2}\left( \omega\right) $ if and only if the testing and one-tailed
Muckenhoupt conditions hold, i.e
\begin{align*}
& \int_{I}\left\vert H\left( \mathbf{1}_{I}\sigma\right) \right\vert
^{2}d\omega\lesssim\int_{I}d\sigma\text{ and }\int_{I}\left\vert H\left(
\mathbf{1}_{I}\omega\right) \right\vert ^{2}d\sigma\lesssim\int_{I
d\omega,\ \ \ \ \ \text{all intervals }I\subset\mathbb{R}^{n},\\
& \left( \int_{\mathbb{R}}\frac{\left\vert I\right\vert }{\left\vert
I\right\vert ^{2}+\left\vert x-c_{I}\right\vert ^{2}}d\sigma\left( x\right)
\right) \left( \frac{1}{\left\vert I\right\vert }\int_{I}d\omega\right)
\lesssim1,\ \text{and its dual},\ \ \ \ \ \text{all intervals }I\subset
\mathbb{R}^{n}.
\end{align*}
This result was itself the culmination of a long history beginning with
groundbreaking works by Hunt, Muckenhoupt and Wheeden \cite{HuMuWh}, Coifman,
Jones and Semmes \cite{CoJoSe}, and by Nazarov, Treil and Volberg (see
\cite{NTV} and \cite{NTV4} and references given there), and the final form was
achieved by Hyt\"{o}nen \cite{Hyt2}.
For fractional Riesz transforms in higher dimensions, it is known that the two
weight norm inequality \emph{with doubling measures} is equivalent to the
fractional one-tailed Muckenhoupt and $T1$ cube testing conditions, see
\cite[Theorem 1.4]{LaWi} and \cite[Theorem 2.11]{SaShUr9}. However, these
results rely on certain `positivity' properties of the gradient of the kernel
(which for the Hilbert transform kernel $\frac{1}{y-x} $ is simply $\frac
{d}{dx}\frac{1}{y-x}>0$ for $x\neq y$), something that is not available even
for strongly elliptic fractional Calder\'{o}n-Zygmund operators.
Finally, we mention\footnote{See e.g. the references at the end of the paper
involving these authors, and those already mentioned elsewhere in the paper,
M. A. Alfonseca, K. Astala, P. Auscher, A. Axelsson, M. Christ, D. Cruz-Uribe,
R. Fefferman, C. Grigoriadis, P. Guan, S. Hofmann, C. Kenig, S. Kim, A.
Lerner, J. M. Martell, H. Martikainen, P. Mattila, A. McIntosh, M. Melnikov,
C. Muscalu, N. Nikolski, S. Ombrosi, L. P\"{a}iv\"{a}rinta, M. Paparizos, F.
Peherstorfer, C. Perez, J. Pipher, M. Prats, E. Saksman, T. Tao, P.
Tchamitchian, C. Thiele, X. Tolsa, T. Toro, J. Verdera, and P. Yuditskii.} the
existence of a large body of related results on one weight function theory,
sufficient bump conditions, local $Tb$ theorems, applications to partial
differential equations, the Painlev\'{e} problem, the Kato problem, and many
others too numerous to list here.
In Section 2 we recall preliminary definitions and results we need from the
literature. In Section 3 we construct the corona decomposition adapted to the
smaller pivotal conditions, and in Section 4 we review the various reductions
of the proof by side length and the shifted corona, appealing to known results
when available. Section 5 contains the main new contribution of this paper,
namely the extension of the `NTV reach' to weighted Alpert wavelets, then in
Section 6, the estimates scattered throughout the paper are collected to
finish the proof. We conclude with Section \ref{T translate section} where we
introduce a \textquotedblleft$T_{\operatorname*{translate}}$
theorem\textquotedblright\ which in particular shows that ball-testing and
cube-testing are equivalent for doubling weights and any elliptic
Calder\'{o}n-Zygmund operator.
\section{Preliminaries}
Denote by $\mathcal{P}^{n}$ the collection of cubes in $\mathbb{R}^{n}$ having
sides parallel to the coordinate axes; all cubes mentioned in this paper will
be elements of $\mathcal{P}^{n}$. A positive locally finite Borel measure
$\mu$ on $\mathbb{R}^{n}$ is said to satisfy the\emph{\ doubling condition} if
$\left\vert 2Q\right\vert _{\mu}\leq C_{\operatorname*{doub}}\left\vert
Q\right\vert _{\mu}$ for all cubes $Q\in\mathcal{P}^{n}$. It is well known
(see e.g. the introduction in \cite{SaUr}) that doubling implies reverse
doubling, which means that there exists a positive constant $\theta_{\mu
}^{\operatorname*{rev}}$, called a \emph{reverse doubling exponent}, such tha
\[
\sup_{Q\in\mathcal{P}^{n}}\frac{\left\vert sQ\right\vert _{\mu}}{\left\vert
Q\right\vert _{\mu}}\leq s^{\theta_{\mu}^{\operatorname*{rev}}
,\ \ \ \ \ \text{for all sufficiently small }s>0.
\]
\subsection{Standard fractional singular integrals and the norm inequality}
Let $0\leq\alpha<n$ and $\kappa\in\mathbb{N}$. We define a standard $\left(
\kappa+\delta\right) $-smooth $\alpha$-fractional Calder\'{o}n-Zygmund kernel
$K^{\alpha}(x,y)$ to be a function $K^{\alpha}:\mathbb{R}^{n}\times
\mathbb{R}^{n}\rightarrow\mathbb{R}$ satisfying the following fractional size
and smoothness conditions: for $x\neq y$, and with $\nabla_{1}$ and
$\nabla_{2}$ denoting gradient in the first and second variables
respectively,
\begin{align}
& \left\vert \nabla_{1}^{j}K^{\alpha}\left( x,y\right) \right\vert \leq
C_{CZ}\left\vert x-y\right\vert ^{\alpha-j-n},\ \ \ \ \ 0\leq j\leq
\kappa,\label{sizeandsmoothness'}\\
& \left\vert \nabla_{1}^{\kappa}K^{\alpha}\left( x,y\right) -\nabla
_{1}^{\kappa}K^{\alpha}\left( x^{\prime},y\right) \right\vert \leq
C_{CZ}\left( \frac{\left\vert x-x^{\prime}\right\vert }{\left\vert
x-y\right\vert }\right) ^{\delta}\left\vert x-y\right\vert ^{\alpha-\kappa
-n},\ \ \ \ \ \frac{\left\vert x-x^{\prime}\right\vert }{\left\vert
x-y\right\vert }\leq\frac{1}{2},\nonumber
\end{align}
and where the same inequalities hold for the adjoint kernel $K^{\alpha,\ast
}\left( x,y\right) \equiv K^{\alpha}\left( y,x\right) $, in which $x$ and
$y$ are interchanged, and where $\nabla_{1}$ is replaced by $\nabla_{2}$.
\subsubsection{Ellipticity of kernels}
Following \cite[(39) on page 210]{Ste}, we say that an $\alpha$-fractional
Calder\'{o}n-Zygmund kernel $K^{\alpha}$ is \emph{elliptic in the sense of
Stein} if there is a unit vector $\mathbf{u}_{0}\in\mathbb{R}^{n}$ and a
constant $c>0$ such tha
\begin{align}
\left\vert K^{\alpha}\left( x,x+t\mathbf{u}_{0}\right) \right\vert \geq
c\left\vert t\right\vert ^{\alpha-n},\ \ \ \ \ \text{for all }t\in\mathbb{R}.
\label{steinelliptic
\end{align}
\subsubsection{Defining the norm inequality\label{Subsubsection norm}}
We follow the approach in \cite[see page 314]{SaShUr9}. So we suppose that
$K^{\alpha}$ is a standard $\left( \kappa+\delta\right) $-smooth $\alpha
$-fractional Calder\'{o}n-Zygmund kernel, and we introduce a family $\left\{
\eta_{\delta,R}^{\alpha}\right\} _{0<\delta<R<\infty}$ of nonnegative
functions on $\left[ 0,\infty\right) $ so that the truncated kernels
$K_{\delta,R}^{\alpha}\left( x,y\right) =\eta_{\delta,R}^{\alpha}\left(
\left\vert x-y\right\vert \right) K^{\alpha}\left( x,y\right) $ are bounded
with compact support for fixed $x$ or $y$, and uniformly satisfy
(\ref{sizeandsmoothness'}). Then the truncated operators
\[
T_{\sigma,\delta,R}^{\alpha}f\left( x\right) \equiv\int_{\mathbb{R}^{n
}K_{\delta,R}^{\alpha}\left( x,y\right) f\left( y\right) d\sigma\left(
y\right) ,\ \ \ \ \ x\in\mathbb{R}^{n},
\]
are pointwise well-defined when $f$ is bounded with compact support, and we
will refer to the pair $\left( K^{\alpha},\left\{ \eta_{\delta,R}^{\alpha
}\right\} _{0<\delta<R<\infty}\right) $ as an $\alpha$-fractional singular
integral operator, which we typically denote by $T^{\alpha}$, suppressing the
dependence on the truncations.
\begin{definition}
\label{def bounded}We say that an $\alpha$-fractional singular integral
operator $T^{\alpha}=\left( K^{\alpha},\left\{ \eta_{\delta,R}^{\alpha
}\right\} _{0<\delta<R<\infty}\right) $ satisfies the norm inequalit
\begin{equation}
\left\Vert T_{\sigma}^{\alpha}f\right\Vert _{L^{2}\left( \omega\right)
\leq\mathfrak{N}_{T^{\alpha}}\left\Vert f\right\Vert _{L^{2}\left(
\sigma\right) },\ \ \ \ \ f\in L^{2}\left( \sigma\right) .
\label{two weight'
\end{equation}
provide
\[
\left\Vert T_{\sigma,\delta,R}^{\alpha}f\right\Vert _{L^{2}\left(
\omega\right) }\leq\mathfrak{N}_{T^{\alpha}}\left( \sigma,\omega\right)
\left\Vert f\right\Vert _{L^{2}\left( \sigma\right) },\ \ \ \ \ f\in
L^{2}\left( \sigma\right) ,0<\delta<R<\infty.
\]
\end{definition}
\begin{description}
\item[Independence of Truncations] \label{independence}In the presence of the
classical Muckenhoupt condition $A_{2}^{\alpha}$, the norm inequality
(\ref{two weight'}) is independent of the choice of truncations used,
including \emph{nonsmooth} truncations as well - see \cite{LaSaShUr3}.
However, in dealing with the Energy Lemma \ref{ener} below, where $\kappa
^{th}$ order Taylor approximations are made on the truncated kernels, it is
necessary to use sufficiently smooth truncations. Similar comments apply to
the Cube Testing conditions.
\end{description}
\subsection{$\kappa$-cube testing and pivotal conditions}
In this subsection we describe a variety of testing and pivotal conditions
that arise in the course of our proof, but which do not appear in the
statement of our main Theorem \ref{main}, where only the classical testing
condition in (\ref{Muck and test}) is used.
The $\kappa$\emph{-cube testing conditions} associated with an $\alpha
$-fractional singular integral operator $T^{\alpha}$ are given b
\begin{align}
\left( \mathfrak{T}_{T^{\alpha}}^{\left( \kappa\right) }\left(
\sigma,\omega\right) \right) ^{2} & \equiv\sup_{Q\in\mathcal{P}^{n}
\max_{0\leq\left\vert \beta\right\vert <\kappa}\frac{1}{\left\vert
Q\right\vert _{\sigma}}\int_{Q}\left\vert T_{\sigma}^{\alpha}\left(
\mathbf{1}_{Q}m_{Q}^{\beta}\right) \right\vert ^{2}\omega<\infty
,\label{def Kappa polynomial'}\\
\left( \mathfrak{T}_{\left( T^{\alpha}\right) ^{\ast}}^{\left(
\kappa\right) }\left( \omega,\sigma\right) \right) ^{2} & \equiv
\sup_{Q\in\mathcal{P}^{n}}\max_{0\leq\left\vert \beta\right\vert <\kappa
\frac{1}{\left\vert Q\right\vert _{\omega}}\int_{Q}\left\vert \left(
T^{\alpha,\ast}\right) _{\omega}\left( \mathbf{1}_{Q}m_{Q}^{\beta}\right)
\right\vert ^{2}\sigma<\infty,\nonumber
\end{align}
where $\left( T^{\alpha,\ast}\right) _{\omega}=\left( T_{\sigma}^{\alpha
}\right) ^{\ast}$, with $m_{Q}^{\beta}\left( x\right) \equiv\left(
\frac{x-c_{Q}}{\ell\left( Q\right) }\right) ^{\beta}$ for any cube $Q$ and
multiindex $\beta$, where $c_{Q}$ is the center of the cube $Q$, and where we
interpret the right hand sides as holding uniformly over all sufficiently
smooth truncations of $T^{\alpha}$. Equivalently, in the presence of
$A_{2}^{\alpha}$, we can take a single suitable truncation, see Independence
of Truncations in Subsubsection \ref{independence}.
We also use the \emph{triple }$\kappa$\emph{-cube testing conditions} in which
the integrals are over the triple $3Q$ of $Q$
\begin{align}
\left( \mathfrak{TR}_{T^{\alpha}}^{\left( \kappa\right) }\left(
\sigma,\omega\right) \right) ^{2} & \equiv\sup_{Q\in\mathcal{P}^{n}
\max_{0\leq\left\vert \beta\right\vert <\kappa}\frac{1}{\left\vert
Q\right\vert _{\sigma}}\int_{3Q}\left\vert T_{\sigma}^{\alpha}\left(
\mathbf{1}_{Q}m_{Q}^{\beta}\right) \right\vert ^{2}\omega<\infty
,\label{full testing}\\
\left( \mathfrak{TR}_{\left( T^{\alpha}\right) ^{\ast}}^{\left(
\kappa\right) }\left( \omega,\sigma\right) \right) ^{2} & \equiv
\sup_{Q\in\mathcal{P}^{n}}\max_{0\leq\left\vert \beta\right\vert <\kappa
\frac{1}{\left\vert Q\right\vert _{\omega}}\int_{3Q}\left\vert \left(
T^{\alpha,\ast}\right) _{\omega}\left( \mathbf{1}_{Q}m_{Q}^{\beta}\right)
\right\vert ^{2}\sigma<\infty.\nonumber
\end{align}
The smaller fractional Poisson integrals $\mathrm{P}_{m}^{\alpha}\left(
Q,\mu\right) $ used here, in \cite{RaSaWi} and elsewhere, are given by
\begin{equation}
\mathrm{P}_{m}^{\alpha}\left( Q,\mu\right) =\int_{\mathbb{R}^{n}}\frac
{\ell\left( Q\right) ^{m}}{\left( \ell\left( Q\right) +\left\vert
y-c_{Q}\right\vert \right) ^{n+m-\alpha}}d\mu\left( y\right)
,\ \ \ \ \ m\geq1. \label{def kappa Poisson
\end{equation}
The $\kappa^{th}$-order fractional Pi\textbf{V}otal Conditions $\mathcal{V
_{2}^{\alpha,\kappa},\mathcal{V}_{2}^{\alpha,\kappa,\ast}<\infty$, $\kappa
\geq1$, are given b
\begin{align}
\left( \mathcal{V}_{2}^{\alpha,\kappa}\left( \sigma,\omega\right) \right)
^{2} & =\sup_{Q\supset\dot{\cup}Q_{r}}\frac{1}{\left\vert Q\right\vert
_{\sigma}}\sum_{r=1}^{\infty}\mathrm{P}_{\kappa}^{\alpha}\left(
Q_{r},\mathbf{1}_{Q}\sigma\right) ^{2}\left\vert Q_{r}\right\vert _{\omega
}\ ,\label{both pivotal k}\\
\left( \mathcal{V}_{2}^{\alpha,\kappa,\ast}\left( \sigma,\omega\right)
\right) ^{2} & =\sup_{Q\supset\dot{\cup}Q_{r}}\frac{1}{\left\vert
Q\right\vert _{\omega}}\sum_{r=1}^{\infty}\mathrm{P}_{\kappa}^{\alpha}\left(
Q_{r},\mathbf{1}_{Q}\omega\right) ^{2}\left\vert Q_{r}\right\vert _{\sigma
}\ ,\nonumber
\end{align}
where the supremum is taken over all subdecompositions of a cube
$Q\in\mathcal{P}^{n}$ into pairwise disjoint dyadic subcubes $Q_{r}$. The
following lemma from \cite[Subsection 4.1 on pages 12-13, especially Remark
15]{Saw6} was the point of departure for freeing the theory from reliance on
energy conditions when the measures are doubling.
\begin{lemma}
[\cite{Saw6}]\label{doub piv}If $\sigma$ is a doubling measure, then for
sufficiently large $\kappa$ depending on the doubling constant of $\sigma$, we
have
\[
\mathcal{V}_{2}^{\alpha,\kappa}\left( \sigma,\omega\right) \leq
CA_{2}^{\alpha}\left( \sigma,\omega\right) .
\]
\end{lemma}
\subsection{Weighted Alpert bases for $L^{2}\left( \mu\right) $ and
$L^{\infty}$ control of projections\label{Subsection Haar}}
We now recall the construction of weighted Alpert wavelets in \cite{RaSaWi},
and refer also to \cite{AlSaUr} for the correction of a small oversight in
\cite{RaSaWi}. Let $\mu$ be a locally finite positive Borel measure on
$\mathbb{R}^{n}$, and fix $\kappa\in\mathbb{N}$. For each cube $Q$, denote by
$L_{Q;\kappa}^{2}\left( \mu\right) $ the finite dimensional subspace of
$L^{2}\left( \mu\right) $ that consists of linear combinations of the
indicators of\ the children $\mathfrak{C}\left( Q\right) $ of $Q$ multiplied
by polynomials of degree less than $\kappa$, and such that the linear
combinations have vanishing $\mu$-moments on the cube $Q$ up to order
$\kappa-1$
\[
L_{Q;\kappa}^{2}\left( \mu\right) \equiv\left\{ f
{\displaystyle\sum\limits_{Q^{\prime}\in\mathfrak{C}\left( Q\right) }}
\mathbf{1}_{Q^{\prime}}p_{Q^{\prime};\kappa}\left( x\right) :\int
_{Q}f\left( x\right) x^{\beta}d\mu\left( x\right) =0,\ \ \ \text{for
}0\leq\left\vert \beta\right\vert <\kappa\right\} ,
\]
where $p_{Q^{\prime};\kappa}\left( x\right) =\sum_{\beta\in\mathbb{Z
_{+}^{n}:\left\vert \beta\right\vert \leq\kappa-1\ }a_{Q^{\prime};\beta
}x^{\beta}$ is a polynomial in $\mathbb{R}^{n}$ of degree less than $\kappa$.
Here $x^{\beta}=x_{1}^{\beta_{1}}x_{2}^{\beta_{2}}...x_{n}^{\beta_{n}}$. Let
$d_{Q;\kappa}\equiv\dim L_{Q;\kappa}^{2}\left( \mu\right) $ be the dimension
of the finite dimensional linear space $L_{Q;\kappa}^{2}\left( \mu\right) $.
Consider an arbitrary dyadic grid $\mathcal{D}$. For $Q\in\mathcal{D}$, let
$\bigtriangleup_{Q;\kappa}^{\mu}$ denote orthogonal projection onto the finite
dimensional subspace $L_{Q;\kappa}^{2}\left( \mu\right) $, and let
$\mathbb{E}_{Q;\kappa}^{\mu}$ denote orthogonal projection onto the finite
dimensional subspac
\[
\mathcal{P}_{Q;\kappa}^{n}\left( \mu\right) \equiv
\mathrm{\operatorname*{Span}}\{\mathbf{1}_{Q}x^{\beta}:0\leq\left\vert
\beta\right\vert <\kappa\}.
\]
For a doubling measure $\mu$, it is proved in \cite{RaSaWi}, that we have the
orthonormal decomposition
\begin{equation}
f=\sum_{Q\in\mathcal{D}}\bigtriangleup_{Q;\kappa}^{\mu}f,\ \ \ \ \ f\in
L_{\mathbb{R}^{n}}^{2}\left( \mu\right) ,\ \ \ \ \ \text{where }\left\langle
\bigtriangleup_{P;\kappa}^{\mu}f,\bigtriangleup_{Q;\kappa}^{\mu}f\right\rangle
=0\text{ for }P\neq Q, \label{Alpert expan
\end{equation}
where convergence holds both in $L_{\mathbb{R}^{n}}^{2}\left( \mu\right) $
norm and pointwise $\mu$-almost everywhere, the telescoping identitie
\begin{equation}
\mathbf{1}_{Q}\sum_{I:\ Q\subsetneqq I\subset P}\bigtriangleup_{I;\kappa
^{\mu}=\mathbb{E}_{Q;\kappa}^{\mu}-\mathbf{1}_{Q}\mathbb{E}_{P;\kappa}^{\mu
}\ \text{ \ for }P,Q\in\mathcal{D}\text{ with }Q\subsetneqq P,
\label{telescoping
\end{equation}
and the moment vanishing condition
\begin{equation}
\int_{\mathbb{R}^{n}}\bigtriangleup_{Q;\kappa}^{\mu}f\left( x\right)
\ x^{\beta}d\mu\left( x\right) =0,\ \ \ \text{for }Q\in\mathcal{D},\text{
}\beta\in\mathbb{Z}_{+}^{n},\ 0\leq\left\vert \beta\right\vert <\kappa\ .
\label{mom con
\end{equation}
We have the bound for the Alpert projections $\mathbb{E}_{I;\kappa}^{\mu}$
(\cite[see (4.7) on page 14]{Saw6}):
\begin{equation}
\left\Vert \mathbb{E}_{I;\kappa}^{\mu}f\right\Vert _{L_{I}^{\infty}\left(
\mu\right) }\lesssim E_{I}^{\mu}\left\vert f\right\vert \leq\sqrt{\frac
{1}{\left\vert I\right\vert _{\mu}}\int_{I}\left\vert f\right\vert ^{2}d\mu
},\ \ \ \ \ \text{for all }f\in L_{\operatorname*{loc}}^{2}\left( \mu\right)
. \label{analogue
\end{equation}
In terms of the Alpert coefficient vectors $\widehat{f}\left( I\right)
\equiv\left\{ \left\langle f,h_{I;\kappa}^{\mu,a}\right\rangle \right\}
_{a\in\Gamma_{I,n,\kappa}}$ for an orthonormal basis $\left\{ h_{I;\kappa
}^{\mu,a}\right\} _{a\in\Gamma_{I,n,\kappa}}$ of $L_{I;\kappa}^{2}\left(
\mu\right) $ where $\Gamma_{I,n,\kappa}$ is a convenient finite index set of
size $d_{Q;\kappa}$, we thus hav
\begin{equation}
\left\vert \widehat{f}\left( I\right) \right\vert =\left\Vert \bigtriangleup
_{I;\kappa}^{\sigma}f\right\Vert _{L^{2}\left( \sigma\right) }\leq\left\Vert
\bigtriangleup_{I;\kappa}^{\sigma}f\right\Vert _{L^{\infty}\left(
\sigma\right) }\sqrt{\left\vert I\right\vert _{\sigma}}\leq C\left\Vert
\bigtriangleup_{I;\kappa}^{\sigma}f\right\Vert _{L^{2}\left( \sigma\right)
}=C\left\vert \widehat{f}\left( I\right) \right\vert . \label{analogue'
\end{equation}
\subsection{The Pivotal Lemma}
For $0\leq\alpha<n$, let $\mathrm{P}^{\alpha}\left( J,\mu\right)
\equiv\mathrm{P}_{1}^{\alpha}\left( J,\mu\right) $ denote the standard
Poisson integral, where $\mathrm{P}_{m}^{\alpha}\left( J,\mu\right) $ is as
defined in (\ref{def kappa Poisson}). The following extension of the `energy
lemma' is due to Rahm, Sawyer and Wick \cite{RaSaWi} in the case the
polynomial $R\left( x\right) $ is constant, and this case is proved in
detail in \cite[Lemmas 28 and 29 on pages 27-30]{Saw6}.
\begin{lemma}
[\textbf{Pivotal Lemma}]\label{ener}Fix $\kappa\geq1$. Let $J\ $be a cube in
$\mathcal{D}$, and let $\Psi_{J}$ be an $L^{2}\left( \omega\right) $
function supported in $J$ with vanishing $\omega$-means of all orders less
than $\kappa$. Let $R\left( x\right) $ be a polynomial of degree less than
$\kappa$ that satisfies $\sup_{x\in J}\left\vert R\left( x\right)
\right\vert \leq1$. Let $\nu$ be a positive measure supported in
$\mathbb{R}^{n}\setminus\gamma J$ with $\gamma>1$, and let $T^{\alpha}$ be a
standard $\alpha$-fractional singular integral operator with $0\leq\alpha<n$.
Then we have the `pivotal' boun
\begin{equation}
\left\vert \left\langle RT^{\alpha}\left( \varphi\nu\right) ,\Psi
_{J}\right\rangle _{L^{2}\left( \omega\right) }\right\vert \lesssim
C_{\gamma}\mathrm{P}_{\kappa}^{\alpha}\left( J,\nu\right) \sqrt{\left\vert
J\right\vert _{\omega}}\left\Vert \Psi_{J}\right\Vert _{L^{2}\left(
\omega\right) }\ ,\label{piv bound
\end{equation}
for any function $\varphi$ with $\left\vert \varphi\right\vert \leq1$.
\end{lemma}
To obtain Lemma \ref{ener} in the case of a general polynomial $R\left(
x\right) $ of degree at most $\kappa$, replace the Taylor expansion of the
function $x\rightarrow K\left( x,y\right) $ that arises in the proof in
\cite[Lemmas 28 and 29 on pages 27-30]{Saw6}, by the Taylor expansion of the
function $x\rightarrow R\left( x\right) K\left( x,y\right) $, and note
that the estimates obtained there remain true with possibly larger constants.
We also recall from \cite[Lemma 33]{Saw6} the following Poisson estimate, that
is a straightforward extension of the case $m=1$ due to NTV in \cite{NTV4}.
This lemma is the key to exploiting the crucial reduction to
$\operatorname{good}$ cubes $J$ that we use below, see \cite{NTV4} and
\cite{NTV}.
\begin{lemma}
\label{Poisson inequality}Fix $m\geq1$. Suppose that $J\subset I\subset K$ and
that $\operatorname*{dist}\left( J,\partial I\right) >2\sqrt{n}\ell\left(
J\right) ^{\varepsilon}\ell\left( I\right) ^{1-\varepsilon}$. The
\begin{equation}
\mathrm{P}_{m}^{\alpha}(J,\sigma\mathbf{1}_{K\setminus I})\lesssim\left(
\frac{\ell\left( J\right) }{\ell\left( I\right) }\right) ^{m-\varepsilon
\left( n+m-\alpha\right) }\mathrm{P}_{m}^{\alpha}(I,\sigma\mathbf{1
_{K\setminus I}). \label{e.Jsimeq
\end{equation}
\end{lemma}
\section{The Calder\'{o}n-Zygmund and $\kappa$-pivotal corona decomposition}
To set the stage for control of the stopping form below in the absence of the
energy condition, we construct the \emph{Calder\'{o}n-Zygmund }$\kappa
$\emph{-pivotal} corona decomposition, in analogy with the energy version used
in the two part paper \{\cite{LaSaShUr3},\cite{Lac}\} and in \cite{SaShUr7},
for a function $f$ in $L^{2}\left( \sigma\right) $ that is supported in a
dyadic cube $F_{1}^{0}$. Fix $\Gamma>1$ and define $\mathcal{G}_{0}=\left\{
F_{1}^{0}\right\} $ to consist of the single cube $F_{1}^{0}$, and define the
first generation $\mathcal{G}_{1}=\left\{ F_{k}^{1}\right\} _{k}$ of
\emph{CZ }$\kappa$\emph{-pivotal stopping children} of $F_{1}^{0}$ to be the
\emph{maximal} dyadic subcubes $I$ of $F_{0}$ satisfyin
\[
\left\{
\begin{array}
[c]{c
\text{either }E_{I}^{\sigma}\left\vert f\right\vert \geq\Gamma E_{F_{1}^{0
}^{\sigma}\left\vert f\right\vert \\
\text{or }\mathrm{P}_{\kappa}^{\alpha}\left( I,\mathbf{1}_{F_{1}^{0}
\sigma\right) ^{2}\left\vert I\right\vert _{\omega}\geq\Gamma\left\vert
I\right\vert _{\sigma
\end{array}
\right. .
\]
Then define the second generation $\mathcal{G}_{2}=\left\{ F_{k}^{2}\right\}
_{k}$ of CZ $\kappa$-pivotal\emph{\ }stopping children of $F_{1}^{0}$ to be
the \emph{maximal} dyadic subcubes $I$ of some $F_{k}^{1}\in\mathcal{G}_{1}$
satisfyin
\[
\left\{
\begin{array}
[c]{c
\text{either }E_{I}^{\sigma}\left\vert f\right\vert \geq\Gamma E_{F_{k}^{1
}^{\sigma}\left\vert f\right\vert \\
\text{or }\mathrm{P}_{\kappa}^{\alpha}\left( I,\mathbf{1}_{F_{k}^{1}
\sigma\right) ^{2}\left\vert I\right\vert _{\omega}\geq\Gamma\left\vert
I\right\vert _{\sigma
\end{array}
\right. .
\]
Continue by recursion to define $\mathcal{G}_{n}$ for all $n\geq0$, and then
set
\[
\mathcal{F\equiv
{\displaystyle\bigcup\limits_{n=0}^{\infty}}
\mathcal{G}_{n}=\left\{ F_{k}^{n}:n\geq0,k\geq1\right\}
\]
to be the set of all CZ $\kappa$-pivotal stopping intervals in $F_{1}^{0}$
obtained in this way.
The $\sigma$-Carleson condition for $\mathcal{F}$ follows as usual from the
first step
\[
\sum_{F^{\prime}\in\mathfrak{C}_{\mathcal{F}}\left( F\right) }\left\vert
F^{\prime}\right\vert _{\sigma}\leq\frac{1}{\Gamma}\sum_{F^{\prime
\in\mathfrak{C}_{\mathcal{F}}\left( F\right) }\left\{ \mathrm{P}_{\kappa
}^{\alpha}\left( F^{\prime},\mathbf{1}_{F}\sigma\right) ^{2}\left\vert
F^{\prime}\right\vert _{\omega}+\frac{1}{E_{F}^{\sigma}\left\vert f\right\vert
}\int_{F^{\prime}}\left\vert f\right\vert d\sigma\right\} \leq\frac{1
{\Gamma}\left( \mathcal{V}_{2}^{\alpha,\kappa}\left( \sigma,\omega\right)
^{2}+1\right) \left\vert F\right\vert _{\sigma},
\]
where we used the $\kappa$-pivotal condition (\ref{both pivotal k}) in the
last inequality. Moreover, if we define
\begin{equation}
\alpha_{\mathcal{F}}\left( F\right) \equiv\sup_{F^{\prime}\in\mathcal{F
:\ F\subset F^{\prime}}E_{F^{\prime}}^{\sigma}\left\vert f\right\vert ,
\label{def alpha
\end{equation}
then in each corona
\[
\mathcal{C}_{F}\equiv\left\{ I\in\mathcal{D}:I\subset F\text{ and
}I\not \subset F^{\prime}\text{ for any }F^{\prime}\in\mathcal{F}\text{ with
}F^{\prime}\varsubsetneqq F\right\} ,
\]
we have, from the definition of the stopping times, both average contro
\begin{equation}
E_{I}^{\sigma}\left\vert f\right\vert <\Gamma\alpha_{\mathcal{F}}\left(
F\right) ,\ \ \ \ \ I\in\mathcal{C}_{F}\text{ and }F\in\mathcal{F},
\label{average control
\end{equation}
and $\kappa$-pivotal contro
\begin{equation}
\mathrm{P}_{\kappa}^{\alpha}\left( I,\mathbf{1}_{F}\sigma\right)
^{2}\left\vert I\right\vert _{\omega}<\Gamma\left\vert I\right\vert _{\sigma
},\ \ \ \ \ I\in\mathcal{C}_{F}\text{ and }F\in\mathcal{F}.
\label{energy control
\end{equation}
Finally, as in \cite{NTV4}, \cite{LaSaShUr3} and \cite{SaShUr7}, we obtain the
Carleson condition and the quasiorthogonality inequality
\begin{equation}
\sum_{F^{\prime}\preceq F}\left\vert F^{\prime}\right\vert _{\mu}\leq
C_{0}\left\vert F\right\vert _{\mu}\text{ for all }F\in\mathcal{F};\text{\ and
}\sum_{F\in\mathcal{F}}\alpha_{\mathcal{F}}\left( F\right) ^{2}\left\vert
F\right\vert _{\mu}\mathbf{\leq}C_{0}^{2}\left\Vert f\right\Vert
_{L^{2}\left( \mu\right) }^{2}. \label{Car and quasi
\end{equation}
Define the two corona projection
\[
\mathsf{P}_{\mathcal{C}_{F}}^{\sigma}\equiv\sum_{I\in\mathcal{C}_{F
}\bigtriangleup_{I;\kappa}^{\sigma}\text{ and }\mathsf{P}_{\mathcal{C
_{F}^{\mathbf{\tau}-\operatorname*{shift}}}^{\omega}\equiv\sum_{J\in
\mathcal{C}_{F}^{\mathbf{\tau}-\operatorname*{shift}}}\bigtriangleup
_{J;\kappa}^{\omega}\ ,
\]
wher
\begin{align}
\mathcal{C}_{F}^{\mathbf{\tau}-\operatorname*{shift}} & \equiv\left[
\mathcal{C}_{F}\setminus\mathcal{N}_{\mathcal{D}}^{\mathbf{\tau}}\left(
F\right) \right] \cu
{\displaystyle\bigcup\limits_{F^{\prime}\in\mathfrak{C}_{\mathcal{F}}\left(
F\right) }}
\left[ \mathcal{N}_{\mathcal{D}}^{\mathbf{\tau}}\left( F^{\prime}\right)
\setminus\mathcal{N}_{\mathcal{D}}^{\mathbf{\tau}}\left( F\right) \right]
;\label{def shift}\\
\text{where }\mathcal{N}_{\mathcal{D}}^{\mathbf{\tau}}\left( F\right) &
\equiv\left\{ J\in\mathcal{D}:J\subset F\text{ and }\ell\left( J\right)
>2^{-\mathbf{\tau}}\ell\left( F\right) \right\} ,\nonumber
\end{align}
and note that $f=\sum_{F\in\mathcal{F}}\mathsf{P}_{\mathcal{C}_{F}}^{\sigma
f$. Thus the corona $\mathcal{C}_{F}^{\mathbf{\tau}-\operatorname*{shift}}$
has the top $\mathbf{\tau}$ levels from $\mathcal{C}_{F}$ removed, and
includes the first $\mathbf{\tau}$ levels from each of its $\mathcal{F
$-children, except if they have already been removed.
The main result we need from \cite{Saw6} regarding these coronas is the
Intertwining Proposition.
\begin{proposition}
[{The Intertwining Proposition \cite[see Subsection 6.4]{Saw6}}
\label{strongly adapted}Suppose that $\mathcal{F}$ satisfies bot
\[
\sum_{F^{\prime}\in\mathcal{F}:F^{\prime}\subset F}\left\vert F^{\prime
}\right\vert _{\mu}\leq C_{0}\left\vert F\right\vert _{\mu}\text{ for all
}F\in\mathcal{F},\text{\ and }\sum_{F\in\mathcal{F}}\alpha_{\mathcal{F
}\left( F\right) ^{2}\left\vert F\right\vert _{\mu}\mathbf{\leq}C_{0
^{2}\left\Vert f\right\Vert _{L^{2}\left( \mu\right) }^{2},
\]
where $\alpha_{\mathcal{F}}\left( F\right) $ is as in (\ref{def alpha}), and
tha
\[
\left\Vert \bigtriangleup_{I;\kappa}^{\sigma}f\right\Vert _{L^{\infty}\left(
\sigma\right) }\leq C\alpha_{\mathcal{F}}\left( F\right) ,\ \ \ \ \ f\in
L^{2}\left( \sigma\right) ,\ I\in\mathcal{C}_{F}.
\]
Then for $\operatorname*{good}$ functions $f\in L^{2}\left( \sigma\right) $
and $g\in L^{2}\left( \omega\right) $, and with $\kappa\geq1$, we hav
\[
\left\vert \sum_{F\in\mathcal{F}}\ \sum_{I:\ I\supsetneqq F}\ \left\langle
T_{\sigma}^{\alpha}\bigtriangleup_{I;\kappa}^{\sigma}f,\mathsf{P
_{\mathcal{C}_{F}^{\mathbf{\tau}-\operatorname*{shift}}}^{\omega
}g\right\rangle _{\omega}\right\vert \lesssim\left( \mathcal{V}_{2
^{\alpha,\kappa}+\sqrt{A_{2}^{\alpha}}+\mathfrak{T}_{T^{\alpha}}^{\kappa
}\right) \ \left\Vert f\right\Vert _{L^{2}\left( \sigma\right) }\left\Vert
g\right\Vert _{L^{2}\left( \omega\right) }.
\]
\end{proposition}
\section{Reduction of the proof to local forms}
The proof of Theorem \ref{main} will require significant new arguments beyond
those in \cite{Saw6}. In particular, we will use the shifted corona
decomposition as in \cite{LaSaShUr3} and \cite{SaShUr7}, instead of the
parallel corona decomposition used in \cite{Saw6}, and construct a paraproduct
- stopping - neighbour decomposition of NTV type for a local Alpert form, but
including a new commutator form that arises from the fact that singular
integrals do not in general commute with multiplication by polynomials.
To prove Theorem \ref{main}, we begin by proving the bilinear form bound
\[
\left\vert \left\langle T_{\sigma}^{\alpha}f,g\right\rangle _{\omega
}\right\vert \lesssim\left( \sqrt{A_{2}^{\alpha}\left( \sigma,\omega\right)
}+\mathfrak{TR}_{T^{\alpha}}^{\left( \kappa\right) }\left( \sigma
,\omega\right) +\mathfrak{TR}_{\left( T^{\alpha}\right) ^{\ast}}^{\left(
\kappa\right) }\left( \omega,\sigma\right) +\mathcal{V}_{2}^{\alpha,\kappa
}+\mathcal{V}_{2}^{\alpha,\kappa,\ast}\right) \ \left\Vert f\right\Vert
_{L^{2}\left( \sigma\right) }\left\Vert g\right\Vert _{L^{2}\left(
\omega\right) },
\]
for $\operatorname{good}$ functions $f$ and $g$ in the sense of Nazarov, Treil
and Volberg, see \cite{NTV} for the treatment we use here\footnote{See also
\cite[Subsection 3.1]{SaShUr10} for a treatment using finite collections of
grids, in which case the conditional probability arguments are elementary.}.
Following the weighted Haar expansions of Nazarov, Treil and Volberg, we write
$f$ and $g$ in weighted Alpert wavelet expansions
\begin{equation}
\left\langle T_{\sigma}^{\alpha}f,g\right\rangle _{\omega}=\left\langle
T_{\sigma}^{\alpha}\left( \sum_{I\in\mathcal{D}}\bigtriangleup_{I;\kappa_{1
}^{\sigma}f\right) ,\left( \sum_{J\in\mathcal{D}}\bigtriangleup
_{J;\kappa_{2}}^{\omega}g\right) \right\rangle _{\omega}=\sum_{I\in
\mathcal{D}\ \text{and }J\in\mathcal{D}}\left\langle T_{\sigma}^{\alpha
}\left( \bigtriangleup_{I;\kappa_{1}}^{\sigma}f\right) ,\left(
\bigtriangleup_{J;\kappa_{2}}^{\omega}g\right) \right\rangle _{\omega}\ .
\label{expand
\end{equation}
Then the sum is further decomposed by first \emph{Cube Size Splitting}, then
using the \emph{Shifted Corona Decomposition}, according to the
\emph{Canonical Splitting}. We assume the reader is reasonably familiar with
the notation and arguments in the first eight sections of \cite{SaShUr7}. The
$n$-dimensional decompositions used in \cite{SaShUr7} are in spirit the same
as the one-dimensional decompositions in \cite{LaSaShUr3}, as well as the
$n$-dimensional decompositions in \cite{LaWi}, but differ in significant details.
A fundamental result of Nazarov, Treil and Volberg \cite{NTV} is that all the
cubes $I$ and $J$ appearing in the bilinear form above may be assumed to be
$\left( \mathbf{r},\varepsilon\right) -\operatorname*{good}$, where a dyadic
interval $K$ is $\left( \mathbf{r},\varepsilon\right) -\operatorname*{good
$, or simply $\operatorname*{good}$, if for \emph{every} dyadic supercube $L$
of $K$, it is the case that \textbf{either} $K$ has side length at least
$2^{1-\mathbf{r}}$ times that of $L$, \textbf{or} $K\Subset_{\left(
\mathbf{r},\varepsilon\right) }L$. We say that a dyadic cube $K$ is $\left(
\mathbf{r},\varepsilon\right) $-\emph{deeply embedded} in a dyadic cube $L$,
or simply $\mathbf{r}$\emph{-deeply embedded} in $L$, which we write as
$K\Subset_{\mathbf{r},\varepsilon}L$, when $K\subset L$ and both
\begin{align}
\ell\left( K\right) & \leq2^{-\mathbf{r}}\ell\left( L\right)
,\label{def deep embed}\\
\operatorname*{dist}\left( K
{\displaystyle\bigcup\limits_{L^{\prime}\in\mathfrak{C}_{\mathcal{D}}L}}
\partial L^{\prime}\right) & \geq2\ell\left( K\right) ^{\varepsilon
\ell\left( L\right) ^{1-\varepsilon}.\nonumber
\end{align}
Here is a brief schematic diagram derived from \cite{SaShUr7} summarizing the
shifted corona decompositions that we will follow using Alpert wavelet
expansions below. We first introduce parameters as in \cite{SaShUr7}. We will
choose $\varepsilon>0$ sufficiently small later in the argument, and then
$\mathbf{r}$ must be chosen sufficiently large depending on $\varepsilon$ in
order to reduce matters to $\left( \mathbf{r},\varepsilon\right)
-\operatorname{good}$ functions by the NTV argument in \cite{NTV}.
\begin{definition}
\label{def parameters}The parameters $\mathbf{\tau}$ and $\mathbf{\rho}$ are
fixed to satisfy
\[
\mathbf{\tau}>\mathbf{r}\text{ and }\mathbf{\rho}>\mathbf{r}+\mathbf{\tau},
\]
where $\mathbf{r}$ is the goodness parameter already fixed.
\end{definition}
\[
\fbox{
\begin{array}
[c]{ccccccc
\left\langle T_{\sigma}^{\alpha}f,g\right\rangle _{\omega} & & & & & & \\
\downarrow & & & & & & \\
\mathsf{B}_{\Subset_{\mathbf{\rho},\varepsilon}}\left( f,g\right) & + &
\mathsf{B}_{_{\mathbf{\rho},\varepsilon}\Supset}\left( f,g\right) & + &
\mathsf{B}_{\cap}\left( f,g\right) & + & \mathsf{B}_{\diagup}\left(
f,g\right) \\
\downarrow & & & & & & \\
\mathsf{T}_{\operatorname*{diagonal}}\left( f,g\right) & + & \mathsf{T
_{\operatorname*{far}\operatorname*{below}}\left( f,g\right) & + &
\mathsf{T}_{\operatorname*{far}\operatorname*{above}}\left( f,g\right) & + &
\mathsf{T}_{\operatorname*{disjoint}}\left( f,g\right) \\
\downarrow & & \downarrow & & & & \\
\mathsf{B}_{\Subset_{\mathbf{\rho},\varepsilon}}^{F}\left( f,g\right) & &
\mathsf{T}_{\operatorname*{far}\operatorname*{below}}^{1}\left( f,g\right) &
+ & \mathsf{T}_{\operatorname*{far}\operatorname*{below}}^{2}\left(
f,g\right) & & \\
\downarrow & & & & & & \\
\mathsf{B}_{\operatorname{stop}}^{F}\left( f,g\right) & + & \mathsf{B
_{\operatorname{paraproduct}}^{F}\left( f,g\right) & + & \mathsf{B
_{\operatorname{neighbour}}^{F}\left( f,g\right) & + & \mathsf{B
_{\operatorname*{commutator}}^{F}\left( f,g\right)
\end{array}
$
\]
\subsection{Cube Size Splitting}
The NTV \emph{Cube Size Splitting} of the inner product $\left\langle
T_{\sigma}^{\alpha}f,g\right\rangle _{\omega}$ given in (\ref{expand}) splits
the pairs of cubes $\left( I,J\right) $ in a simultaneous Alpert
decomposition of $f$ and $g$ into four groups determined by relative position,
and is given b
\begin{align}
\left\langle T_{\sigma}^{\alpha}f,g\right\rangle _{\omega} &
{\displaystyle\sum\limits_{\substack{I,J\in\mathcal{D}\\J\Subset
_{\mathbf{\rho},\varepsilon}I}}}
\left\langle T_{\sigma}^{\alpha}\left( \bigtriangleup_{I;\kappa}^{\sigma
}f\right) ,\left( \bigtriangleup_{J;\kappa}^{\omega}g\right) \right\rangle
_{\omega}
{\displaystyle\sum\limits_{\substack{I,J\in\mathcal{D}\\J_{\mathbf{\rho
},\varepsilon}\Supset I}}}
\left\langle T_{\sigma}^{\alpha}\left( \bigtriangleup_{I;\kappa}^{\sigma
}f\right) ,\left( \bigtriangleup_{J;\kappa}^{\omega}g\right) \right\rangle
_{\omega}\label{cube size}\\
&
{\displaystyle\sum\limits_{\substack{I,J\in\mathcal{D}\\J\cap I=\emptyset
\text{ and }\frac{\ell\left( J\right) }{\ell\left( I\right) }\notin\left[
2^{-\mathbf{\rho}},2^{\mathbf{\rho}}\right] }}}
\left\langle T_{\sigma}^{\alpha}\left( \bigtriangleup_{I;\kappa}^{\sigma
}f\right) ,\left( \bigtriangleup_{J;\kappa}^{\omega}g\right) \right\rangle
_{\omega}
{\displaystyle\sum\limits_{\substack{I,J\in\mathcal{D}\\2^{-\mathbf{\rho}
\leq\frac{\ell\left( J\right) }{\ell\left( I\right) }\leq2^{\mathbf{\rho
}}}}
\left\langle T_{\sigma}^{\alpha}\left( \bigtriangleup_{I;\kappa}^{\sigma
}f\right) ,\left( \bigtriangleup_{J;\kappa}^{\omega}g\right) \right\rangle
_{\omega}\nonumber\\
& =\mathsf{B}_{\Subset_{\mathbf{\rho},\varepsilon}}\left( f,g\right)
+\mathsf{B}_{_{\mathbf{\rho},\varepsilon}\Supset}\left( f,g\right)
+\mathsf{B}_{\cap}\left( f,g\right) +\mathsf{B}_{\diagup}\left( f,g\right)
.\nonumber
\end{align}
The assumption that the cubes $I$ and $J$ in the Alpert supports of $f$ and
$g$ are $\left( \mathbf{r},\varepsilon\right) -\operatorname*{good}$ remains
in force throughout the proof. We have the following bound from \cite[see
Lemma 31]{Saw6}, which in turn followed the NTV arguments for Haar wavelets in
\cite[see the proof of Lemma 7.1]{SaShUr7} (see also \cite{LaSaShUr3})
\begin{align}
\left\vert \mathsf{B}_{\cap}\left( f,g\right) +\mathsf{B}_{\diagup}\left(
f,g\right) \right\vert & \le
{\displaystyle\sum\limits_{\substack{I,J\in\mathcal{D}\\J\cap I=\emptyset
\text{ and }\frac{\ell\left( J\right) }{\ell\left( I\right) }\notin\left[
2^{-\mathbf{\rho}},2^{\mathbf{\rho}}\right] }}}
\left\vert \left\langle T_{\sigma}^{\alpha}\left( \bigtriangleup_{I;\kappa
}^{\sigma}f\right) ,\left( \bigtriangleup_{J;\kappa}^{\omega}g\right)
\right\rangle _{\omega}\right\vert
{\displaystyle\sum\limits_{\substack{I,J\in\mathcal{D}\\2^{-\mathbf{\rho}
\leq\frac{\ell\left( J\right) }{\ell\left( I\right) }\leq2^{\mathbf{\rho
}}}}
\left\vert \left\langle T_{\sigma}^{\alpha}\left( \bigtriangleup_{I;\kappa
}^{\sigma}f\right) ,\left( \bigtriangleup_{J;\kappa}^{\omega}g\right)
\right\rangle _{\omega}\right\vert \label{routine}\\
& \leq C\left( \mathfrak{T}_{T^{\alpha}}^{\kappa}+\mathfrak{T
_{T^{\alpha,\ast}}^{\kappa}+\mathcal{WBP}_{T^{\alpha}}^{\left( \kappa
_{1},\kappa_{2}\right) }\left( \sigma,\omega\right) +\sqrt{A_{2}^{\alpha
}\right) \left\Vert f\right\Vert _{L^{2}\left( \sigma\right) }\left\Vert
g\right\Vert _{L^{2}\left( \omega\right) },\nonumber
\end{align}
where if $\Omega$ is the set of all dyadic grids,
\[
\mathcal{WBP}_{T^{\alpha}}^{\left( \kappa_{1},\kappa_{2}\right) }\left(
\sigma,\omega\right) \equiv\sup_{\mathcal{D}\in\Omega}\sup
_{\substack{Q,Q^{\prime}\in\mathcal{D}\\Q\subset3Q^{\prime}\setminus
Q^{\prime}\text{ or }Q^{\prime}\subset3Q\setminus Q}}\frac{1}{\sqrt{\left\vert
Q\right\vert _{\sigma}\left\vert Q^{\prime}\right\vert _{\omega}}
\sup_{\substack{f\in\left( \mathcal{P}_{Q}^{\kappa_{1}}\right)
_{\operatorname*{norm}}\left( \sigma\right) \\g\in\left( \mathcal{P
_{Q^{\prime}}^{\kappa_{2}}\right) _{\operatorname*{norm}}\left(
\omega\right) }}\left\vert \int_{Q^{\prime}}T_{\sigma}^{\alpha}\left(
\mathbf{1}_{Q}f\right) \ gd\omega\right\vert <\infty
\]
is a weak boundedness constant introduced in \cite{Saw6}. This constant will
be removed in the final section below using the following bound proved in
\cite[see (6.25) in Subsection 6.7 and note that only triple testing is needed
there by choosing $\ell\left( Q^{\prime}\right) \leq\ell\left( Q\right) $
(using duality and $T^{\alpha,\ast}$ if needed)]{Saw6},
\begin{equation}
\mathcal{WBP}_{T^{\alpha}}^{\left( \kappa_{1},\kappa_{2}\right) }\left(
\sigma,\omega\right) \leq C_{\kappa,\varepsilon}\left( \mathfrak{TR
_{T^{\alpha}}^{\left( \kappa\right) }\left( \sigma,\omega\right)
+\mathfrak{TR}_{T^{\alpha,\ast}}^{\left( \kappa\right) }\left(
\omega,\sigma\right) \right) . \label{unif bound'
\end{equation}
Since the \emph{below} and \emph{above} forms $\mathsf{B}_{\Subset
_{\mathbf{\rho},\varepsilon}}\left( f,g\right) ,\mathsf{B}_{_{\mathbf{\rho
},\varepsilon}\Supset}\left( f,g\right) $ are symmetric, matters are reduced
to provin
\[
\left\vert \mathsf{B}_{\Subset_{\mathbf{\rho},\varepsilon}}\left( f,g\right)
\right\vert \lesssim\left( \mathfrak{T}_{T^{\alpha}}^{\kappa_{1
}+\mathfrak{T}_{T^{\alpha,\ast}}^{\kappa_{2}}+\sqrt{A_{2}^{\alpha}}\right)
\left\Vert f\right\Vert _{L^{2}\left( \sigma\right) }\left\Vert g\right\Vert
_{L^{2}\left( \omega\right) }\ .
\]
\subsection{Shifted Corona Decomposition}
For this we recall the \emph{Shifted Corona Decomposition}, as opposed to the
\emph{parallel} corona decomposition used in \cite{Saw6}, associated with the
Calder\'{o}n-Zygmund $\kappa$-pivotal stopping cubes $\mathcal{F}$ introduced
above. But first we must invoke standard arguments, using the $\kappa$-cube
testing conditions (\ref{full testing}),\ to permit us to assume that $f$ and
$g$ are supported in a finite union of dyadic cubes $F_{0}$ on which they have
vanishing moments of order less than $\kappa$.
\subsubsection{The initial reduction using testing}
For this construction, we will follow the treatment as given in
\cite{SaShUr12}. We first restrict $f$ and $g$ to be supported in a large
common cube $Q_{\infty}$. Then we cover $Q_{\infty}$ with $2^{n}$ pairwise
disjoint cubes $I_{\infty}\in\mathcal{D}$ with $\ell\left( I_{\infty}\right)
=\ell\left( Q_{\infty}\right) $. We now claim we can reduce matters to
consideration of the $2^{2n}$ form
\[
\sum_{I\in\mathcal{D}:\ I\subset I_{\infty}}\sum_{J\in\mathcal{D}:\ J\subset
J_{\infty}}\int\left( T_{\sigma}^{\alpha}\bigtriangleup_{I;\kappa}^{\sigma
}f\right) \bigtriangleup_{J;\kappa}^{\omega}gd\omega,
\]
as both $I_{\infty}$ and $J_{\infty}$ range over the dyadic cubes as above.
First we note that when $I_{\infty}$ and $J_{\infty}$ are distinct, the
corresponding form is included in the sum $\mathsf{B}_{\cap}\left(
f,g\right) +\mathsf{B}_{\diagup}\left( f,g\right) $, and hence controlled.
Thus it remains to consider the forms with $I_{\infty}=J_{\infty}$ and use the
cubes $I_{\infty}$ as the starting cubes in our corona construction below.
Indeed, we have from (\ref{Alpert expan}) tha
\[
f=\sum_{I\in\mathcal{D}:\ I\subset I_{\infty}}\bigtriangleup_{I;\kappa
}^{\sigma}f+\mathbb{E}_{I_{\infty};\kappa}^{\sigma}f\ \ \ \ \text{ and
\ \ \ \ }g=\sum_{J\in\mathcal{D}:\ J\subset I_{\infty}}\bigtriangleup
_{J;\kappa}^{\omega}g+\mathbb{E}_{I_{\infty};\kappa}^{\omega}g,
\]
which can then be used to write the bilinear form $\int\left( T_{\sigma
}f\right) gd\omega$ as a sum of the form
\begin{align}
& \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \int\left( T_{\sigma}f\right) gd\omega
=\sum_{I_{\infty}}\left\{ \sum_{I,J\in\mathcal{D}:\ I,J\subset I_{\infty
}\int\left( T_{\sigma}^{\alpha}\bigtriangleup_{I;\kappa}^{\sigma}f\right)
\bigtriangleup_{J;\kappa}^{\omega}gd\omega\right. \label{sum of forms}\\
& \left. +\sum_{I\in\mathcal{D}:\ I\subset I_{\infty}}\int\left( T_{\sigma
}^{\alpha}\bigtriangleup_{I;\kappa}^{\sigma}f\right) \mathbb{E}_{I_{\infty
};\kappa}^{\omega}gd\omega+\sum_{J\in\mathcal{D}:\ J\subset I_{\infty}
\int\left( T_{\sigma}^{\alpha}\mathbb{E}_{I_{\infty};\kappa}^{\sigma
}f\right) \bigtriangleup_{J;\kappa}^{\omega}gd\omega+\int\left( T_{\sigma
}^{\alpha}\mathbb{E}_{I_{\infty};\kappa}^{\sigma}f\right) \mathbb{E
_{I_{\infty};\kappa}^{\omega}gd\omega\right\} ,\nonumber
\end{align}
taken over the $2^{n}$ cubes $I_{\infty}$ above.
The second, third and fourth sums in (\ref{sum of forms}) can be controlled by
the $\kappa$-testing conditions (\ref{full testing}), e.g. using
Cauchy-Schwarz
\begin{align}
& \left\vert \sum_{I\in\mathcal{D}:\ I\subset I_{\infty}}\int\left(
T_{\sigma}^{\alpha}\bigtriangleup_{I;\kappa}^{\sigma}f\right) \mathbb{E
_{I_{\infty};\kappa}^{\omega}gd\omega\right\vert \leq\left\Vert \sum
_{I\in\mathcal{D}:\ I\subset I_{\infty}}\bigtriangleup_{I;\kappa}^{\sigma
}f\right\Vert _{L^{2}\left( \sigma\right) }\left\Vert \mathbb{E}_{I_{\infty
};\kappa}^{\omega}g\right\Vert _{L^{\infty}}\left\Vert \mathbf{1}_{I_{\infty
}T_{\omega}^{\alpha,\ast}\left( \frac{\mathbb{E}_{I_{\infty};\kappa}^{\omega
}g}{\left\Vert \mathbb{E}_{I_{\infty};\kappa}^{\omega}g\right\Vert
_{L^{\infty}}}\right) \right\Vert _{L^{2}\left( \sigma\right)
}\label{top control}\\
& \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \lesssim\left\Vert f\right\Vert
_{L^{2}\left( \sigma\right) }\frac{\left\vert \widehat{g}\left( I_{\infty
}\right) \right\vert }{\sqrt{\left\vert I_{\infty}\right\vert _{\omega}
}\mathfrak{T}_{T_{\omega}^{\alpha,\ast}}^{\kappa}\sqrt{\left\vert I_{\infty
}\right\vert _{\omega}}\leq\mathfrak{T}_{T_{\omega}^{\alpha,\ast}}^{\kappa
}\left\Vert f\right\Vert _{L^{2}\left( \sigma\right) }\left\Vert
g\right\Vert _{L^{2}\left( \omega\right) }\ ,\nonumber
\end{align}
and similarly for the third and fourth sum.
\subsubsection{The shifted corona}
Recall the shifted corona $\mathcal{C}_{\mathcal{F}}^{\mathbf{\tau
}-\operatorname*{shift}}\left( F\right) $ defined in (\ref{def shift}). A
simple but important property is the fact that the $\mathbf{\tau}$-shifted
coronas $\mathcal{C}_{F}^{\mathbf{\tau}-\operatorname*{shift}}$ have overlap
bounded by $\mathbf{\tau}$
\begin{equation}
\sum_{F\in\mathcal{F}}\mathbf{1}_{\mathcal{C}_{F}^{\mathbf{\tau
-\operatorname*{shift}}}\left( J\right) \leq\mathbf{\tau},\ \ \ \ \ J\in
\mathcal{D}. \label{tau overlap
\end{equation}
It is convenient, for use in the canonical splitting below, to introduce the
following shorthand notation for $F,G\in\mathcal{F}$
\[
\left\langle T_{\sigma}^{\alpha}\left( \mathsf{P}_{\mathcal{C}_{F}}^{\sigma
}f\right) ,\mathsf{P}_{\mathcal{C}_{G}^{\mathbf{\tau}-\operatorname*{shift}
}^{\omega}g\right\rangle _{\omega}^{\Subset_{\mathbf{\rho}}}\equiv
\sum_{\substack{I\in\mathcal{C}_{F}\text{ and }J\in\mathcal{C}_{G
^{\mathbf{\tau}-\operatorname*{shift}}\\J\Subset_{\mathbf{\rho}
I}}\left\langle T_{\sigma}^{\alpha}\left( \bigtriangleup_{I;\kappa}^{\sigma
}f\right) ,\left( \bigtriangleup_{J;\kappa}^{\omega}g\right) \right\rangle
_{\omega}\ .
\]
\subsection{Canonical Splitting}
We then proceed with the \emph{Canonical Splitting} of $\mathsf{B
_{\Subset_{\mathbf{\rho},\varepsilon}}\left( f,g\right) $ in
(\ref{cube size}) as in \cite{SaShUr7}, but with Alpert wavelets in place of
Haar wavelets
\begin{align*}
\mathsf{B}_{\Subset_{\mathbf{\rho},\varepsilon}}\left( f,g\right) &
=\sum_{F\in\mathcal{F}}\left\langle T_{\sigma}\left( \mathsf{P
_{\mathcal{C}_{F}}^{\sigma}f\right) ,\mathsf{P}_{\mathcal{C}_{F
^{\mathbf{\tau}-\operatorname*{shift}}}^{\omega}g\right\rangle _{\omega
}^{\Subset_{\mathbf{\rho}}}+\sum_{\substack{F,G\in\mathcal{F}\\G\subsetneqq
F}}\left\langle T_{\sigma}\left( \mathsf{P}_{\mathcal{C}_{F}}^{\sigma
}f\right) ,\mathsf{P}_{\mathcal{C}_{G}^{\mathbf{\tau}-\operatorname*{shift}
}^{\omega}g\right\rangle _{\omega}^{\Subset_{\mathbf{\rho}}}\\
& +\sum_{\substack{F,G\in\mathcal{F}\\G\supsetneqq F}}\left\langle T_{\sigma
}\left( \mathsf{P}_{\mathcal{C}_{F}}^{\sigma}f\right) ,\mathsf{P
_{\mathcal{C}_{G}^{\mathbf{\tau}-\operatorname*{shift}}}^{\omega
}g\right\rangle _{\omega}^{\Subset_{\mathbf{\rho}}}+\sum_{\substack{F,G\in
\mathcal{F}\\F\cap G=\emptyset}}\left\langle T_{\sigma}\left( \mathsf{P
_{\mathcal{C}_{F}}^{\sigma}f\right) ,\mathsf{P}_{\mathcal{C}_{G
^{\mathbf{\tau}-\operatorname*{shift}}}^{\omega}g\right\rangle _{\omega
}^{\Subset_{\mathbf{\rho}}}\\
& \equiv\mathsf{T}_{\operatorname*{diagonal}}\left( f,g\right)
+\mathsf{T}_{\operatorname*{far}\operatorname*{below}}\left( f,g\right)
+\mathsf{T}_{\operatorname*{far}\operatorname*{above}}\left( f,g\right)
+\mathsf{T}_{\operatorname*{disjoint}}\left( f,g\right) .
\end{align*}
The final two forms $\mathsf{T}_{\operatorname*{far}\operatorname*{above
}\left( f,g\right) $ and $\mathsf{T}_{\operatorname*{disjoint}}\left(
f,g\right) $ each vanish just as in \cite{SaShUr7}, since there are no pairs
$\left( I,J\right) \in\mathcal{C}_{F}\times\mathcal{C}_{G}^{\mathbf{\tau
}-\operatorname*{shift}}$ with both (\textbf{i}) $J\Subset_{\mathbf{\rho
},\varepsilon}I$ and (\textbf{ii}) either $F\subsetneqq G$ or $G\cap
F=\emptyset$. The far below form $\mathsf{T}_{\operatorname*{far
\operatorname*{below}}\left( f,g\right) $ is then further split into two
forms $\mathsf{T}_{\operatorname*{far}\operatorname*{below}}^{1}\left(
f,g\right) $ and $\mathsf{T}_{\operatorname*{far}\operatorname*{below}
^{2}\left( f,g\right) $ as in \cite{SaShUr7}
\begin{align*}
& \ \ \ \ \ \ \ \ \ \ \mathsf{T}_{\operatorname*{far}\operatorname*{below
}\left( f,g\right) =\sum_{G\in\mathcal{F}}\sum_{F\in\mathcal{F
:\ G\subsetneqq F}\sum_{\substack{I\in\mathcal{C}_{F}\text{ and
J\in\mathcal{C}_{G}^{\mathbf{\tau}-\operatorname*{shift}}\\J\Subset
_{\mathbf{\rho},\varepsilon}I}}\left\langle T_{\sigma}^{\alpha}\left(
\bigtriangleup_{I;\kappa}^{\sigma}f\right) ,\left( \bigtriangleup_{J;\kappa
}^{\omega}g\right) \right\rangle _{\omega}\\
& =\sum_{G\in\mathcal{F}}\sum_{F\in\mathcal{F}:\ G\subsetneqq F}\sum
_{J\in\mathcal{C}_{G}^{\mathbf{\tau}-\operatorname*{shift}}}\sum
_{I\in\mathcal{C}_{F}\text{ and }J\subset I}\left\langle T_{\sigma}^{\alpha
}\left( \bigtriangleup_{I;\kappa}^{\sigma}f\right) ,\left( \bigtriangleup
_{J;\kappa}^{\omega}g\right) \right\rangle _{\omega}\\
& -\sum_{F\in\mathcal{F}}\sum_{G\in\mathcal{F}:\ G\subsetneqq F}\sum
_{J\in\mathcal{C}_{G}^{\mathbf{\tau}-\operatorname*{shift}}}\sum
_{I\in\mathcal{C}_{F}\text{ and }J\subset I\text{ but }J\not \Subset
_{\mathbf{\rho},\varepsilon}I}\left\langle T_{\sigma}^{\alpha}\left(
\bigtriangleup_{I;\kappa}^{\sigma}f\right) ,\left( \bigtriangleup_{J;\kappa
}^{\omega}g\right) \right\rangle _{\omega}\equiv\mathsf{T
_{\operatorname*{far}\operatorname*{below}}^{1}\left( f,g\right)
-\mathsf{T}_{\operatorname*{far}\operatorname*{below}}^{2}\left( f,g\right)
.
\end{align*}
The second form $\mathsf{T}_{\operatorname*{far}\operatorname*{below}
^{2}\left( f,g\right) $ is easily seen to satisf
\begin{equation}
\left\vert \mathsf{T}_{\operatorname*{far}\operatorname*{below}}^{2}\left(
f,g\right) \right\vert \lesssim C\left( \mathfrak{T}_{T^{\alpha}}^{\kappa
}+\mathfrak{T}_{T^{\alpha,\ast}}^{\kappa}+\mathcal{WBP}_{T^{\alpha}}^{\left(
\kappa_{1},\kappa_{2}\right) }\left( \sigma,\omega\right) +\sqrt
{A_{2}^{\alpha}}\right) \left\Vert f\right\Vert _{L^{2}\left( \sigma\right)
}\left\Vert g\right\Vert _{L^{2}\left( \omega\right) },
\label{second far below
\end{equation}
just as for the analogous inequality in \cite{SaShUr7} for Haar wavelets, by
(\ref{routine}). To control the first and main form $\mathsf{T
_{\operatorname*{far}\operatorname*{below}}^{1}\left( f,g\right) $, we use
the $\kappa$-pivotal Intertwining Proposition \ref{strongly adapted} recalled
from \cite{Saw6} in the earlier section on preliminaries. This proposition
then immediately gives the boun
\begin{equation}
\left\vert \mathsf{T}_{\operatorname*{far}\operatorname*{below}}^{1}\left(
f,g\right) \right\vert \lesssim\left( \mathfrak{T}_{T^{\alpha}}^{\kappa_{1
}+\sqrt{A_{2}^{\alpha}}+\mathcal{V}_{2}^{\alpha,\kappa_{1}}\right) \left\Vert
f\right\Vert _{L^{2}\left( \sigma\right) }\left\Vert g\right\Vert
_{L^{2}\left( \omega\right) }. \label{first far below
\end{equation}
To handle the diagonal term $\mathsf{T}_{\operatorname*{diagonal}}\left(
f,g\right) $, we decompose according to the stopping times $\mathcal{F}$
\begin{equation}
\mathsf{T}_{\operatorname*{diagonal}}\left( f,g\right) =\sum_{F\in
\mathcal{F}}\mathsf{B}_{\Subset_{\mathbf{\rho},\varepsilon}}^{F}\left(
f,g\right) ,\text{ where }\mathsf{B}_{\Subset_{\mathbf{\rho},\varepsilon
}^{F}\left( f,g\right) \equiv\left\langle T_{\sigma}^{\alpha}\left(
\mathsf{P}_{\mathcal{C}_{F}}^{\sigma}f\right) ,\mathsf{P}_{\mathcal{C
_{F}^{\mathbf{\tau}-\operatorname*{shift}}}^{\omega}g\right\rangle _{\omega
}^{\Subset_{\mathbf{\rho}}}, \label{def block
\end{equation}
and it is enough, using Cauchy-Schwarz and quasiorthogonality
(\ref{Car and quasi}) in $f$, together with orthogonality in both $f$ and $g$,
to prove the `below form' bound involving the usual cube testing constant
\begin{equation}
\left\vert \mathsf{B}_{\Subset_{\mathbf{\rho},\varepsilon}}^{F}\left(
f,g\right) \right\vert \lesssim\left( \mathfrak{T}_{T^{\alpha}}+\sqrt
{A_{2}^{\alpha}}\right) \ \left( \alpha_{\mathcal{F}}\left( F\right)
\sqrt{\left\vert F\right\vert _{\sigma}}+\left\Vert \mathsf{P}_{\mathcal{C
_{F}}^{\sigma}f\right\Vert _{L^{2}\left( \sigma\right) }\right)
\ \left\Vert \mathsf{P}_{\mathcal{C}_{F}^{\mathbf{\tau}-\operatorname*{shift
}}^{\omega}g\right\Vert _{L^{2}\left( \omega\right) }\ .
\label{below form bound
\end{equation}
Indeed, this then gives the estimate
\begin{equation}
\left\vert \mathsf{T}_{\operatorname*{diagonal}}\left( f,g\right)
\right\vert \lesssim\left( \mathfrak{T}_{T^{\alpha}}+\sqrt{A_{2}^{\alpha
}\right) \left\Vert f\right\Vert _{L^{2}\left( \sigma\right) }\left\Vert
g\right\Vert _{L^{2}\left( \omega\right) }. \label{diag est
\end{equation}
Thus at this point we have reduced the proof of Theorem \ref{main} to
\begin{enumerate}
\item proving (\ref{below form bound}),
\item and controlling the triple polynomial testing condition
(\ref{full testing}) by the usual cube testing condition and the classical
Muckenhoupt condition (\ref{Muck and test}).
\end{enumerate}
In the next section we address the first issue by proving the inequality
(\ref{below form bound}) for the below forms $\mathsf{B}_{\Subset
_{\mathbf{\rho},\varepsilon}}^{F}\left( f,g\right) $. In the final section,
we address the second issue and complete the proof of Theorem \ref{main} by
drawing together all of the estimates.
\section{The NTV reach for Alpert wavelets}
In this section we give the main new argument of this paper. It will be
convenient to denote our fractional singular integral operators by
$T^{\lambda}$, $0\leq\lambda<n$, instead of $T^{\alpha}$, as $\alpha$ will
denote a multi-index in $\mathbb{Z}_{+}^{n}$. But first, we note that for a
doubling measure $\mu$, a cube $I$ and a polynomial $P$, we have $\left\Vert
P\mathbf{1}_{I}\right\Vert _{L^{\infty}\left( \mu\right) }=\sup_{x\in
I}\left\vert P\left( x\right) \right\vert $; in particular, $\left\Vert
P\mathbf{1}_{I}\right\Vert _{L^{\infty}\left( \sigma\right) }=\left\Vert
P\mathbf{1}_{I}\right\Vert _{L^{\infty}\left( \omega\right) }=\left\Vert
P\mathbf{1}_{I}\right\Vert _{L^{\infty}}$.
We will adapt the classical reach of NTV using Haar wavelet projections
$\bigtriangleup_{I}^{\sigma}$, namely the ingenious `thinking outside the box'
idea of the paraproduct / stopping / neighbour decomposition of Nazarov, Treil
and Volberg \cite{NTV4}. Since we are using weighted Alpert wavelet
projections $\bigtriangleup_{I;\kappa}^{\sigma}$ instead, the projection
$\mathbb{E}_{I^{\prime};\kappa}^{\sigma}\bigtriangleup_{I;\kappa}^{\sigma}f$
onto the child $I^{\prime}\in\mathfrak{C}_{\mathcal{D}}\left( I\right) $
equals $M_{I^{\prime};\kappa}\mathbf{1}_{I_{\pm}}$ where $M=M_{I^{\prime
};\kappa}$ is a polynomial of degree less than $\kappa$ restricted to
$I^{\prime}$, as opposed to a constant in the Haar case, and hence no longer
commutes in general with the operator $T_{\sigma}^{\lambda}$. As mentioned in
the introduction, this results in a new commutator form to be bounded, and
complicates bounding the remaining forms as well. We have from
(\ref{def block}), that $\mathsf{B}_{\Subset_{\mathbf{\rho},\varepsilon
};\kappa}^{F}\left( f,g\right) $ equal
\begin{align*}
& \sum_{\substack{I\in\mathcal{C}_{F}\text{ and }J\in\mathcal{C
_{F}^{\mathbf{\tau}-\operatorname*{shift}}\\J\Subset_{\mathbf{\rho
},\varepsilon}I}}\left\langle T_{\sigma}^{\lambda}\left( \mathbf{1}_{I_{J
}\bigtriangleup_{I;\kappa}^{\sigma}f\right) ,\bigtriangleup_{J;\kappa
}^{\omega}g\right\rangle _{\omega}+\sum_{\substack{I\in\mathcal{C}_{F}\text{
and }J\in\mathcal{C}_{F}^{\mathbf{\tau}-\operatorname*{shift}}\\J\Subset
_{\mathbf{\rho},\varepsilon}I}}\sum_{\theta\left( I_{J}\right)
\in\mathfrak{C}_{\mathcal{D}}\left( I\right) \setminus\left\{
I_{J}\right\} }\left\langle T_{\sigma}^{\lambda}\left( \mathbf{1
_{\theta\left( I_{J}\right) }\bigtriangleup_{I;\kappa}^{\sigma}f\right)
,\bigtriangleup_{J;\kappa}^{\omega}g\right\rangle _{\omega}\\
& \equiv\mathsf{B}_{\operatorname{home};\kappa}^{F}\left( f,g\right)
+\mathsf{B}_{\operatorname*{neighbour};\kappa}^{F}\left( f,g\right) ,
\end{align*}
and we further decompose the home form using
\begin{equation}
M_{I^{\prime}}=M_{I^{\prime};\kappa}\equiv\mathbf{1}_{I^{\prime
}\bigtriangleup_{I;\kappa}^{\sigma}f=\mathbb{E}_{I^{\prime};\kappa}^{\sigma
}\bigtriangleup_{I;\kappa}^{\sigma}f, \label{def M
\end{equation}
to obtai
\begin{align*}
& \mathsf{B}_{\operatorname{home};\kappa}^{F}\left( f,g\right)
=\sum_{\substack{I\in\mathcal{C}_{F}\text{ and }J\in\mathcal{C}_{F
^{\mathbf{\tau}-\operatorname*{shift}}\\J\Subset_{\mathbf{\rho},\varepsilon
I}}\left\langle M_{I_{J}}T_{\sigma}^{\lambda}\mathbf{1}_{F},\bigtriangleup
_{J;\kappa}^{\omega}g\right\rangle _{\omega}-\sum_{\substack{I\in
\mathcal{C}_{F}\text{ and }J\in\mathcal{C}_{F}^{\mathbf{\tau
-\operatorname*{shift}}\\J\Subset_{\mathbf{\rho},\varepsilon}I}}\left\langle
M_{I_{J}}T_{\sigma}^{\lambda}\mathbf{1}_{F\setminus I_{J}},\bigtriangleup
_{J;\kappa}^{\omega}g\right\rangle _{\omega}\\
& +\sum_{\substack{I\in\mathcal{C}_{F}\text{ and }J\in\mathcal{C
_{F}^{\mathbf{\tau}-\operatorname*{shift}}\\J\Subset_{\mathbf{\rho
},\varepsilon}I}}\left\langle \left[ T_{\sigma}^{\lambda},M_{I_{J}}\right]
\mathbf{1}_{I_{J}},\bigtriangleup_{J;\kappa}^{\omega}g\right\rangle _{\omega
}\equiv\mathsf{B}_{\operatorname*{paraproduct};\kappa}^{F}\left( f,g\right)
+\mathsf{B}_{\operatorname*{stop};\kappa}^{F}\left( f,g\right)
+\mathsf{B}_{\operatorname*{commutator};\kappa}^{F}\left( f,g\right) .
\end{align*}
Altogether then we have the weighted Alpert version of the NTV paraproduct
decomposition\footnote{In \cite[see the end of Section 10 on Concluding
Remarks]{Saw6} it was remarked that one cannot extend a nonconstant
polynomial, normalized to a cube $Q$, to a supercube $F$ without destroying
the normalization in general. This obstacle to the paraproduct decomposition
of NTV is overcome here by controlling the commutator form.}
\[
\mathsf{B}_{\Subset_{\mathbf{\rho},\varepsilon};\kappa}^{F}\left( f,g\right)
=\mathsf{B}_{\operatorname*{paraproduct};\kappa}^{F}\left( f,g\right)
+\mathsf{B}_{\operatorname*{stop};\kappa}^{F}\left( f,g\right)
+\mathsf{B}_{\operatorname*{commutator};\kappa}^{F}\left( f,g\right)
+\mathsf{B}_{\operatorname*{neighbour};\kappa}^{F}\left( f,g\right) .
\]
\subsection{The paraproduct form}
We first pigeonhole the sum over pairs $I$ and $J$ arising in the paraproduct
form according to which child $I^{\prime}\in\mathfrak{C}_{\mathcal{D}}\left(
I\right) $ contains $J$
\[
\mathsf{B}_{\operatorname*{paraproduct};\kappa}^{F}\left( f,g\right)
=\sum_{I\in\mathcal{C}_{F}}\sum_{I^{\prime}\in\mathfrak{C}_{\mathcal{D
}\left( I\right) }\sum_{\substack{J\in\mathcal{C}_{F}^{\mathbf{\tau
}-\operatorname*{shift}}:\ J\Subset_{\mathbf{\rho},\varepsilon}I\\J\subset
I^{\prime}}}\left\langle M_{I^{\prime};\kappa}T_{\sigma}^{\lambda
\mathbf{1}_{F},\bigtriangleup_{J;\kappa}^{\omega}g\right\rangle _{\omega}.
\]
This form $\mathsf{B}_{\operatorname*{paraproduct};\kappa}^{F}\left(
f,g\right) $ can be handled as usual, using the telescoping property
(\ref{telescoping}) to sum the restrictions to a cube $J\in\mathcal{C
_{F}^{\mathbf{\tau}-\operatorname*{shift}}$ of the polynomials $M_{I^{\prime
};\kappa}$ over the relevant cubes $I$, to obtain a restricted polynomial
$\mathbf{1}_{J}P_{I^{\prime};\kappa}$ that is controlled by $\alpha
_{\mathcal{F}}\left( F\right) $, and then passing the polynomial
$M_{J;\kappa}$ over to $\bigtriangleup_{J;\kappa}^{\omega}g$. More precisely,
for each $J\in\mathcal{C}_{F}^{\mathbf{\tau}-\operatorname*{shift}}$, let
$I_{J}^{\natural}$ denote the smallest $K\in\mathcal{C}_{F}$ such that
$J\Subset_{\mathbf{\rho},\varepsilon}K$ (and as a consequence $J\Subset
_{\mathbf{\rho},\varepsilon}I$ for all $I\supset I_{J}^{\natural}$), and let
$I_{J}^{\flat}$ denote the $\mathcal{D}$-child of $I_{J}^{\natural}$ that
contains $J$. Then we further consider the two possibilities where
$I_{J}^{\flat}$ is in $\mathcal{C}_{F}$ or not. We have
\begin{equation}
\sum_{I\in\mathcal{C}_{F}:\ I_{J}^{\natural}\subset I}\mathbf{1
_{J}M_{I^{\prime};\kappa}=\mathbf{1}_{J}\sum_{I\in\mathcal{C}_{F
:\ I_{J}^{\natural}\subset I}M_{I^{\prime};\kappa}=\mathbf{1}_{J}\left(
\mathbb{E}_{I_{J}^{\flat};\kappa}^{\sigma}f-\mathbb{E}_{F;\kappa}^{\sigma
}f\right) \equiv\left\{
\begin{array}
[c]{ccc
\mathbf{1}_{J}P_{J;\kappa}^{\operatorname{corona}} & \text{ if } &
I_{J}^{\flat}\in\mathcal{C}_{F}\\
\mathbf{1}_{J}P_{J;\kappa}^{\operatorname*{stopping}} & \text{ if } &
I_{J}^{\flat}\in\mathcal{F
\end{array}
\right. , \label{either event
\end{equation}
and we set $P_{J;\kappa}^{\operatorname{corona}}\equiv0$ if $I_{J}^{\flat
}\not \in \mathcal{C}_{F}$, and $P_{J;\kappa}^{\operatorname*{stopping}
\equiv0$ if $I_{J}^{\flat}\in\mathcal{C}_{F}$. Then we write $\mathsf{B
_{\operatorname*{paraproduct};\kappa}^{F}\left( f,g\right) $ a
\begin{align*}
& \sum_{\substack{I\in\mathcal{C}_{F}\text{, }I_{J}^{\flat}\in\mathcal{C
_{F}\text{ and }J\in\mathcal{C}_{F}^{\mathbf{\tau}-\operatorname*{shift
}\\J\Subset_{\mathbf{\rho},\varepsilon}I}}\left\langle M_{I_{J};\kappa
}T_{\sigma}^{\lambda}\mathbf{1}_{F},\bigtriangleup_{J;\kappa}^{\omega
}g\right\rangle _{\omega}+\sum_{\substack{I\in\mathcal{C}_{F}\text{,
I_{J}^{\flat}\not \in \mathcal{C}_{F}\text{ and }J\in\mathcal{C
_{F}^{\mathbf{\tau}-\operatorname*{shift}}\\J\Subset_{\mathbf{\rho
},\varepsilon}I}}\left\langle M_{I_{J};\kappa}T_{\sigma}^{\lambda
\mathbf{1}_{F},\bigtriangleup_{J;\kappa}^{\omega}g\right\rangle _{\omega}\\
& \equiv\mathsf{B}_{\operatorname*{paraproduct};\kappa}^{F\left(
\operatorname{corona}\right) }\left( f,g\right) +\mathsf{B
_{\operatorname*{paraproduct};\kappa}^{F\left( \operatorname*{stopping
\right) }\left( f,g\right) .
\end{align*}
First consider $\mathsf{B}_{\operatorname*{paraproduct};\kappa}^{F\left(
\operatorname{corona}\right) }\left( f,g\right) $. Then (\ref{analogue})
shows that
\begin{equation}
\left\Vert \mathbf{1}_{J}P_{J;\kappa}^{\operatorname{corona}}\right\Vert
_{L^{\infty}\left( \sigma\right) }\leq\left\Vert \mathbb{E}_{I_{J}^{\flat
};\kappa}^{\sigma}f\right\Vert _{L^{\infty}\left( \sigma\right) }+\left\Vert
\mathbb{E}_{F;\kappa}^{\sigma}f\right\Vert _{L^{\infty}\left( \sigma\right)
}\mathbb{\leq}C\alpha_{\mathcal{F}}\left( F\right) , \label{unif bdd
\end{equation}
which gives
\begin{align*}
& \left\vert \mathsf{B}_{\operatorname*{paraproduct};\kappa}^{F\left(
\operatorname{corona}\right) }\left( f,g\right) \right\vert =\left\vert
\sum_{J\in\mathcal{C}_{F}^{\mathbf{\tau}-\operatorname*{shift}}}\left\langle
\mathbf{1}_{J}\left( \sum_{\substack{I\in\mathcal{C}_{F}:\ I_{J}^{\flat
\in\mathcal{C}_{F}\\J\Subset_{\mathbf{\rho},\varepsilon}I}}\sum
_{\substack{I^{\prime}\in\mathfrak{C}_{\mathcal{D}}\left( I\right)
\\J\subset I^{\prime}\text{ }}}M_{I^{\prime};\kappa}\right) T_{\sigma
}^{\lambda}\mathbf{1}_{F},\bigtriangleup_{J;\kappa}^{\omega}g\right\rangle
_{\omega}\right\vert \\
& =\left\vert \sum_{J\in\mathcal{C}_{F}^{\mathbf{\tau}-\operatorname*{shift
}}\left\langle \mathbf{1}_{J}P_{J;\kappa}^{\operatorname{corona}}T_{\sigma
}^{\lambda}\mathbf{1}_{F},\bigtriangleup_{J;\kappa}^{\omega}g\right\rangle
_{\omega}\right\vert =\left\vert \left\langle T_{\sigma}^{\lambda
\mathbf{1}_{F},\sum_{J\in\mathcal{C}_{F}^{\mathbf{\tau}-\operatorname*{shift
}}P_{J;\kappa}^{\operatorname{corona}}\bigtriangleup_{J;\kappa}^{\omega
}g\right\rangle _{\omega}\right\vert \\
& \leq\left\Vert T_{\sigma}^{\lambda}\mathbf{1}_{F}\right\Vert _{L^{2}\left(
\omega\right) }\alpha_{\mathcal{F}}\left( F\right) \left\Vert \sum
_{J\in\mathcal{C}_{F}^{\mathbf{\tau}-\operatorname*{shift}}}\frac{P_{J;\kappa
}^{\operatorname{corona}}}{\alpha_{\mathcal{F}}\left( F\right)
}\bigtriangleup_{J;\kappa}^{\omega}g\right\Vert _{L^{2}\left( \omega\right)
}\leq\mathfrak{T}_{T^{\lambda}}\sqrt{\left\vert F\right\vert _{\sigma}
\alpha_{\mathcal{F}}\left( F\right) \left\Vert \sum_{J\in\mathcal{C
_{F}^{\mathbf{\tau}-\operatorname*{shift}}}\frac{P_{J;\kappa
^{\operatorname{corona}}}{\alpha_{\mathcal{F}}\left( F\right)
\bigtriangleup_{J;\kappa}^{\omega}g\right\Vert _{L^{2}\left( \omega\right)
}.
\end{align*}
Now we will use an almost orthogonality argument that exploits the fact that
for $J^{\prime}$ small compared to $J$, the function $M_{J^{\prime};\kappa
}\bigtriangleup_{J^{\prime};\kappa}^{\omega}g$ has vanishing $\omega$-mean,
and the polynomial $\mathbf{1}_{J}P_{J;\kappa}^{\operatorname{corona
}\bigtriangleup_{J;\kappa}^{\omega}g$ is relatively smooth at the scale of
$J^{\prime}$, together with the fact that the polynomials $R_{J;\kappa
\equiv\frac{P_{J;\kappa}^{\operatorname{corona}}}{\alpha_{\mathcal{F}}\left(
F\right) }$ of degree at most $\kappa-1$, have $L^{\infty}$ norm uniformly
bounded by (\ref{unif bdd}), to show that
\begin{align}
\left\Vert \sum_{J\in\mathcal{C}_{F}^{\mathbf{\tau}-\operatorname*{shift}
}R_{J;\kappa}\bigtriangleup_{J;\kappa}^{\omega}g\right\Vert _{L^{2}\left(
\omega\right) }^{2} & =\sum_{J\in\mathcal{C}_{F}^{\mathbf{\tau
}-\operatorname*{shift}}}\left\Vert R_{J;\kappa}\bigtriangleup_{J;\kappa
}^{\omega}g\right\Vert _{L^{2}\left( \omega\right) }^{2}+\sum
_{\substack{J,J^{\prime}\in\mathcal{C}_{F}^{\mathbf{\tau
-\operatorname*{shift}}\\J\neq J^{\prime}}}\int\left( R_{J;\kappa
}\bigtriangleup_{J;\kappa}^{\omega}g\right) \left( R_{J^{\prime};\kappa
}\bigtriangleup_{J^{\prime};\kappa}^{\omega}g\right) d\omega
\label{to show that}\\
& \lesssim\sum_{J\in\mathcal{C}_{F}^{\mathbf{\tau}-\operatorname*{shift}
}\left\Vert R_{J;\kappa}\bigtriangleup_{J;\kappa}^{\omega}g\right\Vert
_{L^{2}\left( \omega\right) }^{2}\lesssim\sum_{J\in\mathcal{C
_{F}^{\mathbf{\tau}-\operatorname*{shift}}}\left\Vert \bigtriangleup
_{J;\kappa}^{\omega}g\right\Vert _{L^{2}\left( \omega\right)
^{2}=\left\Vert \mathsf{P}_{\mathcal{C}_{F}^{\mathbf{\tau
-\operatorname*{shift}}}^{\omega}g\right\Vert _{L^{2}\left( \omega\right)
}^{2}.\nonumber
\end{align}
Indeed, if $J^{\prime}$ is small compared to $J$, and $J^{\prime}\subset
J_{J^{\prime}}\subset J$, we hav
\begin{align*}
& \left\vert \int\left( R_{J;\kappa}\bigtriangleup_{J;\kappa}^{\omega
}g\right) \left( R_{J^{\prime};\kappa}\bigtriangleup_{J^{\prime};\kappa
}^{\omega}g\right) d\omega\right\vert \\
& =\left\Vert R_{J;\kappa}\bigtriangleup_{J;\kappa}^{\omega}g\right\Vert
_{L^{\infty}\left( \omega\right) }\left\Vert R_{J^{\prime};\kappa
}\bigtriangleup_{J^{\prime};\kappa}^{\omega}g\right\Vert _{L^{\infty}\left(
\omega\right) }\left\vert \int\left( \frac{R_{J;\kappa}\bigtriangleup
_{J;\kappa}^{\omega}g}{\left\Vert R_{J;\kappa}\bigtriangleup_{J;\kappa
}^{\omega}g\right\Vert _{L^{\infty}\left( \omega\right) }}\right) \left(
\frac{R_{J^{\prime};\kappa}\bigtriangleup_{J^{\prime};\kappa}^{\omega
g}{\left\Vert R_{J^{\prime};\kappa}\bigtriangleup_{J^{\prime};\kappa}^{\omega
}g\right\Vert _{L^{\infty}\left( \omega\right) }}\right) d\omega\right\vert
\\
& \lesssim\left\Vert R_{J;\kappa}\bigtriangleup_{J;\kappa}^{\omega
}g\right\Vert _{L^{\infty}\left( \omega\right) }\left\Vert R_{J^{\prime
};\kappa}\bigtriangleup_{J^{\prime};\kappa}^{\omega}g\right\Vert _{L^{\infty
}\left( \omega\right) }\frac{\ell\left( J^{\prime}\right) }{\ell\left(
J\right) }\sqrt{\frac{\left\vert J^{\prime}\right\vert _{\omega}}{\left\vert
J\right\vert _{\omega}}}\sqrt{\left\vert J^{\prime}\right\vert _{\omega
}\left\vert J\right\vert _{\omega}}\\
& \lesssim\sqrt{\frac{\left\vert J^{\prime}\right\vert _{\omega}}{\left\vert
J\right\vert _{\omega}}}\frac{\ell\left( J^{\prime}\right) }{\ell\left(
J\right) }\left\Vert \bigtriangleup_{J^{\prime};\kappa}^{\omega}g\right\Vert
_{L^{2}\left( \omega\right) }\left\Vert \bigtriangleup_{J;\kappa}^{\omega
}g\right\Vert _{L^{2}\left( \omega\right) },
\end{align*}
by (\ref{analogue'}), i.e
\begin{align*}
\left\Vert R_{J^{\prime};\kappa}\bigtriangleup_{J^{\prime};\kappa}^{\omega
}g\right\Vert _{L^{\infty}\left( \omega\right) }\sqrt{\left\vert J^{\prime
}\right\vert _{\omega}} & \lesssim\left\Vert \bigtriangleup_{J^{\prime
};\kappa}^{\omega}g\right\Vert _{L^{\infty}\left( \omega\right)
\sqrt{\left\vert J^{\prime}\right\vert _{\omega}}\lesssim\left\Vert
\bigtriangleup_{J^{\prime};\kappa}^{\omega}g\right\Vert _{L^{2}\left(
\omega\right) },\\
\left\Vert R_{J;\kappa}\bigtriangleup_{J;\kappa}^{\omega}g\right\Vert
_{L^{\infty}\left( \omega\right) }\sqrt{\left\vert J\right\vert _{\omega}}
& \lesssim\left\Vert \bigtriangleup_{J;\kappa}^{\omega}g\right\Vert
_{L^{\infty}\left( \omega\right) }\sqrt{\left\vert J\right\vert _{\omega
}\lesssim\left\Vert \bigtriangleup_{J;\kappa}^{\omega}g\right\Vert
_{L^{2}\left( \omega\right) }.
\end{align*}
Thu
\begin{align*}
& \sum_{\substack{J,J^{\prime}\in\mathcal{C}_{F}^{\mathbf{\tau
-\operatorname*{shift}}\\J^{\prime}\subsetneqq J}}\left\vert \int\left(
R_{J;\kappa}\bigtriangleup_{J;\kappa}^{\omega}g\right) \left( R_{J^{\prime
};\kappa}\bigtriangleup_{J^{\prime};\kappa}^{\omega}g\right) d\omega
\right\vert \lesssim\sum_{\substack{J,J^{\prime}\in\mathcal{C}_{F
^{\mathbf{\tau}-\operatorname*{shift}}\\J^{\prime}\subsetneqq J}}\sqrt
{\frac{\left\vert J^{\prime}\right\vert _{\omega}}{\left\vert J\right\vert
_{\omega}}}\frac{\ell\left( J^{\prime}\right) }{\ell\left( J\right)
}\left\Vert \bigtriangleup_{J^{\prime};\kappa}^{\omega}g\right\Vert
_{L^{2}\left( \omega\right) }\left\Vert \bigtriangleup_{J;\kappa}^{\omega
}g\right\Vert _{L^{2}\left( \omega\right) }\\
& =\sum_{m=1}^{\infty}2^{-m}\sum_{\substack{J,J^{\prime}\in\mathcal{C
_{F}^{\mathbf{\tau}-\operatorname*{shift}}\\\ell\left( J^{\prime}\right)
=2^{-m}\ell\left( J\right) }}\sqrt{\frac{\left\vert J^{\prime}\right\vert
_{\omega}}{\left\vert J\right\vert _{\omega}}}\left\Vert \bigtriangleup
_{J^{\prime};\kappa}^{\omega}g\right\Vert _{L^{2}\left( \omega\right)
}\left\Vert \bigtriangleup_{J;\kappa}^{\omega}g\right\Vert _{L^{2}\left(
\omega\right) }\\
& \leq\sum_{m=1}^{\infty}2^{-m}\sqrt{\sum_{\substack{J^{\prime}\in
\mathcal{C}_{F}^{\mathbf{\tau}-\operatorname*{shift}}\\\pi^{\left( m\right)
}J^{\prime}\in\mathcal{C}_{F}^{\mathbf{\tau}-\operatorname*{shift}
}}\left\Vert \bigtriangleup_{J^{\prime};\kappa}^{\omega}g\right\Vert
_{L^{2}\left( \omega\right) }^{2}}\sqrt{\sum_{\substack{J^{\prime
\in\mathcal{C}_{F}^{\mathbf{\tau}-\operatorname*{shift}}\\J=\pi^{\left(
m\right) }J^{\prime}\in\mathcal{C}_{F}^{\mathbf{\tau}-\operatorname*{shift}
}}\frac{\left\vert J^{\prime}\right\vert _{\omega}}{\left\vert J\right\vert
_{\omega}}\left\Vert \bigtriangleup_{J;\kappa}^{\omega}g\right\Vert
_{L^{2}\left( \omega\right) }^{2}}\lesssim\sum_{J\in\mathcal{C
_{F}^{\mathbf{\tau}-\operatorname*{shift}}}\left\Vert \bigtriangleup
_{J;\kappa}^{\omega}g\right\Vert _{L^{2}\left( \omega\right) }^{2}\ .
\end{align*}
Altogether we have show
\begin{align*}
\left\vert \mathsf{B}_{\operatorname*{paraproduct};\kappa}^{F\left(
\operatorname{corona}\right) }\left( f,g\right) \right\vert &
\lesssim\mathfrak{T}_{T^{\lambda}}\sqrt{\left\vert F\right\vert _{\sigma
}\alpha_{\mathcal{F}}\left( F\right) \left\Vert \sum_{J\in\mathcal{C
_{F}^{\mathbf{\tau}-\operatorname*{shift}}}\frac{P_{J;\kappa
^{\operatorname{corona}}}{\alpha_{\mathcal{F}}\left( F\right)
\bigtriangleup_{J;\kappa}^{\omega}g\right\Vert _{L^{2}\left( \omega\right)
}\\
& \lesssim\mathfrak{T}_{T^{\lambda}}\sqrt{\left\vert F\right\vert _{\sigma
}\alpha_{\mathcal{F}}\left( F\right) \sqrt{\sum_{J\in\mathcal{C
_{F}^{\mathbf{\tau}-\operatorname*{shift}}}\left\Vert \bigtriangleup
_{J;\kappa}^{\omega}g\right\Vert _{L^{2}\left( \omega\right) }^{2
}=\mathfrak{T}_{T^{\lambda}}\sqrt{\left\vert F\right\vert _{\sigma}
\alpha_{\mathcal{F}}\left( F\right) \left\Vert \mathsf{P}_{\mathcal{C
_{F}^{\mathbf{\tau}-\operatorname*{shift}}}^{\omega}g\right\Vert
_{L^{2}\left( \omega\right) }.
\end{align*}
Finally, we consider the remaining paraproduct form $\mathsf{B
_{\operatorname*{paraproduct};\kappa}^{F\left( \operatorname*{stopping
\right) }\left( f,g\right) $ in which $I_{J}^{\flat}\in\mathcal{F}$ when
$P_{J;\kappa}^{\operatorname*{stopping}}$ is nonzero. From (\ref{either event
) we hav
\begin{align*}
& \left\vert \mathsf{B}_{\operatorname*{paraproduct};\kappa}^{F\left(
\operatorname*{stopping}\right) }\left( f,g\right) \right\vert =\left\vert
\sum_{F^{\prime}\in\mathfrak{C}_{\mathcal{F}}\left( F\right) }\sum
_{I\in\mathcal{C}_{F}\text{, }I\supset I_{J}^{\natural}\text{, }I_{J}^{\flat
}=F^{\prime}\text{ and }J\in\mathcal{C}_{F}^{\mathbf{\tau
-\operatorname*{shift}}}\left\langle M_{I_{J};\kappa}T_{\sigma}^{\lambda
}\mathbf{1}_{F},\bigtriangleup_{J;\kappa}^{\omega}g\right\rangle _{\omega
}\right\vert \\
& =\left\vert \sum_{F^{\prime}\in\mathfrak{C}_{\mathcal{F}}\left( F\right)
}\sum_{\substack{J:\ I_{J}^{\flat}=F^{\prime}\\J\in\mathcal{C}_{F
^{\mathbf{\tau}-\operatorname*{shift}}}}\left\langle \mathbf{1}_{J}\sum
_{I\in\mathcal{C}_{F}\text{, }I\supset I_{J}^{\natural}}M_{I_{J};\kappa
}T_{\sigma}^{\lambda}\mathbf{1}_{F},\bigtriangleup_{J;\kappa}^{\omega
}g\right\rangle _{\omega}\right\vert =\left\vert \sum_{F^{\prime
\in\mathfrak{C}_{\mathcal{F}}\left( F\right) }\sum_{\substack{J:\ I_{J
^{\flat}=F^{\prime}\\J\in\mathcal{C}_{F}^{\mathbf{\tau}-\operatorname*{shift
}}}\left\langle P_{J;\kappa}^{\operatorname*{stopping}}T_{\sigma}^{\lambda
}\mathbf{1}_{F},\bigtriangleup_{J;\kappa}^{\omega}g\right\rangle _{\omega
}\right\vert \\
& =\left\vert \sum_{F^{\prime}\in\mathfrak{C}_{\mathcal{F}}\left( F\right)
}\left\langle T_{\sigma}^{\lambda}\mathbf{1}_{F},\sum_{\substack{J:\ I_{J
^{\flat}=F^{\prime}\\J\in\mathcal{C}_{F}^{\mathbf{\tau}-\operatorname*{shift
}}}P_{J;\kappa}^{\operatorname*{stopping}}\bigtriangleup_{J;\kappa}^{\omega
}g\right\rangle _{\omega}\right\vert \lesssim\sum_{F^{\prime}\in
\mathfrak{C}_{\mathcal{F}}\left( F\right) }\left\Vert \mathbf{1}_{F^{\prime
}}T_{\sigma}^{\lambda}\mathbf{1}_{F}\right\Vert _{L^{2}\left( \omega\right)
}\alpha_{\mathcal{F}}\left( F\right) \left\Vert \sum_{\substack{J:\ I_{J
^{\flat}=F^{\prime}\\J\in\mathcal{C}_{F}^{\mathbf{\tau}-\operatorname*{shift
}}}S_{J;\kappa}\bigtriangleup_{J;\kappa}^{\omega}g\right\Vert _{L^{2}\left(
\omega\right) }\\
& \leq\mathfrak{T}_{T^{\lambda}}\sqrt{\left\vert F\right\vert _{\sigma
}\alpha_{\mathcal{F}}\left( F\right) \left\Vert \sum_{J\in\mathcal{C
_{F}^{\mathbf{\tau}-\operatorname*{shift}}}S_{J;\kappa}\bigtriangleup
_{J;\kappa}^{\omega}g\right\Vert _{L^{2}\left( \omega\right)
\end{align*}
where $S_{J;\kappa}=\frac{P_{J;\kappa}^{\operatorname*{stopping}}
{\alpha_{\mathcal{F}}\left( F\right) }$, and where in the last line we have
used Cauchy-Schwarz in $F^{\prime}$, the pairwise disjointedness of the
$F^{\prime}$, and the testing condition. Now we use that $\sigma$ doubling and
$\pi_{\mathcal{D}}F^{\prime}\in\mathcal{C}_{F}$ imply $E_{F^{\prime
}\left\vert f\right\vert \lesssim\alpha_{\mathcal{F}}\left( F\right) $, and
hence that $\left\Vert S_{J;\kappa}\right\Vert _{L^{\infty}}\leq C$. Thus we
can apply (\ref{to show that}) with $R_{J;\kappa}$ replaced by $S_{J;\kappa}$
to obtain
\[
\left\Vert \sum_{J\in\mathcal{C}_{F}^{\mathbf{\tau}-\operatorname*{shift}
}S_{J;\kappa}\bigtriangleup_{J;\kappa}^{\omega}g\right\Vert _{L^{2}\left(
\omega\right) }^{2}\lesssim\sum_{J\in\mathcal{C}_{F}^{\mathbf{\tau
}-\operatorname*{shift}}}\left\Vert S_{J;\kappa}\bigtriangleup_{J;\kappa
}^{\omega}g\right\Vert _{L^{2}\left( \omega\right) }^{2}\lesssim\sum
_{J\in\mathcal{C}_{F}^{\mathbf{\tau}-\operatorname*{shift}}}\left\Vert
\bigtriangleup_{J;\kappa}^{\omega}g\right\Vert _{L^{2}\left( \omega\right)
}^{2}=\left\Vert \mathsf{P}_{\mathcal{C}_{F}^{\mathbf{\tau
-\operatorname*{shift}}}^{\omega}g\right\Vert _{L^{2}\left( \omega\right)
}^{2},
\]
because the polynomials $S_{J;\kappa}$ satisfy the same estimates as the
polynomials $R_{J;\kappa}$. Combining the estimates for $\mathsf{B
_{\operatorname*{paraproduct};\kappa}^{F\left( \operatorname{corona}\right)
}\left( f,g\right) $ and $\mathsf{B}_{\operatorname*{paraproduct};\kappa
}^{F\left( \operatorname*{stopping}\right) }\left( f,g\right) $, we
obtain
\begin{equation}
\left\vert \mathsf{B}_{\operatorname*{paraproduct};\kappa}^{F}\left(
f,g\right) \right\vert \lesssim\mathfrak{T}_{T^{\lambda}}\alpha_{\mathcal{F
}\left( F\right) \sqrt{\left\vert F\right\vert _{\sigma}}\ \left\Vert
\mathsf{P}_{\mathcal{C}_{F}^{\mathbf{\tau}-\operatorname*{shift}}}^{\omega
}g\right\Vert _{L^{2}\left( \omega\right) } \label{para est
\end{equation}
as required by (\ref{below form bound}).
\subsection{The commutator form}
If $T=H$ is the Hilbert transform on the real line, and if $P_{\ell}\left(
x\right) =x^{\ell}$ with $1\leq\ell\leq\kappa$ , then by the moment vanishing
properties of Alpert projections, we get that $H$ commutes with polynomials
$P$ of degree at most $\kappa$ when acting on a function with vanishing
$\sigma$-means up to order $\kappa-1$, i.e. $\left\langle \left(
HP-PH\right) \bigtriangleup_{I;\kappa}^{\sigma}f,g\right\rangle _{\omega
}=\left\langle 0,g\right\rangle =0$. This motivates the following argument.
Fix $\kappa\geq1$. Assume that $K^{\lambda}$ is a general standard $\lambda
$-fractional kernel in $\mathbb{R}^{n}$, and $T^{\lambda}$ is the associated
Calder\'{o}n-Zygmund operator, and that $P_{\alpha,a,I^{\prime}}\left(
x\right) =\left( \frac{x-a}{\ell\left( I^{\prime}\right) }\right)
^{\alpha}=\left( \frac{x_{1}-a_{1}}{\ell\left( I^{\prime}\right) }\right)
^{\alpha_{1}}...\left( \frac{x_{n}-a_{n}}{\ell\left( I^{\prime}\right)
}\right) ^{\alpha_{n}}$, where $1\leq\left\vert \alpha\right\vert \leq
\kappa-1$ (since when $|\alpha|=0$, $P_{\alpha,a,I^{\prime}}$ commutes with
$T^{\lambda}$) and $I^{\prime}\in\mathfrak{C}_{\mathcal{D}}\left( I\right)
$, $I\in\mathcal{C}_{F}$. We use the well-known formul
\[
x^{\alpha}-y^{\alpha}=\sum_{k=1}^{n}\left( x_{k}-y_{k}\right)
{\displaystyle\sum\limits_{\left\vert \beta\right\vert +\left\vert
\gamma\right\vert =\left\vert \alpha\right\vert -1}}
c_{\alpha,\beta,\gamma}x^{\beta}y^{\gamma},
\]
to writ
\begin{align*}
& \mathbf{1}_{I^{\prime}}\left( x\right) \left[ P_{\alpha,a,I^{\prime
}\ ,T_{\sigma}^{\lambda}\right] \mathbf{1}_{I^{\prime}}\left( x\right)
=\mathbf{1}_{I^{\prime}}\left( x\right) \int K^{\lambda}\left( x-y\right)
\left\{ P_{\alpha,a,I^{\prime}}\left( x\right) -P_{\alpha,a,I^{\prime
}\left( y\right) \right\} \mathbf{1}_{I^{\prime}}\left( y\right)
d\sigma\left( y\right) \\
& =\mathbf{1}_{I^{\prime}}\left( x\right) \int K^{\lambda}\left(
x-y\right) \left\{ \sum_{k=1}^{n}\left( \frac{x_{k}-y_{k}}{\ell\left(
I^{\prime}\right) }\right)
{\displaystyle\sum\limits_{\left\vert \beta\right\vert +\left\vert
\gamma\right\vert =\left\vert \alpha\right\vert -1}}
c_{\alpha,\beta,\gamma}\left( \frac{x-a}{\ell\left( I^{\prime}\right)
}\right) ^{\beta}\left( \frac{y-a}{\ell\left( I^{\prime}\right) }\right)
^{\gamma}\right\} \mathbf{1}_{I^{\prime}}\left( y\right) d\sigma\left(
y\right) \\
& =\sum_{k=1}^{n
{\displaystyle\sum\limits_{\left\vert \beta\right\vert +\left\vert
\gamma\right\vert =\left\vert \alpha\right\vert -1}}
c_{\alpha,\beta,\gamma}\mathbf{1}_{I^{\prime}}\left( x\right) \left[
\int\Phi_{k}^{\lambda}\left( x-y\right) \left\{ \left( \frac{y-a
{\ell\left( I^{\prime}\right) }\right) ^{\gamma}\right\} \mathbf{1
_{I^{\prime}}\left( y\right) d\sigma\left( y\right) \right] \left(
\frac{x-a}{\ell\left( I^{\prime}\right) }\right) ^{\beta},
\end{align*}
where $\Phi_{k}^{\lambda}\left( x-y\right) =K^{\lambda}\left( x-y\right)
\left( \frac{x_{k}-y_{k}}{\ell\left( I^{\prime}\right) }\right) $. Thus
$\left[ P_{\alpha,a,I^{\prime}},T_{\sigma}^{\lambda}\right] \mathbf{1
_{I^{\prime}}\left( x\right) $ is a `polynomial' of degree $\left\vert
\alpha\right\vert -1$ with \emph{variable} coefficients. Integrating the
commutator against $\bigtriangleup_{J;\kappa}^{\omega}g$ for some $J\subset
I^{\prime}$, we get
\begin{align}
& \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \left\langle \left[ P_{\alpha,a,I^{\prime
}\ ,T_{\sigma}^{\lambda}\right] \mathbf{1}_{I^{\prime}},\bigtriangleup
_{J;\kappa}^{\omega}g\right\rangle _{\omega}=\int\left[ P_{\alpha
,a,I^{\prime}}\ ,T_{\sigma}^{\lambda}\right] \mathbf{1}_{I^{\prime}}\left(
x\right) \bigtriangleup_{J;\kappa}^{\omega}g\left( x\right) d\omega\left(
x\right) \label{two pieces}\\
& =\int\left[ P_{\alpha,a,I^{\prime}}\ ,T_{\sigma}^{\lambda}\right]
\mathbf{1}_{I^{\prime}\setminus2J}\left( x\right) \bigtriangleup_{J;\kappa
}^{\omega}g\left( x\right) d\omega\left( x\right) +\int\left[
P_{\alpha,a,I^{\prime}}\ ,T_{\sigma}^{\lambda}\right] \mathbf{1}_{2J}\left(
x\right) \bigtriangleup_{J;\kappa}^{\omega}g\left( x\right) d\omega\left(
x\right) \equiv\operatorname{Int}^{\lambda,\natural}\left( J\right)
+\operatorname{Int}^{\lambda,\flat}\left( J\right) ,\nonumber
\end{align}
where we are suppressing the dependence on $\alpha$. For the first term
$\operatorname{Int}^{\lambda,\natural}\left( J\right) $ we writ
\[
\operatorname{Int}^{\lambda,\natural}\left( J\right) =\sum_{k=1}^{n
{\displaystyle\sum\limits_{\left\vert \beta\right\vert +\left\vert
\gamma\right\vert =\left\vert \alpha\right\vert -1}}
c_{\alpha,\beta,\gamma}\operatorname{Int}_{k,\beta,\gamma}^{\lambda,\natural
}\left( J\right) ,
\]
where with the choice $a=c_{J}$ the center of $J$, we defin
\begin{align}
\operatorname{Int}_{k,\beta,\gamma}^{\lambda,\natural}\left( J\right) &
\equiv\int_{J}\left[ \int_{I^{\prime}\setminus2J}\Phi_{k}^{\lambda}\left(
x-y\right) \left( \frac{y-c_{J}}{\ell\left( I^{\prime}\right) }\right)
^{\gamma}d\sigma\left( y\right) \right] \left( \frac{x-c_{J}}{\ell\left(
I^{\prime}\right) }\right) ^{\beta}\bigtriangleup_{J;\kappa}^{\omega
}g\left( x\right) d\omega\left( x\right) \label{coefficients}\\
& =\int_{I^{\prime}\setminus2J}\left\{ \int_{J}\Phi_{k}^{\lambda}\left(
x-y\right) \left( \frac{x-c_{J}}{\ell\left( I^{\prime}\right) }\right)
^{\beta}\bigtriangleup_{J;\kappa}^{\omega}g\left( x\right) d\omega\left(
x\right) \right\} \left( \frac{y-c_{J}}{\ell\left( I^{\prime}\right)
}\right) ^{\gamma}d\sigma\left( y\right) .\nonumber
\end{align}
While these integrals need no longer vanish, we will show they are suitably
small, using that the function $\left( \frac{x-c_{J}}{\ell\left( I^{\prime
}\right) }\right) ^{\beta}\bigtriangleup_{J;\kappa}^{\omega}g\left(
x\right) $ is supported in $J$ and has vanishing $\omega$-means up to order
$\kappa-\left\vert \beta\right\vert -1$, and that the function $\Phi
_{k}^{\lambda}\left( z\right) $ is appropriately smooth away from $z=0$,
i.e. $\left\vert \nabla^{m}\Phi_{k}^{\lambda}\left( z\right) \right\vert
\leq C_{m,n}\frac{1}{\left\vert z\right\vert ^{m+n-\lambda-1}\ell\left(
I\right) }$.
Indeed, we have the following estimate for the integral in braces in
(\ref{coefficients}), keeping in mind that $y\in I^{\prime}\setminus2J$ and
$x\in J$, where the second line below follows by Taylor's formula and that
$h\left( x\right) =\left( \frac{x-c_{J}}{\ell\left( I^{\prime}\right)
}\right) ^{\beta}\bigtriangleup_{J;\kappa}^{\omega}g\left( x\right) $ has
vanishing $\omega$-means up to order $\kappa-1-\left\vert \beta\right\vert $,
and the third line uses that $h\left( x\right) $ is supported in $J$
\begin{align*}
& \left\vert \int\Phi_{k}^{\lambda}\left( x-y\right) h\left( x\right)
d\omega\left( x\right) \right\vert =\left\vert \int\frac{1}{\left(
\kappa-\left\vert \beta\right\vert \right) !}\left( \left( x-c_{J}\right)
\cdot\nabla\right) ^{\kappa-\left\vert \beta\right\vert }\Phi_{k}^{\lambda
}\left( \eta_{J}^{\omega}\right) h\left( x\right) d\omega\left( x\right)
\right\vert \\
& \lesssim\left\Vert h\right\Vert _{L^{1}\left( \omega\right) }\frac
{\ell\left( J\right) ^{\kappa-\left\vert \beta\right\vert }}{\left[
\ell\left( J\right) +\operatorname*{dist}\left( y,J\right) \right]
^{\kappa-\left\vert \beta\right\vert +n-\lambda-1}\ell\left( I^{\prime
}\right) }\lesssim\left( \frac{\ell\left( J\right) }{\ell\left(
I^{\prime}\right) }\right) ^{\left\vert \beta\right\vert }\frac{\ell\left(
J\right) ^{\kappa-\left\vert \beta\right\vert }}{\left[ \ell\left(
J\right) +\operatorname*{dist}\left( y,J\right) \right] ^{\kappa
-\left\vert \beta\right\vert +n-\lambda-1}\ell\left( I^{\prime}\right)
}\sqrt{\left\vert J\right\vert _{\omega}}\left\vert \widehat{g}\left(
J\right) \right\vert ,
\end{align*}
since
\[
\left\Vert h\right\Vert _{L^{1}\left( \omega\right) }=\int_{J}\left\vert
\left( \frac{x-c_{J}}{\ell\left( I^{\prime}\right) }\right) ^{\beta
}\bigtriangleup_{J;\kappa}^{\omega}g\left( x\right) \right\vert
d\omega\left( x\right) \leq\left( \frac{\ell\left( J\right) }{\ell\left(
I^{\prime}\right) }\right) ^{\left\vert \beta\right\vert }\left\Vert
\bigtriangleup_{J;\kappa}^{\omega}g\right\Vert _{L^{1}\left( \omega\right)
}\lesssim\left( \frac{\ell\left( J\right) }{\ell\left( I^{\prime}\right)
}\right) ^{\left\vert \beta\right\vert }\sqrt{\left\vert J\right\vert
_{\omega}}\left\vert \widehat{g}\left( J\right) \right\vert .
\]
Now recall the orthonormal basis $\left\{ h_{I;\kappa}^{\mu,a}\right\}
_{a\in\Gamma_{I,n,\kappa}}$ of $L_{I;\kappa}^{2}\left( \mu\right) $. We
consider the polynomial
\[
Q_{I^{\prime};\kappa}^{\mu}\equiv\frac{1}{\left\vert \widehat{f}\left(
I\right) \right\vert }1_{I^{\prime}}\bigtriangleup_{I;\kappa}^{\sigma
f=\frac{1}{\left\vert \widehat{f}\left( I\right) \right\vert }M_{I^{\prime
};\kappa},
\]
so that $Q_{I^{\prime};\kappa}^{\mu}$ is a renormalization of the polynomial
$M_{I^{\prime};\kappa}$ introduced earlier. From (\ref{analogue'}) we have
$\left\Vert Q_{I^{\prime};\kappa}^{\sigma}\right\Vert _{\infty}\approx\frac
{1}{\sqrt{\left\vert I\right\vert _{\sigma}}}$. Hence for $c_{J}\in J\subset
I^{\prime}$, if we writ
\[
Q_{I^{\prime};\kappa}^{\sigma}\left( y\right) =\sum_{\left\vert
\alpha\right\vert <\kappa}b_{\alpha}\left( \frac{y-c_{J}}{\ell\left(
I^{\prime}\right) }\right) ^{\alpha}=\sum_{\left\vert \alpha\right\vert
<\kappa}b_{\alpha}P_{\alpha,c_{J},I^{\prime}}\left( y\right) ,
\]
and then rescale to the unit cube and invoke the fact that any two norms on a
finite dimensional vector space are equivalent, we obtai
\begin{equation}
\sum_{\left\vert \alpha\right\vert <\kappa}\left\vert b_{\alpha}\right\vert
\approx\left\Vert Q_{I^{\prime};\kappa}^{\sigma}\right\Vert _{\infty
\approx\frac{1}{\sqrt{\left\vert I\right\vert _{\sigma}}}. \label{we obtain
\end{equation}
We then boun
\begin{align*}
& \left\vert \left\langle \left[ Q_{I^{\prime};\kappa},T_{\sigma}^{\lambda
}\right] \mathbf{1}_{I^{\prime}\setminus2J},\bigtriangleup_{J;\kappa
^{\omega}g\right\rangle _{\omega}\right\vert \leq\sum_{\left\vert
\alpha\right\vert <\kappa}\left\vert b_{\alpha}\left\langle \left[
P_{\alpha,c_{J},I^{\prime}},T_{\sigma}^{\lambda}\right] \mathbf{1
_{I^{\prime}\setminus2J},\bigtriangleup_{J;\kappa}^{\omega}g\right\rangle
_{\omega}\right\vert \\
& \lesssim\frac{1}{\sqrt{\left\vert I\right\vert _{\sigma}}}\max_{\left\vert
\alpha\right\vert <\kappa}\left\vert \left\langle \left[ P_{\alpha
,c_{J},I^{\prime}},T_{\sigma}^{\lambda}\right] \mathbf{1}_{I^{\prime
}\setminus2J},\bigtriangleup_{J;\kappa}^{\omega}g\right\rangle _{\omega
}\right\vert ,
\end{align*}
so that it remains to estimate each inner product $\operatorname{Int
^{\lambda,\natural}\left( J\right) =\left\langle \left[ P_{\alpha
,c_{J},I^{\prime}},T_{\sigma}^{\lambda}\right] \mathbf{1}_{I^{\prime
}\setminus2J},\bigtriangleup_{J;\kappa}^{\omega}g\right\rangle _{\omega}$ as follows
\[
\left\vert \operatorname{Int}^{\lambda,\natural}\left( J\right) \right\vert
=\left\vert \sum_{k=1}^{n
{\displaystyle\sum\limits_{\left\vert \beta\right\vert +\left\vert
\gamma\right\vert =\left\vert \alpha\right\vert -1}}
c_{\alpha,\beta,\gamma}\operatorname{Int}_{k,\beta,\gamma}^{\lambda,\natural
}\left( J\right) \right\vert \lesssim\max_{\left\vert \beta\right\vert
+\left\vert \gamma\right\vert =\left\vert \alpha\right\vert -1}\left\vert
\operatorname{Int}_{k,\beta,\gamma}^{\lambda,\natural}\left( J\right)
\right\vert ,
\]
where $\left\vert \beta\right\vert +\left\vert \gamma\right\vert =\left\vert
\alpha\right\vert -1$. The estimates above imply
\begin{align*}
& \left\vert \operatorname{Int}_{k,\beta,\gamma}^{\lambda,\natural}\left(
J\right) \right\vert \leq\int_{I^{\prime}\setminus2J}\left\vert \int\Phi
_{k}^{\lambda}\left( x-y\right) \left( \frac{x-c_{J}}{\ell\left(
I^{\prime}\right) }\right) ^{\beta}\bigtriangleup_{J;\kappa}^{\omega
}g\left( x\right) d\omega\left( x\right) \right\vert \left\vert \left(
\frac{y-c_{J}}{\ell\left( I^{\prime}\right) }\right) ^{\gamma}\right\vert
d\sigma\left( y\right) \\
& \lesssim\int_{I^{\prime}\setminus2J}\left( \frac{\ell\left( J\right)
}{\ell\left( I^{\prime}\right) }\right) ^{\left\vert \beta\right\vert
}\frac{\ell\left( J\right) ^{\kappa-\left\vert \beta\right\vert }}{\left[
\ell\left( J\right) +\operatorname*{dist}\left( y,J\right) \right]
^{\kappa-\left\vert \beta\right\vert +n-\lambda-1}\ell\left( I\right)
\sqrt{\left\vert J\right\vert _{\omega}}\left\vert \widehat{g}\left(
J\right) \right\vert \left( \frac{\ell\left( J\right)
+\operatorname*{dist}\left( y,J\right) }{\ell\left( I^{\prime}\right)
}\right) ^{\left\vert \gamma\right\vert }d\sigma\left( y\right) \\
& =\int_{I^{\prime}\setminus2J}\left( \frac{\ell\left( J\right)
{\ell\left( I^{\prime}\right) }\right) ^{\left\vert \alpha\right\vert
-1}\left( \frac{\ell\left( J\right) }{\ell\left( J\right)
+\operatorname*{dist}\left( y,J\right) }\right) ^{\kappa-\left\vert
\alpha\right\vert +1}\frac{1}{\left[ \ell\left( J\right)
+\operatorname*{dist}\left( y,J\right) \right] ^{n-\lambda-1}\ell\left(
I^{\prime}\right) }\sqrt{\left\vert J\right\vert _{\omega}}\left\vert
\widehat{g}\left( J\right) \right\vert d\sigma\left( y\right) \\
& =\left( \frac{\ell\left( J\right) }{\ell\left( I^{\prime}\right)
}\right) ^{\left\vert \alpha\right\vert -1}\sqrt{\left\vert J\right\vert
_{\omega}}\left\vert \widehat{g}\left( J\right) \right\vert \left\{
\int_{I^{\prime}\setminus2J}\left( \frac{\ell\left( J\right) }{\ell\left(
J\right) +\operatorname*{dist}\left( y,J\right) }\right) ^{\kappa
-\left\vert \alpha\right\vert +1}\frac{1}{\left[ \ell\left( J\right)
+\operatorname*{dist}\left( y,J\right) \right] ^{n-\lambda-1}\ell\left(
I^{\prime}\right) }d\sigma\left( y\right) \right\} .
\end{align*}
Now we fix $t\in\mathbb{N}$, and estimate the sum of $\left\vert
\operatorname{Int}^{\lambda,\natural}\left( J\right) \right\vert $ over
those $J\subset I^{\prime}$ with $\ell\left( J\right) =2^{-t}\ell\left(
I^{\prime}\right) $ by splitting the integration in $y$ according to the size
of $\ell\left( J\right) +\operatorname*{dist}\left( y,J\right) $, to
obtain the following bound
\begin{align*}
& \sum_{\substack{J\subset I^{\prime}\\\ell\left( J\right) =2^{-t
\ell\left( I\right) }}\left\vert \operatorname{Int}^{\lambda,\natural
}\left( J\right) \right\vert \\
& \lesssim2^{-t\left( \left\vert \alpha\right\vert -1\right)
\sum_{\substack{J\subset I^{\prime}\\\ell\left( J\right) =2^{-t}\ell\left(
I\right) }}\sqrt{\left\vert J\right\vert _{\omega}}\left\vert \widehat
{g}\left( J\right) \right\vert \left\{ \int_{I^{\prime}\setminus2J}\left(
\frac{\ell\left( J\right) }{\ell\left( J\right) +\operatorname*{dist
\left( y,J\right) }\right) ^{\kappa-\left\vert \alpha\right\vert +1
\frac{d\sigma\left( y\right) }{\left[ \ell\left( J\right)
+\operatorname*{dist}\left( y,J\right) \right] ^{n-\lambda-1}\ell\left(
I^{\prime}\right) }\right\} \\
& \lesssim2^{-t\left( \left\vert \alpha\right\vert -1\right)
\sum_{J\subset I^{\prime}:\ \ell\left( J\right) =2^{-t}\ell\left( I\right)
}\sqrt{\left\vert J\right\vert _{\omega}}\left\vert \widehat{g}\left(
J\right) \right\vert \left\{ \sum_{s=1}^{t}\int_{2^{s+1}J\setminus2^{s
J}\left( 2^{-s}\right) ^{\kappa-\left\vert \alpha\right\vert +1
\frac{d\sigma\left( y\right) }{\left( 2^{s}\ell\left( J\right) \right)
^{n-\lambda-1}\ell\left( I^{\prime}\right) }\right\} \\
& \lesssim2^{-t\left\vert \alpha\right\vert }\sum_{J\subset I^{\prime
:\ \ell\left( J\right) =2^{-t}\ell\left( I\right) }\sqrt{\left\vert
J\right\vert _{\omega}}\left\vert \widehat{g}\left( J\right) \right\vert
\sum_{s=1}^{t}\left( 2^{-s}\right) ^{\kappa-\left\vert \alpha\right\vert
+1}2^{-s\left( n-\lambda-1\right) }\frac{\left\vert 2^{s}J\right\vert
_{\sigma}}{\ell\left( J\right) ^{n-\lambda}},
\end{align*}
which, upon pigeonholing the sum in $J$ according to membership in the
grandchildren of $I$ at depth $t-s$, gives:
\begin{align*}
& \sum_{J\in\mathfrak{C}_{\mathcal{D}}^{\left( t\right) }\left( I^{\prime
}\right) }\left\vert \operatorname{Int}^{\lambda,\natural}\left( J\right)
\right\vert \lesssim2^{-t\left\vert \alpha\right\vert }\sum_{J\in
\mathfrak{C}_{\mathcal{D}}^{\left( t\right) }\left( I^{\prime}\right)
}\sqrt{\left\vert J\right\vert _{\omega}}\left\vert \widehat{g}\left(
J\right) \right\vert \sum_{s=1}^{t}\left( 2^{-s}\right) ^{\kappa-\left\vert
\alpha\right\vert +n-\lambda}\frac{\left\vert 2^{s}J\right\vert _{\sigma
}{\ell\left( J\right) ^{n-\lambda}}\\
& =2^{-t\left\vert \alpha\right\vert }\sum_{s=1}^{t}\left( 2^{-s}\right)
^{\kappa-\left\vert \alpha\right\vert }\sum_{K\in\mathfrak{C}_{\mathcal{D
}^{\left( t-s\right) }\left( I^{\prime}\right) }\sum_{J\in\mathfrak{C
_{\mathcal{D}}^{\left( s\right) }\left( K\right) }\sqrt{\left\vert
J\right\vert _{\omega}}\left\vert \widehat{g}\left( J\right) \right\vert
\frac{\left\vert 2^{s}J\right\vert _{\sigma}}{\ell\left( K\right)
^{n-\lambda}}\\
& \lesssim2^{-t\left\vert \alpha\right\vert }\sum_{s=1}^{t}\left(
2^{-s}\right) ^{\kappa-\left\vert \alpha\right\vert }\sum_{K\in
\mathfrak{C}_{\mathcal{D}}^{\left( t-s\right) }\left( I^{\prime}\right)
}\frac{\left\vert 3K\right\vert _{\sigma}}{\ell\left( K\right) ^{n-\lambda
}\sum_{J\in\mathfrak{C}_{\mathcal{D}}^{\left( s\right) }\left( K\right)
}\sqrt{\left\vert J\right\vert _{\omega}}\left\vert \widehat{g}\left(
J\right) \right\vert \\
& \lesssim2^{-t\left\vert \alpha\right\vert }\sum_{s=1}^{t}\left(
2^{-s}\right) ^{\kappa-\left\vert \alpha\right\vert }\sum_{K\in
\mathfrak{C}_{\mathcal{D}}^{\left( t-s\right) }\left( I^{\prime}\right)
}\frac{\left\vert 3K\right\vert _{\sigma}}{\ell\left( K\right) ^{n-\lambda
}\sqrt{\left\vert K\right\vert _{\omega}}\sqrt{\sum_{J\in\mathfrak{C
_{\mathcal{D}}^{\left( s\right) }\left( K\right) }\left\vert \widehat
{g}\left( J\right) \right\vert ^{2}}\\
& \lesssim2^{-t\left\vert \alpha\right\vert }\sqrt{A_{2}^{\lambda}}\sum
_{s=1}^{t}\left( 2^{-s}\right) ^{\kappa-\left\vert \alpha\right\vert
\sum_{K\in\mathfrak{C}_{\mathcal{D}}^{\left( t-s\right) }\left( I^{\prime
}\right) }\sqrt{\left\vert K\right\vert _{\sigma}}\sqrt{\sum_{J\in
\mathfrak{C}_{\mathcal{D}}^{\left( s\right) }\left( K\right) }\left\vert
\widehat{g}\left( J\right) \right\vert ^{2}},
\end{align*}
where we used the $A_{2}^{\lambda}$ condition and doubling for $\sigma$ in the
last inequality. Thus we hav
\begin{align*}
\sum_{J\in\mathfrak{C}_{\mathcal{D}}^{\left( t\right) }\left( I^{\prime
}\right) }\left\vert \operatorname{Int}^{\lambda,\natural}\left( J\right)
\right\vert & \lesssim2^{-t\left\vert \alpha\right\vert }\sqrt
{A_{2}^{\lambda}}\sum_{s=1}^{t}\left( 2^{-s}\right) ^{\kappa-\left\vert
\alpha\right\vert }\sqrt{\left\vert I^{\prime}\right\vert _{\sigma}}\sqrt
{\sum_{J\in\mathfrak{C}_{\mathcal{D}}^{\left( t\right) }\left( I^{\prime
}\right) }\left\vert \widehat{g}\left( J\right) \right\vert ^{2}}\\
& \lesssim2^{-t}\sqrt{A_{2}^{\lambda}}\sqrt{\left\vert I^{\prime}\right\vert
_{\sigma}}\sqrt{\sum_{J\in\mathfrak{C}_{\mathcal{D}}^{\left( t\right)
}\left( I^{\prime}\right) }\left\vert \widehat{g}\left( J\right)
\right\vert ^{2}},
\end{align*}
since $1\leq\left\vert \alpha\right\vert \leq\kappa-1$.
We now claim the same estimate holds for the sum of $\left\vert
\operatorname{Int}^{\lambda,\flat}\left( J\right) \right\vert $ over
$J\subset I^{\prime}$ with $\ell\left( J\right) =2^{-t}\ell\left(
I^{\prime}\right) $. We write
\[
\operatorname{Int}^{\lambda,\flat}\left( J\right) =\sum_{k=1}^{n
{\displaystyle\sum\limits_{\left\vert \beta\right\vert +\left\vert
\gamma\right\vert =\left\vert \alpha\right\vert -1}}
c_{\alpha,\beta,\gamma}\operatorname{Int}_{k,\beta,\gamma}^{\lambda,\flat
}\left( J\right) ,
\]
and estimate
\begin{align*}
& \left\vert \operatorname{Int}_{k,\beta,\gamma}^{\lambda,\flat}\left(
J\right) \right\vert \lesssim\left\vert \int_{J}\left( \int_{2J}\Phi
_{k}^{\lambda}\left( x-y\right) \left( \frac{y-c_{J}}{\ell\left(
I^{\prime}\right) }\right) ^{\gamma}d\sigma\left( y\right) \right)
\left( \frac{x-c_{J}}{\ell\left( I^{\prime}\right) }\right) ^{\beta
}\bigtriangleup_{J;\kappa}^{\omega}g\left( x\right) d\omega\left( x\right)
\right\vert \\
& \leq\int_{J}\left( \int_{2J}\frac{1}{\ell\left( I\right) \left\vert
x-y\right\vert ^{n-\lambda-1}}\left\vert \frac{y-c_{J}}{\ell\left( I^{\prime
}\right) }\right\vert ^{\left\vert \gamma\right\vert }d\sigma\left(
y\right) \right) \left\vert \frac{x-c_{J}}{\ell\left( I^{\prime}\right)
}\right\vert ^{\left\vert \beta\right\vert }\left\vert \bigtriangleup
_{J;\kappa}^{\omega}g\left( x\right) \right\vert d\omega\left( x\right) \\
& \lesssim\left( \frac{\ell\left( J\right) }{\ell\left( I^{\prime
}\right) }\right) ^{\left\vert \gamma\right\vert +\left\vert \beta
\right\vert }\frac{1}{\ell\left( I^{\prime}\right) \sqrt{\left\vert
J\right\vert _{\omega}}}\left\vert \widehat{g}\left( J\right) \right\vert
\int_{J}\int_{2J}\frac{d\sigma\left( y\right) d\omega\left( x\right)
}{\left\vert x-y\right\vert ^{n-\lambda-1}}\lesssim\sqrt{A_{2}^{\lambda}
\frac{\ell\left( J\right) }{\ell\left( I^{\prime}\right) }\left\vert
\widehat{g}\left( J\right) \right\vert \sqrt{\left\vert J\right\vert
_{\sigma}},
\end{align*}
where in the last inequality we used that $\left\vert \beta\right\vert
+\left\vert \gamma\right\vert =\left\vert \alpha\right\vert \geq1$ and the
estimate
\[
\int_{J}\int_{2J}\frac{d\sigma\left( y\right) d\omega\left( x\right)
}{\left\vert x-y\right\vert ^{n-\lambda-1}}\lesssim\sqrt{A_{2}^{\lambda}
\ell\left( J\right) \sqrt{\left\vert J\right\vert _{\sigma}\left\vert
J\right\vert _{\omega}}.
\]
Indeed, in order to estimate the double integral using the $A_{2}^{\lambda}$
condition we cover the band $\left\vert x-y\right\vert \leq C2^{-m}\ell\left(
J\right) $ by a collection of cubes $Q\left( z_{m},C2^{-m}\ell\left(
J\right) \right) \times Q\left( z_{m},C2^{-m}\ell\left( J\right) \right)
$ in $CJ\times CJ$ with centers $\left( z_{m},z_{m}\right) $ and bounded
overlap. Then we hav
\begin{align*}
& \int_{J}\int_{2J}\frac{d\sigma\left( y\right) d\omega\left( x\right)
}{\left\vert x-y\right\vert ^{n-\lambda-1}}\leq\sum_{m=0}^{\infty
{\displaystyle\iint\limits_{\substack{x,y\in2J\\\left\vert x-y\right\vert
\approx2^{-m}\ell\left( J\right) }}}
\frac{d\sigma\left( y\right) d\omega\left( x\right) }{\left( 2^{-m
\ell\left( J\right) \right) ^{n-\lambda-1}}\\
& \approx\sum_{m=0}^{\infty}\sum_{Q\left( z_{m},C2^{-m}\ell\left( J\right)
\right) \times Q\left( z_{m},C2^{-m}\ell\left( J\right) \right)
\int_{Q\left( z_{m},C2^{-m}\ell\left( J\right) \right) \times Q\left(
z_{m},C2^{-m}\ell\left( J\right) \right) }\frac{d\sigma\left( y\right)
d\omega\left( x\right) }{\left( 2^{-m}\ell\left( J\right) \right)
^{n-\lambda-1}}\\
& \leq\frac{1}{\ell\left( J\right) ^{n-\lambda-1}}\sum_{m=0}^{\infty
}2^{m\left( n-\lambda-1\right) }\sum_{z_{m}}\left\vert Q\left(
z_{m},C2^{-m}\ell\left( J\right) \right) \right\vert _{\sigma}\left\vert
Q\left( z_{m},C2^{-m}\ell\left( J\right) \right) \right\vert _{\omega}\\
& \leq\frac{1}{\ell\left( J\right) ^{n-\lambda-1}}\sum_{m=0}^{\infty
}2^{m\left( n-\lambda-1\right) }\sum_{z_{m}}\sqrt{\left\vert Q\left(
z_{m},C2^{-m}\ell\left( J\right) \right) \right\vert _{\sigma}\left\vert
Q\left( z_{m},C2^{-m}\ell\left( J\right) \right) \right\vert _{\omega
}\sqrt{A_{2}^{\lambda}}\left( 2^{-m}\ell\left( J\right) \right)
^{n-\lambda}\\
& \lesssim\sqrt{A_{2}^{\lambda}}\ell\left( J\right) \sum_{m=0}^{\infty
}2^{m\left( n-\lambda-1\right) }2^{-m\left( n-\lambda\right)
\sqrt{\left\vert CJ\right\vert _{\sigma}\left\vert CJ\right\vert _{\omega
}\lesssim\sqrt{A_{2}^{\lambda}}\ell\left( J\right) \sqrt{\left\vert
J\right\vert _{\sigma}\left\vert J\right\vert _{\omega}}.
\end{align*}
No
\[
\sum_{J\in\mathfrak{C}_{\mathcal{D}}^{\left( t\right) }\left( I^{\prime
}\right) }\left\vert \operatorname{Int}^{\lambda,\flat}\left( J\right)
\right\vert \lesssim\sum_{J\in\mathfrak{C}_{\mathcal{D}}^{\left( t\right)
}\left( I^{\prime}\right) }\sqrt{A_{2}^{\lambda}}\frac{\ell\left( J\right)
}{\ell\left( I\right) }\left\vert \widehat{g}\left( J\right) \right\vert
\sqrt{\left\vert J\right\vert _{\sigma}}\leq2^{-t}\sqrt{A_{2}^{\lambda}
\sqrt{\left\vert I^{\prime}\right\vert _{\sigma}}\sqrt{\sum_{J\in
\mathfrak{C}_{\mathcal{D}}^{\left( t\right) }\left( I^{\prime}\right)
}\left\vert \widehat{g}\left( J\right) \right\vert ^{2}},
\]
and so altogether we hav
\begin{align*}
& \sum_{J\in\mathfrak{C}_{\mathcal{D}}^{\left( t\right) }\left( I^{\prime
}\right) }\left\vert \left\langle \left[ P_{\alpha,c_{J},I^{\prime
},T_{\sigma}^{\lambda}\right] \mathbf{1}_{I^{\prime}},\bigtriangleup
_{J;\kappa}^{\omega}g\right\rangle _{\omega}\right\vert =\sum_{J\in
\mathfrak{C}_{\mathcal{D}}^{\left( t\right) }\left( I^{\prime}\right)
}\left\vert \operatorname{Int}^{\lambda}\left( J\right) \right\vert \\
& \leq\sum_{J\in\mathfrak{C}_{\mathcal{D}}^{\left( t\right) }\left(
I^{\prime}\right) }\left\vert \operatorname{Int}^{\lambda,\natural}\left(
J\right) \right\vert +\sum_{J\in\mathfrak{C}_{\mathcal{D}}^{\left( t\right)
}\left( I^{\prime}\right) }\left\vert \operatorname{Int}^{\lambda,\flat
}\left( J\right) \right\vert \lesssim2^{-t}\sqrt{A_{2}^{\lambda}
\sqrt{\left\vert I^{\prime}\right\vert _{\sigma}}\sqrt{\sum_{J\in
\mathfrak{C}_{\mathcal{D}}^{\left( t\right) }\left( I^{\prime}\right)
}\left\vert \widehat{g}\left( J\right) \right\vert ^{2}}.
\end{align*}
Finally then we obtain from this and (\ref{we obtain})
\begin{align*}
\sum_{J\in\mathfrak{C}_{\mathcal{D}}^{\left( t\right) }\left( I^{\prime
}\right) }\left\vert \left\langle \left[ Q_{I^{\prime};\kappa},T_{\sigma
}^{\lambda}\right] \mathbf{1}_{I^{\prime}},\bigtriangleup_{J;\kappa}^{\omega
}g\right\rangle _{\omega}\right\vert & =\sum_{J\in\mathfrak{C}_{\mathcal{D
}^{\left( t\right) }\left( I^{\prime}\right) }\left\vert \sum_{\left\vert
\alpha\right\vert \leq\kappa-1}b_{\alpha}\left\langle \left[ P_{\alpha
,c_{J},I^{\prime}},T_{\sigma}^{\lambda}\right] \mathbf{1}_{I^{\prime
},\bigtriangleup_{J;\kappa}^{\omega}g\right\rangle _{\omega}\right\vert \\
& \lesssim2^{-t}\sqrt{A_{2}^{\lambda}}\sqrt{\sum_{J\in\mathfrak{C
_{\mathcal{D}}^{\left( t\right) }\left( I^{\prime}\right) }\left\vert
\widehat{g}\left( J\right) \right\vert ^{2}}.
\end{align*}
Now using $M_{I_{J};\kappa}=\left\vert \widehat{f}\left( I\right)
\right\vert Q_{I_{J};\kappa}$, and applying the above estimates with
$I^{\prime}=I_{J}$, we can sum over $t$ and $I\in\mathcal{C}_{F}$ to obtai
\begin{align}
& \left\vert \mathsf{B}_{\operatorname*{commutator};\kappa}^{F}\left(
f,g\right) \right\vert \leq\sum_{\substack{I\in\mathcal{C}_{F}\text{ and
}J\in\mathcal{C}_{F}^{\mathbf{\tau}-\operatorname*{shift}}\\J\Subset
_{\mathbf{\rho},\varepsilon}I}}\left\vert \widehat{f}\left( I\right)
\right\vert \left\vert \left\langle \left[ T_{\sigma}^{\lambda
,Q_{I_{J};\kappa}\right] \mathbf{1}_{I_{_{J}}},\bigtriangleup_{J;\kappa
}^{\omega}g\right\rangle _{\omega}\right\vert \label{comm est}\\
& \lesssim\sum_{t=\mathbf{r}}^{\infty}\sum_{I\in\mathcal{C}_{F}}2^{-t
\sqrt{A_{2}^{\lambda}}\left\vert \widehat{f}\left( I\right) \right\vert
\sqrt{\sum_{\substack{J\in\mathfrak{C}_{\mathcal{D}}^{\left( t\right)
}\left( I_{J}\right) \text{ }\\J\in\mathcal{C}_{F}^{\mathbf{\tau
}-\operatorname*{shift}}}}\left\vert \widehat{g}\left( J\right) \right\vert
^{2}}\lesssim\sqrt{A_{2}^{\lambda}}\sum_{t=\mathbf{r}}^{\infty}2^{-t
\sqrt{\sum_{I\in\mathcal{C}_{F}}\left\vert \widehat{f}\left( I\right)
\right\vert ^{2}}\sqrt{\sum_{I\in\mathcal{C}_{F}}\sum_{\substack{J\in
\mathfrak{C}_{\mathcal{D}}^{\left( t\right) }\left( I_{J}\right)
\\J\in\mathcal{C}_{F}^{\mathbf{\tau}-\operatorname*{shift}}}}\left\vert
\widehat{g}\left( J\right) \right\vert ^{2}}\nonumber\\
& \lesssim\sqrt{A_{2}^{\lambda}}\sum_{t=\mathbf{r}}^{\infty}2^{-t}\left\Vert
\mathsf{P}_{\mathcal{C}_{F}}^{\sigma}f\right\Vert _{L^{2}\left(
\sigma\right) }\left\Vert \mathsf{P}_{\mathcal{C}_{F}^{\mathbf{\tau
}-\operatorname*{shift}}}^{\omega}g\right\Vert _{L^{2}\left( \omega\right)
}\lesssim\sqrt{A_{2}^{\lambda}}\left\Vert \mathsf{P}_{\mathcal{C}_{F}
^{\sigma}f\right\Vert _{L^{2}\left( \sigma\right) }\left\Vert \mathsf{P
_{\mathcal{C}_{F}^{\mathbf{\tau}-\operatorname*{shift}}}^{\omega}g\right\Vert
_{L^{2}\left( \omega\right) }.\nonumber
\end{align}
\subsection{The neighbour form}
In this form we obtain the required bound by taking absolute values inside the
sum, and then arguing as in the case of Haar wavelets in \cite[end of
Subsection 8.4]{SaShUr7}. We begin with $M_{I^{\prime};\kappa}=\mathbf{1
_{I^{\prime}}\bigtriangleup_{I;\kappa}^{\sigma}f$ as in (\ref{def M}) to
obtain
\begin{align*}
\left\vert \mathsf{B}_{\operatorname*{neighbour};\kappa}^{F}\left(
f,g\right) \right\vert & \leq\sum_{\substack{I\in\mathcal{C}_{F}\text{ and
}J\in\mathcal{C}_{F}^{\mathbf{\tau}-\operatorname*{shift}}\\J\Subset
_{\mathbf{\rho},\varepsilon}I}}\sum_{\theta\left( I_{J}\right)
\in\mathfrak{C}_{\mathcal{D}}\left( I\right) \setminus\left\{
I_{J}\right\} }\left\vert \left\langle T_{\sigma}^{\lambda}\left(
\mathbf{1}_{\theta\left( I_{J}\right) }\bigtriangleup_{I;\kappa}^{\sigma
}f\right) ,\bigtriangleup_{J;\kappa}^{\omega}g\right\rangle _{\omega
}\right\vert \\
& \leq\sum_{\substack{I\in\mathcal{C}_{F}\text{ and }J\in\mathcal{C
_{F}^{\mathbf{\tau}-\operatorname*{shift}}\\J\Subset_{\mathbf{\rho
},\varepsilon}I}}\sum_{I^{\prime}\equiv\theta\left( I_{J}\right)
\in\mathfrak{C}_{\mathcal{D}}\left( I\right) \setminus\left\{
I_{J}\right\} }\left\vert \left\langle T_{\sigma}^{\lambda}\left(
M_{I^{\prime};\kappa}\mathbf{1}_{I^{\prime}}\right) ,\bigtriangleup
_{J;\kappa}^{\omega}g\right\rangle _{\omega}\right\vert
\end{align*}
Using the pivotal bound (\ref{piv bound}) on the inner product with
$\nu=\left\Vert M_{I^{\prime};\kappa}\right\Vert _{L^{\infty}\left(
\sigma\right) }\mathbf{1}_{I^{\prime}}d\sigma$, and then estimating by the
usual Poisson kernel
\begin{align*}
\left\vert \left\langle T_{\sigma}^{\lambda}\left( M_{I^{\prime};\kappa
}\mathbf{1}_{I^{\prime}}\right) ,\bigtriangleup_{J;\kappa}^{\omega
}g\right\rangle _{\omega}\right\vert & \lesssim\mathrm{P}_{\kappa}^{\lambda
}\left( J,\left\Vert M_{I^{\prime};\kappa}\right\Vert _{L^{\infty}\left(
\sigma\right) }\mathbf{1}_{I^{\prime}}\sigma\right) \sqrt{\left\vert
J\right\vert _{\omega}}\left\Vert \bigtriangleup_{J;\kappa}^{\omega
}g\right\Vert _{L^{2}\left( \omega\right) }\\
& \leq\left\Vert M_{I^{\prime};\kappa}\right\Vert _{L^{\infty}\left(
\sigma\right) }\mathrm{P}^{\lambda}\left( J,\mathbf{1}_{I^{\prime}
\sigma\right) \sqrt{\left\vert J\right\vert _{\omega}}\left\Vert
\bigtriangleup_{J;\kappa}^{\omega}g\right\Vert _{L^{2}\left( \omega\right)
},
\end{align*}
and the estimate $\left\Vert M_{I^{\prime};\kappa}\right\Vert _{L^{\infty
}\left( \sigma\right) }\approx\frac{1}{\sqrt{\left\vert I^{\prime
}\right\vert _{\sigma}}}\left\vert \widehat{f}\left( I\right) \right\vert $
from (\ref{analogue'}), to obtai
\begin{align*}
& \left\vert \mathsf{B}_{\operatorname*{neighbour};\kappa}^{F}\left(
f,g\right) \right\vert \lesssim\sum_{\substack{I\in\mathcal{C}_{F}\text{ and
}J\in\mathcal{C}_{F}^{\mathbf{\tau}-\operatorname*{shift}}\\J\Subset
_{\mathbf{\rho},\varepsilon}I}}\sum_{I^{\prime}\equiv\theta\left(
I_{J}\right) \in\mathfrak{C}_{\mathcal{D}}\left( I\right) \setminus\left\{
I_{J}\right\} }\frac{\left\vert \widehat{f}\left( I\right) \right\vert
}{\sqrt{\left\vert I^{\prime}\right\vert _{\sigma}}}\mathrm{P}^{\lambda
}\left( J,\mathbf{1}_{I^{\prime}}\sigma\right) \sqrt{\left\vert J\right\vert
_{\omega}}\left\Vert \bigtriangleup_{J;\kappa}^{\omega}g\right\Vert
_{L^{2}\left( \omega\right) }\\
& =\sum_{I\in\mathcal{C}_{F}}\sum_{\substack{I_{0},I_{\theta}\in
\mathfrak{C}_{\mathcal{D}}\left( I\right) \\I_{0}\neq I_{\theta}
}\sum_{\substack{J\in\mathcal{C}_{F}^{\mathbf{\tau}-\operatorname*{shift
}\\J\Subset_{\mathbf{\rho},\varepsilon}I\text{ and }J\subset I_{0}
}\frac{\left\vert \widehat{f}\left( I\right) \right\vert }{\sqrt{\left\vert
I_{\theta}\right\vert _{\sigma}}}\mathrm{P}^{\lambda}\left( J,\mathbf{1
_{I_{\theta}}\sigma\right) \sqrt{\left\vert J\right\vert _{\omega}}\left\Vert
\bigtriangleup_{J;\kappa}^{\omega}g\right\Vert _{L^{2}\left( \omega\right)
}\\
& =\sum_{s=\mathbf{r}}^{\infty}\sum_{I\in\mathcal{C}_{F}}\sum
_{\substack{I_{0},I_{\theta}\in\mathfrak{C}_{\mathcal{D}}\left( I\right)
\\I_{0}\neq I_{\theta}}}\sum_{\substack{J\in\mathcal{C}_{F}^{\mathbf{\tau
}-\operatorname*{shift}}\text{ and }\ell\left( J\right) =2^{-s}\ell\left(
I\right) \\J\Subset_{\mathbf{\rho},\varepsilon}I\text{ and }J\subset I_{0
}}\frac{\left\vert \widehat{f}\left( I\right) \right\vert }{\sqrt{\left\vert
I_{\theta}\right\vert _{\sigma}}}\mathrm{P}^{\lambda}\left( J,\mathbf{1
_{I_{\theta}}\sigma\right) \sqrt{\left\vert J\right\vert _{\omega}}\left\Vert
\bigtriangleup_{J;\kappa}^{\omega}g\right\Vert _{L^{2}\left( \omega\right)
}\\
& \lesssim\sum_{s=\mathbf{r}}^{\infty}\sum_{I\in\mathcal{C}_{F}
\sum_{\substack{I_{0},I_{\theta}\in\mathfrak{C}_{\mathcal{D}}\left( I\right)
\\I_{0}\neq I_{\theta}}}\sum_{\substack{J\in\mathcal{C}_{F}^{\mathbf{\tau
}-\operatorname*{shift}}\text{ and }\ell\left( J\right) =2^{-s}\ell\left(
I\right) \\J\Subset_{\mathbf{\rho},\varepsilon}I\text{ and }J\subset I_{0
}}\frac{\left\vert \widehat{f}\left( I\right) \right\vert }{\sqrt{\left\vert
I_{\theta}\right\vert _{\sigma}}}\left\{ \left( 2^{-s}\right)
^{1-\varepsilon\left( n+1-\lambda\right) }\mathrm{P}^{\lambda}\left(
I_{0},\mathbf{1}_{I_{\theta}}\sigma\right) \right\} \sqrt{\left\vert
J\right\vert _{\omega}}\left\Vert \bigtriangleup_{J;\kappa}^{\omega
}g\right\Vert _{L^{2}\left( \omega\right) }\\
& =\sum_{I\in\mathcal{C}_{F}}\sum_{\substack{I_{0},I_{\theta}\in
\mathfrak{C}_{\mathcal{D}}\left( I\right) \\I_{0}\neq I_{\theta}
}\sum_{s=\mathbf{r}}^{\infty}A\left( I,I_{0},I_{\theta},s\right) ,
\end{align*}
wher
\[
A\left( I,I_{0},I_{\theta},s\right) =\left( 2^{-s}\right) ^{1-\varepsilon
\left( n+1-\lambda\right) }\frac{\left\vert \widehat{f}\left( I\right)
\right\vert }{\sqrt{\left\vert I^{\prime}\right\vert _{\sigma}}
\mathrm{P}^{\lambda}\left( I_{0},\mathbf{1}_{I_{\theta}}\sigma\right)
\sum_{\substack{J\in\mathcal{C}_{F}^{\mathbf{\tau}-\operatorname*{shift
}\text{ and }\ell\left( J\right) =2^{-s}\ell\left( I\right) \\J\Subset
_{\mathbf{\rho},\varepsilon}I\text{ and }J\subset I_{0}}}\sqrt{\left\vert
J\right\vert _{\omega}}\left\Vert \bigtriangleup_{J;\kappa}^{\omega
}g\right\Vert _{L^{2}\left( \omega\right) }.
\]
Now recall the following estimate proved in \cite[bottom of page 120 to the
top of page 122]{SaShUr7}
\[
\left\vert \sum_{I\in\mathcal{C}_{F}}\sum_{\substack{I_{0},I_{\theta
\in\mathfrak{C}_{\mathcal{D}}\left( I\right) \\I_{0}\neq I_{\theta}
}\sum_{s=\mathbf{r}}^{\infty}A\left( I,I_{0},I_{\theta},s\right) \right\vert
\lesssim\sqrt{A_{2}^{\lambda}}\left\Vert \mathsf{P}_{\mathcal{C}_{F}}^{\sigma
}\right\Vert _{L^{2}\left( \sigma\right) }\left\Vert \mathsf{P
_{\mathcal{C}_{F}^{\mathbf{\tau}-\operatorname*{shift}}}^{\omega}\right\Vert
_{L^{2}\left( \omega\right) },
\]
where the quantity $A\left( I,I_{0},I_{\theta},s\right) $ was defined (in
our notation) b
\[
\left( 2^{-s}\right) ^{1-\varepsilon\left( n+1-\lambda\right) }\left\vert
E_{I_{\theta}}^{\sigma}\bigtriangleup_{I;1}^{\sigma}f\right\vert
\mathrm{P}^{\lambda}\left( I_{0},\mathbf{1}_{I_{\theta}}\sigma\right)
\sum_{\substack{J\in\mathcal{C}_{F}^{\mathbf{\tau}-\operatorname*{shift
}\text{ and }\ell\left( J\right) =2^{-s}\ell\left( I\right) \\J\Subset
_{\mathbf{\rho},\varepsilon}I\text{ and }J\subset I_{0}}}\sqrt{\left\vert
J\right\vert _{\omega}}\left\Vert \bigtriangleup_{J;\kappa}^{\omega
}g\right\Vert _{L^{2}\left( \omega\right) }.
\]
When $\sigma$ is doubling, the reader can check that $\left\vert E_{I_{\theta
}}^{\sigma}\bigtriangleup_{I;1}^{\sigma}f\right\vert \approx\frac{\left\vert
\widehat{f}\left( I\right) \right\vert }{\sqrt{\left\vert I_{\theta
}\right\vert _{\sigma}}}$, and that the proof in \cite[bottom of page 120 to
the top of page 122]{SaShUr7} applies verbatim to our situation. This proves
the required bound for the neighbour form
\begin{equation}
\left\vert \mathsf{B}_{\operatorname*{neighbour};\kappa}^{F}\left(
f,g\right) \right\vert \sqrt{A_{2}^{\lambda}}\left\Vert \mathsf{P
_{\mathcal{C}_{F}}^{\sigma}\right\Vert _{L^{2}\left( \sigma\right)
}\left\Vert \mathsf{P}_{\mathcal{C}_{F}^{\mathbf{\tau}-\operatorname*{shift}
}^{\omega}\right\Vert _{L^{2}\left( \omega\right) }. \label{neigh est
\end{equation}
\subsection{The stopping form}
To bound the stopping form, we will use the $\kappa$-pivotal condition
together with a variant of the Haar stopping form argument due to Nazarov,
Treil and Volberg \cite{NTV4}. Recall tha
\[
\left\vert \widehat{f}\left( I\right) \right\vert Q_{I^{\prime};\kappa
}=\mathbb{E}_{I^{\prime};\kappa}^{\sigma}f-\mathbf{1}_{I^{\prime}
\mathbb{E}_{I;\kappa}^{\sigma}f.
\]
We begin the proof by pigeonholing the ratio of side lengths of $I$ and $J$ in
the stopping form
\begin{align*}
\mathsf{B}_{\operatorname*{stop};\kappa}^{A}\left( f,g\right) & \equiv
\sum_{I\in\mathcal{C}_{F}}\sum_{I^{\prime}\in\mathfrak{C}_{\mathcal{D}}\left(
I\right) }\sum_{\substack{J\in\mathcal{C}_{F}^{\mathbf{\tau
-\operatorname*{shift}}\\J\subset I^{\prime}\text{ and }J\Subset
_{\mathbf{\rho},\varepsilon}I}}\left\vert \widehat{f}\left( I\right)
\right\vert \left\langle Q_{I^{\prime};\kappa}T_{\sigma}^{\lambda
\mathbf{1}_{F\setminus I^{\prime}},\bigtriangleup_{J;\kappa}^{\omega
}g\right\rangle _{\omega}\\
& =\sum_{s=0}^{\infty}\sum_{I\in\mathcal{C}_{F}}\sum_{I^{\prime
\in\mathfrak{C}_{\mathcal{D}}\left( I\right) }\sum_{\substack{J\in
\mathcal{C}_{F}^{\mathbf{\tau}-\operatorname*{shift}}\text{and }\ell\left(
J\right) =2^{-s}\ell\left( I\right) \\J\subset I^{\prime}\text{ and
}J\Subset_{\mathbf{\rho},\varepsilon}I}}\left\vert \widehat{f}\left(
I\right) \right\vert \left\langle Q_{I^{\prime};\kappa}T_{\sigma}^{\lambda
}\mathbf{1}_{F\setminus I^{\prime}},\bigtriangleup_{J;\kappa}^{\omega
}g\right\rangle _{\omega}\equiv\sum_{s=0}^{\infty}\mathsf{B
_{\operatorname*{stop};\kappa,s}^{A}\left( f,g\right) \ .
\end{align*}
By (\ref{piv bound}) with $R=\frac{Q_{I^{\prime};\kappa}}{\left\Vert
Q_{I^{\prime};\kappa}\right\Vert _{\infty}}$, and then using $\left\Vert
Q_{I^{\prime};\kappa}\right\Vert _{\infty}\lesssim\frac{1}{\sqrt{\left\vert
I^{\prime}\right\vert _{\sigma}}}$ and the Poisson inequality (\ref{e.Jsimeq
), we obtain
\begin{align*}
& \left\vert \mathsf{B}_{\operatorname*{stop};\kappa,s}^{A}\left(
f,g\right) \right\vert \lesssim\sum_{I\in\mathcal{C}_{F}}\sum_{I^{\prime
\in\mathfrak{C}_{\mathcal{D}}\left( I\right) }\sum_{\substack{J\in
\mathcal{C}_{F}^{\mathbf{\tau}-\operatorname*{shift}}\text{and }\ell\left(
J\right) =2^{-s}\ell\left( I\right) \\J\subset I^{\prime}\text{ and
}J\Subset_{\mathbf{\rho},\varepsilon}I}}\left\vert \widehat{f}\left(
I\right) \right\vert \left\Vert Q_{I^{\prime};\kappa}\right\Vert _{\infty
}\ \mathrm{P}_{\kappa}^{\lambda}\left( J,\mathbf{1}_{F\setminus I^{\prime
}\sigma\right) \sqrt{\left\vert J\right\vert _{\omega}}\left\Vert
\bigtriangleup_{J;\kappa}^{\omega}g\right\Vert _{L^{2}\left( \omega\right)
}\\
& \lesssim\sum_{I\in\mathcal{C}_{F}}\sum_{I^{\prime}\in\mathfrak{C
_{\mathcal{D}}\left( I\right) }\sum_{\substack{J\in\mathcal{C
_{F}^{\mathbf{\tau}-\operatorname*{shift}}\text{and }\ell\left( J\right)
=2^{-s}\ell\left( I\right) \\J\subset I^{\prime}\text{ and }J\Subset
_{\mathbf{\rho},\varepsilon}I}}\left\vert \widehat{f}\left( I\right)
\right\vert \frac{1}{\sqrt{\left\vert I^{\prime}\right\vert _{\sigma}
}\ \left( 2^{-s}\right) ^{\kappa-\varepsilon\left( n+\kappa-\lambda\right)
}\mathrm{P}_{\kappa}^{\lambda}\left( I,\mathbf{1}_{F\setminus I^{\prime
}\sigma\right) \sqrt{\left\vert J\right\vert _{\omega}}\left\Vert
\bigtriangleup_{J;\kappa}^{\omega}g\right\Vert _{L^{2}\left( \omega\right)
}\\
& \lesssim\left( 2^{-s}\right) ^{\kappa-\varepsilon\left( n+\kappa
-\lambda\right) }\sqrt{\sum_{I\in\mathcal{C}_{F}}\sum_{I^{\prime
\in\mathfrak{C}_{\mathcal{D}}\left( I\right) }\sum_{\substack{J\in
\mathcal{C}_{F}^{\mathbf{\tau}-\operatorname*{shift}}\text{and }\ell\left(
J\right) =2^{-s}\ell\left( I\right) \\J\subset I^{\prime}\text{ and
}J\Subset_{\mathbf{\rho},\varepsilon}I}}\left\vert \widehat{f}\left(
I\right) \right\vert ^{2}\frac{1}{\left\vert I^{\prime}\right\vert _{\sigma
}\ \mathrm{P}_{\kappa}^{\lambda}\left( I,\mathbf{1}_{F\setminus I^{\prime
}\sigma\right) ^{2}\left\vert J\right\vert _{\omega}}\\
&
\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \times
\sqrt{\sum_{I\in\mathcal{C}_{F}}\sum_{I^{\prime}\in\mathfrak{C}_{\mathcal{D
}\left( I\right) }\sum_{\substack{J\in\mathcal{C}_{F}^{\mathbf{\tau
}-\operatorname*{shift}}\text{and }\ell\left( J\right) =2^{-s}\ell\left(
I\right) \\J\subset I^{\prime}\text{ and }J\Subset_{\mathbf{\rho
,\varepsilon}I}}\left\Vert \bigtriangleup_{J}^{\omega}g\right\Vert
_{L^{2}\left( \omega\right) }^{2}}\ .
\end{align*}
Now we note tha
\[
\sum_{I\in\mathcal{C}_{F}}\sum_{I^{\prime}\in\mathfrak{C}_{\mathcal{D}}\left(
I\right) }\sum_{\substack{J\in\mathcal{C}_{F}^{\mathbf{\tau
-\operatorname*{shift}}\text{and }\ell\left( J\right) =2^{-s}\ell\left(
I\right) \\J\subset I^{\prime}\text{ and }J\Subset_{\mathbf{\rho
,\varepsilon}I}}\left\Vert \bigtriangleup_{J}^{\omega}g\right\Vert
_{L^{2}\left( \omega\right) }^{2}\lesssim\left\Vert \mathsf{P
_{\mathcal{C}_{F}^{\mathbf{\tau}-\operatorname*{shift}}}^{\omega}g\right\Vert
_{L^{2}\left( \omega\right) }^{2}\ ,
\]
and use the stopping control bound $\mathrm{P}_{\kappa}^{\lambda}\left(
I,\mathbf{1}_{F\setminus I^{\prime}}\sigma\right) ^{2}\left\vert J\right\vert
_{\omega}\leq\Gamma\left\vert J\right\vert _{\sigma}\leq2\mathcal{V
_{2}^{\alpha}\left( \sigma,\omega\right) \left\vert J\right\vert _{\sigma}$
in the corona $\mathcal{C}_{F}$, and $\left\Vert \mathsf{P}_{\mathcal{C}_{F
}^{\sigma}f\right\Vert _{L^{2}\left( \sigma\right) }^{2}=\sum_{I\in
\mathcal{C}_{F}}\left\vert \widehat{f}\left( I\right) \right\vert ^{2}$, to
obtai
\begin{align*}
\left\vert \mathsf{B}_{\operatorname*{stop};\kappa,s}^{A}\left( f,g\right)
\right\vert & \lesssim\left( 2^{-s}\right) ^{\kappa-\varepsilon\left(
n+\kappa-\lambda\right) }\mathcal{V}_{2}^{\alpha}\left( \sigma
,\omega\right) \sqrt{\sum_{I\in\mathcal{C}_{F}}\left\vert \widehat{f}\left(
I\right) \right\vert ^{2}\sum_{I^{\prime}\in\mathfrak{C}_{\mathcal{D}}\left(
I\right) }\sum_{\substack{J\in\mathcal{C}_{F}^{\mathbf{\tau
-\operatorname*{shift}}\text{and }\ell\left( J\right) =2^{-s}\ell\left(
I\right) \\J\subset I^{\prime}\text{ and }J\Subset_{\mathbf{\rho
,\varepsilon}I}}\frac{\left\vert J\right\vert _{\sigma}}{\left\vert I^{\prime
}\right\vert _{\sigma}}}\left\Vert \mathsf{P}_{\mathcal{C}_{F}^{\mathbf{\tau
}-\operatorname*{shift}}}^{\omega}g\right\Vert _{L^{2}\left( \omega\right)
}\\
& \lesssim\left( 2^{-s}\right) ^{\kappa-\varepsilon\left( n+\kappa
-\lambda\right) }\mathcal{V}_{2}^{\alpha}\left( \sigma,\omega\right)
\left\Vert \mathsf{P}_{\mathcal{C}_{F}}^{\sigma}f\right\Vert _{L^{2}\left(
\sigma\right) }\left\Vert \mathsf{P}_{\mathcal{C}_{F}^{\mathbf{\tau
}-\operatorname*{shift}}}^{\omega}g\right\Vert _{L^{2}\left( \omega\right)
}.
\end{align*}
Finally then we sum in $s$ to obtai
\begin{align}
& \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \left\vert \mathsf{B}_{\operatorname*{stop
;\kappa}^{A}\left( f,g\right) \right\vert \leq\mathcal{V}_{2}^{\alpha
}\left( \sigma,\omega\right) \sum_{s=0}^{\infty}\left\vert \mathsf{B
_{\operatorname*{stop};\kappa,s}^{A}\left( f,g\right) \right\vert
\label{stop est}\\
& \lesssim\mathcal{V}_{2}^{\alpha}\left( \sigma,\omega\right) \sum
_{s=0}^{\infty}\left( 2^{-s}\right) ^{\kappa-\varepsilon\left(
n+\kappa-\lambda\right) }\left\Vert \mathsf{P}_{\mathcal{C}_{F}}^{\sigma
}f\right\Vert _{L^{2}\left( \sigma\right) }\left\Vert \mathsf{P
_{\mathcal{C}_{F}^{\mathbf{\tau}-\operatorname*{shift}}}^{\omega}g\right\Vert
_{L^{2}\left( \omega\right) }\lesssim\mathcal{V}_{2}^{\alpha}\left(
\sigma,\omega\right) \left\Vert \mathsf{P}_{\mathcal{C}_{F}}^{\sigma
}f\right\Vert _{L^{2}\left( \sigma\right) }\left\Vert \mathsf{P
_{\mathcal{C}_{F}^{\mathbf{\tau}-\operatorname*{shift}}}^{\omega}g\right\Vert
_{L^{2}\left( \omega\right) },\nonumber
\end{align}
if we take $\varepsilon<\frac{\kappa}{n+\kappa-\lambda}$.
\section{Conclusion of the proofs}
Collecting all the estimates proved above, namely (\ref{routine}),
(\ref{top control}), (\ref{second far below}), (\ref{first far below}),
(\ref{unif bound'}), Lemma \ref{doub piv}, (\ref{diag est}), (\ref{para est}),
(\ref{comm est}), (\ref{neigh est}) and (\ref{stop est}), we obtain that for
any dyadic grid $\mathcal{D}$, and any admissible truncation of $T^{\alpha}$
\begin{align*}
\left\vert \left\langle T_{\sigma}^{\alpha}\mathsf{P}_{\operatorname{good
}^{\mathcal{D}}f,\mathsf{P}_{\operatorname{good}}^{\mathcal{D}}g\right\rangle
_{\omega}\right\vert & \leq C\left( \mathfrak{TR}_{T^{\alpha}}^{\left(
\kappa\right) }\left( \sigma,\omega\right) +\mathfrak{TR}_{T^{\alpha,\ast
}^{\left( \kappa\right) }\left( \omega,\sigma\right) +A_{2}^{\alpha
}\left( \sigma,\omega\right) +\varepsilon_{3}\mathfrak{N}_{T^{\alpha
}\left( \sigma,\omega\right) \right) \\
& \times\left\Vert \mathsf{P}_{\operatorname{good}}^{\mathcal{D}}f\right\Vert
_{L^{2}\left( \sigma\right) }\left\Vert \mathsf{P}_{\operatorname{good
}^{\mathcal{D}}g\right\Vert _{L^{2}\left( \omega\right) }.
\end{align*}
Thus for any admissible truncation of $T^{\alpha}$, using the above two
theorems, we obtain
\begin{align}
\mathfrak{N}_{T^{\alpha}}\left( \sigma,\omega\right) & \leq C\sup
_{\mathcal{D}}\frac{\left\vert \left\langle T_{\sigma}^{\alpha}\mathsf{P
_{\operatorname{good}}^{\mathcal{D}}f,\mathsf{P}_{\operatorname{good
}^{\mathcal{D}}g\right\rangle _{\omega}\right\vert }{\left\Vert \mathsf{P
_{\operatorname{good}}^{\mathcal{D}}f\right\Vert _{L^{2}\left( \sigma\right)
}\left\Vert \mathsf{P}_{\operatorname{good}}^{\mathcal{D}}g\right\Vert
_{L^{2}\left( \omega\right) }}\label{absorb}\\
& \leq C\left( \mathfrak{TR}_{T^{\alpha}}^{\left( \kappa\right) }\left(
\sigma,\omega\right) +\mathfrak{TR}_{T^{\alpha,\ast}}^{\left( \kappa\right)
}\left( \omega,\sigma\right) +A_{2}^{\alpha}\left( \sigma,\omega\right)
\right) +C\varepsilon_{3}\mathfrak{N}_{T^{\alpha}}\left( \sigma
,\omega\right) .\nonumber
\end{align}
Our next task is to use the doubling hypothesis to replace the triple $\kappa
$-testing constants by the usual cube testing constants, and we start with a lemma.
\begin{lemma}
Let $m\geq1$. Suppose that $\sigma$ and $\omega$ are locally finite positive
Borel measures on $\mathbb{R}^{n}$, with $\sigma$ doubling. If $T^{\alpha}$ is
a bounded operator from $L^{2}\left( \sigma\right) $ to $L^{2}\left(
\omega\right) $, then\ for every grandchild $I^{\prime}\in\mathfrak{C
_{\mathcal{D}}^{\left( m\right) }\left( I\right) $, and each
$0<\varepsilon_{1}<1$, there is a positive constant $C_{m,\varepsilon_{1}}$
such tha
\[
\sqrt{\int_{3I\setminus I^{\prime}}\left\vert T_{\sigma}^{\alpha
\mathbf{1}_{I^{\prime}}\right\vert ^{2}d\omega}\leq\left\{ C_{m,\varepsilon
_{1}}\sqrt{A_{2}^{\alpha}\left( \sigma,\omega\right) }+\varepsilon
_{1}\mathfrak{N}_{T^{\alpha}}\left( \sigma,\omega\right) \right\}
\sqrt{\left\vert I^{\prime}\right\vert _{\sigma}}\ .
\]
\end{lemma}
\begin{proof}
Given $0<\delta<1$, we split the left hand side as usual
\begin{align*}
\sqrt{\int_{3I\setminus I^{\prime}}\left\vert T_{\sigma}^{\alpha
\mathbf{1}_{I^{\prime}}\right\vert ^{2}d\omega} & \lesssim\sqrt
{\int_{3I\setminus I^{\prime}}\left\vert T_{\sigma}^{\alpha}\mathbf{1
_{\left( 1-\delta\right) I^{\prime}}\right\vert ^{2}d\omega}+\sqrt
{\int_{3I\setminus I^{\prime}}\left\vert T_{\sigma}^{\alpha}\mathbf{1
_{I^{\prime}\setminus\left( 1-\delta\right) I^{\prime}}\right\vert
^{2}d\omega}\\
& \equiv A+B.
\end{align*}
Using the halo estimate in \cite[Lemma 24 on page 24]{Saw6}, we obtai
\[
B\leq\mathfrak{N}_{T^{\alpha}}\left( \sigma,\omega\right) \sqrt{\left\vert
I^{\prime}\setminus\left( 1-\delta\right) I^{\prime}\right\vert _{\sigma
}\leq\sqrt{\frac{C}{\ln\frac{1}{\delta}}}\mathfrak{N}_{T^{\alpha}}\left(
\sigma,\omega\right) \sqrt{\left\vert I^{\prime}\right\vert _{\sigma}}.
\]
For term $A$ we hav
\begin{align*}
A & \lesssim\sqrt{\int_{3I\setminus I^{\prime}}\left\vert \int_{\left(
1-\delta\right) I^{\prime}}\left[ \delta\ell\left( I^{\prime}\right)
\right] ^{\alpha-n}d\sigma\right\vert ^{2}d\omega}=\sqrt{\left( 3\cdot
2^{m}\delta\right) ^{2\left( \alpha-n\right) }\frac{\left\vert \left(
1-\delta\right) I\right\vert _{\sigma}\left\vert 3I\setminus I^{\prime
}\right\vert _{\omega}}{\ell\left( 3I\right) ^{2\left( n-\alpha\right)
}\left\vert \left( 1-\delta\right) I\right\vert _{\sigma}}\\
& \lesssim\sqrt{\left( 2^{m}\delta\right) ^{2\left( \alpha-n\right)
}\frac{\left\vert 3I\right\vert _{\sigma}\left\vert 3I\right\vert _{\omega
}{\ell\left( 3I\right) ^{2\left( n-\alpha\right) }}\left\vert I\right\vert
_{\sigma}}\leq\left( 2^{m}\delta\right) ^{\alpha-n}A_{2}^{\alpha}\left(
\sigma,\omega\right) \sqrt{\left\vert I\right\vert _{\sigma}}.
\end{align*}
Now we choose $0<\delta<1$ so that $\sqrt{\frac{C}{\ln\frac{1}{\delta}}
\leq\varepsilon_{1}$.
\end{proof}
Recall that the triple $\kappa$-cube testing conditions use the $Q$-normalized
monomials $m_{Q}^{\beta}\left( x\right) \equiv\mathbf{1}_{Q}\left(
x\right) \left( \frac{x-c_{Q}}{\ell\left( Q\right) }\right) ^{\beta}$,
for which we have $\left\Vert m_{Q}^{\beta}\right\Vert _{L^{\infty}}\approx1$.
\begin{theorem}
\label{no tails}Suppose that $\sigma$ and $\omega$ are locally finite positive
Borel measures on $\mathbb{R}^{n}$, with $\sigma$ doubling, and let $\kappa
\in\mathbb{N}$. If $T^{\alpha}$ is a bounded operator from $L^{2}\left(
\sigma\right) $ to $L^{2}\left( \omega\right) $, then\ for every
$0<\varepsilon_{2}<1$, there is a positive constant $C\left( \kappa
,\varepsilon_{2}\right) $ such that
\[
\mathfrak{TR}_{T^{\alpha}}^{\left( \kappa\right) }\left( \sigma
,\omega\right) \leq C\left( \kappa,\varepsilon_{2}\right) \left[
\mathfrak{T}_{T^{\alpha}}\left( \sigma,\omega\right) +\sqrt{A_{2}^{\alpha
}\left( \sigma,\omega\right) }\right] +\varepsilon_{2}\mathfrak{N
_{T^{\alpha}}\left( \sigma,\omega\right) \ ,\ \ \ \ \ \kappa\geq1,
\]
and where the constants $C\left( \kappa,\varepsilon_{2}\right) $ depend only
on $\kappa$ and $\varepsilon$, and not on the operator norm $\mathfrak{N
_{T^{\alpha}}\left( \sigma,\omega\right) $.
\end{theorem}
\begin{proof}
Fix a dyadic cube $I$. If $P$ is an $I$-normalized polynomial of degree less
than $\kappa$ on the cube $I$, i.e. $\left\Vert P\right\Vert _{L^{\infty
}\approx1$, then we can approximate $P$ by a step function
\[
S\equiv\sum_{I^{\prime}\in\mathfrak{C}_{\mathcal{D}}^{\left( m\right)
}\left( I\right) }a_{I^{\prime}}\mathbf{1}_{I^{\prime}},
\]
satisfyin
\[
\left\Vert S-\mathbf{1}_{I}P\right\Vert _{L^{\infty}\left( \sigma\right)
}<\frac{\varepsilon_{2}}{2}\ ,
\]
provided we take $m\geq1$ sufficiently large depending on $n$ and $\kappa$,
but independent of the cube $I$. Then using the above lemma with $C2^{\frac
{m}{2}}\varepsilon_{1}\leq\frac{\varepsilon_{2}}{2}$, and the estimate
$\left\vert a_{I^{\prime}}\right\vert \lesssim\left\Vert P\right\Vert
_{L^{\infty}}\lesssim1$, we hav
\begin{align*}
& \sqrt{\int_{3I}\left\vert T_{\sigma}^{\alpha}\mathbf{1}_{I}P\right\vert
^{2}d\omega}\leq\sqrt{\int_{3I}\left\vert \sum_{I^{\prime}\in\mathfrak{C
_{\mathcal{D}}^{\left( m\right) }\left( I\right) }a_{I^{\prime}}T_{\sigma
}^{\alpha}\mathbf{1}_{I^{\prime}}\right\vert ^{2}d\omega}+\sqrt{\int
_{3I}\left\vert T_{\sigma}^{\alpha}\left[ \left( S-P\right) \mathbf{1
_{I}\right] \right\vert ^{2}d\omega}\\
& \leq C\sum_{I^{\prime}\in\mathfrak{C}_{\mathcal{D}}^{\left( m\right)
}\left( I\right) }\sqrt{\int_{3I\setminus I^{\prime}}\left\vert
a_{I^{\prime}}T_{\sigma}^{\alpha}\mathbf{1}_{I^{\prime}}\right\vert
^{2}d\omega}+C\sum_{I^{\prime}\in\mathfrak{C}_{\mathcal{D}}^{\left( m\right)
}\left( Q\right) }\sqrt{\int_{I^{\prime}}\left\vert a_{I^{\prime}}T_{\sigma
}^{\alpha}\mathbf{1}_{I^{\prime}}\right\vert ^{2}d\omega}+\frac{\varepsilon
_{2}}{2}\mathfrak{N}_{T^{\alpha}}\left( \sigma,\omega\right) \sqrt
{\left\vert I\right\vert _{\sigma}}\\
& \leq C\sum_{I^{\prime}\in\mathfrak{C}_{\mathcal{D}}^{\left( m\right)
}\left( Q\right) }\left\{ C_{m,\varepsilon_{1}}\sqrt{A_{2}^{\alpha}\left(
\sigma,\omega\right) }+\varepsilon_{1}\mathfrak{N}_{T^{\alpha}}\left(
\sigma,\omega\right) +\mathfrak{T}_{T^{\alpha}}\left( \sigma,\omega\right)
\right\} \sqrt{\left\vert I^{\prime}\right\vert _{\sigma}}+\frac
{\varepsilon_{2}}{2}\mathfrak{N}_{T^{\alpha}}\left( \sigma,\omega\right)
\sqrt{\left\vert I\right\vert _{\sigma}}\\
& \leq C\left\{ \mathfrak{T}_{T^{\alpha}}\left( \sigma,\omega\right)
+\sqrt{A_{2}^{\alpha}\left( \sigma,\omega\right) }\right\} \sqrt{\left\vert
I\right\vert _{\sigma}}+\left( C2^{\frac{m}{2}}\varepsilon_{1}+\frac
{\varepsilon_{2}}{2}\right) \mathfrak{N}_{T^{\alpha}}\left( \sigma
,\omega\right) \sqrt{\left\vert I\right\vert _{\sigma}}.
\end{align*}
\end{proof}
Combining this with (\ref{absorb}) we obtai
\[
\mathfrak{N}_{T^{\alpha}}\left( \sigma,\omega\right) \leq C\left(
\mathfrak{T}_{T^{\alpha}}\left( \sigma,\omega\right) +\mathfrak{T
_{T^{\alpha,\ast}}\left( \omega,\sigma\right) +\sqrt{A_{2}^{\alpha}\left(
\sigma,\omega\right) }\right) +C\left( \varepsilon_{2}+\varepsilon
_{3}\right) \mathfrak{N}_{T^{\alpha}}\left( \sigma,\omega\right) .
\]
Since $\mathfrak{N}_{T^{\alpha}}\left( \sigma,\omega\right) <\infty$ for
each truncation, we may absorb the final summand on the right into the left
hand side provided $C\left( \varepsilon_{2}+\varepsilon_{3}\right) <\frac
{1}{2}$, to obtai
\[
\mathfrak{N}_{T^{\alpha}}\left( \sigma,\omega\right) \lesssim\mathfrak{T
_{T^{\alpha}}\left( \sigma,\omega\right) +\mathfrak{T}_{T^{\alpha,\ast
}\left( \omega,\sigma\right) +\sqrt{A_{2}^{\alpha}\left( \sigma
,\omega\right) }.
\]
By the definition of boundedness of $T^{\alpha}$ in Definition
(\ref{def bounded}), this completes the proof of the first assertion in
Theorem \ref{main}. The second assertion in Theorem \ref{main} follows
immediately from the following lemma.
\begin{lemma}
If $K^{\alpha}$ is elliptic in the sense of Stein, then
\[
\sqrt{A_{2}^{\alpha}\left( \sigma,\omega\right) }\lesssim\mathfrak{N
_{T^{\alpha}}\left( \sigma,\omega\right) .
\]
\end{lemma}
\begin{proof}
The argument in \cite[see the proof of Proposition 7 page 210]{Ste2} shows
that there is $\varepsilon>0$, depending only on the constant $C_{CZ}$ in
(\ref{sizeandsmoothness'}), such that
\[
\frac{\left\vert Q\right\vert _{\omega}\left\vert Q^{\prime}\right\vert
_{\sigma}}{\ell\left( Q\right) ^{2\left( n-\alpha\right) }}\lesssim
\mathfrak{N}_{T^{\alpha}
\]
when $\operatorname*{dist}\left( Q,Q^{\prime}\right) \approx\ell\left(
Q\right) =\ell\left( Q^{\prime}\right) $, and $\left\vert \frac
{x-y}{\left\vert x-y\right\vert }-\mathbf{u}_{0}\right\vert <\varepsilon$
whenever $x\in Q$ and $y\in Q^{\prime}$. Now given a cube $Q_{0}$, we choose
subcubes $Q$ and $Q^{\prime}$ satisfying the previous conditions to obtai
\[
\frac{\left\vert Q_{0}\right\vert _{\omega}\left\vert Q_{0}\right\vert
_{\sigma}}{\ell\left( Q_{0}\right) ^{2\left( n-\alpha\right) }
\lesssim\frac{\left\vert Q\right\vert _{\omega}\left\vert Q^{\prime
}\right\vert _{\sigma}}{\ell\left( Q\right) ^{2\left( n-\alpha\right)
}\lesssim\mathfrak{N}_{T^{\alpha}}\ ,
\]
upon using that both $\sigma$ and $\omega$ are doubling.
\end{proof}
Thus in the presence of the classical Muckenhoupt condition, we have shown the
equivalence of $\left( a\right) $ and $\left( b\right) $ in Theorem
\ref{main}. The equivalence of these conditions with $\left( c\right) $ and
$\left( d\right) $ now follows from \cite[Section 9]{Saw6}, where it is
shown that $\left( c\right) \Longrightarrow\left( d\right) \Longrightarrow
\left( b\right) $ and $\left( a\right) \Longrightarrow\left( c\right) $
for doubling measures in the presence of the classical Muckenhoupt condition.
\section{A $T_{\operatorname*{translate}}$ theorem for doubling measures}
\label{T translate section}
The arguments used above rely crucially on the nested property of dyadic
cubes, and break down completely even for balls. Here we instead use a fairly
elementary argument to show that testing over indicators of cubes
$\mathbf{1}_{Q}$ in the doubling theorem above can be replaced by testing over
indicators of balls $\mathbf{1}_{B}$. In fact, we can replace cubes or balls
by translates and dilates of any fixed bounded set having positive Lebesgue
measure. The theorem below extends this to translates of fixed bounded
functions $b_{k},b_{k}^{\ast}$ with integral $1$ at each length scale $2^{k}$,
which we refer to as a \textquotedblleft$T_{\operatorname*{translate}}$
theorem\textquotedblright\ \emph{because at any given scale, we test a
function \textquotedblleft b}$_{k}$\emph{\textquotedblright\ and all of its
translates at that scale.}
More precisely, suppose that for each $t\in\left( 0,\infty\right) $, we are
given a bounded complex-valued function $B_{t}$ on $\mathbb{R}^{n}$ satisfying
\begin{enumerate}
\item $\operatorname*{Supp}B_{t}\subset Q_{t}\equiv\left[ -\frac{t}{2
,\frac{t}{2}\right) $,
\item $\left\vert B_{t}\left( x\right) \right\vert \leq\frac{C}{\left\vert
Q_{t}\right\vert }$,
\item $\int B_{t}\left( x\right) dx=1$.
\end{enumerate}
\begin{definition}
\label{translate testing}Set $\mathcal{B}\left( x,t\right) \equiv
B_{t}\left( x\right) $ where the functions $B_{t}$ are as above, and let
$b_{Q}\left( x\right) \equiv\left\vert Q\right\vert B_{\ell\left( Q\right)
}\left( x-c_{Q}\right) $ be a translation and normalization of
$B_{\ell\left( Q\right) }$ (so that it satisfies estimates similar to
$\mathbf{1}_{Q}$). For $\sigma$ and $\omega$ locally finite positive Borel
measures on $\mathbb{R}^{n}$, define the $\mathcal{B}$\emph{-testing constant}
and the $\delta$\emph{-full }$\mathcal{B}$\emph{-testing constant} for the
operator $T^{\alpha}$ b
\begin{align*}
\mathfrak{T}_{T^{\alpha}}^{\mathcal{B}}\left( \sigma,\omega\right) &
\equiv\sup_{Q}\frac{\sqrt{\int_{Q}\left\vert T_{\sigma}^{\alpha}b_{Q}\left(
x\right) \right\vert ^{2}d\omega\left( x\right) }}{\sqrt{\int_{Q}\left\vert
b_{Q}\left( y\right) \right\vert ^{2}d\sigma\left( y\right) }},\\
\mathfrak{F}_{\delta}\mathfrak{T}_{T^{\alpha}}^{\mathcal{B}}\left(
\sigma,\omega\right) & \equiv\sup_{Q}\frac{\sqrt{\int_{\frac{2}{\delta
Q}\left\vert T_{\sigma}^{\alpha}b_{Q}\left( x\right) \right\vert ^{2
d\omega\left( x\right) }}{\sqrt{\int_{Q}\left\vert b_{Q}\left( y\right)
\right\vert ^{2}d\sigma\left( y\right) }}.
\end{align*}
\end{definition}
If we take $B_{t}(x)=\frac{1}{t^{n}}\mathbf{1}_{B\left( 0,t\right) }\left(
x\right) $, then the $\mathcal{B}$-testing constant is simply the
ball-testing constant.
Fix $0<\delta<1$. Given a cube $Q\in\mathcal{P}$, define the $\delta
$-convolutio
\begin{align*}
B_{Q,\delta}\left( y\right) & \equiv\mathbf{1}_{Q}\ast B_{\delta
\ell\left( Q\right) }\left( y\right) =\int_{\mathbb{R}^{n}}\mathbf{1
_{Q}\left( y-z\right) B_{\delta\ell\left( Q\right) }\left( z\right) dz\\
& =\int_{Q}B_{\delta\ell\left( Q\right) }\left( y-z\right) dz=\int
_{Q}\tau_{z}B_{\delta\ell\left( Q\right) }\left( y\right)
dz,\ \ \ \ \ x\in\mathbb{R}^{n}.
\end{align*}
Now we comput
\begin{align*}
& \sqrt{\int_{Q}\left\vert T_{\sigma}^{\alpha}\mathbf{1}_{Q}\left( x\right)
\right\vert ^{2}d\omega\left( x\right) }\leq\sqrt{\int_{Q}\left\vert
T_{\sigma}^{\alpha}B_{Q,\delta}\left( x\right) \right\vert ^{2
d\omega\left( x\right) }+\sqrt{\int_{Q}\left\vert T_{\sigma}^{\alpha}\left(
\mathbf{1}_{Q}-B_{Q,\delta}\right) \left( x\right) \right\vert ^{2
d\omega\left( x\right) }\\
& \leq\sqrt{\int_{Q}\left\vert T_{\sigma}^{\alpha}\left( B_{Q,\delta}\left(
y\right) \right) \left( x\right) \right\vert ^{2}d\omega\left( x\right)
}+\mathfrak{N}_{T^{\alpha}}\sqrt{\int_{Q}\left\vert \left( \mathbf{1
_{Q}-B_{Q,\delta}\right) \left( x\right) \right\vert ^{2}d\sigma\left(
x\right) },
\end{align*}
wher
\begin{align*}
& \sqrt{\int_{Q}\left\vert T_{\sigma}^{\alpha}\left( B_{Q,\delta}\left(
y\right) \right) \left( x\right) \right\vert ^{2}d\omega\left( x\right)
}=\left\vert Q\right\vert \sqrt{\int_{Q}\left\vert T_{\sigma}^{\alpha}\left(
\int_{Q}\tau_{z}B_{\delta\ell\left( Q\right) }\left( y\right) \frac
{dz}{\left\vert Q\right\vert }\right) \left( x\right) \right\vert
^{2}d\omega\left( x\right) }\\
& \leq\left\vert Q\right\vert \int_{Q}\sqrt{\int_{Q}\left\vert T_{\sigma
}^{\alpha}\left( \tau_{z}B_{\delta\ell\left( Q\right) }\right) \left(
x\right) \right\vert ^{2}d\omega\left( x\right) }\frac{dz}{\left\vert
Q\right\vert }\\
& \leq\int_{Q}\mathfrak{F}_{\delta}\mathfrak{T}_{T^{\alpha}}^{B_{\delta
\ell\left( Q\right) }}\sqrt{\int_{Q}\left\vert \left( \tau_{z}B_{\delta
\ell\left( Q\right) }\right) \left( y\right) \right\vert ^{2
d\sigma\left( y\right) }dz\\
& \leq\mathfrak{F}_{\delta}\mathfrak{T}_{T^{\alpha}}^{B_{\delta\ell\left(
Q\right) }}\frac{C\left\vert Q\right\vert }{\left\vert Q_{\delta\ell\left(
Q\right) }\right\vert }\sqrt{\left\vert Q\right\vert _{\sigma}}\leq
C_{\delta}\mathfrak{F}_{\delta}\mathfrak{T}_{T^{\alpha}}^{B_{\delta\ell\left(
Q\right) }}\sqrt{\left\vert Q\right\vert _{\sigma}},
\end{align*}
an
\[
\sqrt{\int_{Q}\left\vert \left( \mathbf{1}_{Q}-B_{Q,\delta}\right) \left(
x\right) \right\vert ^{2}d\sigma\left( x\right) }\leq C\sqrt{\frac{1
{\ln\frac{1}{\delta}}\left\vert Q\right\vert _{\sigma}}.
\]
Thus altogether we obtai
\begin{equation}
\mathfrak{T}_{T^{\alpha}}=\sup_{Q}\frac{\sqrt{\int_{Q}\left\vert T_{\sigma
}^{\alpha}\mathbf{1}_{Q}\left( x\right) \right\vert ^{2}d\omega\left(
x\right) }}{\sqrt{\left\vert Q\right\vert _{\sigma}}}\leq C_{\delta
}\mathfrak{F}_{\delta}\mathfrak{T}_{T^{\alpha}}^{B_{\delta\ell\left(
Q\right) }}+C\sqrt{\frac{1}{\ln\frac{1}{\delta}}}\mathfrak{N}_{T^{\alpha}}\ ,
\label{alt
\end{equation}
and now using $A_{2}^{\alpha}$ to control $\mathfrak{F}_{\delta
\mathfrak{T}_{T^{\alpha}}^{B_{\delta\ell\left( Q\right) }}$ by
$\mathfrak{T}_{T^{\alpha}}^{B_{\delta\ell\left( Q\right) }}$, we have shown
that a $T1$ theorem for $T^{\alpha}$ over cubes follows from a $Tb$ theorem
where the functions $b_{Q}$ are translates of the fixed function
$B_{\ell\left( Q\right) }$. In particular we can take $B_{t}=\frac{1}{t^{n
}\mathbf{1}_{B\left( 0,t\right) }\left( x\right) $ to get ball-testing
implies cube-testing.
More precisely, we have the following theorem.
\begin{theorem}
\label{trans}Let $\sigma$ and $\omega$ be doubling measures on $\mathbb{R
^{n}$. Then with $T^{\alpha}$ and $\mathcal{B}$ as above we have bot
\begin{align}
\mathfrak{F}_{\delta}\mathfrak{T}_{T^{\alpha}}^{\mathcal{B}}\left(
\sigma,\omega\right) & \leq C_{\delta}\mathfrak{T}_{T^{\alpha
}^{\mathcal{B}}\left( \sigma,\omega\right) +C_{\delta}\sqrt{A_{2}^{\alpha
}\left( \sigma,\omega\right) }+C\sqrt{\frac{1}{\ln\frac{1}{\delta}
}\mathfrak{N}_{T^{\alpha}}\left( \sigma,\omega\right) ,\label{both}\\
\mathfrak{T}_{T^{\alpha}}\left( \sigma,\omega\right) & \leq C_{\delta
}\mathfrak{F}_{\delta}\mathfrak{T}_{T^{\alpha}}^{\mathcal{B}}\left(
\sigma,\omega\right) +C\sqrt{\frac{1}{\ln\frac{1}{\delta}}}\mathfrak{N
_{T^{\alpha}}\left( \sigma,\omega\right) .\nonumber
\end{align}
Altogether then we hav
\[
\mathfrak{T}_{T^{\alpha}}\left( \sigma,\omega\right) \leq C_{\delta
}\mathfrak{T}_{T^{\alpha}}^{\mathcal{B}}\left( \sigma,\omega\right)
+C_{\delta}\sqrt{A_{2}^{\alpha}\left( \sigma,\omega\right) }+C\sqrt{\frac
{1}{\ln\frac{1}{\delta}}}\mathfrak{N}_{T^{\alpha}}\left( \sigma
,\omega\right) ,
\]
and hence
\[
\mathfrak{N}_{T^{\alpha}}\left( \sigma,\omega\right) \leq C\left(
\mathfrak{T}_{T^{\alpha}}^{\mathcal{B}}+\mathfrak{T}_{\left( T^{\alpha
}\right) ^{\ast}}^{\mathcal{B}}+\sqrt{A_{2}^{\alpha}}\right) .
\]
\begin{proof}
The bound for $\mathfrak{F}_{\delta}\mathfrak{T}_{T^{\alpha}}^{\mathcal{B
}\left( \sigma,\omega\right) $ in the first line of (\ref{both}) is in
Theorem \ref{no tails}, and the bound for $\mathfrak{T}_{T^{\alpha}}\left(
\sigma,\omega\right) $ in the second line of (\ref{both}) is in (\ref{alt}).
An absorption argument proves the final display.
\end{proof}
\end{theorem}
Note that it follows from this theorem that the cubical annuli in the
cancellation conditions can be replaced with the classical spherical annuli.
Indeed, one simply mimics the proof for the cubical case.
\begin{remark}
The only property of cubes used in Theorem \ref{trans}, was that their
boundaries are not charged by doubling measures (the specific decay
$\sqrt{\frac{1}{\ln\frac{1}{\delta}}}$ of the halo can be replaced by any
decay to zero). Thus we can replace cube testing $\mathfrak{T}_{T^{\alpha
}\left( \sigma,\omega\right) $ in Theorem \ref{trans} by ball testing, or
any other `shape testing' in which the boundary of the `shape' is not charged
by doubling measures.
\end{remark}
|
\section*{Abstract}
{\bf
One of the features of many-body quantum systems with Hilbert-space fragmentation are stationary states manifesting quantum jamming. It was recently shown that these are ``states with memory'', in which, e.g., measuring a localised observable has everlasting macroscopic effects. We study such a measurement catastrophe with an example that stands out for its clarity. We show in particular that at late times the expectation value of a charge density becomes a nontrivial function of the ratio between distance and time notwithstanding the corresponding current approaching zero.
}
\vspace{10pt}
\noindent\rule{\textwidth}{1pt}
\tableofcontents\thispagestyle{fancy}
\noindent\rule{\textwidth}{1pt}
\vspace{10pt}
\section{Introduction}
Complexity is behind the most intriguing phenomena observed in many-body systems: phases of matter, spontaneous symmetry breaking, relaxation, etc.
On the other hand, our personal understanding of a phenomenon is almost inevitably reductive, being based on simplifications and analogies.
Undressing a problem of the complicated details that do not affect the behaviour under investigation is one of the most important steps towards the development of a decent physical picture. Just as the framework of renormalisation group allows one to identify what is relevant at a phase transition, so the conception of a simple model
helps identify what is essential.
Exactly solvable models are arguably the main source of procedural knowledge in physics: by solving a problem exactly, we can grasp the significance of what could be seen in nature.
These models, which are generally integrable, are clearly special. As such, they exhibit exceptional features that could be absent in real systems. A phenomenon identified in an exactly solvable model is nevertheless always potentially observable, if not exactly like predicted in the model, as a somehow related phenomenon. Pre-thermalisation is a remarkable example of this kind: integrable systems are known to relax differently than generic ones~\cite{Kinoshita2006,rigol2007,Cazalilla2006,Eisert2015,Polkovnikov2011}, and yet such special, atypical dynamics leave a mark also on the relaxation of generic systems~\cite{Bertini2015}, which can exhibit pre-thermalisation plateaux~\cite{Langen2016}.
In fact, even working in the framework of integrability can present formidable complications that undermine our qualitative understanding. This is one of the motivations behind the recent interest in identifying the ``simplest'' interacting integrable systems. In this respect, a spin-$1/2$ chain model with a three-site interaction termed the ``dual folded XXZ model'' was recently highlighted~\cite{folded:1,folded:2,pozsgay2021}. It corresponds to a special point of a rather complicated integrable system called Bariev model~\cite{Bariev1991,Bariev1992}, and is equivalent to the celebrated Heisenberg XXZ model in the limit of large anisotropy. The dual folded XXZ model was recently solved by two of us from scratch by means of a coordinate Bethe Ansatz that has been adapted to the special symmetries emerging in the large-anisotropy limit of the XXZ model~\cite{folded:1}. In particular, the Hilbert space splits into an exponentially large number of sectors with a given configuration of particles, each particle being associated with a spin up in either an even or an odd position. Similar structures arise in what was recently dubbed ``Hilbert-space fragmentation''~\cite{sala2020,khemani2020,moudgalya2021}.
In each sector the model is almost free, with a diagonal scattering matrix in which only one diagonal element is different from $-1$ and equal to $-e^{i q}$, where $q$ is the difference between the momenta of the scattering particles, in a remarkable analogy with the scattering phase modification factor in hard-rod deformed models (see, e.g., Refs~\cite{pozsgay2021hardrod,cardy2021toverline}). A striking feature found in fragmented models, such as the folded XXZ one, is the existence of an exponentially large sector consisting of jammed states: particles are stuck and can not move~\cite{moudgalyaKrylov2021,HSF,Turner2018,katja2020,bastianello2021fragmentation,Menon1997}. In Ref.~\cite{measurement2021} it was shown that such states have memory: the effect of a localised perturbation does not fade away but rather becomes visible at macroscopic scales for arbitrarily long times.
This is similar to the memory effects observed when perturbing ground states with spontaneously broken symmetries~\cite{Zauner2015,Eisler2020} (see also Refs~\cite{Eisler2020,Eisler2016,Gruber2021,Eisler2021} for a perturbation local in the fermionic basis), where however the dynamics are effectively constrained in a finite-dimensional space. In the setting of Ref.~\cite{measurement2021}, instead, an exponentially large number of states are involved.
Considering how dramatic the change in the state is, it seems appropriate to refer to it as a ``measurement catastrophe''. In particular, this effect was reported in a class of states that allow for an exact analytic analysis but are, however, not simple enough to make us gain a clear understanding of the phenomenon. This paper is first of all a solution to that problem.
We consider a family of jammed states for which it is possible to write down the time evolving state after a spin flip in a simple, elegant way. We then focus on observables whose expectation values can be predicted with minimal information about the initial state. Among them there will be the densities of two conserved charges: the magnetisation and the staggered magnetisation. They will be our key monitor of the measurement catastrophe: indeed, they turn out to be affected by the original spin flip in a macroscopic way. This sets the example herein apart from the one considered in Ref.~\cite{measurement2021}, where instead, in the long time limit, charge densities and currents become independent of their positions. The central part of the paper will be devoted to devising a physical picture that could reconcile the nontrivial profile of the (staggered) magnetisation at the Euler/ballistic scale with the fact that the corresponding currents vanish in the limit of infinite time. First we will present a set of rules that allow us to predict the asymptotic behaviour of any spin
solely from the knowledge of the initial configuration of few neighbouring spins.
Using the rules we will then show that the asymptotic state of the system is locally jammed and that the currents of the magnetisation and the staggered magnetisation, as well as all other conventional charges of the folded XXZ model, vanish in it. By computing the statistics of the particles' positions, we will show that the macroscopic reorganisation of the spin configuration coincides with small correlated microscopic fluctuations of particles (spins up) that do not exceed two lattice sites. Finally, we will focus on a special class of weakly interacting initial jammed states, which allow for the preparation of correlated spin pairs in an antiferromagnetic background. We will study, in particular, the entanglement between spins.
\subsection{Overview of the model}
The dual folded XXZ model describes the time evolution of states in the large-anisotropy limit of the Heisenberg model. It is obtained in the ``folded picture'', which was defined in Ref.~\cite{folded:1} as a representation of time evolution with Hamiltonians that have a large coupling constant. Based on the strong coupling expansion of Ref.~\cite{macdonald1988}, the ``folded picture'' refers to a decomposition of time evolution into a periodic high-frequency evolution of operators and a ``gentle'' evolution of the state with a time-independent Hamiltonian.
The dual folded XXZ Hamiltonian reads
\begin{align}
\label{eq:Hamiltonian}
H=\mathcal{J}\sum_{\ell}\frac{1- \sigma_{\ell+1}^z}{2}(\sigma_\ell^x \sigma_{\ell+2}^x+ \sigma_\ell^y \sigma_{\ell+2}^y)\, .
\end{align}
It is integrable and describes constrained hopping on a lattice, summarised by the following local dynamical rule: $\ket{\uparrow\downarrow\downarrow}\,\leftrightarrow\,\ket{\downarrow\downarrow\uparrow}$. Such a rule preserves the sequence of parities of the positions of spins up. This in turn implies conservation of the configuration of two species of particles defined by that parity. The presence of multiple species of particles typically signals that an integrable model is solvable via a complicated nested Bethe Ansatz. And indeed, the dual folded XXZ model corresponds to a strong-repulsion regime of the two-component Bariev model~\cite{Bariev1991,Bariev1992}, which is of that type. Nevertheless, at the special point described by Eq.~\eqref{eq:Hamiltonian} the nested structure of the Bariev model is reduced to such a degree that Bethe equations can be solved explicitly. Perhaps the most striking remnant of the nested structure is the conservation of the configuration, which results in the breakdown of one-site shift invariance into two-site shift invariance.
The constrained hopping leads to a very rich Hilbert-space structure combined with an exceptional simplicity of the Bethe equations. It enables exact analytical and numerical analysis of various phenomena that have recently been in the spotlight: Hilbert-space fragmentation~\cite{HSF}, pre-relaxation~\cite{Fagotti2014}, generalised hydrodynamics~\cite{folded:2}, spin transport phenomena~\cite{sarang2021}, the effect of non-abelian symmetries, time-translation symmetry breaking~\cite{pozsgay2021}, and the macroscopic effects of local measurements~\cite{measurement2021}. Most of these phenomena are accessible in sectors that allow for an even simpler description; we mention, for example, the non-interacting sectors, where only one species of particles is present, or the exponentially large jammed sector, where nearest-neighbour pairs of spins down are prohibited and the basis states can be chosen to be eigenstates of all $\sigma^z_\ell$ simultaneously.
The local conserved charges of the dual folded XXZ model have been classified in Ref.~\cite{folded:1}. Their place in the Yang-Baxter integrability structure of the model~\cite{zhou96,shiroishi97,Zhang2020} was described in Refs.~\cite{balazs2021a,balazs2021b}. We mention that the fragmentation of the Hilbert space in the folded XXZ model could be related to the existence of additional non-abelian conservation laws, potentially associated with the emergence of the integrable analogue (cf. Refs~\cite{zadnik16,medenjak20,miao21}) of quantum many-body scars~\cite{moudgalya2021,vafek2017,Beri2019,sanjay2021,sanjay2020,mark2020}.
\section{Initial state and spin-flip protocol}\label{s:protocol}
We consider a protocol that consists of the following steps:
\begin{enumerate}
\item Prepare the system in a \emph{jammed} product state with spins aligned along the $z$-axis (in either direction).
In a jammed state fragments of the form $\cdots\downarrow\downarrow\cdots$ are prohibited. We use the notations of Ref.~\cite{folded:1}: spins up correspond to particles of two species determined by the parity of their positions. Specifically, a spin up at site $\ell$ corresponds to a particle of species $b=2\lceil\ell/2\rceil-\ell\in\{0,1\}$ at macrosite $\ell'=\lceil\ell/2\rceil$. A jammed state is completely characterised by the configuration $\underline b=(\ldots, b_n,b_{n+1},\ldots)$ of spins up as follows:
\begin{align}
\ket{\underline b}=\prod_{j}\sigma_{2\ell_j'-b_j}^+\ket{\downarrow\cdots\downarrow}\, .
\end{align}
Here, the macrosites satisfy the recurrence relation
\begin{align}
\label{eq:recurrence}
\ell'_{j+1}=\ell'_{j}+1-b_{j}(1-b_{j+1})\, ,
\end{align}
according to which two subsequent particles of species $1$ and $0$, respectively, occupy the same macrosite. This prevents state fragments of the form $\cdots\underline{\uparrow\downarrow}\,\underline{\downarrow\uparrow}\cdots$ (underlined are the macrosites), which violate the jamming condition.
\item Flip a spin up that was close to a spin down (otherwise the state remains jammed). This step can be interpreted as a result of a double projective measurement, e.g., measuring $\sigma^z$ after $\sigma^x$.
The spin flip destroys a particle. We choose its index to be $0$, so that the state now reads
\begin{equation}
\Bigl(\prod_{j\le-1}\sigma^+_{2\ell'_j-b_j}\Bigr)\Bigl(\prod_{j\ge 1}\sigma^+_{2\ell'_j-b_j}\Bigr)\ket{\downarrow\cdots\downarrow}\, .
\end{equation}
Recurrence~\eqref{eq:recurrence} can now be solved with respect to the position of the spin flipped. First we note that either $b_{-1}$ or $b_1$ should be equal to $b_0$: we flipped a spin close to a spin down (the site of the next spin up has the same parity as the site of the spin flipped). We choose the macrosites in such a way that at $\ell'_0=0$ there are two spins down. As a result we either have $b_{-1}=b_0=0$ or $b_{1}=b_0=1$.
In fact, since we removed a particle (a spin up), it is convenient to redefine the index of the particles, $j$, so as to fill the hole at $j=0$. We choose the convention $j_{\rm new}=j_{\rm old}+\theta(j_{\rm old}<0)$ and then find
\begin{equation}\label{eq:mapping0}
\ell'_j=j-\theta(j\leq 0)+\begin{cases}
-\sum^{j-1}_{m=1}b^{(0)}_{m}(1-b^{(0)}_{m+1})\,, &\text{if }j>0\\
\sum^{|j|}_{m=1}(1-b_{1-m}^{(0)}) b_{-m}^{(0)}\,, &\text{if }j\leq 0\, .
\end{cases}
\end{equation}
Here, $\underline b^{(0)}$ is the configuration obtained by removing $b_0$ from $\underline b$: $b^{(0)}_j=b_{j-\theta(j\leq 0)}$.
We indicate this state by
\begin{equation}
\ket{\Psi(t=0)}=\ket{0;\underline b^{(0)}}\, ,
\end{equation}
where $\ket{n;\underline c}$ denotes a state which is jammed everywhere except between the particles $c_n$ and $c_{n+1}$, where either {\em two} or {\em three} consecutive spins down are present. We call this excess of spins down ``impurity''.
\item Time evolve with $H$.
While the folded XXZ model is interacting, time evolution is particularly simple in the sector of our initial state. As shown in Appendix~\ref{app:time_evolution_of_state}, this sector is spanned by states $\ket{n;\underline b^{(0)}}$, in which the particle macrosites read
\begin{align}
\label{eq:mapping1}
\ell^{\prime(n)}_j=j-\theta(j\leq n)+\begin{cases}
-\sum^{j-1}_{m=1}b^{(0)}_{m}(1-b^{(0)}_{m+1})\,, &\text{if }j>0\\
\sum^{-j}_{m=1}(1-b_{1-m}^{(0)}) b_{-m}^{(0)}\,, &\text{if }j\leq 0\, .
\end{cases}
\end{align}
They are obtained from Eq.~\eqref{eq:mapping0} by noting that, when the impurity jumps across a particle, the macrosite of the latter shifts by $1$ in the opposite direction.
\end{enumerate}
\section{Exact solution to the dynamics}
The configuration of particles in the state is preserved under the action of the Hamiltonian~\eqref{eq:Hamiltonian}. As a result, the impurity can be interpreted as a fermion freely hopping in the background $\underline{b}^{(0)}$ of particles. As shown in Appendix~\ref{app:time_evolution_of_state}, the state at time $t$ reads
\begin{equation}
\ket{\Psi(t)}=\sum_n (-i)^n J_n(4\mathcal{J} t)\ket{n;\underline b^{(0)}}\, ,
\end{equation}
where $J_n(x)$ are Bessel functions.
Let us denote the expectation value $\bra{\Psi(t)}A\ket{\Psi(t)}$ by $\braket{A}_t$ for any operator $A$ at arbitrary time $t$. For the expectation values of operators $D$ that are diagonal in the basis of states $\ket{n;\underline{b}^{(0)}}$, for example $D=\sigma_\ell^z$, we now have
\begin{equation}
\label{eq:diagonal_expectation}
\braket{D}_t=\sum_n |J_n(4\mathcal{J} t)|^2\braket{n;\underline b^{(0)}|D|n;\underline b^{(0)}}\, .
\end{equation}
If $D^2=1$, like for a string of Pauli matrices $\sigma_\ell^z$, the expectation values $\braket{n;\underline b^{(0)}|D|n;\underline b^{(0)}}$ can be either $1$ or $-1$.
The subtle difference between these dynamics and the ones in the XX model, in which the initial state has a single spin up, is hidden in the mapping~\eqref{eq:mapping1} between the positions in the chain and those in the configuration. As will be shown in the following, this mapping has striking effects on the dynamics of local observables, such as the $z$-component of the spin.
\subsection{Macrosite fluctuations of particles}
\label{sec:fluctuations}
We embrace here the Lagrangian perspective of following the motion of particles. A particle is characterised by its species and its macro-position. The species is fixed whereas the macro-position is a dynamical variable. In the following we compute the statistics of the particles' macro-positions. To that aim, let us consider the macrosite~\eqref{eq:mapping1} corresponding to the particle labeled by $j$ (i.e., the $j$-th spin up). This can be thought of as the eigenvalue of an operator that is diagonal in the basis of states $\ket{n;\underline{b}^{(0)}}$, in which it assumes values $\theta(j>n)+c(j)$, where $c(j)$ denotes is its state-independent part.
Using Eq.~\eqref{eq:diagonal_expectation} we can now compute the characteristic function $\braket{e^{i k \ell'_j}}_t$. We find
\begin{equation}
\label{eq:characteristic_fn}
\braket{e^{i k \ell'_j}}_t=e^{i k c(j)}[(e^{i k}-1)f(j,t)+1]\, ,
\end{equation}
where we introduced the auxiliary quantity
\begin{align}
\label{eq:f_function}
f(x,t):=\sum_n \theta(x>n)|J_n(4\mathcal{J}t)|^2\sim\frac{1}{2}+\frac{1}{\pi}\arcsin\Big(\frac{x}{4\mathcal{J} t}\Big)\, .
\end{align}
The characteristic function~\eqref{eq:characteristic_fn}, in particular, implies
\begin{align}
\mathrm{Prob}(\ell'_j)=\begin{cases}
1-f(j,t)\,,&\text{if }\ell'_j=c(j)\\
f(j,t)\,,&\text{if }\ell'_j=c(j)+1\, ,
\end{cases}
\end{align}
as well as
\begin{align}
\braket{\ell'_j}_t=f(j,t)+c(j)\,,\qquad \braket{\ell_j'^2}_t^{\rm c}=f(j,t)-f(j,t)^2\,,
\end{align}
where $\langle A B\rangle^{\rm c}_t:=\langle AB\rangle_t-\langle A\rangle_t\langle B\rangle_t$ denotes the connected correlation function.
The particles therefore experience fluctuations of order $O(1)$ around their average positions. The fluctuations of different particles are moreover coherent, as can be inferred from the following non-negative equal-time connected correlation function
\begin{align}
\braket{\ell'_m\ell'_n}_t^{\rm c}=\theta(m\ge n) f(n,t)+\theta(n>m)f(m,t)-f(m,t)f(n,t)\, ,
\end{align}
plotted along with $\braket{\ell_j'^2}_t^{\rm c}$ in Fig.~\ref{fig:fluctuations}.
\begin{figure}[ht!]
\centering
\hspace{-1.75em}\includegraphics[width=0.5\textwidth]{fluctuations.pdf}
\caption{Fluctuations of the macrosite corresponding to the $j$-th particle, and the equal-time connected correlation function between the macrosites of particle $0$ and particle $j$. Both are plotted as functions of the ballistic ray $j/(\mathcal{J}t)$.}
\label{fig:fluctuations}
\end{figure}
\subsection{Local magnetisation}
We now adopt an Eulerian point of view and study the properties of the state at a given position or ray.
The Hamiltonian commutes with $\prod_j\sigma_j^z$ and the initial state is an eigenstate of that operator, therefore $\braket{\sigma_\ell^x}_t=\braket{\sigma_\ell^y}_t=0$. The local magnetisation along the $z$-axis instead exhibits a nontrivial ballistic-scale profile. Indeed, as shown in Appendices~\ref{app:matrix_elements} and \ref{app:profiles_magnetisation}, the asymptotic behaviour of the local magnetisation along the $z$-axis is determined by the following rules:
\begin{align}
\label{eq:asymptotic_magnetisation}
\braket{\sigma_{\ell}^z}_t\sim \begin{cases}
1\,,&\text{for }\ket{\cdots \bullet\cdots\uparrow_\ell\updownarrow\uparrow\cdots}\vee \ket{\cdots\uparrow\updownarrow\uparrow_\ell\cdots\bullet\cdots}\\
\frac{2}{\pi}\arcsin\big(\tfrac{x(\ell)}{4\mathcal{J} t}\big)\,,&\text{for }\ket{\cdots\bullet\cdots\uparrow_\ell\uparrow\downarrow\cdots}\vee \ket{\cdots\uparrow\uparrow\downarrow_\ell\cdots\bullet\cdots}\\
-\frac{2}{\pi}\arcsin\big(\tfrac{x(\ell)}{4\mathcal{J}t}\big)\,,&\text{for }\ket{\cdots\bullet\cdots\downarrow_\ell\uparrow\uparrow\cdots}\vee \ket{\cdots\downarrow\uparrow\uparrow_\ell\cdots\bullet\cdots}\\
-1\,,&\text{for }\ket{\cdots\bullet\cdots\downarrow_\ell\uparrow\downarrow\cdots}\vee\ket{\cdots\downarrow\uparrow\downarrow_\ell\cdots\bullet\cdots}\,.
\end{cases}
\end{align}
Here, $\ket{\cdots\bullet\cdots\updownarrow_\ell\cdots}$ and $\ket{\cdots\updownarrow_\ell\cdots\bullet\cdots}$ stand for fragments of the initial state in which $\ell$ is positive or negative, respectively. For large-enough $|\ell|$, $x(\ell)$ can be computed from
\begin{align}
\label{eq:x_to_ell}
\frac{1}{\ell}\int_0^{x(\ell)}\mathrm d n\, \xi(n)=\frac{1}{2}\, ,
\end{align}
where $\xi(n)$ is the coarse-grained macro-distance between particles, namely the average of $1-b^{(0)}_j(1-b^{(0)}_{j+1})$ over a large-enough interval of $j$ including
$n$. For example, $\xi(n)=3/4$ for the fragment $\ldots,1,1,0,0,1,1,0,0,1,1,0,0,\ldots$ The remarkable feature of Eq.~\eqref{eq:asymptotic_magnetisation} is that the asymptotic value of the local magnetisation at site $\ell$ is determined solely by a configuration of three consecutive spins around that site. The only sites in which the spins are asymptotically not aligned along the $z$-axis are those that start in one of the following spin configurations: $\bullet\cdots\underline{\uparrow}\!\uparrow\downarrow$, $\bullet\cdots\underline{\downarrow}\!\uparrow\uparrow$, $\uparrow\uparrow\!\underline{\downarrow}\cdots\bullet$, and $\downarrow\uparrow\!\underline{\uparrow}\cdots\bullet$. Their common feature is the existence of both species of particles and therefore the presence of interaction~\cite{folded:1}. Indeed, as shown in Fig.~\ref{fig:Folded_magnetization}, initial states with a single species of particles do not exhibit nontrivial ballistic profiles.
\begin{figure}[t!]
\hspace{-1.em}
\centering
\includegraphics[width=0.95\textwidth]{3panel_Folded_magn_t10_Neel_2.pdf}
\caption{Panel (a) shows $\braket{\sigma^z_\ell}_t$ as a function of the ballistic ray coordinate $\ell/(\mathcal{J}t)$. The initial state (immediately after the spin flip) is $\ket{\cdots\uparrow\uparrow\downarrow\uparrow\uparrow\downarrow\uparrow \downarrow \downarrow\uparrow\uparrow\downarrow\uparrow\uparrow\downarrow\cdots}$. Blue points correspond both to numerical implementation of Eq.~\eqref{eq:diagonal_expectation} with $D=\sigma_\ell^z$, as well as to the DMRG data (they are in perfect agreement). The orange line corresponds to the asymptotic behaviour~\eqref{eq:asymptotic_magnetisation}, in which $x(\ell)=2\ell/3$, as follows from Eq.~\eqref{eq:x_to_ell}.
Panel (b) shows $\braket{\sigma^z_{\ell}}_t$ for $\mathcal{J}t=10$ for a state with one species of particles, i.e., a noninteracting state. For the initial configuration we chose a N\'eel state with one of the spins up flipped down $\ket{\cdots\uparrow\downarrow \uparrow\downarrow \downarrow\downarrow \uparrow\downarrow \uparrow\downarrow\cdots}$. Panel (c) shows $\braket{\sigma^z_{\ell}}_t$ for $\mathcal{J}t=10$ in the state given in Panel (b) with complementary flipped up spin at $\ell = 10$. Two blue points describe the spins not aligned with the $z$-axis.
}
\label{fig:Folded_magnetization}
\end{figure}
\subsection{Emergence of a locally quasi-jammed state}
Since the spin flip at time $t=0$ creates an excess of spins down, for $t>0$ the state is not jammed anymore. However, the weaker condition
\begin{align}\label{eq:LQJcond}
P_{\downarrow\downarrow}(\ell)=\bigl\langle\frac{1-\sigma_\ell^z}{2}\frac{1-\sigma_{\ell+1}^z}{2}\bigr\rangle_t\xrightarrow{\mathcal{J}t\rightarrow\infty}0\, ,\quad \text{for all $\ell$}\, ,
\end{align}
used in Ref.~\cite{measurement2021} to define locally quasi-jammed states (LQJS), still holds. Expressed in the basis of states $\ket{n;\underline{b}^{(0)}}$, the expectation value used in the LQJS condition reads (see Appendix~\ref{app:correlators})
\begin{align}
\label{eq:jamming2}
P_{\downarrow\downarrow}(\ell)=\sum_n|J_n(4\mathcal{J}t)|^2\braket{n;\underline{b}^{(0)}|\frac{1-\sigma_{\ell}^z}{2}|n;\underline{b}^{(0)}}\braket{n;\underline{b}^{(0)}|\frac{1-\sigma_{\ell+1}^z}{2}|n;\underline{b}^{(0)}}\, .
\end{align}
From Eq.~\eqref{eq:jamming2} it follows that the LQJS condition is satisfied up to corrections of order $O(1/t)$, a feature corroborated by exact numerics -- see Fig.~\ref{fig:Folded_Jamming_condition} (an exact analytical calculation in a weakly interacting scenario is presented in Section~\ref{s:special_protocol}). Indeed,
only few terms of the sum in Eq.~\eqref{eq:jamming2} are different from zero. Let us call $j(\ell)$ the index of the particle represented by a spin up at site $\ell$, or at one of the neighbouring sites if the spin at $\ell$ is down. For $|n-j(\ell)|> n_0$, with $n_0$ a finite positive integer, the matrix elements of $\sigma^z_\ell$ and $\sigma_{\ell+1}^z$ have simple asymptotic forms derived in Appendix~\ref{app:matrix_elements}. Using those, we can compute the eigenvalues of $\tfrac{1}{2}(1-\sigma^z_\ell)$ and
$\tfrac{1}{2}(1-\sigma^z_{\ell+1})$ for all possible configurations of spins in the vicinity of $\ell$: they are shown in Table~\ref{tab:matrix_elements} for a positive $\ell$. Observing that their product is always zero, we conclude that $P_{\downarrow\downarrow}(\ell)$ is a finite sum of squared Bessel functions, which decay to $0$ as $1/t$.
\begin{table}[h!]
\begin{enumerate}[(a)]
\item $n-j(\ell)>n_0$:\\
\begin{tabular}{l || c | c | c | c | c | c | c | c}
& $\uparrow_\ell \uparrow \downarrow \uparrow$
& $ \uparrow_\ell \uparrow \uparrow \downarrow$
& $\uparrow_\ell \uparrow \uparrow \uparrow$
& $\uparrow_\ell \downarrow \uparrow \uparrow$
& $\uparrow_\ell \downarrow \uparrow \downarrow$
& $ \downarrow_\ell \uparrow \uparrow \uparrow$
&$\downarrow_\ell \uparrow \uparrow \downarrow$
& $ \downarrow_\ell \uparrow \downarrow \uparrow$ \\
\hline
$\braket{n;\underline{b}^{(0)}|\tfrac{1}{2}(1\!-\!\sigma_{\ell}^z)|n;\underline{b}^{(0)}}$
& 1 & 0 & 0 & 0 & 0 & 0 & 0 & 1 \\
\hline
$\braket{n;\underline{b}^{(0)}|\tfrac{1}{2}(1\!-\!\sigma_{\ell+1}^z)|n;\underline{b}^{(0)}}$
& 0 & 1 & 0 & 0 & 1 & 0 & 1 & 0
\end{tabular}
\item $n-j(\ell)<-n_0$:\\
\begin{tabular}{l || c | c | c | c | c | c | c | c}
& $\uparrow_\ell \uparrow \downarrow \uparrow$
& $ \uparrow_\ell \uparrow \uparrow \downarrow$
& $\uparrow_\ell \uparrow \uparrow \uparrow$
& $\uparrow_\ell \downarrow \uparrow \uparrow$
& $\uparrow_\ell \downarrow \uparrow \downarrow$
& $ \downarrow_\ell \uparrow \uparrow \uparrow$
&$\downarrow_\ell \uparrow \uparrow \downarrow$
& $ \downarrow_\ell \uparrow \downarrow \uparrow$ \\
\hline
$\braket{n;\underline{b}^{(0)}|\tfrac{1}{2}(1\!-\!\sigma_{\ell}^z)|n;\underline{b}^{(0)}}$
& 0 & 0 & 0 & 0 & 0 & 1 & 1 & 1 \\
\hline
$\braket{n;\underline{b}^{(0)}|\tfrac{1}{2}(1\!-\!\sigma_{\ell+1}^z)|n;\underline{b}^{(0)}}$
& 0 & 0 & 0 & 1 & 1 & 0 & 0 & 0
\end{tabular}
\end{enumerate}
\caption{The expectation values of the projectors $\tfrac{1}{2}(1-\sigma_\ell^z)$ and $\tfrac{1}{2}(1-\sigma_{\ell+1}^z)$ in the states $\ket{n;\underline{b}^{(0)}}$ depend only on the initial configuration of the four spins on sites $\ell$, $\ell+1$, $\ell+2$, and $\ell+3$. For $|n-j(\ell)|> n_0$, with $n_0$ a finite positive integer, the products of these expectation values are always zero.}
\label{tab:matrix_elements}
\end{table}
\begin{figure}[t!]
\hspace{-1.5em}\centering
\includegraphics[width=0.6\textwidth]{Jamming_rescaled.pdf}
\caption{Rescaled jamming condition profile $\mathcal{J}t\mathcal{P}_{\downarrow\downarrow}(\ell,t)$ for $\mathcal{J}t=25$ (blue) and $\mathcal{J}t=50$ (orange) in the initial state $\ket{\cdots\uparrow\uparrow\downarrow\uparrow\uparrow\downarrow\uparrow \downarrow \downarrow\uparrow\uparrow\downarrow\uparrow\uparrow\downarrow\cdots}$. The DMRG data confirm the $O(1/t)$ approach towards an LQJS: $\mathcal{J}tP_{\downarrow\downarrow}(\ell,t)$ forms a profile with a scale invariant envelope, conjectured to be $a/\sqrt{1-[\zeta/v_\pm]^2}$, where $a\approx 0.16$, $v_\pm = \pm 8\mathcal{J}/\left(1+\tfrac{2}{L}\braket{S^z}_0\right) = \pm 6$ ($\braket{S^z}_0$ is the total magnetisation in the initial state) -- see Ref.~\cite{measurement2021}.
}
\label{fig:Folded_Jamming_condition}
\end{figure}
\subsection{Local conservation laws and currents}
\label{sec:charges_currents}
The spin along the $z$-axis is related to the macrosite densities of two local conservation laws
\begin{align}
S^z=\sum_{\ell'}\tfrac{1}{2}(\sigma_{2\ell'-1}^z+\sigma_{2\ell'}^z)\, ,\qquad S_-^z=\sum_{\ell'}\tfrac{1}{2}(\sigma_{2\ell'}^z-\sigma_{2\ell'-1}^z)\, ,
\end{align}
i.e., the magnetisation and the staggered magnetisation, respectively. Using the asymptotic rules~\eqref{eq:asymptotic_magnetisation} we obtain
\begin{equation}
\Bigl\langle\frac{\sigma_{2\ell'}^z+\sigma_{2\ell'-1}^z}{2}\Bigr\rangle_t\sim\begin{cases}
1\,,&\text{for }\ket{\cdots \bullet\cdots(\uparrow\uparrow)_{\ell'}\uparrow\uparrow\cdots}\vee \ket{\cdots\uparrow\uparrow(\uparrow\uparrow)_{\ell'}\cdots\bullet\cdots}\,\\
\frac{1}{\pi}\arccos\big(-\tfrac{\tilde x(\ell')}{4\mathcal{J}t}\big)\,,&\text{for }\begin{cases}
\ket{\cdots \bullet\cdots(\uparrow\uparrow)_{\ell'}\uparrow\downarrow\cdots}\vee\ket{\cdots\uparrow\uparrow(\uparrow\downarrow)_{\ell'}\cdots\bullet\cdots}\\
\ket{\cdots \bullet\cdots(\uparrow\uparrow)_{\ell'}\downarrow\uparrow\cdots}\vee\ket{\cdots\uparrow\uparrow(\downarrow\uparrow)_{\ell'}\cdots\bullet\cdots}
\end{cases}\\
\frac{1}{\pi}\arccos\big(\tfrac{\tilde x(\ell')}{4\mathcal{J}t}\big)\,,&\text{for }\begin{cases}\ket{\cdots \bullet\cdots(\uparrow\downarrow)_{\ell'}\uparrow\uparrow\cdots}\vee\ket{\cdots \uparrow\downarrow(\uparrow\uparrow)_{\ell'}\cdots\bullet\cdots}\\
\ket{\cdots \bullet\cdots(\downarrow\uparrow)_{\ell'}\uparrow\uparrow\cdots}\vee\ket{\cdots\downarrow\uparrow(\uparrow\uparrow)_{\ell'}\cdots\bullet\cdots}
\end{cases}\\
0\,,&\text{for }\begin{cases}\ket{\cdots \bullet\cdots(\uparrow\downarrow)_{\ell'}\uparrow\downarrow\cdots}\vee
\ket{\cdots\uparrow\downarrow(\uparrow\downarrow)_{\ell'}\cdots\bullet\cdots}\\
\ket{\cdots \bullet\cdots(\downarrow\uparrow)_{\ell'}\uparrow\downarrow\cdots}\vee\ket{\cdots\downarrow\uparrow(\uparrow\downarrow)_{\ell'}\cdots\bullet\cdots}\\
\ket{\cdots \bullet\cdots(\downarrow\uparrow)_{\ell'}\downarrow\uparrow\cdots}\vee\ket{\cdots\downarrow\uparrow(\downarrow\uparrow)_{\ell'}\cdots\bullet\cdots}
\end{cases}
\end{cases}
\end{equation}
for the magnetisation, and
\begin{equation}
\Bigl\langle\frac{\sigma_{2\ell'}^z-\sigma_{2\ell'-1}^z}{2}\Bigr\rangle_t\sim\begin{cases}
-\frac{1}{\pi}\arccos\big(\tfrac{\tilde x(\ell')}{4\mathcal{J}t}\big)\,,&\text{for }\ket{\cdots \bullet\cdots(\uparrow\uparrow)_{\ell'}\uparrow\downarrow\cdots}\vee
\ket{\cdots\uparrow\uparrow(\uparrow\downarrow)_{\ell'}\cdots\bullet\cdots}\\
-\frac{1}{\pi}\arccos(-\tfrac{\tilde x(\ell')}{4\mathcal{J}t})\,,&\text{for }\ket{\cdots \bullet\cdots(\uparrow\downarrow)_{\ell'}\uparrow\uparrow\cdots}\vee
\ket{\cdots\uparrow\downarrow(\uparrow\uparrow)_{\ell'}\cdots\bullet\cdots}\\
\frac{1}{\pi}\arccos(\tfrac{\tilde x(\ell')}{4\mathcal{J}t})\,,&\text{for }\ket{\cdots \bullet\cdots(\uparrow\uparrow)_{\ell'}\downarrow\uparrow\cdots}\vee
\ket{\cdots \uparrow\uparrow(\downarrow\uparrow)_{\ell'}\cdots\bullet\cdots}\\
\frac{1}{\pi}\arccos(-\tfrac{\tilde x(\ell')}{4\mathcal{J}t})\,,&\text{for }\ket{\cdots \bullet\cdots(\downarrow\uparrow)_{\ell'}\uparrow\uparrow\cdots}\vee
\ket{\cdots\downarrow\uparrow(\uparrow\uparrow)_{\ell'}\cdots\bullet\cdots}\\
\frac{2}{\pi}\arcsin(\tfrac{\tilde{x}(\ell')}{\mathcal{J} t})\,,&\text{for }\ket{\cdots \bullet\cdots(\downarrow\uparrow)_{\ell'}\uparrow\downarrow\cdots}\vee
\ket{\cdots\downarrow\uparrow(\uparrow\downarrow)_{\ell'}\cdots\bullet\cdots}\\
-1\,,&\text{for }\ket{\cdots \bullet\cdots(\uparrow\downarrow)_{\ell'}\uparrow\downarrow\cdots}\vee
\ket{\cdots\uparrow\downarrow(\uparrow\downarrow)_{\ell'}\cdots\bullet\cdots}\\
0\,,&\text{for }\ket{\cdots \bullet\cdots(\uparrow\uparrow)_{\ell'}\uparrow\uparrow\cdots}\vee
\ket{\cdots\uparrow\uparrow(\uparrow\uparrow)_{\ell'}\cdots\bullet\cdots}\\
1\,,&\text{for }
\ket{\cdots\bullet\cdots(\downarrow\uparrow)_{\ell'}\downarrow\uparrow\cdots}\vee
\ket{\cdots\downarrow\uparrow(\downarrow\uparrow)_{\ell'}\cdots\bullet\cdots}\end{cases}
\end{equation}
for the staggered magnetisation: in both cases $\tilde x(\ell')\sim x(2\ell')\sim x(2\ell'-1)$ is the coarse-grained position in the chain.
The currents of the magnetisation and of the staggered magnetisation
are obtained from the time derivative of $\sigma_\ell^z$ in the Heisenberg picture
\begin{align}
\frac{{\rm d}}{{\rm d}t}\sigma^z_\ell(t)\Bigr|_{t=0}=-\mathcal{J} D_{\ell,\ell+2}\frac{1-\sigma_{\ell+1}^z}{2}+\mathcal{J} D_{\ell-2,\ell}\frac{1-\sigma_{\ell-1}^z}{2}\, ,
\end{align}
where $D_{\ell,m}=\sigma_\ell^x\sigma_m^y-\sigma_\ell^y\sigma_m^x$. In a jammed state, each term on the right-hand side vanishes. We can therefore conclude that the two currents approach zero in any locally quasi-jammed state in which the jamming condition is satisfied up to $O(1/t)$ corrections.
The reader could be surprised at this result as we showed earlier that the corresponding charges have a nontrivial profile in the ballistic scaling limit $x\sim \mathcal{J} t$: in these situations, using the continuity equation, one could infer that also the differences of currents should be functions of $x/t$
\begin{equation}
\partial_t q(\tfrac{x}{t})+\partial_x j(x,t)=0\quad\Rightarrow\quad j(x,t)-j(0,t)=\int_0^{x/t}{\rm d} \zeta q'(\zeta)\zeta\,.
\end{equation}
Here, $q$ is the expectation value of the charge density and $j$ that of the corresponding current.
This argument can however be applied only if the charge density is a (piece-wise) continuous function of $x/t$. In our case we have an additional microscopic dependence on the lattice position $q=q(\ell,t)$: the charge expectation value depends on the pattern of spins around the site $\ell$ -- cf.~\eqref{eq:asymptotic_magnetisation}. As we move from site to site, $q$ jumps between four smooth asymptotic curves that can be recognised only on large (ballistic) scales. The continuity equation on the chain then reads
\begin{equation}
\partial_t q(\ell,t)=j(\ell+1,t)-j(\ell,t)
\end{equation}
and, due to the microscopic dependence on $\ell$, it now becomes possible to have a nontrivial ballistic profile for the charge notwithstanding the corresponding current approaching zero. The macroscopic reconfiguration of the spin profiles after a localised perturbation is thus not a typical transport phenomenon: as shown in Section~\ref{sec:fluctuations}, the positions of the spins up only experience ``collective'' fluctuations of magnitude $O(1)$. The phenomenon is somewhat reminiscent of the travelling surface waves in which the particles of water undergo only a short-scale oscillatory motion. The distinguishing and perhaps most surprising feature of the phenomenon shown here is that it is a result of an instantaneous perturbation.
\begin{figure}[h!]
\hspace{-1.5em}\centering
\includegraphics[width=0.95\textwidth]{3panel_Q1_minus_Q2_plus.pdf}
\caption{Spin current $q_1^-(\ell,t)$ and higher charge $q_2^+(\ell,t)$ for $\mathcal{J}t=25$ (blue points) and $\mathcal{J}t=50$ (orange points), obtained from the DMRG data for the initial state $\ket{\cdots\uparrow\uparrow\downarrow\uparrow\uparrow\downarrow\uparrow \downarrow \downarrow\uparrow\uparrow\downarrow\uparrow\uparrow\downarrow\cdots}$. In the large-time limit both profiles vanish as $1/t$, whereas the charges $q_1^+$ and $q_2^-$ are strictly zero during the time evolution.}
\label{fig:Q1_minus_Q2_plus}
\end{figure}
Using the thermodynamic Bethe Ansatz description developed in Ref.~\cite{folded:2}, it can also be shown that the expectation values of the remaining (known) local conservation laws (reported in Refs~\cite{folded:1,folded:2}) asymptotically vanish. In the thermodynamic limit their expectation values per macrosite read~\cite{folded:2}
\begin{align}
\label{eq:charge_expect_value}
q_n^\pm(x,t)=\int{\rm d}p\, \rho_{x,t}(p) \mathcal{Q}_n^\pm(p)\,,
\end{align}
where the root density $\rho_{x,t}(p)$ encodes the distribution of the occupied momenta at $x$ and $t$, while $\mathcal{Q}_n^+(p)=4\mathcal{J}\cos(n p)$ and $\mathcal{Q}_n^-(p)=4\mathcal{J}\sin(np)$ are the single-particle charge expectation values. The root density $\rho$ and the total density $\rho_{\rm t}$ of vacancies (i.e., particles and holes) in the momentum space satisfy $\int{\rm d}p\,\rho(p)=\xi^{-1}$ and $2\pi\rho_{\rm t}=1+\mu/\xi$, respectively,
where $\xi^{-1}$ is the particle density per macrosite, while $\mu$ is the thermodynamic limit of the ratio $\sum_{j}b_j(1-b_{j+1})/N$, $N$ being the number of spins up (see Ref.~\cite{folded:2}). In a jammed state of an infinite system one has $\xi+\mu=1$, whence $\int{\rm d}p\,\rho(p)=1/(1-\mu)$ and $\rho_{\rm t}=1/(2\pi[1-\mu])$~\cite{folded:1}. The occupancy ratio $n=\rho(p)/\rho_{\rm t}$ then satisfies $\int {\rm d}p\,n(p)=2\pi$ and, since $0\le n(p)\le 1$, one has $n(p)=1$. This implies that $\rho(p)$ is constant, whence all local charge expectation values, expect for the magnetisation (and the staggered one, albeit for a different reason\footnote{The staggered magnetisation is not related to the root density $\rho$, so the argument above does not pertain to it.}), vanish:
\begin{align}
q_n^\pm(x,t)=0,\quad\text{for all $n>0$}\, .
\end{align}
A similar conclusion can be drawn for the currents, whose expectation values read
\begin{align}
j_n^\pm(x,t)=\int{\rm d}p\, \rho_{x,t}(p)v_{x,t}(p) \mathcal{Q}_n^\pm(p) + \text{const.}\,,
\end{align}
where $v(p)=-4 \mathcal{J} \sin p -4\mathcal{J}\mu\int\frac{\mathrm d k}{2\pi}n(k)(\sin k-\sin p)$ is the dressed velocity~\cite{folded:2}. By restricting to jammed states, we immediately see that all currents become independent of the state and hence approach trivially constant profiles in locally quasi-jammed states like those emerging in our setting.
\subsection{Entanglement in a weakly interacting setting}
\label{s:special_protocol}
In this section we identify both classical and pure quantum contributions to the two-point spin correlation functions. We focus on a weakly interacting scenario in which the background consists of jammed particles that are of the same species almost everywhere except for those in a special region. In the latter, both species of particles are present in the maximal number.
In other words, the initial jammed state is obtained from the N\'eel state by flipping up a sequence of $M$ spins down. We denote the corresponding macrosites by $m',m'+1,\ldots,m'+M-1$, so we hav
\begin{equation}
\label{eq:weakly_interacting_state}
\ket{\Psi(0^-)}=\ket{\cdots\downarrow\uparrow\downarrow\uparrow \underset{m'}{\underline{\uparrow\uparrow}}\uparrow\uparrow \cdots \uparrow\uparrow \underset{m'+M-1}{\underline{\uparrow\uparrow}} \downarrow\uparrow\downarrow\uparrow\cdots}\,,
\end{equation}
where we are going to assume $m'>0$. Note that for this specific state $M=\sum_{j}b_j(1-b_{j+1})$. In the light of the Bethe Ansatz solution of Ref.~\cite{folded:1}, a nonzero sum $\sum_{j}b_j(1-b_{j+1})$ is always associated with interaction. Since, however, the sum is finite and we are in the thermodynamic limit, the state~\eqref{eq:weakly_interacting_state} can be considered weakly interacting. By flipping the spin in position $0$ according to the protocol described in Section~\ref{s:protocol}, we end up with the state
\begin{equation}
\label{eq:weakly_interacting_state_2}
\ket{\Psi(0)}= \ket{\cdots\downarrow\uparrow\underset{0}{\underline{\downarrow\downarrow}}\downarrow\uparrow \cdots\downarrow\uparrow\downarrow\uparrow
\underset{m'}{\underline{\uparrow\uparrow}}\uparrow\uparrow \cdots \uparrow\uparrow
\underset{m'+M-1}{\underline{\uparrow\uparrow}} \downarrow\uparrow\downarrow\uparrow\cdots}\,,
\end{equation}
which is no longer jammed, but evolves in time towards a locally quasi-jammed state. One can indeed readily show (see Appendix~\ref{app:special_protocol})
\begin{equation}
P_{\downarrow\downarrow}(\ell) =
\left\{\begin{aligned}
&J^2_{\lceil\ell/2\rceil}(4 \mathcal{J} t)\,, &&\text{if } \lfloor\ell/2\rfloor<m'-1 \\
&J^2_{\ell-m'+2}(4 \mathcal{J} t)\,, &&\text{if } m'-1\le \lfloor\ell/2\rfloor <m'+M-2 \\
&J^2_{\lceil\ell/2\rceil+M}(4 \mathcal{J} t)\,, &&\text{if } \lfloor\ell/2\rfloor\ge m'+M-2\,, \\
\end{aligned}\right.
\end{equation}
hence the jamming condition is asymptotically satisfied up to $O(1/t)$ corrections -- cf.~\eqref{eq:LQJcond}.
For the local magnetisation we obtain
\begin{align}
\begin{aligned}
\label{eq:Neel_magnetization}
\braket{\sigma^z_{2\ell'-1}}_t&=
\left\{
\begin{aligned}
&2f(m'+2M-2,t)-1 \,,&&\text{if } \ell'=m'+M-1
\\
&1-2f(m',t) \,,&&\text{if } \ell'=m'-1
\\
&1-2J^2_{m'-2\ell'-1}(4 \mathcal{J} t)-2J^2_{m'-2\ell'}(4 \mathcal{J}t) \,,
\phantom{ii}
&& \text{if } m'\le\ell'\le m'+M-2
\\
&-1 \,,&& \text{otherwise}\,,
\end{aligned}
\right. \\
\braket{\sigma^z_{2\ell'}}_t&=
\left\{
\begin{aligned}
&1
-2J^2_{\ell'}(4 \mathcal{J} t)
\,,&&\text{if } \ell'<m'-1
\\&1
-2J^2_{\ell'+M}(4 \mathcal{J} t)
\,,&&\text{if } \ell'>m'+\Delta-2
\\&1
-2 J^2_{m'-2\ell'-2}(4 \mathcal{J} t)
-2 J^2_{m'-2\ell'-1}(4 \mathcal{J} t)
\,,&& \text{if } m'-1\le \ell'\le m'+M-2
\,,
\end{aligned}
\right.
\end{aligned}
\end{align}
where $f(x,t)$ is defined in Eq.~\eqref{eq:f_function}, while $\braket{\sigma^x_\ell}_t=0$ and $\braket{\sigma^y_\ell}_t=0$ (see Appendix~\ref{app:special_protocol_SpinProfiles} for details). Note that, except for the sites $2m'-3$ and $2m'+2M-3$, the long-time limit of the expectation value of $\sigma^z$ is either $1$ or $-1$.
Using the asymptotic expression for the function $f$, reported in Eq.~\eqref{eq:f_function}, we instead obtain
\begin{equation}
\label{eq:special_trajectories}
\begin{aligned}
\braket{\sigma^z_{2m'-3}}_t&\sim -\frac{2}{\pi}\arcsin\Bigl(\frac{m'}{4 \mathcal{J} t}\Bigr),\qquad
\braket{\sigma^z_{2m'+2M-3}}_t&\sim \frac{2}{\pi}\arcsin\Bigl(\frac{m'}{4 \mathcal{J} t}\Big[1+2\frac{M-1}{m'}\Big]\Bigr)\,,
\end{aligned}
\end{equation}
for the $z$-components of the spins at the sites $2m'-3$ and $2m'+2M-3$ (see Fig.~\ref{fig:Neel_magnetization}). This implies that, in the infinite-time limit, the reduced density matrix of any spin with site index $j\nin\{2m'-3,2m'+2M-3\}$ is pure. The reduced density matrix of each of the two spins at sites $2m'-3$ and $2m'+2M-3$ is instead maximally mixed (proportional to the identity). Hence, the spins $2m'-3$ and $2m'+2M-3$ are maximally entangled with the rest of the system (the entanglement entropy of their reduced density matrices is maximal).
On the other hand, the large-time limit of the correlation $\braket{\sigma^z_{2m'-3}\sigma^z_{2m'+2M-3}}_t$ is $-1$, i.e., the two spins are asymptotically maximally correlated. One might be tempted to conclude that these two spins are entangled with one another even in the infinite-time limit, i.e., that they are in a cat-state. Instead, it can be shown that their entanglement goes to zero as $\log(\mathcal{J}t)/(\mathcal{J}t)^2$, as does the entanglement between any other two spins. To see this, we compute the entanglement of formation~\cite{Wootters1997,Wootters1998} $E$ between two given spins:
\begin{equation}
E(\hat \rho)=h\Bigl(\frac{1+\sqrt{1-C^2(\hat \rho)}}{2}\Bigr)\,.
\end{equation}
Here, $\hat \rho$ is the reduced density matrix of the two spins,
\begin{equation}
h(x)= -x\log_2x-(1-x)\log_2(1-x)\, ,
\end{equation}
while $C(\hat \rho)$ is the so-called concurrence, defined as
\begin{equation}
C(\hat \rho)=\max\{0,\lambda_1-\lambda_2-\lambda_3-\lambda_4\},
\end{equation}
where $\lambda_j$ are the square roots of the eigenvalues of $\hat \rho(\sigma^y\otimes\sigma^y)\hat \rho^\ast(\sigma^y\otimes\sigma^y)$ in descending order, and $\hat\rho^\ast$ is the complex conjugate of $\hat\rho$. Even though $\hat\rho(\sigma^y\otimes\sigma^y)\hat\rho^\ast(\sigma^y\otimes\sigma^y)$ is not necessarily self-adjoint, its eigenvalues are real and non-negative (it is the product of two positive semi-definite matrices).
\begin{figure}[t!]
\hspace{-1.5em}\centering
\includegraphics[width=0.65\textwidth]{Neel_2times.pdf}
\caption{Magnetisation profile $\braket{\sigma^z_{\ell}}_t$ for the initial state described in Eq.~\eqref{eq:weakly_interacting_state_2} with $m'=9$ and $M=10$. Colors of the points correspond to different fixed times $\mathcal{J}t=10$ (green) and $\mathcal{J}t=20$ (blue). The spins at sites $2m'-3$ and $2m'+2M-3$ are the only ones for which the $z$-components of the magnetisation are not asymptotically equal to $\pm 1$. Instead, as time $t$ is increased, they move towards the zero ray (time increases while the lattice positions are fixed) along the trajectories (orange dashed lines) given in Eq.~\eqref{eq:special_trajectories}.
}
\label{fig:Neel_magnetization}
\end{figure}
Note that the concurrence (and hence the entanglement) is automatically zero whenever the density matrix is diagonal because the matrix $\hat\rho(\sigma^y\otimes\sigma^y)\hat{\rho}^\ast(\sigma^y\otimes\sigma^y)$ has two pairs of degenerate positive eigenvalues. One can show that the off-diagonal elements of the two-spin reduced density matrix $\hat \rho$ go to zero as $1/t$ for large times, thus lifting the degeneracy of the matrix $\hat \rho(\sigma^y\otimes\sigma^y)\hat{\rho}^\ast(\sigma^y\otimes\sigma^y)$ by a term of order $O(1/t)$. This makes the entanglement of formation approach zero as $\log(\mathcal{J}t)/(\mathcal{J}t)^2$.
Nonetheless, the two spins at sites $2m'-3$ and $2m'+2M-3$ are special, since they are maximally correlated: in the infinite-time limit their reduced density matrix reads
\begin{equation}
\hat{\rho}\xrightarrow{t\rightarrow\infty} \frac{\ket{\uparrow\downarrow}\bra{\uparrow\downarrow} +\ket{\downarrow\uparrow}\bra{\downarrow\uparrow}}{2}\,,
\end{equation}
which is a classical mixture. Note that the macro-distance $M$ between the correlated spins can be arbitrarily large.
In general, the spread of the rescaled spin-spin entanglement with time is shown in Fig.~\ref{fig:delta_protocol_different_times}. At $t=0$ there are no entangled spins (the initial state is a product state). For $t>0$ the region containing entangled spins grows linearly in time, consistently with the light-cone dynamics. At first the entanglement spreads from the initial position of the impurity in all directions -- cf. Fig.~\ref{fig:delta_protocol_different_times}(a) -- with the maximal velocity $8J$.
When the edge of the light-cone reaches macro-site $m$, a ``leakage'' of the entanglement between the wedges occurs through a corridor, in which only the next-nearest neighbour spins become entangled -- cf. Fig.~\ref{fig:delta_protocol_different_times}(b,c).
We observe that the couples of spins $(i,j)$ with $i,j<2m'-1$ are not affected by the presence of the domain wall in the initial state: their behaviour is the same as if the initial state were a N\'eel state. Instead, the entanglement properties of the pairs of spins $(i,j)$ with $i,j>2m'+M-5$ are shifted: their entanglement corresponds to that in the N\'eel state in which the impurity is initially located at macro-site $-M$. Finally, in our case, all the spins outside those two regions are decoupled unless they are next-nearest neighbours, which is not the case when the dynamics start from the N\'eel state.
\begin{figure}[t!]
\centering
\begin{tikzpicture
\node at (-12.6,0) {\includegraphics[width=.3\textwidth]{DeltaProtDelta7size120emme18time1point5.png}};
\node at (-12.6,-2.7) {(a) $\mathcal{J}t=1.5$};
\node at (-7.8,0) {\includegraphics[width=.3\textwidth]{DeltaProtDelta7size120emme18time5point1.png}};
\node at (-7.8,-2.7) {(b) $\mathcal{J}t=5.1$};
\node at (-3,0) {\includegraphics[width=.3\textwidth]{DeltaProtDelta7size120emme18time9point9.png}};
\node at (-3,-2.7) {(c) $\mathcal{J}t=9.9$};
\node (scale) at (-.4,0) {\includegraphics[height=.2\textwidth]{scale.pdf}};
\node[above =.84cm] at (.1,0) {$0.6$};
\node[below =1.2cm] at (.1,0) {$0.0$};
\node at (.1,-.18) {$0.3$};
\end{tikzpicture}
\caption{Spin-spin entanglement for an initial state with $m'=12$, $M=9$ at different times, rescaled by $\log (\mathcal{J}t)/(\mathcal{J}t)^2$.
The numbers on both axes refer to spin indices: the element $(i,j)$ of the plot characterises the entanglement between the spins at the sites $i$ and $j$. The plot is symmetric w.r.t. the bisector, i.e., the line containing the points $(i,i)$, since the entanglement between the spins $i$ and $j$ equals that between the spins $j$ and $i$.
The blue (orange) lines refer to the spin $2m'-1$ (resp. $2m'+2M-5$), i.e., the one close to the start (resp. end) of the domain wall of spins up. They delimit the nonzero entanglement regions. Panels (a), (b), and (c) show the spread of the entanglement before the domain of spins up is reached by the light cone, during the spread of the entanglement through the domain, and after the edge of the light cone has passed through the domain, respectively.
}
\label{fig:delta_protocol_different_times}
\end{figure}
\section{From the dual folded XXZ to the XXZ model}
The dual folded XXZ model was originally envisioned as an effective model describing the large-anisotropy limit of the Heisenberg XXZ spin-$1/2$ chain. In this respect it is natural to wonder how well it acts as such. Specifically, is there a sign of the measurement catastrophe studied herein and in Ref.~\cite{measurement2021} in the XXZ model with large anisotropy? To answer this question, one needs to consider the inverse of the duality transformation and invert an additional unitary transformation that acts on the initial state and characterises the so-called ``folded picture'' (see Ref.~\cite{folded:1}). The latter unitary transformation only contributes $O(1/\Delta)$ corrections and will be neglected in the following discussion. The inverse duality transformation maps $\sigma_\ell^z\mapsto \sigma_\ell^z\sigma_{\ell+1}^z$, so it essentially acts as follows:
\begin{align}
\label{eq:spin_mapping}
\ket{\cdots\uparrow\uparrow\downarrow\uparrow\uparrow\downarrow\uparrow\downarrow\downarrow\uparrow\uparrow\downarrow\uparrow\uparrow\downarrow\cdots}\xrightarrow[\text{transformation}]{\text{inverse duality}} \ket{\cdots\uparrow\uparrow\uparrow\downarrow\downarrow\downarrow\uparrow\uparrow\downarrow\uparrow\uparrow\uparrow\downarrow\downarrow\downarrow\uparrow\cdots}\, .
\end{align}
Time-evolving this initial state with the Heisenberg Hamiltonian $H_{\rm XXZ}(\Delta)$ is expected to yield a scaling profile of $\braket{\sigma_\ell^z\sigma_{\ell+1}^z}_t$ converging, as $\Delta$ is increased, to the one of $\braket{\sigma_\ell^z}_t$ in the dual folded XXZ model. This is indeed observed in numerical simulations based on tensor network techniques -- cf. Fig.~\ref{fig:XXZ_vs_Folded}, and supports the validity of the strong-coupling expansion described in Ref.~\cite{folded:1}.
Besides that, transforming back to XXZ helps us see why the local measurement could have such a relevant effect: in the XXZ setting, it creates a domain-wall structure. If the two halves of the chain had different integrals of motion, the theory of generalised hydrodynamics~\cite{castro16,bertini16,vir17} would predict nontrivial ballistic profiles. This is not the case. One could then wonder whether we are seeing the effects of quasi-conserved operators that are exactly conserved only in the large-anisotropy limit. So far, our investigations have been inconclusive, so the question is still open.
\begin{figure}[t!]
\centering
\includegraphics[width=0.95\textwidth]{3panel_XXZ_vs_Folded.pdf}
\caption{Panel (a) shows correlation function $\braket{\sigma^z_\ell \sigma^z_{\ell+1}}_t$ in the Heisenberg XXZ model (blue and orange points) compared to the magnetization of the folded model Eq.~\eqref{eq:asymptotic_magnetisation} (black line). Data points quickly converge to the prediction of the folded model for a relatively large values of $\Delta$, whereas for small anisotropy they exhibit a disagreement with the prediction during the whole time evolution. Panel (b) shows the difference $\braket{\sigma^z_\ell}_{\rm folded} -\braket{\sigma^z_\ell \sigma^z_{\ell+1}}_{\rm XXZ}$ outside of the light cone as a function of time. For short times the predictions in both models linearly deviate from each other. In the inset we plot a slope $\alpha$ of this difference as a function of the anisotropy $\Delta$. Data points are fitted with $a\Delta^{-2} + b\Delta^{-3}$ , where $a\approx0.241$, $b\approx1.31$.}
\label{fig:XXZ_vs_Folded}
\end{figure}
\section{Discussion}
The simplicity of the calculations carried out in this paper is rather surprising in view of the fact that we are dealing with a genuinely interacting model. It goes even beyond the simplicity of the Bethe equations of the dual folded XXZ model~\cite{folded:1}. It seems rather related to the setting considered -- that of locally perturbed jammed states, in which particles are closely packed. Such a scenario can be mapped into the problem of a free particle moving in a jammed background. We have exploited this mapping to solve the measurement catastrophe discussed in Ref.~\cite{measurement2021}, i.e., the macroscopic change of spin profiles after a localised perturbation. We have established that no genuine spin transport occurs: the particles' positions exhibit small fluctuations around their expectation values, and the current of the local magnetisation is asymptotically zero. This is a result of the emergent scale separation in which the smooth asymptotic scaling of the sublattice local magnetisation coexists with an underlying discontinuous dependence of $\braket{\sigma^z_\ell}$ on the position $\ell$. Interactions play a key role in the setting we studied: the phenomenon is not observed if particles do not interact. On the other hand, a finite number of interacting particles is enough to trigger the effect. In particular, perturbing a weakly interacting initial state enables isolation of classically correlated pairs of spins.
Jammed states typically arise in kinetically constrained models, and it remains unclear whether the measurement catastrophe described in this work is a universal feature of such models~\cite{garrahan2010kinetically}.
During the finalisation of this work and in the attempt to unveil the physical mechanism behind the measurement catastrophe, one of us has realised that an analogous sensitivity to localised perturbations can be present also after global quenches~\cite{Fagotti2021Global}. In that case the existence of infinitely many ``semilocal charges'' was identified as crucial. That mechanism, however, does not seem to account for the phenomenon discussed in this paper, which appears to be more related to the exceptionality of the jammed sector, consisting of states with memory. Indeed, to the best of our knowledge the dual folded Hamiltonian only has a single semilocal charge, first identified in Ref.~\cite{folded:1}, which corresponds to the total spin along the $z$-axis when mapping back to the XXZ Heisenberg model. That charge has a nontrivial profile in some of the examples that we considered, but it is irrelevant in others, including in the example investigated in Ref.~\cite{measurement2021}.
\section*{Acknowledgements}%
\paragraph{Funding information.} %
This work was supported by the European Research Council under the Starting Grant No. 805252 LoCoMacro.
|
\section{Introduction}
Charmless non-leptonic $B_{(s)}$-meson decays play a prominent role in testing the CKM mechanism of quark flavour mixing, in determining the angles of the unitarity triangle and -- closely related -- in quantifying the amount of CP violation in the quark flavour sector of the Standard Model (SM), which represents the only established source of CP violation to date. Moreover, these decays offer numerous observables such as branching ratios, direct CP asymmetries, and polarization fractions, which entail a rich and interesting phenomenology, in particular if precision can be achieved both on the experimental and the theoretical sides.
On the experimental side numerous non-leptonic decay channels have been scrutinized at the B-factories~\cite{BaBar:2014omp} and at hadron collider experiments in the past years. Ongoing and future programs at the upgraded LHC and at Belle II will further increase the experimental precision of non-leptonic $B$ decays~\cite{LHCb:2018roe,Belle-II:2018jsg} which will be particularly important since quite a number of decay channels and observables -- especially direct CP asymmetries -- are not yet measured or are only poorly constrained.
On the theoretical side the phenomenology of charmless non-leptonic $B_{(s)}$-decays is governed by the interplay between CKM factors (weak phases), Wilson coefficients, and non-perturbative hadronic matrix elements, where the computation of the latter has been the bottleneck to precision predictions for quite some time due to the appearance of QCD effects from many different scales that arise in the purely hadronic initial and final states, and due to the fact that these matrix elements cannot be directly computed with non-perturbative methods such as lattice QCD at present (however, see~\cite{RBC:2015gro} for an interesting study on kaon decays). The hadronic matrix element is oftentimes split up into the so-called topological amplitudes such as the colour-allowed, the colour-suppressed tree, annihilation, the QCD and electroweak penguin amplitudes, etc.~\cite{Chau:1986jb,Chau:1987tk,Chau:1990ay,Gronau:1994rj,Gronau:1995hm,Cheng:2014rfa}.
There are various strategies that have been developed to get a handle on the hadronic matrix elements, each having its virtues and drawbacks. These approaches are mainly based on flavour symmetries of the light quarks (see e.g.~\cite{Zeppenfeld:1980ex} for an early work) or on factorization, like PQCD~\cite{Keum:2000ph,Lu:2000em} or QCD factorization (QCDF)~\cite{Beneke:1999br,Beneke:2000ry,Beneke:2001ev}. The latter provides a rigorous and systematic framework to disentangle short- from long-distance physics in the heavy quark limit, and at leading power in $\Lambda_{\rm QCD}/m_b$ has arrived at a highly sophisticated level~\cite{Beneke:2003zv,Beneke:2006hg,Beneke:2005vv,Kivel:2006xc,Pilipp:2007mg,Beneke:2006mk,Bell:2007tv,Bell:2009nk,Bell:2009fm,Beneke:2009ek,Kim:2011jm,Bell:2014zya,Bell:2015koa,Bell:2020qus,Beneke:2020vnb,Beneke:2021jhp}. One of the shortcomings of the QCDF approach in its present use is the failure to compute sub-leading power corrections in the heavy-quark expansion from first field-theoretical principles (see~\cite{Khodjamirian:2005wn} for a determination of annihilation topologies for
$B\rightarrow \pi \pi$ using QCD-sum
rules). Unfortunately, this results in sizeable uncertainties that in many observables spoil the precision achieved at leading power (see e.g.~\cite{Bell:2015koa}). However, it is fair to say that combinations of observables can be built that are robust against hadronic uncertainties such as the combination $\Delta A_{\rm CP}$ in the $K\pi$ channel~\cite{Gronau:2005kz}, which still persists as the so-called $K\pi$ puzzle (see e.g.~\cite{Buras:2003yc,Chiang:2004nm,Mishima:2004um,Cheng:2004ru,Fleischer:2007mq,Bell:2015koa}).
Flavour symmetries based on the approximate $SU(3)$ invariance in the $(u,d,s)$ flavour space or one of its SU(2) subgroups isospin, U-spin and V-spin have been used extensively in the Literature and have been applied both to non-leptonic bottom~\cite{Savage:1989ub,Gronau:1990ka,Deshpande:1994ii,Gronau:1995hm,He:1998rq,He:2000ys,Fu:2003fy,Chiang:2006ih,Chiang:2008zb,Chiang:2008vc,Grossman:2013lya,Cheng:2014rfa,Hsiao:2015iiu} and charm decays~\cite{Rosner:1999xd,Lipkin:2000sf,Gronau:2000ru,Bhattacharya:2008ss,Hiller:2012xm,Feldmann:2012js,Muller:2015lua,Cheng:2016ejf}. The advantages of this approach are the fact that hardly any assumption about the scales of QCD effects are needed, and that it relates different decay channels, thereby reducing the number of parameters, for instance in a global analysis. While the isospin subgroup of the full flavour $SU(3)$ can be considered unbroken, the U and V-spin subgroups as well as the full flavour $SU(3)$ are severely broken by the splitting between the down/up and the strange quark masses. Although some progress has been made in quantifying the amount of flavour breaking~\cite{Gronau:1995hm,Jung:2009pb,Cheng:2012xb,Grossman:2012ry,Muller:2015lua}, the failure of its rigorous implementation can still be regarded as the main drawback of this approach.
One obvious idea is to combine the different approaches, with the goal to benefit from the various advantages which each one of them offers, while at the same time minimizing the sensitivities to their individual drawbacks. Aspects of factorization combined with the topological amplitudes were put forward in the so-called factorization-assisted topological amplitude approach (FAT)~\cite{Li:2012cfa,Qin:2013tje,Zhou:2015jba,Wang:2017hxe,Jiang:2017zwr}, while combinations of factorization and flavour symmetries were studied e.g.\ in~\cite{Gronau:1995hn,Descotes-Genon:2006spp,Cheng:2011qh,Hsiao:2015iiu}. Also numerous other phenomenological studies with aspects from several approaches can be found in the literature~\cite{Fleischer:1993gr,Buras:1995pz,Gronau:1995hm,Fleischer:1997um,Ciuchini:1997hb,Buras:1997cv,Neubert:1998pt,Neubert:1998jq,Gronau:1998fn,Buras:1998ra,Ali:1998eb,Ali:1998gb,Fleischer:1999jv,Ciuchini:2001gv,Grossman:2003qp,Zenczykowski:2004tw,Buras:2004th,Grossman:2005jb,Lipkin:2005pb,Cheng:2005bg,Imbeault:2006ss,Ali:2007ff,Cheng:2009cn,Cheng:2009mu,Imbeault:2011jz,Bobeth:2014rra,AlvarengaNogueira:2016qyt,Fleischer:2016jbf,Fleischer:2017vrb,Fleischer:2018bld,Bediaga:2020qxg,Fleischer:2021cct,Fleischer:2021cwb,Wang:2020gmn}.
Recently the amplitudes of the topological and $SU(3)$-invariant methods were revisited in~\cite{He:2018php}, whose main findings are that the number of amplitudes in these two descriptions is equal and that the relations between them are linear. In the present work, we make use of the relations between the topological, the $SU(3)$-irreducible decomposition and the factorization approach. We start from the $SU(3)$-irreducible amplitude which we first fit to current experimental data on branching ratios and direct CP asymmetries by performing a $\chi^2$-fit with minimal theoretical bias. We determine our best fit point using random sampling and then evaluate the two-dimensional $1\sigma$ confidence regions for each of the relevant complex amplitudes in the $SU(3)$-irreducible approach. Subsequently we translate the $SU(3)$ amplitudes into those from QCD factorization, with the goal of obtaining the corresponding confidence regions, in particular for the annihilation amplitudes which are currently inaccessible from first QFT principles. Our $\chi^2$-fit is performed under the assumption of exact flavour $SU(3)$. We leave the implementation of symmetry breaking and further theoretical input, e.g.\ from QCDF at leading power, to future work.
This paper is organised as follows. After specifying the effective Hamiltonian in section~\ref{sec:Heff}, we give the decay amplitudes in the topological and $SU(3)$-irreducible frameworks in section~\ref{sec:TopIrr}. Section~\ref{sec:QCDF} deals with the QCD factorization amplitudes, while their connection to the topological ones are given in section~\ref{sec:QCDFTop}. In sections~\ref{sec:prepfit} and~\ref{sec:fit} we give details on the input parameters and statistical analysis, specifically on the $\chi^2$-fit. In section~\ref{sec:results} we present the results, including best-fit points, confidence regions and the value and uncertainty bands for branching fractions and CP asymmetries. We conclude in section~\ref{sec:conclusion}.
\section{Effective Hamiltonian basis}
\label{sec:Heff}
Our starting point is the effective weak Hamiltonian~\cite{Buchalla:1995vs,Chetyrkin:1997gb}. Since our analysis is to a large extent independent of the precise operator basis we adopt the convention of~\cite{Beneke:2003zv}.
\begin{align}
\mathcal{H}^q_{\rm eff}&= \frac{G_F}{\sqrt{2}}\sum_{p=u,c}\lambda^{(q)}_p\Biggl[C_1 Q^{q,p}_1 + C_2 Q^{q,p}_2 + \sum^{10}_{i=3} C_i Q^q_i \Biggl]+\rm{h.c.} \nonumber\\
&=\frac{G_F}{\sqrt{2}}\Biggl[\lambda^{(q)}_u \bigl[ C_1 (Q^{q,u}_1-Q^{q,c}_1) + C_2 (Q^{q,u}_2-Q^{q,c}_2) \bigl] \nonumber\\
& \hspace*{40pt}-\lambda^{(q)}_t \biggl( C_1 Q^{q,c}_1 + C_2 Q^{q,c}_2 + \sum^{10}_{i=3} C_i Q^q_i \biggl)\Biggl]+\rm{h.c.}, \label{eq:EffH}
\end{align}
where $q\in\{d,s\}$, $\lambda^{(q)}_p=V_{pb} V^*_{pq}$ and we have moreover made use of the unitarity relation
\begin{equation}
\lambda^{(q)}_u+\lambda^{(q)}_c+\lambda^{(q)}_t=0.
\label{eq:CKMrel}
\end{equation}
The four-quark operators are given as follows
\begin{align}
\allowdisplaybreaks
Q_{1}^{q, \, p}&=\bigl(\bar{p}_{\beta}b_{\alpha}\bigl)_{V-A}
\bigl(\bar{ q }_{\alpha} p_{\beta}\bigl)_{V-A}
\, , &
Q_{2}^{q, \, p}=& \bigl(\bar{p} b\bigl)_{V-A} \bigl(\bar{q} p\bigl)_{V-A}\, ,
\nonumber \\
Q^q_{3}&=\bigl(\bar{q}b\bigl)_{V-A} \sum_k\bigl(\bar{k}k\bigl)_{V-A} \, ,
&
Q^q_{4}=&\bigl(\bar{q}_\alpha b_\beta\bigl)_{V-A}\sum_k\bigl(\bar{k}_\beta k_\alpha\bigl)_{V-A}\, ,
\nonumber\\
Q^q_{5}&=\bigl(\bar{q} b\bigl)_{V-A} \sum_k\bigl(\bar{k}k\bigl)_{V+A} \, ,
& Q^q_{6}=&\bigl(\bar{q}_\alpha b_\beta\bigl)_{V-A}\sum_k\bigl(\bar{k}_\beta k_\alpha\bigl)_{V+A}\, ,
\nonumber\\
Q^q_{7}&=\bigl(\bar{q} b\bigl)_{V-A}\sum_k\frac{3}{2}e_k\bigl(\bar{k} k\bigl)_{V+A} \, ,
& Q^q_{8}=&\bigl(\bar{q}_\alpha b_\beta\bigl)_{V-A}\sum_k\frac{3}{2}e_k\bigl(\bar{k}_\beta k_\alpha\bigl)_{V+A}\, ,
\nonumber\\
Q^q_{9 }&=\bigl(\bar{q} b \bigl)_{V-A}\sum_{k}\frac{3}{2}e_k\bigl(\bar{k} k \bigl)_{V-A}\, ,
&Q^q_{10}=&\bigl(\bar{q}_\alpha b_\beta \bigl)_{V-A}\sum_{k}\frac{3}{2}e_k\bigl(\bar{k}_\beta k_{\alpha} \bigl)_{V-A}\,.
\label{eq:mainbasis}
\end{align}
Colour indices in singlet currents are suppressed to ease the reading. The sum over $k$ includes the five lightest quark flavours, and $e_k$ is the electric charge of quark $k$ in units of the positron charge. Currently, the Wilson Coefficients $C_i$ have been computed to high orders in different bases~\cite{Buchalla:1995vs,Gorbahn:2004my}.
\section{Topological and $SU(3)$-irreducible representation}
\label{sec:TopIrr}
The physical amplitudes of the processes $B\rightarrow P P$, $P$ being a light pseudoscalar meson, are traditionally written using two representations, the topological and the $SU(3)$-irreducible one. The equivalence between both approaches was recently discussed in \cite{He:2018php}. To make the present paper self-consistent, we summarize the most important relations below.
\subsection{Topological representation}
\label{sec:Top}
Usually, the following notation is used to describe the different topological amplitudes
\begin{itemize}
\item $T$: referring to the color-allowed tree amplitude.
\item $C$: denoting the color-suppressed tree diagram.
\item $E$: referring to the $W$-exchange diagram.
\item $P$, denoting QCD penguin contributions.
\item $S$, representing the flavor QCD singlet penguin.
\item $A$, denoting annihilation diagrams.
\end{itemize}
Using the notation above, the labels for other topologies can be written, for example $T_{AS}$ is used to denote a tree-annihilation singlet topology. Then, the topological decomposition of the physical amplitudes reads
\begin{eqnarray}
{\cal A}^{TDA} = i\frac{G_{F}}{\sqrt{2}}\Bigl[\mathcal{T}^{TDA} + \mathcal{P}^{TDA}\Bigl],
\label{eq:TDA}
\end{eqnarray}
with
\begin{eqnarray}
{\cal T}^{TDA} &=& T~B_i (M)^{i}_j \bar H^{jl}_k (M)^k_l + C~B_i (M)^{i}_j \bar H^{lj}_k (M)^k_l
+ A~B_i \bar H^{il}_j (M)^j_k (M)^{k}_l\nonumber \\
&&+ E~B_i \bar H^{li}_j (M)^j_k (M)^{k}_l
+ T_{ES} B_i \bar H^{ij}_{l} (M)^{l}_j (M)^k_k
+T_{AS} B_i \bar H^{ji}_{l} (M)^{l}_j
(M)^k_k\nonumber\\
&&+ T_{S} B_i (M)^{i}_j \bar H^{lj}_{l} (M)^k_k + T_{PA} B_i \bar H^{li}_{l} (M)^j_k (M)^{k}_j +T_{P} B_i (M)^{i}_j (M)^j_k \bar H^{lk}_{l}\nonumber\\
&& + T_{SS} B_i \bar H^{li}_{l} (M)^j_j (M)^{k}_{k},\label{eq:new_tree}
\end{eqnarray}
and
\begin{eqnarray}
{\cal P}^{TDA}&=& P~B_i (M)^{i}_{j} (M)^{j}_k \tilde H^k
+ P_{T}~B_i (M)^{i}_j \tilde H^{jl}_k (M)^k_l
+ S~B_i (M)^{i}_{j} \tilde H^j (M)^{k}_k\nonumber\\
&&+ P_C~B_i (M)^{i}_j \tilde H^{lj}_k (M)^k_l
+ P_{TA} B_i \tilde H^{il}_j (M)^j_k (M)^{k}_l
+ P_A~B_i \tilde H^i (M)^{j}_{k} (M)^{k}_j \nonumber\\
&& + P_{TE} B_i \tilde H^{ji}_k (M)^k_l (M)^{l}_j
+P_{AS} B_i \tilde H^{ji}_{l} (M)^{l}_j (M)^k_k
+P_{SS} B_i \tilde H^i (M)^{j}_{j} (M)^{k}_k\nonumber\\
&& +P_{ES} B_i \tilde H^{ij}_{l} (M)^{l}_j (M)^k_k,
\label{eq:new_peng}
\end{eqnarray}
where all the indices can assume the values $1,2,3$.
The $B$-meson vector with components $(B_i)$ reads $B=(B^+,B^0_d,B^0_s)$ and the light-meson matrix $(M)^i_j$ has the structure
\begin{equation}
\label{eq:Mmatrix}
\begin{aligned}
M &= \left( \begin{array}{ccc}
\frac{\pi^0}{\sqrt 2} + \frac{\eta_q}{\sqrt 2}
+ \frac{\eta'_q}{\sqrt 2} & \pi^- & K^- \\
\pi^+ & - \frac{\pi^0}{\sqrt2} + \frac{\eta_q}{\sqrt 2}
+ \frac{\eta'_q}{\sqrt 2} & \bar K^0 \\
K^+ & K^0 & \eta_s + \eta'_s \end{array} \right).
\end{aligned}
\end{equation}
The states $\eta_{q}$, $\eta'_{q}$, $\eta_{s}$ and $\eta'_{s}$ appearing in the diagonal of the meson matrix in Eq.~(\ref{eq:Mmatrix}) correspond to the flavor states used to describe $\eta$~--~$\eta'$ mixing in the Feldmann–Kroll–Stech scheme, for more details see section~\ref{Sec:etamixing}.
The graphical representation of the topological amplitudes presented in Eq.~(\ref{eq:new_tree}) is given in Fig.~\ref{fig:Prop}.
Here, we have used the symbols $\tilde{H}$ and $\bar{H}$ to denote the flavour tensors, which allow to construct flavour singlets when contracted with the meson vector and matrices. Unlike the standard conventions here we are absorbing the different CKM-components inside the flavour tensors, thus we are using the symbol $``-"$ to indicate the presence of the component $\lambda^{(q)}_u$ and the symbol $``\sim"$ to indicate the presence of $\lambda^{(q)}_t$, for $q=d, s$.
\begin{figure}[tp]
\begin{center}
\includegraphics[width=1.0\textwidth]{Topo_1.pdf}
\includegraphics[width=1.0\textwidth]{Topo_2.pdf}
\includegraphics[width=0.5\textwidth]{Topo_3.pdf}
\caption{Diagrammatic representation of the tree-like topological amplitudes in Eq.~(\ref{eq:new_tree}).}
\label{fig:Prop}
\end{center}
\end{figure}
The non-zero elements of $\tilde{H}^{ij}_{k}$ and $\bar{H}^{ij}_{k}$ are then
\begin{eqnarray}
\bar{H}^{12}_1=\lambda^{(d)}_{u},&~&
\bar{H}^{13}_1=\lambda^{(s)}_{u},\nonumber\\
\tilde{H}^{12}_1=\lambda^{(d)}_{t},&~&
\tilde{H}^{13}_1=\lambda^{(s)}_{t}.
\end{eqnarray}
Analogously, for the vectors $\bar{H}^i$ and $\tilde{H}^i$ the non-zero components are
\begin{eqnarray}
\bar{H}^{2}=\lambda^{(d)}_{u},&~&
\bar{H}^{3}=\lambda^{(s)}_{u},\nonumber\\
\tilde{H}^{2}=\lambda^{(d)}_{t},&~&
\tilde{H}^{3}=\lambda^{(s)}_{t}.
\end{eqnarray}
\subsection{$SU(3)$-irreducible representation}
\label{sec:Irr}
Alternatively to the topological decomposition in Eq.~(\ref{eq:TDA}), the $SU(3)$-irreducible decomposition of the physical amplitudes reads
\begin{eqnarray}
{\cal A}^{IRA} =i\frac{G_{F}}{\sqrt{2}}\Bigl[ \mathcal{T}^{IRA} + \mathcal{P}^{IRA}\Bigl].
\label{eq:IRA}
\end{eqnarray}
The irreducible representation for the ``tree'' contribution to the amplitude for the $B\rightarrow P P$ decay processes is expressed as
\begin{eqnarray}\label{eq:su3basis}
\mathcal{T}^{IRA} &=&A_3^T B_i (\bar{H}_{\bar 3})^i (M)_k^j(M)_j^k +C_3^T B_i (M)^i_j (M)^j_k (\bar{H}_{\bar 3})^k+B_3^T B_i (\bar{H}_3)^i (M)_k^k(M)_j^j \nonumber\\
&&
+D_3^T B_i (M)^i_j (\bar{H}_{\bar 3})^j (M)^k_k+A_6^T B_i (H_{ 6})^{ij}_k (M)_j^l(M)_l^k
+C_6^T B_i (M)^i_j (\bar{H}_{ 6})^{jl}_k (M)_l^k \nonumber\\
&&
+B_6^T B_i (\bar{H}_{ 6})^{ij}_k (M)_j^k(M)_l^l+A_{15}^T B_i (\bar{H}_{\overline{15}})^{ij}_k (M)_j^l(M)_l^k
+C_{15}^T B_i (M)^i_j (\bar{H}_{\overline{15}})^{jk}_l (M)_k^l\nonumber\\
&&+B_{15}^T B_i (\bar{H}_{\overline{15}})^{ij}_k (M)_j^k(M)_l^l.
\end{eqnarray}
with
\begin{eqnarray}
(\bar{H}_3)^2=\lambda^{(d)}_u,\quad
(\bar{H}_6)^{12}_1=-(\bar{H}_6)^{21}_1=(\bar{H}_6)^{23}_3=-(H_6)^{32}_3&=&\lambda^{(d)}_u \, , \nonumber\\
2(\bar{H}_{15})^{12}_1=2(\bar{H}_{15})^{21}_1=-3(\bar{H}_{15})^{22}_2=-6(\bar{H}_{15})^{23}_3=-6(\bar{H}_{15})^{32}_3&=&6\lambda^{(d)}_u \, ,
\end{eqnarray}
and
\begin{eqnarray}
(\bar{H}_3)^3=\lambda^{(s)}_u,\quad
(\bar{H}_6)^{13}_1=-(\bar{H}_6)^{31}_1=(\bar{H}_6)^{32}_3=-(H_6)^{23}_3&=&\lambda^{(s)}_u \, , \nonumber\\
2(\bar{H}_{15})^{13}_1=2(\bar{H}_{15})^{31}_1=-3(\bar{H}_{15})^{33}_3=-6(\bar{H}_{15})^{32}_3=-6(\bar{H}_{15})^{23}_3&=&6\lambda^{(s)}_u \, ,
\end{eqnarray}
where the subindices in the different tensors denote the $SU(3)$-multiplet under consideration. The penguin components of the amplitude $\mathcal{P}^{IRA}$ can be obtained from $\mathcal{T}^{IRA}$ via the replacements
\begin{eqnarray}
A^T_{j}\rightarrow A^P_{j},\quad C^T_{j}\rightarrow C^P_{j}, \quad B^T_{j}\rightarrow B^P_{j} \quad D^T_{j}\rightarrow D^P_{j}\label{eq:treepengSU3}
\end{eqnarray}
and
\begin{eqnarray}
\bar{H}_3\rightarrow
\tilde{H}_3,\quad
\bar{H}_6\rightarrow
\tilde{H}_{6},\quad
\bar{H}_{15}\rightarrow
\tilde{H}_{15}\, ,
\end{eqnarray}
with the non-zero components given by
\begin{eqnarray}
(\tilde{H}_3)^2=\lambda^{(d)}_t,\quad
(\tilde{H}_6)^{12}_1=-(\tilde{H}_6)^{21}_1=(\tilde{H}_6)^{23}_3=-(\tilde{H}_6)^{32}_3&=&\lambda^{(d)}_t \, ,\nonumber\\
2(\tilde{H}_{15})^{12}_1=2(\tilde{H}_{15})^{21}_1=-3(\tilde{H}_{15})^{22}_2=-6(\tilde{H}_{15})^{23}_3=-6(\tilde{H}_{15})^{32}_3&=&6\lambda^{(d)}_t \, ,
\end{eqnarray}
and
\begin{eqnarray}
(\tilde{H}_3)^3=\lambda^{(s)}_t,\quad
(\tilde{H}_6)^{13}_1=-(\bar{H}_6)^{31}_1=(\bar{H}_6)^{32}_3=-(H_6)^{23}_3&=&\lambda^{(s)}_t \, , \nonumber\\
2(\tilde{H}_{15})^{13}_1=2(\tilde{H}_{15})^{31}_1=-3(\tilde{H}_{15})^{33}_3=-6(\tilde{H}_{15})^{32}_3=-6(\tilde{H}_{15})^{23}_3&=&6\lambda^{(s)}_t.
\end{eqnarray}
Interestingly, once the physical amplitudes are computed, the component $A^T_6$ appears always in the combinations $C^T_6-A^T_6$ and $B^T_6 + A^T_6$ and consequently it can always be absorbed according to the following rules \cite{Paz:2002ev},
\begin{eqnarray}
C^T_6-A^T_6\rightarrow C^T_6,\nonumber\\
B^T_6 + A^T_6 \rightarrow B^T_6,
\label{eq:reparameterization}
\end{eqnarray}
and analogously for the case of the corresponding penguin amplitudes,
\begin{eqnarray}
C^P_6-A^P_6\rightarrow C^P_6,\nonumber\\
B^P_6 + A^P_6 \rightarrow B^P_6.
\label{eq:reparameterizationpeng}
\end{eqnarray}
In other words, we redefine $C_6$ and $B_6$ to include $A_6$ for ``trees'' and ``penguins'', respectively. Here, we follow~\cite{He:2018php} in the decomposition of the color allowed and color suppressed electroweak penguin operators $P_{EW}$ and $P^{C}_{EW}$ according to the rule
\begin{eqnarray}
\bar{q}b\sum_{q'} e_{q'}\bar{q}'q'=\bar{q}b\bar{u}u+\bar{q}b\bar{c}c-\frac{1}{3}\bar{q}b\sum_{q'}\bar{q}'q',\label{eq:splitEW}
\end{eqnarray}
where the first two terms and their color suppressed analog lead to the amplitudes $P_T$ and $P_C$, respectively. The remaining part of Eq.~(\ref{eq:splitEW}) is then absorbed into the penguin amplitude $P$. Thus, we have
\begin{eqnarray}
P_{EW}&=&P_T -\frac{1}{3}P',\\
P^{C}_{EW}&=&P_{C}-\frac{1}{3}P^{'C}.
\end{eqnarray}
We proceed by counting the total number of real degrees of freedom. In total we have 20 complex coefficients which appear in the tree and penguin contributions. This translates into a total of 40 real quantities. Once the parameter shifts in Eq.~(\ref{eq:reparameterization}) and Eq.~(\ref{eq:reparameterizationpeng}) are performed we are left with 36 real quantities. However, an overall phase can always be absorbed and is irrelevant for the modulus of the physical amplitudes. Hence we are left with 35 real coefficients. Once we include the $\eta-\eta'$ mixing angle, the total count increases to 36.
\begin{table}[tp]
\begin{center}
\scalebox{0.85}{\begin{tabular}{|c|c|c|c|c|c|c|c|c|c|c|}
\hline
\textbf{Channel} & $A_{3}^T$& $C_{3}^T$&$A_{6}^T$&$C_{6}^T$&$A_{15}^T$&$C_{15}^T$&$B_{3}^T$&$B_{6}^T$&$B_{15}^T$ &$D_{3}^T$ \\
\hline
$B^-\rightarrow \pi^0\pi^-$ & $0$ & $0$& $0$&$0$&$0$&$4\sqrt{2}$&$0$&$0$&$0$&$0$\\
$B^-\rightarrow K^0 K^-$ & $0$ &$1$ &$1$&$-1$&$3$&$-1$&$0$&$0$&$0$&$0$\\
$B^0\rightarrow \pi^+ \pi^-$& $2$ & $1$ &$-1$ &$1$ &$1$&$3$&$0$&$0$&$0$&$0$\\
$B^0\rightarrow \pi^0 \pi^0$& $2$&$1$&$-1$&$1$&$1$&$-5$&$0$&$0$&$0$&$0$ \\
$B^0\rightarrow K^+ K^-$& $2$&$0$&$0$&$0$&$2$&$0$&$0$&$0$&$0$&$0$\\
$B^0\rightarrow K^0 \bar{K}^0$& $2$& $1$& $1$& $-1$& $-3$& $-1$& $0$& $0$&$0$&$0$\\
$B_s\rightarrow \pi^0 K^0$& 0&$-\frac{1}{\sqrt{2}}$&$\frac{1}{\sqrt{2}}$&$-\frac{1}{\sqrt{2}}$&
$\frac{1}{\sqrt{2}}$&$\frac{5}{\sqrt{2}}$&$0$&$0$&$0$&$0$ \\
$B_s\rightarrow \pi^- K^+$&$0$&$1$&$-1$&$1$&$-1$&$3$&$0$&$0$&$0$&$0$\\
$B^-\rightarrow \pi^0 K^-$& 0&$\frac{1}{\sqrt{2}}$& $\frac{1}{\sqrt{2}}$& $-\frac{1}{\sqrt{2}}$&$\frac{3}{\sqrt{2}}$&$\frac{7}{\sqrt{2}}$&$0$&$0$&$0$&$0$\\
$B^-\rightarrow \pi^- K^0$& $0$&$1$&$1$&$-1$&$3$&$-1$&$0$&$0$&$0$&$0$\\
$B^0\rightarrow \pi^+ K^-$&$0$&$1$&$-1$&$1$&$-1$&$3$&$0$&$0$&$0$&$0$\\
$B^0\rightarrow \pi^0 K^0$&0&$-\frac{1}{\sqrt{2}}$&
$\frac{1}{\sqrt{2}}$&$-\frac{1}{\sqrt{2}}$&$\frac{1}{\sqrt{2}}$&$\frac{5}{\sqrt{2}}$&$0$&$0$&$0$&$0$\\
$B_s\rightarrow \pi^+ \pi^-$&$2$&$0$&$0$&$0$&$2$&$0$&$0$&$0$&$0$&$0$\\
$B_s\rightarrow \pi^0\pi^0$&$2$&$0$&$0$&$0$&$2$&$0$&$0$&$0$&$0$&$0$\\
$B_s\rightarrow K^+ K^-$&$2$&$1$&$-1$&$1$&$1$&$3$&$0$&$0$&$0$&$0$\\
$B_s\rightarrow K^0 \bar{K}^0$& $2$&$1$&$1$&$-1$&$-3$&$-1$&$0$&$0$&$0$&$0$\\
\hline
\end{tabular}}
\caption{$B\rightarrow P P$ amplitudes decomposition in the $SU(3)$-basis. By virtue of Eq.~(\ref{eq:treepengSU3}) the corresponding numbers for the penguin components coincide with the tree ones. Where applicable, the coefficients agree with~\cite{He:2018php}. \label{tab:tableAmplBPP}}
\end{center}
\end{table}
To make the present paper self-contained we give in Tables~\ref{tab:tableAmplBPP} and \ref{tab:tableAmplBPPeta} the decay amplitudes for $B\rightarrow P P$ transitions in the $SU(3)$-de\-composition.
\begin{table}[tp]
\begin{center}
\scalebox{0.85}{\begin{tabular}{|c|c|c|c|c|c|c|c|c|c|c|}
\hline
\textbf{Channel} & $A_{3}^T$& $C_{3}^T$&$A_{6}^T$&$C_{6}^T$&$A_{15}^T$&$C_{15}^T$&$B_{3}^T$&$B_{6}^T$&$B_{15}^T$ &$D_{3}^T$\\
\hline
$B^-\rightarrow \eta_q \pi^-$&0& $\sqrt{2}$& $\sqrt{2}$& 0& $3\sqrt{2}$& $2\sqrt{2}$&
0& $\sqrt{2}$& $3\sqrt{2}$& $\sqrt{2}$ \\
$B^-\rightarrow \eta_s \pi^-$&$0$& $0$& $0$& $1$& $0$& $-1$& $0$& $1$& $3$& $1$\\
$B^0\rightarrow \eta_q \pi^0$&0& $-1$& $-1$& $0$& $5$& $2$& $0$& $-1$& $5$& $-1$\\
$B^0\rightarrow \eta_s \pi^0$&$0$& $0$& $0$& $-\frac{1}{\sqrt{2}}$& 0& $\frac{1}{\sqrt{2}}$& 0&$-\frac{1}{\sqrt{2}}$&$\frac{5}{\sqrt{2}}$& $-\frac{1}{\sqrt{2}}$\\
$B_s\rightarrow \eta_q K^0$&0& $\frac{1}{\sqrt{2}}$& $-\frac{1}{\sqrt{2}}$& $-\frac{1}{\sqrt{2}}$&$-\frac{1}{\sqrt{2}}$& $\frac{1}{\sqrt{2}}$& 0& $-\sqrt{2}$&$-\sqrt{2}$& $\sqrt{2}$\\
$B_s\rightarrow \eta_s K^0$&$0$& $1$& $-1$& $0$& $-1$& $-2$& $0$& $-1$& $-1$& $1$\\
$B^-\rightarrow \eta_q K^-$&$0$& $\frac{1}{\sqrt{2}}$& $\frac{1}{\sqrt{2}}$& $\frac{1}{\sqrt{2}}$&$\frac{3}{\sqrt{2}}$& $\frac{5}{\sqrt{2}}$& $0$& $\sqrt{2}$& $3\sqrt{2}$&$\sqrt{2}$\\
$B^-\rightarrow \eta_s K^-$&$0$& $1$& $1$& $0$& $3$& $-2$& $0$& $1$& $3$& $1$\\
$B^0\rightarrow \eta_q K^0$&$0$& $\frac{1}{\sqrt{2}}$& $-\frac{1}{\sqrt{2}}$& $-\frac{1}{\sqrt{2}}$&$-\frac{1}{\sqrt{2}}$& $\frac{1}{\sqrt{2}}$& $0$& $-\sqrt{2}$&
$-\sqrt{2}$& $\sqrt{2}$\\
$B^0\rightarrow \eta_s K^0$&$0$&$1$&$-1$& $0$& $-1$& $-2$& $0$& $-1$& $-1$& $1$\\
$B_s\rightarrow \eta_q \pi^0$&$0$& $0$& $-2$& $0$& $4$& $0$& $0$& $-2$& $4$& $0$\\
$B_s\rightarrow \eta_s \pi^0$&$0$& $0$& $0$& $-\sqrt{2}$& $0$& $2\sqrt{2}$& $0$&$-\sqrt{2}$& $2\sqrt{2}$& $0$ \\
$B^0\rightarrow \eta_q \eta_q$&1&$\frac{1}{2}$& -$\frac{1}{2}$& $-\frac{1}{2}$& $\frac{1}{2}$&$\frac{1}{2}$& $2$& $-1$& $1$& $1$\\
$B^0\rightarrow \eta_q \eta_s$&$0$& $0$& $0$& $\frac{1}{\sqrt{2}}$& $0$& $-\frac{1}{\sqrt{2}}$& $2\sqrt{2}$& $\frac{1}{\sqrt{2}}$&-$\frac{1}{\sqrt{2}}$& $\frac{1}{\sqrt{2}}$ \\
$B^0\rightarrow \eta_s\eta_s$&$1$& $0$& $1$& $0$& $-1$& $0$& $1$& $1$& $-1$& $0$\\
$B_s\rightarrow \eta_q\eta_q$&$1$& $0$& $0$& $0$& $1$& $0$& $2$& $0$& $2$& $0$\\
$B_s\rightarrow \eta_q\eta_s$&$0$& $0$& $0$& $0$& $0$& $\sqrt{2}$& $2\sqrt{2}$& $0$&$-\sqrt{2}$& $\sqrt{2}$\\
$B_s\rightarrow \eta_s\eta_s$&$1$& $1$& $0$& $0$& $-2$& $-2$& $1$& $0$& $-2$& $1$\\
\hline
\end{tabular}}
\caption{$B\rightarrow P P$ amplitudes decomposition in the $SU(3)$-basis including singlets. Also here the corresponding numbers for the penguin components coincide with the tree ones.\label{tab:tableAmplBPPeta}}
\end{center}
\end{table}
\subsection{Connecting $SU(3)$-invariant and topological decomposition}
\label{sec:connect}
The transformation rules between the topological basis and the $SU(3)$-invariant one for trees are given by~\cite{He:2018php}
\begin{align}
A_3^T&= -\frac{A}{8} + \frac{3E}{8}+T_{PA}, &
B_3^T&= T_{SS} +\frac{3T_{AS}-T_{ES}}{8},\nonumber\\
C_3^T&= \frac{1}{8} ({3A-C-E+3T})+T_P, & D_3^T&= T_{S} +\frac{1}{8} (3C-T_{AS}+3T_{ES}-T), \nonumber\\
A_6^T&= \frac{1}{4}(A-E), &
B_6^T&= \frac{1}{4}(T_{ES}-T_{AS}),\nonumber\\
C_6^T&= \frac{1}{4}(-C+T), &
A_{15}^T&= \frac{A+E}{8}, \nonumber\\
B_{15}^T&= \frac{T_{ES}+T_{AS}}{8}, &
C_{15}^T&= \frac{C+T}{8},
\end{align}
and for penguins by
\begin{align}
A_3^P&= -\frac{P_{TA}}{8} + \frac{3 P_{TE}}{8}+P_{A}, &
B_3^P&= P_{SS} +\frac{3P_{AS}-P_{ES}}{8},\nonumber\\
C_3^P&= \frac{1}{8} ({3 P_{TA}-P_{C}-P_{TE}+3P_{T}})+ P, & D_3^P&= P_{S} +\frac{1}{8} (3P_{C}-P_{AS}+3P_{ES}-P), \nonumber\\
A_6^P&= \frac{1}{4}(P_{A}-P_{E}), &
B_6^P&= \frac{1}{4}(P_{ES}-P_{AS}),\nonumber\\
C_6^P&= \frac{1}{4}(-P_{C}+P_{T}), &
A_{15}^P&= \frac{P_A+P_E}{8}, \nonumber\\
B_{15}^P&= \frac{P_{ES}+P_{AS}}{8}, &
C_{15}^P&= \frac{P_C+P_T}{8}.
\end{align}
For convenience we also give the inverse rules which allow to convert the topological basis into the $SU(3)$-invariant one. For the tree-like amplitudes, they read,
\begin{eqnarray}
\label{eq:ToptoSU3}
T&=& 2 C_6^T +4C_{15}^T, \;\; C= 4C_{15}^T -2C_{6}^T, \;\; A= 2A_6^T +4A_{15}^T,\;\; E= 4A_{15}^T-2A_6^T,\nonumber\\
T_{P}&=& -A_{6}^T-A_{15}^T+C_{3}^T-C_{6}^T-C_{15}^T, \;\; T_{PA}= A_{3}^T+A_{6}^T -A_{15}^T,\;\; T_{AS}= 4B_{15}^T -2B_{6}^T, \nonumber\\
T_{ES}&=& 2B_6^T +4B_{15}^T, \;\; T_{SS}=B_{3}^T+B_{6}^T -B_{15}^T,\;\;T_{S}=-B_{6}^T-B_{15}^T+C_{6}^T -C_{15}^T +D_{3}^T.\nonumber\\
\end{eqnarray}
In the case of the penguins the rules can easily be derived if the following substitutions are applied
\begin{align}
T&\rightarrow P_{T}, & C & \rightarrow P_{C}, &
A&\rightarrow P_{TA},& T_{P}&\rightarrow P, &
E&\rightarrow P_{TE},\nonumber\\
T_{PA}&\rightarrow P_{A}, &
T_{AS}&\rightarrow P_{AS},&
T_{ES}&\rightarrow P_{ES}, &
T_{SS}&\rightarrow P_{SS}, &
T_{S}&\rightarrow P_{S}.
\end{align}
\section{Amplitudes in QCD factorization}
\label{sec:QCDF}
The calculation of the amplitude for $B\rightarrow M_1 M_2$ in QCD factorization can be obtained by applying the general formula~\cite{Beneke:2003zv}
\begin{eqnarray}
&&\hspace{-1.0truecm}
\mathcal{A}^{\rm{QCDF}}=i\frac{G_{F}}{\sqrt{2}}\sum_{p=u,c}\,A_{M_1 M_2}\,\bigg\{
B M_1 \left( \alpha_1\delta_{pu}\hat{U} +
\alpha_4^p\hat{I}
+ \alpha_{4,EW}^p\hat{Q} \right) M_2\,
\Lambda_p \nonumber\\
&&\qquad\mbox{}+ B M_1 \Lambda_p\cdot
\mbox{Tr}\left[\left( \alpha_2\delta_{pu} \hat{U} + \alpha_3^p\hat{I}
+ \alpha_{3,EW}^p\,\hat {Q} \right) M_2\right] \nonumber\\
&&\qquad\mbox{}+ B \left( \beta_2\delta_{pu} \hat{U} + \beta_3^p\hat{I}
+ \beta_{3,EW}^p\,\hat{Q} \right)
M_1 M_2\Lambda_p \nonumber\\
&&\qquad\mbox{}+ B \Lambda_p\cdot
\mbox{Tr}\left[\left( \beta_1 \delta_{pu} \hat{U} + \beta_4^p\hat{I}
+ b_{4,EW}^p\,\hat{Q} \right) M_1 M_2\right]
\nonumber\\
&&\qquad\mbox{}+ B \left( \beta_{S2}\delta_{pu}\hat{U} + \beta_{S3}^p\hat{I}
+ \beta_{S3,EW}^p\,\hat{Q} \right)
M_1\Lambda_p\cdot\mbox{Tr}M_2 \nonumber\\
&&\qquad\mbox{}+ B\Lambda_p\cdot
\mbox{Tr}\left[\left( \beta_{S1}\delta_{pu}\hat{U} + \beta_{S4}^p\hat{I}
+ b_{S4,EW}^p\hat{Q} \right) M_1\right]\cdot
\mbox{Tr}M_2 \bigg\} \,,
\label{eq:mastereq}
\end{eqnarray}
where
\begin{align}
\Lambda_p &= \left( \begin{array}{c}
0 \\ \lambda_p^{(d)} \\ \lambda_p^{(s)}
\end{array} \right), & \hat{U} &= \left( \begin{array}{ccc}
1~ & 0~ & 0 \\
0~ & 0~ & 0 \\
0~ & 0~ & 0
\end{array} \right) , \\[0.2em]
\hat {Q} &= \frac{3}{2} Q = \left( \begin{array}{ccc}
1 & 0 & 0 \\
0 & -\frac12 & 0 \\
0 & 0 & -\frac12
\end{array} \right), &
\hat{I} &=\left( \begin{array}{ccc}
1~ & 0~ & 0 \\
0~ & 1~ & 0 \\
0~ & 0~ & 1
\end{array} \right),
\end{align}
and the meson vector $B$ and matrix $M$ coincide with the ones introduced in section~\ref{sec:Top}. In view of the connection to the $SU(3)$-invariant decomposition that we establish below, we drop the dependence of the QCDF amplitudes $\{\alpha_i,\alpha^{p}_{i}, \beta_{i},b_i\}$ on the final-state mesons from now on. The coefficient $A_{M_1 M_2}$ corresponds to
\begin{eqnarray}
A_{M_1 M_2}&=& M^2_B \, F^{B\rightarrow M_1}_0(0) \, f_{M_2}.
\label{eq:AM1M2}
\end{eqnarray}
It is relevant for our purposes to note that the general formula Eq.~(\ref{eq:mastereq}) consists of the contraction of the meson matrices $B$, $M_1$ and $M_2$ with the generic structure
\begin{eqnarray}
\tilde{\hat{C}}_{r}&=&\sum_{p=u,c}\Bigl[(\tilde{T}_{r}\delta_{pu} \hat{U} + \tilde{P}^{(1),p}_{r} \hat{I} +\tilde{P}^{(2),p}_{r}\hat{Q})\otimes\Lambda_p\Bigl]
\label{eq:amplstr}
\end{eqnarray}
where
\begin{eqnarray}
\tilde{T}_{k}&\in&\{\alpha_1,\alpha_2,\beta_2,\beta_1, \beta_{S2}, \beta_{S1}\},\nonumber\\
\tilde{P}^{(1), p}_{k}&\in&\{\alpha^p_4, \alpha^p_3, \beta^p_3, \beta^p_4, \beta^p_{S3},\beta^p_{S4}\},\nonumber\\
\tilde{P}^{(2), p}_{k} &\in& \{\alpha^p_{4, EW}, \alpha^p_{3, EW},\beta^p_{3, EW}, b^p_{4, EW},\beta^p_{S3, EW}, b^p_{S4,EW}\}.\quad\quad\quad
\end{eqnarray}
Thus, Eq.~(\ref{eq:mastereq}) can be written in terms of indices as
\begin{eqnarray}
\mathcal{A}^{\rm{QCDF}}&=&i\frac{G_{F}}{\sqrt{2}}
A_{M_1 M_2}\Bigl\{B_i M^{i}_{j}(\tilde{\hat{C}}_{1})^{j l}_{k}M^{k}_{l}
+
B_i M^{i}_{j}(\tilde{\hat{C}}_{2})^{lj }_{k}M^{k}_{l}
+
B_i(\tilde{\hat{C}}_{3})^{ij }_{k}M^{k}_{l}M^{l}_{j}\nonumber\\
&&\qquad\qquad+B_i (\tilde{\hat{C}}_{4})^{li}_{k} M^{k}_{r}M^{r}_{l} + B_i (\tilde{\hat{C}}_{5})^{ij}_{k}M^{k}_{i}M^{l}_{l}+ B_i (\tilde{\hat{C}}_{6})^{ji}_{k})M^{k}_{j}M^{l}_{l}\Bigl\},
\end{eqnarray}
where we have omitted the indices $1$ and $2$ in the meson matrices and we have identified
\begin{eqnarray}
(\tilde{\hat{C}}_{r})^{ij}_{k}&=&\sum_{p=u,c}\Bigl[(\tilde{T}_{r}\delta_{pu} \hat{U} + \tilde{P}^{(1),p}_{r} \hat{I} +\tilde{P}^{(2),p}_{r}\hat{Q})^{i}_{k}\Lambda^{j}_p\Bigl].
\label{eq:Cmatrix}
\end{eqnarray}
\section{Equivalence of QCDF and the topological basis}
\label{sec:QCDFTop}
Our aim is to establish a connection between the amplitudes $\{\alpha_i,\alpha^{p}_{i}, \beta_{i},b_i\}$ appearing in the QCDF formalism inside Eq.~(\ref{eq:mastereq}) and the topological/$SU(3)$-irreducible representations as presented in Eqs.~(\ref{eq:new_tree}), (\ref{eq:new_peng}) and (\ref{eq:su3basis}). The desired transformation formulas are most conveniently obtained first between the topological and the QCDF expressions. Once this is done, the relationships between the QCDF amplitudes and the $SU(3)$-irreducible ones can be easily calculated since we already know the transformation rules between the topological and the $SU(3)$ case.
To obtain the desired conversion rules we
first decompose the matrix $\hat{Q}$ in terms of $\hat{U}$ and $\hat{I}$
\begin{eqnarray}
\hat{Q}&=&\frac{3}{2}\hat{U}-\frac{1}{2}\hat{I}.
\label{eq:Qdecomposition}
\end{eqnarray}
Then, we notice that one of the key differences between the topological and the QCDF decomposition are the CKM factors. In the first case the amplitudes are written in terms of $\lambda^{(q)}_u$ and $\lambda^{(q)}_t$, for $q=d,s$. However in the second case this is done through the factors $\lambda^{(q)}_u$ and $\lambda^{(q)}_c$ encoded inside the vectors $\Lambda_u$ and $\Lambda_c$. Therefore, we translate the unitarity rule in Eq.~(\ref{eq:CKMrel}) in terms of the CKM $\Lambda$-vectors and express $\Lambda_t$ in terms of $\Lambda_u$ and $\Lambda_c$ via
\begin{eqnarray}
\Lambda_t=-\Lambda_u - \Lambda_c.
\label{eq:CKMLambda}
\end{eqnarray}
Finally, we substitute Eq.~(\ref{eq:Qdecomposition}) and Eq.~(\ref{eq:CKMLambda}) inside the amplitude in Eq.~(\ref{eq:amplstr}) to arrive at
\begin{eqnarray}
\tilde{\hat{C}}_{r}&=&\Bigl[\tilde{T} +\frac{3}{2}\tilde{P}^u_2 - \frac{3}{2}\tilde{P}^c_2\Bigl]\hat{U}\otimes\Lambda_u + \Bigl[\tilde{P}^u_1-\tilde{P}^c_1-\frac{1}{2}\Bigl\{\tilde{P}^u_2-\tilde{P}^c_2\Bigl\}\Bigl]\hat{I}\otimes\Lambda_u\nonumber\\
&&-\frac{3}{2}\tilde{P}^c_2 \hat{U}\otimes\Lambda_t -\Bigl[ \tilde{P}^c_1 -\frac{\tilde{P}^c_2}{2}\Bigl]\hat{I}\otimes\Lambda_t,
\label{eq:amplstrexpanded}
\end{eqnarray}
or in terms of components
\begin{eqnarray}
(\tilde{\hat{C}}_{r})^{ij}_{k}&=&\Bigl[\tilde{T} +\frac{3}{2}\tilde{P}^u_2 - \frac{3}{2}\tilde{P}^c_2\Bigl]\hat{U}^{i}_{k}(\Lambda_u)^{j} + \Bigl[\tilde{P}^u_1-\tilde{P}^c_1-\frac{1}{2}\Bigl\{\tilde{P}^u_2-\tilde{P}^c_2\Bigl\}\Bigl]\delta^{i}_{k}(\Lambda_u)^{j}\nonumber\\
&&-\frac{3}{2}\tilde{P}^c_2 \hat{U}^{i}_{k}(\Lambda_t)^{j} -\Bigl[ \tilde{P}^c_1 -\frac{\tilde{P}^c_2}{2}\Bigl]\delta^{i}_{k}(\Lambda_t)^{j}.
\label{eq:amplstrexpandedcomp}
\end{eqnarray}
We are now ready to establish the equivalence between the QCD-factorization little amplitude basis and the topological decomposition basis discussed in Sec.~\ref{sec:TopIrr}. This can easily be done if we
make the following identifications with the different tensors appearing in Eq.~(\ref{eq:new_tree}) and Eq.~(\ref{eq:new_peng})
\begin{eqnarray}
U^i_k(\Lambda_u)^j=\bar{H}^{ij}_k,\quad\quad
U^i_k(\Lambda_t)^j=\tilde{H}^{ij}_k,
\quad\quad
(\Lambda_t)^i=\tilde{H}^i.
\end{eqnarray}
Our set of transformation rules finally reads
\begin{align}
T&=A_{M_{1}M_{2}}\Bigl[\alpha_1 + \frac{3}{2}\alpha^{u}_{4,EW}
-\frac{3}{2}\alpha^{c}_{4,EW}\Bigl],&
C&=A_{M_{1}M_{2}}\Bigl[\alpha_2+\frac{3}{2}\alpha^{u}_{3,EW}
-\frac{3}{2}\alpha^{c}_{3,EW}\Bigl],\nonumber\\
E&=A_{M_1 M_2}\Bigl[\beta_1+\frac{3}{2}b^{u}_{4,EW}-\frac{3}{2}b^{c}_{4,EW}\Bigl], &
A&=A_{M_1 M_2}\Bigl[\beta_{2}+\frac{3}{2}\beta^{u}_{3,EW}-\frac{3}{2}\beta^{c}_{3,EW}\Bigl],\nonumber\\
T_{AS}&=A_{M_1 M_2}\Bigl[\beta_{S1}+\frac{3}{2}b^{u}_{S4,EW}-\frac{3}{2}b^{c}_{S4,EW} \Bigl],&
T_{ES}&=A_{M_1 M_2}\Bigl[\beta_{S2}+\frac{3}{2}\beta^u_{S3, EW}-\frac{3}{2}\beta^c_{S3, EW}\Bigl],\nonumber
\end{align}
\begin{eqnarray}
T_{PA}&=&A_{M_1 M_2}\Bigl[\beta^{u}_{4}-\beta^{c}_{4}-\Bigl(\frac{b^{u}_{4,EW}}{2}-\frac{b^{c}_{4,EW}}{2}\Bigl)\Bigl], \nonumber\\
T_{SS}&=&A_{M_1 M_2}\Bigl[\beta^{u}_{S4}-\beta^{c}_{S4}-\Bigl(\frac{b^{u}_{S4,EW}}{2}-\frac{b^{c}_{S4,EW}}{2}\Bigl)\Bigl],\nonumber\\
T_{S}&=&A_{M_1 M_2}\Bigl[\alpha^{u}_{3}-\alpha^{c}_{3}-\Bigl(\frac{\alpha^{u}_{3,EW}}{2}-\frac{\alpha^{c}_{3,EW}}{2}\Bigl)+\Bigl(\beta^{u}_{S3}-\beta^{c}_{S3}\Bigl)-\Bigl(\frac{\beta^{u}_{S3,EW}}{2}-\frac{\beta^{c}_{S3,EW}}{2}\Bigl)\Bigl],\nonumber\\
T_{P}&=&A_{M_1 M_2}\Bigl[\alpha^{u}_{4}-\alpha^{c}_{4}-\Bigl(\frac{\alpha^{u}_{4, EW}}{2}-\frac{\alpha^{c}_{4, EW}}{2}\Bigl)+\Bigl(\beta^{u}_{3}-\beta^{c}_{3}\Bigl)-\Bigl(\frac{\beta^{u}_{3,EW}}{2}-\frac{\beta^{c}_{3,EW}}{2}\Bigl)\Bigl],\nonumber\\
S&=&-A_{M_1 M_2}\Bigl[\alpha^{c}_{3}+\beta^{c}_{S3}-\frac{\alpha^{c}_{3, EW}}{2}-\frac{\beta_{S3, EW}}{2}\Bigl],\nonumber\\
P&=&-A_{M_1 M_2}\Bigl[\alpha^{c}_{4}+\beta^{c}_{3}-\frac{\alpha^{c}_{4,EW}}{2}-\frac{\beta^{c}_{3,EW}}{2}\Bigl],\nonumber
\end{eqnarray}
\begin{eqnarray}
P_{A}=-A_{M_1 M_2}\Bigl[\beta^{c}_{4}-\frac{b^{c}_{4, EW}}{2}\Bigl],&~~&P_{SS}=-A_{M_1 M_2}\Bigl[\beta^{c}_{S4}-\frac{b^{c}_{S4, EW}}{2}\Bigl],\nonumber\\
P_{C}=-\frac{3}{2}A_{M_1 M_2}\alpha^{c}_{3,EW},&~~&P_{T}=-\frac{3}{2}A_{M_1 M_2}\alpha^{c}_{4,EW},\nonumber\\
P_{TA}=-\frac{3}{2}A_{M_1M_2}\beta^{c}_{3,EW},&~~&P_{TE}=-\frac{3}{2}A_{M_1M_2}b^{c}_{4,EW},\nonumber\\
P_{AS}=-\frac{3}{2}A_{M_1M_2}b^{c}_{S4,EW},&~~&P_{ES}=-\frac{3}{2}A_{M_1M_2}\beta^{c}_{S3, EW}.
\label{eq:QCDFtotopological}
\end{eqnarray}
One of the interesting features of the formulas in Eq.~(\ref{eq:QCDFtotopological}) is the fact that the flavour structure combines amplitudes with different physical content. For instance, consider the case of the structures $(\alpha_4-\alpha_{4,EW}/2)$ and $(\beta_3-\beta_{3,EW}/2)$ which appear multiplying the same factor $B_i M^{i}_{j} M^{j}_{k}\bar{H}^{lk}_{l}$,
where the first factor refers to penguin topologies, whereas the second to annihilation ones.\\
The formulas in Eqs.~(\ref{eq:QCDFtotopological}) can be simplified if we consider the following NLO results \cite{Beneke:2003zv}
\begin{eqnarray}
\alpha^{u}_{3}=\alpha^{c}_{3}=\alpha_{3},
\quad
\alpha^{u}_{3,EW}=\alpha^{c}_{3,EW}=\alpha_{3,EW},
\quad
\beta^{u}_{i}=\beta^{c}_{i}=\beta_{i},
\quad
b^{u}_i=b^{c}_i=b_i.
\label{eq:QCDFconditions}
\end{eqnarray}
Moreover, at NLO the amplitudes $\alpha^{p}_{4,EW}$ are at the permille level, such that the differences $|\alpha^{c}_{4,EW}-\alpha^{u}_{4,EW}|$ are at most $\mathcal{O}(10^{-3})$. Analogously for $\alpha^{p}_{4}$ the magnitudes are about $0.1$, and the differences $|\alpha^{c}_{4}-\alpha^{u}_{4}|$ are about $0.02$. Then, we simplify Eqs.~(\ref{eq:QCDFtotopological}) as follows
\begin{align}
T&=A_{M_1 M_2}\alpha_1, &
C&= A_{M_1 M_2} \alpha_2, &
E&=A_{M_1 M_2} \beta_1, \nonumber\\
A&=A_{M_1 M_2} \beta _2, &
T_{AS}&=A_{M_1 M_2}\beta_{S1},&
T_{ES}&=A_{M_1 M_2}\beta_{S2},
\nonumber
\end{align}
\begin{eqnarray}
S&=&-A_{M_1 M_2}\Bigl[\alpha_{3}+\beta_{S3}-\frac{\alpha_{3, EW}}{2}-\frac{\beta_{S3, EW}}{2}\Bigl],\nonumber\\
P&=&-A_{M_1 M_2}\Bigl[\alpha^{c}_{4}+\beta_{3}-\frac{\alpha^{c}_{4,EW}}{2}-\frac{\beta_{3,EW}}{2}\Bigl],\nonumber
\end{eqnarray}
\begin{align}
P_A&=-A_{M_1 M_2}(\beta_4-\frac{b_{4, EW}}{2}), &
P_{SS}&=-A_{M_1 M_2}(\beta_{S4}-\frac{b_{S4, EW}}{2}),\nonumber\\
P_C&=-\frac{3}{2}A_{M_1 M_2}\alpha_{3, EW}, &
P_{T}&=-\frac{3}{2}A_{M_1 M_2}\alpha^{c}_{4, EW},\nonumber\\
P_{TA}&=-\frac{3}{2}A_{M_1 M_2}\beta_{3, EW}, &
P_{TE}&=-\frac{3}{2}A_{M_1 M_2}b_{4, EW},\nonumber\\
P_{AS}&=-\frac{3}{2}A_{M_1 M_2}b_{S4, EW},&
P_{ES}&=-\frac{3}{2}A_{M_1 M_2}\beta_{S3, EW},\nonumber
\end{align}
\begin{align}
T_{PA}&=0, & T_{SS}&=0, &T_{S}&=0, & |T_{P}|&<0.02.
\label{eq:TopologicalQCDF}
\end{align}
We realize that the expressions in Eq.~(\ref{eq:QCDFconditions}) simplify drastically the relationships between the topological and the QCDF decomposition of the physical amplitudes. For instance there are multiple cancellations between the annihilation
contributions $\beta_i$. Moreover, most of the extra contributions of the form
$\alpha^{c}_i-\alpha^{u}_i$ are expected to be subleading and after applying model dependent information can in most of the cases be neglected, with the only exception $\alpha^c_4-\alpha^u_4$.\\
We can now invert and solve for the QCDF little amplitudes in terms of the topological decomposition coefficients, and find the following results
\begin{align}
A_{M_1 M_2}\alpha_1&= T, & A_{M_1 M_2}\alpha_2&= C, & A_{M_1 M_2}\beta_1&= E, & A_{M_1 M_2}\beta_2 &= A,\nonumber\\
A_{M_1 M_2}\beta_{S1} &= T_{AS}, & A_{M_1 M_2}\beta_{S2}&=T_{ES}, &
A_{M_1 M_2}\alpha_{3, EW}&=-\frac{2}{3}P_C, &
A_{M_1 M_2}\alpha^{c}_{4, EW}&=-\frac{2}{3}P_{T},\nonumber
\end{align}
\begin{align}
A_{M_1 M_2}\beta_{3, EW}&=-\frac{2}{3}P_{TA},&
A_{M_1 M_2}b_{4, EW}&=-\frac{2}{3}P_{TE},&
A_{M_1 M_2}\beta_4&=-P_{A}+\frac{1}{3}P_{TE},\nonumber\\
A_{M_1 M_2}\beta_{S3, EW}&=-\frac{2}{3}P_{ES},&
A_{M_1 M_2}b_{S4, EW}&=-\frac{2}{3}P_{AS},&
A_{M_1 M_2}\beta_{S4}&=-P_{SS}+\frac{1}{3}P_{AS},\nonumber
\end{align}
\begin{align}
A_{M_1 M_2}\Bigl(\alpha_3 + \beta_{S3}\Bigl)&=-S-\frac{P_{C}}{3}-\frac{P_{ES}}{3},&
A_{M_1 M_2}\Bigl(\alpha^{c}_{4}+\beta_{3}\Bigl)&=-P-\frac{P_T}{3}-\frac{P_{TA}}{3}.
\label{eq:SU£toTop}
\end{align}
So far, we have decomposed the physical amplitudes in terms of the CKM factors $\lambda_u^{(q)}$ and $\lambda_t^{(q)}$, or equivalently $\Lambda_u$ and $\Lambda_t$. However, we can also try an alternative parameterization in terms of $\Lambda_u$ and $\Lambda_c$. Thus, we proceed in an analogous way as we did previously during the decomposition of the $\tilde{\hat{C}}_r$ factors in Eq.~(\ref{eq:amplstrexpanded}), but this time we do not use the CKM-unitary relationship. Our result is then
\begin{eqnarray}
\tilde{\hat{C}}_r&=&\Bigl[\tilde{T}_{r} + \frac{3}{2}\tilde{P}^{(2), u}_r\Bigl]\hat{U}\otimes \Lambda_{u}+\Bigl[\tilde{P}^{(1), u}_r-\frac{\tilde{P}^{(2), u}_r}{2}\Bigl]\hat{I}\otimes \Lambda_{u}\nonumber\\
&&+\frac{3}{2} \tilde{P}^{(2), c}_r \hat{U}\otimes \Lambda_{c}+ \Bigl[\tilde{P}^{(1), c}_r-\frac{\tilde{P}^{(2), c}_r}{2} \Bigl] \hat{I}\otimes \Lambda_{c}.
\end{eqnarray}
For the physical amplitude we write
\begin{eqnarray}
\mathcal{A}^{\rm QCDF}=i\frac{G_{F}}{\sqrt{2}} \, \sum_{p=u,c}&&
\Bigl(c^{{\rm QCDF},p}_{1}B M_1 \hat{U} M_2 \Lambda_p
+ c^{{\rm QCDF},p}_{2}B \hat{U} M_1 M_2 \Lambda_p \nonumber\\
&&+c^{{\rm QCDF},p}_{3} B M_1 M_2 \Lambda_p + c^{{\rm QCDF},p}_{4}
B M_1 \Lambda_p \mbox{Tr}\Bigl[\hat{U} M_2\Bigl] \nonumber\\
&&+ c^{{\rm QCDF},p}_5 B\hat{U} M_1 \Lambda_p \mbox{Tr}\Bigl[M_2\Bigl]
+ c^{{\rm QCDF},p}_6 B M_1 \Lambda_p \mbox{Tr}\Bigl[M_2\Bigl]\nonumber\\
&&+c^{{\rm QCDF}, p}_7 B \Lambda_p \mbox{Tr}\Bigl[\hat{U} M_1 M_2\Bigl]
+c^{{\rm QCDF}, p}_8 B\Lambda_p \mbox{Tr}\Bigl[\hat{U} M_1\Bigl]
\mbox{Tr}\Bigl[M_2\Bigl]\nonumber\\
&&+c^{{\rm QCDF}, p}_{9} B\Lambda_p \mbox{Tr}\Bigl[M_1 M_2\Bigl]
+ c^{{\rm QCDF}, p}_{10} B\Lambda_p \mbox{Tr}\Bigl[M_1\Bigl] \mbox{Tr}\Bigl[M_2\Bigl]\Bigl) \, .
\end{eqnarray}
We can identify the following set of independent coefficients proportional to $\Lambda_{u}$,
\begin{align}
c^{{\rm QCDF}, u}_{1}&=A_{M_1 M_2}\Bigl(\alpha_{1}+\frac{3}{2}\alpha^{u}_{4, EW}\Bigl), &
c^{{\rm QCDF}, u}_{7}&=A_{M_1 M_2}\Bigl(\beta_{1}+\frac{3}{2}b^{u}_{4, EW}\Bigl),\nonumber\\
c^{{\rm QCDF}, u}_{2}&=A_{M_1 M_2}\Bigl(\beta_{2}+\frac{3}{2}\beta^{u}_{3, EW}\Bigl), &
c^{{\rm QCDF}, u}_{8}&=A_{M_1 M_2}\Bigl(\beta_{S1}+\frac{3}{2}b^{u}_{S4, EW}\Bigl),\nonumber\\
c^{{\rm QCDF}, u}_{3}&= A_{M_1 M_2}\Bigl(\alpha^{u}_{4}-\frac{\alpha^{u}_{4, EW}}{2}+\beta^{u}_3-\frac{\beta^{u}_{3, EW}}{2}\Bigl), &
c^{{\rm QCDF}, u}_{9}&=A_{M_1 M_2}\Bigl(\beta^{u}_{4}-\frac{b^u_{4, EW}}{2}\Bigl),\nonumber\\
c^{{\rm QCDF}, u}_{4}&=A_{M_1 M_2}\Bigl(\alpha_2 +\frac{3}{2}\alpha^{u}_{3, EW}\Bigl), &
c^{{\rm QCDF}, u}_{10}&=A_{M_1 M_2}\Bigl(\beta^u_{S4}-\frac{b^u_{S4, EW}}{2}\Bigl),\nonumber\\
c^{{\rm QCDF}, u}_{5}&=A_{M_1 M_2}\Bigl(\beta_{S2}+\frac{3}{2}\beta^{u}_{S3, EW}\Bigl), & &\nonumber\\
c^{{\rm QCDF}, u}_{6}&=A_{M_1 M_2}\Bigl(\alpha^{u}_{3}-\frac{\alpha^{u}_{3, EW}}{2}+\beta^{u}_{S3}-\frac{\beta^{u}_{S3, EW}}{2}\Bigl), & & \label{eq:QCDF1}
\end{align}
and analogously for $\Lambda_{c}$,
\begin{align}
\allowdisplaybreaks
c^{{\rm QCDF}, c}_{1}&=\frac{3}{2}A_{M_1 M_2}\alpha^{c}_{4, EW}, &
c^{{\rm QCDF}, c}_{7}&=\frac{3}{2}A_{M_1 M_2} b^c_{4, EW}, \nonumber\\
c^{{\rm QCDF}, c}_{2}&=\frac{3}{2}A_{M_1 M_2}\beta^c_{3, EW}, &
c^{{\rm QCDF}, c}_{8}&=\frac{3}{2}A_{M_1 M_2}b^c_{S4, EW}, \nonumber\\
c^{{\rm QCDF}, c}_{3}&=A_{M_1 M_2}\Bigl(\alpha^{c}_{4}-\frac{\alpha^{c}_{4, EW}}{2}+\beta_{3}-\frac{\beta_{3, EW}}{2}\Bigl), &
c^{{\rm QCDF}, c}_{9}&=A_{M_1 M_2}\Bigl(\beta^{c}_{4}-\frac{b^c_{4, EW}}{2}\Bigl), \nonumber\\
c^{{\rm QCDF}, c}_{4}&=\frac{3}{2}A_{M_1 M_2}\alpha^{c}_{3, EW}, &
c^{{\rm QCDF}, c}_{10}&=A_{M_1 M_2}\Bigl(\beta^c_{S4}-\frac{b^c_{S4, EW}}{2}\Bigl).\nonumber\\
c^{{\rm QCDF}, c}_{5}&=\frac{3}{2}A_{M_1 M_2}\beta^c_{S3, EW}, && \nonumber\\
c^{{\rm QCDF}, c}_{6}&=A_{M_1 M_2}\Bigl(\alpha^{c}_{3}-\frac{\alpha^{c}_{3, EW}}{2}+\beta^c_{S3}-\frac{\beta^c_{S3, EW}}{2}\Bigl), && \label{eq:QCDF2}
\end{align}
Without considering any further relations between the amplitudes $c^{{\rm QCDF}, u}_{i}$ and $c^{{\rm QCDF}, c}_{i}$ we have in total 20 complex quantities and in principle 40 real degrees of freedom. However, to describe the phenomenological processes we only need 18 complex amplitudes.
To prove this, we consider the coefficient matrix resulting from expanding each one of the 34 transitions for $B\rightarrow P P$ in terms of the numerical coefficients $c^{{\rm QCDF}, u}_{i}$ analogous to those in Tables \ref{tab:tableAmplBPP} and \ref{tab:tableAmplBPPeta} given for the coefficients in the $SU(3)$ irreducible representation. It is easy to see that the rank of this matrix is $9$, meaning that once the phenomenology is taken into account not all the $c^{{\rm QCDF}, u}_{i}$ coefficients are independent and only 9 linear combinations of them are required. Analogous considerations can be done for the set of coefficients $c^{{\rm QCDF}, c}_{i}$ with the same outcome. Thus the total number of complex amplitudes required is $18$, implying 36 real degrees of freedom. This is in line with~\cite{beneketalk} where the fact that QCDF describes $B\to PP$ transitions in terms of $18$ independent complex amplitudes was already pointed out. The linear relations between the QCDF and topological amplitudes, however, are to the best of our knowledge worked out for the first time in the present article.
As argued before, an additional overall phase can be absorbed since it is not physically relevant. This can be done by taking one of the coefficients $c^{{\rm QCDF},u}_{i}$ or $c^{{\rm QCDF},c}_{i}$ to be real. Hence, if we consider Eqs.~(\ref{eq:QCDF1}) and~(\ref{eq:QCDF2}) and the absorption of the global phase, we see that the total number of meaningful real parameters is $35$. The counting leads to 36 real values once the $\eta-\eta'$ mixing angle is added. This is the same result as was obtained when doing the analysis in the $\Lambda_{u}$, $\Lambda_{t}$ decomposition.
Using Eqs. (\ref{eq:QCDFconditions}) we can simplify Eqs.~(\ref{eq:QCDF1}) and (\ref{eq:QCDF2}), we then obtain two extra conditions
\begin{eqnarray}
c^{{\rm QCDF}, c}_{9}=c^{{\rm QCDF}, u}_{9} \, ,& &c^{{\rm QCDF}, c}_{10}=c^{{\rm QCDF}, u}_{10}.
\end{eqnarray}
For completeness we invert the simplified expressions derived from Eqs.~(\ref{eq:QCDF1}) and~(\ref{eq:QCDF2}) leading to
\begin{align}
A_{M_1 M_2}\alpha^{c}_{4, EW}&=\frac{2}{3}c^{{\rm QCDF},c}_{1}, &
A_{M_1 M_2}\beta_{3, EW}&=\frac{2}{3}c^{{\rm QCDF},c}_{2}, &
A_{M_1 M_2}\alpha^{c}_{3, EW}&=\frac{2}{3}c^{{\rm QCDF}, c}_{4}, \nonumber\\
A_{M_1 M_2}\beta_{S3, EW}&=\frac{2}{3}c^{{\rm QCDF}, c}_{5}, &
A_{M_1 M_2}b_{4, EW}&=\frac{2}{3}c^{{\rm QCDF}, c}_{7}, &
A_{M_1 M_2}b_{S4, EW}&=\frac{2}{3}c^{{\rm QCDF}, c}_{8}, \nonumber
\end{align}
\begin{align}
A_{M_1 M_2}\beta_1&=c^{{\rm QCDF}, u}_{7} - c^{{\rm QCDF}, c}_{7}, &
A_{M_1 M_2}\beta_{S1}&=c^{{\rm QCDF}, u}_{8} - c^{{\rm QCDF}, c}_{8}, \nonumber\\
A_{M_1 M_2}\beta_{2}&=c_{2}^{{\rm QCDF}, u}-c_{2}^{{\rm QCDF}, c}, &
A_{M_1 M_2}\beta_{4}&=c^{{\rm QCDF}, u}_{9}+\frac{1}{3}c^{{\rm QCDF}, c}_{7},\nonumber\\
A_{M_1 M_2}\beta_{S4}&=c^{{\rm QCDF}, u}_{10}+\frac{1}{3}c^{{\rm QCDF}, c}_{8}, &
A_{M_1 M_2}\Bigl(\alpha_{1}+\frac{3}{2}\alpha^{u}_{4, EW}\Bigl)&=c^{{\rm QCDF}, u}_{1},\nonumber
\end{align}
\begin{align}
A_{M_1 M_2}\Bigl(\alpha^{c}_{4}+\beta_3 \Bigl)&=\frac{1}{3}c^{{\rm QCDF}, c}_{1}+\frac{1}{3}c^{{\rm QCDF}, c}_{2}+c^{{\rm QCDF}, c}_{3}, \nonumber\\
A_{M_1 M_2}\Bigl(\alpha^{c}_3+\beta_{S3}\Bigl)&=
\frac{1}{3}c^{{\rm QCDF}, c}_{4}+\frac{1}{3}c^{{\rm QCDF}, c}_{5}+c^{{\rm QCDF}, c}_{6},\nonumber\\
A_{M_1 M_2}\Bigl(\alpha^{u}_3+\beta_{S3}-\frac{1}{2}\alpha^{c}_{3, EW}\Bigl)&=
c^{{\rm QCDF}, u}_{6}+\frac{1}{3}c^{{\rm QCDF}, c}_{5}, \nonumber\\
A_{M_1 M_2}\Bigl(\alpha^{u}_4+\beta_3-\frac{1}{2}\alpha^{u}_{4, EW}\Bigl)&=c^{{\rm QCDF}, u}_{3}+\frac{1}{3}c^{{\rm QCDF}, c}_{2}.
\end{align}
We can realize how the $\Lambda_u$~--~$\Lambda_t$ parameterization in Eq.~(\ref{eq:TopologicalQCDF}) has advantages with respect to the $\Lambda_u$~--~$\Lambda_c$ one in Eqs.~(\ref{eq:QCDF1}) and~(\ref{eq:QCDF2}) when it comes to disentangle the QCDF amplitudes $\{\alpha_i,\alpha^{p}_{i}, \beta_{i},b_i\}$. For instance, once we account for the expressions in Eq.~(\ref{eq:QCDFconditions}), there are multiple cancellations between the up and charm versions of a given little subamplitude, leading to simple relationships between the elements of the topological basis and the QCDF one as shown in Eq.~(\ref{eq:TopologicalQCDF}). On the other hand, in the $\Lambda_u$~--~$\Lambda_c$ decomposition in Eqs.~(\ref{eq:QCDF1}) and~(\ref{eq:QCDF2}) these cancellations do not arise, leaving the coefficients $c^{{\rm QCDF}, p}_{i}$ in terms of linear combinations of multiple $\{\alpha_i,\alpha^{p}_{i}, \beta_{i},b_i\}$ quantities.
\section{Setting the stage for the $\chi^2$-fit}
\label{sec:prepfit}
\subsection{Relevant physical quantities}
We extract the values for the $SU(3)$-irreducible parameters using branching fractions and CP-asymmetries for the different $B\rightarrow P P$ transitions, where $P$ is a pseudoscalar meson.
The basic ingredient behind our phenomenological determinations is the decay rate
\begin{eqnarray}
\Gamma(\bar{B}\rightarrow M_1 M_2)&=&\frac{S}{16\pi M_B}|\mathcal{A}_{B\rightarrow M_1 M_2}|^2,
\end{eqnarray}
where $S$ is a symmetry factor with values $S=1$ if $M_1$ and $M_2$ are different and $S=1/2$ if $M_1$ and $M_2$ are identical. Then, we determine the CP-averaged branching fraction of the process $\bar{B}\rightarrow \bar{f}$ via the formula
\begin{eqnarray}
\mathcal{B}(\bar{B}\rightarrow \bar{f})&=&\frac{1}{2}\tau_{B} \Bigl[\Gamma(\bar{B}\rightarrow \bar{f}) + \Gamma(B\rightarrow f)\Bigl].
\end{eqnarray}
In the case of the CP asymmetries, our theoretical evaluations are performed using
\begin{eqnarray}
\mathcal{A}_{\rm CP}(\bar{B}\rightarrow \bar{f})&=&\frac{\Gamma(\bar{B}
\rightarrow \bar{f})-\Gamma(B\rightarrow f)}{\Gamma(\bar{B}\rightarrow \bar{f})+\Gamma(B\rightarrow f)},
\end{eqnarray}
where we have followed the conventions introduced in \cite{Beneke:2003zv}.
The list of processes considered, together with the corresponding state of the art experimental status are shown in Tables~\ref{tab:tableBr_combined} and~\ref{tab:tableAs_combined}, whose presentation we relegate to section~\ref{sec:results} in order to have the experimental input and the $\chi^2$-fit result at a glance.
\subsection{The $\eta$-$\eta'$ system}
\label{Sec:etamixing}
During our analysis, we account for $\eta$-$\eta'$ mixing. In particular, we adopt the Feldmann–Kroll–Stech (FKS) scheme~\cite{Feldmann:1998vh}, which is based on one single mixing angle $\theta_{\rm FKS}$. In the FKS-scheme, the flavor states are related to the physical states $\eta$, $\eta'$ according to \cite{Beneke:2002jn}
\begin{equation}
\left(
\begin{array}{cc}
\ket{\eta} \\
\ket{\eta'} \\
\end{array}
\right)
=
\left(
\begin{array}{cc}
\cos\theta_{\rm FKS} & -\sin\theta_{\rm FKS} \\
\sin\theta_{\rm FKS} & \cos\theta_{\rm FKS} \\
\end{array}
\right)
\left(
\begin{array}{cc}
\ket{\eta_{q}} \\
\ket{\eta_{s}} \\
\end{array}
\right),
\end{equation}
which leads to the following equations for the decay constants
\begin{align}
f^q_{\eta}&=f_q\cos\theta_{\rm FKS},& f^q_{\eta'}&=f_q \sin\theta_{\rm FKS},\nonumber\\
f^s_{\eta}&=-f_s\sin\theta_{\rm FKS},& f^s_{\eta'}&=f_s \cos\theta_{\rm FKS}.
\end{align}
Therefore, for the amplitudes of the processes $B\rightarrow M \eta$, $B\rightarrow M \eta'$ with $M\neq \eta, \eta'$ we obtain
\begin{eqnarray}
A(B\rightarrow M \eta)&=&
\tilde{A}(B\rightarrow M \eta_q)\cos(\theta_{\rm FKS})
-
\tilde{A}(B\rightarrow M \eta_s)\sin(\theta_{\rm FKS}),\nonumber\\
A(B\rightarrow M \eta')&=&
\tilde{A}(B\rightarrow M \eta_q)\sin(\theta_{\rm FKS})
+
\tilde{A}(B\rightarrow M \eta_s)\cos(\theta_{\rm FKS}),
\end{eqnarray}
where $\tilde{A}(B\rightarrow M \eta_{q})\propto F^{B\rightarrow M}_0(0) f_q$ and $\tilde{A}(B\rightarrow M \eta_{s})\propto F^{B\rightarrow M}_0(0) f_s$.\\
Since the form factors for the transitions $B\rightarrow \eta$ and $B\rightarrow \eta'$ can be parameterized as \cite{Beneke:2002jn}
\begin{eqnarray}
F^{B\rightarrow \eta^{(\prime)}}_{0}&=&
F_1\frac{f^q_{\eta^{(\prime)}}}{f_{\pi}}
+F_{2}\frac{\sqrt{2}f^{q}_{\eta^{(\prime)}} +f^{s}_{\eta^{(\prime)}}}{\sqrt{3}f_{\pi}},
\label{eq:Feta}
\end{eqnarray}
and considering $F_1=F_0^{B\rightarrow \pi}(0)$ and $F_2=0$, just as in \cite{Beneke:2002jn}, we write for the amplitudes of $B$-meson decays with final states consisting of pairs $\eta$, $\eta'$
\begin{eqnarray}
A(B\rightarrow \eta \eta)&=&
\tilde{A}(B\rightarrow \eta_q\eta_q)\cos^2\theta_{\rm FKS} +\tilde{A}(B\rightarrow \eta_s\eta_s)\sin^2\theta_{\rm FKS}\nonumber\\
&&- 2\tilde{A}(B\rightarrow \eta_q\eta_s)\cos\theta_{\rm FKS}\sin\theta_{\rm FKS},\nonumber\\
A(B\rightarrow \eta' \eta')&=&
\tilde{A}(B\rightarrow \eta'_q\eta'_q)\cos^2\theta_{\rm FKS} +\tilde{A}(B\rightarrow \eta'_s\eta'_s)\sin^2\theta_{\rm FKS}\nonumber\\
&&+ 2\tilde{A}(B\rightarrow \eta'_q\eta'_s)\cos\theta_{\rm FKS}\sin\theta_{\rm FKS},\nonumber\\
A(B\rightarrow \eta \eta')&=&\tilde{A}(B\rightarrow \eta_q \eta'_q)\cos\theta_{\rm FKS}\sin\theta_{\rm FKS}+\tilde{A}(B\rightarrow \eta_q \eta'_s)\cos^2\theta_{\rm FKS}\nonumber\\
&&-\tilde{A}(B\rightarrow \eta_s \eta'_q) \sin^2\theta_{\rm FKS} -\tilde{A}(B\rightarrow \eta_s\eta'_s)\sin\theta_{\rm FKS}\cos\theta_{\rm FKS}.
\end{eqnarray}
To perform our numerical estimates we need the value of the mixing angle $\theta_{\rm FKS}$, together with the values for the decay constants $f_q$ and $f_s$. They have been obtained from fits to data involving the processes $J/\psi\rightarrow \eta (\eta') \rho$, $\eta'\rightarrow \rho \gamma$, $\rho \rightarrow \eta \gamma$, $p\bar{p}\rightarrow \eta(\eta')M$, $J/\psi\rightarrow \eta\gamma$, $\eta(\eta')\rightarrow \gamma \gamma$ and $T\rightarrow P_1 P_2$, where $T$ is a $2^{++}$ tensor meson and $P_i$ is a pseudoscalar meson \cite{Feldmann:1998vh},
\begin{eqnarray}
\theta_{\rm FKS}=39.3^{\circ}\pm 1.0^{\circ},
\quad
f_{q}=(1.07\pm 0.02) f_{\pi},
\quad
f_{s}=(1.34\pm 0.06) f_{\pi}.
\label{eq:decay_constants}
\end{eqnarray}
For the purpose of our numerical determinations we calculate the coefficient $A_{M_1 M_2}$ in Eq.~(\ref{eq:AM1M2}) considering the form factors and decay constants included in Table~\ref{tab:FormFactors}. Moreover, the quantities relevant to the $\eta$ and $\eta'$ contributions are evaluated using Eq.~(\ref{eq:Feta}) and the inputs given in Eq.~(\ref{eq:decay_constants}). Finally, we find
\begin{eqnarray}
A_{M_1 M_2}=(1.25\pm 0.02)~{\rm GeV}^3.
\label{eq:AM1M@Average}
\end{eqnarray}
\begin{table}[htp]
\begin{center}
\begin{tabular}{ |c|c|c||c|c|c| }
\hline
&&&&&\\
Parameter& Value & Reference & Parameter & Value & Reference\\
&&&&&\\
\hline
& $0.258\pm 0.031$ & \cite{Ball:2004ye} &&&\\
& $0.25\pm 0.05$ & \cite{Khodjamirian:2006st} &&$0.331\pm 0.041$& \cite{Ball:2004ye} \\
& $0.28\pm 0.05$ & \cite{Khodjamirian:2011ub} &&$0.31\pm 0.04$& \cite{Khodjamirian:2006st}\\
$F^{B\rightarrow \pi}_{0}(0)$ & $0.21\pm 0.07$ & \cite{Gubernari:2018wyi}& $F^{B\rightarrow K}_{0}(0)$& $0.27\pm 0.08$&\cite{Gubernari:2018wyi}\\
& $0.31\pm 0.02$ & \cite{SentitemsuImsong:2014plu}&& $0.395\pm 0.033$&\cite{Khodjamirian:2017fxg}\\
& $0.281\pm 0.038$ & \cite{Wang:2015vgv}&&$0.364\pm 0.05$& \cite{Lu:2018cfc}\\
& $0.301\pm 0.023$& \cite{Khodjamirian:2017fxg}&&& \\
\hline
&&&&&\\
$f_{\pi}~[\rm{GeV}]$
&
$0.1302\pm 0.0008$
&
\cite{FlavourLatticeAveragingGroup:2019iem}
&
$f_{K}~[\rm{GeV}]$&$0.1557\pm 0.0007$& \cite{FlavourLatticeAveragingGroup:2019iem}\\
&&&&&\\
\hline
\end{tabular}
\end{center}
\caption{Form factors for the transitions $B\rightarrow \pi$ and $B\rightarrow K$ and decay constants $f_{\pi}$ and $f_{K}$.
\label{tab:FormFactors}}
\end{table}
\section{Details on the $\chi^2$-fit}
\label{sec:fit}
To assess the set of values for the $SU(3)$-invariant amplitudes compatible with experimental data we perform a $\chi^2$-fit.
Thus, we establish the compatibility between the experimental data and our theoretical determinations by evaluating the function
\begin{eqnarray}
\chi^2=\sum_{i} \Bigl(\frac{\mathcal{O}_i^{\rm Theo}-\mathcal{O}_i^{\rm Exp}}{\sigma_i^{\rm Exp}}\Bigl)^2,
\label{eq:chi2}
\end{eqnarray}
where $\mathcal{O}$ is either a branching fraction or a CP-asymmetry and the labels ``Theo'' and ``Exp'' make reference to the corresponding theoretical or experimental determinations. Moreover the sum runs over all the physical observables which have been measured. Our list of experimental inputs are presented in Tables \ref{tab:tableBr_combined} and~\ref{tab:tableAs_combined}. We exclude from the $\chi^2$-function quantities which either are not experimentally available, or where only an upper bound exists. Actually, the upper bounds will be considered as constraints during our fitting procedure, more details on this below.
We start by determining our relevant degrees of freedom. As discussed before there are in principle $10$ complex tree-like amplitudes and $10$ complex penguin-like amplitudes leading to a total of 40 real quantities, however $A^T_6$ can be absorbed into $B^T_6$ and $C^T_6$, correspondingly $A^P_6$ can be absorbed into $B^P_6$ and $C^P_6$. Moreover we can get rid of a global phase by taking $C^P_3$ as a real parameter. Finally, the study of $\eta-\eta'$ mixing requires an additional real quantity, the mixing angle $\theta_{FKS}$. Consequently, our parameter space consist of 36-dimensional vectors.
There are two stages when it comes to determine the values of the $SU(3)$-invariant quantities from experimental data. First, we extract our best fit-point and second we will calculate the allowed regions about this particular point at a given confidence level.
During the first stage we follow the standard approach and parameterize the tree and penguin amplitudes in terms of their modulus $|A_3|$, $|C_3|$, $|C_6|$, $|A_{15}|$, $|C_{15}|$, $|B_3|$, $|B_6|$, $|B_{15}|$, $|D_3|$ and their corresponding phases $\delta_{A_3}$, $\delta_{C_3}$, $\delta_{C_6}$, $\delta_{A_{15}}$, $\delta_{C_{15}}$, $\delta_{B_3}$, $\delta_{B_6}$, $\delta_{B_{15}}$, $\delta_{D_3}$ plus the $\eta-\eta'$ mixing angle $\theta_{FKS}$. To determine the best fit point we use random sampling, i.e.\ we generate randomly $10^{9}$ points in our 36-dimensional space obeying a flat probability distribution. We then calculate the $\chi^2$ function for each one of these points. We divide our $10^{9}$ evaluations into $1000$ individual runs of $10^{6}$-points each. For each individual slot of $10^{6}$ points we select the top 5 points with the lowest $\chi^2$ determined previously. Then for each one of our $1000$ slots, we apply the Sequential Least Square Programming algorithm, SLSQP, available in python, taking as starting points those with the corresponding 5 minimal $\chi^2$. Finally, the overall minimum of all the SLSQP-minimizations is taken as our best-fit point.
Based on different numerical tests we find that the random sampling procedure converges quickly to the region of minimum values if we generate points for $\{|C^{T}_3|,$
$|C^{T}_6|, |C^{T}_{15}|,$ $|B^{T}_{6}|, |B^{T}_{15}|, |D^{T}_{3}|, |C^{P}_{3}|, |C^{P}_{6}|, |C^{P}_{15}|, |B^{P}_6|, |B^{P}_{15}|, |D^{P}_3|\}$
inside the interval $[0, 0.3]$ and for $\{|A^{T}_{3}|, |A^{T}_{15}|,|B^{T}_{3}|, |A^{P}_{3}|, |A^{P}_{15}|, |B^{P}_{3}|\}$ inside the region $[0, 0.1]$. Finally, for the phases we sample random values within $[-\pi, \pi]$.
Performing the $\chi^2$-fit based only on the experimental data
leads to a rather unconstrained scenario with a large number of fit parameters. To improve the situation we opt for introducing a minimal set of extra conditions arising within the context of QCDF, while at the same time keeping the theoretical bias from that approach as marginal as possible. Currently, the color-allowed tree amplitude $\alpha_1$ has been calculated including corrections up to NNLO leading for the $B\rightarrow \pi \pi$ process to \cite{Beneke:2009ek}
\begin{eqnarray}
\alpha_{1}(\pi\pi)=1.000^{+0.029}_{-0.069}+(0.011^{+0.023}_{-0.050})i.
\label{eq:alpha1}
\end{eqnarray}
It can be seen from Eq.~(\ref{eq:alpha1}) that the real part is known with a precision better than $7\%$, and the study in~\cite{Beneke:2009ek} further reveals that it is robust against large higher-order perturbative corrections. Hence, during our statistical analysis we impose that our determination for the real part of $\alpha_{1}(\pi\pi)$ from the $SU(3)$ parameters is constrained to lie within the region
\begin{eqnarray}
\Re(\alpha_{1})=1.000^{+0.138}_{-0.138},
\label{eq:alpharegions}
\end{eqnarray}
where we have adopted a conservative approach by considering a symmetric uncertainty which amounts to twice the theoretical value presented in Eq.~(\ref{eq:alpha1}).
According to Eq.~(\ref{eq:SU£toTop}), to map out the QCDF amplitude $\alpha_{1}$ into the $SU(3)$ evaluations we have to take into account the extra factor $A_{M_1 M_2}$ whose numerical value is presented in Eq.~(\ref{eq:AM1M@Average}). Thus, our full constraint reads
\begin{eqnarray}
\Re(A_{M_1 M_2}\alpha_1)=(1.245\pm 0.173)~{\rm GeV}^3.
\label{eq:consalpha1}
\end{eqnarray}
Additionally, based on state of the art experimental determinations, we require that the following bounds are obeyed during the minimization procedure
\begin{eqnarray}
Br(B_s\rightarrow \pi^0 \pi^0)<2.10\times 10^{-4},&~~~&
Br(B_s\rightarrow \eta \pi^0)< 10^{-3},\nonumber\\
Br(B^0\rightarrow \eta\eta)<10^{-6},&~~~&
Br(B^0\rightarrow \eta'\eta')<1.7\times 10^{-6},\nonumber\\
Br(B^0\rightarrow \eta' \eta)<1.2\times 10^{-6},&~~~&A_{CP}(B_s\rightarrow \eta K^0)<10^{-3}.
\label{eq:inequalities}
\end{eqnarray}
Our main goal is to evaluate the size of the QCDF amplitudes $\beta_{i}$ as allowed by data, without introducing too much a priori information on the size of these quantities. We then perform a study including as constraints Eqs.~(\ref{eq:alpharegions})~--~(\ref{eq:inequalities}). In addition, based on the results in Eq.~(\ref{eq:TopologicalQCDF}) we impose the following conditions (dimensionful quantities are given in units of GeV${}^3$)
\begin{eqnarray}
T_{PA}=T_{SS}=T_{S}=0,&&|T_{P}|<0.1.
\label{eq:constraints}
\end{eqnarray}
With the help of Eqs.~(\ref{eq:ToptoSU3}) we can easily translate Eqs.~(\ref{eq:constraints}) into constraints for the $SU(3)$-invariant parameters which are the quantities that we are actually fitting to experimental data. We would like to highlight the fact that we are being conservative with respect to the bound imposed over $T_{P}$ by enlarging the upper limit established by the magnitudes of the differences $\alpha^c_4-\alpha^u_4$ and $\alpha^c_{4, EW}-\alpha^u_{4, EW}$ by a factor of five.
For the determination of the confidence intervals of our $SU(3)$ parameters, we perform a likelihood ratio test and determine the $p$ value applying Wilk's theorem. In principle we can establish the confidence intervals for the modulus and phases of the relevant $SU(3)$-invariant quantities. However, we find that it is more informative to study the allowed regions in the space expanded by the real and the imaginary components of each one of the $SU(3)$-invariant amplitudes. Hence, the confidence level is estimated considering
\begin{eqnarray}
1-\frac{\Gamma(\Delta\chi^2/2, \nu/2 )}{\Gamma(\nu/2)} = 1-p,
\end{eqnarray}
where $\Gamma$ is the normalised upper incomplete Gamma function, $\nu$ represents the number of degrees of freedom and
\begin{eqnarray}
\Delta \chi^2=\chi^2_{0}-\chi_{min}^2.
\label{eq:Deltachi}
\end{eqnarray}
In Eq.~(\ref{eq:Deltachi}) $\chi^2_{min}$ is the global minimum for the $\chi^2$ function and $\chi^2_{0}$ is the minimum subjected to the null-hypothesis over the value of one of the $SU(3)$-invariant parameters.
Since we are fitting for the real and imaginary parts of a complex quantity simultaneously $\nu=2$. Moreover, we take $1-p=0.68$, implying that for our confidence regions
\begin{eqnarray}
\Delta \chi^2 \leq 2.3.
\end{eqnarray}
\section{Results}
\label{sec:results}
\subsection{$SU(3)$-irreducible amplitudes}
Let us now give the results of our $\chi^2$-fit. We start with the best-fit point of the $SU(3)$-irreducible amplitudes in polar coordinates (the modulus are given in units of GeV${}^3$),
\begin{align}
|A^T_{3}|&=0.029, \qquad &
\delta_{A^T_3}&=-3.083, \qquad\qquad&
|C^T_{3}|&=0.258, \qquad &
\delta_{C^T_{3}}&=-0.105,\nonumber\\
|C^T_{6}|&=0.235, &
\delta_{C^T_{6}}&=-0.079, &
|A^T_{15}|&=0.029, &
\delta_{A^T_{15}}&=-3.083,\nonumber\\
|C^T_{15}|&=0.151, &
\delta_{C^T_{15}}&=0.061, &
|B^T_3|&=0.034,&
\delta_{B^T_3}&=3.087,\nonumber\\
|B^T_6|&=0.033, &
\delta_{B^T_6}&=-0.286, &
|B^T_{15}|&=0.008, &
\delta_{B^T_{15}}&=-1.892,\nonumber\\
|D^T_3|&=0.055, &
\delta_{D^T_3}&=2.942, & &&&
\end{align}
\begin{align}
|A^P_{3}|&=0.014,&\qquad
\delta_{A^P_3}&=-1.328, \qquad\qquad &
|C^P_{6}|&=0.145, \qquad&
\delta_{C^P_{6}}&=-2.881,\nonumber\\
|A^P_{15}|&=0.003, &
\delta_{A^P_{15}}&=2.234, &
|C^P_{15}|&=0.003, &
\delta_{C^P_{15}}&=-0.608,\nonumber\\
|B^P_3|&=0.043,&
\delta_{B^P_3}&=2.367,&
|B^P_6|&=0.099,&
\delta_{B^P_6}&=0.353,\nonumber\\
|B^P_{15}|&=0.031,&
\delta_{B^P_{15}}&=-0.690,&
|D^P_{3}|&=0.030,&
\delta_{D^P_{3}}&=0.477,\nonumber\\
|C^P_3|&=0.008,&
\theta_{FKS}&=0.628. & &&&
\end{align}
\begin{figure}[htp]
\centering
\includegraphics[width=7cm]{SU3_0_0.pdf}
\includegraphics[width=7cm]{SU3_1_0.pdf}
\includegraphics[width=7cm]{SU3_3_0.pdf}
\includegraphics[width=7cm]{SU3_4_0.pdf}
\includegraphics[width=7cm]{SU3_5_0.pdf}
\includegraphics[width=7cm]{SU3_6_0.pdf}
\includegraphics[width=7cm]{SU3_7_0.pdf}
\includegraphics[width=7cm]{SU3_8_0.pdf}
\caption{$SU(3)$ fit, part1. Amplitudes carry units of GeV${}^3$.\label{fig:SU3_1_1}}
\end{figure}
\begin{figure}[htp]
\centering
\includegraphics[width=7cm]{SU3_9_0.pdf}
\includegraphics[width=7cm]{SU3_10_0.pdf}
\includegraphics[width=7cm]{SU3_11_0.pdf}
\includegraphics[width=7cm]{SU3_13_0.pdf}
\includegraphics[width=7cm]{SU3_14_0.pdf}
\includegraphics[width=7cm]{SU3_15_0.pdf}
\includegraphics[width=7cm]{SU3_16_0.pdf}
\includegraphics[width=7cm]{SU3_17_0.pdf}
\caption{$SU(3)$ fit, part2. Amplitudes carry units of GeV${}^3$.\label{fig:SU3_2_1}}
\end{figure}
\begin{figure}[htp]
\centering
\includegraphics[width=7cm]{SU3_18_0.pdf}
\includegraphics[width=7cm]{SU3_19_0.pdf}
\caption{$SU(3)$ fit, part3. Amplitudes carry units of GeV${}^3$.\label{fig:SU3_3_1}}
\end{figure}
Additionally in Figs.~\ref{fig:SU3_1_1}, \ref{fig:SU3_2_1} and~\ref{fig:SU3_3_1} we present the two-dimensional confidence regions for the real and imaginary parts of the complex $SU(3)$-invariant quantities. Note that whereas our best-fit point is determined in polar coordinates, the two-dimensional confidence regions for each complex amplitude are obtained in the Cartesian basis.
In consistency with previous studies \cite{Hsiao:2015iiu,Fu:2003fy}, we see that the absolute value of the real and imaginary components of the annihilation contributions $A_i$ are always below $0.05$. On the other hand, the absolute values of the complex components for the rest of the amplitudes $C_i$, $B_i$ and $D_i$ are at most $0.32$. Notice however, that our estimation of the confidence regions are done simultaneously for the real and imaginary components of the individual amplitudes, i.e.\ they are performed for two degrees of freedom. This is a central difference with respect to the analyses in~\cite{Hsiao:2015iiu, Fu:2003fy} where to the best of our understanding the confidence regions are delivered taking into account uni-dimensional intervals for the magnitudes and phases of the different parameters involved. To assess the goodness of our fit we compute the $\chi^2$ per degree of freedom obtaining
\begin{eqnarray}
\chi^2/d.o.f.&=&0.851.
\end{eqnarray}
\begin{table}[htp]
\begin{center}
\scalebox{0.85}{\begin{tabular}{|c|c|ccc||c|c|ccc|}
\hline
\multirow{4}{*}{Channel} & \multicolumn{4}{|c||}{\textbf{Branching ratio}}&
\multirow{4}{*}{Channel} & \multicolumn{4}{|c|}{\textbf{Branching ratio}}\\
&\multicolumn{4}{|c||}{\textbf{in units of $10^{-6}$}} &&
\multicolumn{4}{|c|}{\textbf{in units of $10^{-6}$}}\\
\cline{2-5}\cline{7-10}
&Experimental& \multicolumn{3}{|c||}{Theoretical} &
&Experimental& \multicolumn{3}{|c|}{Theoretical}\\
\hline
\rule{0pt}{14pt}$B^-\rightarrow \pi^0\pi^-$ & $5.5\pm 0.4$ & &$6.04^{+2.42}_{-2.51}$ &&
$B^-\rightarrow \eta \pi^-$& $4.02\pm 0.27$&&$3.80^{+1.25}_{-1.55}$ &\\[0.4em]
$B^-\rightarrow K^0 K^-$ & $1.31\pm 0.17$ &&$1.36^{+0.17}_{-0.16}$&&
$B^-\rightarrow \eta' \pi^-$&$2.7 \pm 0.9$&&$3.55^{+4.49}_{-1.67}$&\\[0.4em]
$\bar{B}^0\rightarrow \pi^+ \pi^-$& $5.12\pm 0.19$&&$6.31^{+0.61}_{-0.50}$&&
$\bar{B}^0\rightarrow \eta \pi^0$& $0.41\pm 0.17$&&$0.41^{+8.90}_{-4.08}$&\\[0.4em]
$\bar{B}^0\rightarrow \pi^0 \pi^0$& $1.59\pm 0.26$&&$1.01^{+1.30}_{-0.51}$&&
$\bar{B}^0\rightarrow \eta' \pi^0$& $1.2\pm 0.6$&&$1.20^{+3.62}_{-1.19}$&\\[0.4em]
$\bar{B}^0\rightarrow K^+ K^-$& $0.078\pm 0.015$&&$0.13^{+0.08}_{-0.07}$&&
$\bar{B}_s\rightarrow \eta K^0$& Not available&&$0.13^{+0.11}_{-0.08}$&\\[0.4em]
$\bar{B}^0\rightarrow K^0 \bar{K}^0$& $1.21 \pm 0.16$&&$1.13^{+0.83}_{-0.91}$&&
$\bar{B}_s\rightarrow \eta' K^0$& Not available &&$6.65^{+1.48}_{-1.65}$&\\[0.4em]
$\bar{B}_s\rightarrow \pi^- K^+$& $5.8\pm 0.7$&&$7.75^{+0.63}_{-0.09}$&&
$B^-\rightarrow \eta K^-$&$2.4\pm 0.4$&&$2.34^{+1.39}_{-1.67}$&\\[0.4em]
$B^-\rightarrow \pi^0 K^-$& $12.9\pm 0.5$&&$12.78^{+1.75}_{-1.94}$&&
$B^-\rightarrow \eta' K^-$&$70.4\pm 2.5$&&$70.82^{+11.16}_{-11.53}$&\\[0.4em]
$B^-\rightarrow \pi^- \bar{K}^0$& $23.7\pm 0.8$&&$23.85^{+2.23}_{-2.31}$&&
$\bar{B}^0\rightarrow \eta K^0$&$1.23\pm 0.27$&&$1.38^{+1.15}_{-0.36}$&\\[0.4em]
$\bar{B}^0\rightarrow \pi^+ K^-$& $19.6 \pm 0.5$&&$19.47^{+1.72}_{-2.24}$&&
$\bar{B}^0\rightarrow \eta' K^0$& $6.6\pm 0.4$&&$6.65^{+1.48}_{-1.65}$&\\[0.4em]
$\bar{B}^0\rightarrow \pi^0 \bar{K}^0$& $9.9 \pm 0.5$&&$10.17^{+2.00}_{-2.30}$&&
$\bar{B}_s\rightarrow \eta \pi^0$& $<10^{3}$&&$31.15^{+39.05}_{-31.14}$&\\[0.4em]
$\bar{B}_s\rightarrow \pi^+ \pi^-$& $0.7 \pm 0.1$&&$0.57^{+0.40}_{-0.42}$&&
$\bar{B}_s\rightarrow \eta' \pi^0$& Not available&&$11.13^{+74.75}_{-11.12}$&\\[0.4em]
$\bar{B}_s\rightarrow \pi^0\pi^0$& $< 210$&&$0.28^{+0.20}_{-0.21}$&&
$\bar{B}^0\rightarrow \eta \eta$&$<1$&&$0.30^{+0.70}_{-0.30}$&\\[0.2em]
$\bar{B}_s\rightarrow K^+ K^-$& $26.6\pm 2.2$&&$20.63^{+6.80}_{-8.09}$&&
$\bar{B}_s\rightarrow \eta \eta$& $<1.5\times 10^{3}$&&$2.58^{+36.53}_{-2.57}$&\\[0.4em]
$\bar{B}_s\rightarrow K^0 \bar{K}^0$& $20\pm 6$&&$24.64^{+18.84}_{-21.14}$&&
$\bar{B}^0\rightarrow \eta'\eta'$&$<1.7$&&$1.14^{+0.57}_{-1.07}$&\\[0.4em]
$\bar{B}_s\rightarrow \pi^0 K^0$& Not available&&$0.71^{+1.47}_{-0.27}$&&
$\bar{B}_s\rightarrow \eta'\eta'$& $33\pm 7$&&$33.00^{+24.52}_{-31.74}$&\\[0.4em]
& & & &&
$\bar{B}^0\rightarrow \eta'\eta$& $<1.2$&&$0.61^{+0.59}_{-0.60}$&\\[0.4em]
& & & &&
$\bar{B}_s\rightarrow \eta'\eta$& Not available&&$0.61^{+0.59}_{-0.60}$&\\[0.4em]
\hline
\end{tabular}}
\caption{Experimental input and fit results for CP-averaged $B\to PP$ branching fractions.\label{tab:tableBr_combined}}
\end{center}
\end{table}
\begin{table}[htp]
\begin{center}
\scalebox{0.85}{\begin{tabular}{|c|c|ccc||c|c|ccc|}
\hline
\multirow{4}{*}{Channel} & \multicolumn{4}{|c||}{\textbf{CP asymmetries}}&
\multirow{4}{*}{Channel} & \multicolumn{4}{|c|}{\textbf{CP asymmetries}} \\
&\multicolumn{4}{|c||}{\textbf{in percent}}&&
\multicolumn{4}{|c|}{\textbf{in percent}}\\
\cline{2-5}\cline{7-10}
&Experimental& \multicolumn{3}{|c||}{Theoretical}&
&Experimental& \multicolumn{3}{|c|}{Theoretical}\\
\hline
\rule{0pt}{14pt}$B^-\rightarrow \pi^0\pi^-$ & $3\pm 4$ &&$5.45^{+22.02}_{-20.60}$&&
$B^-\rightarrow \eta \pi^-$& $-14\pm 7$&&$-11.37^{+14.49}_{-26.90}$&\\[0.4em]
$B^-\rightarrow K^0 K^-$ & $4\pm 14$&&$18.82^{+36.93}_{-30.83}$&&
$B^-\rightarrow \eta' \pi^-$&$6\pm 16$&&$4.71^{+59.79}_{-57.97}$&\\[0.4em]
$\bar{B}^0\rightarrow \pi^+ \pi^-$& $32\pm 4$&&$35.01^{+3.19}_{-22.29}$&&
$\bar{B}_s\rightarrow \eta K^0$&$<0.1$&&$0.10^{+0.00}_{-100.07}$&\\[0.4em]
$\bar{B}^0\rightarrow \pi^0 \pi^0$& $33\pm 22$&&$-10.58^{+40.69}_{-89.40}$&&
$\bar{B}_s\rightarrow \eta' K^0$&Not available&&$-0.58^{+100.57}_{-79.58}$&\\[0.4em]
$\bar{B}^0\rightarrow K^0 \bar{K}^0$&$-60\pm 70$&&$-6.88^{+85.39}_{-81.37}$&&
$B^-\rightarrow \eta K^-$&$-37\pm 8$&&$-42.23^{+42.23}_{-16.00}$&\\[0.4em]
$\bar{B}_s\rightarrow \pi^- K^+$& $22.1\pm 1.5 $&&$20.84^{+2.39}_{-2.57}$&&
$B^-\rightarrow \eta' K^-$&$0.4\pm 1.1$&&$0.63^{+3.98}_{-4.30}$&\\[0.4em]
$B^-\rightarrow \pi^0 K^-$& $3.7\pm 2.1$&&$3.72^{+7.19}_{-4.35}$&&
$\bar{B}^0\rightarrow \eta K^0$&Not available&&$-0.01^{+40.07}_{-0.02}$&\\[0.4em]
$B^-\rightarrow \pi^- K^0$&$-1.7\pm 1.6$&&$-1.08^{+1.76}_{-2.32}$&&
$\bar{B}^0\rightarrow \eta' K^0$&$-6\pm 4$ &&$0.03^{+4.82}_{-11.69}$&\\[0.4em]
$\bar{B}^0\rightarrow \pi^+ K^-$& $-8.3\pm 0.4$&&$-8.38^{+8.38}_{-1.01}$&&
$\bar{B}^0\rightarrow \eta \pi^0$&Not available&&$-27.39^{+127.11}_{-72.58}$&\\[0.4em]
$\bar{B}^0\rightarrow \pi^0 \bar{K}^0$& $0\pm 13$&&$-0.97^{+19.35}_{-3.20}$&&
$\bar{B}^0\rightarrow \eta' \pi^0$&Not available&&$-43.67^{+143.63}_{-56.33}$&\\[0.4em]
$\bar{B}_s\rightarrow K^+ K^-$& $-14\pm 11$&&$-10.58^{+10.58}_{-3.60}$&&
$\bar{B}_s\rightarrow \eta \pi^0$& Not available&&$0.88^{+94.98}_{-98.70}$&\\[0.4em]
$\bar{B}_s\rightarrow \pi^+ \pi^-$& Not available&&$17.56^{+11.84}_{-38.25}$&&
$\bar{B}_s\rightarrow \eta' \pi^0$& Not available&&$1.57^{+77.56}_{-95.66}$&\\[0.4em]
$\bar{B}_s\rightarrow \pi^0\pi^0$& Not available&&$17.56^{+11.84}_{-38.25}$&&
$\bar{B}^0\rightarrow \eta \eta$&Not available&&$3.46^{+96.50}_{-103.45}$&\\[0.4em]
$\bar{B}_s\rightarrow K^0 \bar{K}^0$& Not available&&$0.31^{+5.07}_{-4.59}$&&
$\bar{B}_s\rightarrow \eta \eta$& Not available&&$14.29^{+76.81}_{-113.09}$&\\[0.4em]
$\bar{B}^0\rightarrow K^+ K^-$& Not available&&$-78.45^{+161.99}_{-20.78}$&&
$\bar{B}^0\rightarrow \eta'\eta'$&Not available&&$42.41^{+57.55}_{-142.41}$&\\[0.4em]
$\bar{B}_s\rightarrow \pi^0 K^0$& Not available&&$13.74^{+29.49}_{-113.73}$&&
$\bar{B}_s\rightarrow \eta'\eta'$&Not available&&$-2.05^{+15.29}_{-13.44}$&\\[0.4em]
& & & &&
$\bar{B}^0\rightarrow \eta'\eta$&Not available&&$-12.32^{+112.32}_{-87.67}$&\\[0.4em]
& & & &&
$\bar{B}_s\rightarrow \eta'\eta$&Not available&&$3.43^{+96.36}_{-103.22}$&\\[0.4em]
\hline
\end{tabular}}
\caption{Experimental input and fit results for $B\to PP$ CP asymmetries.
\label{tab:tableAs_combined}}
\end{center}
\end{table}
\begin{figure}[htp]
\centering
\includegraphics[width=7cm]{QCDF_0_0.png}
\includegraphics[width=7cm]{QCDF_1_0.png}
\includegraphics[width=7cm]{QCDF_2_0.png}
\includegraphics[width=7cm]{QCDF_4_0.png}
\includegraphics[width=7cm]{QCDF_5_0.png}
\includegraphics[width=7cm]{QCDF_6_0.png}
\includegraphics[width=7cm]{QCDF_7_0.png}
\includegraphics[width=7cm]{QCDF_8_0.png}
\caption{QCDF fit, part1. Amplitudes carry units of GeV${}^3$.\label{fig:QCDF_1_1}}
\end{figure}
\begin{figure}[htp]
\centering
\includegraphics[width=7cm]{QCDF_10_0.png}
\includegraphics[width=7cm]{QCDF_11_0.png}
\includegraphics[width=7cm]{QCDF_12_0.png}
\includegraphics[width=7cm]{QCDF_13_0.png}
\includegraphics[width=7cm]{QCDF_14_0.png}
\includegraphics[width=7cm]{QCDF_15_0.png}
\caption{QCDF fit, part2. Amplitudes carry units of GeV${}^3$.\label{fig:QCDF_2_0}}
\end{figure}
One of the consequences of the relatively large confidence intervals are bigger uncertainties for the evaluations of the branching fractions and CP asymmetries, which we show in Tables~\ref{tab:tableBr_combined} and~\ref{tab:tableAs_combined}. We can see that most of our results agree with the experimental determinations within the $1\sigma$ range. There are, however, two exceptions: $\mathcal{B}(B^0\rightarrow \pi^+ \pi^-)$ and $\mathcal{B}(B_s\rightarrow \pi^- K^+)$ where the discrepancies are at the level of $2\sigma$ and $2.5\sigma$ respectively. We can identify the source of these discrepancies in the relatively high value of the pre-factor $A_{M_1 M_2}$ appearing in Eq.~(\ref{eq:AM1M@Average}), which is in turn driven by the $B \to \eta^{(\prime)}$ form factors. Moreover, the value of $A_{M_1 M_2}$ has an additional impact on the fit through the constraint from Eq.~(\ref{eq:consalpha1}). As a matter of fact, data prefers a relatively low value for the product $A_{M_1 M_2}\alpha_1\sim 1.0$~GeV${}^3$. We checked this by performing the $\chi^2$-fit with an increased uncertainty to the number in Eq.~(\ref{eq:consalpha1}), and also by determining $A_{M_1 M_2}$ including only the form factors and decay constants of the transitions $B\to \pi$ and $B\to K$. Both scenarios lead to a reduced value of the $\chi^2$ per degree of freedom. However, we prefer to keep the setup of the analysis in the way described, since we consider that determining $A_{M_1 M_2}$ from all $B \to P$ form factors is the most transparent way of carrying out our analysis which consistently relies on the $SU(3)$ symmetry.
\subsection{QCDF amplitudes}
Translating our best-fit point to the QCDF amplitudes we obtain (all dimensionful numbers are given in GeV${}^3$)
\begin{align}
A_{M_1 M_2}\alpha_1&=1.072+5.596\times 10^{-5} i,
& A_{M_1 M_2}\alpha_2&=0.136+0.073i, \nonumber\\
A_{M_1 M_2} \beta_1&=-0.117-0.007 i, &
A_{M_1 M_2} \beta_2&=A_{M_1 M_2}\beta_1, \nonumber\\
A_{M_1 M_2} \beta_{S1}&=-0.074-0.0112i, &
A_{M_1 M_2} \beta_{S2}&=0.054-0.049 i, \nonumber\\
A_{M_1 M_2} \alpha_{3, EW}&= -0.193-0.045 i, &
A_{M_1 M_2} \alpha^{c}_{4, EW}&=0.181+0.053 i,\nonumber\\
A_{M_1 M_2} \beta_{3, EW}&= 0.005-0.006 i, &
A_{M_1 M_2} b_{4, EW}&=A_{M_1 M_2} \beta_{3, EW},\nonumber\\
A_{M_1 M_2} \beta_{S3, EW}&=-0.188+0.007 i, &
A_{M_1 M_2} b_{S4, EW}&=0.061+0.098 i, \nonumber\\
A_{M_1 M_2}\beta_4&=-0.003+0.013 i, &
A_{M_1 M_2}\beta_{S4}&=0.031-0.030 i, \nonumber\\
A_{M_1 M_2}\Bigl(\alpha_3 +\beta_{S3}\Bigl)&=0.230+0.067 i, &
A_{M_1 M_2}\Bigl(\alpha_4 +\beta_{3}\Bigl)&=-0.242-0.062 i.
\end{align}
Using the relations in~(\ref{eq:TopologicalQCDF}) we can map out the confidence regions of the $SU(3)$-irreducible amplitudes into the QCDF ones presented in Figs.~\ref{fig:QCDF_1_1} and~\ref{fig:QCDF_2_0}. Note that in those figures we plot the product of $A_{M_1M_2}$ times the QCDF amplitudes, whereas in the lines to follow we refer to the QCDF amplitudes themselves. One of our main results is the size of the annihilation amplitudes $\beta_i$ and $b_i$. While some of them such as $\beta_4$ and $\beta_{S4}$ get constrained around or below $0.1$, others can be up to $\sim 0.3$ as in the case of $\beta_{S1}$ and $\beta_{S2}$. Concerning $\alpha_1$, the best fit for the real part is at the lower end of the interval defined by Eq.~(\ref{eq:alpharegions}), and its imaginary component can be up to $0.08$. The imaginary part of $\alpha_2$ can grow to values of up to $-0.5$. Within our analysis we can only evaluate the combinations $\alpha_3 + \beta_{S3}$ and $\alpha_4 + \beta_{3}$, and is not possible to isolate any of the amplitudes independently without introducing extra information. While many of the uncertainty bands are still rather sizeable, our results can act as a valuable guide to get a handle on the size of annihilation contributions in foreseen phenomenological studies on two-body charmless non-leptonic $B_{(s)}$ decays, and be considered as a first step towards improving the treatment of annihilation amplitudes with respect to the early QCDF studies~\cite{Beneke:2001ev,Beneke:2003zv}.
The extraction of the annihilation contributions can in principle also be carried out by fitting to data Eqs.~(\ref{eq:QCDF1}) and (\ref{eq:QCDF2}), however the confidence regions are in general bigger than those obtained in Figs.~\ref{fig:QCDF_1_1} and~\ref{fig:QCDF_2_0} since,
as discussed before, the cancellations between the different up and charm penguin contributions
arising in the $\Lambda_u-\Lambda_t$ parameterization does not occur in this
particular case, leading to the coefficients $c^{{\rm QCDF}, p}_i$ being expressed as linear combinations of the QCDF amplitudes $\{\alpha_i,\alpha_i^{p},\beta_i,b_i\}$, which makes it relatively difficult to disentangle the pure annihilation contribution.
\subsection{Estimating the size of $SU(3)$ breaking}
So far our discussion has been limited to the situation where the flavour $SU(3)$-symmetry holds. A proper treatment of the $SU(3)$ breaking situation requires a full and careful analysis where this effect is incorporated in the theoretical formulas leading to the extraction of information from data. This will correspond to an entire new study on which we comment briefly in the outlook section. We can, however, carry out a preliminary estimation of the implications of giving up the $SU(3)$-flavour symmetry, in particular on the weak annihilation amplitudes. One way of accounting for $SU(3)$-breaking is by estimating the coefficient $A_{M_1 M_2}$ introduced in Eq.~(\ref{eq:AM1M2}) in a channel-dependent fashion. Consider for instance the transitions $B\rightarrow K K$ for which the global factor extracted from $KK$ final states only reads
\begin{eqnarray}
A_{K K}&=&1.54\pm 0.09,
\end{eqnarray}
and consequently, with the $SU(3)$-symmetric $A_{M_1 M_2}$ from~(\ref{eq:AM1M@Average}),
\begin{eqnarray}
A_{K K}/A_{M_1 M_2}&=&1.23 \pm 0.07.
\end{eqnarray}
\begin{figure}[t]
\centering
\includegraphics[width=10cm]{QCDF_Beta_1_B_to_KK}
\caption{Comparison of the $SU(3)$ symmetric (blue, best fit point in red) and $SU(3)$ breaking (green, best fit point in white) cases for the QCDF amplitude $\beta_1$. \label{fig:SU3_breaking_comparison}}
\end{figure}
Therefore, the maximum deviation with respect to the $SU(3)$ symmetric case is $30\%$ for this particular situation, a value that is within the ballpark of numbers quoted in the literature. We can then study the impact of this deviation on the regions for the weak annihilation amplitudes. Consider for example the case of $A_{M_1 M_2}\beta_1$ shown in Fig.~\ref{fig:QCDF_1_1} which now gets modified according to Fig.~\ref{fig:SU3_breaking_comparison}. We can see how the imaginary part grows up to values slightly above $\pm 0.1$ and the real part shifts to values in the range $[-0.21, -0.08]$. As stated before this is a very preliminary study of the $SU(3)$-flavour symmetry breaking effects which will be properly extended in a forthcoming analysis.
\section{Conclusion and Outlook}
\label{sec:conclusion}
There are two well-known model independent descriptions of non-leptonic $B$ decays, the topological and the $SU(3)$-irreducible one. In addition, the QCDF framework offers a systematic way of parameterizing pertubative and non-perturbative contributions. Even though QCDF offers a theoretically solid technique to address the perturbative quantities, the weak annihilation amplitudes, which belong to the non-perturbative category, are challenging to estimate in quantum field theory. In the present work we have determined the possible regions of weak annihilation amplitudes from a data-driven approach.
We first establish the transformation rules between the topological and the QCDF decomposition. We then perform a global fit to the $SU(3)$-irreducible amplitudes with a few additional mild assumptions, and give predictions for CP-averaged branching fractions and direct CP asymmetries. Then, assuming flavour $SU(3)$ to be unbroken, we exploit the connection between the QCDF and the $SU(3)$-irreducible decomposition to translate the fit results into $1\sigma$ confidence regions of the real and imaginary parts of the QCDF amplitudes.
One of our main results is that the size of some of the annihilation amplitudes (such as $\beta_4$ and $\beta_{S4}$) get constrained very well around or below $0.1$, others can take values up to $\sim 0.3$ (such as $\beta_{S1}$ and $\beta_{S2}$). While the confidence regions obtained from the fit are in many cases still sizeable they can be regarded as an upper bound, and provide valuable information on non-perturbative input parameters in future phenomenological studies on two-body charmless non-leptonic $B_{(s)}$ decays. Moreover, with improved experimental data the situation will further improve.
We would like to highlight the fact that we consider an $SU(3)$-invariant approach because this reduces the number of parameters to be fitted to a manageable level. However, in future studies a first step towards a more complete description should include an implementation of $SU(3)$ breaking. This can, for instance, be done by taking the leading-power results from QCDF, where the $SU(3)$ breaking is incorporated naturally in form factors, decay constants and light-cone distribution amplitudes. This will allow for a direct fit to the QCDF weak annihilation amplitudes. We leave such a study to future work.
\section*{Acknowledgments}
We would like to thank Thorsten Feldmann and Thomas Mannel for useful discussions. This research was supported by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under grant 396021762 — TRR 257 ``Particle Physics Phenomenology after the Higgs Discovery''. We also acknowledge the use of the TTP computer cluster at KIT in Karlsruhe.
\bibliographystyle{utphys}
|
\section{Introduction}
\epigraph{“I have no way of knowing whether the events that I am about to narrate are effects or causes.”}{\textit{Jorge Luis Borges \\ Ficciones}}
\numberwithin{equation}{subsection}
The aim of mechanics is to describe a physical system. In modern physics, there are two main ways of dealing with this. Classical mechanics, first formalised by Newton in the $17^{\text{th}}$ century, gives a deterministic description of physics on the macroscopic scale. Quantum mechanics, discovered in the early $20^{\text{th}}$ century, gives a microscopic description of physics that is often interpreted probabilistically. To motivate the subsequent discussion, we will briefly recap some key facts about these theories.
\subsection{Background}
One of the two standard descriptions of classical mechanics on $\mathbb{R}^n$ is the Hamiltonian formalism. Here the phase space is parametrized by canonical coordinates, consisting of generalized coordinates $\textbf{q}:=(q_1,...q_n)$ and conjugate generalized momenta $\textbf{p}:=(p_1,...,p_n)$. The system's energy is represented by a Hamiltonian function $H(\textbf{q},\textbf{p})$; the time evolution is described by Hamilton's equations. Given functions $f(\textbf{q},\textbf{p})$ and $g(\textbf{q},\textbf{p})$, the Poisson bracket is defined in this setting as:
\begin{equation*}\{f,g\} := \frac{\partial f}{\partial q^i}\frac{\partial g}{\partial p_i} - \frac{\partial f}{\partial p_i}\frac{\partial g}{\partial q^i},
\end{equation*}
where we have adopted Einstein notation for repeated indices, as we will for the remainder of the essay. With this construction, the time evolution of $f(\textbf{q},\textbf{p},t)$ can be easily expressed in the form:
\begin{equation*} \frac{\text{d}f}{\text{d}t} = \{f,H\}+\frac{\partial f}{\partial t}.
\end{equation*}
This setup can then be extended to a general phase space. In classical mechanics, the states often form a smooth manifold $M$, typically equipped with a Poisson bracket, making it a Poisson structure. Here, the observables are the real-valued smooth functions $C^{\infty}(M)$; the algebra of observables is commutative. The motivating example is $M=T^{\ast}N$, the cotangent bundle of the space of configurations $N$. This has a coordinate-free Poisson bracket, making it into a Poisson manifold.
On the other hand, in quantum mechanics, the states are rays in a Hilbert space. The observables are then self-adjoint operators acting on the states. The evolution of the quantum system is governed by the Schr{\"o}dinger equation:
\begin{equation*}i\hbar \frac{\text{d}}{\text{d}t}\psi(t)=H\psi(t),
\end{equation*}
where $H$ denotes the quantum observable Hamiltonian. Clearly, quantum observables do not necessarily commute.
Following the formation of quantum theory, many physicists and mathematicians sought to reconcile these systems by \emph{quantizing} classical mechanics. This may seem logically strange, as in principle we should derive classical behaviours from quantum theory. Unfortunately, our only existing methods for modelling quantum systems start with classical descriptions.
\subsection{Quantization}
A natural idea for quantizing a classical system is to convert its observables into operators on a Hilbert space. We call this Hilbert space quantization, where a linear map $Q$ maps functions in $C^{\infty}(M)$ to self-adjoint operators. However, it is clear that classical and quantum observables are incompatible in many ways. For example, we have the problem of operator ordering: $Q(f)$ and $Q(g)$ do not need to commute, but in the classical case $fg = gf$ always. This makes our mapping inconsistent.
A natural idea, proposed by P.A.M. Dirac in \cite{Dirac}, is that $Q$ satisfies the condition:
\begin{equation*}Q(\{f,g\})=\frac{1}{i\hbar}[Q(f),Q(g)].
\end{equation*}
When $f$ and $g$ are canonical coordinates and momenta, we see that this becomes the Heisenberg commutation relation. Here Dirac considered $\hbar$ as a numerical value, Planck's constant, rather than as a formal variable in $\mathbb{R}[[\hbar]]$, a distinction that will be important later in our exposition. It was famously shown by H.J. Groenewold in \cite{Groenewold} that a map satisfying the above condition, which maps arbitrary polynomials in $x$ and $p$ to operator polynomials in $X$ and $P$, does not exist (if we also send the constant function $1$ to the identity).
There are various ways of overcoming this problem, but we must discard some of the assumptions of Groenewold's theorem, for example by restricting the space of functions to be quantized. One standard way of doing this is geometric quantization \cite{Woodhouse}, a type of Hilbert space quantization. A concrete example is the two-sphere:
\begin{equation*} x^2+y^2+z^2=j^2,
\end{equation*}
viewed as a symplectic manifold. In this case, even the simple observables $x,y,z$ can only be quantized for certain $j$. This is because: (i) any geometric quantization maps into a finite-dimensional Hilbert space, and (ii) the two-sphere's Poisson structure induces on the linear span of $x,y,z$ a Lie algebra structure isomorphic to the Lie algebra of angular momentum operators \cite{Hawkins}. In fact, here this implies that $j\in \frac{1}{2}\mathbb{N}$.
\subsection{Deformation quantization}
An alternative approach, and the subject of this essay, is \emph{deformation quantization}. The idea is to deform the commutative algebra of functions in $C^{\infty}(M)$ to a non-commutative, associative algebra. The deformation depends on a formal parameter $\hbar$, so that our new product reduces to the original commutative pointwise product when $\hbar=0$. In \cite{Kontsevich}, Kontsevich showed that a deformed \emph{star product} exists for every finite-dimensional Poisson manifold. He furthermore gave an explicit graphical formula for constructing a star product when the Poisson manifold is an open subset of $\mathbb{R}^d$.
Note the difference with geometric quantization, or Hilbert space quantization in general: we are just deforming the algebra of the observables, without actually introducing a Hilbert space to act on. In this sense, deformation quantization is an abstract, mathematical affair, with ambiguous physical applicability. Everything is formal; we do not even worry about convergence. Returning to the two-sphere, Kontsevich has shown a deformation quantization exists for any $j$, whilst we already know a geometric quantization only exists for $j\in \frac{1}{2} \mathbb{N}$.
\subsection{Outline of work}
In this essay, we will present and explain Kontsevich's results. Along the way they touch on many different (and beautiful) areas of mathematics; we aim to give the reader some appreciation for the surprising connections that emerge.
Kontsevich's main result comes from embedding star products and Poisson structures into a richer mathematical structure, as is often the case in mathematics. We will see that Poisson structures are just deformations of the zero Poisson structure, and that any star product is obviously a deformation of function multiplication:
\begin{equation*}
f \star g = fg + B_1(f,g)\hbar + B_2(f,g){\hbar}^2+...
\end{equation*}
Since we are working with deformation problems, we will need to introduce the notion of a differential graded Lie algebra, or DGLA: specifically, we need a map between the DGLAs of multivector fields and multidifferential operators. A first attempt will be the HKR map of Hochschild, Konstant and Rosenberg \cite{Hochschild}. However, it does not respect the multiplicative structures on either side, a problem because our deformations correspond to solutions of the Maurer-Cartan equation:
\begin{equation*}
\textnormal{d}a+\frac{1}{2}[a,a]=0.
\end{equation*}
Ultimately, we will see that Kontsevich solved this issue with a different kind of mapping, known as an $L_{\infty}$-quasi-isomorphism:
\begin{equation*}
U: T_{poly}(M) \rightarrow D_{poly}(M).
\end{equation*}
We then examine Kontsevich's graphical formula in more depth. For associativity there must be weight coefficients in front of the star product's operators. Remarkably, these real numbers are expressible as rational linear combinations of multiple zeta values, a fact proven constructively in \cite{Panzer}. We will discuss this recent work, as well as other mysterious relationships between this (physically motivated) subject and number theory.
We perform a few original calculations. Firstly, in Section \ref{amend15}, we show Proposition \ref{prop1} using elementary techniques and our method. Using this, we give a simple proof to the degree one part of the HKR theorem in Proposition \ref{prop5}. This is the most relevant case for the purposes of deformation quantization, because Kontsevich's broader formality theorem only modifies the bijection between Maurer-Cartan sets at $O(\hbar^2)$; to our knowledge, an elementary proof did not previously exist in the literature. Finally, Appendix \ref{amend14} contains an original proof of Lemma \ref{lem2}, utilising the algorithm of \cite{Panzer} in an elementary setting (with heavy input from our supervisor). All figures are produced originally.
\subsection{Acknowledgements}
Firstly, I would like to sincerely thank my supervisor, Dr Erik Panzer, who proposed the topic of the dissertation this work is largely drawn from. His help and direction were invaluable throughout the writing of this essay, and were all the more appreciated during an unusual year. Most corrections were based on his scrupulously careful reading and insightful commentary, both of which went far beyond the standard call of marking duty.
I would also like to thank those closest to me during the writing of this essay: my father, for providing intellectual guidance; my mother, for her emotional support and encouragement; and finally my girlfriend, Sophie, and brother, Niall, for both generally putting up with me.
\section{Basics}
Here we introduce the basic algebra necessary to define deformation quantization. This section's material is drawn in large part from \cite{Cattaneo}, \cite{Esposito}, and \cite{Kontsevich}.
\subsection{Poisson structures}
We want to formulate what it means for classical mechanics to exist on a smooth manifold $M$, not just $\mathbb{R}^n$. We begin with:
\begin{defn}[Poisson algebra]
A Poisson algebra is a vector space $V$ over a field $\mathbb{K}$ equipped with two $\mathbb{K}$-bilinear products. Firstly, a commutative product that makes $V$ into an commutative and associative $\mathbb{K}$-algebra. Secondly, a Poisson bracket $\{.,.\}:V\times V \rightarrow V$, that makes $V$ into a Lie algebra, and furthermore satisfies the Leibniz rule:
\begin{equation*} \{x,yz\}=y\{x,z\}+\{x,y\}z.\end{equation*}
\end{defn}
We saw in the introduction the standard Poisson bracket for $\mathbb{R}^n$; this generalises for arbitrary $M$ with:
\begin{defn}[Poisson manifold]
A Poisson manifold is a smooth manifold M, equipped with a product $\{.,.\}:C^{\infty}(M)\times C^{\infty}(M) \rightarrow C^{\infty}(M)$ that makes $(C^{\infty}(M),\{.,.\})$ into a Poisson algebra.
\end{defn}
Here, the pointwise multiplication on $C^{\infty}(M)$ is the commutative product. To further understand this construction, we consider bivector fields. We want the Poisson bracket to satisfy the Leibniz rule in each component separately, in local coordinates $\{x^1,...,x^n\}$. So the map $f\rightarrow \{f,g\}$ is a derivation, with the same true for the other argument. It follows that there exists a unique $ \alpha \in \Gamma(\wedge^{2}TM) \; \text{with} \; \{f,g\} = \alpha(\text{d}f,\text{d}g) = (\text{d}f \otimes \text{d}g) (\alpha)$,
where $\Gamma(\wedge^{2}TX)$ denotes the space of skew-symmetric bivector fields. We can write:
\begin{equation*} \label{ref3} \alpha=\alpha^{ij}\partial_i\otimes \partial_j,\end{equation*}
with $\alpha^{ij}=\{x^i,x^j\}$.
Of course, the fact that $\{.,.\}$ satisfies the Jacobi identity should also impose a condition on the $\alpha^{ij}$. This is most easily formulated in terms of the \emph{Schouten-Nijenhuis bracket}, which acts between multivector fields. Let $\mathfrak{X}^k(M)=\Gamma({\wedge}^k TM)$, with $\mathfrak{X}^0(M)=C^{\infty}(M)$: we will use this notation for the remainder of this essay. On $\mathfrak{X}^1(M)$ there is the usual Lie bracket, which defines the \emph{Lie derivative}:
\begin{equation} \label{ref2}
\mathcal{L}_{X}(Y):=[X,Y]=XY-YX.
\end{equation}
\begin{defn}[Schouten-Nijenhuis bracket]
The Lie bracket (\ref{ref2}) can be extended to an operation: $[ .,.]_S: \mathfrak{X}^a(M) \otimes \mathfrak{X}^b(M) \rightarrow \mathfrak{X}^{a+b-1}(M)$, the Schouten-Nijenhuis bracket, defined on homogeneous elements by:
\begin{multline} \label{ref15}
[X_1\wedge...\wedge X_a,Y_1\wedge...\wedge Y_b]_S := \\ \sum_{i=1}^{a}\sum_{j=1}^{b}(-1)^{i+j}[X_i,Y_j]\wedge X_1 \wedge ... \wedge X_{i-1} \wedge X_{i+1} \wedge ... \wedge X_a \wedge Y_1 \wedge ... \wedge Y_{j-1} \wedge Y_{j+1} \wedge ... \wedge Y_b,
\end{multline}
and extended linearly to all multivector fields.
\end{defn}
Here the wedge product denotes the anti-symmetrization of the tensor product. When $a=1$ this defines the Lie derivative $\mathcal{L}_{X_1}(Y_1 \wedge ... \wedge Y_b)$ of multivector fields. We will later naturally extend these definitions to formal multivector fields without comment. Armed with $[.,.]_S$, we have the easy condition:
\begin{prop}
The bivector $\alpha$ satisfies the Jacobi identity if and only if:
\begin{equation} \label{ref4}
[\alpha,\alpha]_S = 0.
\end{equation}
So this is also the condition necessary for it to define a Poisson structure.
\end{prop}
To prove this, one can show that the Jacobi identity and (\ref{ref4}) are both equivalent to a condition in $\{x^i\}$, but the full calculation is not illuminating (see \cite{Esposito} for details). A final note about this bracket: it makes the space of multivector fields into a differential graded Lie algebra, with a shift of grading (this will be discussed further when we address the formality theorem).
We now consider some simple examples. Firstly, there are constant Poisson structures on $\mathbb{R}^n$, when the $\alpha^{ij}$ are constant functions with $\alpha_{ij}=-\alpha_{ji}$; these will later correspond to the Moyal product.
\begin{exmp}[Poisson structure on $\mathbb{R}^2$]
On $\mathbb{R}^2$ it is easily checked that all smooth Poisson structures are of the form:
\begin{equation*}
\{f,g\}=q(x,y)\left(\frac{\partial f}{\partial x} \frac{\partial g}{\partial y} - \frac{\partial f}{\partial y}\frac{\partial g}{\partial x}\right),
\end{equation*}
where $q: \mathbb{R}^2 \rightarrow \mathbb{R}$ is an arbitrary smooth function.
\end{exmp}
For a smooth Poisson structure on $\mathbb{R}^3$, we can think of $\mathbb{R}^3$ as the dual of the Lie algebra $so(3)$. In fact, given an arbitrary real, finite-dimensional Lie algebra $\mathfrak{g}$, we can easily construct a Poisson structure on the dual $\mathfrak{g}^*$ by identifying $\mathfrak{g}$ with its double dual $\mathfrak{g}^{**}$ \cite{Laurent}. This is because the Lie bracket is an element of $(\Lambda^2 \mathfrak{g}^*) \otimes \mathfrak{g}$.
\begin{exmp}[Poisson structure on $\mathbb{R}^3$]
Using this recipe for $\mathbb{R}^3$, we get a Poisson bracket on $C^{\infty}(\mathbb{R}^3) \times C^{\infty}(\mathbb{R}^3)$, defined by:
\begin{equation*}
\{.,.\} := x \frac{\partial}{\partial y} \wedge \frac{\partial}{\partial z} + y \frac{\partial }{\partial z} \wedge \frac{\partial }{\partial x} + z \frac{\partial }{\partial x} \wedge \frac{\partial }{\partial y}.
\end{equation*}
\end{exmp}
\noindent We have now seen some easy Poisson structures, corresponding to the classical systems we aim to deform. The Poisson bracket will play an important role in bridging classical and quantum mechanics; it is already introducing some non-commutativity into the classical realm.
\subsection{Star products}
Recall that we aim to quantize our system by deforming the pointwise multiplication on $C^{\infty}(M)$ into a non-commutative, associative product. This should depend on a formal parameter $\hbar$, so that $\star$ reduces to pointwise multiplication when $\hbar=0$. To preserve a sense of the quantum mechanical origins of the theory, we use $\hbar$ as our formal variable, but this is not Planck's constant. If we ever wished to actually interpret the results, we would use $\hbar \rightarrow i \hbar$.
\begin{defn}[Star product]
A star product in $C^{\infty}(M)$ is an associative, $\mathbb{R}[\![\hbar ] \!]$ bilinear product:
\begin{equation*} \star: C^{\infty}(M)[\![\hbar ] \!] \times C^{\infty}(M)[\![\hbar ] \!] \rightarrow C^{\infty}(M)[\![\hbar ] \!], \end{equation*}
acting on $f,g\in C^{\infty}(M) \subset C^{\infty}(M)[\![\hbar ] \!] $ as:
\begin{equation}
\label{ref14}
f \star g = \sum_{i=0}^{\infty} B_i(f,g) {\hbar}^i,
\end{equation}
and extended $\mathbb{R}[\![\hbar ] \!]$ linearly. The $B_i$ are bidifferential operators and we require also that $B_0(f,g)=fg$.
\end{defn}
If $\hbar$ was a number, we would hope that the series on the right of (\ref{ref14}) would converge and define a function on M. To avoid worrying about this we just map into $C^{\infty}(M)[\![\hbar ] \!]$, where $\hbar$ is now regarded as a formal variable. Then for associativity we clearly need to take star products between formal power series also, so $\star$ must act on
$ C^{\infty}(M)[\![\hbar ] \!] \times C^{\infty}(M)[\![\hbar ] \!]$.
We can now decompose $B_1$ into symmetric and anti-symmetric parts as:
\begin{equation*} \label{ref8}
B_1^-=\frac{1}{2}(B_1(f,g)-B_1(g,f)), \; \;
B_1^+=\frac{1}{2}(B_1(f,g)+B_1(g,f)).
\end{equation*}
\begin{lemma}
Given a star product $f\star g= fg +B_1(f,g)\hbar+B_2(f,g){\hbar}^2...$, $B_1^-$ defines a Poisson structure on $M$.
\end{lemma}
\begin{proof}
Associativity of $\star$ is equivalent to the condition:
\begin{equation}
\label{ref6}
\sum_{i+j=n} B_i(B_j(f,g),h) = \sum_{i+j=n} B_i(f,B_j(g,h)), \; \; \forall f,g,h \in C^{\infty}(M).
\end{equation}
For $n=1$ this implies the relation:
\begin{equation}
\label{ref7}
fB_1(g,h)-B_1(fg,h)+B_1(f,gh)-B_1(f,g)h=0,
\end{equation}
with the same identity for permutations of $f,g$ and $h$ (we will call this the cocycle condition for reasons that will become clear later in the exposition). Now, adding (\ref{ref7}) to one cyclic and one non-cyclic permutation implies that $B_1^-$ satisfies the Leibniz rule. If we take (\ref{ref6}) for $n=2$:
\begin{equation*}
B_2(fg,h)-B_2(f,gh)+B_1(B_1(f,g),h)-B_1(f,B_1(g,h))+B_2(f,g)h-fB_2(g,h)=0,
\end{equation*}
add the two cyclic permutations, then subtract the three non-cyclic permutations, we get exactly that $B_1^-$ satisfies the Jacobi identity.
\end{proof}
We can see a relationship emerging between Poisson structures and star products, but it needs to be made more mathematically precise. This leads to:
\begin{defn}[Deformation quantization] \label{amend13}
A deformation quantization of a Poisson manifold is a star product for which the induced bracket satisfies $B_1^-(.,.)=\frac{1}{2}\{.,.\}$, where $\{.,.\}$ denotes the original Poisson bracket.
\end{defn}
From this definition, it is immediate that the ordinary product on functions extended $\mathbb{R}[\![\hbar]\!]$ linearly to $C^{\infty}(M)[\![\hbar]\!]$, which is commutative, corresponds to the zero Poisson structure. Note that the factor of $\frac{1}{2}$ means that the commutator of the star product satisfies Dirac's original quantization condition, at order $\hbar$.
\section{Kontsevich's result} \label{amend15}
The material in the next two sections draws on \cite{Cattaneo}, \cite{Keller}, \cite{Esposito}, \cite{Gutt} and \cite{Kontsevich}.
\subsection{Formal star products}
Let $J$ be the set of automorphisms of $C^{\infty}(M)[\![\hbar ] \!]$, taken as a $\mathbb{R} [\![\hbar ] \!]$ module, of the following form on $f \in C^{\infty}(M)$:
\begin{equation*}
T(f) = \sum_{n=0}^{\infty}T_n(f){\hbar}^n.
\end{equation*}
Here the $T_n$ are linear differential operators. We require that they vanish on constants and also that $T_0$ is the identity. Given $T$ in this form, we can easily solve for the inverse, $D=T^{-1}$, in the same form,
where the $D_n$ are inductively defined by:
\begin{equation*}
D_0=\mathds{1}, \; \; D_n=-\sum_{m=0}^{n-1} T_m D_{n-m}.
\end{equation*} The fact that $T_0$ is the identity implies that the shifting is only happening in the formal portions of $C^{\infty}(M)[\![\hbar]\!]$. Now we have:
\begin{defn}[Equivalent star products] \label{ref24}
If there exists $T \in J$ such that:
\begin{equation*}
T(u\star v) = T(u)\star' T(v), \; \; \forall u,v \in C^{\infty}(M) [\![\hbar]\!],
\end{equation*}
then $\star$ and $\star'$ belong to the same star product equivalence class. We also say that they are gauge equivalent.
\end{defn}
We now state and prove two important results about gauge equivalence. Firstly:
\begin{lemma} \label{lem1}
The $B_1^-$ Poisson bracket depends only on the star product equivalence class.
\end{lemma}
\begin{proof}
\emph{(Taken from \cite{Esposito})}. Given an arbitrary star product $\star$, apply $T\in J$, obtaining a gauge equivalent star product $\star'$:
\begin{equation*} \label{ref37}
\sum_{i=0}^{\infty} B_i(f,g) {\hbar}^i \rightarrow_T
\sum_{i=0}^{\infty} {C_i} (f,g) {\hbar}^i.
\end{equation*}
Using the formal expansion of $T$, this translates to the following condition for $T_1$:
\begin{equation} \label{ref38}
B_1(f,g)+T_1(fg)=C_1(f,g)+T_1(f)g+fT_1(g).
\end{equation}
So $B_1(f,g)-C_1(f,g)$ is symmetric in $f$ and $g$ and does not contribute to the induced bracket.
\end{proof}
Secondly:
\begin{prop} \label{prop1}
Any star product is gauge equivalent to a star product such that $B_1^+=0$.
\end{prop}
Kontsevich provides a highly abstract proof for this statement when he proves his own version of the HKR theorem, the entry point to the formality theorem, but the machinery used goes beyond the scope of this essay. Going through the calculation in depth is useful, as the ideas will later generalise to the HKR discussion. The existing proofs we found in the literature were either abstract or very sketchy with the details. So now we provide our own.
\begin{proof}
Starting with an arbitrary star product, apply $T\in J$:
\begin{equation*}
\sum_{i=0}^{\infty} B_i(f,g) {\hbar}^i \rightarrow_T
\sum_{i=0}^{\infty} {C_i} (f,g) {\hbar}^i.
\end{equation*}
Then, (\ref{ref38}) gives us an explicit form for $C_1(f,g)$:
\begin{equation} \label{amend4}
C_1(f,g)=B_1(f,g)-fT_1(g)+T_1(fg)-T_1(f)g.
\end{equation}
Let $(x_k)$ be local coordinates for $M$.
Say we start with the bidifferential operator $B_1(f,g)$:
\begin{equation*}
B_1= \sum_{I,J} a_{I,J} \partial_I \otimes \partial_J,
\end{equation*}
where the $I,J$ are coordinate multi-indices, such that $a_\mathbf{0}$ is non-zero, i.e. there is a term with no derivatives. Firstly choosing $T_1(f)= a_\mathbf{0} f$, we have that $C_1(f,g)=B_1(f,g)-a_\mathbf{0} f$, so we can assume up to gauge equivalence that $B_1$ has no order zero term.
We now want to enforce:
\begin{equation} \label{amend5}
C_1(f,g)+C_1(g,f)=0, \; \; \forall f,g \in C^\infty(M)
\end{equation}
Using (\ref{amend4}), (\ref{amend5}) is now equivalent to the $T_1$ condition:
\begin{equation} \label{ref30}
T_1(fg)=fT_1(g)+T_1(f)g-B_1^+(f,g), \; \; \forall f,g \in C^\infty(M).
\end{equation}
We will use this to define $T_1$ locally on smooth functions.
Assume that $T_1$ is defined separately on monomials containing only $\{x_1,...,x_{n-1}\}$ and monomials containing only ${x_n}$, such that (\ref{ref30}) holds true in either case. Define $T_1$ on monomials of $\{x_1,...x_n\}$, expressible as $x_n^{k_1} p_1$, $p_1=p_1(x_1,...,x_{n-1})$, as:
\begin{equation} \label{ref31}
T(x_n^{k_1}p_1):=T(x_n^{k_1})p_1+x_n^{k_1}T(p_1)-B_1^+(x_n^{k_1},p_1).
\end{equation}
Take two arbitrary such monomials $x_n^{k_1}p_1$ and $x_n^{k_2}p_2$. We want to show that:
\begin{equation} \label{ref32}
T(x_n^{k_1} p_1 x_n^{k_2} p_2) = T(x_n^{k_1}p_1)x_n^{k_2} p_2+x_n^{k_1}p_1T(x_n^{k_2} p_2)-B_1^+(x_n^{k_1}p_1,x_n^{k_2} p_2).
\end{equation}
Now using our definition (\ref{ref31}), we have that:
\begin{equation} \label{ref33}
T_1(x_n^{k_1} p_1 x_n^{k_2} p_2) = T_1(x_n^{k_1+k_2})p_1p_2+x_n^{k_1+k_2}T_1(p_1 p_2)-B_1^+(x_n^{k_1+k_2},p_1p_2).
\end{equation}
Now reapplying (\ref{ref31}) to (\ref{ref32}) and (\ref{ref33}), we see that their equivalence can be simplified to:
\begin{multline*}
p_1p_2B_1^+(x_n^{k_1},x_n^{k_2})+x_n^{k_1+k_2}B_1^+(p_1,p_2)+B_1^+(x_n^{k_1+k_2},p_1p_2) \\
= B_1^+(x_n^{k_1},p_1)x_n^{k_2}p_2+x_n^{k_1}p_1B_1^+(x_n^{k_2},p_2)+B_1^+(x_n^{k_1}p_1,x_n^{k_2}p_2).
\end{multline*}
But then this just comes from the associativity of the star product for different bracketing of four terms, restricting operators to the symmetric part.
By induction we need only define $T_1$ for monomials of the form $x_k^n$, again so that (\ref{ref30}) holds. Before doing this, recall that:
\begin{equation} \label{amend9}
B_1^+= \sum_{I,J} a_{I,J} \partial_{I} \otimes \partial_J.
\end{equation}
Now consider the cocycle condition (\ref{ref7}) for $B_1$; noting that it still holds when the arguments are swapped, it must also hold for the operator $B_1^+$. Imposing $h=1$ gives us:
\begin{equation} \label{amend8}
fB_1^+(g,1)-B_1^+(fg,1)=0, \; \; \forall f,g \in C^\infty(M),
\end{equation}
where we have cancelled terms by the symmetry of the operator.
Substituting (\ref{amend9}) into (\ref{amend8}), note that all terms except those with $\partial_J=\mathds{1}$ vanish, so that:
\begin{equation*}
f\sum_{I} a_{I,0}\partial_I(g)-\sum_{I}a_{I,0}\partial_{I}(fg)=0, \; \; \forall f,g \in C^\infty(M).
\end{equation*}
In other words, the operator $\sum_{I} a_{I,0} \partial_I$ is function linear, and hence is simply multiplication by $a_{0,0}$, but by gauge equivalence we have already chosen this to be $0$. Thus $B_1^+$ has no terms of the form $a_{I,0} \partial_I \otimes \mathds{1}$, and since $a_{I,J}=a_{J,I}$, there are no terms of the form $a_{0,I}\mathds{1}\otimes \partial_{I}$ either.
We will now define $T_1$ on monomials of the form $x^n$. Take $T_1$ to vanish on constant functions and linear functions. We then define:
\begin{equation} \label{ref35}
T_1(x^n) := T_1(x^{n-1})x-B_1^+(x^{n-1},x), \; \; n \geq 2.
\end{equation}
We want to show that:
\begin{equation} \label{ref34}
T_1(x^{n} x^{m})=T_1(x^{n})x^{m}+x^{n} T_1(x^{m})-B_1^+(x^{n},x^{m}), \; \; \forall n,m.
\end{equation}
Fix $n$ arbitrarily; use induction on $m$. Say $T_1$ is defined such that (\ref{ref34}) holds for all $m \leq k$. We have that:
\begin{equation*}
T_1(x^{n+k+1})=T_1(x^{n+k})x-B_1^+(x^{n+k},x).
\end{equation*}
This is equivalent to:
\begin{equation*}
T_1(x^{n})x^{k+1}+x^nT_1(x^{k+1})-B_1^+(x^{n},x^{k+1}),
\end{equation*}
if and only if the condition:
\begin{equation*}
xB_1^+(x^n,x^k)+B_1^+(x^{n+k},x)=B_1^+(x^n,x^{k+1})+x^nB_1^+(x,x^{k}),
\end{equation*}
is satisfied. But this is just simple associativity of the star product, again restricting to the symmetric part.
$T_1$ is now defined for local monomials, and by linearity for polynomials. The condition (\ref{ref30}) holds for polynomials $f$ and $g$ as both sides of the equality are bilinear in $(f,g)$. It is stated in \cite{Cattaneo} that this is enough to define a differential operator with the desired properties on all local smooth functions. Unconvinced with the reasoning, we provide our own explanation.
It is sufficient to show that our $T_1$ is expressible as a differential operator on all monomials. Start with one variable, say $x_k$. Using (\ref{ref35}), we have that:
\begin{equation} \label{ref36}
T_1(x_k^n)=x_kT_1(x_k^{n-1})-B_1^+(x_k,x_k^{n-1})=...=-\sum_{i=1}^{n-1}x_k^{n-1-i}B_1^+(x_k,x_k^i).
\end{equation}
Recall the decomposition (\ref{amend9}) for the operator $B_1^+$. We have already shown, through the cocycle condition, that it does not contain terms of the form: $\alpha_j(\mathds{1}\otimes \partial_k^j +\partial_k^j \otimes \mathds{1})$. But then $B_1^+(x_k,x_k^i)$ can only be non-zero for a special case, namely, for terms of the form:
\begin{equation} \label{amend11}
\beta_j(\partial_k \otimes \partial_k^j+\partial_k^j \otimes \partial_k),
\end{equation}
restricting to $1 \leq j \leq n-1$. For $j=1$ (\ref{ref36}) becomes:
\begin{equation*}
-\sum_{i=1}^{n-1} 2 \beta_1 i x_k^{n-1-i} x_k^{i-1}= -\sum_{i=1}^{n-1} 2 \beta_1 i x_k^{n-2}= -{n(n-1)}\beta_1 x_k^{n-2}=-\beta_1(\partial_k)^2(x_k^n):= D_1^1(x_k^n),
\end{equation*}
whilst for $2\leq j \leq n-1$ it becomes:
\begin{multline*}
-\sum_{i=1}^{n-1} \beta_j x_k^{n-1-i} \frac{i!}{(i-j)!}x^{i-j}= -\beta_j \left( \sum_{i=j}^{n-1} \frac{i!}{(i-j)!} \right) x_k^{n-j-1} \\ = - \beta_j \frac{n(n-1)...(n-j)}{j+1} x_k^{n-j-1}=-\frac{\beta_j}{j+1}(\partial_{k})^{j+1}(x^n_k):=D_j(x_k^n),
\end{multline*}
where we have used a combinatorial identity for the sum of partial factorials to treat the summed term in brackets.
Defining $D_{x_k}:=\sum_{j} D_j$, these neat simplifications give us an expression for $T_1$ on single coordinate monomials as a single differential operator.
For a monomial of $x_k$ and $x_l$, $k \neq l$, say $x_k^i x_l^j$, we use the the condition (\ref{ref30}) and identical reasoning. We have:
\begin{multline*}
T_1(x_k^i x_l^j)=T_1(x_k^i)x_l^j+x_k^iT_1(x_l^j)-B_1^+(x_k^i,x_l^j)
=\frac{1}{2}(D_{x_k}+D_{x_l})(x_k^ix_l^j)+\sum_{\gamma, \delta} \alpha_{\gamma,\delta}(\partial_k)^\gamma\otimes(\partial_l)^\delta(x_k^i,x_l^j)
\\= \frac{1}{2}(D_{x_k}+D_{x_l})(x_k^ix_l^j)+\sum_{\gamma, \delta} \alpha_{\gamma, \delta} (\partial_k)^\gamma(\partial_l)^\delta(x_k^ix_l^j):=D_{x_k, x_l}(x_k^i x_l^j).
\end{multline*}
Arguing inductively on variables we have an expression for $T_1$ as a differential operator $D_{x_1...x_n}$ on all variables with the required properties.
We have found a differential operator $D$ satisfying (\ref{ref30}) on all polynomials of $\{x_1,...,x_n\}$; we need to check that (\ref{ref30}) holds for all smooth functions $f$ and $g$. Because this an equality of functions, we only need to check at points, where both sides only depend on polynomials of sufficiently large Taylor expansions of $f$ and $g$, to which we apply our previous arguments.
More precisely, given a point $p \in M$ and $f, g\in C^\infty(M)$, consider the set of all polynomial functions $h$ such that $h$ agrees exactly with the first $n$ terms of the Taylor series of $f$ at $p$; let $f_n$ denote any arbitrary element of this set. As $D$ and $B_1^+$ are differential operators with some highest-order component, by Taylor's theorem, we know that there exists $k \in \mathbb{N}$ such that:
\begin{equation*}
D(f)(p) = D(f_n)(p), \; \; B_1^+(f,g)(p) = B_1^+(f_n,g_n)(p), \; \; \forall n \geq k.
\end{equation*}
We moreover know that there exists $m \geq k$ such that:
\begin{equation*}
D(f_n g_n)(p) = D((fg)_k)(p)=D(fg)(p), \; \; \forall n \geq m,
\end{equation*}
with $m$ chosen large enough for the Taylor series of $fg$ to be covered up to order $k$ by $f_n g_n$. Now we are done, using the fact that we have already proved (\ref{ref30}) for all polynomials; we have that:
\begin{multline*}
D(fg)(p) = D(f_m g_m)(p) \\ =f_m (p)D(g_m)(p)+ D(f_m)(p)g_m(p)-B_1^+(f_m,g_m)(p)= f(p)D(g)(p) +D(f)(p)g(p)-B_1^+(f,g)(p).
\end{multline*}
Finally, choose a coordinate covering for our manifold: \begin{equation*}
M=\bigcup_i U_i.
\end{equation*} Now using the previous arguments for $\mathbb{R}^n$ locally, we obtain a collection of local differential operators $D_i$. Choose a partition of unity $\{\rho_i\}_{i \in I}$ subordinate to the cover $\bigcup_i U_i$, and define $D:= \sum_i \rho_i D_i $, noting that multiplication by $\rho_i$ makes $D_i$ into a global operation. Using that $\sum_i \rho_i =1$, we obtain the global relationship:
\begin{equation*}
D(fg)=fD(g)+D(f)g-B_1^+(f,g),
\end{equation*}
and as $D$ is locally a differential operator, it is a global differential operator with all the required properties.
\end{proof}
Because we are only considering star products up to gauge equivalence, Proposition \ref{prop1} implies that if we can quantize a Poisson structure with $B_1^-=\{.,.\}$, we can quantize with $B_1=\{.,.\}$. From here on we will just consider this.
\subsection{Formal Poisson structures}
We can now do something similar for Poisson structures:
\begin{defn}[Formal Poisson structures]
A formal Poisson structure is a formal power series $\pi_{\hbar}= \sum_{i=0}^{\infty} \hbar^i \pi_i$, where the $\pi_i$ are skew-symmetric bivector fields. Also the Schouten-Nijenhuis bracket of $\pi_{\hbar}$ with itself must vanish at each $\hbar$ order.
\end{defn}
Clearly, we do not need each $\pi_i$ to be a Poisson structure, only the leading order term $\pi_0$. For the actual structures we will deform, $\pi_0=0$. Then the formal power series can be thought of as deformations of the zero Poisson structure. When are formal Poisson structures equivalent? The recipe is as follows:
\begin{enumerate}
\item Given a diffeomorphism $\phi: M \rightarrow N$, we have the pushforward $\phi_*$, where $\phi_* [X,Y]_S=[\phi_* X, \phi_* Y ]_S$.
\item Because the $[.,.]_S$ bracket of a Poisson structure $\pi$ with itself is invariant of the pushforward, the group of diffeomorphisms of $M$ acts on Poisson structures, with $\pi_{\phi}=\phi_* \pi$.
\item We can now generalise this action to formal Poisson structures.
\end{enumerate}
\begin{defn}[Equivalent formal Poisson structures] \label{ref22}
The formal Poisson structures $\pi_{\hbar}$ and $\pi_{\hbar}'$ are equivalent if there exists a formal diffeomorphism between them, so that:
\begin{equation*}
\pi_{\hbar}= \sum_{n=0}^{\infty} \frac{{\hbar}^n}{n!} (\mathcal{L}_X)^n \pi_{\hbar}^{'} = \text{exp}(\hbar \mathcal{L}_X) \pi_{\hbar}'.
\end{equation*}
Here $X$ is a formal vector field lying in $\mathfrak{X}^1(M)[\![\hbar]\!]$ and $\mathcal{L}_X$ is the Lie derivative.
\end{defn}
\subsection{Kontsevich's theorem}
We now have everything necessary to formulate Kontsevich's result.
\begin{thm}[Kontsevich, \cite{Kontsevich}] \label{ref27}
Every Poisson manifold has a deformation quantization. Furthermore, the set of gauge equivalence classes of star products on a smooth manifold $M$ can be naturally identified with the set of equivalence classes of Poisson structures depending formally on $\hbar$:
\begin{equation*}
\alpha_{\hbar} = \alpha(\hbar) = \alpha_1 \hbar + \alpha_2 {\hbar}^2+... \in \Gamma (\wedge^2 TM) [\![\hbar]\!], \; \; [\alpha,\alpha]_S = 0 \in \Gamma (\wedge^3 TM) [\![\hbar]\!],
\end{equation*}
modulo the group of formal diffeomorphisms.
\end{thm}
As an example, we have a deformation quantization of our constant Poisson structure on $\mathbb{R}^n$, namely:
\begin{exmp}[The Moyal product]
\begin{equation} \label{ref13}
f \star g:= \sum_{n=0}^{\infty} \frac{{\hbar}^n}{n!} \left( \prod_{k=1}^{n} \alpha^{i_k j_k} \right) \left( \prod_{k=1}^{n} \partial_{i_k} f \right) \left( \prod_{k=1}^{n} \partial_{j_k} g \right).
\end{equation}
\end{exmp}
This was known before Kontsevich, but we will see later how it is rederived from his formula. It is the simplest non-trivial example we can consider.
\section{The formality theorem}
The goal of the following sections is to state Kontsevich's formality theorem, via describing the richer mathematical structure referred to in the introduction.
\subsection{Differential graded Lie algebras}
Recall that a \emph{graded vector space} $V$ is just a vector space that breaks up into a direct sum of the form: $V=\bigoplus_{n \in \mathbb{Z}} V^n$. If $v \in V^n$, we say $\text{deg}(v)=\bar{v}=n$. We now have:
\begin{defn}[Differential graded Lie algebra] \label{ref18}
A differential graded Lie algebra $(V, \textnormal{d}, [.,.])$ is a graded vector space $V=\bigoplus_{n \in \mathbb{Z}} V^n$, together with a linear map $\textnormal{d}:V \rightarrow V$, and a bilinear map $[.,.]: V^i \otimes V^j \rightarrow V^{i+j}$, satisfying all the following conditions for any homogeneous $x,y,z \in V$: \begin{enumerate}
\item \textnormal{d}, the differential, makes $(V,\textnormal{d})$ into a chain complex: $\textnormal{d}(V^n) \in V^{n+1}$ and $\textnormal{d}^2=0$.
\item $[.,.]$ is graded skew-symmetric: $[x,y]=(-1)^{\bar{x}\bar{y}+1}[y,x]$.
\item $[.,.]$ satisfies the graded Jacobi identity: $[x,[y,z]]=[[x,y],z]+(-1)^{\bar{x} \bar{y}}[y,[x,z]]$.
\item $[.,.]$ and \textnormal{d} satisfy the graded Leibniz rule: $\textnormal{d}[x,y]=[\textnormal{d}x,y]+(-1)^{\bar{x}}[x,\textnormal{d}y]$.
\end{enumerate}
\end{defn}
Simply put, a DGLA is a complex with a compatible graded Lie algebra structure. If there is no differential, but just a bracket satisfying the above rules, this is a \emph{graded Lie algebra}, or GLA. Any GLA can be made into a trivial DGLA using the differential d $=0$. We say that elements in the kernel of d are \emph{closed}, or \emph{cocycles}, and that elements in image of d are \emph{exact}, or \emph{coboundaries}.
\begin{defn}[Cohomology complex]
Given a DGLA $V$, there is the cohomology quotient group:
\begin{equation*}
H^i(V) := \left. \textnormal{Ker}(\textnormal{d}: V^i \rightarrow V^{i+1}) \middle/ \textnormal{Im}(\textnormal{d}: V^{i-1} \rightarrow V^i ) \right..
\end{equation*}
Then the $H := \bigoplus_{n} H^i(V)$ is itself a DGLA. It has an obvious graded vector space structure; an inherited bracket defined on equivalences classes $|x|,|y|$ by $[|x|,|y|]_H$ $:= |[x,y]_V|$; and finally a differential \textnormal{d} just set to be $0$.
\end{defn}
Now a \emph{DGLA morphism} is a linear map $\phi: V_1 \rightarrow V_2$ of degree $0$, i.e. $\phi(V_1^i) \subset V_2^i$, that commutes with the brackets and differentials. Obviously a DGLA morphism $\phi:V_1 \rightarrow V_2$ induces a morphism $H(\phi): H_1 \rightarrow H_2$, between the cohomologies.
\begin{defn}[Quasi-isomorphism] \label{ref19}
A quasi-ismorphism of DGLAs is a morphism of DGLAs that induces an isomorphism between the cohomologies.
\end{defn}
It is important to note that this morphism of DGLAs is compatible with the brackets, of both the DGLAs and cohomologies. On the other hand, a \emph{quasi-isomorphism of complexes} only respects the differential. This distinction will soon be important.
\subsection{DGLA examples}
As stated in the introduction, deformations of mathematical structures are controlled by DGLAs \cite{Manetti}. Here we are concerned with deformations of function multiplication and the zero Poisson structure, so it is natural DGLAs arise. Kontsevich identified two key DGLAs associated to these deformation problems.
\begin{defn}[$\mathbf{T_{poly}(M)}$]
Let $T_{poly}^n(M):=\mathfrak{X}^{n+1}(M)$. Take the graded vector space:
\begin{equation*}
T_{poly}(M) := \bigoplus_{n=-1}^{\infty} T_{poly}^n(M).
\end{equation*}
The Schouten-Nijenhuis bracket for homogeneous elements is already defined by (\ref{ref15}). $T_{poly}(M)$ is a DGLA with $[.,.]_S$ and the differential \textnormal{d} $=0$, the DGLA of multivector fields.
\end{defn}
Note that we need to shift the degree in the direct sum to respect the definition of the bracket in Definition (\ref{ref18}). The cohomology just coincides with the DGLA itself. For our second example we must first define the Hochschild DGLA corresponding to an associative unital algebra $A$ (over a ring $\mathbb{K}$); it is just the DGLA on the complex of multilinear maps from $A$ to itself.
\begin{defn}[Hochschild DGLA] \label{amend1}
Let $C^n(A):=\text{Hom}_{\mathbb{K}}(A^{\otimes(n+1)},A)$. Take the graded vector space:
\begin{equation*}
C:=\bigoplus_{n=-1}^{\infty}C^n.
\end{equation*}
Define a composition between $\phi \in C^{m_1}(A)$ and $\psi \in C^{m_2}(A)$ as:
\begin{equation} \label{ref16}
(\phi \circ \psi)(f_0,....,f_{{m_1}+{m_2}}):= \sum_{i=0}^{m_1} (-1)^{(m_2)(i)}\phi(f_0,...,f_{i-1},\psi(f_i,...,f_{i+m_2}),f_{i+m_2+1},...,f_{m_1+m_2}),
\end{equation}
for arbitrary $f_i, 0 \leq i \leq m_1+m_2+1$, elements of $A$. The Gerstenhaber bracket on homogeneous elements: $[.,.]_G:C^{m_1}\otimes C^{m_2} \rightarrow C^{m_1+m_2}$, is given by:
\begin{equation*}
[\phi,\psi]_G:=\phi \circ \psi - (-1)^{m_1 m_2} \psi \circ \phi.
\end{equation*}
We also have a differential \textnormal{d}: $C^n\rightarrow C^{n+1}$. An explicit formula can be written as:
\begin{multline} \label{ref17}
(-1)^n(\textnormal{d} \phi) (f_0,...,f_{n}) := \sum_{i=0}^{n-1}(-1)^{i+1}\phi(f_0,...,f_{i-1},f_if_{i+1},...,f_{n})\\ +f_0\phi(f_1,...,f_{n})+(-1)^{n-1}\phi(f_0,...,f_{n-1})f_{n}=(-\phi \circ \mu+(-1)^n \mu \circ \phi )(f_0,...,f_{n})=-[\phi,\mu]_G(f_0,...,f_{n}),
\end{multline}
where $\mu \in C^1(A)$ denotes the ordinary product for $A$. Now $C$ is a DGLA with $[.,.]_G$ and the differential (\ref{ref17}).
\end{defn}
The fact this is a DGLA can be proved, for example in \cite{Cattaneo}, where they also provide a nice pictorial representation of the composition map (\ref{ref16}). There is again a shift in degree; we naturally interpret $C^{-1}(A)$ as just $A$ itself. Note also that the differential has been simply expressed in terms of the Gerstenhaber bracket. We now have:
\begin{lemma} \label{amend12}
An element of $C^1(A)$ is an associative multiplication of the algebra $A$ if and only if the Gerstenhaber bracket with itself vanishes.
\end{lemma}
\begin{proof}
\emph{(Taken from \cite{Cattaneo}).} Given $\phi \in C^1(A)$, it is clear by Definition \ref{amend1} that $\phi$ is an associative multiplication of $A$ if and only if it satisfies the associativity condition:
\begin{equation} \label{amend2}
\phi(\phi(f,g),h)-\phi(f, \phi(g,h))=0, \; \; \forall f,g,h \in A.
\end{equation}
But now:
\begin{equation} \label{amend3}
[\phi, \phi]_G(f,g,h)=2\phi \circ \phi (f,g,h)=2( \phi(\phi(f,g),h)-\phi(f,\phi(g,h))).
\end{equation}
Clearly, (\ref{amend3}) vanishes $\forall f,g,h \in A$ if and only if condition (\ref{amend2}) holds.
\end{proof}
Let us now define $D_{poly}(M)$, simply restricting our last example to multidifferential operators. From here on $A$ denotes the specific algebra $C^{\infty}(M)$ and take $\mathbb{K}=\mathbb{R}$.
\begin{defn}[$\mathbf{D_{poly}(M)}$]
Let $D_{poly}^n(M)$, $n \geq -1$, be the subset of $\text{Hom}(A^{\otimes(i+1)},A)$ containing only the multidifferential operators. An expression for the general form of these maps, taken from \cite{Kontsevich}, is:
\begin{equation}
f_0 \otimes... \otimes f_n \rightarrow \sum_{(I_0,...,I_n)} C^{I_0,...,I_n}\partial_{I_0}(f_0)...\partial_{I_n}(f_n),
\end{equation}
where the sum is finite, the $I_k$ denote multi-indices, and the $f_k$, $C^{I_0,...I_n}$ are functions in the local coordinates. Now take the graded vector space:
\begin{equation*}
D_{poly}(M):= \bigoplus_{n=-1}^{\infty} D_{poly}^n(M).
\end{equation*}
$D_{poly}(M)$ is closed under the Gerstenhaber bracket and the Hochschild differential. It is therefore a DGLA, with $[.,.]_G$ and \textnormal{d}, (\ref{ref17}).
\end{defn}
Why have we defined these DGLAs? Several relationships can already be stated. Given a star product $f\star g= fg+B_1(f,g)\hbar+...$, the bidifferential operator $B_1$ is in $D^1_{poly}(M)$. If we use a gauge transformation $T$ to transform $\star$ into $\star'$, as in (\ref{ref37}), the formula for $C_1$ gives:
\begin{equation} \label{ref41}
C_1(f,g)-B_1(f,g)=T_1(fg)-T_1(f)g-fT_1(g)=\text{d}(-T_1)(f,g),
\end{equation}
an expression for gauge equivalence at order $\hbar$ in terms of the differential. So the gauge equivalence classes of closed bidifferential operators are just $H^1(D_{poly}(M))$.
Clearly Poisson structures on $M$ lie in $T_{poly}(M)$. Now $D_{poly}(M)[\![\hbar]\!]$ and $T_{poly}(M)[\![\hbar]\!]$ are also DGLAs, just tensoring with $\mathbb{R}[\![\hbar]\!]$ and extending operations $\mathbb{R}[\![\hbar]\!]$ linearly. Then formal Poisson structures that are deformations of the zero Poisson structure lie in $\hbar T_{poly}^1(M)[\![\hbar]\!]$ and star products correspond to elements of $\hbar D_{poly}^1(M)[\![\hbar]\!]$.
\subsection{The Maurer-Cartan equation}
We now explain our assertion in the introduction that Poisson structures and star products correspond to solutions of the Maurer-Cartan equation.
\begin{defn}[The Maurer-Cartan equation]
Given a DGLA $V$ there is the Maurer-Cartan equation:
\begin{equation*}
\textnormal{d}a+\frac{1}{2}[a,a]_V=0, \; \; a \in {V^1}.
\end{equation*}
For $V=D_{poly}(M)[\![\hbar]\!]$ or $T_{poly}(M)[\![\hbar]\!]$, the Maurer-Cartan set $MC(V)$ equals $\gamma \in \hbar V^1$ such that \textnormal{d}$\gamma+\frac{1}{2}[\gamma,\gamma]=0$.
\end{defn}
Consider the Maurer-Cartan sets of our two important DGLAs:
\begin{exmp}[$\mathbf{MC(T_{poly}(M))}$]
We know that $\pi$ is a Poisson structure $\iff$ $[\pi,\pi]_S=0$. Since for $\hbar T_{poly}(M)[\![\hbar]\!]$, $\textnormal{d}=0$, we see that formal Poisson structures with $\pi_0=0$ coincide with $MC(T_{poly}^1(M)[\![\hbar]\!])$.
\end{exmp}
\begin{exmp}[$\mathbf{MC(D_{poly}(M))}$]
If $B \in \hbar D_{poly}^1(M)[\![\hbar]\!]$, by Lemma \ref{amend12} we know that the deformation $\mu+B$ is associative $\iff$ $[\mu+B,\mu+B]_G=0$. Now $[\mu,\mu]_G=0$, and as $\bar \mu = \bar B = 1$, we also have that $\text{d}(B)=[\mu,B]_G$ and $[\mu,B]_G=[B,\mu]_G$. Thus $B$ defines a star product $\mu+B \iff B \in MC(D_{poly}^1(M)[\![\hbar]\!])$.
\end{exmp}
\subsection{Gauge groups}
Finally, we also have gauge groups for our DGLAs. These further translate the previous discussions of gauge equivalence into our new framework. Take $G^{0}( T_{poly}(M)[\![\hbar]\!])$ to be the group of formal diffeomorphisms given in Definition \ref{ref22}; the group structure comes from the Baker-Campbell-Hausdorff formula:
\begin{equation*}
\textnormal{exp}(\hbar X) \textnormal{exp}(\hbar Y):= \textnormal{exp}\left( \hbar X + \hbar Y = \frac{1}{2} \hbar [X,Y]+...\right).
\end{equation*}
Because the exponential of a derivation is an automorphism of the DGLA, the bracket and Maurer-Cartan set is preserved by $G^0(T_{poly}(M)[\![\hbar]\!])$. We can thus define the quotient space:
\begin{equation*} \label{ref23}
\overline{MC}(T_{poly}(M)[\![\hbar]\!]) := \frac{MC(T_{poly}(M)[\![\hbar]\!])}{G^0( T_{poly}(M)[\![\hbar]\!])},
\end{equation*}
corresponding with formal Poisson structure equivalence classes. Similarly, define:
\begin{equation*}
G^0(D_{poly}(M)[\![\hbar]\!]) \\ := \{ \phi: D_{poly}(M)[\![\hbar]\!] \rightarrow D_{poly}(M)[\![\hbar]\!] : \phi = e^{\hbar[S,.]},\; \; S \in D_{poly}^0(M)[\![\hbar]\!] \}.
\end{equation*}
Then taking the quotient set $\overline{MC}(D_{poly}(M)[\![\hbar]\!])$, elements in the same equivalence class correspond to gauge equivalent star products, in the sense of Definition \ref{ref24} (for a proof see \cite{Waldmann}). In fact, these constructions exist for any DGLA, but for simplicity we restrict to our examples.
\subsection{The HKR map}
From here on, we may sometimes be somewhat informal in our definitions, as anything defined for $T_{poly}(M)$ or $D_{poly}(M)$ can just be extended linearly to the tensors with $\mathbb{R}[\![\hbar]\!]$. Of course, our goal is ultimately to associate our two deformation problems. The first step towards achieving this is the HKR theorem.
\begin{thm}[Hochschild-Konstant-Rosenberg, \cite{Hochschild}] \label{ref25}
Define the $A$-linear map $U_1: T_{poly}(M) \rightarrow D_{poly}(M)$ on homogeneous elements in $T_{poly}(M)$ as :
\begin{equation*} \label{ref29}
(U_1(X_0 \wedge ... \wedge X_n))(f_0,...,f_n) := \frac{1}{(n+1)!}\sum_{\sigma \in S_{n+1}} \text{sgn}(\sigma) \prod_{i=0}^{n} X_{\sigma(i)}(f_i),
\end{equation*}
with $f_0,...,f_n$ functions on $M$ and $\sigma$ denoting a permutation in the set of permutations of $n$ elements, $S_n$. Then this map is a quasi-isomorphism of complexes.
\end{thm}
This was stated originally in \cite{Hochschild} in a different setting. Kontsevich gives a proof for smooth manifolds, in \cite{Kontsevich}, but the proof is quite abstract; the relation to the original theorem is not even completely clear. Here we will only prove the HKR theorem at degrees zero and one.
\begin{prop}
\begin{equation*}
T_{poly}^0(M) \cong H^0(D_{poly}(M)).
\end{equation*}
\end{prop}
\begin{proof}
\begin{itemize}
\item For a morphism of complexes, any $B \in T_{poly}^0(M)$ must be closed in $D_{poly}^0(M)$. Take $f,g$ arbitrary in $C^\infty(M)$. We have the cocycle condition:
\begin{equation*}
\textnormal{d}(B)(f,g)=-B(fg)+fB(g)+B(f)g=0.
\end{equation*}
This is just the statement that $B$ is a derivation, so we are done as $B$ is a vector field.
\item In this situation, injectivity is immediate. Surjectivity requires that every Hochschild cocycle is a vector field, but this follows by the argument above, noting that a Hochschild cocycle is necessarily a derivation.
\end{itemize}
\end{proof}
\begin{prop} \label{prop5}
\begin{equation*}
T_{poly}^1(M)\cong H^1(D_{poly}(M)).
\end{equation*}
\end{prop}
Using Lemma \ref{lem1} and Proposition \ref{prop1} this is already nearly proven: we just fill in the remaining details ourselves here.
\begin{proof}
\begin{itemize}
\item For a morphism of complexes, any $B \in T_{poly}^1(M)$ must be closed in $D_{poly}^1(M)$. Take $f,g,h$ arbitrary in $C^{\infty}(M)$. We have the cocycle condition:
\begin{equation*}
\text{d}(B)(f,g,h)=fB(g,h)-B(fg,h)+B(f,gh)-B(f,g)h=0.
\end{equation*}
It is easy and sufficient to check this for $B=X\otimes Y- Y \otimes X$, $X$ and $Y$ arbitrary vector fields, by direct computation.
\item For injectivity, any $B_1,B_2 \in T_{poly}^1(M)$, such that $B_1 \neq B_2$, must not be in the same cohomology class. This is just Lemma \ref{lem1}.
\item For surjectivity, every Hochschild cocycle must be cohomologous to a bivector field. Proposition \ref{prop1} together with the gauge transformation in terms of d, (\ref{ref41}), gives us a cohomologous skew-symmetric Hochschild cocycle.
\end{itemize}
\ \\
If every skew-symmetric Hochschild cocycle $B$ is a bivector field, the proof is complete. It is sufficient to check that $B$ is a derivation in one variable. We want to show that:
\begin{equation*}
B(f,gh)=gB(f,h)+B(f,g)h, \; \; \forall f,g,h.
\end{equation*}
Recall the cocycle condition:
\begin{equation} \label{ref39}
fB(g,h)+B(f,gh)=B(fg,h)+B(f,g)h.
\end{equation}
Cyclically permuting $f,g,h$ then using skew-symmetry, we also have:
\begin{equation} \label{ref40}
gB(h,f)+B(g,hf)=B(g,h)f-B(f,gh).
\end{equation}
Subtracting (\ref{ref40}) from (\ref{ref39}) then gives the relation:
\begin{equation} \label{ref42}
2B(f,gh)=B(fg,h)+B(f,g)h-gB(h,f)-B(g,hf).
\end{equation}
Reapplying the cocycle condition implies that:
\begin{equation*}
B(fg,h)-B(g,fh)=gB(f,h)+B(f,g)h.
\end{equation*}
Substituting this into (\ref{ref42}) and using skew-symmetry, we are done.
\end{proof}
We have constructed a map from the DGLA controlling deformations of the zero Poisson structure to the DGLA controlling deformations of function multiplication. However, this is not the full story. We have just seen that formal Poisson structures and star products correspond to the Maurer-Cartan sets of $T_{poly}^1(M)[\![\hbar]\!]$ and $D_{poly}^1(M)[\![\hbar]\!]$: we therefore need a mapping that respects the DGLA structures.
The HKR theorem only gives us a quasi-isomorphism of complexes. Because the multiplicative structure associated to $[.,.]_G$ is not mapped to $[.,.]_S$, it does not give a bijection of Maurer-Cartan sets. Ideally we would want a DGLA quasi-isomorphism, but a weaker map is enough: the $L_{\infty}$-quasi-morphism. Although the HKR theorem was insufficient, we will soon see that it was a necessary first step.
\subsection{L-infinity morphisms}
Recall the notion of a DGLA quasi-isomorphism $\phi_1:V_1 \rightarrow V_2$, given in Definition \ref{ref19}. The map acts from $V_1$ to $V_2$; we do not require that it has an inverse.
Nevertheless, we may believe that an equivalence between $V_1$ and $V_2$ of the form:
\begin{equation*}
V_1 \rightarrow V_3 \leftarrow V_4 \rightarrow ... \leftarrow V_n \leftarrow V_2,
\end{equation*}
where the arrows denote quasi-isomorphisms, should still have meaning. In fact, if $V_1$ and $V_2$ are related in the above way, there exists $V_3$ such that:
\begin{equation*}
V_1 \leftarrow V_3 \rightarrow V_2,
\end{equation*}
(see \cite{Keller}). If $V_1$ and $V_2$ satisfy this condition, we say that they are \emph{homotopy equivalent}. We will soon see that the condition of homotopy equivalence is closely related to our new type of mapping, the $L_{\infty}$-morphism. These are `weaker' versions of DGLA morphisms, where we relax the necessary conditions.
\begin{defn}[$\mathbf{L_{\infty}}$\textbf{-morphism}]
An $L_{\infty}$-morphism of DGLAs, $f: V_1\rightarrow V_2$ is a sequence of maps:
\begin{equation*}
f_n: {V_1}^{\otimes n} \rightarrow V_2, \; \;n\geq 1,
\end{equation*}
with each map homogeneous of degree $1-n$, satisfying the following conditions on arbitrary homogeneous elements $x_1,...,x_n$:
\begin{enumerate}
\item The $f_n$ are graded anti-symmetric:
\begin{equation*}
f_n(x_1\otimes...\otimes x_i \otimes x_{i+1}\otimes ...\otimes x_n)= -(-1)^{\overline x_i \overline x_{i+1} }f_n(x_1\otimes...\otimes x_{i+1} \otimes x_i\otimes...\otimes x_n).
\end{equation*}
\item For $n\geq1$:
\begin{multline} \label{ref20}
\textnormal{d}f_n(x_1 \wedge x_2 \wedge ... \wedge x_n)-\sum_{i=1}^{n} \pm f_n(x_1 \wedge ... \wedge \textnormal{d}x_i \wedge ... \wedge x_n)\\
=\frac{1}{2}\sum_{k,l\geq 1, k+l=n}\frac{1}{k!l!}\sum_{\sigma \in S_n} \pm [f_k(x_{\sigma_1}\wedge...\wedge x_{\sigma_k}),f_l(x_{\sigma_{k+1}},\wedge...\wedge x_{\sigma_n})]+
\sum_{i<j} \pm f_{n-1}([x_i,x_j]\wedge x_1\wedge...\wedge x_n).
\end{multline}
\end{enumerate}
\end{defn}
Here the notation of the wedge product is used because the value on the tensor product only depends on its equivalence class in the graded wedge product. Homogeneous of degree $1-n$ means that for homogeneous elements we sum the degrees, then add $1-n$: $f_n$ maps into this grading. For $n=1$, (\ref{ref20}) implies $f_1$ is a \emph{morphism of complexes}: $f_1 \circ d = d \circ f_1$. For $n=2$, it implies that:
\begin{equation*} \label{ref21}
f_1([x_1,x_2])-[f_1(x_1),f_1(x_2)]=\textnormal{d}(f_2(x_1,x_2))+f_2(\textnormal{d}(x_1), x_2)+(-1)^{\overline x_1}f_2(x_1, \textnormal{d}(x_2)),
\end{equation*}
for all homogeneous $x_1$ and $x_2$. The bracket structure is respected by $f_1$, up to the right hand side. In fact, if $x_1$ and $x_2$ are closed, the right hand side is exact and quotiented out by the image of d. So $f_1$ induces a morphism of cohomologies, considered themselves as DGLAs.
$L_{\infty}$-morphisms generalise DGLA morphisms. We can now consider:
\begin{defn}[$L_{\infty}$-quasi-isomorphism]
An $L_{\infty}$-quasi-isomorphism is an $L_{\infty}$-morphism such that the first term, $f_1$, is a quasi-isomorphism of complexes.
\end{defn}
Interestingly, this is the same as homotopy equivalence.
\begin{prop}
The following statements are equivalent:
\begin{enumerate}
\item There exists $V_3$ such that: $V_1 \leftarrow V_3 \rightarrow V_2$.
\item There exists an $L_{\infty}$-quasi-isomorphism from $V_1 \rightarrow V_2$.
\end{enumerate}
\end{prop}
For more discussion of this, see \cite{Keller}. At the cost of more theoretical machinery, we have a convenient single map, rather than the messy setup of multiple DGLA morphisms in different directions.
\subsection{The formality theorem}
Before giving the main result, we state:
\begin{thm} \label{ref26}
Given an $L_{\infty}$-quasi isomorphism $f:T_{poly}(M)[\![\hbar]\!] \rightarrow D_{poly}(M)[\![\hbar]\!]$, the map:
\begin{equation*} \label{ref28}
x \rightarrow \sum_{n=1}^{\infty} \frac{1}{n!} f_n(x,...,x),
\end{equation*}
induces a bijection:
\begin{equation*}
\overline{MC}(T_{poly}(M)[\![\hbar]\!]) \simeq \overline{MC}(D_{poly}(M)[\![\hbar]\!]).
\end{equation*}
\end{thm}
This is in fact true for any two DGLAs, with a bijection induced between the two corresponding quotient sets previously alluded to. With this setup, we have everything necessary to understand Kontsevich's formality theorem.
\begin{thm}[Kontsevich, \cite{Kontsevich}]
There exists an $L_{\infty}$-quasi-isomorphism:
\begin{equation*}
U: T_{poly}(M)[\![\hbar]\!] \rightarrow D_{poly}(M)[\![\hbar]\!],
\end{equation*}
with the first component $U_1$ given by the HKR quasi-isomorphism of complexes of Theorem \ref{ref25}.
\end{thm}
The proof of these results go beyond the scope of this essay. Nevertheless, we have established sufficient machinery for the reader to understand that together they give us the natural identification of Theorem \ref{ref27}. In particular, we have a deformation quantization of any given Poisson manifold, what we wanted all along. Note that if we want to quantize an ordinary Poisson structure $\alpha$, we input $\frac{\hbar \alpha}{2}$, again so that the star product's commutator has the right form at order $\hbar$.
The HKR map was indeed the first step in constructing the $L_{\infty}$-quasi-isomorphism. Moreover, this bijection is a deformation quantization, because $f_1$, the HKR map, sends $\hbar \alpha$ to itself, and all other components map into $\hbar$ order two and higher. In the final chapter, we will return to these ideas when we discuss the broader mathematical connections. One can formulate a notion of homotopy equivalence between the so-called \emph{stable} $L_{\infty}$-quasi-isomorphisms: then the real Grothendieck-Teichm{\"u}ller group $GT(\mathbb{R})$ acts on equivalence classes \cite{Dolgushev}, \cite{Dolgushev1}.
\section{Kontsevich's formula}
Armed with Kontsevich's main theoretical result, we proceed to something more concrete. The $L_{\infty}$-quasi-isomorphism gives us a simple, explicit formula for finding a star product given a Poisson structure on a manifold that is an open subset of $\mathbb{R}^d$. This section's background material is drawn largely from \cite{Cattaneo}, \cite{Esposito} and \cite{Kontsevich}; the later discussion is based on \cite{Panzer}.
\subsection{Admissible graphs}
We are going to write our star product formula in terms of a family of \emph{admissible graphs}. These can be defined as follows.
\begin{defn}[Oriented graphs]
An oriented graph $\Gamma$ is a pair $(V_{\Gamma}, E_{\Gamma},)$ such that $E_{\Gamma}$ is a subset of $V_{\Gamma} \times V_{\Gamma}$.
\end{defn}
$V_{\Gamma}$ are the vertices of $\Gamma$ and $E_{\Gamma}$ are the edges of $\Gamma$; the edge $e=(v_1,v_2)$ starts at $v_1$ and ends at $v_2$. Formally, a labelled, oriented graph $\Gamma$ is admissible if:
\begin{enumerate}
\item $\Gamma$ has $n+2$ vertices, $\{1,...,n,L,R\}$, where $L,R$ are just symbols.
\item $\Gamma$ has $2n$ edges.
\item The edges are labelled by $e_1^1,e_1^2,...,e_n^1,e_n^2$; the edges $e_k^1$, $e_k^2$ start at the vertex $k$.
\item The edges $e_k^1=(v_k,v_1)$ and $e_k^2=(v_k,v_2)$ are such that $v_1 \neq v_2$.
\item There is no vertex $v$ such that the $(v,v) \in E_{\Gamma}$.
\end{enumerate}
We are not allowing infinite graphs or multiple edges. To restate the rules simply, two edges go out from the first $n$ vertices $\{1,...,n\}$; these edges must end at different vertices from each other. The vertices $\{L,R\}$ have no edges going out. Finally, we cannot have loops. Let $G$ be the set of admissible graphs and $G_n$ be the set of admissible graphs with $n+2$ vertices. If $\Gamma \in G_n$, we say $|\Gamma|=n$; note that $G_n$ contains $(n(n+1))^n$ elements. Now Kontsevich's formula is a sum over the graphs.
\begin{thm}[Kontsevich, \cite{Kontsevich}]
Given $\pi$, a Poisson structure on an open subset of $\mathbb{R}^d$, the formula:
\begin{equation} \label{ref9}
f \star g := \sum_{\Gamma \in G} \frac{{\hbar}^n}{n!} \omega_{\Gamma} B_{\Gamma, \pi}(f,g),
\end{equation}
defines an associative product. If we change coordinates, we obtain a gauge-equivalent star product.
\end{thm}
Collecting terms, the operator at order $h^n$ is given by:
\begin{equation*}
B_n(f,g) = \frac{1}{n!} \sum_{\Gamma \in G_n} {\omega_{\Gamma}} B_{\Gamma, \pi}(f,g).
\end{equation*}
Of course we need to explain the terms in the sum. Given a Poisson structure $\pi$, we assign a bidifferential operator $B_{\Gamma, \pi}$ to each graph $\Gamma$, depending on both $\Gamma$ and $\pi$. In front of this there is a weight, $\omega_{\Gamma} \in \mathbb{R}$, depending only on $\Gamma$. The star product of $f$ and $g$ is then given by the formal power series (\ref{ref9}).
\begin{figure}
\includegraphics[width=12cm]{Images/graph3.pdf}
\centering
\caption{Unlabelled admissible graphs for (i) $n=1$ and (ii) $n=2$.}
\centering
\label{fig:2}
\end{figure}
\subsection{Bidifferential operators}
Given a graph and Poisson bivector field, $\{f,g\}=\pi^{ij}\partial_i f \partial_j g$, how do we calculate $B_{\Gamma, \pi}(f,g)$? The recipe is as follows.
\begin{enumerate}
\item Use a map $I:E_{\Gamma} \rightarrow \{1,...,d\}: (e_1^1,e_1^2,...,e_n^1,e_n^2) \rightarrow (i_1,...,i_d)$, changing the labels of the edges.
\item Assign $f$ to $L$ and $g$ to $R$. Assign the tensor $\pi^{I(e^1_k) I(e^2_k)}$ to the vertex $k$.
\item For every edge labelled $i_j$, apply a partial derivative with respect to $i_j$ to the function or tensor associated with the vertex endpoint.
\item Following this setup, we can define a general formula, summing over maps $I$:
\end{enumerate}
\begin{equation*}
B_{\Gamma, \pi} := \sum_{I:E_{\Gamma} \rightarrow \{1,...,d\}} \left[ \prod_{k=1}^{n} \left( \prod_{e\in E_{\Gamma}, e=(*,k)} \partial_{I(e)} \right) \pi^{I(e_k^1) I(e_k^2)}\right] \times \left( \prod_{e\in E_{\Gamma}, e=(*,L)} \partial_{I(e)} \right)f\left( \prod_{e\in E_{\Gamma}, e=(*,R)} \partial_{I(e)} \right)g.
\end{equation*}
In summary, each vertex is assigned a tensor or function, and edges define partial derivatives on their endpoints. Multiplying these together, we then sum over the edge's labels. To illustrate the rule, consider the graph in Figure \ref{fig:1}. The corresponding operator is:
\begin{equation*}
B_{\Gamma, \pi}(f,g)=\sum_{1 \leq i_1...i_6 \leq d} \pi^{i_5 i_6} \partial_{i_5}\partial_{i_2}(\pi^{i_3 i_4})\partial_{i_3}(\pi^{i_1 i_2})
\partial_{i_1} \partial_{i_4}(f) \partial_{i_6} (g).
\end{equation*}
\begin{figure}
\includegraphics[width=6cm]{Images/graph1.pdf}
\centering
\caption{An admissible graph of order 3.}
\label{fig:1}
\end{figure}
\subsection{Weight integrals}
For the weights $\omega_\Gamma$, as with the formula, we first give the definition, then explain the meaning of the terms. The weights $\omega_\Gamma$ are defined as integrals of differential $2n$-forms:
\begin{equation} \label{ref10}
\omega_{\Gamma}:= \int_{{\mathcal{H}}_n} \text{d}\phi_{e_1} \wedge ... \wedge \text{d}\phi_{e_{2n}}.
\end{equation}
These coefficients depend only on the graphs; they need to be calculated once and for all to quantize any Poisson structure. We use the following recipe:
\begin{enumerate}
\item Let $\mathcal{H}$ denote the upper half complex plane, $\text{Im}(z)>0$.
\item Given $p,q \in \mathcal{H}$, define the function:
\begin{equation*}
\phi(p_1,p_2):=\frac{1}{2 \pi} \text{arg}\left(\frac{p_2-p_1}{p_2-\overline p_1}\right).
\end{equation*}
This function is best thought of as an angle, normalised by $2 \pi$. It can also be interpreted geometrically in terms of geodesics of $\mathcal{H}$ endowed with the Lobachevsky metric. This function can be extended continuously to the set the set of points $(p,q)$ with $Im(q) \geq 0, \; p \neq q$. For a fuller discussion of this see \cite{Esposito}. \item Lets $\mathcal{H}_n$ be the set of pairwise distinct $n$-tuples of points in $\mathcal{H}$:
\begin{equation*}
\mathcal{H}_n:=\{(p_1,...,p_n)|p_i \in \mathcal{H}, \; p_i \neq p_j \; \text{if} \; i \neq j\}.
\end{equation*}
This called the \emph{space of configurations}. It gets an orientation coming from the natural complex structure.
\item Given a graph $\Gamma \in G_n$ and a configuration $(p_1,...,p_n)$, draw $\Gamma$ on $\mathbb{R}^2$ by assigning $p_k \in \mathcal{H}$ to the vertex $k$, $0$ to the vertex $L$, and $1$ to the vertex $R$.
\item Each edge $e_k$ now defines an ordered pair of points $(p_i,p_j)$ and an angle $\phi_{e_k}:=\phi(p_i,p_j)$.
\end{enumerate}
As the points $p_i$ move around $\mathcal{H}$, the $\phi_{e_i}$ define functions on $\mathcal{H}_n$ via projection. Then the integral in (\ref{ref10}) gives us the weight $\omega_{\Gamma}$; it is shown in \cite{Kontsevich} that it converges absolutely. Everything in (\ref{ref9}) is now defined.
We now consider some simple weight integrals and star products. If $e_k$ is the edge from vertex $p_i$ to vertex $p_j$, we can write $\phi_{e_k}$ alternatively as $\phi_{p_i\rightarrow p_j}$. We will first calculate the most basic wedge integral, corresponding to a graph with three vertices and two edges in $\mathcal{H}$. We refer to this setup as an `out-out' vertex.
\begin{lemma}[General wedge integral] \label{lem2}
We have that:
\begin{equation} \label{ref11}
\int_{x \in \mathcal{H}}\textnormal{d}\phi_{x \rightarrow y} \wedge \textnormal{d}\phi_{x \rightarrow z} = \frac{1}{2 \pi i} \textnormal{log} \left( \frac{y-\bar{z}}{z-\bar{y}} \right).
\end{equation}
\end{lemma}
This lemma is stated in \cite{Panzer}, but the proof there is more complicated. Drawing heavy inspiration from the work of our supervisor, we have performed the calculation originally in a simpler way. A full proof is given in Appendix \ref{amend14}.
\begin{proof}
\emph{(Sketch)}. Apply the following algorithm: \begin{enumerate}
\item Separate the integrand differential form into holomorphic and anti-holomorphic factors.
\item Add on any antiholomorphic differentials to the holomorphic factor, to enable a single-valued primitive to be found.
\item Apply Stokes' theorem to our single-valued integrand.
\item Express the remaining contributions from Stokes' in terms of single valued functions depending on the remaining points.
\item Iterate this as many times as necessary.
\end{enumerate}
\end{proof}
Note we are not yet enforcing that two of the vertices lie on the real line, as in Kontsevich's actual construction. One can show the continuity of the integral (\ref{ref11}), so that in the limit as $y \rightarrow 0$, $z \rightarrow 1$, it implies that:
\begin{equation*}
\int_{x \in \mathcal{H}}\textnormal{d}\phi_{x \rightarrow 0} \wedge \textnormal{d}\phi_{x \rightarrow 1} = \frac{1}{2}.
\end{equation*}
In $G_1$ there are two graphs that differ by swapping the edges; they are exactly the `out-out' vertex graphs. Thus $\omega_1=\frac{1}{2}$ and by skew-symmetry of the differential form, $\omega_2=-\frac{1}{2}$. So $f \star g$ has contribution at order $\hbar$ given by:
\begin{equation*}
\frac{\hbar}{2}{\pi}^{ij}({\partial}_i f {\partial}_j g -{\partial_j} f {\partial_i} g) = \hbar {\pi}^{ij}\partial_i f \partial_j g,
\end{equation*}
and if we input $\frac{1}{2}\{.,.\}$, we get an induced $B_1^-$ bracket of $\frac{1}{2}\{.,.\}$, exactly what was stated in Definition \ref{amend13}.
Having performed this integral, we can now return to the Moyal product.
\begin{prop}
Applying the Kontsevich formula (\ref{ref9}) to a constant Poisson structure gives us the Moyal product (\ref{ref13}).
\end{prop}
\begin{proof}
\begin{itemize}
\item Consider the constant Poisson structure with ${\pi}^{ij} \in \mathbb{R}$. Any graph with an edge ending in a vertex other than $L$ or $R$ gives contribution $0$, as there is a term ${\partial}_i {\pi}^{jk}=0$.
\begin{figure}
\includegraphics[width=4.5cm]{Images/graph2.pdf}
\centering
\caption{A graph of order $n$ with non-zero contribution to the Moyal product.}
\centering
\label{fig:3}
\end{figure}
\item The only non-zero contribution graphs have vertices with all edges going into $L$ or $R$. There are $2^n$ such graphs in $G_n$, differing by swapping pairs of edges from each vertex.
\item Swapping the edges does not change the contribution to the star product, as we saw for ${\Gamma_1}, {\Gamma_2}$, because ${\pi}^{ij}=-{\pi}^{ji}$.
\end{itemize}
\ \\
Let $\Gamma$ be the graph of this form such that the first labelled edge from each vertex goes into L. Because the contribution from all graphs is the same, at order $n$ we have:
\begin{equation*}
B_n(f,g) = \frac{2^n}{n!} {\omega_{\Gamma}} B_{\Gamma,\pi}(f,g) = \frac{2^n}{n!} {\omega}_{\Gamma} \left( \prod_{k=1}^{n} {\pi}^{i_k j_k} \right) \left( \prod_{k=1}^{n} \partial_{i_k} f \right) \left( \prod_{k=1}^{n} \partial_{j_k} g \right).
\end{equation*}
For $\omega_{\Gamma}$, the differential $2n$-form decomposes trivially, so that:
\begin{equation*}
\int_{\mathcal{H}^n} \text{d} \phi_{e_1} \wedge...\wedge \text{d} \phi_{e_n} = \left( \int_{x \in \mathcal{H}} \phi_{x \rightarrow 0} \wedge \phi_{x \rightarrow 1} \right)^n = \frac{1}{2^n}.
\end{equation*}
Putting this together, the $2^n$ in the numerator and denominator cancel so that:
\begin{equation*}
f \star g = \sum_{n} \frac{{\hbar}^n}{n!} \left( \prod_{k=1}^{n} {\pi}^{i_k j_k} \right) \left( \prod_{k=1}^{n} \partial_{i_k} f \right) \left( \prod_{k=1}^{n} \partial_{j_k} g \right),
\end{equation*}
exactly what we want, i.e. equation (\ref{ref13}).
\end{proof}
We can see that breaking up graphs into smaller pieces is a good technique for dealing with more complicated graphs. If we first enforce $y=0,z=1$, the integral (\ref{ref11}) is considerably simpler: see, for example, the proof given in \cite{Esposito}, which uses an easy trick using the range of integration. However, we have presented a more involved calculation for general $y$ and $z$, because the algorithm generalises when one considers more involved weight integral calculations.
We are always using the principle branch of $\text{log}(z)$ and our method means that the argument is never allowed to be negative. As an example, the argument in our answer, $\frac{y-\bar{z}}{z-\bar{y}}$, is never in $\mathbb{R}^-$. If one considers larger graphs, where further integration is be needed (for example of $y$ and $z$), the use of single-valued functions throughout means that one does not need to carry data about choices of log branches forward in the calculation. In the next chapter, we discuss the general algorithm of \cite{Panzer} for computing weight integrals. It is just the same as one presented above, generalised to a more sophisticated theoretical setting; running through Appendix \ref{amend14} therefore gives the reader some insight into how the results are achieved.
\section{Multiple zeta values}
To conclude, we explore the weight integrals in further depth, focusing in particular on the recent work \cite{Panzer}, of Erik Panzer, Peter Banks and Brent Pym. We also consider the connections of the subject to other areas of mathematics. This section draws on \cite{Zagier}, \cite{Waldschmidt}, \cite{MZVs} for background material.
\subsection{Periods}
We traditionally think of our number systems as falling in the hierarchy:
\begin{align*}
\mathbb{N} \subset \mathbb{Z} \subset \mathbb{Q} \subset \overline{\mathbb{Q}}\\ \cap \: \quad \cap \\
\mathbb{R} \subset \mathbb{C}
\end{align*}
(Taken from \cite{Zagier}). There is, however, another class of numbers between $\overline{\mathbb{Q}}$ and $\mathbb{C}$, known as the \emph{periods}. Intuitively, these can be thought of as complex numbers expressible with a finite amount of algebraic information; they are a natural extension of $\overline{\mathbb{Q}}$. More precisely, we can state:
\begin{defn}[Period]
A period is a complex number, such that the both the real and imaginary parts are absolutely convergent integrals of rational functions with rational coefficients, over domains in $\mathbb{R}^n$ given by rational coefficient polynomial inequalities.
\end{defn}
A particular class of periods, the \emph{multiple zeta values}, or MZVs, is particularly relevant for the present discussion. Recall the standard definition of the \emph{Riemann zeta function} for $s\in \mathbb{C}$:
\begin{equation*} \label{ref43}
\zeta(s):=\sum_{n=1}^{\infty}n^{-s}.
\end{equation*}
This leads to:
\begin{defn}[Multiple zeta values]
Generalising the zeta values we have:
\begin{equation*}
\zeta(s_1,...,s_k) := \sum_{0<n_1<...<n_k} \frac{1}{{n_1}^{s_1}...{n_k}^{s_k}},
\end{equation*}
for $s:=\{s_1,...,s_k\},s_i \in \mathbb{Z}^+$, $s_k>1$. All MZVs are convergent.
\end{defn}
For a given MZV, we say that $|s|=s_1+...+s_k$ is the weight. There are many interesting relations between the MZVs; for a broader discussion see \cite{MZVs}.
\subsection{MZVs and weight integrals}
Consider the admissible graphs of Figure \ref{fig:2}. We have already calculated the weight of $(i)$ as $\frac{1}{2}$; considering the graphs in $(ii)$ from left to right, one can calculate the weights as: $\frac{1}{4}, \frac{1}{12}, -\frac{1}{12}, \frac{1}{24}$ respectively (The first is just a Moyal graph). We may think we see a simple pattern, but this idea is quickly dispelled by considering $\Gamma$ in Figure \ref{fig:4}. The corresponding weight is:
\begin{equation} \label{ref44}
\omega_{\Gamma}=-\frac{1}{6048}+\frac{9}{128}\frac{{\zeta(3)}^2}{{\pi}^6}.
\end{equation}
\begin{figure}
\includegraphics[width=6cm]{Images/finalnew1.pdf}
\centering
\caption{An admissible graph of order 6.}
\label{fig:4}
\end{figure}
See \cite{Website}. The MZV $\zeta(3)$ has emerged from a weight integral calculation. In fact, Kontsevich conjectured himself that the weights are given by rational linear combinations of MZVs. This was recently proven in \cite{Panzer}:
\begin{thm}[Banks, Panzer, Pym, \cite{Panzer}]
Given a star product graph $\Gamma$ at order $\hbar^n$, the weight $\omega_{\Gamma}$ can be expressed as:
\begin{equation*}
\omega_{\Gamma}=\sum_{|s|\leq n}\frac{\zeta(s)}{(i \pi)^{|s|}}\mathbb{Q}.
\end{equation*}
\end{thm}
Because $\omega_{\Gamma} \in \mathbb{R}$, the $i$ in the denominator implies that some of the rational coefficients must be zero. In fact, the authors of \cite{Panzer} go further, as the constructive proof gives a general algorithm for computing any $\omega_{\Gamma}$.
\subsection{Polylogarithms}
To understand some ideas behind the proof, we introduce:
\begin{defn}[Polylogarithm] Define the $n^{th}$ polylogarithm function, $n \geq 1$ as the Chen iterated integral:
\begin{equation*}
Li_{n}(x) := \int_{0}^{x} \frac{\textnormal{d}z}{1-z} \left( \frac{\textnormal{d}z}{z}\right)^{n-1}.
\end{equation*}
This depends on a choice of path, but only up to homotopy equivalence: it is therefore a multi-valued function like a logarithm. By choosing the branch cut $(-\infty,0] \cup [1,\infty)$, we get a single-valued function. This is an improper integral, but by choosing a suitable limit we can find the value for paths with endpoints $0$ and $1$.
\end{defn}
There are differing conventions for ordering the differential forms in the integrand: we assume that the leftmost integral is performed first. This now leads to:
\begin{lemma} \label{lem5} Polylogarithms can be related to the Riemann zeta function with the identity:
\begin{equation*}
Li_n(1)=\zeta(n).
\end{equation*}
\end{lemma}
\begin{proof}
By choosing the branch cut $(-1,0]$, we have a simply-connected domain $D=\{|x|<1\}\setminus(-1,0]$. So all paths in $D$ are homotopy equivalent and $Li_n(x)$ is independent of the path and single-valued. Now:
\begin{equation*}
\int_{0}^{x} \frac{\textnormal{d}z}{1-z} \left( \frac{\textnormal{d}z}{z}\right)^{n-1}=\int_{0}^{x} -\textnormal{log}(1-z) \left( \frac{\textnormal{d}z}{z}\right)^{n-1}.
\end{equation*}
Using the convergent Taylor series for $\textnormal{log}(1-z)$ in $|z|<1$, this is equal to:
\begin{equation*}
\int_{0}^{x}\left( z+ \frac{z^2}{2}+\frac{z^3}{3}+...\right) \left( \frac{\textnormal{d}z}{z}\right)^{n-1} =\int_{0}^{x}\left( z+ \frac{z^2}{2^2}+\frac{z^3}{3^2}+...\right) \left( \frac{\textnormal{d}z}{z}\right)^{n-2}
=...=\sum_{i=1}^{\infty}\frac{x^i}{i^n}.
\end{equation*}
Taking the limit as $x\rightarrow 1$ we obtain $\zeta(n)$ as desired.
\end{proof}
Now, we can easily generalise this construction to:
\begin{defn}[Multiple polylogarithm] The single-variable multiple polylogarithm is given by the Chen iterated integral:
\begin{equation*}
Li_{(s_1,...,s_n)}(x):=\int_{0}^{x}\frac{\textnormal{d}z}{1-z}\left(\frac{\textnormal{d}z}{z} \right)^{s_1-1}\frac{\textnormal{d}z}{1-z}...\frac{\textnormal{d}z}{1-z}\left(\frac{\textnormal{d}z}{z} \right)^{s_n-1},
\end{equation*}
for $s_1,...,s_n \in \mathbb{Z^+}$. This is a multi-valued function on $\mathbb{C} \setminus \{0,1\}$.
\end{defn}
The relevance is immediate from:
\begin{lemma}
The MZVs can be expressed with the identity:
\begin{equation*} \label{lem4}
Li_{(s_1,...,s_n)}(1)= \sum_{n_1>...>n_k>0} \frac{1}{{n_1}^{s_1}...{n_k}^{s_k}}.
\end{equation*}
\end{lemma}
\begin{proof} \emph{(Sketch)}. The idea is exactly the same as the proof of Lemma \ref{lem5}, expanding the integral using Taylor series and integrating term-by-term. See \cite{MZVs}
for full details.
\end{proof}
For a broader discussion of these functions, and their relations to MZVs, see \cite{Waldschmidt}. We note a final interesting fact about MZVs. There is a long-standing `folklore conjecture' that the odd integer values of the zeta function:
\begin{equation*}
\pi, \; \zeta(3), \; \zeta(5), \; \zeta(7),...
\end{equation*}
are algebraically independent \cite{Brown}. This suggests that certain Kontsevich weights, such as $\omega_{\Gamma}$ in (\ref{ref44}), are, in fact, transcendental.
\subsection{A constructive proof}
Recall the sketch proof of Lemma \ref{lem2}; we just used the algorithm of \cite{Panzer}. However, the authors generalise the ideas we presented there by finding real-analytic, single-valued primitives in terms of multiple polylogarithms rather than ordinary logs. Integrating out variables and applying (regularized) Stokes' theorem \cite{Panzer}, MZVs are eventually obtained exactly through Lemma \ref{lem4}.
Importantly, we can now theoretically give a deformation quantization of a Poisson structure up to any desired $\hbar$ order. In fact, the authors of \cite{Panzer} have created open source software for this, available at \cite{Website}, where there is also a full list of weights up to ${\hbar}^6$. Prior work used various tricks for calculating weight integrals in special cases. The \emph{Bernoulli graphs} were treated previously in \cite{Kathotia}, where the coefficients were found to be easily expressible in terms of Bernoulli numbers. The software was checked for consistency against these calculations (and others like it), producing the same results.
The significance of \cite{Panzer} is clear when one considers the large amount of previous work (see also \cite{merkulov}, \cite{Shoikhet}, \cite{VandenBergh}) that has focused on these calculations. Bringing this all together, the authors have, in a sense, `closed' a long-standing area of research.
\subsection{Final remarks}
\begin{figure}
\includegraphics[width=8.5cm]{Images/ideas.pdf}
\centering
\caption{Diagrammatic representation of the key ideas, figures and their relations.}
\label{fig5}
\centering
\end{figure}
We began this essay with a simple, physically motivated aim: the quantization of classical mechanics. We saw that an early historical guess for this was given by Dirac, who suggested the quantization condition:
\begin{equation*}Q(\{f,g\})=\frac{1}{i\hbar}[Q(f),Q(g)].
\end{equation*}
After introducing the idea of deformation quantization, there was finally a mysterious connection of this physics problem to MZVs, typically associated with number theory. This is not the only strange link. Recall that the Kontsevich formality theorem gives us an $L_{\infty}$-quasi-isomorphism:
\begin{equation*}
U: T_{poly}(M)\rightarrow D_{poly}(M).
\end{equation*}
Consider the space of stable $L_{\infty}$-quasi-isomorphisms from $T_{poly}(M)$ to $D_{poly}(M)$ (see \cite{Dolgushev1}), denoted by:
\begin{equation*}\text{Maps}^{L_{\infty}}_{q.i.}(T_{poly}(M), D_{poly}(M)).\end{equation*} We can introduce a notion of homotopy equivalence between the maps in this set; then equivalent maps induce the same bijection of Maurer-Cartan sets, in the sense of Theorem \ref{ref26} (see \cite{Dolgushev}).
Remarkably, the \emph{real Grothendieck-Teichm{\"u}ller group} $GT(\mathbb{R})$ now acts on the set of homotopy equivalence classes, $\pi_0 \text{Maps}^{L_{\infty}}_{q.i.}(T_{poly}(M), D_{poly}(M))$ (see \cite{Dolgushev}, \cite{Dolgushev1}, \cite{Grothendieck}). We know that there is a homomorphism from the \emph{absolute Galois group of} $\mathbb{Q}$, Gal$(\overline{\mathbb{Q}}/\mathbb{Q})$, to the \emph{profinite} $\widehat{GT}$ \cite{Drinfeld}, and therefore to the $p$-adic points of the \emph{algebraic} $GT$:
\begin{equation*}
\text{Gal}(\overline{\mathbb{Q}} / \mathbb{Q}) \rightarrowtail \widehat{GT} \rightarrow GT(\mathbb{Q}_p).
\end{equation*}
So Gal$(\overline{\mathbb{Q}}/\mathbb{Q})$ would act on the space of $p$-adic deformation quantizations, if such a notion made sense, another puzzling connection with number theory.
In fact, $GT(\mathbb{R})$ also acts on another construction, the space of \emph{Drinfeld associators}, Drin, which are formal noncommututative power series in two variables satisfying a complex set of algebraic relations. Drinfeld gives a specific associator, $KZ$, with MZVs as coefficients \cite{Drinfeld}; it is hard to escape the feeling that none of this is a coincidence:
\begin{equation*}
KZ \in \text{Drin} \; \reflectbox{$\lcirclearrowright$} \; GT \lcirclearrowright \pi_0 \text{Maps}^{L_{\infty}}_{q.i.}(T_{poly}(M), D_{poly}(M)) \; \reflectbox{$\in$} \; \text{$L_{\infty}$ Kont}.
\end{equation*}
In the introduction, we stated that deformation quantization has ambiguous physical applicability, because there is no Hilbert space to act on and convergence is ignored. Whilst this criticism may be fair in some sense, the strange and beautiful links of the subject to many disparate areas of mathematics and physics (see Figure \ref{fig5}) lead us to believe that it is not the end of the story. To conclude, we pose three (possibly very naive) questions.
\begin{enumerate}
\item How can deformation quantization be used for Hilbert space quantization?
\item What are the underlying relations between deformation quantization and other mathematical structures where MZVs appear?
\item Are arithmetic geometry and physics related on some fundamental level?
\end{enumerate}
|
\section{Introduction}
This document is a model and instructions for \LaTeX.
Please observe the conference page limits.
\section{Ease of Use}
\subsection{Maintaining the Integrity of the Specifications}
The IEEEtran class file is used to format your paper and style the text. All margins,
column widths, line spaces, and text fonts are prescribed; please do not
alter them. You may note peculiarities. For example, the head margin
measures proportionately more than is customary. This measurement
and others are deliberate, using specifications that anticipate your paper
as one part of the entire proceedings, and not as an independent document.
Please do not revise any of the current designations.
\section{Prepare Your Paper Before Styling}
Before you begin to format your paper, first write and save the content as a
separate text file. Complete all content and organizational editing before
formatting. Please note sections \ref{AA}--\ref{SCM} below for more information on
proofreading, spelling and grammar.
Keep your text and graphic files separate until after the text has been
formatted and styled. Do not number text heads---{\LaTeX} will do that
for you.
\subsection{Abbreviations and Acronyms}\label{AA}
Define abbreviations and acronyms the first time they are used in the text,
even after they have been defined in the abstract. Abbreviations such as
IEEE, SI, MKS, CGS, ac, dc, and rms do not have to be defined. Do not use
abbreviations in the title or heads unless they are unavoidable.
\subsection{Units}
\begin{itemize}
\item Use either SI (MKS) or CGS as primary units. (SI units are encouraged.) English units may be used as secondary units (in parentheses). An exception would be the use of English units as identifiers in trade, such as ``3.5-inch disk drive''.
\item Avoid combining SI and CGS units, such as current in amperes and magnetic field in oersteds. This often leads to confusion because equations do not balance dimensionally. If you must use mixed units, clearly state the units for each quantity that you use in an equation.
\item Do not mix complete spellings and abbreviations of units: ``Wb/m\textsuperscript{2}'' or ``webers per square meter'', not ``webers/m\textsuperscript{2}''. Spell out units when they appear in text: ``. . . a few henries'', not ``. . . a few H''.
\item Use a zero before decimal points: ``0.25'', not ``.25''. Use ``cm\textsuperscript{3}'', not ``cc''.)
\end{itemize}
\subsection{Equations}
Number equations consecutively. To make your
equations more compact, you may use the solidus (~/~), the exp function, or
appropriate exponents. Italicize Roman symbols for quantities and variables,
but not Greek symbols. Use a long dash rather than a hyphen for a minus
sign. Punctuate equations with commas or periods when they are part of a
sentence, as in:
\begin{equation}
a+b=\gamma\label{eq}
\end{equation}
Be sure that the
symbols in your equation have been defined before or immediately following
the equation. Use ``\eqref{eq}'', not ``Eq.~\eqref{eq}'' or ``equation \eqref{eq}'', except at
the beginning of a sentence: ``Equation \eqref{eq} is . . .''
\subsection{\LaTeX-Specific Advice}
Please use ``soft'' (e.g., \verb|\eqref{Eq}|) cross references instead
of ``hard'' references (e.g., \verb|(1)|). That will make it possible
to combine sections, add equations, or change the order of figures or
citations without having to go through the file line by line.
Please don't use the \verb|{eqnarray}| equation environment. Use
\verb|{align}| or \verb|{IEEEeqnarray}| instead. The \verb|{eqnarray}|
environment leaves unsightly spaces around relation symbols.
Please note that the \verb|{subequations}| environment in {\LaTeX}
will increment the main equation counter even when there are no
equation numbers displayed. If you forget that, you might write an
article in which the equation numbers skip from (17) to (20), causing
the copy editors to wonder if you've discovered a new method of
counting.
{\BibTeX} does not work by magic. It doesn't get the bibliographic
data from thin air but from .bib files. If you use {\BibTeX} to produce a
bibliography you must send the .bib files.
{\LaTeX} can't read your mind. If you assign the same label to a
subsubsection and a table, you might find that Table I has been cross
referenced as Table IV-B3.
{\LaTeX} does not have precognitive abilities. If you put a
\verb|\label| command before the command that updates the counter it's
supposed to be using, the label will pick up the last counter to be
cross referenced instead. In particular, a \verb|\label| command
should not go before the caption of a figure or a table.
Do not use \verb|\nonumber| inside the \verb|{array}| environment. It
will not stop equation numbers inside \verb|{array}| (there won't be
any anyway) and it might stop a wanted equation number in the
surrounding equation.
\subsection{Some Common Mistakes}\label{SCM}
\begin{itemize}
\item The word ``data'' is plural, not singular.
\item The subscript for the permeability of vacuum $\mu_{0}$, and other common scientific constants, is zero with subscript formatting, not a lowercase letter ``o''.
\item In American English, commas, semicolons, periods, question and exclamation marks are located within quotation marks only when a complete thought or name is cited, such as a title or full quotation. When quotation marks are used, instead of a bold or italic typeface, to highlight a word or phrase, punctuation should appear outside of the quotation marks. A parenthetical phrase or statement at the end of a sentence is punctuated outside of the closing parenthesis (like this). (A parenthetical sentence is punctuated within the parentheses.)
\item A graph within a graph is an ``inset'', not an ``insert''. The word alternatively is preferred to the word ``alternately'' (unless you really mean something that alternates).
\item Do not use the word ``essentially'' to mean ``approximately'' or ``effectively''.
\item In your paper title, if the words ``that uses'' can accurately replace the word ``using'', capitalize the ``u''; if not, keep using lower-cased.
\item Be aware of the different meanings of the homophones ``affect'' and ``effect'', ``complement'' and ``compliment'', ``discreet'' and ``discrete'', ``principal'' and ``principle''.
\item Do not confuse ``imply'' and ``infer''.
\item The prefix ``non'' is not a word; it should be joined to the word it modifies, usually without a hyphen.
\item There is no period after the ``et'' in the Latin abbreviation ``et al.''.
\item The abbreviation ``i.e.'' means ``that is'', and the abbreviation ``e.g.'' means ``for example''.
\end{itemize}
An excellent style manual for science writers is \cite{b7}.
\subsection{Authors and Affiliations}
\textbf{The class file is designed for, but not limited to, six authors.} A
minimum of one author is required for all conference articles. Author names
should be listed starting from left to right and then moving down to the
next line. This is the author sequence that will be used in future citations
and by indexing services. Names should not be listed in columns nor group by
affiliation. Please keep your affiliations as succinct as possible (for
example, do not differentiate among departments of the same organization).
\subsection{Identify the Headings}
Headings, or heads, are organizational devices that guide the reader through
your paper. There are two types: component heads and text heads.
Component heads identify the different components of your paper and are not
topically subordinate to each other. Examples include Acknowledgments and
References and, for these, the correct style to use is ``Heading 5''. Use
``figure caption'' for your Figure captions, and ``table head'' for your
table title. Run-in heads, such as ``Abstract'', will require you to apply a
style (in this case, italic) in addition to the style provided by the drop
down menu to differentiate the head from the text.
Text heads organize the topics on a relational, hierarchical basis. For
example, the paper title is the primary text head because all subsequent
material relates and elaborates on this one topic. If there are two or more
sub-topics, the next level head (uppercase Roman numerals) should be used
and, conversely, if there are not at least two sub-topics, then no subheads
should be introduced.
\subsection{Figures and Tables}
\paragraph{Positioning Figures and Tables} Place figures and tables at the top and
bottom of columns. Avoid placing them in the middle of columns. Large
figures and tables may span across both columns. Figure captions should be
below the figures; table heads should appear above the tables. Insert
figures and tables after they are cited in the text. Use the abbreviation
``Fig.~\ref{fig}'', even at the beginning of a sentence.
\begin{table}[htbp]
\caption{Table Type Styles}
\begin{center}
\begin{tabular}{|c|c|c|c|}
\hline
\textbf{Table}&\multicolumn{3}{|c|}{\textbf{Table Column Head}} \\
\cline{2-4}
\textbf{Head} & \textbf{\textit{Table column subhead}}& \textbf{\textit{Subhead}}& \textbf{\textit{Subhead}} \\
\hline
copy& More table copy$^{\mathrm{a}}$& & \\
\hline
\multicolumn{4}{l}{$^{\mathrm{a}}$Sample of a Table footnote.}
\end{tabular}
\label{tab1}
\end{center}
\end{table}
\begin{figure}[htbp]
\centerline{\includegraphics{fig1.png}}
\caption{Example of a figure caption.}
\label{fig}
\end{figure}
Figure Labels: Use 8 point Times New Roman for Figure labels. Use words
rather than symbols or abbreviations when writing Figure axis labels to
avoid confusing the reader. As an example, write the quantity
``Magnetization'', or ``Magnetization, M'', not just ``M''. If including
units in the label, present them within parentheses. Do not label axes only
with units. In the example, write ``Magnetization (A/m)'' or ``Magnetization
\{A[m(1)]\}'', not just ``A/m''. Do not label axes with a ratio of
quantities and units. For example, write ``Temperature (K)'', not
``Temperature/K''.
\section*{Acknowledgment}
The preferred spelling of the word ``acknowledgment'' in America is without
an ``e'' after the ``g''. Avoid the stilted expression ``one of us (R. B.
G.) thanks $\ldots$''. Instead, try ``R. B. G. thanks$\ldots$''. Put sponsor
acknowledgments in the unnumbered footnote on the first page.
\section*{References}
Please number citations consecutively within brackets \cite{b1}. The
sentence punctuation follows the bracket \cite{b2}. Refer simply to the reference
number, as in \cite{b3}---do not use ``Ref. \cite{b3}'' or ``reference \cite{b3}'' except at
the beginning of a sentence: ``Reference \cite{b3} was the first $\ldots$''
Number footnotes separately in superscripts. Place the actual footnote at
the bottom of the column in which it was cited. Do not put footnotes in the
abstract or reference list. Use letters for table footnotes.
Unless there are six authors or more give all authors' names; do not use
``et al.''. Papers that have not been published, even if they have been
submitted for publication, should be cited as ``unpublished'' \cite{b4}. Papers
that have been accepted for publication should be cited as ``in press'' \cite{b5}.
Capitalize only the first word in a paper title, except for proper nouns and
element symbols.
For papers published in translation journals, please give the English
citation first, followed by the original foreign-language citation \cite{b6}.
\section{Introduction}
The push towards decarbonization of the electric grid has seen an explosive growth in renewable energy installation.
The state of California generates approximately 25\% of its electricity demand from solar resources, resulting in the characteristic ``duck curve'' in net system load \cite{duck}. This new operating condition, where dispatchable bulk resources must quickly meet the large and rapid change in electricity demand, introduces challenges to grid operators. Storage offers one solution, as a flexible and low-inertia resource, however remains too expensive for utility-scale system-wide adoption. Instead, we propose system operators look towards the distribution grid to provide some support. Small-scale consumer owned distributed energy resources (DERs) which include rooftop PV with inverters, demand response, and batteries can provide support to the bulk grid.
Recent works in this area are \cite{sheha2020solving} which proposes a game theoretic framework for dynamic pricing, and \cite{calero2021duck} which shifts thermal cooling loads of residential units to mitigate duck curve effects. Where these works rely on local approaches with assets locally providing the mitigation mechanisms, our proposed work takes a systems-approach. We leverage distributed optimization to deploy resources over a larger spacial area of the distribution grid to enhance mitigation strategies while suitably accommodating global network constraints.
The main contribution of the paper pertains to the distributed coordination of DERs across a large region of the distribution grid to mitigate the duck curve. This coordination is accomplished by leveraging the variable power factor setting of PV inverters, flexible loads to reduce consumption, and distributed storage devices such as community batteries. This central contribution is realized by formulating the global optimization problem as an optimal power flow, and employing a current-injection (CI) based linear model describing the power physics of the unbalanced distribution grid. The CI-based OPF is solved in a distributed fashion, using a Proximal Atomic Coordination (PAC) approach \cite{RomvaryTAC} with Nesterov acceleration, called NST-PAC. We evaluate the performance on a modified IEEE-34 node network, using load and generation data for San Francisco. Our results show the successful mitigation of the duck curve through the coordination of spatially distributed DERs.
In Sections~\ref{sec:CI} and \ref{sec:PAC} we introduce the CI power flow model for the unbalanced distribution grid, and the accelerated distributed optimization algorithm called NST-PAC. In Section~\ref{sec:case} we carry out a numerical case study, comparing local and distributed paradigms. We provide conclusions in Section~\ref{sec:conclusion}.
\section{Current Injection Model} \label{sec:CI}
The distribution grid is a highly unbalanced network, with many single- and two-phase lines and corresponding loads. As DER penetration increases, many of which are added to single-phase lines, the network will become more unbalanced. In light of this, power flow models which assume balanced flow \cite{gan2014exact} or those which are valid for only a small range of angle imbalances \cite{robbins2015optimal} cannot suitably describe the future grid. Further, the faster timescales of operation due to renewable variability and flexibility requirements need a model which is simple and results in OPF problems which are computationally tractable. For this reason, there is interest in developing scalable linear models for the unbalanced grid, and leveraging distributed computation and algorithms to solve the OPF problem.
In this work, we utilize the Current Injection (CI) model, a linear model for multi-phase unbalanced distribution grids. The CI model takes a similar approach to the Bus Injection Model, wherein all nodal power injections are modeled as current injections, and all phasor variables are represented in Cartesian coordinates. The 3-phase impedance matrix is used to describe the self and mutual inductance between phases to model the coupling of phases that are common to a distribution grid. The non-convexity of the AC-OPF and the subsequent nonlinearity of SOCP and SDP convexification strategies typically used are avoided in the CI approach by leveraging McCormick Envelope (MCE) based convex relaxation \cite{mccormick_1976} for the bilinear power relations. The MCE uses the convex hull representation of bilinear terms to render a linear OPF model. Suitable pre-processing techniques can be utilized to determine adequate bounds on the nodal voltages and currents to ensure a tight convex relaxation.
\subsection{Problem Formulation}
We denote a general distribution network as a graph $\Gamma(\mathcal{N},\mathcal{E})$, where $\mathcal{N}\coloneqq\{1,...,N\}$ denotes the set of nodes, $\mathcal{E}\coloneqq\{(m,n)\}$ denotes the set of edges, and each phase is expressed as $\phi\in \mathcal{P}, \; \mathcal{P}=\{a,b,c\}$. The CI-OPF for the network is then written as\footnote{For brevity, we omit the McCormick Envelope relaxation of the bilinear terms and the pre-processing step. See \cite{FerroCIM_IFAC,FerroThesis} for details}:
\begingroup\makeatletter\def\f@size{9.5}\check@mathfonts
\def\maketag@@@#1{\hbox{\m@th\normalfont#1}}
\begingroup
\allowdisplaybreaks
\begin{subequations} \label{eq:CI}
\begin{align}
\underset{x}{\mathop{\min }}\, f & \left(x\right) \label{eq:CIOPF_obj} \\
AV &= ZI_{\textrm{flow}} \label{eq:CIOPF_ohm} \\
{{I}^{R}} &= \operatorname{Re}\left( A^TI_{\textrm{flow}} \right),\quad {{I}^{I}} = \operatorname{Im}\left( A^TI_{\textrm{flow}} \right) \label{eq:CIOPF_Idef} \\
P_{j}^{\phi} &= V_{j}^{\phi,R}I_{j}^{\phi,R}+V_{j}^{\phi,I}I_{j}^{\phi,I} \label{eq:CIOPF_Pdef} \\
Q_{j}^{\phi} &= -V_{j}^{\phi,R}I_{j}^{\phi,I}+V_{j}^{\phi,I}I_{j}^{\phi,R} \label{eq:CIOPF_Qdef} \\
\underline{P}_{j}^{\phi} &\le P_{j}^{\phi} \le \overline{P}_{j}^{\phi}, \quad \underline{Q}_{j}^{\phi} \le Q_{j}^{\phi}\le \overline{Q}_{j}^{\phi} \label{eq:CIOPF_powerlim} \\
\underline{V}_{j}^{\phi,R} &\le V_{j}^{\phi,R} \le \overline{V}_{j}^{\phi,R}, \quad \underline{V}{j}^{\phi,I} \le V_{j}^{\phi,I} \le \overline{V}_{j}^{\phi,I} \label{eq:CIOPF_Vlim}\\
\underline{I}_j^{\phi,R} &\le I_j^{\phi,R}\le \overline{I}_j^{\phi,R}, \quad \underline{I}_j^{\phi,I} \le I_j^{\phi,I}\le \overline{I}_j^{\phi,I} \label{eq:Ilim}
\end{align}
\end{subequations}
\endgroup \endgroup
where $x = \left[{{I}^{R}}\;{{I}^{I}}\;{{V}^{R}}\;{{V}^{I}}\;P\;Q\;I_{\textrm{flow}}^R\;I_{\textrm{flow}}^I\right]$ is the decision vector for the CI-OPF problem; $I, V, P, Q$ denote the vector of nodal current injections, voltages, and real/reactive power injections respectively; $I_{\textrm{flow}}$ denotes the vector of line currents; $A \in \mathbb{R}^{3N\times3N}$ is the 3-phase graph incidence matrix; $Z$ is the system impedence matrix. We use $x^R$ and $x^I$ to denote the real and imaginary components of a complex number $x$; overbar $\overline{{x}}$ and underbar $\underline{{x}}$ denote the upper and lower limits of a variable ${x}$; $\operatorname{Re}(\cdot)$ and $\operatorname{Im}(\cdot)$ denote the real and imaginary components of a complex number. Constraint \eqref{eq:CIOPF_ohm} describes the generalized Ohm's law, \eqref{eq:CIOPF_Idef} describe Kirchhoff's Current Law, and \eqref{eq:CIOPF_Pdef}-\eqref{eq:CIOPF_Qdef} are the definitions of real and reactive power. Constraints \eqref{eq:CIOPF_Pdef}-\eqref{eq:Ilim} are for all nodes $j \in \mathcal{N}$, per each phase $\phi \in \mathcal{P}$, and for all time $t$.
The convex relaxation of the CI-OPF problem through McCormick envelopes adds additional constraints to the problem namely \eqref{eq:Ilim}, concerning the minimum and maximum limits of injected current at each phase of each node. The determination of these upper and lower bounds on nodal current is not trivial. An effective heuristic to define these values is presented in \cite{FerroThesis}.
\subsection{Model of DERs}
We use active sign convention, such that nodal injections are positive. The PV units are modelled as a generator with variable power factor (\textit{pf}), equipped with a multiphase inverter, where the ratio of P and Q determine the \textit{pf} setting:
\begingroup\makeatletter\def\f@size{9.5}\check@mathfonts
\def\maketag@@@#1{\hbox{\m@th\normalfont#1}}
\begin{align}
P_{j}^{\phi} \tan(\cos^{-1}(-{pf})) &\leq Q_j^{\phi} \leq P_j^{\phi} \tan(\cos^{-1}({pf})) & \label{eq:CIOPF_pf}
\end{align}
\endgroup
Flexible loads are reductions in real power demand (reactive power loads are fixed), and are modelled as a percentage reduction from the forecasted load. The power constraints are modified as $\overline{P}_j(t) = \underline{P}_j(t)*(1-\alpha_j^\text{DR}(t))$, by recalling the use of active sign convention. For inflexible loads, $\underline{P}_j=\overline{P}_j, \overline{P}_j \leq 0$. Prosumers, which are nodes where both load and generation are present, are modelled with additional variables representing both load and generation, $P^L$ and $P^G$ respectively. For prosumers, the inverter is modelled on $P^G$ and $Q^G$, and load flexibility is modelled on $P^L$. Variables $P^L$ and $P^G$ are both nonnegative, and the same for reactive power.
\vspace{-10pt}
\begingroup\makeatletter\def\f@size{9.5}\check@mathfonts
\def\maketag@@@#1{\hbox{\m@th\normalfont#1}}
\begin{subequations} \label{eq:prosumer}
\begin{align}
P_j &= P_j^G-P_j^L, \quad P_j^G \geq 0, P_j^L \geq 0 \\
Q_j &= Q_j^G-Q_j^L, \quad Q_j^L \geq 0 \\
\overline{P}_j &= \overline{P}_j^G-\underline{P}_j^L, \quad \underline{P}_j = \underline{P}_j^G-\overline{P}_j^L \\
\overline{Q}_j &= \overline{Q}_j^G-\underline{Q}_j^L, \quad \underline{Q}_j = \underline{Q}_j^G-\overline{Q}_j^L
\end{align}
\end{subequations}
\endgroup
\noindent Battery storage devices are modelled using the power charge and discharge, $P_{j}^\text{sc}(t)$ and $P_{j}^\text{sd}(t)$ respectively for node $j$ and time $t$. These are nonnegative variables. The state of charge, $b_j(t)$, is calculated as an integral constraint using the actions of the previous period and the initial state of charge, $b_j^0 \coloneqq b_j(t=0)$. We model charge and discharge efficiencies ($\eta_j^C$ and $\eta_j^D$), self-discharge rate ($\eta_j^\text{self}$), and impose a minimum state of charge ($\underline{b}_j$) to ensure battery health. We assume all batteries operate at unity power factor, i.e. $Q_{j} = 0\quad \forall t$.
\begingroup\makeatletter\def\f@size{9.5}\check@mathfonts
\def\maketag@@@#1{\hbox{\m@th\normalfont#1}}
\begin{subequations}
\begin{align}
P_j &= \frac{1}{\eta_j^\text{D}} P_{j}^\text{sd}(t) - \eta_j^\text{C} P_{j}^\text{sc}(t) \\
0 &\leq P_{j}^\text{sd} \leq \overline{P}_{j}^\text{sd}, \quad 0 \leq P_{j}^\text{sc} \leq \overline{P}_{j}^\text{sc} \\
b_{j}(t) &= (1-\eta_j^\text{self})b_{j}(t-1) + \eta_j^\text{C} P_{j}^\text{sc}(t) - \frac{1}{\eta_j^\text{D}} P_{j}^\text{sd}(t) \\
\underline{b}_{j} &\leq b_{j}(t) \leq \overline{b}_{j}
\end{align}
\end{subequations}
\endgroup
The storage model introduces a dependency on control action in one period to previous periods. The other DERs and loads do not have inter-temporal constraints, so constraints \eqref{eq:CI}-\eqref{eq:prosumer} are simply replicated for each time step $t$.
\section{PAC-Based Distributed Implementation} \label{sec:PAC}
In this section we introduce a PAC-based distributed optimization algorithm to coordinate a large number of spatially distributed DERs at the grids edge. Consider a global optimization problem composed of equality and inequality constraints, which may be coupled in time
\begin{gather}
\underset{x}{\text{min}} \sum_{i=1}^S f_i(x) \nonumber\\
\text{s. t.}\quad Gx=b, \quad Hx\le d\vspace{-0.02in}
\label{eq:global}
\end{gather}
where $\sum_{{i} \in S} f_i\left({x}\right)$ represents the total objective function. Problem \eqref{eq:global} can be distributed into $K=\left\{1,...,k,...K\right\}$ separate coupled optimization problems, denoted as \textit{atoms}. We use a decomposition profile which separates the vector of variables $x$ into two sets: $\mathsf{L} = \left\{ L_{j}, \quad \forall j \in K\right\}$ and $\mathsf{O} = \left\{ O_{{j}}, \quad \forall j \in K\right\}$, which represent the partition of decision variables ``owned'' and ``copied'' by atom $j$. The set of total variables (owned and copied) by an atom is denoted as $\mathsf{T}$. The decomposition profile also separates the constraints into sets owned by each atom, as $\mathsf{C} = \left\{ C_{{j}}, \quad \forall j \in K \right\}$. The notion of variables copies are used to satisfy the coupling in constraints and/or objective function. In the context of the CI model, the power physics of the grid \eqref{eq:CIOPF_ohm}-\eqref{eq:CIOPF_Idef} result in these coupling constraints. Note that the CI does not have coupling introduced by inequality constraints, however the decomposition can be trivially extended to coupled inequality constraints. Using the decomposition profile, we obtain:
\begin{equation} \label{eq:sof:aso}
\begin{array}{c}
\underset {{a}_{{j}}}{\text{{ min}}} \quad \sum_{{j} \in K}{f}_{{j}}\left({a}_{{j}}\right) \\
\begin{array}{rl}
\text{subj. to:} & \left\{\begin{array}{rl}
{{G}}_{{j}} {a}_{{j}} = {b}_{j}, & \text{ for all } {j} \in K\\
{{H}}_{{j}} {a}_{{j}} \le {d}_{j}, & \text{ for all } {j} \in K\\
B_j a = {0}, & \text{ for all } {j} \in K
\end{array}\right.
\end{array}
\end{array}
\end{equation}
where $a_j$ is atom $j$'s variables (both owned and copied), $f_j \left(a_j\right) $ is the atomic objective function, and $G_j$, $b_j$, $H_j$, and $d_j$ represent the submatrix or subvector of $G, b, H$, and $d$ respectively. Finally, $B$ is in incidence matrix over the owned and copied atomic variables, defined as
\begin{equation*}
B_{i}^{m} \triangleq \left\{\begin{array}{rl} -1, & \text{if $i$ is `owned` and $m$ a related `copy`} \\
1, & \text{if $m$ is `owned` and $i$ a related `copy`} \\
0, & \text{otherwise}\end{array}\right.
\end{equation*}
We then use $B_j$ ($B^j$) to denote the relevant incoming (out-going) edges of the directed graph for atom-$j$. To fully parallelize the optimization, we introduce \textit{coordination constraints}, which must be satisfied for every atom. These require all atomic copied variables in a given ${j}$th atom to equal the value of their corresponding owned in $i$th atom, $i\neq j$:
\begin{align}
B_j {a} = 0 \quad \forall {j} \in K
\end{align}
\subsection{An Accelerated Algorithm: NST-PAC}
We next present an accelerated variant of the PAC algorithm in \cite{RomvaryTAC}, which includes time-varying gains and Nesterov type acceleration \cite{nesterov1983method}, called NST-PAC. The NST-PAC is a primal-dual method with $\ell_2$ and proximal regularization, Nesterov type acceleration for both primal and dual variables, and privacy-preserving features. We begin by forming the atomic Lagrangian function:
\begingroup\makeatletter\def\f@size{9.5}\check@mathfonts
\def\maketag@@@#1{\hbox{\m@th\normalfont#1}}
\begin{align}
\mathcal{L} \left({a},{\mu},{\nu}\right) &= \sum_{{j} \in K} \left[{f}_{{j}} \left({a}_{{j}}\right) + {\mu}_{{j}}^{T} ({{G}}_{{j}} {a}_{{j}}-b_j) + {\nu}_{{j}}^{T} B_j {a} \right] \nonumber \nonumber \\
&= \sum_{{j} \in K} \left[{f}_{{j}} \left({a}_{{j}}\right) + {\mu}_{{j}}^{T} ({{G}}_{{j}} {a}_{{j}}-b_j) + {\nu}^{T} B^j {a}_{j} \right] \nonumber \\
&\triangleq \sum_{{j} \in K} \mathcal{L}_{{j}} \left({a}_{{j}}, {\mu}_{{j}}, {\nu}\right). \label{eq:pac:alg:1:1}
\end{align}
\endgroup
The algorithm is carried out as below:
\begingroup\makeatletter\def\f@size{9.5}\check@mathfonts
\def\maketag@@@#1{\hbox{\m@th\normalfont#1}}
\begin{align}
{a}_{{j}} \left[ \tau + 1 \right] &= \underset{{{a}_{{j}} \in \mathbb{R}^{|T_j|}}}{\text{{ argmin}}} \left\{\mathcal{L}_{{j}}\left({a}_{{j}}, \hat{{\mu}}_{{j}} \left[\tau\right], \hat{{\nu}}\left[ \tau \right]\right)\right.
+\frac{\rho_j\gamma_j}{2}\left\| {{G}_{j}}{{a}_{j}}-{{b}_{j}} \right\|_{2}^{2}\nonumber\\
&\quad \left.+ \frac{\rho_j\gamma_j}{2}\left\| {{B}_{j}}{{a}_{j}} \right\|_{2}^{2}+ \frac{1}{2\rho_j} \left\Vert {a}_{{j}} - {a}_{{j}} \left[ \tau \right] \right\Vert_{2}^{2}\right\},\label{pacnst_beg} \\
{{\hat{a}}_{j}}\left[ \tau +1 \right] &= {{a}_{j}}\left[ \tau +1 \right]+{{\alpha }_{j}}[\tau +1]({{a}_{j}}\left[ \tau +1 \right]-{{a}_{j}}\left[ \tau \right])\label{pacnst_acc0} \\
{{\mu }_{j}}\left[ \tau +1 \right] &= {{\hat{\mu }}_{j}}\left[ \tau \right]+\rho_j\gamma_j({{G}_{j}}{{\hat{a}}_{j}}\left[ \tau +1 \right]-{{b}_{j}}) \\
{{\hat{\mu }}_{j}}\left[ \tau +1 \right] &= {{\mu }_{j}}\left[ \tau +1 \right]+{{\phi }_{j}}[\tau +1]({{\mu }_{j}}\left[ \tau +1 \right]-{{\mu }_{j}}\left[ \tau \right])\label{pacnst_acc1} \\
\text{Comm}&\text{unicate } {{\hat{a}}_{j}}\text{for all }j\in \left[ K \right]\text{ with neighbors} \\
{{\nu }_{j}}\left[ \tau +1 \right] &= {{\hat{\nu }}_{j}}\left[ \tau \right] + \rho_j\gamma_j{{B}_{j}}{{\hat{a}}_{j}}\left[ \tau +1 \right] \\
{{\hat{\nu }}_{j}}\left[ \tau +1 \right] &= {{\nu }_{j}}\left[ \tau +1 \right]+{{\theta }_{j}}[\tau +1]({{\nu }_{j}}\left[ \tau +1 \right]-{{\nu }_{j}}\left[ \tau \right])\label{pacnst_acc2} \\
\text{Comm}&\text{unicate } {{\hat{\nu }}_{j}}\text{ for all }j\in \left[ K \right]\text{with}\,\text{neighbors} \label{pacnst_end}
\end{align}
\endgroup
\noindent where $\rho_j$, $\gamma_j$ are atom-varying over-relaxation and step-size parameters, respectively. The proposed NST-PAC uses $\ell2$ regularization terms rather than the prox-linear variant in PAC. Further, both primal and dual variables are accelerated using Nesterov type acceleration, to speed up convergence. Further, we extend the privacy-preserving feature of the PAC algorithm to both the primal and dual variables, by using three iteration-varying and atom specific parameters for the accelerated terms, ${{\alpha }_{j}}\left[ \tau \right]>\alpha _{j}^{\min }$, ${{\phi }_{j}}\left[ \tau \right]>\phi _{j}^{\min }$ and ${{\theta }_{j}}\left[ \tau \right]>\theta _{j}^{\min }$. In the original algorithm privacy is kept only for dual variables. A detailed analysis of convergence rate, communication and computational complexity, and privacy are provided in \cite{RomvaryTAC}.
\section{Case Study} \label{sec:case}
We consider a case study of San Francisco, California, using the IEEE-34 node network as a proxy for the distribution grid. The load data from the IEEE datasheet serves as the daily average load, and the 24-hr load profiles are obtained from the ODEI dataset from NREL for the Typical Meteorological Year \cite{odei_dataset}. All loads are assumed to have a constant power factor of 0.95. The network loads are classified as residential or commercial loads based on the size of the load, by matching the load levels of the IEEE-34 network with the TMY data. Commercial loads include retail space, small and medium office buildings, primary school, medium and large restaurants, and a hospital. The network is modified to include DERs which include clusters of rooftop photovoltaic (PV) units, flexible loads, and three battery storage units. The penetration of PV is 38\%, as measured by the ratio of nameplate capacity to average system load. This high DER penetration scenario is a reasonable projection given the RPS initiatives in California. Each PV unit is assumed to be equipped with an inverter with corresponding power electronic control, which can be operated at variable power factor in the range of 0.8 to 1. PV curtailment is not considered. The flexible loads are modelled as typical residential cooling loads for California \cite{LBNL_CAISO_DR}. Variations in nodal demand response are obtained by shifting the baseline profile obtained from \cite{LBNL_CAISO_DR} in time and space, with both following zero-mean Gaussian distributions with variances of 0.075 and 0.1 respectively. The three battery units are a 450kW-120kWh community unit\footnote{Modeled on the Ellenbrook unit from the PowerBank trail, Australia \cite{tagabe_2020}} at node 6, a cluster of 40 Tesla Powerwall+ batteries (each 13.5kW-5kWh) at node 19, and a 800kW-185kWh hospital unit at node 27. Any load not met by local generation (or storage) is assumed to be served by the bulk grid at the point of common coupling (PCC), at node 1 in the network ($j = 1$). Figure~\ref{fig:ieee_case} shows the network topology.
\begin{figure}
\centering
\includegraphics[scale=0.4]{fig/ieee34_topology_withlegend.pdf}
\caption{Topology of IEEE-34 network. PV units and flexible loads are present throughout the network.} \vspace{-0.2in}
\label{fig:ieee_case}
\end{figure}
\subsection{Scenarios}
We consider three scenarios, as below:
\begin{itemize}
\item \underline{Scenario A:} \textit{Baseline}. All PV inverters operate at unity pf, and batteries and flexible loads are not present.
\item \underline{Scenario B:} \textit{Local control}. Each DER owner operates its devices and manages its loads.
\item \underline{Scenario C:} \textit{Distributed control}. All devices and loads are coordinated using the NST-PAC algorithm.
\end{itemize}
Scenario A quite trivially is the characteristic duck curve, where the high PV generation at unity power factor results in large ramping requirements of transmission-level generators. Scenario B is a local approach, where each agent will minimize its peak load throughout the day. This serves as a very rough approximation of reducing the ramping requirements of bulk resources, by noticing that the largest ramp typically coincides with the peak demand in the evening. This action can be motivated by the fact that consumers can be charged based on their peak energy consumption. To leverage the capabilities of the storage devices, neighbouring nodes are clustered with the battery. Residential loads (and corresponding DERs) at nodes 3, 4, and 5, share the community battery at node 6. Residential loads at node 20 and the primary school at node 21 share the cluster of Tesla Powerwall+ batteries at node 19. The hospital at node 26 is assumed to own and operate the battery at node 27. All remaining nodes are treated as independent agents. After clustering, there are a total of 26 agents in the network, each managing its own consumption and generation, to minimize its peak load. The multi-period optimization problem solved by each agent is a simple power balance, where any load in excess of local generation is assumed to be served by the bulk system. The power physics between nodes within a cluster is not modelled. The objective function is $f_\text{local}(y) = \text{max}_{t} \left\{ -P(t)\right\}$,
where $y=\left[P, Q,P^L,P^G,Q^L,Q^G,P^\text{sd},P^\text{sc} \right]$.
Scenario C is the PAC-based approach which requires coordination between the devices. This approach accommodates system-level constraints including grid power physics, and minimizes the ramping requirements at the PCC. In Scenario C, the PAC agents solve the CI-based multi-phase unbalanced OPF, to minimize the objective function $f(x)$ which minimizes the difference in power supplied by the PCC from one hour to the next. The function is $f_\text{PAC}(x) = \left\| \sum_{\phi\in \mathcal{P}} P_1^\phi(t) - P_1^\phi(t-1) \right\|$.
\subsection{Results and Discussion}
Each scenario was simulated on the IEEE-34 node network. Figure~\ref{fig:res_netload} shows the net load served by the bulk system. The load curve for Scenario A shows the characteristic high ramps down and up when solar generation begins and ends. Scenario B provides minimal improvement, reducing the load for most hours of the day, while Scenario C effectively leverages the DERs to reduce the ramping requirements. Note that the objective function reduces the hour-to-hour change in load, and to do so, increases load during the hours of peak PV generation (roughly 10am to 4pm) to charge the batteries, which are then discharged in the late evening to reduce the net load. The magnitude of hour-to-hour ramping is shown in Figure~\ref{fig:res_ramping}. Notably, the local optimization of Scenario B, which minimizes the peak load throughout the day as a proxy for minimizing the load ramp in the evening, is not able to suitably reduce the ramping requirement. The distributed optimization approach, on the other hand, is able to leverage system-wide information and coordinate the DERs to provide grid-level support, as needed by the bulk system. This coordinated approach is able to reduce ramping requirements throughout the day, with a 23\% reduction in ramping requirements at the 4pm peak. Table~\ref{tab:results} presents the total ramping reduction for Scenarios B and C, as compared to the baseline in A, and the computational run times. As expected, the proposed distributed coordination significantly outperforms the local approach, with 28\% reduction in ramping required. The local approach is unable to provide any reduction. Unsurprisingly, the local approach takes less time to reach a decision (albeit an inferior one), while the distributed approach takes considerably longer (completing 1000 iterations). However, computational time of 20s is still well within the acceptable time-frame for decision making, and enables DERs to provide bulk-level support.
\begin{figure*}[]
\centering
\subfloat[][Minimum initial SOC] {\includegraphics[scale=0.26,trim=1.5cm 7cm 0cm 6.5cm,clip]{fig/SOC_battzerocap_minramp_Nov7.pdf}\label{fig:res_storagePAC_zero}}
\subfloat[][Mid initial SOC] {\includegraphics[scale=0.26,trim=1.5cm 7cm 0cm 6.5cm,clip]{fig/SOC_battmidcap_minramp_Nov7.pdf}\label{fig:res_storagePAC_mid}}
\subfloat[][Full initial SOC] {\includegraphics[scale=0.26,trim=1.5cm 7cm 0cm 6.5cm,clip]{fig/SOC_battfullcap_minramp_Nov7.pdf}\label{fig:res_storagePAC_full}}
\caption{State of charge of each storage device for Scenario C, with different initial state of charge for each battery.}
\label{fig:res_storagePAC} \vspace{-0.2in}
\end{figure*}
\begin{figure}[]
\centering
\includegraphics[scale=0.28,trim=1.5cm 7cm 0cm 6.5cm,clip]{fig/netload_battfullcap_minramp_Nov7.pdf}
\caption{Net load serviced by the bulk system, for each scenario. These plots do not include power loss over lines.}
\label{fig:res_netload} \vspace{-0.2in}
\end{figure}
\begin{figure}[]
\centering
\includegraphics[scale=0.28,trim=1.5cm 7cm 0cm 6.5cm,clip]{fig/ramping_battfullcap_minramp_Nov7.pdf}
\caption{Magnitude of ramping required by bulk system generators for each scenario.}
\label{fig:res_ramping} \vspace{-0.35in}
\end{figure}
We next investigate the impact of the battery units' initial state of charge. We run each Scenario for (1) Minimum SOC where initial capacity is at 45, 0, and 160 kWh; (2) Mid SOC where initial capacity is at 120, 400, and 400 kWh; and (3) Full SOC where initial capacity is at 450, 540, and 800 kWh, respectively for batteries at nodes 6, 19, and 27.
Figure~\ref{fig:res_storagePAC} plots the SOC of each battery unit for the three cases. The usage pattern in each of the cases is very similar, with charging in the early morning and afternoon to build up storage capacity for the evening. Interestingly, the final SOC are non-zero for the three cases, and are quite similar, suggesting the batteries may be able to retain a higher minimum charge to be used as backup power in the case of emergencies. The batteries provide flexibility to increase or decrease load throughout the day, as required by the grid. The plots for net load are quite similar for all three cases, and so have been omitted.
\begin{table}[]
\begin{tabular}{l|ccc}
& \multicolumn{1}{l}{A (Baseline)} & \multicolumn{1}{l}{B (Local)} & \multicolumn{1}{l}{C (Proposed)} \\ \hline
Total ramping need (kW) & 3923.7 & 3941.0 & 2839.4 \\
Ramping reduction & - & -0.44\% & 27.63\% \\
Mean run time per agent & - & 0.0947s & 16.97s
\end{tabular}
\caption{Summary of results for Scenarios A thru C. Scenario A is baseline with no decision making, so no ramping reduction or computational time to report.}
\label{tab:results}
\end{table}
\section{Conclusion} \label{sec:conclusion}
We presented and evaluated a distributed approach to coordinate DERs to provide services to the bulk system. We leverage a CI-based linear model of the unbalanced grid, and an accelerated PAC-based algorithm called NST-PAC. Our case study on a modified IEEE-34 node network shows how distributed techniques can leverage information from different resources to successfully mitigate the duck curve, reducing ramping requirements of bulk system generators by up to 23\%. This framework can be extended to include electric vehicles, by modeling the vehicle-to-grid capabilities and corresponding cost of battery cycling and lifetime degradation in the optimization problem. Using such an approach throughout the distribution grid can reduce the challenges of the new operating condition resulting from high solar and renewable penetration, reduce system costs, and improve renewable integration. Future work will concern the development of faster distributed algorithms applicable to nonconvex costs and constraints, with discrete and continuous controls.
\bibliographystyle{ieeetr}
|
\section{Introduction}
\label{sec:introduction}
In the
\textsc{Online Vector Balancing}\xspace problem introduced by Spencer~\cite{Spencer77}, vectors
$a_1, a_2, \ldots, a_T \in [-1,1]^n$ arrive online, and the algorithm
irrevocably assigns a sign $\sigma(a_t)$ immediately upon seeing $a_t$,
with the goal of minimizing the \emph{discrepancy of the signed sum},
i.e., $\| \sum_t \sigma(a_t)\cdot a_t \|_{\infty}$. Following a sequence of works~\cite{BS20,BJSS20,BJMSS-SODA21}, the state-of-the-art bounds for this problem is an elegant randomized algorithm that maintains a
discrepancy of $O(\sqrt{n} \log (nT))$~\cite{ALS-STOC21}. Their result assumes an
\emph{oblivious} adversary, so that the choice of arriving vectors does not
depend on the internal state of the algorithm.
Indeed, if we allow \emph{adaptive}
adversaries then every online algorithm incurs $\Omega(\sqrt{T})$ discrepancy~\cite{Spencer77}.
We think of $T \gg n$, so $\Omega(\sqrt{T})$ is much larger than $O(\sqrt{n} \log (nT))$.
We initiate the study of \textsc{Fully-Dynamic Vector Balancing}\xspace, where vectors can both arrive
or depart at each time step, and the algorithm must always maintain a
low-discrepancy signing of the vectors present in the system at all
times. Since it is easy to construct examples where no algorithm can
guarantee non-trivial discrepancy bounds if it is forced to commit to
the sign of a vector upon arrival, we study the problem where the
algorithm can \emph{re-sign the vectors from time to time}. Indeed,
many real-world applications that motivate such discrepancy-based
methods (such as in fair allocations, sparsification routines, etc.)
have a fully-dynamic flavor to them, with the corresponding inputs
being dynamic in nature due to both insertions and deletions.
\textbf{Problem} (\textsc{Fully-Dynamic Vector Balancing}\xspace). \textit{
We start with an empty collection of \emph{active} vectors $A(0)$. At each time/update
$t \in [T]$, an adaptive adversary either inserts a new vector
$a_t \in [-1,1]^n$, i.e., $A(t) = A(t-1) \cup \{a_t\}$, or removes
an existing vector $a \in A(t-1)$, i.e.,
$A(t) = A(t-1) \setminus \{a\}$. The goal is to maintain
\emph{signings} $\sigma_t: A(t) \rightarrow \{\pm 1\}$ to minimize
the norm $\| \sum_{a \in A(t)} \sigma_t(a) \cdot a \|_{\infty}$. The
algorithm can reassign the sign of a vector $a$ (i.e., set
$\sigma_t(a) \neq \sigma_{t-1}(a)$), and the total \emph{recourse} is the sum total of the reassignments.
}
Two trivial solutions exist: (a)~recomputing low-discrepancy signings
on the active set of vectors after every update operation incurs optimal offline discrepancy guarantees with a recourse of $\Theta(T)$ per
update, and (b) an independent and uniformly random signing of every
new vector
maintains at any time $t$ a signing of discrepancy
$\Theta(\sqrt{T \log n})$ w.h.p., while performing no recourse
whatsoever.
Since $T \gg n$, this is much larger than the
optimal offline discrepancy bounds of $O(\sqrt{n})$ for any collection of $T$ vectors in
$[-1,1]^n$~\cite{Spencer85,Bansal-FOCS10,Lovett-Meka-SICOMP15}.
We ask: \emph{can we get near-optimal\footnote{In this paper, we use ``near-optimal'' to mean optimal up to poly-logarithmic factors.} discrepancy bounds with a small amount of
recourse?}
\subsection{Our Results and Techniques}
\paragraph{Fully-Dynamic Vector Balancing.}
Our first main contribution is the design of an algorithm which maintains low-discrepancy
signings for the fully-dynamic problem that nearly matches the offline discrepancy
bounds while giving an amortized recourse that is only logarithmic in the
sequence length $T$.
\begin{theorem}[Fully-Dynamic: General Vector Balancing]
\label{thm:main1}
There is an efficient algorithm for \textsc{Fully-Dynamic Vector Balancing}\xspace with update
vectors in $[-1,1]^n$ which maintains signings
$\sigma_t(\cdot)$ with discrepancy $O(\sqrt{n})$ and an amortized recourse of $O(n \log T)$ per update, even against adaptive adversaries. For Komlos' setting, i.e., if all the updates vectors have $\ell_2$ length at most $1$ (instead of $\ell_\infty$ length), the algorithm achieves discrepancy $O(\sqrt{\log(n)})$ with an amortized recourse of $O(n \log T)$ per update.
\end{theorem}
Since in this theorem we are competitive against \emph{adaptive adversaries}, it illustrates the power of recourse: in the absence of recourse, we get $\Omega(\sqrt{T})$ lower bounds
on the discrepancy even for arrival-only sequences of $2$-dimensional vectors. This is because the adversary can always make the next vector to be
orthogonal to the current signed sum.
At a very high level, our algorithm divides the instance into many
parts of size $O(n)$, obtains a good partial signing for each part
(such that all but $n$ vectors are signed), and recurses on the
residual instance. The algorithm imposes a tree-like hierarchy on
these parts, so that it can easily adapt to inserts or deletes with
bounded recourse by only re-running the computations on the part
suffering the insertion/deletion, and on any internal node on the
corresponding root-leaf path from that part to the root. If we are not
careful, the discrepancy of the overall vector can be proportional to
the number of parts, since we could accrue error in each
part. However, we use linear algebraic ideas inspired
by~\cite{BaranyGrinberg81} to \emph{couple all the parts}, thereby
always ensuring that the sum of the partial signings across all nodes
of the tree (except the root) is zero.
\paragraph{Fully-Dynamic Edge Orientation/Carpooling and Local Search.}
Given the general result above, next we focus on the special case of
orienting edges of a graph to minimize the maximum imbalance between
the in- and out-degrees.
Fagin and Williams~\cite{Fagin83}
posed the \emph{carpooling problem},
which corresponds to vector balancing with vectors of the form
$(0, \ldots, 0, 1, 0, \ldots, 0, -1, 0, \ldots) \in \mathbb{R}^n$, and the
graph discrepancy objective is precisely the $\| \cdot \|_{\infty}$ of
the signed sum of vectors.
\cite{Fagin83,Ajtai98} use this problem to model fairness in
scheduling, where edges represent shared commitments (such as
carpooling), orientations give primary and secondary partners of the
commitment (e.g., driver and co-driver), and hence the discrepancy
measures fairness for individuals, in terms of how many commitments
he/she is the primary partner for, relative to the total number of
commitments he/she is a part of.
Somewhat surprisingly, \cite{Ajtai98} showed that any algorithm must
suffer $\Omega(n)$ discrepancy on some worse-case adaptive sequence of
edge arrivals. On the other hand, for an oblivious sequence of edge
arrivals, it is easy to maintain orientations with
$O(\sqrt{n \log n})$ discrepancy by simply orienting edges randomly
(while always orienting repeated parallel edges $(u,v)$ oppositely).
To mitigate such strong lower bounds, \cite{Ajtai98} and recently
Gupta et al.~\cite{GuptaKKS20} study a stochastic version of the
problem where the arriving edges are sampled from a known distribution:
they design algorithms to maintain $\operatorname{poly}\!\log(n,T)$-discrepancy. The
recent algorithm of Alweiss et al.~\cite{ALS-STOC21} also extends to
this special case giving $O(\log (nT))$ discrepancy bounds for any
oblivious sequence of edge arrivals, not just stochastic ones. None of these prior algorithms extend to a
fully-dynamic input consisting of both insertions and deletions.
Moreover, \Cref{thm:main1} guarantees near-optimal discrepancy only with $O(n \log T)$ amortized recourse.
In this paper, we give \emph{deterministic} near-optimal discrepancy algorithms with near-optimal amortized recourse.
\begin{theorem}[Fully-Dynamic Edge Orientation]
\label{thm:main2}
There is an efficient deterministic algorithm that maintains an
orientation of $\operatorname{poly}\!\log n$ discrepancy while performing an
amortized recourse of $\operatorname{poly}\!\log n$ per update.
\end{theorem}
Since this algorithm is deterministic,
the guarantees also hold against
adaptive adversaries: there are $\Omega(n)$
discrepancy bounds for no-recourse algorithms against such adversaries, even
for the setting of only arrivals.
At a high level, our algorithm can be seen as a composition of two
modules. Firstly, we consider a simple local-search procedure, which
flips an edge from $u \rightarrow v$ to $v \rightarrow u$ if the
current discrepancy of $v$ exceeds that of $u$ by more than
$2$. Clearly, this reduces the discrepancy of the maximum of these two
vertices. Our crucial observation is that this process always maintains
low-discrepancy signings when the graph is an \emph{expander}. We find
this interesting, since we can show that there are bad local optima with
$\operatorname{poly}(n)$ discrepancy for general graphs.
Secondly, we show how to dynamically maintain a partitioning of the
edge set of an arbitrary graph $G$ into a disjoint collection of
expanders $G_1, G_2, \ldots, G_{\ell}$ with each vertex appearing in
at most $\operatorname{poly}\!\log n$ many expanders, such that the \emph{amortized
number of changes} to $G_1$, $G_2$, $\ldots, G_\ell$, per
update to $G$ is bounded. (This expander decomposition can be viewed
as a ``preconditioning'' step.)
We build on ideas recently developed for dynamic graph
algorithms~\cite{saranurak2019expander,bernstein2020fullydynamic}: our challenge is to show
that dynamic expander decomposition can be done along with
local search on the individual expanders, and specifically to control
the potential functions that guide our proofs.
Indeed, using the above two modules to obtain~\Cref{thm:main2}
requires new ideas.
When an update (insertion or deletion) occurs
to $G$, we first modify our expander decomposition,
and re-run local search \emph{starting from the prior local optima} in
each expander. While this ensures good discrepancy bounds, it could lead to many local search moves.
In order to bound the latter quantity,
our idea
is to use a potential
function in each expander to bound the recourse, such that each step
of local search always decreases the potential by at least a constant. This
is somewhat delicate: a single update in $G$ can change any particular
expander $G_i$ by a lot
(even though the amortized
recourse is bounded), and hence the single-step potential change can
be huge. We show how to maintain some Lipschitzness properties for our
potential function under inserts and deletes, which gives us the final bounds
of $\operatorname{poly}\!\log n$ on both the discrepancy and the
recourse.
Along the way, we also develop a better understanding of the strengths
and limitations of local search as a technique for discrepancy minimization
problems, both for graphs and for general vectors.
\begin{theorem}[Informal: Discrepancy of Local Optima]
\label{thm:main3}
For edge orientation in expanders, any locally optimal solution for
local search using the simple potential
$\Phi = \sum_{v \in V} disc(v)^2$ has discrepancy $O(\log
n)$. For arbitrary graphs, however, the discrepancy can be as bad as
$\Omega(n^{1/3})$. For general vectors in $\{\pm 1\}^n$ (and in
$[-1,1]^n$), the local search bound using the simple potential $\Phi = \sum_{i \in [n]} \big(\sum_t \sigma(a_t)\cdot a_t(i) \big)^2$
deteriorates to $\Omega(2^n)$ (and to
$\Omega(\sqrt{T})$).
\end{theorem}
\medskip \noindent {\bf Signing $s$-Sparse Vectors for Online Arrivals.} Finally,
we consider the problem with $s$-sparse vectors, which interpolates
between the graphical case of $s=2$ and the general case.
In the offline setting, the classical linear-algebraic algorithm of
Beck and Fiala~\cite{BeckFiala-DAM81} constructs a signing with
disrepancy $2s-1$ (independent of $n$ and $T$). Subsequent works by
Banaszczyk~\cite{Banaszczyk-Journal98} and Bansal, Dadush, and Garg~\cite{BansalDG-FOCS16} develop techniques to get
discrepancy $O(\sqrt{s \log n})$, and a long-standing question in
discrepancy theory is to improve this bound to $O(\sqrt{s})$.
Here we study the \textsc{Online Vector Balancing}\xspace problem only with arrivals. In this setting, the algorithm of \cite{ALS-STOC21} maintains signings
of discrepancy $O(\sqrt{s} \log (nT))$ without recourse, but against
an \emph{oblivious} adversary. In \Cref{sec:insert-only} we give a generic reduction that
can maintain near-optimal discrepancy for \textsc{Online Vector Balancing}\xspace
against adaptive adversaries, with small recourse.
\begin{theorem}[Arrivals Only: Online Vector Balancing with Recourse]
\label{thm:main4}
There is an efficient algorithm for \textsc{Online Vector Balancing}\xspace with $s$-sparse vectors
that achieves $O(\sqrt{s \log n} \log T)$ discrepancy and $O(\log T)$
amortized recourse per update against an adaptive adversary.
\end{theorem}
\subsection{Further Related Work}
\label{sec:related-work}
Discrepancy theory is a rich and vibrant area of research~\cite{Chazelle-Book01,Matousek-Book09}. While some
initial works~\cite{Spencer77,Barany79} focused on the online
discrepancy problem, the majority of research dealt with the offline
setting, where the $T$ vectors are given upfront. Near-optimal
results are known for settings such as discrete set
systems~\cite{Spencer85,Bansal-FOCS10,Lovett-Meka-SICOMP15} (i.e., vectors in $\{0,1\}^n$), \emph{sparse}
set systems~\cite{BeckFiala-DAM81} ($s$-sparse binary vectors),
and general vectors in the unit ball~\cite{Banaszczyk-Journal98,Beck-Combinatorica81,Giannopoulos,Rothvoss14,BansalDGL18}.
There has been a renewed interest in the online discrepancy setting, where many of
the techniques developed for the offline setting no longer
extend. Most of the results for online vector discrepancy deal with
\emph{stochastic settings} of the problem where the arriving vectors
satisfy some distributional
assumptions~\cite{Ajtai98,BS20,BJMSS-SODA21}. A
recent breakthrough work~\cite{ALS-STOC21} gives a very elegant
randomized algorithm with near-optimal discrepancy for general vectors
arriving online. However, to the best of our knowledge, none of these
ideas easily extend to the fully-dynamic setting where vectors can
also depart---which is the focus of this paper. In fact, we do not
know how to adapt existing ideas to establish non-trivial results for
even the simple \emph{deletions-only} setting: starting with $T$
vectors, a uniformly random subset of $\nicefrac{T}{2}$ of these vectors are
deleted one-by-one. Can we always maintain a low-discrepancy signing
of the remaining vectors with small recourse?
The study of dynamic algorithms also has a rich history, both in the
\emph{recourse} model, which measures the number of updates made the
algorithm per update, and the \emph{update-time} model, which measures
the running time of the algorithm per update. Apart from graph
problems, these models have been studied in a variety of settings such
network design~\cite{imase1991dynamic, gupta2014online, gu2016power,
lkacki2015power}, clustering \cite{guo2020power, cohen2019fully},
matching \cite{grove1995online,chaudhuri2009online, bosek2014online},
and scheduling \cite{phillips1993online, westbrook2000load,
andrews1999improved, sanders2009online, skutella2010robust,
epstein2014robust, gupta2014maintaining}, and set cover
\cite{bhattacharya2018deterministic, bhattacharya2017deterministic,
bhattacharya2019deterministically, abboud2019dynamic,
bhattacharya2019new, bhattacharya2020improved,
Gupta:2017:ODA:3055399.3055493}.
A different version of
edge-orientation, commonly known as graph balancing, involves
minimizing just the maximum in-degree (see,
e.g.,~\cite{MR1734116,MR2313458,MR3238400}): the techniques used for
that version seem quite different from those needed here.
\paragraph{Paper Outline.} We present the results for
\textsc{Fully-Dynamic Vector Balancing}\xspace and specifically \Cref{thm:main1} in \S\ref{sec:distrib-BG}. The
results for graph balancing appear in \S\ref{sec:expand-decomp}. Other results for local-search algorithms appear in
\S\ref{section-lower-bounds-local-search} and
\S\ref{sec:local-search-upper}. We close with an insertion-only algorithm
for sparse vectors, and conclusions and open problems in \S\ref{sec:insert-only}.
\iffalse
Tentative story
\begin{enumerate}
\item Vector Balancing for $\{\pm 1\}^d$ vectors has been greatly studied in both Offline and Online models.
\item For many applications, the input is dynamic and vectors can both arrive and depart. Cite works from Online Algos with Recourse to motivate the Recourse model. Existing algorithmic techniques for discrepancy cannot handle adaptive insertions and or (random) deletions.
\item The linear algeraic approach of Barany-Grinberg can be made to work using a Hierarchical Tree to get near-optimal discrepancy with $O(n)$ amortised recourse.
\item Another approach to try is Local Search. Unfortunately, it loses exponential in $d$, which is easy to get since there are only $2^d$ vectors (Remark: we can also get $d^d$ using local search).
\item Since $O(n)$ recourse due to Baranyi-Grinberg is a lot, especially for $s$ sparse vectors, a natural question is that can we do much better for $2$ sparse vectors? This is precisely the discrepancy problem on graphs. Linear Algebraic approach doesn't give anything better since the problem is still $n$ dimensional.
Local search can also be shown to be quite bad.
\item However, we observe the Local Search works well on Expanders. This brings us to our Expander Decomposition approach, which gets us near-optimal bounds for both discrepancy and recourse.
\end{enumerate}
\fi
\section{Fully-Dynamic Vector Balancing}
\label{sec:distrib-BG}
In this section, we prove~\Cref{thm:main1}. Given a set of vectors
$a_1, a_2, \ldots, a_T \in [-1,1]^n$, the \textsf{B\'{a}r\'{a}ny-Grinberg}\xspace algorithm signs
them such that the discrepancy of the signed sum is at most
$2n$. However, this signing is highly sensitive to insert or delete
operations. We address this issue by recursively dividing the input
sequence such that we lose only $O(n)$ discrepancy at each level of
this recursion tree---we call this the distributed \textsf{B\'{a}r\'{a}ny-Grinberg}\xspace
algorithm. We then show how it can easily handle insert and delete
operations with low recourse.
The main idea underlying the
\textsf{B\'{a}r\'{a}ny-Grinberg}\xspace algorithm~\cite{BaranyGrinberg81} is the following linear algebraic lemma.
\begin{lemma}[Rounding Lemma~\cite{BaranyGrinberg81}]
\label{lem:BG}
Let $a_1, a_2, \ldots, a_T \in [-1,1]^n$ be the columns of matrix
$A \in [-1,1]^{n \times T}$. For any initial fractional signing
$x \in [-1,1]^T$, there exists a (near-integral) signing $y$ with
all but $n$ variables being $\pm 1$ such that $A y = A
x$.
\end{lemma}
The signing $y$ is obtained by moving to a basic feasible solution
(BFS) of the following set of linear constraints
$ \{Ay = Ax, \; y \in [-1,1]^T\}$, where $x$ is treated as being
fixed. Based on \Cref{lem:BG}, B\'ar\'any and
Grinberg~\cite{BaranyGrinberg81} gave the following offline algorithm:
starting with the all-zeros vector as the fractional signing (i.e.,
$x = \bf{0}$), let $y$ be the almost-integral vector satisfying
$A y = 0$. Now
randomly rounding the fractional
variables (with bias given by the $y_i$ values) and using
concentration bounds shows a discrepancy of $O(\sqrt{n \log n})$, or
using sophisticated rounding schemes can give the tight $O(\sqrt{n})$
discrepancy~\cite{Spencer85,Bansal-FOCS10,Lovett-Meka-SICOMP15}.
\subsection{An Equivalent, Recursive Viewpoint}
\label{sec:dist-BG}
A natural question is: \emph{can we extend the above \textsf{B\'{a}r\'{a}ny-Grinberg}\xspace algorithm to
the dynamic case}? Naively using the rounding lemma does not work,
since the rounded solutions $y$ and $y'$ for matrices $A$ and $A'$
differing in one column could be very different. Our idea is to
simulate the \textsf{B\'{a}r\'{a}ny-Grinberg}\xspace algorithm in a distributed and recursive manner. We divide the sequence $\{1, \ldots, T\}$ into
sub-sequences of length $2n$ each, which gives us a set of $m:= T/2n$
sub-sequences (assume w.l.o.g., e.g., by padding, that $T/2n$ is a
power of 2). Let $P_1, \ldots, P_m$ denote these sub-sequences ordered
from left to right. We build a binary tree $\mathcal{T}$ of height
$\log_2 m$ on $m$ leaves, where leaf $j$ corresponds to the
sub-sequence $P_j$. Similarly, for an internal node $v$, define $P_v$
as the sub-sequence formed by taking the union of $P_j$ over all leaves
$j$ below $v$.
The signing algorithm ${\textproc{DBG}}\xspace(v)$, where $v$ is a node of $\mathcal{T}$ is
shown in~\Cref{algorithm:recursivebg}. It assigns values
$y_i^v \in [-1,1]$ to the vectors $a_i$ for $i \in P_v$ such that the
following two conditions are satisfied:
\begin{enumerate} [partopsep=0pt,topsep=0pt,parsep=0pt]
\item[(I1)] $\sum_{i \in P_v} y_i^v a_i = 0$, and
\item[(I2)] all but at most $n$
variables $y_i^v, i \in P_v$ are either $+1$ or $-1$.
\end{enumerate}
Applying this
property to the root node yields~\Cref{lem:BG}. While the end result
is identical to the one-shot \textsf{B\'{a}r\'{a}ny-Grinberg}\xspace algorithm, this yields some
crucial advantages in the dynamic setting. Indeed, when a vector is
inserted/deleted, only a single leaf's sub-sequence changes. We will
show that this leads to making changes in the signing assigned by the
ancestors of just this leaf, giving a total recourse of
$\widetilde{O}(n)$ per update!
For a subset $I$ of indices, let $A_I$ denote the submatrix of $A$ given by the columns corresponding to $I$. Similarly, for a vector $z$ indexed by $P_v$ and a subset $F$ of $P_v$, define $z|_F$ as the restriction of $z$ to $F$. The algorithm ${\textproc{DBG}}\xspace(v)$ begins by recursively assigning values to the two sub-sequences corresponding to its two children. Since these assignments, satisfy the two invariant conditions above, combining the two solutions into a new solution $x$ (in line~\ref{l:BGcombine}) leads to at most $2n$ fractional variables. Using~\Cref{lem:BG}, we reduce the number of fractional variables to $n$.
Finally, we can maintain a (integral) signing $\sigma$ by randomly assigning signs to the fractional variables $F_r$ at the root $r$ and retaining the values $y^r$ for rest of the vectors. We now show by induction that the two invariant properties are satisfied, the proof is deferred to~\Cref{sec:appsec2}.
\begin{restatable}{lemma}{dbginv}
\label{lem:dbginvariant}
The variables $y_i^v, i \in P_v$ satisfy the invariant properties~(I1) and (I2) at the end of ${\textproc{DBG}}\xspace(v)$.
\end{restatable}
\begin{algorithm}
\caption{Distributed-\textsf{B\'{a}r\'{a}ny-Grinberg}\xspace: ${\textproc{DBG}}\xspace(v)$}
\label{algorithm:recursivebg}
\textbf{Input:} A node $v$ of $\mathcal{T}$. \\
\textbf{Output:} $(y^v, F_v)$: an assignment $y^v_i \in [-1,1]$ for each $i \in P_v$, and $F_v \subseteq P_v$ is the index set of ``fractionally'' signed vectors, i.e., indices $i$ such that $-1 < y^v_i < 1$.
\begin{algorithmic}[1]
\If{$v$ is not a leaf}
\State Let $v_L$ and $v_R$ be the left and the right children of $v$ respectively.
\State $(y^{v_L}, F_{v_L}) \leftarrow {\textproc{DBG}}\xspace(v_L), (y^{v_R}, F_{v_R}) \leftarrow {\textproc{DBG}}\xspace(v_R).$
\State Define $F := F_{v_L} \cup F_{v_R}$, $x_i := y^{v_L}_i$ for all $i \in P_{v_L}, x_i := y^{v_R}_i$ for all $i \in P_{v_R}$. \label{l:BGcombine}
\Else
\State Define $F:= P_v, x_i = 0$ for all $i \in P_v$.
\EndIf
\State Using~\Cref{lem:BG} find a vector $y' \in [-1,1]^{|F|}$
such that (i) $A_{F} \cdot y'= A_F \cdot x|_F$, and (ii) there
are at most $n$ indices (denoted by the set $F_v \subseteq F$) having $-1 < y_i' < 1.$
\label{l:bg}
\State Define $y^v_i = x_i$ for $i \in P_v \setminus F$ and $y^v_i = y'_i$ for $i \in F$.
\State Return $(y^v, F_v)$.
\end{algorithmic}
\end{algorithm}
\subsection{Dealing with Update Operations}
Before describing the insert/delete operations, we describe a useful subroutine {\textproc{UpdateVector}}\xspace, which given an assignment $y^{old}$ to $a_1, \ldots, a_T$, updates it to a new assignment $y^{new}$ when one of the vectors $a_j$ in the sequence changes. The algorithm is very similar to~\Cref{algorithm:recursivebg}, but it needs to recurse on only one child of $v$. the one containing the index $j$.
As a result, the vectors $y^{old}$ and $y^{new}$ differ in at most $O(n\log T)$ coordinates. Details are deferred to~\Cref{sec:appsec2}.
\paragraph{Dynamic Insert and Delete.} We now discuss the algorithm when an insert or delete operation happens. The algorithm works in phases:
a new phase starts when the number of vectors becomes $2^\ell$ for some $\ell$, and ends when this quantity reaches $2^{\ell-1}$ or $2^{\ell+1}$. Whenever a new phase starts, we run ${\textproc{DBG}}\xspace$ algorithm to find an assignment $y$. During a phase, we always maintain exactly $2^{\ell+1}$ vectors -- this can be ensured at the beginning of this phase by padding with $2^\ell$ zero vectors. This ensures that the tree $\mathcal{T}$ does not change during a phase.
When a delete operation happens, we call {\textproc{DBGUpdate}}\xspace, where the deleted vector gets updated to the zero vector. Similarly, when an insert operation happens, we update one of the zero vectors to the inserted vector. Thus, we get the following result:
\begin{lemma}
The amortized recourse of this fully-dynamic algorithm is
$O(n \log N)$ per update operation, where $N$ is the maximum number
of active vectors at any point in time.
\end{lemma}
\begin{proof}
The work done at the beginning of a phase can be charged to the length of the input sequence at this time. This results in $O(1)$ amortized recourse.
We show in~\Cref{cor:BG} that the amortized recourse after each update during a phase is $O(n \log N)$. This proves the overall amortized recourse bound.
\end{proof}
Finally, since there are only $n$ fractional variables at the root,
we can use any state-of-the-art offline discrepancy minimization algorithm to
sign these vectors, e.g., to get $O(\sqrt{n})$ discrepancy for vectors
with unit $\ell_\infty$-norm \cite{Spencer85,Bansal-FOCS10,Lovett-Meka-SICOMP15}, or to get
$O(\sqrt{\log n})$ discrepancy for vectors with unit $\ell_2$-norm~\cite{BansalDG-FOCS16,BansalDGL18}.
This proves \Cref{thm:main1}.
\section{Fully-Dynamic Edge Orientation}
\label{sec:expand-decomp}
We next consider the case of dynamically orienting edges in a graph to
maintain bounded discrepancy. In this problem, at each time/update an
adaptive adversary either inserts a new edge $e_t = (u,v)$ or removes
an existing edge $e$ from a graph $G(t)$. Assigning an \emph{orientation} to each edge
$(u,v)$ as $u \rightarrow v$ or $v \rightarrow u$, the discrepancy of
a vertex $v$ is
${\mathsf{disc}}(v) = |~ |\delta_{\rm in}(v)| - |\delta_{\rm out}(v)| ~ |$,
where $\delta_{\rm in}(v)$ and $\delta_{\rm out}(v)$ are the sets of
in- and out-edges incident at $v$. Our goal is to minimize $\max_{v \in V} {\mathsf{disc}}(v)$. The algorithm is allowed to
re-orient any edge $e$, and the \emph{amortized recourse} is the
average number of re-orientations per edge insertion/deletion. We now
present the first fully-dynamic algorithms with ${\rm polylog}(n)$
discrepancy and recourse.
\paragraph{Useful Notation}
For an undirected graph $G$ and any set $S \subseteq V$, define $E(S)$ as
the set of edges whose endpoints are both in $S$; for sets $S, T$,
define $E(S,T) = \{ e \in E \mid |e \cap S| = |e \cap T| = 1\}$.
Define the \emph{volume} of a set $S$ to be
$\mathsf{vol}(S) = \sum_{v \in S} \deg(v)$.
\begin{definition}[$\phi$-Expander]
\label{def:expander}
A graph $G$ is a \emph{$\phi$-expander} if for all subsets $S\subseteq V$,
\[
|E(S,V-S)|\geq \phi \cdot \min(\mathsf{vol}(S),\mathsf{vol}(V-S)) \enspace .
\]
In this case, we also say the graph $G$ has \emph{conductance} at least $\phi$.
\end{definition}
\begin{definition}[$\gamma$-Weak-Regularity]
\label{def:weak-reg}
For $\gamma \in [0,1]$, an undirected graph $G$ is
\emph{$\gamma$-weakly-regular} if the minimum degree of any vertex is at
least $\gamma$ times the average degree $2m/n$.
\end{definition}
\subsection{High Level Overview} \label{sec:graph-overview}
We now provide a detailed overview of our algorithm, and then delve into the individual components.
A natural algorithm for the edge orientation problem is a {\em local search}
procedure: while there exists an edge $(u,v)$ currently oriented $u \rightarrow v$ such that
${\rm disc}(v) > {\rm disc}(u) + 2$, flip its orientation to $v \rightarrow u$.
Although locally optimal orientations
could have discrepancy $\Omega(n^{1/3})$ for general graphs (see an example in~\Cref{sec:local-search-graph-lb}), our first crucial result is that
they always have low discrepancy on \emph{expanders}.
\begin{theorem}
\label{theorem:local-search-expander}
Let $G(V,E)$ be a $\gamma$-weakly-regular $\phi$-expander. Then the
discrepancy of any solution produced by \textproc{Local-Search}\xspace is
$O\big(\frac{\log m}{\phi \gamma}\big)$.
\end{theorem}
The proof of this theorem appears in~\Cref{sec:localExpander}. In
order to apply our local search algorithm to arbitrary graphs, our
plan is to use the powerful idea of expander decompositions (see,
e.g.,~\cite{ST-STOC04,saranurak2019expander,bernstein2020fullydynamic}).
At a high level, such schemes decompose any graph $G$ into a disjoint union of expanders with each vertex appearing in a small number of them. For concreteness, we use the following result
from~\cite[Theorem~19]{GuptaKKS20}\footnote{For ease of
exposition, we use a result that runs in exponential time;
using approximate low-conductance cuts gives a polynomial runtime
with additional
logarithmic factors.}.
\begin{theorem}[Decomposition into Weakly-Regular Expanders]
\label{theorem:gkks20}
Any graph $G = (V,E)$ can be decomposed into an edge-disjoint union
of smaller graphs $G_1, G_2, \ldots,G_k$ such that: (a) each vertex
appears in at most $O(\log^2 n)$ many smaller graphs, and (b) each
of the smaller subgraphs $G_i$ is a $\phi/4$-weakly-regular
$\phi$-expander, where $\phi=\Theta(1/\log n)$.
\end{theorem}
In order to make this into a dynamic decomposition, our algorithm follows a natural idea of maintaining $\log m$ \emph{levels/scales}, and placing each edge of the current graph $G$ at one of these levels. We use $G_i$ to denote the
subgraph formed by the level-$i$ edges; crucially, we ensure that $G_i$ has at
most $2^i$ edges. For each level $i$, we maintain the expander decomposition of $G_i$ into $\bigcup_{j} \, G_{i,j}$ where $G_{i,j}$ represents the $j^{th}$
expander in this decomposition. Since each vertex appears in at most $\log^2 n$ expanders at every level, overall any vertex will appear in $O(\log^3 n)$
expanders. Hence, our goal is to maintain a low-discrepancy signing for each expander, with bounded number of re-orientations as the expander changes due to updates. Next we discuss how insertions are easier to handle, but deletions require several new ideas.
{\bf Insertions.} When edges are inserted into $G$, we insert it into $G_1$ (the lowest scale) and orient it arbitrarily. Whenever a level $i$ becomes full, i.e., $|G_i| > 2^i$, we remove all edges and add them to the higher level $i+1$, and \emph{recompute the expander decomposition using~\Cref{theorem:gkks20} from scratch for the graph consisting of all edges in this level}. We also recompute an optimal offline low-discrepancy discrepancy orientation for each expander\footnote{It is easy to optimally orient any graph in the offline setting: we consistently orient the edges of all cycles, to be left with a forest. We can then again orient all the maximal paths between pairs of leaves in a consistent manner, to end up with an orientation where every vertex has discrepancy in $\{-1,0,1\}$.}. Of course, we may need to cascade to higher levels if the next level also overflows. However, the total cost of all these edge reorientations can be easily charged to the recent arrivals that caused the overflow.
{\bf Deletions.} Our insertion procedure guarantees that an expander $G_{i,j}$ only observes deletions in its lifetime (before the expander decomposition at its level is recomputed). So when the adversary deletes an edge from $G$ (called a {\em primary deletion}), we can remove it from the corresponding expander $G_{i,j}$ it belongs to, and simply re-run local search from the current orientation if it continues to have expansion at least, say $\phi/6$. We can then bound the recourse by tracking the changes to the associated $\ell_2$ potential $\Phi$ for this graph. However, \emph{what do we do when $G_{i,j}$ ceases to be an expander?}
Our idea is to simply identify a cut of sparsity $< \phi/6$ and remove the smaller side $\Delta P$ from the graph $G_{i,j}$, and repeat if necessary.
This is called the $\textproc{Prune}\xspace$ procedure and we formally describe it in~\Cref{subsection:dynamic-expander-pruning}. The edges which are incident to $\Delta P$ are re-inserted into the system using the insertion algorithm. In~\Cref{theorem:pruning-ravi}, we bound the number of pruned edges (also called \emph{secondary deletions}) in terms of the number of actual adversarial edge deletions which caused the drop in expansion, and so we are able to amortize the recourse of re-inserting these pruned edges back into our algorithm.
\begin{theorem}
\label{theorem:pruning-ravi}
Let $G_0 = (V_0, E_0)$ be a $\phi$-expander with $m$ edges, $n$
vertices, and minimum degree $\delta$.
For a subset $S \subseteq V_0$, let $\mathsf{vol}_0(S)$ denote its initial volume
in $G_0$. There is an algorithm called \textproc{Prune}\xspace (described in~\Cref{subsection:dynamic-expander-pruning}), which for every adversarial deletion of any edge in $G_0$, outputs a (possibly empty) set of vertices $\Delta P$ to be pruned/removed which satisfies the following properties.
Let $P_t$ denote the aggregate set of vertices pruned over a sequence of $t$ adversarial deletions inside $G_0$, i.e., $V_t := V_0 \setminus P_t$ and $G_t$ is the graph with the undeleted edges of $E_0$ that are induced on $V_t$. Then, for each $1 \leq t \leq \phi^2 m/20$:
\begin{enumerate} [label=(\roman*),partopsep=0pt,topsep=0pt,parsep=0pt,itemsep=0pt]
\item $P_t \subseteq P_{t+1}$. \label{cond:PIncreases}
\item $G_t$ is a $\phi/6$-``strong expander'', i.e., for any subset
$A \subseteq V_t$,
$$ |E_t(A, V_t \setminus A)| \geq (\phi/6)\cdot
\min\big(\mathsf{vol}_0(A), \mathsf{vol}_0(V_t \setminus A) \big). $$ Hence
the minimum degree of a vertex in $V_t$ is at least
$\phi\delta/6$.
\item $\mathsf{vol}_0(P_t) \leq 6t/(5\phi)$. \label{cond:volP}
\end{enumerate}
\end{theorem}
Similar ideas have recently been used for dynamic graph algorithms, e.g. in~\cite{saranurak2019expander,bernstein2020fullydynamic}, but our algorithms and analyses are more direct since we are concerned only with the amortized recourse rather than the update time. However, new challenges appear due to our discrepancy minimization setting.
{\medskip \noindent \bf A `Potential' Problem. }
While the above procedure identifies the set of edges to prune, so that the residual graph remains an expander, we still need to maintain a low-discrepancy orientation on the expander as it undergoes deletions and prunings. Indeed, the above ideas essentially allow us to cleanly reduce the fully dynamic problem to the follow special case of only handling deletions on expanders: let $G = (V,E)$ be a $\phi/6$-expander, currently oriented according to local search. Then, suppose $e$ is an adversarial deletion, and suppose $\Delta P$ is the set of vertices to be removed as computed by the $\textproc{Prune}\xspace$ procedure. Then, how many flips would we need to end-up at a locally-optimal orientation on $G[V \setminus \Delta P]$, which we know has bounded discrepancy since $G[V \setminus \Delta P]$ is an expander? If we can bound this in terms of the number of edges incident to $\Delta P$, then we would be done, since these are precisely the number of secondary deletions, which are in turn bounded in terms of adversarial deletions.
A natural attempt is to simply re-run local search on $G[V \setminus \Delta P]$ starting from the current orientation. While this will converge to a low-discrepancy solution because $G[V \setminus \Delta P]$ is an expander, our recourse analysis proceeds by tracking a quadratic potential function, and this could increase a lot if we suddenly remove all edges incident to $\Delta P$ en masse. Removing the edges one by one is also also an issue as the intermediate graphs won't satisfy the desired expansion to argue both discrepancy as well as recourse (which indirectly depends on having good discrepancy bounds to control the potential). To resolve this issue, we craft a collection of ``fake'' intermediate graphs that interpolate between the graphs $G$ and $G[V \setminus \Delta P]$ which ensure that (i) all of them
have good expansion properties, and (ii) the potential change in moving from one to another is bounded. Our overall algorithm is to then repeatedly re-run local search after moving to each intermediate graph, until we end up with the final orientation on $G[V \setminus \Delta P]$.
We now formalize this in the following theorem, which bounds the recourse needed to move from a locally optimal orientation in $G[V]$ to one in $G[V \setminus \Delta P]$. Let $H$ denote any graph with a current orientation represented by $\dir{H}$. We then define the following potential
\[ \Phi(\dir{H}) := \sum_{v \in V(H)} {\mathsf{disc}}(v)^2 \, .\]
\begin{theorem} \label{thm:iterated-local-search}
Let $G_{t-1} = (V_{t-1}, E_{t-1})$ be a $\phi/6$-expander as maintained by our algorithm, and suppose the adversary deletes an edge $e_t \in E_{t-1}$. Moreover, suppose an associated set of vertices $\Delta P \subseteq V_{t-1}$ are pruned by $\textproc{Prune}\xspace$ to obtain the graph $G_{t} = (V_t, E_t)$ which is a $\phi/6$-expander, where $V_t = V_{t-1} \setminus \Delta P$ and $E_{t}$ is the subset of $E_{t-1} \setminus \{e_t\}$ induced on $V_t$.
Then, starting from a locally optimal orientation $\dir{G}_{t-1}$ we can compute a locally optimal orientation $\dir{G}_t$ by performing at most $L_t$ flips satisfying
$$ L_t \leq \left(\frac{ \log m}{\phi^2 \gamma} + \frac{\mathsf{vol}_0(\Delta P) \log m}{\phi^2 \gamma}\right) + \Phi(\dir{G}_{t-1}) - \Phi(\dir{G}_{t}) \, . $$
\end{theorem}
With this our algorithm description is complete. For the discrepancy analysis, note that our algorithm at all times maintains a locally optimal orientation in each expander at each level, and every vertex appears in at most ${\rm polylog}(n)$ expanders from~\Cref{theorem:gkks20}, giving us an overall discrepancy of ${\rm polylog}(n)$ by combining with~\Cref{theorem:local-search-expander}. For the recourse analysis, any time the insertion algorithm overflows and a rebuild happens in the higher level, we can charge the recourse to the adversarial insertions as well as re-insertions of the edges removed by $\textproc{Prune}\xspace$. The latter is in turn bounded in terms of the adversarial deletions by~\Cref{theorem:pruning-ravi}. Finally, we bound the total recourse within an expander, as parts of it are pruned out, for which we appeal to~\Cref{thm:iterated-local-search}. Since~\Cref{theorem:pruning-ravi}\,(iii) ensures that the total volume of all the sets which are pruned can be bounded in terms of $O(1/\phi)$ times the number of adversarial deletions, we get that the total number of flips done over a sequence of adversarial deletions in any expander is at most $\frac{\log m}{\phi^3 \gamma}$ times the number of adversarial deletions plus the potential $\Phi(\dir{G}_0)$ of the initial expander, which is small since we start with an optimal orientation where each vertex has discrepancy at most $1$ when the expander is formed.
\subsection{Local-Search for Weakly-Regular Expanders} \label{sec:localExpander}
In this section we prove \Cref{theorem:local-search-expander} that local search ensures low discrepancy on any weakly-regular expander.
Recall that the local search flips an edge $(u,v)$ oriented from $u$ to $v$ whenever ${\mathsf{disc}}(v) > {\mathsf{disc}}(u) + 2$.
\begin{algorithm}
\caption{\textproc{Local-Search}\xspace}
\label{algorithm:local}
\textbf{Input:} Graph $G=(V,E)$ and an initial partial orientation.\\
\textbf{Output:} Revised orientation which is a local optimum.
\begin{algorithmic}[1]
\State Arbitrarily direct any undirected edges in $G$.
\State While there exists a directed edge $(u,v)$ such
that flipping it decreases $\Phi:= \sum_u {\mathsf{disc}}(u)^2$, flip
it.
\end{algorithmic}
\end{algorithm}
\begin{proof}[Proof of~\Cref{theorem:local-search-expander}]
Let $\vec{G}=(V, \vec{E})$ be the directed graph
corresponding to a local optimum. Consider the node $v$ with
largest discrepancy $k$; without loss of generality, assume
$k \geq 0$. We perform a breadth-first-search (BFS) in $\vec{G}$
starting from $v$, but only following the \emph{incoming} edges at each step. Let
$L_i$ be the vertices at level $i$ during this BFS, i.e., $L_i$ is
the set of vertices $w$ for which the shortest path in $\vec{G}$ to
$v$ contains $i$ edges. Let $S_i$ denote the set of vertices up to
level $i$, i.e., $S_i := \bigcup_{i'=0}^i L_{i'}$. The fact
that $\vec{G}$ is a local optimum means there are no improving flips,
and hence the discrepancy of any vertex in $L_i$ is at least
$k-2i$. In turn, this implies that there are at least $k/2$ layers,
and the discrepancy of any vertex in $S_{k/4}$ is at least $k/2$.
We now show that the volume of $S_{k/4}$'s complement is large.
\begin{claim} \label{claim:largeVol}
$\mathsf{vol}(V \setminus S_{k/4}) \geq 2\gamma m/3$.
\end{claim}
\begin{proof}
Each node in $S_{k/4}$ has discrepancy at least $k/2$, and each
node in $V\setminus S_{k/4}$ has discrepancy at least $-k$. Since
the total discrepancy of all the vertices in $V$ is 0, it follows
that
\[
0 ~=~ \sum_{v \in V} {\mathsf{disc}}(v) ~\geq~ |S_{k/4}| \cdot k/2 - |V \setminus S_{k/4}| \cdot k.
\]
This implies that $|V \setminus S_{k/4}| \geq n/3.$ Now
$\gamma$-weak-regularity implies each vertex in $G$ has degree at
least $\gamma \cdot \frac{2m}{n}$, and hence the sum of the degrees
of the vertices in $V \setminus S_{k/4}$ is at least $2m
\gamma/3$.
\end{proof}
We now show that the size of the edge set $E(S_i)$ increases geometrically.
\begin{claim}
\label{cl:expinc}
For any $i \leq k/4$, $|E(S_{i+1})| \geq (1 + \phi \gamma/3) |E(S_i)|.$
\end{claim}
\begin{proof}
Given a directed graph $\vec{G}$ and a subset $X$ of vertices, let
$\delta^-(X)$ and $\delta^+(X)$ denote the set of incoming edges
into $X$ (from $V \setminus X$), and the set of outgoing edges from
$X$ (to $V \setminus X$) respectively. Since the discrepancy of each
vertex in $S_i$ is positive,
\[
0 ~\leq~ \sum_{w \in S_i} {\mathsf{disc}}(w) ~=~ \sum_{w \in S_i}
\big( |\delta^-(w)| - |\delta^+(w)| \big)
~=~|\delta^-(S_i)| - |\delta^+(S_i)|.
\]
The expansion property now implies that
\begin{align}
\label{eq:expansion1}
|\delta^-(S_i)| ~\geq~ \tfrac12 |\delta(S_i)| ~\geq~ \tfrac{\phi}{2}\cdot \min\big(\mathsf{vol}(S_i), \mathsf{vol}(V-S_i) \big).
\end{align}
Since all edges in $\delta^-(S_i)$ are directed from $L_{i+1}$ to $L_i$,
\begin{align*}
|E(S_{i+1})| ~ \geq~ |E(S_i)| + |\delta^-(S_i)|
\; &~\stackrel{\eqref{eq:expansion1}}{\geq} ~
|E(S_i)| + \frac{\phi}{2} \min\big(\mathsf{vol}(S_i), \mathsf{vol}(V-S_i) \big) \\
&~\geq~ |E(S_i)| + \frac{\phi}{2} \min\big(|E(S_i)|, 2\gamma m/3\big),
\end{align*}
where we used $\mathsf{vol}(S_i) \geq |E(S_i)|$ and \Cref{claim:largeVol}
for the two terms of the last inequality. Since $m \geq |E(S_i)|$ and
$\gamma \leq 1$, the RHS above is at least $ |E(S_i)| \cdot (1 + \phi \gamma/3). $
\end{proof}
Finally the fact that $S_0 = \{v\}$ implies that
$|E(S_1)| \geq |\delta^-(S_0)| \geq k$, and using \Cref{cl:expinc} we get
$$|E(S_{k/4})| \geq k \; \big( 1 + \nicefrac{\phi \gamma}{3} \big)^{k/4-1}. $$ Since $|E(S_{k/4})| \leq m$, we get
$k = O\big(\frac{\log m}{\phi\gamma}\big)$.
\end{proof}
\begin{figure}
\centering
\includegraphics{images/local-search/expander-local-search.pdf}
\caption{local-search for expanders.}
\label{fig:expanders-local-search}
\end{figure}
The weak-regularity property was used only in \Cref{claim:largeVol}
above; it is easy to alter the proof to show that
$\mathsf{vol}(V \setminus S_{k/4}) \geq \nicefrac{\gamma m}{4}$ even if all but $\nicefrac{n}{7}$
vertices in $G$ satisfy the weak-regularity property. This proves~\Cref{theorem:local-search-expander}.
\begin{corollary}
\label{cor:weakerlocalsearch}
Let $G$ be a $\phi$-expander, such that degree of every vertex,
except perhaps a subset of at most $\nicefrac{n}{7}$ vertices, is at least
$ 2\gamma m/n$. Then the discrepancy of \textproc{Local-Search}\xspace is
$O \big( \frac{\log m}{\phi \gamma} \big)$.
\end{corollary}
The expansion plays a crucial role here:
in \Cref{section-lower-bounds-local-search} we show that locally-optimum
solutions of \textproc{Local-Search}\xspace can have large discrepancy for general
graphs.
\subsection{Dynamic Expander Pruning}
\label{subsection:dynamic-expander-pruning}
In this section we prove~\Cref{theorem:pruning-ravi}. We first recall the setup. Suppose we start with a $\phi$-expander $G_0=(V_0, E_0)$, and at each
step $t$ some edge $e_t$ is deleted by the adversary (call these \emph{primary
deletions}). The goal is to remove a small portion of the graph so that the remaining portion continues to be, say, a $\phi/6$-expander.
Since the graph may may violate the expansion requirement due to deletions, we perform additional \emph{secondary
deletions} at each step to maintain a slightly smaller subgraph
$G_t = (V_t, E_t)$ which is $\nicefrac{\phi}{6}$-expanding, such that
$\{e_1, \ldots, e_t\} \cap E_t = \emptyset$. Crucially, the number of
secondary deletions is only a factor $\nicefrac1\phi$ more than the
number of primary deletions until that point. The idea of this greedy {\em expander pruning algorithm} is simple:
whenever the edge deletion $e_t$ creates a sparse cut in $G_{t-1}$, we
iteratively remove the smaller side of such a sparse cut from the
current graph until we regain expansion. (See~\Cref{algo:prune} for
the formal definition. Again we assume we can find low-conductance
cuts; using an approximation algorithm would lose logarithmic factors.)
\begin{algorithm}
\caption{$\textproc{Prune}\xspace(G_{t-1} , e_t)$}
\label{algo:prune}
\textbf{Input:} Graph $G_{t-1} = (V_{t-1},E_{t-1})$ and an edge $e_t \in E_{t-1}$ which gets deleted at this step.\\
\textbf{Output:} A set of vertices $\Delta P$ that have to be pruned from $G_{t-1}$ to get $G_t$.
\begin{algorithmic}[1]
\State Define $G_t$ with edges $E_t \gets E_{t-1} \setminus e_t$ and vertices $V_t \gets V_{t-1}$.
\State $\Delta P \gets \emptyset$.
\While{there is a subset $A \subseteq V_t$ with $|E_t(A, V_t \setminus A)| < (\phi/6) \cdot \mathsf{vol}_0(A)$}
\label{l:sparse}
\State Assume that $A$ is the smaller side of the cut
\State Remove $A$ from $G_t$, i.e., $V_{t} \leftarrow V_t \setminus A$, $E_{t} \leftarrow G_t[V_t \setminus A]$, and $\Delta P \leftarrow \Delta P \cup A$.
\EndWhile
\State Return $\Delta P$
\end{algorithmic}
\end{algorithm}
For a subset $S \subseteq V_0$, let $\mathsf{vol}_0(S)$ denote its initial volume
in $G_0$. Observe that the expansion in line~\ref{l:sparse} above is
measured with respect to $\mathsf{vol}_0(A)$, the volume of the set $A$ in
$G_0$, which is a stronger condition than comparing to the volume in
the current graph.
Define $P_t$ to be the set of vertices pruned in the first $t$
iterations, i.e., $P_t = V_0 \setminus V_{t}$, so that
$P_0 = \emptyset$. We now show that this algorithm maintains a
``strong expansion'' property at all times (i.e., the expansion
property holds with respect to the initial volume $\mathsf{vol}_0$).
\begin{proof}[Proof of~\Cref{theorem:pruning-ravi}]
The first property uses that $P_t$ is the set of vertices removed in
the first $t$ steps. The second property uses the stopping condition
of the algorithm, and the fact that $\mathsf{vol}_0(\{v\}) \geq \delta$ for
all vertices.
To prove the final property, let the subsets pruned by the algorithm
be $A^1, A^2, \ldots, A^s$ in the order they are pruned. For an
index $\ell$, let $B^\ell := \cup_{\ell'=1}^\ell A^{\ell'}$. Let
$E_t' := E_0 \setminus \{e_1, \ldots, e_{t} \}$. The following claim
bounds the number of edges leaving $B^\ell$.
\begin{claim} \label{claim:CutPEdges} If set $A^\ell$ is pruned in
iteration $t$, then
$|E_t'(B^\ell,V_0-B^\ell)| \leq \frac{\phi}{6}\mathsf{vol}_0(B^\ell)$.
Also, $\mathsf{vol}_0(B^\ell) \leq \mathsf{vol}_0(V_0 \setminus B^\ell)$.
\end{claim}
Before we prove this claim, we use it to prove the third property: note that
$ |E_0(B^\ell, V_0 \setminus B^\ell)| \geq \phi \cdot
\mathsf{vol}_0(B^\ell) $ because $G_0$ is a $\phi$-expander, and moreover
$\mathsf{vol}_0(B^\ell) \leq \mathsf{vol}_0(V_0 \setminus B^\ell)$ by the second
part of \Cref{claim:CutPEdges}. Moreover,
the first part of the claim implies that
$ |E_t'(B^\ell, V_0 \setminus B^\ell)| \leq
\frac{\phi}{6}\mathsf{vol}_0(B^\ell)$. Since $E_t'$ changes by at most
one edge per deletion, it follows that
\begin{gather}
t ~\geq~ |E_0(B^\ell, V_0 \setminus B^\ell)| - |E_t'(B^\ell, V_0
\setminus B^\ell)| \geq \phi \cdot \mathsf{vol}_0(B^\ell) -
\frac{\phi}{6}\cdot \mathsf{vol}_0(B^\ell) ~=~ \frac{5 \phi}{6}
\mathsf{vol}_0(B^\ell). \label{eq:1}
\end{gather}
Since $P_t$ is the same as $B^\ell$ for some $\ell$, this completes
the proof.
\end{proof}
\begin{proof}[Proof of \Cref{claim:CutPEdges}]
\begin{figure}
\centering
\includegraphics[width=4.5in]{images/local-search/prune.pdf}
\caption{Dynamic expander pruning.}
\label{fig:expander-pruning}
\end{figure}
We proceed by induction on $\ell$. When $\ell=0$ this is trivial
since $B^\ell = \emptyset$. Suppose the claim holds for
$B^{\ell-1}$, and we need to prove it for
$B^\ell = B^{\ell-1} \cup A^\ell$. For sake of brevity, we denote
$B' = B^{\ell - 1}, B = B^\ell$ and $A = A^\ell$ (see
\Cref{fig:expander-pruning}). Any edge in $E_{t}'(B, V_{0} \setminus B)$ either lies in
$E_t'(B', V_0 \setminus B) \subseteq E_t'(B', V_0 \setminus B')$ or in
$E_{t}'(A, V_0 \setminus B)$. By the induction hypothesis, the
former is at most $\frac{\phi}{6}\mathsf{vol}_0(B')$. The latter, by the
condition in line~\ref{l:sparse}, is at most
$(\phi/6) \cdot \mathsf{vol}_0(A)$. Summing the two,
we get $|E_{t}'(B, V_{0} \setminus B)| \leq (\phi/6) \mathsf{vol}_0(B') +
(\phi/6) \mathsf{vol}_0(A)$, which proves the first part of the claim.
To prove second part of the claim, it suffices to show
$\mathsf{vol}_0(B) \leq m$. By the induction hypothesis,
$\mathsf{vol}_0(B') \leq m$. Further, by~(\ref{eq:1}) we have
$\mathsf{vol}_0(B') \leq \frac{6t}{5 \phi}$. Now two cases arise:
\begin{enumerate} [partopsep=0pt,topsep=0pt,parsep=0pt]
\item $\mathsf{vol}_0(B')\geq \frac{\phi}{6}\mathsf{vol}_0(A)$: In this case,
\[ t~\stackrel{(\ref{eq:1})}{\geq}~ \frac{5\phi}{6}\mathsf{vol}_0(B')
~\geq~ \frac{5\phi}{12}\mathsf{vol}_0(B') + \frac{5\phi^2}{72}\mathsf{vol}_0(A)
~\geq~ \frac{5\phi^2}{72}\mathsf{vol}_0(B'+A) \; = \; \frac{5\phi^2}{72}
\mathsf{vol}_0(B) \enspace.
\]
Since $t \leq \nicefrac{\phi^2 m}{20},$ it follows that
$\mathsf{vol}_0(B) \leq m$.
\item $\mathsf{vol}_0(B')< \frac{\phi}{6}\mathsf{vol}_0(A)$: Consider the cut
$(A,V_0 \setminus B)$. Now,
\begin{align*}
|E_0(A,V_0 \setminus B)| &~=~ |E_0(A,V_0 \setminus
A)|-|E_0(A,B')| ~\geq~ \phi
\cdot \mathsf{vol}_0(A)- \mathsf{vol}_0(B') ~\geq~ \nicefrac56 \cdot \mathsf{vol}_0(A).
\end{align*}
On the other hand, $A$ is pruned by our algorithm because it is a
sparse cut, i.e.,
$|E'_t(A,V_0 \setminus B)| < \frac{\phi}{6} \cdot
\mathsf{vol}_0(A)$. Therefore, the total number of deletions $t$ is at
least
$\frac{5\phi}{6}\mathsf{vol}_0(A)-\frac{\phi}{6}\mathsf{vol}_0(A)=\frac{4\phi}{6}\mathsf{vol}_0(A)$. Now
we argue as in the first case. We get
\[ t~>~ \frac{3\phi}{6}\mathsf{vol}_0(A) + \frac{\phi}{6}\mathsf{vol}_0(A) ~>~
\frac{3\phi}{6}\mathsf{vol}_0(A) + \mathsf{vol}_0(B') ~ \geq ~ \frac{3\phi}{6}
\mathsf{vol}_0(B'+A) ~=~ \frac{\phi}{2} \mathsf{vol}_0(B) \enspace.
\]
Since $t \leq \nicefrac{\phi^2 m}{20}$, we again get $\mathsf{vol}_0(B) \leq m.$ \qedhere
\end{enumerate}
\end{proof}
\subsection{Dynamic Local-Search on Expanders with Deletions}
\label{subsection:local-search-pruning-expanders}
In this section we show how to dynamically maintain a locally-optimal orientation of an expander, as parts of it are pruned out over time, thereby proving~\Cref{thm:iterated-local-search}. The algorithm appears
as~\Cref{algorithm:prune-fake-recolor}.
We assume that the expander $G_{t-1}$ is maintained by dynamic pruning procedure $\textproc{Prune}\xspace$ and satisfies the expansion properties of~\Cref{theorem:pruning-ravi}. We also assume that we have a locally optimal orientation $\dir{G}_{t-1}$ inductively maintained by~\Cref{algorithm:prune-fake-recolor}. Then, when the adversary deletes an edge $e_t$ and $\textproc{Prune}\xspace$ computes a set $\Delta P$ of vertices to remove from $G_{t-1}$ to obtain a graph $G_t$, we show how to compute a locally optimal orientation $\dir{G}_t$ with a bounded number of flips.
Recall that we do this via a potential function argument. For any graph $H$ and current orientation $\dir{H}$, the potential of this orientation is $\Phi(\dir{H}) := \sum_{v \in V(H)} {\mathsf{disc}}(v)^2$. Indeed, the main issue is that there could be some vertices in $V_{t-1} \setminus \Delta P$ which are incident to many edges from $\Delta P$. Hence, if we remove $\Delta P$ in one shot, the potential of the residual graph could increase a lot. To resolve this, we replace $\Delta P$ by a set $F$ of an equal number $|\Delta P|$ of \emph{fake vertices}, and replace all the edges between $V_t := V_{t-1} \setminus \Delta P$ and $\Delta P$ with edges between $V_t$ and $F$ in a balanced round-robin manner to preserve the discrepancy of every vertex of $V_t$ w.r.t. its discrepancy in $\dir{G}_{t-1}$. Due to this balanced way of distributing the edges, we can show that the potential of the fake graph over $V_t \cup F$ is no more than that of $\dir{G}_{t-1}$, and moreover, even after deleting a subset $F' \subseteq F$ of fake vertices, $G_t \cup (F \setminus F')$ is an expander. These properties motivate running the following algorithm: transition from $G_{t} \cup F$ to $G_t$ by removing the fake vertices (and its incident edges) one-by-one, and re-running local search \emph{after each deletion}.
\begin{algorithm}[H]
\caption{\textproc{Prune-and-Reorient}$(\dir{G}_{t-1}, e_t, \Delta P)$}
\label{algorithm:prune-fake-recolor}
\textbf{Input:} Graph $G_{t-1}=(V_{t-1}, E_{t-1})$ with orientation $\dir{G}_{t-1}$, deleted edge
$e_t$, and pruned set $\Delta P \subseteq V_{t-1}$.\\
\textbf{Output:} A low-discrepancy orientation $\dir{G}_t$ for $G_t = (V_t, E_t)$ where $V_t := V_{t-1} \setminus \Delta P$ and $E_t$
is the subset of $E_{t-1} \setminus \{e_t\} $ induced on $V_t$.
\begin{algorithmic}[1]
\State Create fake vertices $F := \{f_1, \ldots , f_{N}\}$, where
$N=|\Delta P|$, and define $H_t := (V_t \cup F, E_t)$.
\State Let $E^+$ be edges in $E_{t-1}(V_{t}, \Delta P) \setminus \{e_t\}$
oriented from $V_{t}$ to $\Delta P$. Denote $E^+ = \{e_1, \ldots,
e_r\}$, such that all edges incident to a vertex in $V_{t}$ appear consecutively.
\For{$i=1, \ldots, r$}
\State For edge $e_i = (v_i, p_i) \in E^+$, add edge $(v_i, f_{(i
\bmod N)+1})$ into $H_t$ oriented from $v_i$ to $f_{(i \bmod N)+1}$.
\EndFor
\State Repeat above loop for edges $E^-$ in $E_{t-1}(V_t, \Delta P) \setminus \{e_t\}$
oriented into $V_t$; adds more edges to $H_t$.
\State Run \textproc{Local-Search}\xspace on $H_t$. \label{l:localsearchH}
\For{each $1 \leq j\leq N$}
\State Remove vertex $f_j$ and incident edges from $H_t$. \label{l2:remfake}
\State Run \textproc{Local-Search}\xspace on (the current graph) $H_t$. \label{l:localsearchj}
\EndFor
\State Define the final orientation $\dir{G}_t$ to be the final orientation $\dir{H}_{t}$ of $H_t$ (there are no fake vertices).
\end{algorithmic}
\end{algorithm}
\begin{proof}[Proof of~\Cref{thm:iterated-local-search}]
Let $H_t^j$ denote the graph $H_t$ after the
removal of the fake vertices $f_1, \ldots, f_{j}$, so that
$H_t^0 = H_t$. Since $t$ is fixed, we suppress the subscript $t$ for
the rest of this discussion.
We begin with some useful claims towards bounding the total number of flips $L_t$.
Firstly, we show that each of the intermediate graphs is a reasonable expander. The idea is that the pruned set $\Delta P$ is small compared to $V_{t}$, because whenever it becomes sufficiently large, the dynamic expander decomposition algorithm rebuilds the expander and we can charge the recourse to the adversarial deletions. As a result, both the number of fake vertices as well as their volume is substantially smaller than that of the ``real'' vertices $V_t$, and so the expansion properties of $V_t$ are approximately retained in each of the intermediate graphs $H^j$.
\begin{lemma}
\label{lemma:expansion-fake-vertices}
For each $j \in \{0, \ldots, N\}$, the graph $H^j$ is a $\phi/36$-expander.
\end{lemma}
\begin{proof}
Let $F^j := \{f_{j+1}, \ldots, f_{N} \}$ be a suffix of the fake
vertices. The vertex set of $H^j$ is $W^j := V_t \cup F^j$. Let
$E^j$ denote the set of edges in $H^j$---these are the union of
edges between $V_t$, i.e., $E_{t-1}(V_t)$, and those going between
$F^j$ and $V_t$. Let $S$ be a subset of vertices in $H^j$; we need
to show that
\begin{align}
\label{eq:expproof}
|E^j(S, W^j \setminus S)|~ \geq ~ \frac{\phi}{36} \min(\mathsf{vol}^j(S), \mathsf{vol}^j(W^j \setminus S)),
\end{align}
where $\mathsf{vol}^j$ denotes the volume with respect to $E^j$. Let
$\mathsf{vol}_t$ denote the volume with respect to the edges
$E_t = E_{t-1}(V_t)$. Recall that $\mathsf{vol}_0$ denotes the volume with
respect to $G_0$ (i.e., the expander graph before any deletions were
performed).
Without loss of generality, assume that
$\mathsf{vol}^j(S) \leq \mathsf{vol}^j(W^j \setminus S)$. Let $S_r$ and $S_f$
(``real'' and ``fake'') denote $S \cap A_t$ and $S \cap F^j$
respectively (see \Cref{fig:fake-graphs-expansion}).
\begin{figure}
\centering
\includegraphics{images/local-search/fake.pdf}
\caption{Expansion property of intermediate graphs $W^j$.}
\label{fig:fake-graphs-expansion}
\end{figure}
\begin{claim}
\label{cl:compl}
$\mathsf{vol}_t(S_r) \leq \frac{2}{2-\phi} \cdot \mathsf{vol}_t(V_t \setminus S_r)$.
\end{claim}
\begin{proof}
We know that $V_0 = P_t \cup V_t$. Since
$t \leq \phi^2 m/20$, \Cref{theorem:pruning-ravi}(iii)
shows that $\mathsf{vol}_0(P_t) \leq \frac{6}{5\phi} \cdot \frac{\phi^2 m}{20} \leq \phi m/4$. Therefore, $\mathsf{vol}^j(V_t)$ is at least $2\left(m - \frac{\phi m}{4}\right)=2m-\frac{\phi m}{2}$.
Now,
$$\mathsf{vol}^j(S_r) ~\leq~ \mathsf{vol}^j(S) ~\leq~ \mathsf{vol}^j(W^j \setminus S) ~\leq~ \mathsf{vol}^j(V_t \setminus S_r) + \mathsf{vol}_0(P_t). $$
Using $\mathsf{vol}_0(P_t) \leq \phi m/2$, this gives $\mathsf{vol}^j(S_r)-\mathsf{vol}^j(V_t\setminus S_r)\leq \phi m/2$. We also know that
$\mathsf{vol}^j(S_r) + \mathsf{vol}^j(V_t \setminus S_r) = \mathsf{vol}^j(V_t) \geq 2m -
\frac{\phi m}{2}$.
Eliminating $m$ from these two inequalities gives the claim.
\end{proof}
\Cref{theorem:pruning-ravi} shows that
$$|E^j(S_r, V_t \setminus S_r)| ~\geq~ \frac{\phi}{6}
\min(\mathsf{vol}_0(S_r),\mathsf{vol}_0(V_t \setminus S_r)) ~\geq~ \frac{\phi}{6} \cdot \frac{2-\phi}{2} \cdot \mathsf{vol}^j(S_r)~\geq~ \frac{\phi}{12} \mathsf{vol}^j(S_r),$$
where the second-last inequality follows from \Cref{cl:compl}, and the last inequality uses $\phi\leq 1$.
We now consider two cases:
\begin{enumerate} [partopsep=0pt,topsep=0pt,parsep=0pt]
\item $|E^j(S_f, V_t \setminus S_r)| \geq \mathsf{vol}^j(S_f)/2$: In this
case,
$$ |E^j(S, W^j \setminus S)| ~\geq~ |E^j(S_r, V_t \setminus S_r)| +
|E^j(S_f, V_t \setminus S_r)| ~\geq~ \frac{\phi}{12} \mathsf{vol}^j(S_r) +\frac{1}{2}\mathsf{vol}^j(S_f) ~\geq~ \frac{\phi}{12} \mathsf{vol}^j(S). $$
\item $|E^j(S_f, V_t \setminus S_r)| \leq \mathsf{vol}^j(S_f)/2$: This
implies $\mathsf{vol}^j(S_r) \geq \mathsf{vol}^j(S_f)/2$. Therefore,
$$ |E^j(S, W^j \setminus S)| ~\geq~ |E^j(S_r, V_t \setminus S_r)|
~\geq~ \frac{\phi}{12} \mathsf{vol}^j(S_r) ~=~ \frac{\phi}{36} (\mathsf{vol}^j(S_r) +
2\mathsf{vol}^j(S_r)) ~\geq~ \frac{\phi}{36} \mathsf{vol}^j(S).$$
\end{enumerate}
Hence the proof of \Cref{lemma:expansion-fake-vertices} follows.
\end{proof}
Next we show that \textproc{Local-Search}\xspace gives low discrepancy on the graphs $H^j$,
\emph{even though they may not be weakly-regular}.
\begin{lemma}
\label{lemma:discrepancy-fake-vertices}
For every $j \in \{0, \ldots, N\}$, the discrepancy of $H^j$ at a
local optimum is $O \big(\frac{\log m}{\phi^2 \gamma} \big)$.
\end{lemma}
\begin{proof}
We apply~\Cref{cor:weakerlocalsearch} to
$H^j$. \Cref{lemma:expansion-fake-vertices} implies that $H^j$ is
$\nicefrac{\phi}{36}$-expander, so it suffices to show that a large fraction of
the vertices of $H^j$ have large degree.
First of all, since we allow at most
$D = \nicefrac{\phi^2 m}{20}$ deletions, $\mathsf{vol}_0(P_t)\leq \frac{6}{5\phi}
\cdot \frac{\phi^2 m}{20}=\frac{3\phi m}{50}$ by \Cref{theorem:pruning-ravi}(iii). Using $\gamma$-weak-regularity of $G_0$ and $\gamma=\phi/4$ (\Cref{theorem:gkks20}), this implies $|P_t|\leq \frac{n}{2m\gamma}\cdot \frac{3\phi m}{50} = \frac{3n}{25}$. So $H^j$ has at least $22n/25$ vertices. It follows that $|P_t|\leq 3n'/22 \leq n'/7$, where $n'$ denotes the number of vertices in $H^j$. Now, \Cref{theorem:pruning-ravi}(ii) implies that the degree of any vertex belonging to set $V_t$ in the graph $G_t$
is at least $\phi \delta/6$, where $\delta \geq 2\gamma m/n$. (This
uses that $G_0$ is $\gamma$-weakly-regular). Thus, the degree of any
vertex in $V_t$ in the intermediate graph $H^j$ is also at least
$$ \frac{\phi\delta}{6} ~=~ \frac{2 \gamma \phi m}{6n} ~\geq~ \frac{22 \gamma \phi m'}{75n'} ~=~ \left(\frac{11 \gamma \phi}{75}\right)\cdot \frac{2m'}{n'}, $$
where $m',n'$ denote the number of edges and vertices in $H^j$
respectively (since $m' \leq m$ and $n' \geq 22 n/25$ as shown
above). The desired result now follows
from~\Cref{cor:weakerlocalsearch}.
\end{proof}
We are now ready to conduct the potential-based analysis for bounding the number of flips. We bound the recourse by studying the $\ell_2$-potential $ \Phi(\dir{H}^j) := \sum_{w \in V(H^j)} {\mathsf{disc}}(w)^2 $ as we transition from $G_{t-1}$ to $G_t$. Indeed, note that a flip made by local search decreases the potential by at least 1, so the recourse is at
most the \emph{total increase} in the potential. This increase happens
during~\Cref{algorithm:prune-fake-recolor} when we replace $\Delta P$
by $F$ to get the graph $H_t$ (with its resulting orientation), and
when we remove the fake vertex $f_j$ from $H_t^{j-1}$ to get $H_t^j$
(in line~\ref{l2:remfake}). We bound the potential increase during each
of these steps.
We give some notation first. Let
$\dir{G}_{t-1}, \dir{G}_t, \dir{H}_t^j$ be the subgraphs
$G_{t-1}, G_t, H_t^j$ oriented after running \textproc{Local-Search}\xspace respectively. Let $\dir{G}'_{t-1}$ denote $\dir{G}_{t-1} \setminus \{e_t\}$. Recall that $E^+$, $E^-$ (and therefore $H_t$) are defined using edges of $\dir{G}'_{t-1}$. Let
$\dirp{H_t}$ be the orientation of $H_t$ just after we replace
$\Delta P$ by $F$ in $\dir{G}'_{t-1}$ (i.e., before
line~\ref{l:localsearchH}). Similarly let $\dirp{H^j}$ (we again
suppress the subscript $t$ for ease of notation) be the orientation of
$H^j$ just after we remove $f_j$ but before we run \textproc{Local-Search}\xspace on it (in
line~\ref{l:localsearchj}).
Since edges are added in a round-robin manner between $V_t$ and $\Delta P$ in
$H_t$, there are no parallel edges.
\begin{claim}
\label{cl:potential0}
$\Phi(\dir{G}'_{t-1}) - \Phi(\dir{G}_{t-1})$ is at most $O\big( \frac{\log m}{\phi^2 \gamma}\big)$.
\end{claim}
\begin{proof}
Recall that $\dir{G}'_{t-1}$ is obtained by removing $e_t$ from $\dir{G}_{t-1}$. Let $d$ be the maximum discrepancy of a vertex in $\dir{G}_{t-1}$. \Cref{theorem:pruning-ravi} implies that $G_{t-1}$ is an $\Omega(\phi\gamma)$-weakly-regular $\Omega(\phi)$-expander, so \Cref{theorem:local-search-expander} implies that the discrepancy $d$ is $O \big( \frac{\log m}{\phi^2\gamma} \big).$ Hence the removal of $e_t$ from $\dir{G}_{t-1}$ can increase the potential by at most $2((d+1)^2 - d^2) = 2(2d+1) = O\big( \frac{\log m}{\phi^2 \gamma}\big)$, thus proving the claim.
\end{proof}
Next, we show that the potential cannot increase while going from $\dir{G}'_{t-1}$ to $\dirp{H_t}$. This uses the fact that we essentially re-distributed all the edges in $E^+$ and $E^-$ in a balanced round-robin manner.
\begin{claim}
\label{cl:potential1}
$\Phi(\dirp{H_t}) - \Phi(\dir{G}'_{t-1}) \leq 0$.
\end{claim}
\begin{proof}
For a given sum $s$ and variables satisfying $\sum_{i=1}^N x_i = s$, the optimal (w.r.t. $\ell_2$ norm) integer assignment of variables has $x_i \in \{\lfloor \frac{a_+ - a_-}{N} \rfloor, \lfloor \frac{a_+ - a_-}{N} \rfloor + 1\}$ for all $i$ and is unique up to permutations. For our problem, $s=a_+ - a_-$ and the $x_i$'s denote the discrepancies of the fake vertices. So it suffices to prove the following:
\begin{claim}
For each addition of an edge $(v,f)\in (E^+\cup E^-)$ in \Cref{algorithm:prune-fake-recolor}, $\exists\, d'$ such that just after the addition, $\{{\mathsf{disc}}(f')\,|\,f'\in F\}\subseteq \{d',d'+1\}$. In particular, after the addition of all edges, $\{{\mathsf{disc}}(f')\,|\,f'\in F\}\subseteq \{\lfloor \frac{a_+ - a_-}{N} \rfloor,\lfloor \frac{a_+ - a_-}{N} \rfloor+1\}$.
\end{claim}
\begin{proof}
Recall that we first add the edges in $E^+$. Since they are added in round robin fashion, the claim is trivially true up to this point. At this point, there will be some prefix of vertices $F'\subseteq F$ with discrepancy $d'+1$ and the rest have discrepancy $d'$. Now consider the addition of edges in $E^-$. If $|E^-|\leq |F'|$, then nodes in $F\setminus F'$ remain unchanged and the discrepancy of some nodes in $F'$ will become $d'$, thus still satisfying the desired property. If $|E^-| > |F'|$, then after $|F'|$ insertions, all nodes will have discrepancy $d'$, and after this point, discrepancies decrease by $1$ in a round-robin fashion, thus maintaining the desired property. In particular, after the insertion of all edges, we have $d'=\lfloor \frac{a_+ - a_-}{N} \rfloor$. This is because if $\frac{a_+ - a_-}{N}$ is integral, then all vertices in $F$ will have discrepancy $\frac{a_+-a_-}{N}=\lfloor \frac{a_+-a_-}{N} \rfloor$ and if $\frac{a_+ - a_-}{N}$ is non-integral, since it is the average discrepancy, it is a convex combination of $d'$ and $d'+1$, implying $d' = \lfloor \frac{a^- - a^+}{N} \rfloor$.
\end{proof}
As explained in the beginning of the proof, the claim immediately implies that $\Phi(\dirp{H_t}) \leq \Phi(\dir{G}'_{t-1})$.
\end{proof}
\begin{claim}
\label{cl:potential2}
For any $j \in \{1, \ldots, N\}$, if $\delta(f_j)$ is the degree of
$f_j$ in $H_t$, the potential change is
\[ \Phi(\dirp{H}^{j}) - \Phi(\dir{H}^{j-1}) ~\leq~ O\bigg(
\frac{\delta(f_j) \; \log m}{\phi^2 \gamma}\bigg). \]
\end{claim}
\begin{proof}
Let $d$ be the maximum discrepancy of a vertex in
$\dir{H}^{j-1}$. When we remove the fake vertex $f_j$ from
$\dir{H}^{j-1}$, the discrepancy of the neighbors of $f_j$ changes
by 1, and so the potential increases by at most
$2d \delta(f_j)$. \Cref{lemma:discrepancy-fake-vertices} shows that
$d$ is $O \big( \frac{ \log m}{\phi^2 \gamma} \big)$.
\end{proof}
\Cref{cl:potential0,cl:potential1,cl:potential2} show that the total increase in the
potential due to deletion of $e_t$, creation of $H_t$ and deletion of a fake vertices is at
most
$ \frac{ \log m}{\phi^2 \gamma} + \frac{\mathsf{vol}_0(\Delta P) \log
m}{\phi^2 \gamma}.$
If $L_t$ denotes the number of flips performed by \textproc{Local-Search}\xspace during
\textproc{Prune-and-Recolor}$(G_{t-1}, e_t)$, then
$$ \Phi(\dir{G}_t) - \Phi(\dir{G}_{t-1}) ~\leq~ \left(\frac{ \log m}{\phi^2 \gamma} + \frac{\mathsf{vol}_0(\Delta P) \log m}{\phi^2 \gamma}\right) - L_t. $$
This completes the proof of~\Cref{thm:iterated-local-search}.
\end{proof}
We end this section by using~\Cref{thm:iterated-local-search} in an aggregate sense, over a sequence of $t$ adversarial deletions.
\begin{theorem}
\label{theorem:local-search-recourse}
Let $G_0 = (V_0,E_0)$ be a $\gamma$-weakly-regular $\phi$-expander
with $m$ edges and $n$ vertices. Suppose at most
$D = \nicefrac{\phi^2 m}{20}$ edges are deleted adversarially. Then
for any $t \leq D$, the total number of edge flips performed by
\Cref{algorithm:prune-fake-recolor} during the first $t$ deletions
is at most $O\big(\frac{\log m}{\phi^3 \gamma}\cdot t + m\big)$.
\end{theorem}
\begin{proof}
The proof is to simply combine \Cref{theorem:pruning-ravi,thm:iterated-local-search} over the sequence of adversarial deletions. Indeed, we can use the facts that the total volume of the pruned set is at most $\mathsf{vol}_0(P_t) \leq 6t/(5\phi)$ along with $\Phi(\dir{G}_0)\leq m$ (optimal offline orientation of $G_0$ has discrepancy at most $n \leq m$), and $\Phi(\dir{G}_t)\geq 0$ to complete the proof.
\end{proof}
\subsection{Putting Everything Together}
\label{sec:overall}
We now formally describe our overall algorithms and analyses. To keep track of the internal states of the algorithms, we maintain an internal clock which is initialized at \emph{\text{moment}\xspace} $\tau=0$ (but eventually $\tau$ will exceed time $t$).
At any
\text{moment}\xspace $\tau$, we maintain a decomposition of the current graph
$\currG{\tau}$ into several subgraphs $\{\levelG{i}{\tau}\}_{i \geq 0}$, where $\levelG{i}{\tau}$ is the {\em level-$i$} subgraph of $\currG{\tau}$. These subgraphs maintain the following invariants:
\begin{enumerate} [partopsep=0pt,topsep=0pt,parsep=0pt]
\item[(I1)] For each \text{moment}\xspace $\tau$ and level $i$, the graph $\levelG{i}{\tau}$ has at most $2^i$ edges.
\item[(I2)] For every $\tau$ and $i$, subgraph $\levelG{i}{\tau}$ has
a {\em creation \text{moment}\xspace} $\tau_0$ which is at most $\tau$. Graph $\levelG{i}{\tau}$ is a
subgraph of $\levelG{i}{\tau_0}$, i.e., we only delete edges from this level between $\tau_0$ and $\tau$.
\item[(I3)] For each $\tau$ and $i$, we maintain a decomposition of
$\levelG{i}{\tau}$ into subgraphs $\expG{i}{j}{\tau}$ for $j \geq 1$, such that any vertex appears in at most $\log^2 n$ of these subgraphs. Moreover, if $\tau_0$ is the creation \text{moment}\xspace of $\levelG{i}{\tau},$ then $\expG{i}{j}{\tau_0}$ is $\gamma$-weakly-regular $\phi$-expander for all $j$, and $\expG{i}{j}{\tau}$ is a subgraph of $\expG{i}{j}{\tau_0}$ for all $j$.
\end{enumerate}
Although not mentioned explicitly in the invariants, the subgraph
$\expG{i}{j}{\tau}$ also has expansion and the weak-regularity
properties given by~\Cref{theorem:pruning-ravi}: in the notation of
this theorem, $\expG{i}{j}{\tau} = G_t$, where $G_0$ is the
corresponding subgraph at the creation \text{moment}\xspace $\tau_0$ of
$\levelG{i}{\tau}$.
\medskip\textbf{Edge Insertions.} We first consider the (easier) case of adversarial edge insertions.
The algorithm appears in \Cref{algorithm:insert}. We first insert the
edge $e$ into level-$1$. Whenever a level-$j$ subgraph overflows
(i.e., has more than $2^j$ edges), we empty this level and move all
the edges to the subsequent level. If this process stops at level $j$,
we build a new expander decomposition of the graph at this level using~\Cref{theorem:gkks20}, and also recompute an optimal offline low-discrepancy discrepancy orientation for each expander. As mentioned before, it is easy to optimally orient any graph in the offline setting: we consistently orient the edges of all cycles, to be left with a forest. We can then again orient all the maximal paths between pairs of leaves in a consistent manner, to end up with an orientation where every vertex has discrepancy in $\{-1,0,1\}$. Note that since it is the optimal discrepancy solution, it is also a locally optimal orientation.
\begin{algorithm}[H]
\caption{\textproc{Insert}$(e, \tau)$}
\label{algorithm:insert}
\textbf{Input:} Edge $e$ to be inserted in $\currG{\tau}$.\\
\textbf{Output:} Graph $\currG{\tau+1}$ with decomposition into levels.
\begin{algorithmic}[1]
\State Find the smallest $i$ such that $\{e\}\cup \levelG{1}{\tau}\cup \cdots \cup \levelG{i}{\tau}$ has at most $2^i$ edges. \label{l:i}
\State Set $\levelG{i'}{\tau+1} = \emptyset$ for $i' =1, \ldots, i-1$. \label{l:empty}
\State Set $\levelG{i'}{\tau+1} \leftarrow \levelG{i'}{\tau},$ for
all $i' > i$ and
$\levelG{i}{\tau+1} \leftarrow \{e\}\cup \levelG{1}{\tau}\cup \cdots \cup \levelG{i}{\tau}.$
\State Let the expander decomposition of $\levelG{i}{\tau+1}$ (using \Cref{theorem:gkks20})
return subgraphs $\expG{i}{j}{\tau+1}, j \geq 1$; define their
\emph{creation \text{moment}\xspace} to be $\tau+1$. \label{l:creation}
\State Find a discrepancy-at-most-$1$ orientation for each $\expG{i}{j}{\tau+1},
j \geq 1$. \label{l:disc-one}
\State $\tau\gets \tau+1$
\end{algorithmic}
\end{algorithm}
\medskip\textbf{Edge Deletions.} For the case of adversarial edge
deletions, when an edge $e$ is deleted from subgraph
$\expG{i}{j}{\tau}$, we first check if $\tfrac{\phi^2 m_{i,j}}{20}$
edges have been deleted from $\expG{i}{j}{\tau_0}$, where $\tau_0$ is
the creation \text{moment}\xspace of $\levelG{i}{\tau}$ and $m_{i,j}$ was the number
of edges in $\expG{i}{j}{\tau_0}$. If so, we remove the subgraph
$\expG{i}{j}{\tau}$ and re-insert these edges (these are called
\emph{internal inserts}). Otherwise, we run \Cref{algo:prune} on
$\expG{i}{j}{\tau}$ and edge $e$ to get subset $\Delta P$, and then
call~\Cref{algorithm:prune-fake-recolor} which removes $\Delta P$ (via
\emph{secondary deletes}) and reorients edges of
$\expG{i}{j}{\tau}$. Finally, the edges of $\Delta P$ are re-inserted
(causing more \emph{internal inserts}). The algorithm is shown
formally in \Cref{algorithm:delete}.
\begin{algorithm}
\caption{\textproc{Delete}$(e, \tau)$}
\label{algorithm:delete}
\textbf{Input:} Edge $e$ to be deleted from $\currG{\tau}$.\\
\textbf{Output:} Graph after deletion of edge $e$.
\begin{algorithmic}[1]
\State Find the level $i$ and index $j$ such that $e$ belongs to $\expG{i}{j}{\tau}.$
\State Let $\tau_0\gets$ creation \text{moment}\xspace of $\levelG{i}{\tau}$, and
$m_{i,j} \gets$ number of edges in $\expG{i}{j}{\tau_0}$.
\State Let $T_{i,j}$ be the set of $\tau' \in [\tau_0, \tau]$ at which an adversarial edge deletion happened in $\expG{i}{j}{\tau'}$.
\If{$|T_{i,j}| \geq \nicefrac{\phi^2 m_{i,j}}{20}$}
\State Remove all edges in $\expG{i}{j}{\tau}$, re-insert all
except $e$
one-by-one using \Cref{algorithm:insert} while incrementing $\tau$. \label{l:ins1}
\Else
\State $\Delta P \leftarrow \textproc{Prune}\xspace(\expG{i}{j}{\tau}, e)$ (see
\Cref{algo:prune}, where $t = |T_{i,j}|$, and $G_{t-1} = \expG{i}{j}{\tau}$).
\State $G_{i,j}(\tau + 1) \gets \textproc{Prune-and-Reorient}(\expG{i}{j}{\tau},e,
\Delta P$) (see \Cref{algorithm:prune-fake-recolor}) and $\tau\gets \tau + 1$. \label{l:prunerecolor}
\State Reinsert edges of $\Delta P$ except $e$ one-by-one while incrementing $\tau$
using \Cref{algorithm:insert}. \label{l:ins2}
\EndIf
\end{algorithmic}
\end{algorithm}
We are now ready to analyze the discrepancy of $\currG{\tau}$ for all
$\tau$, as well as the amortized recourse. We will prove the following quantitative version of \Cref{thm:main2}.
\begin{theorem}[Main Theorem: Graph Orientation]
\label{theorem:overall}
Suppose we start with the empty graph on $n$ vertices and it
undergoes adversarial edge insertions and deletions. There is an algorithm that maintains discrepancy of $O(\log^7 n)$ with an amortized
recourse of $O(\log^5 n)$ per update.
\end{theorem}
\begin{proof}
Firstly, we can assume w.l.o.g. that we never have parallel edges, as we can handle repetitions in the following black-box manner. Let $E$ denote the set of active edges and let $E'$ denote the set of active edges in the no-repetitions black-box. For the copies of an edge $e$ in $E$ (call it $T_e$), we will maintain the following invariants: (a) if $|T_e|$ is even, then $e \notin E'$ and half of them are signed $+1$ and, (b) if $|T_e|$ is odd, then $e\in E'$ and the $1$'s and $-1$'s in $T_e$ differ by at most one such that overall the signs add up to $\sigma'(e)$. These invariants ensure that the discrepancy for $E$ is equal to that for $E'$. To maintain these invariants:
\begin{enumerate}[partopsep=0pt,topsep=0pt,parsep=0pt]
\item If $|T_e|$ is even and $e$ is added/deleted in $E$, then call insert procedure with $e$ into $E'$ and for each edge $e'$ whose orientation changes in $E'$, you have to flip exactly one copy in $T_{e'}$ to satisfy the invariant.
\item If $|T_e|$ is odd and $e$ is added/deleted in $E$, then you have to re-orient at most one of edge in $T_e$ to ensure that $+1$'s and $-1$'s are equal in $T_e$. Then call delete procedure on $e$ from $E'$. Again, for each edge $e'$ flipped in $E'$, you have to flip exactly one copy in $T_{e'}$ to satisfy the invariant.
\end{enumerate}
For the rest of the proof we assume that there are no parallel edges.
We first bound the recourse of the algorithm.
There are two sources of recourse:
\begin{enumerate} [label=(\alph*),partopsep=0pt,topsep=0pt,parsep=0pt]
\item While performing a discrepancy-at-most-$1$ orientation
in~\Cref{l:disc-one} of \textproc{Insert} (\Cref{algorithm:insert}):
this could happen due to adversarial insert or {\em internal} inserts, i.e., due to lines~\ref{l:ins1} or~\ref{l:ins2} in~\Cref{algorithm:delete}.
\item During \textproc{Local-Search}\xspace performed inside procedure \textproc{Prune-and-Reorient}, called in line~\ref{l:prunerecolor} of~\Cref{algorithm:delete}.
\end{enumerate}
We first bound the number of calls to the \textproc{Insert} procedure. Let $T$ denote the total number of adversarial inserts and deletes.
\begin{claim}
\label{cl:inserttotal}
The total number of calls to \textproc{Insert} procedure is at most $\frac{2T}{\phi^2 \gamma}.$
\end{claim}
\begin{proof}
Clearly, the number of adversarial inserts is at most $T$. First consider the internal inserts caused by line~\ref{l:ins1} of \Cref{algorithm:delete}. These can be charged to the $\phi^2 \gamma m_{i,j}$ adversarial deletes in the set $T_{i,j}$. Therefore, the number of such calls to \textproc{Insert} procedure is at most $\frac{T}{\phi^2 \gamma}.$ Similarly, the number of inserts in line~\ref{l:ins2} of \Cref{algorithm:delete} is at most $\mathsf{vol}_0(\Delta P)$, the total number of such internal inserts corresponding to a fixed expander graph which undergoes $D'\leq D$ adversarial edge deletions is at most $\frac{6D'}{5 \phi}$ (by~\Cref{theorem:pruning-ravi}). Summing over all expanders, this quantity is at most $\frac{6T}{5 \phi}$. Thus, the overall number of calls to \textproc{Insert} is at most $\frac{2T}{\phi^2 \gamma}.$
\end{proof}
We now bound the recourse caused by rebuilding of levels due to insertions.
\begin{claim}
\label{cl:recourse1}
The total re-orientations due to \Cref{l:disc-one} of \textproc{Insert} is at most $\frac{4T \log m}{\phi^2 \gamma}$.
\end{claim}
\begin{proof}
For a fixed level ${i^\star}$, let $T'$ be the set of $\tau$ when we call \textproc{Insert} and the index $i$ selected in line~\ref{l:i} in \Cref{algorithm:insert} happens to be ${i^\star}$. For any such \text{moment}\xspace $\tau$, the total number of edges added to $\levelG{i}{\tau+1}$ is at most
$2^{i^\star}$ and at least $2^{{i^\star} -1}$. It follows that the number of re-orientations due to \Cref{l:disc-one} of \textproc{Insert} at this \text{moment}\xspace is also at most $2^{i^\star}$. Note that we empty the levels $1, \ldots, {i^\star}-1$ at this \text{moment}\xspace. Therefore, between any two consecutive \text{moments}\xspace in $T'$, we must have inserted at least $2^{{i^\star}-1}$ edges (these could be either adversarial or internal inserts). Thus, the total number of re-orientations due to \Cref{l:disc-one} of \textproc{Insert} for all \text{moments}\xspace in $T'$ is at most twice the total number of calls to \textproc{Insert}, which is at most $\frac{2T}{\phi^2 \gamma}$ (by~\Cref{cl:inserttotal}).
Since there are at most $\log m$ levels, the desired result follows.
\end{proof}
We will next bound the recourse due to local-search steps in the \textproc{Prune-and-Reorient} procedure.
\begin{claim}
\label{cl:recourse2}
The total number of re-orientations due to \textproc{Local-Search}\xspace called in line~\ref{l:localsearchH} of \Cref{algorithm:prune-fake-recolor} is $O \left( \frac{T \log n}{\gamma \phi^3} \right)$.
\end{claim}
\begin{proof}
Consider any particular expander graph $\expG{i}{j}{\tau_0}$ created at \text{moment}\xspace $\tau_0$. Consider the calls to the \textproc{Prune-and-Reorient} procedure where the deleted edge belongs to $\expG{i}{j}{\tau'}$ for some $\tau' \geq \tau_0$.
We know that $D' \leq D$, this inequality could be strict because we may remove all the level $i$ edges at some \text{moment}\xspace because of line~\ref{l:empty} in~\Cref{algorithm:insert}.~\Cref{theorem:local-search-recourse} shows that the total number of re-orientations due to \textproc{Local-Search}\xspace called in line~\ref{l:localsearchH} of \textproc{Prune-and-Reorient} procedure during these $D'$ \text{moments}\xspace is at most (a constant factor of)
$ \frac{D' \log m}{\gamma \phi^3} + m_{i,j}, $
where $m_{i,j}$ is the number of edges in $\expG{i}{j}{\tau}$. When we add the above for all expanders, the first term is at most
$\frac{T \log m}{\gamma \phi^3}$. The second term is the sum over all expanders that get created during the algorithm of the number of edges in the expander. Expanders are created in line~\ref{l:creation} of \Cref{algorithm:insert}, and so their total size can be bounded in the same manner as the argument used in the proof of~\Cref{cl:recourse1}. Hence this quantity is at most $\frac{4T \log m}{\phi^2 \gamma}$.
\end{proof}
Since we use $\phi,\gamma = \Theta(1/\log n)$, the above results show that the amortized recourse is
at most $O \left( \gamma^{-1}\phi^{-3} \log n \right)= O\left(\log n \cdot \log^3 n \cdot \log n\right) = O(\log^5 n)$. We now bound the discrepancy of any vertex.
\begin{claim}
The discrepancy of any vertex is bounded by $O(\log^7 n)$ at all times.
\end{claim}
\begin{proof}
From \Cref{lemma:discrepancy-fake-vertices}, discrepancy of a vertex in any expander is $O(\phi^{-2}\gamma^{-1}\log m)$, which is $O(\log^4 n)$ since we use $\gamma, \phi = \Theta(1/\log n)$. Since each vertex appears in at most $O(\log^3 n)$ expanders, we get that the discrepancy is bounded by $O(\log^7 n)$ at all times.
\end{proof}
This completes the proof of \Cref{theorem:overall}.
\end{proof}
\section{Lower Bounds for Local Search}
\label{section-lower-bounds-local-search}
In this section, we will show that for general vectors and
general graphs, typical $\ell_2$-potential local search procedures do not guarantee low
discrepancy.
\subsection{The $\ell_2$-Potential for General Vectors}
The $\ell_2$-potential for a signing $\{\varepsilon_i\}$ for vectors
$a_i$ is $\sum_j (\sum_i \varepsilon_i a_{ij})^2 = \| S \|^2$, where
$S := \sum_i \varepsilon_i a_i$. Hence we are at a local optimum if for each
$i$, the potential change $\|S-2\varepsilon_i a_i\|_2 -\|S\|_2$
due to flipping $a_i$ is non-negative. We will show that there exist locally optimal solutions on $T$ vectors with discrepancy $\Omega(\sqrt{T})$.
Consider the following set of vectors in 2 dimensions: $T/2$ vectors of the form $(1,1/\sqrt{T})$ and $T/2$ vectors of the form $(-1,1/\sqrt{T})$ with signing $\varepsilon = \mathbf{1}$. Adding these vectors, we get $S = (0,\sqrt{T})$. Now, for any vector $a$ (w.l.o.g. $a=(1,1/\sqrt{T})$) in the collection,
\[
||S-2a||_2^2 - ||S||_2^2 ~=~ ||(-2,\sqrt{T}-\frac{2}{\sqrt{T}})||_2^2-||(0,\sqrt{T})||_2^2 ~=~ 4 + \left(\sqrt{T}-\frac{2}{\sqrt{T}}\right)^2-T ~=~ \frac{4}{T} ~>~ 0.
\]
Hence, this is indeed a local optimum and has discrepancy $\sqrt{T}$.
\subsection{The $\ell_2$-Potential for $\{\pm 1\}$-Vectors}
The $\ell_2$-potential for a signing $\{\varepsilon_i\}$ for $\pm 1$-vectors
$a_i$ is $\sum_j (\sum_i \varepsilon_i a_{ij})^2 = \| S \|^2$, where
$S := \sum_i \varepsilon_i a_i$. Hence we are at a local optimum if for each
$i$, the potential change $\|S-2\varepsilon_i a_i\|_2 -\|S\|_2$
due to flipping $a_i$ is non-negative.
Since $\|a_i\|_2^2 = n$, the above condition is equivalent to showing
\begin{gather}
S^\intercal (\varepsilon_i a_i) \leq n \label{eq:l2-lopt}
\end{gather}
for all $i$. We can show
an $\Omega(2^{n/2})$ locality gap in this case.
\begin{lemma}
\label{lemma:lower-bound-vectors}
There is a family of instances of $\{\pm 1\}$ vectors, one for each
$n$ that is a multiple of $8$, having local optima with discrepancy $\Omega(2^{n/2})$ but
global optima having zero discrepancy.
\end{lemma}
\begin{proof}
We construct a $\{\pm 1\}$ matrix $M$ with $n$ columns and
$2\cdot \sum_{i=1}^{n/2} r_i$ rows, where $r_i$ is set later. We prove that
giving signs $\varepsilon=1$ to the rows of this matrix is a local
optimum with large discrepancy. Let $S = \pmb{\varepsilon}M$ denote the sum
of the rows of $M$. Our construction consists of \emph{repeating
units}, where the $i^{\text{th}}$ repeating unit (for $i=1,\ldots ,n/2$) is the following
$2\times n$ sub-matrix:
\[
\begin{bmatrix}
\begin{bmatrix}
-1 & -1 \\
-1 & -1
\end{bmatrix} \text{ repeated $i-1$ times},
\begin{bmatrix}
1 & 1 \\
1 & 1
\end{bmatrix},
\begin{bmatrix}
1 & -1 \\
-1 & 1
\end{bmatrix} \text{ repeated $\frac{n}{2} - i$ times}
\end{bmatrix}_{2\times n} .
\]
This unit is repeated $r_i$ times. We will later set $r_i$ to an even number, implying that any vector appears an even number of times. Therefore, by signing these even number of copies in an alternating fashion, we get that the global optimum has discrepancy $0$. By
construction, $S = (s_1,s_1,s_2,s_2,...,s_{n/2},s_{n/2})$ for some
integers $s_j$. Define $\vec{s} := (s_1,s_2, \ldots,s_{n/2})$.
\begin{claim}
\label{claim:pm1-lower-bound}
Let $B$ be the $\frac{n}{2}\times \frac{n}{2}$ lower-triangular
matrix with $1$s on the diagonal and $-1$s in the lower
triangle. Then using
$\vec{r} := (r_1,\ldots ,r_{n/2})^\intercal = \frac{n}{4}
(B^{-1})^\intercal B^{-1} \mathbf{1}$ results in $M$
whose row-sum $S = (s_1,s_1,s_2,s_2,\ldots ,s_{n/2},s_{n/2})$ satisfies
$\vec{s}:=(s_1,s_2,\ldots ,s_{n/2})^\intercal = 2B^\intercal \vec{r}$. Moreover,
$\epsilon_i = 1$ for all $i$ is a local optimum.
\end{claim}
\begin{proof}
The row sum $(s_1,s_1,s_2,s_2,\ldots ,s_{n/2},s_{n/2})$ satisfies:
\begin{align*}
2r_1 - \ldots -2 r_{n/2-1} - 2 r_{n/2}&=s_1\\
\vdots \\
2 r_{n/2-1} - 2 r_{n/2} &=s_{n/2-1}\\
2 r_{n/2} &=s_{n/2}
\end{align*}
which implies $\vec{s} = 2B^\intercal \vec{r}$. Next, we check the
condition~(\ref{eq:l2-lopt}) for local optimality: for any vector
$a$ in the $i^{\text{th}}$ repeating unit,
\[\langle S, a \rangle = -2s_1 -2s_2 -\ldots -2s_{i-1} + 2s_i = 2(B\vec{s})_i = 2(B\cdot 2B^\intercal \vec{r})_i = n. \qedhere
\]
\end{proof}
Putting the facts from \Cref{claim:pm1-lower-bound} together, the
discrepancy vector
$\vec{s} = 2B^\intercal \vec{r} = \frac{n}{2} B^{-1} \mathbf{1}$. We explicitly write down the
inverse of the lower-triangular matrix as follows:
\[
\begin{bmatrix}
1 & & & & \\
-1 & \ddots & & & \\
\vdots & \ddots & \ddots & & \\
\vdots & & \ddots & \ddots & \\
-1 & \cdots & \cdots & -1 & 1
\end{bmatrix}_{k\times k}^{-1}
=
\begin{bmatrix}
1 & & & & \\
1 & \ddots & & & \\
2 & \ddots & \ddots & & \\
\vdots & \ddots & \ddots & \ddots & \\
2^{k-2} & \cdots & 2 & 1 & 1
\end{bmatrix},
\]
Using this, we get that $\vec{s}$ has entries of value
$\Omega(2^{n/2})$, which proves \Cref{lemma:lower-bound-vectors}.
\end{proof}
\begin{remark}
Since our example has repetitions, we could try to use this
structure to assign opposite signs to these multiple copies, and
thereby get low discrepancy. However, it is easy to extend this to
avoid repetitions. Since
$(r_1,\ldots ,r_{n/2}) = \frac{n}{4} (B^{-1})^\intercal
B^{-1}\mathbf{1}$, the total number of rows
$R=\sum_i r_i$ in our original example is even and at most
$2^{4n}$. Take the original matrix $M$, append $2^{4n}-R$ rows of
$(1,-1,1,-1,\ldots )$ and $(-1,1,-1,1,\ldots )$ in alternation to
obtain a $2^{4n} \times n$ matrix $M'$. Moreover, append a
$2^{4n}\times 4n$ matrix of all possible $\{\pm 1\}^{4n}$ vectors to
the right of $M'$ to get the final $2^{4n} \times 5n$ matrix
$M''$. The row sum is now $S''=(S,\mathbf{0}^{4n})$, and any vector
in the first $R$ rows continues to satisfy
$\langle S'', a \rangle = n \leq 5n$ by construction. For any vector after
that, $\langle S'', a \rangle = 0 \leq 5n$, because
$ \langle (s_1,s_1,\ldots ,s_{n/2},s_{n/2}), (1,-1,1,-1,\ldots ) \rangle = 0$. So
$M''$ has rows in $\{\pm 1\}^{5n}$, and it is a local optimum with
discrepancy $\Omega(2^{n/2})$.
\end{remark}
\subsection{The $\ell_2$-Potential for General Graphs} \label{sec:local-search-graph-lb}
We saw that the local search with the $\ell_2$ potential was effective
on expanders: however, it fails for general graphs.
We now show instances with $n$ vertices and local optima having
discrepancy $\Omega(n^{1/3})$.
\begin{lemma}
\label{lemma:lower-bound-graphs}
There is an infinite family of graph instances with local optima for
the $\ell_2$-potential having discrepancy $\Omega(n^{1/3})$.
\end{lemma}
\begin{figure}[H]
\centering
\includegraphics{images/local-search/graph-local-search-lower-bound.pdf}
\caption{Lower bound of $\Omega(n^{1/3})$ discrepancy for local search on graphs.}
\label{fig:graphs-lower-bound-simple}
\end{figure}
\begin{proof}
For even integer $k = \Omega(n^{1/3})$, construct a layered digraph
with $k$ layers (see \Cref{fig:graphs-lower-bound-simple}). Denote
the vertices in layer $i$ by $L_i$. For every $u\in L_i$ and
$v\in L_{i+1}$, add a directed edge from $v$ to $u$. The number of
vertices $n_i$ in layer $L_i$ is chosen so that the root has
discrepancy $k$ and each node in $L_i$ has discrepancy $k-2i$. Since
a node $u\in L_i$ has incoming edges from every $v\in L_{i+1}$ and
outgoing edges to every $v'\in L_{i-1}$, it suffices to have
$n_{i+1}-n_{i-1} = {\mathsf{disc}}(u) = k-2i$. The base cases are $n_0 = 1$,
$n_1=k$. Since $k$ is even, this recurrence results in a symmetric
instance with a zero-discrepancy layer at the center. There are $k$
layers, and increase in size from $n_{i-1}$ to $n_{i+1}$ is at most
$k$. So the total number of nodes (up to constant factors) is at
most $k + 2k + \ldots + (k)k = (1+2+\ldots +k)k = O(k^3)$.
Finally, each node in $L_i$ has discrepancy equal to $k-2i$ by
construction. Since all edges (in the directed graph) are of the
form $u\rightarrow v$ with $u\in L_j, v\in L_{j-1}$ for some $j$, we
have that ${\mathsf{disc}}(v)-{\mathsf{disc}}(u) = (k-2(j-1))-(k-2j) = 2$, and hence the
orientation is indeed a local optimum.
\end{proof}
\section{Upper-Bounds for Local Search on Unstructured Graphs and Vectors}
\label{sec:local-search-upper}
In \Cref{sec:expand-decomp} we saw how local search with the
$\ell_2$-potential on expander graphs results in small discrepancy;
and \Cref{section-lower-bounds-local-search} gave strong lower bounds
for general graphs, and general collections of vectors. In this
section we give an upper bound for general graphs that matches the
lower bounds; for vectors, we get an exponential-in-$n$ (but
quantitatively weaker) upper bound.
\subsection{Local Search Upper Bound of $n^{O(n)}$ for $\{\pm 1\}$ vectors}
\label{sec:LB-general}
Let us consider a locally optimal configuration. Without loss of generality, $\varepsilon = 1$. For each participating vector $a$, using the fact that $\|a\|_2^2 =n$, we can rewrite the condition $\|S-2a\|_2\geq \|S\|_2$ as $S^T a\leq n$. So for the worst case we are interested in the following program.
\begin{align*}
\max_{a_1,\ldots ,a_T\in \{\pm 1\}^n, \,S=\sum_i a_i} \|S\|_\infty\\
\text{s.t. } \langle S, a_i \rangle \leq n \,\,\forall i\in[T].
\end{align*}
We can show that for $\{\pm 1\}$ vectors, discrepancy at a local optimum is bounded by a function of $n$, independent of $T$. Without loss of generality, $S$ has all positive coordinates. For any vector $u\geq 1$,
$$\|S\|_\infty \leq \langle S, u \rangle$$
Suppose there is such a $u$ of the form $\sum x_i a_i$ with $x_i \geq 0$, then we will get $\|S\|_\infty\leq n\cdot(\sum x_i)$ since $\langle S, a_i \rangle \leq n$. Clearly $x_i = 1$ is a feasible solution, but instead let us optimize it as follows.
\begin{alignat*}{2}
\min &\textstyle \sum_{i=1}^T x_i &&\\
x_i &\geq 0 &\qquad\qquad&\text{for $i=1,\ldots ,T$ ($T$ constraints)}\\
\textstyle \sum_{i=1}^T x_i a_i &\geq 1 &&\text{($n$ constraints)}
\end{alignat*}
Let $x^*$ be a corner in the feasible region. By definition of a
corner, we will have $T$ linearly independent tight constraints at
$x^*$. Let $n'\leq n$ of them be of the second kind and $T-n'$ of
the first kind. That is, there are $n'$ non-zero $x^*_i$'s. Then
we will have $\sum_{i:x^*_i\neq 0} x^*_i a'_i = 1$, where $a'_i$
is obtained from $a_i$ by retaining only the $n'$ coordinates
corresponding to tight constraints. Arranging these
$n'$-dimensional vectors as columns of a matrix, we get a
$n'\times n'$ matrix $V$ with $V\beta = 1$, where $\beta$ is a
$n'$ dimensional vector containing the nonzero $x^*_i$'s. $V$ is
full rank since if the rows had a non-trivial linear combination
giving zero, then using those coefficients for the $n'$ type-2
tight constraints will give a vector that is non-zero only at
positions $i$ where $x^*_i = 0$. So it is a linear combination of
the $T-n'$ type 1 tight constraints of the form $x_i = 0$. This
contradicts the linear independence of the tight constraints at a
vertex. Hence $V$ is full rank, which implies $\beta = V^{-1}\mathbf{1}\xspace$.
\begin{claim}
\label{cl:beta}
The entries of $\beta$ are bounded by $n^{O(n)}$.
\end{claim}
\begin{proof}
For a $k\times k$ matrix $M$, $\det(A)=\sum_{\pi\in S_k}sign(\pi)m_{1,\pi(1)}m_{2,\pi(2)}...m_{k,\pi(k)}$.
Since there are at most $k^k$ permutations, any $k\times k$ matrix with $\pm 1$ entries has determinant at most $k^k$.
We know that $V^{-1} = \frac{adj(V)}{\det(V)}$ and each entry of $adj(V)$ is the determinant of a $(n'-1)\times (n'-1)$ submatrix of $V$ (i.e., by removing a row and column).
Also since $V$ is an invertible $\pm 1$ matrix, $|\det(V)|\geq 1$. Hence we have that entries of $V^{-1}$ are bounded by $(n'-1)^{n'-1} = O(n^n)$, and since $\beta = V^{-1}\mathbf{1}\xspace$, the entries of $\beta$ are also bounded by $n^{O(n)}$.
\end{proof}
Now recall that $\|S\|_\infty \leq n\cdot(\sum_{i=1}^T \alpha_i) = n \cdot (\sum_{j=1}^{n'} \beta_j)$ since $\beta$ is a $n'$ dimensional vector containing the non-zero $x_i$'s. \Cref{cl:beta} implies that this quantity is at most $n\cdot(\sum_{j=1}^{n'} n^{O(n)}) \leq n\cdot (n'\cdot n^{O(n)}) = n^{O(n)}$.
\subsection{Local Search for General Graphs: Upper Bounds}
\label{subsection:graphs-L-local-search}
In this section, we will show upper bounds for a simple variant of
\textproc{Local-Search}\xspace involving flips along directed paths instead of single
edges. We will refer to it by \textproc{Path-Local-Search}\xspace with parameter $L$
(\Cref{algorithm:path-local}). This is simpler than the method
involving expander decompositions (\Cref{theorem:overall}), but does
not guarantee logarithmic bounds. (In the next section, we show our analysis is tight.)
\begin{algorithm}
\caption{\textproc{Path-Local-Search}\xspace}
\label{algorithm:path-local}
\textbf{Input:} Graph $G=(V,E)$ and an initial partial coloring, Parameter $L$.\\
\textbf{Output:} Revised orientation which is a local optimum.
\begin{algorithmic}[1]
\State Arbitrarily direct any undirected edges in $G$.
\State While there exists a directed path $(u_0,\ldots ,u_l)$ with $l\leq L$ such that ${\mathsf{disc}}(u_l) > {\mathsf{disc}}(u_0) + 2$, flip all the edges in the directed path.
\end{algorithmic}
\end{algorithm}
\begin{figure}
\centering
\includegraphics{images/local-search/path-local-search.pdf}
\caption{Discrepancy of a local optimum of \textproc{Path-Local-Search}\xspace}
\label{fig:graphs-upper-bound}
\end{figure}
\begin{theorem}
\label{theorem:local-search-L-tradeoff}
Suppose we start with the empty graph on $n$ vertices and it undergoes adversarial edge insertions and deletions such that at any point, the graph does not have multiple edges. Then there is a deterministic algorithm that achieves $O(D)$ discrepancy with $O(\sqrt{n/D})$ amortized recourse for any $\Omega(1)\leq D \leq O(n)$.
\end{theorem}
\begin{proof}
Recall that \textproc{Path-Local-Search}\xspace uses the following rule: if there is a directed path $a\rightarrow b$ of length $\leq L$ such that ${\mathsf{disc}}(b)>{\mathsf{disc}}(a)+2$, then flip all the edges in the path. Let us bound the discrepancy at a local optimum.
Let $\vec{G}=(V, \vec{E})$ be the directed graph
corresponding to a local optimum. Consider the node $v$ with
largest discrepancy $k$; without loss of generality, assume
$k \geq 0$. We perform BFS in $\vec{G}$
starting from $v$, following \emph{incoming} edges at each step. Let
$L_i$ be the vertices at level $i$ during this BFS, i.e., $L_i$ is
the set of vertices $w$ for which the shortest path in $\vec{G}$ to
$v$ contains $i$ edges. Let $S_i$ denote the set of vertices in the
first $i$ layers, i.e., $S_i := \bigcup_{i'=0}^i L_{i'}$.
The fact
that $\vec{G}$ is a local optimum means there are no improving flips,
and hence the discrepancy of any vertex in $L_i$ is at least
$k-2\left(\lfloor \frac{i-1}{L}\rfloor + 1\right)$ (see \Cref{fig:graphs-upper-bound}).
In turn, this implies that there are at least $kL/2$ layers,
and the discrepancy of any vertex in $S_{kL/4}$ is at least $k/2$.
We now prove the following claim about the rate at which the number of nodes grows.
\begin{claim}
\label{claim:L-local-search-nodes}
Let $n_i$ be the number of nodes in layer $i$ for $0\leq i\leq \frac{kL}{4}$, then $|S_i| = \sum_{j=0}^i n_j \geq \frac{i^2 k}{8}$.
\end{claim}
\begin{proof}
By induction. Base case: Trivial since $|S_0| = 1$.
Induction step: Given a directed graph $\vec{G}$ and a subset $X$ of vertices, let $\delta^-(X)$ denote the set of incoming edges
into $X$ (from $V \setminus X$). Since there are no repeated edges and all edges in $\delta^-(S_i)$ are directed from $L_{i+1}$ to $L_i$, we have $|\delta^-(S_i)| \leq n_i n_{i+1}$.
Since all vertices in $S_i$ have discrepancy at least $k/2$, the number of incoming edges $|\delta^-(S_i)|$ must be at least $\frac{k}{2} \cdot |S_{i}|$. Combining the two inequalities, we have
$n_i n_{i+1}\geq \frac{k}{2} \cdot |S_{i}|.$
Now using this this with the definition of $S_{i+1}$,
\begin{align*}
|S_{i+1}| ~=~ |S_{i-1}|+n_i+n_{i+1}
~&\geq \frac{(i-1)^2 k}{8} + 2\sqrt{n_i n_{i+1}} \quad \text{(induction hypothesis and AM-GM inequality)}\\
&\geq \frac{(i-1)^2 k}{8} + 2\sqrt{\frac{k}{2}\cdot|S_i|} \\
&\geq \frac{(i-1)^2 k}{8} + 2\sqrt{\frac{k}{2}\cdot\frac{i^2 k}{8}} \quad \text{(induction hypothesis)}\\
&\geq \frac{(i-1)^2 k}{8} + \frac{ik}{2}
~=~ \frac{(i+1)^2 k}{8} \enspace,
\end{align*}
which completes the proof of \Cref{claim:L-local-search-nodes}.
\end{proof}
Now, \Cref{claim:L-local-search-nodes} implies $|S_{kL/4}|\geq (kL/4)^2 k/8 = k^3 L^2/128$. Since we also have $|S_{kL/4}|\leq n$, we get $k\leq O(n^{1/3}/L^{2/3})$.
To get the amortized recourse, let us track the $l_2$ potential $\Phi = \sum_u {\mathsf{disc}}(u)^2$, which is initially zero. Before each insertion/deletion, we are at a local optimum, which has a discrepancy of at most $n^{1/3}/L^{2/3}$. So $\Phi$ can change by at most $O(n^{1/3}/L^{2/3})$ when you insert/delete. In each step of local search, $\Phi$ decreases by at least $1$. Since $\Phi \geq 0$, the total number of local search steps is at most $T \cdot \frac{n^{1/3}}{L^{2/3}}$. Since a local search step flips at most $L$ edges, the total recourse is at most $L$ times the number of local search steps. This implies that the amortized recourse is at most $ \frac{n^{1/3}}{L^{2/3}} \cdot L = n^{1/3}L^{1/3}$. This proves \Cref{theorem:local-search-L-tradeoff} in the range $\Omega(1)\leq D\leq O(n^{1/3})$.
For the range $\Omega(n^{1/3}) \leq D \leq n$, we use the following lemma.
\begin{lemma}
\label{lemma:delta-local-search}
Perform local search with $L=1$ but perform a step when ${\mathsf{disc}}(v)-{\mathsf{disc}}(u)>\delta$. Then a local optimum has discrepancy at most $n^{1/3}\delta^{2/3}$.
\end{lemma}
\begin{proof}
Very similar to the previous analysis. There will now be $k/2\delta$
layers, which will imply that total number of vertices is at least
$(k/2\delta)^2 k/8 = \Omega(k^3/\delta^2)$. Since this must be at most
$n$, this implies that the discrepancy $k=O(n^{1/3}\delta^{2/3})$. The
potential drops by at least $\delta$ in each local search step, so amortized recourse will be at most $O(n^{1/3}\delta^{2/3}/\delta) = O(n^{1/3}/\delta^{1/3})$.
\end{proof}
This completes the proof of \Cref{theorem:local-search-L-tradeoff}.
\end{proof}
As an aside, setting $L=1$ and $\delta = 2$ in
\Cref{lemma:delta-local-search} shows the local search in
\Cref{sec:expand-decomp} achieves $O(n^{1/3})$ discrepancy when
performed on general graphs instead of on expanders; this matches the
lower bound in~\Cref{lemma:lower-bound-graphs}.
\subsection{Tightness of discrepancy bound of \textproc{Path-Local-Search}\xspace}
\begin{figure}
\centering
\includegraphics{images/local-search/graph-path-local-search-lower-bound.pdf}
\caption{Example showing discrepancy bound at a local optimum of \textproc{Path-Local-Search}\xspace in \Cref{subsection:graphs-L-local-search} is tight.}
\label{fig:graphs-lower-bound}
\end{figure}
We now provide a family of instances to show that the discrepancy
bound for \textproc{Path-Local-Search}\xspace given in \Cref{subsection:graphs-L-local-search}
is tight. Note that \textproc{Path-Local-Search}\xspace with $L=1$ is the same as \textproc{Local-Search}\xspace, and
so using $L=1$ in the lemma below reproduces the $\Omega(n^{1/3})$
lower bound of \Cref{lemma:lower-bound-graphs}.
\begin{lemma}
For \textproc{Path-Local-Search}\xspace with parameter $L$, there is a family of instances of increasing size that are each a local optima and have discrepancy $k$ and number of nodes $O(k^3 L^2)$.
\end{lemma}
\begin{proof}
We will construct a layered directed graph with $O(kL)$ layers (see \Cref{fig:graphs-lower-bound}). The instance will be symmetric and we will use even $k$ and odd $L$. Let us denote layer $i$ by $L_i$. For every $u\in L_i$, $v\in L_{i+1}$, there is a directed edge from $v$ to $u$. We will set up the number of vertices $n_i$ in each layer so that the root has discrepancy $k$ and each node in $L_i$ has discrepancy $k-2\left(\lfloor \frac{i-1}{L}\rfloor + 1\right)$. Let us look at a node $u\in L_i$. It has incoming edges from every $v\in L_{i+1}$ and has outgoing edges to every $v'\in L_{i-1}$. So it suffices to have $n_{i+1}-n_{i-1} = {\mathsf{disc}}(u) = k-2\left(\lfloor \frac{i-1}{L}\rfloor + 1\right)$. The base cases are $n_0 = 1$, $n_1=k$. It is not hard to see that since $k$ is even and $L$ is odd, this recurrence will result in a symmetric instance with an odd number of zero discrepancy layers in the center.
There are $O(kL)$ layers and from $n_{i-1}$ to $n_{i+1}$, the increase is at most $k$. So the total number of nodes (up to constant factors) is at most $k + 2k + \ldots + (kL)k = (1+2+\ldots +kL)k = O(k^3 L^2)$.
\end{proof}
One can get tightness for the other side of the tradeoff (\Cref{lemma:delta-local-search}) by the same idea.
\subsection{Local Search for Forests}
If we are promised that at each time $t$ the underlying graph is a forest (in the undirected sense), then we can obtain constant discrepancy with logarithmic recourse using the variant of local search that flips along directed paths.
\begin{figure}
\centering
\includegraphics{images/local-search/trees.pdf}
\caption{Discrepancy of a local optimum in the forest case}
\label{fig:trees}
\end{figure}
\begin{lemma}
For forests, performing \textproc{Path-Local-Search}\xspace with $L=\log n$ gives $O(1)$ discrepancy and $O(\log n)$ amortized recourse.
\end{lemma}
\begin{proof}
First, we bound the discrepancy at a local optimum by $3$. For this, suppose for contradiction, the discrepancy (wlog it is positive) is at least 4. Consider the node with discrepancy at least 4 (see \Cref{fig:trees}). It will have a at least 4 incoming edges. Due to local optimum condition with $L=\log n$, each of these children will have discrepancy at least $2$, and therefore will have at least $2$ children. Since $L=\log n$ and forests do not have cycles, we can continue this binary-tree like growth argument till $\log n$ layers. Now the number of nodes will be at least $2^{\log n + 1} - 1 = 2n-1$, which is a contradiction. So we get $O(1)$ discrepancy and by the potential based argument as in previous proofs (e.g., \Cref{theorem:local-search-L-tradeoff}), we get $O(\log n)$ amortized recourse.
\end{proof}
\section{Conclusions and Insert-Only Algorithms} \label{sec:insert-only}
In this paper we initiate the study of fully dynamic discrepancy problems, where vectors/edges can both arrive
or depart at each time step, and the algorithm must always maintain a low-discrepancy signing. Prior algorithms for online discrepancy could only handle arrivals, and that too only for an oblivious adversary. We obtain near-optimal discrepancy bounds for both the edge-orientation and the vector balancing cases. We achieve the former in near-optimal $\tilde{O}(1)$ amortized recourse, and the latter in $\title{O}(n)$ amortized recourse (which is exponentially better in $T$ than the naive algorithm which recolors after each update).
The main open question left by our work is whether we can achieve near-optimal discrepancy for vector balancing in $\tilde{O}(1)$ amortized recourse per update. If true, this would imply our edge-orientation results as a special case.
\begin{open} \label{openPb:FullDyn}
For fully-dynamic vector balancing with vectors of $\ell_2$ length at most $1$ arriving or departing, can we achieve $\operatorname{poly}\!\log(nT)$ discrepancy in $\operatorname{poly}\!\log(nT)$ amortized recourse per update?
\end{open}
Currently, we don't know how to achieve near-optimal discrepancy even using $o(n)$ amortized recourse. However, in the case of insertions only (i.e., no departures), we can answer the above question affirmatively using the following \textproc{Dyadic Resigning}\xspace algorithm. Note that prior works, e.g.~\cite{ALS-STOC21,LSS-arXiv21}, could only achieve this against an oblivious adversary.
The \textproc{Dyadic Resigning}\xspace algorithm is simple: when the $t^{th}$ vector arrives,
let $\ell$ be the largest power of $2$ that divides $t$. Use any
offline algorithm $\ensuremath{\mathcal{A}}$ to construct a fresh signing of the most
recently arrived $2^\ell$ vectors.
\begin{theorem}
For any sequence of $T$ adaptive inserts, suppose that an offline algorithm
$\ensuremath{\mathcal{A}}$ can produce a signing of discrepancy at most $D$ when given
any subset of these vectors. Then the \textproc{Dyadic Resigning}\xspace algorithm achieves
at each timestep $t \leq T$ discrepancy at most
$D \lceil\log_2 t \rceil$. Moreover, each vector is assigned a new
sign at most $\lceil \log_2 T\rceil$ times.
\end{theorem}
\begin{proof}
Let $t = 2^{a_1} + 2^{a_2} + \ldots + 2^{a_s}$, where
$a_1 > a_2 > \ldots > a_s \geq 0$. Let
$\tau_i = \sum_{j \leq i} 2^{a_j}$. To prove the discrepancy bound,
the main observation is that at each timestep $t$, the current
signing consists of the output of $\ensuremath{\mathcal{A}}$ on $\lceil \log_2 t \rceil$
different subintervals of the input sequence:
$\{v_1, \ldots, v_{\tau_1}\}$, $\{v_{\tau_1+1}, \ldots, v_{\tau_2}\}$,
all the way down to $\{v_{\tau_{s-1}+1}, \ldots, v_{\tau_s} =
v_t\}$. (This can be proved using an inductive argument.) The
discrepancy for each of these logarithmically-many is at most $D$,
by our assumption on the algorithm $\ensuremath{\mathcal{A}}$, which proves the first
claim. The second claim uses that each time a vector is given a new
sign, it belongs to an subinterval of twice the length; this can
happen only $O(\log T)$ times.
\end{proof}
Using the algorithm of~\cite{Banaszczyk-Journal98,BansalDG-FOCS16} gives us the following result.
\begin{corollary} \label{Cor:insertOnly}
There is an algorithm for the insert-only setting that ensures a
discrepancy of $O(\sqrt{\log n} \log T)$ for any sequence of vectors
of $\ell_2$ length at most $1$ (and hence discrepancy of
$O(\sqrt{s \log n} \log T)$ for any sequence of $s$-sparse vectors with entries in $[-1,1]$) in $O(\log T)$ amortized recourse per update.
\end{corollary}
Another interesting future direction is to get near-optimal discrepancy for small \emph{worst-case recourse} per update (instead of amortized recourse). E.g., in the setting of \Cref{openPb:FullDyn}, can we achieve $\tilde{O}(1)$ discrepancy in $\tilde{O}(1)$ worst-case recourse per update? It will also be interesting to improve \Cref{Cor:insertOnly} to get $O(1)$ amortized recourse per update, or to even get $O(1)$ worst-case recourse per update.
|
\section{Preamble}
\subsection{The Bourgain--Brezis--Mironescu--Maz'ya--Shaposhnikova limit formulae}
In their celebrated paper \cite{BBM}, Bourgain--Brezis--Mironescu studied limits
of Gagliardo semi-norms defined by\footnote{We assume that $p>1,$ since, as in
\cite{Braz}, our main focus will be the fractional Sobolev spaces defined via
$(-\Delta)^{\frac{t}{2}}.$ $\ $The case $p=1$ requires a separate treatment.}%
\begin{equation}
\Vert f\Vert_{\dot{W}^{s,p}({\mathbb{R}}^{n})}=\left( \int_{{\mathbb{R}}%
^{n}}\int_{{\mathbb{R}}^{n}}\frac{\left\vert f(x)-f(y)\right\vert ^{p}%
}{\left\vert x-y\right\vert ^{n+sp}}dxdy\right)^{1/p},\qquad s\in(0,1),\quad p\in
(1,\infty), \label{gag1}%
\end{equation}
and showed that while%
\begin{equation}
\lim_{s\rightarrow1^{-}}\Vert f\Vert_{\dot{W}^{s,p}({\mathbb{R}}^{n})}<\infty\iff
f\quad\text{is constant}, \label{Constancy}%
\end{equation}
the Gagliardo seminorms can be normalized so that the $L^{p}$ norm of the
gradient can be recovered: For $f\in W_{p}^{1}({\mathbb{R}}^{n})$, it holds
\begin{equation}
\lim_{s\rightarrow1^{-}}(1-s)^{\frac{1}{p}}\Vert f\Vert_{\dot{W}^{s,p}%
({\mathbb{R}}^{n})}=C(p,n)\Vert\nabla f\Vert_{L^{p}({\mathbb{R}}^{n})},
\label{BBM}%
\end{equation}
where $C(p,n)$ is a constant that depends only on $p$ and $n,$ and can be
computed explicitly.
Later, Maz'ya--Shaposhnikova \cite{Mazya} used a different normalization to
deal with the case where the (smoothness) parameter $s$ tends to zero, and
proved that
\begin{equation}
\lim_{s\rightarrow0^{+}}s^{\frac{1}{p}}\Vert f\Vert_{\dot{W}^{s,p}({\mathbb{R}%
}^{n})}=c(p,n)\Vert f\Vert_{L^{p}({\mathbb{R}}^{n})}. \label{MS}%
\end{equation}
These results have sparked intense research interest and have been used and
extended in many different directions (for recent related material,
applications and references cf. \cite{Braz}, \cite{BN}, \cite{BGT}, \cite{cejas}, \cite{GT}, \cite{polam}, \cite{Xu}). The
original proofs were somewhat complicated, and the process of finding the
correct normalizations involved ad-hoc considerations. For example, the
correct normalization for (\ref{MS}) was found using the sharp form of the
Sobolev embedding and, conversely, the limit result (\ref{BBM}) suggested the
sharp Fractional Sobolev inequality obtained in \cite{BBM1}.
In \cite{Milman}, \cite{KaradzhovMilmanXiao}, it was observed that these
limiting results can be naturally understood in the more general context of
scales of interpolation spaces. Given a pair of compatible spaces
$(X_{0},X_{1}),$ interpolation methods (cf. \cite{BerghLofstrom}) give rise to
scales of spaces, in particular, the scale $(X_{0},X_{1})_{s,q},$
$s\in(0,1),q\in\lbrack1,\infty],$ produced by the real method\footnote{See
Section \ref{SectionMethod} for background information on interpolation and
extrapolation.}, will play an important r\^{o}le in this paper. We
colloquially refer to $X_{0}$ and $X_{1}$ as \textquotedblleft end point
spaces\textquotedblright,\ and we think that, as the parameter $s$ varies from
$0$ to $1,$ we are moving from one end point space in the scale, to the other.
In this setting, the limit theorems take the form of a continuity result that
can be informally interpreted as saying that \textquotedblleft with
appropriate normalizations we can recover the end point
spaces\textquotedblright\ (cf. Theorem \ref{teomarkao}),
\begin{equation}
\lim_{s\rightarrow0^{+}} c_{s,q} \left\Vert f\right\Vert _{(X_{0},X_{1}%
)_{s,q}}=\sup_{t>0}K(t,f;X_{0},X_{1}); \quad \text{ }\lim_{s\rightarrow1^{-}}%
c_{s,q}\left\Vert f\right\Vert _{(X_{0},X_{1})_{s,q}}=\sup_{t>0}%
\frac{K(t,f;X_{0},X_{1})}{t}. \label{m1}%
\end{equation}
For many pairs of spaces $(X_{0},X_{1}),$ we actually have estimates: For
$f\in X_{0}\cap X_{1},$\footnote{Throughout the paper, the symbol $f \lesssim g$ indicates the existence of a constant $c > 0$, which is independent of all essential parameters, such that $f \leq c \, g$. The symbol $f \approx g$ means that $f \lesssim g$ and $g \lesssim f$.}
\begin{equation}
\sup_{t>0}K(t,f;X_{0},X_{1})\approx\left\Vert f\right\Vert _{X_{0}},\text{
\ \ \ \ \ \ }\sup_{t>0}\frac{K(t,f;X_{0},X_{1})}{t}\approx\left\Vert
f\right\Vert _{X_{1}}. \label{m4}%
\end{equation}
Furthermore, if the pair is
\textquotedblleft exact\textquotedblright \, then one can replace the equivalences signs in \eqref{m4} by equalities\footnote{From this point of view, the constants that appear in (\ref{BBM})
and (\ref{MS}) are connected to the computation of the underlying
$K$-functionals.} (cf. \cite{Milman}, \cite{KaradzhovMilmanXiao}).
From the interpolation point of view, (\ref{BBM}) and (\ref{MS}) are not
separate limiting results, but part of the recovery of end point norms of a
particular interpolation scale. Moreover, we note that the limiting formulae
(\ref{m1}) are somewhat stronger, in the sense that we are free to choose the
index $q,$ which indeed, is not dependent on the original pair of spaces, as
it is the case in (\ref{BBM}) or (\ref{MS}), where $q=p.$
In fact, the normalization constants produced by interpolation are
\textquotedblleft universal\textquotedblright, in the sense that they are
independent of the originating pair of spaces and have an explicit
formula\footnote{There is a method to define and compute the normalizations in
order insure the continuity of the interpolation scales. In particular, the normalizations
for other methods of interpolation are also known. For example, for the
complex method of Calder\'{o}n, $c_{s}=1;$ for the $J-$method, $c_{s,q}%
=\left( (1-s)sq^{\prime}\right) ^{-1/q^{\prime}},$ where $\frac{1}{q}%
+\frac{1}{q^{\prime}}=1.$ In the case of quasi-Banach spaces some adjustments
are necessary (cf. \cite[pag. 35]{JawerthMilman} for more details).}%
\begin{equation}
c_{s,q}=\left\{
\begin{array}
[c]{cc}%
\left( s(1-s)q\right) ^{\frac{1}{q}}, & \quad s\in(0,1),\quad q\in\lbrack1,\infty),\\
1, & q=\infty.
\end{array}
\right. \label{m2}%
\end{equation}
The correct formulae for the normalizations originate from considerations
arising in the extrapolation theory of \cite{JawerthMilman}, where one needs
to work with continuous scales, and must pay close attention to the constants
that appear in the estimates. It turns out that normalized interpolation norms
have certain crucial monotonicity properties that will play a role in what
follows. In particular, it was shown in \cite[(3.2), page 19]{JawerthMilman}
that, if $q\leq r,$ then%
\[
c_{s,r}\left\Vert f\right\Vert _{(X_{0},X_{1})_{s,r}}\leq c_{s,q} \left\Vert
f\right\Vert _{(X_{0},X_{1})_{s,q}}.
\]
Furthermore, the usual reiteration formulae of interpolation theory, as
exemplified by Holmstedt's reiteration formula (cf. \cite[Theorem 2.1, page
307]{BennettSharpley}, \cite[(3.15), page 33]{JawerthMilman}, \cite[Lemmas
2.1--2.3, pages 61--63]{KaradzhovMilman}, \cite{KaradzhovMilmanXiao}, etc.) can be
sharpened, when the norms have been normalized according to \eqref{m2}.
Combining (\ref{m1}) and the fact that%
\begin{equation}
(L^{p}(\mathbb{R}^{n}),\dot{W}_{p}^{1}(\mathbb{R}^{n}))_{s,p}=\dot{W}%
^{s,p}(\mathbb{R}^{n}), \label{m3}%
\end{equation}
with hidden equivalence constants independent of $s,$ we can derive
versions of (\ref{BBM}) and (\ref{MS}). The proof of (\ref{m3}) hinges upon
the known computation of the underlying $K$-functional in terms of a
$p-$modulus of continuity (cf. \cite[Theorem 4.12, page 339]{BennettSharpley}
and the references therein),%
\begin{align*}
K(u,f;L^{p}(\mathbb{R}^{n}),\dot{W}_{p}^{1}(\mathbb{R}^{n})) &
:=\inf_{f=f_{0}+f_{1}}\{\left\Vert f_{0}\right\Vert _{L^{p}(\mathbb{R}^{n}%
)}+u\left\Vert \nabla f_{1}\right\Vert _{L^{p}(\mathbb{R}^{n})}\}\\
& \approx\sup_{\left\vert h\right\vert \leq u}\| \Delta_{h}%
^{1}f\|_{L^{p}(\mathbb{R}^{n})},
\end{align*}
where%
\[
\Delta_{h}^{1}f(x)=\Delta_{h}f(x)=f(x+h)-f(x),\qquad x, h\in{\mathbb{R}}^{n}.
\]
The higher order differences $\Delta_{h}^{k}, \, k \in \mathbb{N},$ are defined inductively,
\[
\Delta_{h}^{k+1}=\Delta_{h}\Delta_{h}^{k},
\]
and we can write
\begin{equation}
\Delta_{h}^{k}f(x)=\sum_{j=0}^{k}(-1)^{j}{\binom{k}{j}}f(x+(k-j)h).
\label{ClassicDiff}%
\end{equation}
The corresponding $K$-functionals for higher order Sobolev spaces $\dot{W}^k_p(\R^n)$ are given by (cf. \cite[Theorem
4.12, pag. 339]{BennettSharpley}, \cite{KaradzhovMilmanXiao}, and the
references therein)
\begin{align}
K(u^k,f;L^{p}(\mathbb{R}^{n}),\dot{W}^{k}_{p}(\mathbb{R}^{n})) & \approx
\sup_{\left\vert h\right\vert \leq u} \| \Delta_{h}%
^{k}f\|_{L^{p}(\mathbb{R}^{n})},\label{mod1}\\
K(u^k,f;L^{p}(\mathbb{R}^{n}),W^{k}_{p}(\mathbb{R}^{n})) & \approx
\min\{1,u^k\}\left\Vert f\right\Vert _{L^{p}(\mathbb{R}^{n})}+\sup_{\left\vert
h\right\vert \leq u} \| \Delta_{h}^{k}f\| _{L^{p}%
(\mathbb{R}^{n})}. \label{mod2}%
\end{align}
Combining (\ref{mod1}) and (\ref{mod2}) with (\ref{m1}), we can prove
Bourgain--Brezis--Mironescu--Maz'ya--Shaposhnikova higher order limit results in
a unified fashion\footnote{The proof of property (\ref{m4}) related to $(L^p(\R^n), \dot{W}^k_p(\R^n))$ can be seen, e.g., in
\cite[Proposition 3, pag 139]{Stein}.} (cf. \cite{KaradzhovMilmanXiao}).
Very recently, Brazke, Schikorra and Yung \cite{Braz}, using heavy Harmonic
Analysis machinery, obtained a new approach to (\ref{BBM}) and suggested a
possible extension to fractional Sobolev spaces, through the use of
Triebel--Lizorkin spaces. The paper \cite{Braz} contains a number of
interesting new inequalities and poses a number of intriguing open problems.
It was their paper that provided the initial impetus for the present work.
However, in order to solve some of the main questions asked in \cite{Braz} we
will rely on interpolation and extrapolation methods. Interestingly, as we
hope to show, interpolation theory will provide us with the crucial spaces,
and the appropriate scalings, to resolve the outstanding issues.
The first problem proposed in \cite[page 5]{Braz} that we shall discuss here
is the extension of (\ref{BBM}) to the fractional case. The interpolation
method provides the strategy, although to implement it requires the
introduction of new spaces as we now explain.
Let $t>0,$ and consider the fractional Sobolev spaces via the fractional Laplacian\footnote{$\dot{H}%
^{t,p}(\mathbb{R}^{n})$ is also known as the space of Riesz potentials, while
$H^{t,p}(\mathbb{R}^{n})$ is the space of Bessel potentials (cf.
\cite{Stein}).}%
\[
\dot{H}^{t,p}(\mathbb{R}^{n})=\{f:\left\Vert f\right\Vert _{\dot{H}%
^{t,p}(\mathbb{R}^{n})}=\| (-\Delta)^{\frac{t}{2}}f\|_{L^{p}%
(\mathbb{R}^{n})}<\infty\},
\]%
\[
H^{t,p}(\mathbb{R}^{n})=\{f:\left\Vert f\right\Vert _{H^{t,p}(\mathbb{R}^{n}%
)}=\left\Vert f\right\Vert _{L^{p}(\mathbb{R}^{n})}+\left\Vert f\right\Vert
_{\dot{H}^{t,p}(\mathbb{R}^{n})}<\infty\},
\]
and the corresponding interpolation pairs $(L^{p}({\mathbb{R}}^{n}%
),H^{t,p}({\mathbb{R}}^{n})),$ $(L^{p}({\mathbb{R}}^{n}),\dot{H}%
^{t,p}({\mathbb{R}}^{n})).$ As a consequence of $L^p$-boundedness of the Riesz transforms,
\begin{equation}\label{R}
\|(-\Delta)^{\frac{1}{2}} f\|_{L^p(\R^n)} \approx \|\nabla f\|_{L^p(\R^n)}, \qquad p \in (1, \infty),
\end{equation}
and thus $\dot{H}^{1, p} (\R^n) = \dot{W}^1_p(\R^n)$ and $H^{1, p} (\R^n) = W^1_p(\R^n)$.
In view of the previous discussion the program we shall follow should be
clear. Indeed, a direct application of (\ref{m1}) yields the following
extension of the Bourgain--Brezis--Mironescu--Maz'ya--Shaposhnikova limits: Let
$t>0,s\in(0,1),p>1,$ then for $f\in H^{t,p}({\mathbb{R}}^{n})$,
\begin{equation}
\lim_{s\rightarrow1^{-}}(1-s)^{\frac{1}{p}}\Vert f\Vert_{(L^{p}({\mathbb{R}}%
^{n}),\dot{H}^{t,p}({\mathbb{R}}^{n}))_{s,p}}\approx\|(-\Delta
)^{\frac{t}{2}}f\| _{L^{p}({\mathbb{R}}^{n})},\lim_{s\rightarrow0^{+}}%
s^{\frac{1}{p}}\Vert f\Vert_{(L^{p}({\mathbb{R}}^{n}),\dot{H}^{t,p}%
({\mathbb{R}}^{n}))_{s,p}}\approx\Vert f\Vert_{L^{p}({\mathbb{R}}^{n})},
\label{agreg1}%
\end{equation}
with a corresponding result for the inhomogeneous pair $(L^{p}({\mathbb{R}}^{n}%
),H^{t,p}({\mathbb{R}}^{n}))$:%
\begin{equation}
\lim_{s\rightarrow1^{-}}(1-s)^{\frac{1}{p}}\Vert f\Vert_{(L^{p}({\mathbb{R}}%
^{n}),H^{t,p}({\mathbb{R}}^{n}))_{s,p}}\approx\left\Vert f\right\Vert
_{H^{t,p}(\mathbb{R}^{n})}, \lim_{s\rightarrow 0^{+}}s^{\frac{1}{p}}\Vert
f\Vert_{(L^{p}({\mathbb{R}}^{n}),H^{t,p}({\mathbb{R}}^{n}))_{s,p}}\approx\Vert
f\Vert_{L^{p}({\mathbb{R}}^{n})}. \label{agreg2}%
\end{equation}
As we noted before, the case $t=1$ in (\ref{agreg1}) (resp. (\ref{agreg2}))
corresponds to (\ref{BBM}) and (\ref{MS}), while integer values can be seen to
correspond to the higher order limit theorems of \cite{KaradzhovMilmanXiao}.
However, to obtain meaningful statements for arbitrary $t>0,$ we are required
to find explicit descriptions of the indicated interpolation spaces, and
control the corresponding interpolation norms within constants independent of
the parameter $s.$ The difficulty here is due to the fact that the nonlocal operator
$(-\Delta)^{\frac{t}{2}}$ is involved. We will overcome these obstructions using a
tool from Approximation Theory: Fractional differences. Fractional differences
give an appropriate extension of (\ref{ClassicDiff}) and lead to a natural
extension of (\ref{mod1}) and (\ref{mod2}).
We now introduce fractional differences and a class of seminorms that extends
the classical Gagliardo seminorms \eqref{gag1} to the fractional case.
\subsection{Interpolation and fractional differences: The Butzer seminorms}
Butzer and his collaborators (cf. \cite{Butzer, ButzerW} and the references therein)
apparently introduced the idea of modifying (\ref{ClassicDiff}) in order to
define fractional differences of any order $t>0$ so that a related fractional differentiation approach can be developed. For $t>0,$ we let
$\Delta_{h}^{t}$ denote the Butzer operator\footnote{The concept of fractional
differences was already used by Butzer et al. \cite{Butzer} to introduce
fractional order moduli of smoothness, which has recently become a powerful
tool in approximation theory, e.g., in the study of sharp inequalities between
moduli of smoothness (Jackson--Marchaud--Ulyanov inequalities), cf.
\cite{Tikhonov}, \cite{Trebels}, \cite{Kolomoitsev}, \cite{Kolomoitsev21} and
the references therein.} given by%
\begin{equation}
\Delta_{h}^{t}f(x)=\sum_{j=0}^{\infty}(-1)^{j}{\binom{t}{j}}f(x+(t-j)h),
\label{FractDiff}%
\end{equation}
where by definition
\[
{\binom{t}{j}}=\frac{t(t-1)\ldots(t-j+1)}{j!},\qquad {\binom{t}{0}}=1.
\]
Obviously if $t=k\in{{\mathbb{N}}}$, then ${\binom{t}{j}=0}$ for all $j\geq
k+1,$ and (\ref{FractDiff}) coincides with the classical differences
\eqref{ClassicDiff}. Remarkably, the $K$-functionals, we are seeking to
compute in relation with \eqref{agreg1}-\eqref{agreg2}, can be formally obtained\footnote{For the details of the computation
we refer to \cite{Kolomoitsev}, with \cite{Butzer} and \cite{Wilmes} as a
forerunners.} by replacing $\Delta_{h}^{k}$ by $\Delta_{h}^{t}$ on the right
hand side of (\ref{mod1}) and (\ref{mod2}), to obtain
\begin{equation*}
K(u^{t},f;L^{p}(\mathbb{R}^{n}),\dot{H}^{t,p}(\mathbb{R}^{n}))\approx
\sup_{\left\vert h\right\vert \leq u}\left\Vert \Delta_{h}^{t}f\right\Vert
_{L^{p}(\mathbb{R}^{n})},
\end{equation*}%
\begin{equation*}
K(u^{t},f;L^{p}(\mathbb{R}^{n}),H^{t,p}(\mathbb{R}^{n}))\approx\min
\{1,u^{t}\}\left\Vert f\right\Vert _{L^{p}(\mathbb{R}^{n})}+\sup_{\left\vert
h\right\vert \leq u}\left\Vert \Delta_{h}^{t}f\right\Vert _{L^{p}%
(\mathbb{R}^{n})}.
\end{equation*}
Armed with the formulae for the $K$-functionals we can give a complete
description of the interpolation spaces $(L^{p},\dot{H}^{t,p}({\mathbb{R}}%
^{n}))_{s,q}$ (resp. $(L^{p},H^{t,p}({\mathbb{R}}^{n}))_{s,q}$) (cf. Lemma
\ref{LemmaIntFractGS} below) in terms of the following (semi)norms.
\begin{definition}
Let\footnote{In this paper we shall be interested in the case
$1<p<\infty.$} $1\leq p\leq\infty,0<s<t,$ then we define the Butzer seminorms,
as follows\footnote{Note also that \eqref{DefFractGS} enables us to introduce
Sobolev spaces of any order $s>0$ (not necessarily $s<1$ like in
\eqref{Gagliardo}), via higher order differences. For instance, one can
introduce for $s\in(0,2)$,
\[
\Vert f\Vert_{\dot{W}^{s,p}({\mathbb{R}}^{n}),2}=\bigg(\int_{{\mathbb{R}}^{n}%
}\int_{{\mathbb{R}}^{n}}\frac{|f(x+2h)-2f(x+h)+f(x)|^{p}}{|h|^{sp+N}%
}\,dx\,dh\bigg)^{\frac{1}{p}}.
\]
}
\begin{equation}
\Vert f\Vert_{\dot{W}^{s,p}({\mathbb{R}}^{n}),t}:=\left\{
\begin{array}
[c]{cc}%
\big(\int_{{\mathbb{R}}^{n}}\int_{{\mathbb{R}}^{n}}\frac{|\Delta_{h}%
^{t}f(x)|^{p}}{|h|^{sp+n}}\,dx\,dh \big)^{\frac{1}{p}}, &\quad p<\infty\\
\sup_{(x,h)} \frac{|\Delta_{h}^{t}f(x)|}{|h|^{s}}, & \quad p=\infty.
\end{array}
\right. \label{DefFractGS}%
\end{equation}
We let\footnote{The appearence of the prefactor $(s(t-s)p)^{-\frac{1}{p}}$ in front of $\|f\|_{L^p(\R^n)}$ is dictated by standard normalizations in interpolation theory. This will become clear later, cf. Lemma \ref{LemmaIntFractGS}.}
\begin{equation}\label{DefFractGSIn}%
\|f\|_{W^{s, p}(\R^n), t} := (s(t-s)p)^{-\frac{1}{p}} \|f\|_{L^p(\R^n)} + \|f\|_{\dot{W}^{s, p}(\R^n), t}.
\end{equation}
\end{definition}
Observe that when $t=1$ we recover the classical Gagliardo seminorms (cf. \eqref{gag1})
\begin{equation}
\Vert f\Vert_{\dot{W}^{s,p}({\mathbb{R}}^{n}),1}=\Vert f\Vert_{\dot{W}%
^{s,p}({\mathbb{R}}^{n})},\qquad s\in(0,1). \label{FractModuInt}%
\end{equation}
Furthermore, if $s<t<1$ then $\Vert\cdot\Vert_{\dot{W}^{s,p}({\mathbb{R}}%
^{n}),t}$ defines an equivalent semi-norm on $\dot{W}^{s,p}({\mathbb{R}}%
^{n})$, i.e.,
\begin{equation}
\Vert f\Vert_{\dot{W}^{s,p}({\mathbb{R}}^{n}),t}\approx\Vert f\Vert_{\dot
{W}^{s,p}({\mathbb{R}}^{n})}. \label{EquivalenceGSnorms}%
\end{equation}
Note, however, that the constants of equivalence blow up, as $s$ approaches
$t$ (cf. Lemma \ref{LemmaFractInt} below):
\begin{equation}
\frac{1}{(t-s)^{\frac{1}{\max\{p,2\}}}}\Vert f\Vert_{\dot{W}^{s,p}({\mathbb{R}}^{n}%
)}\lesssim\Vert f\Vert_{\dot{W}^{s,p}({\mathbb{R}}^{n}),t}\lesssim\frac
{1}{(t-s)^{\frac{1}{\min\{p,2\}}}}\Vert f\Vert_{\dot{W}^{s,p}({\mathbb{R}}^{n})}.
\label{later}%
\end{equation}
By direct computation we characterize the interpolation norms (cf. Lemma \ref{LemmaIntFractGS}
below): Suppose that $1<p<\infty,0<s<1$ and $t>0$ then%
\begin{equation*}
\left\Vert f\right\Vert _{(L^{p}({\mathbb{R}}^{n}),\dot{H}^{t,p}({\mathbb{R}%
}^{n}))_{s,p}}\approx\left\Vert f\right\Vert _{\dot{W}^{st,p}({\mathbb{R}}%
^{n}),t},
\end{equation*}
with hidden constants of equivalence independent of $s$ (but depending on
$n,p$ and $t$).
In particular, to derive the fractional
Bourgain--Brezis--Mironescu--Maz'ya--Shaposhnikova limit formulae, we simply
replace the interpolation seminorms $\|\cdot\|_{(L^{p},\dot
{H}^{t,p}({\mathbb{R}}^{n}))_{s,p}}$ by $\|\cdot\|_{\dot
{W}^{st,p}({\mathbb{R}}^{n}),t}$ in (\ref{agreg1}) (resp., replace $\| \cdot\|_{(L^{p},H^{t,p}({\mathbb{R}}^{n}))_{s,p}}$ with $\|\cdot\|_{\dot{W}^{st,p}({\mathbb{R}}^{n}),t}+ (s(1-s))^{-\frac{1}{p}} \|\cdot\|
_{L^{p}({\mathbb{R}}^{n})}$ in (\ref{agreg2})$).$ For example, the general
fractional homogeneous version reads now
\begin{equation}
\lim_{s\rightarrow t^{-}}(t-s)^{\frac{1}{p}}\Vert f\Vert_{\dot{W}^{s,p}%
({\mathbb{R}}^{n}),t}\approx \| (-\Delta)^{\frac{t}{2}} f\|_{L^p
(\mathbb{R}^{n})}, \qquad \lim_{s\rightarrow0^{+}}s^{\frac{1}{p}}\Vert f\Vert_{\dot
{W}^{s,p}({\mathbb{R}}^{n}),t}\approx\Vert f\Vert_{L^{p}({\mathbb{R}}^{n})}.
\label{bbmf}%
\end{equation}
This answers completely the question raised in \cite[page 5]{Braz} asking for a
possible fractional extension of (\ref{BBM}) (see also \eqref{R}). Note that
(\ref{bbmf}) and (\ref{later}) show that we cannot use the classical Gagliardo
seminorms for this purpose.
Further evidence of the important r\^{o}le of the Butzer seminorms in our
project is provided by the following result, which answers another question
raised in \cite{Braz}.
\subsection{The fractional Bourgain--Brezis--Mironescu convergence theorem via
Butzer seminorms}
We consider the extension to fractional Sobolev spaces of the
Bourgain--Brezis--Mironescu convergence theorem \cite{BBM},
\begin{theorem}
\label{convergencebbm}Let $p\in(1,\infty)$, assume that $f_{k}\in
\mathcal{S}({\mathbb{R}}^{n})$, $k\in\mathbb{N}$, and%
\[
f_{k}\rightharpoonup f\quad\text{weakly in}\quad L^{p}({\mathbb{R}}^{n}%
)\quad\text{as}\quad k\rightarrow\infty.
\]
Let $\{s_{k}\}_{k\in{{\mathbb{N}}}}\subset(0,1)$ be such that $s_{k}\uparrow1$
and, moreover, assume that
\[
\Lambda:=\sup_{k\in{{\mathbb{N}}}}\Big(\Vert f_{k}\Vert_{L^{p}({\mathbb{R}%
}^{n})}+(1-s_{k})^{\frac{1}{p}}\Vert f_{k}\Vert_{\dot{W}^{s_{k},p}%
({\mathbb{R}}^{n})}\Big)<\infty.
\]
Then $f\in W_{p}^{1}({\mathbb{R}}^{n})$ and there exists $C=C(p,n)>0$ such
that
\[
\Vert f\Vert_{L^{p}({\mathbb{R}}^{n})}+\Vert\nabla f\Vert_{L^{p}({\mathbb{R}%
}^{n})}\leq C\Lambda.
\]
In fact, $f_{k}\rightarrow f$ strongly in $L_{loc}^{p}({\mathbb{R}}^{n}).$
\end{theorem}
In \cite{Braz} the authors propose a plan of attack for a possible fractional
extension of Theorem \ref{convergencebbm}. The idea is based on the use of the
Triebel-Lizorkin spaces $\dot{F}^s_{p, q}(\R^N)$ (we refer to Section \ref{secfunctionspaces} for the definitions). Indeed, based on this idea, a new proof of Theorem
\ref{convergencebbm} was presented in \cite{Braz}, using the following sharp Sobolev-type
inequality:
\begin{theorem}
\label{Theorem343} Let $s\in(0,1),p\in(1,\infty)$ and $\Lambda>1$. Assume
$1-s\leq\frac{1}{2\Lambda}$ and let $1-\bar{r}=\Lambda(1-s)$. Then there
exists $C$, which depends only on $n,p$ and $\Lambda$, such that
\[
\Vert f\Vert_{\dot{F}_{p,2}^{r}({\mathbb{R}}^{n})}\leq C\Big(\Vert
f\Vert_{L^{p}({\mathbb{R}}^{n})}+(1-s)^{\frac{1}{p}}\Vert f\Vert_{\dot
{W}^{s,p}({\mathbb{R}}^{n})}\Big),\qquad r\in\lbrack0,\bar{r}].
\]
\end{theorem}
With this result at their disposal the proof of Theorem \ref{convergencebbm}
given in \cite{Braz} now proceeds by combining Theorem \ref{Theorem343}
(taking limits as $r\rightarrow1^{-})$, with \eqref{R} and the Littewood--Paley estimate
\begin{equation*}
\Vert f\Vert_{\dot{F}_{p,2}^{t}({\mathbb{R}}^{n})}\approx\Vert(-\Delta
)^{\frac{t}{2}}f\Vert_{L^{p}({\mathbb{R}}^{n})}, \qquad1<p<\infty,
\end{equation*}
and then applying the Rellich--Kondrachov theorem.
In view of this, the following conjecture was formulated in \cite{Braz}.
\begin{op}
[{\cite[Question 1.11]{Braz}}]\label{Question1.11} Let $p\in(1,\infty)$,
assume that
\[
f_{k}\rightharpoonup f\quad\text{weakly in}\quad L^{p}({\mathbb{R}}^{n}%
)\quad\text{as}\quad k\rightarrow\infty.
\]
Let $t\in(0,1)$ and $(s_{k})_{k\in{{\mathbb{N}}}}\subset(0,t)$ such that
$s_{k}\uparrow t$ and assume that
\[
\Lambda:=\sup_{k\in{{\mathbb{N}}}}\Big(\Vert f_{k}\Vert_{L^{p}({\mathbb{R}%
}^{n})}+(t-s_{k})^{\frac{1}{p}}\Vert f_{k}\Vert_{\dot{W}^{s_{k},p}%
({\mathbb{R}}^{n})}\Big)<\infty.
\]
Then $f\in H^{t,p}({\mathbb{R}}^{n})$ and there exists $C=C(p,n,t)>0$ such
that
\[
\lim_{\bar{t}\uparrow t}\limsup_{k\rightarrow\infty}\Vert(-\Delta)^{\frac
{\bar{t}}{2}}f_{k}\Vert_{L^{p}({\mathbb{R}}^{n})}\leq C\Lambda.
\]
\end{op}
Note that in the case $t=1,$ the conjectured result is exactly Theorem
\ref{convergencebbm}. However, as we have indicated above, the standard Gagliardo
seminorms are \textquotedblleft too far" from the \textquotedblleft
exact\textquotedblright\ interpolation seminorms $\|\cdot\|_{\dot{W}^{s_{k},p}({\mathbb{R}%
}^{n}), t} \,$ (cf. \eqref{DefFractGS}) for $0<t<1$. We use this insight to give a counterexample to Open Problem \ref{Question1.11} (cf.
Section \ref{SectionMR}, Proposition \ref{PropQuestion1.11} for the precise
statement, and Section \ref{SectionProofs} for a proof).
The problem thus remains: What spaces should we use to formulate and prove a
fractional extension of Theorem \ref{convergencebbm}? Once again
interpolation, via the Butzer seminorms, comes to our rescue and we are
essentially able to prove that the conjectured result in Open Problem \ref{Question1.11} is true if in its
statement we replace \textquotedblleft Gagliardo seminorms\textquotedblright%
\ by \textquotedblleft Butzer seminorms\textquotedblright\ (cf. Section
\ref{SectionMR}, Theorem \ref{ThemQuestion1.11} for the precise statement, and
Section \ref{SectionProofs} for a complete proof).
\subsection{Sharp Sobolev inequalities}
Another problem proposed in \cite{Braz} deals with Sobolev type inequalities.
Let $0<s<t<1$, an elementary version of Sobolev's inequality asserts that
\begin{equation}
\Vert f\Vert_{\dot{W}^{s,2}({\mathbb{R}}^{n})}\leq\gamma_{s,t}\Big(\Vert
f\Vert_{L^{2}({\mathbb{R}}^{n})}+\Vert f\Vert_{\dot{W}^{t,2}({\mathbb{R}}%
^{n})}\Big). \label{SobolevHilbert}%
\end{equation}
Here $\gamma_{s,t}$ is a positive constant which depends, in particular, on
the smoothness parameters $s$ and $t$. Furthermore, at least formally,
\eqref{SobolevHilbert} is still valid in the limiting values $s=0$ and $t=1$.
According to \eqref{Constancy}, \eqref{BBM} and \eqref{MS}, the validity of
the previous assertion should be reflected in the behavior of the equivalence
constant $\gamma_{s,t}$ that appears in \eqref{SobolevHilbert}, in terms of
the blows up $\min\{t,1-t\}^{\frac{1}{2}}$ and $\min\{s,1-s\}^{\frac{1}{2}}$.
Indeed, the following result was obtained in \cite[Corollary 1.3]{Braz}.
\begin{theorem}
\label{ThmIntro1} Let $0<s<t<1$ and $f\in\mathcal{S}({\mathbb{R}}^{n})$. Then
there exists $C=C(n)>0$, such that
\[
\min\{s,1-s\}^{\frac{1}{2}}\Vert f\Vert_{\dot{W}^{s,2}({\mathbb{R}}^{n})}\leq
C\Big(\Vert f\Vert_{L^{2}({\mathbb{R}}^{n})}+\min\{t,1-t\}^{\frac{1}{2}}\Vert
f\Vert_{\dot{W}^{t,2}({\mathbb{R}}^{n})}\Big).
\]
\end{theorem}
It is well known that the $L^{p}$-counterpart of \eqref{SobolevHilbert} is
also true, i.e.,
\begin{equation*}
\Vert f\Vert_{\dot{W}^{s,p}({\mathbb{R}}^{n})}\leq C_{s,t}\Big(\Vert
f\Vert_{L^{p}({\mathbb{R}}^{n})}+\Vert f\Vert_{\dot{W}^{t,p}({\mathbb{R}}%
^{n})}\Big)
\end{equation*}
provided that $0<s<t<1$ and $1<p<\infty$. Accordingly, a similar result, in
the spirit of Theorem \ref{ThmIntro1}, was expected. However, the set of
harmonic analysis techniques developed in \cite{Braz} do not work outside the
Hilbert setting given by $p=2$. This leads to the following question, explicitly raised in
\cite{Braz}:
\begin{op}
\label{OpenProblemIntro1} Let $p\in(1,\infty),0<\theta<s<t<1$ and
$f\in\mathcal{S}({\mathbb{R}}^{n})$. Does there exist $C=C(n,p,\theta)>0$,
such that%
\begin{equation}
\min\{s,1-s\}^{\frac{1}{p}}\Vert f\Vert_{\dot{W}^{s,p}({\mathbb{R}}^{n})}\leq
C\Big(\Vert f\Vert_{L^{p}({\mathbb{R}}^{n})}+\min\{t,1-t\}^{\frac{1}{p}}\Vert
f\Vert_{\dot{W}^{t,p}({\mathbb{R}}^{n})}\Big)? \label{OpenProblemIntro1Eq}%
\end{equation}
\end{op}
We provide a positive answer to this problem. Indeed, in Section
\ref{SectionMR}, Theorem \ref{ConjectureBSY}, we state a somewhat stronger
result which is then proved in Section \ref{SectionProofs}.
We now turn to explain the organization of the paper. The brief Section
\ref{secfunctionspaces} contains the basic background and notation concerning
the function spaces we shall consider; Section \ref{SectionMR} contains the
statements of the main results, which we then prove in Section
\ref{SectionProofs}. The interpolating Section \ref{SectionMethod} contains
detailed information with complete proofs concerning the interpolation and
extrapolation tools that constitute the basic methods we use in this paper.
Finally, in the Appendix, we collect further results and applications. In particular, we prove an extension of the classical fractional Sobolev inequalities of Bourgain--Brezis--Mironescu \cite{BBM1} and Maz'ya--Shaposhnikova \cite{Mazya}, that were one of the original motivations for the limiting theorems \eqref{BBM} and \eqref{MS}.
\vspace{4mm}
\textbf{Acknowledgments.} We are grateful to Petru Mironescu for his constant interest and encouragement during the preparation of this paper and to Sergey Tikhonov for precious information concerning fractional differences and moduli of smoothness.
The first named author has been partially supported by the LABEX MILYON (ANR-10-LABX-0070) of Universit\'e de Lyon, within the program ``Investissement d'Avenir" (ANR-11-IDEX-0007) operated by the
French National Research Agency (ANR), and by MTM2017-84058-P (AEI/FEDER, UE) .
\section{Background on function spaces\label{secfunctionspaces}}
Let $s\in(0,1)$ and $p\in(1,\infty)$, the \emph{(fractional) Sobolev space} $\dot
{W}^{s,p}({\mathbb{R}}^{n})$ is equipped with the \emph{Gagliardo seminorm}
\begin{equation}
\Vert f\Vert_{\dot{W}^{s,p}({\mathbb{R}}^{n})}:=\bigg(\int_{{\mathbb{R}}^{n}%
}\int_{{\mathbb{R}}^{n}}\frac{|f(x)-f(y)|^{p}}{|x-y|^{sp+n}}%
\,dx\,dy\bigg)^{1/p}. \label{Gagliardo}%
\end{equation}
The \emph{Riesz potential space} $\dot{H}^{s,p}({\mathbb{R}}^{n})$ is endowed with
\[
\Vert f\Vert_{\dot{H}^{s,p}({\mathbb{R}}^{n})}:=\Vert(-\Delta)^{\frac{s}{2}%
}f\Vert_{L^{p}({\mathbb{R}}^{n})}.
\]
The spaces $\dot{H}^{s,p}({\mathbb{R}}^{n})$ make sense for any $s\in
{\mathbb{R}}$. In particular, $\dot{H}^{k,p}({\mathbb{R}}^{n}),\,k\in
{{\mathbb{N}}},$ coincides with the \emph{classical Sobolev space} $\dot{W}_{p}%
^{k}({\mathbb{R}}^{n})$ and
\begin{equation}
\Vert f\Vert_{\dot{H}^{k,p}({\mathbb{R}}^{n})}\approx\Vert f\Vert_{\dot{W}%
_{p}^{k}({\mathbb{R}}^{n})}:=\|\nabla^{k}f\|_{L^{p}({\mathbb{R}}^{n}%
)},\qquad p\in(1,\infty). \label{SobRiesz}%
\end{equation}
We let $H^{s,p}({\mathbb{R}}^{n})=L^{p}({\mathbb{R}}^{n})\cap\dot{H}%
^{s,p}({\mathbb{R}}^{n})$ (the \emph{space of Bessel potentials}) and $\|f\|_{H^{s,p}({\mathbb{R}}^{n})} = \|f\|_{L^p(\R^n)} + \Vert(-\Delta)^{\frac{s}{2}%
}f\Vert_{L^{p}({\mathbb{R}}^{n})}$.
The \emph{(homogeneous) Triebel--Lizorkin space} $\dot{F}_{p,q}^{s}({\mathbb{R}}%
^{n}),\,s\in{\mathbb{R}},\,p\in(0,\infty),\,q\in(0,\infty]$, is formed by all
$f\in\dot{\mathcal{S}}^{\prime}({\mathbb{R}}^{n})$ such that
\[
\Vert f\Vert_{\dot{F}_{p,q}^{s}({\mathbb{R}}^{n})}:=\bigg\|\bigg(\sum
_{j=-\infty}^{\infty}2^{jsq}|\Delta_{j}f|^{q}\bigg)^{1/q}\bigg\|_{L^{p}%
({\mathbb{R}}^{n})}<\infty,
\]
with the obvious modification if $q=\infty$. Here, $\{\Delta_{j}%
f:j\in{{\mathbb{Z}}}\}$ is the standard dyadic Littlewood--Paley decomposition
of $f$. The inhomogeneous counterparts, $F_{p,q}^{s}({\mathbb{R}}^{n}),$ can
be introduced similarly.
Interchanging the roles of $L^{p}({\mathbb{R}}^{n})$ and $\ell_{q}%
({{\mathbb{Z}}})$ in the definition of $\dot{F}_{p,q}^{s}({\mathbb{R}}^{n})$
we arrive at the Besov spaces. Specifically, the \emph{(homogeneous) Besov space}
$\dot{B}_{p,q}^{s}({\mathbb{R}}^{n}),\,s\in{\mathbb{R}},p,q\in(0,\infty]$, is
formed by all $f\in\dot{\mathcal{S}}^{\prime}({\mathbb{R}}^{n})$ such that
\[
\Vert f\Vert_{\dot{B}_{p,q}^{s}({\mathbb{R}}^{n})}:=\bigg(\sum_{j=-\infty
}^{\infty}2^{jsq}\Vert\Delta_{j}f\Vert_{L^{p}({\mathbb{R}}^{n})}%
^{q}\bigg)^{1/q}<\infty,
\]
with the obvious modifications if $q=\infty$.
We shall use the following standard notations. Given a quasi-Banach space $X$
and $\lambda>0$, we will denote by $\lambda X$ the space endowed with
\[
\Vert f\Vert_{\lambda X}=\lambda\Vert f\Vert_{X}, \qquad f \in X.
\]
Let $Y$ be a quasi-Banach space and let $\beta>0$. We will use the notation
$\lambda X\hookrightarrow\beta Y$ to indicate that there exists a positive
constant $C$, independent of $\lambda$ and $\beta$, such that
\[
\beta\Vert f\Vert_{Y}\leq C\lambda\Vert f\Vert_{X},\qquad f\in X.
\]
In particular, if $\lambda=\beta=1$ then we simply write $X\hookrightarrow Y$.
By $X=Y$ we mean that $X\hookrightarrow Y$ and $Y\hookrightarrow X$.
We collect some well-known relations between the function spaces described
above (cf. \cite{BerghLofstrom}, \cite{Stein} and \cite{Triebel83}).
\begin{lemma}
\label{TableCoincidences}
\begin{enumerate}
[\upshape(i)]
\item Let $s\in{\mathbb{R}},\,p\in(0,\infty)$ and $q\in(0,\infty]$. Then
\[
\dot{B}_{p,\min\{p,q\}}^{s}({\mathbb{R}}^{n})\hookrightarrow\dot{F}_{p,q}%
^{s}({\mathbb{R}}^{n})\hookrightarrow\dot{B}_{p,\max\{p,q\}}^{s}({\mathbb{R}%
}^{n}).
\]
In particular,
\[
\dot{F}_{p,p}^{s}({\mathbb{R}}^{n})=\dot{B}_{p,p}^{s}({\mathbb{R}}^{n}).
\]
\item Let $s\in(0,1)$ and $p\in(1,\infty)$. Then
\[
\dot{W}^{s,p}({\mathbb{R}}^{n})=\dot{B}_{p,p}^{s}({\mathbb{R}}^{n}).
\]
\item Let $s\in{\mathbb{R}}$ and $p\in(1,\infty)$. Then
\[
\dot{F}_{p,2}^{s}({\mathbb{R}}^{n})=\dot{H}^{s,p}({\mathbb{R}}^{n}).
\]
\end{enumerate}
\end{lemma}
\section{Main results}
\label{SectionMR}
In this section we provide a list of statements of the main results we have
obtained. The proofs will be given in Section \ref{SectionProofs}.
We start outlining some details of a counterexample that provides a negative
answer to Open Problem \ref{Question1.11}. To avoid technicalities, we switch
temporarily from ${\mathbb{R}}^{n}$ to the unit circle ${\mathbb{T}}$\footnote{Spaces of periodic functions are defined similarly as their analogs on $\R^n$, simply replacing $L^p(\R^n)$ by $L^p(\T)$.}.
However, our construction can be easily modified to deal with ${\mathbb{R}%
}^{n}$ and $p>\frac{2n}{n+1}$; in this connection see Remark
\ref{RemarkCounterexample} below.
\begin{proposition}
\label{PropQuestion1.11} Let $p\in(1,\infty)$ and $t\in(0,1)$. Let $f$ be
formally associated to a Fourier series as follows,
\begin{equation}
f(x)\sim\sum_{\nu=1}^{\infty}\nu^{-t-1+\frac{1}{p}}\cos(\nu x),\qquad
x\in{\mathbb{T}}.\label{FourierSeries}%
\end{equation}
Then, $f\in L^{p}({\mathbb{T}})$ but $f\not \in H^{t,p}({\mathbb{T}})$.
Furthermore, given any sequence $(s_{k})_{k\in{{\mathbb{N}}}}\subset(0,t)$
with $s_{k}\uparrow t$,
\[
\sup_{k\in{{\mathbb{N}}}}\,(t-s_{k})^{\frac{1}{p}}\Vert f\Vert_{\dot{W}%
^{s_{k},p}({\mathbb{T}})}<\infty.
\]
\end{proposition}
In fact, we are able to go far beyond, and we show that a correct
formulation of Open Problem \ref{Question1.11} is obtained by means of replacing
the classical seminorms $\|\cdot\|_{\dot{W}^{s,p}({\mathbb{R}}^{n})}$
by the Butzer seminorms $\|\cdot\|_{\dot{W}^{s,p}({\mathbb{R}}^{n}%
),t}$ (cf. \eqref{DefFractGS} and \eqref{EquivalenceGSnorms}).
\begin{theorem}
\label{ThemQuestion1.11} Let $p\in(1,\infty)$, assume that
\[
f_{k}\rightharpoonup f\quad\text{weakly in}\quad L^{p}({\mathbb{R}}^{n}%
)\quad\text{as}\quad k\rightarrow\infty.
\]
Let $t\in(0,1)$ and $(s_{k})_{k\in{{\mathbb{N}}}}\subset(0,t)$ be such that
$s_{k}\uparrow t.$ Assume that
\[
\Lambda:=\sup_{k\in{{\mathbb{N}}}}\Big(\Vert f_{k}\Vert_{L^{p}({\mathbb{R}%
}^{n})}+(t-s_{k})^{\frac{1}{p}}\Vert f_{k}\Vert_{\dot{W}^{s_{k},p}%
({\mathbb{R}}^{n}),t}\Big)<\infty.
\]
Then $f\in H^{t,p}({\mathbb{R}}^{n})$ and there exists $C=C(n, p, t)>0$ such
that
\[
\lim_{\bar{t}\uparrow t}\limsup_{k\rightarrow\infty}\Vert(-\Delta)^{\frac
{\bar{t}}{2}}f_{k}\Vert_{L^{p}({\mathbb{R}}^{n})}\leq C\Lambda.
\]
\end{theorem}
To prove Theorem \ref{ThemQuestion1.11} we need an extension of Theorem
\ref{Theorem343} formulated in terms of the Butzer seminorms
$\Vert\cdot\Vert_{\dot{W}^{s,p}({\mathbb{R}}^{n}),t}$. It turns out that our
method is flexible enough to incorporate both, the case $s\rightarrow0+$ , as
well as to deal with inhomogeneous Triebel--Lizorkin spaces.
\begin{theorem}
\label{Theorem343Fract} Let $0 < s < t \leq1, \, p \in(1, \infty)$ and
$\Lambda> 1$.
\begin{enumerate}
\item[(1)] Assume $t-s\leq\frac{t}{2\Lambda}$ and let $t-\bar{r}=\Lambda
(t-s)$. Then, there exists $C$, which depends only on $n,p,t$ and $\Lambda$,
such that
\begin{equation}
\Vert f\Vert_{F_{p,2}^{\bar{r}}({\mathbb{R}}^{n})}\leq C\Big(\Vert
f\Vert_{L^{p}({\mathbb{R}}^{n})}+(t-s)^{\frac{1}{p}}\Vert f\Vert_{\dot
{W}^{s,p}({\mathbb{R}}^{n}),t}\Big).\label{Aux61}%
\end{equation}
There is a corresponding result for homogeneous Triebel--Lizorkin spaces,
\begin{equation}
\Vert f\Vert_{\dot{F}_{p,2}^{r}({\mathbb{R}}^{n})}\leq C\Big(\Vert
f\Vert_{L^{p}({\mathbb{R}}^{n})}+(t-s)^{\frac{1}{p}}\Vert f\Vert_{\dot
{W}^{s,p}({\mathbb{R}}^{n}),t}\Big),\qquad r\in\lbrack0,\bar{r}].\label{Aux6}%
\end{equation}
\item[(2)] Assume $s<\frac{1}{2\Lambda}$ and let $\bar{r}=\frac{1}{\Lambda}s$.
Then there exists $C$, which depends only on $n,p,t$ and $\Lambda$, such that
\begin{equation}
\Vert f\Vert_{F_{p,2}^{\bar{r}}({\mathbb{R}}^{n})}\leq C\Big(\Vert
f\Vert_{L^{p}({\mathbb{R}}^{n})}+s^{\frac{1}{p}}\Vert f\Vert_{\dot{W}%
^{s,p}({\mathbb{R}}^{n}),t}\Big). \label{Aux62}%
\end{equation}
Likewise, for homogeneous Triebel--Lizorkin spaces we have
\begin{equation}
\Vert f\Vert_{\dot{F}_{p,2}^{r}({\mathbb{R}}^{n})}\leq C\Big(\Vert
f\Vert_{L^{p}({\mathbb{R}}^{n})}+s^{\frac{1}{p}}\Vert f\Vert_{\dot{W}%
^{s,p}({\mathbb{R}}^{n}),t}\Big),\qquad r\in\lbrack0,\bar{r}]. \label{Aux6new}%
\end{equation}
\end{enumerate}
\end{theorem}
We apply our method to provide an extension of \eqref{BBM}-\eqref{MS} in terms
of the fractional Laplacian and Butzer seminorms.
\begin{theorem}
\label{TheoremFBBM} Let $t\in(0,1]$ and $p\in(1,\infty)$. Assume $f\in
H^{t,p}({\mathbb{R}}^{n})$. Then
\begin{equation}
\lim_{s\rightarrow t^{-}}(t-s)^{\frac{1}{p}}\Vert f\Vert_{\dot{W}^{s,p}%
({\mathbb{R}}^{n}),t}\approx\Vert(-\Delta)^{\frac{t}{2}}f\Vert_{L^{p}%
({\mathbb{R}}^{n})} \label{FractGSAssertion}%
\end{equation}
and
\begin{equation}
\lim_{s\rightarrow0^{+}}s^{\frac{1}{p}}\Vert f\Vert_{\dot{W}^{s,p}({\mathbb{R}%
}^{n}),t}\approx\Vert f\Vert_{L^{p}({\mathbb{R}}^{n})}.
\label{FractGSAssertionMS}%
\end{equation}
The corresponding results for inhomogeneous spaces also hold true:
\begin{equation}
\lim_{s\rightarrow t^{-}}(t-s)^{\frac{1}{p}}\Vert f\Vert_{W^{s,p}%
({\mathbb{R}}^{n}),t}\approx \|f\|_{H^{t, p}(\R^n)} \label{FractGSAssertionIn}%
\end{equation}
and
\begin{equation}
\lim_{s\rightarrow0^{+}}s^{\frac{1}{p}}\Vert f\Vert_{W^{s,p}({\mathbb{R}%
}^{n}),t}\approx\Vert f\Vert_{L^{p}({\mathbb{R}}^{n})}.
\label{FractGSAssertionMSIn}%
\end{equation}
\end{theorem}
\begin{remark}
Observe that \eqref{FractGSAssertion} and \eqref{FractGSAssertionMS} with $t=1$ give back, up to equivalence constants, the classical formulae \eqref{BBM} and \eqref{MS}.
\end{remark}
As indicated by Theorems \ref{ThemQuestion1.11} and \ref{TheoremFBBM}, the
classical seminorms $\Vert\cdot\Vert_{\dot{W}^{s,p}({\mathbb{R}}^{n})}$ are not the
optimal choices for dealing with Sobolev-type inequalities involving
$H^{t,p}({\mathbb{R}}^{n})$. This phenomenon is illustrated in the following result.
\begin{theorem}
\label{ThmSobolevBSY} Let $1<p<\infty$ and $0<r<t<1$. Then there exists
$C=C(n,p,r,t)>0,$ such that
\begin{equation}
\Vert f\Vert_{\dot{W}^{s,p}({\mathbb{R}}^{n})}\leq C((s-r)(t-s))^{\frac
{1}{\max\{p,2\}}}\bigg(\frac{1}{(s-r)^{\frac{1}{p}}}\Vert f\Vert_{\dot
{F}_{p,2}^{r}({\mathbb{R}}^{n})}+\frac{1}{(t-s)^{\frac{1}{p}}}\Vert
f\Vert_{\dot{F}_{p,2}^{t}({\mathbb{R}}^{n})}\bigg), \label{rst}%
\end{equation}
for every $s\in(r,t)$. In the limiting cases, $r=0$ and $t=1$, we have
\begin{equation}
\Vert f\Vert_{\dot{W}^{s,p}({\mathbb{R}}^{n})}\leq C(t-s)^{\frac{1}%
{\max\{p,2\}}}\bigg(\frac{1}{s^{\frac{1}{p}}}\Vert f\Vert_{L^{p}({\mathbb{R}%
}^{n})}+\frac{1}{(t-s)^{\frac{1}{p}}}\Vert(-\Delta)^{\frac{t}{2}}f\Vert
_{L^{p}({\mathbb{R}}^{n})}\bigg) \label{WLD}
\end{equation}
for $0 < t < 1$ and
\begin{equation}
\Vert f\Vert_{\dot{W}^{s,p}({\mathbb{R}}^{n})}\leq C(s-r)^{\frac{1}%
{\max\{p,2\}}}\bigg(\frac{1}{(s-r)^{\frac{1}{p}}}\Vert(-\Delta)^{\frac{r}{2}%
}f\Vert_{L^{p}({\mathbb{R}}^{n})}+\frac{1}{(1-s)^{\frac{1}{p}}}\Vert\nabla
f\Vert_{L^{p}({\mathbb{R}}^{n})}\bigg) \label{WLD2}%
\end{equation}
for $0 < r < 1$, respectively.
\end{theorem}
As an immediate consequence of \eqref{WLD} and \eqref{WLD2}, we obtain
\begin{corollary}\label{CorSharp}
Let $0<r<t<1$. Then
\begin{equation}
\sup_{s\in\lbrack r,t)}(t-s)^{\frac{1}{p}-\frac{1}{\max\{p,2\}}}\Vert
f\Vert_{\dot{W}^{s,p}({\mathbb{R}}^{n})}\leq C\Big(\Vert f\Vert_{L^{p}%
({\mathbb{R}}^{n})}+\Vert(-\Delta)^{\frac{t}{2}}f\Vert_{L^{p}({\mathbb{R}}%
^{n})}\Big) \label{Sharp}%
\end{equation}
and
\begin{equation}
\sup_{s\in(r,t]}(s-r)^{\frac{1}{p}-\frac{1}{\max\{p,2\}}}\Vert f\Vert_{\dot
{W}^{s,p}({\mathbb{R}}^{n})}\leq C\Big(\Vert(-\Delta)^{\frac{r}{2}}%
f\Vert_{L^{p}({\mathbb{R}}^{n})}+\Vert\nabla f\Vert_{L^{p}({\mathbb{R}}^{n}%
)}\Big). \label{Sharp2}%
\end{equation}
\end{corollary}
\begin{remark}
\label{RemarkOptim} Theorem \ref{ThmSobolevBSY} is a considerable improvement of Theorem 1.6 in \cite{Braz}: Let $1 < p < \infty$ and $0 \leq r < t \leq 1$. Then there exists $C = C(n, p) > 0$ such that
\begin{equation}
\Vert f\Vert_{\dot{W}^{s,p}({\mathbb{R}}^{n})}\leq C\, \bigg(\frac{1}{(s-r)^{\frac{1}{p}}}\Vert f\Vert_{\dot
{F}_{p,2}^{r}({\mathbb{R}}^{n})}+\frac{1}{(t-s)^{\frac{1}{p}}}\Vert
f\Vert_{\dot{F}_{p,2}^{t}({\mathbb{R}}^{n})}\bigg), \qquad s \in (r, t). \label{rsnew}%
\end{equation}
The estimates provided in Theorem \ref{ThmSobolevBSY} sharpens \eqref{rsnew} due to the appearance of one of the additional prefactors $((s-r)(t-s))^{\frac{1}%
{\max\{p,2\}}}, (t-s)^{\frac{1}{\max\{p, 2\}}}$ and $(s-r)^{\frac{1}{\max\{p, 2\}}}$. Note that the constant $C$ in \eqref{rsnew} does not depend on $r$ and $t$, but this is not the case in \eqref{rst}. The reason behind this additional flexibility in \eqref{rsnew} is explained by its non-optimality.
On the other hand, \eqref{Sharp} sharpens the following inequality obtained in Corollary 1.7 in \cite{Braz}: Let $1 < p < \infty$ and $0 < r < t \leq 1$. Then there exists $C = C(n, p, r) > 0$ such that
\begin{equation}
\sup_{s\in\lbrack r,t)}(t-s)^{\frac{1}{p}}\Vert
f\Vert_{\dot{W}^{s,p}({\mathbb{R}}^{n})}\leq C\Big(\Vert f\Vert_{L^{p}%
({\mathbb{R}}^{n})}+\Vert(-\Delta)^{\frac{t}{2}}f\Vert_{L^{p}({\mathbb{R}}%
^{n})}\Big). \label{1.5}
\end{equation}
In particular, this inequality with $t=1$ gives one of the estimates in the classical Bourgain--Brezis--Mironescu formula \eqref{BBM}. However, \eqref{1.5} is not optimal if $t \in (0, 1)$. In fact, as was already observed in \cite[Remark 1.8]{Braz}, the inequality \eqref{1.5} is only useful with $p < 2$. Specifically, if $p \geq 2$ and $0 < t < 1$, it is well known that
\begin{equation}\label{Tri}
\sup_{s\in\lbrack r,t)} \Vert
f\Vert_{\dot{W}^{s,p}({\mathbb{R}}^{n})}\leq C\Big(\Vert f\Vert_{L^{p}%
({\mathbb{R}}^{n})}+\Vert(-\Delta)^{\frac{t}{2}}f\Vert_{L^{p}({\mathbb{R}}%
^{n})}\Big).
\end{equation}
This obstruction can be overcome via the inequality \eqref{Sharp}. Specifically, if $p\in(1,2)$ then the prefactor $(t-s)^{\frac{1}{p}}$ on the
left-hand side of \eqref{1.5} is now improved by $(t-s)^{\frac{1}{p}-\frac
{1}{2}}$ in \eqref{Sharp}. On the other hand, if $p\in\lbrack2,\infty),$ then
\eqref{Sharp} coincides with the optimal estimate \eqref{Tri}. On the other hand, inequality \eqref{Sharp2}, with $p<2,$ seems to be new.
\end{remark}
We give a positive answer to Open Problem \ref{OpenProblemIntro1}. In fact, we
obtain a stronger version by means of removing the $\theta$-dependance from
the constant $C,$ and working with inhomogeneous norms on both sides of
\eqref{OpenProblemIntro1Eq}. Furthermore, our method works with the more
general family of seminorms formed by $\Vert\cdot\Vert_{\dot{W}%
^{s,p}({\mathbb{R}}^{n}),\alpha}$ (cf. \eqref{DefFractGS}) for arbitrary order of smoothness $\alpha > 0$.
\begin{theorem}
\label{ConjectureBSY} Let $p\in(1,\infty),\,\alpha>0,$ and $0<s<t<\alpha$.
Then there exists a positive constant $C$, independent of $s$ and $t$, such
that
\[
\Vert f\Vert_{L^{p}({\mathbb{R}}^{n})}+\min\{s,\alpha-s\}^{\frac{1}{p}}\Vert
f\Vert_{\dot{W}^{s,p}({\mathbb{R}}^{n}),\alpha}\leq C\Big(\Vert f\Vert
_{L^{p}({\mathbb{R}}^{n})}+\min\{t,\alpha-t\}^{\frac{1}{p}}\Vert f\Vert
_{\dot{W}^{t,p}({\mathbb{R}}^{n}),\alpha}\Big).
\]
In particular, if $\alpha=1$ then (cf. \eqref{FractModuInt})
\[
\Vert f\Vert_{L^{p}({\mathbb{R}}^{n})}+\min\{s,1-s\}^{\frac{1}{p}}\Vert
f\Vert_{\dot{W}^{s,p}({\mathbb{R}}^{n})}\leq C\Big(\Vert f\Vert_{L^{p}%
({\mathbb{R}}^{n})}+\min\{t,1-t\}^{\frac{1}{p}}\Vert f\Vert_{\dot{W}%
^{t,p}({\mathbb{R}}^{n})}\Big)
\]
for all $0<s<t<1$.
\end{theorem}
\section{Interpolation/extrapolation: A primer}
\label{SectionMethod}
The goal of this section is to present our interpolation/extrapolation based
methodology. These methods will be applied in Section \ref{SectionProofs} to prove all the results that were
stated in Section \ref{SectionMR}.
\subsection{Revisiting some classical interpolation formulae}
All the results contained in this subsection are well known. However, for the
purposes of this paper, it is important to establish these results with sharp
control on the dependance with respect to the parameters involved. For the
convenience of the reader, who may not be familiar with interpolation theory,
we give a self-contained exposition, since it is hard to find the material organized in a manner that fits our needs in this paper.
Let $(A_{0},A_{1})$ be a quasi-semi-normed pair. The \emph{real interpolation space}
$(A_{0},A_{1})_{\theta,q},\,\theta\in(0,1),\,q\in(0,\infty]$, is formed by all
the elements $f\in A_{0}+A_{1},$ such that
\begin{equation}
\Vert f\Vert_{(A_{0},A_{1})_{\theta,q}}:=\bigg(\int_{0}^{\infty}\left(
t^{-\theta}K(t,f;A_{0},A_{1})\right) ^{q}\frac{dt}{t}\bigg)^{1/q}%
<\infty,\label{DefInt}%
\end{equation}
where $K(t,f;A_{0},A_{1})$ denotes the \emph{$K$-functional} for the pair
$(A_{0},A_{1}),$ defined for $t>0,$ by
\[
K(t,f;A_{0},A_{1}):=\inf_{f=f_{0}+f_{1}}\Big(\Vert f_{0}\Vert_{A_{0}}+t\Vert
f_{1}\Vert_{A_{1}}\Big).
\]
We shall simply write $K(t,f)$ when the underlying pair of spaces $(A_0, A_1)$ is
understood from the context. We refer to the standard references on
interpolation theory \cite{BennettSharpley, BerghLofstrom, Triebel}. Sometimes
it is more convenient to work with the modified $K$-functional given by
\[
K_{p}(t,f;A_{0},A_{1}):=K_{p}(t,f)=\inf_{f=f_{0}+f_{1}}\Big(\Vert f_{0}%
\Vert_{A_{0}}^{p}+t^{p}\Vert f_{1}\Vert_{A_{1}}^{p}\Big)^{\frac{1}{p}}%
\]
for $p\in(0,\infty)$. Clearly $K(t,f)\approx K_{p}(t,f)$ with equivalence
constants depending only on $p$. Furthermore, it is plain that
\begin{equation}
(A_{0},A_{1})_{\theta,q}=(A_{1},A_{0})_{1-\theta,q}%
\label{OrderedCouplesCommute}%
\end{equation}
with equality of norms.
Let $A$ be a Banach space and $p\in(0,\infty)$. The vector-valued Lebesgue
spaces $L^{p}({\mathbb{R}}^{n};A)$ are formed by all strongly measurable
functions $f:{\mathbb{R}}^{n}\rightarrow A$ such that
\[
\Vert f\Vert_{L^{p}({\mathbb{R}}^{n};A)}:=\bigg(\int_{{\mathbb{R}}^{n}}\Vert
f(x)\Vert_{A}^{p}\,dx\bigg)^{1/p}<\infty.
\]
The next result is well known.
\begin{lemma}
\label{LemmaInterpolationLpVector} Let $(A_{0},A_{1})$ be a pair of Banach
spaces and let $s\in(0,1),p\in(0,\infty)$. Then
\begin{equation*}
(L^{p}({\mathbb{R}}^{n};A_{0}),L^{p}({\mathbb{R}}^{n};A_{1}))_{s,p}%
=L^{p}({\mathbb{R}}^{n};(A_{0},A_{1})_{s,p})
\end{equation*}
and
\begin{equation*}
\Vert f\Vert_{(L^{p}({\mathbb{R}}^{n};A_{0}),L^{p}({\mathbb{R}}^{n}%
;A_{1}))_{s,p}}\approx\Vert f\Vert_{L^{p}({\mathbb{R}}^{n};(A_{0},A_{1}%
)_{s,p})}
\end{equation*}
where the hidden equivalence constants depend only on $p$.
\end{lemma}
\begin{proof}
It is plain that
\[
K(t,f;L^{p}({\mathbb{R}}^{n};A_{0}),L^{p}({\mathbb{R}}^{n};A_{1}%
))^{p}\approx\int_{{\mathbb{R}}^{n}}K_{p}(t,f(x);A_{0},A_{1})^{p}\,dx.
\]
Therefore using Fubini's theorem,
\begin{align*}
\Vert f\Vert_{(L^{p}({\mathbb{R}}^{n};A_{0}),L^{p}({\mathbb{R}}^{n}%
;A_{1}))_{s,p}}^{p} & \approx\int_{0}^{\infty}t^{-sp}\int_{{\mathbb{R}}^{n}%
}K(t,f(x);A_{0},A_{1})^{p}\,dx\,\frac{dt}{t}\\
& \approx\int_{{\mathbb{R}}^{n}}\Vert f(x)\Vert_{(A_{0},A_{1})_{s,p}}%
^{p}\,dx.
\end{align*}
\end{proof}
For $s\in{\mathbb{R}}$ and $p\in(0,\infty)$, the space $\ell_{p}%
^{s}({{\mathbb{Z}}})$ is formed by all the scalar-valued sequences $\xi=(\xi_{j}%
)_{j\in{{\mathbb{Z}}}},$ such that
\[
\Vert\xi\Vert_{\ell_{p}^{s}({{\mathbb{Z}}})}:=\bigg(\sum_{j=-\infty}^{\infty
}2^{jsp}|\xi_{j}|^{p}\bigg)^{1/p}<\infty.
\]
Similarly, we can define the spaces $\ell_{p}^{s}({{\mathbb{N}}}_{0})$ where
${{\mathbb{N}}}_{0}={{\mathbb{N}}}\cup\{0\}$.
\begin{lemma}
\label{Lemma4} Let $-\infty<s_{0}<s_{1}<\infty,\,\theta\in(0,1),s=(1-\theta
)s_{0}+\theta s_{1},$ and let $p,q\in(0,\infty)$. Then
\[
(\ell_{q}^{s_{0}}({{\mathbb{Z}}}),\ell_{q}^{s_{1}}({{\mathbb{Z}}}))_{\theta
,p}=\ell_{p}^{s}({{\mathbb{Z}}}),
\]
with
\[
\bigg(\frac{1}{\theta^{\frac{1}{\max\{p,q\}}}}+\frac{1}{(1-\theta)^{\frac
{1}{\max\{p,q\}}}}\bigg)\Vert\xi\Vert_{\ell_{p}^{s}({{\mathbb{Z}}})}%
\lesssim\Vert\xi\Vert_{(\ell_{q}^{s_{0}}({{\mathbb{Z}}}),\ell_{q}^{s_{1}%
}({{\mathbb{Z}}}))_{\theta,p}}\lesssim\bigg(\frac{1}{\theta^{\frac{1}%
{\min\{p,q\}}}}+\frac{1}{(1-\theta)^{\frac{1}{\min\{p,q\}}}}\bigg)\Vert
\xi\Vert_{\ell_{p}^{s}({{\mathbb{Z}}})}%
\]
uniformly w.r.t. $\theta$. The corresponding result also holds true for
${{\mathbb{N}}}_{0}$-indexed sequences.
\end{lemma}
\begin{proof}
It is well known and easy to see that
\[
K_{q}(t,\xi;\ell_{q}^{s_{0}}({{\mathbb{Z}}}),\ell_{q}^{s_{1}}({{\mathbb{Z}}%
}))=\bigg(\sum_{j=-\infty}^{\infty}[\min\{2^{js_{0}},2^{js_{1}}t\}|\xi
_{j}|]^{q}\bigg)^{1/q}.
\]
By the monotonicity properties of the $K$-functional, the following estimates
hold uniformly w.r.t. $\theta\in(0,1)$,
\begin{align}
\Vert\xi\Vert_{(\ell_{q}^{s_{0}}({{\mathbb{Z}}}),\ell_{q}^{s_{1}}%
({{\mathbb{Z}}}))_{\theta,p}} & \approx\bigg(\sum_{l=-\infty}^{\infty
}2^{l\theta(s_{1}-s_{0})p}K(2^{-l(s_{1}-s_{0})},\xi;\ell_{q}^{s_{0}%
}({{\mathbb{Z}}}),\ell_{q}^{s_{1}}({{\mathbb{Z}}}))^{p}\bigg)^{1/p}\nonumber\\
& \approx\bigg(\sum_{l=-\infty}^{\infty}2^{l\theta(s_{1}-s_{0})p}%
\bigg(\sum_{j=-\infty}^{\infty}[2^{js_{0}}\min\{1,2^{(j-l)(s_{1}-s_{0})}%
\}|\xi_{j}|]^{q}\bigg)^{p/q}\bigg)^{1/p}\nonumber\\
& \approx A+B,\label{A+B}%
\end{align}
where
\[
A:=\bigg(\sum_{l=-\infty}^{\infty}2^{l(\theta-1)(s_{1}-s_{0})p}\bigg(\sum
_{j=-\infty}^{l}2^{js_{1}q}|\xi_{j}|^{q}\bigg)^{p/q}\bigg)^{1/p}%
\]
and
\[
B:=\bigg(\sum_{l=-\infty}^{\infty}2^{l\theta(s_{1}-s_{0})p}\bigg(\sum
_{j=l}^{\infty}2^{js_{0}q}|\xi_{j}|^{q}\bigg)^{p/q}\bigg)^{1/p}.
\]
We consider now two possible cases. Assume first $p\geq q.$
Therefore
\begin{align*}
A & \geq\bigg(\sum_{l=-\infty}^{\infty}2^{l(\theta-1)(s_{1}-s_{0})p}%
\sum_{j=-\infty}^{l}2^{js_{1}p}|\xi_{j}|^{p}\bigg)^{1/p}\\
& =\bigg(\sum_{j=-\infty}^{\infty}2^{js_{1}p}|\xi_{j}|^{p}\sum_{l=j}^{\infty
}2^{l(\theta-1)(s_{1}-s_{0})p}\bigg)^{1/p}\\
& \approx\frac{1}{(1-\theta)^{1/p}}\bigg(\sum_{j=-\infty}^{\infty}2^{jsp}%
|\xi_{j}|^{p}\bigg)^{1/p}.
\end{align*}
By the sharp version of Hardy's inequality (see, e.g., \cite[p. 196]%
{SteinWeiss}),
\begin{align*}
A & =\bigg(\sum_{l=-\infty}^{\infty}\bigg(2^{l(\theta-1)(s_{1}-s_{0})q}%
\sum_{j=-\infty}^{l}2^{js_{1}q}|\xi_{j}|^{q}\bigg)^{p/q}\bigg)^{1/p}\\
& \lesssim\frac{1}{(1-\theta)^{1/q}}\bigg(\sum_{l=-\infty}^{\infty}%
2^{lsp}|\xi_{l}|^{p}\bigg)^{1/p}.
\end{align*}
Hence the term $A$ can be estimated by
\begin{equation}
\frac{1}{(1-\theta)^{1/p}}\bigg(\sum_{j=-\infty}^{\infty}2^{jsp}|\xi_{j}%
|^{p}\bigg)^{1/p}\lesssim A\lesssim\frac{1}{(1-\theta)^{1/q}}\bigg(\sum
_{j=-\infty}^{\infty}2^{j s p}|\xi_{j}|^{p}\bigg)^{1/p}.\label{A}%
\end{equation}
Furthermore, similar estimates also hold for $B$, namely,
\begin{equation}
\frac{1}{\theta^{1/p}}\bigg(\sum_{j=-\infty}^{\infty}2^{jsp}|\xi_{j}%
|^{p}\bigg)^{1/p}\lesssim B\lesssim\frac{1}{\theta^{1/q}}\bigg(\sum
_{j=-\infty}^{\infty}2^{jsp}|\xi_{j}|^{p}\bigg)^{1/p}.\label{B}%
\end{equation}
Combining \eqref{A+B}, \eqref{A} and \eqref{B}, we arrive at the desired
result under $p\geq q$.
Suppose now $p<q$. Then
\begin{align*}
A & \leq\bigg(\sum_{l=-\infty}^{\infty}2^{l(\theta-1)(s_{1}-s_{0})p}%
\sum_{j=-\infty}^{l}2^{js_{1}p}|\xi_{j}|^{p}\bigg)^{1/p}\\
& \approx\frac{1}{(1-\theta)^{1/p}}\bigg(\sum_{j=-\infty}^{\infty}2^{jsp}%
|\xi_{j}|^{p}\bigg)^{1/p}.
\end{align*}
On the other hand, by a sharp version of Copson's inequality (see, e.g., \cite{Leindler}) we find that
\[
A\gtrsim\frac{1}{(1-\theta)^{1/q}}\bigg(\sum_{j=-\infty}^{\infty}2^{jsp}%
|\xi_{j}|^{p}\bigg)^{1/p}.
\]
Hence
\[
\frac{1}{(1-\theta)^{1/q}}\bigg(\sum_{j=-\infty}^{\infty}2^{jsp}|\xi_{j}%
|^{p}\bigg)^{1/p}\lesssim A\lesssim\frac{1}{(1-\theta)^{1/p}}\bigg(\sum
_{j=-\infty}^{\infty}2^{jsp}|\xi_{j}|^{p}\bigg)^{1/p},
\]
with hidden constants of equivalence uniform w.r.t. $\theta\in(0,1)$. The term
$B$ can be estimated similarly. Combining estimates yields the desired result for $p < q$.
\end{proof}
We close this subsection by recalling the well-known characterization of
$\dot{B}_{p,q}^{s}({\mathbb{R}}^{n})$ as an interpolation space between
$L^{p}({\mathbb{R}}^{n})$ and $\dot{W}^{k}_{p}({\mathbb{R}}^{n})$; see, e.g.,
\cite[Chapter 5, Corollary 4.13, page 341]{BennettSharpley} and \cite[Theorem
6.3.1, page 147]{BerghLofstrom}.
\begin{lemma}
\label{LemmaIntBesov} Let $s\in(0,1),\,k\in{{\mathbb{N}}}$ and $p\in
(1,\infty)$. Then\footnote{The limiting values $p=1,\infty$ are also
adimissible.}
\[
(L^{p}({\mathbb{R}}^{n}),\dot{W}_{p}^{k}({\mathbb{R}}^{n}))_{s,q}=\dot
{B}_{p,q}^{sk}({\mathbb{R}}^{n}).
\]
\end{lemma}
\subsection{Extrapolation theory}
Generally speaking, the extrapolation theory of Jawerth and Milman
\cite{JawerthMilman} develops methods to recover the endpoint spaces of a
given parametrized scale of quasi-Banach spaces $\{A_{\theta}:\theta
\in(0,1)\}.$ In particular, it contains a careful analysis of the asymptotic
behavior of the related quasi-norms $\Vert\cdot\Vert_{A_{\theta}}$ as
$\theta\rightarrow0^{+}$ or $\theta\rightarrow1^{-}$ . In \cite{JawerthMilman} it
is shown that $(\theta(1-\theta)p)^{\frac{1}{p}}$ is the appropriate
normalization factor for the real interpolation method $(A_{0},A_{1}%
)_{\theta,p}$ (cf. \eqref{DefInt}). This assertion is well-illustrated by the
following results.
\begin{lemma}
\label{LemmaInterpolation12} Let $p \in(0, \infty)$. Then there exists $C >
0$, which depends only on $p$, such that, for every $\theta\in(0, 1)$ and $f
\in A_{0} \cap A_{1}$,
\[
\|f\|_{(A_{0}, A_{1})_{\theta, p}} \leq C \Big( \frac{1}{\theta^{1/p}}
\|f\|_{A_{0}} + \frac{1}{(1-\theta)^{1/p}} \|f\|_{A_{1}} \Big).
\]
\end{lemma}
\begin{proof}
Since $K(t,f)\leq\min\{\Vert f\Vert_{A_{0}},t\Vert f\Vert_{A_{1}}\}$, it
follows that
\begin{align*}
\Vert f\Vert_{(A_{0},A_{1})_{\theta,p}}^{p} & =\int_{0}^{1}t^{-\theta
p}K(t,f)^{p}\frac{dt}{t}+\int_{1}^{\infty}t^{-\theta p}K(t,f)^{p}\frac{dt}%
{t}\\
& \leq\Vert f\Vert_{A_{1}}^{p}\int_{0}^{1}t^{(1-\theta)p}\frac{dt}{t}+\Vert
f\Vert_{A_{0}}^{p}\int_{1}^{\infty}t^{-\theta p}\frac{dt}{t}\\
& =\Vert f\Vert_{A_{1}}^{p}\frac{1}{(1-\theta)p}+\Vert f\Vert_{A_{0}}%
^{p}\frac{1}{\theta p}.
\end{align*}
\end{proof}
\begin{lemma}
\label{LemmaExtrapolSharp}
\begin{enumerate}
[\upshape(i)]
\item Assume $\theta\in(0, 1)$ and $0 < p \leq q < \infty$. Then
\[
(\theta(1-\theta) p)^{1/p} (A_{0}, A_{1})_{\theta, p} \hookrightarrow
(\theta(1-\theta) q)^{1/q} (A_{0}, A_{1})_{\theta, q}.
\]
\item Assume further that $(A_{0},A_{1})$ is ordered (i.e., $A_{1}%
\hookrightarrow A_{0}$). Let $0<\theta<\eta<1$ and let $p,q\in(0,\infty)$.
Then, the norm of the embedding\footnote{The assumption that the couple
$(A_{0},A_{1})$ is ordered is not restrictive, i.e., a corresponding embedding
result still holds for general couples $(A_{0},A_{1})$. However, this
embedding is simplified when dealing with ordered couples, which will be the
only case of interest in this paper.}
\[
(A_{0},A_{1})_{\eta,p}\hookrightarrow(A_{0},A_{1})_{\theta,q}%
\]
does not exceed
\[
\left\{
\begin{array}
[c]{cl}%
\frac{(\eta(1-\eta))^{1/p}}{(\theta(1-\theta))^{1/q}} & \text{if}\quad p\leq
q,\\
& \\
\frac{1}{(\eta-\theta)^{1/q-1/p}}+\frac{\eta^{1/p}(1-\eta)^{1/p}}{\theta
^{1/q}} & \text{if}\quad p>q,
\end{array}
\right.
\]
uniformly w.r.t. $\eta$ and $\theta$.
\end{enumerate}
\end{lemma}
\begin{proof}
The formula that was stated in (i) can be found in \cite[(3.2), p.
19]{JawerthMilman}. Concerning (ii) with $p=q$, i.e.,
\begin{equation}
(\eta(1-\eta))^{1/p}(A_{0},A_{1})_{\eta,p}\hookrightarrow(\theta
(1-\theta))^{1/p}(A_{0},A_{1})_{\theta,p},\label{KMOrder}%
\end{equation}
we refer to \cite[Corollary 2.1]{KaradzhovMilman}. Furthermore the case $p\leq
q$ in (ii) can be obtained as a direct consequence of (i) and \eqref{KMOrder}.
Next we concentrate on the case $p > q$ in (ii). Since $A_{1} \hookrightarrow
A_{0}$, it is plain to see that $K(t, f) \approx\|f\|_{A_{0}}$ for $t > 1$.
Then
\begin{equation}
\label{LemmaEmbConstantInterpolationProof1}\|f\|_{(A_{0}, A_{1})_{\theta, q}}
\approx\bigg(\int_{0}^{1} (t^{-\theta} K(t, f))^{q} \frac{dt}{t} \bigg)^{1/q}
+ \frac{1}{\theta^{1/q}} \|f\|_{A_{0}}.
\end{equation}
On the one hand, applying H\"{o}lder's inequality, we can estimate the
integral given on the right-hand side of
\eqref{LemmaEmbConstantInterpolationProof1} by
\begin{equation}
\bigg(\int_{0}^{1}(t^{-\theta}K(t,f))^{q}\frac{dt}{t}\bigg)^{1/q}\lesssim
\frac{1}{(\eta-\theta)^{1/q-1/p}}\bigg(\int_{0}^{1}(t^{-\eta}K(t,f))^{p}%
\frac{dt}{t}\bigg)^{1/p}. \label{LemmaEmbConstantInterpolationProof2}%
\end{equation}
On the other hand, using monotonicity properties of the $K$-functional, the
second term in the right-hand side of
\eqref{LemmaEmbConstantInterpolationProof1} can be dominated as follows
\begin{equation}
\Vert f\Vert_{A_{0}}\lesssim\min\{\eta,1-\eta\}^{1/p}\bigg(\int_{0}^{\infty
}(t^{-\eta}K(t,f))^{p}\frac{dt}{t}\bigg)^{1/p}.
\label{LemmaEmbConstantInterpolationProof3}%
\end{equation}
Inserting \eqref{LemmaEmbConstantInterpolationProof2} and
\eqref{LemmaEmbConstantInterpolationProof3} into
\eqref{LemmaEmbConstantInterpolationProof1}, we get
\[
\Vert f\Vert_{(A_{0},A_{1})_{\theta,q}}\lesssim\bigg(\frac{1}{(\eta
-\theta)^{1/q-1/p}}+\frac{\min\{\eta,1-\eta\}^{1/p}}{\theta^{1/q}}\bigg)\Vert
f\Vert_{(A_{0},A_{1})_{\eta,p}}.
\]
\end{proof}
The previous result tells us that the standard interpolation norm $\Vert
\cdot\Vert_{(A_{0},A_{1})_{\theta,p}}$ given in \eqref{DefInt} can be
renormalized by $(\theta(1-\theta)p)^{1/p}\Vert\cdot\Vert_{(A_{0}%
,A_{1})_{\theta,p}}$ so that sharp estimates w.r.t. the interpolation
parameter $\theta$ can now be achieved. Another important example of this
phenomenon occurs in reiteration formulas. Indeed, we recall the following
sharp versions of reiteration properties for the real method.
\begin{lemma}
[{\cite[Theorems 2.11 and 2.12]{KaradzhovMilman} and \cite[Theorem
3]{KaradzhovMilmanXiao}}]\label{LemmaKMX2} Let $s_{0}, s_{1}, \theta\in(0, 1)$
and $p, q \in(0, \infty)$.
\begin{enumerate}
[\upshape(i)]
\item Let $s=(1-\theta)s_{0}+\theta s_{1}$. The following embeddings hold
uniformly w.r.t. $\theta$
\begin{align*}
(\theta(1-\theta))^{-1/\min\{p,q\}}(A_{0},A_{1})_{s,p} & \hookrightarrow
((A_{0},A_{1})_{s_{0},q},(A_{0},A_{1})_{s_{1},q})_{\theta,p}\\
& \hspace{-4cm}\hookrightarrow(\theta(1-\theta))^{-1/\max\{p,q\}}(A_{0}%
,A_{1})_{s,p}.
\end{align*}
\item The following embeddings hold uniformly w.r.t. $s_{1}$ and $\theta$
\begin{align*}
s_{1}^{1/p-1/q} [(1-s_{1})^{-1/q} + (1-\theta)^{-1/\min\{p, q\}}] (A_{0},
A_{1})_{\theta s_{1}, p} & \\
& \hspace{-8cm}\hookrightarrow(A_{0}, (A_{0}, A_{1})_{s_{1}, q})_{\theta, p}
\hookrightarrow s_{1}^{1/p} (1-\theta)^{-1/\max\{p, q\}} (A_{0}, A_{1})_{
\theta s_{1}, p}.
\end{align*}
\end{enumerate}
\end{lemma}
A key role in our arguments is played by the continuity properties of general
interpolation scales obtained in \cite{Milman} (see also
\cite{KaradzhovMilmanXiao}). In particular, we will make use of the following
result for the real interpolation method.
\begin{theorem}
\label{teomarkao}Let $p>0$ and $f\in A_{0}\cap A_{1}$. Then
\begin{equation}
\lim_{\theta\rightarrow1^{-}}(\theta(1-\theta)p)^{\frac{1}{p}}\Vert
f\Vert_{(A_{0},A_{1})_{\theta,p}}=\sup_{t>0}\frac{K(t,f;A_{0},A_{1})}{t}
\label{ProofFractGS1}%
\end{equation}
and
\begin{equation}
\lim_{\theta\rightarrow0^{+}}(\theta(1-\theta)p)^{\frac{1}{p}}\Vert
f\Vert_{(A_{0},A_{1})_{\theta,p}}=\sup_{t>0}K(t,f;A_{0},A_{1}).
\label{ProofFractGS1New}%
\end{equation}
\end{theorem}
\subsection{Butzer seminorms via interpolation}
A key result in our approach is that the family of semi-norms $\Vert\cdot
\Vert_{\dot{W}^{s,p}({\mathbb{R}}^{n}),t}$ introduced in \eqref{DefFractGS}
can be generated (with sharp constants) via interpolation of the classical
pair $(L^{p}({\mathbb{R}}^{n}),\dot{H}^{t,p}({\mathbb{R}}^{n}))$. This is the
assertion contained in the following
\begin{lemma}
\label{LemmaIntFractGS} Let $\theta\in(0,1),t\in(0,\infty)$ and $p\in
(1,\infty)$. Then
\begin{equation}
\Vert f\Vert_{(L^{p}({\mathbb{R}}^{n}),\dot{H}^{t,p}({\mathbb{R}}%
^{n}))_{\theta,p}}\approx\Vert f\Vert_{\dot{W}^{\theta t,p}({\mathbb{R}}%
^{n}),t} \label{LemmaIntFractGSEq1}%
\end{equation}
and
\begin{equation}
\Vert f\Vert_{(L^{p}({\mathbb{R}}^{n}),H^{t,p}({\mathbb{R}}^{n}))_{\theta,p}%
}\approx \Vert f\Vert_{W^{\theta t,p}({\mathbb{R}}^{n}),t}
\label{LemmaIntFractGSEq2}%
\end{equation}
where hidden equivalence constants are independent of $\theta$ (but depending
on $n,p$ and $t$).
\end{lemma}
\begin{remark}
Letting $t=1$ in the previous result (cf. \eqref{SobRiesz} and
\eqref{FractModuInt}), we have
\begin{equation}
\Vert f\Vert_{(L^{p}({\mathbb{R}}^{n}),\dot{W}_{p}^{1}({\mathbb{R}}%
^{n}))_{\theta,p}}\approx\Vert f\Vert_{\dot{W}^{\theta,p}({\mathbb{R}}^{n})}
\label{IntSob1}%
\end{equation}
and, cf. \eqref{DefFractGSIn},
\[
\Vert f\Vert_{(L^{p}({\mathbb{R}}^{n}),W_{p}^{1}({\mathbb{R}}^{n}))_{\theta
,p}}\approx\frac{1}{(\theta(1-\theta)p)^{\frac{1}{p}}}\Vert f\Vert
_{L^{p}({\mathbb{R}}^{n})}+\Vert f\Vert_{\dot{W}^{\theta,p}({\mathbb{R}}^{n}%
)}.
\]
In particular, \eqref{IntSob1} was already shown in \cite[Lemma 1]{Milman}.
\end{remark}
\begin{proof}
[Proof of Lemma \ref{LemmaIntFractGS}]We shall employ the characterizations of
the corresponding $K$-functional given by
\begin{equation}
K(u^{t},f;L^{p}({\mathbb{R}}^{n}),\dot{H}^{t,p}({\mathbb{R}}^{n}))\approx
\frac{1}{u^{n}}\int_{|h|<u}\Vert\Delta_{h}^{t}f\Vert_{L^{p}({\mathbb{R}}^{n}%
)}\,dh\approx\sup_{|h|<u}\Vert\Delta_{h}^{t}f\Vert_{L^{p}({\mathbb{R}}^{n})}
\label{ProofFractGS1*}%
\end{equation}
for $u>0$; this characterization is well known in the classical case
$t=k\in{{\mathbb{N}}}$ (see, e.g., \cite[Chapter 5, (4.42), p. 341]%
{BennettSharpley} and \cite[Theorem 6.7.3]{BerghLofstrom}), for the general
case $t\in(0,\infty)$ we refer to \cite[(1.10)]{Kolomoitsev} (with \cite{Butzer} and
\cite{Wilmes} as a forerunners).
Inserting the first equivalence in \eqref{ProofFractGS1*} into the definition
of interpolation space (cf. \eqref{DefInt}), we have
\begin{equation*}
\Vert f\Vert_{(L^{p}({\mathbb{R}}^{n}),\dot{H}^{t,p}({\mathbb{R}}%
^{n}))_{\theta,p}}\approx\bigg(\int_{0}^{\infty}u^{-\theta tp-np}%
\bigg(\int_{|h|<u}\Vert\Delta_{h}^{t}f\Vert_{L^{p}({\mathbb{R}}^{n}%
)}\,dh\bigg)^{p}\frac{du}{u}\bigg)^{1/p}.
\end{equation*}
Consequently, the desired result \eqref{LemmaIntFractGSEq1} would be
established once we are able to prove (cf. \eqref{DefFractGS})
\begin{equation}
\Vert f\Vert_{\dot{W}^{\theta t,p}({\mathbb{R}}^{n}),t}\approx\bigg(\int%
_{0}^{\infty}u^{-\theta tp-np}\bigg(\int_{|h|<u}\Vert\Delta_{h}^{t}%
f\Vert_{L^{p}({\mathbb{R}}^{n})}\,dh\bigg)^{p}\frac{du}{u}\bigg)^{1/p}%
\label{ClaimDefFractGS}%
\end{equation}
uniformly w.r.t. $\theta\in(0,1)$.
To deal with the estimate $\gtrsim$ in \eqref{ClaimDefFractGS}, we apply
H\"{o}lder's inequality and change the order of integration so that
\begin{align*}
\int_{0}^{\infty}u^{-\theta tp-np}\bigg(\int_{|h|<u}\Vert\Delta_{h}^{t}%
f\Vert_{L^{p}({\mathbb{R}}^{n})}\,dh\bigg)^{p}\frac{du}{u} & \lesssim\int%
_{0}^{\infty}u^{-\theta tp-n}\int_{|h|<u}\Vert\Delta_{h}^{t}f\Vert
_{L^{p}({\mathbb{R}}^{n})}^{p}\,dh\frac{du}{u}\\
& \hspace{-6cm}\approx\int_{{\mathbb{R}}^{n}}|h|^{-\theta tp-n}\Vert
\Delta_{h}^{t}f\Vert_{L^{p}({\mathbb{R}}^{n})}^{p}\,dh=\Vert f\Vert_{\dot
{W}^{\theta t,p}({\mathbb{R}}^{n}),t}^{p}.
\end{align*}
Concerning the estimate $\lesssim$ in \eqref{ClaimDefFractGS}, we use basic
monotonicity properties and the second estimate given in
\eqref{ProofFractGS1*} to get
\begin{align*}
\Vert f\Vert_{\dot{W}^{\theta t,p}({\mathbb{R}}^{n}),t}^{p} & =\int%
_{{\mathbb{R}}^{n}}\frac{\Vert\Delta_{h}^{t}f\Vert_{L^{p}({\mathbb{R}}^{n}%
)}^{p}}{|h|^{\theta tp+n}}\,dh\\
& =\sum_{j=-\infty}^{\infty}\int_{2^{j-1}<|h|\leq2^{j}}\frac{\Vert\Delta
_{h}^{t}f\Vert_{L^{p}({\mathbb{R}}^{n})}^{p}}{|h|^{\theta tp+n}}\,dh\\
& \lesssim\sum_{j=-\infty}^{\infty}2^{-j\theta tp}\sup_{|h|\leq2^{j}}%
\Vert\Delta_{h}^{t}f\Vert_{L^{p}({\mathbb{R}}^{n})}^{p}\\
& \approx\sum_{j=-\infty}^{\infty}2^{-j\theta tp}\bigg(\frac{1}{2^{j n}}%
\int_{|h|<2^{j}}\Vert\Delta_{h}^{t}f\Vert_{L^{p}({\mathbb{R}}^{n}%
)}\,dh\bigg)^{p}.
\end{align*}
Next we focus on \eqref{LemmaIntFractGSEq2}. The $K$-functional for the pair
$(L^{p}({\mathbb{R}}^{n}),H^{t,p}({\mathbb{R}}^{n}))$ can be estimated as (see
\cite[(4.2)]{Wilmes})
\[
K(u,f;L^{p}({\mathbb{R}}^{n}),H^{t,p}({\mathbb{R}}^{n}))\approx\min
\{1,u\}\Vert f\Vert_{L^{p}({\mathbb{R}}^{n})}+\sup_{|h|\leq u^{\frac{1}{t}}%
}\Vert\Delta_{h}^{t}f\Vert_{L^{p}({\mathbb{R}}^{n})}%
\]
for $u>0$ and $f\in L^{p}({\mathbb{R}}^{n})$. Therefore, by
\eqref{ProofFractGS1*},
\[
K(u,f;L^{p}({\mathbb{R}}^{n}),H^{t,p}({\mathbb{R}}^{n}))\approx\min
\{1,u\}\Vert f\Vert_{L^{p}({\mathbb{R}}^{n})}+K(u,f;L^{p}({\mathbb{R}}%
^{n}),\dot{H}^{t,p}({\mathbb{R}}^{n})).
\]
By elementary computations we find
\begin{align*}
\Vert f\Vert_{(L^{p}({\mathbb{R}}^{n}),H^{t,p}({\mathbb{R}}^{n}))_{\theta,p}%
}^{p} & \approx\bigg(\int_{0}^{\infty}u^{-\theta p}\min\{1,u\}^{p}\frac
{du}{u}\bigg)\Vert f\Vert_{L^{p}({\mathbb{R}}^{n})}^{p}+\Vert f\Vert
_{(L^{p}({\mathbb{R}}^{n}),\dot{H}^{t,p}({\mathbb{R}}^{n}))_{\theta,p}}^{p}\\
& =\frac{1}{\theta(1-\theta)p}\Vert f\Vert_{L^{p}({\mathbb{R}}^{n})}%
^{p}+\Vert f\Vert_{(L^{p}({\mathbb{R}}^{n}),\dot{H}^{t,p}({\mathbb{R}}%
^{n}))_{\theta,p}}^{p}.
\end{align*}
Therefore, by \eqref{LemmaIntFractGSEq1},
\[
\Vert f\Vert_{(L^{p}({\mathbb{R}}^{n}),H^{t,p}({\mathbb{R}}^{n}))_{\theta,p}%
}^{p}\approx\frac{1}{\theta(1-\theta)p}\Vert f\Vert_{L^{p}({\mathbb{R}}^{n}%
)}^{p}+\Vert f\Vert_{\dot{W}^{\theta t,p}({\mathbb{R}}^{n}),t}^{p}.
\]
\end{proof}
The following result shows that the seminorm $\Vert f\Vert_{\dot{W}%
^{s,p}({\mathbb{R}}^{n}),t}$ is equivalent to the classical Gagliardo seminorm
$\Vert f\Vert_{\dot{W}^{s,p}({\mathbb{R}}^{n})}$, but the constants of
equivalence blow up as $s\rightarrow t^{-}$.
\begin{lemma}
\label{LemmaFractInt} Let $0<s<t<1$ and $1<p<\infty$. Then
\begin{equation}
\frac{1}{(t-s)^{\frac{1}{\max\{p,2\}}}}\Vert f\Vert_{\dot{W}^{s,p}({\mathbb{R}}^{n}%
)}\lesssim\Vert f\Vert_{\dot{W}^{s,p}({\mathbb{R}}^{n}),t}\lesssim\frac
{1}{(t-s)^{\frac{1}{\min\{p,2\}}}}\Vert f\Vert_{\dot{W}^{s,p}({\mathbb{R}}^{n})},
\label{LemmaFractInt1}%
\end{equation}
where the hidden equivalence constants are independent of $s$.
\end{lemma}
\begin{proof}
Recall the relationships between $\dot{H}^{t,p}({\mathbb{R}}^{n})$ and
$\dot{B}_{p,q}^{t}({\mathbb{R}}^{n})$ (cf. Lemma \ref{TableCoincidences}),
\[
\dot{B}_{p,\min\{p,2\}}^{t}({\mathbb{R}}^{n})\hookrightarrow\dot{H}%
^{t,p}({\mathbb{R}}^{n})\hookrightarrow\dot{B}_{p,\max\{p,2\}}^{t}%
({\mathbb{R}}^{n}),
\]
therefore, by interpolation, the following embeddings hold uniformly w.r.t.
$\theta\in(0,1)$
\begin{equation}
(L^{p}({\mathbb{R}}^{n}),\dot{B}_{p,\min\{p,2\}}^{t}({\mathbb{R}}%
^{n}))_{\theta,p}\hookrightarrow(L^{p}({\mathbb{R}}^{n}),\dot{H}%
^{t,p}({\mathbb{R}}^{n}))_{\theta,p}\hookrightarrow(L^{p}({\mathbb{R}}%
^{n}),\dot{B}_{p,\max\{p,2\}}^{t}({\mathbb{R}}^{n}))_{\theta,p}%
.\label{LemmaFractInt1Proof}%
\end{equation}
Furthermore, by Lemma \ref{LemmaIntBesov}, we can rewrite
\[
(L^{p}({\mathbb{R}}^{n}),\dot{B}_{p,q}^{t}({\mathbb{R}}^{n}))_{\theta
,p}=(L^{p}({\mathbb{R}}^{n}),(L^{p}({\mathbb{R}}^{n}),\dot{W}_{p}%
^{1}({\mathbb{R}}^{n}))_{t,q})_{\theta,p},\qquad q\in(0,\infty),
\]
with related equivalence constants independent of $\theta$. Combining this and
Lemma \ref{LemmaKMX2}(ii), we obtain the embeddings
\begin{align*}
(1-\theta)^{-1/\min\{p,q\}}(L^{p}({\mathbb{R}}^{n}),\dot{W}_{p}^{1}%
({\mathbb{R}}^{n}))_{\theta t,p} & \hookrightarrow(L^{p}({\mathbb{R}}%
^{n}),\dot{B}_{p,q}^{t}({\mathbb{R}}^{n}))_{\theta,p}\\
& \hspace{-4cm}\hookrightarrow(1-\theta)^{-1/\max\{p,q\}}(L^{p}({\mathbb{R}%
}^{n}),\dot{W}_{p}^{1}({\mathbb{R}}^{n}))_{\theta t,p}.%
\end{align*}
In light of Lemma \ref{LemmaIntFractGS} (see also \eqref{IntSob1}), the
previous embeddings turn out to be equivalent to
\begin{equation}
(1-\theta)^{-1/\max\{p,q\}} \|f\|_{\dot{W}^{\theta
t,p}({\mathbb{R}}^{n})} \lesssim \|f\|_{(L^{p}({\mathbb{R}}^{n}),\dot{B}_{p,q}^{t}({\mathbb{R}}%
^{n}))_{\theta,p} } \lesssim (1-\theta)^{-1/\min\{p,q\}} \|f\|_{\dot{W}^{\theta t,p}({\mathbb{R}}^{n})}.\label{LemmaFractInt2}%
\end{equation}
Combining \eqref{LemmaFractInt1Proof} and \eqref{LemmaFractInt2}, we find%
$$
(1-\theta)^{-1/\max\{p,2\}} \|f\|_{\dot{W}^{\theta
t,p}({\mathbb{R}}^{n})} \lesssim \|f\|_{(L^{p}({\mathbb{R}}^{n}),\dot{H}^{t,p}({\mathbb{R}}%
^{n}))_{\theta,p}} \lesssim (1-\theta)^{-1/\min\{p,2\}} \|f\|_{\dot{W}^{\theta t,p}({\mathbb{R}}^{n})}.
$$
Equivalently (cf. \eqref{LemmaIntFractGSEq1})
\[
(1-\theta)^{-1/\max\{p,2\}}\Vert f\Vert_{\dot{W}^{\theta t,p}({\mathbb{R}}%
^{n})}\lesssim\Vert f\Vert_{\dot{W}^{\theta t,p}({\mathbb{R}}^{n}),t}%
\lesssim(1-\theta)^{-1/\min\{p,2\}}\Vert f\Vert_{\dot{W}^{\theta
t,p}({\mathbb{R}}^{n})}
\]
which, after the change of variables $\theta\leftrightarrow\frac{s}{t},$
yields the desired result \eqref{LemmaFractInt1}.
\end{proof}
\section{Proofs of the main results}
\label{SectionProofs}
\subsection{Proof of Proposition \ref{PropQuestion1.11}}
We now recall some results that we need from the theory of Fourier series with
monotonic coefficients.
Suppose that the Fourier series of $f\in L^{1}({\mathbb{T}})$ is given by%
\begin{equation}
f(x)\sim\sum_{\nu=1}^{\infty}c_{\nu}\cos(\nu x),\qquad x\in{\mathbb{T}%
},\label{GenFS}%
\end{equation}
with
\begin{equation}
c_{\nu}\geq c_{\nu+1}\geq\cdots \geq 0\qquad\text{and}\qquad c_{\nu}\rightarrow
0.\label{MonCond}%
\end{equation}
A well-known theorem of Hardy--Littlewood (see, e.g., \cite[p. 129, Vol.
II]{Zygmund}) asserts that $f\in L^{p}({\mathbb{T}})$ if and only if
\[
\sum_{\nu=1}^{\infty}\nu^{p-2}c_{\nu}^{p}<\infty.
\]
Furthermore
\begin{equation}
\Vert f\Vert_{L^{p}({\mathbb{T}})}^{p}\approx\sum_{\nu=1}^{\infty}\nu
^{p-2}c_{\nu}^{p}.\label{HLtheorem}%
\end{equation}
This result has been further extended in \cite{DominguezTikhonov} in order to
deal with other spaces of smooth functions (namely, Besov spaces and Sobolev
spaces), as well as more general classes of Fourier series with monotone-type
coefficients (the so called, general monotone class). In particular, it was
shown in \cite[Theorem 4.25]{DominguezTikhonov} that if $f$ is given by
\eqref{GenFS} with coefficients satisfying \eqref{MonCond} then it belongs to
$H^{t,p}({\mathbb{T}}),\,t\in{\mathbb{R}},\,p\in(1,\infty),$ if and only if
\[
\sum_{\nu=1}^{\infty}\nu^{tp+p-2}c_{\nu}^{p}<\infty.
\]
Moreover,
\begin{equation}
\Vert f\Vert_{H^{t,p}({\mathbb{T}})}^{p}\approx\sum_{\nu=1}^{\infty}%
\nu^{tp+p-2}c_{\nu}^{p}.\label{HLSob}%
\end{equation}
Let $t > 0$ and assume $c_{\nu}=\nu^{-t-1+\frac{1}{p}},\,\nu\in{{\mathbb{N}}}$. By virtue of
\eqref{HLtheorem} and \eqref{HLSob}
\begin{equation}
\Vert f\Vert_{L^{p}({\mathbb{T}})}^{p}\approx\sum_{\nu=1}^{\infty}\nu^{p-2}%
\nu^{-tp-p+1}=\sum_{\nu=1}^{\infty}\nu^{-tp-1}<\infty\label{Counter1*}%
\end{equation}
and
\begin{equation*}
\Vert f\Vert_{H^{t,p}({\mathbb{T}})}^{p}\approx\sum_{\nu=1}^{\infty}%
\nu^{tp+p-2}\nu^{-tp-p+1}=\sum_{\nu=1}^{\infty}\nu^{-1}=\infty,
\end{equation*}
respectively. Consequently, the Fourier series $f$ defined by
\eqref{GenFS} belongs to $L^{p}({\mathbb{T}})$ but not to
$H^{t,p}({\mathbb{T}})$.
To conclude, it remains to show that
\begin{equation}
\Vert f\Vert_{\dot{W}^{s_{k},p}({\mathbb{T}})}\leq C(t-s_{k})^{-\frac{1}{p}%
}\label{ClaimCounterEx}%
\end{equation}
uniformly w.r.t. $k\in{{\mathbb{N}}}$. To proceed, we argue as follows. In
view of \eqref{IntSob1},
\begin{equation}
\Vert f\Vert_{\dot{W}^{s_{k},p}({\mathbb{T}})}\approx\bigg(\int_{0}^{\infty
}u^{-s_{k}p}K(u,f;L^{p}({\mathbb{T}}),\dot{W}_{p}^{1}({\mathbb{T}}))^{p}%
\frac{du}{u}\bigg)^{\frac{1}{p}},\label{Counter1}%
\end{equation}
with hidden constants of equivalence independent of $k$. Next we
estimate $K(u,f;L^{p}({\mathbb{T}}),\dot{W}_{p}^{1}({\mathbb{T}}))$.
For $u>0$, the classical moduli of smoothness of $f\in L^{p}({\mathbb{T}})$ is
defined by
\begin{equation}
\omega(f,u)_{p}:=\sup_{|h|\leq u}\Vert\Delta_{h}f\Vert_{L^{p}({\mathbb{T}}%
)}.\label{DefModuli}%
\end{equation}
Assume further that $f$ satisfies \eqref{GenFS} and \eqref{MonCond}. Then it
is known that $\omega(f,u)_{p}$ can be estimated in terms of the Fourier
coefficients of $f$. Namely,
\begin{equation}
\omega(f,u)_{p}\approx\bigg(\sum_{\nu=1}^{\infty}\min\{1,\nu u\}^{p}\nu
^{p-2}c_{\nu}\bigg)^{\frac{1}{p}},\qquad u\in(0,1);\label{GTGM}%
\end{equation}
(cf. \cite[Theorem 6.2]{Tikhonov} and \cite[Theorem 6.1]{GorbachevTikhonov}
where this is actually done even in the more general context of general
monotone Fourier coefficients.)
Specializing \eqref{GTGM} to $f$ given by \eqref{GenFS}, yields
\[
\omega(f,l^{-1})_{p}\approx l^{-1}\bigg(\sum_{\nu=1}^{l}\nu^{(1-t)p-1}%
\bigg)^{\frac{1}{p}}+\bigg(\sum_{\nu=l}^{\infty}\nu^{-tp-1}\bigg)^{\frac{1}%
{p}}\approx\frac{1}{(t(1-t))^{\frac{1}{p}}}l^{-t}%
\]
for every $l\in{{\mathbb{N}}}$ and $t\in(0,1)$. According to
\eqref{ProofFractGS1*} (with $t=1$) and \eqref{DefModuli}, the previous estimate is equivalent to
\begin{equation}
K(l^{-1},f;L^{p}({\mathbb{T}}),\dot{W}_{p}^{1}({\mathbb{T}}))\approx\frac
{1}{(t(1-t))^{\frac{1}{p}}}l^{-t}.\label{Counter2}%
\end{equation}
Furthermore, we have the trivial estimate
\begin{equation}
K(u,f;L^{p}({\mathbb{T}}),\dot{W}_{p}^{1}({\mathbb{T}}))\leq\Vert
f\Vert_{L^{p}({\mathbb{T}})}.\label{Counter3}%
\end{equation}
Putting together \eqref{Counter1}, \eqref{Counter2}, \eqref{Counter3},
applying monotonicity properties of $K$-functionals, and taking into account
that $s_{k}\uparrow t$, we obtain
\begin{align*}
\Vert f\Vert_{\dot{W}^{s_{k},p}({\mathbb{T}})}^{p} & \lesssim\int_{0}%
^{1}u^{-s_{k}p}K(u,f;L^{p}({\mathbb{T}}),\dot{W}_{p}^{1}({\mathbb{T}}%
))^{p}\frac{du}{u}+\frac{1}{s_{k}p}\Vert f\Vert_{L^{p}({\mathbb{T}})}^{p}\\
& \approx\sum_{l=1}^{\infty}l^{s_{k}p-1}K(l^{-1},f;L^{p}({\mathbb{T}}%
),\dot{W}_{p}^{1}({\mathbb{T}}))^{p}+\Vert f\Vert_{L^{p}({\mathbb{T}})}^{p}\\
& \approx\sum_{l=1}^{\infty}l^{-(t-s_{k})p-1}+\Vert f\Vert_{L^{p}%
({\mathbb{T}})}^{p}\\
& \approx\frac{1}{t-s_{k}}+\Vert f\Vert_{L^{p}({\mathbb{T}})}^{p}\\
& \lesssim\frac{1}{t-s_{k}}.
\end{align*}
This concludes the proof of \eqref{ClaimCounterEx}. \qed
\begin{remark}
\label{RemarkCounterexample} Similar ideas can be used to establish the
counterpart of Proposition \ref{PropQuestion1.11} in ${\mathbb{R}}^{n}$. We
will only sketch the proof. Consider the monotone function
\[
F_{0}(u)=\left\{
\begin{array}
[c]{cl}%
u^{-t-n+\frac{n}{p}} & \text{if}\quad u>1,\\
& \\
1 & \text{if}\quad u\in(0,1],
\end{array}
\right.
\]
and define the radial function $f(x)=f_{0}(|x|),\,x\in{\mathbb{R}}^{n},$ where
$f_{0}$ is the inverse Fourier--Hankel transform of $F_{0}$, i.e.,
\[
f_{0}(u)=\frac{2}{\Gamma\big(\frac{n}{2}\big)(2\sqrt{\pi})^{n}}\int%
_{0}^{\infty}F_{0}(\xi)j_{n/2-1}(u\xi)\xi^{n-1}\,d\xi,
\]
where $j_{\alpha}(u)=\Gamma(\alpha+1)(u/2)^{-\alpha}J_{\alpha}(u)$ is the
normalized Bessel function ($j_{\alpha}(0)=1$), $\alpha\geq-1/2$, and
$J_{\alpha}$ is the classical Bessel function of the first kind of order
$\alpha$. Applying the analogue of Hardy--Littlewood theorem for
$L^{p}({\mathbb{R}}^{n}),\,p>\frac{2 n}{n+1},$ (cf. \cite[Theorem
1]{GorbachevLiflyandTikhonov} and \cite[(4.10)]{GorbachevTikhonov}), we have
\[
\Vert f\Vert_{L^{p}({\mathbb{R}}^{n})}^{p}\approx\int_{0}^{\infty}%
u^{n p-n-1}F_{0}(u)^{p}\,du=\int_{0}^{1}u^{n p-n-1}\,du+\int_{1}^{\infty
}u^{-tp-1}\,du<\infty,
\]
which gives the Euclidean setting counterpart of \eqref{Counter1*}. On the
other hand, using the counterpart of \eqref{HLSob} for $H^{t,p}({\mathbb{R}%
}^{n})$, which may be found in \cite[Theorem 4.8]{DominguezTikhonov}, we have
\begin{align*}
\Vert f\Vert_{H^{t,p}({\mathbb{R}}^{n})}^{p} & \approx\int_{0}^{1}%
u^{n p-n-1}F_{0}^{p}(u)\,du+\int_{1}^{\infty}u^{tp+n p-n-1}F_{0}(u)^{p}\,du\\
& =\int_{0}^{1}u^{n p-n-1}\,du+\int_{1}^{\infty}\frac{du}{u}=\infty.
\end{align*}
The analog of \eqref{GTGM} for the moduli of smoothness $\omega(f,u)_{p}%
=\sup_{|h|\leq u}\Vert\Delta_{h}f\Vert_{L^{p}({\mathbb{R}}^{n})}$ (cf.
\eqref{DefModuli}) was obtained in \cite[Corollary 4.1 and (7.6)]%
{GorbachevTikhonov}, namely,
\[
\omega(f,u)_{p}^{p}\approx u^{p}\int_{0}^{1/u}\xi^{p+n p-n-1}F_{0}^{p}%
(\xi)\,d\xi+\int_{1/u}^{\infty}\xi^{n p-n-1}F_{0}^{p}(\xi)\,d\xi.
\]
In particular (since $t\in(0,1)$),
\[
\omega(f,u)_{p}\approx u^{t}\qquad\text{for}\qquad u\in(0,1).
\]
Following now line by line the arguments given in the proof of Proposition
\ref{PropQuestion1.11}, one can show that there exists $C>0$ such that
\[
\Vert f\Vert_{\dot{W}^{s_{k},p}({\mathbb{R}}^{n})}\leq C(t-s_{k})^{-\frac
{1}{p}}\qquad\text{for every}\qquad k\in{{\mathbb{N}}}.
\]
\end{remark}
\subsection{Proof of Theorem \ref{Theorem343Fract}}
We give a unified proof in the inhomogeneous setting (i.e., \eqref{Aux61} and
\eqref{Aux62}). It follows from Lemma \ref{LemmaIntFractGS} that
\[
(s(t-s))^{\frac{1}{p}}\Vert f\Vert_{(L^{p}({\mathbb{R}}^{n}),H^{t,p}%
({\mathbb{R}}^{n}))_{\frac{s}{t},p}}\approx\Vert f\Vert_{L^{p}({\mathbb{R}%
}^{n})}+(s(t-s))^{\frac{1}{p}}\Vert f\Vert_{\dot{W}^{s,p}({\mathbb{R}}^{n}%
),t}.
\]
Consequently, both \eqref{Aux61} \& \eqref{Aux62} will be established if we
are able to show that
\begin{equation}
\Vert f\Vert_{F_{p,2}^{\bar{r}}({\mathbb{R}}^{n})}\lesssim(s(t-s))^{\frac
{1}{p}}\Vert f\Vert_{(L^{p}({\mathbb{R}}^{n}),H^{t,p}({\mathbb{R}}%
^{n}))_{\frac{s}{t},p}}.\label{51}%
\end{equation}
To prove this inequality, we invoke the method of retractions (cf.
\cite{Triebel}) which shows that the interpolation space $(L^{p}({\mathbb{R}%
}^{n}),H^{t,p}({\mathbb{R}}^{n}))_{\frac{s}{t},p}$ can be isomorphically
identified (with constants of equivalence independent of $s$) with
$(L^{p}({\mathbb{R}}^{n};\ell_{2}({{\mathbb{N}_0}})),L^{p}({\mathbb{R}}^{n}%
;\ell_{2}^{t}({{\mathbb{N}_0}})))_{\frac{s}{t},p}$. In view of the interpolation
formula given in Lemma \ref{LemmaInterpolationLpVector},
\[
\Vert f\Vert_{(L^{p}({\mathbb{R}}^{n};\ell_{2}({{\mathbb{N}_0}})),L^{p}%
({\mathbb{R}}^{n};\ell_{2}^{t}({{\mathbb{N}_0}})))_{\frac{s}{t},p}}\approx\Vert
f\Vert_{L^{p}({\mathbb{R}}^{n};(\ell_{2}({{\mathbb{N}_0}}),\ell_{2}%
^{t}({{\mathbb{N}_0}}))_{\frac{s}{t},p})},
\]
which reduces\footnote{Here we must take into account that $F_{p,2}^{\bar{r}%
}({\mathbb{R}}^{n})$ can be identified with the vector-valued space
$L^{p}({\mathbb{R}}^{n};\ell_{2}^{\bar{r}}({{\mathbb{N}_0}}))$. } \eqref{51} to
the following sharp embedding at the level of sequence spaces
\begin{equation}
(s(t-s))^{\frac{1}{p}}(\ell_{2}({{\mathbb{N}_0}}),\ell_{2}^{t}({{\mathbb{N}_0}%
}))_{\frac{s}{t},p}\hookrightarrow\ell_{2}^{\bar{r}}({{\mathbb{N}_0}%
}).\label{ClaimSeqSpaces}%
\end{equation}
Since (cf. Lemma \ref{Lemma4})
\begin{equation*}
(\bar{r}(t-\bar{r}))^{\frac{1}{2}}(\ell_{2}({{\mathbb{N}_0}}),\ell_{2}%
^{t}({{\mathbb{N}_0}}))_{\frac{\bar{r}}{t},2}=\ell_{2}^{\bar{r}}({{\mathbb{N}_0}%
}),
\end{equation*}
the embedding \eqref{ClaimSeqSpaces} turns out to be equivalent to
\begin{equation}
(s(t-s))^{\frac{1}{p}}(\ell_{2}({{\mathbb{N}_0}}),\ell_{2}^{t}({{\mathbb{N}_0}%
}))_{\frac{s}{t},p}\hookrightarrow(\bar{r}(t-\bar{r}))^{\frac{1}{2}}(\ell
_{2}({{\mathbb{N}_0}}),\ell_{2}^{t}({{\mathbb{N}_0}}))_{\frac{\bar{r}}{t}%
,2}.\label{ClaimSeqSpacesNew}%
\end{equation}
It remains to show the validity of \eqref{ClaimSeqSpacesNew}. To do this, we
distinguish two possible cases. Suppose first that $p\leq2$, then by Lemma
\ref{LemmaExtrapolSharp}(ii) (note that $\bar{r}<s$),
\begin{equation*}
(s(t-s))^{\frac{1}{p}}(\ell_{2}({{\mathbb{N}_0}}),\ell_{2}^{t}({{\mathbb{N}_0}%
}))_{\frac{s}{t},p}\hookrightarrow(\bar{r}(t-\bar{r}))^{\frac{1}{2}}(\ell
_{2}({{\mathbb{N}_0}}),\ell_{2}^{t}({{\mathbb{N}_0}}))_{\frac{\bar{r}}{t},2}.
\end{equation*}
Suppose now that $p>2$. Applying again Lemma \ref{LemmaExtrapolSharp}(ii), we
have
\begin{equation}
\bigg(\frac{1}{(s-\bar{r})^{1/2-1/p}}+\frac{s^{1/p}(t-s)^{1/p}}{\bar{r}^{1/2}%
}\bigg)(\ell_{2}({{\mathbb{N}_0}}),\ell_{2}^{t}({{\mathbb{N}_0}}))_{\frac{s}{t}%
,p}\hookrightarrow(\ell_{2}({{\mathbb{N}_0}}),\ell_{2}^{t}({{\mathbb{N}_0}%
}))_{\frac{\bar{r}}{t},2}.\label{G2}%
\end{equation}
Furthermore, it is easy to check that under the assumptions (1) \& (2) in Theorem \ref{Theorem343Fract}, the
constant in \eqref{G2} behaves like
\begin{equation*}
\frac{1}{(s-\bar{r})^{1/2-1/p}}+\frac{s^{1/p}(t-s)^{1/p}}{\bar{r}^{1/2}%
}\approx(s(t-s))^{\frac{1}{p}}(\bar{r}(t-\bar{r}))^{-\frac{1}{2}%
},
\end{equation*}
which yields the desired embedding \eqref{ClaimSeqSpacesNew}.
For the homogeneous setting (i.e., \eqref{Aux6} and \eqref{Aux6new}), observe
that it is enough to deal with $r=\bar{r}$, since for $r\in(0,\bar{r})$ we
have
\[
\Vert f\Vert_{\dot{F}_{p,2}^{r}({\mathbb{R}}^{n})}\leq\Vert f\Vert_{\dot
{F}_{p,2}^{0}({\mathbb{R}}^{n})}+\Vert f\Vert_{\dot{F}_{p,2}^{\bar{r}%
}({\mathbb{R}}^{n})}\leq C(\Vert f\Vert_{L^{p}({\mathbb{R}}^{n})}+\Vert
f\Vert_{\dot{F}_{p,2}^{\bar{r}}({\mathbb{R}}^{n})}).
\]
Henceforth, we concentrate on \eqref{Aux6} and \eqref{Aux6new} with $r=\bar
{r}$. The methodology for the inhomogeneous setting given above (i.e.,
\eqref{Aux61} and \eqref{Aux62}) yields (cf. \eqref{51})
\begin{equation}
\Vert f\Vert_{\dot{F}_{p,2}^{\bar{r}}({\mathbb{R}}^{n})}\lesssim
(s(t-s))^{\frac{1}{p}}\Vert f\Vert_{(L^{p}({\mathbb{R}}^{n}),H^{t,p}%
({\mathbb{R}}^{n}))_{\frac{s}{t},p}}.\label{FrHom}%
\end{equation}
Indeed, in this case the analog of \eqref{ClaimSeqSpaces} reads
\begin{equation}
(s(t-s))^{\frac{1}{p}}(\ell_{2}({{\mathbb{Z}}}),\ell_{2}^{t}({{\mathbb{Z}}%
}))_{\frac{s}{t},p}\hookrightarrow\ell_{2}^{\bar{r}}({{\mathbb{Z}}%
})\label{EmbNonOrder}%
\end{equation}
where
\[
\bar{r}=\left\{
\begin{array}
[c]{cl}%
t-\Lambda(t-s), & \text{if}\quad s\rightarrow t^{-},\\
& \\
\frac{1}{\Lambda}s, & \text{if}\quad s\rightarrow0^{+}.
\end{array}
\right.
\]
Here the couple $(\ell_{2}({{\mathbb{Z}}}),\ell_{2}^{t}({{\mathbb{Z}}}))$ is
not ordered and we cannot invoke Lemma \ref{LemmaExtrapolSharp}(ii) directly
to deal with \eqref{EmbNonOrder}. However, this difficulty can be overcome as
follows. We write ${{\mathbb{Z}}}={{\mathbb{N}}}_{0}\cup{{\mathbb{N}}}_{-}$
where ${{\mathbb{N}}}_{-}:=\{-j:j\in{{\mathbb{N}}}\}$. Since $\ell_{2}%
^{t}({{\mathbb{N}}}_{0})\hookrightarrow\ell_{2}({{\mathbb{N}}}_{0})$ and
$\ell_{2}({{\mathbb{N}}}_{-})\hookrightarrow\ell_{2}^{t}({{\mathbb{N}}}_{-})$,
it follows from \eqref{ClaimSeqSpaces} that
\begin{equation}
(s(t-s))^{\frac{1}{p}}(\ell_{2}({{\mathbb{N}}}_{0}),\ell_{2}^{t}({{\mathbb{N}%
}}_{0}))_{\frac{s}{t},p}\hookrightarrow\ell_{2}^{\bar{r}}({{\mathbb{N}}}%
_{0})\label{Decomp1}%
\end{equation}
and, by \eqref{OrderedCouplesCommute}, and a simple change of variables,
\begin{equation}
(s(t-s))^{\frac{1}{p}}(\ell_{2}({{\mathbb{N}}}_{-}),\ell_{2}^{t}({{\mathbb{N}%
}}_{-}))_{\frac{s}{t},p}\hookrightarrow\ell_{2}^{\bar{r}}({{\mathbb{N}}}%
_{-}).\label{Decomp2}%
\end{equation}
Let $\xi=(\xi)_{j\in{{\mathbb{Z}}}}$ and consider the related sequences
$\xi^{1}=(\xi_{j})_{j\in{{\mathbb{N}}}_{0}}$ and $\xi^{2}=(\xi_{j}%
)_{j\in{{\mathbb{N}}}_{-}}$. By triangle inequality and \eqref{Decomp1},
\eqref{Decomp2}, we derive
\begin{align*}
\Vert\xi\Vert_{\ell_{2}^{\bar{r}}({{\mathbb{Z}}})} & \leq\Vert\xi^{1}%
\Vert_{\ell_{2}^{\bar{r}}({{\mathbb{N}}}_{0})}+\Vert\xi^{2}\Vert_{\ell
_{2}^{\bar{r}}({{\mathbb{N}}}_{-})}\\
& \leq C(s(t-s))^{\frac{1}{p}}(\Vert\xi^{1}\Vert_{(\ell_{2}({{\mathbb{N}}%
}_{0}),\ell_{2}^{t}({{\mathbb{N}}}_{0}))_{\frac{s}{t},p}}+\Vert\xi^{2}%
\Vert_{(\ell_{2}({{\mathbb{N}}}_{-}),\ell_{2}^{t}({{\mathbb{N}}}_{-}%
))_{\frac{s}{t},p}})\\
& \leq2C(s(t-s))^{\frac{1}{p}}\Vert\xi\Vert_{(\ell_{2}({{\mathbb{Z}}}%
),\ell_{2}^{t}({{\mathbb{Z}}}))_{\frac{s}{t},p}},
\end{align*}
where the last step follows immediately from the interpolation property
applied to the canonical projections $\xi\mapsto\xi^{i},\,i=1,2$. The proof of
\eqref{EmbNonOrder} is complete establishing that \eqref{FrHom} holds. Now the
rest of the proof follows line by line the arguments provided for the
inhomogeneous case. \qed
\subsection{Proof of Theorem \ref{ThemQuestion1.11}}
Without loss of generality, we may assume that $t-s_{k}\leq\frac{t}{4}$.
According to \eqref{Aux6} there is $C=C(n,p,t)>0$ such that, for every $r\leq
t-2(t-s_{k})$,
\[
\Vert f_{k}\Vert_{\dot{F}_{p,2}^{r}({\mathbb{R}}^{n})}\leq C\Big(\Vert
f_{k}\Vert_{L^{p}({\mathbb{R}}^{n})}+(t-s_{k})^{\frac{1}{p}}\Vert f_{k}%
\Vert_{\dot{W}^{s_{k},p}({\mathbb{R}}^{n}),t}\Big)\leq C\Lambda.
\]
Taking limits as $k\rightarrow\infty,$ and noting that $\lim_{k\rightarrow
\infty}t-2(t-s_{k})=t$, the previous estimate yields
\[
\limsup_{k\rightarrow\infty}\Big(\Vert f_{k}\Vert_{L^{p}({\mathbb{R}}^{n}%
)}+\Vert f_{k}\Vert_{\dot{F}_{p,2}^{r}({\mathbb{R}}^{n})}\Big)\lesssim
\Lambda\qquad\text{for all}\qquad r\in(0,t).
\]
Therefore, one can apply \cite[Lemma 2.6]{Braz} to derive $f\in H^{t,p}%
({\mathbb{R}}^{n})$ and
\[
\Vert(-\Delta)^{\frac{t}{2}}f\Vert_{L^{p}({\mathbb{R}}^{n})}\lesssim\Lambda.
\]
\qed
\begin{remark}
It may be instructive to revisit the counterexample provided in Proposition
\ref{PropQuestion1.11} to show the important role played by the seminorms
$\Vert\cdot\Vert_{\dot{W}^{s,p}({\mathbb{R}}^{n}),t}$ in Theorem
\ref{ThemQuestion1.11}. Consider the Fourier series $f$ defined by
\eqref{FourierSeries}. It was shown in Proposition \ref{PropQuestion1.11} that
the condition
\[
\sup_{k\in{{\mathbb{N}}}}\,(t-s_{k})^{\frac{1}{p}}\Vert f\Vert_{\dot{W}%
^{s_{k},p}({\mathbb{T}})}<\infty
\]
is satisfied but $f\not \in H^{t,p}({\mathbb{T}})$. Next we check that this
example does not contradict Theorem \ref{ThemQuestion1.11} since
\begin{equation}
\limsup_{k\rightarrow\infty}\,(t-s_{k})^{\frac{1}{p}}\Vert f\Vert_{\dot
{W}^{s_{k},p}({\mathbb{T}}),t}=\infty. \label{ClaimCounter}%
\end{equation}
Indeed, applying \cite[Theorem 6.2]{Tikhonov} one can estimate, for every
$l\in{{\mathbb{N}}}$,
\begin{align*}
K(l^{-t},f;L^{p}({\mathbb{T}}),\dot{H}^{t,p}({\mathbb{T}})) & \approx
l^{-t}\bigg(\sum_{\nu=1}^{l}\nu^{tp+p-2}\nu^{-tp-p+1}\bigg)^{1/p}%
+\bigg(\sum_{\nu=l}^{\infty}\nu^{p-2}\nu^{-tp-p+1}\bigg)^{1/p}\\
& \approx l^{-t}(1+\log l)^{1/p}%
\end{align*}
and thus, by Lemma \ref{LemmaIntFractGS},
\begin{align*}
\Vert f\Vert_{\dot{W}^{s_{k},p}({\mathbb{T}}),t}^{p} & \approx\Vert
f\Vert_{(L^{p}({\mathbb{T}}),\dot{H}^{t,p}({\mathbb{T}}))_{\frac{s_{k}}{t},p}%
}^{p}\\
& \gtrsim\sum_{l=0}^{\infty}2^{ls_{k}p}K(2^{-lt},f;L^{p}({\mathbb{T}}%
),\dot{H}^{t,p}({\mathbb{T}}))^{p}\\
& \approx\sum_{l=0}^{\infty}2^{-l(t-s_{k})p}(1+l)\\
& \gtrsim(t-s_{k})^{-1}\sum_{l=\left\lfloor {\frac{1}{t-s_{k}}}\right\rfloor
}^{\infty}2^{-l(t-s_{k})p}\\
& \approx\frac{(t-s_{k})^{-1}}{1-2^{-(t-s_{k})p}}\approx(t-s_{k})^{-2}.
\end{align*}
Hence
\[
(t-s_{k})^{1/p}\Vert f\Vert_{\dot{W}^{s_{k},p}({\mathbb{T}}),t}\gtrsim
(t-s_{k})^{-1/p}%
\]
which yields \eqref{ClaimCounter} (since $s_{k}\uparrow t$).
\end{remark}
\subsection{Proof of Theorem \ref{TheoremFBBM}}
Let $(A_{0},A_{1})=(L^{p}({\mathbb{R}}^{n}),\dot{H}^{t,p}({\mathbb{R}}^{n}))$.
Recall that for this couple for $f\in$ $H
^{t,p}({\mathbb{R}}^{n}),$ (cf. \cite[Corollary 10, page 75]{Wilmes},
\cite[Section 1.3, Property 1]{Kolomoitsev})
\[
\sup_{t>0}\frac{K(t,f;L^{p}({\mathbb{R}}^{n}),\dot{H}^{t,p}({\mathbb{R}}%
^{n}))}{t}\approx\Vert(-\Delta)^{\frac{t}{2}}f\Vert_{L^{p}({\mathbb{R}}^{n})}%
\]
and
\[
\sup_{t>0}K(t,f;L^{p}({\mathbb{R}}^{n}),\dot{H}^{t,p}({\mathbb{R}}%
^{n}))\approx\Vert f\Vert_{L^{p}({\mathbb{R}}^{n})}.
\]
Therefore \eqref{FractGSAssertion} (respectively, \eqref{FractGSAssertionMS})
is an immediate consequence of Lemma \ref{LemmaIntFractGS} and
\eqref{ProofFractGS1} (respectively, \eqref{ProofFractGS1New}).
Concerning the inhomogeneous setting (i.e., \eqref{FractGSAssertionIn} and \eqref{FractGSAssertionMSIn}), the proofs are immediate consequences of their homogeneous counterparts (i.e., \eqref{FractGSAssertion} and \eqref{FractGSAssertionMS}) applied to \eqref{DefFractGSIn}.
\qed
\subsection{Proof of Theorem \ref{ThmSobolevBSY}}
Assume $0<r<t<1$. Applying Lemma \ref{LemmaInterpolation12} to the
Triebel--Lizorkin pair $(A_{0},A_{1})=(\dot{F}_{p,2}^{r}({\mathbb{R}}%
^{n}),\dot{F}_{p,2}^{t}({\mathbb{R}}^{n}))$, one has
\[
\Vert f\Vert_{(\dot{F}_{p,2}^{r}({\mathbb{R}}^{n}),\dot{F}_{p,2}%
^{t}({\mathbb{R}}^{n}))_{\theta,p}}\lesssim\frac{1}{\theta^{1/p}}\Vert
f\Vert_{\dot{F}_{p,2}^{r}({\mathbb{R}}^{n})}+\frac{1}{(1-\theta)^{1/p}}\Vert
f\Vert_{\dot{F}_{p,2}^{t}({\mathbb{R}}^{n})}%
\]
for every $\theta\in(0,1)$. Since (cf. Lemma \ref{TableCoincidences}(i))
\[
\dot{F}_{p,2}^{r}({\mathbb{R}}^{n})\hookrightarrow\dot{B}_{p,\max\{p,2\}}%
^{r}({\mathbb{R}}^{n})\qquad\text{and}\qquad\dot{F}_{p,2}^{t}({\mathbb{R}}%
^{n})\hookrightarrow\dot{B}_{p,\max\{p,2\}}^{t}({\mathbb{R}}^{n}),
\]
the previous inequality implies
\begin{equation}
\Vert f\Vert_{(\dot{B}_{p,\max\{p,2\}}^{r}({\mathbb{R}}^{n}),\dot{B}%
_{p,\max\{p,2\}}^{t}({\mathbb{R}}^{n}))_{\theta,p}}\lesssim\frac{1}%
{\theta^{1/p}}\Vert f\Vert_{\dot{F}_{p,2}^{r}({\mathbb{R}}^{n})}+\frac
{1}{(1-\theta)^{1/p}}\Vert f\Vert_{\dot{F}_{p,2}^{t}({\mathbb{R}}^{n})}.
\label{ProofThmSobolevBSY1}%
\end{equation}
Next we compute the interpolation norm given in the left-hand side of
\eqref{ProofThmSobolevBSY1}. Indeed, since $r,t\in(0,1)$, we can invoke Lemma
\ref{LemmaIntBesov} to state
\[
\dot{B}_{p,\max\{p,2\}}^{r}({\mathbb{R}}^{n})=(L^{p}({\mathbb{R}}^{n}),\dot
{W}_{p}^{1}({\mathbb{R}}^{n}))_{r,\max\{p,2\}}%
\]
and
\[
\dot{B}_{p,\max\{p,2\}}^{t}({\mathbb{R}}^{n})=(L^{p}({\mathbb{R}}^{n}),\dot
{W}_{p}^{1}({\mathbb{R}}^{n}))_{t,\max\{p,2\}}.
\]
Therefore, in light of Lemmas \ref{LemmaKMX2}(i) and \eqref{IntSob1}, we get,
uniformly w.r.t. $\theta$,
\begin{align*}
\Vert f\Vert_{(\dot{B}_{p,\max\{p,2\}}^{r}({\mathbb{R}}^{n}),\dot{B}%
_{p,\max\{p,2\}}^{t}({\mathbb{R}}^{n}))_{\theta,p}} & \approx\Vert
f\Vert_{((L^{p}({\mathbb{R}}^{n}),\dot{W}_{p}^{1}({\mathbb{R}}^{n}%
))_{r,\max\{p,2\}},(L^{p}({\mathbb{R}}^{n}),\dot{W}_{p}^{1}({\mathbb{R}}%
^{n}))_{t,\max\{p,2\}})_{\theta,p}}\\
& \hspace{-4cm}\gtrsim(\theta(1-\theta))^{-1/\max\{p,2\}}\Vert f\Vert
_{(L^{p}({\mathbb{R}}^{n}),\dot{W}_{p}^{1}({\mathbb{R}}^{n}))_{(1-\theta
)r+\theta t,p}}\\
& \hspace{-4cm}\approx(\theta(1-\theta))^{-1/\max\{p,2\}}\Vert f\Vert
_{\dot{W}^{(1-\theta)r+\theta t,p}({\mathbb{R}}^{n})}.
\end{align*}
Combining this and \eqref{ProofThmSobolevBSY1}, we arrive at
\begin{equation}
\frac{1}{(\theta(1-\theta))^{1/\max\{p,2\}}}\Vert f\Vert_{\dot{W}%
^{(1-\theta)r+\theta t,p}({\mathbb{R}}^{n})}\lesssim\frac{1}{\theta^{1/p}%
}\Vert f\Vert_{\dot{F}_{p,2}^{r}({\mathbb{R}}^{n})}+\frac{1}{(1-\theta)^{1/p}%
}\Vert f\Vert_{\dot{F}_{p,2}^{t}({\mathbb{R}}^{n})}.
\label{ProofThmSobolevBSY2}%
\end{equation}
Given $s\in(r,t)$, we choose $\theta\in(0,1)$ such that $s=(1-\theta)r+\theta
t$ and make the corresponding change of variables in
\eqref{ProofThmSobolevBSY2}, i.e.,
\[
\frac{1}{((s-r)(t-s))^{1/\max\{p,2\}}}\Vert f\Vert_{\dot{W}^{s,p}({\mathbb{R}%
}^{n})}\lesssim\frac{1}{(s-r)^{1/p}}\Vert f\Vert_{\dot{F}_{p,2}^{r}%
({\mathbb{R}}^{n})}+\frac{1}{(t-s)^{1/p}}\Vert f\Vert_{\dot{F}_{p,2}%
^{t}({\mathbb{R}}^{n})}.
\]
The proofs in the limiting cases $r=0$ and $t=1$ (i.e., \eqref{WLD} and \eqref{WLD2} respectively) are easier and we
omit further details. \qed
\subsection{Proof of Theorem \ref{ConjectureBSY}}
Since $H^{\alpha,p}({\mathbb{R}}^{n})\hookrightarrow L^{p}({\mathbb{R}}^{n})$,
it follows from Lemma \ref{LemmaExtrapolSharp}(ii) that
\[
(s(\alpha-s))^{1/p}\Vert f\Vert_{(L^{p}({\mathbb{R}}^{n}),H^{\alpha
,p}({\mathbb{R}}^{n}))_{\frac{s}{\alpha},p}}\lesssim(t(\alpha-t))^{1/p}\Vert
f\Vert_{(L^{p}({\mathbb{R}}^{n}),H^{\alpha,p}({\mathbb{R}}^{n}))_{\frac
{t}{\alpha},p}}.
\]
Combining with \eqref{LemmaIntFractGSEq2} yields
\begin{align*}
\Vert f\Vert_{L^{p}({\mathbb{R}}^{n})}+(s(\alpha-s))^{1/p}\Vert f\Vert
_{\dot{W}^{s,p}({\mathbb{R}}^{n}),\alpha} & \approx (s(1-s))^{1/p}\Vert
f\Vert_{(L^{p}({\mathbb{R}}^{n}),H^{\alpha,p}({\mathbb{R}}^{n}))_{\frac
{s}{\alpha},p}}\\
& \hspace{-5cm}\lesssim(t(1-t))^{1/p}\Vert f\Vert_{(L^{p}({\mathbb{R}}%
^{n}),H^{\alpha,p}({\mathbb{R}}^{n}))_{\frac{t}{\alpha},p}}\\
& \hspace{-5cm}\approx\Vert f\Vert_{L^{p}({\mathbb{R}}^{n})}+(t(\alpha
-t))^{1/p}\Vert f\Vert_{\dot{W}^{t,p}({\mathbb{R}}^{n}),\alpha}.
\end{align*}
The desired result now follows since $s(\alpha-s)\approx\min\{s,\alpha-s\}$. \qed
\section*{Appendix A: Sharp versions of fractional Sobolev inequalities}
For $p \in (0, \infty)$ and $q \in (0, \infty]$, the \emph{Lorentz space} $L^{p, q}(\R^n)$ is formed by all measurable functions $f$ defined on $\R^n$ such that
$$
\|f\|_{L^{p, q}(\R^n)} := \bigg(\int_0^\infty \big(u^{\frac{1}{p}} f^*(u) \big)^q \frac{du}{u} \bigg)^{\frac{1}{q}} < \infty
$$
(with the usual modification if $q=\infty$).
As usual, $f^*$ denotes the non-increasing rearrangement of $f$. By $f^{**}$ we denote the maximal function given by $f^{**}(u) := \frac{1}{u} \int_0^{u} f^*(v) \, dv$. We refer to \cite{BennettSharpley} and \cite{BerghLofstrom} for detailed accounts on Lorentz spaces.
Let $0 < s < 1, \, 1 \leq p < \frac{n}{s}$ and $p^* = \frac{n p}{n- s p}$. The Bourgain--Brezis--Mironescu--Maz'ya--Shasposhnikova formula (cf. \cite{BBM1} and \cite{Mazya}) claims that there exists $C= C(n, p) > 0$ such that
\begin{equation}\label{BBMnew}
\|f\|^p_{L^{p^{*}, p}(\R^n)} \leq C \frac{s (1-s)}{(n- s p)^{p}} \|f\|_{\dot{W}^{s, p}(\R^n)}^p;
\end{equation}
see also \cite{KaradzhovMilmanXiao}.
This inequality can be considered as a sharp version of the classical Sobolev embedding
\begin{equation}\label{Sob}
\dot{W}^1_p(\R^n) \hookrightarrow L^{\frac{n p}{n- p}, p}(\R^n), \qquad 1 \leq p < n.
\end{equation}
Indeed, \eqref{Sob} follows from \eqref{BBMnew} by taking limits as $s \to 1^{-}$ (cf. \eqref{BBM}).
Let $t \in (0, 1)$ be fixed. Note that the formula \eqref{BBMnew} does not provide any insight if $s \to t^{-}$. So the natural question here is: can we obtain an analogue of \eqref{BBMnew} for fractional smoothness $t$? Or equivalently, what is the fractional counterpart of \eqref{BBMnew} related to the classical embedding
\begin{equation}\label{HLSFract}
\dot{H}^{t, p}(\R^n) \hookrightarrow L^{\frac{n p}{n- t p}, p}(\R^n), \qquad 1 < p < \frac{n}{t}?
\end{equation}
The answer is given again in terms of the Butzer seminorms $\|\cdot\|_{\dot{W}^{s, p}(\R^n), t}$ (cf. \eqref{DefFractGS}).
\begin{theorem}\label{ThmBBMt}
Let $0 < s < t \leq 1, 1 < p < \frac{n}{s}$ and $p^{*} = \frac{n p}{n-s p}$. Assume $f \in W^{s, p}(\R^n)$. Then there exists $C = C(n, p, t) > 0$ such that
\begin{equation}\label{BBMFract}
\|f\|^p_{L^{p^{*}, p}(\R^n)} \leq C \frac{s (t-s)}{(n- s p)^{p}} \|f\|_{\dot{W}^{s, p}(\R^n), t}^p.
\end{equation}
In particular, if $t=1$ then one recovers \eqref{BBMnew}.
\end{theorem}
\begin{remark}
Let $t > 0$ and $1 < p < \frac{n}{t}$. Taking limits as $s \to t^{-}$ in \eqref{BBMFract} and applying Theorem \ref{TheoremFBBM}, we recover \eqref{HLSFract}.
\end{remark}
\begin{proof}[Proof of Theorem \ref{ThmBBMt}]
\textsc{Case 1}: Assume $t < \frac{n}{p}$ and $s \in (0, t)$. Under these assumptions, the Hardy--Littlewood--Sobolev theorem (cf. \cite[p. 139]{Oneil}) asserts
$$
\dot{H}^{t, p}(\R^n) \hookrightarrow L^{\frac{n p}{n- t p}, p}(\R^n).
$$
If we interpolate this embedding with the trivial one $L^p(\R^n) \hookrightarrow L^p(\R^n)$, we arrive at
\begin{equation}\label{ForClaim-1}
\|f\|_{ (L^p(\R^n), L^{\frac{n p}{n- t p}, p}(\R^n))_{\frac{s}{t}, p}} \lesssim \|f\|_{(L^p(\R^n), \dot{H}^{t, p}(\R^n))_{\frac{s}{t}, p}}
\end{equation}
uniformly with respect to $s \in (0, t)$. In light of Lemma \ref{LemmaIntFractGS},
\begin{equation}\label{ForClaim0}
\|f\|_{(L^p(\R^n), \dot{H}^{t, p}(\R^n))_{\frac{s}{t}, p}} \approx \|f\|_{\dot{W}^{s, p}(\R^n), t}.
\end{equation}
Next we show that
\begin{equation}\label{ForClaim}
\|f\|_{L^{p^{*}, p}(\R^n)} \approx (s (t-s))^{\frac{1}{p}} \|f\|_{ (L^p(\R^n), L^{\frac{n p}{n- t p}, p}(\R^n))_{\frac{s}{t}, p}}.
\end{equation}
Indeed, since $L^{\frac{n p}{n- t p}, p}(\R^n) = (L^p(\R^n), L^\infty(\R^n))_{\frac{t p}{n}, p}$ (see, e.g., \cite[Theorem 5.2.1, p. 109]{BerghLofstrom}), we can write
$$
(L^p(\R^n), L^{\frac{n p}{n- t p}, p}(\R^n))_{\frac{s}{t}, p} = (L^p(\R^n), (L^p(\R^n), L^\infty(\R^n))_{\frac{t p}{n}, p})_{\frac{s}{t}, p}
$$
with hidden constants of equivalence independent of $s$. Invoking now the sharp reiteration formula given in Lemma \ref{LemmaKMX2}(ii), we get
\begin{equation}\label{For1}
(L^p(\R^n), L^{\frac{n p}{n- t p}, p}(\R^n))_{\frac{s}{t}, p} = (t-s)^{-\frac{1}{p}} (L^p(\R^n), L^\infty(\R^n))_{\frac{s p}{n}, p}.
\end{equation}
Furthermore, using the well-known characterization of the $K$-functional (cf. \cite[Theorem 5.2.1, p. 109]{BerghLofstrom})
$$
K(u, f; L^p(\R^n), L^\infty(\R^n)) \approx \bigg(\int_0^{u^p} (f^*(v))^p \, dv \bigg)^{\frac{1}{p}}
$$
and applying Fubini's theorem, we obtain
\begin{align*}
\|f\|_{(L^p(\R^n), L^\infty(\R^n))_{\frac{s p}{n}, p}}^p & = \int_0^\infty ( u^{-\frac{s p}{n}} K(u, f; L^p(\R^n), L^\infty(\R^n)))^p \frac{du}{u} \\
&\approx \int_0^\infty u^{-\frac{s p}{n}} \int_0^{u} (f^*(v))^p \, dv \frac{du}{u} \\
& = \int_0^\infty (f^*(v))^p \int_v^\infty u^{-\frac{s p}{n}} \frac{du}{u} dv \\
& = \frac{n}{s p} \|f\|_{L^{p^*, p}(\R^n)}^p.
\end{align*}
Inserting this into \eqref{For1}, we arrive at the desired estimate \eqref{ForClaim}.
Combining now \eqref{ForClaim-1}--\eqref{ForClaim} we achieve \eqref{BBMFract}.
\\
\textsc{Case 2:} Assume that either\footnote{In particular, this argument with $t= \frac{n}{p} \in \N$ completes the proof of \cite[Theorem 8]{KaradzhovMilmanXiao}.}
$$t > \frac{n}{p} \quad \text{and} \quad s \in \Big(0, \frac{n}{p} \Big)$$
or
$$
t= \frac{n}{p} \quad \text{and} \quad s \to 0^{+}.
$$
We will make use of the well-known embedding (see, e.g., \cite[p. 164]{Stein})
$$
\dot{H}^{\frac{n}{p}, p} (\R^n) \hookrightarrow \text{BMO}(\R^n)
$$
where $\text{BMO}(\R^n)$ is the space of bounded mean oscillation functions of John--Nirenberg \cite{JN} endowed with the seminorm
$$
\|f\|_{\text{BMO}(\R^n)} := \|f^{\#}\|_{L^\infty(\R^n)}
$$
and $f^{\#}$ is the sharp maximal function of $f \in L^1_{\text{loc}}(\R^n)$, i.e.,
$$
f^{\#} (x) := \sup_{Q \ni x} \frac{1}{|Q|} \int_Q |f(x)-f_Q| \, dx, \qquad x \in \R^n,
$$
where the supremum runs over all cubes $Q$ in $\R^n$ with sides parallel to the axes of coordinates and $f_Q = \frac{1}{|Q|} \int_Q f$.
In this case, the analog of \eqref{ForClaim-1} reads as
\begin{equation}\label{ForClaim-1new}
\|f\|_{ (L^p(\R^n), \text{BMO}(\R^n))_{\frac{s p}{n}, p}} \lesssim \|f\|_{(L^p(\R^n), \dot{H}^{\frac{n}{p}, p}(\R^n))_{\frac{s p}{n}, p}}
\end{equation}
for all $s \in (0, \frac{n}{p})$.
We make the following claim, uniformly with respect to $s \in (0, \frac{n}{p})$,
\begin{equation}\label{ClaimBMO}
(L^p(\R^n), \text{BMO}(\R^n))_{\frac{s p}{n}, p} \hookrightarrow s^{-\frac{1}{p}} (n- s p) L^{p^{*}, p}(\R^n).
\end{equation}
Assuming momentarily the validity of this claim, then \eqref{BBMFract} follows easily from \eqref{ForClaim-1new} and \eqref{ForClaim0} (with $t = \frac{n}{p}$).
It remains to show \eqref{ClaimBMO}. To proceed with, we will make use of the known estimate (cf. \cite[Corollary 3.3]{JawerthTorchinsky})
$$
K(u, f; L^p(\R^n), \text{BMO}(\R^n)) \approx \bigg(\int_0^{u^p} (f^{\#*}(v))^p \, dv \bigg)^{\frac{1}{p}}.
$$
Accordingly, by Fubini's theorem,
\begin{align}
\|f\|_{(L^p(\R^n), \text{BMO}(\R^n))_{\frac{s p}{n}, p}}^p &= \int_0^\infty (u^{-\frac{s p}{n}} K(u, f; L^p(\R^n), \text{BMO}(\R^n)) )^p \frac{du}{u} \nonumber \\
& \approx \int_0^\infty u^{-\frac{s p}{n}} \int_0^{u} (f^{\#*}(v))^p \, dv \frac{du}{u} \label{10} \\
& = \frac{n}{s p} \int_0^\infty (v^{\frac{1}{p^{*}}}f^{\#*}(v))^p \frac{dv}{v}. \nonumber
\end{align}
We can compare $f^{\#}$ and $f^{**}$ via the following weak-type estimate (cf. \cite[Proposition 8.10, p. 398]{BennettSharpley})
$$
(f-f_\infty)^{**}(u) \lesssim \int_u^\infty f^{\#*}(v) \frac{dv}{v}
$$
where $f_\infty := \lim_{|Q| \to \infty} \frac{1}{|Q|} \int_Q f$. Note that $f_\infty = 0$ since $f \in L^p(\R^n)$. Accordingly, by Hardy's inequality (see, e.g., \cite[p. 196]{SteinWeiss}),
\begin{align*}
\|f\|_{L^{p^{*}, p}(\R^n)} &\leq \bigg(\int_0^\infty [u^{\frac{1}{p^*}} f^{**}(u)]^p \frac{du}{u} \bigg)^{\frac{1}{p}} \\
& \lesssim \bigg(\int_0^\infty \bigg(u^{\frac{1}{p^*}} \int_u^\infty f^{\#*}(v) \frac{d v}{v} \bigg)^p \frac{du}{u} \bigg)^{\frac{1}{p}} \\
& \lesssim p^* \bigg(\int_0^\infty u^{\frac{p}{p^*}} (f^{\#*}(u))^p \frac{du}{u} \bigg)^{\frac{1}{p}} \\
& \approx s^{\frac{1}{p}} p^{*} \|f\|_{(L^p(\R^n), \text{BMO}(\R^n))_{\frac{s p}{n}, p}}
\end{align*}
where we have used \eqref{10} in the last step.
The proof of \eqref{ClaimBMO} is complete.
\\
\textsc{Case 3:} Assume $t = \frac{n}{p}$ and $s \to t^{-}$. Let $t_0 \in (0, t)$ be fixed and choose $\theta \in (0, 1)$ such that $s = (1-\theta) t_0 + \theta t$. Then there exists a constant $C >0$, which is independent of $s$, such that
\begin{equation}\label{3.1}
\|f\|_{(\dot{H}^{t_0, p}(\R^n), \dot{H}^{t, p}(\R^n))_{\theta, p}} \leq C \, \|f\|_{\dot{W}^{s, p}(\R^n), t}
\end{equation}
Indeed, since (see, e.g., \cite[Section 6.4, pp. 149--153]{BerghLofstrom})
$$\dot{B}^{t_0}_{p, \min\{p, 2\}} (\R^n) = (L^p(\R^n), \dot{H}^{t, p}(\R^n))_{\frac{t_0}{t}, \min\{p, 2\}}$$
and
$$\dot{B}^{t_0}_{p, \min\{p, 2\}} (\R^n) \hookrightarrow \dot{H}^{t_0, p} (\R^n),$$
we have
\begin{equation}\label{3.2}
( (L^p(\R^n), \dot{H}^{t, p(}\R^n))_{\frac{t_0}{t}, \min\{p, 2\}}, \dot{H}^{t, p}(\R^n))_{\theta, p} \hookrightarrow (\dot{H}^{t_0, p}(\R^n), \dot{H}^{t, p}(\R^n))_{\theta, p}
\end{equation}
with related embedding constant independent of $s$. Furthermore, in virtue of the sharp version of the reiteration formula given in Lemma \ref{LemmaKMX2}(ii) and \eqref{OrderedCouplesCommute} (and taking into account that $\theta \to 1^{-}$), the following holds
$$
(L^p(\R^{n}), \dot{H}^{t, p}(\R^n))_{1- (1-\theta)(1-\frac{t_0}{t}), p} \hookrightarrow ( (L^p(\R^n), \dot{H}^{t, p}(\R^n))_{\frac{t_0}{t}, \min\{p, 2\}}, \dot{H}^{t, p}(\R^n))_{\theta, p}
$$
uniformly with respect to $\theta$, or equivalently, by Lemma \ref{LemmaIntFractGS},
\begin{equation}\label{3.3}
\|f\|_{( (L^p(\R^n), \dot{H}^{t, p}(\R^n))_{\frac{t_0}{t}, \min\{p, 2\}}, \dot{H}^{t, p}(\R^n))_{\theta, p}} \lesssim \|f\|_{\dot{W}^{s, p}(\R^n), t}.
\end{equation}
Putting together \eqref{3.2} and \eqref{3.3}, the estimate \eqref{3.1} is achieved.
By the well-known fact (see, e.g., \cite[Theorem 1(i), Section 5.2.3, p. 242]{Triebel83}) that $(-\Delta)^{\frac{t_0}{2}}$ acts as an isomorphism from $\dot{H}^{t_0, p}(\R^n)$ onto $L^p(\R^n)$ and from $\dot{H}^{t, p}(\R^n)$ onto $\dot{H}^{t-t_0, p}(\R^n)$, we easily derive that
$$
K(u, f; \dot{H}^{t_0, p}(\R^n), \dot{H}^{t, p}(\R^n)) \approx K(u, (-\Delta)^{\frac{t_0}{2}} f; L^p(\R^n), \dot{H}^{t-t_0, p}(\R^n))
$$
and thus
$$
\|f\|_{(\dot{H}^{t_0, p}(\R^n), \dot{H}^{t, p}(\R^n))_{\theta, p}} \approx \|(-\Delta)^{\frac{t_0}{2}} f\|_{(L^p(\R^n), \dot{H}^{t-t_0}_p(\R^n))_{\theta, p}}
$$
where the hidden constants are independent of $\theta$. According to Lemma \ref{LemmaIntFractGS}, the last estimate can be equivalently rewritten as
\begin{equation}\label{3.4}
\|f\|_{(\dot{H}^{t_0, p}(\R^n), \dot{H}^{t, p}(\R^n))_{\theta, p}} \approx \|(-\Delta)^{\frac{t_0}{2}} f\|_{\dot{W}^{\theta (t-t_0), p}(\R^n), t-t_0}.
\end{equation}
Since $0 < \theta (t-t_0) < t-t_0 < t = \frac{n}{p}$, we can apply Case 1 so that (take into account that $f \in W^{s, p}(\R^n)$ and, in particular, $f \in H^{t_0, p}(\R^n)$ since $t_0 < s$)
\begin{equation}\label{3.5}
\|(-\Delta)^{\frac{t_0}{2}} f \|_{L^{\frac{n p}{n-(s-t_0)p} , p}(\R^n)} \lesssim (1-\theta)^{\frac{1}{p}} \|(-\Delta)^{\frac{t_0}{2}} f\|_{\dot{W}^{\theta (t-t_0), p}(\R^n), t-t_0}.
\end{equation}
As usual, we denote by $I_\alpha, \, \alpha \in (0, n),$ the Riesz potential operator
$$
(I_\alpha f)(x) := c_{n, \alpha} \int_{\R^n} \frac{f(y)}{|x-y|^{n-\alpha}} \, dy, \qquad x \in \R^n.
$$
By well-known mapping properties of $I_\alpha$ (cf. \cite[Chapter V, Section 1.2]{Stein})
\begin{equation}\label{3.6}
I_{t_0}: L^{\frac{n p}{n-(s-t_0)p}, p}(\R^n) \to L^{p^{*}, p}(\R^n).
\end{equation}
However, our method requires to control the norm of this operator with respect to $s$. Accordingly, to make the exposition self-contained, we provide below a detailed proof of \eqref{3.6} using standard techniques. It follows from O'Neil's inequality \cite[Lemma 1.5]{Oneil} that
$$
(I_{t_0} f)^*(u) \lesssim \int_u^\infty f^{**}(v) v^{\frac{t_0}{n}-1} \, dv
$$
and thus, by Hardy's inequalities (cf. \cite[p. 196]{SteinWeiss}),
\begin{align*}
\|I_{t_0} f\|_{L^{p^{*}, p}(\R^n)} & \lesssim \bigg(\int_0^\infty u^{\frac{p}{p^*}} \bigg(\int_u^\infty f^{**}(v) v^{\frac{t_0}{n}-1} \, dv \bigg)^{p} \frac{du}{u} \bigg)^{\frac{1}{p}} \\
& \lesssim p^{*} \bigg(\int_0^\infty \big(u^{\frac{n-(s-t_0) p}{n p}} f^{**}(u) \big)^p \frac{du}{u} \bigg)^{\frac{1}{p}} \\
& \lesssim \frac{p^{*}}{1-\frac{1}{p} + \frac{s-t_0}{n}} \bigg(\int_0^\infty (u^{\frac{n-(s-t_0) p}{n p}} f^*(u))^p \frac{du}{u} \bigg)^{\frac{1}{p}} \\
& \approx \frac{1}{n- s p} \|f\|_{L^{\frac{n p}{n-(s-t_0) p}, p}(\R^n)}
\end{align*}
where we have used $s \to t^{-} = \big(\frac{n}{p} \big)^{-}$ in the last step. This proves (cf. \eqref{3.6})
\begin{equation}\label{3.7}
\|I_{t_0}\|_{L^{\frac{n p}{n-(s-t_0) p}, p}(\R^n) \to L^{p^{*}, p}(\R^n)} \lesssim (n- s p)^{-1}
\end{equation}
uniformly with respect to $s \to t^{-}$.
Since $I_{t_0} (-\Delta)^{\frac{t_0}{2}} f = f$, in light of \eqref{3.7}, \eqref{3.5}, \eqref{3.4} and \eqref{3.1} we get
\begin{align*}
\|f\|_{L^{p^{*}, p}(\R^n)} &\lesssim (n- s p)^{-1} \| (-\Delta)^{\frac{t_0}{2}} f \|_{L^{\frac{n p}{n-(s-t_0) p}, p}(\R^n)} \\
& \lesssim (n- s p)^{\frac{1}{p}-1} \|(-\Delta)^{\frac{t_0}{2}} f\|_{\dot{W}^{\theta (t-t_0), p}(\R^n), t-t_0} \\
& \approx (n- s p)^{\frac{1}{p}-1} \|f\|_{(\dot{H}^{t_0, p}(\R^n), \dot{H}^{t, p}(\R^n))_{\theta, p}} \\
& \lesssim (n- s p)^{\frac{1}{p}-1} \|f\|_{\dot{W}^{s, p}(\R^n), t}.
\end{align*}
The proof is finished.
\end{proof}
As an immediate consequence of Theorem \ref{ThmBBMt} and the well-known inequality between Lorentz spaces (cf. \cite[p. 192]{SteinWeiss})
$$
\Big(\frac{q}{p} \Big)^{\frac{1}{q}} \|f\|_{L^{p, q}(\R^n)} \leq \Big(\frac{r}{p} \Big)^{\frac{1}{r}} \|f\|_{L^{p, r}(\R^n)}, \qquad r < q,
$$
we obtain the following
\begin{corollary}
Let $0 < s < t \leq 1, 1 < p < \frac{n}{s}$ and $p^{*} = \frac{n p}{n-s p}$. Assume $f \in W^{s, p}(\R^n)$. Then there exists $C = C(n, p, t) > 0$ such that
\begin{equation*}
\|f\|^p_{L^{p^{*}}(\R^n)} \leq C \, s (t-s) (n- s p)^{1-p} \|f\|_{\dot{W}^{s, p}(\R^n), t}^p.
\end{equation*}
\end{corollary}
\section*{Appendix B: Sharp relationships between Sobolev--Triebel--Lizorkin
spaces\label{SectionAppendix}}
The methodology proposed in \cite{Braz} relies on sharp comparison estimates
between $\dot{W}^{s,p}({\mathbb{R}}^{n}),\dot{F}_{p,2}^{s}({\mathbb{R}}^{n})$
and $\dot{F}_{p,p}^{s}({\mathbb{R}}^{n})$ using tools from harmonic analysis.
We now provide an approach to these results relying on the interpolation
techniques presented in Section \ref{SectionMethod}.
Let $s\in(0,1)$ and $p\in(1,\infty)$. Recall that (cf. Lemma
\ref{TableCoincidences})
\begin{equation}
\dot{W}^{s,p}({\mathbb{R}}^{n})=\dot{F}_{p,p}^{s}({\mathbb{R}}^{n}%
),\label{GTL}%
\end{equation}%
\[
\dot{F}_{p,2}^{s}({\mathbb{R}}^{n})\hookrightarrow\dot{W}^{s,p}({\mathbb{R}%
}^{n}),\qquad p\geq2,
\]
and
\[
\dot{W}^{s,p}({\mathbb{R}}^{n})\hookrightarrow\dot{F}_{p,2}^{s}({\mathbb{R}%
}^{n}),\qquad p\leq2.
\]
The sharp behavior of the related equivalence constants in these embeddings
plays a key role in the Bourgain--Brezis--Mironescu method. In particular, note
that the family of semi-norms $\{\Vert\cdot\Vert_{\dot{F}_{p,p}^{s}%
({\mathbb{R}}^{n})}:s\in(0,1)\}$ converges to $\Vert\cdot\Vert_{\dot{F}%
_{p,p}^{1}({\mathbb{R}}^{n})}$ as $s\rightarrow1^{-}$, however this is not the
case for $\{\Vert\cdot\Vert_{\dot{W}^{s,p}({\mathbb{R}}^{n})}:s\in(0,1)\}$
(cf. \eqref{Constancy}). In particular, the hidden equivalence constants in
\eqref{GTL} must exhibit a certain blow up as $s$ approaches the limiting
values $0$ and $1$. The precise answer is contained in the following
\begin{theorem}
\label{TheoremBSY1} Let $0 < s < t < \infty$ and $p \in(1, \infty)$. Then
there exists a positive constant $C$, which is independent of $s$, such that
\begin{enumerate}
\item[(a)]
\begin{align}
C^{-1}\bigg(\frac{1}{s^{\frac{1}{\max\{p,2\}}}}+\frac{1}{(t-s)^{\frac{1}%
{\max\{p,2\}}}}\bigg)\Vert f\Vert_{\dot{F}_{p,p}^{s}({\mathbb{R}}^{n})} &
\leq\Vert f\Vert_{\dot{W}^{s,p}({\mathbb{R}}^{n}),t}\nonumber\\
& \hspace{-4cm}\leq C\bigg(\frac{1}{s^{\frac{1}{\min\{p,2\}}}}+\frac
{1}{(t-s)^{\frac{1}{\min\{p,2\}}}}\bigg)\Vert f\Vert_{\dot{F}_{p,p}%
^{s}({\mathbb{R}}^{n})}, \label{BSYEstim}%
\end{align}
\item[(b)]
\[
C\bigg(\frac{1}{s^{\frac{1}{p}}}+\frac{1}{(t-s)^{\frac{1}{p}}}\bigg)\Vert f\Vert_{\dot
{F}_{p,2}^{s}({\mathbb{R}}^{n})}\leq\Vert f\Vert_{\dot{W}^{s,p}({\mathbb{R}%
}^{n}),t},\qquad p\leq2,
\]
\item[(c)]
\[
\Vert f\Vert_{\dot{W}^{s,p}({\mathbb{R}}^{n}),t}\leq C\bigg(\frac{1}{s^{\frac{1}{p}}%
}+\frac{1}{(t-s)^{\frac{1}{p}}}\bigg)\Vert f\Vert_{\dot{F}_{p,2}^{s}({\mathbb{R}}%
^{n})},\qquad p\geq2.
\]
\end{enumerate}
\end{theorem}
\begin{remark}
The previous result can be understood as a generalization to higher order
smoothness of \cite[Theorems 1.2 and 1.5]{Braz} (which corresponds to the
classical case $t=1$, cf. \eqref{FractModuInt}).
\end{remark}
\begin{proof}
[Proof of Theorem \ref{TheoremBSY1}](a): By Lemma \ref{LemmaIntFractGS}, for
every $s\in(0,t)$,
\begin{equation}
\Vert f\Vert_{\dot{W}^{s,p}({\mathbb{R}}^{n}),t}\approx\Vert f\Vert
_{(L^{p}({\mathbb{R}}^{n}),\dot{H}^{t,p}({\mathbb{R}}^{n}))_{\frac{s}{t},p}}.
\label{ThmProof1}%
\end{equation}
The interpolation space given in the right-hand side of the previous estimate
can be explicitly computed via the retraction method. Specifically,
$(L^{p}({\mathbb{R}}^{n}),\dot{H}^{t,p}({\mathbb{R}}^{n}))$ can be identified
with the vector-valued pair $(L^{p}({\mathbb{R}}^{n};\ell_{2}({{\mathbb{Z}}%
})),\ell^{p}({\mathbb{R}}^{n};\ell_{2}^{t}({{\mathbb{Z}}})))$ and thus, by
Lemma \ref{LemmaInterpolationLpVector},
\begin{equation}
\Vert(f_{j})\Vert_{(L^{p}({\mathbb{R}}^{n};\ell_{2}({{\mathbb{Z}}}%
)),L^{p}({\mathbb{R}}^{n};\ell_{2}^{t}({{\mathbb{Z}}})))_{\frac{s}{t},p}%
}\approx\Vert(f_{j})\Vert_{L^{p}({\mathbb{R}}^{n};(\ell_{2}({{\mathbb{Z}}%
}),\ell_{2}^{t}({{\mathbb{Z}}}))_{\frac{s}{t},p})}. \label{ThmProof1new}%
\end{equation}
Furthermore, according to Lemma \ref{Lemma4},
\begin{align}
\bigg(\frac{1}{s^{\frac{1}{\max\{p,2\}}}}+\frac{1}{(t-s)^{\frac{1}%
{\max\{p,2\}}}}\bigg)\Vert\xi\Vert_{\ell_{p}^{s}({{\mathbb{Z}}})} &
\lesssim\Vert\xi\Vert_{(\ell_{2}({{\mathbb{Z}}}),\ell_{2}^{t}({{\mathbb{Z}}%
}))_{\frac{s}{t},p}}\nonumber\\
& \hspace{-4cm}\lesssim\bigg(\frac{1}{s^{\frac{1}{\min\{p,2\}}}}+\frac
{1}{(t-s)^{\frac{1}{\min\{p,2\}}}}\bigg)\Vert\xi\Vert_{\ell_{p}^{s}%
({{\mathbb{Z}}})}. \label{ThmProof1new2}%
\end{align}
As a combination of \eqref{ThmProof1new} and \eqref{ThmProof1new2},
\begin{align}
\bigg(\frac{1}{s^{\frac{1}{\max\{p,2\}}}}+\frac{1}{(t-s)^{\frac{1}%
{\max\{p,2\}}}}\bigg)\Vert(f_{j})\Vert_{L^{p}({\mathbb{R}}^{n};\ell_{p}%
^{s}({{\mathbb{Z}}}))} & \lesssim\Vert(f_{j})\Vert_{(L^{p}({\mathbb{R}}%
^{n};\ell_{2}({{\mathbb{Z}}})),L^{p}({\mathbb{R}}^{n};\ell_{2}^{t}%
({{\mathbb{Z}}})))_{\frac{s}{t},p}}\nonumber\\
& \hspace{-5cm}\lesssim\bigg(\frac{1}{s^{\frac{1}{\min\{p,2\}}}}+\frac
{1}{(t-s)^{\frac{1}{\min\{p,2\}}}}\bigg)\Vert(f_{j})\Vert_{L^{p}({\mathbb{R}%
}^{n};\ell_{p}^{s}({{\mathbb{Z}}}))}. \label{ThmProof1new3}%
\end{align}
By the well-known fact that $\dot{F}_{p,p}^{s}({\mathbb{R}}^{n})$ is a retract
of $L^{p}({\mathbb{R}}^{n};\ell_{p}^{s}({{\mathbb{Z}}}))$, we can put together
\eqref{ThmProof1new3} and \eqref{ThmProof1} to arrive at \eqref{BSYEstim}.
(b): Let $p\leq2$. Applying Lemmas \ref{Lemma4} and \ref{LemmaExtrapolSharp}%
(i),
\begin{equation}
\Vert\xi\Vert_{\ell_{2}^{s}({{\mathbb{Z}}})}\approx(s(t-s))^{\frac{1}{2}}%
\Vert\xi\Vert_{(\ell_{2}({{\mathbb{Z}}}),\ell_{2}^{t}({{\mathbb{Z}}}%
))_{\frac{s}{t},2}}\lesssim(s(t-s))^{\frac{1}{p}}\Vert\xi\Vert_{(\ell
_{2}({{\mathbb{Z}}}),\ell_{2}^{t}({{\mathbb{Z}}}))_{\frac{s}{t},p}}.
\label{ThmProof1new4}%
\end{equation}
Taking into account that $\dot{F}_{p,2}^{s}({\mathbb{R}}^{n})$ is a retract of
$L^{p}({\mathbb{R}}^{n};\ell_{2}^{s}({{\mathbb{Z}}}))$, it follows now from
\eqref{ThmProof1}, \eqref{ThmProof1new} and \eqref{ThmProof1new4} that
\[
(s(t-s))^{-\frac{1}{p}}\Vert f\Vert_{\dot{F}_{p,2}^{s}({\mathbb{R}}^{n}%
)}\lesssim\Vert f\Vert_{\dot{W}^{s,p}({\mathbb{R}}^{n}),t}.
\]
The proof of (c) follows similar ideas as those given in (b).
\end{proof}
|
\section{Introduction}
The problem we focus on this work is the Guillotine 2D Knapsack Problem with orthogonal (and unrestricted) cuts, constrained demand, unlimited stages, and no rotation.
We will refer to this specific variant as G2KP.
If we further qualify the G2KP, we only mean to discard the qualifiers above that directly conflict with the extra qualifiers, if any.
The G2KP is an NP-hard problem~\cite{russo:2020}.
The work also focuses on obtaining optimal solutions for this problem through Mixed-Integer Linear Programming (MILP).
We propose two simple but effective enhancements regarding a state-of-the-art MILP formulation for the G2KP (which may also benefit some closely related problem variants).
\subsection{Explanation of the problem and some close variants}
An instance of the G2KP consists of: a rectangle of length~\(L\) and width~\(W\) (hereafter called \emph{original plate}); a set of rectangles~\(\bar{J}\) (also referred to as \emph{pieces}) where each rectangle~\(j \in \bar{J}\) has a length~\(l_j\), a width~\(w_j\), a profit~\(p_j\), and a demand~\(u_j\)
We assume, without loss of generality, that all such values are positive integers.
The G2KP seeks to maximise the profit of the pieces obtained by cutting the original plate.
The \emph{guillotine} qualifier means every cut always go from one side of a plate to other; a cut never stops or starts from the middle of a plate.
A consequence of this rule is that we often do not obtain the pieces directly from the original plate.
We cut the original plate into intermediary plates \(j \in J\), \(J \supseteq \bar{J}\), which we further cut following the same rule.
If we do not cut a plate further, then it is either: thrown away as trim/waste for no profit; or, if it has the same size as a piece, sold by the piece profit value.
\emph{Orthogonal cuts} are always parallel to one side of a plate (and perpendicular to the other).
Consequently, any intermediary plate~\(j\) is always a rectangle, and have a well-defined~\(l_j\) and~\(w_j\).
\emph{Unrestricted cuts} mean we are allowed to make horizontal (vertical) cuts different from the width (length) of a piece.
We will mention the G2KP with restricted cuts further in the text, as solving it exactly is a costly but high-quality primal heuristic for the G2KP.
\emph{Constrained demand} means we can sell at most~\(u_j\) copies of piece~\(j\).
The G2KP with \emph{unconstrained demand} is not NP-hard; exact algorithms of pseudo-polynomial time complexity exist~\cite{beasley:1985}.
Consequently, interesting G2KP instances have~\(u_j < \lceil L / l_j \rceil \times \lceil W / w_j \rceil \) for at least one piece~\(j\) (if not for all pieces).
\emph{Unlimited stages} means there is no limit to the number of times the guillotine switches between horizontal and vertical orientations.
In the exact \(k\)-staged G2KP, the guillotine is switched at most \(k-1\) times.
Consequently, a 2-staged G2KP has all cuts in some orientation before any cuts in the other orientation.
The non-exact \(k\)-staged G2KP adds one extra stage in which the only cuts allowed are the ones that trim plates to the size of pieces.
The \emph{no-rotation} qualifier means we never switch length and width during the cutting process; especially, we cannot sell a plate~\(j\) as a piece of length~\(w_j\) and width~\(l_j\).
The literature further distinguishes between \emph{weigthed} and \emph{unweighted} problem variants.
In the weighted variant, pieces have an arbitrary profit value, while in the unweighted variant the profit value is always equivalent to the piece area.
Consequently, the unweighted variant is equivalent to minimising waste and is a particular case of the weighted variant.
Any algorithm that solves the weighted variant (as is our case) can solve the unweighted variant by setting the piece profit values to their areas.
While our work focuses on this specific problem, the enhanced formulation we present may be readily adapted to, at least, the Guillotine 2D version of the following problems: the Cutting Stock Problem (and the Bin Packing Problem); the Strip Packing Problem; the Multiple Knapsack Problem; the Orthogonal Packing Problem; and the variant allowing rotation for all previously mentioned problems.
See~\cite{furini:2016} for more details.
We do not define or further discuss these problems or variants in this work.
\subsection{Motivation}
The G2KP and its closely related variants are of undisputable interest of the industry, especially wood, paper, metal, and glass cutting industries.
The vast and growing literature on the subject examined by~\cite{iori:2020} and by~\cite{russo:2020} is enough proof of such interest.
To pick a single recent case study see~\cite{clautiaux:2019}, which solves a unique variant of the Guillotine 2D Cutting Stock Problem for a glass factory manufacturing double-paned windows.
We focus on MILP as the solving method (instead of \emph{ad hoc} solutions) because its adaptability amplifies the value of any enhancements we obtain.
A better MILP formulation means:
a better solving procedure for the many (already mentioned) closely related problem variants;
a better continuous relaxation for computing an optimistic guess on the objective value of all these variants (some \emph{ad hoc} algorithms of the literature use MILP solvers to compute their bounds);
not only a better exact method but also a better base for heuristics or anytime procedures;
an immediate benefit from parallelisation, automatic problem decomposition, and solver-implemented heuristics;
and, finally, better ageing of the method over the years through the current trends of multiple-cores processors and ever-advancing solver performance.
\subsection{Contributions and paper outline}
The main contributions of this work are:
an enhanced MILP formulation based on a previous state-of-the-art formulation, its proof of correctness, and empirical evidence of its better performance;
a straigthforward adaptation of a previously known reduction procedure for both the original and the enhanced formulations, and empirical evidence of its positive impact on their performance;
finally, we present new upper and lower bounds, as well as optimal values, for many recently proposed hard instances from~\cite{velasco:2019}.
For such, we reimplement a state-of-the-art MILP formulation and an optional pricing procedure used by it.
This reimplementation allows us to compare both approaches fully.
All code used is available in the first author's repository ({\small\url{https://github.com/henriquebecker91/GuillotineModels.jl/tree/0.2.4}}).
We organise the rest of the paper the following way:
\autoref{sec:related_work} analyses how our work interacts with the pre-existing literature;
\autoref{sec:psn} introduces some mathematical concepts and explains the reduction we adapted from the literature;
\autoref{sec:enhanced_model} describes our enhanced formulation and briefly explains how it differs from the state-of-the-art formulation it is based on;
\autoref{sec:experimental_results} presents our experiments and the empirical results we derive from them;
\autoref{sec:conclusions} delivers our conclusions and suggests future work.
\section{Related work}
\label{sec:related_work}
We do not intend to provide a full overview of the literature, instead we:
refer to surveys; discuss only closely related works and how they interact with our contributions; and opportunely point out missing connections between related works.
Two relevant surveys have come out recently.
\cite{iori:2020} catalogues exact methods and relaxations for 2D cutting problems including guillotine problems.
\cite{russo:2020} reviews the literature of our particular problem at length -- there G2KP is referred to as Constrained 2D Cutting or C2DC.
Moreover, \cite{russo:2020}~points out three strategies employed by previous exact solving methods which cause loss of optimality.
Our work does not employ any of these three strategies.
One of these strategies is a dominance rule that is valid for the unconstrained case but not for the constrained case.
In 1972, \cite{herz:1972}~proposed a dominance rule for the G2KP with unconstrained demand based on the same principle and warned about the possibility of misusing the rule in the constrained case.
The first MILP formulation dealing with guillotine cuts and unlimited stages was proposed by~\cite{messaoud:2008} in 2008.
The problem considered by~\cite{messaoud:2008} is the Strip Packing Problem, but adapting the formulation to the knapsack variant would not change its fundamentals.
Previously, \cite{lodi:2003}~had proposed two MILP formulations for 2-staged G2KP.
As noted by~\cite{belov_thesis:2003}, modeling \(k\)-staged cuts for \(k \geq 3\) (unlimited stages included) was considered difficult at the time.
The size of most \(k\)-staged formulations is exponential on the number of stages (i.e.,~\(k\)).
The formulation of~\cite{messaoud:2008} had about \(3n^4/4\) variables and \(2n^4\) constraints (where \(n\) is the number of pieces) it also employed, according to the authors, a ``very loose linear relaxation'' due to which ``the practical interest of this formulation is still limited''.
The characterization of guillotine cuts proposed by~\cite{messaoud:2008} seems to have been simultaneously proposed by~\cite{pisinger:2007}.
The first MILP formulation specifically for the G2KP was proposed by~\cite{furini:2016} in 2016.
An extended version of~\cite{furini:2016} appears in~\cite{dimitri_thesis} (a PhD thesis).
Their formulation has pseudo-polynomial size, \(O((L + W) \times L \times W)\)~variables and \(O(L \times W)\) constraints, and its relaxation provides a stronger bound than~\cite{messaoud:2008}.
It was the first formulation able to solve medium-sized instances of the literature.
Besides the formulation, \cite{furini:2016}~proposes two reductions and one pricing procedure; all of these are reimplemented by our work.
They also present and prove a theorem to assure the correctness of one of their reductions~(\emph{Cut-Position}).
A similar theorem and proof appear in~\cite{song:2010}.
In this work, we propose an enhanced formulation based on the one from \cite{furini:2016} mentioned above.
A significant advantage of our enhancement is to avoid the enumeration of any cuts after the middle of a plate.
This advantage appears in many works since~\cite{herz:1972}.
Recently, \cite{delorme:2019} adapted a formulation for the one-dimensional Cutting Stock Problem to obtain this same advantage.
However, the way \cite{delorme:2019}~changes their formulation to obtain this advantage is not the same as our approach.
The most recent MILP formulations for the G2KP come from three works by Martin et alii~\cite{martin:2020:models,martin:2020:bottom,martin:2020:top}.
These formulations are compared against the formulation of~\cite{furini:2016}.
We base our enhanced formulation on~\cite{furini:2016} and also compare against it.
The formulations of ~\cite{martin:2020:models,martin:2020:bottom,martin:2020:top}
have a looser relaxation bound compared to~\cite{furini:2016}, but perform better than~\cite{furini:2016} in instances for which~\cite{furini:2016} has a much larger number of variables.
Considering the instances used in~\cite{furini:2016}, our enhanced formulation dominates the formulation of~\cite{furini:2016}.
Our formulation also dramatically improves the running times of instances in which the formulation of~\cite{furini:2016} performed worse than \cite{martin:2020:models,martin:2020:bottom,martin:2020:top} (e.g., the gcut1--gcut12 instances).
Consequently, while it may be interesting for completeness sake, we do not compare against the formulations proposed in~\cite{martin:2020:models,martin:2020:bottom,martin:2020:top}.
\section{Notation, Discretization, and Plate-Size Normalization}
\label{sec:psn}
The performance of solving methods for cutting and packing problems often heavily depends on the number of (cut/packing) positions considered.
Since the seminal works of~\cite{cw:1977} and~\cite{herz:1972}, solving methods avoid considering each possible position, but instead consider only a subset necessary to guarantee optimality.
The literature includes many such subsets, which are often referred to as \emph{discretizations}.
The most common way of computing these discretizations are Dynamic Programming (DP) algorithms.
These DP algorithms usually only take a small fraction of the running time, but the size of the position subset outputted by them strongly affects the time spent by the rest of the solving method.
Both~\cite{furini:2016} and our enhanced formulation have one constraint for each attainable distinctly-sized plate and one variable for each potential cut over each of these plates.
Therefore, eliminating a single cutting position has the following effects:
\textbf{(i)} it removes one variable for each distinctly-sized plate that allowed that cutting position;
\textbf{(ii)} if that cutting position was the only way to produce some distinctly-sized plates\footnote{Note that the same cutting position, when applied to distinctly-sized plates, may generate different children.}, then it also removes the constraints associated with these plates;
\textbf{(iii)} if (ii) excludes one or more constraints/plates, then it also excludes all variables representing possible cuts over the excluded plates;
\textbf{(iv)} finally, if (iii) eliminates one or more variables/cuts, then it may trigger (ii) again (i.e., other plates stop being attainable), cyclically.
In this work, the only cut subset (discretization) considered are the canonical dissections of~\cite{herz:1972}, hereafter referred to as \emph{normal cuts} instead.
We acknowledge the existence of stricter discretizations: the raster points of~\cite{terno:1987,guntram:1966}, the regular normal patterns of~\cite{boschetti:2002} (named this way by~\cite{cote:2018}), and the Meet-in-the-Middle (MiM) of~\cite{cote:2018}.
The reasons for our choice of discretization are numerous:
it works well with the \emph{Plate-Size Normalization} procedure we describe below;
it is the same discretization employed by~\cite{furini:2016} (from which we base our enhanced formulation on);
MiM main gain is reducing the number of cut positions after the middle of a plate, which our enhanced formulation already discards anyway;
the regular normal patterns compute a distinct subset-sum for each pair of plate and piece, which we consider excessive (there may exist hundreds of thousands of intermediary plate possibilities);
finally, the raster points complicate our proofs and our \emph{Plate-Size Normalization} weakens its benefits.
The set~\(O = \{v, h\}\) denotes the cut orientation: \(v\) is vertical (parallel to width, perpendicular to length); \(h\) is horizontal (parallel to length, perpedicular to width).
Let us recall that the demand of a piece~\(i \in \bar{J}\) is denoted by~\(u_i\).
If we define the set of pieces fitting a plate~\(j\) as~\(I_j = \{i \in \bar{J} : l_i \leq l_j \land w_i \leq w_j \}\), we can define~\(N_{jo}\) (i.e., the set of the normal cuts of orientation~\(o\) over plate~\(j\)) as:
\begin{equation}
N_{jo}= \left\{
\begin{array}{lllr}
\{q: 0 < q < l_j; & \forall i \in I_j, \exists n_i \in [0 \isep u_i], q = \sum_{i\in I_j} n_i l_i \} & \quad \text{if } o = v,\\
\{q: 0 < q < w_j; & \forall i \in I_j, \exists n_i \in [0 \isep u_i], q = \sum_{i\in I_j} n_i w_i \} & \quad \text{if } o = h.
\end{array}\right.
\end{equation}
The sets defined above never include cuts at the plate extremities (i.e., \(0\), \(l_j\) for \(N_{jv}\), and \(w_j\) for \(N_{jh}\)).
Any of these cuts will always create (i)~a~zero-area plate and (ii)~a~copy of the plate that is being cut.
Consequently, these cuts only add symmetries and may be disregarded.
The goal of the \emph{Plate-Size Normalization} procedure we propose is to reduce the number of distinctly-sized plates considered.
Fewer distinctly-sized plates mean fewer constraints and trigger the same cascading effect described by items (ii)--(iv) above.
The property exploited by the procedure is already known and similarly exploited by~\cite{alvarez:2009} and by~\cite{dolatabadi:2012}.
We state the property as:
\begin{proposition}
\label{pro:normalization}
Given a plate~\(j\), \(l_j\) may always be replaced by \(l^\prime_j = max\{q : q \in N_{kv}, q \leq l_j\}\) in which \(w_k = w_j\) but \(l_k > l_j\), without loss of optimality.
The analogue is valid for the width.
\end{proposition}
We do not replicate any proof here. We can then define:
\begin{definition}
The length of a plate~\(j\) is considered normalized if, and only if, \(l_j = l^\prime_j\).
The analogue is valid for the width.
The size of a plate is normalized if, and only if, both its length and its width are normalized.
\end{definition}
The \emph{Plate-Size Normalization} procedure we propose consists only of replacing every non-size-normalized plate enumerated by their normalized counterpart.
The number of distinctly-sized plates diminishes because the procedure replaces many plates of distinct but similar dimensions by a single plate.
The only extra effort added by \emph{Plate-Size Normalization} consists of binary searches over~\(N_{ko}\) sets for each plate~\(j\).
A suitable \(N_{ko}\) set for each plate~\(j\) was already computed by the plate enumeration procedure before introducing the \emph{Plate-Size Normalization} (no extra effort required).
\begin{remark}
If a normal cut~\(q\) divides the size-normalized plate~\(j\), the first child is always size normalized, but the second child may not be size normalized.
\end{remark}
\begin{example}
Given two pieces with \(l = [5, 7]\), \(u = [2, 3]\), and a size-normalized plate of length~\(21\), a normal cut at~\(12\) creates a non-normalized second child of length~\(9\).
\end{example}
\begin{comment}
\subsection{Expanded example for thesis proposal}
I need an original plate and about 3~5 pieces.
Ideally the original plate should already be size-normalized, to be fair.
The smallest piece needs to be distant of one in absolute terms, but cannot be on the relatively large side, because this makes harder for an intermediary plate have many replacements.
We may use only squares, but this is kinda boring.
Probably the easiest way is to make a branch of the code in which the code enumerating plates saves which non-normalized plates were replaced by each normalized plate and outputs them.
\end{comment}
\section{Our changes to Furini's model}
\label{sec:enhanced_model}
The formulation proposed in~\cite{furini:2016} is elegant: the pieces are just intermediary plates that may be sold.
Our contribution consists of changes to both the preprocessing step and to the formulation.
These changes significantly reduce the number of variables.
Differently, these changes deepen the distinction between plates and pieces and, consequently, may be regarded as sacrificing some elegance for performance.
The essentials of the formulation remain the same and, for this reason, we consider the model presented here as an enhanced model, not an entirely new model.
The cut enumeration in~\cite{furini:2016} excludes some symmetric cuts; that is, if two different cuts create the same set of two child plates, then the symmetric cut in the second half of the plate may be ignored.
Differently,~\cite{nicos:1977} disregards \emph{all} cuts after the middle of the plate because of symmetry.
If~\cite{furini:2016} would do the same as~\cite{nicos:1977} it could become impossible to trim a plate to the size of a piece.
For example, if there was a piece with length larger than half the length of a plate, and such plate has no normal cut with the exact length of the needed trim, then the piece could not be extracted from the plate, even if the piece fits the plate.
The goal of our changes is to reduce the number of cuts (i.e., model variables) by getting closer to the symmetry-breaking rule used in~\cite{nicos:1977} without loss of optimality.
\subsection{The enhanced formulation}
\label{sec:enhanced}
Our changes to the formulation are restricted to replacing the set of integer variables~\(y_j, i \in \bar{J},\) with a new set of variables~\(e_{ij}, (i, j) \in E, E \subseteq \bar{J} \times J\), and the necessary adaptations to accomodate this change.
In the original formulation, \(y_i\) denoted the number of times a plate~\(i\) was sold as the piece~\(i\), in this case, the plate always had the exact size of the piece.
Our \emph{extraction variables}~\(e_{ij}\) denote a piece~\(i\) was extracted from plate~\(j\), which size may differ from the size of the piece.
For convenience, we also define \(E_{i*} = \{ j : \exists~(i, j) \in E \}\) and \(E_{*j} = \{i : \exists~(i, j) \in E \}\).
The set \(O = \{h, v\}\) denotes the horizontal and vertical cut orientations.
The set \(Q_{jo}\) (\(\forall j \in J, o \in O\)) denotes the set of possible cuts (or cut positions) of orientation~\(o\) over plate~\(j\).
The parameter~\(a\) is a byproduct of the plate enumeration process.
If cutting a plate~\(k \in J\) with a cut of orientation~\(o \in O\) at position~\(q \in Q_{jo}\) adds a plate~\(j \in J\) to the stock, then~\(a^o_{qkj} = 1\); otherwise~\(a^o_{qkj} = 0\).
The description of this parameter in~\cite{furini:2016} has a typo, as pointed out by~\cite{martin:2020}:
``[...] there is a typo in their definition of parameter~\(a^o_{qkj}\), as the indices~\(j\) and~\(k\) seem to be exchanged.''.
In a valid solution, the value of \(x^o_{qj}\) is the number of times a plate~\(j \in J\) is cut with orientation~\(o \in O\) at position~\(q \in Q_{jo}\); while the value of~\(e_{ij}\) is the number of sold pieces of type~\(i \in \bar{J}\) that were extracted from plates of type~\(j \in J\).
The plate~\(0 \in J\) is the original plate, and it may also be in~\(\bar{J}\), as there may exist a piece of the same size as the original plate.
\begin{align}
\mbox{max.} &\sum_{(i, j) \in E} p_i e_{ij} \label{eq:objfun}\\
\mbox{s.t.} &\sum_{o \in O}\sum_{q \in Q_{jo}} x^o_{qj} + \sum_{i \in E_{*j}} e_{ij} \leq \sum_{k \in J}\sum_{o \in O}\sum_{q \in Q_{ko}} a^o_{qkj} x^o_{qk} \hspace*{0.05\textwidth} & \forall j \in J, j \neq 0,\label{eq:plates_conservation}\\
& \sum_{o \in O}\sum_{q \in Q_{0o}} x^o_{q0} + \sum_{i \in E_{*0}} e_{i0} \leq 1 &,\label{eq:just_one_original_plate}\\
& \sum_{j \in E_{i*}} e_{ij} \leq u_i & \forall i \in \bar{J},\label{eq:demand_limit}\\
& x^o_{qj} \in \mathbb{N}^0 & \forall j \in J, o \in O, q \in Q_{jo},\label{eq:trivial_x}\\
& e_{ij} \in \mathbb{N}^0 & \forall (i, j) \in E.\label{eq:trivial_e}
\end{align}
The objective function maximizes the profit of the extracted pieces~\eqref{eq:objfun}.
Constraint~\eqref{eq:plates_conservation} guarantees that for every plate~\(j\) that was further cut or had a piece extracted from it (left-hand side), there must be a cut making available a copy of such plate (right-hand side).
One copy of the original plate is available from the start~\eqref{eq:just_one_original_plate}.
The amount of extracted copies of some piece type must respect the demand for that piece type (a piece extracted is a piece sold)~\eqref{eq:demand_limit}.
Finally, the domain of all variables is the non-negative integers~\eqref{eq:trivial_x}-\eqref{eq:trivial_e}.
\subsection{The revised variable enumeration}
\label{sec:var_enum}
The variable enumeration described in~\cite{furini:2016} employs some rules to reduce the number of variables; they are symmetry-breaking, \emph{Cut-Position}, and \emph{Redundant-Cut}.
The two last rules are not discussed here; \cite{furini:2016}~proves their correctness and they do not conflict with the enhanced model.
The use of the \(x\)~variables does not change from the original formulation to our revised formulation -- however, the size of the enumerated set of variables changes.
Our revised enumeration does not create any variable~\(x^o_{jq}\) in which \((o = h \land q > \lceil w_j / 2 \rceil) \lor (o = v \land q > \lceil l_j / 2 \rceil)\).
The original formulation has variables~\(y_i\), \(i \in \bar{J}\), while the revised formulation replaces them with variables~\(e_{ij}\), \((i, j) \in E\), \(E \subseteq \bar{J} \times J\).
Set~\(\bar{J} \times J\) is orders of magnitude larger than~\(\bar{J}\).
Consequently, set~\(E\) must be a small subset to avoid having a revised model with more variables than the original.
A suitable subset may be obtained by a simple rule: \((i, j) \in E\) if, and only if, packing piece~\(i\) in plate~\(j\) does not allow any other piece to be packed in~\(j\).
\subsection{The proof of correctness}
The previous section presented a detailed explanation of the changes to the formulation and variable enumeration.
This section proves such changes do not affect the correctness of the model.
In~\cite{furini:2016}, only the perfect symmetries described below are removed.
Our changes may be summarized to:
\begin{enumerate}
\item There is no variable for any cut that occurs after the middle of a plate.
\item A piece may be obtained from a plate if, and only if, the piece fits the plate, and the plate cannot fit an extra piece (of any type).
\end{enumerate}
The second change alone cannot affect the model correctness.
The original formulation was even more restrictive in this aspect:
a piece could only be sold if a plate of the same dimensions existed.
In our revised formulation there will always exist an extraction variable in such case:
if a piece and plate match perfectly, there is no space for any other piece, fulfilling our only criteria for the existence of extraction variables.
Consequently, what needs to be proved is that:
\begin{theorem}
\label{the:enhanced_correctness}
Without changing the pieces obtained from a packing, we may replace any normal cut after the middle of a plate by a combination of piece extractions and cuts at the middle of a plate or before it.
\end{theorem}
\begin{proof}
This is a proof by exhaustion. The set of all normal cuts after the middle of a plate may be split into the following cases:
\begin{enumerate}
\item The cut has a perfect symmetry. \label{case:perfectly_symmetric}
\item The cut does not have a perfect symmetry.
\begin{enumerate}
\item Its second child can fit at least one piece. \label{case:usable_second_child}
\item Its second child cannot fit a single piece.
\begin{enumerate}
\item Its first child packs no pieces. \label{case:no_pieces}
\item Its first child packs a single piece. \label{case:one_piece}
\item Its first child packs two or more pieces. \label{case:many_pieces}
\end{enumerate}
\end{enumerate}
\end{enumerate}
We believe to be self-evident that the union of~\cref{case:perfectly_symmetric,case:usable_second_child,case:no_pieces,case:one_piece,case:many_pieces} is equal to the set of all normal cuts after the middle of a plate. We present an individual proof for each of these cases.
\begin{description}
\item[\Cref{case:perfectly_symmetric} -- \textbf{The cut has a perfect symmetry.}]
If two distinct cuts have the same children (with the only difference being the first child of one cut is the second child of the other cut, and vice-versa), then the cuts are perfectly symmetric.
Whether a plate is the first or second child of a cut does not make any difference for the formulation or for the problem.
If the cut is in the second half of the plate, then its symmetry is in the first half of the plate.
Consequently, both cuts are interchangeable, and we may keep only the cut in the first half of the plate.
\item[\Cref{case:usable_second_child} -- \textbf{Its second child can fit at least one piece.}]
\autoref{pro:normalization} allows us to replace the second child by a size-normalized plate that can pack any demand-abiding set of pieces the original second child could pack.
The second child of a cut that happens after the middle of the plate is smaller than half a plate, and its size-normalized counterpart may only be the same size or smaller.
So the size-normalized plate could be cut as the first child by a normal cut in the first half of the plate.
Moreover, the old first child (now second child) have stayed the same size or grown (because the size-normalization of its sibling), which guarantee this is possible.
\item[\Cref{case:no_pieces} -- \textbf{Its first child packs no piece.}]
If both children of a single cut do not pack any pieces, then the cut may be safely ignored.
\item[\Cref{case:one_piece} -- \textbf{Its first child packs a single piece.}]
First, let us ignore this cut for a moment and consider the plate being cut by it (i.e., the parent plate).
The parent plate either: can fit an extra piece together with the piece the first child would pack, or cannot fit any extra pieces.
If it cannot fit any extra pieces, this fulfills our criteria for having an extraction variable, and the piece may be obtained through it.
The cut in question can then be disregarded (i.e., replaced by the use of such extraction variable).
However, if it is possible to fit another piece, then there is a normal cut in the first half of the plate that would separate the two pieces, and such cut may be used to shorten the plate.
This kind of normal cuts may successively shorten the plate until it is impossible to pack another piece, and the single piece that was originally packed in the first child may then be obtained employing an extraction variable.
\item[\Cref{case:many_pieces} -- \textbf{Its first child packs two or more pieces.}]
If the first child packs two or more pieces, but the second child cannot fit a single piece (i.e., it is waste), then the cut separating the first and second child may be omitted and any cuts separating pieces inside the first child may still be done.
If some of the plates obtained by such cuts need the trimming that was provided by the omitted cut, then these plates will be packing a single piece each, and they are already considered in~\cref{case:one_piece}.
\end{description}
Given the cases cover every cut after the middle of a plate, and each case has a proof, then follows that \autoref{the:enhanced_correctness} is correct. \qed
\end{proof}
\section{Experimental results}
\label{sec:experimental_results}
There are three formulation implementations that provide data used in our comparisons:
\emph{original} refers to the implementation presented in~\cite{furini:2016,dimitri_thesis};
\emph{faithful} refers to our reimplementation of \emph{original};
\emph{enhanced} refers to our enhanced formulation presented in~\autoref{sec:enhanced_model}.
The \emph{original} implementation was not available\footnote{
We asked the authors of~\cite{furini:2016} for the \emph{original} implementation and Dimitri Thomopulos informed us it was not available.
}.
Consequently, all data relative to \emph{original} presented in this work comes from~\cite{dimitri_thesis}.
Both \emph{faithful} and \emph{enhanced} data were obtained by runs using the setup described in~\autoref{sec:setup}.
Each formulation may be modified by applying any combination of the following optional procedures:
\emph{priced} -- refer to the pricing procedure described in~\cite{furini:2016,dimitri_thesis};
\emph{normalized} -- the plate-size normalization procedure described in~\autoref{sec:psn};
\emph{warmed} -- the MIP models solved were warm-started with a solution found by a previous step;
\emph{Cut-Position} and \emph{Redundant-Cut} -- are reduction procedures described in~\cite{furini:2016,dimitri_thesis}, that may be enabled and disabled individually.
For each experiment described in the next sections, if we do not mention a procedure, then it is disabled.
The term \emph{restricted priced} refers to the model for the restricted version of the problem that is solved inside the pricing procedure mentioned above.
Consequently, for each run of a \emph{priced} variant, there will be a \emph{restricted priced} run with the same combination of optional procedures.
The differences between the \emph{restricted priced} and the (unrestricted) \emph{priced} models are mainly that:
(i) the \emph{restricted priced} model never has a horizontal (vertical) cut that does not match the width (length) of a piece;
(ii) the \emph{restricted priced} model is MIP-started with the solution of an heuristic (described in~\cite{furini:2016}) while the \emph{priced} model is MIP-started with the solution of the \emph{restricted priced} model;
(iii) the distinct solutions used to MIP-start the respective models are also used as the lower bound for the pricing procedure (details in~\cite{furini:2016}).
The goal of the pricing procedure is to remove unneeded variables from the model.
However, the priced model often ends up with unneeded constraints and variables due to pricing.
This effect is similar to the one described by items (ii)--(iv) in~\autoref{sec:psn}: if some variables (i.e., cuts) are removed, then some plates are never produced (i.e., some constraints just fix their variables to zero), consequently other variables/cuts become impossible, recursiverly.
The effort to remove such unnecessary variables and constraints is negligible.
The algorithm used is similar to finding the connected subgraph in the directed hypergraph defined by the variables/cuts (edges) and constraints/plates (nodes) starting from the original plate.
In \emph{priced} variants of \emph{faithful} and \emph{enhanced} this \emph{purge} procedure is done unless stated otherwise.
Our experiments will show that this \emph{purge} drastically reduces the number of variables and constraints, but has almost no effect on the running times.
Nonetheless, we encourage future comparisons to implement this \emph{purge} procedure, as it helps determine the real size of the solved models.
Each experiment fills a gap for the next experiments:
\autoref{sec:lp_method} explains the choice of LP algorithms made in all remaining experiments;
\autoref{sec:faithful_reimplementation} provides evidence that \emph{faithful} is on par with \emph{original}, allowing us to use it as a replacement;
\autoref{sec:comparison} compares \emph{faithful} to \emph{enhanced} and shows the value of our contributions (namely, the \emph{normalize} procedure and the \emph{enhanced} formulation);
\autoref{sec:new_results} applies the methods with best results in the last experiment to prove new optimal values and bounds for harder instances.
\subsection{Setup}
\label{sec:setup}
Every experiment in this work uses the following setup unless stated otherwise.
The CPU was an AMD\textsuperscript{\textregistered} Ryzen\textsuperscript{TM} 9 3900X 12-Core Processor (3.8GHz, cache: L1 -- 768KiB, L2 -- 6 MiB, L3 -- 64 MiB) and 32GiB of RAM were available (2 x Crucial Ballistix Sport Red DDR4 16GB 2.4GHz).
The operating system used was Ubuntu 20.04 LTS (Linux 5.4.0-42-generic).
Hyper-Threading was disabled.
Each run executed on a single thread, and no runs executed simultaneously.
The computer did not run any other CPU bound task during the experiments.
The exact version of the code used is available online (\url{https://github.com/henriquebecker91/GuillotineModels.jl/tree/0.2.4}), and it was run using Julia 1.4.2~\cite{julia} with JuMP 0.20.1~\cite{JuMP} and Gurobi 9.0.2~\cite{gurobi}.
The following Gurobi parameters had non-default values: \verb+Threads+~\(= 1\); \verb+Seed+~\(= 1\); \verb+MIPGap+~\(= 10^{-6}\) (to guarantee optimality); and \verb+TimeLimit+~\(= 10800\) (i.e., three hours).
The next section explains the rationale for using \verb+Method+~\(= 2\) (i.e., barrier) to solve the root node relaxation of the final built model; and \verb+Method+~\(= 1\) (i.e., dual simplex) inside pricing (if pricing is enabled).
\subsection{The choice of LP algorithm}
\label{sec:lp_method}
Both \cite{furini:2016} and \cite{dimitri_thesis} do not specify the algorithm used for solving the MILP root node relaxation and, if pricing is enabled, for solving some LP models (upper bound computation) and the MILP root node relaxation of the \emph{restricted priced} model.
As we use Gurobi, we are discussing the \verb+Method+ parameter (for LP models and MILP root node relaxations), and not the \verb+NodeMethod+ parameter (for non-root nodes).
The choice of the algorithm can drastically impact running times.
A preliminary experiment included all LP algorithms available in Gurobi.
\autoref{tab:lp_method_comparison} presents the data of the two algorithms selected for use.
They are the \emph{Dual Simplex} and the \emph{Barrier}.
The runs use the \emph{faithful} implementation, with \emph{Cut-Position} and \emph{Redundant-Cut} enabled, in its \emph{priced} (Priced PP-G2KP in~\cite{furini:2016,dimitri_thesis}) and \emph{not priced} (PP-G2KP in~\cite{furini:2016,dimitri_thesis}) variants.
For convenience, we limited the experiment to a few instances.
This subset consists of all instances for which the \emph{Complete PP-G2KP Model} finds the optimal solution within the time limit in~\cite{furini:2016} (Table 2).
If pricing is disabled, the root node relaxation contributes for the majority of the running time.
This characteristic makes them a good choice for this experiment.
\begin{table}
\caption{Comparison of LP-solving algorithms used inside solving procedure.}
\begin{tabular}{@{\extracolsep{4pt}}lrrrrrrr@{}}
\hline\hline
Instance & \multicolumn{3}{c}{Dual Simplex} & \multicolumn{3}{c}{Barrier} & DS + B \\\cline{2-4}\cline{5-7}
& N. P. & R. \% & Priced & N. P. & R. \% & Priced & Priced \\\hline
CU1 & 27.37 & 92.11 & 3.79 & 24.18 & 94.68 & 3040.82 & \textbf{3.58} \\
STS4 & 93.49 & 89.88 & 48.80 & 49.94 & 77.32 & 7851.30 & \textbf{47.75} \\
STS4s & 103.20 & 94.92 & 39.29 & 43.74 & 86.34 & 8470.41 & \textbf{38.36} \\
gcut9 & 226.68 & 72.29 & \textbf{3.92} & 51.48 & 85.77 & 2060.04 & 4.01 \\
okp1 & 51.95 & 84.18 & 38.89 & \textbf{32.41} & 67.78 & -- & 38.79 \\
okp4 & 98.25 & 93.35 & 144.30 & \textbf{72.09} & 92.31 & -- & 141.53 \\
okp5 & 178.13 & 89.89 & 252.09 & \textbf{96.38} & 67.24 & -- & 239.44 \\\hline\hline
\end{tabular}
\label{tab:lp_method_comparison}
\end{table}
In \autoref{tab:lp_method_comparison}, \emph{Dual Simplex} and \emph{Barrier} indicate the respective algorithm was used for all LPs and root node relaxations;
and \emph{DS + B} means that \emph{Dual Simplex} was used to solve all LPs inside the pricing phase and \emph{Barrier} was used to solve the root node relaxation of the final model.
The columns \emph{N. P.} (\emph{Not Priced}) and \emph{Priced} display the time to solve (in seconds) using the aforementioned variant.
The columns \emph{R.\%} refer to the per cent of the time spent by \emph{Not Priced} in the root node relaxation of the final model.
The following conclusions can be derived from \autoref{tab:lp_method_comparison}.
Using the \emph{Barrier} algorithm in the pricing phase is not viable.
This impracticality happens because the pricing phase includes an iterative variable pricing phase.
This iterative phase repeatedly adds variables to one LP model and solve it again.
The \emph{Barrier} algorithm solves every LP from scratch;
the \emph{Dual Simplex} reuses the previous basis and saves considerable effort.
However, \emph{Barrier} performs better if there is no previous base to reuse.
Consequently, the configuration chosen was \emph{Dual Simplex} for the pricing phase, and \emph{Barrier} for the root relaxation of the final model.
\subsection{Comparison of \emph{faithful} against \emph{original}}
\label{sec:faithful_reimplementation}
Without a reimplementation of \emph{original}, any comparison would need to be made directly against the data in~\cite{dimitri_thesis}.
However, such comparison would hardly be fair, as it compares across machines, solvers, and programming languages.
Also, for example, it does not allow us to assess the benefits of applying the \emph{plate-size normalization} procedure to the \emph{original} formulation.
The purpose of this section is to show that \emph{faithful} may be fairly used in place of \emph{original}.
For this purpose, \autoref{tab:faithful_reimplementation} compares the number of model variables and number of plates of the diverse model variants presented in~\cite{furini:2016,dimitri_thesis} (using the same 59 instances).
The number of enumerated plates has a strong correlation to the number of constraints in the model.
Both~\cite{furini:2016} and~\cite{dimitri_thesis} present the number of plates and not the number of constraints.
To simplify the comparison, we do the same.
The \emph{Priced PP-G2KP} runs in~\cite{furini:2016,dimitri_thesis} had three time limits of one hour to solve: the restricted model (i.e., obtaining a lower bound); the iterative variable pricing (i.e., obtaining an upper bound); the final model.
Such configuration always generates a final model.
However, it also has two drawbacks:
(i) the computer performance may define the answer given in the first two phases, affecting the size of the final model (and making it harder to make a fair comparison);
(ii) if the restricted model, or the iterated variable pricing, cannot be done in one hour, then the final model will probably hit the time limit too -- in~\cite{furini:2016}, every run that hits one of the two first time limits also hits the third time limit.
We chose to use a single three-hour time limit.
\autoref{tab:faithful_reimplementation} references the names used in~\cite{furini:2016,dimitri_thesis}.
The \emph{Complete PP-G2KP} is the formulation with all optional procedures disabled, while the \emph{PP-G2KP} mean both \emph{Cut-Position} and \emph{Redundant-Cut} are enabled.
\emph{Restricted PP-G2KP} and its priced version are solved inside \emph{Priced PP-G2KP} runs.
The \emph{original} had no \emph{purge} phase after pricing.
Consequently, for the columns that refer to \emph{original}, the last row just repeats the data of the row above.
The sum of columns \emph{T. L.} (Time Limit) and \emph{E. R.} (Early Return) gives the number of instances excluded from consideration in the respective row.
Column \emph{T. L.} has the number of instances for which \emph{faithful} reached the time limit without generating the respective model variant -- these instances are: Hchl7s, okp2, and okp3.
The column \emph{E. R.} has the number of instances for which our reimplementation found an optimal solution before generating the respective model variant\footnote{
If the lower and upper bounds found during pricing are the same, then the optimal solution was found before generating the final model.
The instances in which this happened for an unrestricted solution are 3s, A1s, CU1, CU2, W, cgcut1, and wang20.
The instance A1s presented this behaviour already in the pricing of the restricted model.
}.
Columns \emph{O. \#v} and \emph{O. \#v} refer to \emph{original}.
Column \emph{O. \#v} (\emph{O. \#p}) presents the sum of variables (plates) for the instances in which \emph{faithful} generated a model.
Columns \emph{F. \%v} and \emph{F. \%p} refer to \emph{faithful}.
Column \emph{R. \%v} (\emph{R. \%p}) has the sum of variables (plates) in the generated models, as a percentage of the quantity obtained by the original implementation.
\begin{table}
\caption{Comparison of \emph{faithful} against \emph{original}.}
\begin{tabular}{lccrrrr}
\hline\hline
Variant & T. L. & E. R. & O. \#v & F. \%v & O. \#p & F. \%p\\\hline
Complete PP-G2KP & 0 & 0 & 156,553,107 & 100.00 & 1,882,693 & 100.00\\
Complete +Cut-Position & 0 & 0 & 103,503,930 & 99.99 & 1,738,263 & 100.01\\
Complete +Redundant-Cut & 0 & 0 & 121,009,381 & 109.94 & 1,882,693 & 100.00\\
PP-G2KP (CP + RC) & 0 & 0 & 74,052,541 & 120.05 & 1,738,263 & 100.01\\
Restricted PP-G2KP & 0 & 0 & 5,335,976 & 99.28 & 306,673 & 99.99\\
Priced Restricted PP-G2KP & 0 & 1 & 3,904,683 & 102.20 & 305,690 & 99.99\\
(no purge) Priced PP-G2KP & 3 & 7 & 14,619,460 & 93.74 & 1,642,382 & 100.01\\
Priced PP-G2KP & 3 & 7 & 14,619,460 & 31.92 & 1,642,382 & 25.55\\\hline\hline
\end{tabular}
\label{tab:faithful_reimplementation}
\end{table}
The following conclusions can be derived from \autoref{tab:faithful_reimplementation}.
All variants, except \emph{Priced PP-G2KP}, are within \(\pm0.01\)\% of the expected number of plates (and, consequently, of constraints).
The \emph{Complete PP-G2KP}, \emph{Complete +Cut-Position}, and \emph{Restricted PP-G2KP} are within \(\pm1\)\% of the expected number of variables.
The number of variables in both \emph{Complete +Redundant-Cut} and \emph{PP-G2KP (CP + RC)} is \(10\sim20\)\% larger than expected.
Our reimplementation of \emph{Redundant-cut} reduction seems responsible for both deviations.
However, it follows closely the description given in~\cite{dimitri_thesis}.
The number of variables and plates in \emph{Priced} variants is not entirely deterministic.
The number of variables of \emph{Priced} variants is either slightly above (\(+2\)\%) or lower (\emph{\(-6\sim68\)\%}).
For all non-\emph{priced} variants, the fraction of the running time spent in the model generation is negligible.
Consequently, the comparison presented in~\autoref{tab:faithful_reimplementation} is sufficient.
We cannot say the same for the \emph{priced} variants.
\cite{furini:2016,dimitri_thesis} does not report the size of the multiple LP models solved inside the iterative pricing (a phase of the pricing).
For instances in which \emph{original} and \emph{faithful} executed all phases of pricing and solved the final model, the \emph{original} spent 34.35\% of its time in the iterative pricing phase, while \emph{faithful} spent 61.69\%.
It is hard to pinpoint the source of this discrepancy.
One possible explanation is that, in \emph{original}, other phases took more time than they took in \emph{faithful}.
For example, \emph{faithful} uses the \emph{barrier} algorithm for the root node relaxation of the final model, which reduces the percentage of time spent in this phase.
Nevertheless, for the subset of the instances aforementioned, the total time spent by \emph{faithful} was about 13\% of the time spent by \emph{original}.
While the difference between machines and solvers does not allow us to infer much from that figure, we believe that the magnitude of the difference guarantees that we are not making a gross misrepresentation.
\subsection{Comparison of \emph{faithful} against \emph{enhanced}}
\label{sec:comparison}
The primary purpose of this section is to evaluate our contributions to the state of the art.
Our contributions are the \emph{normalize} reduction (i.e., the plate-size normalization presented in~\autoref{sec:psn}) and the \emph{enhanced} formulation (presented in \autoref{sec:enhanced}).
The state of the art consists in a formulation (\emph{Complete PP-G2KP}), two reductions (\emph{Cut-Position} and \emph{Redundant-Cut}), and a pricing procedure presented in~\cite{furini:2016,dimitri_thesis}.
In this section, we use our reimplementation of \emph{Complete PP-G2KP} named \emph{faithful} (to distinguish from the data of the \emph{original}).
We also reimplemented the reductions and the pricing procedure, but as \emph{enhanced} may also enable them, we avoid labelling these procedures as \emph{faithful} as to avoid confusion.
The \emph{faithful} and \emph{enhanced} formulations cannot be combined.
However, both allow enabling any combination of the optional procedures.
The only exception is \emph{Redundant-Cut}, which is unnecessary for \emph{enhanced} and, therefore, never applied to it.
Outside of this exception, in this section, \emph{Redundant-Cut} and \emph{Cut-Position} are always enabled.
These reductions never increase the number of variables (or constraints), cost a negligible amount of computational effort, and were already discussed in~\cite{furini:2016,dimitri_thesis}.
We also examine the effects of our \emph{purge} procedure and warm-starting the non-\emph{priced} model.
The deterministic heuristic used to MIP-start the non-\emph{priced} models is the same used in the \emph{restricted priced} model solved inside the pricing procedure.
The meaning of the columns in~\autoref{tab:contribution} follow:
\emph{T. T.} (Total Time) -- sum of the time spent in all instances including timeouts, in seconds;
\emph{\#e} (early) -- number of instances in which pricing found an optimal solution (and, consequently, did not generate a final model);
\emph{\#m} (modeled) -- number of instances that generated a final model;
\emph{\#s} (solved) -- number of solved instances;
\emph{\#b} (best) -- number of instances that the respective variant solved faster than any other variant;
\emph{S. T. T.} (Solved Total Time) -- same as Total Time but excluding runs ended by time or memory limit;
\emph{\#variables} (\emph{\#plates}) -- sum of the variables (plates) in all generated final models (see column~\emph{\#m}).
The first row (Faithful) has two runs that ended in memory exhaustion.
We count the time of these runs as they were timeouts.
\begin{table}
\rowcolors{1}{white}{gray-table-row}
\caption{Comparison of \emph{faithful} vs. \emph{enhanced} over the 59 instances used in~\cite{dimitri_thesis}.}
\begin{tabular}{lrrrrrrrr}
\hline\hline
Variant & T. T. & \#e & \#m & \#s & \#b & S. T. T. & \#variables & \#plates \\\hline
Faithful & 106,057 & -- & 59 & 53 & 0 & 41,257 & 88,901,964 & 1,738,366 \\
Enhanced & 25,538 & -- & 59 & 58 & 2 & 14,738 & 3,216,774 & 231,836 \\
F. +Normalizing & 60,078 & -- & 59 & 56 & 0 & 27,678 & 60,316,964 & 610,402 \\
E. +Normalizing & 14,169 & -- & 59 & 59 & 52 & 14,169 & 2,733,125 & 145,157 \\
F. +N. +Warming & 60,542 & -- & 59 & 56 & 0 & 28,142 & 60,316,964 & 610,402 \\
E. +N. +Warming & 9,778 & -- & 59 & 59 & 4 & 9,778 & 2,733,125 & 145,157 \\
Priced F. +N. +W. & 49,919 & 8 & 50 & 55 & 0 & 6,719 & 3,210,857 & 174,214 \\
Priced E. +N. +W. & 9,108 & 8 & 51 & 59 & 1 & 9,108 & 600,778 & 64,904 \\
P. F. +N. +W. -Purge & 50,054 & 8 & 50 & 55 & 0 & 6,854 & 8,072,810 & 544,892 \\
P. E. +N. +W. -Purge & 9,209 & 8 & 51 & 59 & 0 & 9,209 & 1,021,526 & 134,102 \\\hline\hline
\end{tabular}
\label{tab:contribution}
\end{table}
Considering the data from~\autoref{tab:contribution} we can state that:
\begin{enumerate}
\item \emph{enhanced} solves more instances than \emph{faithful} (using at most 24\% of its time);
\item the number of variables of `Enhanced' is almost the same as `Priced F. +N. +W.';
\item between `Enhanced' and `Priced F. +N. +W.' the former has better results;
\item \emph{normalize} further reduces variables by \(14\sim32\)\% and plates by \(37\sim65\)\%;
\item MIP-starting \emph{enhanced} makes its slightly slower in 52 instances;
\item MIP-starting \emph{enhanced} saves more than one hour in the other 7 instances;
\item any benefit from MIP-start in `F. +N. +Warming' was negated by its timeouts;
\item \emph{purge} greatly reduces the model size but has almost no effect on running time;
\item the effects of applying \emph{pricing} to \emph{enhanced} are not much better than \emph{purge};
\item applying \emph{pricing} to \emph{faithful} is positive overall but loses one solved instance.
\end{enumerate}
In~\autoref{tab:time_fractions}, \emph{Time} is the sum of all time (in seconds) spent in the 47 instances that had all phases executed by all four variants considered.
These are the same 47 indicated in row \emph{Priced F. +N. +W.} of \autoref{tab:contribution}.
From the 59 instances dataset, 4 had timeout (Hchl4s, Hchl7s, okp2, and okp3), and 8 found an optimal solution inside pricing (3s, A1s, CU1, CU2, W, cgcut1, okp4, and wang20).
All remaining columns present percentages of the time spent in a specific phase:
\emph{E} -- enumeration of cuts and plates (and all reductions);
\emph{H} -- restricted heuristic used to warm-start the restricted priced model;
\emph{RP} -- restricted pricing (not including the heuristic time);
\emph{IP} -- iterative pricing;
\emph{FP} -- final pricing;
\emph{LP} -- root node relaxation of the final model;
\emph{BB} -- branch-and-bound over the final model.
\begin{table}
\rowcolors{1}{white}{gray-table-row}
\caption{Fraction of the total time spent in each step (only runs that executed all steps).}
\begin{tabular}{lrrrrrrrrr}
\hline\hline
Variant & Time & E~\% & H~\% & RP~\% & IP~\% & FP~\% & LP~\% & BB~\% \\\hline
Priced Faithful +N. +W. & 6,632 & 0.12 & 0.38 & 26.16 & 57.36 & 2.91 & 4.56 & 8.29 \\
Priced Enhanced +N. +W. & 1,178 & 0.03 & 2.18 & 50.89 & 23.66 & 0.46 & 2.70 & 19.95 \\
P. F. +N. +W. -Purge & 6,766 & 0.11 & 0.37 & 26.00 & 57.03 & 2.81 & 5.12 & 8.45 \\
P. E. +N. +W. -Purge & 1,185 & 0.03 & 2.18 & 50.70 & 23.64 & 0.46 & 2.83 & 20.09 \\\hline\hline
\end{tabular}
\label{tab:time_fractions}
\end{table}
Considering the data from~\autoref{tab:time_fractions} we can state that:
\begin{enumerate}
\item both \emph{BB} and \emph{LP} phases are slightly faster with \emph{purge} as expected;
\item both \emph{E} and \emph{H} phases are almost negligible (at most 2\% with \emph{H} in \emph{enhanced});
\item together the \emph{RP} and \emph{IP} phases account for \(74.5\sim83.5\)\%;
\item \emph{RP} and \emph{IP} swap percentages between \emph{enhanced} and \emph{faithful};
\item \emph{faithful} shows some overhead in all phases strongly affected by model size.
\end{enumerate}
\subsection{Evaluating \emph{enhanced} against harder instances}
\label{sec:new_results}
The purposes of the experiment described in this section are:
(i) to show the limitations of the \emph{enhanced} formulation against more challenging instances;
(ii) to provide better bounds and new proven optimal values for such instances.
\cite{velasco:2019} proposes a set of 80 hard instances to test the limitations of their bounding procedures; we use these instances in this section.
Only two variants were executed for this experiment, the \emph{priced} and non-\emph{priced} versions of \emph{enhanced} with \emph{Cut-Position}, \emph{normalize}, and \emph{MIP-start} enabled.
We also present the results for the \emph{restricted priced} variant because it executes inside \emph{priced} (the same reductions apply to it).
\autoref{tab:velasco_summary} presents a summary of all runs, and \autoref{tab:velasco_new_results} presents the improved bounds and solved instances.
For this experiment, Gurobi was allowed to use the 12 physical cores of our machine.
Gurobi distributes the effort of the B\&B phase equally among all cores.
However, solving an LP (as a root node relaxation, or not) calls barrier, primal simplex, and dual simplex.
Each of these three uses a single thread, and Gurobi stops when the first of them finish.
\autoref{tab:velasco_summary} columns are:
\emph{C.} -- instance class (described in~\cite{velasco:2019}, 20 instances each);
\emph{Variant} -- the solving method employed;
\emph{\#m} (modeled) -- number of instances in which the model was built before timeout;
\emph{Avg. \#v} and \emph{Avg. \#p} -- the average number of variables and plates in the \emph{\#m} instances that generated a final model for the respective variant;
\emph{T. T.} (Total Time) -- sum of the time spent in all instances in seconds, including timeouts;
\emph{\#s} (solved) -- number of instances solved;
\emph{Avg. S. T.} (Avg. Solved Time) -- as total time but excludes timeouts and divides by \emph{\#s}.
Averages were used instead of simple sums because the very different number of generated and solved models made the sums misleading.
\begin{table}
\caption{Summary table for the instances proposed in~\cite{velasco:2019}.}
\begin{tabular}{lrrrrrrr}
\hline\hline
C. & Variant & \#m & Avg. \#v & Avg. \#p & T. T. & \#s & Avg. S. T. \\\hline
\multirow{3}{*}{1} & Not Priced & 20 & 1,787,864.55 & 22,316.50 & 172,574 & 5 & 2,114.85 \\
& Restricted Priced & 13 & 467,692.15 & 17,139.00 & 180,051 & 5 & 3,610.29 \\
\vspace{1.5mm} & Priced & 5 & 264,315.80 & 11,978.40 & 196,733 & 3 & 4,377.77 \\
\multirow{3}{*}{2} & Not Priced & 20 & 1,533,490.70 & 18,638.50 & 167,973 & 5 & 1,194.68 \\
& Restricted Priced & 20 & 453,159.70 & 18,638.30 & 155,184 & 8 & 3,198.11 \\
\vspace{1.5mm} & Priced & 8 & 394,613.88 & 9,735.50 & 178,812 & 4 & 1,503.01 \\
\multirow{3}{*}{3} & Not Priced & 20 & 2,895,300.75 & 33,249.40 & 171,155 & 5 & 1,831.11 \\
& Restricted Priced & 10 & 431,913.00 & 15,895.80 & 174,569 & 5 & 2,513.80 \\
\vspace{1.5mm} & Priced & 5 & 372,597.00 & 13,287.80 & 179,712 & 4 & 1,728.08 \\
\multirow{3}{*}{4} & Not Priced & 20 & 3,201,374.45 & 35,197.10 & 167,776 & 7 & 3,910.89 \\
& Restricted Priced & 10 & 497,802.20 & 17,011.00 & 197,047 & 2 & 1,323.65 \\
& Priced & 2 & 211,093.00 & 14,227.00 & 199,477 & 2 & 2,538.79 \\\hline\hline
\end{tabular}
\label{tab:velasco_summary}
\end{table}
Concerning the data from~\autoref{tab:velasco_summary}, we want to highlight some unexpected results:
(i) the total number of instances solved by the \emph{restricted priced} was slightly smaller than non-\emph{priced}, even with non-\emph{priced} solving the harder \emph{unrestricted} problem;
(ii) many runs reached time limit without solving the continuous relaxation of the \emph{restricted} model (necessary for creating \emph{restricted priced} model);
(iii) non-\emph{priced} solved more instances than \emph{priced} in all cases.
Ideally, the pricing procedure would significantly reduce the size of the model and, consequently, the root node relaxation and B\&B phases would take much less time to solve.
However, the gain in decreasing the size of the (already reduced) \emph{enhanced} model further does not seem to compensate for the cost of solving hard LPs more than once.
Also, previous sections have shown that reducing the model size does not guarantee that the running time will be reduced by the same magnitude.
The purpose of \autoref{tab:velasco_new_results} is to allow querying the exact values for specific instances.
Even so, there are some gaps to fill.
For the instances presented in \autoref{tab:velasco_new_results},
the min / mean / max gap between the heuristic lower bound and the final lower bound were: 0.38 / 18.08 / 37.03 (non-\emph{priced}); 0.68 / 20.62 / 37.29 (\emph{restricted priced}); 9.17 / 19.38 / 32.24 (\emph{priced}).
In other words, no solution, or best bound, was given by the heuristic, and most of the time, its solution was considerably improved.
For the reader convenience, we can also summarize that our experiment has:
proved 22 unrestricted optimal values (5 already proven by~\cite{velasco:2019}, confirming their results);
proved 22 restricted optimal values (in an overlapping but distinct subset of the instances);
improved lower bounds for 25 instances;
improved upper bounds for 58 instances.
\autoref{tab:velasco_new_results} groups lower and upper bounds that are valid for the unrestricted problem.
Column \emph{RP UB} (restricted priced upper bound) is kept separate as it is not a valid bound for the unrestricted problem.
Bold indicates the best unrestricted bounds for the instance.
For the same instance and variant, if the LB and the UB are the same, both values are underlined.
The sub-headers mean:
\emph{RP} -- Restricted Priced (solved inside \emph{P} runs);
\emph{P} -- Priced;
\emph{NP} -- Not Priced;
\emph{V\&U} -- obtained by Velasco and Uchoa in~\cite{velasco:2019}.
\begin{table}
\let\mc\multicolumn
\rowcolors{3}{white}{gray-table-row}
\caption{Instances solved (restricted or unrestricted) or with improved bounds.}
\begin{tabular}{lrrrrrrrr}
\hline\hline
\hiderowcolors
Instance & \mc4c{Lower Bounds for Unrestricted} & RP UB & \mc3c{Upper Bounds for Unr.} \\\cline{2-5}\cline{7-9}
& \mc1c{RP} & \mc1c{P} & \mc1c{NP} & \mc1c{V\&U} & & \mc1c{P} & \mc1c{NP} & \mc1c{V\&U} \\\hline
\showrowcolors
P1\_100\_200\_25\_1 & \underline{\textbf{27,251}} & \underline{\textbf{27,251}} & \underline{\textbf{27,251}} & \textbf{27,251} & \underline{27,251} & \underline{\textbf{27,251}} & \underline{\textbf{27,251}} & 27,340 \\
P1\_100\_200\_25\_2 & \underline{\textbf{25,090}} & \textbf{25,090} & \textbf{25,090} & 24,870 & \underline{25,090} & 25,403 & \textbf{25,389} & 25,522 \\
P1\_100\_200\_25\_3 & \underline{\textbf{25,730}} & \textbf{25,730} & \textbf{25,730} & \textbf{25,730} & \underline{25,730} & 25,974 & \textbf{25,909} & 26,088 \\
P1\_100\_200\_25\_4 & \underline{26,732} & \underline{\textbf{26,896}} & \underline{\textbf{26,896}} & 26,769 & \underline{26,732} & \underline{\textbf{26,896}} & \underline{\textbf{26,896}} & 27,051 \\
P1\_100\_200\_25\_5 & \textbf{26,152} & -- & \textbf{26,152} & 25,772 & 26,565 & -- & \textbf{26,617} & 26,857 \\
P1\_100\_200\_50\_1 & 28,388 & -- & \underline{\textbf{28,440}} & 28,388 & 28,504 & -- & \underline{\textbf{28,440}} & 28,558 \\
P1\_100\_200\_50\_2 & \underline{\textbf{26,276}} & \underline{\textbf{26,276}} & \underline{\textbf{26,276}} & \textbf{26,276} & \underline{26,276} & \underline{\textbf{26,276}} & \underline{\textbf{26,276}} & 26,326 \\
P1\_100\_200\_50\_3 & \textbf{27,192} & -- & \textbf{27,192} & 27,165 & 27,536 & -- & \textbf{27,483} & 27,679 \\
P1\_100\_200\_50\_4 & 28,058 & -- & \textbf{28,095} & 27,977 & 28,345 & -- & \textbf{28,340} & 28,388 \\
P1\_100\_200\_50\_5 & \textbf{27,722} & -- & \underline{\textbf{27,722}} & 27,603 & 27,930 & -- & \underline{\textbf{27,722}} & 28,009 \\
P1\_100\_400\_25\_1 & 53,247 & -- & 53,008 & \textbf{53,904} & 54,540 & -- & \textbf{54,707} & 55,038 \\
P1\_100\_400\_25\_2 & -- & -- & 41,275 & \textbf{44,581} & -- & -- & \textbf{47,091} & 47,097 \\
P1\_100\_400\_25\_3 & 42,748 & -- & 46,222 & \textbf{47,455} & \textbf{\large \textasteriskcentered} & -- & \textbf{49,371} & 49,473 \\
P1\_100\_400\_25\_4 & -- & -- & 38,567 & \textbf{40,517} & -- & -- & \textbf{46,069} & 46,078 \\
P1\_100\_400\_25\_5 & 44,482 & -- & \textbf{53,220} & 53,205 & \textbf{\large \textasteriskcentered} & -- & 54,120 & \textbf{54,063} \\
P1\_100\_400\_50\_1 & -- & -- & 53,831 & \textbf{55,856} & -- & -- & \textbf{56,897} & 57,074 \\
P1\_100\_400\_50\_2 & -- & -- & 40,440 & \textbf{48,373} & -- & -- & \textbf{51,754} & 51,893 \\
P1\_100\_400\_50\_4 & -- & -- & \textbf{55,107} & 52,708 & -- & -- & \textbf{55,654} & 55,661 \\
P1\_100\_400\_50\_5 & -- & -- & \textbf{53,749} & 53,502 & -- & -- & \textbf{55,005} & 55,454 \\
P2\_200\_100\_25\_1 & \underline{\textbf{21,494}} & \underline{\textbf{21,494}} & \underline{\textbf{21,494}} & \underline{\textbf{21,494}} & \underline{21,494} & \underline{\textbf{21,494}} & \underline{\textbf{21,494}} & \underline{\textbf{21,494}} \\
P2\_200\_100\_25\_2 & \underline{25,244} & \underline{\textbf{25,413}} & \underline{\textbf{25,413}} & \textbf{25,413} & \underline{25,244} & \underline{\textbf{25,413}} & \underline{\textbf{25,413}} & 25,648 \\
P2\_200\_100\_25\_3 & \underline{25,282} & \textbf{25,397} & \textbf{25,397} & \textbf{25,397} & \underline{25,282} & \textbf{25,640} & 25,647 & 25,723 \\
P2\_200\_100\_25\_4 & 25,729 & -- & \textbf{25,734} & 25,437 & 26,181 & -- & \textbf{26,239} & 26,898 \\
P2\_200\_100\_25\_5 & \underline{26,211} & \textbf{26,413} & \underline{\textbf{26,413}} & 26,220 & \underline{26,211} & 26,728 & \underline{\textbf{26,413}} & 26,898 \\
P2\_200\_100\_50\_1 & \textbf{25,679} & -- & 25,626 & 25,627 & 26,233 & -- & \textbf{26,282} & 26,447 \\
P2\_200\_100\_50\_2 & \underline{\textbf{27,801}} & \underline{\textbf{27,801}} & \underline{\textbf{27,801}} & 27,789 & \underline{27,801} & \underline{\textbf{27,801}} & \underline{\textbf{27,801}} & 27,943 \\
P2\_200\_100\_50\_3 & \underline{27,435} & \textbf{27,453} & \textbf{27,453} & \textbf{27,453} & \underline{27,435} & 27,584 & \textbf{27,579} & 27,596 \\
P2\_200\_100\_50\_4 & 27,395 & -- & \textbf{27,439} & 27,362 & 27,668 & -- & \textbf{27,704} & 27,718 \\
P2\_200\_100\_50\_5 & \underline{\textbf{29,386}} & \underline{\textbf{29,386}} & \underline{\textbf{29,386}} & \underline{\textbf{29,386}} & \underline{29,386} & \underline{\textbf{29,386}} & \underline{\textbf{29,386}} & \underline{\textbf{29,386}} \\
P2\_400\_100\_25\_1 & 49,327 & -- & \textbf{49,947} & 49,026 & 50,218 & -- & \textbf{50,365} & 51,006 \\
P2\_400\_100\_25\_2 & 48,312 & -- & \textbf{48,542} & 47,773 & 49,268 & -- & \textbf{49,315} & 49,908 \\
P2\_400\_100\_25\_3 & \textbf{46,970} & -- & 46,860 & 45,406 & 47,113 & -- & \textbf{47,204} & 48,938 \\
P2\_400\_100\_25\_4 & \textbf{51,051} & -- & 49,847 & 49,521 & 51,526 & -- & \textbf{51,600} & 52,229 \\
P2\_400\_100\_25\_5 & \textbf{49,620} & -- & 48,832 & 47,403 & 50,440 & -- & \textbf{50,580} & 54,248 \\
P2\_400\_100\_50\_1 & \underline{54,550} & 54,550 & \textbf{54,679} & 52,890 & \underline{54,550} & 54,981 & \textbf{54,916} & 55,629 \\
P2\_400\_100\_50\_2 & \textbf{54,821} & -- & 54,768 & 53,492 & 55,183 & -- & \textbf{55,181} & 55,543 \\
P2\_400\_100\_50\_3 & 54,141 & -- & \textbf{54,747} & 54,216 & 55,537 & -- & \textbf{55,709} & 56,065 \\
P2\_400\_100\_50\_4 & 53,375 & -- & \textbf{54,240} & 48,649 & 54,857 & -- & \textbf{54,987} & 55,604 \\
P2\_400\_100\_50\_5 & \textbf{53,763} & -- & 53,541 & 50,047 & 54,893 & -- & \textbf{54,918} & 55,471 \\
P3\_150\_150\_25\_1 & \underline{29,896} & \underline{\textbf{29,989}} & \underline{\textbf{29,989}} & 29,896 & \underline{29,896} & \underline{\textbf{29,989}} & \underline{\textbf{29,989}} & 30,005 \\
P3\_150\_150\_25\_2 & \textbf{29,345} & -- & 29,196 & 29,101 & 29,906 & -- & 29,965 & \textbf{29,961} \\
P3\_150\_150\_25\_3 & \underline{\textbf{30,286}} & \underline{\textbf{30,286}} & \underline{\textbf{30,286}} & \textbf{30,286} & \underline{30,286} & \underline{\textbf{30,286}} & \underline{\textbf{30,286}} & 30,327 \\
P3\_150\_150\_25\_5 & \underline{\textbf{31,332}} & \textbf{31,332} & \textbf{31,332} & 30,924 & \underline{31,332} & 31,715 & \textbf{31,682} & 31,839 \\
P3\_150\_150\_50\_1 & \underline{31,377} & \underline{\textbf{31,701}} & \underline{\textbf{31,701}} & \textbf{31,701} & \underline{31,377} & \underline{\textbf{31,701}} & \underline{\textbf{31,701}} & 31,892 \\
P3\_150\_150\_50\_2 & 30,846 & -- & \textbf{30,884} & \textbf{30,884} & 31,110 & -- & \textbf{31,008} & 31,115 \\
P3\_150\_150\_50\_3 & \underline{32,037} & \underline{\textbf{32,121}} & \underline{\textbf{32,121}} & 32,050 & \underline{32,037} & \underline{\textbf{32,121}} & \underline{\textbf{32,121}} & 32,240 \\
P3\_150\_150\_50\_4 & \textbf{31,925} & -- & \underline{\textbf{31,925}} & \textbf{31,925} & 32,210 & -- & \underline{\textbf{31,925}} & 32,070 \\
P3\_150\_150\_50\_5 & \textbf{31,631} & -- & 31,521 & 31,448 & 31,857 & -- & \textbf{31,896} & 31,901 \\
P3\_250\_250\_25\_1 & -- & -- & 51,027 & \textbf{58,480} & -- & -- & \textbf{60,548} & 60,611 \\
P3\_250\_250\_25\_2 & -- & -- & 63,646 & \textbf{68,070} & -- & -- & \textbf{73,316} & 73,339 \\
P3\_250\_250\_50\_1 & -- & -- & 59,072 & \textbf{67,603} & -- & -- & \textbf{76,117} & 76,341 \\
P3\_250\_250\_50\_2 & -- & -- & 62,772 & \textbf{75,569} & -- & -- & \textbf{82,644} & 82,666 \\
P4\_150\_150\_25\_1 & 30,870 & -- & \underline{\textbf{30,923}} & \textbf{30,923} & 31,094 & -- & \underline{\textbf{30,923}} & 31,130 \\
P4\_150\_150\_25\_2 & 30,576 & -- & \underline{\textbf{30,687}} & 30,460 & 30,786 & -- & \underline{\textbf{30,687}} & 30,931 \\
P4\_150\_150\_25\_3 & 30,257 & -- & \underline{\textbf{30,352}} & \underline{\textbf{30,352}} & 30,501 & -- & \underline{\textbf{30,352}} & \underline{\textbf{30,352}} \\
P4\_150\_150\_25\_4 & \underline{30,055} & \underline{\textbf{30,106}} & \underline{\textbf{30,106}} & \underline{\textbf{30,106}} & \underline{30,055} & \underline{\textbf{30,106}} & \underline{\textbf{30,106}} & \underline{\textbf{30,106}} \\
P4\_150\_150\_25\_5 & \textbf{30,582} & -- & 30,102 & \textbf{30,582} & 30,952 & -- & \textbf{31,228} & 31,286 \\
P4\_150\_150\_50\_1 & \underline{\textbf{31,673}} & \underline{\textbf{31,673}} & \underline{\textbf{31,673}} & \underline{\textbf{31,673}} & \underline{31,673} & \underline{\textbf{31,673}} & \underline{\textbf{31,673}} & \underline{\textbf{31,673}} \\
P4\_150\_150\_50\_2 & 32,302 & -- & \underline{\textbf{32,317}} & \textbf{32,317} & 32,434 & -- & \underline{\textbf{32,317}} & 32,423 \\
P4\_150\_150\_50\_3 & 30,906 & -- & \textbf{30,913} & 30,882 & 31,500 & -- & \textbf{31,519} & 31,756 \\
P4\_150\_150\_50\_4 & 31,912 & -- & \underline{\textbf{31,961}} & 31,912 & 32,206 & -- & \underline{\textbf{31,961}} & 32,140 \\
P4\_150\_150\_50\_5 & \textbf{32,027} & -- & 31,845 & 31,864 & 32,331 & -- & \textbf{32,308} & 32,484 \\
P4\_250\_250\_25\_4 & -- & -- & 69,530 & \textbf{79,476} & -- & -- & \textbf{81,634} & 81,839 \\
P4\_250\_250\_50\_2 & -- & -- & 67,675 & \textbf{77,206} & -- & -- & \textbf{87,314} & 87,331 \\
P4\_250\_250\_50\_4 & -- & -- & 69,063 & \textbf{78,359} & -- & -- & \textbf{86,941} & 87,069 \\\hline\hline
\end{tabular}
\textbf{\large \textasteriskcentered} These runs hit the time limit at the very start of the upper bound computation and, consequently, they produced only large and irrelevant upper bounds, which we omit to keep the table formatting.
\label{tab:velasco_new_results}
\end{table}
\section{Conclusions}
\label{sec:conclusions}
The present work advances the state of the art on MILP formulations for the G2KP.
We improve the performance of one of the most competitive MILP formulations for the G2KP by at least one order of magnitude.
In the instance set selected by the original formulation, our enhanced formulation dominates the original formulation.
Concerning other competitive MILP formulations in the literature, we keep the advantage of tighter bounds the original formulation had over them, and greatly reduce the model size and running times for instances that these other formulations had the advantage.
In the experiments, we have already discussed some elementary inferences, for example: the limitations (and partial success) of our improved formulation against the most recent and challenging instances in the literature; and the impact on the performance caused by the LP-solving algorithm, by the specific changes we made, by MIP-starting the models, and by some procedures proposed together with the original model (i.e., pricing and some preprocessing reductions).
Here we present more general conclusions from a broader perspective.
\emph{We believe symmetry-breaking plays a significant part in the success of our enhanced formulation.}
In our experiments, we focus on the significant reduction of the model size because it is easier to measure.
However, in \autoref{sec:comparison}, by comparing formulations with and without the \emph{purge} procedure, we see that a significant reduction of the model size does not always lead to a significant reduction of running times.
In the case of the variables removed by the \emph{purge} procedure (which could never assume a nonzero value), it seems clear the solver was able to disregard them without the need of our explicit removal.
The same does not apply to the variables removed by our enhanced model, which could assume nonzero values and compose symmetric solutions.
A single extraction variable may replace many distinct sequences of cuts that would extract the same piece from the same slightly-larger plate.
We also believe our results suggest that clever dominance rules may considerably improve pseudo-polynomial models (which often have tight bounds but large formulations) before resorting to more complicated techniques (as the pricing procedure proposed in~\cite{furini:2016} or column generation techniques)
\emph{Limited parallelisation of solving LP models is becoming a bottleneck.}
Obtaining tighter bounds, even at the cost of larger model size, is often valuable.
Some recent examples of this trade-off are pseudo-polynomial models like ours, but exponential-sized models solved by column generation are a pervasive and older example of the same trade-off.
In our experiment focusing on finding new optimal solutions for hard instances, it became clear that this approach shifts computational effort from the massively parallelisable B\&B phase to the almost serial root node relaxation phase.
This effect postpones finding the first primal solution and diminishes the value in massive computer clusters.
Our suggestions for future works follow: adapt the formulation for closely related problem variants and compare to their state-of-the-art solving procedure; expand on the symmetry-breaking; search for more parallelisable ways of solving LPs; consider other frameworks besides the pricing framework of~\cite{furini:2016}.
\begin{comment}
\pagebreak
|
\section{Introduction}
Several real-world problems often involve a series of selfish agents optimizing their benefits while mutually affecting their decisions. The concept of Nash equilibrium \citep{nash_equilibrium_1950,nash_non-cooperative_1951} revolutionized the understanding of the agents' strategic behavior by proposing a flexible and interpretable solution, with consequences and applications in many different contexts. The Nash equilibrium constitutes a \emph{stable} solution, meaning that no single agent has an incentive to defect from it profitably. Nash equilibria, however, may intrinsically differ in their features, for instance, in terms of a given welfare function measuring the common good for the collectivity of the agents. Above all, the quality of equilibria often does not match the quality of the \emph{social optimum}, i.e., the best possible solution for the collectivity. In general, the social optimum is not a stable solution and, therefore, does not emerge naturally from the agents' interactions. Nevertheless, in numerous contexts, a central authority may suggest solutions to the agents, preferably ensuring that such solutions satisfy two foremost properties. First, the authority should ensure that little to no incentives exist for the agents to refuse the proposed solution. Second, the solution should be sufficiently close -- in terms of quality -- to the social optimum. The best trade-off between these two properties is the \emph{best} Nash equilibrium, i.e., a solution that optimizes a welfare function among the equilibria. Often, the main focus is on selecting a Pure Nash Equilibrium (\emph{PNE}\xspace), a stable solution where each agent selects one alternative with probability one (in contrast to a Mixed-Strategy equilibrium, where agents randomize over the set of their alternatives).
In this context, the Algorithmic Game Theory (\emph{AGT}\xspace) community pioneered the study of the interplay between Game Theory and algorithms with a focus on equilibria's efficiency \citep{nisan_algorithmic_2008}. The discipline attracted significant attention from the computer science and optimization communities, especially to study games where agents solve optimization problems (e.g., \citet{facchinei_finite-dimensional_2004}).
Several recent works \citep{hemmecke_nash-equilibria_2009,koppe_rational_2011,sagratella_computing_2016,carvalho_computing_2017,cronert_equilibrium_2020,schwarze_branch-and-prune_2022,harks_generalized_2022,dragotto_zero_2021} considered Integer Programming Games (\emph{IPG}s\xspace), namely games where the agents solve parametrized integer programs. In this work, we focus on a class of \emph{simultaneous} and \emph{non-cooperative} \emph{IPG}s\xspace among $n$ players (agents), as described in \cref{def:Game}, where each player controls $m$ integer variables. %
\begin{definition}[\emph{IPG}\xspace]
Each player $i=1,2,\dots,n$ solves \eqref{eq:def:game}, where $u^i(x^i, x^{-i})$ -- given $x^{-i}$ -- is a function in $x^i$ with integer coefficients, $A^i \in \mathbb{Z}^{r\times m}$, $b^i \in \mathbb{Z}^r$.
\begin{align}
\max_{x^i} \{ u^i(x^i, x^{-i}): x^i \in \mathcal{X}^i \}, \; \mathcal{X}^i := \{ A^i x^i \le b^i, x^i \in \mathbb{Z}^{m} \}.
\label{eq:def:game}
\end{align}
\label{def:Game}
\end{definition}
\noindent As standard game-theory notation, let $x^i$ denote the vector of variables of player $i$, and let the operator $\left( \cdot \right)^{-i}$ be $\left( \cdot \right)$ except $i$. The vector $x^{-i}=(x^1, \dots, x^{i-1},$ $x^{i+1}, \dots x^n)$ represents the variables of $i$'s \emph{opponents} (all players but $i$), and the set of linear constraints $A^i x^i \le b^i$ defines the feasible region $\mathcal{X}^i$ of player $i$. We assume all integer variables are lower and upper bounded, and thus that $\mathcal{X}^i$ is finite. In \emph{IPG}s\xspace, the strategic interaction occurs in the players' objective functions, and not within their feasible regions.
Specifically, players choose their strategy simultaneously, and each player $i$'s utility (or payoff) $u^i(x^i, x^{-i})$ is a function in $x^i$ parametrized in $i$'s opponents variables $x^{-i}$. Without loss of generality, we assume the entries of $A^i$ and $b^i$ and the coefficients of $u^i(x^i, x^{-i})$ are integers. Further, considering the space of all players' variables $(x^1,\dots,x^n)$, we assume one can always linearize the non-linear terms in each $u^i$ with a finite number of inequalities and auxiliary variables (e.g., \citet{sherali_reformulation-linearization_1999,vielma_mixed_2015}). We remark that this assumption is not restrictive; on the contrary, it enables us to tackle several games where the players' utilities are not linear (see \cref{sec:games}).
Besides, we assume \begin{enumerate*} \item players have \emph{complete information} about the structure of the game, i.e., each player knows the other players' optimization problems via their feasible regions and objectives, \item each player is \emph{rational}, namely it always selects the best possible strategy given the information available on its opponents, and \item \emph{common knowledge of rationality}, namely each player knows its opponents are rational, and there is complete information \end{enumerate*}.
\emph{IPG}s\xspace extend traditional resource-allocation tasks and combinatorial optimization problems to a multi-agent setting, and their modeling power lies precisely in the discrete variables and game dynamics they can model. Indeed, in several real-world applications, requirements such as indivisible quantities and fixed production costs often require the use of discrete variables (see, for instance, \citet{bikhchandani_competitive_1997}). Several recent works explored the application of \emph{IPG}s\xspace in various contexts. To name a few, \citet{gabriel_solving_2013} modeled energy production games,
\citet{david_fuller_alternative_2017} proposed discrete unit commitment problem with fixed production costs,
\citet{anderson_supplier_2017} modeled a game where firms reserve discrete blocks of capacities from their suppliers, \citet{federgruen_multi-product_2015} proposed a price competition framework with $n$ competitors offering a discrete number of substitutable products, and \citet{carvalho_nash_2017} exploited \emph{IPG}s\xspace in the context of kidney exchange programs.
Despite the high potential impact of \emph{IPG}s\xspace in many domains, practitioners and researchers often make restrictive assumptions about the game's structure to guarantee that solutions are unique or computationally tractable. This is mainly due to the lack of a general, scalable and reliable methodology to select efficient solutions in \emph{IPG}s\xspace, which could potentially open new opportunities in terms of applications.
This lack is the core motivation behind our work: providing a general-purpose algorithmic framework to optimize over the solutions of \emph{IPG}s\xspace. Specifically, we focus on optimizing over the set of \emph{PNE}s\xspace for the \emph{IPG}s\xspace defined above and on characterizing the polyhedral structure of the set containing the \emph{PNE}s\xspace. The algorithmic framework possesses a solid theoretical foundation, and it integrates with the existing tools from the theory and practice of integer programming and combinatorial optimization. From a computational perspective, it is highly flexible, and it generally outperforms the algorithms available in the present literature.
Our framework is problem-agnostic and general, yet, it can be customized to address problem-specific needs.
\paragraph{Literature. } \citet{koppe_rational_2011} pioneered \emph{IPG}s\xspace by laying down their first formal definition. The authors also provided an algorithmic framework to enumerate \emph{PNE}s\xspace when the players' utilities are differences of piecewise linear convex payoff functions. Although their approach is theoretically well-grounded, there is no computational evidence of its effectiveness.
Indeed, even in some 2-player games (e.g., normal-form \citep{rosenberg_enumeration_2005,avis_enumeration_2010} and bimatrix \citep{audet_enumeration_2006} games) there are considerable computational challenges involved in the design of efficient algorithms for computing and selecting equilibria.
\citet{sagratella_computing_2016} proposed a branching method to enumerate the \emph{PNE}s\xspace in \emph{IPG}s\xspace where each $u^i(x^i,x^{-i})$ is convex in $x^i$. More recently, \citet{schwarze_branch-and-prune_2022} extended the work of \citet{sagratella_computing_2016} by proposing an improved branch-and-prune scheme that also drops the convexity assumption on the players' utilities. \citet{del_pia_totally_2017} focused on totally-unimodular congestion games, namely \emph{IPG}s\xspace where players have totally-unimodular constraint sets $\mathcal{X}^i$. They propose a strongly-polynomial time algorithm to find a \emph{PNE}\xspace and derive some computational complexity results. Their results have been extended by \citet{kleer_computation_2021}. More recently, \citet{Dragotto_2021_CNP} proposed a general-purpose cutting plane algorithm to compute a \emph{PNE}\xspace in \emph{IPG}s\xspace where each player utility is linear in their variables and bilinear with respect to the other players' variables. However, their approach does not handle equilibria selection, and requires a specific structure on the players' objectives to derive the Karush–Kuhn–Tucker conditions associated with the linear relaxation of their optimization problems.
An important family of techniques for computing Mixed-Strategy equilibria is the one of support enumeration algorithms. The core idea is to determine if an equilibrium with a given support for each player -- e.g., a subset of its strategies -- exists in a normal-form game by solving a linear system of inequalities. \citet{porter_simple_2008} and \citet{sandholm_mixed-integer_2005} exploited this idea in the context of $n$-players normal-form games. Since equilibria in such games tend to have small supports, as proved theoretically by \citet{mclennan_expected_2005}, support enumeration algorithms tend to be practically efficient in normal-form games.
Inspired by the approach of \citet{porter_simple_2008}, \citet{carvalho_computing_2017} introduced the sample generation method ($SGM$) to compute an equilibrium in separable \emph{IPG}s\xspace (i.e., where each player's payoff takes the form of a sum-of-products) where players have bounded strategy sets. Their algorithm iteratively refines a sample of players' supports to compute an equilibrium or a correlated equilibrium (i.e., a generalization of the Nash equilibrium). However, the $SGM$ does not handle the enumeration or selection of equilibria, nor can it prove that no equilibrium exists. \citet{cronert_equilibrium_2020} modified the $SGM$ -- extending the work of \citet{carvalho_computing_2017} -- by proposing an enumerative algorithm to compute all the equilibria with the additional assumptions that all the players' variables are integer. They further complemented their approach with some considerations stemming from the theory of equilibria selection of \citet{harsanyi_new_1995}.
Nevertheless, identifying the correct samples leading to equilibria in \emph{IPG}s\xspace could be computationally cumbersome. While our approach shares a few elements with \citet{cronert_equilibrium_2020}, it does not require any sampling in order to compute and select equilibria. This fundamental aspect leads to significant differences in terms of practical effectiveness and performance of the algorithms (see \cref{sec:games}).
Although the previous methodological works provide an insightful perspective on the computability and the selection of equilibria in \emph{IPG}s\xspace, there are other significant intrinsic questions concerning the general nature of equilibria.
Indeed, from the \emph{AGT}\xspace standpoint, not all equilibria are created equal. Three paradigmatic questions in \emph{AGT}\xspace and Game Theory are often: \begin{enumerate*} \item Does at least one \emph{PNE}\xspace exist? \item How good (or bad) is a \emph{PNE}\xspace compared to the social optimum? \item If more than one equilibrium exists, can one select the best \emph{PNE}\xspace according to a given measure of quality? \end{enumerate*}
Establishing that a \emph{PNE}\xspace does not exist may turn out to be a difficult task \citep{daskalakis_complexity_2009}. Nash proved that there always exists a Mixed-Strategy equilibrium in finite games, i.e., games with a finite number of strategies and players.
In \emph{IPG}s\xspace, where the set of players' strategies is large, deciding if a \emph{PNE}\xspace exists is generally a $\Sigma^p_2$-hard\xspace decision problem in the polynomial hierarchy \citep{vaz_existence_2018}.
To measure the efficiency of equilibria, \citet{goos_worst-case_1999} introduced the concept of \emph{Price of Anarchy} (\emph{PoA}\xspace), the ratio between the welfare value of the worst-possible equilibrium and the welfare value of a social optimum. Similarly, \citet{schulz_performance_2003,anshelevich_near-optimal_2003} introduced the \emph{Price of Stability} (\emph{PoS}\xspace), the ratio between the welfare value of the best-possible equilibrium and a social optimum's one. In the \emph{AGT}\xspace literature, many works focus on providing theoretical bounds for the \emph{PoS}\xspace and the \emph{PoA}\xspace, often by exploiting the game's structural properties \citep{anshelevich_near-optimal_2003,anshelevich_price_2008,chen_network_2006,nisan_algorithmic_2008,roughgarden_bounding_2004}. However, in practice, one may be interested in establishing the exact values of such prices in order to characterize the efficiency of equilibria in specific applications. This further highlights the need for general and effective algorithmic frameworks to select equilibria.
\paragraph{Contributions. }In this work, we shed new light on the intersection between \emph{AGT}\xspace and integer programming. We propose a new theoretical and algorithmic framework to efficiently and reliably compute, enumerate, and select \emph{PNE}s\xspace for the \emph{IPG}s\xspace in \cref{def:Game}.
We summarize our contributions as follows:
\begin{enumerate}
\item From a theoretical perspective, we provide a polyhedral characterization of the convex hull of the \emph{PNE}s\xspace. We adapt the concepts of valid inequality, closure, and separation oracle to the domain of Nash equilibria. Specifically, we introduce the concept of \emph{equilibrium inequality} to guide the exploration of the set of \emph{PNE}s\xspace. With this respect, we provide a general class of equilibrium inequalities and prove -- through the concept of \emph{equilibrium closure} -- they are sufficient to define the convex hull of the \emph{PNE}s\xspace.
From a game-theory standpoint, we explore the interplay between the concept of rationality and cutting planes through the \emph{equilibrium inequalities}. Since in any game, a player $i$ may never play some of its strategies due to their induced payoffs, it is reasonable to think that player $i$ would only pick its strategies from a \emph{rational} subset of $\mathcal{X}^i$. In other words, we provide an interpretable criterion -- in the form of a cutting plane -- for a player to play or not some strategies. In this sense, what we propose constitutes an analytical and geometrical characterization of the sets of equilibria providing a novel perspective on equilibria selection.
%
\item From a practical perspective, we design a cutting plane algorithm -- \emph{ZERO Regrets} -- that computes the most efficient \emph{PNE}\xspace for a given welfare function. This algorithm is flexible and scalable, it can potentially enumerate \emph{all} the \emph{PNE}s\xspace and compute approximate \emph{PNE}s\xspace. The algorithm exploits an \emph{equilibrium separation oracle}, a procedure separating non-equilibrium strategies from \emph{PNE}s\xspace through general and problem-specific equilibrium inequalities. Furthermore, our framework smoothly integrates with existing mathematical programming solvers, allowing practitioners to exploit the capabilities of the available optimization technologies.
%
\item We evaluate our algorithmic framework on a range of applications and problems from the relevant works in the literature. We provide a solid benchmark against the existing approaches and show the flexibility and effectiveness of \emph{ZERO Regrets}. The classes of games we select derive from practical applications (e.g., competitive facility locations, network design) and methodological studies and the associated benchmark instances (e.g., games among quadratic programs). First, we consider the Knapsack Game, an \emph{IPG}\xspace where each player solves a binary knapsack problem. For this problem, we also provide theoretical results on the computational complexity of establishing the existence of \emph{PNE}s\xspace and two problem-specific equilibrium inequalities. Second, we focus on a Network Formation Game, a well-known and intensely investigated problem in \emph{AGT}\xspace, where players build a network over a graph via a cost-sharing mechanism. Third, we consider a Competitive Facility Location and Design game, where several sellers strategically decide the location and design of their facilities in order to maximize their revenues. Finally, we test our algorithm on a game where players solve integer problems with convex and non-convex quadratic objectives. \emph{ZERO Regrets} outperforms any baseline, proving to be highly efficient in both enumerating and selecting \emph{PNE}s\xspace.
\end{enumerate}
We remark that our framework can be extended to the non-linear case, i.e., when $u^i$ is non-linearizable. However, we focus on the linear case
\begin{enumerate*}
\item to provide geometrical, polyhedral, and combinatorial insights on the structure of Nash equilibria in \emph{IPG}s\xspace, and
\item to foster the interaction with existing streams of research in Combinatorial Optimization.
\end{enumerate*}
We structure the paper as follows. In \cref{sec:definitions}, we introduce the fundamental definitions and terminology. In \cref{sec:eqInequalities} we introduce the theoretical elements of our algorithmic framework. In \cref{sec:CompPNE}, we describe our cutting plane algorithm and its separation oracle and their extensions to compute approximate equilibria. In \cref{sec:games} we present an extensive computational campaign on the applications mentioned above, and, in \cref{sec:concluding}, we provide some concluding remarks.
\section{Definitions}
\label{sec:definitions}
We assume the reader is familiar with basic concepts of polyhedral theory and integer programming \citep{conforti_integer_2014}. We introduce the notation and definitions related to an \emph{IPG}\xspace instance $G$, where we omit explicit references to $G$ when unnecessary. Let $\mathcal{X}^i$ be the \emph{set of feasible strategies} (or the feasible set) of player $i$, and let any strategy $\bar{x}^i \in \mathcal{X}^i$ be a \emph{(pure) strategy} for $i$.
Any $\bar{x}=(\bar{x}^1,\dots,\bar{x}^n)$ -- with $\bar{x}^i \in \mathcal{X}^i$ for any $i$ -- is a \emph{strategy profile}.
Let the vector $x^{-i}=(x^1,\dots, x^{i-1}, x^{i+1}, \dots x^n)$ denote the vector of \emph{the $i$'s opponents (pure) strategies}.
The \emph{payoff} for $i$ under the profile $\bar{x}$ is $u^i(\bar{x}^i,\bar{x}^{-i})$. We define $S(\bar{x})=\sum_{i=1}^n u^i(\bar {x}^i,\bar{x}^{-i})$ as the \emph{social welfare} corresponding to a given strategy profile $\bar{x}$.
\paragraph{Equilibria and Prices.}
A strategy $\bar{x}^i$ is a best-response strategy for player $i$ given its opponents' strategies $\bar{x}^{-i}$ if $u^i(\bar{x}^i,\bar{x}^{-i}) \ge u^i(\hat{x}^i,\bar{x}^{-i})$ for any $\hat{x}^i \in \mathcal{X}^i$; equivalently, we say $i$ cannot profitably deviate to any $\hat{x}^i$ from $\bar{x}^i$. The difference $u^i(\bar{x}^i,\bar{x}^{-i}) - u^i(\hat{x}^i,\bar{x}^{-i})$ is called the \emph{regret} of strategy $\hat{x}^i$ under $\bar{x}^{-i}$.
Let $\mathcal{BR}(i,\bar{x}^{-i})=\{ x^i \in \mathcal{X}^i: u^i(x^i,\bar{x}^{-i})\ge u^i(\hat{x}^i,\bar{x}^{-i}) \; \forall \; \hat{x}^i \in \mathcal{X}^i \}$ be the set of best-responses for $i$ under $\bar{x}^{-i}$. A strategy profile $\bar{x}$ is a \emph{PNE}\xspace if, for any player $i$ and any strategy $\hat{x}^i \in \mathcal{X}^i$, $u^i(\bar{x}^i,\bar{x}^{-i}) \ge u^i(\hat{x}^i,\bar{x}^{-i})$, i.e. any $\bar{x}^i$ is a best-response to $\bar{x}^{-i}$ (all regrets are $0$). Equivalently, in a \emph{PNE}\xspace, no player $i$ can unilaterally improve its payoff by deviating from its strategy $\bar{x}^i$. We define the \emph{optimal social welfare} as $OSW=\max_{x^1,\dots,x^n}\{ S(x) : x^i \in \mathcal{X}^i \; \forall i=1,2,\dots,n\}$. Given $G$, we denote as $\mathcal{N}=\{x=(x^1,\dots,x^n): x \text{ is a \emph{PNE}\xspace for } G\}$ the set of its \emph{PNE}s\xspace. Also, let $\mathcal{N}^i:=\{ x^i : (x^i,x^{-i}) \in \mathcal{N} \}$, with $\mathcal{N}^i \subseteq \mathcal{X}^i$ be the set of \emph{equilibrium strategies} for $i$, namely the strategies of $i$ appearing in at least a \emph{PNE}\xspace. If $\mathcal{N}$ is not empty, let: \begin{enumerate*} \item $\dot{x} \in \mathcal{N}$ be so that $S(\dot{x})\le S(\bar{x})$ for any $\bar{x} \in \mathcal{N}$ (i.e., the \emph{PNE}\xspace with the \emph{worst} welfare), and \item $\ddot{x} \in \mathcal{N}$ be so that $S(\ddot{x})\ge S(\bar{x})$ for any $\bar{x} \in \mathcal{N}$ (i.e., the \emph{PNE}\xspace with the \emph{best} welfare) \end{enumerate*}. Assuming w.l.o.g. $OSW > 0$ and $S(\ddot{x}) > 0$, the \emph{PoA}\xspace of $G$ is $\frac{OSW}{S(\dot{x})}$, and the \emph{PoS}\xspace is
$\frac{OSW}{S(\ddot{x})}$. The definitions of \emph{PoA}\xspace and \emph{PoS}\xspace hold when agents maximize a welfare function. Otherwise, when agents minimize their costs (e.g., the costs of routing packets in a network), we exchange numerator and denominator in both the \emph{PoA}\xspace and the \emph{PoS}\xspace.
\paragraph{Polyhedral Theory.} For a set $S$, let $\operatorname{conv}(S)$ be its convex hull. Let $P$ be a polyhedron: $\operatorname{bd}(P)$, $\operatorname{ext}(P)$, $\int(P)$, are the boundary, the set of vertices (extreme points), and the interior of $P$, respectively. Let $P \subseteq \mathbb{R}^p$ and $\tilde{x} \notin P$ a point in $\mathbb{R}^p$. A \emph{cut} is a valid inequality $\pi^\top x \le \pi_0$ for $P$ violated by $\tilde{x}$, i.e., $\pi^\top \tilde{x} > \pi_0$ and $\pi^\top x \le \pi_0$ for any $x \in P$.
Given a point $\hat{x} \in \mathbb{R}^p$ and $P$, we define the \emph{separation problem} as the task of determining that either \begin{enumerate*} \item $\hat{x} \in P$, or \item $\hat{x} \notin P$ and returning a cut $\pi^\top x \le \pi_0$ for $P$ and $\hat{x}$ \end{enumerate*}. For each player $i$, the set $\operatorname{conv}(\mathcal{X}^i)$ is the \emph{perfect formulation} of $\mathcal{X}^i$, namely an integral polyhedron whose vertices are in $\mathcal{X}^i$.
\section{Lifted Space and Equilibrium Inequalities}
\label{sec:eqInequalities}
Cutting plane methods are attractive tools for integer programs, both from a theoretical and an applied perspective. The essential idea is to iteratively refine a relaxation of the original problem by cutting off fractional solutions via valid inequalities for the integer program's perfect formulation.
Nevertheless, in an \emph{IPG}\xspace where the solution paradigm is the Nash equilibrium, we argue there exist stronger families of cuts, yet, not necessarily \emph{valid} for each player's perfect formulation $\operatorname{conv}(\mathcal{X}^i)$.
In fact, for any player $i$, some of its best-responses in $\operatorname{bd}(\operatorname{conv}(\mathcal{X}^i))$ may never appear in a \emph{PNE}\xspace, since no equilibrium strategies $\mathcal{N}^{-i}$ of $i$'s opponents induce $i$ to play such best-responses. In this work, we introduce a general class of inequalities to characterize the nature of $\operatorname{conv}(\mathcal{N})$. Such inequalities play a pivotal role in the cutting plane algorithm of Section \ref{sec:CompPNE}.
\paragraph{Dominance and Rationality. } We ground our reasoning in the concepts of \emph{rationality} and \emph{dominance} \citep{ pearce_rationalizable_1984,bernheim_rationalizable_1984}.
Given two strategies $\bar{x}^i \in \mathcal{X}^i$ and $\hat{x}^i \in \mathcal{X}^i$ for player $i$, $\bar{x}^i$ is strictly \emph{dominated} by $\hat{x}$ if, for any choice of opponents strategies $x^{-i}$, then $u^i(\hat{x},x^{-i}) > u^i(\bar{x},x^{-i})$. Then, a \emph{rational} player will never play dominated strategies. This also implies no player $i$ would play any strategy in $\int(\operatorname{conv}(\mathcal{X}^i))$. Since dominated strategies -- by definition -- are never best-responses, they will never be part of any \emph{PNE}\xspace. In \cref{ex:KPGBoundDominance}, the set $\mathcal{X}^2$ is made of $3$ strategies $(x^2_1,x^2_2)=(0,0)$, $(1,0)$, $(0,1)$.
Yet, $(x^2_1,x^2_2)=(0,0)$ is dominated by $(x^2_1,x^2_2)=(0,1)$, and the latter is dominated by $(x^2_1,x^2_2)=(1,0)$. However, when considering player 1, we need the assumption of \emph{common knowledge of rationality} to conclude which strategy the player will play. Player 1 needs to know that player 2 would never play $x^2_2 = 1$ to declare $(x^1_1,x^1_2)=(0,1)$ being dominated by $(x^1_1,x^1_2)=(1,0)$. When searching for a \emph{PNE}\xspace in this example, it follows that $\mathcal{N}^1=\{(x^1_1,x^1_2)=(1,0)\}$ and $\mathcal{N}^2=\{(x^2_1,x^2_2)=(1,0)\}$. This inductive (and iterative) process of removal of strictly dominated strategies is known as the \emph{iterated elimination of dominated strategies} (\emph{IEDS}\xspace). This process produces tighter sets of strategies and never excludes any \emph{PNE}\xspace from the game \citep[Ch.4]{tadelis_game_2013}.
\begin{example}
Consider the \emph{IPG}\xspace where player 1 solves $\max_{x^1}\{ 6x^1_1 + x^1_2 - 4x^1_1x^2_1 +6x^1_2x^2_2 : 3x^1_1 + 2x^1_2 \le 4, x^1\in\{0,1\}^2 \}$, and player 2 solves $\max_{x^2}\{ 4x^2_1 + 2x^2_2 -x^2_1x^1_1 -x^2_2x^1_2 : 3x^2_1 + 2x^2_2 \le 4, x^2\in\{0,1\}^2 \}$. The only \emph{PNE}\xspace is $(\bar{x}^1_1,\bar{x}^1_2)=(1,0)$, $(\bar{x}^2_1,\bar{x}^2_2)=(1,0)$ with a welfare of $S(\bar{x})=5$, $u^1(\bar{x}^1,\bar{x}^2)=2$, and $u^2(\bar{x}^2,\bar{x}^1)=3$.
\label{ex:KPGBoundDominance}
\end{example}
In the same fashion of \emph{IEDS}\xspace, we propose a family of inequalities that cuts off -- from each player's feasible set -- the strategies that never appear in a \emph{PNE}\xspace. Thus, from an \emph{IPG}\xspace instance $G$, we aim to derive an instance $G'$ where $\mathcal{N}^i$ replaces each player's feasible set $\mathcal{X}^i$. Note that, since all $\mathcal{X}^i$ are finite sets, all $\mathcal{N}^i$ are finite as well as the number of \emph{PNE}s\xspace.
\subsection{A Lifted Space}
Given the social welfare $S(x)$, we aim to find the \emph{PNE}\xspace maximizing it, namely, we aim to perform equilibria selection. In this context, the first urgent question is what space should we work in. Since mutually optimal strategies define \emph{PNE}s\xspace, a natural choice is to consider a space of all players' variables $x$. As mentioned in the introduction, we assume the existence of a higher-dimensional (lifted) space where we linearize the non-linear terms in any $u^i(\cdot)$ via auxiliary variables $z$ and corresponding constraints (e.g., \citet{sherali_reformulation-linearization_1999,vielma_mixed_2015}). Although our scheme holds for an arbitrary $f(x):\prod_{i=1}^n\mathcal{X}^i\rightarrow \mathbb{R}$ we can linearize to $f(x,z)$, we focus on $S(x)$ and the corresponding higher-dimensional $S(x,z)$ defined in the lifted space. Let $\mathcal{L}$ be the set of (i.) linear constraints necessary to linearize the non-linear terms, and (ii.) the integrality requirements and bounds on the $z$ variables. The lifted space is then
\begin{align}
\mathcal{K}= \{ (x^1,\dots,x^n,z) \in \mathcal{L}, x^i \in \mathcal{X}^i \text{ for any } i=1,\dots,n \}.
\label{eq:def:LiftedGame}
\end{align}
Any vector $x^1,\dots,x^n,z$ in (\ref{eq:def:LiftedGame}) corresponds to a unique strategy profile $x=(x^1,\dots,x^n)$, since $x$ induces $z$. $\mathcal{K}$ is then a set defined by linear constraints and integer requirements, and thus it is reasonable to deal with $\operatorname{conv}(\mathcal{K})$ and some of its projections. For brevity, let $\operatorname{proj}_{x}\operatorname{conv}(\mathcal{K})= \{x=(x^1,\dots,x^n) : \exists z \text{ s.t. } (x^1,\dots,x^n,z) \in \operatorname{conv}(\mathcal{K}) \}$, and let $u^i(x^i,x^{-i})$ include the $z$ variables when working in the space of $\operatorname{conv}(\mathcal{K})$.
\subsection{Equilibrium Inequalities}
The integer points in $\operatorname{proj}_{x}(\operatorname{conv}(\mathcal{K}))$ encompass all the game's strategy profiles. However, we need to focus on $\mathcal{E} =\{ (x^1,\dots,x^n,z) \in \operatorname{conv}(\mathcal{K}) : (x^1,\dots,x^n) \in \operatorname{conv}(\mathcal{N}) \}$, since projecting out $z$ yields the convex hull of \emph{PNE}\xspace profiles $\operatorname{conv}(\mathcal{N})$.
By definition $\mathcal{E} $ is a polyhedron, and $\operatorname{proj}_{x^i}(\mathcal{E}) = \operatorname{conv}(\mathcal{N}^i)$.
The role of $\mathcal{E}$ is similar to the one of a perfect formulation for an integer program. As optimizing a linear function over a perfect formulation results in an integer optimum, optimizing a linear function $S(x,z)$ over $\mathcal{E}$ results in a \emph{PNE}\xspace. For this reason, we call $\mathcal{E}$ the \emph{perfect equilibrium formulation} for $G$. Also, the equivalent of the integrality gap in integer programming is the \emph{PoS}\xspace, namely the ratio between the optimal value of $f(x,z)$ over $\operatorname{conv}(\mathcal{K})$ and $\mathcal{E}$, respectively.
All considered, we establish the concept of \emph{equilibrium inequality}, a valid inequality for $\mathcal{E}$.
\begin{definition}[Equilibrium Inequality]
Consider an \emph{IPG}\xspace instance $G$.
An inequality is an \emph{equilibrium inequality} for $G$ if it is a valid inequality for $\mathcal{E}$.
\label{def:EqIneq}
\end{definition}
\paragraph{A Class of Equilibrium Inequalities. } We introduce a generic class of equilibrium inequalities that are linear in the space of $\operatorname{conv}(\mathcal{K})$. Consider any strategy $\tilde{x}^i \in \mathcal{X}^i$ for $i$: for any $i$'s opponents' strategy $x^{-i}$, $u^i(\tilde{x}^i,x^{-i})$ provides a lower bound on $i$'s payoff since $\tilde{x}^i\in \mathcal{X}^i$ (i.e., $\tilde{x}^i$ is a feasible point). Then, $u^i(\tilde{x}^i,x^{-i}) \le u^i(x^i,x^{-i})$ holds for every player $i$. We introduce such inequalities in \cref{pro:Inequalities}.
\begin{proposition}
Consider an \emph{IPG}\xspace instance $G$. For any player $i$ and $\tilde{x}^i \in \mathcal{X}^i$, the inequality $u^i(\tilde{x}^i,x^{-i}) \le u^i(x^i,x^{-i})$ is an \emph{equilibrium inequality}.
\label{pro:Inequalities}
\end{proposition}
\iftoggle{ARXIV}{
\begin{proof}
}{
%
\proof{Proof of \cref{pro:Inequalities}.}
}
If a point $(\bar{x},\bar{z}) \in \mathcal{E}$, then $\bar x \in \operatorname{conv}(\mathcal{N})$. First, consider the case where $\bar{x} \in \operatorname{ext}(\operatorname{conv}(\mathcal{N}))$, namely $\bar{x} \in \mathcal{N}$ by definition. Assume $(\bar{x},\bar{z})$ violates the inequality associated with at least a player $i$, then, $u^i(\tilde{x}^i,\bar{x}^{-i}) > u^i(\bar{x}^i,\bar{x}^{-i})$. Therefore, $i$ can profitably deviate from $\bar{x}^i$ to $\tilde{x}^i$ under $\bar{x}^{-i}$, which contradicts $\bar{x} \in \mathcal{N}$ and $(\bar{x},\bar{z}) \in \mathcal{E}$. Thus, no point $(\bar{x},\bar{z}) \in \mathcal{E}$ with $\bar{x} \in \operatorname{ext}(\operatorname{conv}(\mathcal{N}))$ violates the inequality. Since we can represent any point $(\bar{x},\bar{z}) \in \mathcal{E}$ as a convex combination of the extreme points of $\operatorname{conv}(\mathcal{N})$, the proposition holds by iterating the previous reasoning for each extreme point in the support of $(\bar{x},\bar{z})$.
\iftoggle{ARXIV}{
\end{proof}
}{
%
\Halmos
\endproof
}
A fundamental issue is whether the inequalities of \cref{pro:Inequalities} are sufficient to define the set $\mathcal{E}$. By modulating the concept of closure introduced by \citet{chvatal_edmonds_1973}, we prove this is indeed the case. We define the \emph{equilibrium closure} as the convex hull of the points in $\mathcal{K}$ satisfying the equilibrium inequalities of \cref{pro:Inequalities}.
\begin{theorem}
Consider an \emph{IPG}\xspace instance $G$ where $|\mathcal{N}| \neq 0$. Let the \emph{equilibrium closure} given by the equilibrium inequalities of \cref{pro:Inequalities} be
$$
P^{e}:= \operatorname{conv} \Big (\left\{ (x,z) \in \mathcal{K} : \begin{array}{l}
u^i(\tilde{x}^i, x ^{-i}) \le u^i(x^i,x^{-i}) \\
\forall \tilde{x}: \tilde{x}^i \in \mathcal{BR}(i,\tilde{x}^{-i}), \; i =1,\dots,n
\end{array}\right\} \Big),
$$
where the equilibrium inequalities consider only the best-responses $\tilde{x}^i$ for any player $i$. Then, \begin{enumerate*} \item $P^{e}$ is a rational polyhedron, \item there exists no point $(x,z) \in \int(P^{e})$ such that $x \in \mathbb{Z}^{nm}$, \item $P^{e} = \mathcal{E}$ \end{enumerate*}.
\label{thm:Closure}
\end{theorem}
\iftoggle{ARXIV}{
\begin{proof}
}{
\proof{Proof of \cref{thm:Closure}.}
}
\emph{Proof of (i.) } The set $\mathcal{K}$ is finite since any $\mathcal{X}^i$ is finite, the number of best-responses and, correspondingly, of equilibrium inequalities, is finite.
Both equilibrium inequalities and the inequalities defining $\mathcal{X}^i$ have integer coefficients. Therefore, $P^{e}$ is a rational polyhedron.
\emph{Proof of (ii.) } Assume there exists a point $(\bar x, \bar z) \in \int(P^e)$ such that $\bar x \in \mathbb{Z}^{nm}$. By definition of Nash equilibrium, $\bar x \in \mathcal{N}$ since $(\bar x, \bar z)$ satisfies all the equilibrium inequalities in $P^e$. However, since $(\bar x, \bar z) \in \int(P^e)$, then no equilibrium inequality can be tight, contradicting the fact $\bar x$ is a \emph{PNE}\xspace. Therefore, there cannot exist any $(\bar x, \bar z) \in \int(P^e)$ such that $\bar x \in \mathbb{Z}^{nm}$. This also implies that all \emph{PNE}s\xspace lie on the boundary of $P^e$.
\emph{Proof of (iii.) }
Since $P^e$ contains all the equilibrium inequalities generated by the players' best-responses, then any $(\bar x, \bar z) \in \mathcal{E}$ belongs to $P^e$ as of \cref{pro:Inequalities}, and $\mathcal{E} \subseteq P^e$.
Let $(\hat x, \hat z)$ be a point in $\operatorname{ext}(P^e)$. By definition, $(\hat x, \hat z)$ is an integer point, and it corresponds to a \emph{PNE}\xspace. Indeed, non-equilibria integer points cannot belong to $P^e$ since they would violate at least one equilibrium inequality associated with the players' best-responses.
Equivalently, for any $(\hat x, \hat z) \in \operatorname{ext}(P^e)$, its projection $\operatorname{proj}_x = \hat x$ is in $\mathcal{N}$.
Since all \emph{PNE}s\xspace are on the boundary of $P^e$, $P^{e} = \mathcal{E}$ necessarily.
\iftoggle{ARXIV}{
\end{proof}
}{
%
\Halmos
\endproof
}
Throughout the proof of \cref{thm:Closure}, we show that $P^e$ yields indeed the perfect equilibrium formulation $\mathcal{E}$. Although the description of $P^e$ may contain an exponential number of possibly redundant equilibrium inequalities, it precisely describes the set of \emph{PNE}s\xspace in the lifted space.
In \cref{ex:epsilon}, we showcase the construction $P^e$ via \cref{thm:Closure} for a small \emph{IPG}\xspace.
\begin{example}
\label{ex:epsilon}
Consider an \emph{IPG}\xspace where player 1 solves $\max_{x^1}\{ x^1_1 + 3x^1_2 + 7x^1_3 -6x^1_1x^2_1 +3x^1_2x^2_2 +2x^1_3x^2_3: 6x^1_1 + 4x^1_2 + 5x^1_3\le 7, x^1\in\{0,1\}^3 \}$, and player 2 solves $\max_{x^2}\{ 9x^2_1 + 9x^2_2 + 2x^2_3 -6x^2_1x^1_1 +5x^2_2x^1_2 +7x^2_3x^1_3 : 4x^2_1 + 2x^2_2 + 5x^2_3 \le 5, x^2\in\{0,1\}^3 \}$.
There are $4$ feasible strategies for each player $i$, namely, $(x^i_1,x^i_2,x^i_3)=(0,0,0)\lor(0,0,1)\lor(0,1,0)\lor(1,0,0)$. The $3$ \emph{PNE}s\xspace of this game are:
\begin{enumerate*}
\item $\bar{x}^1=(0,0,1)$ and $\bar{x}^2=(0,0,1)$ with $u^1(\bar{x}^1,\bar{x}^2)=9$ and $u^2(\bar{x}^2,\bar{x}^1)=9$, %
\item $\bar{x}^1=(0,0,1)$ and $\bar{x}^2=(0,1,0)$ with $u^1(\bar{x}^1,\bar{x}^2)=7$ and $u^2(\bar{x}^2,\bar{x}^1)=9$, %
\item $\bar{x}^1=(0,0,1)$ and $\bar{x}^2=(1,0,0)$ with $u^1(\bar{x}^1,\bar{x}^2)=7$ and $u^2(\bar{x}^2,\bar{x}^1)=9$.
\end{enumerate*}
We linearize the game by introducing $3$ variables $z_j \in \{0,1\}$ for any player's variable $j \in \{1,2,3\}$ such that $z_j=1$ if and only if $x^1_j=x^2_j=1$. We model these implications through the constraints $z_j \le x^i_j$ and $z_j \ge x^1_j + x^2_j -1$ for any player $i$ and variable $j$. Hence,
\begin{align}
\mathcal{K}= \left\{
x^1 \in \{0,1\}^3, x^2\in \{0,1\}^3, z\in \{0,1\}^3 :
\begin{array}{l}
6x^1_1 + 4x^1_2 + 5x^1_3\le 7, \;
4x^2_1 + 2x^2_2 + 5x^2_3 \le 5 \\
z_j \le x^1_j, \; z_j \le x^2_j, \; z_j \ge x^1_j + x^2_j -1 \; \; \forall j \in \{1,2,3\}
\end{array}
\right\}. \nonumber
\end{align}
Correspondingly, the two players' utility functions in the linearized space are given by the two linear expressions $u^1(x^1,x^2)=x^1_1 + 3x^1_2 + 7x^1_3 -6z_1 +3z_2 +2z_3$ and $u^2(x^2,x^1)=9x^2_1 + 9x^2_2 + 2x^2_3 -6z_1 +5z_2 +7z_3$, respectively.
On the one hand, the best-response of player $1$ to any of player $2$'s feasible strategies is $\tilde{x}^1 = (0,0,1)$, i.e., $\mathcal{BR}(1,\tilde{x}^2) = \{ (0,0,1)\}$ for any feasible strategy $\tilde{x}^2$. The equilibrium inequality associated with $\tilde{x}^1=(0,0,1)$ is $7 + 2x^2_3 \le x^1_1 + 3x^1_2 + 7x^1_3 -6z_1 +3z_2 +2z_3$. The left-hand side of the inequality represents $u^1(\tilde{x}^1,x^{2})$, namely player 1's utility function evaluated on $\tilde{x}^1$.
On the other hand, player $2$'s best-responses and the associated equilibrium inequalities are: \begin{enumerate*}
\item $\tilde{x}^2=(1,0,0)$ with the inequality $9-6x^1_1 \le 9x^2_1 + 9x^2_2 + 2x^2_3 -6z_1 +5z_2 +7z_3$,
\item $\tilde{x}^2=(0,1,0)$ with the inequality $9+5x^1_2 \le9x^2_1 + 9x^2_2 + 2x^2_3 -6z_1 +5z_2 +7z_3$,
\item $\tilde{x}^2=(0,0,1)$ with the inequality $2+7x^1_3 \le 9x^2_1 + 9x^2_2 + 2x^2_3 -6z_1 +5z_2 +7z_3$.
\end{enumerate*}
Therefore,
\begin{align}
P^{e}= \operatorname{conv} \Big(\left\{ (x,z) \in \mathcal{K} \ : \begin{array}{l}
7 + 2x^2_3 \le x^1_1 + 3x^1_2 + 7x^1_3 -6z_1 +3z_2 +2z_3\\
9-6x^1_1 \le 9x^2_1 + 9x^2_2 + 2x^2_3 -6z_1 +5z_2 +7z_3 \\
9+5x^1_2 \le9x^2_1 + 9x^2_2 + 2x^2_3 -6z_1 +5z_2 +7z_3 \\
2+7x^1_3 \le 9x^2_1 + 9x^2_2 + 2x^2_3 -6z_1 +5z_2 +7z_3
\end{array}\right\} \Big) \nonumber.
\end{align}
By explicitly computing the above convex hull, we obtain
\begin{align}
P^e= \left\{ (x,z) \ :\begin{array}{l}
x^2_1 \ge 0,\; x^2_2 \ge 0,\; x^2_3\ge0,\; x^1_1 = 0,\; x^1_2 = 0,\; x^1_3 = 1, \\ x^2_1 + x^2_2 + x^2_3 = 1,\;
z_1 = 0,\; z_2 = 0,\; x^2_1 + x^2_2 + z_3 = 1
\end{array}\right\} \nonumber.
\end{align}
The projections onto the $x$ space of the extreme points of $P^e$ correspond to the $3$ \emph{PNE}s\xspace, and thus $P^e = \mathcal{E}$.
\end{example}
\section{The Cutting Plane Algorithm and its Oracle}
\label{sec:CompPNE}
If an oracle gives us $\mathcal{E}$ in the form of a set of linear inequalities, then an optimal solution to $\max_{x^1,\dots,x^n,z}\{ f(x,z) : (x,z) \in \mathcal{E} \}$ (i.e., a linear program) that is also an extreme point of $\mathcal{E}$ is a \emph{PNE}\xspace for $G$ for any function $f(x,z)$. However, there are two major issues. First, $\mathcal{E} \subseteq \operatorname{conv}(\mathcal{K})$, and $\operatorname{conv}(\mathcal{K})$ is a perfect formulation described by a possibly large number of inequalities. Second, retrieving $\mathcal{E}$ through \cref{thm:Closure} may still require a large number of inequalities.
In practice, we actually do not need $\mathcal{E}$ nor $\operatorname{conv}(\mathcal{K})$: a more reasonable goal is to get a polyhedron containing $\operatorname{conv}(\mathcal{K})$ over which we can optimize $f(x,z)$ efficiently and obtain an integer solution (i.e., $x \in \mathcal{K}$) that is also a \emph{PNE}\xspace. The first step is to obtain an integer solution. We could deploy branching schemes and known families of integer programming cutting planes, which are also equilibrium inequalities since they are valid for $\mathcal{E}$. Equivalently, we can exploit a Mixed-Integer Programming (\emph{MIP}\xspace) solver to solve $\max_{x^1,\dots,x^n,z}\{ f(x,z) : (x,z) \in \mathcal{K} \}$. If the maximizer is a \emph{PNE}\xspace, the algorithm terminates. Otherwise, the second step is to cut off such maximizer, since it is not a \emph{PNE}\xspace, by separating at least an equilibrium inequality of \cref{pro:Inequalities}.
\paragraph{Equilibrium Separation Oracle. } Given a point $(\tilde x,\tilde z)$, for instance, the point returned by a \emph{MIP}\xspace solver, the central question is to decide if $\tilde x \in \mathcal{N}$, and, if not, to derive an equilibrium inequality to cut off $(\tilde x,\tilde z)$. If we use the equilibrium inequalities of \cref{pro:Inequalities}, the process terminates in a finite number of iterations, since \cref{thm:Closure}. In the spirit of \citet{grotschel_ellipsoid_1981,karp_linear_1982}, we define a separation oracle for the equilibrium inequalities and $\mathcal{E}$. The \emph{equilibrium separation oracle} solves the \emph{equilibrium separation problem} of \cref{def:SeparationProblem}.
\begin{definition}[Equilibrium Separation Problem]
Consider an \emph{IPG}\xspace instance $G$. Given a point $(\bar{x}, \bar{z})$, the \emph{equilibrium separation problem} is the task of determining that either: \begin{enumerate*} \item $(\bar{x}, \bar{z}) \in \mathcal{E}$, or \item $(\bar{x}, \bar{z}) \notin \mathcal{E}$ and return an equilibrium inequality violated by $(\bar{x}, \bar{z})$\end{enumerate*}.
\label{def:SeparationProblem}
\end{definition}
\noindent \cref{Alg:ESO} presents our separation oracle for the inequalities of \cref{pro:Inequalities}. Given $(\bar{x}, \bar{z})$ and an empty set of linear inequalities $\phi$, the algorithm outputs either (i.) \texttt{yes} if $(\bar{x}, \bar{z}) \in \mathcal{E}$, or (ii.) \texttt{no} and a set of violated equilibrium inequalities $\phi$ if $(\bar{x}, \bar{z}) \notin \mathcal{E}$. The algorithm separates at most one inequality for any player $i$. By definition, $\bar{x}^i$ should be a best-response to be in a \emph{PNE}\xspace. Therefore, for any player $i$, the algorithm solves $\max_{x^i} \{ u^i(x^i,\bar{x}^{-i}) : A^i x^i \le b^i, x^i \in \mathbb{Z}^m \}$, where $\hat{x}^i$ is one of its maximizers. If $u^i(\bar{x}^i,\bar{x}^{-i}) = u^i(\hat{x}^i,\bar{x}^{-i})$, $\bar{x}^i$ is also a best-response.
However, if $u^i(\hat{x}^i,\bar{x}^{-i}) > u^i(\bar{x}^i,\bar{x}^{-i})$, the algorithm adds to $\phi$ an equilibrium inequality $u^i(\hat{x}^i,x^{-i})$ $\le u^i(x^i,x^{-i})$ violated by $(\bar{x}, \bar{z})$. After considering all players, if $|\phi|=0$, then $\bar{x}$ is a \emph{PNE}\xspace and the answer is \texttt{yes}. Otherwise, the algorithm returns \texttt{no} and $\phi \neq \emptyset$, i.e., at least an equilibrium inequality cutting off $(\bar{x}, \bar{z})$.
\begin{algorithm}[!ht]
\caption{Equilibrium Separation Oracle \label{Alg:ESO}}
\KwData{An \emph{IPG}\xspace instance $G$, a point $(\bar x, \bar z)$, and a set of cuts $\phi=\emptyset$.}
\KwResult{ Either: (i.)
\texttt{yes} if $(\bar{x}, \bar{z}) \in \mathcal{E}$, or (ii.) \texttt{no} and $\phi$.}
\For{$i\gets1$ \KwTo $n$}{
$\hat{x}^i \gets \max_{x^i} \{ u^i(x^i,\bar{x}^{-i}) : A^i x^i \le b^i, x^i \in \mathbb{Z}^m \}$ \;
\If{$ u^i(\hat{x}^i,\bar{x}^{-i}) > u^i(\bar{x}^i,\bar{x}^{-i})$}
{
Add $u^i(\hat{x}^i,x^{-i})\le u^i(x^i,x^{-i})$ to $\phi$\;
}
}
\lIf{$|\phi|=0$} { \KwRet{\texttt{yes}} }
\lElse { \KwRet{\texttt{no} and $\phi$} }
\end{algorithm}
\paragraph{ZERO Regrets. } We present our cutting plane algorithm \emph{ZERO Regrets} in \cref{Alg:ZERO}.
The inputs are an instance $G$, and a function $f(x)$, while the output is either the \emph{PNE}\xspace $\ddot x$ maximizing $f(x)$, or a certificate that no \emph{PNE}\xspace exists. Let $\Phi$ be a set of equilibrium inequalities, and $\mathcal{Q}=\max_{x^1,\dots,x^n,z} \{ f(x,z) : (x,z) \in \mathcal{K}, (x,z) \; s.t. \; \Phi \}$. We assume $\mathcal{Q}$ is feasible and bounded. Otherwise, there would be no rationale behind getting a \emph{PNE}\xspace with an arbitrarily bad welfare. At each iteration, we compute an optimal solution $(\bar x, \bar z)$ of $\mathcal{Q}$. Afterwards, the equilibrium separation oracle (\cref{Alg:ESO}) evaluates $(\bar x, \bar z)$. If the oracle returns \texttt{yes}, then $\ddot{x}=\bar x$ is the \emph{PNE}\xspace maximizing $f(x)$ in $G$. Otherwise, the oracle returns a set $\phi$ of equilibrium inequalities cutting off $(\bar x, \bar z)$, and the algorithm adds $\phi$ to $\Phi$. Therefore, the process restarts by solving $\mathcal{Q}$ with the additional set of constraints. If at any iteration $\mathcal{Q}$ becomes infeasible, then $G$ has no \emph{PNE}\xspace. We remark that \cref{thm:Closure} implies both correctness and finite termination of \cref{Alg:ZERO}.
\begin{algorithm}[!ht]
\caption{ZERO Regrets \label{Alg:ZERO}}
\KwData{An \emph{IPG}\xspace instance $G$, and a function $f(x)$.}
\KwResult{ Either: (i.) the \emph{PNE}\xspace $\ddot{x}$ maximizing $f(x)$, or (ii.) \texttt{no} \emph{PNE}\xspace }
$\Phi =\{ 0 \le 1\}$, and $\mathcal{Q}=\max_{x^1,\dots,x^n,z} \{ f(x,z) : (x,z) \in \mathcal{K}, (x,z) \; s.t. \; \Phi \}$\;
\While{true}{
\lIf{$\mathcal{Q}$ is infeasible}{\KwRet{\texttt{no} \emph{PNE}\xspace} }
$(\bar x, \bar z)= \arg\max \mathcal{Q}$; $\phi=\emptyset$ \;
\If{\texttt{EquilibriumSeparationOracle$(G,(\bar x, \bar z), \phi)$} is \texttt{yes}}
{
\KwRet{$\ddot{x}=\bar x;$} \label{Alg:ZERO:ReturnPNE}
}
\lElse
{
add $\phi$ to $\Phi$
}
}
\end{algorithm}
Although it is sufficient to add just one equilibrium inequality in $\phi$ cutting off the incumbent solution $(\bar x, \bar z)$, we expect that a good trade-off between $|\phi|=1$ and $|\phi|=n$ may speed up the convergence of the algorithm. This includes, for instance, also adding non-violated equilibrium inequalities. In \cref{ex:ZERORegrets}, we provide a toy example of \emph{ZERO Regrets}.
\begin{example}
Consider the game in \emph{Example 1} where player 1 solves $\max_{x^1}\{ 6x^1_1 + x^1_2 - 4x^1_1x^2_1 +6x^1_2x^2_2 : 3x^1_1 + 2x^1_2 \le 4, x^1\in\{0,1\}^2 \}$, and player 2 solves $\max_{x^2}\{ 4x^2_1 + 2x^2_2 -x^2_1x^1_1 -x^2_2x^1_2 : 3x^2_1 + 2x^2_2 \le 4, x^2\in\{0,1\}^2 \}$.
As in \emph{Example 2}, to linearize the players' utility functions, we introduce two binary variables $z_1$ and $z_{2}$ equal to $1$ if both players select items $1$ and $2$, respectively. The linearization constraints are $z_1 \le x^1_1$, $z_1 \le x^2_1$, $z_1 \ge x^1_1 + x^2_1 - 1$, $z_2 \le x^1_2$, $z_2 \le x^2_2$, $z_2 \ge x^1_2 + x^2_2 - 1$. Thus, player 1's utility function is $6x^1_1 + x^1_2 - 4z_1 +6z_2$ and player 2's utility function is $4x^2_1 + 2x^2_2 - z_1 - z_2$.
Correspondingly, problem $\mathcal{Q}$ maximizing the social welfare is
\begin{align}
\max_{(x^1, \;x^2, \;z)} \qquad & 6x^1_1 + x^1_2 + 4x^2_1 + 2x^2_2 -5z_1 + 5z_2 \nonumber \\
s.t. \qquad & 3x^1_1 + 2x^1_2 \le 4,\, 3x^2_1 + 2x^2_2 \le 4 \nonumber \\
& z_j \le x^1_j,\, z_j \le x^2_j,\, z_j \ge x^1_j + x^2_j - 1 \quad j=1,2. \nonumber \\
& x^1_j,\, x^2_j,\, z_j \in\{0,1\} \quad j=1,2. \nonumber
\end{align}
An optimal solution of the problem is $(\bar{x}^1_1,\bar{x}^1_2)=(1,0)$, $(\bar{x}^2_1,\bar{x}^2_2)=(0,1)$, $\bar{z}_1 = \bar{z}_2 = 0$. The social welfare is $8$, and the players' utility values are $6$ and $2$, respectively. However, this solution is not a \emph{PNE}\xspace. In fact, while a best-response to $\bar{x}^2$ for player 1 is $\bar{x}^1$, the best-response to $\bar{x}^1$ for player 2 is $(\hat{x}^2_1,\hat{x}^2_2)=(1,0)$ with an utility value of $3$.
Therefore, from player 2, we derive the equilibrium inequality
$4 - x^{1}_1 \le 4x^2_1 + 2x^2_2 -z_1 - z_2$
cutting off $(\bar{x}, \bar{z})$.
By adding the equilibrium inequality to $\mathcal{Q}$, the optimal solution is then $(\bar{x}^1_1,\bar{x}^1_2)=(1,0)$, $(\bar{x}^2_1,\bar{x}^2_2)=(1,0)$, $\bar{z}_1 = 1$, $\bar{z}_2 = 0$, with utility values $2$ and $3$ and a welfare of $5$. Since $\bar{x}$ is a \emph{PNE}\xspace, the algorithm terminates by finding a \emph{PNE}\xspace with a \emph{PoS}\xspace of $8/5$.
\label{ex:ZERORegrets}
\end{example}
\paragraph{Game-theoretical Interpretation.} We provide a straightforward game-theoretical interpretation of \emph{ZERO Regrets}. The algorithm acts as a central authority (e.g., a central planner) when optimizing $f(x,z)$ over $\mathcal{K}$, producing a solution that optimizes the welfare.
Afterward, it proposes the solution to each player, who evaluates it through the equilibrium separation oracle. The latter acts as a \emph{rationality} blackbox, in the sense that the oracle advises each player $i$ whether the proposed strategy is acceptable or not. In other words, the rationality blackbox tells the player $i$ if it should selfishly (and rationally) deviate to a better strategy, ignoring the central authority's advice. On the one hand, if the rationality blackbox says the solution is acceptable for player $i$, then the player knows through the oracle that it should accept the proposed strategy. On the other hand, if at least one player $i$ refuses the proposed solution, the central authority should exclude such a solution and formulate a new proposal. Namely, it should cut off the non-equilibrium strategy and compute a new solution maximizing the welfare.
\subsection{Extensions}
\label{sec:extensions}
We showcase the flexibility of our algorithmic framework by proposing two extensions to \emph{ZERO Regrets}. Specifically, to address broader practical needs, we propose two extensions for enumerating \emph{PNE}s\xspace and computing approximate \emph{PNE}s\xspace.
\paragraph{Enumerating \emph{PNE}s\xspace. } We can easily tune \emph{ZERO Regrets} to enumerate all the \emph{PNE}s\xspace in $\mathcal{N}$ as follows. In \cref{Alg:ZERO:ReturnPNE} of \cref{Alg:ZERO}, instead of terminating and returning $\ddot x$, we memorize $\ddot x$ and add an (invalid) inequality cutting off the \emph{PNE}\xspace from $\mathcal{E}$. Since all $x$ variables are integer-constrained, such inequality can be, for instance, an hamming distance from $\bar x$. The algorithm will possibly compute a new \emph{PNE}\xspace, cut it off (e.g., through a hamming distance constraint), and move the search towards the next equilibrium. Eventually, $\mathcal{Q}$ will become infeasible, thus certifying that the algorithm enumerated all the existing \emph{PNE}s\xspace.
\paragraph{Approximating \emph{PNE}s\xspace. } An absolute $\epsilon$-\emph{PNE}\xspace is a \emph{PNE}\xspace where each player can deviate at most by a value $\epsilon$ for any best-response \citep{nisan_algorithmic_2008}, namely, where the regret for each player is at most $\epsilon$. Absolute $\epsilon$-\emph{PNE}s\xspace may be a reasonable compromise whenever no \emph{PNE}\xspace exists. Although any \emph{PNE}\xspace is an absolute $\epsilon$-\emph{PNE}\xspace with $\epsilon=0$, one may be interested in computing an absolute $\epsilon$-\emph{PNE}\xspace with an upper bound on $\epsilon$ while maximizing $f(x,z)$. We can adapt our algorithmic framework to compute an absolute $\epsilon$-\emph{PNE}\xspace as follows. We introduce a bounded continuous variable $\epsilon$ in $\mathcal{Q}$, and we let \cref{Alg:ESO} separate equilibrium inequalities in the form of $u^i(\hat{x}^i,x^{-i}) - \epsilon \le u^i(x^i,x^{-i})$. Depending on the application of interest, one may still optimize the function $f(x,z)$ or minimize $\epsilon$ without affecting the correctness of the algorithm. A similar modification enables the algorithm to handle relative $\epsilon$-\emph{PNE}\xspace, i.e., a profile of strategies where the payoff of each player's strategy is at least $\epsilon$ times the best-response payoff. Given a constant $\epsilon$, the corresponding equilibrium inequalities are $u^i(\hat{x}^i,x^{-i}) \epsilon \le u^i(x^i,x^{-i})$.
\section{Applications}
\label{sec:games}
We evaluate \emph{ZERO Regrets} on a wide range of problems from relevant works in the literature. We aim to provide a solid benchmark against the existing solution approaches and show the effectiveness of \emph{ZERO Regrets} in selecting and enumerating equilibria. The games we select stem from practical applications (e.g., competitive facility locations, network design) and methodological studies with the associated benchmark instances (e.g., games among quadratic programs).
Specifically, we consider the following games:
\begin{enumerate}
\item The Knapsack Game (\emph{KPG}\xspace) \citep{Dragotto_2021_CNP,carvalho_computing_2017,carvalho_computation_2016}, where each player solves a binary knapsack problem. We select the equilibrium maximizing the social welfare, and we provide theoretical results on the complexity of deciding whether a \emph{PNE}\xspace exists. We also introduce two problem-specific equilibrium inequalities.
\item The Network Formation Game (\emph{NFG}\xspace) \citep{chen_network_2006,anshelevich_price_2008} -- a paradigmatic game in \emph{AGT}\xspace with plenty of applications in network design -- where players seek to build a network through a cost-sharing mechanism. We select the equilibrium maximizing the social welfare.
\item The Competitive Facility Location and Design game (\emph{CFLD}\xspace) \citep{cronert_equilibrium_2020,aboolian_competitive_2007}, where each player decides both the location of its facilities and their “design” (i.e., the facilities' features) while competing for customer demand. As in the \emph{KPG}\xspace and the \emph{NFG}\xspace, we focus on finding the \emph{PNE}\xspace maximizing the social welfare.
\item The Quadratic \emph{IPG}s\xspace (\emph{qIPG}s\xspace) introduced by \citet{sagratella_computing_2016} and recently considered in \citet{schwarze_branch-and-prune_2022}, where each player optimizes a (non-convex) quadratic function over box constraints and integrality requirements. As in the original papers, we focus on enumerating all the existing \emph{PNE}s\xspace, or determine that none exists.
\end{enumerate}
In what follows, we briefly describe the previous games and present the associated computational results\footnote{We performed our tests on an \emph{Intel Xeon Gold 6142} equipped with $128$GB of RAM and $8$ threads, employing $Gurobi$ 9.5 as \emph{MIP}\xspace solver for \cref{Alg:ZERO}.}.
\subsection{The Knapsack Game}
The \emph{KPG}\xspace is an \emph{IPG}\xspace among $n$ players, where each player $i$ solves a binary knapsack problem with $m$ items in the form of
\begin{align}
\max_{x^i} \Big\{ \sum_{j=1}^{m}p^{i}_jx^i_j + \sum_{k=1, k\neq i}^{n}\sum_{j=1}^{m} C^i_{k,j} x^i_j x^k_j : \sum_{j=1}^{m} w^i_j x^{i}_j \leq b^i, x^i \in \{0,1\}^m \Big\}.
\label{eq:KPG}
\end{align}
As in the classical knapsack problem \citep{kellerer_knapsack_2004}, we assume that the profits $p^{i}_j$, weights $w^i_j$ and capacities $b^i$ are in $\mathbb{Z}^+_0$. The selection of an item $j$ by a player $k \neq i$ impacts either negatively or positively the profit $p^{i}_j$ for player $i$ through integer coefficients $C^i_{k,j}$. Clearly, given the strategies of the other players $x^{-i}$, computing a corresponding best-response for player $i$ is $\mathcal{NP}$-hard\xspace. We can apply our algorithmic framework by linearizing the bilinear products $x^i_j x^k_j$ (for any $i,k,j$) with $\mathcal{O}(m n^2)$ auxiliary variables and additional constraints (see \cref{ex:epsilon}).
\citet{carvalho_computation_2016} introduced the game with $n=2$ and $p^i_j = 0$ $\forall j=1, \dots, m,\, i=1,2$. \citet{Dragotto_2021_CNP,carvalho_computing_2017} consider a more general game variant allowing $p^i_j$ and $w^i_j$ to take negative integer values. However, their algorithms focus on Mixed-Strategy equilibria and cannot perform exact equilibria selection.
In \cref{thm:IPGSigma2p}, we formally prove that deciding if a \emph{KPG}\xspace instance has a \emph{PNE}\xspace\xspace -- even with two players -- is $\Sigma^p_2$-complete\xspace in the polynomial hierarchy, matching the result of \citet{carvalho_computing_2017} for general \emph{IPG}s\xspace.
\begin{theorem}
Deciding if a \emph{KPG}\xspace instance has a \emph{PNE}\xspace is a $\Sigma^p_2$-complete\xspace problem.
\label{thm:IPGSigma2p}
\end{theorem}
\noindent The proof, where we perform a reduction from the $\Sigma^p_2$-complete\xspace \emph{DeNegre Bilevel Knapsack Problem} \citep{caprara_study_2014,denegre_thesis_2011, dellacroce_exact_2020}, is in the appendix.
Furthermore, we show that when at least one \emph{PNE}\xspace exists, the \emph{PoS}\xspace and \emph{PoA}\xspace can be arbitrarily bad.
\begin{proposition}
The \emph{PoA}\xspace and the \emph{PoS}\xspace in \emph{KPG}\xspace can be arbitrarily bad.
\label{prop:PoaPos}
\end{proposition}
\iftoggle{ARXIV}{
\begin{proof}
}{
%
\proof{Proof of \cref{prop:PoaPos}.}
}
Consider the following \emph{KPG}\xspace instance with $n=2$: player 1 solves the problem $\max_{x^1}\{ Mx^1_1 + x^1_2 - (M-2)x^1_1x^2_1 -x^1_2x^2_2 : 3x^1_1 + 2x^1_2 \leq 4, x^1\in\{0,1\}^2 \}$ where $M$ is an arbitrarily large value; player 2 solves $\max_{x^2}\{ 4x^2_1 + x^2_2 -x^2_1x^1_1 -x^2_2x^1_2 : 3x^2_1 + 2x^2_2 \leq 4, x^2\in\{0,1\}^2 \}$. The only \emph{PNE}\xspace is $(\bar{x}^1_1,\bar{x}^1_2,\bar{x}^2_1,\bar{x}^2_2)=(1,0,1,0)$, with $u^1(\bar{x}^1,\bar{x}^{2})=2$, $u^2(\bar{x}^2,\bar{x}^{1})=3$, $S(\bar{x})=5$. The maximum welfare $OSW=M+1$ is given by $(\hat{x}^1_1,\hat{x}^1_2,\hat{x}^2_1,\hat{x}^2_2)=(1,0,0,1)$, i.e., $OSW$ is arbitrarily large and there are no bounds on both the \emph{PoA}\xspace and the \emph{PoS}\xspace.
\iftoggle{ARXIV}{
\end{proof}
}{
%
\Halmos
\endproof
}
\subsubsection{Strategic Inequalities. }
We further strengthen our cutting plane algorithm by introducing two classes of problem-specific equilibrium inequalities for the \emph{KPG}\xspace.
\paragraph{Strategic Dominance Inequalities.} In the binary knapsack problem, a well-known hierarchy of dominance relationships exists among items, as we formalize in \cref{def:KPDominance}.
\begin{definition}[Dominance Rule]
Given two items $j$ and $j^\prime$ with profits $\bar{p}_j$ $\bar{p}_{j^\prime}$ and weights $w_j$, $w_{j^\prime}$, if $w_j \le w_{j^\prime}$ and $\bar{p}_j > \bar{p}_{j^\prime}$, then we say item $j$ \emph{dominates} item $j^\prime$.
\label{def:KPDominance}
\end{definition}
The above concept of dominance implies that, in any optimal knapsack solution, if one packs a dominated item $j^\prime$, then it should also pack item $j$. Otherwise, one would be able to improve the solution by selecting $j$ instead of $j^\prime$. This reasoning translates to the inequality $x_{j^\prime} \le x_j$, which is always valid for any optimal knapsack solution. We aim to extend this concept of dominance to the \emph{KPG}\xspace by incorporating the strategic interactions among players. In order to derive such inequalities, we reason about how, for any player $i$, the decisions of $x^{-i}$ affect the profits of $i$'s items. More formally, for any player $i$ and item $j$, let $p^{min}_j$, and $p^{max}_j$ be the minimum and maximum profit the strategies of the other players can induce, respectively. We claim the dominance rule of \cref{def:KPDominance} extends to the one of \cref{prop:StrictDom} in the \emph{KPG}\xspace.
\begin{proposition}
For each player $i$, if the dominance rule applies for two items $j$ and $j^\prime$ with $\bar{p}_j = p^{min}_j$ and $\bar{p}_{j^\prime} = p^{max}_{j^\prime}$, then the inequality $x^i_{j^\prime} \le x^i_{j}$ is an \emph{equilibrium inequality}.
\label{prop:StrictDom}
\end{proposition}
\iftoggle{ARXIV}{
\begin{proof}
}{
%
\proof{Proof of \cref{prop:StrictDom}.}
}
Since all best-responses of player $i$ cannot select the dominated item $j^\prime$ without selecting item $j$ for any $x^{-i}$, the claim holds.
\iftoggle{ARXIV}{
\end{proof}
}{
%
\Halmos
\endproof
}
We denote the inequalities of \cref{prop:StrictDom} as \emph{Strategic Dominance Inequalities}. We further extend the previous reasoning to derive other forms of dominance inequalities by evaluating how the strategic interaction (i.e., the items that the other players select) affects the items' profits for each player $i$. In other words, we derive equilibrium inequalities that incorporate the strategic interaction by including the variables of multiple players. For instance, consider the case with two players. If the profits of two items $j$ and $j^\prime$ for player $1$ fulfill the dominance rule when player $2$ selects item $j$ and does not select item $j^\prime$, then
$$ x^1_{j^\prime} \le x^1_{j} + (1 - x^2_j) + x^2_{j^\prime}$$
is an equilibrium inequality. Namely, if there exists a \emph{PNE}\xspace with $x^2_j = 1$ and $x^2_{j^\prime} = 0$, the dominance rule between item $j$ and $j^\prime$ applies for player $1$, otherwise the inequality is not binding.
\paragraph{Strategic Payoff Inequalities. } We introduce a second class of \emph{strategic} inequalities by exploiting two observations on the knapsack problem. For any player $i$, the strategy of all-zeros $x^i=(0,\dots,0)$ is always feasible under the packing constraint, and yields a payoff of $0$. Therefore, for any player $i$ and item $j$, if $p^i_j + \sum_{k=1, k \neq i}^n C^i_{k,j} <0$, player $i$ may not select $j$ depending on its opponent choices $x^{-i}$. More generally, let $\mathcal{S}^{i}_j$ be the \emph{interaction set} of $i$'s opponents inducing a negative profit for item $j$, namely, a set so that
\begin{align}
p^i_j + \sum_{k \in \mathcal{S}^{i}_j} C^i_{k,j} <0.
\label{eq:StrategicPayoff1}
\end{align}
The interaction is \emph{minimal} if, for any proper subset $\bar{\mathcal{S}}^{i}_j$ of $\mathcal{S}^{i}_j$, then $p^i_j + \sum_{k \in \bar{\mathcal{S}}^{i}_j} C^i_{k,j} >0$. The inequality (\ref{eq:StrategicPayoff1}) implies that if $x^k_j=1$ for any $k \in \mathcal{S}^{i}_j$, then $x^i_j=0$. In general, this implies that for any interaction set, the inequality
$$
x^i_j + \sum_{k \in \mathcal{S}^{i}_j} x^k_j \le |\mathcal{S}^{i}_j|
$$
is an equilibrium inequality. We define the latter inequality as \emph{Strategic Payoff Inequality}. In practice, the inequalities generated by minimal interaction sets are stronger than those generated by non-minimal interaction sets since they generally involve more variables.
Clearly, the effort to separate and include all the previous strategic inequalities may not be negligible when $n$ and $m$ increase. In practice, at each iteration of \cref{Alg:ZERO}, we separate and add to $\mathcal{Q}$ only the inequalities violated by the incumbent solution $(\bar x, \bar z)$.
\paragraph{Computational Results. } We generate \emph{KPG}\xspace instances with $n= 2, 3$ and $m= 25, 50, 75,$ $100$, and with $p^{i}_j$ and $w^i_j$ being random integers uniformly distributed in $[1, 100]$ for any $i$. The knapsack capacities $b^i$ are equal to $0.2\sum_{j=1}^{m} w^i_j$, $0.5\sum_{j=1}^{m} w^i_j$, or $0.8\sum_{j=1}^{m} w^i_j$, respectively. For what concerns the strategic interaction, we focus on three different distributions for the integer interaction coefficients $C^i_{k,j}$. For any player $i$, they can be: \emph{A)} equal and uniformly distributed in $[1,100]$, or \emph{B)} random and uniformly distributed in $[1,100]$, or \emph{C)} random and uniformly distributed in $[-100,100]$. In \cref{tab:KPG}, we present the results for the $72$ resulting instances. For any given number of players $n$, items $m$ and distribution of coefficients $C^i_{k,j}$ ($(n, m, d$)), we report the performance over the $3$ instances with different capacities, in terms of:
\begin{enumerate*}
\item the average number of Equilibrium Inequalities of \cref{pro:Inequalities} added (\emph{\#EI}),
\item the average number of Strategic Payoff Inequalities (\emph{\#EI\_P}) (which we only compute for the instances with distribution \emph{C}),
\item the average number of Strategic Dominance Inequalities (\emph{\#EI\_D}),
\item the average computational time (\emph{Time}),
\item the average computational time (\emph{Time-\nth{1}}) to find the first \emph{PNE}\xspace (if any),
\item the average \emph{PoS}\xspace (\emph{PoS}\xspace) for the best \emph{PNE}\xspace (if any), and
\item the number of time limit hits (\emph{Tl})
\end{enumerate*}. The average values in \emph{\#EI}, \emph{\#EI\_P}, \emph{\#EI\_D}, \emph{Time} and \emph{Time-\nth{1}} also consider the instances where we hit the time limit\footnote{ We remark that the same observation holds on all our experiments.}, which we set to $1800$ seconds.
\emph{ZERO Regrets} solves almost all instances with $n=2$, regardless of the type of strategic interaction. Both running times and the number of equilibrium inequalities are significantly modest for a $\Sigma^p_2$-hard\xspace game. The \emph{PoS}\xspace is generally low and increases with distribution $C$ due to the nature of the complex strategic interactions stemming from both negative and positive $C^i_{k,j}$ coefficients. We remind that a \emph{PoS}\xspace close to $1$ does not mean the instance is computationally “easy”. On the contrary, a $\emph{PoS}\xspace \approx 1$ highlights the existence of a high-quality \emph{PNE}\xspace (i.e., with a welfare close to the one of the $OSW$) and also provides further evidence concerning the urgency of selecting efficient \emph{PNE}s\xspace.
\emph{ZERO Regrets} performs robustly even in large instances, establishing a significant computational advantage over the previously developed approaches in the literature. \citet{carvalho_computing_2017,Dragotto_2021_CNP} consider up to $m=40$ items with $n=3$ by just computing an equilibrium, while \citet{cronert_equilibrium_2020} only perform equilibria selection with $m<5$.
\begin{table}[!ht]
\setlength\tabcolsep{1em}
\centering
\resizebox{0.9\textwidth}{!}{%
\begin{tabular}{>{\bf}c@{\hspace{3em}}lllllll}
\toprule
\textbf{($\bm n$, $\bm m$, $\bm d$)} & \textbf{\#EI} & \textbf{\#EI\_P} & \textbf{\#EI\_D} & \textbf{Time} & \textbf{Time-\nth{1}} & \textbf{\emph{PoS}\xspace} & \textbf{Tl} \\
\midrule
(2, 25, A) & 14.67 & 0.00 & 3.00 & 0.06 & 0.05 & 1.07 & 0/3 \\
(2, 25, B) & 17.33 & 0.00 & 3.67 & 0.12 & 0.09 & 1.02 & 0/3 \\
(2, 25, C) & 29.33 & 9.67 & 7.67 & 0.39 & 0.04 & 1.06 & 0/3 \\
(2, 50, A) & 20.00 & 0.00 & 2.67 & 0.21 & 0.21 & 1.02 & 0/3 \\
(2, 50, B) & 26.67 & 0.00 & 19.67 & 0.51 & 0.39 & 1.01 & 0/3 \\
(2, 50, C) & 72.67 & 27.00 & 11.33 & 6.34 & 0.92 & 1.08 & 0/3 \\
(2, 75, A) & 38.00 & 0.00 & 31.00 & 0.60 & 0.44 & 1.00 & 0/3 \\
(2, 75, B) & 100.67 & 0.00 & 34.00 & 8.35 & 5.71 & 1.02 & 0/3 \\
(2, 75, C) & 112.67 & 38.33 & 67.00 & 47.75 & 3.96 & 1.08 & 0/3 \\
(2, 100, A) & 25.33 & 0.00 & 14.67 & 0.76 & 0.58 & 1.01 & 0/3 \\
(2, 100, B) & 205.33 & 0.00 & 79.67 & 220.42 & 143.45 & 1.01 & 0/3 \\
(2, 100, C) & 697.33 & 55.33 & 119.67 & 1205.29 & 11.33 & 1.05 & 2/3 \\
(3, 25, A) & 31.00 & 0.00 & 9.33 & 0.21 & 0.21 & 1.01 & 0/3 \\
(3, 25, B) & 44.00 & 0.00 & 14.67 & 0.33 & 0.33 & 1.02 & 0/3 \\
(3, 25, C) & 91.00 & 29.67 & 33.67 & 29.78 & 5.64 & 1.26 & 0/3 \\
(3, 50, A) & 95.00 & 0.00 & 24.33 & 18.39 & 11.68 & 1.03 & 0/3 \\
(3, 50, B) & 206.00 & 0.00 & 44.33 & 626.45 & 167.01 & 1.01 & 1/3 \\
(3, 50, C) & 148.00 & 63.00 & 224.67 & 382.24 & - & - & 0/3 \\
(3, 75, A) & 64.00 & 0.00 & 119.00 & 4.65 & 2.07 & 1.02 & 0/3 \\
(3, 75, B) & 278.00 & 0.00 & 92.67 & 982.97 & 272.69 & 1.01 & 1/3 \\
(3, 75, C) & 173.00 & 87.33 & 319.67 & 658.77 & - & - & 1/3 \\
(3, 100, A) & 261.00 & 0.00 & 144.67 & 1200.65 & 666.13 & 1.00 & 2/3 \\
(3, 100, B) & 479.00 & 0.00 & 168.33 & tl & - & - & 3/3 \\
(3, 100, C) & 184.00 & 171.00 & 1019.67 & 1200.31 & - & - & 2/3 \\
\bottomrule \vspace{-\aboverulesep}
\end{tabular} %
}
\caption{Results overview for the \emph{KPG}\xspace. The complete tables of results are in the Appendix (\cref{tab:KPG_Full2} and \cref{tab:KPG_Full3}).}
\label{tab:KPG}
\end{table}
\subsection{The Network Formation Game}
Network design games are paradigmatic problems in Algorithmic Game Theory \citep{chen_network_2006,anshelevich_price_2008,nisan_algorithmic_2008}.
Their natural application domain is often the one of computer networks and the Internet itself, where several selfish agents opportunistically decide how to share a scarce resource, for instance, the bandwidth. \citet{tardos_network_2004} accurately claims that the impact and future of the complex technology we develop through the Internet critically depend on the ability to balance the diverse needs of the selfish agents in the network. We consider a (weighted) \emph{NFG}\xspace -- similar to the one of \citet{chen_network_2006} -- where $n$ players are interested in building a computer network. Let $G(V, E)$ be a directed graph representing a network layout, where $V$, $E$ are the sets of vertices and edges, respectively. Each edge $(h,l) \in E$ has a construction cost $c_{hl} \in \mathbb{Z}^+$, and each player $i$ wants to connect an origin $s^i$ with a destination $t^i$ while minimizing its construction costs. A cost-sharing mechanism determines the cost of each edge $(h,l)$ for a player as a function of the number of players crossing $(h,l)$. Arguably, the most common and widely-adopted mechanism is the \emph{Shapley cost-sharing mechanism}, where players using $(h,l)$ equally share its cost $c_{hl}$. The goal is to find a \emph{PNE}\xspace minimizing the sum of construction costs for each player or determine that no \emph{PNE}\xspace exists.
We model the \emph{NFG}\xspace as an \emph{IPG}\xspace as follows. For any player $i$ and edge $(h,l)$, let the binary variables $x^i_{hl}$ be $1$ if $i$ uses the edge. We employ standard flow constraints to model a path between $s^i$ and $t^i$.
For conciseness, we represent these constraints and binary requirements with a set $\mathcal{F}^i$ for each $i$. Thus, each player $i$ solves:
\begin{align}
\min_{x^i} \{ \sum_{(h,l)\in E} \frac{c_{hl}x^i_{hl}}{\sum_{k=1}^{n} x^k_{hl}}: x^i \in \mathcal{F}^i \}.
\label{eq:NFG}
\end{align}
For any player $i$, the cost contribution of each edge $(h,l)$ to the objective is not linear in $x$ and may not be defined for some choices of $x$ (i.e., $\sum_{k=1}^{n} x^k_{hl}=0$). However, we can linearize the fractional terms and eliminate the indefiniteness. For instance, consider a game with $n=3$ and the objective of player $1$. Let the binary variable $z^{j, \dots, k}_{hl}$ be $1$ if only players $j, \dots, k$ select the edge $(h,l)$. Then, $x^1_{hl} = z^1_{hl} + z^{12}_{hl} + z^{13}_{hl} + z^{123}_{hl}$, $x^2_{hl} = z^2_{hl} + z^{12}_{hl} + z^{23}_{hl} + z^{123}_{hl}$, $x^3_{hl} = z^3_{hl} + z^{13}_{hl} + z^{23}_{hl} + z^{123}_{hl}$ along with a clique constraint $ z^1_{hl} + z^2_{hl} + z^3_{hl} + z^{12}_{hl} + z^{13}_{hl} + z^{23}_{hl} + z^{123}_{hl} \le 1$. The term for edge $(h,l)$ in the objective of player $1$ is then $c_{hl}z^1_{hl} + \frac{c_{hl}}{2}(z^{12}_{hl} + z^{13}_{hl}) + \frac{c_{hl}}{3}z^{123}_{hl}$. In our tests, we consider the general weighted \emph{NFG}\xspace \citep{chen_network_2006}, where each player $i$ has a weight $w^i$, and the cost share of each selected $(h,l)$ is $w^i c_{hl}$ divided by the weights of all players using $(h,l)$. Specifically, we consider the $3$-player weighted \emph{NFG}\xspace, where a \emph{PNE}\xspace may not exist, and selecting one if multiple equilibria exist is generally an $\mathcal{NP}$-hard\xspace problem \citep{anshelevich_price_2008,chen_network_2006}.
\paragraph{Computational Results. } In order to tackle challenging instances, we consider the \emph{NFG}\xspace with $n = 3$ on grid-based directed graphs $G(V, E)$, where each $i$ has to cross the grid from left to right to reach its destination. Compared to a standard grid graph, we randomly add some edges between adjacent layers to increase the number of paths, and to facilitate the interaction among players. The instances are so that $|V| \in [50,500]$, and the costs $c_{hl}$ for each edge $ (h,l)$ are random integers uniformly distributed in $[20, 100]$.
We consider three distributions of player's weights: \begin{enumerate*}
\item the Shapely-mechanism with $w^1=w^2=w^3=1$, where a \emph{PNE}\xspace always exists, yet selecting the most efficient \emph{PNE}\xspace is $\mathcal{NP}$-hard\xspace, or
\item $w^1=0.6$, $w^2=0.2$, and $w^3=0.2$, or
\item $w^1=0.45$, $w^2=0.45$, and $w^3=0.1$.
\end{enumerate*}
\cref{tab:NFG} reports the results, where we average over the distributions of the players' weights. For each graph, the table reports the graph size ($|V|,|E|$), whereas the other columns have the same meaning of the ones of \cref{tab:KPG}. Our algorithm effectively solves all the instances but $3$ within a time limit of $1800$ seconds and consistently selects high-efficiency \emph{PNE}s\xspace. Further, our algorithm finds the first \emph{PNE}\xspace in considerably limited computing times. Generally, the previous literature does not consider this problem from a computational perspective, but only provides theoretical and possibly pessimistic bounds on the \emph{PoS}\xspace and \emph{PoA}\xspace. Nevertheless, we can compute efficient \emph{PNE}s\xspace even in large-size graphs (i.e., $\emph{PoS}\xspace \approx 1$), with a limited number of equilibrium inequalities and modest running times, showing the practical effectiveness of our algorithm in a paradigmatic \emph{AGT}\xspace problem.
\begin{table}[!ht]
\centering
\resizebox{\textwidth}{!}{%
\begin{tabular}{>{\bf}c@{\hspace{1em}}r@{\hspace{1em}}r@{\hspace{1em}}r@{\hspace{1em}}r@{\hspace{3em}}r@{\hspace{1em}}>{\bf}c@{\hspace{1em}}r@{\hspace{1em}}r@{\hspace{1em}}r@{\hspace{1em}}r@{\hspace{1em}}r@{\hspace{1em}}}
\toprule
\textbf{($\bm{|V|}$, $\bm{|E|}$)} & \textbf{\#EI} & \textbf{Time} & \textbf{Time-\nth{1}}& \textbf{\emph{PoS}\xspace} & \textbf{Tl} & \textbf{($\bm{|V|}$, $\bm{|E|}$))} & \textbf{\#EI} & \textbf{Time} & \textbf{Time-\nth{1}}& \textbf{\emph{PoS}\xspace} & \textbf{Tl} \\
\midrule
(50, 99) & 6.00 & 0.04 & 0.04 & 1.12 & 0/3 & (300, 626) & 21.00 & 12.11 & 2.64 & 1.04 & 0/3 \\
(100, 206) & 2.33 & 0.05 & 0.04 & 1.00 & 0/3 & (350, 730) & 19.00 & 13.92 & 7.42 & 1.01 & 0/3 \\
(150, 308) & 6.00 & 0.64 & 0.25 & 1.01 & 0/3 & (400, 822) & 248.67 & 694.95 & 228.69 & 1.08 & 1/3 \\
(200, 416) & 11.67 & 3.28 & 1.11 & 1.06 & 0/3 & (450, 934) & 394.67 & 1199.98 & 2.61 & 1.11 & 2/3 \\
(250, 517) & 64.67 & 63.50 & 16.07 & 1.02 & 0/3 & (500, 1060) & 35.67 & 87.07 & 7.25 & 1.00 & 0/3 \\
\bottomrule \vspace{-\aboverulesep} \\
\end{tabular} %
}
\caption{Results overview for the \emph{NFG}\xspace. The complete table of results is in the Appendix (\cref{tab:NFG_Full}).}
\label{tab:NFG}
\end{table}
\subsection{The Competitive Facility Location and Design Game}
The \emph{CFLD}\xspace \citep{aboolian_competitive_2007} is a game where sellers (players) compete for the demand of customers located in a given geographical area. Each seller makes two fundamental choices: where to open its selling facilities and the product assortment of such facilities, i.e., their “design”. Symmetrically, the customers select their favorite facilities depending on the relative distance from a facility and its attractiveness in terms of design.
We consider a variant recently presented by \citet{cronert_equilibrium_2020}, where $n$ competitors simultaneously choose the location and design of their facilities. Let $L$ be the set of potential facility locations, $J$ be the set of customers, and let $R_l$ denote the set of design alternatives for each location $l \in L$. Each player $i$ has an available budget $B^i$ and incurs in a fixed cost $f_{lr}^i$ when opening a facility at location $l \in L$ and with the design $r \in R_l$. Each player $i$ acquires a share of the demand $w_j$ of a customer $j \in J$ according to a utility $u_{ljr}^i$, whose value depends upon the distance of customer $j$ from a facility in location $l$ and the design choice of such facility (see \citet{cronert_equilibrium_2020} for more details). The \emph{CFLD}\xspace formulates as an \emph{IPG}\xspace where each player $i$ solves
\begin{maxi!}
{x^i}{
%
\sum_{j\in J} w_j \frac{\sum_{l \in L} \sum_{r \in R_l} u_{ljr}^ix^i_{lr}}
{\sum_{k=1}^n \sum_{l \in L} \sum_{r \in R_l} u_{ljr}^kx^k_{lr}}
%
\protect\label{eq:CFLD:Obj}}
{\label{eq:CFLD}}{}
%
\addConstraint{\sum_{l \in L} \sum_{r \in R_l} f_{lr}^ix^i_{lr}}{ \le B^i \protect\label{eq:CFLD:Budget}, \quad}{}{}
%
\addConstraint{\sum_{r \in R_l} x^i_{lr} }{\le 1 \protect\label{eq:CFLD:sos} \quad \quad }{\forall l \in L,}
%
\addConstraint{ x^i_{lr} \in \{0, 1\}}{ \protect\label{eq:CFLD:Vars} \quad }{\forall l \in L, \forall r \in R_l.}
\end{maxi!}
\noindent The binary variable $x^i_{lr}$ is $1$ if player $i$ opens a facility in the location $l \in L$ with a design $r \in R_l$.
The objective function \eqref{eq:CFLD:Obj} represents the share of customer demands player $i$ maximizes, the constraint \eqref{eq:CFLD:Budget} is the budget constraint for player $i$, and the constraints \eqref{eq:CFLD:sos} enforce that player $i$ can open only one facility in a location $l$. As in the \emph{NFG}\xspace, the objective is not linear in $x$, and the denominator can be zero; however, we can linearize it with tailored fractional programming techniques.
\paragraph{Computational Results. }
We test \emph{ZERO Regrets} on a representative set of instances from \citet{cronert_equilibrium_2020} to which we refer for the details concerning the distributions of locations and customers, and the entries $w_j$, $u_{ljr}$, $f_{lr}$. The resulting $64$ instances with $n=2,3$ have $50$ locations and $50$ customers, with budgets $B^1 \in [10, 40]$ and $B^2 = B^1, B^1 +10, \dots, 100$, $B^3 = 10$. \cref{tab:CFLD} summarizes the results, where we aggregate and average over the values of $B^1$. We benchmark our results against the performance of \emph{eSGM-WM} from \citet[Table 2]{cronert_equilibrium_2020}, which ran on a machine with similar hardware characteristics. Although the authors compute both pure and mixed welfare-maximizing equilibria, we focus on computing only the welfare-maximizing \emph{PNE}\xspace. Generally, \emph{ZERO Regrets} outperforms algorithm \emph{eSGM-WM} even in instances where only \emph{PNE}s\xspace exist. Our algorithm solves $62$ instances over $64$ within a time limit of $3600$ seconds. The running times of \emph{ZERO Regrets} are sensibly limited compared to those of \emph{eSGM-WM}, and never hit the time limit on the instances with $n=2$. Occasionally, the running times are dramatically smaller, e.g., the instance with $n=2$, $B^1 = 40$, $B^2 = 80$ where only one \emph{PNE}\xspace exists: our algorithm finds the most efficient \emph{PNE}\xspace in about $1636$ seconds, while \emph{eSGM-WM} requires $163315$ seconds.
\begin{table}[!ht]
\centering
\resizebox{\textwidth}{!}{%
\begin{tabular}{>{\bf}c@{\hspace{1em}}r@{\hspace{1em}}r@{\hspace{1em}}r@{\hspace{1em}}r@{\hspace{3em}}r@{\hspace{1em}}>{\bf}c@{\hspace{1em}}r@{\hspace{1em}}r@{\hspace{1em}}r@{\hspace{1em}}r@{\hspace{1em}}r@{\hspace{1em}}}
\toprule
\textbf{($\bm{n}$, $\bm{B^1}$)} &
\textbf{\#EI} &
\textbf{Time} &
\textbf{Time-\nth{1}}&
\textbf{\emph{PoS}\xspace} &
\textbf{Tl} &
\textbf{($\bm{n}$, $\bm{B^1}$)} &
\textbf{\#EI} &
\textbf{Time} &
\textbf{Time-\nth{1}}&
\textbf{\emph{PoS}\xspace} &
\textbf{Tl} \\
\midrule
(2, 10) & 4.00 & 1.76 & 1.76 & 1.01 & 0/10 & (3, 10) & 6.30 & 3.56 & 2.11 & 1.02 & 0/10 \\
(2, 20) & 5.11 & 7.39 & 3.03 & 1.01 & 0/9 & (3, 20) & 9.00 & 27.72 & 7.29 & 1.03 & 0/9 \\
(2, 30) & 9.25 & 339.35 & 59.39 & 1.03 & 0/8 & (3, 30) & 18.38 & 754.84 & 555.78 & 1.05 & 1/8 \\
(2, 40) & 16.40 & 682.00 & 294.92 & 1.07 & 0/5 & (3, 40) & 25.20 & 1863.92 & 739.63 & 1.06 & 1/5 \\
\bottomrule \vspace{-\aboverulesep}
\end{tabular} %
}
\caption{Results overview for the \emph{CFLD}\xspace from the instances of \citet[Table 2]{cronert_equilibrium_2020}. The complete table of results is in the Appendix (\cref{tab:CFLD_Full}).}
\label{tab:CFLD}
\end{table}
\subsection{The Quadratic Game}
The \emph{qIPG}\xspace is a simultaneous non-cooperative \emph{IPG}\xspace introduced by \citet{sagratella_computing_2016}, where each player $i$ solves the problem
\begin{align}
\min_{x^i} \{ \frac{1}{2}(x^i)^\top Q^i x^i + (C^i x^{-i})^\top x^i + (c^i)^\top x^i : LB \le x^i \le UB, \; x^i \in \mathbb{Z}^m \}.
\label{eq:QuadratiGame}
\end{align}
Specifically, each player $i$ controls $m$ integer variables bounded by the vectors $LB$ and $UB$. The strategic interaction involves the term $(C^i x^{-i})^\top x^i$, while the linear and quadratic terms solely depend on each player's choices. While \citet{sagratella_computing_2016} considers only instances with positive-definite $Q^i$ matrices (i.e., the problem is convex in $x^i$ for any $i$), \citet{schwarze_branch-and-prune_2022} consider arbitrary matrices $Q^i$ (i.e., non-convex objectives). In particular, the latter generalizes the former by dropping the convexity requirement w.r.t. $x^i$ on the payoffs $u^i(x^i,x^{-i})$. In contrast with the aforementioned applications, we let the \emph{MIP}\xspace solver manage the linearization of the quadratic terms in each player's payoff in order to fully integrate \emph{ZERO Regrets} with the features of the existing \emph{MIP}\xspace technology. As in \citet{sagratella_computing_2016,schwarze_branch-and-prune_2022}, we setup \emph{ZERO Regrets} to enumerate all \emph{PNE}s\xspace or to certify that no \emph{PNE}\xspace exists.
\paragraph{Computational Results.} We test our algorithm on both convex and non-convex benchmarks of the \emph{qIPG}\xspace.
First, we consider the \emph{qIPG}\xspace from \citet{schwarze_branch-and-prune_2022}, and test our algorithm on the same instance set. We refer to the original paper for the details on instance generation. Besides the bounds on the $x^i$ variables, these instances also include $m$ non-redundant linear inequalities $A^i x^i \le b^i$ for each player $i$. \cref{tab:Quad_SS} reports an overview of the results with a similar notation to the one of the previous tables. In the first column, we report the tuple $(n, m, t)$, where $t$ is either $C$ when each player's problem is convex or $NC$ otherwise. We additionally report the average number of \emph{PNE}s\xspace in the column $\#EQs$. We solve all the $56$ instances in less than $416$ seconds globally, with the most computationally-difficult instance requiring $56$ seconds.
Similar to the previous tests, our algorithm strongly outperforms the one of \citet{schwarze_branch-and-prune_2022}.
Specifically, their algorithm runs out of time in $25$ instances (time limit of $3600$ seconds) and solves the remaining $31$ instances with non-negligible computational times (i.e., about $1302$ seconds on average).
\begin{table}[!ht]
\setlength\tabcolsep{0.7em}
\centering
\resizebox{0.67\textwidth}{!}{%
\begin{tabular}{>{\bf}crrrrrrr}
\toprule
\textbf{($\bm{n}, \bm{m}, \bm{t}$)} &
\textbf{\#EI} &
\textbf{\#EQs} &
\textbf{Time} &
\textbf{Time-\nth{1}} &
\textbf{\emph{PoS}\xspace} &
\textbf{\emph{PoA}\xspace} &
\textbf{Tl} \\
\midrule
(2, 2, C) & 14.00 & 1.67 & 0.35 & 0.18 & 1.17 & 1.43 & 0/4 \\
(2, 3, C) & 26.75 & 1.60 & 1.13 & 0.42 & 5.56 & 6.10 & 0/8 \\
(2, 4, C) & 35.00 & 1.00 & 3.40 & 1.19 & 1.32 & 1.32 & 0/4 \\
(2, 5, C) & 56.00 & 1.50 & 35.88 & 7.72 & 2.19 & 4.74 & 0/4 \\
(3, 2, C) & 42.00 & 2.00 & 1.21 & 0.65 & 1.83 & 2.40 & 0/4 \\
(3, 3, C) & 108.75 & 1.00 & 35.62 & 7.09 & 3.79 & 3.79 & 0/4 \\
(2, 2, NC) & 16.00 & 1.33 & 0.43 & 0.30 & 2.20 & 2.20 & 0/4 \\
(2, 3, NC) & 20.25 & 1.75 & 1.07 & 0.40 & 1.33 & 1.68 & 0/8 \\
(2, 4, NC) & 19.50 & 1.00 & 1.51 & 0.82 & 1.02 & 1.02 & 0/4 \\
(2, 5, NC) & 28.50 & 1.67 & 14.03 & 2.38 & 1.31 & 1.53 & 0/4 \\
(3, 2, NC) & 30.00 & 2.67 & 1.03 & 0.37 & 1.03 & 1.60 & 0/4 \\
(3, 3, NC) & 46.50 & 1.33 & 6.11 & 2.43 & 1.44 & 1.45 & 0/4 \\
\bottomrule \vspace{-\aboverulesep}
\end{tabular} %
}
\caption{Results overview for the \emph{qIPG}\xspace from the instances of \citet{schwarze_branch-and-prune_2022}. The complete table of results is in the Appendix (\cref{tab:Quad_SS_Full}).}
\label{tab:Quad_SS}
\end{table}
To get a broader perspective, we also consider the convex instances of the game generated according to the scheme proposed in \citet{sagratella_computing_2016}. The matrices $Q^i$ and $C^i$ are a random positive-definite and a random matrix, respectively, with rational entries in the range $[-25, 25]$, while the entries of $c^i$ are integer numbers in the range $[-5, 5]$. We generate our instances with $n \in [1,6]$, $m \in [1,10]$, $LB \in [-1000,0]$, and $UB \in[5,1000]$ similarly to \citet{sagratella_computing_2016}.
We report the average results in \cref{tab:Quad_S}, where we aggregate for $n$.
\emph{ZERO Regrets} finds the first \emph{PNE}\xspace in less than a second on average, and manages to solve any instance in less than $12$ seconds, even when more than $10$ \emph{PNE}s\xspace exist (see \cref{tab:Quad_Sa_Full}). Although we note that the algorithm of \citet{sagratella_computing_2016} ran on a less performing machine, the results in \cref{tab:Quad_S} highlight the remarkable effectiveness of \emph{ZERO Regrets}. The speedup in the performance seems to be considerably larger than the improvement associated with different hardware and software specifications (i.e., our algorithm is $100$ times faster in terms of computing times).
\begin{table}[!ht]
\setlength\tabcolsep{0.7em}
\centering
\resizebox{0.50\textwidth}{!}{%
\begin{tabular}{>{\bf}clllll}
\toprule
\textbf{$\bm{n}$} &
\textbf{\#EI} &
\textbf{\#EQs} &
\textbf{Time} &
\textbf{Time-\nth{1}} &
\textbf{Tl} \\
\midrule
2 & 81.33 & 3.67 & 2.21 & 0.58 & 0/12 \\
3 & 115.13 & 4.00 & 2.44 & 0.60 & 0/8 \\
4 & 119.00 & 10.25 & 3.90 & 0.97 & 0/4 \\
6 & 79.50 & 3.50 & 0.96 & 0.38 & 0/4 \\
\bottomrule \vspace{-\aboverulesep}
\end{tabular} %
}
\caption{Results for the \emph{qIPG}\xspace from the instances of \citet{sagratella_computing_2016}. The complete table of results is in the Appendix (\cref{tab:Quad_Sa_Full}).}
\label{tab:Quad_S}
\end{table}
\section{Concluding Remarks}
\label{sec:concluding}
This paper presents a general framework to compute, enumerate and select equilibria for a class of \emph{IPG}s\xspace. These games are a fairly natural multi-agent extension of traditional problems in Operations Research, such as resource allocation, pricing, and combinatorial problems, and are powerful modeling tools for various applications.
We provide a theoretical characterization of our framework through the concepts of \emph{equilibrium inequality} and \emph{equilibrium closure}. We explore the interplay between rationality and cutting planes by introducing a series of general and special-purpose classes of equilibrium inequalities and provide an interpretable criterion to frame the strategic interaction among players. The algorithm we introduce is general and it smoothly integrates with the existing optimization technology. Practically, we apply our framework to various problems from the relevant literature and significant application domains. We perform an extensive computational campaign and demonstrate the high efficiency and scalability of \emph{ZERO Regrets}. Our computational results also provide evidence of the existence of efficient \emph{PNE}s\xspace, further motivating the need for suitable algorithms to select or enumerate them. We also remark that our algorithm could practically work -- up to a numerical tolerance -- even when some of the players' variables are bounded and continuous, e.g., by dropping the integrality requirement on some variables.
We are prudently optimistic about the impact our framework may have in applied domains and the future methodological research directions it may open. We envision the potential for a series of theoretical contributions regarding the structure of new classes of general and problem-specific equilibrium inequalities and computational methods to further improve the algorithm's performance. Above all, we hope our framework will foster future academic research and clear the way for novel and impactful applications of \emph{IPG}s\xspace.
\iftoggle{ARXIV}{
\subsubsection*{Acknowledgements}
We would like to thank Ulrich Pferschy for the valuable discussions concerning our work.
}{
\ACKNOWLEDGMENT{We would like to thank Ulrich Pferschy for useful discussions on our work.}
}
\newpage
|
\section{Introduction}
\label{sec:Intro}
Axions~\cite{Weinberg:1977ma,Wilczek:1977pj} are the direct and testable prediction following from the Peccei-Quinn (PQ) mechanism~\cite{Peccei:1977hh,Peccei:1977ur} proposed as a possible solution to the \textit{strong CP problem} of the Standard Model (SM) of particle physics.
This problem, still one of the most puzzling in modern particle physics, concerns the apparent absence of CP-violating effects in Quantum Chromodynamics (QCD).
To emphasize their role in explaining the observed QCD behavior,
such particles are also dubbed \emph{QCD axions}.
Other kinds of pseudoscalar particles, with properties very similar to the QCD axion but no relation to the strong CP problem, emerge naturally in several extensions of the SM, in particular in compactified string theories~\cite{Witten:1984dg,Conlon:2006tq,Arvanitaki:2009fg,Acharya:2010zx, Higaki:2011me,Cicoli:2012sz,Demirtas:2018akl,Mehta:2021pwf}.
To distinguish them from the QCD axion, these are often called axion-like particles (ALPs).
The considerations discussed in this paper apply, in large part, to the QCD axion as well as to ALPs, except when specific QCD axion models are discussed.
In the interest of brevity, we will refer to both QCD axions and ALPs as ``axions'' throughout this paper.
The axion phenomenology and experimental landscape are discussed in several recent reviews~\cite{Irastorza:2018dyq,DiVecchia:2019ejf,DiLuzio:2020wdo,Agrawal:2021dbo,Sikivie:2020zpn}.
In summary, axions are expected to couple to the SM fields with model dependent couplings,
\begin{equation}
\label{eq:L_int}
\mathcal{L}_{\rm int} = -\frac{1}{4} g_{a\gamma} a F_{\mu\nu} \tilde F^{\mu\nu}
- \sum_f i\,g_{af}\, a\,\bar f \gamma_5 f \, ,
\end{equation}
where $F$ is the electromagnetic field tensor and $f$ are the SM fermions (for the present work, it is sufficient to consider couplings to protons, neutrons and electrons, so we can assume $f=p,n,e$).
Particularly appealing is the axion coupling to photons, since it allows very promising experimental strategies for the axion search.
However, the couplings to electrons and nucleons are also employed in experimental axion searches.
In this work, we examine the axion coupling to nucleons (for a recent calculation see~\cite{GrillidiCortona:2015jxo}) and explore the potential to probe it with axion helioscopes~\cite{Sikivie:1983ip} which, as suggested by the name, search for axions produced in the sun.
From a theoretical point of view the nucleon coupling is of particular interest because it receives, similar to the photon coupling, an unavoidable contribution from the defining coupling of QCD axions to gluons.
While it is possible to conceive models where this is (partially) cancelled, thereby having QCD axions with a suppressed coupling to nucleons~\cite{DiLuzio:2017ogq}, a large nuclear coupling is expected in most models which solve the strong CP problem and in certain cases can be considerably enhanced (see, e.g., Ref.~\cite{Darme:2020gyx}).
A sizable nucleon coupling is therefore an expected feature of a QCD axion.
In this sense, measuring the axion-nucleon coupling would be a good indication of the QCD axion nature.
Furthermore, such a detection
by a helioscope would likely be accompanied by a signal from Primakoff (and possibly Compton/Bremsstrahlung) axions.
All these spectra have well defined shapes and so, given enough statistics and energy resolution, the data analysis might help understand
specific properties of the detected particle
(see discussion in Ref.~\cite{Jaeckel:2018mbn}
for a similar analysis in the case of the axion-photon and axion-electron couplings).
The axion coupling to nucleons can be probed indirectly in astrophysics, through the effects on the cooling of neutron stars (NS)~\cite{Keller:2012yr,Sedrakian:2015krq,Hamaguchi:2018oqw,Beznogov:2018fda,Sedrakian:2018kdm,Leinson:2021ety} and from the analysis of the observed neutrino signal from SN 1987A~\cite{Turner:1987by,Burrows:1988ah,Raffelt:1987yt,Raffelt:1990yz,Carenza:2019pxu,Carenza:2020cis,Fischer:2021jfm}.
A direct detection is also possible, in principle, through experiments such as
CASPEr-gradient~\cite{Garcon:2017ixh} or ARIADNE~\cite{Arvanitaki:2014dfa}.\footnote{Note, however, that both experiments rely on important assumptions which strongly affect the predicted sensitivities.
The CASPEr detection potential depends on the axion primordial abundance, and it is strongly reduced if axions constitute only a
small fraction of the total dark matter in the galaxy.
The ARIADNE potential
depends on the CP-violating axion-scalar coupling,
which is presently very poorly known (see.\ e.g., Ref.~\cite{Bertolini:2020hjc}).}
Axions coupled to nucleons could also be produced in the sun, for example through the decay of excited nuclear states.
As pointed out already in Weinberg's seminal paper~\cite{Weinberg:1977ma}, being pseudoscalars, axions can be emitted in magnetic nuclear transitions.
The best known example, and the one adopted here, is the decay of the first excited state of $^{57}$Fe with the emission of a 14.4\,keV axion. Other nuclear transitions turn out to generate a substantially smaller axion flux (see also Appendix~\ref{app:lines}).
The search for these $^{57}$Fe axions has a long history.
The currently most powerful helioscope, the CERN Axion Solar Telescope (CAST)~\cite{CAST:2009jdc}, as well as CUORE~\cite{CUORE:2012ymr} and, more recently, XENON1T~\cite{XENON:2020rca}, have
searched for axions produced in this transition and provided constraints on the axion-nucleon coupling.
In this paper, we assess the potential of the next generation of axion helioscopes, BabyIAXO~~\cite{BabyIAXO:2020mzw,Abeln:2020eft}
and IAXO~\cite{Irastorza:2011gs,IAXO:2019mpb,Giannotti:2016drd}, to detect such axions.
An important motivating factor is that with BabyIAXO~ the transition from concept to real experiment is imminent, as its construction is expected to start in
2022.
We therefore believe it is of great importance to assess its potential for an axion detection in all possible channels.
Furthermore, the BabyIAXO~ and IAXO potential to probe this coupling are expected to be considerably superior to that of CAST~\cite{CAST:2009jdc}, allowing to probe large regions of additional axion parameter space.
We provide a guide for the best setups required to maximize the efficiency to detect axions from $^{57}$Fe.
We also use this opportunity to include new theoretical developments.
In particular, the matrix elements for the relevant transition have recently been revised~\cite{Avignone:2017ylv}, and show a $\sim$\,30\,\% increase in the axion emission rate.
In addition, we use a more recent solar model than previous publications.
Finally, we discuss the model dependence of the $^{57}$Fe~ signal
and identify a class of QCD axion models
which yield an enhanced signal
compared to the standard
KSVZ~\cite{Kim:1979if,Shifman:1979if}
and
DFSZ~\cite{Zhitnitsky:1980tq,Dine:1981rt}
axion models.
The paper is organized as follows.
In section~\ref{sec:ALPs_from_nuclear_transitions}, we revisit the problem of the solar axion production in nuclear transitions, provide an updated expression for the flux from $^{57}$Fe~
and discuss its dependence on the specific axion model;
in section~\ref{sec:stars}, we
present a brief discussion of the current stellar bounds;
section~\ref{sec:sensitivity_estimates} is dedicated to the sensitivity estimates for Baby\-IAXO and IAXO, assuming different experimental setups;
finally, in section~\ref{sec:Conclusion} we provide our conclusions.
A discussion of other nuclei that could contribute spectral lines to the solar axion flux can be found in Appendix~\ref{app:lines}.
\section{Axions from nuclear transitions in the sun}
\label{sec:ALPs_from_nuclear_transitions}
Axion helioscopes~\cite{Sikivie:1983ip} such as the planned International Axion Observatory (IAXO)~\cite{Irastorza:2011gs, BabyIAXO:2020mzw, IAXO:2019mpb,Giannotti:2016drd,Abeln:2020eft} are searching for very light (sub-eV) axions produced in the sun.
Axions can be produced in the solar plasma by various processes.
Helioscopes search primarily for axions originating from the Primakoff effect (cf, e.g.~\cite{Raffelt:1985nk,Raffelt:1987yu, Raffelt:2006cw,Carosi:2013rla,Ayala:2014pea,Giannotti:2015kwo,Giannotti:2017hny}) or from axion-electron interactions~\cite{Redondo:2013wwa,Hoof:2021mld,Jaeckel:2019xpa,Jaeckel:2018mbn}.
This makes them sensitive to both the axion-photon coupling $g_{a\gamma}$ and to the product of photon and electron couplings $g_{a\gamma}g_{ae}$~\cite{Barth:2013sma,Jaeckel:2018mbn}.
Axions can also be emitted in solar nuclear processes.
There are two noteworthy mechanisms for this: nuclear fusion and decays as well as thermal excitation and subsequent de-excitation of the nuclei of stable isotopes.
Axions produced in nuclear fusion and nuclear decay processes typically carry an energy of the order of $\sim$\,MeV.
The axion flux from the $p+d\to$ $^{3}{\rm He}+a$\,(5.5\,MeV) reaction,
which provides one of the most intense axion fluxes from nuclear reactions, has been experimentally searched for by Borexino~\cite{Borexino:2012guz}.
More recently, the SNO data~\cite{Bhusal:2020bvx} has also been scrutinized for traces of these 5.5\,MeV axions.
Helioscopes can, in principle, also be equipped with a
$\gamma$-ray detector to look for such axions.
For example, the CAST helioscope installed a $\gamma$-ray calorimeter for some time to gain sensitivity to these high-energy axions~\cite{CAST:2009klq}.
However, in general all these analyses provided somewhat
weak bounds on the axion-nucleon coupling,
since the solar flux from nuclear reactions is
not expected to be very large.
Indeed, Ref.~\cite{CAST:2009klq} estimated an axion flux of the order of $10^{10} (g^{3}_{aN})^{2} /({\rm cm^2s})$ for the reaction mentioned above.
This is more than 10 orders of magnitude smaller than the flux we will find below, Eq.~\eqref{eq:totalflux}.
A perhaps more promising direction is to look for low lying nuclear excitations of stable isotopes with a significant abundance inside the sun that can be thermally excited.
Two candidates have been proposed for this in the past, $^{57}$Fe \cite{Moriyama:1995bz,Krcmar:1998xn} and $^{83}$Kr \cite{Gavrilyuk:2014mch}.
The former has a first nuclear excitation energy $E^*$ of 14.4\,keV and the latter of 9.4\,keV.
In the solar core, at temperatures $T\sim1.3$\,keV, these excited states have a small but non-vanishing occupation number that can be calculated from the Boltzmann distribution.
The amount of axions produced is proportional to the occupation number, the isotope abundance and the inverse lifetime of the excited state.
By combining a list of possible elements and their nuclear transitions \cite{Roehlsberger:2004} with the solar abundances in \cite{Asplund},
it becomes clear that for IAXO it is the $^{57}$Fe transition that would produce the strongest signal (see Appendix~\ref{app:lines} for more details).
\subsection{Effective axion coupling in the $^{57}$Fe nuclear transition}
\label{sec:models}
To describe the axion interactions with nuclei it is convenient to rewrite the relevant terms in the Lagrangian~\eqref{eq:L_int} as
\begin{align}
\label{eq:geffdef2}
\mathcal{L}_{aN}
= - i a \bar N \gamma_5 \left( g^0_{aN} + g^3_{aN} \tau^3 \right) N \,.
\end{align}
Here, $N = (p, n)^T$ is the nucleon doublet,
$g^0_{aN}$, $g^3_{aN}$ are the iso-scalar and iso-vector couplings, respectively,
and $\tau^3$ is the Pauli matrix.
The axion-to-photon branching ratio for the decay rates
of the first excited state of $^{57}$Fe can then be expressed as~\cite{Haxton:1991pu,Avignone:1988bv}
\begin{align}
\label{eq:Gamma_Ratio_general}
\frac{\Gamma_a}{\Gamma_\gamma}=
\left( \frac{k_a}{k_\gamma} \right)^3
\frac{1}{2\pi\alpha}\frac{1}{1+\delta^2}
\left[ \frac{\beta \, g_{aN}^{0} + g_{aN}^{3}}
{\left( \mu_0-\frac12 \right) \beta + \mu_3 -\eta} \right]^2
,
\end{align}
where $k_a,~k_\gamma$ are the axion and photon momenta,
$\mu_0$ and $\mu_3$ are the isoscalar and isovector nuclear magnetic moments (expressed in nuclear magnetons),
$\delta$ is the E2/M1 mixing ratio for the $^{57}$Fe nuclear transition,
while $\beta$ and $\eta$ are constants dependent on the nuclear structure~\cite{Avignone:1988bv}.
The constant $\delta\simeq 0.002$ can be safely neglected to the level of precision required in this work.
The other constants were recently reevaluated in Ref.~\cite{Avignone:2017ylv}.
Assuming $(k_a/k_\gamma)\simeq 1$, which applies to ultrarelativistic axions, and adopting the most recent
values in Ref.~\cite{Avignone:2017ylv} for the relevant nuclear constants, we find
\begin{align}
\label{eq:Gamma_Ratio_specific}
\frac{\Gamma_a}{\Gamma_\gamma}
&=2.32
\left( -1.31\, g_{aN}^0 + g_{aN}^3 \right)^2\\
&=2.32 \left( 0.16\, g_{ap} +1.16\, g_{an} \right)^2\,,
\label{eq:ratio_Avignone}
\end{align}
where the second line is expressed in terms of the more common axion couplings to neutrons,
\begin{equation}
g_{an} = g_{aN}^0 - g_{aN}^3\;,
\end{equation}
and to protons,
\begin{equation}
g_{ap} = g_{aN}^0 + g_{aN}^3\;.
\end{equation}
From the form of Eq.~\eqref{eq:ratio_Avignone},
it is convenient to define the effective nucleon coupling as\footnote{We stress that this combination is specific to the 14.4 keV transition in iron. However, as this is by far the most promising transition we use it in the following without a specific label.}
\begin{align}
\label{eq:g_N_eff}
g_{aN}^{{\rm eff}}=0.16\, g_{ap} +1.16\, g_{an}\,,
\end{align}
which is the coupling combination that controls the axion emission rate in this transition.
Notice that the updated branching ratio is 27\,\% larger than the one
found in Ref.~\cite{Haxton:1991pu} and used in the previous experimental analyses, including CAST~\cite{CAST:2009jdc}, CUORE~\cite{CUORE:2012ymr} and more recently XENON1T~\cite{XENON:2020rca}.
Furthermore, the relative importance of the coupling to protons has strengthened in this new analysis even though, as evident from Eq.~\eqref{eq:ratio_Avignone}, it is still considerably less relevant than the coupling to neutrons.
\subsection{Axion model dependence
of the $^{57}$Fe~ transition rate}\label{subsec:modeldep}
Before proceeding with the experimental sensitivity, it is interesting to look at the axion emission rate for some
benchmark axion models.
Expressing the axion-nucleon
couplings in terms of the
dimensionless
axion-quark
coefficients $c^0_q$ \cite{diCortona:2015ldu}
(defined via the Lagrangian term
$\frac{\partial_\mu a}{2 f_a} c^0_q
\bar q \gamma^\mu \gamma_5 q$,
with $f_a$ the axion decay constant), Eq.~\eqref{eq:ratio_Avignone} can be cast as\footnote{In deriving this equation, we use the standard relation (cf., e.g.,~\cite{GrillidiCortona:2015jxo})
\begin{align}
m_a\simeq 5.7 \, \mu {\rm eV}\times \left( \frac{10^{12} \, {\rm GeV}}{f_a} \right)
\end{align}
between the axion mass and decay constant.
}
\begin{align}
\label{eq:Gamma_Ratio_quarkcoupl}
&\frac{\Gamma_a}{\Gamma_\gamma}
=5.81 \times 10^{-16}
(
1
+3.28 \, c^0_u
-9.97 \, c^0_d
+0.52 \, c^0_s \nonumber \\
&
+0.16 \, c^0_c
+0.12 \, c^0_b
+0.048 \, c^0_t
)^2
\left( \frac{m_a}{1 \, \text{eV}} \right)^2 \,,
\end{align}
where $m_a$ is the axion mass.
In the KSVZ model \cite{Kim:1979if,Shifman:1979if}
one has $c^0_q = 0$, yielding
\begin{equation}
\label{eq:Gamma_Ratio_KSVZ}
\left. \frac{\Gamma_a}{\Gamma_\gamma} \right|_{\rm KSVZ}
=5.81 \times 10^{-16}
\left( \frac{m_a}{1 \, \text{eV}} \right)^2 \,,
\end{equation}
while in
the DFSZ
model \cite{Zhitnitsky:1980tq,Dine:1981rt},
$c^0_{u,c,t} = \frac{1}{3} \cos^2\beta$
and $c^0_{d,s,b} = \frac{1}{3} \sin^2\beta$
corresponding to
\begin{align}
\label{eq:Gamma_Ratio_DFSZ}
&\left.\frac{\Gamma_a}{\Gamma_\gamma}\right|_{\rm DFSZ}
=5.81 \times 10^{-16} \\
&\times (
1
+ 1.16 \cos^2\beta - 3.11 \sin^2\beta
)^2
\left( \frac{m_a}{1 \, \text{eV}} \right)^2 \,,
\nonumber
\end{align}
with $\tan\beta$
defined in the perturbative domain $\in[0.25, 170]$ \cite{DiLuzio:2020wdo}.
Note that in KSVZ models the axion emission rate is accidentally suppressed as in these models the neutron coupling is very small. In DSFZ models it can get enhanced by
up to a factor of $\sim$\,4 with respect to the KSVZ.
Eq.~\eqref{eq:Gamma_Ratio_quarkcoupl}
suggests that a strong enhancement of the
axion emission rate can be achieved
for ``down-philic'' axions, with
$c^0_d \gg c^0_u$.
More generally this holds if
a sizable cancellation with the
model-independent factor normalized to 1
in Eq.~\eqref{eq:Gamma_Ratio_quarkcoupl}
is avoided.
This last possibility naturally happens in a class
of non-universal
DFSZ models with $c^0_u + c^0_d = 1$
that have been analyzed in
Refs.~\cite{DiLuzio:2017ogq,Bjorkeroth:2018ipq,Bjorkeroth:2019jtx}
(for a summary of axion couplings in those
models see Table~3 in \cite{DiLuzio:2021ysg}).
For instance, in the non-universal M1 model
of Ref.~\cite{DiLuzio:2017ogq}
one has
$c^0_u = c^0_c = \sin^2\beta$,
$c^0_t = - \cos^2\beta$,
$c^0_d = c^0_s = \cos^2\beta$,
$c^0_b = - \sin^2\beta$,
with
$\tan\beta \in [0.25, 170]$
(also, $c^0_e = - \sin^2\beta$ and $E/N = 2/3$),
leading to
\begin{align}
\label{eq:Gamma_Ratio_M1}
&\left.\frac{\Gamma_a}{\Gamma_\gamma}\right|_{\rm M1}
=5.81 \times 10^{-16} \\
&\times (
1
+ 3.32 \sin^2\beta - 9.50 \cos^2\beta
)^2
\left( \frac{m_a}{1 \, \text{eV}} \right)^2 \,,
\nonumber
\end{align}
which at small $\beta$
yields an $\mathcal{O}(60)$
enhancement
of the axion emission rate
with respect to the KSVZ model.\footnote{Again keeping $\beta$ in the perturbative range.}
Other non-universal DFSZ models,
among those mentioned above,
feature a similar enhancement
of the $g^{\rm eff}_{aN}$
coupling, but they have different
values for the axion-photon coupling
(that is important
in detection for the IAXO setup).
In particular,
the non-universal
model $\mathcal{T}_2^{(u)}$
of Ref.~\cite{Bjorkeroth:2018ipq}
features the largest
axion coupling to photons
among the general
class of non-universal DFSZ models with two Higgs doublets
(see Table 5 in \cite{DiLuzio:2021ysg}).
\subsection{Solar $^{57}$Fe axion flux and limits}
\label{sec:axion_flux}
To calculate the resulting flux from thermal excitation and subsequent de-excitation in detail,
we follow the derivation by Moriyama \cite{Moriyama:1995bz}, also used by the CAST collaboration in their search for this source \cite{CAST:2009jdc},
but adopt updated nuclear matrix elements derived in Ref.~\cite{Avignone:2017ylv}.
The axion emission rate per unit mass of solar matter is given by
\begin{equation}
\label{eq:axionspersolarmatter}
\mathcal{N}_a=\mathcal{N}\,\omega_1(T)\, \frac{1}{\tau_0}\frac{1}{1+\alpha} \frac{\Gamma_a}{\Gamma_\gamma},
\end{equation}
where $\mathcal{N}$ is the $^{57}$Fe number density per solar mass, $\omega_1$ the occupation number of the first excited state, $\tau_0$ the lifetime of the excited state, $\alpha$ the internal conversion coefficient and $\frac{\Gamma_a}{\Gamma_\gamma}$ the branching ratio of axion to photon emission, given in Eq.~\eqref{eq:ratio_Avignone}.
A detailed description of these parameters is given in Appendix~\ref{app:lines}.
The axion signal from this nuclear transition is expected to be very narrow.
The natural line width is negligible compared to the Doppler broadening $\sigma$ of
\begin{equation}
\sigma(T)=E^*\sqrt{\frac{T}{m_{\text{Fe57}}}}\sim 2 \ \text{eV}\,,
\label{eq:width}
\end{equation}
corresponding to a full width at half maximum (FWHM) of $2.35\,\sigma \sim 5$ eV.
Consequently, the spectral axion flux at a distance of one astronomical unit $d_\odot$ from the sun is an integral of a Gaussian peak over the solar radius $R_\odot$,\footnote{We use the solar axion flux library~\cite{Hoof:2021mld} which is publicly available at \url{https://github.com/sebhoof/SolarAxionFlux} to perform all axion flux calculations.}
\begin{align}
\frac{\dif\Phi_a(E_a)}{\dif E_a}=&\frac{1}{4\pi d_\odot^2} \int_{0}^{R_\odot} \mathcal{N}_a(T(r)) \frac{1}{\sqrt{2\pi}\sigma(T(r))} \nonumber \\
&\times \exp \left(-\frac{(E_a-E^*)^2}{2\sigma(T(r))^2}\right)\rho(r)4\pi r^2 \text{d}r.
\end{align}
Although, depending on the optics adopted, IAXO's field of view may cover only part of the sun,
we integrate over the entire solar radius since -- as
illustrated in Fig~\ref{fig:flux-vs-radius} -- more than 99\,\% of the total flux originates from a circle around the solar centre with radius 0.15\,$R_\odot$.
\begin{figure}
\centering
\includegraphics[width=1\linewidth]{figures/flux_accumulation.pdf}
\caption{Radial dependence of the solar axion flux from $^{57}$Fe transitions and the Primakoff effect. We show the fraction of the flux inside the field of view which is a circle around the solar center with radius $r$. }
\label{fig:flux-vs-radius}
\end{figure}
The solar model (in our case B16-AGSS09~\cite{Vinyoles:2016djt}) fixes temperature $T$, density $\rho$ and iron abundance at every radius.
It is noteworthy that this axion source is particularly sensitive to the temperature. This is because of the thermal occupation number $\omega_1 \propto e^{-E^*/T}$ of the excited nuclear state. For instance, the latest high and low metallicity solar models (B16-GS98 and B16-AGSS09 \cite{Vinyoles:2016djt}) only differ by about 1\,\% in their respective core temperature but this alone results in a difference of 12\,\% in the total flux.
We do not expect the narrow, Doppler-broadened peak to be resolved.\footnote{Metallic magnetic calorimeters~\cite{Unger2021} may reach an energy resolution of few eV which would enable them to coarsely resolve the Doppler peak.} Hence, only the total flux is of interest and the integral over $E_a$ can be performed to get the total solar axion flux from the $^{57}$Fe nuclear transition \cite{CAST:2009jdc},
\begin{align}
\label{eq:totalflux}
\Phi_a=&\frac{1}{4\pi d_\odot^2} \mathcal{N}
\frac{1}{\tau_0}\frac{1}{1+\alpha}
\frac{\Gamma_a}{\Gamma_\gamma}\nonumber \\&\times \int_{0}^{R_\odot} \omega_1(T(r))\rho(r)4\pi r^2 \text{d}r\\
=& 5.06 \times 10^{23}\ (g_{aN}^{{\rm eff}})^2 \ \rm{cm}^{-2}\rm{s}^{-1} \ .
\end{align}
This numerical result is larger than the one previously derived by the CAST collaboration~\cite{CAST:2009jdc} because we worked with the updated nuclear matrix elements. In addition, we integrated over a more recent solar model, namely B16-AGSS09~\cite{Vinyoles:2016djt}, whose core temperature and iron abundance is smaller compared to the values adopted previously. This explains why the overall flux is not equally enhanced as the updated nuclear matrix elements.
The following calculations are all done with this updated axion flux and the cited bounds derived by the CAST collaboration in~\cite{CAST:2009jdc} were rescaled accordingly.
\bigskip
\section{Confronting stellar bounds}
\label{sec:stars}
Let us start by quickly updating the energy loss constraint.
Using the revised axion rate, Eq~\eqref{eq:ratio_Avignone}, and the solar model B16-AGSS09 we find a total axion luminosity (energy loss rate),
\begin{equation}
L_a= 8.38\times 10^9 (g_{aN}^{{\rm eff}})^2\,L_{\odot}
\end{equation}
via the $^{57}$Fe transition.
The recent study in Ref.~\cite{Vinyoles:2015aba} constrains an exotic energy loss to a maximum of 3\,\% of the standard solar luminosity $L_{\odot}$.
This leads to an improved bound on the axion effective coupling with nucleons,
\begin{equation}
g_{aN}^{{\rm eff}}\leq 1.89\times 10^{-6}\,.
\label{eq:solar_bound}
\end{equation}
Notice that the result in Eq.~\eqref{eq:solar_bound}, known as the \emph{solar bound} on $g_{aN}^{{\rm eff}}$,
is about a factor of two more stringent than the
previous constraint, $g_{aN}^{{\rm eff}}\leq 3.6 \times 10^{-6}$~\cite{CAST:2009jdc}.
Besides the enhanced emission rate~\cite{Avignone:2017ylv} and the updated solar model~\cite{Vinyoles:2016djt}, this is also due to Ref.~\cite{CAST:2009jdc} having excluded only $L_a > 0.1 \ L_\odot$.
\bigskip
The axion-nuclear coupling can also be constrained from other stellar observations.
In particular, strong bounds were derived from X-ray observations of various NS~\cite{Keller:2012yr,Sedrakian:2015krq,Hamaguchi:2018oqw,Beznogov:2018fda,Sedrakian:2018kdm,Leinson:2021ety}.
These bounds are, however, subject to several uncertainties
and do not always agree with each other, not even when referring to the same star~\cite{DiLuzio:2021ysg}.
In any case, all these analyses suggest
a limit
of $\sim 10^{-9}$ on some combination of axion-nucleon couplings.\footnote{In most of these analyses, the limit applies only to the axion-neutron coupling.}
A similar bound can be deduced from the analysis of the neutrino signal observed in coincidence with the SN 1987A event~\cite{Raffelt:1987yt,Turner:1987by,Raffelt:1990yz,Turner:1991ax,
Raffelt:1993ix,Keil:1996ju,Hanhart:2000ae,Fischer:2016cyd,Carenza:2019pxu,Fischer:2021jfm}.
Here, we refer specifically to the most recent analysis~\cite{Carenza:2019pxu}, which derived the bound
\begin{equation}
g_{an}^2+ 0.6\, g_{ap}^2 + 0.5\, g_{an}\,g_{ap}\lesssim 8.3 \times 10^{-19} \,.
\label{eq:gan_gap_SN_bound}
\end{equation}
This is a very strong constraint and, as we will see, only advanced setups may allow the exploration of smaller couplings.
It is worth noticing, however,
that the
coupling expected in the $^{57}$Fe transition differs from
the coupling in Eq.~\eqref{eq:gan_gap_SN_bound} and that, for some
specific models, the coupling relevant for the solar axion searches may be enhanced or suppressed with respect to what is constrained by the supernova (SN) argument (see Fig.~\ref{fig:models_vs_SN}).
For now, we will not investigate this argument further but
we stress that, due to the peculiar effective coupling
appearing in the $^{57}$Fe transition,
the comparison with the astrophysical bounds is model dependent.
\begin{figure}
\centering
\includegraphics[width=1\linewidth]{figures/SN_M1new.pdf}
\caption{
Effective axion coupling entering in the $^{57}$Fe transition as a function of the angle $\beta$, which defines the model dependent couplings (Cf.~Sec.~\ref{subsec:modeldep}).
In the figure, the colored lines indicate the value of the coupling $g_{aN}^{\rm eff}$
calculated assuming that the specific axion model saturates the SN bound, given in Eq.~\eqref{eq:gan_gap_SN_bound}.
}
\label{fig:models_vs_SN}
\end{figure}
That said, Eq.~\eqref{eq:gan_gap_SN_bound} can be translated into a limit on $g^{\rm eff}_{aN}$ by choosing the ratio between the proton and neutron couplings such that the left hand side of this equation is minimal, $g_{ap}/g_{an}\approx -0.2$ while keeping $g^{\rm eff}_{aN}$ constant.
This yields,
\begin{equation}
\label{eq:SNbound}
\left|g^{\rm eff}_{aN}\right|\lesssim 1.1\times 10^{-9}.
\end{equation}
The only model in Fig.~\ref{fig:models_vs_SN} which saturates this bound is the M1 model.
\section{Sensitivity estimates}
\label{sec:sensitivity_estimates}
In this section, we present a detailed discussion of the helioscopes potential to detect the $^{57}$Fe axion flux.
A particular emphasis is given to BabyIAXO,
which is expected to
start construction soon at DESY in Hamburg.
Estimates for the more advanced IAXO and IAXO+ configurations~\cite{IAXO:2019mpb} will also be presented.
Before going into specific experimental aspects let us make a few general considerations that can provide some help with the experimental design.
To detect the $^{57}$Fe line we want to maximize the signal to noise ratio in the single relevant energy bin which contains all of the signal events.
There are two main contributions to the background for the measurement of the nucleon line.
One is the usual background rate of the detector itself (e.g., cosmic rays, environmental gammas, intrinsic detector radioactivity) that usually grows linearly with the area of the detector and it is typically proportional to the spectral size of the signal bin.
This means that, in the case of the expected narrow signals, it can be reduced by making use of good energy resolution detectors. Second, there is the physics background due to Primakoff production. This grows with the photon coupling. Importantly, as the Primakoff spectrum is continuous, it also grows linearly with worsening energy resolution.
Combining these two effects leads us to the following figure of merit,
\begin{equation}
\label{eq:merit}
f\propto \frac{S}{\sqrt{B}}\propto \frac{\epsilon_{o}\epsilon_{d}\ g^2_{a\gamma}}{\sqrt{\Delta E_d}\sqrt{b a+ g^{4}_{a\gamma}\kappa \epsilon_{o}\epsilon_{d}}}.
\end{equation}
Here, $\epsilon_{o,d}$ are the optics and detector efficiencies, $\Delta E_d$ is the energy resolution
of the detector\footnote{In the numerical calculations in Sects.~\ref{sec:massless} and \ref{sec:sensitivity_massive} the detector resolution is implemented by a sharp signal bin of width $\Delta E_d$ centered around the $^{57}$Fe peak.}, $b$ is the (spectral) background rate
per area, and $a$ is the signal spot area on the detector. $\kappa$ quantifies the Primakoff flux in the $^{57}$Fe signal bin and is implicitly defined in Eq.~\eqref{eq:kappa}.
From this we can identify the parameters to be optimized.
While good energy resolution is critical this should not be offset by too large a background. Similarly, with focusing optics we can reduce the detector area and therefore the background contribution. However, again there is a balance because such X-ray optics may have an efficiency significantly smaller than $1$. We will see all this more concretely when considering explicit setups below.
\subsection{BabyIAXO}
\label{sec:BabyIAXO}
Conceived as an intermediate stage towards the full IAXO experiment, BabyIAXO~\cite{BabyIAXO:2020mzw,Abeln:2020eft} is nevertheless expected to substantially advance the exploration of the axion parameter space.
With its sensitivity, the experiment will be able to study QCD axion models, and to investigate stellar cooling hints and other well motivated sections of the parameter space~\cite{IAXO:2019mpb,DiLuzio:2020wdo}.
The BabyIAXO~ experiment is mainly designed to measure Primakoff axions in the energy range from 1\,keV to 10\,keV with the peak of the solar axion flux spectrum at $\sim$3\,keV.
BabyIAXO~ will consist of two magnetic bores of 10\,m length and 70\,cm diameter, each with an average magnetic field strength of about 2\,T.
Together with newly developed X-ray optics and detector systems providing higher energy resolution and lower background, BabyIAXO~ will be the first helioscope to exceed the sensitivity of the CAST experiment.
The magnet bores of BabyIAXO~ are of similar diameter as those of IAXO (60 cm) and IAXO+ (80 cm),
thus the experience from the optics and detector development for BabyIAXO~ can later be applied to IAXO \cite{BabyIAXO:2020mzw}.
The first detection system for BabyIAXO is chosen to be microbulk micromegas technology. This detector technology has proven background levels as low as
$10^{-7}/(\textrm{keV}\, \textrm{cm}^2\, \textrm{s})$ \,\cite{Aune_2014}
and a high detection efficiency for Primakoff photons $<$10\,keV.
A variety of other detector types, like silicon drift detectors (SDD), metallic magnetic calorimeters (MMC) and transition edge sensors (TES), are also studied for BabyIAXO~ aiming to optimize the energy resolution for precision measurements of the axion spectrum\,\cite{BabyIAXO:2020mzw}.
\subsection{BabyIAXO~configurations for $^{57}$Fe~detection}
For the measurement of $^{57}$Fe axions at 14.4\,keV, the detection efficiency in the baseline BabyIAXO~ configuration is not optimal
and an enhancement of the
detection system is necessary.
Indeed, GEANT4 simulations of the Micromegas detector to be used in the baseline BabyIAXO~ configuration
show that the ionization probability of 14.4\,keV photons at a conversion length of 3\,cm at the given Argon gas mixture and pressure is only around~15\,\%.
Additionally, the current designs of the BabyIAXO~telescopes are not optimized for energies above $10$~keV~\cite{Armengaud:2014gea}.
The expected optics efficiency at 14.4\,keV is a mere $1.3\,\%$ (cf.~``baseline''/BabyIAXO$_0$ configuration in Tab.~\ref{tab:setups}).
In the following we discuss different adaptations to the detector system of BabyIAXO~ for the measurement of 14.4\,keV photons with enhanced efficiency, energy resolution and lower background.
Let us start with a relatively minimal modification.
The sensitivity of the Micromegas detector for 14.4\,keV photons can be enhanced by changing the gas mixture and adjusting the pressure. Gas mixtures with an inert gas of higher Z-value, like Xenon, show a higher conversion efficiency for 14.4\,keV photons, compared to Argon.
A suitable high ionization probability for 14.4\,keV photons of $>$90\,\% should be reachable with a 10\,cm photon conversion length and a Xenon based gas mixture at atmospheric pressure.
Currently, microbulk Micromegas~\cite{microbulk:2010} are developed to have a square active area of 25\,cm~x~25\,cm~\cite{Castel:2019ykn} and the whole bore opening can be covered with a few detector tiles.
As BabyIAXO~ features two bores, one could be used for operation without X-ray optics and this detector could be operated in parallel on the second magnetic bore.
Removing the optics and covering of the whole area improves the detection efficiency, as this avoids the losses from inefficient X-ray optics at this energy.
However, the detector will have a higher background
due to the larger conversion volume
than the smaller micromegas.
This ``no optics'' configuration is denoted as ``no optics''/BabyIAXO$_1$ in Tab.~\ref{tab:setups} and in the figures.
The second detection concept for 14.4\,keV photons at BabyIAXO~ is based on silicon drift detectors (SDD). This detector consists of a thick negative doped layer, that is fully depleted by a negative bias voltage, and positive doped contacts and strips on both sides of the layer. The incoming X-ray radiation generates electrons in the depleted zone that are then drifted towards the anode at the end of the layer.
An SDD detector of 300\,µm thickness may reach 50\,\% sensitivity for 15\,keV photons~\cite{sdd:2003}. This detector type is also considered as an additional detector for the baseline measurements at BabyIAXO~ \cite{BabyIAXO:2020mzw}. Efficient X-ray optics focusing the photons on the small detector structure are mandatory, as SDDs
are only produced in small pixels with a detector size in the range of millimeters.
Preliminary studies showed that the optimization of the X-ray optical system for high energy X-rays is possible,
using multilayer coating techniques~\cite{nustarcoat:2011}.
The Nuclear Spectroscopic Telescope Array (NuSTAR)~\cite{nustar:2013} has already used X-ray optics sensitive to photons in the range of 5-80\,keV.
A dedicated instrument with non-imaging optics can further boost the throughput at 14.4\,keV.
A realistic figure for the optics efficiency is
$\epsilon_0=0.3$, which is the value adopted in our analysis.
In our study here, we assume that the optimized optics will be used in just one of the two BabyIAXO~ bores.
This assumption will be lifted in our analysis of IAXO and IAXO+.
With this included we obtain the ``optimized optics''/BabyIAXO$_2$ configuration (cf.~Tab.~\ref{tab:setups}).
Finally, a third detection concept suitable for the detection of 14.4\,keV photons at BabyIAXO~ is based on Cadmium-Zinc-Telluride (CZT) semiconductors.
The detection principle is similar to silicon-based ionization detectors.
With a higher Z-value compared to silicon, CZT provides a higher ionization probability for the photons and CZT of only 300\,µm thickness have an ionization probability for 14.4\,keV of $>$99\,\%.
Indeed CZT detectors are already used in experiments focusing on the detection of hard X-rays like NuSTAR~\cite{nustar:2013}.
These detectors can reach an energy resolution down to 2\,\% in the relevant energy range \cite{czt_eres:2005},
which makes them optimal for
discrimination of the Primakoff background. However, the good energy resolution comes with the price of a somewhat increased background.
Furthermore, currently it is only possible to produce detectors with an active area of up to 2\,cm~x~2\,cm~\cite{czt:2005}, which again makes an efficient X-ray optic necessary for the use of CZT detectors at BabyIAXO.
Coupled with the same NuSTAR-like optics adopted in BabyIAXO$_2$, this is our ``energy resolution''/BabyIAXO$_3$ configuration.
Notice that, just like in the case of BabyIAXO$_2$, the detector and optics discussed above will be implemented in only one of the two BabyIAXO~ bores (see Tab.~\ref{tab:setups}).
\subsection{IAXO and IAXO+}
The full scale helioscope IAXO
is expected to begin construction during the BabyIAXO~ data-taking period.
It adopts realistic components which will allow a considerably better performance than BabyIAXO~ in the entire mass range.
IAXO+ is a more aggressive setup which will allow to
increase the sensitivity and a deep exploration of physically motivated axion parameters (see. e.g., Refs.~\cite{IAXO:2019mpb,DiLuzio:2021ysg}).
Here, we consider two possible setups for IAXO/IAXO+, summarized in table~\ref{tab:setups}.
The first, indicated as IAXO$_\textrm{b}$,
uses the benchmark configuration parameters
discussed in the most recent IAXO publication~\cite{IAXO:2019mpb}.
In this case, we assume an energy resolution of 2\,\% in the
relevant energy range, which is a very realistic figure for CZT detectors, as discussed in Sec.~\ref{sec:BabyIAXO}.
In addition we assume, somewhat optimistically, that significant improvements in the backgrounds of these detectors can be achieved.
Alternatively, we consider a further configuration, indicated as IAXO$_\textrm{r}$, with
optimized energy resolution.
Though generally realistic, we nevertheless note that the energy resolution in this configuration is non-trivial
and may require new technologies such as microcalorimeters operated at mK temperatures~\cite{Unger2021,osti_21266472}.
Magnetic microcalorimeters (MMCs) are also studied in the IAXO collaboration for precision measurements of the axion spectrum in the energy range $<$10$\,$keV. First measurements have shown an energy resolution of 6.1$\,$eV (FWHM) at the 5.9$\,$keV $^{55}$Fe-peak~\cite{Unger2021}.
To achieve the sufficient energy resolution,
advanced cryogenic setups operating in the milli-Kelvin regime will need to be implemented and a sufficiently low background still needs to be established. We therefore have allowed for a somewhat larger background rate.
For IAXO+ we anticipate also additional improvements of the detection system combined with the enhanced magnetic field, area etc.~outlined in~\cite{IAXO:2019mpb}.
\subsection{Results (massless axions)}
\label{sec:massless}
For our sensitivity study in the coupling space spanned by axion-photon $g_{a\gamma}$ and effective axion-nucleon coupling $g^{\text{eff}}_{aN}$,
we are going to assume that the axion is very light ($\lesssim$20\,meV) or massless. We will briefly comment on the effects of the mass later, in Sec.~\ref{sec:sensitivity_massive}. For the moment, let us nevertheless note that in case of a discovery of an axion with a non-vanishing mass one can employ a suitable amount of buffer gas so that further measurements such as the one of the nucleon coupling can essentially be done as in the case of vanishing mass.
In the massless case, the conversion probability of an axion to a photon $P_{a \rightarrow \gamma}$ is energy independent,
\begin{equation}
P_{a \rightarrow \gamma}=\frac{g_{a\gamma}^2 B^2 L^2}{4}, \label{eq:conversionmassless}
\end{equation}
where $B$ is the magnetic field and $L$ the length of the conversion volume. Other experimental parameters that we used in the calculation include the cross-section of the magnet bores $A$, the total observation time $t$, optics and detector efficiencies, $\epsilon_o$ and $\epsilon_d$, the size of the focal spot $a$, the background level $b$ and the relative energy resolution at 14.4\,keV $r_\omega$.
We have considered several different combinations of parameters which are listed in Tab.~\ref{tab:setups}.
The benchmark values for the BabyIAXO, IAXO and IAXO+ magnets were taken from the most recent IAXO publication~\cite{IAXO:2019mpb}.
Because the FWHM of the Doppler-broadened iron peak is only $\sim5$\,eV, it can be assumed that the whole signal is always in one energy bin.\footnote{This is only approximately true for the configurations IAXO$_{\rm r}$ and IAXO$^+_{\rm r}$ in Tab.~\ref{tab:setups} for which we assumed an energy resolution equal to the FWHM of the emission peak.} We can therefore calculate the expected number of signal events $\mu_{\text{signal}}$ directly from the total flux given in Eq.~\eqref{eq:totalflux}
\begin{align}
\mu_{\text{signal}} &= \Phi_a\ P_{a \rightarrow \gamma}\ A\ t\ \epsilon_o\ \epsilon_d \\
&\propto (g_{a\gamma}g_{aN}^{{\rm eff}})^2.
\end{align}
As already discussed at the beginning of the section, to find the $^{57}$Fe detection sensitivity, we also have to take all possible background sources into account. First, the detector background, which is quantified by the background level $b$ and which can be measured accurately at times when the magnet bores are not pointed at the sun. Second, the tail of the Primakoff spectrum, which may act as an additional background. The expected background events $\mu_{\rm back}$ are therefore given by
\begin{align}
\mu_{\rm back}=& \int_{E^*-\frac{\Delta E_d}{2}}^{E^*+\frac{\Delta E_d}{2}}\left( \frac{\textrm{d}\Phi_a^{\rm P}}{\textrm{d}\omega} \ \epsilon_o \epsilon_d\right)\textrm{d}\omega\ P_{a \rightarrow \gamma} A t \nonumber\\& + b a t \Delta E_d \\
\simeq & \left(g_{a\gamma}^4\kappa \epsilon_o \epsilon_d +b a \right)\Delta E_d t \;. \label{eq:kappa}
\end{align}
The Primakoff flux $\Phi_a^{\rm P}$ as well as the efficiencies $\epsilon_o$ and $\epsilon_d$ are in general functions of $\omega$. In case of a sufficiently small energy resolution $\Delta E_d$, we can average over the energy and describe the Primakoff background using the constant $\kappa$, as done in Eq.\eqref{eq:kappa}.
If the Primakoff background to the $^{57}$Fe-peak is detectable at 14.4\,keV, it is clear that there will be a much stronger Primakoff signal at smaller energies and we will have measured $g_{a\gamma}$ very precisely. Therefore, we either know the expected contribution from Primakoff axions to the number of background events $\mu_{\text{back}}$ or it is negligible compared to the intrinsic detector background.
The $p$-value of the Poisson-distributed observed number of counts $k$ in the signal bin is given by
\begin{align}
p=&\sum_{n=k}^{\infty} \frac{\mu_{\textrm{back}}^ne^{-\mu_{\textrm{back}}}}{n!}\ ,
\end{align}
where $k\in \mathbb{N}$. In order to find the sensitivity in parameter space, we have to calculate the expectation value of $p$ assuming a Poisson distribution for $k$ with mean $\mu = \mu_{\textrm{back}} + \mu_{\textrm{signal}}$ for each possible value of the two couplings. Note that $\mu_{\textrm{back}}$ is the sum of the usual detector background and Primakoff background.
We regard the experiment as sensitive to a set of axion couplings if the expected $p$-value, $\langle p\rangle_k$, is smaller than 0.05.\footnote{A measurement with $ p=0.05$ would strictly speaking only amount to a 2$\sigma$ anomaly. Nevertheless, it is common to define the sensitivity in this way because it coincides with the expected exclusion limits in the case of a null result.} The resulting sensitivity curves are plotted in Fig.~\ref{fig:sensitivity}.
\begin{table*}[!t]
\renewcommand{\arraystretch}{1.15}
\centering
\begin{tabular}{l|cccc|cc|cc}
\textbf{} & \multicolumn{4}{c|}{\textbf{BabyIAXO~}} &\multicolumn{2}{c|}{\textbf{IAXO}} & \multicolumn{2}{c}{\textbf{IAXO+}} \\
\hline
& \footnotesize{baseline} &\makecell{\footnotesize{no} \\ \footnotesize{optics}}& \makecell{\footnotesize{optimized} \\ \footnotesize{optics}} &\makecell{\footnotesize{high energy} \\ \footnotesize{resolution}}
& \makecell{\footnotesize{low} \\ \footnotesize{background}}& \makecell{\footnotesize{high energy} \\ \footnotesize{resolution}}&\makecell{\footnotesize{low} \\ \footnotesize{background}}& \makecell{\footnotesize{high energy} \\ \footnotesize{resolution}}\\
Label& \footnotesize{BabyIAXO$_0$} & \footnotesize{BabyIAXO$_1$} & \footnotesize{BabyIAXO$_2$} & \footnotesize{BabyIAXO$_3$} & \footnotesize{IAXO$_{\rm b}$} &\footnotesize{IAXO$_{\rm r}$}& \footnotesize{IAXO$^+_{\rm b}$}&\footnotesize{IAXO$^+_{\rm r}$}\\
\hline \hline
$B$\ [T] & 2 & 2 & 2 & 2 & 2.5 & 2.5 & 3.5 & 3.5 \\
$L$\ [m] & 10 & 10 & 10 & 10 & 20 & 20 & 22 & 22 \\
$A$\ [m$^2$] & 0.77 & 0.38 & 0.38 & 0.38 & 2.3 & 2.3 & 3.9 & 3.9\\
$t$\ [year] & 0.75 & 0.75 & 0.75 & 0.75 & 1.5 & 1.5 & 2.5 & 2.5 \\
$b$\ [$\frac{1}{\textrm{keV} \textrm{cm}^2 \textrm{s}}$] & $10^{-7}$ & $10^{-6}$ & $10^{-7}$ & $10^{-5}$ & $10^{-8}$ & $10^{-6}$ & $10^{-9}$ & $10^{-6}$ \\
$\epsilon_d$ & 0.15 & 0.9 & 0.5 & 0.99 & 0.99 & 0.99 & 0.99 & 0.99 \\
$\epsilon_0$ & 0.013 & 1 & 0.3 & 0.3 & 0.3 & 0.3 & 0.3 & 0.3 \\
$a$\ [cm$^2$] & 0.6 & 3800 & 0.3 & 0.3 & 1.2 & 1.2 & 1.2 & 1.2 \\
$r_\omega=\frac{\Delta E_{d}}{14.4\,{\rm keV}}$ & 0.12 & 0.12 & 0.12 & 0.02 & 0.02 & $\frac{5}{14400}$& 0.02 & $\frac{5}{14400}$ \\
\end{tabular}
\caption{
List of experimental parameters adopted for all helioscope configurations which are considered in Figs.~\ref{fig:sensitivity} and~\ref{fig:massive}. As usual $B$ is the magnetic field of the helioscope, $L$ its length and $A$ the area. $t$ is the time that the helioscope is pointed at the sun.
For these parameters we use values based on~\cite{IAXO:2019mpb}.
As already mentioned below Eq.~\eqref{eq:merit} $\epsilon_{o,d}$ are the efficiencies of the optics and detector, $b$ is the spectral background rate per detector area, and $r_{\omega}$ is the relative spectral resolution of the detector.
Setup~BabyIAXO$_0$
is the baseline BabyIAXO, BabyIAXO$_1$ is a version without optics, BabyIAXO$_{2,3}$ assume optics optimized for the 14.4\,keV line, with BabyIAXO$_3$ including also a good energy resolution.
In addition we show parameters from more advanced setups of IAXO and IAXO+.
}
\label{tab:setups}
\end{table*}
\begin{figure}[t]
\centering
\includegraphics[width=1\linewidth]{Fe57babyIAXO.pdf}
\caption{Model independent prediction for the sensitivity to the axion couplings for light axions ($\lesssim20$\,meV).
The different regions refer to the setups presented in Tab.~\ref{tab:setups}.
The dark red region is the solar bound discussed in the text (cf.~Eq.~\eqref{eq:solar_bound}).
The dark blue region represents the latest CAST exclusion
regions from searches for the Primakoff flux~\cite{CAST:2017uph} and the $^{57}$Fe~ peak~\cite{CAST:2009jdc}, which we have rescaled to match the updated axion flux from $^{57}$Fe~ transitions. The dashed horizontal green line indicates the expected sensitivity to the pure Primakoff flux. The supernova limit, Eq.~\eqref{eq:SNbound}, $g^{\rm eff}_{aN}\lesssim 1.1\times 10^{-9}$ is not shown.}
\label{fig:sensitivity}
\end{figure}
Since we have assumed massless axions for these sensitivity estimates, we do not show the parameter space of DFSZ and M1 models (cf.~Sec.~\ref{sec:models}). These
models require large masses -- $m_a\gtrsim 100\,$meV -- at those couplings, and
helioscopes quickly lose sensitivity above $\sim$\,20\,meV (cf.~Fig.~\ref{fig:massive}).
This problem could be eased by filling the helioscope bore with a buffer gas~\cite{vanBibber:1988ge}.
Our analysis shows the potential of BabyIAXO~ to study areas
of the parameter space, well beyond the solar bound and the region probed by CAST.
The different green shaded areas show the experimental potential for different configurations,
summarized in Tab.~\ref{tab:setups}.
The most efficient setups for BabyIAXO~ are the ones with optimized optics (labeled BabyIAXO$_{2,3}$ in Tab.~\ref{tab:setups}).
As evident from the table, these setups allows for an enormous reduction of the total background by limiting the focal spot area $a$ by $\sim$\,4 orders of magnitude with respect to the no-optics solution.
Adopting this setting,
BabyIAXO~ would be able to extend its detection potential also to regions of the parameter space below the Primakoff sensitivity, for
$g_{aN}^{\rm eff} \gtrsim 10^{-7}$.
It is, therefore, possible, at least in principle, that BabyIAXO~ could discover axions through the $^{57}$Fe~ channel, before the Primakoff flux can be detected.
If, on the other hand, axions have couplings in the green shaded area above the Primakoff sensitivity line (dashed green),
one might have the opportunity to extract both couplings and derive information about the underlying axion model.
The more optimistic IAXO and IAXO+ configurations can explore an even larger area of parameter space. A noteworthy feature of these exclusion curves is their behaviour at different values of $g_{a\gamma}$. At values of $g_{a\gamma}\lesssim 10^{-11}\ \textrm{GeV}^{-1}$ the detector background dominates over the Primakoff background. Therefore the figure of merit in Eq.~\eqref{eq:merit} becomes
$f\propto \frac{\epsilon_{o}\epsilon_{d}\ g^2_{a\gamma}}{\sqrt{\Delta E_db a}}$. With the parameters given in Table~\ref{tab:setups}, the configurations with minimized background slightly outperform the ones with optimized energy resolution in this regime (cf.~Fig.~\ref{fig:sensitivity}). However, at $g_{a\gamma}\gtrsim 10^{-11}\ \textrm{GeV}^{-1}$ the Primakoff background starts to play a role and eventually dominates. In this regime the figure of merit is given by $f\propto \sqrt{\frac{\epsilon_{o}\epsilon_{d} }{\Delta E_d\kappa }}$. The detector background becomes negligible and the configurations with optimized energy resolution are significantly more sensitive than the ones with minimized background. Therefore, the ideal detector for the $^{57}$Fe line crucially depends on the value of $g_{a\gamma}$. If BabyIAXO~ detects Primakoff axions, a detector with good energy resolution may be required to supress the Primakoff background to the $^{57}$Fe line. If, on the other hand, BabyIAXO~ only puts a stronger bound on $g_{a\gamma}$, the energy resolution becomes less important and the low background detectors may be advantageous.
\subsection{Effects of a finite axion mass}
\label{sec:sensitivity_massive}
On the production side, the axion mass only becomes relevant at scales of the solar temperature $\sim$\,keV. Therefore, we can safely regard the solar axion flux as independent of the mass.
However, on the detection side, a finite axion mass can cause decoherence between the photon and axion wave functions inside the magnet bores and lead to a signal suppression.
The full expression for the conversion probability $P_{a \rightarrow \gamma}$ of axions into photons in the helioscope reads~\cite{Raffelt:1987im,vanBibber:1988ge}
\begin{equation}
\label{eq:conversionmassive}
P_{a \rightarrow \gamma}=\frac{g_{a\gamma}^2 B^2 L^2}{4} \times \frac{2(1-\cos(qL))}{(qL)^2} \ ,
\end{equation}
where $q$ is the transferred momentum given by $m_a^2/(2\omega)$ and $\omega$ is the energy of the axion.
The suppression becomes relevant for masses above $\sim$\,20\,meV, the exact value depending on the length of the magnet.
The decoherence effect can be compensated at the cost of some absorption by feeding a buffer gas into the bores~\cite{vanBibber:1988ge,CAST:2008ixs}.
This is why the sensitivity study to effectively massless (i.e.\ $m_a \lesssim 20\ \mathrm{meV}$) axions in the previous section serves as a good benchmark.
Nevertheless, we also want to explicitly investigate the sensitivity
to massive axions without a buffer gas.
To do this, we assume that the background from Primakoff axions is negligible and that instead the detector background dominates. In this case, the background is independent of any axion properties and the signal depends on the product of the two couplings, $g_{a\gamma} g^\mathrm{eff}_{aN}$, as well as the mass.
The statistical analysis is equivalent to the one in the massless case.
From the expected signal and background we calculate the expected $p$-value and draw the sensitivity curves where $p$ is expected to be smaller than 0.05.
The results for a selection of viable setups are plotted in Fig.~\ref{fig:massive}.
The regions shaded in yellow indicate the coupling relations for the DFSZ, M1 and $\mathcal{T}_2^{(u)}$ models (cf.~Sec.~\ref{sec:models}).
The sensitivity curves are very similar to typical helioscope exclusion plots in the coupling vs.\ mass plane with two noteworthy exceptions. Because the $^{57}$Fe line is highly energetic at 14.4\,keV, the transferred momentum $q$ is smaller than for axions of the same mass from other solar processes. As a result the decoherence effect only becomes relevant at slightly higher masses in comparison to -- for instance -- Primakoff axions. To illustrate this effect, we have plotted the expected sensitivity of the IAXO$^+_{\rm b}$ setup with a decoherence factor from a Primakoff spectrum as a dashed black line in Fig.~\ref{fig:massive}. Furthermore, the oscillations of the form factor for large $qL$ are clearly visible in the $^{57}$Fe exclusion lines while they are washed out in the case of the broadband Primakoff spectrum.
\begin{figure}
\centering
\includegraphics[width=1\linewidth]{Fe_bound.pdf}
\caption{Model independent prediction of the IAXO sensitivity to the $^{57}$Fe peak,
assuming that axions are produced only through the axion coupling to nucleons. In contrast to Fig.~\ref{fig:sensitivity}, we directly show the sensitivity of the various setups to the coupling combination $g_{a\gamma} g_{aN}^\textrm{eff}$ under the assumption that the Primakoff background is negligible.
The oscillations at higher masses are due to the form factor in the conversion probability.
For comparison we show the effect of decoherence for a Primakoff spectrum as a dashed black line.
The dark blue region represents the rescaled
CAST result~\cite{CAST:2009jdc}.
The dark yellow region indicates the parameter space expected for the DFSZ model. In brighter shades of yellow, we show the flavor non-universal DFSZ models M1 and $\mathcal{T}_2^{(u)}$.
The dashed yellow
line, on the other hand, shows the
expected coupling for a nucleophilic QCD axion model of the kind presented in Ref.~\cite{Darme:2020gyx}, with $n=3$.
All models with $n>3$ would be already accessible to BabyIAXO.
Note that the experimental sensitivity estimates here do not assume the use of a buffer gas, which would extend the sensitivity to higher masses.}
\label{fig:massive}
\end{figure}
\section{Discussion and conclusion}
\label{sec:Conclusion}
In this work, we have presented the first dedicated investigation of the BabyIAXO~ and IAXO potential to detect 14.4\,keV axions from $^{57}$Fe~ transitions in the sun.
The analysis is based on a recent reevaluation of the matrix elements for the
nuclear transition~\cite{Avignone:2017ylv} and an updated solar model~\cite{Vinyoles:2016djt}.
We have carefully considered different realistic setups for the detectors and optics
to evaluate which combination gives the best performance for 14.4\,keV axions.
The configuration parameters are listed in Tab.~\ref{tab:setups}.
Our results, summarized in Fig.~\ref{fig:sensitivity} and~\ref{fig:massive}, show that already BabyIAXO~ will be able to study a large section of interesting axion parameter space, well beyond the region accessed by CAST, particularly
in the configurations with an optimized X-ray optics.
The potential will be greatly improved with IAXO and IAXO+.
The figures also show representative QCD axion models.
The yellow band in Fig.~\ref{fig:massive} represents the parameter area spanned by generalized DFSZ models~\cite{DiLuzio:2017ogq,DiLuzio:2021ysg} (QCD axion models with two Higgs doublets plus a singlet scalar field, allowing for flavor non-universality).
There, we present the specific examples of the classical DFSZ axion model, and of the M1 and $\mathcal{T}_2^{(u)}$ models discussed in Sec.~\ref{sec:models}.
We are not showing the well known KSVZ axion model since
an accidental cancellation reduces its effective coupling to nucleons relevant in the $^{57}$Fe~ transition (see Sec.~\ref{sec:models}).
Although
BabyIAXO~ is expected to have enough sensitivity to
explore large sections of the parameter space for these models~\cite{DiLuzio:2021ysg},
a sizable
axion flux from $^{57}$Fe~ transitions requires large axion masses,
where BabyIAXO~ loses sensitivity.
This problem can be eased with the use of a buffer gas~\cite{vanBibber:1988ge}, a technique already tested in CAST~\cite{CAST:2017uph}.
The sensitivities shown in the figures do not account for this option in BabyIAXO~ nor in its scaled up versions.
A dedicated study may show if a buffer gas may allow to probe the M1 or other DFSZ-like models through the $^{57}$Fe~ line in the near future.
Less minimal models for QCD axions may present larger couplings to nucleons and be better accessible through the $^{57}$Fe~ channel.
For example, the nucleophilic QCD axion models presented in Ref.~\cite{Darme:2020gyx} (and shown in Fig.~\ref{fig:massive}), have exponentially large couplings to nucleons\footnote{In model A of Ref.~\cite{Darme:2020gyx},
one has
\begin{align}
g_{a\gamma}g_{aN}^{\rm eff}\sim
\frac{2^{2n}\alpha_{\rm em}}{2\pi\,f_a}
\frac{m_n}{f_a}
\sim 2^{2n+2}\cdot 10^{-17}\,
\left(\frac{m_a}{\rm eV}\right)^2 \,{\rm GeV^{-1}}
\end{align}
where $n+1$ is the number of Higgs doublets in the model.
The parameter $n$ is constrained to $n\lesssim 50$
in order to avoid sub-Planckian Landau poles \cite{DiLuzio:2017pfr}.
}
and are efficiently produced in $^{57}$Fe~ transitions even at lower axion mass.
As shown in Fig.~\ref{fig:massive}, practically, the entire class of these models will be accessible already to BabyIAXO, even without the need for a buffer gas.
One should nevertheless keep in mind that most of the region shown in the figures is in tension with astrophysical considerations, in particular, SN1987A (cf.~Eq.~\eqref{eq:SNbound}). As these are affected by their own uncertainties (e.g. relying on a single supernova event) it would nevertheless be comforting to have independent confirmation in more controlled setups.
In addition, the IAXO+ setup shown in Fig.~\ref{fig:sensitivity} approaches a level of sensitivity comparable to Eq.~\eqref{eq:SNbound}. This shows a pathway to pushing beyond the astrophysical limits.
In conclusion, helioscopes of the next generation may offer a unique chance to probe an interesting range of the $g_{aN}^{\rm eff}$-$g_{a\gamma}$ parameter space.
While this potential is expected to be greatly improved with its scaled up versions, IAXO and IAXO+,
already BabyIAXO~ will have enough sensitivity to detect axions with couplings to nucleons over an order of magnitude below the solar bound (see, Fig.~\ref{fig:sensitivity}).
Furthermore, our analysis shows that already BabyIAXO, especially if equipped with optimized optics,
has the potential to detect through the $^{57}$Fe~ channel axions too weakly coupled to photons to give a sizable Primakoff flux (region below the dashed green line in Fig.~\ref{fig:sensitivity}).
In a more likely scenario, a detection of axions through $^{57}$Fe~ will be accompanied by a (larger) signal from Primakoff axions,
allowing to extract important information about its couplings to both photons and nucleons.
Ending on an optimistic outlook, we note that discovery of an axion and its nucleon-coupling induced lines could perhaps also shed light on properties of the sun.\footnote{See~\cite{Jaeckel:2019xpa} for an opportunity to use axions with electron couplings to measure the solar metallicity.}
For example, the strong temperature dependence may make this a good way to measure the sun's core temperature.
\section*{Acknowledgements}
We would like to thank José Crespo López-Urrutia for helpful comments and discussions. LT is funded by the Graduiertenkolleg \textit{Particle physics beyond the Standard Model} (GRK 1940).
The work of M.G.~is supported by funding from a grant provided by the Fulbright U.S.~Scholar Program and by a grant from the Fundación Bancaria Ibercaja y Fundación CAI.
M.G.~thanks the Departamento de Física Teórica and the Centro de Astropartículas y Física de Altas Energías (CAPA) of the Universidad de Zaragoza for hospitality during the completion of this work. I.G.I and J. G. acknowledge support from the European Research Council (ERC) under the European Union’s Horizon 2020
research and innovation programme, grant agreement ERC-2017-AdG788781 (IAXO+), as
well as the Spanish Agencia Estatal de Investigación under grant PID2019-108122GB.
J.J.~is grateful to be part of the ITN HIDDeN which is supported within the European Union’s Horizon 2020 research and innovation programme grant agreement No 860881.
Part of this work was performed under the auspices of the US Department of Energy by Lawrence Livermore National Laboratory under Contract No. DE-AC52-07NA27344.
\bigskip
|
\section{Additional Ablation Study}
We discuss further ablation results regarding: (1)~The effect of the entropy coefficient $\alpha$ on the performance as well as convergence speed for learning our visual policy. (2)~The effect of using our proposed weight initialization strategy for the asymmetric actor-critic framework. (3)~The effect of behavioral cloning for trajectory smoothing over the motion planner augmented RL trajectories. (4)~The dependency of our method on the success of MoPA-RL.
\subsection{Effect of varying the entropy coefficient $\alpha$}
\label{sec:alpha}
In Section \mysecref{sec:visual_policy_training}, we discuss the role of tuning the entropy coefficient $\alpha$ for maintaining the trade-off between entropy maximization as well as reward maximization for the SAC objective. For our visual policy learning using asymmetric actor-critic framework, we mainly focus on the exploration achieved by our state-based agent using MoPA-RL \citep{yamada2020mopa}. This helps us leverage our state agent's experience for providing guided exploration rather than exploring the entire state space again for visual policy learning. Therefore, we focus on the reward maximization objective by choosing a very small value of $\alpha$.
\mytbref{tab:mopa-rl-log-alpha} reports the $\log (\alpha)$ values for our learned state-agent using MoPA-RL. For accelerating our visual policy learning, we use smaller $\alpha$ values to further emphasize on reward maximization. In \myfigref{fig:ablation_alpha}, we show that using smaller values of $\alpha$ assists in faster convergence and improving sample efficiency as compared to using higher $\alpha$ values. Moreover, smaller $\alpha$ values consistently converge for all the environments whereas for larger $\alpha$ values, we can see unstable performance varying across different tasks. We also observe that large values of $\alpha$ undergo large perturbations during training making learning unstable whereas smaller values of $\alpha$ remain stable throughout training, as shown in \myfigref{fig:ablation_alpha_change}.
\begin{table}[ht]
\centering
\begin{tabular}{|c|c|c|c|}
\toprule
& Sawyer Push & Sawyer Lift & Sawyer Assembly \\
\midrule
$\log (\alpha)$ & -0.63 & -2.7 & -0.29 \\
\bottomrule
\end{tabular}
\vspace{4mm}
\caption{
Values of the final $\log( \alpha)$ after MoPA-RL training for all three environments.}
\label{tab:mopa-rl-log-alpha}
\end{table}
\begin{figure*}[ht]
\begin{minipage}{\textwidth}
\centering
\begin{subfigure}[t]{0.3\textwidth}
{\includegraphics[width=\textwidth]{images/ablation/SawyerPush-Abl-Alpha-ASR.pdf}}
\caption{Sawyer Push}
\end{subfigure}
\begin{subfigure}[t]{0.3\textwidth}
{\includegraphics[width=\textwidth]{images/ablation/SawyerLift-Abl-Alpha-ASR.pdf}}
\caption{Sawyer Lift}
\end{subfigure}
\begin{subfigure}[t]{0.3\textwidth}
{\includegraphics[width=\textwidth]{images/ablation/SawyerAssembly-Abl-Alpha-ASR.pdf}}
\caption{Sawyer Assembly}
\end{subfigure}
\end{minipage}
\caption{Ablation curves for showing the effect of the entropy coefficient $\alpha$ for visual policy learning via asymmetric actor-critic framework. As the trends show, smaller $\alpha$ values lead to a stable and faster training for all three environments.}
\label{fig:ablation_alpha}
\end{figure*}
\begin{figure*}[ht]
\begin{minipage}{\textwidth}
\centering
\begin{subfigure}[t]{0.3\textwidth}
{\includegraphics[width=\textwidth]{images/ablation/SawyerPush-Abl-Alpha-Change.pdf}}
\caption{Sawyer Push}
\end{subfigure}
\begin{subfigure}[t]{0.3\textwidth}
{\includegraphics[width=\textwidth]{images/ablation/SawyerLift-Abl-Alpha-Change.pdf}}
\caption{Sawyer Lift}
\end{subfigure}
\begin{subfigure}[t]{0.3\textwidth}
{\includegraphics[width=\textwidth]{images/ablation/SawyerAssembly-Abl-Alpha-Change.pdf}}
\caption{Sawyer Assembly}
\end{subfigure}
\end{minipage}
\caption{Ablation curves for showing the change in the entropy coefficient $\alpha$ during training for visual policy learning via asymmetric actor-critic framework. As the trends show, larger $\alpha$ values fluctuate during training leading to unstable learning, whereas smaller values of $\alpha$ do not show large variations throughout training for all three environments.}
\label{fig:ablation_alpha_change}
\end{figure*}
\subsection{Effect of weight initialization}
\label{sec:weight_init}
As discussed in Section \mysecref{sec:evaluation_exp}, we note that our weight initialization technique for the actor-critic networks before visual policy learning significantly aids in achieving optimal performance and also facilitates improving learning speed. In \myfigref{fig:ablation_init}, we see that without appropriate initialization for the asymmetric critic using our state agent and the visual actor using behavioral cloning, the agent achieves minimal or no success rate across all environments up till 1.2M environment steps. On the other hand, our method with initialization converges to a success rate of $\approx$ 1.0 much earlier than 1.0M environment steps for all three tasks.
\begin{figure*}[ht]
\begin{minipage}{\textwidth}
\centering
\begin{subfigure}[t]{0.3\textwidth}
{\includegraphics[width=\textwidth]{images/ablation/SawyerPush-Abl-Init-ASR.pdf}}
\caption{Sawyer Push}
\end{subfigure}
\begin{subfigure}[t]{0.3\textwidth}
{\includegraphics[width=\textwidth]{images/ablation/SawyerLift-Abl-Init-ASR.pdf}}
\caption{Sawyer Lift}
\end{subfigure}
\begin{subfigure}[t]{0.3\textwidth}
{\includegraphics[width=\textwidth]{images/ablation/SawyerAssembly-Abl-Init-ASR.pdf}}
\caption{Sawyer Assembly}
\end{subfigure}
\end{minipage}
\caption{Learning curves comparing our method with and without actor-critic initialization for all three environments. Without initialization, the agent achieves minimal or no success rate, whereas it quickly learns to achieve optimal performance with weight initialization across all the tasks.}
\label{fig:ablation_init}
\end{figure*}
\subsection{Behavioral cloning for trajectory smoothing}
As mentioned in \mysecref{sec:traj_smoothing}, we use BC to smooth out jittery motion planner augmented trajectories and store them in an augmented expert replay buffer $\mathcal{R}_e$. In \myfigref{fig:bc_vs_mopa_traj}, we qualitatively compare BC and MoPA-RL's random rollouts with the same start and goal positions. As we can visualize, the BC trajectory path is smoother compared to the trajectory obtained via the MoPA-RL policy. This is indeed significant for a refined transition from a state-based policy to a visual policy in terms of performance as well as convergence speed and sample efficiency (see \myfigref{fig:learning_curves_comparison}).
\begin{figure*}[ht]
\begin{minipage}{\textwidth}
\centering
\begin{subfigure}[t]{0.3\textwidth}
{\includegraphics[width=\textwidth]{images/ablation/bc_mopa_trajs_sawyer_push.png}}
\caption{Sawyer Push}
\end{subfigure}
\begin{subfigure}[t]{0.3\textwidth}
{\includegraphics[width=\textwidth]{images/ablation/bc_mopa_trajs_sawyer_lift.png}}
\caption{Sawyer Lift}
\end{subfigure}
\begin{subfigure}[t]{0.3\textwidth}
{\includegraphics[width=\textwidth]{images/ablation/bc_mopa_trajs_sawyer_assembly.png}}
\caption{Sawyer Assembly}
\end{subfigure}
\end{minipage}
\caption{Visualization of end-effector position for a randomly chosen rollout comparing the smoothness of behavioral cloning trajectories with respect to MoPA-RL trajectories with the same start state (marked in \textcolor{green}{green}) and final end-effector position (marked in \textcolor{red}{red}) for all three tasks.}
\label{fig:bc_vs_mopa_traj}
\end{figure*}
\subsection{Sub-optimal training for MoPA-RL}
To analyze the dependence of our approach on the success of the MoPA-RL training in stage one, we experiment with sub-optimal models of MoPA-RL. We train on lesser environment interactions and then use the critic for weight initialization, followed by our asymmetric visual agent training. We train for the same number of environment steps as Table 1 results, i.e., 2M environment steps. We report the results in \mytbref{tab:suboptimal_mopa_push} and \mytbref{tab:suboptimal_mopa_lift} for Sawyer Push and Sawyer Lift tasks, respectively.
As reported in \mytbref{tab:suboptimal_mopa_push} and \mytbref{tab:suboptimal_mopa_lift}, we observe that although our method benefits from initialization using the state-based agent compared to random initialization, it does not heavily depend on the state-based agent's success. This shows that even when MoPA-RL cannot completely solve the task, our framework for distilling the state-based policy into the visual policy achieves optimal performance.
\begin{table*}[t]
\small
\begin{subtable}[ht]{1.0\textwidth}
\begin{center}
\begin{tabular}{|l|c|c|c|c|c|c|}
\toprule
& \multicolumn{6}{c|}{\textbf{Environment Steps}} \\
& \multicolumn{3}{c|}{\textbf{0.2M}} & \multicolumn{3}{c|}{\textbf{1.0M}} \\
& \textbf{ASR~$\uparrow$} & \textbf{AEL~$\downarrow$} & \textbf{ADR~$\uparrow$} &
\textbf{ASR~$\uparrow$} & \textbf{AEL~$\downarrow$} & \textbf{ADR~$\uparrow$} \\
\midrule
MoPA-RL \citep{yamada2020mopa} & 69.2 & 102.7 & 41.0 & 98.0 & 105.4 & 54.5 \\
BC-Visual & 93.2 & 137.4 & 37.1 & 98.8 & 57.4 & 97.2 \\
Ours & 100.0 & 34.4 & 108.2 & 100 & 32.0 & 110.8 \\
\bottomrule
\end{tabular}
\end{center}
\caption{Sawyer Push}
\label{tab:suboptimal_mopa_push}
\end{subtable}
\begin{subtable}[ht]{1.0\textwidth}
\begin{center}
\begin{tabular}{|l|c|c|c|c|c|c|c|c|c|}
\toprule
& \multicolumn{9}{c|}{\textbf{Environment Steps}} \\
& \multicolumn{3}{c|}{\textbf{0.5M}} & \multicolumn{3}{c|}{\textbf{0.52M}} & \multicolumn{3}{c|}{\textbf{1.0M}} \\
& \textbf{ASR~$\uparrow$} & \textbf{AEL~$\downarrow$} & \textbf{ADR~$\uparrow$} &
\textbf{ASR~$\uparrow$} & \textbf{AEL~$\downarrow$} & \textbf{ADR~$\uparrow$} & \textbf{ASR~$\uparrow$} & \textbf{AEL~$\downarrow$} & \textbf{ADR~$\uparrow$} \\
\midrule
MoPA-RL \citep{yamada2020mopa} & 41.4 & 153.8 & 21.0 & 60.6 & 138.3 & 29.2 & 95 & 109.2 & 52.7\\
BC-Visual & 48 & 189.4 & 14.1 & 55.2 & 146.6 & 24.2 & 63 & 109.4 & 34.8 \\
Ours & 99.0 & 42.0 & 101.6 & 99.4 & 42.9 & 101.0 &
99.0 & 42.0 & 101.7 \\
\bottomrule
\end{tabular}
\end{center}
\caption{Sawyer Lift}
\label{tab:suboptimal_mopa_lift}
\end{subtable}
\caption{Average success rate (ASR), episode length (AEL), and discounted return (ADR) of our method compared with MoPA-RL and BC using sub-optimal and optimal MoPA-RL models. Even with sub-optimal training, when our state-based agent does not learn to completely solve the task, our method for distillation into a visual policy achieves high performance and successfully solves the task in lesser steps (smaller AEL).}
\end{table*}
\section{Implementation Details}
\label{sec:implementation}
\subsection{Network architecture and hyperparameter selection}
For training our state-based agent in step one using MoPA-RL \citep{yamada2020mopa}, we use three layered fully connected neural networks with 256 hidden units and ReLU activation for actor $\pi_\phi$ and critic $\mathcal{Q}_{\psi}$ networks and MP implementation similar to \citep{yamada2020mopa} using the RRT-Connect algorithm. For our asymmetric agent's critic, we retain the critic network same as $\mathcal{Q}_{\psi}$ and the visual actor $\pi_\theta$ is a 3-layered convolution neural network followed by three fully connected layers with 256 hidden units and Leaky ReLU activation and another set of fully connected layers outputting the mean and the standard deviation of the Gaussian distribution over an action space. The behavioral cloning agent shares the network architecture with the asymmetric visual actor $\pi_\theta$. We specify other hyperparameter details for SAC and behavioral cloning training in \mytbref{tab:hyperparameters_sac}, \mytbref{tab:hyperparameters_bc} and \mytbref{tab:hyperparameters_env}. \\
\begin{minipage}{\textwidth}
\begin{minipage}[t]{0.42\textwidth}
\vspace{0.2em}
\makeatletter\def\@captype{table}
\begin{tabular}{|c|c|}
\toprule
Parameter & Value \\
\midrule
Optimizer & Adam \\
Learning rate & 1e-5\\
Discount factor ($\gamma$) & 0.99 \\
Replay buffer size & $10^6$ \\
Image size & 32x32 \\
Minibatch size & 256 \\
Nonlinearity & ReLU \\
No. of expert trajectories & 100 \\
Network update per env. step & 1 \\
\bottomrule
\end{tabular}
\vspace{1mm}
\caption{SAC hyperparameters shared across all environments}
\label{tab:hyperparameters_sac}
\end{minipage}
\hspace{1.0em}
\begin{minipage}[t]{0.4\textwidth}
\vspace{0.2em}
\makeatletter\def\@captype{table}
\begin{tabular}{|c|c|}
\toprule
Parameter & Value \\
\midrule
Optimizer & Adam \\
Learning rate & 5e-4\\
\# observation-action pairs & $\approx$ 1M \\
Train-test split & 9:1 \\
Image size & 32x32 \\
Minibatch size & 512 \\
Nonlinearity & LeakyReLU \\
Scheduler step size & 5 \\
Scheduler decay rate & 0.99 \\
\bottomrule
\end{tabular}
\vspace{1mm}
\caption{Behavioral cloning hyperparameters shared across all environments}
\label{tab:hyperparameters_bc}
\end{minipage}
\end{minipage}
\begin{table}[ht]
\centering
\vspace{4mm}
\begin{tabular}{|c|c|c|c|}
\toprule
& Sawyer Push & Sawyer Lift & Sawyer Assembly \\
\midrule
Action dimension & 7 & 8 & 7 \\
Reward Scale & 0.8 & 0.5 & 1.0 \\
\bottomrule
\end{tabular}
\vspace{4mm}
\caption{Environment-specific parameters for SAC training}
\label{tab:hyperparameters_env}
\end{table}
\subsection{Choosing the optimal model for behavioral cloning}
For behavioral cloning training, we use validation success rate for choosing the optimal model weights rather than using loss on the test dataset. As seen in the \myfigref{fig:bc_training}, we see that a lower error on the test set does not necessarily provide the best validation accuracy. Therefore, we pick the model on the first such epoch which receives 100\% validation accuracy. For measuring the validation accuracy after each training epoch, we use 5 episodes each on 6 randomly chosen seeds. We train our BC agent for a total of 140 epochs and select the earliest model with the best validation accuracy for BC trajectory smoothing as well as asymmetric visual actor's weight initialization. We also report the training time and number of epochs (for optimal convergence) required to account for the additional steps required in learning the BC agent in \mytbref{tab:bc_epochs}.
\begin{figure*}[ht]
\begin{minipage}{\textwidth}
\centering
\begin{subfigure}[t]{0.45\textwidth}
{\includegraphics[width=\textwidth]{images/ablation/SawyerAssembly-Abl-Optimal-BC-Loss.pdf}}
\caption{Loss curves for behavioral cloning}
\end{subfigure}
\begin{subfigure}[t]{0.45\textwidth}
{\includegraphics[width=\textwidth]{images/ablation/SawyerAssembly-Abl-Optimal-BC-VSR.pdf}}
\caption{Validation Accuracy for behavioral cloning}
\end{subfigure}
\end{minipage}
\caption{Learning curves for behavioral cloning showing loss on the train set, loss on the test set as well as the validation accuracy per epoch for Sawyer Assembly environment.}
\label{fig:bc_training}
\end{figure*}
\begin{table}[ht]
\centering
\begin{tabular}{|c|c|c|c|}
\toprule
& Sawyer Push & Sawyer Lift & Sawyer Assembly \\
\midrule
Number of epochs & 12 & 13 & 24 \\
Wall-clock time & 30 & 39 & 120 \\
\bottomrule
\end{tabular}
\vspace{4mm}
\caption{Number of epochs and wall-clock time (in minutes) for training a BC agent used for weight initialization and BC Smoothing.}
\label{tab:bc_epochs}
\end{table}
\section{Domain Randomization}
For domain randomization, we randomly sample a variation for the texture, color and lighting conditions for each episode during training. An illustration of the domain randomization samples for each environment is shown in \myfigref{fig:dr_envs}. Training our method with domain randomized environments helps in learning robust policies that are capable of transfer to unseen domains and distractor objects as shown in Section \myfigref{sec:policy_transfer}.
\begin{figure}
\centering
\begin{subfigure}{0.31\textwidth}
\includegraphics[width=\textwidth]{images/push_dr.pdf}
\caption{Sawyer Push}
\label{fig:dr_env_sawyer_push}
\end{subfigure}
\hfill
\begin{subfigure}{0.31\textwidth}
\includegraphics[width=\textwidth]{images/lift_dr.pdf}
\caption{Sawyer Lift}
\label{fig:dr_env_sawyer_lift}
\end{subfigure}
\hfill
\begin{subfigure}{0.31\textwidth}
\includegraphics[width=\textwidth]{images/assembly_dr.pdf}
\caption{Sawyer Assembly}
\label{fig:dr_env_sawyer_assembly}
\end{subfigure}
\caption{Illustration for random samples of domain randomization for each environment,}
\label{fig:dr_envs}
\end{figure}
\section{Environment Details}
\label{sec:appendix_env}
We simulate all of our 3D environments using MuJoCo physics engine~\citep{todorov2012mujoco}. Subsequently, we describe the observation details for all three environments. We maintain all other specifics regarding the reward functions, success criteria as well as environment intial states same as explained in \citet{yamada2020mopa}.
\subsection{Sawyer Push}
The task is to reach an object placed inside a box and push it towards the goal region using a Sawyer Arm. We define the positions of the goal, object and the end-effector as $p_\text{goal}$, $p_\text{obj}$ and $p_\text{eef}$ respectively.
\paragraph{Observations.} Each state observation $s_{t}$ comprises of the joint state $(\sin\theta, \cos\theta)$, angular joint velocity, quaternion end-effector coordinates, $p_\text{eef}$ the position of the object, the position of the goal $p_\text{goal}$, distance between the end-effector and the object, and the distance between the object and goal position. This acts as an input to the critic in the asymmetric framework.
The visual observation $o_{t}$ comprises of the simulated image corresponding to $s_{t}$ and the robot joint space information. This is used as an input to the actor in the asymmetric framework for learning the visual policy.
\subsection{Sawyer Lift}
For Sawyer Lift, the task is to reach an object placed inside a box, grasp it using the gripper hand and then lift it up above the box height.
\paragraph{Observations.} The state observation $s_{t}$ comprises of the joint state $(\sin\theta, \cos\theta)$, angular joint velocity, the position of the object position and quaternion, end-effector coordinates, $p_\text{eef}$ , the position of the goal $p_\text{goal}$, and distance between the end-effector and the object. This acts as an input to the critic in the asymmetric framework.
For an input to the asymmetric actor for visual policy learning, the visual observation $o_{t}$ comprises of the simulated image corresponding to $s_{t}$ and the robot joint space information.
\subsection{Sawyer Assembly}
In Sawyer Assembly, the task is to assemble the fourth leg of a tabletop in its gripper to its corresponding hole where the other three legs are already in place. This needs to be done while avoiding collisions with the obstructions posed by the other three table legs that are already assembled since the table is free to move under collisions.
\paragraph{Observations.} The state observation $s_t$ comprises of the joint state $(\sin\theta, \cos\theta)$, angular velocity, the position of the hole, the head and tail positions of the leg in the gripper hand and its quaternion.
The input to the asymmetric actor for visual policy learning comprises of the simulated image $o_t$ and the robot joint state information.
\section{Conclusion}
In this paper, we introduce a two-step distillation method for learning manipulation tasks in obstructed environments in the visual domain. In step one, we use a MP-augmented RL policy as the state-based expert and subsequently learn a visual BC agent from it, removing the motion planner dependency. In step two, we learn a vision-based agent using an asymmetric actor-critic framework. This step is further expedited via proper weight initialization, BC trajectory-guided RL training, and entropy coefficient tuning, making our method highly sample-efficient. Our visual policy combined with domain randomization demonstrates successful zero-shot transfer to unseen environments with new visual domains and distractors. Beyond zero-shot transfer, fine-tuning the policy in the real world by learning a vision-based critic or applying simulation-to-real techniques~\citep{Zhang2021PolicyTA} is our definite future work to realize simulation-to-real transfer.
\section{Experiments}
\label{sec:experiments}
In this paper, we propose to distill a motion planner augmented policy into a visual control policy for complex manipulation tasks in obstructed environments.
In our experiments, we aim to answer the following questions: (1) Does IL efficiently learn policies for obstructed environments in image space? Moreover, is naively combining IL with RL sufficient for solving complex tasks? (2) Is distillation better than directly learning a visual policy using MP? (3) Does our approach, MoPA-PD, efficiently learn a visual policy using prior state-based exploration knowledge? (4) Is our visual policy capable of domain transfer and robust to unseen distractors?
\subsection{Environments}
\begin{figure}
\centering
\begin{subfigure}{0.3\textwidth}
\includegraphics[width=\textwidth]{images/sawyerpush_env.png}
\caption{Sawyer Push}
\label{fig:env_sawyer_push}
\end{subfigure}
\hfill
\begin{subfigure}{0.3\textwidth}
\includegraphics[width=\textwidth]{images/sawyerlift_env.png}
\caption{Sawyer Lift}
\label{fig:env_sawyer_lift}
\end{subfigure}
\hfill
\begin{subfigure}{0.3\textwidth}
\includegraphics[width=\textwidth]{images/sawyerassembly_env.png}
\caption{Sawyer Assembly}
\label{fig:env_sawyer_assembly}
\end{subfigure}
\caption{
Manipulation tasks in obstructed environments from \citet{yamada2020mopa}. (a)~Sawyer Push: Push the red cube in the box to the green goal. (b)~Sawyer Lift: Lift out the can inside the box. (c)~Sawyer Assembly: Insert the table leg in the hole in the table top.
}
\label{fig:envs}
\end{figure}
We evaluate our approach on three obstructed environments (see \myfigref{fig:envs}) from \citet{yamada2020mopa}, simulated using the MuJoCo physics engine~\citep{todorov2012mujoco}. We use a 32x32 image as a visual observation.
\begin{itemize}[leftmargin=2em]
\item \textbf{Sawyer Push:} A Rethink Sawyer robot arm with 7~DoF, initially located outside the box, needs to reach an object placed inside the box and push it to the goal location within the box.
\item \textbf{Sawyer Lift:} The Sawyer arm must reach an object placed inside a box, grasp, and lift it outside the box, avoiding collisions. The Sawyer arm is always initialized outside the box.
\item \textbf{Sawyer Assembly:} The Sawyer arm needs to assemble the fourth leg (already attached to the gripper) of a table into the vacant hole while avoiding collisions to other three legs. This environment is built upon the IKEA furniture assembly environment~\citep{lee2021ikea}.
\end{itemize}
For Sawyer Push and Sawyer Assembly, the agent receives a reward proportional to the distance between the end-effector and object, or the object and goal position when the distance is less than $\epsilon$. We use $\epsilon=0.1$ for Sawyer Push and $\epsilon=0.3$ for Sawyer Assembly while the distance between initial and goal state is around $1.2$. An episode is considered successful when the distance between the cube and goal (Sawyer Push) or the peg-head and goal (Sawyer Assembly) is less than 0.05. For Sawyer Lift, we use the reward function similar to \citet{fan2018surreal}, where a reward is defined for all three intermediate stages of the task: \emph{reach}, \emph{grasp} and \emph{lift}. A bonus reward signal of $150$ is received upon successful task completion in all the environments. We refer readers to \mysecref{sec:appendix_env} for more details.
\begin{table*}[t]
\small
\centering
\renewcommand{\tabcolsep}{4pt}
\begin{tabular}{|l|c|c|c|c|c|c|c|c|c|}
\toprule
& \multicolumn{3}{c|}{\textbf{Sawyer Push}} & \multicolumn{3}{c|}{\textbf{Sawyer Lift}} & \multicolumn{3}{c|}{\textbf{Sawyer Assembly}}\\
& \textbf{ASR~$\uparrow$} & \textbf{AEL~$\downarrow$} & \textbf{ADR~$\uparrow$} &
\textbf{ASR~$\uparrow$} & \textbf{AEL~$\downarrow$} & \textbf{ADR~$\uparrow$} &
\textbf{ASR~$\uparrow$} & \textbf{AEL~$\downarrow$} & \textbf{ADR~$\uparrow$} \\
\midrule
MoPA-RL & 98.2 & 111.0 & 52.7 & 95.0 & 109.2 & 52.7 & 99.8 & 63.3 & 82.4 \\
\hline
BC-Visual & 99.4 & 118.0 & 46.9 & 62.0 & 108.8 & 34.6 & 97.0 & 115.1 & 50.4 \\
Asym. SAC & 0.0 & 250.0 & 0.0 & 0.0
& 250.0 & 0.2 & 0.0 & 250.0 & 3.4 \\
MoPA-Asym. SAC & 0.0 & 250.0 & 0.0 & 0.0 & 250.0 & 0.0 & 0.0 & 250.0 & 0.0 \\
CoL & 0.0 & 250.0 & 0.0 & 29.8 & 173.3 & 9.7 & 0.0 & 250.0 & 0.0 \\
CoL w/ BC smoothing & 0.0 & 250.0 & 0.0 & 0.0 & 250.0 & 5.1 & 0.0 & 250.0 & 0.0 \\
Ours w/o BC smoothing & \textbf{100.0} & 34.0 & 108.7 & \textbf{99.4} & 43.6 & 100.7 & 0.0 & 250.0 & 0.0 \\
Ours & \textbf{100.0} & \textbf{32.0} & \textbf{110.8} & 99.0 & \textbf{42.0} & \textbf{101.7} & \textbf{100.0} & \textbf{61.7} & \textbf{84.5} \\
\bottomrule
\end{tabular}
\caption{
Average success rate (ASR), episode length (AEL), and discounted return (ADR) of our method and baselines averaged over five seeds. Each method is evaluated after 3M environment steps. MoPA-RL~\citep{yamada2020mopa} is the expert agent trained with state-based policy. All baselines below horizontal line are trained with asymmetric actor-critic for fair comparison. Maximum horizon is 250 for all tasks.
}
\label{table_baselines}
\end{table*}
\subsection{Baselines}
We compare our method with the following baselines to evaluate the merits of our approach:
\begin{itemize}[leftmargin=2em]
\item \textbf{MoPA-RL} \citep{yamada2020mopa}: A MP-augmented policy, for learning large displacement actions with MP and smaller actions with RL. This policy also serves as our state-based expert agent.
\item \textbf{Visual Behavioral Cloning (BC-Visual)}: A behavioral cloning policy trained on the image-action pairs collected from MoPA-RL trajectories.
\item \textbf{Asym. SAC} \citep{pinto2018asymmetric}: An asymmetric actor-critic method, trained using SAC where the critic is learned in the state space and the actor is learned in the image space.
\item \textbf{MoPA Asym. SAC}: MoPA-RL policy~\citep{yamada2020mopa} learned using the asymmetric framework. Note that this method still uses a motion planner with an augmented action space. This method is a direct attempt at learning a visual policy using a MP-augmented framework~\citep{yamada2020mopa}.
\item \textbf{CoL}~\citep{Goecks2019IntegratingBC}: A policy learned using the Cycle-of-Learning framework, which is the state-of-the-art algorithm for learning from demonstrations (LfD).
\item \textbf{CoL (w BC Smoothing)}~\citep{Goecks2019IntegratingBC}: A policy similar to CoL~\citep{Goecks2019IntegratingBC}, with BC for trajectory smoothing.
\item \textbf{Ours (w/o BC Smoothing)}: A policy learned using our approach described in \mysecref{sec:experiments} without BC trajectory smoothing, i.e., we directly use MoPA-RL trajectories in $\mathcal{R}_e$.
\end{itemize}
\begin{figure}[t]
\centering
\begin{subfigure}[t]{0.31\textwidth}
\includegraphics[width=\textwidth]{images/evaluation/Rebuttal-SawyerPush-ADR.pdf}
\caption{Sawyer Push}
\end{subfigure}
\quad
\begin{subfigure}[t]{0.31\textwidth}
\includegraphics[width=\textwidth]{images/evaluation/Rebuttal-SawyerLift-ADR.pdf}
\caption{Sawyer Lift}
\end{subfigure}
\quad
\begin{subfigure}[t]{0.31\textwidth}
\includegraphics[width=\textwidth]{images/evaluation/Rebuttal-SawyerAssembly-ADR.pdf}
\caption{Sawyer Assembly}
\end{subfigure}
\caption{Learning curves of our method compared to baselines. All methods are trained for 3M environment steps.
For the methods that require MoPA-RL, we train MoPA-RL for 1M steps and then train the methods for 2M steps (total 3M) for a fair comparison.
Our method solves all the tasks with the highest average discounted return. Our method w/o BC smoothing can solve two tasks with slower convergence compared to Ours, but fails on Sawyer Assembly.}
\label{fig:learning_curves_comparison}
\end{figure}
\subsection{Evaluation}
\label{sec:evaluation_exp}
We compare our approach against baselines on the following evaluation metrics averaged over five random seeds and $100$ unseen episodes per seed:
\begin{itemize}[leftmargin=2em]
\item \textbf{Average Success Rate (ASR)} is the average number of successful episodes.
\item \textbf{Average Episode Length (AEL)} is the average length of successful episodes.
\item \textbf{Average Discounted Return (ADR)} is the average discounted sum of rewards $\sum_{t=0}^{T-1} \gamma^t R(s_t, a_t)$, with $T$ being the episode horizon. An episode completed in more time steps has a lower discounted reward, due to exponentially discounted reward with $\gamma=0.99$.
\end{itemize}
\subsection{Results}
\paragraph{Comparisons with baselines} \myfigref{fig:learning_curves_comparison} illustrates the learning curves of our method and all other baselines using discounted rewards against the number of training steps. As per the trend, our method is far more sample-efficient and outperforms all the baselines. Asym. SAC~\citep{pinto2018asymmetric} fails to directly learn a visual policy for Sawyer Lift and Sawyer Assembly in obstructed environments, where exploration is hard. Moreover, MoPA Asym. SAC, which is a direct attempt at learning a visual policy while using MP, does not successfully learn to solve the tasks. CoL~\citep{Goecks2019IntegratingBC} which has a straightforward combination of RL and BC objective for actor optimization receives partial rewards for some tasks at around 1.5-2M steps, much slower than our method's convergence.
As reported in \mytbref{table_baselines}, BC achieves decent performance for Sawyer Push and Assembly, but not for the Sawyer Lift task. However, it achieves high AEL and low ADR for all tasks, showing its inefficiency to solve the task fast. This is because BC is bound to perform as good as the expert demonstrations (here from MoPA-RL) at its best. In contrast, our method optimizes trajectories beyond expert signals and achieves significantly lower AEL and higher ADR compared to baselines.
\paragraph{Ablation on BC trajectory smoothing}
Our method without BC smoothing has higher variance across seeds and converges slower than our method for two tasks (see \myfigref{fig:learning_curves_comparison}). For Sawyer Assembly, it does not learn to solve the task at all. In short, BC smoothing of the motion planner trajectories is important for our method to work across all environments. This is because the motion planner based trajectories are usually jittery and non-smooth. Using behavioral cloning refines each transition, thereby making the entire trajectories much smoother and help in efficiently training the RL agent.
\paragraph{Ablation on weight initialization}
We compare our method with and without actor-critic initialization in appendix, \myfigref{fig:ablation_init} and do not observe any episode success until 1.2M environment steps for the latter. This shows the importance of our proposed weight initialization for learning in the visual domain. We further elaborate the ablation results in \mysecref{sec:weight_init}.
\paragraph{Ablation on entropy coefficient tuning} We examine the effect of different values of $\alpha$ on the trade-off between entropy maximization and reward maximization for the SAC objective as shown in \mysecref{sec:alpha}. Compared to higher $\alpha$ values, smaller values of $\alpha$ improve the sample efficiency during visual policy learning. Since we utilize prior knowledge from the state-based agent, we use a smaller alpha to exploit the previously acquired knowledge instead of exploring the entire state space again.
\subsection{Policy Transfer to Different Domains}
\label{sec:policy_transfer}
\begin{figure}[t]
\centering
\begin{subfigure}{0.24\textwidth}
\includegraphics[width=\textwidth]{images/dr_env_push.png}
\caption{Domain randomization}
\label{fig:sim2real:env0}
\end{subfigure}
\hfill
\begin{subfigure}{0.24\textwidth}
\includegraphics[width=\textwidth]{images/sawyer_push_original_env_start.PNG}
\caption{Original environment}
\label{fig:sim2real:env1}
\end{subfigure}
\hfill
\begin{subfigure}{0.24\textwidth}
\includegraphics[width=\textwidth]{images/dr_sawyer_push.PNG}
\caption{Scenario 1}
\label{fig:sim2real:env2}
\end{subfigure}
\hfill
\begin{subfigure}{0.24\textwidth}
\includegraphics[width=\textwidth]{images/dr_sawyer_push_2.PNG}
\caption{Scenario 2}
\label{fig:sim2real:env3}
\end{subfigure}
\caption{
Illustration of environments for domain transfer. (a)~Domain randomized environments are used for training; For testing, we use (b)~original environment without randomization, (c)~Scenario~1 with distractors such as unseen cubes (blue and magenta), walls, floor, and furniture, and (d)~Scenario~2 similar to Scenario~1 along with variation in size of the table and texture of the box and table.
}
\label{fig:sim2real_simulation}
\end{figure}
\begin{table}[t]
\small
\begin{center}
\renewcommand{\tabcolsep}{4pt}
\begin{tabular}{|l|c|c|c|c|c|c|c|c|c|}
\toprule
& \multicolumn{3}{c|}{\textbf{Sawyer Push}} & \multicolumn{3}{c|}{\textbf{Sawyer Lift}} & \multicolumn{3}{c|}{\textbf{Sawyer Assembly}}\\
& \textbf{ASR~$\uparrow$} & \textbf{AEL~$\downarrow$} & \textbf{ADR~$\uparrow$} &
\textbf{ASR~$\uparrow$} & \textbf{AEL~$\downarrow$} & \textbf{ADR~$\uparrow$} &
\textbf{ASR~$\uparrow$} & \textbf{AEL~$\downarrow$} & \textbf{ADR~$\uparrow$} \\
\midrule
Original env & 99.7 & 39.1 & 103.1 & 99.3 & 37.3 & 106.5 & 100.0 & 49.7 & 93.4 \\
Scenario 1 & 100.0 & 40.4 & 102.2 & 96.7 & 38.3 & 103 & 100.0 & 47.9 & 94.8 \\
Scenario 2 & 99.3 & 43.4 & 99.04 & 97.3 & 37.0 & 104.7 & 100.0 & 47.7 & 95.0 \\
\bottomrule
\end{tabular}
\end{center}
\caption{Evaluation metrics for domain transfer to unseen environments, illustrated in \myfigref{fig:sim2real_simulation}.}
\label{tab:sim_transfer_table}
\end{table}
In this experiment, we learn our policy with domain randomization (DR) to verify its domain transfer capabilities. DR is a promising approach for modelling transferable policies using a multitude of variations in simulation during training~\citep{Tobin2017DomainRF,Sheckells2019UsingDD}, which makes the policy invariant to changes that are insignificant for the task. For this work, we randomize the simulation environment using different textures, colors, and lighting conditions for training as shown in \myfigref{fig:sim2real_simulation} and \myfigref{fig:dr_envs}. We then test our DR policy on three unseen scenarios, as illustrated in \myfigref{fig:sim2real_simulation} comprising of (1)~the original environment without any randomization; (2)~Scenario~1 with realistic background distractions (e.g.,\ ~furniture, walls, and floor) and unseen distractor cubes (blue and magenta) near the target cube (red); and (3)~Scenario~2 with additional changes in the texture of the table and box, and different size of the table.
Our policy attains more than 96\% ASR in all three unseen scenarios for all the tasks (see \mytbref{tab:sim_transfer_table}). This illustrates the robustness of our proposed framework in learning transferable visual policies, robust to unseen distractors in a sample-efficient manner.
\section{Introduction}
Solving complex manipulation tasks in obstructed environments is a challenging problem in deep reinforcement learning (RL) since it requires precise object interactions as well as collision-free movement across obstacles.
To tackle this problem, prior works~\citep{yamada2020mopa, Xia2020ReLMoGenLM,10.1007/978-3-030-61616-8_24} have proposed to combine the strengths of motion planning (MP) and RL -- safe collision-free maneuvers of MP and sophisticated contact-rich interactions of RL, demonstrating promising results.
However, MP requires access to the geometric state of an environment for collision checking, which is often not available in the real world, and is also computationally expensive for a real-time control. To deploy such agents in realistic settings, we need to resolve the dependency on the state information and costly computation of MP, such that the agent can perform a task in the visual domain.
To this end, we propose a two-step distillation framework, motion planner augmented policy distillation (MoPA-PD), that transfers the state-based motion planner augmented RL policy (MoPA-RL~\citep{yamada2020mopa}) into a visual control policy, thereby removing the motion planner and the dependency on the state information.
Concretely, our framework consists of two stages: (1)~visual behavioral cloning (BC~\citep{pomerleau1989alvinn}) with trajectories collected using the MoPA-RL policy and (2)~vision-based RL training with the guidance of smoothed trajectories from the BC policy.
The first step, visual BC, removes the dependency on the motion planner and the resulting visual BC policy generates smoother behaviors compared to the motion planner's jittery behaviors. Then, in the second step, we further improve the visual policy using asymmetric actor-critic RL~\citep{pinto2018asymmetric} directly from the image observations to overcome sub-optimality in the MoPA-RL policy through self-exploration.
For efficient vision-based RL training, our method leverages the experience of the BC policy and the state-based critic from MoPA-RL by initializing the critic network with the MoPA-RL critic and the actor network through BC pre-training; BC-trajectory guided RL training using a separate expert replay buffer; tuning the entropy coefficient to encourage exploitation (i.e.,\ maximizing reward) over exploration (i.e.,\ maximizing entropy).
In summary, our contributions are as follows:
\begin{itemize}[leftmargin=2em]
\item We propose a novel framework to learn a visual control policy in cluttered scenarios by distilling a state-space policy that uses a motion planner into an image-space policy by removing the motion planner dependency.
\item Our asymmetric visual policy leaning ensures high sample-efficiency using weight initialization followed by BC trajectory-guided RL with entropy coefficient tuning.
\item The distilled visual policy is capable of zero-shot domain transfer to unseen environments with visual domain randomization during training.
\item Our method outperforms the state-of-the-art methods in terms of success rate, sample efficiency, and path length on three manipulation tasks in obstructed environments.
\end{itemize}
\begin{figure}[t]
\centering
\includegraphics[width=1\textwidth]{images/overview.pdf}
\caption{
Our two-step framework distills the state-based MoPA-RL agent into a visual control policy. In stage 1, we learn a visual BC policy using MoPA-RL trajectories, thereby ensuring smooth trajectories for the expert replay buffer $\mathcal{R}_e$, while removing the motion planner. In stage 2, we learn a BC trajectory-guided vision-based RL agent with an asymmetric actor-critic algorithm using both the expert replay buffer $\mathcal{R}_e$ and agent replay buffer $\mathcal{R}_\pi$. To improve sample efficiency, we initialize the actor and the critic of the vision-based agent with the weights of the BC policy and the MoPA-RL critic, respectively.
}
\label{fig:model}
\end{figure}
\section{Method}
\label{sec:methodology}
Our goal is to solve complex manipulation tasks in obstructed environments with visual inputs. While MP-based methods can efficiently solve such tasks, they are restricted to learning a state-based policy which is difficult to transfer to real environments. Moreover, sampling-based motion planning incurs significant computational costs, making it hard to integrate them into real-time controllers. Thus, we propose a method that distills an MP-augmented state-based agent into a visual policy, removing the MP and state dependency. We formally define our problem and introduce the MP-augmented RL in \mysecref{sec:preliminaries}. Then, we describe our two-step visual distillation approach in \mysecref{sec:visual_policy_training}.
\subsection{Preliminaries}
\label{sec:preliminaries}
\paragraph{Problem formulation}
We formulate our problem as a Markov Decision Process (MDP) $M$ defined by a tuple $(\mathcal{S}, \mathcal{O}, \mathcal{A}, P, R, \rho_0, \gamma)$ that consists of a state space $\mathcal{S}$, partial observation space $\mathcal{O}$ (visual inputs corresponding to states), action space $\mathcal{A}$, transition function $P:\mathcal{S}\times\mathcal{A}\times\mathcal{S} \rightarrow \mathbb{R}$, reward function $R$, initial state distribution $\rho_0$, and discount factor $\gamma$. The agent is represented as a policy $\pi(a_t | o_t)$, which takes an action $a_t$ under a visual observation $o_t$ and receives a reward $r_t=R(s_t, a_t)$, with the state transitioning to $s_{t+1}$. The goal of the agent is to maximize the expected discounted sum of rewards $\mathbb{E}_{(s,a) \sim \pi} \sum_{t=0}^{T-1} \gamma^t r_t$, where $T$ is the episode horizon.
\paragraph{Motion Planner-Augmented RL (MoPA-RL)}
To tackle manipulation tasks in cluttered environments, \citet{yamada2020mopa} proposed to incorporate MP and RL via an MP-augmented MDP $\tilde{M}=(\mathcal{S}, \tilde{\mathcal{A}}, \tilde{P}, \tilde{R}, \rho_0, \gamma)$ that consists of the state space $\mathcal{S}$, enlarged action space $\tilde{\mathcal{A}}$ augmenting the direct action space $\mathcal{A}$ with the MP action space $\hat{\mathcal{A}}$,\footnote{
MoPA-RL~\citep{yamada2020mopa} defines the direct action space as $\mathcal{A} = [-\Delta q_\text{step}, \Delta q_\text{step}]^d$, where $\Delta q_\text{step}$ is the maximum joint displacement, and the enlarged MP-augmented action space as $\tilde{\mathcal{A}} = [-\Delta q_\text{MP}, \Delta q_\text{MP}]^d$, where $\Delta q_\text{MP}$ is the motion planner action limit with $\Delta q_\text{MP} > \Delta q_\text{step}$. In MoPA-RL, actions in the direct action space $a \in \mathcal{A}$ are directly applied as a joint torque while other large actions $\hat{a} \in \hat{\mathcal{A}} = \tilde{\mathcal{A}} \setminus \mathcal{A}$ invoke motion planning. We refer the readers to \citet{yamada2020mopa} for more details.
}
augmented transition function $\tilde{P}:\mathcal{S}\times\tilde{\mathcal{A}}\times\mathcal{S} \rightarrow \mathbb{R}$, augmented reward function $\tilde{R}(s, \tilde{a})$, initial state distribution $\rho_0$, and discount factor $\gamma$.
MoPA-RL~\citep{yamada2020mopa} learns a policy $\pi_\phi(\tilde{a}_t |s_t)$ on the augmented MDP $\tilde{M}$ with an off-policy RL algorithm, SAC~\citep{Haarnoja2018SoftAO}. Precisely, given a state $s_t$, the policy predicts an action $\tilde{a}_t$, which is defined as a robot joint displacement $\Delta q_t$.
If the action lies within the direct action space $\mathcal{A}$, it is directly executed by the controller as the agent performs sophisticated and contact-rich manipulations. However, when the actions are larger in magnitude (i.e.~$\tilde{a} \in \hat{\mathcal{A}}$), the probability of collisions in the presence of obstacles increases. Therefore, a sampling-based motion planner, RRT-Connect~\citep{Kuffner2000RRTconnectAE}, is called to realize such actions with large joint displacements by computing collision-free paths. This allows the agent to efficiently explore obstructed environments while avoiding collision~\citep{yamada2020mopa}.
\subsection{Motion Planner Augmented Policy Distillation}
\label{sec:visual_policy_training}
Despite the advantages of MoPA-RL, its applicability to the real world is limited due to the high computational cost of MP and dependency on fully observable environment states.
To this end, we propose a visual distillation method, motion planner augmented policy distillation (MoPA-PD), that learns an image-based control policy (without MP) from a state-based MP-augmented policy by leveraging its rollouts and the learned state-based critic as a guidance.
Concretely, our proposed method consists of two stages, as illustrated in \myfigref{fig:model}. Given a state-based MoPA-RL agent, we first use BC to train a vision-based actor with trajectories collected from the MoPA-RL policy (\mysecref{sec:traj_smoothing}), and then we further improve the vision-based agent via BC trajectory-guided asymmetric actor-critic RL~\citep{pinto2018asymmetric} with the visual BC actor and MoPA-RL critic (\mysecref{sec:visual_rl}).
\subsubsection{Stage 1: Visual Behavioral Cloning and Trajectory Smoothing}
\label{sec:traj_smoothing}
The MoPA-RL policy is learned with access to complete information about the environment states and also utilizes MP for executing large action steps without collisions. In this paper, we aim to utilize demonstrations from this planner-based policy and distill it into a visual control policy, thereby deducting expensive MP computations and training the actor in visual domain. With the learned MoPA-RL policy, we first collect multiple transitions $d_i$ in low-level (direct) action space and store them into the MoPA-RL demonstration dataset $\mathcal{D}_\text{mp} = \{d_1, d_2, d_3, ...\}$, where $d_i=(s_i, o_i, a_i, r_i, s_{i+1})$. Then, we train a visual BC actor $\pi_\theta(a_t | o_t)$ using observation-action pairs $(o_i, a_i)$ from the dataset $\mathcal{D}_\text{mp}$ by minimizing the mean squared error.
Distilling the MoPA-RL trajectories using BC not only enables the BC actor to work directly on visual inputs but also reduces jerky motion planning behaviors of MoPA-RL, which occur due to motion planner's priority on obstacle avoidance over trajectory smoothness~\citep{Zhu2015ACO}.
By removing unnecessary, jerky movements through visual BC, the resulting trajectories become smoother and even shorter, and thus help learning consistent and smooth motions.
However, the BC actor often fails when it encounters states not seen during training due to covariate shift~\citep{Ross2011ARO}. To improve the robustness of the policy, we further train the visual BC actor using RL with additional environment interactions. In other words, the BC actor can be a good starting policy for visual RL training and the trajectories collected from the BC actor can effectively guide exploration.
\subsubsection{Stage 2: Vision-based RL with Asymmetric Actor-Critic}
\label{sec:visual_rl}
In the second stage, we further train the visual policy using RL directly on image observations to enhance the robustness of the policy and overcome sub-optimality in the planner-based policy.
For efficient RL training, we adopt an asymmetric actor-critic architecture~\citep{pinto2018asymmetric}, where the actor acts based on environment images and robot joint states while the critic learns from environment states.
This architecture is motivated by learning transferable visual policies, which do not require state information during inference but benefit from them while learning in simulation. Our training procedure comprises of the following components:
\paragraph{Weight initialization of actor and critic networks}
Initializing actor and critic networks with suitable weights can accelerate RL training by providing more optimal rollouts from the actor and informative learning signals from the critic compared to the randomly initialized actor and critic~\citep{Cruz2017PretrainingNN, Anderson2015FasterRL}.
Especially when learning from pixels, suitable initialization can significantly improve sample efficiency~\citep{Rajeswaran-RSS-18, Goecks2019IntegratingBC}.
To this end, we propose to leverage the state-based agent's experience by utilizing its weights for initializing our asymmetric visual agent's networks, thereby bringing their initial distributions closer.
Thus, the visual actor network $\pi_\theta(a_t | o_t)$ is initialized with the weights of the visual BC actor learned in \mysecref{sec:traj_smoothing}.
We also initialize the critic $Q_\psi(s_t, a_t)$ with the critic of MoPA-RL. Note that the action spaces of two critic networks are different, but initialization is still feasible since the action space of MoPA-RL is a superset of the direct action space.
\paragraph{BC trajectory-guided asymmetric RL}
After initializing the actor and critic networks, we collect agent rollouts into the agent replay buffer $\mathcal{R}_\pi$. As an RL algorithm, we adopt Soft Actor-Critic (SAC~\citep{Haarnoja2018SoftAO}), an off-policy model-free RL algorithm for continuous control. We optimize our asymmetric visual policy $\pi_{\theta}$ by maximizing the following objective:
\begin{equation}
J(\pi_\theta) = \sum_{t=0}^T \mathbb{E}_{(\mathbf{s_t},\mathbf{a_{t}}) \sim \rho_{\pi_\theta}}\left[Q_{\psi} \left(\mathbf{s}_{t}, \mathbf{a}_{t}\right)+\alpha \mathcal{H}\left(\pi_\theta\left(\cdot \mid \mathbf{o}_{t}\right)\right)\right],
\label{eq:sac_objective}
\end{equation}
where the temperature parameter $\alpha$ balances between exploration and exploitation using entropy $\mathcal{H}$.
RL training with a BC-initialized policy often quickly deviates from the original policy. To prevent this problem and guide exploration during RL, we update our RL agent not only with agent trajectories but also with BC policy (i.e. expert) trajectories~\citep{Hester2018DeepQF}. Thus, we collect smoothed trajectories from our visual BC policy, $(s_i, o_i, a_i, r_i, s_{i+1})$, in an expert replay buffer $\mathcal{R}_e$. Then, for each RL training iteration, we sample 1:3 transitions from $\mathcal{R}_e$ and $\mathcal{R}_\pi$, respectively.
A separate expert replay buffer ensures guided exploration from the expert and also circumvents catastrophic forgetting after weight initialization. Note that we use the BC trajectories $\mathcal{R}_e$ instead of the MoPA-RL data $\mathcal{D}_\text{mp}$ because jittery motion planning paths in the MoPA-RL data make RL training difficult and sub-optimal.
\paragraph{Entropy coefficient tuning}
The performance of SAC is known to be sensitive to $\alpha$~\citep{Wang2020MetaSACAT}.
Since our asymmetric agent already starts with the well-trained BC actor and MoPA-RL critic, we initialize $\alpha$ to values lower than the final $\alpha$ obtained in MoPA-RL. This is to ensure that with prior knowledge in the state-based agent, the visual agent focuses more on maximizing rewards over exploration. We describe the hyperparameter choice and implementation details in \mysecref{sec:alpha} and \mysecref{sec:implementation}.
In summary, we propose a two-step visual distillation method for a state-based MoPA-RL agent using visual BC followed by BC trajectory-guided vision-based RL to remove the dependency on MP and environment states. For efficient RL training, our method leverages weight initialization of the actor and critic, the expert replay buffer of BC smoothed trajectories, and entropy coefficient tuning.
\section{Related Work}
The objective of our work is to learn complex manipulation tasks in realistic and obstructed environments. Motion planner augmented RL methods~\citep{yamada2020mopa, Xia2020ReLMoGenLM,10.1007/978-3-030-61616-8_24} have shown promising results in solving complex tasks in obstructed environments by combining MP and RL. However, these methods cannot be deployed in real-world settings due to their dependency on state information. To remove this dependency, state-based policies must be transferred to the visual domain.
A typical approach to distill state-based policies into vision-based policies is behavioral cloning (BC~\citep{pomerleau1989alvinn}). However, BC often suffers when encountering states unseen during training~\citep{Ross2011ARO}. Therefore, prior works have proposed various methods to improve BC agents using offline RL~\citep{Lee2021BPP} and online RL~\citep{Jeong2020SSSR}. However, these works do not involve the distillation of motion planners.
Several recent efforts have been made towards distilling MP algorithms into neural motion planners using learning-based methods~\citep{Qureshi2018DeeplyIN,Qureshi2019MotionPN,Bency2018TowardsNN,Pfeiffer2017FromPT,Jurgenson2019HarnessingRL}. Most of these efforts can be broadly categorized into imitation learning (IL) and RL paradigms. For IL, supervised learning on MP trajectories has been used to learn neural network-based planners~\citep{Qureshi2019MotionPN,Bency2018TowardsNN,Pfeiffer2017FromPT,Jurgenson2019HarnessingRL,strudelnmp2020}. However, the performance of such supervised learning approaches are limited by the collected dataset and demonstrator's performance.
To discover better policies with additional interactions, various off-policy RL methods have been studied for superseding motion planners with neural network policies~\citep{Jurgenson2019HarnessingRL,ha2020multiarm,Luo2021SelfImitationLB}.
These approaches utilize expert trajectories stored in the replay buffer for guided exploration, which leads to better policies than experts~\citep{Nair2018OvercomingEI, Hester2018DeepQF, Vecerk2017LeveragingDF, Rajeswaran-RSS-18}.
However, these works learn either in unobstructed environments~\citep{Nair2018OvercomingEI, Hester2018DeepQF, Vecerk2017LeveragingDF, Rajeswaran-RSS-18} or on simpler tasks not involving object manipulation~\citep{Jurgenson2019HarnessingRL}. They also assume fully observable environments and learn policies in state space. In this work, we focus on complex manipulation tasks in obstructed environments using visual observations.
Moreover, most offline RL approaches~\citep{Nair2020AWAC,Endrawis2021ESF} in sparse reward scenarios are coupled with hindsight experience replay (HER~\citep{Andrychowicz2017HindsightER}).
They empirically show that not using HER significantly hurts the agent's performance
~\citep{Jurgenson2019HarnessingRL, pinto2018asymmetric, Luo2021SelfImitationLB, Nair2018OvercomingEI}. However, HER is designed to efficiently train goal-conditioned policies for multi-goal RL, which makes it unsuitable for tasks with multiple sequential steps without explicitly conditioning on goals. In contrast, our method successfully learns composite manipulation tasks in cluttered environments without using HER for tackling sparse rewards.
|
\section{Introduction}
\label{sec:intro}
Two-frequency observations of the Viking spacecraft in 1976 (lander
and orbiter) were once used for analysis of radio signal propagation delays
due to solar wind plasma
\citep{Callahan1977,Muhleman1981}.
Also, there were plans for dual-frequency ranging for the Mars Express
spacecraft~\citep{Patzold2004,MEX-Corona}, but no information is known about
the outcome.
Apart from those, the spacecraft tracking measurements are
made on single frequency, thus determining the delay in solar wind
directly from the measurements themselves is not possible. While, in
principle, other radio observations could help, such as
VLBI \citep{Aksim2019} or observations of scintillations or radio
sources \citep{Shishov2016}, in practice there are too few of them, and
even fewer, if any, on the dates of spacecraft tracking measurements.
Three options remain for accounting for delay in solar wind when determining
planetary orbits for ephemeris:
\begin{enumerate}
\item Using a simplified model of solar wind electron density distribution,
and then determining its single free parameter (a reference electron density) from
the spacecraft tracking observations.
\item Using \textit{in situ}\xspace electron density measurements done by
spacecraft such as the Advanced Composition Explorer (ACE) and
\textit{Wind}, to account for large-scale
temporal variations of the said electron density in the
interplanetary space, while keeping a simplified symmetric model for
electron density distribution at any moment.
\item Using a more complicated dynamical model of solar wind, which
accounts for corotating interaction regions existing in solar wind.
Such models of the solar wind are often built on top of models
of solar corona, whose parameters are fit to magnetograms, either
daily or mean (averaged over one solar cycle). This approach
does not necessarily preclude using the \textit{in situ}\xspace data.
\end{enumerate}
The symmetric models most often include the term of electron density
proportional to $1/r^2$, where $r$ is the distance to the center of the
Sun. This term assumes constant radial velocity of electrons%
\footnote{
The connection between density distribution $N_{\mathrm{e}}(r)$ and radial velocity
$v_r(r)$ is expressed by the equation $r^2 v_r(r) N_{\mathrm{e}}(r) = \mathrm{const}$,
which follows from the continuity equation \citep{Aksim2019}.
}.
The factor of this term (reference electron density) is a free
parameter of the model that is determined from spacecraft ranging
observations.
While all modern planetary ephemeris (DE, EPM, INPOP) use the $1/r^2$
density, their approaches differ in details. DE200
\citep{Standish1990} used the following model for solar wind
correction of Mariner ranges:
$$ N_{\mathrm{e}} = \frac{A}{r^6} + \frac{B}{r^2},$$
\noindent where $A$ is a fitted parameter and $B$ depends on
solar latitude $\beta$:
$$B = \frac{ab}{a^2 \sin^2\beta + b^2 \cos^2\beta}$$
\noindent ($a$ and $b$ are fitted parameters as well).
DE430 \citep{DE430TR} presumably uses a similar model, where the
parameters are determined per-planet or per-spacecraft.
In EPM2004 \citep{Pitjeva2004} and subsequent releases of EPM, the following
model was used:
\begin{equation}\label{eq:bbdot}
N_{\mathrm{e}} = \frac{A}{r^6} + \frac{B + \dot{B} t}{r^2},
\end{equation}
\noindent where $A$ was fixed to the value determined in DE200, and
$B$ with its linear drift $\dot{B}$ were determined from observations,
per-planet, per-year.
INPOP presumably has two models that are used interchangeably
\citep{Verma2013}: (i) a two-parameter model with one term proportional
to $1/r^2$ and the other proportional to $1/r^4$; (ii) a two-parameter
model with the term proportional to $1/r^\epsilon$, where $\epsilon$
is the second determined parameter. The parameters are fit per each
solar conjunction of each planet of interest, and also separately for
spacecraft being in the fast-wind zone and in the slow-wind zone.
Outside planetary ephemeris, other solar wind electron density
models exist; see e.g. \citet{Leblanc98} where $1/r^2$, $1/r^4$, and
$1/r^6$ terms are combined. However, the $1/r^4$ and higher order
terms have negligible effect on spacecraft tracking signals that pass
farther than 15 solar radii ($15 R_{\odot}$) from the center of the Sun.
The signals that pass closer than that cannot be used for precise orbit
determination anyway because of the temporal and spatial instabilities
of the delay in solar wind in such proximity.
There exist other models with fractional powers of $r$ close to 2, see
e.g. \citet{Bougeret1984} with $r^{-2.10}$, \citet{Mann1999} with
$r^{-2.16}$, or \citet{Verma2013} with values of $\epsilon$ ranging
from 2.10 to 2.50. Such models in general do not possess physical
meaning and, as we will now show, do not offer prospects of more
accurate representation of spacecraft signal delay.
A major problem in determination of solar wind parameters from
spacecraft ranges (aside from the temporal, latitudinal, and longitudinal
variations) is that not only the factor of the $1/r^2$ (or
$1/r^\epsilon$) term must be determined from observations, but the
whole set of parameters that form the ephemeris solution. Those
include parameters of orbits of the planets, mass of the Sun, masses
of dozens of asteroids, and, most importantly, signal delays (biases)
that come from spacecraft transponder going out of calibration and
from systematic errors of signal registration in radio observatories
\citep{Kuchynka2012}.
In fact, just having one factor and one offset to fit for the
delay of spacecraft ranges in solar wind already makes the power of $r$
largely irrelevant. \Cref{fig:1r2_vs_1r25} in \Cref{sec:difficulties}
provides an example where $1/r^{2.5}$ model is fitted to data generated
using the $1/r^2$ model well enough for practical purposes.
We emphasise that there is no power law that can be preferable in a
symmetric model of solar wind electron density --- either from the
position of physical principles or from the position of processing of
signal delays.
\subsection{Medium-term density variations}\label{sec:meduim-term}
Medium-term variations in the solar wind are ones that have duration similar
to the period of solar rotation (Sun's synodic rotation period is 27.2753 days).
One obvious problem with the approach of using the $1/r^2$ model
is the assumption that solar wind is radially symmetric.
The asymmetry of the solar wind can be clearly seen, for instance,
in the LASCO coronagraph images%
\footnote{\url{http://www.swpc.noaa.gov/products/lasco-coronagraph}}.
The fact that spacecraft ranging
observations are sensitive enough to be affected by this asymmetry
was reported by \citet{Kuchynka2012}, who analyzed 10-day rolling averages
of the observation residuals and noticed the presence of periodic variations
with a period of 27 days. (See \Cref{sec:residuals} for brief introduction
to radio ranging residuals).
The way to account for these periodic variations is to adopt, instead of the
$1/r^2$ model, a three-dimensional numerical model of the solar wind that
would simulate (to some level of agreement) the corotating structures of the solar wind
based on observations of the Sun.
In this work, we examine the results obtained with the WSA-ENLIL model,
which consists of two parts:
the ENLIL solar wind model \citep{Odstrcil2003} and
the WSA \citep[Wang--Sheeley--Arge, ][]{Wang1990,Arge2000,Sheeley2017} model
of the solar corona.
ENLIL is a magnetohydrodynamic (MHD) solar wind model that calculates
3D distributions of solar wind's density, velocity, and magnetic field
by solving a system of differential MHD equations using a coronal solution
given by WSA as a boundary condition at 0.1~AU.
WSA is a semi-empirical solar corona model that uses line-of-sight
magnetic field measurements of the solar photosphere assembled into synoptic
magnetograms and simulates the solar wind plasma outflow up to 0.1~AU.
\begin{figure}
\centering
\includegraphics[height=5cm]{img/enlil_20170124.png}
\includegraphics[height=5cm]{img/enlil_20170510.png}
\caption{
Normalized solar wind electron number density calculated by
WSA-ENLIL with daily-updated GONGz magnetograms.
The data are plotted in the helio-equatorial plane with latitude of 0\degree\ in
the HEEQ coordinate system \citep[see][]{Thompson2006}.
}
\label{fig:enlil_sample_map}
\end{figure}
The primary purpose of WSA-ENLIL is prediction of solar wind
structures reaching Earth (and, with the help of the CONE model,
prediction of geomagnetic storms caused by coronal mass
ejections). However, there is no reason the model cannot be used for
other purposes, such as processing of spacecraft ranging observations.
\Cref{fig:enlil_sample_map} shows WSA-ENLIL density distributions
in the solar equatorial plane for January 24 and May 10, 2017.
The spiral shapes in the map are corotating solar wind structures.
\subsection{Long-term density variations}
Along with previously mentioned 27-day periodic variations in solar wind's
density, there are also long-term variations that happen on a larger
scale, i.e. on a scale of years.
These long-term variations can be seen by analysing \textit{in situ}\xspace density
measurements applying a rolling average filter with a window size
large enough to smooth out any medium-term data variations.
\Cref{fig:density_1au} shows 365-day rolling average of the solar wind electron
density at 1~AU calculated twice: (i) from the OMNI dataset
\citep{OMNI2014,WindACE2005} using values of proton density and alpha--proton
ratio and assuming charge neutrality (see \Cref{sec:insitu-data}),
and (ii) from the ENLIL simulations (see \Cref{sec:enlil_data}).
The plot reveals that the average electron density in the long term can deviate
up to 20\% from its mean value, so the assumption of some ``nominal'' electron
density for all time would be unjustified in the processing of spacecraft radio ranges.
Our suggestion is that these long-term variations must be taken into account
regardless of the particular solar wind model used.
Specifically, if the power-law model like $N_0/r^2$ is used,
one constant $N_0$ multiplier would result in constant density at 1~AU, which
would not be in agreement with \textit{in situ}\xspace observations.
Therefore, $N_0$ must be made proportional to \textit{in situ}\xspace density smoothed in
a way that would eliminate any medium-term disturbances and variations,
such as 27-day periodic ones resulting from the Sun's rotation.
In the case of ENLIL, we have to take into account the inner workings of the
model~\citep{ENLILCode}.
First, we must note that ENLIL assumes that the electron and proton densities are equal.
Also, it is important that to calculate the density
at the inner boundary, ENLIL requires that at $r = 0.1$~AU the condition
of constant momentum flux holds:
\begin{equation}
N_{\mathrm{e}}(\vartheta, \varphi)\, v_r^2(\vartheta, \varphi) =
d_\mathrm{fast}^{\phantom{1}} \cdot v_\mathrm{fast}^2 = \mathrm{const},
\end{equation}
in which $v_r(\vartheta, \varphi)$ is the radial velocity distribution produced
by the WSA solar corona model, $N_{\mathrm{e}}(\vartheta, \varphi)$ is the density
distribution further used as the boundary condition, and
$d_\mathrm{fast}$ and $v_\mathrm{fast}$ are free model parameters named
\texttt{dfast} and \texttt{vfast}.
Presumably, parameters $d_\mathrm{fast}$ and $v_\mathrm{fast}$ should be
adjusted so that the solution matches observational data.
If great precision is not a requirement, this adjustment may be performed
post factum, i.e. by directly scaling the solution $N_{\mathrm{e}}(r, \vartheta, \varphi)$.
The WSA-ENLIL average density curve shown in \Cref{fig:density_1au}
was calculated with fixed values of $d_\mathrm{fast}$ and $v_\mathrm{fast}$,
i.e. with one constant value of momentum flux.
The figure shows that when momentum flux is fixed, WSA-ENLIL's average electron
density at 1 AU (i) is not constant and (ii) does not match the OMNI numbers.
Quite obviously, with average density being non-constant, simply setting
$d_\mathrm{fast}$ proportional to electron density derived from OMNI would do very little
to bring the two curves in agreement.
Instead, knowing that variations in density are defined by variations
in WSA velocity, we presume that the first step should be to use values of
$v_\mathrm{fast}$ based on WSA velocity distribution $v_r(\vartheta, \varphi)$,
possibly choosing the distribution's average as the parameter's value.
Doing so would produce solutions with near-constant average density at 1~AU,
and, at that stage, $d_\mathrm{fast}$ may be set to values proportional to OMNI
density. Such a setup should, in theory, produce solutions that match
observational \textit{in situ}\xspace data.
However, since ENLIL simulations require time to complete and
solutions may be scaled post factum, the simplest (and the one used in this work)
way to bring ENLIL data in agreement with OMNI is to remove
the density trend shown in \Cref{fig:density_1au} directly from the solution
$N_{\mathrm{e}}(t, r, \vartheta, \varphi)$ and to apply to the data the same procedure
as we do to $N_0$ parameter of the $N_0/r^2$ model.
\begin{figure}
\centering
\includegraphics{tikzimg/density_1au.pdf}
\caption{
Solar wind electron density at 1 AU (365-day rolling average),
years 2006--2017.
}
\label{fig:density_1au}
\end{figure}
For a detailed description of the \textit{in situ}\xspace\ data used, see \Cref{sec:insitu-data}.
\section{Radio signal propagation in solar wind}
Radio signal propagation velocity in plasma is defined by
the group refractive index \citep{Aksim2019}
\begin{equation}
\label{eq:n_group}
n_{\mathrm{gr}}(\vecc r, \omega) \approx 1 + \frac{N_{\mathrm{e}}(\vecc r)}{2 \, N_{\mathrm{crit}}(\omega)},
\end{equation}
where $\omega$ is the signal's angular frequency,
$N_{\mathrm{e}}(\vecc r)$ is the plasma's electron number density,
and $N_{\mathrm{crit}}(\omega) = m_\mathrm{e} \, \varepsilon_0 \, \omega^2/e^2$ is
critical plasma density ($e$ is the charge of electron, and $\varepsilon_0$
is the vacuum permittivity).
To calculate the time%
\footnote{Conforming to the tradition of measuring time delays in units of length,
by ``time delay $\tau$'' we mean not \emph{time delay} specifically,
but rather \emph{optical path length}, which is measured in meters.}
it takes for a radio signal of frequency $\omega$ to travel
through the solar wind plasma from the point given by position vector $\vecc r_a$
to the point $\vecc r_b$, one has to integrate the plasma's group refractive
index $n_{\mathrm{gr}}(\vecc r, \omega)$ along the straight line connecting $\vecc r_a$
and $\vecc r_b$, that is, to evaluate the integral
\begin{equation}
\label{eq:tau_integral}
\begin{split}
\tau &= \int_{\vecc r_a}^{\vecc r_b} n_{\mathrm{gr}}(s, \omega) \, \mathrm{d} s = \\
&=
|\vecc r_b - \vecc r_a | +
\frac{1}{2 \, N_{\mathrm{crit}}(\omega)}\int_{\vecc r_a}^{\vecc r_b} N_{\mathrm{e}}(s) \, \mathrm{d} s
= \tau_\mathrm{dist} + \tau_\mathrm{cor}.
\end{split}
\end{equation}
The term $\tau_\mathrm{dist}$ in the above expression corresponds
to the time delay
due to the distance between $\vecc r_a$ and $\vecc r_b$,
i.e. the time it would take for the signal to travel the same path in vacuum,
and $\tau_\mathrm{cor}$ is the additional dispersive time delay due to the lower
propagation velocity in solar wind.
In the case of a power-law electron density model $N_{\mathrm{e}}(r) \propto r^\alpha$
the integral in \Cref{eq:tau_integral} may be expressed in closed form
\citep[see][]{Aksim2019}.
In the general case, specifically if the electron density is defined by a grid
of numerical values, which is the case with ENLIL and other numerical models,
the integral has to be evaluated numerically.
To calculate the integral numerically, we used the Simpson's rule
\citep[sec. 5.1]{Kahaner1989} with step size of 0.00625~AU,
equal to the radial grid spacing of the used simulation data
(see \Cref{sec:enlil_data}), which allows to exclude possible error
originating from the grid.
\section{Data}
Despite this work being focused on Mars spacecraft observations, each
of the considered solar wind models was tested as part of its
respective planetary solution. Each solution was fit not only to Mars
spacecraft observations, but to the full set of observations usually
required to build ephemeris, including not only spacecraft ranging, but
also optical observations and other kinds of
observations. For details, we refer the reader to \citep{Pitjeva2013,PitjevaPitjev2014}.
\subsection{Spacecraft ranges}\label{sec:spacecraft-ranges}
The Mars spacecraft ranges that were crucial to this work were taken
from the webpage of Solar System Dynamics (SSD) group at NASA
JPL\footnote{\url{https://ssd.jpl.nasa.gov/planets/obs_data.html}}, which includes
Mars Global Surveyor (MGS, 1999--2006),
Odyssey (2002--2013), and Mars Reconnaissance Orbiter (MRO, 2006--2013).
The extended set of ranges for MRO and Odyssey up to
the end of 2017 was kindly provided by William Folkner. The provided
ranges are normal points (NPs) made from raw observations. A normal point
represents the distance from an Earth radio observatory to the center of Mars,
corrected for the delay in ionosphere and troposphere.
Following the decision from \citep{Kuchynka2012}, pre-2005 Odyssey and
MRO ranges acquired by Deep Space Network stations 26 and 43 were
removed from analysis, as well as ranges acquired by station 54 before
October 2008.
MGS, Odyssey, and MRO ranges that pass closer than 30$R_{\odot}$
to the Sun suffer too much from solar wind instabilities and have
been removed from analysis. Ranges that pass farther than
30$R_{\odot}$, but closer than 60$R_{\odot}$ have been treated
separately (but equally with other ranges).
Mars Express (MEX) and Venus Express (VEX) ranges \citep{MorleyBudnik} that are
also sensitive to the delay in solar wind were downloaded from the Geoazur
website\footnote{\url{http://www.geoazur.fr/astrogeo/?href=observations/base}}.
They too are published as observatory--Mars distances corrected for
ionosphere and troposphere, but otherwise they are ``raw'' observations, so normal
points had to be made from them before processing. VEX ranges between 23.08.2010 and
31.10.2010 were excluded from analysis due to poor residuals, the reason
believed to be bad orbit determination due to trajectory control
maneuvers being more frequent than usual\footnote{This information
was obtained in personal communication from Elena Pitjeva, who
in turn obtained it from Trevor Morley.};
VEX ranges between 02.10.2011 and 05.12.2011 were excluded too,
assuming the same reason. Due to generally big noise in VEX residuals,
the decision was made to exclude VEX ranges that pass closer
than $90R_{\odot}$ to the Sun, so that they do not interfere
with less noisy observations in the comparison of the solar wind models.
The MEX ranges were treated similarly to Odyssey and MRO,
with the cutoff values of $30R_{\odot}$ and $60R_{\odot}$.
MESSENGER \citep[2011--2014, ][]{Park2017} and Cassini (2004--2014)
spacecraft ranges (normal
points) were also taken from the aforementioned SSD webpage; however,
the published values have been already reduced for the delay in solar wind
(calculated by the JPL DE solar wind model), so they did not
participate directly in the analysis of solar wind models. However,
those observations are very important for Solar system ephemeris as a
whole; in particular, they reduce the uncertainty of the mass of the
Sun.
In this work, different models of solar wind plasma electron density are
compared in their application to building planetary ephemeris. One of the models,
WSA-ENLIL, depends (in its WSA part) on the GONG magnetograms
that did not exist before September 2006 (see
\Cref{sec:enlil_data}), so the comparison is done for 2006--2017. The
spacecraft observations that touch this time span are summarized in
\Cref{tbl:data-spacecraft}. The reception (Rx) and transmission (Tx)
frequencies are listed so as to emphasise that the frequency
is important for calculation of the delay of radio signal in plasma,
see \Cref{eq:n_group}.
\begin{table*}
\begin{threeparttable}
\caption{Spacecraft observations used in the comparison of solar wind models.
NPs = normal points, Rx = reception, Tx = transmission,
$\sigma$ = \textit{a priori} error}
\label{tbl:data-spacecraft}
\begin{tabular}{lcrrrcc}
\hline
\textbf{Spacecraft} &
\textbf{Time span} &
\multicolumn{2}{c}{\textbf{\# of NPs}} &
\textbf{$\sigma$, one-way} &
\multicolumn{2}{c}{\textbf{Frequency}} \\
& &
\textbf{$> 60 R_{\odot}$} & \textbf{$30..60 R_{\odot}$} &
&
\textbf{Rx} & \textbf{Tx} \\
\hline
Odyssey & 2002--2017 & 7988 & 1784 & 0.5 m & 7155 MHz & 8407 MHz \\
MRO & 2006--2017 & 1924 & 561 & 0.5 m & 7183 MHz & 8439 MHz \\
MEX & 2005--2015 & 2888 & 718 & 1.5 m & 7100 MHz & 8400 MHz\\
VEX & 2006--2013 & 1294$^\dagger$\hspace{-0.42em} & --- & 3 m & 7166 MHz & 8419 MHz \\ \hline
\hline
\multicolumn{7}{l}{
$^\dagger$ Sun--signal distance for all VEX normal points
is greater than $90 R_\odot$.
}
\end{tabular}
\end{threeparttable}
\end{table*}
\subsection{\textit{In situ}\xspace solar wind measurements}
\label{sec:insitu-data}
Low-resolution (daily) version of the NASA/GSFC's OMNI
dataset\footnote{\url{http://spdf.gsfc.nasa.gov/pub/data/omni/low_res_omni}}
was used as the source of \textit{in situ}\xspace\ data in this work. The dataset is
combined\footnote{Methods used for the combination are outlined at
\url{http://omniweb.gsfc.nasa.gov/html/omni_min_data.html}} from
\textit{in situ}\xspace data measured by multiple spacecraft. The ones that serve as sources
of the particle density
data for 2006--2017 are the ACE \citep{ACE1998}
and \textit{Wind} \citep{Wind2021}.
OMNI provides proton density $N_{\mathrm{p}}$ and alpha--proton ratio $N_{\mathrm{a}}/N_{\mathrm{p}}$.
To calculate electron density $N_{\mathrm{e}}$,
we used the plasma quasi-neutrality condition $N_{\mathrm{e}} = N_{\mathrm{p}} + N_{\mathrm{a}}/2$.
The daily density data provided by OMNI are volatile because of the
nature of the solar wind structures; see the ``raw'' electron density
data derived from OMNI on \Cref{fig:omni-savgol}. Since we are
interested in the (average) electron density along the signal path,
our analysis can not benefit from the information about the daily
variations at the end of the signal path (near Earth). On the other
hand, we assume that a sufficiently smoothed electron density curve
will represent long-term variations of electron density of the solar
wind in the whole inner Solar system, including the Earth--Mars line.
The Savitzky--Golay filter \citep{SavGol1964} was chosen as one of the
most widely used filters for time series. The parameters of the filter
were chosen in an ad hoc manual procedure: the window of 511 days and
the polynomial order 5. The filtered data are shown on \Cref{fig:omni-savgol}.
\begin{figure}
\centering
\includegraphics{tikzimg/omni-savgol.pdf}
\caption{
OMNI data of electron density near Earth, years 2006--2017. The filter used is Savitzky--Golay
with a window of 511 days and the polynomial order 5.
}
\label{fig:omni-savgol}
\end{figure}
Alternatively to OMNI, \textit{in situ}\xspace\ electron density data are also provided
by the ESA/NASA Solar and Heliospheric Observatory \citep[SOHO,
][]{hovestadt1995celias}.
This data\footnote{\url{http://umtof.umd.edu/pm/crn}} is not part
of the OMNI combination.
The results obtained with the SOHO data (not shown in the text) turned out
to be somewhat worse than the ones obtained with OMNI.
\subsection{Numerical solar wind density maps}
\label{sec:enlil_data}
WSA-ENLIL data were produced by request via the ``Runs on Request'' service on
the Community Coordinated Modeling Center (CCMC)
website\footnote{\url{http://ccmc.gsfc.nasa.gov}}.
The whole dataset covers the dates between September 2006 and December 2017
and consists of 12 separate runs, each one covering one year.
The grid size in the WSA-ENLIL runs was set to
$256 \times 10 \times 90\ (r \times \vartheta \times \phi)$
with grid boundaries 0.1...1.7~AU, $-$20\degree...+20\degree\ latitude,
0\degree...360\degree\ longitude (in HEEQ coordinate system).
The grid spacing, therefore, is 0.00625~AU in radial direction and 4\degree\
in latitudinal and longitudinal directions.
The output frequency for the simulations was set to once every 12 hours.
The solar surface magnetic field data used for WSA coronal solutions
consisted of daily updated synoptic maps produced by the Global Oscillation
Network Group \citep[GONG, ][]{Hill2018GONG}.
It must be noted that there are two kinds of GONG magnetograms, called
GONGb and GONGz. They were supposed to be identical with the only
difference that GONGz has non-solar magnetic field bias
removed. However, it was found in late 2015 that GONGz
magnetograms suffered from over-subtraction in the polar regions. The
correction of the over-subtraction gradually went into effect starting
from August 2016 and reaching all GONG stations in April
2017. Unfortunately, it is not possible to retroactively correct the
past GONGz data~\citep[sec. 5.4]{CCMCReport}. So the numerical
simulations in 2017 and later are expected to be more accurate when
GONGz is used; before 2017, the expected improvement of GONGz is not
reached and GONGb is more accurate.
\section{Analysis}
As said above, a full planetary ephemeris solution were obtained for
each of the considered solar wind models. All solutions were based on
EPM and ERA-8 software \citep{PavlovSkripnichenko}.
\subsection{Dynamical model}
\label{sec:dynmodel}
EPM has a single dynamical model of the Solar system that includes all
planets, the Moon, the Sun, asteroids (see
section~\ref{sec:asteroids}), discrete model of Kuiper belt
(160 uniformly distributed points of equal mass)
and 30 individual trans-Neptunian objects \citep{PitjevaKuiperBelt}.
Sixteen bodies
(the Sun, the planets, Pluto, Ceres, Pallas, Vesta, Iris, Bamberga)
obey Einstein--Infeld--Hoffmann equations of motion. Other bodies,
for the sake of performance, interact with those 16 bodies with only
Newtonian forces, and do not interact with each other. It has been
checked that taking into account the full set of relativistic
interactions affects the Earth--Mars distance by a few centimeters,
which is negligible with the present observations.
Apart from point-mass interactions, the model includes additional
accelerations from solar oblateness and Lense--Thirring effect.
Earth also gets ``point mass--figure'' accelerations that come from the Sun,
Venus, Mars, Jupiter, and the Moon \citep{pavlov2016}.
The planetary part of the EPM model has over 600 parameters
that are determined simultaneously in the least-squares
fit (see \Cref{sec:residuals}).
Here are the parameters that are important for the analysis of
radio ranges from Earth to Mars orbiters:
\begin{itemize}
\item orbital elements for all planets at epoch,
\item solar oblateness factor,
\item masses of 279 asteroids,
\item distance scale parameter, historically known as correction to the value
of AU,
\item single factor of electron density, which applies to delay in solar wind plasma,
\item biases to compensate for calibration errors or clock offsets on Earth
or on spacecraft: 15 per-station biases for MGS/Odyssey, 14 per-station
biases for MRO, single transponder delay for MGS, single transponder delay
for Mars Express%
\footnote{It is not possible, nor necessary, to determine a transponder
delay for MRO, because it is absorbed by the per-station biases
in the solution. For the same reason, of the MGS/Odyssey pair,
only the MGS transponder delay is determined.}.
\end{itemize}
The distance scale parameter represents the correction to $GM_{\odot}$---the mass of
the Sun. It is numerically harder to determine the mass of the Sun
directly because of correlations with semi-major axes of planets.
Determination of $GM_{\odot}$ (and, in specific solutions, its time derivative)
allows to study the physical properties of the Sun and to test general
relativity \citep{Pitjeva2021}.
It is important to process different kinds of observations
simultaneously to get realistic values of parameters and error
estimates. For example, MESSENGER observations help greatly to
determine the solar oblateness factor, because Mercury is close to the
Sun. The solar oblateness factor affects orbits of Earth and Mars. To
determine the distance scaling factor, one has to take ranging data
for MESSENGER, MEX, Mars orbiters, and Cassini. Earth's orbit, whose
precision is obviously important for processing of the Earth--Mars
ranges, depends on the Moon's orbit around Earth, which is determined
from lunar laser ranging observations. Optical observations are
important for determination of orbits of Uranus and Neptune,
who in turn affect Mars's orbit on long time spans.
\subsection{Perturbing asteroids}\label{sec:asteroids}
An important part of the model, especially in regard to the Mars
spacecraft observations, are the point-masses that represent the
asteroids in the Main Asteroid Belt. In EPM2017, 301 largest
asteroids were present in the dynamical model; the remaining part of
the Main Asteroid Belt was modelled as a discrete rotating annulus
consisting of 180 uniformly distributed points of equal mass
\citep{Pitjeva2018,EPM2017}. Masses of 16 individual asteroids were
fixed to the values determined from observations of spacecraft or
natural satellites orbiting them; 30 masses were determined as part of
ephemeris solution (i.e. by the perturbations that they inflict on
orbits of inner planets); other asteroids' masses were determined
as mean densities of three taxonomic classes (C, S, M).
In the updated ephemeris dynamical model used in this work, the number
of individual asteroids is 279. The source list of asteroids was
merged (with obvious removal of duplicates) from 343 asteroids of the
DE430 model \citep{DE430TR} and 287 asteroids from \citep[Table
A.1]{Kuchynka2010}. The former is believed to be the list of most
perturbing asteroids, while the latter is believed to be the list of
most ``non-ring-like-acting'' asteroids, i.e. their cumulative effect
on inner planets cannot be modelled by a uniform ring. The resulting
list contained 379 asteroids. Then, 100 asteroids whose masses were
determined negative (though always within uncertainty) were excluded
from the model. Such negative masses are an artifact of the
least-squares method and they come from the fact that there is not
enough observational data to properly determine them.
None of the 279 masses were fixed in planetary solutions in this work;
rather, all of them were determined along with other planetary
parameters by a weighted least-squares method enhanced with Tikhonov
regularization \citep{Kuchynka2013,KanPavlov}.
The \textit{a priori} masses and uncertainties used in
the regularization were taken from different sources. The masses of 17
asteroids are known with small uncertainties, thanks to observations
of spacecraft or their natural satellites (comparing to EPM2017,
binary asteroid Euphrosyne was added to the list). For 77 asteroids,
the \textit{a priori} values were the ones determined from observed
deflections of orbits of other asteroids during close approach. For
the remaining 285 asteroids, the estimates were taken based on
diameters observed in infrared \citep{IRAS,Masiero2011} and densities
assigned to the three taxonomic classes.
\subsection{Solutions}
Three solutions were made, with identical sets of observations and parameters,
differing only in solar wind delay model for the time span of 2006--2017.
\begin{itemize}
\item In Solution I, the delays were calculated in accordance with the
simplest model of electron density, symmetric and stationary: $N_{\mathrm{e}} =
C N_0 / r^2$, where $N_0$ (the reference electron density at $r=1$
AU) was made equal to the 2006--2017 time-average of OMNI data, which
was found to be 5.97 cm$^{-3}$. The formal error of this average value
was calculated to be 0.07 cm$^{-3}$ from the formal errors of the proton
density and alpha/proton ratio: $\sigma(N_\mathrm{p})$ and $\sigma(N_{\mathrm{a}}/N_{\mathrm{p}})$.
The single electron density factor $C$ (which works as
the factor of delay, see the $\tau_\mathrm{cor}$ term
in \Cref{eq:tau_integral}) was fit to observations.
Ideally, $C$ would be equal to 1; the purpose of $C$ is to account
for uncertainty in the observed electron density, and also
for the fact that the assumption of stationary electron
density in 2006--2017 is not very realistic.
\item In Solution II, the delays were calculated in accordance with a
symmetric, but non-stationary electron density model: $N_{\mathrm{e}} = C
N_1(t) / r^2$, where the function $N_1(t)$ is is equal to the
smoothed electron density values derived from OMNI (see \Cref{sec:insitu-data} and \Cref{fig:omni-savgol}).
The single factor $C$ was fit to observations.
\item In Solution III, the delays were calculated in accordance with the
WSA-ENLIL model of electron density, scaled so that the
(365-day rolling average of)
electron density near Earth is equal to the smoothed OMNI values.
Similarly to Solutions I and II, a single factor $C$ of the electron
density was fit to observations.
\end{itemize}
The values of the fitted electron density factor $C$ in the three
solutions are listed in \Cref{tbl:results-wrms}. While its reference
value is 1 in all solutions, there is a difference in how it is fitted.
In Solution I, the $C$ was
not constrained, since the assumption of stationary electron
density in 2006--2017 is already not very realistic. In Solutions II
and III, the Tikhonov regularization was used to effectively constrain
$C$ so that its \textit{a priori} standard deviation is 1.17\%.
The value was taken equal to the relative standard deviation of
the electron density in the OMNI data (see the description of the
Solution I above).
Pre-2006 spacecraft ranges, which include MGS and parts of Odyssey and
MEX, were processed with the symmetric stationary solar wind
model, with different factors fit for different periods of time
between Sun--Mars conjunctions. One factor was fit for
2002/01/01---2003/12/31, another for 2004/01/01---2006/09/29.
A different factor was fit for pre-2002 MGS ranges. The same factor
was used for pre-1990 radar observations of Mars, Mercury, and Venus,
although they are hardly sensitive to delay in solar wind because of
low accuracy.
\section{Results}
\label{sec:results}
\Cref{fig:oc-mro-ody-1r2} shows the residual (postfit) differences of
Odyssey (blue points) and MRO (red points) ranges plus the values of
the model delay in solar wind. For better visibility, only
ranges from mid-2013 to mid-2017 are shown. The model delay in solar wind
is shown separately with green curves, so that one can see how the model
delay curve is fitted to observations.
The black curves show the time-averaged spacecraft range residuals.
They have visible medium-term
variations with period close to 27 days (the period changes as Earth
and Mars move on their orbits; the vertical lines on \Cref{fig:oc-mro-ody-1r2}
are separated by the solar rotation period). The amplitude of
variations increases near Sun--Mars conjunctions, which is expected
because during the conjunctions, Earth--Mars signals pass close to the Sun,
and at small distances from the Sun the electron density is higher than
at large distances. Observations with
signal passing closer than $30R_{\odot}$ to the Sun are not shown.
On the bottom plot of \Cref{fig:oc-mro-ody-1r2} it can be seen that
the WSA-ENLIL solar wind delay model expectedly has 27-day variations,
too. In some cases there is a good match between peaks in the model
and in the residuals, but often it goes ``out of sync'' or the
amplitudes become very different.
One explanation for this is the unfortunate fact that all the GONG
magnetograms are obtained from Earth---and the daily-updated map of
the magnetic field that is fed to the ENLIL's MHD equations is correct
only on the visible side of the Sun. On the nonvisible side, it is a
prediction, often not a very accurate one.
Before 2017, we can also attribute the differences between WSA-ENLIL
model data and real spacecraft delays to the GONGb magnetograms
containing non-solar magnetic field bias (see
\Cref{sec:enlil_data}). For 2017, when GONGz magnetograms were used,
we support the first explanation with anecdotal evidence. A good match
between WSA-ENLIL model and spacecraft data can be clearly seen on the
bottom plot of \Cref{fig:oc-mro-ody-1r2} in the first three months of
2017, while in the following three months, this is no longer the case.
\Cref{fig:enlil_sample_map} shows two example maps for the two
quarters of 2017. One can see that on the first map, the corotating
structures that intersect the Earth--Mars line originate from the
visible half of the Sun's surface; while on the second map, the
Earth--Mars line has changed in the way that it is also intersected by
a corotating structure whose origin is on the nonvisible half. (If the
pictures were animated, the structures would rotate anticlockwise in
time.) We admit that this can not be accepted as strong evidence and
hope for a future possibility to analyze post-2017 spacecraft
data together with WSA-ENLIL simulations based on properly fixed GONGz
magnetograms.
The Mars Express data are shown in \Cref{fig:oc-mex}. One can see,
apart from the expected peaks with a period of approximately 27
days, long-term variations that far exceed the peaks. The cause of the
variations certainly cannot be an inaccuracy of the EPM's dynamical
model of motion of Mars, or reductions, or solar wind delay model,
because there are no variations of that magnitude in the Odyssey and MRO
residuals. It is also unlikely that the variations come from a radio
observatory on Earth: same Deep Space Network stations observe Mars
Express, Odyssey and MRO spacecraft. The most reasonable explanation
would be that there are problems in orbit determination for the spacecraft.
In any case, the comparison of solar wind delay models against the MEX
observations currently makes little scientific sense, at least with
the ranges that pass farther than 30 $R_{\odot}$ from the Sun.
\begin{figure*}
\centering
\includegraphics{tikzimg/oc-mro-ody-1r2-1.pdf}
\includegraphics{tikzimg/oc-mro-ody-1r2-2.pdf}
\includegraphics{tikzimg/oc-mro-ody-1r2-3.pdf}
\caption{
Odyssey and MRO range residuals without correction for delay in solar wind
(blue and red points, respectively),
their 5-day rolling averages (black curve),
and model values (green curve) of delays in solar wind fitted
to observations via the factor $C$. The top plot corresponds to Solution I
($C N_0/r^2$ model), the middle plot corresponds to Solution II ($C N_1(t)/r^2$ model),
and the bottom plot corresponds to Solution III (WSA-ENLIL model).
The time span is from mid-2013 to
mid-2017. The vertical grid lines are placed 27.2753 days apart so as to
assist in examining the periodic variations due to the Sun's rotation. }
\label{fig:oc-mro-ody-1r2}
\end{figure*}
\begin{figure*}
\centering
\includegraphics{tikzimg/oc-mex.pdf}
\caption{
MEX range residuals without correction for delay in solar wind (red points)
their 5-day rolling averages (black curve),
and model values (green curve) of solar wind delays fitted
to observations via the factor $C$. The plot corresponds to Solution III (WSA-ENLIL model).
The time span is from mid-2013 to
mid-2017. The vertical grid lines are placed 27.2753 days apart so as to
assist in examining the periodic variations due to the Sun's rotation. }
\label{fig:oc-mex}
\end{figure*}
The statistics of the postfit results and values of important
parameters are given in \Cref{tbl:results-wrms}. Solution II
(time-dependent symmetric model) is slightly better than Solution I
(stationary symmetric model) overall by the reduced $\chi^2$ metric
and particularly better near conjunctions.
Solution III (WSA-ENLIL-based solar wind delay model) is no better than
Solution II, most probably because of the aforementioned problems with
magnetograms.
\begin{table}
\centering
\caption{Weighted root-mean-square (WRMS) statistics and values of
important parameters in the three solutions. The listed uncertainties are $3\sigma$.}
\label{tbl:results}
\begin{tabular}{lrrr}
\hline
&
\textbf{Solution I} &
\textbf{Solution II} &
\textbf{Solution III} \\
& $1/r^2$ & $1/r^2$ + OMNI & WSA-ENLIL \\
\hline
\textbf{WRMS} & & & \\
\hspace{1em} Odyssey & 0.564 m & 0.563 m & 0.563 m \\
\hspace{1em} --- conj.$^\dagger$ & 1.16 m & 1.15 m & 1.15 m \\
\hspace{1em} MRO & 0.619 m & 0.620 m & 0.619 m \\
\hspace{1em} & 1.20 m & 1.19 m & 1.19 m \\
\hspace{1em} MEX & 2.46 m & 2.46 m & 2.46 m \\
\hspace{1em} & 1.96 m & 1.96 m & 1.95 m \\
\hspace{1em} VEX & 6.24 m & 6.04 m & 6.04 m \\
Density & $1.31 \pm 0.81$ & $1.037 \pm 0.03$ & $1.037 \pm 0.03$ \\[-0.5ex]
factor $C$ & & & \\
$\Delta GM_{\odot}^\ddagger,\ \mathrm{km}^3/\mathrm{s}^2$ & $0.972 \pm 0.526$
& $1.457 \pm 0.463$
& $1.470 \pm 0.463$ \\
Variance of & 1.1127 & 1.1124 & 1.1124 \\[-0.5ex]
unit weight & & & \\[-0.5ex]
(reduced $\chi^2$) & & & \\ \hline \hline
\multicolumn{4}{p{8cm}}{
$^\dagger$ ``conj.'' (conjunction) observations are those where
the closest distance from the Earth--Mars signal to the Sun
is between 30 and 60 solar radii
}\\
\multicolumn{4}{l}{
$^\ddagger$ $\Delta GM_{\odot} =
GM_{\odot} - 132712440041\ \mathrm{km}^3/\mathrm{s}^2$
}
\end{tabular}
\label{tbl:results-wrms}
\end{table}
In Solutions II and III, the electron density delay factor $C$ was
constrained by Tikhonov regularization to be close to the reference
value of 1. An additional experiment (not shown) was made where the
Tikhonov regularization was not applied to $C$ in Solution II. The
resulting value of $C$ turned out to be $1.35 \pm 0.81$ (as opposed to
$1.037 \pm 0.03$ in ordinary Solution II). This confirms that
proper usage of absolute \textit{in situ} measurements of
electron density really helps to bring the solution, including the
$GM_{\odot}$ estimate, closer to reality. Also, the constraint on the
factor helps to reduce the uncertainty of $GM_{\odot}$ by $\sim 12$\%
(the uncertainty is 0.526 in Solution I and 0.463 in Solution II).
Another important improvement of Solution II over previous EPM
solutions is that only a single parameter related to solar wind is fit in
the solution, as compared to more than 50 in EPM2017
(see $B$ and $\dot B$ in \cref{eq:bbdot}, per-year, per-planet)
or 10 ($B$ per-conjunction from 2002 to 2018 and single $B$ prior to 2002)
in a later EPM version that was not released \citep{Pavlov2019,Pitjeva2021}.
\section{Conclusion and future work}
Conclusion:
\begin{itemize}
\item Interplanetary spacecraft ranging observations are sensitive
to signal delay due to solar wind and, in particular, to its medium-term
variations caused by the rotation of the Sun, and to its long-term
variations caused by changes in Sun's activity.
\item The ENLIL model of solar wind, built on daily magnetograms
done by the GONG network, can serve as a basis for a numerical
solar wind delay model. Such a model, though, can have or not
have a good match to the MRO and Odyssey ranges. Spacecraft ranges
that pass through the parts of the electron density map that are
effectively built from visible disk magnetograms have
(with otherwise equal conditions) better residuals than
the ranges that pass through the parts of the map that
are built from older magnetograms rotated to the present date.
\item Mathematical model of ephemeris can benefit from taking
into account a solar wind delay model that has
long-term variability determined from the OMNI dataset of
\textit{in situ} observations.
\item Constraining the (usually unconstrained) electron density
factor in a planetary ephemeris solution to a reasonable range
allows to reduce the error in the determined gravitational
parameter of the Sun by $\sim$ 12\%.
\item Using the WSA-ENLIL-based solar wind delay model to account
for medium-term variations can improve the residuals in individual
cases, but does not improve the solution overall.
\end{itemize}
Interplanetary spacecraft ranging observations are sensitive enough to
electron density so that they are used in this work to perform
comparative analysis and validation of solar wind models.
They can be used for validation of future solar wind models;
and potentially they can be used as a source of data for solar wind models.
\section*{Acknowledgements}
The authors would like to thank their colleagues from the IAA RAS:
Elena Pitjeva for continuous support of this work and a lot of
helpful advice regarding the planetary ephemeris solutions, and
Margarita Kan for implementing the Tikhonov regularization
method in the ERA-8 software and putting together the list of
\textit{a priori} masses of asteroids.
Planetary observations, including spacecraft ranges that are essential for this
work, were collected at the NASA SSD webpage supported by William Folkner
and at the Geoazur website supported by Agn{\`e}s Fienga.
The authors acknowledge use of NASA/GSFC's Space Physics Data
Facility's OMNIWeb service, and OMNI data, and are thankful to the
OMNIWeb team.
Solar wind simulation results have been provided by
the Community Coordinated Modeling Center at Goddard Space Flight Center
through their public ``Runs on Request'' system.
The numerical heliospheric code ENLIL has been developed by Dusan
Odstrcil who is currently at the George Mason University.
The WSA model, which gives the boundary conditions for the ENLIL
model, was developed by Yi-Ming Wang and Neil R. Sheeley Jr. at the
Naval Research Laboratory and later improved by Nick Arge, who worked at
the National Oceanic and Atmospheric Administration in Boulder at the
time and now works at the Goddard Space Flight Center.
The authors are indebted to Dusan Odstrcil for dedicated numerical
experiments and detailed explanations of inner workings of the ENLIL
model, to Leila Mays for organizing the numerical simulations at the
CCMC, and to Maria Kuznetsova for leading the CCMC itself.
\section*{Data Availability}
\label{sec:data_availability}
The Mars Global Surveyor (MGS), Mars Reconnaissance Orbiter (MRO),
and Odyssey spacecraft ranging data underlying this article
are available at the webpage of Solar System Dynamics (SSD) group at NASA
JPL\footnote{\url{https://ssd.jpl.nasa.gov/planets/obs_data.html}}.
The extended set of ranges for MRO and Odyssey up to
the end of 2017 was kindly provided by William Folkner of the SSD group
by permission and may be shared on reasonable request to the SSD team
of NASA JPL.
Mars Express (MEX) and Venus Express (VEX) ranges were downloaded from the Geoazur
website\footnote{\url{http://www.geoazur.fr/astrogeo/?href=observations/base}}.
The WSA-ENLIL data were produced by request via the ``Runs on Request'' service
at the Community Coordinated Modeling Center (CCMC) website%
\footnote{\url{http://ccmc.gsfc.nasa.gov}}.
The data are available for download through the ``Runs on Request'' service%
\footnote{\url{http://ccmc.gsfc.nasa.gov/ungrouped/SH/Solar_main.php}},
the names of the runs are:
\begin{itemize}
\item \href{http://ccmc.gsfc.nasa.gov/database_SH/Dan_Aksim_122319_SH_2.php}{Dan\_Aksim\_122319\_SH\_2} (2006)
\item \href{http://ccmc.gsfc.nasa.gov/database_SH/Dan_Aksim_122319_SH_3.php}{Dan\_Aksim\_122319\_SH\_3} (2007)
\item \href{http://ccmc.gsfc.nasa.gov/database_SH/Dan_Aksim_122319_SH_4.php}{Dan\_Aksim\_122319\_SH\_4} (2008)
\item \href{http://ccmc.gsfc.nasa.gov/database_SH/Dan_Aksim_122319_SH_5.php}{Dan\_Aksim\_122319\_SH\_5} (2009)
\item \href{http://ccmc.gsfc.nasa.gov/database_SH/Dan_Aksim_122319_SH_6.php}{Dan\_Aksim\_122319\_SH\_6} (2010)
\item \href{http://ccmc.gsfc.nasa.gov/database_SH/Dan_Aksim_122319_SH_7.php}{Dan\_Aksim\_122319\_SH\_7} (2011)
\item \href{http://ccmc.gsfc.nasa.gov/database_SH/Dan_Aksim_122319_SH_8.php}{Dan\_Aksim\_122319\_SH\_8} (2012)
\item \href{http://ccmc.gsfc.nasa.gov/database_SH/Dan_Aksim_122319_SH_9.php}{Dan\_Aksim\_122319\_SH\_9} (2013)
\item \href{http://ccmc.gsfc.nasa.gov/database_SH/Dan_Aksim_120920_SH_1.php}{Dan\_Aksim\_120920\_SH\_1} (2014)
\item \href{http://ccmc.gsfc.nasa.gov/database_SH/Dan_Aksim_122319_SH_11.php}{Dan\_Aksim\_122319\_SH\_11} (2015)
\item \href{http://ccmc.gsfc.nasa.gov/database_SH/Dan_Aksim_122319_SH_12.php}{Dan\_Aksim\_122319\_SH\_12} (2016)
\item \href{http://ccmc.gsfc.nasa.gov/database_SH/Dan_Aksim_112119_SH_3.php}{Dan\_Aksim\_112119\_SH\_3} (2017)
\end{itemize}
The OMNI \textit{in situ}\xspace data are available publicly through the OMNIWeb website%
\footnote{\url{http://omniweb.gsfc.nasa.gov}}.
Specifically, the file ``omni\_01\_av.dat'', containing daily averages
of solar wind parameters, comes from the low-resolution version%
\footnote{\url{http://spdf.gsfc.nasa.gov/pub/data/omni/low_res_omni/}}
of the OMNI dataset.
\bibliographystyle{mnras}
|
\section{Introduction}\label{intro}
\subsection{Two classical models in phase transitions vs.\ minimal surfaces}\label{sub-1.1}
The Peierls-Nabarro model was introduced in the early 1940's in the context of crystal dislocations \cite{Peierls, Nabarro} and also arises in the study of phase transitions with line-tension effects \cite{ABS} and boundary vortices in thin magnetic films \cite{K}. This model concerns the energy functional
\[
I_\varepsilon(u) :=
\frac \varepsilon 4[u]_{H^{1/2}(\mathbb{R}^n)}^2 +\negmedspace\int_{\mathbb{R}^n} W(u)\,dx, \ \quad [u]_{H^{1/2}(\mathbb{R}^n)}^2 := \iint_{\mathbb{R}^n\times \mathbb{R}^n}\negmedspace \frac{\big|u(x)-u(\bar x)\big|^2}{|x-\bar x|^{n+1}} \,dx \,d\bar x,
\]
where $u: \mathbb{R}^n \rightarrow (-1,1)$ and $W(u):= 1+\cos(\pi u)$.
The very related Allen-Cahn functional, introduced later, in 1958, within the context of the Van Der Walls-Cahn-Hilliard theory for phase transitions in fluids \cite{CahnHil}, is also tightly connected to the Ginzburg-Landau theory of superconductivity. It is defined as
\[
J_\varepsilon(u) := \frac{\varepsilon^2}{2} [u]_{H^{1}(\mathbb{R}^n)}^2 +\int_{\mathbb{R}^n} W(u)\,dx, \qquad [u]_{H^{1}(\mathbb{R}^n)}^2 := \int_{\mathbb{R}^n} |\nabla u|^2\,dx,
\]
where $u: \mathbb{R}^n \rightarrow (-1,1)$ and $W(u):= \frac 1 4 (1-u^2)^2$.
In both models $\varepsilon>0$ is a parameter and $W(u)$ is a so-called {\em double-well potential}, namely, a function with two minima (or wells) at the values $u=-1$ and $u= +1$ which correspond to two ``stable phases''.
Critical points $u\in C^2(\mathbb{R}^n)$ of $I_\varepsilon$ and $J_\varepsilon$ (more precisely, of their localized versions presented later in Subsection~\ref{subsectionstable}) solve,\footnote{Up to modifying the multiplicative constants in the definitions of $[\,\cdot\,]_{H^s}$ and $(-\Delta)^s$ in order to make them consistent.} respectively, the Peierls-Nabarro and Allen-Cahn equations:
\[
\varepsilon(-\Delta)^{1/2} u + W'(u) = 0 \qquad \mbox{and} \qquad \varepsilon^2(-\Delta) u + W'(u) = 0.
\]
A deep link between any of the two models and minimal surfaces is found when investigating the asymptotic behaviour of sequences of minimizers of $I_\varepsilon$ and $J_\varepsilon$ as $\varepsilon\downarrow 0$. Indeed, as a consequence of $\Gamma$-convergence results of Alberti, Bouchitt\'e, and Seppecher \cite{ABS} and of Modica and Mortola \cite{MM}, respectively, the following holds:
\vspace{4pt}
{\it If $u_{\varepsilon_k}:\mathbb{R}^n \rightarrow \mathbb{R}$ is a sequence of minimizers (in every bounded set) of either $I_{\varepsilon_k}$ or $J_{\varepsilon_k}$
and $\varepsilon_k\downarrow 0$, then (up to a subsequence)
\begin{equation}\label{assymp}
u_{\varepsilon_k} \, \stackrel{L^1_{\rm loc}}{\longrightarrow} \, \chi_{E}-\chi_{\mathbb{R}^n\setminus E}, \qquad \mbox{where $E\subset \mathbb{R}^n$ is a minimizer of the perimeter.}
\end{equation}
}In other words, for every $\lambda\in(-1,1)$ the level sets $\{u_\varepsilon=\lambda\}$ converge, as $\varepsilon\downarrow 0$ and up to a subsequence, to an area minimizing hypersurface (in particular minimal, i.e., with zero mean curvature).
\subsection{The fractional Allen-Cahn energies}\label{sub-1.2}
The classical functionals $I_\varepsilon$ and $J_\varepsilon$ introduced above belong to the more general family of Allen-Cahn energies
\[
E_{s, \varepsilon} (u) := \frac{\varepsilon^s}{4} [u]_{H^{s/2}(\mathbb{R}^n)}^2 +\int_{\mathbb{R}^n} W(u)\,dx ,
\]
where $s\in (0,2]$ and
\[
\qquad [u]_{H^{s/2}(\mathbb{R}^n)}^2 := (2-s)\iint_{\mathbb{R}^n\times \mathbb{R}^n} \frac{\big|u(x)-u(\bar x)\big|^2}{|x-\bar x|^{n+s}} \,dx\,d\bar x.
\]
Note that the normalization factor $2-s$ in the definition of $[u]_{H^{s/2}(\mathbb{R}^n)}^2$ guarantees that
\[[u]_{H^{s/2}(\mathbb{R}^n)}^2 \rightarrow \int_{\mathbb{R}^n} |\nabla u|^2\,dx\qquad \mbox{as $s\uparrow 2$}\]
(up to a dimensional multiplicative constant).
The functionals $E_{s,\varepsilon}$ have been extensively studied in the literature; see among others \cite{C-SM, SV-1, C-Si1, CSi, CC2, S-new} and references therein.
Note that the classical functionals $I_{\varepsilon}$ and $J_\varepsilon$ correspond to the cases $s=1$ and $s=2$ of $E_{s,\varepsilon}$.
Critical points $u: \mathbb{R}^n \to (-1,1)$ of $E_{s,\varepsilon}$ (in fact, of the natural localized version of $E_{s,\varepsilon}$ given in Subsection~\ref{subsectionstable}) satisfy the fractional Allen-Cahn equation
\begin{equation}\label{fract}
\varepsilon^s(-\Delta)^{s/2} u + W'(u) =0, \quad |u|\le 1,
\end{equation}
in $\mathbb{R}^n$ (up to a positive multiplicative constant in front of $(-\Delta)^{s/2} u$).
Similarly to the cases $s=1$ and $s=2$, for every $s\in [1,2]$ the energies $E_{s, \varepsilon}$ (suitably localized and renormalized) $\Gamma$-converge to the classical perimeter as $\varepsilon\downarrow 0$ ---see \cite{SV-gamma}.
As a consequence, if $u_{\varepsilon_k}$ is a sequence of minimizers (in every bounded set) of $E_{s,\varepsilon_k}$ with $s\in [1,2]$, then \eqref{assymp} holds (up to a subsequence).
Interestingly, a new qualitative behaviour is found for $s\in(0,1)$, where the asymptotic analysis of $E_{s, \varepsilon}$ as $\varepsilon \downarrow 0$ leads to the so-called $s$-{\em minimal surfaces} (or {\em nonlocal $s$-minimal surfaces}), introduced by Caffarelli, Roquejoffre, and Savin in \cite{CRS}. These new geometric objects generalize classical minimal surfaces (which are recovered as a limit case when $s\uparrow 1$) and share with them several structural properties; see \cite{CRS, FV, CSV, CCS} for the precise definition and several results on $s$-minimal surfaces.
\subsection{Bernstein's and De Giorgi's conjectures} \label{sub-1.3}
Let us recall the classical Bernstein's problem:
{\it If the graph of a function defined in all of $\mathbb{R}^{n-1}$ is a minimal surface in $\mathbb{R}^n$, is the function necessarily linear? }
This question has deep connections with the regularity theory for the (co-dimen-\-sion one) Plateau problem in any dimension and, more precisely, with the optimal dimensional bounds on the singular set of minimal surfaces ---as shown by De Giorgi in \cite{DG-Bern}. Moreover, its study contributed to striking developments in the theory of minimal surfaces. The question was completely answered in the late 1960's in a series of groundbreaking works, which established that:
\begin{enumerate}
\item[(i)] Hyperplanes are the only graphs of functions defined on $\mathbb{R}^{n-1}$ which are minimal surfaces in $\mathbb{R}^n$ as long as $n\le 8$ (Bernstein \cite{Bern}, Fleming \cite{Fle}, De Giorgi \cite{DG1, DG-Bern}, Almgren \cite{Alm}, and Simons \cite{Simons}).
\item[(ii)] There exist entire minimal graphs in $\mathbb{R}^n$ which are not hyperplanes in dimensions $n\ge 9$ (Bombieri, De Giorgi, and Giusti \cite{BDG}).
\end{enumerate}
Bernstein's problem finds a counterpart for certain ``critical points'' of $J_\varepsilon$ in a famous conjecture by De Giorgi (1978):
\begin{conjecture}[\cite{DG}]\label{DGconj}{\it
Let $u\in C^2(\mathbb{R}^{n})$, $|u|<1$, be a solution of $-\Delta u=u-u^3$ in~$\mathbb{R}^n$ satisfying $\partial_{x_{n}} u >0$.
Then, if $n\le 8$, $u$ must be 1D, that is, all its level sets $\{u=\lambda\}$ must be hyperplanes.}
\end{conjecture}
Conjecture \ref{DGconj} was first proved, about twenty years after it was raised, in dimensions $n=2$ and $n=3$, by Ghoussoub and Gui \cite{GG} and Ambrosio and Cabr\'e \cite{AmbrC}, respectively.
Almost ten years later, in the celebrated paper \cite{Savin}, Savin adressed the conjecture in the dimensions $4\le n\le 8$, and he succeeded in proving it under the additional assumption
\begin{equation}\label{limits}
\lim_{x_{n}\to \pm \infty} u = \pm 1.
\end{equation}
Short after, Del Pino, Kowalczyk, and Wei \cite{dPKW} established the existence of a counterexample to the conjecture in dimensions $n\ge 9$.
Since for $s\in [1,2)$ the functionals $J_\varepsilon$ and $E_{s,\varepsilon}$ have the identical asymptotic behaviour \eqref{assymp}, there are no heuristic reasons to prefer $J_\varepsilon$ to $E_{s,\varepsilon}$ when $s\in [1,2)$, in the statement of the De Giorgi conjecture. This motivates
\begin{conjecture}\label{DGconj2}
Conjecture \ref{DGconj} also holds for $(-\Delta)^{s/2} u=u-u^3$ when $s\in [1,2)$.
\end{conjecture}
On the other hand, for nonlocal $s$-minimal surfaces the analogue of the Bernstein's problem is well understood for $s\in (0,1)$ sufficiently close to $1$. Indeed, it follows, by
combining the results in \cite{CRS}, \cite{CV}, \cite{SV}, and \cite{FV} that:
\vspace{6pt}
{
\em For $ n\le 8$, there is a dimensional constant $s_*\in [0,1)$ such that, if $s\in [s_*,1)$, hyperplanes are the only graphs $\{x_n = \phi(x_1, \dots, x_{n-1})\}$ which are $s$-minimal surfaces in $\mathbb{R}^n$ ---for $n=2$ and $n=3$ this is known with $s_*=0$.
}
\vspace{6pt}
The heuristics thus suggest that the De Giorgi conjecture should be true for monotone critical points of $E_{\varepsilon,s}$ also in the range $s\in (s_*,1)$. That is:
\begin{conjecture}\label{DGconj2b}{\it
Conjecture \ref{DGconj} also holds for $(-\Delta)^{s/2} u=u-u^3$ when $s\in (s_*,1)$, where $s_*\in[0,1)$ is a dimensional constant.
}
\end{conjecture}
The articles by Cabr\'e and Sol\`a-Morales \cite{C-SM}, Cabr\'e and Sire \cite{CSi}, and Sire and Valdinoci \cite{Si-V} proved Conjectures \ref{DGconj2} and \ref{DGconj2b} in dimension $n=2$, for $s$ in the whole range $(0,2)$ ---that is $s_*=0$.
Later, Cabr\'e and Cinti \cite{CC1,CC2} established Conjecture~\ref{DGconj2} for $n=3$, $s\in [1,2)$. {{In \cite{DFV}, Dipierro, Farina, and Valdinoci proved Conjecture 1.3 for $n = 3$ and $s_* = 0$.}}
Very recently, Figalli and Serra \cite{FS} have proved it for $n=4$, $s=1$.
For all $s\in (1,2)$, the existence of monotone solutions to \eqref{fract} which are not $1D$ in dimensions $n\ge9$ has been announced in \cite{CLW} ---to appear in a work by Chan, D\'avila, del Pino, Liu, and Wei \cite{CDPLW}.
In this paper we will prove Conjecture \ref{DGconj2b} in dimension $n=4$ for $s\in (s_*,1)$, for some $s_*\in (0,1)$. We will also give a sufficient condition (in terms of a rigidity statement on $s$-minimal cones) guaranteeing that Conjecture \ref{DGconj2b} holds for some pair $(n, s)$, with $s\in (0,1)$ and $n\ge 3$.
Under the additional assumption \eqref{limits}, Conjecture~\ref{DGconj2} has been proved by Savin \cite{S-new,S-new2} for $4\le n\le 8$, $s\in [1,2)$.
Finally, Conjectures \ref{DGconj2} and \ref{DGconj2b} have been proven, also under the additional assumption \eqref{limits}, in Dipierro, Serra, and Valdinoci \cite{dPSV} in two situations: for $n=3$ and $s_*=0$, as well as for $4\le n\le 8$ and $s\in (s_*,1)$ with $s_*<1$ is sufficiently close to $1$.
\subsection{Monotone vs.\ stable solutions} \label{sub-1.4}
The assumption $\partial_{x_{n}} u >0$ easily yields that $u$ is a {\em stable solution}, i.e., a critical point of the localized version of $E_{s,\varepsilon}$ (presented below in Subsection~\ref{subsectionstable}) with nonnegative second variation.
The additional assumption \eqref{limits} on limits at infinity is only used in \cite{Savin, S-new, S-new2, dPSV} to guarantee that $u$ is, in addition, a minimizer of $E_{s,\varepsilon}$ in every bounded set
of~$\mathbb{R}^n$.
To address the conjecture without the additional assumption \eqref{limits}, it is natural to introduce the two limit functions $u^{\pm} := \lim_{x_{n} \to \pm\infty} u$. These functions depend only on the first $n-1$ variables $x_1, \dots, x_{n-1}$ and are stable solutions of \eqref{fract} in $\mathbb{R}^{n-1}$.
It is not difficult to prove (see Proposition \ref{implication}),
that if $u^\pm$ are 1D then $u$ is a minimizer. As a consequence, the following implication holds for all $s\in (0,2]$:
$$
\left.
\begin{array}{c}
\text{Stable sol'ns of \eqref{fract} in $\mathbb{R}^{n-1}$ are 1D}
\\
\text{and}\\
\text{Minimizers of \eqref{fract} in $\mathbb{R}^n$ are 1D}
\end{array}
\right\}
\Rightarrow \text{Monotone sol'ns of \eqref{fract} in $\mathbb{R}^n$ are 1D.}
$$
The difficult problem of classifying stable solutions to \eqref{fract} is connected to the following well-known conjecture for minimal surfaces:
\begin{conjecture}\label{conjstableminimal}{\it
Stable embedded minimal hypersurfaces in $\mathbb{R}^n$ are hyperplanes as long as $n\le 7$.}
\end{conjecture}
A positive answer to this conjecture is known for $n=3$, by a result of Fischer-Colbrie and Schoen \cite{FishS} and of Do Carmo and Peng \cite{dCP} from the late seventies, and for $n=4$ by a very recent paper of Chodosh and Li~\cite{CL}. It remains open for $5\le n\le 7$.
Instead, the analogue of Conjecture \ref{conjstableminimal} for area-minimizing (a stronger notion than stability) hypersurfaces is completely understood in every dimension: it holds, indeed, if and only if $n\leq 7$.
This shows that extending a result for minimizers to stable solutions may be a very difficult problem.
Still, Conjecture~\ref{conjstableminimal} suggests the so-called ``stability conjecture'':
\begin{conjecture}\label{DGconj3}{\it
Let $u\in C^2(\mathbb{R}^{n})$, $|u|<1$, be a stable critical point of $E_{s, \varepsilon}$.
Then, if $n\le 7$, $u$ must be $1D$ provided $s \in (s_* ,2]$, where $s_*\in [0,1)$ is a dimensional constant.}
\end{conjecture}
{
As explained before, once Conjecture \ref{DGconj3} is known to hold for some $(n-1,s)$ then Conjecture \ref{DGconj} / \ref{DGconj2} / \ref{DGconj2b} for $(n,s)$ can be reduced to the question of proving 1D symmetry of minimizers, which is well-understood thanks to the results in \cite{Savin, S-new, S-new2, dPSV}.
}
With the exception of the case $n=3$ and $s=1$ considered in \cite{FS}, before our work
Conjecture~\ref{DGconj3} was open in every dimension $n\ge 3$, with the case of the Laplacian ($s=2$) being a long standing open problem. The case $n=2$ is simpler (thanks to a certain ``parabolic manifold'' type property, it can be proved through some Liouville theorems initiated in \cite{BCN, AmbrC}) and has been established for the whole range $s\in (0,2]$: for $s=2$ in \cite{GG, AmbrC,AAC}, for $s=1$ in \cite{C-SM}, and for $0<s<2$ in \cite{CSi,Si-V}.
In this paper we establish Conjecture \ref{DGconj3} when $n=3$ and $s\in (s_*,1)$, for some $s_*<1$. From this and a result for minimizers from~\cite{dPSV}, we deduce Conjecture \ref{DGconj2b} when $n=4$ and $s\in (s_*,1)$.
\subsection{Main result: a new classification theorem}\label{sub-1.5}
We study stable solutions of the fractional Allen-Cahn equation
\begin{equation}\label{theequation}
(-\Delta)^{s/2}u + W'(u)=0, \quad |u|<1\quad \mbox{in }\mathbb{R}^n,
\end{equation}
with $s\in(0,1)$, where
\begin{equation} \label{quarticpot}
W(u):= \frac 1 4 (1-u^2)^2
\end{equation}
is the standard quartic double-well potential with wells at $\pm 1$. Note that this is equation \eqref{fract} with $\varepsilon=1$ (we can always assume this value of the parameter after scaling).
For the precise definition of {\em stable solution} to \eqref{theequation}, see Subsection \ref{subsectionstable} below.
The main goal of the paper is to establish the following classification result.
For brevity, we use in its statement the following terminology. Recall that, as mentioned before, stable solutions to the fractional Allen-Cahn equation in all of $\mathbb{R}^2$ have been already classified.
\begin{defi}\label{defiwhatcones}
For $n\ge 3$, we say that {\em hyperplanes are the only stable $s$-minimal cones in~$\mathbb{R}^n\setminus \{0\}$} when the following holds:
if $ 3\le m \le n$ and $\Sigma\subset \mathbb{R}^m$ is a stable
$s$-minimal cone in $\mathbb{R}^m\setminus \{0\}$, whose boundary $\partial \Sigma$ is nonempty and smooth away from~$0$, then necessarily $\partial \Sigma$ is a hyperplane (up to sets of measure zero).
The notion of (smooth away from $0$) stable $s$-minimal cone in $\mathbb{R}^m\setminus \{0\}$ is exactly that of \cite[Definition 1.1]{CCS}, which we recall in Definition \ref{defstablecone} below.\footnote{In this paper and in \cite{CCS} the perturbations to define stability do not modify the cone in a neighborhood of the origin (even if in \cite{CCS} we used the terminology ``stable $s$-minimal cones in~$\mathbb{R}^m$'').}
\end{defi}
{
\begin{rem}
Let us comment on Definition \ref{defiwhatcones}, which will be an assumption of our main result, Theorem \ref{thmclas} below.
\begin{itemize}
\item First, we will not be assuming that stable $s$-minimal cones in $\mathbb{R}^2\setminus \{0\}$ must be flat: the integer $m$ in the definition is at least 3.
This is important because (similarly as it happens in the classical case $s=1$) certain $s$-minimal symmetric cones like the ``cross'' $\{x_1 x_2>0\}$ are stable in $\mathbb{R}^2 \setminus \{0\}$, at least if $s$ is sufficiently close to $1$.
\item Second, the reason why, given a dimension $n$ we need an hypothesis in lower dimensions $3\le m\le n$, is the use of a dimension reduction argument of Federer type in the proof of Theorem \ref{thmclas}.
This is also why Definition \ref{defiwhatcones} concerns only cones with smooth trace on the sphere (the dimension reduction argument will always allow us to suppose this).
Now, in this dimension reduction, one needs to consider the case $m=2$ (and not only $m\ge3$). However, for the case $m=2$ we do not need to assume any flatness hypothesis since we will actually prove in Lemma \ref{hwioheoithe} the following property, which holds true for all $s\in(0,1)$: any cone in $\mathbb{R}^n$ of the form $\tilde\Sigma \times \mathbb{R}^{n-2}$ and such that $\chi_\Sigma- \chi_{\Sigma^c}$ can be obtained as limit of some sequence of stable critical points of $E_{s,\varepsilon}$ (with vanishing parameter $\varepsilon$) must be a half-space.
\end{itemize}
\end{rem}
}
The following is our main result.
\begin{thm}\label{thmclas}
Assume that, for some pair $(n,s)$ with $n\ge 3$ and $s\in (0,1)$, hyperplanes are the only stable $s$-minimal cones in $\mathbb{R}^n\setminus \{0\}$.
Then, every stable solution of \eqref{theequation}-\eqref{quarticpot} is a 1D layer solution, namely, $u(x)= \phi(e\cdot x)$ for some direction $e\in S^{n-1}$ and increasing function $\phi: \mathbb{R} \rightarrow (-1,1)$.
\end{thm}
\begin{rem}\label{otherW}
As it will be clear from the proofs, throughtout the paper we do not need $W$ to be given exactly by $W(u)= \frac 1 4 (1-u^2)^2$, nor to be even. It can be replaced everywhere by any function $W\in C^3([-1,1])$ satisfying $W(\pm1)=0$, $W>0$ in $(-1,1)$, $W\le W(t_0)$ {in $(-1,1)$} for some $t_0\in(-1,1)$, $\{t\in [-1,1]\,:\,W'(t)=0\} =\{-1,t_0,1\}$, $W''(\pm1) >0$, and $W''(t_0)<0$. For instance, $W$ can be replaced everywhere by the Peierls-Nabarro potential $1+\cos (\pi u)$. However, for simplicity, we will write the results for the potential \eqref{quarticpot}.
Note also that the results of Subsection~\ref{sub-2.1} apply to any $C^3$ potential $W$.
\end{rem}
Thanks to the theorem,
to establish Conjecture \ref{DGconj3} in $\mathbb{R}^n$, $n\geq 3$, it suffices to prove that hyperplanes are the only stable $s$-minimal cones in $\mathbb{R}^n\setminus \{0\}$ for $s\in (s_*,1)$. Recently, in~\cite{CCS} we have
proved that planes are the only stable $s$-minimal cones in $\mathbb{R}^3\setminus \{0\}$ for $s\in (s_*,1)$, for some $s_*<1$. As a consequence of this last result, we obtain the two following corollaries.
First, we can prove the stability conjecture in $\mathbb{R}^3$ for $s\in (0,1)$ sufficiently close to $1$.
\begin{cor}\label{corclas2}
Conjecture \ref{DGconj3} holds in $\mathbb{R}^3$
---that is, every stable critical point of $E_{s, \varepsilon}$ in $\mathbb{R}^3$ is $1D$ for $s\in (s_*,1)$ with $s_*<1$ sufficiently close to 1.
\end{cor}
Using Corollary \ref{corclas2}, we establish the De Giorgi type conjecture in $\mathbb{R}^4$ for $s\in (0,1)$ sufficiently close to $1$.
\begin{cor}\label{corclas3}
Conjecture \ref{DGconj2b} holds in $\mathbb{R}^4$
---that is, every monotone solution of \eqref{theequation} in $\mathbb{R}^4$ is $1D$ for $s\in (s_*,1)$ with $s_*<1$ sufficiently close to 1.
\end{cor}
We recall that in \cite{FS} Figalli and Serra proved Conjecture \ref{DGconj2b} in $\mathbb{R}^4$ for $s=1$.
\subsection{Definitions: localized energy, minimizers, and stable solutions}\label{subsectionstable}
Thro-\-ughout the paper we consider solutions $u: \mathbb{R}^n\to (-1,1)$ to integro-differential equations
\begin{equation}\label{equK}
L_K u+ W'(u )=0 \quad \mbox{in } \Omega,
\end{equation}
where $W\in C^3(\mathbb{R})$, $\Omega\subset \mathbb{R}^n$ is an open set, and
\begin{equation}\label{LK}
L_Ku(x):= \int_{\mathbb{R}^n} \big(u(x)-u(\bar x)\big)K(x-\bar x)\,d\bar x.
\end{equation}
Although in some parts of the paper we will take $L_K = (-\Delta)^{s/2}$, i.e., $K(z) = (2-s) |z|^{-n-s}$,
some other parts will hold and will be stated for more general kernels~$K$. Throughout the article,
$K$ belongs to the ellipticity class $\mathcal L_2$ of Caffarelli and Silvestre~\cite{CS3}. That is, we assume $s\in (0,2)$,
\begin{equation}\label{L0}
K(z) =K(-z),
\quad
\frac{(2-s)\lambda}{|z|^{n+s}} \le K(z) \le \frac{(2-s)\Lambda}{|z|^{n+s}},
\end{equation}
and that $K$ is of class $C^2$ away from the origin with derivatives satisfying
\begin{equation}\label{L2}
\max\bigl\{|z|\, |\partial_{e} K(z)| \,,\, |z|^2 |\partial_{ee} K(z) |\bigr\} \le \frac{(2-s)\Lambda}{|z|^{n+s}},
\end{equation}
for all $z\in\mathbb{R}^n\setminus \{0\}$ and $e\in S^{n-1}$. Here $\lambda$ and $\Lambda$ are given positive constants.
By a solution of \eqref{equK}, we mean a bounded function in all of $\mathbb{R}^n$ which is, in $\Omega$, a $C^2$ strong solution.
Note that, since $K$ is in the class $\mathcal L_2$, $W\in C^3(\mathbb{R})$, and $L_K$ is translation invariant, any measurable function $u:\mathbb{R}^n \to (-1,1)$ which solves \eqref{equK} in the sense of distributions, i.e., $\int uL_K \xi+ W'(u )\xi=0$ for all $\xi \in C^\infty_c(\Omega)$, belongs to $C^2(\Omega)$ and hence it is a strong solution. Indeed, this follows from the existing regularity theory for nonlocal elliptic equations, as explained in Appendix~\ref{app-C}.
We now recall the standard definitions of localized energy, minimizers, and stable solution in this more general framework.
We first point out that \eqref{equK} is the Euler-Lagrange equation of the localized energy functional
\begin{equation}\label{energyW}
\mathcal E_{\Omega} (v) := \mathcal E_{\Omega}^{\rm Sob} (v) + \mathcal E_{\Omega}^{\rm Pot} (v) ,
\end{equation}
where
\[
\mathcal E_{\Omega}^{\rm Sob} (v) := \frac{1}{4}\iint_{(\mathbb{R}^n\times\mathbb{R}^n)\setminus (\Omega^c \times \Omega^c)}
|v(x)-v(\bar x)|^2 K(x-\bar x) \,dx\,d\bar x, \quad \mathcal E_{\Omega}^{\rm Pot} (v) :=\int_{\Omega} W(v)\,dx.
\]
Here and throughout the paper, $\Omega^c$ denotes the complement of $\Omega$ in $\mathbb{R}^n$.
Note that $\mathcal E_{\Omega}(v)$ is finite for every function $v$ which is bounded in $\mathbb{R}^n$ and Lipschitz in a neighborhood of $\overline\Omega$, since $s\in (0,2)$.
We say that $u$ is a {\em minimizer} of $\mathcal E_\Omega$ (or a minimizer of \eqref{equK} in $\Omega$) if $\mathcal E_\Omega(u)<\infty$ and $\mathcal E_\Omega(u)\leq \mathcal E_\Omega(v)$ for all $v$ satisfying $v=u$ outside of $\Omega$.
If $u$ is a minimizer of $\mathcal E_\Omega$, then
\begin{equation}\label{secondvar}
\frac{d^2}{dt^2} \Big|_{t=0} \mathcal E_\Omega(u + t \xi) \ge 0 \quad \mbox{for all Lipschitz function } \xi \text{ with } \xi =0 \text{ in } \Omega^c.
\end{equation}
This motivates the definition of stable solution.
Within this framework, $u$ is a stable solution if and only if $u$ is a critical point of $\mathcal E_\Omega$ for which
the second variation of $\mathcal E_\Omega$ at $u$ ---with respect to compactly supported perturbations vanishing outside $\Omega$--- is nonnegative. In other words:
\begin{defi}[\textbf{Stability}]\label{stability}
Given an open set $\Omega\subset\mathbb{R}^n$, we say that a solution $u$ of \eqref{equK} is {\em stable} if \eqref{secondvar} holds, or equivalently,
\begin{equation}\label{stable}
\frac{1}{2}\iint_{\mathbb{R}^n\times\mathbb{R}^n} |\xi(x)-\xi(\bar x)|^2 K(x-\bar x)\,dx\,d\bar x + \int_{\mathbb{R}^n} W''(u)\xi^2\,dx\geq 0
\end{equation}
for all Lipschitz functions $\xi$ in $\mathbb{R}^n$ with compact support in $\mathbb{R}^n$ and such that $\xi=0$ in $\Omega^c$.
As a particular case, we say that $u$ is a stable solution of $\eqref{theequation}$ if \eqref{stable} holds with $K(z)=(2-s) |z|^{-n-s}$ for all Lipschitz functions $\xi$ with compact support in~$\mathbb{R}^n$.
\end{defi}
We conclude this section giving the notion of fractional $s$-perimeter, as introduced by Caffarelli, Roquejoffre, and Savin in \cite{CRS}. Let $s\in (0,1)$. Given a set $E$ in $\mathbb{R}^n$ and a bounded open set $\Omega\subset \mathbb{R}^n$, we define the fractional $s$-perimeter of $E$ in $\Omega$ as
\begin{equation}\label{def-per}
\hspace{-.181cm} P_{s}(E,\Omega):=\iint_{(E\cap \Omega) \times (\mathbb{R}^n\setminus E)}\frac{1}{|x-\bar x|^{n+s}}\,dx \,d\bar x + \iint_{(\Omega\setminus E)\times(E \setminus \Omega)}\frac{1}{|x-\bar x|^{n+s}}\,dx \,d\bar x.
\end{equation}
Observe that, in our notation, $P_s(E,\Omega)=2\mathcal E^{\rm Sob}_\Omega(\chi_E)$, where $\chi_E$ denotes the characteristic function of the set $E$ and we choose here $K(z)=|z|^{-n-s}$.
\section{Further new results and organization of the paper} \label{sec-2}
Theorem \ref{thmclas} will follow as a combination of several new results ---which in addition are of independent interest--- on stable solutions to integro-differential semilinear equations. They are valid in any dimension $n\ge 2$ and include:
\begin{itemize}
\item A sharp $BV$ estimate;
\item A bound of $\mathcal E^{\rm Pot}$ by $\mathcal E^{\rm Sob}$, a cube decomposition, and an energy estimate;
\item A density estimate;
\item The convergence of blow-downs to stable $s$-minimal cones.
\end{itemize}
These results are presented in detail in the next subsections.
\subsection{$BV$ estimate} \label{sub-2.1}
For $s\in (0,1)$, in Section \ref{sec-3} we establish $BV$ and Sobolev estimates for stable solutions to the semilinear equation $L_K(u)+W'(u)=0$ in a ball~$B_{2R}$. Their proof follow the techniques introduced in \cite{CSV} within the context of stable nonlocal $s$-minimal surfaces. Our estimates are optimal in their dependence on the radius $R$
and universal in the sense that they are independent of the potential~$W$. Note that a universal $BV$ estimate is somehow surprising since it corresponds to a quantity of higher order than the one of the equation, which is $s$.
By the same method and independently of our work, the $BV$ and Sobolev estimates \eqref{BVest} and \eqref{energy-est-eq} have also been established by Gui and Li~\cite[Proposition~1.7]{GL}. In the current paper we additionally keep track on how the constants in the estimates blow-up as $s \uparrow1$. Such information could be useful in some applications.\footnote{This was the case in our previous paper \cite{CCS}, where we needed the dependence on $s$ (as $s\uparrow 1$) in the $BV$ estimate from \cite{CSV} for stable $s$-minimal sets.}
\begin{thm}\label{BV}
Given $s_0>0$, let $n\ge 2$, $s\in(s_0,1)$, $W$ be any $C^3(\mathbb{R})$ function, and $K$ satisfy \eqref{L0} and \eqref{L2}. Let $R>0$ and $u:\mathbb{R}^n\to (-1,1)$ be a stable solution of $L_K u+W'(u)=0$ in $B_{2R}\subset\mathbb{R}^n$.
Then,
\begin{equation}\label{BVest}
\int_{B_R}|\nabla u|\,dx \le \frac{C}{1-s} R^{n-1},\end{equation}
where $C$ is a constant which depends only on $n$, $\lambda$, $\Lambda$, and $s_0$.
\end{thm}
{
Two remarks on this result are in order.
\begin{rem}\label{remscaling}We emphasize that, in Theorem \ref{BV}, the constant $C$ is completely independent of the potential $W$, which is an arbitrary $C^3$ function ---in particular it is not necessarily a double-well potential.
As a consequence, the estimate \eqref{BVest} is scale invariant: the estimate for any $R>0$ follows from the case $R=1$ applied to the rescaled function $\tilde u = u(R\,\cdot\,)$, which solves $L_K \tilde u+\widetilde W'(\tilde u)=0$ in $B_{2}$, for $\widetilde W := R^sW$.
\end{rem}
\begin{rem} In the case of solutions $u_\varepsilon :\mathbb{R}^n\to \mathbb{R}$ of the fractional Allen-Cahn equation \eqref{fract} (recall that $\varepsilon>0$ is a small parameter), Theorem \ref{BV} reads:
\[
u_\varepsilon\mbox{ stable in $B_2$} \quad \Rightarrow \quad \int_{B_1}|\nabla u_\varepsilon|\,dx \le \frac{C}{1-s}.
\]
We emphasize that the constant in the right-hand side of the previous estimate is independent of $\varepsilon$. We do not expect the blow-up rate $(1-s)^{-1}$ as $s\uparrow 1$ to be sharp, but for such uniform in $\varepsilon$ estimate to hold, the right-hand side needs to blow up at least at the rate $(1-s)^{-1/2}$ ---we actually expect the power $-{1/2}$ to be the sharp rate.
Indeed, by the argument sketched next, for all $n\ge 2$ and $s\in (0,1)$ there exists $\varepsilon = \varepsilon(s)>0$ and a stable solution $u_\varepsilon: \mathbb{R}^n\to \mathbb{R}$ of \eqref{fract} in $B_2$ satisfying $\int_{B_1}|\nabla u_\varepsilon|\,dx \ge c(1-s)^{-1/2}$, where $c$ stays bounded away from zero as $s\uparrow 1$.
To show this, consider the set \[E_* : = \bigcup_{k\in \mathbb Z} \,\big\{ 2kC_*(1-s)^{1/2} \le x_n \le (2k+1)C_*(1-s)^{1/2}\big\},\] whose boundary consists of parallel hyperplanes arranged in a $C_*(1-s)^{1/2}$-periodic fashion. Let us prove first that $E_*$ is stable $s$-minimal set in $(-1,1)^n\supset B_1$, provided $C_*$ is chosen large enough.
To verify this, we can use the stability criterium found in \cite{DDPW,FFFMM} involving the nonlocal analogue $c_s^2$ of the squared norm of the second fundamental form of $\partial E_*$. To this end, one first checks that $c_s^2$ behaves as the quantity $(C_*(1-s)^{1/2})^{-1-s}$.
Hence, the stability inequality in $(-1,1)^n$ reads
\begin{equation}\label{wheiothewohw}
\sum_{j} C_*^{-1-s} \int_{\Gamma_j}\eta^2 d\mathcal H^{n-1}_x \le C(1-s)^{\frac{1+s}{2}}\sum_{i,j} \int_{\Gamma_i} \int_{\Gamma_j} \frac{\big|\eta(x)-\eta(\bar{x})\big|^2}{|x-\bar{x}|^{n+s}}d\mathcal H^{n-1}_xd\mathcal H^{n-1}_{\bar{x}},
\end{equation}
for all $\eta\in C^{0,1}_c([-1,1]^n)$ where $\Gamma_j : = \{x_n = j C_*(1-s)^{1/2} \}$.
By the fractional Poincar\' e inequality for the $H^{(1+s)/2}$ seminorm in $[-1,1]^{n-1}$ (applied to the functions $\eta_{|_{\Gamma_j}} $, $\Gamma_j \cong \mathbb{R}^{n-1}$) we have
\[
\int_{\Gamma_j}\eta^2 d\mathcal H^{n-1}_x \le C (1-s) \int_{\Gamma_j} \int_{\Gamma_j} \frac{\big|\eta(x)-\eta(\bar{x})\big|^2}{|x-\bar{x}|^{n+s}}d\mathcal H^{n-1}_xd\mathcal H^{n-1}_{\bar{x}},
\]
for all $\eta\in C^{0,1}_c([-1,1]^n)$, where $C$ is uniformly bounded as $s\uparrow 1$. Therefore, since ${(1-s)}/{(1-s)^{\frac{1+s}{2}}} \to 1$ as $s \uparrow 1$, \eqref{wheiothewohw} will hold provided that we choose $C_*$ large enough. This makes $E_*$ stable.
At the same time, it is known \cite{Cetal} that the fractional Allen-Cahn equation \eqref{fract} admits periodic 1D solutions for all large enough periods (relatively to $\varepsilon$). Hence, given $s\in (0,1)$, for every $0<\varepsilon \le c(1-s)^{1/2}$ there exists a periodic 1D solution $u_{*,\varepsilon}(x_n)$ of \eqref{fract} such that $\{u_{*,\varepsilon}=0\} = \partial E_*$.
Now, as $\varepsilon\downarrow 0$ we have, by construction, $u_{*,\varepsilon} \to \chi_{E_*} - \chi_{E_*^c} $ and hence one can show that, for $\varepsilon$ small enough depending on $s$,
$u_{*,\varepsilon}$ inherits from $E_*$ the stability in $(-1,1)^n$. Finally, note that for such construction $\int_{B_1}|\nabla u_\varepsilon|\,dx$ blows up at the rate $(1-s)^{-1/2}$ (by the coarea formula), as claimed.
\end{rem}
}
As a simple consequence of Theorem \ref{BV} we deduce the following estimate for the Sobolev part of the energy of stable solutions to \eqref{equK}.
\begin{cor}\label{energy-est}
Given $s_0>0$, let $n\ge 2$, $s\in(s_0,1)$, $W$ be any $C^3(\mathbb{R})$ function, and $K$ satisfy \eqref{L0} and \eqref{L2}. Let $R>0$ and $u:\mathbb{R}^n\to (-1,1)$ be a stable solution of $L_K u+W'(u)=0$ in $B_{2R}\subset \mathbb{R}^n$.
Then,
\begin{equation}\label{energy-est-eq}
\mathcal E^{\rm Sob}_{B_R}(u)\le \frac{C}{(1-s)^2} R^{n-s},
\end{equation}
where $C$ is a constant which depends only on $n$, $\lambda$, $ \Lambda$, and $s_0$.
\end{cor}
\subsection{Control of $\mathcal E^\text{Pot }$ by $\mathcal E^\text{Sob}$}\label{pot0}
In Section \ref{sec-4} we need to assume $W$ to be a double-well potential. For simplicity we take $W$ to be the quartic potential \eqref{quarticpot}, although all proofs apply to the more general class of double-well potentials described in Remark~\ref{otherW}.
The following result states that $\mathcal E^{\rm Sob}$ in $B_R$ controls $\mathcal E^{\rm Pot}$ in a slightly smaller ball.
Here we can take $s\in (0,2]$ and thus we include the case of the Laplacian. To our knowledge, this result is new even for the Laplacian.{\footnote{For $s=2$ and in the particular case of stable solutions in all of $\mathbb{R}^n$, Villegas \cite[Proposition~1.5]{Vill} has recently proved that the quotient of Sobolev and potential energies in $B_R$ tends to $1$ as $R\uparrow \infty$.}}
To state our result for $s=2$,
we recall that if $K_s(z) \asymp (2-s)|z|^{-n-s}$ (where $X\asymp Y$ here means $X\le CY$ and $Y \le CX$) we have $L_{K_s} v \to a_{ij}\partial_{ij} v$ for all $v\in C^2_c(\mathbb{R}^n)$ as $s\uparrow 2$ (see the computations in \cite[Section 6] {CS-reg}), where $a_{ij} \asymp {\rm Id}$ (that is $a_{ij}\partial_{ij}$ is some translation invariant second order elliptic operator).
In the following proposition, $L_{K}$ is meant to be, when $s=2$, this limiting second order elliptic operator.
\begin{prop}\label{DircontrolsPot2}
Given $s_0>0$, let $n\ge 2$, $s\in(s_0,2]$, $W(u)=\frac 1 4 (1-u^2)^2$, and $K$ satisfy \eqref{L0} and \eqref{L2}. Let $R>0$ and $u:\mathbb{R}^n\rightarrow (-1,1)$ be a stable solution of $L_{K}u + W'(u)=0$ in~$B_{2R}$.
Then,
\begin{equation}\label{Sobcontrols} \mathcal E^{\rm Pot}_{B_{R-R_0}}(u) \le C\mathcal E^{\rm Sob}_{B_R}(u)\end{equation}
whenever $R> R_0$, where $C$ and $R_0$ are positive constants which depend only on $n$, $\lambda$, $\Lambda$, and $s_0$.
\end{prop}
To prove Proposition \ref{DircontrolsPot2} we cover the ball $B_{R-R_0}$ by cubes of size comparable to $R_0$ and with controlled overlapping, and we show that the contribution of each of these cubes to the total Sobolev energy controls its contribution to the potential energy.
We believe that this cube decomposition, which holds also for the classical Allen-Cahn equation, may be of independent interest.
In Appendix~\ref{app-B}, we present an alternative proof ---only in the case $s=2$, for brevity--- of a weaker version of the estimate of Proposition \ref{DircontrolsPot2} in which an error term appears on the right-hand side of \eqref{Sobcontrols}. The error is of order $R^{n-s}$ when $s\in(0,1)$, $R^{n-1}$ for $s\in(1,2]$, and $R^{n-1}\log R$ for $s=1$. Even if the result is weaker, we give this alternative proof since it is new even for the Laplacian (to the best of our knowledge), very short, and interesting in itself. It relies on taking a suitable test function on the stability inequality and on elementary estimates.
Combining Proposition \ref{DircontrolsPot2} with Corollary \ref{energy-est} we obtain the following bound for the full energy of stable solutions in a ball.
\begin{thm}\label{thm1}
Given $s_0>0$, let $n\ge 2$, $s\in(s_0,1)$, $W(u)=\frac 1 4 (1-u^2)^2$, and $K$ satisfy \eqref{L0} and \eqref{L2}. Let $R\geq 1$ and $u:\mathbb{R}^n\rightarrow (-1,1)$ be a stable solution of $L_K u+W'(u)=0$ in $B_{2R}\subset\mathbb{R}^n$.
Then,
\begin{equation}\label{egest}
\mathcal E_{B_R}(u) \le \frac{C}{(1-s)^2}R^{n-s},
\end{equation}
where $C$ is a constant which depends only on $n$, $\lambda$, $\Lambda$, and $s_0$.
\end{thm}
For $s\in [1,2]$, \cite{CC1, CC2} established the related energy estimates $\mathcal E_{B_R}(u) \le C R^{n-1}$, for $s\in (1,2]$, and $\mathcal E_{B_R}(u) \le C R^{n-1}\log R$, for $s=1$, in the case of minimizers in $B_{2R}\subset\mathbb{R}^n$. However, such strong estimates are not expected to hold if one considers stable solutions ---assuming stability only in the double ball $B_{2R}$ as in Theorem \ref{thm1}, not in the whole $\mathbb{R}^n$--- instead of minimizers.
\subsection{Density estimates} \label{sub-2.3}
Section \ref{sec-5} deals with density estimates for stable solutions of $(-\Delta)^{s/2}u + W'(u)=0$ in all of $\mathbb{R}^n$.
We need to assume the operator $L_K$ to be the fractional Laplacian $(-\Delta)^{s/2}$ since an important ingredient in their proof is the monotonicity formula from \cite{CC2}, which uses the extension property of $(-\Delta)^{s/2}$ and thus is not available for other kernels. In addition, our proof uses crucially \eqref{BVest} and thus needs $s<1$.
\begin{prop} \label{density}
Let $n\ge 2$, $s\in (0,1)$, and $W(u)= \tfrac 1 4 (1-u^2)^2$. Let $u:\mathbb{R}^n\rightarrow (-1,1)$ be a stable solution of $(-\Delta)^{s/2}u + W'(u)=0$ in $\mathbb{R}^n$.
Then, for every $\bar c\in (0,1)$ there exist positive constants $\omega_0$ and $R_0$, which depend only on $\bar c$, $n$, and $s$, such that the following holds. For every $R\geq R_0$, if
\begin{equation}\label{hp-density}
R^{-n}\int_{B_R} |1+u| \, dx \leq \omega_0\qquad \left(\mbox{respectively,}\quad R^{-n}\int_{B_R} |1-u| \, dx \leq \omega_0\right),
\end{equation}
then
\begin{equation}\label{th-density}
\left\{ u\ge -\bar c\right\}\cap B_{R/2} = \varnothing\qquad \left(\mbox{respectively,}\quad \left\{ u\le \bar c\right\}\cap B_{R/2} = \varnothing\right).
\end{equation}
\end{prop}
\subsection{Convergence of blow-downs} \label{sub-2.4}
The goal of Section \ref{sec-6} is to establish that the blow-downs of entire stable solutions to $(-\Delta)^{s/2}u + W'(u)=0$ in $\mathbb{R}^n$ converge, up to a subsequence, to the characteristic function of a stable nonlocal $s$-minimal cone. This convergence is local, both in the $L^1$-sense and in the sense of the Hausdorff distance between the level sets and the (boundary of the) cone. {To be precise, following \cite[Definitions 1.1 and 2.2]{CCS}, we recall now two notions of stability for the fractional perimeter $P_s$ defined in $\eqref{def-per}$.
The first one is the notion of stability for a cone that is smooth away from the origin. It is the notion used in Definition \ref{defiwhatcones} above.
\begin{defi}\label{defstablecone}
Let $\Sigma\subset \mathbb{R}^{n}$ be a cone with nonempty boundary of class $C^2$ away from the origin.
We say that $\Sigma$ is a stable cone for the $s$-perimeter in $\mathbb{R}^n\setminus \{0\}$, or stable $s$-minimal cone in $\mathbb{R}^n\setminus \{0\}$, if
\begin{equation}\label{stabilityweak}
\liminf_{t\to 0} \frac{1}{t^2} \big(P_s(\phi^t_X(\Sigma), B_1)-P_s(\Sigma, B_1))\ge 0
\end{equation}
for all vector fields $X\in C^\infty_c(B_1\setminus\{0\}, \mathbb{R}^n)$. Here $\phi^t_X:\mathbb{R}^n\to \mathbb{R}^n$ denotes the integral flow of $X$ at time $t$ (which is a smooth diffeomorphism for $t$ small).
\end{defi}
The following is the notion of {\em weak stability} for a general set $E$ with finite $s$-perimeter, as given in \cite[Definition 2.2]{CCS}.
\begin{defi}\label{def-weakly}
A set $E\subset \mathbb{R}^n$ with $P_s(E,\Omega)<\infty$ is said to be {\em weakly stable} in $\Omega$ for the $s$-perimeter if for every given vector field $X=X(x,t)\in C^\infty_c(\Omega\times(-1,1); \mathbb{R}^n)$ we have
\[
\liminf_{t\to 0} \frac{1}{t^2} \big(P_s(\phi^t_X(E), \Omega)-P_s(E, \Omega)\big)\ge 0,
\]
where $\phi^t_X $ denotes the integral flow of $X$ at time $t$ (with $0$ as initial time).
\end{defi}
In contrast with this last notion, Definition \ref{defstablecone} assumes the vector field $X=X(x)$ to be autonomous. However, since a cone $\Sigma$ in Definition \ref{defiwhatcones} is smooth outside of the origin, it is simple to see that $\Sigma$ is a stable $s$-minimal cone in $\mathbb{R}^n\setminus\{0\}$ (in the sense of Definition \ref{defstablecone}) if and only if $\Sigma$ is a weakly stable set in $\mathbb{R}^n\setminus\{0\}$ for the fractional perimeter $P_s$ (in the sense of Definition \ref{def-weakly}) .}
In our following result, we will prove the limiting cone $\Sigma$ to be weakly stable not only in $\mathbb{R}^n\setminus\{0\}$, but in $\mathbb{R}^n$. In its statement we use the notation
\[
\overline D(E;x) = \limsup_{r\to 0} \frac{|E\cap B_r{(x)}|}{|B_r|} \qquad \mbox{and}\qquad \underline D(E;x) = \liminf_{r\to 0} \frac{|E\cap B_r{(x)}|}{|B_r|}
\]
for the upper and lower densities of a set $E\subset\mathbb{R}^n$ at a point $x\in \mathbb{R}^n$.
\begin{thm}\label{thm2}
Let $n\ge 2$, $s\in (0,1)$, and $W(u)=\frac 1 4 (1-u^2)^2$. Let $u:\mathbb{R}^n\rightarrow (-1,1)$ be a stable solution of $(-\Delta)^{s/2}u + W'(u)=0$ in $\mathbb{R}^n$.
Then, for every given blow-down sequence $u_{R_j}(x)= u(R_jx)$ with $R_j\uparrow \infty$, there is a subsequence $R_{j_k}$ such that
\[ u_{R_{j_k}} \rightarrow \chi_{\Sigma}-\chi_{\Sigma^c} \quad \mbox{ in }L^1(B_1)\]
for some cone $\Sigma$ which is a weakly stable set in $\mathbb{R}^n$ for the fractional perimeter~$P_s$ and which is nontrivial (not equal to $\mathbb{R}^n$ or $\varnothing$ up to sets of measure zero).
In addition, up to changing $\Sigma$ in a set of measure zero, we have
\begin{equation}\label{wngioewhtioeh1}
x\in \partial \Sigma \qquad \Leftrightarrow\qquad 0<\underline D(x; \Sigma) \le \overline D(x; \Sigma)< 1
\end{equation}
and, for all given $c\in(-1,1)$ and $\rho\ge 1$, we have
\begin{equation}\label{wngioewhtioeh2}
d_{\rm Hausdorff}\big(\, \{u_{R_{j_k}} \ge c\}\cap B_\rho\,,\, \Sigma \cap B_\rho\,\big) \rightarrow 0 \qquad \text{as } k\uparrow\infty,
\end{equation}
where $d_{\rm Hausdorff}(X,Y) = \inf\{ d>0 \,: \, X\subset Y+B_d \ \mbox{and} \ Y\subset X+B_d\}$ denotes the standard Hausdorff distance between subsets of $\mathbb{R}^n$.
\end{thm}
The proof of Theorem \ref{thm2} puts together all our results stated above in this section.
More precisely, the first part of Theorem \ref{thm2} ($L^1$-convergence) is the content of Proposition~\ref{conv-L^1}, the proof of which uses the $BV$ and energy estimates from Theorems \ref{BV} and \ref{thm1}, as well as the monotonicity formula from \cite{CC2}.
The second part of the statement (uniform convergence) follows then from the density estimates of Proposition \ref{density}.
\subsection{Proofs of the main result and its corollaries} \label{sub-2.5}
In Section \ref{sec-7} we prove Theorem \ref{thmclas} by combining Theorem \ref{thm2} and the ``improvement of flatness'' results established by Dipierro, Serra, and Valdinoci in \cite{dPSV}.
There are some nontrivial technical details involved in our proof, like a dimension reduction argument that allows us to assume that the stable cones obtained after blow-down are smooth away from $0$ (as required in Definition \ref{defiwhatcones}). It is important to deal with this smoothness issue to guarantee the applicability of Theorem \ref{thmclas} in concrete cases. For instance, from \cite{CCS}, in $\mathbb{R}^3$ we only know how to classify stable cones that are smooth away from $0$.
Finally, also in Section \ref{sec-7} we will give the straightforward proofs of Corollaries~\ref{corclas2} and \ref{corclas3}.
\subsection{Smooth stable $s$-minimal surfaces in $\mathbb{R}^3$ are planes when $s\sim 1$}\label{sub-2.6}
In Appendix \ref{app-A}, the arguments of Section \ref{sec-7} will be easily modified to establish the flatness of $C^2$ stable $s$-minimal surfaces in all of $\mathbb{R}^n$ whenever all stable $s$-minimal cones are known to be hyperplanes up to dimension $n$. As a consequence we deduce that every $C^2$ stable $s$-minimal surface in $\mathbb{R}^3$ must be a plane if $s\in (s_*,1)$ for some $s_* < 1$, since our main result in \cite{CCS} states that planes are the only stable $s$-minimal cones in $\mathbb{R}^3\setminus\{0\}$ (smooth away from $0$) for all $s \in(s_*, 1)$. This was already announced without proof in \cite[Corollary 1.3]{CCS}.
More precisely, we have the following result:
\begin{thm}\label{thmRn}
Assume that, for some pair $(n,s)$ with $n\ge 3$ and $s\in (0,1)$, hyperplanes are the only stable $s$-minimal cones in $\mathbb{R}^n\setminus\{0\}$.
Let $E$ be an open subset of $\mathbb{R}^n$, with $\partial E$ nonempty and of class $C^2$, and such that $E$ is a weakly stable set for the $s$-perimeter in $\mathbb{R}^n$ (as defined in {Definition~\ref{def-weakly}}).
Then, $E$ is a half-space.
\end{thm}
\begin{cor}\label{corR3}
Let $E\subset \mathbb{R}^3$ be a $C^2$ open subset, with $\partial E\not = \varnothing$, which is weakly stable for the $s$-perimeter in $\mathbb{R}^3$ (as defined in {Definition~\ref{def-weakly}}).
If $s$ is sufficiently close to 1, then $E$ is a half-space.
\end{cor}
Theorem \ref{thmRn} will follow from a blow-down procedure. Although the general strategy is similar to the classical one for minimizers of the perimeter, to deal with stable solutions one finds several technical issues that are completely analogous to those in Section \ref{sec-7} for the semilinear equation. This close analogy was our reason to postpone the proof of Corollary 1.3 in \cite{CCS} (i.e., Corollary \ref{corR3} here) to Appendix~\ref{app-A} of the current paper.
\section{$BV$ and Sobolev energy estimates}\label{sec-3}
Throughout the paper, $L_K$ denotes the operator \eqref{LK}, where $K$ satisfies \eqref{L0} and \eqref{L2}.
This section extends some results and techniques introduced in \cite{CSV}, within the context of stable nonlocal minimal surfaces, to stable solutions of semilinear nonlocal problems.
Following \cite{SV-mon,CSV}, we consider translations of the solution $u$ in some direction $\boldsymbol {v}\in S^{n-1}$ and we compare the energies of $u$ and of $u(\cdot+t\boldsymbol v)$.
Since we need to consider perturbations vanishing outside the domain, following the notation of \cite{CSV} we introduce a Lipschitz radial cut-off function $\varphi_4$ such that $\varphi_4\equiv 1$ in $B_{2}$, $\varphi_4\equiv 0$ outside~$B_4$,
and $|\nabla \varphi_4| \leq 1/2$. For this we take $\varphi_4$ to be linear, as a function of the radius $|x|$, in $B_4\setminus B_2$.
For $\boldsymbol v\in S^{n-1}$ and $t\in(-1,1)$ we consider the map
\begin{equation}\label{perturb}
\Psi_{t}(y):= y + t\varphi_4 (y)\boldsymbol v \qquad\text{for $y\in\mathbb{R}^n$.}
\end{equation}
Clearly, $\Psi_t$ is a Lipschitz diffeomorphism from $\mathbb{R}^n$ onto itself. In particular, both $\Psi_t$ and $\Psi_t^{-1}$ are Lipschitz and coincide with the identity outside $B_4$. Given a function $u$ defined in all of~$\mathbb{R}^n$, we introduce
\[ u_{t}(x) := u\bigl(\Psi_{t}^{-1}(x)\bigr).\]
Even if we do not denote its dependence on $\boldsymbol v$, the function $u_t$ depends on both $\boldsymbol v$ and $t$. We finally set
$$
M_t(x):=\max\{u(x),u_t(x)\} \quad \text{ and } \quad m_t(x):=\min\{u(x),u_t(x)\}.
$$
Notice that $u_t$, $M_t$, and $m_t$ all coincide with $u$ outside $B_4$.
We now prove now the analogue of Lemma 2.1 in \cite{CSV} for the energy \eqref{energyW} associated to \eqref{equK}.
\begin{lem}\label{lem2A}
Let $n\ge 2$, $s\in (0,1)$, let $W$ be any continuous function in $\mathbb{R}$, and $K$ satisfy \eqref{L0} and \eqref{L2}. Let $u\in C^1(B_6) \cap L^\infty(\mathbb{R}^n)$ be a given function.
Then, for all $t\in(-1,1)$ we have that
\begin{equation}\label{eqlemAstables}
\mathcal E_{B_4}(u_{t}) + \mathcal E_{B_4}(u_{-t}) - 2\mathcal E_{B_4}(u) \le C t^2 \mathcal E^{\rm Sob}_{B_4}(u),
\end{equation}
where $C$ is a constant which depends only on $n$, $\lambda$, and $\Lambda$.
\end{lem}
\begin{proof}
We set $A_4:=(\mathbb{R}^n\times \mathbb{R}^n) \setminus (B_4^c\times B_4^c)$. We have
\[
\mathcal E_{B_4}(u_{\pm t}) = \frac{1}{4}\iint_{A_4} |u(\Psi_{\pm t}^{-1}(x))-u(\Psi_{\pm t}^{-1}(\bar x))|^2K(x-\bar x)\,dx\,d\bar x + \int_{B_4} W(u(\Psi_{\pm t}^{-1}(x)))\,dx.
\]
Changing variables $y= \Psi^{-1}_{\pm t}(x)$, $\bar y= \Psi^{-1}_{\pm t}(\bar x)$ in the integrals above (recall that $\Psi^{-1}_{\pm t}$ sends $B_4$ and $B_4^c$ onto themselves), we obtain the following expressions for the Sobolev and potential energies:
\[
\mathcal E^{\rm Sob}_{B_4}(u_{\pm t}) = \frac{1}{4}\iint_{A_4} |u(y)-u(\bar y)|^2 K\bigl(\Psi_{\pm t}(y)-\Psi_{\pm t}(\bar y)) \,J_{\pm t}(y) \,J_{\pm t}(\bar y)\,dy \,d\bar y\,
\]
and
\[\mathcal E^{\rm Pot}_{B_4}(u_{\pm t})=\int_{B_4} W(u(y))\,J_{\pm t}(y)\,dy,\]
where $J_{\pm t}$ are the Jacobians.
They are easily seen to be
\[ J_{\pm t}(y)= 1 \pm t\partial_{\boldsymbol v}\varphi_4 (y) ,\]
which are positive quantities since $|t|<1$ and $|\partial_{\boldsymbol v}\varphi_4|\leq 1/2$.
Clearly,
\begin{equation}\label{pot-pm}\mathcal E^{\rm Pot}_{B_4}(u_{ t})+\mathcal E^{\rm Pot}_{B_4}(u_{- t})= \int_{B_4} W(u(y))(J_{t}(y)+J_{-t}(y))\,dy= 2\mathcal E^{\rm Pot}_{B_4}(u).
\end{equation}
To estimate the sum of the two Sobolev energies $\mathcal E^{\rm Sob}_{B_4}(u_{ t})+\mathcal E^{\rm Sob}_{B_4}(u_{- t})$, we first observe that, since $K\in \mathcal L_2$, then it satisfies the following estimates for its first and second derivatives:
$$\max\left\{|z||\partial_eK(z)|, |z|^2\sup _{|y-z|\leq|z|/2} |\partial_{ee}K(y)|\right\}\leq 2^{n+s+2}\frac{\Lambda}{\lambda} K(z)\leq 2^{n+3}\frac{\Lambda}{\lambda} K(z)$$
for every $z\in \mathbb{R}^n\setminus \{0\}$ and $e\in S^{n-1}$.
Using this fact and performing the same computations as in the proof of Lemma 2.1 in \cite{CSV} (we are taking there, with the notations of \cite{CSV}, $R=4$ and $K^*=2^{n+3}(\Lambda/\lambda)K$), we deduce that
\begin{equation*}
\mathcal E^{\rm Sob}_{B_4}(u_t)+ \mathcal E^{\rm Sob}_{B_4}(u_{-t}) \leq 2\mathcal E^{\rm Sob}_{B_R}(u) + C t^2 \iint_{A_4} |u(y)-u(\bar y)|^2 K(y-\bar y)\,dy\,d\bar y,
\end{equation*}
where $C$ depends only on $n$, $\lambda$, and $\Lambda$.
This, combined with \eqref{pot-pm}, concludes the proof of the lemma.
\end{proof}
\begin{lem}\label{inner-stable}
Let $n\ge 2$, $s\in (0,1)$, let $W$ be any $C^3([-1,1])$ function, and $K$ satisfy \eqref{L0} and \eqref{L2}. Let $u: \mathbb{R}^n\to (-1,1)$ be a stable solution of $L_Ku+W'(u)=0$ in $B_6\subset\mathbb{R}^n$.
Then, given
$\nu>0$, there exists $t_0\in (0,1)$ $($possibly depending on $\nu$, $W$, and $u$$)$, such that
\begin{equation}\label{inner-stability}
\mathcal E_{B_4}(M_t) + \mathcal E_{B_4}(m_t) -2\mathcal E_{B_4}(u)\geq -\nu t^2\qquad \mbox{for all\;\;} t\in(-t_0,t_0).
\end{equation}
\end{lem}
\begin{proof}
As shown in Appendix~\ref{app-C}, we know that $u\in C^{2}(B_6)$. Thus, the fractional Sobolev semi-norm $\mathcal E_{B_4}$ of $u$ is finite.
We claim that, given any Lipschitz function $\xi$ vanishing outside $B_4$ and $t\in (-1,1)$ such that $|u+t\xi|\le 1$, we have
\begin{equation}\label{2variation}
\mathcal E_{B_4}(u+t\xi)-\mathcal E_{B_4}(u)\geq - C \|\xi\|_{L^\infty(B_4)}^3|t|^3
\end{equation}
for some constant $C$ depending only on $W$.
Indeed, we have \begin{equation*}
\begin{split}
&\mathcal E_{B_4}(u+t\xi)-\mathcal E_{B_4}(u)=\\
&\hspace{1em}= \frac{t}{2}\iint_{\mathbb{R}^n\times \mathbb{R}^n} (u(x)-u(\bar x))(\xi(x)-\xi(\bar x))K(x-\bar x)\,dx\,d\bar x \\
&\hspace{3em}+\frac{t^2}{4}\iint_{\mathbb{R}^n\times \mathbb{R}^n}|\xi(x)-\xi(\bar x)|^2K(x-\bar x)\,dx\,d\bar x +\int_{B_4} \big(W(u+t\xi)-W(u)\big)\,dx\\
&\hspace{1em}= \frac{t}{2}\iint_{\mathbb{R}^n\times \mathbb{R}^n} (u(x)-u(\bar x))(\xi(x)-\xi(\bar x))K(x-\bar x)\,dx\,d\bar x +\int_{B_4} W'(u)t\xi \, dx\\
&\hspace{3em}+\frac{ t^2}{4} \iint_{\mathbb{R}^n\times \mathbb{R}^n} |\xi(x)-\xi(\bar x)|^2K(x-\bar x)\,dx\,d\bar x+\int_{B_4} W''(u)\frac{(t\xi)^2}{2}\,dx \\
&\hspace{3em} +\int_{B_4} W'''(u^*(t,x)) \frac{(t\xi )^3}{6}\,dx\\
\end{split}
\end{equation*}
for some $u^*(t,x)$ in the interval with endpoints $u(x)$ and $u(x)+ t\xi(x)$ (and thus satisfying $|u^*|\leq 1$). Hence, using that $u$ is a stable critical point and $W\in C^3$,
we deduce \eqref{2variation}.
We now choose $\xi:=\frac{(u_t-u)_+}{t}$ and $\tilde \xi:=-\,\frac{(u_t-u)_-}{t}$ in the above computations, where $(\cdot)_+$ and $(\cdot)_-$ denote the positive and negative parts. Observe that $M_t=u+t\xi$ and $m_t=u+t\tilde\xi$. Also, since $u$ and $u_t$ agree outside $B_4$ and $u$ is $C^1$ in $\overline B_4$, the $L^\infty$-norms of $\xi$ and $\tilde\xi$ are bounded by a constant independent of $t$ (in fact, depending only on $u$, given the cut-off function $\varphi_4$ defining the diffeormorphism $\Psi_t$). Therefore, from \eqref{2variation} applied to these choices $\xi$ and $\tilde\xi$, adding both inequalities we conclude \eqref{inner-stability}.
\end{proof}
With this consequence of stability in hands, we can now state the following result, which is the analogue of Lemma 2.4 in \cite{CSV}.
We recall that we have set
$$A_4=(\mathbb{R}^n\times \mathbb{R}^n) \setminus (B_4^c\times B_4^c).$$
\begin{lem}\label{lemEFtdelta}
Let $n\ge 2$, $s\in (0,1)$, let $W$ be any $C^3([-1,1])$ function, and $K$ satisfy \eqref{L0} and \eqref{L2}. Let $u: \mathbb{R}^n\to (-1,1)$ be a stable solution of $L_Ku+W'(u)=0$ in $B_6\subset\mathbb{R}^n$.
Then, for every $\nu>0$ there exists $t_0>0$ $($possibly depending on $\nu$, $W$, and $u$$)$ such that
\begin{eqnarray*}
&&\min\left\{ \iint_{A_4}(u(x)-u_t(x))_+(u(\bar x)-u_t(\bar x))_-K(x-\bar x)\,dx \,d\bar x,\right.\\
&&\hspace{2.2em}\left.\iint_{A_4}(u(x)-u_{-t}(x))_+(u(\bar x)-u_{-t}(\bar x))_-K(x-\bar x)\,dx \,d\bar x \right\}\leq \left(\eta+2\nu\right)t^2
\end{eqnarray*}
holds for $t\in(-t_0,t_0)$, where $\eta= C\mathcal E^{\rm Sob}_{B_4}(u)$ and $C$ depends only on $n$, $\lambda$, and $\Lambda$. Here, $(\cdot)_+$ and $(\cdot)_-$ denote the positive and negative parts, respectively, of a function.
\end{lem}
\begin{proof}
We first observe that, since $u$ is stable and is $C^2(B_6)$ (by Appendix~\ref{app-C}), it satisfies both estimates \eqref{eqlemAstables} and \eqref{inner-stability}.
Recall also that $M_t$, $m_t$ (as defined in the beginning of the section), $u_t$, and $u$ all coincide outside $B_4$. Moreover, it holds
\begin{equation}\label{pot}
\mathcal E^{\rm Pot}_{B_4}(M_t)+\mathcal E^{\rm Pot}_{B_4}(m_t)=\mathcal E^{\rm Pot}_{B_4}(u)+\mathcal E^{\rm Pot}_{B_4}(u_t).
\end{equation}
We consider now the Sobolev energy. We claim that
\begin{equation}\label{claimA-ii}
\begin{split}
&|M_t(x)-M_t(\bar x)|^2+|m_t(x)-m_t(\bar x)|^2 - |u(x)-u(\bar x)|^2 - |u_t(x)-u_t(\bar x)|^2\\
&\hspace{2em}=-2(u(x)-u_t(x))_+(u(\bar x)-u_t(\bar x))_-.
\end{split}
\end{equation}
Indeed, we first observe that, if $(u(x)-u_t(x))
(u(\bar x)-u_t(\bar x))\geq 0$, then the right-hand side of \eqref{claimA-ii} vanishes and that the equality is clear.
Assume now that $(u(x)-u_{t}(x))(u(\bar x)-u_{t}(\bar x))< 0$.
Then, by symmetry between $x$ and $\bar x$, we may assume $u(x)>u_{ t}(x)$, $u(\bar x)<u_{t}(\bar x)$. Now, a simple computation shows that
\begin{equation*}
\begin{split}
&|M_t(x)-M_t(\bar x)|^2+|m_t(x)-m_t(\bar x)|^2 - |u(x)-u(\bar x)|^2 - |u_t(x)-u_t(\bar x)|^2\\
&\hspace{1em} = -2 u(x)u_t(\bar x) - 2 u_t(x)u(\bar x) + 2u(x)u(\bar x) +2 u_t(x)u_t(\bar x)\\
&\hspace{1em} = -2(u(x)-u_t(x))(u_t(\bar x)-u(\bar x)).
\end{split}
\end{equation*}
This concludes the proof of \eqref{claimA-ii}.
Using \eqref{pot} and \eqref{claimA-ii}, we deduce that
\begin{eqnarray*}
&&\mathcal E_{B_4}(M_t)+\mathcal E_{B_4}(m_t) +\frac{1}{2} \iint_{A_4} (u(x)-u_{t}(x))_+(u(\bar x)-u_{t}(\bar x))_- K(x-\bar x)\,dx\,d\bar x\nonumber\\
&&\hspace{1em} = \mathcal E_{B_4}(u)+\mathcal E_{B_4}(u_{ t})\nonumber
\end{eqnarray*}
for every $t\in (0,1)$. As noticed in \cite{CSV}, the third term on the left-hand side is the important novelty with respect to the analogue equality in the local case (in which it does not appear). It will be responsible of the $BV$ estimate.
Indeed, analogously we have
\begin{eqnarray*}
&&\mathcal E_{B_4}(M_{-t})+\mathcal E_{B_4}(m_{-t})+ \frac{1}{2}\iint_{A_4} (u(x)-u_{-t}(x))_+(u(\bar x)-u_{- t}(\bar x))_- K(x-\bar x)dxd\bar x\nonumber\\
&&\hspace{1em} =\,\, \mathcal E_{B_4}(u)+\mathcal E_{B_4}(u_{ -t}).\nonumber
\end{eqnarray*}
Adding the last two equalities, and using the key bounds \eqref{eqlemAstables} and \eqref{inner-stability}, we deduce
\begin{eqnarray}\label{step2}
&&\mathcal E_{B_4}(M_t)+\mathcal E_{B_4}(m_t)+\mathcal E_{B_4}(M_{-t})+\mathcal E_{B_4}(m_{-t}) + \nonumber \\
&&\hspace{2em} + \frac{1}{2}\iint_{A_4} (u(x)-u_{t}(x))_+(u(\bar x)-u_{t}(\bar x))_- K(x-\bar x)\,dx\,d\bar x +\\
&&\hspace{2em} + \frac{1}{2}\iint_{A_4} (u(x)-u_{-t}(x))_+(u(\bar x)-u_{- t}(\bar x))_- K(x-\bar x)\,dx\,d\bar x \nonumber\\
&&\hspace{1em} = 2\mathcal E_{B_4}(u)+\mathcal E_{B_4}(u_t)+\mathcal E_{B_4}(u_{-t})\nonumber\\
&&\hspace{1em} \leq 4\mathcal E_{B_4}(u)+\eta t^2\nonumber\\
&&\hspace{1em} \leq \mathcal E_{B_4}(M_t)+\mathcal E_{B_4}(m_t)+\mathcal E_{B_4}(M_{-t})+\mathcal E_{B_4}(m_{-t}) + \big(\eta +2\nu)t^2
\end{eqnarray}
for $t\in(-t_0,t_0)$, with $t_0>0$ small enough (possibly depending on $\nu$, $W$, and $u$). From this, the lemma follows immediately.
\end{proof}
The following result is the analogue of Lemma 2.5 in \cite{CSV}.
\begin{lem}\label{key}
Let $n\geq 2$, $\eta>0$, and $u\in C^1(B_2)$ satisfy $\|u\|_{L^\infty(\mathbb{R}^n)}\le 1$. Assume that for every $\boldsymbol v\in S^{n-1}$, there exists a sequence $t_k\rightarrow 0$, such that
\begin{equation}\label{hp-lemkey}
\limsup_{k\rightarrow \infty}\frac{\|\big(u(\cdot)-u(\cdot-t_k\boldsymbol v)\big)_+\|_{L^1(B_1)}\ \|\big(u(\cdot)-u(\cdot-t_k\boldsymbol v)\big)_-\|_{L^1(B_1)}}{t_k^2}\leq \eta.
\end{equation}
Then, the following estimates hold:
\begin{equation}\label{min-v}
\min\left\{\int_{B_1}(\partial_{\boldsymbol v} u)_+\,dx\,,\,\int_{B_1}(\partial_{\boldsymbol v} u)_-\,dx\right\}\leq \sqrt \eta,
\end{equation}
\begin{equation}\label{max-v}
\max\left\{\int_{B_1}(\partial_{\boldsymbol v} u)_+\,dx\,,\,\int_{B_1}(\partial_{\boldsymbol v} u)_-\,dx\right\}\leq 2 |B_1^{(n-1)}|+\sqrt\eta,
\end{equation}
and
\begin{equation}\label{grad}
\int_{B_1}|\nabla u|\,dx\leq 2n \left( |B_1^{(n-1)}|+\sqrt \eta\right),
\end{equation}
where $B_1^{(n-1)}$ denotes the unit ball of $\mathbb{R}^{n-1}$.
\end{lem}
\begin{proof}
The proof is the same as the one of Lemma 2.5 in \cite{CSV}. Here the situation is even simpler since, being $u\in C^1$, the directional derivatives of $u$ exist in the classical sense.
We sketch the main steps of the proof, for the convenience of the reader.
By assumption \eqref{hp-lemkey}, we have that
$$
\limsup_{k\rightarrow \infty}\frac{\min\left\{\|\big(u(\cdot)-u(\cdot-t_k\boldsymbol v)\big)_+\|_{L^1(B_1)},\, \|\big(u(\cdot)-u(\cdot-t_k\boldsymbol v)\big)_-\|_{L^1(B_1)}\right\}}{|t_k|}\leq \sqrt\eta.
$$
Passing to the limit as $t_k\rightarrow 0$ we immediately get \eqref{min-v}.
To prove \eqref{max-v}, we simply observe that
\[ \int_{B_1}\partial_{\boldsymbol v}u \,dx = \int_{B_1}\lim_{k\rightarrow \infty}\frac{u(x+t_k{\boldsymbol v})-u(x)}{t_k} \,dx =\lim_{k\rightarrow \infty}\frac{\int_{B_1+t_k\boldsymbol v} u \,dx-\int_{B_1}u\,dx}{t_k} \]
and hence, since $\|u\|_{L^\infty(\mathbb{R}^n)}\le 1$,
\[\left|\int_{B_1}\partial_{\boldsymbol v} u\,dx\right|\le \limsup_{k\rightarrow \infty}\frac{\bigl|(B_1+t_k{\boldsymbol v})\setminus B_1\big|
+\bigl|B_1\setminus (B_1+t_k{\boldsymbol v})\big|
}{|t_k|} \le 2 |B_1^{(n-1)}|.\]
This, together with \eqref{min-v}, leads to \eqref{max-v}, since $\partial_{\boldsymbol v}u=(\partial_{\boldsymbol v}u)_+ - (\partial_{\boldsymbol v}u)_-$. Moreover, using that $|\partial_{\boldsymbol v}u|=(\partial_{\boldsymbol v}u)_+ + (\partial_{\boldsymbol v}u)_-$, we also deduce that
\[\int_{B_1}|\partial_{\boldsymbol v} u|\,dx\le 2\left( |B_1^{(n-1)}| + \sqrt \eta \right).\]
Finally, since $|\nabla u|\leq |\partial _{e_1} u|+\ldots + |\partial _{e_n} u|$, we conclude \eqref{grad}.
\end{proof}
Before giving the proof of Theorem \ref{BV}, we recall the following abstract result due to L. Simon \cite{Simon} (see also Lemma 3.1 in \cite{CSV}).
\begin{lem}[\cite{Simon}]\label{lem_abstract}
Let $\beta\in \mathbb{R}$ and $C_0>0$. Let $S: \mathcal B \rightarrow [0,+\infty)$ be a nonnegative function defined on the class $\mathcal B$ of open balls contained in the unit ball $B_1$ of $\mathbb{R}^n$ and satisfying the following subadditivity property:
\[ S(B)\le \sum_{j=1}^N S(B^j) \quad \mbox{ whenever } N\in\mathbb{Z}^+, \{B^j\}_{j=1}^N \subset \mathcal B, \text{ and } B \subset \bigcup_{j=1}^N B^j. \]
It follows that there exists a constant $\delta>0$, depending only on $n$ and $\beta$, such that if
\begin{equation}\label{hp-lem}
\rho^\beta S\bigl(B_{\rho/4}(x_0)\bigr) \le \delta \rho^\beta S\bigl(B_\rho(x_0)\bigr)+ C_0\quad \mbox{whenever }B_\rho(x_0)\subset B_1,
\end{equation}
then
\[ S(B_{1/2}) \le CC_0\]
for some constant $C$ which depends only on $n$ and $\beta$.
\end{lem}
We can now give the proof of Theorem \ref{BV}.
\begin{proof}[Proof of Theorem \ref{BV}]
We divide the proof into two steps.
\vspace{3pt}
{\em -Step 1.} We show that if $s\in (s_0, 1)$ and $u: B_6 \to (-1,1)$ is a stable solution of the semilinear equation $L_Ku + W'(u) =0$ in $B_6$ then, for any given $\delta >0$, we have the estimate
\begin{equation}\label{wnhiotheoiht}
\int_{B_1}|\nabla u| \,dx \le \frac{C_\delta}{1-s} + \delta\, \int_{B_4}|\nabla u| \,dx
\end{equation}
where $C_\delta$ depends only on $\delta$, $n$, $\lambda$, $\Lambda$, and $s_0$ (in particular, it does not depend on~$W$).
Indeed, note that in this setting Lemmas \ref{lem2A}, \ref{lemEFtdelta}, and \ref{key} apply to $u$.
Hence, by Lemma \ref{lemEFtdelta}, for every $\nu >0$ there exists $t_0>0$ such that
\begin{eqnarray*}
&&\min\left\{ \iint_{A_4}(u(x)-u_t(x))_+(u(\bar x)-u_t(\bar x))_-K(x-\bar x)\,dx \,d\bar x,\right.\\
&&\hspace{4em}\left.\iint_{A_4}(u(x)-u_{-t}(x))_+(u(\bar x)-u_{-t}(\bar x))_-K(x-\bar x)\,dx \,d\bar x \right\}\leq \left(\eta+2\nu\right)t^2
\end{eqnarray*}
holds for every $t\in (0,t_0)$, where
\begin{equation}\label{eta}
\eta=C\mathcal E^{\rm Sob}_{B_4}(u)
\end{equation}
and $C$ depends only on $n$, $\lambda$, and $\Lambda$.
Now, by \eqref{L0} and since $s\in (0,1)$, we have $K\ge (2-s) 2^{-n-s}\lambda \ge 2^{-n-1}\lambda$ in~$B_2$. We deduce that there is some sequence $t_k\in(-1,1)$ with
$t_k \rightarrow 0$ such that
$$
\limsup_{k\rightarrow \infty}\frac{\|\big(u(\cdot)-u(\cdot-t_k\boldsymbol v)\big)_+\|_{L^1(B_1)}\ \|\big(u(\cdot)-u(\cdot-t_k\boldsymbol v)\big)_-\|_{L^1(B_1)}}{t_k^2}\leq \eta ,
$$
after changing the value of $C$ in \eqref{eta}.
We can now apply Lemma \ref{key} and, thanks to \eqref{grad}, we arrive at
\begin{equation}\label{key1}
\int_{B_1}|\nabla u|\,dx\leq C \left( 1+ \sqrt{\mathcal E^{\rm Sob}_{B_4}(u)}\right)<\infty,
\end{equation}
where $C$ depends on $n$, $\lambda$, and $\Lambda$.
In order to keep track of the precise dependence of the constants on $s$, as $s\uparrow 1$, in what follows $C$ will denote (possibly different) positive constants which depend only on $n$, $\lambda$, $\Lambda$, and $s_0$.
Defining $V(z):= |\nabla u(z)|$ for $z\in B_4$ and $V(z):= 0$ for $z\in \mathbb{R}^n\setminus B_4$, and given $x$ and $\bar x$ both in $B_4$, note that we have
$$
|u(x)-u(\bar x)| =\left| \int_0^1 (\bar x -x ) \cdot \nabla u(x+t(\bar x -x))\,dt\right|
\leq |x -\bar x| \int_0^1 V(x+t(\bar x -x))\,dt.
$$
Using this, that $K\in \mathcal L_2$, and $\|u\|_{L^\infty(\mathbb{R}^n)}\leq 1$, we deduce (thanks to the presence of the factor $1-s$ on the next left-hand side) that
\begin{eqnarray}\label{key2}
(1-s)\mathcal E^{\rm Sob}_{B_4}(u)
&\le & \frac{1}{2} \Lambda(1-s) \bigg(\iint_{B_4\times B_4} \frac{|u(x)-u(\bar x)|^2}{|x-\bar x|^{n+s}} \,dx\,d\bar x \nonumber \\
& & \hspace{4cm}
+ 2\iint_{B_4\times B_4^c} \frac{2^2}{|x-\bar x|^{n+s}}\,dx \,d\bar x\bigg)\nonumber
\\
& \le& \frac{1}{2} \Lambda (1-s)\iint_{B_4\times B_4} \frac{|u(x)-u(\bar x)|^2}{|x-\bar x|^{n+s}} \,dx\,d\bar x + C\nonumber
\\
&\le & \Lambda (1-s)\iint_{B_4\times B_4} \frac{| u(x)-u(\bar x)|}{|x-\bar x|^{n+s}} \,dx\,d\bar x + C\nonumber
\\
&\leq & \Lambda(1-s)\int_{B_4} dx \int_ {B_8} dy \, |y|^{1-n-s} \int_0^{1} dt \,V(x + ty) +C\nonumber
\\
&\leq& \Lambda(1-s)\int_ {B_8} dy \, |y|^{1-n-s} \int_0^{1} dt \int_{\mathbb{R}^n} dx \,V(x + ty) +C\nonumber
\\
& = & \Lambda(1-s)\int_ {B_8} dy \, |y|^{1-n-s} \int_{\mathbb{R}^n}dz \, V(z) +C\nonumber \\
&\le & C\int_{\mathbb{R}^n}V(z)\,dz +C \nonumber \\
&=& C\left( 1 + \int_{B_4}|\nabla u|\,dx \right),
\end{eqnarray}
where $C$ depends only on $n$, $\lambda$, $\Lambda$, and $s_0$; in particular, it stays bounded as $s\uparrow 1$.
Hence, \eqref{key1}, \eqref{key2}, and the Cauchy-Schwarz inequality lead to
\begin{equation}\label{key3}
\begin{split}
\int_{B_1}|\nabla u| \,dx &\le C\left(1+ \frac{1}{(1-s)^{1/2}}\left(1+\int_{B_4}|\nabla u| \,dx \right)^{1/2}\right)
\\
& \le C\left(1+\frac{1}{\delta(1-s)}+\delta\right) + \delta\, \int_{B_4}|\nabla u| \,dx
\end{split}
\end{equation}
for all $\delta>0$.
\vspace{3pt}
{{\em -Step 2.} Now, in order to conclude the proof of the theorem we observe that, since its statement is scaling invariant (it is independent of $W$), we may assume without loss generality that $R=1$. So, let now $u: \mathbb{R}^n \to (-1,1)$ be a stable solution of $L_Ku+W'(u)=0$ in $B_{2R}=B_2$. Given $B_\rho(x_0)\subset B_1$ then the rescaled function $\tilde u(y) =u\left( x_0 +\rho y/4\right)$ is a stable solution in $B_6$ (since $x_0+\rho B_1 \subset B_1 \Rightarrow x_0+\frac{\rho}{4} B_6 \subset B_2$)} of the equation $L_{\tilde K}\tilde u+(\rho/4)^s W'(\tilde u)=0$,
where
\[ \tilde K(z) : = {(\rho/4)}^{n+s} K(\rho z/4) \quad \mbox{belongs again to }\mathcal L_2(s,\lambda, \Lambda).\]
Thus, rescaling the estimate \eqref{wnhiotheoiht}, which applies to $\tilde u$, we obtain
\begin{equation}
\rho^{1-n} \,\int_{B_{\rho/4}(x_0)}|\nabla u| \,dx \le \delta \,\rho^{1-n}\, \int_{B_{\rho}(x_0)}|\nabla u|\,dx + \frac{C_\delta}{1-s}
\end{equation}
for all $\delta>0$, where $C_\delta$ depends only on $\delta$, $n$, $\lambda$, $\Lambda$, and $s_0$.
Therefore, considering the subadditive function
\[S(B):=\int_{B}|\nabla u|\,dx \]
on the class of balls,
and taking $\beta:=1-n$ and $\delta$ as given by Lemma \ref{lem_abstract} (and hence depending only on $n$), we find that
\[ \int_{B_{1/2}}|\nabla u| \,dx = S(B_{1/2}) \le \frac{C}{1-s},\]
where $C$ is a constant which depends only on $n$, $\lambda$, $\Lambda$, and $s_0$.
By scaling and using a standard covering argument, we obtain the same estimate with $B_{1/2}$ replaced by $B_1=B_R$, concluding the proof.
\end{proof}
\begin{proof}[Proof of Corollary \ref{energy-est}]
Let $s\in (0,1)$. Let $u_R: = u(R\, \cdot\,)$. We combine the estimate $\int_{B_1}|\nabla u_R|\,dx \le C(1-s)^{-1}$ of Theorem \ref{BV} with the interpolation type inequality
\[(1-s)\mathcal E^{\rm Sob}_{B_1}(u_R) \le C\bigl(1+\int_{B_1}|\nabla u_R|\,dx\bigr)\] ---which we proved, in a different ball, in \eqref{key2}--- to obtain $\mathcal E^{\rm Sob}_{B_1}(u_R) \le C (1-s)^{-2}$. Now, the corollary follows after rescaling.
\end{proof}
\section{Sobolev energy controls potential energy}\label{sec-4}
In this section we give the proof of Proposition \ref{DircontrolsPot2}. In particular, we will have $s\in (0,2]$ and we include the case of the classical Allen-Cahn equation. All results in this section require the cubes to have large enough diameter and, in particular, the equation to be posed in a sufficiently large ball (as in the statement Proposition~\ref{DircontrolsPot2}).
The proof is based on a suitable cube decomposition ---which may be of independent interest--- and covering arguments.
We identify two types of cubes: of `type I'' and of ``type~II''.
Cubes of type I will be those containing at least one point where $u\sim 0$.
Cubes of type II are cubes in which either $u\sim1$ or $u\sim -1$ in the whole cube.
By cube we always mean a set of the form $Q=x_0+(-l,l)^n$ for some $x_0\in\mathbb{R}^n$ and $l>0$.
The following three lemmas are used in the proof of Proposition \ref{DircontrolsPot2}.
Essentially, the first one (Lemma \ref{tuttookintype1}) is used to show that the contribution of a cube of type~I to the total Sobolev energy controls its contribution to the potential energy. The third lemma (Lemma \ref{tuttookintype2}) will prove this same property for cubes of type II.
The second lemma states that if a cube is not of type I (i.e., it contains no points where $u\sim0$) then the corresponding half-cube is of type II (i.e., we have either $u\sim 1$ or $u\sim -1$ in all of it). Thus, through an appropriate covering argument, this second lemma establishes that there is essentially a dichotomy between cubes of type I and cubes of type II.
Throughout this section $W$ is the double-well potential $\frac{1}{4}(1-u^2)^2$. It satisfies
\begin{equation}\label{WWW}
\begin{cases}
-W''(t)= -W''(-t) \ge \nu_0 \quad & \mbox{for } 0\le t\le c_0
\\
W''(t)= W''(-t) \ge \nu_0 \quad &\mbox{for } 1-c_0 \le t\le 1
\\
-W'(t) = W'(-t) \ge \nu_1 \quad & \mbox{for } \frac{c_0}{2}\le t\le 1-c_0
\end{cases}
\end{equation}
for some constants $c_0\in (0,\frac{2}{3})$, $\nu_0>0$, and $\nu_1>0$ which are totally universal. To check quickly their existence, note that once $c_0$ and $\nu_0$ have been chosen small enough to guarantee the first two properties, the third condition will be satisfied for $\nu_1$ small enough.
\begin{lem}\label{tuttookintype1}
Given $s_0>0$, let $n\ge 2$, $s\in (s_0,2]$, $W(u)=\frac 1 4 (1-u^2)^2$, and $K$ satisfy \eqref{L0} and \eqref{L2}. Let $R>0$ and $u:\mathbb{R}^n\rightarrow (-1,1)$ be a stable solution of $L_{K}u + W'(u)=0$ in~$B_{2R}$.
Then, there exist constants $D_0\ge 1$ and $d_0>0$, which depend only on $n$, $\lambda$, $\Lambda$, and $s_0$, such that the following statement holds true.
For every cube $Q\subset B_R$ satisfying ${\rm diam}(Q)\ge D_0$ and
\[
\{ |u|\le c_0/2\}\cap Q\neq\varnothing
\]
for the constant $c_0$ in \eqref{WWW}, there are two cubes $Q^{(1)}$, $Q^{(2)}$ contained in $Q$ and with diameters $d_0$ for which
\begin{equation}\label{goalQi}
\inf_{Q^{(1)}} u - \sup_{Q^{(2)}} u \ge c_0/4.
\end{equation}
\end{lem}
\begin{proof}
Within the proof, $C$ will denote different positive constants which depend only $n$, $\lambda$, $\Lambda$, and $s_0$. The constants $D_0$ and $d_0$ will be chosen to have this same dependence.
Let $Q\subset B_R$ be some cube satisfying ${\rm diam}(Q)=D \ge D_0$.
\vspace{3pt}
-{\em Step 1.} Let us prove that ${\rm osc }_Q u \ge c_0/2$.
Indeed, arguing by contradiction assume that
\[
{\rm osc }_Q u < c_0/2, \quad \mbox{and recall that we have }\{ |u|\le c_0/2\}\cap Q\neq\varnothing.
\]
Then, $|u|\le c_0$ in $Q$ and using \eqref{WWW} and the stability inequality \eqref{stable}, we obtain
\begin{eqnarray}\label{ineq2-s}
\nu_0 \int_{\mathbb{R}^n} \xi^2\,dx &\le &-\int_{\mathbb{R}^n} W''(u) \xi^2\,dx \nonumber
\\
&\le & \frac{1}{2}\iint_{\mathbb{R}^n\times \mathbb{R}^n} \big|\xi(x)-\xi(\bar x)\big|^2 K(x-\bar x)\,dx \,d\bar x\nonumber
\\
&\le& C(2-s) \iint_{\mathbb{R}^n\times \mathbb{R}^n} \frac{\big|\xi(x)-\xi(\bar x)\big|^2}{|x-\bar x|^{n+s}} \,dx \,d\bar x
\end{eqnarray}
for every $\xi\in C^1_c(Q)$.
Now, if the diameter $D$ of $Q$ is large enough we may contradict this inequality just by scaling. Indeed, denote by $Q_d$ the cube centered at $0$ and with diameter $d$, and let $x_0$ be the center of $Q$ and $D$ its diameter. Choose (universally) a function $\eta\not\equiv 0$ in $C^\infty_c(Q_1)$. Take now $\xi(x)= \eta((x-x_0)/D)$ and notice that
\begin{equation}\label{ineq2-sbis}
(2-s)\iint_{\mathbb{R}^n\times \mathbb{R}^n} \frac{\big|\xi(x)-\xi(\bar x)\big|^2}{|x-\bar x|^{n+s}} \,dx \,d\bar x \le C D^{n-s} \Vert\nabla\eta\Vert_{L^\infty(Q_1)}^2.
\end{equation}
while
\begin{equation}\label{ineq2-sbis2}
\int_{\mathbb{R}^n} \xi^2\,dx = cD^n
\end{equation}
Choosing $D\geq D_0\geq 1$ large enough (note that $D_0^{n-s}\leq D_0^{n-s_0}$), we contradict \eqref{ineq2-s}. This proves Step 1.
Note that the same argument applies in the case $s=2$ when we replace $2-s$ times the double integral by the classical Dirichlet norm.
\vspace{3pt}
-{\em Step 2}. We now use Step 1, and regularity estimates for the equation,
to show that there are two cubes $Q^{(1)}$, $Q^{(2)}$ contained in $Q$ and with diameters $d_0$ such that $\inf_{Q^{(1)}} u - \sup_{Q^{(2)}} u \ge c_0/4$.
Indeed, first note that, as shown in Appendix~\ref{app-C}, we have the estimate $|\nabla u|\le \bar C$ in $B_R$, for some constant $\bar C$ depending only on the quantities stated in the beginning of the proof; here we use that $Q\subset B_R$ and thus $ \textrm{diam}(B_R) \ge \textrm{diam}(Q)=D\ge D_0\ge 1$. Let
$$
d_0= \min \left\{ D_0, \frac{c_0}{8 \bar C}\right\}.
$$
Now, by Step 1 there are two points $x_1,x_2\in \overline{Q}$ such that $u(x_1)-u(x_2)\ge c_0/2$.
Let $Q^{(i)}$ be any two cubes with diameter $d_0$ such that $x_i\in \overline{Q^{(i)}}\subset \overline Q$ (recall that $d_0\le D_0\le \textrm{diam}(Q)$). Then we readily show that \eqref{goalQi} is satisfied by these cubes.
\end{proof}
We can now state the second lemma.
\begin{lem}\label{notype1impliestype2}
Given $s_0>0$, let $n\ge 2$, $s\in (s_0,2]$, $W(u)=\frac 1 4 (1-u^2)^2$, and $K$ satisfy \eqref{L0} and \eqref{L2}. Let $R>0$ and $u:\mathbb{R}^n\rightarrow (-1,1)$ be a stable solution of $L_{K}u + W'(u)=0$ in~$B_{2R}$.
Then, there exists a constant $D_0\ge 1$, which depends only on $n$, $\lambda$, $\Lambda$, and $s_0$, such that the following statement holds true.
For every cube $Q\subset B_R$ with ${\rm diam}(Q)\ge D_0$ we have
\begin{equation}\label{firstcase}
u \ge 1-c_0 \mbox{ in }Q' \quad\text{ if } \, u > c_0/2 \mbox{ in }Q \
\end{equation}
and
\begin{equation}\label{secondcase}
u \le -1+c_0 \mbox{ in }Q' \quad\text{ if } \, u <-c_0/2 \mbox{ in }Q ,
\end{equation}
where $c_0$ is the constant in \eqref{WWW} and $Q'$ is the cube with the same center as $Q$ and half its diameter.
\end{lem}
\begin{proof}
Within the proof, the constant $C$ will depend only $n$, $\lambda$, $\Lambda$, and $s_0$. The constant $D_0$ will be chosen to have this same dependence.
Let $Q\subset B_R$ be some cube satisfying ${\rm diam}(Q)=D \ge D_0$.
We prove only \eqref{firstcase} since, by the even symmetry of $W$, \eqref{secondcase} follows applying \eqref{firstcase} to $-u$.
Assume that $u > c_0/2$ in $Q$. Using \eqref{WWW} we have
\[ L_Ku =- W'(u)\ge \nu_1\quad\mbox{in }Q\setminus \{u\ge 1-c_0\}.\]
Let $\eta\in C^\infty(\mathbb{R}^n)$ with $\eta \equiv 1-c_0$ in $Q_{1/2}$, $-1\le \eta\le 1-c_0$ in $Q_1$, and $\eta \equiv -1$ outside $Q_1$ (recall that $Q_r$ denotes the cube centered at $0$ and with diameter $r$). Consider
\[\tilde\eta(x)= \eta((x-x_0)/D)\]
where $x_0$ is the center of $Q$ and $D$ its diameter. Note that $D\ge D_0$ and that, by taking $D_0$ large enough, we will have
\[ \big| L_K \tilde\eta \big| \le C D^{-s}\le C D_0^{-s} \le \nu_1 \]
with $C$ and $D_0$ as stated in the beginning of the proof.\footnote{{Here we use the presence of the factor $2-s$ on the upper bound for the kernel $K$ as in \eqref{ineq2-sbis}. Instead, when $s=2$, the bound is obvious.}}
Let us then show that
\begin{equation}\label{goal}
u\ge \tilde \eta \quad\mbox{in }Q.
\end{equation}
Indeed, let $U= Q\setminus \{u\ge 1-c_0\}$. Since $ \tilde \eta\le 1-c_0$ in $Q$ and $\tilde \eta\equiv -1$ outside of $Q$ we have $\tilde\eta\le u$ outside of $U$. On the other hand
\[ L_K (u-\tilde \eta) \ge L_K u - \big| L_K \tilde\eta\big| \ge 0 \quad \mbox{in }U,\]
and thus the maximum principle leads to \eqref{goal}. Finally, since by construction $\tilde \eta \equiv 1-c_0$ in $Q'$ the lemma follows.
\end{proof}
{
Finally, we can state our last auxiliary lemma in this section.
\begin{lem}\label{tuttookintype2}
Given $s_0>0$, let $n\ge 2$, $s\in (s_0,2]$, $W(u)=\frac 1 4 (1-u^2)^2$, and $K$ satisfy \eqref{L0} and \eqref{L2}. Let $R>0$ and $u:\mathbb{R}^n\rightarrow (-1,1)$ be a stable solution of $L_{K}u + W'(u)=0$ in~$B_{2R}$.
Assume that for some cube $Q\subset B_R$ we have that
\[
1-|u| \le c_0 \quad \mbox{in } Q,
\]
where $c_0$ is the constant in \eqref{WWW}.
Then
\begin{equation}\label{notseq2}
(2-s)\int_Q \int_{\mathbb{R}^n} \frac{|u(x)-u(\bar x)|^2}{|x-\bar x|^{n+s}} \,dx\,d\bar x \ge \kappa_0\int_{Q} (1-|u|)^2\,dx,
\end{equation}
where $\kappa_0>0$ is a constant depending only on $n$, $\lambda$, $\Lambda$, $s_0$, and ${\rm diam}(Q)$.
When $s=2$, the left hand side of \eqref{notseq2} must be replaced by $\int_Q |\nabla u|^2\,dx$.
\end{lem}
\begin{proof}
We will do the proof in the case $1-c_0\le u\le 1$ in $Q$ (the case $-1\le u \le -1+c_0$ is similar).
Since for simplicity the lemma is stated for the explicit quartic potential $W(u) = \frac 1 4 (1-u^2)^2$ we may take $c_0 := 1 -1/\sqrt 2$, and hence
\begin{equation} \label{21huwgiuw}
W'' = 3u^2-1 \in [1/2,2] \quad \mbox{for } u\in [1-c_0, 1].
\end{equation}
Let $v : = 1-u$. By assumption $v\ge 0$ in all of $\mathbb{R}^n$ and $v\le c_0$ in $Q$. Hence, using \eqref{21huwgiuw} and
\[
L_K v = -L_K u = W'(u) = W'(u)-W'(1)
\]
we obtain
\begin{equation}
\label{wjhtiohweio}
-2 v \le L_K v \le -\frac{1}{2}v \quad \mbox{in }Q.
\end{equation}
Notice that
\[
(2-s)\int_{Q} \int_{\mathbb{R}^n} \frac{|u(x)-u(\bar x)|^2}{|x-\bar x|^{n+s}} \,dx\,d\bar x = (2-s)\int_{Q} \int_{\mathbb{R}^n} \frac{|v(x)-v(\bar x)|^2}{|x-\bar x|^{n+s}} \,dx\,d\bar x
\]
and
\[
\int_{Q} W(u) dx \le \int_{Q} \frac { 2^2(1-u)^2}{4} dx = \int_{Q} v^2dx.
\]
Hence, to prove the lemma we need to show that there exists a constant $C$ depending only on $n$, $\lambda$, $\Lambda$, $s_0$, and ${\rm diam}(Q)$, such that
\[
C (2-s)\int_{Q} \int_{\mathbb{R}^n} \frac{|v(x)-v(\bar x)|^2}{|x-\bar x|^{n+s}} \,dx\,d\bar x \ge \int_{Q} v^2dx.
\]
To prove this, let $\tilde v : = v/ \|v\|_{L^2(Q)}$ and let
\[
\kappa: = (2-s)\int_{Q} \int_{\mathbb{R}^n} \frac{|\tilde v(x)-\tilde v(\bar x)|^2}{|x-\bar x|^{n+s}} \,dx\,d\bar x.
\]
In the remaining of the proof we bound $\kappa$ by below.
First, by the fractional Poincar\'e inequality we have
\[
\kappa \ge (2-s) \int_{Q} \int_{Q} \frac{|\tilde v(x)-\tilde v(\bar x)|^2}{|x-\bar x|^{n+s}} \,dx\,d\bar x
\ge c_0 \int_{Q} (\tilde v-t)^2,
\]
where $t = \frac 1{|Q|}\int_{Q} \tilde v\, dx$ and $c_0=c_0(n,s_0, {\rm diam}(Q))>0$.
Note that, by the triangle inequality, we have
\begin{equation}
\begin{split}
1- (\kappa/c_0)^{1/2} &\le \|\tilde v\|_{L^2(Q)}- \|\tilde v-t\|_{L^2(Q)} \\&
\le t|Q|^{1/2} \le \|\tilde v\|_{L^2(Q)} + \|\tilde v-t\|_{L^2(Q)} \le 1+ (\kappa/c_0)^{1/2}.
\end{split}
\end{equation}
Hence, if $\kappa$ is sufficiently small (which we may of course assume) we have
\begin{equation}\label{wtiowhtoiwh}
\frac 1 2| Q|^{-1/2}\le t \le 2|Q|^{-1/2}.
\end{equation}
On the other hand, $\tilde v$ satisfies $L_K \tilde v \le -\frac 1 2\tilde v$. Thus, if we fix $\xi \in C^\infty_c (Q')$ such that $\int_{\mathbb{R}^n} \xi \,dx =1$ and $\xi\ge0$, where $Q'$ is the cube with the same center as $Q$ and half its diameter, we have
\begin{equation}\label{wtnwowneoktgbeowb}
\begin{split}
\int_{\mathbb{R}^n} \tilde v L_K\xi \,dx &\le -\frac{1}{2} \int_{\mathbb{R}^n} \tilde v \xi \,dx \le -\frac 1 2 \int_{\mathbb{R}^n} t \xi \,dx + \int_{\mathbb{R}^n} |\tilde v -t| \xi \,dx \\ &\le - \frac t 2 + \|\xi\|_{L^2(Q)} \|\tilde v -t\|_{L^2(Q)}
\le - \frac t 2 + C (\kappa/c_0)^{1/2}.
\end{split}
\end{equation}
At the same time, since $\int_{\mathbb{R}^n} L_K\xi \,dx =0$ we have
\[
\int_{\mathbb{R}^n} \tilde v L_K\xi \,dx= \int_{Q} (\tilde v-t) L_K\xi \,dx + \int_{\mathbb{R}^n\setminus Q} (\tilde v-t) L_K\xi \,dx.
\]
Now similarly as before
\[
\bigg|\int_{Q} (\tilde v-t) L_K\xi \,dx\bigg| \le \|L_K\xi\|_{L^2(Q)} \|\tilde v -t\|_{L^2(Q)} \le C (\kappa/c_0)^{1/2}.
\]
Also, since $|L_K\xi (x)| \le C (2-s) {\rm dist }(x,Q')^{-n-s}$ for $x\in \mathbb{R}^n\setminus Q$, we have
\[
\bigg|\int_{\mathbb{R}^n\setminus Q} (\tilde v-t) L_K\xi \,dx \bigg| \le C (2-s) \int_{\mathbb{R}^n\setminus Q} \frac{|\tilde v-t|}{ {\rm dist }(x,Q')^{n+s}} \,dx.
\]
Now, since
\[
\int_{\mathbb{R}^n\setminus Q} \frac{2-s}{ {\rm dist }(x,Q')^{n+s}}\,dx \le C
\]
and
\[
{\rm dist }(x,Q')^{-n-s} \le C |x-\bar x|^{-n-s} \quad \mbox{for all } \bar x \in Q
\]
we obtain, using also the definition of $t$,
\[
\begin{split}
(2-s) \int_{\mathbb{R}^n\setminus Q} \frac{|\tilde v(x) -t|} { {\rm dist }(x,Q')^{n+s}} \,dx &\le (2-s) \frac 1 {|Q|} \int_Q d\bar x \int_{\mathbb{R}^n\setminus Q}\,dx \, (2-s)\frac{|\tilde v(x) -\tilde v(\bar x)|} { {\rm dist }(x,Q')^{n+s}}
\\
&\le C \bigg( \int_Q d\bar x \int_{\mathbb{R}^n\setminus Q}\,dx \, (2-s)\frac{|\tilde v(x) -\tilde v(\bar x)|^2} { {\rm dist }(x,Q')^{n+s}} \bigg)^{1/2}
\\
&\le C \bigg( \int_Q \int_{\mathbb{R}^n\setminus Q} (2-s) \frac{|\tilde v(x)-\tilde v(\bar x)|^2}{|x-\bar x|^{n+s}} \,dx\,d\bar x \bigg)^{1/2}
\\
& \le C \kappa^{1/2}.
\end{split}
\]
Hence, we have shown
\[
\bigg| \int_{\mathbb{R}^n} \tilde v L_K\xi \, dx\bigg| \le C \kappa^{1/2}.
\]
Using also \eqref{wtnwowneoktgbeowb}, we deduce
\[
\frac t 2 \le -\int_{\mathbb{R}^n} \tilde v L_K\xi + C\kappa^{1/2} \le 2C\kappa^{1/2}.
\]
Hence, recalling \eqref{wtiowhtoiwh}, this shows the desired lower bound on $\kappa$.
\end{proof}
}
We can now give the
\begin{proof}[Proof of Proposition \ref{DircontrolsPot2}]
Throughout the proof, all the constants will depend only $n$, $s$, $\lambda$, $\Lambda$, and $s_0$. Let $D_0$ and $d_0$ be constants for which Lemmas \ref{tuttookintype1} and \ref{notype1impliestype2} hold ---for this, take them to be, respectively, the largest of the constants $D_0$ in the lemmas and the smallest of the constants $d_0$.
Let $\mathcal F$ denote the family of cubes $Q$ with center in the lattice $(D_0/\sqrt{n})\, \mathbb Z^n$ and side-length $2D_0/\sqrt{n}$, that is, of the form
$$
Q=\frac{D_0}{\sqrt{n}}\,z_0+\left( -\frac{D_0}{\sqrt{n}},\frac{D_0}{\sqrt{n}} \right)^n \quad \text{for some } z_0\in\mathbb{Z}^n.
$$
Let $\mathcal F_R$ be the family of those cubes $Q\in \mathcal F$ such that $Q\subset B_R$.
Given a cube $Q$, denote by $Q'$ the cube with the same center as $Q$ and half its diameter. Finally, define
$$
R_0:=\frac{3}{2}D_0,
$$
and recall that we assume $R>R_0$.
It is then easy to check that:
\begin{itemize}
\item[(a)]
For all $Q\in \mathcal F_R$, we have $Q\subset B_R$ and $\textrm{diam}(Q)=2D_0$.
\vspace{1mm}
\item[(b)] Each point of $\mathbb{R}^n$ belongs to at most $2^n$ cubes of the family $\mathcal F_R$.
\vspace{1mm}
\item[(c)] The union $\bigcup_{Q\in {\mathcal F_R}} \, Q'$ is disjoint and covers $B_{R-R_0}$ except for a set of measure zero.
\end{itemize}
Throughout the proof, it is easy to check that all arguments hold true when $s=2$ by replacing, within the double integrals, $2-s$ times one of the integrals by the pointwise gradient squared.
We first notice that, by properties (a) and (b),
\begin{equation}\label{chain}
\begin{split}
\mathcal E^{\rm Sob}_{B_R}(u) &\ge \frac{1}{4} \int_{B_R}\int_{\mathbb{R}^n} |u(x)-u(\bar x)|^2 K(x-\bar x)\,dx\,d\bar x
\\
&\ge
\frac{\lambda(2-s)}{4} \int_{B_R}\int_{\mathbb{R}^n} \frac{ |u(x)-u(\bar x)|^2 }{ |x-\bar x|^{n+s}} \,dx\,d\bar x
\\
&\ge \frac{\lambda(2-s)}{4\cdot2^n} \sum_{Q\in \mathcal F_R} \int_{Q}\int_{\mathbb{R}^n} \frac{ |u(x)-u(\bar x)|^2 }{ |x-\bar x|^{n+s}} \,dx\,d\bar x.
\end{split}
\end{equation}
Next, for each $Q\in \mathcal F_R$ we have the following dicothomy. Either
\begin{equation}\label{type1}
\{ |u|\le c_0/2\}\cap Q\neq\varnothing
\end{equation}
or
\begin{equation}\label{type2}
\{ |u|\le c_0/2\}\cap Q =\varnothing.
\end{equation}
Now, if $Q\in \mathcal F_R$ satisfies \eqref{type1} then by Lemma \ref{tuttookintype1} (since $Q$ has diameter $2D_0$) there are two cubes $Q^{(1)}$ and $Q^{(2)}$ contained in $Q$, and with diameters $d_0$, such that
\[
\inf_{Q^{(1)}}u -\sup_{Q^{(2)}} u \ge c_0/4.
\]
It thus follows, using the fractional Poincar\'e inequality and \[\int_{Q'} W(u)\,dx\le |Q'| \max_{[-1,1]} W \le C,\] that
\begin{equation}\label{caseI}
\begin{split}
C (2-s)\int_{Q}\int_{Q} \frac{ |u(x)-u(\bar x)|^2 }{ |x-\bar x|^{n+s}}\,dx\,d\bar x &\ge \inf_{t\in \mathbb{R}} \int_{Q} |u(x)-t|^2 \,dx
\\ &
\ge \left|\frac{(\inf_{Q^{(1)}}u -\sup_{Q^{(2)}} u)^2 }{2}\right|^2 \big| Q^{(i)}\big|
\\
&\ge \left(\frac{c_0}{8}\right)^2 \bigg( \frac{d_0}{\sqrt n}\bigg)^n \ge c\int_{Q'} W(u) \, dx
\end{split}
\end{equation}
for some constant $c>0$.
On the other hand, if $Q$ satisfies \eqref{type2} then by Lemma \ref{notype1impliestype2} we have that
\[ |u|\ge 1-c_0 \quad \mbox{in }Q'.\]
Hence, using Lemma \ref{tuttookintype2} (with $Q$ replaced by $Q'$) we have
\begin{equation}\label{caseII}
(2-s)\int_{Q}\int_{\mathbb{R}^n} \frac{ |u(x)-u(\bar x)|^2 }{ |x-\bar x|^{n+s}}\,dx\,d\bar x \ge c\int_{Q'} (1-|u|)^2\,dx \ge c\int_{Q'} W(u)\,dx.
\end{equation}
Finally, recalling property (c) above, we combine \eqref{caseI} and \eqref{caseII} with \eqref{chain} to obtain
\[
\begin{split}
\mathcal E^{\rm Sob}_{Q_R}(u) &\ge \frac{\lambda(2-s)}{2^n} \sum_{Q\in \mathcal F_R} \int_{Q}\int_{\mathbb{R}^n} \frac{ |u(x)-u(\bar x)|^2 }{ |x-\bar x|^{n+s}} \,dx\,d\bar x
\\
&\ge c \sum_{Q\in \mathcal F_R} \int_{Q''} W(u) \, dx
\\
&\ge c \int_{B_{R-R_0}} W(u) \, dx
\\
&= c\ \mathcal E^{\rm Pot}_{B_{R-R_0}}(u),
\end{split}
\]
that finishes the proof.
\end{proof}
Let us give now the proof of the bound for the full energy when $s\in (0,1)$.
\begin{proof} [Proof of Theorem \ref{thm1}]
Let $s\in (s_0,1)$. By Corollary \ref{energy-est} we have that
\begin{equation}\label{est-Dir}\mathcal E_{B_{R}}^{\rm Sob}(u)\leq \frac{C}{(1-s)^2}R^{n-s}.
\end{equation}
Now, using Proposition \ref{DircontrolsPot2} we deduce from \eqref{est-Dir} the same estimate for ${\mathcal E}_{B_{R/2}}^{\rm Pot}(u)$, provided $R\ge 2 R_0$ (since $R-R_0\geq R/2$). The same bound for ${\mathcal E}_{B_{R/2}}^{\rm Pot}(u)$ is obvious when $1\le R\le 2R_0$.
Finally, a standard covering and scaling argument allows to control the full energy in $B_R$ instead of $B_{R/2}$.
\end{proof}
We close this section with a useful lemma concerning the decay towards $\pm 1$ of solutions $u_\varepsilon$ of $(-\Delta)^s u_\varepsilon +\varepsilon^{-s}W'(u_\varepsilon)=0$ for points $x$ away from $\{|u_\varepsilon| \le \frac 9{10}\}$.
It will be used in the proof of Proposition \ref{propnew}.
\begin{lem}\label{lemdecay}
Given $s_0>0$, let $n\ge 2$, $s\in (s_0,2]$, $W(u)=\frac 1 4 (1-u^2)^2$, and $K$ satisfy \eqref{L0} and \eqref{L2}. Let $u_\varepsilon :\mathbb{R}^n\rightarrow (-1,1)$ be a solution of $L_{K}u + \varepsilon^{-s}W'(u)=0$ in~$B_{2r}$, with $r\ge \varepsilon>0$, satisfying $1-|u_\varepsilon| \le c_0$ in~$B_{2r}$, where $c_0$ is the constant in \eqref{WWW}.
Then,
\[
0\le 1-|u_\varepsilon| \le C \Big(\frac{\varepsilon}{r}\Big)^s \quad \mbox{in } B_r,
\]
where $C$ depends only on $n$, $\lambda$, $\Lambda$, and $s_0$.
\end{lem}
\begin{proof}
By scaling ---i.e., replacing $u$ by $u(\varepsilon\,\cdot\,)$--- it is enough to consider the case $\varepsilon=1$.
As in the proof of Lemma \ref{tuttookintype2}, for the explicit quartic potential $W$ we may take $c_0 := 1 -1/\sqrt 2$ and suppose $1-c_0\le u\le 1$ in $B_{2r}$. Hence, \eqref{21huwgiuw} holds and,
similarly as in \eqref{wjhtiohweio}, the function $v: = 1-u$ satisfies
\begin{equation}
\label{wjthoihwhe}
L_K v \le -\tfrac{1}{2}v \quad \mbox{in }B_{2r}.
\end{equation}
Fix now some smooth radial ``bowl-type'' function $\xi$ satisfying $\chi_{\mathbb{R}^n \setminus B_1} \ge \xi \ge \chi_{\mathbb{R}^n \setminus B_2}$, and let $\xi_r(x) := \xi (x/r)$.
By scaling we have $|L_K \xi_r | \le C_0 r^{-s}$ in $\mathbb{R}^n$.
Let now $w :=2 \xi_r + b$, with $b := 4C_0 r^{-s}$. We have $L_K w \ge -2C_0 r^{-s} = - \frac 1 2 b \ge - \frac 1 2 w$ in $B_{2r}$. Hence, recalling \eqref{wjthoihwhe} and using that $w\ge 2+b \ge 2\ge v$ in
$\mathbb{R}^n\setminus B_{2r}$ we obtain ---by the maximum principle--- $w\ge v$ in $\mathbb{R}^n$. Hence using that $\xi_r\equiv 0$ in $B_r$ we have shown $b\ge v$ in $B_r$, and the lemma follows.
\end{proof}
\section{Density estimates}\label{sec-5}
In this section we establish density estimates when $s\in(0,1)$. We need to restrict our attention to the case of the fractional Laplacian since a crucial ingredient in the proof is the monotonicity formula from \cite{CC2}. This monotonicity formula is known to hold only for equations involving the fractional Laplacian, since its proof relies on the so-called Caffarelli-Silvestre extension~\cite{CafSi}, that we recall here below.
By the result in \cite{CafSi}, $u: \mathbb{R}^n \to (-1,1)$ is a solution of $(-\Delta)^{s/2}u+W'(u)=0$ in~$\mathbb{R}^n$, for $s\in (0,2)$, if and only if $u(x)=U(x,0)$ where $U: \mathbb{R}^{n+1}_+\to (-1,1)$, defined in the half-space $\mathbb{R}^{n+1}_+:= \{(x,y) \, : \, x\in \mathbb{R}^n, y>0\}$, is a solution of
\begin{equation}\label{ext}
\begin{cases}
\mbox{div}(y^{1-s}\nabla U(x,y))=0 & \mbox{in}\;\;\mathbb{R}^{n+1}_+\\
-d_s \lim_{y\rightarrow 0}y^{1-s}\partial_y U(x,y)= -W'(u(x))& \mbox{on}\;\;\mathbb{R}^{n}
\end{cases}
\end{equation}
and $d_s=2^{s-1}\Gamma(\frac{s}{2})/\Gamma(1-\frac{s}{2})>0$.
Here $W$ is any $C^3$ potential and, for simplicity of notation, $\nabla$ denotes the full gradient ---i.e., with respect to the $(x,y)$ variables--- in~$\mathbb{R}^{n+1}_+$.
\begin{defi}\label{defextension}
Given a bounded function $u$ defined on $\mathbb{R}^n$, we will call the unique bounded extension $U$ of $u$ in $\mathbb{R}^{n+1}_+$ satisfying $\mbox{div}(y^{1-s}\nabla U(x,y))=0$, the \textit{$s$-extension} of $u$ (the existence and properties of this extension were studied in \cite{CafSi}; see also~\cite{C-Si1}).
\end{defi}
Let $\widetilde B^+_r((x_0,0))=\{(x,y)\in \mathbb{R}^{n+1}_+\, :\, |(x,y)-(x_0,0)|<r\}$ and $\widetilde B^+_r=\widetilde B^+_r((0,0))$ ---the tilde is used to distinguish balls in $\mathbb{R}^{n+1}$ from balls in $\mathbb{R}^n$. The energy associated to problem \eqref{ext} in a half-ball $\widetilde B_R^+$ is given by
\begin{equation}\label{energy-ext}
\widetilde{\mathcal E}_{R}(U)=\frac{d_s}{2}\int_{\widetilde B_R^+}y^{1-s} |\nabla U(x,y)|^2 \,dx\,dy + \int_{B_R}W(U(x,0))\,dx.
\end{equation}
As before, we distinguish between the Sobolev and the potential energies using the following notations:
\begin{equation*}
\widetilde{\mathcal E}_{R}^{\rm Sob}(U)=\frac{d_s}{2}\int_{\widetilde B_R^+}y^{1-s} |\nabla U(x,y)|^2 \,dx\,dy \quad \mbox{and} \quad \widetilde{\mathcal E}_{R}^{\rm Pot}(U)=\int_{B_R}W(U(x,0))\,dx.
\end{equation*}
We recall now a result from \cite{CRS} which allows to control the Sobolev energy of the $s$-extension $U$ of $u$ in $\mathbb{R}^{n+1}_+$ by the local contribution of the $H^{s/2}$-seminorm of~$u$ itself. We write the result for the specific case of half-balls in $\mathbb{R}^{n+1}$, since this is what we will need
later on. The result is true for a general Lipschitz domain $\Omega$, as one can see in Proposition 7.1 of \cite{CRS}.
\begin{lem}[Proposition 7.1 in \cite{CRS}]\label{up-down}
Let $s\in (0,2)$, $\widetilde B^+_{\rho}((x_0,0))$ be, as before, a half-ball in $\mathbb{R}^{n+1}_+$ centered at $(x_0,0)$ and with radius $\rho$, and let $U$ be the $s$-extension of $u$.
Then,
$$\int_{\widetilde B^+_{\rho}((x_0,0))} y^{1-s} |\nabla U(x,y)|^2\,dx\,dy \leq C \iint_{(\mathbb{R}^n\times \mathbb{R}^n )\setminus (B_{2\rho}^c(x_0)\times B_{2\rho}^c(x_0))} \frac{|u(x)-u(\bar x)|^2}{|x-\bar x|^{n+s}}\,dx\,d\bar x
$$
for some constant $C$ which depends only on $n$ and $s$.
\end{lem}
We also recall the monotonicity formula established in \cite{CC2}.
\begin{prop}[Proposition 3.2 in \cite{CC2}]\label{mon}
Let $s\in (0,2)$, $W\in C^3([-1,1])$ be nonnegative, and $U:\mathbb{R}^{n+1}_+\to (-1,1)$ be a solution of \eqref{ext}.
Then,
$$\Phi(R):=\frac{1}{R^{n-s}} \widetilde{\mathcal E}_{R}(U)$$
is a nondecreasing function of $R>0$.
\end{prop}
The main ingredients in the proof of our density estimate are the previous monotonicity formula and the $BV$ estimate for stable solutions established in Theorem~\ref{BV}.
Before giving the proof of Proposition \ref{density}, we
state the following easy lemma that allows to interpolate between $L^1$ and $W^{1,1}$.
\begin{lem}[Theorem 1 in \cite{Brezis1}]\label{interpo} Let $s\in (0,1)$ and $R>0$.
Let $u$ be a $C^1$ function in $B_R\subset \mathbb{R}^n$ satisfying
$$
R^{-n}\int_{B_R} |u + k|\,dx\leq V \quad \mbox{and}\quad R^{1-n}\int_{B_R} |\nabla u|\,dx\leq P
$$
for some $k\in \mathbb{R}$ and constants $V$ and $P$.
Then,
$$
R^{s-n}\int_{B_R}\int_{B_R}\frac{|u(x)-u(\bar x)|}{|x-\bar x|^{n+s}} \,dx\,d\bar x \leq C V^{1-s}P^{s}
$$
for some constant $C$ which depends only on $n$ and $s$.
\end{lem}
\begin{proof}
It follows, after scaling, from \cite[Theorem 1]{Brezis1} used with $p_1=p_2=p=1$, $s_1=0<1=s_2$, $\theta=1-s$, and $\Omega = B_1$.
\end{proof}
In the proof of Proposition \ref{density}, we will apply Lemma \ref{interpo} above with $k=\pm 1$, $V=\omega_0$, where $\omega_0$ is given in \eqref{hp-density}, and $P=C(1-s)^{-1}$, where $C(1-s)^{-1}$ comes from the $BV$ bound \eqref{BVest}.
We can now give the proof of our density estimate.
\begin{proof}[Proof of Proposition \ref{density}]
We argue by contradiction. Assume that there exists $\bar c\in (0,1)$ for which $R^{-n}\int_{B_R} |1+u| \,dx \le \omega_0$ and $\{u\ge -\bar c\}\cap B_{R/2}\neq \varnothing$.
Throughout the proof, all constants will depend only on $\bar c$, $n$, and $s$.
First, by continuity of $u$ and by taking $\omega_0<R^{-n} |B_{R/2}|$, there will be a point
$x_0\in B_{R/2}$ for which $|u(x_0)|\le \bar c$.
Moreover, by the uniform continuity of $u$ (recall that $|\nabla u| \le C$ in $\mathbb{R}^n$; see Appendix~\ref{app-C}), we will have that $|u|\leq \frac{1+\bar c}{2}$ in some ball of radius $r>0$ centered at $x_0$ (we emphasize that we can take $r$, as well as all other constants in the rest of the proof, to depend only on $\bar c$, $n$, and $s$). Using that $W(u)=\frac{1}{4}(1-u^2)^2$, we deduce
$$r^{s-n}\int_{B_r(x_0)} W(u) \,dx\geq \theta>0$$
for some positive constant $\theta$.
Let now $U$ be the $s$-extension of $u$ in $\mathbb{R}^{n+1}_+$. The previous lower bound on the potential energy in $B_r(x_0)$ leads to
$$r^{s-n}\widetilde{\mathcal E}_{r}(U)=r^{s-n}\left(\frac{d_s}{2}\int_{\widetilde B^+_r(x_0)}y^{1-s}|\nabla U|^2\,dx\,dy + \int_{B_r(x_0)} W(u)\,dx\right) \geq \theta,$$
where, for simplicity of notation, we keep denoting by $\widetilde{\mathcal E_r}$ the energy in the half-ball $\widetilde B^+_r((x_0,0))$ centered at $(x_0,0)$ (instead of at the origin).
Applying the monotonicity formula of Proposition \ref{mon}, we deduce that
\begin{equation}\label{upper-bound}
\rho^{s-n} \widetilde{\mathcal E}_{\rho}(U)\geq \theta\quad \mbox{for every } \rho\ge r.
\end{equation}
We use now Lemma \ref{up-down} to translate the bound in \eqref{upper-bound} for the energy of the extension $U$ into a lower bound for the energy of $u$. We get
\begin{eqnarray}\label{bound2}
\theta&\leq &\rho^{s-n} \widetilde{\mathcal E}_{\rho}(U) \nonumber \\
& \leq &C\rho^{s-n} \left(\iint_{(\mathbb{R}^n\times \mathbb{R}^n )\setminus ( B^c_{2\rho}(x_0)\times B^c_{2\rho}(x_0))} \frac{|u(x)-u(\bar x)|^2}{|x-\bar x|^{n+s}}\,dx\,d\bar x +\int_{B_{\rho}(x_0)} W(u)\,dx \right) \nonumber\\
&\leq &C \rho^{s-n} \iint_{B_{2\rho}(x_0)\times \mathbb{R}^n} \frac{|u(x)-u(\bar x)|^2}{|x-\bar x|^{n+s}}\,dx\,d\bar x ,
\end{eqnarray}
where in the last inequality we have used Proposition \ref{DircontrolsPot2} to bound the potential energy by the Sobolev energy in the larger ball, which requires to take $\rho\ge R_0$ (with $R_0$ being the constant in that proposition) since, then, $\rho+R_0\leq 2\rho$.
We aim now to use assumption \eqref{hp-density} and Lemma \ref{interpo} in order to find a contradiction with \eqref{bound2}. To this end, we need to introduce a larger radius $R$, with $R\ge 4\rho$, and observe that the set of integration in \eqref{bound2} satisfies
\[
B_{2\rho}(x_0)\times \mathbb{R}^n
\subset \big(B_R(x_0)\times B_R(x_0)\big)\cup \big(B_{2\rho}(x_0)\times B_R^c(x_0)\big).
\]
Now, applying Lemma \ref{interpo}, we are able to control the $W^{1,s}$-seminorm of $u$ in $B_R$ by a positive power of $\omega_0$. More precisely, by \eqref{hp-density}, the quantity $R^{-n}\|u+1\|_{L^1(B_R)}$ is controlled by $\omega_0$. On the other hand, our $BV$ estimate \eqref{BVest} gives a bound for $R^{1-n}\|\nabla u\|_{L^1(B_R)}$ by a constant $C(1-s)^{-1}$. Thus, using Lemma \ref{interpo}, $|u|\le 1$, and $R-2\rho\ge R/2$ (since we take $R\ge 4\rho$), we deduce
\begin{eqnarray}\label{bound3}
\theta&\leq &\rho^{s-n} \widetilde{\mathcal E}_{\rho}(U) \leq C\left(\frac{\rho}{R}\right)^{s-n}\,R^{s-n}\iint_{B_{R}(x_0)\times B_R(x_0)}\frac{|u(x)-u(\bar x)|}{|x-\bar x|^{n+s}}\,dx\,d\bar x \nonumber \\
& & \hspace{3cm}+ C\rho^{s-n}\iint_{B_{2\rho}(x_0)\times B_R^c(x_0)} \frac{|u(x)-u(\bar x)|}{|x-\bar x|^{n+s}}\,dx\,d\bar x \nonumber \\
&\leq &C \left(\frac{\rho}{R}\right)^{s-n} \omega_0^{1-s} + C\rho^{s-n}\iint_{B_{2\rho}(x_0)\times B_R^c(x_0)} \frac{dx\,d\bar x}{|x-\bar x|^{n+s}} \nonumber \\
&\leq &C \left(\frac{\rho}{R}\right)^{s-n} \omega_0^{1-s} + C\rho^{s}\int_{\frac{R}{2}}^{+\infty}\frac{r^{n-1}}{r^{n+s}}\,dr \nonumber \\
&\leq & C_1 \left[\left(\frac{\rho}{R}\right)^{s-n}\omega_0^{1-s} +\left(\frac{\rho}{R}\right)^s\right].
\end{eqnarray}
We now take $R$ and $\rho$ such that $\frac{\rho}{R}=(\frac{\theta}{4C_1})^{1/s}$. We may ensure $R\ge 4\rho$ (as required before) by making larger the constant $C_1$ in \eqref{bound3}, if necessary. Since we needed $\rho\ge \max \{r,R_0\}$ within the proof, with $R_0$ being the radius from Proposition~\ref{DircontrolsPot2}, this gives a lower bound for $R$ which becomes our final choice of radius $R_0$ in the statement of Proposition \ref{density}.
Finally, with this choice of $\rho/R$, \eqref{bound3} becomes
\begin{equation}\label{final2}\theta \leq C_1 \left(\frac{\theta}{4C_1}\right)^{\frac{s-n}{s}} \omega_0^{1-s} + \frac{\theta}{4}.\end{equation}
Therefore, choosing $\omega_0$ small enough, we obtain a contradiction and conclude the proof.
\end{proof}
\section{Convergence results}\label{sec-6}
The goal of this section is to prove the following convergence result, which will allow us to give the Proof of Theorem \ref{thm2}.
\begin{prop}\label{conv-L^1}
Let $n\ge 2$, $s\in(0,1)$, and $W(u)=\frac 1 4 (1-u^2)^2$.
Let $u:\mathbb{R}^n\rightarrow (-1,1)$ be a stable solution of $(-\Delta)^{s/2} u +W'(u)=0$ in $\mathbb{R}^n$.
Then, for every sequence $R_j\uparrow \infty$ there exists a subsequence $R_{j_k}$ such that, defining $u_{R} (x):= u(R x)$, we have
\[
u_{R_{j_k}} \rightarrow u_\infty: = \chi_\Sigma-\chi_{ \Sigma^c} \quad \mbox{ in } L^1_{\rm loc}(\mathbb{R}^n)\]
for some cone $\Sigma\subset \mathbb{R}^n$ which is nontrivial $($i.e., it is not equivalent to $\mathbb{R}^n$ nor to $\varnothing$ up to sets of zero measure$)$ and which is a weakly stable set in $\mathbb{R}^n$ for the fractional perimeter~$P_s$.
Moreover, we have the following convergence of the localized energies:
\begin{equation}\label{whoiwhoiehr}
\mathcal E^{\rm Sob}_{B_{R'}} (u_{R_{j_k}}) \to \mathcal E^{\rm Sob}_{B_{R'}}(u_\infty) \quad \mbox{and} \quad R_{j_k}^s \int_{B_{R'}} W(u_{R_{j_k}})\,dx \to 0= \int_{B_{R'}} W(u_\infty)\,dx,
\end{equation}
as $k\uparrow\infty$, in every ball $B_{R'}\subset \mathbb{R}^n$.
\end{prop}
\begin{proof}We divide the proof into two steps.
\vspace{3pt}
-{\em Step 1.} We start by proving that some subsequence of $u_{R_{j}}$ converges in $L^1_{\rm loc}(\mathbb{R}^n)$ to $u_\infty : = \chi_\Sigma-\chi_{ \Sigma^c}$, for some nontrivial cone $\Sigma$, and that the convergence of energies \eqref{whoiwhoiehr} holds.
Throughout the proof, $C$ will denote (possibly different) positive constants which depend only on $n$ and $s$.
Let us take a radius $R' \geq 1$. Theorem \ref{BV} and Corollary \ref{energy-est} yield that
\[
\int_{B_R} |\nabla u|\,dx \le C R^{n-1}\quad \mbox{and}\quad \mathcal E_{B_R} (u) \le C R^{n-s} \quad \mbox{for all }R\ge1.
\]
Thus, for $R\ge 1$, by rescaling we deduce
\begin{equation}\label{W11uR}
\int_{B_{R'}} |\nabla u_R|\,dx \le C_{R'} \quad \mbox{and} \quad \mathcal E^{\rm Sob}_{B_{R'}} (u_R) \le C_{R'},
\end{equation}
where $C_{R'}$ denote constants which depend only on $R'$, $n$, and $s$.
Let $U$ be, as in the previous section, the $s$-extension of $u$ in $\mathbb{R}^{n+1}_+$.
By Lemma \ref{up-down}, and since the potential energy is nonnegative, we have
\begin{equation*}
\widetilde {\mathcal E}^{\rm {Sob}}_{R} (U)
\leq C\iint_{(\mathbb{R}^n\times \mathbb{R}^n)\setminus( B^c_{2R}\times B^c_{2R})} \frac{|u(x)-u(\bar x)|^2}{|x-\bar x|^{n+s}}\,dx\,d\bar x= 4C\, \mathcal E^{\rm {Sob}}_{B_{2R}} (u) \le C R^{n-s}
\end{equation*}
for $R\ge 1$. In addition, by the monotonicity formula of Proposition \ref{mon}, the quantity
\[\Phi(R) = R^{s-n} \widetilde{\mathcal E}_R(U)\]
is monotone nondecreasing. At the same time, by the previous bound on the Sobolev energy and by Proposition \ref{DircontrolsPot2} (which gives control on the potential energy in $B_R$ by the Sobolev energy in $B_{2R}$ if we take $R+R_0\leq 2R$), we deduce that $\Phi$ is bounded above by a finite constant.
We deduce that
\begin{equation}\label{mono-lim}
\Phi(R'R)-\Phi(\bar R) \rightarrow 0 \quad \text{ as } R'R\ge \bar R \uparrow \infty.
\end{equation}
Now, from the proof of the monotonicity formula, which is based on a Pohozaev identity ---see the proof of Proposition 3.2 in \cite{CC2}--- we have that
\begin{equation*}
\Phi'(\rho) = \frac{d_s}{\rho^{n-s}} \int_{\partial ^+ \widetilde B_\rho^+} y^{1-s} (\partial_{r} U)^2 \,d{\mathcal H}^n(x,y) + \frac{s}{\rho^{n-s+1}}\int_{B_\rho} W(u) \,dx,
\end{equation*}
where $\partial^+$ denotes the part of the boundary contained in the open half-space $\{y>0\}$ and $\partial_r$ denotes the radial derivative in $\mathbb{R}^{n+1}_+$.
After rescaling, this becomes
\begin{equation}\label{wnoiwoieht}
\Phi'(R\tilde\rho) = \frac{d_s}{R\tilde\rho^{n-s}} \int_{\partial^+ \widetilde B_{\tilde\rho}^+} \tilde y^{1-s} (\partial_{r} U_R)^2 \,d{\mathcal H}^n(\tilde x,\tilde y) + \frac{s}{R^{1-s}\,{\tilde\rho}^{n-s+1}}\int_{B_{\tilde\rho}} W(u_R) \,d\tilde x,
\end{equation}
where $U_R$ is the $s$-extension of $u_R$.
We integrate now with respect to $ \rho$ to obtain, for $R'/2\geq \bar R/R$,
\begin{equation}\label{wnoiwoieht2}
\begin{split}
\Phi(R' R)-\Phi(\bar R) &=\int_{\bar R}^{R' R} \Phi'(\rho)\,d\rho=R\int_{\bar R/R}^{R'}\Phi'(R{\tilde\rho})\,d{\tilde\rho}\\
&=
d_s \int_{\bar R/R}^{R'} d{\tilde\rho} \, {\tilde\rho}^{s-n} \int_{\partial ^+ \widetilde B_{\tilde\rho}^+} \tilde y^{1-s} (\partial_{r} U_R)^2 \,d{\mathcal H}^n(\tilde x,\tilde y)\\
&\hspace{3em} + sR^{s}\int_{\bar R/R}^{R'} d{\tilde\rho}\,{\tilde\rho}^{s-n-1}\int_{B_{\tilde\rho}}W(u_R)\,d\tilde x\\
&\ge d_s (R')^{s-n} \int_{\widetilde B^+_{R'}\setminus \widetilde B^+_{\bar R/R}} \tilde y^{1-s} (\partial_{r} U_R)^2 \,d \tilde x\,d\tilde y,\\
&\hspace{3em} { + sR^{s}(R')^{s-n-1} \frac{R'}{2} \int_{B_{R'/2}} W(u_R)\,d\tilde x }\quad\text{ if } R\geq 2\bar R/R'.
\end{split}
\end{equation}
Note that, given $R'$ and $\bar R$, we have
$$
\int_{\widetilde B^+_{\bar R/R}} \tilde y^{1-s} (\partial_{r} U_R)^2 \,d \tilde x\,d\tilde y \leq C R^2 \int_{\widetilde B^+_{\bar R/R}} \tilde y^{1-s} \,d \tilde x\,d\tilde y\leq C \bar R^{2-s+n}R^{s-n} \to 0 \quad\text{ as }R\uparrow\infty.
$$
This together with \eqref{mono-lim} and \eqref{wnoiwoieht2}, leads to
\begin{equation}\label{limithomog}
\int_{\widetilde B^+_{R'}} y^{1-s} (\partial_{r} U_R)^2 \,d\tilde x\,d\tilde y \rightarrow 0 \quad \mbox{as }R\uparrow \infty
\end{equation}
and
\begin{equation}\label{limitpm1}
R^s \int_{B_{R'/2}} W(u_{R}) \,d\tilde x\rightarrow 0 \quad \mbox{as }R\uparrow \infty
\end{equation}
for every $R'\ge 1$.
Next, choose any $\sigma \in(s,1)$ and let $N\ge 1$ be an integer.
By the $W^{1,1}$ estimate \eqref{W11uR} applied with $R'=N$, and since $|u|\le 1$, Lemma~\ref{interpo} (applied with $s$ replaced by $\sigma$) leads to
\[ \int_{B_{N}\times B_{N}} \frac{\big|u_R(x)-u_R(\bar x)\big|^2}{|x-\bar x|^{n+\sigma}} \,dx\,d\bar x\leq C_{N,\sigma}\]
for all $R\ge 1$,
where $C_{N,\sigma}$ is a constant depending only on $N$, $\sigma$, $n$, and $s$.
Hence, using that $|u_R|\le 1$ and the compactness of $W^{\sigma/2,2}$ inside $W^{s/2,2}$, there exists a subsequence $u_{R_{j_k}}$, that we still denote by $u_{R_k}$, and a function $u_\infty$ such that
\begin{equation}\label{strong}
\big\| u_{R_{k}} -u_\infty\big\|_{L^2(B_{N})} + \int_{B_{N}\times B_{N}} \frac{\big| (u_{R_{k}} -u_\infty)(x)- (u_{R_{k}}-u_\infty)(\bar x)\big|^2}{|x-\bar x|^{n+s}} \,dx\,d\bar x\rightarrow 0
\end{equation}
as $k\uparrow\infty$.
In addition, letting $N\uparrow\infty$, taking further subsequences, and using a Cantor diagonal argument, we obtain a new subsequence converging in $L^2$ in every ball of $\mathbb{R}^n$.
Now, given $R'\ge 1$, we use once more Lemma \ref{up-down} (now applied to $U_{R_k}-U_\infty$ in $\widetilde B_{R'}^+$, where $U_\infty$ denotes the $s$-extension of $u_\infty$) to control
$$
\int_{\widetilde B_{R'}^+} y^{1-s}\big| \nabla U_{R_k}- \nabla U_\infty\big|^2\,dx\,dy
$$
by the double integral in \eqref{strong} computed now in $B_{2R'}\times \mathbb{R}^n$. Taking $N>2R'$, since \eqref{strong} gives control on the integrals computed over $B_{2R'}\times B_N$, it only remains to make the double integral on $B_{2R'}\times (\mathbb{R}^n\setminus B_N)$ arbitrary small. Such bound is obvious, since $|u_{R_{k}} -u_\infty| \le2$, by taking $N$ large enough. Therefore, we conclude
\begin{equation}\label{hwiohrhee1}
\int_{\widetilde B_{R'}^+} y^{1-s}\big| \nabla U_{R_k}- \nabla U_\infty\big|^2\,dx\,dy \rightarrow 0
\end{equation}
as $k\uparrow\infty$.
From this strong convergence and the local uniform convergence of $u_{R_k}$, passing to the limit in \eqref{limithomog} and \eqref{limitpm1} we obtain
\[ \int_{\widetilde B_{R'/2}^+}y^{1-s} (\partial_r U_\infty)^2 \,dx\,dy = 0 \quad \mbox{and}\quad \int_{B_{R'/2}}W(u_\infty)\,dx =0\]
for every $R'\ge 1$. Therefore, $\partial_r U_\infty = 0$ in $\mathbb{R}^{n+1}_+$ and $W(u_\infty)= 0$ in $\mathbb{R}^n$.
In other words, $U_\infty$ and its trace $u_\infty$ are homogeneous of degree $0$, and $u_\infty$ takes
values $\pm 1$ ---the two wells of $W$. Equivalently, we have that
\[u_\infty=\chi_\Sigma-\chi_{ \Sigma^c}\quad \mbox{in }\mathbb{R}^n\]
for some cone $\Sigma$. In addition, by the same convergences for $u_{R_k}$ that we have just used, we see that \eqref{whoiwhoiehr} holds.
Finally, using the monotonicity of $\Phi$, \eqref{hwiohrhee1}, and \eqref{limitpm1}, we obtain
\begin{equation*}
\begin{split}
0&<\Phi(1)\le \Phi(R)=R^{s-n}\widetilde{\mathcal E}_R(U)=\widetilde{\mathcal E}^{\rm {Sob}}_1(U_R)+R^s\int_{B_1} W(u_R) \,dx)\\
&\le \lim_{R\uparrow \infty}\Big(\widetilde{\mathcal E}^{\rm {Sob}}_1(U_R)+R^s\int_{B_1} W(u_R) \,dx \Big)\\
&=\widetilde{\mathcal E}^{\rm {Sob}}_1(U_\infty).
\end{split}
\end{equation*}
Thus ${U_\infty}$ and ${u_\infty}$ have positive energy, and hence $\Sigma$ is nontrivial (i.e. it is not equal to $\mathbb{R}^n$ nor $\varnothing$ up to sets of measure zero).
\vspace{3pt}
-{\em Step 2}. It remains to prove that $\Sigma$ is a weakly stable set in $\mathbb{R}^n$ for the fractional perimeter~$P_s$.
To this end, let us take an arbitrary smooth vector field $X=X(x,t)$ which is compactly supported in $B_1\times(-1,1)$ (by scaling, since $\Sigma$ is a cone, we may take the support to be the unit ball) and let $\Psi = \Psi_t(x)$ denote the map $(x,t) \mapsto \phi_X^t(x)$, where $\phi_X^t$ is the integral flow defined by the ODE
\[
\frac{d}{dt} \phi_X^t(x) = X(\phi_X^t(x),t) \quad \mbox{with initial condition }\quad \phi_X^0(x) = x.
\]
Note that $\Psi_0= {\rm Id}$ in $\mathbb{R}^n$, $\Psi_t = {\rm Id}$ outside of $B_1$, and that $\Psi_t:\mathbb{R}^n\to\mathbb{R}^n$ is a diffeomorphism for $|t|$ small.
Let us introduce the rescaled energy functional (of which $u_R$ is a stable critical point)
\[
\mathcal E^R_{B_1} (v) := \frac{1}{4} \iint_{(\mathbb{R}^n\times \mathbb{R}^n) \setminus (B_1^c\times B_1^c)}\, \frac{| v(y)-v(\bar y) |^2}{
| y-\bar y|^{n+s}} \,dy\, d\bar y + R^s\int_{B_1} W(v(y))\,dy \, ,\]
We first show that the function $u_{R,t} := u_R\circ \Psi_t^{-1}$ satisfies
\begin{equation}\label{theexpansion}
\mathcal E^R_{B_1} (u_{R,t}) \ge \mathcal E^R_{B_1} (u_R) - C_{\Psi} t^3, \quad \mbox{for }t\in(-T_{\Psi}, T_{\Psi}),
\end{equation}
where $T_{\Psi}>0$ and $C_{\Psi}$ will be, from now on, different positive constants which depend only on $X$, $n$, and $s$ ---in particular, they are independent of~$R$.
To prove this, and as in the proof of Lemma \ref{lem2A}, we make the change of variables $y= \Psi^{-1}_{ t}(x)$, $\bar y= \Psi^{-1}_{t}(\bar x)$ for $|t|$ small. Since $\Psi_{t}^{-1}$ sends $B_1$ and $B_1^c$ onto themselves, setting $A_1:=(\mathbb{R}^n\times \mathbb{R}^n)\setminus (B_1^c\times B_1^c)$ we have
\begin{equation}\label{Eg-urt}
\mathcal E^R_{B_1}(u_{R,t}) = \frac{1}{4}\iint_{A_1}\, \frac{| u_R(y)-u_R(\bar y) |^2}{
\big| \Psi_{t}(y)-\Psi_{t}(\bar y)\big|^{n+s}} \,J_t(y) \,J_t(\bar y)\,dy \,d\bar y + R^{-s} \negthinspace\int_{B_1} W(u_R(y))\,J_t(y)\,dy,
\end{equation}
where $J_{t}$ is the Jacobian ${\rm det} D\Psi_t$.
A Taylor expansion for $J_t$ yields
\begin{equation} \label{expansionJac}
\big| J_t(y) - 1 - h_1(y) t - h_2(y) t^2 \big| \le C_{\Psi}t^3 \quad \mbox{for }t\in(-T_{\Psi}, T_{\Psi}),
\end{equation}
with $ \| h_1\|_{L^\infty(\mathbb{R}^n)} +\| h_2\|_{L^\infty(\mathbb{R}^n)} \le C_{\Psi}$.
Similarly, we have
\[
\left|\, \Psi_{t}(y) - \Psi_{t}(\bar y) - ( y-\bar y) - |y-\bar y|\big( { g} _1(y,\bar y) t + { g}_2(y,\bar y) t^2 \big)\,\right| \le C_{\Psi} |y-\bar y| t^3
\]
for $t\in(-T_{\Psi}, T_{\Psi}) $, where
$\| { g}_1\|_{L^\infty (\mathbb{R}^n\times\mathbb{R}^n )} +\| g_2\|_{L^\infty (\mathbb{R}^n\times\mathbb{R}^n )} \le C_{\Psi}$.
Therefore
\begin{equation} \label{expansionK}
\begin{split}
\big| \Psi_{t}(y)-\Psi_{t}(\bar y)\big|^{-n-s}
&= \big| y-\bar y + |y-\bar y|\,\big( { g}_1(y,\bar y) t + { g}_2(y,\bar y) t^2 +O(t^3)\big)\big|^{-n-s}
\\
&\hspace{-2cm}= |y-\bar y|^{-n-s} \big( 1+ tk_1(y,\bar y) + t^2 k_2(y,\bar y) + O(t^3) \big),
\end{split}
\end{equation}
where $\| k_1\|_{L^\infty (\mathbb{R}^n\times\mathbb{R}^n )} +\| k_2\|_{L^\infty (\mathbb{R}^n\times\mathbb{R}^n )} \le C_{\Psi}$.
Using \eqref{expansionJac} and \eqref{expansionK} in \eqref{Eg-urt} we obtain
\[
\big| \mathcal E^R_{B_1}(u_{R,t}) - \big( \mathcal E^R_{B_1}(u_{R}) + a_1t + a_2 t^2 \big) \big| \le C_{\Psi} \,t^3 \mathcal E^R_{B_1}(u_{R})
\]
for some constants $a_1$ and $a_2$, since the quantity
\[t^3\mathcal E^R_{B_1}(u_{R}) =t^3\left( \frac{1}{4} \iint_{A_1}\, \frac{| u_R(y)-u_R(\bar y) |^2}{
| y-\bar y|^{n+s}} \,dy\, d\bar y + R^s\int_{B_1} W(u_R(y))\,dy \right)\]
controls the error terms which are cubic in the variable $t$.
Now, since by assumption $u_R$ is a stable solution, it must be $a_1 =0$ and $a_2\ge 0$, and thus
\[
\mathcal E^R_{B_1} (u_{R,t}) \ge \mathcal E^R_{B_1} (u_R) - C_{\Psi} \mathcal E^R_{B_1}(u_{R}) t^3, \quad \mbox{for }t\in(-T_{\Psi}, T_{\Psi}).
\]
Finally, thanks to \eqref{W11uR} and \eqref{limitpm1} we have $\mathcal E^R_{B_1} (u_R) \le C$, with $C$ depending only on $n$ and $s$. This concludes the proof of \eqref{theexpansion}.
Now, recalling the convergence of the energies, \eqref{whoiwhoiehr}, we have that \eqref{theexpansion} passes to the limit and we deduce
\[
\mathcal E^{\rm Sob}_{B_1} (u_{\infty,t}) \ge \mathcal E^{\rm Sob}_{B_1} (u_{\infty}) - C_{\Psi} t^3 \quad \mbox{ for }t\in(-T_{\Psi}, T_{\Psi}),
\]
where $u_{\infty,t} = u_{\infty}\circ \Psi^{-1}_t$.
Since $u_\infty = \chi_\Sigma-\chi_{ \Sigma^c}$, we have
\[
\mathcal E^{\rm Sob}_{B_1} (u_\infty) = 2 P_s(\Sigma,B_1) \quad \mbox{and}\quad \mathcal E^{\rm Sob}_{B_1} (u_{\infty,t}) =2P_s(\Psi_t(\Sigma), B_1).
\]
Thus,
\[
P_s(\Psi_t(\Sigma),B_1) \ge P_s(\Sigma, B_1) - C_{\Psi} t^3, \quad \mbox{for }t\in(-T_{\Psi}, T_{\Psi}).
\]
Recalling Definition \ref{def-weakly} and since the smooth compactly supported vector field~$X$ defining $\Psi$ was arbitrarily, we have shown that $\Sigma$ is a weakly stable set in $B_1$ for the fractional perimeter $P_s$. Finally, using that $\Sigma$ is a cone, we easily deduce, by scaling, that it is in fact weakly stable in all of $\mathbb{R}^n$.
\end{proof}
We can now give the
\begin{proof} [Proof of Theorem \ref{thm2}]
The first part of the statement on the $L^1$-convergence has just been proven in Proposition \ref{conv-L^1} above. We have also proved that $\Sigma$ is nontrivial.
The last part of the statement, i.e., that \eqref{wngioewhtioeh1} and \eqref{wngioewhtioeh2} hold after choosing the representative of $\Sigma$ (for which every point of $\Sigma$ with density $1$ belongs to its interior and every point of density 0 belongs to its complement) follows, as usual, from the local $L^1$-convergence and the density estimate of Proposition \ref{density}.
\end{proof}
In Proposition \ref{conv-L^1} we showed that the potential energies of sequences of blow-downs converge to zero.
This was a consequence of the monotonicity formula.
To end this section we now give a stronger property (which will be useful in Section~\ref{sec-7}): a quantitative convergence of the potential energy to zero, as $\varepsilon\downarrow 0$, for stable solutions of $(-\Delta)^{s/2}u_\varepsilon + \varepsilon^{-s} W'(u_\varepsilon)=0$ with $s\in (0,1)$.
\begin{prop} \label{propnew}
Let $n\ge 2$, $s\in (0,1)$, and $W(u)= \frac 1 4 (1-u^2)^2$. For $\varepsilon>0$, let $u_\varepsilon:\mathbb{R}^n\rightarrow (-1,1)$ be a stable solution of $(-\Delta)^{s/2}u_\varepsilon + \varepsilon^{-s} W'(u_\varepsilon)=0$ in $\mathbb{R}^n$.
If $R\ge \varepsilon$, then
\[
\int_{B_{R}} (\varepsilon/R)^{-s} W(u_\varepsilon)\,dx \le C R^n (\varepsilon/R)^{\beta},
\]
where $\beta := \min\big(\frac{1-s}{2}, s \big) >0$ and $C$ depends only on $n$ and $s$.
\end{prop}
\begin{proof}
By scaling we may (and do) assume without loss of generality that $R=1$.
Given $x\in B_1$, let
\[
r_x : = \max ( \min\big( \tfrac 1 8 , \tfrac 1 2 {\rm dist} (x, \{|u|\le \tfrac {9}{10}\}), C_0\varepsilon),
\]
where $C_0>0$ is a large enough constant, depending only on $n$ and $s$, to be chosen later.
Note that if $\varepsilon$ satisfies $1 \le 8C_0\varepsilon$, then
\[
\int_{B_{1}} \varepsilon^{-s} W(u_\varepsilon)\,dx \le (8C_0)^s ({\textstyle \max_{[-1,1]} }W ) |B_{1}| \le C \le C C_0^{\beta} \varepsilon^{\beta}.
\]
Thus, we may (and do) assume that $\frac{1}{8}> C_0\varepsilon$. In particular, $r_x\in [C_0\varepsilon,\frac{1}{8}]$ for all $x\in B_1$.
We will take $C_0$ satisfying $C_0 \ge \max (R_0,1)$, where $R_0$ is the constant of Proposition \ref{density} for $\bar c = \tfrac{9}{10}$. Then, for the constant $\omega_0$ of Proposition \ref{density}, we claim that
\begin{equation}\label{whtiohwioh}
\min\bigg( \int_{B_{4r_x} (x)} |u_\varepsilon -1|\,dx, \int_{B_{4r_x} (x)} |u_\varepsilon +1|\,dx\bigg) \ge \omega_0 (2r_x)^n
\end{equation}
for all $x\in B_1$ such that $r_x< \frac 1 8$.
Indeed, since $r_x<\frac 1 8$, there exists $z\in \{|u|\le \tfrac {9}{10}\} \cap\overline{B}_{1+ \frac 1 4}$ such that $|x-z| \leq 2r_x$. Hence, by Proposition \ref{density} ---applied to $u = u_\varepsilon (\varepsilon \,\cdot\,)$---,
\[
\min\bigg( \int_{B_{2r_x} (z)} |u_\varepsilon -1|\,dx, \int_{B_{2r_x} (z)} |u_\varepsilon +1|\,dx\bigg) \ge \omega_0 (2r_x)^n,
\]
and \eqref{whtiohwioh} follows since $B_{2r_x} (z)\subset B_{4r_x} (x)$.
Note also that $B_{4r_x} (x) \subset B_{3/2}$ for all $x\in B_1$.
On the other hand, thanks to the potential energy estimate in Theorem \ref{thm1} (rescaled) we have (using $4r_x/\varepsilon\ge C_0\ge 1$)
\begin{equation}\label{wngthwiohw}
C_0^s \average\int_{B_{4r_x}} \frac{1}{4} (1-|u_\varepsilon|)^2 \,dx \le (\varepsilon/r_x)^{-s} \average\int_{B_{4r_x}} W(u_\varepsilon)\,dx \le C,
\end{equation}
where $\average\int$ denotes the average.
Let us next show that Poincar\'e's inequality, \eqref{whtiohwioh}, and \eqref{wngthwiohw} for $C_0$ sufficiently large, yield
\begin{equation} \label{whtoiwhoewih}
\int_{B_{4r_x} (x)} |\nabla u_\varepsilon| \ge c r_x^{n-1} \quad \mbox{whenever } r_x<\tfrac 1 8,
\end{equation}
for some constant $c>0$ depending only on $n$ and $s$.
Indeed, if $\int_{B_{4r_x} (x)} |\nabla u_\varepsilon| =: \kappa \, r_x^{n-1}$, then by Poincar\'e's inequality we have
\[
\average\int_{B_{4r_x} (x)} |u_\varepsilon-t| \le C\kappa \quad \mbox{for some }t\in [-1,1].
\]
But then using \eqref{wngthwiohw}, we have
\[
\begin{split}
|1-|t||^2 &= \average\int_{B_{4r_x} (x)} |1-|t||^2 \le 2 \average\int_{B_{4r_x}} (1-|u_\varepsilon|)^2 + 2\average\int_{B_{4r_x}} |u_\varepsilon -t|^2
\\
& \le CC_0^{-s} + 4\average\int_{B_{4r_x}} |u_\varepsilon -t| \le C(C_0^{-s} + \kappa).
\end{split}
\]
Recalling now \eqref{whtiohwioh} we obtain
\[
\begin{split}
\frac{2^n \omega_0}{|B_4|} &\le \min\bigg( \average\int_{B_{4r_x} (z)} |u_\varepsilon -1|\,dx, \average\int_{B_{4r_x} (z)} |u_\varepsilon +1|\,dx\bigg)
\\& \le \average\int_{B_{4r_x} (z)} |u_\varepsilon -t|\,dx + |1-|t||
\le C(\kappa + (C_0^{-s} + \kappa)^{1/2}),
\end{split}
\]
and this gives a lower bound for $\kappa$ provided the $C_0$ is chosen sufficiently large. Thus, \eqref{whtoiwhoewih} is now proved.
We now produce a covering of $B_1$, by some of the balls $B_{r_x}(x)$, as follows.
Given $k \le -4$, let $X_k := \{x\in B_1 \,: \,r_x \in (2^k, 2^{k+1}]\}$ and let $\{ x_j^k\}_{j\in \mathcal J_k}$ be a maximal subset of $X_k$ with the property that the balls $B_{\frac 1 4 r_{x_j^k}}(x_j^k)$ are disjoint.
It then follows (using that all radii $r_x$ belong to $(2^k, 2^{k+1}]$ for $x\in X_k$) that
$$
X_k \subset \bigcup_{j\in \mathcal J_k} B_{r_{x_j^k}}(x_j^k)
$$
and that the family of quadruple balls
$$
\{ B_{4r_{x_j^k}}(x_j^k) \}_{j\in \mathcal J_k}
$$
has (dimensional) finite overlapping.\footnote{That is, every point $x\in \mathbb{R}^n$ belongs to at most $N$ of these balls, with $N$ depending only on~$n$. This is easy to check: if $x\in \mathbb{R}^n$ belonged to $N$ of such balls, we would have the existence of $N$ points $x^k_j$ in $B_{4\cdot 2^{k+1}}(x)$ such that the balls $B_{\frac{1}{4} 2^k}(x^k_j)$ are disjoint and contained in $B_{9\cdot 2^{k}}(x)$.}
Note also that, by construction, the union of the sets $X_k$ when $k$ runs on $\{\lfloor\log_2 (R_0\varepsilon)\rfloor \le k \le -4\}$ covers all of $B_1$.
Now, on the one hand, the BV estimate $\int_{B_{3/2}} |\nabla u_\varepsilon| \,dx\le C$ (which follows from Theorem \ref{BV}) yields, for all $k \le -4$,
\begin{equation} \label{whtoiwhoewih2}
\#\mathcal J_k \le C(2^k)^{1-n}.
\end{equation}
Indeed, this follows using that the balls $B_{ 4 r_{x_j^k}}(x_j^k)$ have finite overlapping and are contained in $B_{3/2}$: when $k<-4$ then $r_{x_j^k} <\frac 1 8$ and hence all the balls satisfy \eqref{whtoiwhoewih} and are contained in $B_{3/2}$ by construction; while for $k=-4$ the radius of the balls is at least $\frac 1 {16}$ so their number must be bounded.
On the other hand, we claim that Lemma \ref{lemdecay} yields
\[
\int_{B_{r_x}(x)} \varepsilon^{-s} W(u_\varepsilon) \,dx\le \int_{B_{r_x}(x)} \varepsilon^{-s} (1-|u_\varepsilon|)^2 \,dx\le C \varepsilon^{-s} \Big(\frac{\varepsilon}{r_x}\Big)^\alpha r_x^n
\]
for any given $\alpha \in [0,2s]$. Indeed, note that if $r_x= C_0 \varepsilon$ the previous estimate is trivial, while if $r_x> C_0\varepsilon$ then $r_x\le \tfrac 1 2 {\rm dist} (x, \{|u|\le \tfrac {9}{10}\})$ and hence we may apply Lemma \ref{lemdecay}
(recall that $r_x\ge C_0\varepsilon\ge \varepsilon$).
Therefore, choosing $\alpha := \min\big( \frac{1+s}{2}, 2s\big)\in (0,1)$ we obtain ---using \eqref{whtoiwhoewih2}---
\[
\begin{split}
\int_{B_1} \varepsilon^{-s} W(u_\varepsilon)\,dx &\le C \sum_{k= \lfloor\log_2 (R_0\varepsilon)\rfloor}^{-4} \sum_{j\in\mathcal J_k} \int_{B_{r_{x^k_j}}(x^k_j)} \varepsilon^{-s} W(u_\varepsilon) \,dx
\\
& \le C \sum_{k= \lfloor\log_2 (R_0\varepsilon)\rfloor}^{-4} \sum_{j\in\mathcal J_k} \varepsilon^{-s} \Big(\frac{\varepsilon}{r_{x_j}}\Big)^\alpha r_{x_j}^n
\\
& \le C \sum_{k= \lfloor\log_2 (R_0\varepsilon)\rfloor}^{-4} \varepsilon^{-s} \Big(\frac{\varepsilon}{2^k}\Big)^\alpha (2^{k+1})^n\, \#\mathcal J_k
\\
& \le C \sum_{k= \lfloor \log_2 (R_0\varepsilon)\rfloor}^{-4} \varepsilon^{\alpha-s} (2^k)^{n-\alpha} (2^k)^{1-n} \le C \varepsilon^{\alpha-s}\sum_{k= -\infty}^{-4} (2^k)^{1-\alpha}
\\
& \le C\varepsilon^{\beta},
\end{split}
\]
as we wanted to show.
\end{proof}
\section{Proofs of the classification results}\label{sec-7}
In this section we give the proof of our classification results. In order to prove Theorem \ref{thmclas}, we will need some preliminary ingredients.
We start by recalling the main results in \cite{dPSV}, which are a consequence of an improvement of flatness theory for phase transitions in the ``genuinely nonlocal'' regime (meaning that the order~$s$ of the operator is less than 1). The first one will be used to conclude one-dimensionality of solutions.
\begin{thm}[Theorem 1.2 in \cite{dPSV}]\label{123}
Let $n\ge 2$, $s\in(0,1)$, and $W(u)=\frac 1 4 (1-u^2)^2$. Let $u:\mathbb{R}^n\rightarrow (-1,1)$ be a solution of $(-\Delta)^{s/2}u + W'(u)=0$ in $\mathbb{R}^n$.
Assume that there exists a function $a:(1,\infty) \rightarrow (0,1]$ such that $a(R)\downarrow 0$ as $R\uparrow\infty$ and
such that, for all $R>0$, we have
\begin{equation}\label{ASS-R}
\{ e_R\cdot x\le -a(R)R\} \subset \big\{u\le -{\textstyle \frac 4 5} \big\}\subset \big\{u\le {\textstyle \frac 4 5}\big \} \subset \{e_R\cdot x\le a(R)R\} \quad \mbox{in }B_{R}
\end{equation}
for some $e_R\in S^{n-1}$
which may depend on $R$.
Then, $u(x)=\phi(e\cdot x)$ for some direction $e\in S^{n-1}$ and an increasing function $\phi:\mathbb{R}\to (-1,1)$. \end{thm}
We next prove a corollary of Theorem 1.1 in \cite{dPSV} which will be useful in the sequel. It is an``iterated version'' of Theorem 1.1 in \cite{dPSV} in the particular case $L= (-\Delta)^s$ and $f(u)= - W'(u) = u-u^3$.
\begin{prop}\label{cor:improvement}
Let $s\in(0,1)$ and $n\ge2$. There exist constants~$\alpha_0\in (0,s/2)$, $
p_0\in(2,\infty)$, and~$a_0\in(0,1/4)$, depending only on $n$ and $s$, such that the following statement holds.
\smallskip
Let $a\in(0,a_0]$ and let
\begin{equation}\label{ja}
j_a := \left\lfloor \frac{\log a}{\log(2^{-\alpha_0})} \right\rfloor \in \mathbb N.
\end{equation}
Let $\varepsilon>0$ and $k\in \mathbb N$ satisfy $ 2^{k-1}\varepsilon \le \big(2^{-\alpha_0(k-1)}a \big)^{p_0}$ and let $u_\varepsilon: \mathbb{R}^n \rightarrow (-1,1)$ be a solution
of
\[
(-\Delta)^su_\varepsilon + \varepsilon^{-s}W'(u_\varepsilon) =0 \quad \mbox{in }B_{2^{j_a}} \subset \mathbb{R}^n
\]
satisfying $0\in \big\{-\frac 3 4 \le u \le \frac 3 4 \big\}$ and
\begin{equation} \label{wethiowehtowh}
\big\{\omega_j \cdot x\le -a 2^{j(1+\alpha_0)} \big\}\, \subset\, \big\{u\le -{\textstyle \frac 3 4}\big\} \,\subset \,\big\{u\le {\textstyle \frac 3 4}\big\} \,\subset \, \big\{\omega_j\cdot x\le a 2^{j(1+\alpha_0)} \big\} \quad \mbox{in }B_{2^j},
\end{equation}
for $0\le j\le j_a$, for some $\omega_j\in S^{n-1}$ .
Then, for all $i= 1,2, \dots, k$ we have
\begin{equation} \label{hwoeithwoih}
\left\{\omega_{-i}\cdot x \le -
\frac{a}{2^{(1+\alpha_0)i}} \right\} \,\subset \, \big\{u\le -{\textstyle \frac 3 4} \big\} \,\subset\, \big\{u\le {\textstyle \frac 3 4} \big\} \,\subset\, \left\{\omega_{-i}\cdot x\le \frac{a}{2^{(1+\alpha_0)i}} \right\} \;\; \mbox{in }B_{2^{-i}},
\end{equation}
for certain $\omega_{-i}\in S^{n-1}$.
\end{prop}
\begin{proof}
The proof will apply inductively Theorem 1.1 of \cite{dPSV} to $u^{(i)}(x) : = u_\varepsilon(2^{i-1}x)$.
Indeed, recall first that ---see \eqref{21huwgiuw}--- we have $-W''(t) \in [-2,-1/2]$ for $|t|\ge \frac{1}{\sqrt 2}$. Since $\frac 3 4 \ge \frac 1 {\sqrt 2}$ we may take the constant $\kappa$ from \cite{dPSV} equal to $1/4$.
Notice that the case $i=1$ of \eqref{hwoeithwoih} follows directly from Theorem 1.1 in \cite{dPSV} since $ 2^{k-1}\varepsilon \le \big(2^{-\alpha_0(k-1)}a \big)^{p_0}$ and $k\ge 1$ guarantee $\varepsilon\le a^{p_0}$.
Assume now that \eqref{hwoeithwoih} holds for $ i = 1, 2, \dots, i_\circ-1$ and that $i_\circ\le k$. Then $u^{(i_\circ)}(x)$ satisfies the assumptions of Theorem 1.1 in \cite{dPSV} with
and $a$ replaced by $2^{-\alpha_0(i-1)}a$ and $\varepsilon$ replaced by $2^{i-1}\varepsilon$ (too see this it may be useful to notice that, by the definition of $j_a$ in \eqref{ja}, we have $ j_{2^{-\alpha_0(i-1)}a} = j_a+ (i-1)$), since we have
\[
2^{(i_\circ-1)}\varepsilon \le 2^{k-1}\varepsilon \le \big(2^{-\alpha_0(k-1)}a \big)^{p_0}\le \big(2^{-\alpha_0(i_\circ-1)}a \big)^{p_0}.
\]
Hence, $u^{(i_\circ)}$ satisfies the conclusion Theorem 1.1 in \cite{dPSV} so, after rescaling, we obtain that \eqref{hwoeithwoih} also holds also for $i = i_\circ$.
\end{proof}
The second result is an easy consequence of Proposition \ref{cor:improvement}: flatness implies a $C^{1,\alpha}$ type result.
\begin{thm}\label{thm63utbg}
Let $n\ge 2$, $s\in(0,1)$, and $W(u)=\frac 1 4 (1-u^2)^2$. Given $\tilde a>0$ there exist positive constants $\sigma_0$, $\delta_0$, $\alpha_0$, and $\varrho_0$, depending only on $\tilde a$, $n$, and $s$, such that the following holds.
Assume that $u_{\tilde \varepsilon}$ is a solution of $(-\Delta)^{s/2} u_{\tilde \varepsilon} + {\tilde \varepsilon}^{-s} W'(u_{\tilde \varepsilon})=0$ in~$B_1$ satisfying
\begin{equation}\label{xxxeee11}
\{x_n< -\sigma_0\} \subset \{u_{\tilde \varepsilon}<-{\textstyle \frac 3 4} \} \subset \{u_{\tilde \varepsilon}<{\textstyle \frac 3 4} \} \subset \{x_n< \sigma_0\} \quad \mbox{in } B_1.
\end{equation}
Then, for all $z\in \{u_{\tilde \varepsilon}=0\} \cap B_{3/4}$ and $k\ge 2$ satisfying $2^{-k} \ge \tilde \varepsilon^{\,\delta_0}$ we have
\begin{equation}\label{xxxeee22}
\begin{split}
\{\omega_{z,k}\cdot(x-z)< -\tilde a 2^{-(1+\alpha_0)k}\varrho_0 \} \subset \{u_{\tilde \varepsilon} <-{\textstyle \frac 3 4} \} \subset \\
\subset \{u_{\tilde \varepsilon}<\textstyle \frac 3 4\} \subset \{\omega_{z,k}\cdot(x-z)< \tilde a 2^{-(1+\alpha_0)k}\varrho_0\}
\end{split}
\end{equation}
in $B_{2^{-k} \varrho_0}(z)$, for some $\omega_{z,k}\in S^{n-1}$.
\end{thm}
\begin{proof}
Let $a_0$, $\alpha_0$, $p_0$ be the constants from Proposition \ref{cor:improvement} (which depend only on $n$ and $s$) and define $\delta_0 := \frac{1}{2+\alpha_0p_0}$.
It remains to choose $\sigma_0>0$ depending only on $\tilde a$, $n$, and $s$.
Note that we may assume
\begin{equation}\label{hweiothiowehoiw}
\tilde a \tilde \varepsilon^{(1+\alpha_0)\delta_0} \le \sigma_0
\end{equation}
since otherwise \eqref{xxxeee22} follows immediately from \eqref{xxxeee11}.
Choose
\begin{equation}\label{ewhtiowhiowh}
a : = \min\{\tilde a, a_0\} \qquad \mbox{and} \qquad \varrho_0:= 2^{-j_a-2},
\end{equation}
where $j_a$ was defined as in \eqref{ja}.
Now, for any $z\in \{u_{\tilde \varepsilon} =0\} \cap B_{3/4}$, let
\[ v^z : = u_{\tilde \varepsilon}( z + \varrho_0\,\cdot\,).\]
Note that $v^{z}$ satisfies $(-\Delta)^{s/2} v^{z} + (\tilde\varepsilon/\varrho_0)^{-s} W'(v^{z})=0$ in $B_{\frac{1}{4\varrho_0}} = B_{2^{j_a}}$.
Choose now $\sigma_0>0$ small so that $\sigma_0/\varrho_0 \le a$. Then it is immediate to verify that, thanks to \eqref{xxxeee11}, $v^{z}$ satisfies the assumption \eqref{wethiowehtowh} of Proposition \ref{cor:improvement} with $\omega_j = e_n$ for all $j=0,\dots, j_a$.
Hence, defining $\varepsilon := \tilde \varepsilon/\varrho_0$, provided
\begin{equation} \label{whoithowi}
2^{k-1}\varepsilon \le (2^{-\alpha_0(k-1)} a)^{p_0},
\end{equation}
Proposition \ref{cor:improvement} yields
\begin{equation}\label{xxxeee3}
\{\omega_{z,k}\cdot x < -\tilde a 2^{-(1+\alpha_0)k} \} \subset \{u_{\tilde \varepsilon} <-{\textstyle \frac 3 4} \} \subset
\{u_{\tilde \varepsilon}<\textstyle \frac 3 4\} \subset \{\omega_{z,k}\cdot x< \tilde a 2^{-(1+\alpha_0)k}\}
\end{equation}
in $B_{2^{-k}}$, for some $\omega_{z,k}\in S^{n-1}$. Note that \eqref{xxxeee3} immediately yields \eqref{xxxeee22} after scaling.
Thus, it only remains to show that \eqref{whoithowi} is satisfied thanks to our assumption $2^{-k} \ge \tilde \varepsilon^{\,\delta_0}$ and our choice of $\delta_0$.
Indeed,
\[
\eqref{whoithowi} \quad \Leftrightarrow \quad \frac{\tilde \varepsilon}{\varrho_0} \le ( 2^{1-k})^{\alpha_0p_0+1} a^{p_0}\quad \Leftrightarrow \quad 2^{-k} \ge \frac{(\varrho_0 a^{p_0})^{-\frac{1}{\alpha_0p_0 +1}}}{2} \tilde \varepsilon^{\frac{1}{1+\alpha_0p_0}} = c_a\tilde \varepsilon^{\frac{1}{1+\alpha_0p_0}}.
\]
But since we choose $\delta_0< \frac{1}{1+\alpha_0p_0}$, recalling \eqref{hweiothiowehoiw} we can absorb the multiplicative constant after possibly decreasing $\sigma_0$ in order to ensure that $\tilde \varepsilon$ is sufficiently small ---recall that $a$ was fixed in \eqref{ewhtiowhiowh}.
\end{proof}
Now, we introduce a class of ``good'' sets. By definition, the characteristic function of a ``good'' set must be the limit of stable solutions to the fractional Allen-Cahn equation with parameter $\varepsilon$, as a sequence of $\varepsilon$ tends to $0$. As we will show in Proposition \ref{good} below, these sets are ``good'' in the sense that they inherit several good properties from the approximating sequence, such as $BV$ and energy estimates, a monotonicity formula, density estimates, and the improvement of flatness. As approximating sequence, we will take later the blow-downs of an entire stable solution to the fractional Allen-Cahn equation.
\begin{defi}\label{defA}
We say that a set $E\subset \mathbb{R}^n$ belongs to the class $\mathcal A$ when
there exists a sequence of functions $u_j$, with $|u_j|< 1$, which are stable solutions of
\begin{equation}\label{approx-AC}(-\Delta)^{s/2}u_j+\varepsilon_j^{-s} W'(u_j)=0 \quad \mbox{in } \mathbb{R}^n, \quad \mbox{with}\quad \varepsilon_j\downarrow 0 \;\;\mbox{as}\;\;j\uparrow \infty,\end{equation}
where $W(u) = \frac 1 4 (1-u^2)^2$, and such that
$$u_j \stackrel{L^1_{\rm loc}}{\longrightarrow} \chi_E-\chi_{E^c}\;\;\mbox{as}\;\;j\uparrow \infty.$$
\end{defi}
\begin{prop}\label{good}
Any set $E\in\mathcal A$ satisfies the following properties.
\begin{enumerate}
\item \textbf{BV and energy estimates}.
\begin{equation}\label{estimates-sets}{\rm Per}(E,B_R)\leq CR^{n-1}\quad \mbox{and}\quad P_s(E,B_R)\leq CR^{n-s},\end{equation}
where $C$ is a constant which depends only on $n$ and $s$.
\vspace{0.5em}
\item \textbf{Monotonicity formula}. Let $\bar u :=\chi_E-\chi_{E^c}$ and let $\bar U$ be its $s$-extension in $\mathbb{R}^{n+1}_+$ (see Definition \ref{defextension}). We set
$$\Phi_E(R)=\frac{1}{ R^{n-s}}\int_{\widetilde B_{R}^+} y^{1-s}|\nabla \bar U(x,y)|^2\,dx\,dy.$$
Then, $\Phi_E$ is a nondecreasing function of $R$ and $\Phi_E(R)$ is constant if and only if $E$ is a cone (i.e., $\bar U$ is homogeneous of degree $0$).
\vspace{0.5em}
\item \textbf{Density estimate}. For some positive constant $\omega_0$, which depends only on $n$ and $s$, we have that
if
\begin{equation*}\label{hp-densitysets}
R^{-n}|E\cap B_R| \leq \omega_0 \quad (\mbox{respectively, } R^{-n}| E^c\cap B_R| \leq \omega_0)
\end{equation*}
for some $R>0$, then
\begin{equation*}\label{th-densitysets}
|E \cap B_{R/2}| = 0 \quad (\mbox{respectively, } | E^c \cap B_{R/2}| = 0).
\end{equation*}
\item \textbf{Improvement of flatness}. There exists $\sigma_0>0$, which depends only on $n$ and $s$, such that
$$\mbox{if}\quad \partial E\cap B_2 \subset \{x\in \mathbb{R}^n\,:\, |x\cdot e_n|\leq \tilde \sigma_0\},$$
then $\partial E\cap B_{1/2}$ is a $C^{1,\alpha}$ graph in the $e_n$ direction.
\vspace{0.5em}
\item \textbf{Blow-up}. Let $E_{r_i,x_0}:=\frac{E-x_0}{r_i}$ with $r_i\downarrow 0$ as $i\uparrow \infty$. If $E_{r_i,x_0}\stackrel{L^1_{\rm loc}}{\longrightarrow} E_*$, then $E_* \in \mathcal A$.
\end{enumerate}
\end{prop}
\begin{rem}
None of the properties (1)-(5) in Proposition \ref{good} are known to hold within the class of all weakly stable sets for the fractional perimeter $P_s$.
This is the reason that brings us to introduce the class $\mathcal A$.
\end{rem}
The results established in our paper allow us to give the
\begin{proof}[Proof of Proposition \ref{good}]
Recall that (see Definition \ref{defA}) $E\in \mathcal A$ if there exists a sequence $u_j$ of stable solutions to the fractional Allen-Cahn equation with parameter $\varepsilon_j\downarrow 0$ such that $u_j \to \chi_E- \chi_{E^c}$ as $j\uparrow \infty$.
(1) The first estimate in \eqref{estimates-sets} follows easily passing to the limit the $BV$ estimate $u_j$ established in \eqref{BVest} (the total variation is lower-semicontinuous).
We emphasize again ---see Remark \ref{remscaling}--- that the $BV$ estimate \eqref{BVest} is independent of the potential $W$ in the statement of Theorem \ref{BV}. We are strongly using this here since $u_j$ satisfies $(-\Delta)^{s/2}u_j+\varepsilon_j^{-s} W'(u_j)=0$ and hence the associated potential $\varepsilon_j^{-s} W(u)$ converges to infinity (since $\varepsilon_j \downarrow 0$) except at $u=\pm1$. Similarly, the estimate on the fractional perimeter follows passing to the limit the estimate of Corollary \ref{energy-est}.
\smallskip
(2) Denote $\bar u : = \chi_{E}-\chi_{E^c}$. Let $U_j$ be the $s$-extension of $u_j$ (as defined in Section~\ref{sec-5}). By Proposition \ref{mon} (rescalled), we know that
$$\Phi_j(R):= \frac{1}{R^{n-s}} \Big\{ \frac{d_s}{2} \int_{\tilde B_R^+} y^{1-s} |\nabla U_j (x,y)|^2 \,dx\,dy + \int_{B_R} \varepsilon_j^{-s} W(u_j(x))\,dx \Big\}
$$
is a nondecreasing function of $R$.
Hence, we have a sequence $\Phi_j(R)$ of nondecreasing functions which is uniformly bounded (in $R$ and $j$) by the energy estimate \eqref{egest} of Theorem \ref{thm1}.
Moreover, arguing as in the proof of Proposition \ref{conv-L^1} ---see \eqref{strong}---, the convergence $u_j\to \bar u$ in $L^1_{\rm loc}(\mathbb{R}^n)$ can be upgraded to a strong convergence in $W^{s/2,2}_{\rm loc}(\mathbb{R}^n)$ and also $U_j\to \bar U$ in $W^{1,2}_{\rm loc}(\mathbb{R}^{n+1}_+, y^{1-s})$.
Also, thanks to Proposition \ref{propnew}, the potential term in $\Phi_j(R)$ converges to zero (as $\varepsilon_j \downarrow 0$) for any fixed $R>0$.
Hence, passing to the limit as $j\to \infty$, we deduce that this monotonicity property is satisfied by the limiting function $\Phi_E(R)$. Moreover similarly as in \eqref{wnoiwoieht}-\eqref{wnoiwoieht2} we obtain
\begin{equation}
\Phi_j(R_2)-\Phi_j(R_1) \ge d_s \int_{B_{R_2}\setminus B_{R_1}} \frac{y^{1-s}}{( |x|^2 +y^2)^{\frac{n-s}{2}}}(\partial_{r} U_j)^2 \,dx\,dy.
\end{equation}
Since $U_j\to \bar U$ strongly $W^{1,2}_{\rm loc}(\mathbb{R}^{n+1}_+, y^{1-s})$ we also obtain that if $\Phi_E$ is constant for $\bar U$ then $E$ must be a cone.
\smallskip
(3) The density estimate follows easily by passing to the limit the corresponding density estimate (established in Proposition \ref{density}) for the approximating sequence $u_j$.
\smallskip
(4) It follows from Theorem \ref{thm63utbg} and the density estimates in Proposition \ref{density}. Indeed, fix $\tilde a>0$ and assume that $\partial E\cap B_2 \subset \{x\in \mathbb{R}^n\,:\, |x\cdot e_n|\leq \tilde \sigma_0\}$. Then the density estimates imply the convergence of $\{u_j>t\}$ in Hausdorff distance towards $E$ for all $t\in (-1,1)$ {(see the last part of the statement of Theorem \ref{thm2})} we have, for $\sigma_0 = 8\tilde \sigma$,
\[
\big\{x_n< - \sigma_0\big\} \subset \{u_j<-{\textstyle \frac 3 4 } \} \subset \{u_j<{\textstyle \frac 3 4 }\} \subset \big\{x_n< \sigma_0\big\} \quad \mbox{in } B_{1}.
\]
Then, Theorem \ref{thm63utbg} (rescaled) yields that for all $z\in \{u_{j}=0\} \cap B_{\frac{3}{4}}$ and $k\ge 1$ satisfying $2^{-k} \ge \varepsilon_j^{\,\delta_0}$ we have
\[
\begin{split}
\{\omega_{z,k}\cdot(x-z)< -\tilde a 2^{-(1+\alpha_0)k} \varrho_0\} \subset \{u_{\tilde \varepsilon_j} <-{\textstyle \frac 3 4 } \} \subset \\
\subset \{u_{\tilde \varepsilon}<{\textstyle \frac 3 4 }\} \subset \{\omega_{z,k}\cdot(x-z)< \tilde a 2^{-(1+\alpha_0)k}\varrho_0\}
\end{split}
\]
in $B_{2^{-k}\varrho_0}(z)$, for some $\omega_{z,k}\in S^{n-1}$. After passing this information to the limit (using again Hausdorff convergence), we deduce that for all $z\in \partial E \cap B_{3/4}$ we have
\begin{equation}\label{wtnoiwhtiowhwo}
\{\omega_{z,k}\cdot(x-z) \le -\tilde a 2^{-(1+\alpha_0)k}\varrho_0\} \subset E \subset \{\omega_{z,k}\cdot(x-z)< \tilde a 2^{-(1+\alpha_0)k}\varrho_0\}
\end{equation}
in $B_{2^{-k}\varrho_0}(z)$ for all $k \ge 0$.
Similarly as for classical minimal surfaces, this implies that $\partial E$ is a $C^{1,\alpha}$ graph inside $B_{1/2}$ provided $\tilde a$ is chosen small enough depending on $\alpha_0$ (note that \eqref{wtnoiwhtiowhwo} yields $|\omega_{z,k} -\omega_{z,k+1}| \le C_0 \tilde a 2^{-\alpha_0 k}$ and hence, by triangle inequality and summing a geometric series $|e_n - \omega_{z,k}| \le C_0 \tilde a \frac{1}{1-2^{-\alpha_0}}<\frac {1}{10}$ for all $k\ge 0$, provided $\tilde a$ is chosen small).
\smallskip
(5) Since $E^i:=E_{r_i,x_0}$ belongs to $\mathcal A$ for every $i\in \mathbb N$, then it can be approximated by a sequence $u^i_j$ as in \eqref{approx-AC}. By assumption for each $m\in \mathbb N$ there exists $i_m$ such that $| (E^{i_m}\setminus E_*) \cup (E_*\setminus E^{i_m} ))\cap B_m| \le \frac 1 m$. Also given this $i_m$ there exists $j_m$ such that
$$
\int_{B_m} \big|u^{i_m}_{j_m} - (\chi_{E^{i_m}}-\chi_{(E^{i_m})^c}) \big| \le \frac 1 m.
$$
Hence, $u^{i_m}_{j_m} \stackrel{L^1_{\rm loc}}{\longrightarrow} \chi_{E_*}-\chi_{E^c_*}$ and thus $E_*\in \mathcal A$.
\end{proof}
We also need the following classification theorem for cones in $\mathcal A$ which are translation invariant in all directions but two of them.
\begin{lem}\label{hwioheoithe}
Assume that some nontrivial cone $\Sigma\subset \mathbb{R}^n$ belongs to $\mathcal A$ and is of the form
\[\widetilde\Sigma \times \mathbb{R}^{n-2}\]
for some cone $\widetilde\Sigma\subset \mathbb{R}^2$.
Then, $\Sigma$ is a half-space.
\end{lem}
\begin{proof}
If $\Sigma$ (which is nontrivial) is not a half-space, then $\tilde \Sigma$ is not a half-plane (and is also nontrivial). Hence, $\partial\widetilde\Sigma$ must contain at least two non-aligned rays.
Recall that, thanks to the density estimates for the class $\mathcal A$, we can chose a representative among sets that differ from $\Sigma$ for a set of measure zero such that every point of the topological boundary of $\Sigma$ has positive density for both $\Sigma$ and $\Sigma^c$.
{Also, thanks to the improvement of flatness property for the class $\mathcal A$, the outwards normal vectors (in~$\mathbb{R}^2$) $\nu_1$ and $\nu_2$ to these two non-aligned rays need to form some positive angle, that is, $|\nu_1-\nu_2|^2\ge c>0$, for some positive constant $c$ depending only on $n$ and $s$.
Let us denote by $\widetilde H_1$ and $\widetilde H_2$ these two non aligned rays, that is for $i=1,2$, we set
$$\widetilde H_i:=\{\widetilde x \in \mathbb{R}^2\,|\,\widetilde x=t\,\omega_i,\,t>0\},$$
for some vectors $\omega_1,\,\omega_2\in \mathbb{R}^2$ such that $\omega_i \cdot \nu_i=0$ for $i=1,2$.
Moreover, we set
$$H_i =\widetilde H_i\times \mathbb{R}^{n-2},\quad i=1,\,2.$$
With these notations, we have that
$$H_1\times H_2 \subset \partial \Sigma \times \partial \Sigma.$$}
We fix a non-increasing cutoff function $\xi \in C^\infty_c([0,3))$ such that $\xi \equiv 1$ in $[0,2]$, and $\xi \le 1$, and we define
\begin{equation}\label{psi}
\psi(x) = \xi\bigg(\sqrt{x_1^2+x_2^2}\bigg) \xi(|x_3|)\cdots \xi(|x_n|).
\end{equation}
In what follows, we change notation and we denote points in $\mathbb{R}^n \times \mathbb{R}^n$ by $(x,y)$ ---instead of the usual $(x, \bar x)$.
We claim that, for every $0<r<1$, the following estimate holds:
\begin{equation}\label{I1}
I_1 : = \iint_{\partial\Sigma \times \partial\Sigma} \frac{\big|\nu_\Sigma (x) -\nu_\Sigma(y)\big|^2}{(r^2 +|x-y|^2)^{\frac{n+s}{2}}}\, \psi^2(x) \psi^2(y)d\mathcal H^{n-1}(x) d\mathcal H^{n-1}(y) \ge C(n,s) r^{-s},
\end{equation}
for some positive constant $C(n,s)$ depending on $n$ and $s$.
To prove the claim, we use the notation $x=(\widetilde x,x')\in \mathbb{R}^2 \times \mathbb{R}^{n-2}$. We have that
\begin{equation*}
\begin{split}
I_1 &\ge c\int_{H_1}\int_{H_2}\frac{\psi^2(x) \psi^2(y)}{(r^2+|x-y|^2)^{\frac{n+s}{2}}}d\mathcal H^{n-1}(x) d\mathcal H^{n-1}(y)\\
&\ge c\iint_{\mathbb{R}^{n-2}\times \mathbb{R}^{n-2}}dx'\,dy'\int_{\widetilde H_1}d\mathcal H^1({\widetilde x})\int_{\widetilde H_2}d\mathcal H^1({\widetilde y})\frac{\psi^2(x) \psi^2(y)}{(r^2+|x-y|^2)^{\frac{n+s}{2}}}.
\end{split}
\end{equation*}
Using the change of variables $X=(0,|\widetilde x|)$, $Y=(0,-|\widetilde y|)$ and the triangle inequality $|\widetilde x -\widetilde y| \le |\widetilde x|+|\widetilde y|=|X-Y|$, we get
\begin{equation*}
\begin{split}
I_1 &\ge c \iint_{\mathbb{R}^{n-2}\times \mathbb{R}^{n-2}}dx'\,dy'\int_{\{0\}\times\mathbb{R}^{+}}dX\int_{\{0\}\times\mathbb{R}^{-}}dY\frac{\psi^2(X,x')\psi^2(Y, y')}{(r^2+|X-Y|^2+|x'-y'|^2)^{\frac{n+s}{s}}}\\
&\ge c \int_{B_1^{n-2}}dx'\int_{\{y'\in \mathbb{R}^{n-2}\,|\,|x'-y'|<1\}}dy'\int_0^1 dX\int_{-1}^0dY\frac{1}{(r^2+|X-Y|^2+|x'-y'|^2)^{\frac{n+s}{s}}}\\
&\ge C c|B_1^{n-2}| \int_0^1d\rho\int_0^1 dX\int_{-1}^0dY\frac{\rho^{n-3}}{(r^2+|X-Y|^2+\rho^2)^{\frac{n+s}{s}}},
\end{split}
\end{equation*}
where we are identifying a point $(0,X)\in \{0\}\times \mathbb{R}^+$ with the real number $X\in \mathbb{R}^+$ and the integration over $\{0\}\times \mathbb{R}^+$ with integration over $\mathbb{R}^+$ (analogously for $Y\in \mathbb{R}^-$), and in the last inequality we have used polar coordinates in $\mathbb{R}^{n-2}$.
Finally, using the change of variables $\bar X=X/r$, $\bar Y=Y/r$, $\bar \rho=\rho/r$, and recalling that $0<r<1$, we deduce
\begin{equation*}
\begin{split}
I_1 &\ge C(n,s)\frac{1}{r^{n+s}}\int_0^1d\rho\int_0^1 dX\int_{-1}^0dY\frac{\rho^{n-3}}{\left(1+\frac{|X-Y|^2}{r^2}+\frac{\rho^2}{r^2}\right)^{\frac{n+s}{2}}}\\
&=C(n,s)\frac{1}{r^{n+s}}\int_0^{1/r}d\bar\rho\int_0^{1/r} d\bar X\int_{-1/r}^0 d\bar Y \frac{r^n\cdot \bar \rho^{n-3}}{(1+|\bar X-\bar Y|^2+\bar \rho^2)^{\frac{n+s}{2}}}\\
&\ge C(n,s)r^{-s}\int_0^{1}d\bar\rho\int_0^{1} d\bar X\int_{-1}^0d\bar Y \frac{ \bar \rho^{n-3}}{(1+|\bar X-\bar Y|^2+\bar \rho^2)^{\frac{n+s}{2}}}=C(n,s) r^{-s},
\end{split}
\end{equation*}
which concludes the proof of \eqref{I1}.
Recall now that since $\Sigma$ belongs to the class $\mathcal A$, there exists, by definition, a sequence of functions $u_j: \mathbb{R}^n \to (-1,1)$ which are stable solutions of $(-\Delta)^{s/2}u_j+\varepsilon_j^{-s} W'(u_j)=0$
and such that
$$u_j \stackrel{L^1_{\rm loc}}{\longrightarrow} \chi_\Sigma-\chi_{\Sigma^c}\;\quad \mbox{as}\;\;j\uparrow \infty.$$
We claim now that the following inequality holds:
\begin{equation}\label{I2-I3}
\begin{split}
I_2: = \iint_{\mathbb{R}^n \times \mathbb{R}^n} &\frac{\big|n_j (x) -n_j(y)\big|^2}{|x-y|^{n+s}} \psi(x)^2 |\nabla u_j|(x) dx\, |\nabla u_j|(y) dy
\\
&\qquad \le \iint_{\mathbb{R}^n \times \mathbb{R}^n} \frac{|\psi(x) -\psi(y)|^2}{|x-y|^{n+s}} |\nabla u_j|(x) dx\, |\nabla u_j|(y) dy =:I_3,
\end{split}
\end{equation}
where
\begin{equation}\label{nu}
n_j(x) : =
\begin{cases}
\frac{\nabla u_j}{|\nabla u_j|} &\quad \mbox{where } \nabla u \neq 0 \\
0 & \quad\mbox{where } \nabla u = 0.
\end{cases}
\end{equation}
Let us prove \eqref{I2-I3}. We start by observing that the stability condition \eqref{stable} (with $\Omega=\mathbb{R}^n$), written for the functions $u_j$, is equivalent to requiring that
\begin{equation}\label{stable2}
\int_{\mathbb{R}^n} \xi(x)(-\Delta)^{s/2}\xi(x)\,dx + \int_{\mathbb{R}^n}\varepsilon^{-s}_j W''(u_j)\xi^2(x)\,dx\ge 0
\end{equation}
for any Lipschitz function $\xi$ which is compactly supported in $\mathbb{R}^n$.
Let us now choose $\xi$ of the form $\xi=\eta \cdot \psi$, where $\eta$ is a Lipschitz function and $\psi \in C^\infty_0(\mathbb{R}^n)$.
A simple computation gives that
$$(-\Delta)^{s/2}\xi(x)=\psi(x)(-\Delta)^{s/2}\eta(x) + \int_{\mathbb{R}^n}\eta(y) \frac{\psi(x)-\psi(y)}{|x-y|^{n+s}}\,dy,$$
which implies
\begin{equation*}
\begin{split}
\int_{\mathbb{R}^n}\xi(x)(-\Delta)^{s/2}\xi(x)\,dx &= \int_{\mathbb{R}^n} \psi^2(x)\eta(x) (-\Delta)^{s/2}\eta(x)\,dx\\
&\hspace{1em}+\iint_{\mathbb{R}^n \times \mathbb{R}^n}\eta(x)\eta(y)\psi(x) \frac{\psi(x)-\psi(y)}{|x-y|^{n+s}}\,dx\,dy\\
&=\int_{\mathbb{R}^n} \psi^2(x)\eta(x) (-\Delta)^{s/2}\eta(x)\,dx\\
&\hspace{1em}+\frac{1}{2}\iint_{\mathbb{R}^n \times \mathbb{R}^n}\eta(x)\eta(y)\frac{|\psi(x)-\psi(y)|^2}{|x-y|^{n+s}}\,dx\,dy.
\end{split}
\end{equation*}
Hence, the stability condition \eqref{stable2} becomes
\begin{equation}\label{stable3}
\begin{split}
&\int_{\mathbb{R}^n} \psi^2(x)\eta(x) (-\Delta)^{s/2}\eta(x)\,dx
+\frac{1}{2}\iint_{\mathbb{R}^n \times \mathbb{R}^n}\eta(x)\eta(y)\frac{|\psi(x)-\psi(y)|^2}{|x-y|^{n+s}}\,dx\,dy\\
&\hspace{2em} + \int_{\mathbb{R}^n}\varepsilon^{-s}_j W''(u_j)\eta^2(x)\psi^2(x)\,dx \ge 0.
\end{split}
\end{equation}
We use now the fact that, for any $i=1,\dots,n$, $\partial_{x_i}u_j$ satisfies the linearized equation $(-\Delta)^{s/2}v+\varepsilon_j^{-s}W''(u)v=0$. By multiplying the (vectorial) equation satisfied by $\nabla u_j$ by $\nabla u_j$ itself, we deduce that
\begin{equation}\label{eq-grad}
\nabla u_j \cdot(-\Delta)^{s/2}\nabla u_j + \varepsilon_j^{-s}W''(u)|\nabla u_j|^2=0.
\end{equation}
Let us now choose $\eta=|\nabla u_j|$ in the stability inequality \eqref{stable3} (this is an admissible choice by the regularity results of Appendix \ref{app-C}) and use \eqref{eq-grad}, to obtain
\[
\begin{split}
&\int_{\mathbb{R}^n}\psi^2(x)|\nabla u_j|(x)(-\Delta)^{s/2}|\nabla u_j|(x)\,dx \\
&\hspace{2em}+\frac{1}{2}\iint_{\mathbb{R}^n \times \mathbb{R}^n}|\nabla u_j(x)||\nabla u_j(y)|\frac{|\psi(x)-\psi(y)|^2}{|x-y|^{n+s}}\,dx\,dy\\
&\hspace{2em}-\int_{\mathbb{R}^n}\psi^2(x)\nabla u_ j(x) \cdot (-\Delta)^{s/2} \nabla u_j(x)\,dx \ge 0.
\end{split}
\]
To conclude the proof of \eqref{I2-I3}, it is enough to observe that
\[
\begin{split}
&\int_{\mathbb{R}^n}\psi^2(x)|\nabla u_j|(x)(-\Delta)^{s/2}|\nabla u_j|(x)\,dx -\int_{\mathbb{R}^n}\psi^2(x)\nabla u_ j(x) \cdot (-\Delta)^{s/2} \nabla u_j(x)dx\\
&=\negmedspace \iint_{\mathbb{R}^n \times \mathbb{R}^n}\negmedspace\psi^2(x)\negmedspace\left(|\nabla u_j|(x) \frac{|\nabla u_j|(x)-|\nabla u_j|(y)}{|x-y|^{n+s}}-\nabla u_j(x)\cdot \frac{\nabla u_j(x)-\nabla u_j(y)}{|x-y|^{n+s}}\right)\negmedspace dx dy\\
&=\negmedspace \iint_{\mathbb{R}^n \times \mathbb{R}^n}\negmedspace\psi^2(x)\frac{\nabla u_j(x)\cdot \nabla u_j(y) - |\nabla u_j|(x)|\nabla u_j|(y)}{|x-y|^{n+s}}dxdy.
\end{split}
\]
This, together with the definition \eqref{nu} of $n_j$, proves \eqref{I2-I3}.
By our uniform $BV$ estimates, we have that $\nabla u_j \to -2D\chi_{\Sigma}$ weakly$^*$ as Radon measures (see, e.g., Proposition 3.13 and formula (3.11) in \cite{AFP}). Here and in the following, we denote by $D\chi_{\Sigma}$ the perimeter measure and by $|D\chi_{\Sigma}|$ its total variation.
Let us show that
\[ I_1 \le I_2\]
for $j$ large enough, where these quantities were defined in \eqref{I1} and \eqref{I2-I3}.
Indeed, we start by observing that
\[
\begin{split}
I_2 &= 2 \iint_{ \mathbb{R}^n\times \mathbb{R}^n} \frac{ \big( |\nabla u_j|(x) |\nabla u_j|(y) - \nabla u_j(x)\cdot \nabla u_j(y) \big)}{|x-y|^{n+s}} \psi^2(x)\, dx\, dy \\
&\ge 2 \iint_{ \mathbb{R}^n\times \mathbb{R}^n} \frac{ \big( |\nabla u_j|(x) |\nabla u_j|(y) - \nabla u_j(x)\cdot \nabla u_j(y) \big)}{(r^2+ |x-y|^2)^{\frac{n+s}{2}}} \psi^2(x)\psi^2(y) \, dx\, dy.
\end{split}
\]
Now, we claim that:
\begin{equation}\label{conv1}
\begin{split}
&\lim_{j\rightarrow \infty} \iint_{ \mathbb{R}^n\times \mathbb{R}^n} \frac { \nabla u_j(x) \cdot\nabla u_j(y)} {(r^2+ |x-y|^2)^{\frac{n+s}{2}}} \psi^2(x)\psi^2(y) \, dx\, dy \\
&\hspace{2em}= \iint_{ \mathbb{R}^n\times \mathbb{R}^n} \frac{ 2 D \chi_{\Sigma}(dx) \cdot 2 D \chi_\Sigma (dy)} {(r^2+ |x-y|^2)^{\frac{n+s}{2}}} \psi^2(x)\psi^2(y)
\end{split}
\end{equation}
and
\begin{equation}\label{conv2}
\begin{split}
&\liminf_{j\rightarrow \infty} \iint_{ \mathbb{R}^n\times \mathbb{R}^n} \frac { |\nabla u_j|(x) |\nabla u_j|(y)} {(r^2+ |x-y|^2)^{\frac{n+s}{2}}} \psi^2(x)\psi^2(y) \, dx\, dy \\
&\hspace{2em} \ge\iint_{ \mathbb{R}^n\times \mathbb{R}^n} \frac{ 2|D \chi_{\Sigma}|(dx) 2|D \chi_\Sigma |(dy)} {(r^2+ |x-y|^2)^{\frac{n+s}{2}}} \psi^2(x)\psi^2(y).
\end{split}
\end{equation}
We first prove \eqref{conv1}.
For every $y\in \mathbb{R}^n$, we define
$$G_j(y):=\int_{\mathbb{R}^n}\nabla u_j(x)\frac{\psi^2(x)\psi^2(y)}{(r^2+|x-y|^{2})^{\frac{n+s}{2}}}\,dx=\int_{\mathbb{R}^n}\nabla u_j(x)\Psi(x,y)\,dx.$$
Using the BV estimate for $u_j$ (which is uniform in $j$) and that, for $y$ fixed, the function $\Psi$ is smooth and compactly supported (as function of $x$), we deduce that the family of functions $\{G_j\}_{j\in \mathbb N}$ is equibounded and equicontinuous. This, combined with the weak$^*$-convergence of $\nabla u_j$ to $-2D\chi_{\Sigma}$, gives that, as $j\rightarrow \infty$, $G_j$ converges uniformly to $G$, where
$$G(y):= -2\int_{\mathbb{R}^n}D\chi_{\Sigma}(dx)\Psi(x,y).$$
Hence, we have that
\begin{equation*}
\begin{split}
&\iint_{ \mathbb{R}^n\times \mathbb{R}^n} \nabla u_j(x) \cdot\nabla u_j(y)\Psi(x,y)\, dx\, dy\\
&\hspace{2em}-\iint_{ \mathbb{R}^n\times \mathbb{R}^n} 2 D \chi_{\Sigma}(dx) \cdot 2 D \chi_\Sigma (y)\Psi(x,y) dy \\
&\hspace{1em}=\int_{\mathbb{R}^n}\nabla u_j(y)\cdot G_j(y)\, dy +2\int_{R^{n}}D\chi_{\Sigma}(dy)\cdot G(y)\\
&\hspace{1em}=\int_{\mathbb{R}^n}\nabla u_j(y)\left(G_j(y)-G(y)\right)\,dy+\int_{\mathbb{R}^n}G(y)\left(\nabla u_j(y)+2 D\chi_{\Sigma}(dy)\right)\rightarrow 0,
\end{split}
\end{equation*}
where the first term tends to zero thanks to the uniform convergence of $G_j$ to $G$ and the uniform BV estimate (on compact sets) for $u_j$, and the second term also vanishes in the limit since $\nabla u_j$ weak$^*$-converge to $-2 D\chi_{\Sigma}$ and $G$ is smooth and compactly supported.
To get \eqref{conv2}, we reason similarly as before and use the lower-semicontinuity of the total variation.
Hence, from \eqref{conv1} and \eqref{conv2}, we get, as $j\to \infty$,
\[
I_2\ge 4I_1 -o(1) \ge I_1.
\]
Finally, let us show that
\[
I_3 \le C_1(n,s),
\]
where $I_3$ was defined in \eqref{I2-I3}.
Indeed, by Theorem \ref{BV} we have $\int_{B_\varrho (x)} |\nabla u_j(y)|\,dy \le C(n,s) \varrho^{n-1}$ for all $x\in \tilde B_3\times[-3,3]^{n-2}$ and $\rho >0$ (here $\tilde B_3$ denote the ball in $\mathbb{R}^2$ centered at $0$ and with radius $3$). Hence
for all $x\in \tilde B_3\times[-3,3]^{n-2}$, defining $A_j(x)= B_{2^{j+1}}(x) \setminus B_{2^j}(x)$ and using that $|\psi(x) -\psi(y)|^2 \le C(n)(|x-y|^2\wedge 1)$, we have
\[
\begin{split}
\int_{\mathbb{R}^n} \frac{(|x-y|^2\wedge 1)}{|x-y|^{n+s}} |\nabla u_j (y)| dy &\le C\sum_{j\in \mathbb{Z}} \frac{2^{2j}\wedge 1}{2^{j(n+s)}} 2^{j (n-1)}
\\
&= C \bigg(\sum_{j<0} 2^{(1-s)j} + \sum_{j\ge 0} 2^{-(1+s)j}\bigg) \le C_1(n,s).
\end{split}
\]
Hence we have shown that, for $j$ large enough,
\[C(n,s) r^{-s}\le I_1 \le I_2 \le I_3\le C_1(n,s). \]
Choosing $r>0$ small we obtain a contradiction.
\end{proof}
We have now all the ingredients to prove our main theorem.
\begin{proof} [Proof of Theorem \ref{thmclas}]
By our convergence result Theorem \ref{thm2} for any given blow-down sequence $u_{R_j}(x)= u(R_jx)$ with $R_j\uparrow \infty$, there is a subsequence $R_{j_\ell}$ such that
\[ u_{R_{j_\ell}} \rightarrow \chi_{\Sigma}-\chi_{\Sigma^c} \quad \mbox{ in }L^1(B_1),\]
where $\Sigma$ is a cone which is a weakly stable set in $\mathbb{R}^n$ for the $s$-perimeter, which belongs to the class $\mathcal A$ (by definition of $\mathcal A$), and which is nontrivial.
We next prove that under our assumption, i.e., that the half-spaces are the only smooth (away from $0$) stable nonlocal $s$-minimal cones in $\mathbb{R}^m\setminus\{0\}$ for any $3\leq m\leq n$, $\Sigma$ must be a half-space. The proof follows Federer's dimension reduction argument and is done by contradiction. Indeed, assume that $\Sigma$ is not a half-space, then, by the just mentioned assumption, there exists at least one point $p_1\in \partial \Sigma\cap S^{n-1}$ at which $\partial \Sigma$ is not smooth in any neighbourhood of $p_1$.
Let us consider the blow-up of $\Sigma$ at $p_1$,
$$\Sigma_{p_1,r}:=\frac{\Sigma-p_1}{r}.$$
Using the energy estimate and the monotonicity formula (points (1) and (2) in Proposition \ref{good}) we have that, up to a subsequence,
$$\Sigma_{p_1,r}\stackrel{L^1_{\rm loc}}{\longrightarrow} \Sigma_1,$$
where $\Sigma_1$ belongs to $\mathcal A$ by point (5) in Proposition \ref{good}. Moreover, by the density estimate (point (3) in Proposition \ref{good}), the convergence of blow-ups in the $L^1_{\rm loc}$-sense can be upgraded to a local uniform convergence (i.e., locally in Hausdorff distance).
Now, if $\Sigma_1$ were a half-space, then by the improvement of flatness property (point (4) of Proposition \ref{good}) we would deduce that $\partial\Sigma$ is smooth in some neighbourhood of $p_1$, reaching a contradiction. Hence, $\Sigma_1$ is not a half-space. Now, $\Sigma_1$ being the blow-up of a cone at a point $p_1\neq 0$, we find that it must be translation invariant in the direction $p_1$.\footnote{This is done exactly as in the case of minimal surfaces:
\[x\in \Sigma_{p_1,r} \ \Leftrightarrow \ (p_1 + rx) \in \Sigma \ \Leftrightarrow \ \frac{\lambda}{r}(p_1 + rx) \in \Sigma \ \Leftrightarrow \ \big( {\textstyle \frac{\lambda-1}{r}}p_1 + \lambda x\big)\in \Sigma_{p_1, r} \]
and hence, taking $\lambda = 1+tr$ and sending $r\to 0$, we obtain $x\in \Sigma_{1} \ \Leftrightarrow\ x+ tp_1\in \Sigma_1$.
} Hence, up to a rotation, $\Sigma_1$ it must be of the form
$$\Sigma_1=\widetilde \Sigma_1\times \mathbb{R},$$
where $\widetilde \Sigma_1 \subset \mathbb{R}^{n-1}$ is a nontrivial cone different from a half-space (since we proved that $\Sigma_1$ cannot be a half-space).
We can now iterate the same argument: if $\tilde \Sigma_1$ were smooth then it should be flat by our assumption. Hence $\widetilde \Sigma_1$ is not smooth and thus there exists $p_2 = (\tilde p_2, 0)$, with $\tilde p_2\in \widetilde \Sigma_1 \cap S^{n-2}$, such that the blow-up $\Sigma_2$ of $\Sigma_1$ at the point $p_2$ belongs to $\mathcal A$ is nontrivial, is not a half-space, and is translation invariant in the direction $x_1$ and $x_2$. In other words, up to a rotation this second blow-up must be some set $\Sigma_2$ in $\mathcal A$ of the form
$$\Sigma_2=\widetilde \Sigma_2\times \mathbb{R}^2,$$
where $\widetilde \Sigma_2 \subset \mathbb{R}^{n-2}$ is a nontrivial cone different from a half-space.
After $n-2$ iterations we arrive at a blow-up $\Sigma_{n-2}$ which belongs to $\mathcal A$ and must be of the form
$$\Sigma_{n-2}=\widetilde \Sigma_{n-2}\times \mathbb{R}^{n-2},$$
where $\widetilde \Sigma_{n-2} \subset \mathbb{R}^{2}$ is a nontrivial cone different from a half-space. Hence, using Lemma \ref{hwioheoithe}, we reach a contradiction, proving that the initial cone $\Sigma$ must be a half-space.
Having proved that $\Sigma$ must be a half-space, we now recall that (by Theorem \ref{thm2}) the convergence of sub-levelsets of $u_{R_{j_\ell}}$ to the half-space $\Sigma$ (in $B_1$) holds also in the sense of the Hausdorff distance. As a consequence $u$ satisfies the asymptotic flatness assumption of Theorem \ref{123} and hence it follows that $u(x)=\phi(e\cdot x)$ for some direction $e\in S^{n-1}$ and some increasing function $\phi:\mathbb{R}\to (-1,1)$.
\end{proof}
\begin{rem}\label{thesame}
The following will be used in Appendix \ref{app-A} to deal with global stable $s$-minimal sets.
Notice that, reasoning exactly as in the proof of Theorem \ref{thmclas}, one can prove that under the same assumption, i.e. that, for some pair $(n,s)$ with $n\ge 3$ and $s\in(0,1)$, hyperplanes are the only stable $s$-minimal cones in $\mathbb{R}^n\setminus\{0\}$, then any set $E\subset \mathbb{R}^n$ belonging to the class $\mathcal A$ is necessarily a half-space. Indeed, after doing a blow-down of $E$, we reduce to a cone $\Sigma$ which is stable and belongs to $\mathcal A$. Hence, by the exact same argument as before, one get that $\Sigma$ is an half-space and finally, by the improvement of flatness property of $\mathcal A$, that so is $E$. More generally, an analogue abstract result which reduces the classification of a stable set $E$ to the classification of stable cones (smooth away from the origin), holds whenever $E$ belongs to a class for which the properties listed in Proposition \ref{good} are satisfied (BV and energy estimates, monotonicity formula, density estimates, improvement of flatness, and blow-up).
\end{rem}
We can now easily deduce our rigidity result in $\mathbb{R}^3$.
\begin{proof} [Proof of Corollary \ref{corclas2}]
Thanks to Theorem \ref{thmclas} we just need to show that for $s$ sufficiently close to 1, any weakly stable $s$-minimal cone in $\mathbb{R}^{3}\setminus\{0\}$ whose boundary is smooth away from $0$ must be either trivial or a half-space. This result was obtained in our previous paper \cite{CCS}.
\end{proof}
Instead, to prove Corollary \ref{corclas3} on monotone solutions in $\mathbb{R}^4$ we first need to establish the following result, which holds in any dimension.
\begin{prop}\label{implication}
Let $n\ge 2$, $s\in(0,1)$, and $W(u)=\frac 1 4 (1-u^2)^2$.
Assume that $u:\mathbb{R}^n \to (-1,1)$ is a solution of \eqref{theequation} satisfying $\partial_{x_n} u>0$. Define $u^{\pm} := \lim_{x_{n} \to \pm\infty} u$.
If each $u^+$ and $u^-$ is either a increasing 1D solution or is identically $\pm 1$, then $u$ is a minimizer in $\mathbb{R}^n$.
\end{prop}
The result for $s=2$ was proven in \cite[Theorem 1.1]{FariVal}. Here we show that the same type of argument works also for $s\in (0,2)$.
\begin{proof}[Proof of Proposition \ref{implication}]
Let $\Omega \subset \mathbb{R}^n$ be a bounded domain and let $v$ be a minimizer of $\mathcal E_\Omega$ with exterior datum $u$ outside of $\Omega$.
Let us first show that $u^-\le v\le u^+$ in $\mathbb{R}^n$. Indeed, in the complement of $\Omega$ the inequality holds since $v=u$ there and $u^-\le u\le u^+$ in $\mathbb{R}^n$.
To show that $u^-\le v\le u^+$ in $\Omega$, let us consider $\overline w^\pm = \max(v, u^\pm)$ and $\underline w^\pm = \min(v, u^\pm)$.
Assume by contradiction that $v>u^+$ at some point in $\Omega$. Then, arguing similarly as in \eqref{claimA-ii}, we would have
\begin{equation}\label{wniogoihb}
\mathcal E_\Omega(\overline w^+)+\mathcal E_\Omega(\underline w^+)<\mathcal E_\Omega(v)+\mathcal E_\Omega(u^+).
\end{equation}
But now since on the one hand $v$ is a minimizer with exterior datum equal to $u$ in the complement of $\Omega$ we have $\mathcal E_\Omega(v) \le \mathcal E_\Omega(\underline w^+)$. On the other hand, since $u^+$ is either $+1$ or $1D$ and increasing, it follows\footnote{A simple way of proving this consists of using the standard argument which involves the foliation of $\mathbb{R}^{n}\times(-1,1)$ given by the horizontal translations of the graph of $u^+$; see \cite[Proof of Proposition~6.2]{CC2}.} that $u^+$ is also a minimizer. Hence, noticing that $\overline w^+$ and $u^+$ coincide outside of $\Omega$, we obtain $\mathcal E_\Omega( u^+)\le \mathcal E_\Omega(\overline w^+)$. We therefore reach a contradiction with \eqref{wniogoihb}.
A similar argument using $\overline w^-$ and $\underline w^-$ shows $u^-\le v$.
Finally, using the standard ``foliation'' $\big\{u(x', x_n +t), \ t\in \mathbb{R}\big\}$, unless $v\equiv u$ we may find a translation of the graph of $u$ touching by above (or by below) the graph of $v$ as some point interior point in $\Omega$, and this contradicts the strong maximum principle. Hence the only possibility is that $v\equiv u$ and thus $u$ is a minimizer in $\Omega$. Since $\Omega$ is an arbitrary bounded domain, we conclude that $u$ is a minimizer in~$\mathbb{R}^n$.
\end{proof}
We can finally give the
\begin{proof} [Proof of Corollary \ref{corclas3}]
By Corollary \ref{corclas2} the limits $u^{\pm} := \lim_{x_{4} \to \pm\infty} u$ (which are stable solutions in $\mathbb{R}^3$) must be either $\pm 1$ or increasing $1D$ solutions.
Thus, by Proposition \ref{implication}, $u$ is a minimizer in $\mathbb{R}^4$. Now, since $s$ is sufficiently close to 1, the corollary follows from Theorem 1.5 in \cite{dPSV}.
\end{proof}
|
\section{Introduction}
\label{introduction}
The experimental realization of two-dimensional fluids of hard polygonal particles to study their phase behaviors
is an active line of research. For example, lithographic techniques applied to prepare nonoverlapping
particles of a specific polygonal shape, and their adsorption or confinement, produce
single monolayers of Brownian particles that diffuse in two dimensions \cite{Chaikin,Mason,Mason2}. A vast body of experimental studies
on these systems have clarified the importance of entropic particle interactions to stabilize different liquid-crystal
and solid phases with exotic symmetries beyond the standard uniaxial nematic (N) symmetry.
In particular, tetratic (T) and triatic (TR) liquid-crystal phases
were found when the particles have cross sections with rectangular \cite{Chaikin}, or triangular \cite{Mason}
geometry. Experiments with particles of square sections did not find the T phase \cite{Mason2}, but this
was due to the roundness of the corners, as proved by Monte Carlo (MC) simulations \cite{Escobedo}.
The stability of T \cite{Schlacken,Frenkel,MR,Donev,Selinger,Sabi} and TR \cite{Dijkstra,MR2} phases as a function of
particle shape was also confirmed by theoretical and simulation works on two-dimensional hard particle fluids.
Experiments on monolayers of vibrated granular cylinders \cite{Menon1,Dani,GP1,GP2}
and squares \cite{Menon2} also showed the presence of stationary nonequilibrium T-like textures in the
arrangement of particles. These results suggested that entropic interactions are also important to determine
the orientational ordering patterns observed in dissipative systems. Experiments
conducted on vibrated granular rods \cite{Ariel} and equilibrated colloidal silica rods \cite{Arts}
under annular confinement showed the presence of topological defects and domains walls between regions
of different orientational and spatial ordering. This in turn suggests similarities
between dissipative and equilibrium systems in situations where entropic interactions play a dominant
role, i.e. at high packing fractions. Recently the T phase of kite-like particles was
also found \cite{Mason3,MR3}. How regular polygons order in liquid-crystal and crystalline phases
as density is varied strongly depends on the number of polygonal sides, an issue that was intensively
studied by MC simulations \cite{Glotzer}.
Recently Gantapara et al. \cite{Dijkstra} have conducted MC simulations on hard particles consisting of
equilateral and right isosceles
triangles. For the latter,
the authors found the presence of a seemingly exotic liquid-crystal phase,
which they called Rhombic (R), between the isotropic (I) and its crystalline counterpart (the rhombic crystal).
This phase seems to exhibit strong octatic (O) correlations, with a high value for corresponding order parameter, $Q_8$.
In an O phase the orientation distribution function, i.e. the probability density of a particle to
orient with respect to one of the equivalent directors at an angle $\phi$, has eightfold
symmetry, $h(\phi)=h(\phi+\pi/4)$, in contrast with the T phase where the symmetry is $h(\phi)=h(\phi+\pi/2)$.
Fig. \ref{fig0m} shows a schematic of the I, O, T
and N liquid-crystal phases of hard right triangles. One can anticipate the importance of
particle clustering in this system, with particles easily forming square-like dimers or tetramers
that build up O or T ordering, corresponding to eight or four equivalent directors, respectively.
Note that the axis of a right triangle is defined by a unit vector connecting the
barycenter and the right-angled vertex.
The phase discovered by Gantapara et al. is probably the standard T phase since,
as shown below, there are good reasons to expect that
the phase cannot be purely octatic. In any case, the main focus of the
authors of Ref. \cite{Dijkstra} was on the study of chirality in the crystal phase of hard equilateral and
right-angled triangles.
\begin{figure*}
\epsfig{file=fig1.eps,width=6.in}
\caption{\label{fig0m}Sketches of different liquid-crystal phases of hard right
triangles: I, O, T and N, as labelled in the figure. The equivalent directors along which
particle axes of triangles (defined in the top) are aligned are also indicated.}
\end{figure*}
We have recently applied a Density-Functional Theory (DFT) based on Scaled-Particle Theory (SPT)
to study the phase behavior of a fluid of hard isosceles triangles as a function of their opening angle \cite{MR2}.
This theory is close in spirit and qualitatively similar in results to the standard Onsager theory \cite{Onsager} and to
extensions of Parsons-Lee-type theories \cite{Lee} applied to two dimensional particles.
For the particular case of right-angled triangles, we found that the only stable liquid-crystal phase predicted by
the theory was the standard uniaxial N phase, with the symmetry $h(\phi)=h(\phi+\pi)$.
This is in stark disagreement with the simulations of Gantapara et al. and points to a essential
problem of the above theory and, by extension, of the (second-order virial) Onsager theory.
We remind the reader that, by contrast, the SPT-based theory correctly predicts the existence of the exotic T phase in
fluids of hard rectangles with low aspect ratio and of the TR phase for a fluid of hard equilateral triangles \cite{Schlacken,MR,MR2}.
The failure of DFT to correctly describe, at least qualitatively, the thermodynamically
stable symmetry of an oriented fluid is unusual. It means that the important
angular correlations that drive the system to the stable symmetry are not present at the level of
two-particle correlations but in higher-order terms. This failure brings to mind the inability of DFT to reproduce
the intermediate hexatic phase in systems of hard discs, although the origin of the problem is
completely different.
In this case even the most sophisticated, albeit still approximate,
versions of DFT applied to the hard disk fluid
predict a first-order transition between the fluid and a crystal with particles at the nodes of
a triangular lattice \cite{Roth}. However, simulations show an intermediate hexatic phase between
fluid and crystal exhibiting quasi-long-range bond-orientational ordering \cite{Strandburg,Mak,Engel},
associated to two-particle positional correlations. One can speculate on the possibility to construct
an accurate hard-disk density functional based on two-body density functions, as proposed for example
in Ref. \cite{Iyetomi,Schiper}: coupling between first nearest-neighbor bonds and spatial coordinates
would be contained explicitly in the theory by construction, and phases with bond orientational ordering
could in principle be stabilised.
In the present case, however, the order is orientational and characterised by a
one-particle distribution function, namely the orientational distribution function $h(\phi)$, not
by a bond-orientational order. The origin of the failure is therefore different. Retrospectively,
it is remarkable how Onsager theory, based solely on two-particle correlations, i.e. on the lowest-order
virial expansion of the free energy, can describe oriented phases of fluids and fluid phase transitions.
Obviously this is because, for fairly elongated particles, the second-order term is by far the
dominant one in the virial expansion. It is therefore even more remarkable that, in the last few
decades, Onsager theory and its extensions (e.g. SPT), all based on two-particle orientational correlations,
have been applied with success even in situations where the latter condition is clearly not validated,
for example in 2D and/or in the case of anisotropic but not very elongated particles.
It is plausible that the limits of validity of the theory may be revealed in some cases,
and the fluid of hard right triangles is, to our knowledge, the first example of this breakdown.
Motivated by this negative finding, in the present article we report on the results
obtained using an extended SPT theory that goes beyond two-body correlations by incorporating
three-body orientational correlations through the third virial coefficient. This is the obvious
step to take in an attempt to remedy the deficiencies of the standard theory.
From the lessons learned in the hard-rectangle case \cite{MR1}, it is known that higher-order correlations
are important to improve the predictions on the stability of high-symmetry orientational
phases, since they more correctly account for particle configurations typical of these phases. However, the
numerical implementation of such a theory is not easy: the third virial coefficient is
in fact a functional of the orientational distribution function, $h(\phi)$, with respect to which
the free energy has to be functionally minimised. In order to make calculations
feasible, we consider a projection of $h(\phi)$ on Fourier space, and represent the third-virial
functional by means of particular low-order moments that dominate the behavior in the neighbourhood of the
bifurcation point (from the isotropic to the oriented phase). A combination of MC integration and
Gaussian quadratures is used to evaluate these moments with reasonable accuracy. Using these techniques,
a bifurcation analysis and the full minimization of the resulting DFT were implemented, which produced
valid results at least for packing fractions close to the bifurcation point.
Interestingly, the results from the extended theory
are qualitatively similar to the predictions of the standard SPT-based theory:
the only orientationally-ordered stable phase is the uniaxial N phase, and the structure of the orientational
distribution function has no indication whatsoever of T or O correlations. The obvious conclusion is that
third-order correlations are still not the important ones in the stability of these phases. Due to the
difficulties associated with the implementation of even higher-order terms, the virial approach
suggested by Onsager theory in this system seems hopeless in practical terms.
To understand the problem in more depth, especially in connection to the stability of high-symmetry
orientational phases, we have also performed $NVT$-MC simulations of a fluid of
hard right-angled triangles. In line with the findings of Gantapara et al. \cite{Dijkstra}, we found that,
under compression from the isotropic fluid, the system becomes ordered in configurations where
particles exhibit strong O correlations. Under expansion from a perfect crystal
formed by tetramers of triangles arranged in squares, Fig. \ref{clust}(a) (this is the stable crystal phase
according to \cite{Dijkstra}),
we found a (possibly discontinuous) melting transition to a T phase (with fourfold symmetry and a high number
of tetramers). This T phase is stable (or metastable) for approximately the same range of packing fractions
as that of the O phase obtained on compression,
and under further expansion transforms into an isotropic phase via a discontinuous
phase transition). All of these symmetries are confirmed from the behaviour of suitable order parameters that
describe the T and O symmetries, and are visualised very directly by looking at the orientational distribution
function $h(\phi)$.
Although we have not traced out the stability boundaries of the I, O and T phases,
these results definitively answer the question as to the inadequacy of a third-virial DFT to describe the
orientational ordering symmetries of a fluid of right-angled triangles. Obviously the inclusion of
even more high-order virial coefficients could improve the description of the fluid
phase behavior. However, as mentioned above,
the numerical implementation of such a theory would constitute a huge numerical
task, much more demanding than standard MC simulations on reasonably sized systems. Clearly a radically
new approach is needed. One possibility is the formulation of new models based on the
self-assembling of particles into clusters or super-particles of different shapes and sizes;
these clusters would in turn be oriented in such a way that the final orientation of monomers will
exhibit the new exotic symmetries. As recently shown \cite{Glotzer2}, the entropic hard-particle
interactions enjoy some similarities with chemical-bonding interactions, which eventually give rise to the
formation of clusters of particles or supra-molecular aggregates, respectively. This view could be
fruitful in the present system and might represent a worthy activity for the future.
The article is organized as follows. Sec. \ref{B3} is devoted to introducing the theoretical model for
a fluid of hard right-angled triangles, a model based on the extended third-virial DFT.
In Secs. \ref{bifurca} and \ref{minimin} we implement a bifurcation analysis and the full minimization of the model,
which is valid for densities close to the bifurcation point. Also, in Sec. \ref{MonteCarlo}
we present MC simulations which confirm
the presence of orientational symmetries different from uniaxial.
Finally some conclusions are drawn in Sec. \ref{conclusions}.
\section{Third-virial DFT}
\label{B3}
We have already pointed out that the SPT-based DFT (an effective second-order virial theory), fails to
predict the stability of the T and/or O liquid-crystal phases. These phases are characterised by
an orientational distribution function $h(\phi)$ with four- and eightfold symmetries, respectively:
$h(\phi)=h(\phi+n\pi/4)$ ($n=1,2$). By contrast, MC simulations \cite{Dijkstra} clearly point to the existence
of both these symmetries in the region between the isotropic and crystal phases, even though the nature of the
stable phase and the relative stability of the two phases still demand clarification (see below).
It is well known that the third and higher-order virial coefficients, $B_n$ with $n>2$, cannot be neglected
if we are to obtain a quantitatively correct description of the phase behavior of hard elongated particles
in two dimensions \cite{Talbot,Rigby}. This can be explained in terms of the nonvanishing limit of the ratio $B_3/B_2^2$
as particle elongation goes to infinity (the so-called Onsager limit). Although the triangular geometry
studied here cannot be defined in such terms, we expect the effect of the third virial coefficient to
be sufficiently important as to merit its inclusion in a DFT approach. A possible theory was already
proposed and applied by us in a study of the hard-rectangle-fluid \cite{MR1}, where the effect of the
third virial coefficient was found to be important (although not crucial to determine phase symmetry).
The essential idea is to approximate the excess free-energy density functional per particle
(in reduced thermal units) as
\begin{eqnarray}
&&\varphi_{\rm exc}[h]=-\log(1-\eta)+\frac{\eta}{1-\eta}b_2[h]\nonumber\\
&&+\left(\frac{\eta}{1-\eta}+\log(1-\eta)\right)\left(b_3[h]-2b_2[h]\right),
\label{Eq1}
\end{eqnarray}
where $\eta=\rho a$ is the packing fraction, defined as the product of mean number density,
$\rho$, and particle area of the right isosceles triangle, $a=l^2/2$, with $l$ the
length of the equally-sized triangle sides. In Eqn. (\ref{Eq1}) the coefficients $b_k[h]$ are defined as
\begin{eqnarray}
b_k[h]=\frac{B_k[h]}{a^{k-1}}-1,
\label{lab}
\end{eqnarray}
where $B_k[h]$, $k=2$ and 3, are the second and third virial coefficients, manifestly
functionals of the orientational distribution function $h(\phi)$:
\begin{eqnarray}
&&B_k[h]=\frac{1}{k}\left(\prod_{i=1}^k\int_0^{2\pi}d\phi_ih(\phi_i)\right) {\cal K}^{(k)}
\left(\boldsymbol{\phi}\right), \nonumber\\
&&\quad \boldsymbol{\phi}=\left(\phi_1,\cdots,\phi_k\right).
\label{Eq5}
\end{eqnarray}
In turn, the kernels ${\cal K}^{(k)}(\boldsymbol{\phi})$ are spatial integrals of
products of Mayer functions $f({\bm r}_{ij},\phi_{ij})$ associated to particles $i$
and $j$ with relative positions and orientations ${\bm r}_{ij}={\bm r}_j-{\bm r}_i$
and $\phi_{ij}=\phi_j-\phi_i$, respectively:
\begin{eqnarray}
&&{\cal K}^{(2)}(\boldsymbol{\phi})=
-\frac{1}{A}\left(\prod_{i=1}^2 \int_A d{\bm r}_i\right)
f({\bm r}_{12},\phi_{12})\nonumber\\
&&=-\int_A d{\bm r} f({\bm r},\phi)
=A_{\rm excl}(\phi), \label{Eq2}\\
&&{\cal K}^{(3)}(\boldsymbol{\phi})=-\frac{1}{A}\left(\prod_{i=1}^3
\int_A d{\bm r}_i\right)
f({\bm r}_{12},\phi_{12})f({\bm r}_{23},\phi_{23})\nonumber\\
&&\times f({\bm r}_{13},\phi_{13})\nonumber\\
&&=-\int_A d{\bm r} \int_A d{\bm r}' f({\bm r},\phi)f({\bm r}',\phi')f({\bm r}'-{\bm r},\phi'-\phi),
\label{Eq3}
\end{eqnarray}
with $A$ the total area. In the above expressions we have implemented the change of variables
${\bm r}\equiv {\bm r}_{12}$ and $\phi\equiv \phi_{12}$ (first and second integrals), and
${\bm r}'\equiv {\bm r}_{13}$ and $\phi'\equiv \phi_{13}$ (second integral).
The object $A_{\rm excl}(\phi_{12})$ is the excluded area between two particles.
Using these expressions, Eqn. (\ref{Eq5}) for $k=2$ and 3 becomes
\begin{eqnarray}
&&B_2[h]=\frac{1}{2}\int_0^{2\pi} d\phi \Psi_2(\phi)
{\cal K}^{(2)}(\phi), \\
&&\Psi_2(\phi)=\int_0^{2\pi}d\phi_1 h(\phi_1)h(\phi_1+\phi),
\label{sub2}\\
&&B_3[h]=\frac{1}{3}\int_0^{2\pi} d\phi\int_0^{2\pi}
d\phi'\Psi_3(\phi,\phi'){\cal K}^{(3)}(\phi,\phi'),\\
&&\Psi_3(\phi,\phi')=\int_0^{2\pi}d\phi_1 h(\phi_1)
h(\phi_1+\phi)h(\phi_1+\phi') \label{sub3}
\end{eqnarray}
The density expansion of Eqn. (\ref{Eq1}) gives
$\displaystyle{\varphi_{\rm exc}[h]=B_2[h]\rho+\frac{1}{2}B_3[h]\rho^2}+\cdots$,
and consequently the pressure is $\beta p=\rho+\rho^2\partial \varphi_{\rm exc}/\partial\rho
=\rho+B_2[h]\rho^2+B_3[h]\rho^3+\cdots$ (with $\beta$ the Boltzmann factor). The truncated
expressions provide the exact low-density limit up to third order in density.
Thus our proposed theory treats two- and three-body correlations exactly.
The ideal free-energy density functional per particle is, as usual,
\begin{eqnarray}
\varphi_{\rm id}[h]=\log \eta-1+\int_0^{2\pi} d\phi h(\phi)\log{\left[2\pi h(\phi)\right]},
\label{id}
\end{eqnarray}
where the thermal volume term has been dropped. Therefore our theory is completely defined as
$\varphi[h]=\varphi_{\rm id}[h]+\varphi_{\rm ex}[h]$.
\section{Bifurcation analysis}
\label{bifurca}
In this section we perform a bifurcation analysis of the theory presented above.
The bifurcation defines the instability of the I phase against
orientational fluctuations of some particular symmetry. We begin by considering
the first-order Fourier expansion of the $h(\phi)$ function:
\begin{eqnarray}
h(\phi)\simeq\frac{1}{2\pi}\left(1+h_n \cos{2n\phi}\right),
\end{eqnarray}
with $h_n$ the first-order Fourier amplitudes. The indexes
$n=1,2,3,4$ account for uniaxial N, tetratic T, triatic TR, and octatic O symmetries,
respectively. Substituting this expression into Eqns. (\ref{sub2}) and (\ref{sub3}) gives, to
lowest order:
\begin{eqnarray}
&&\Psi_2(\phi)\simeq \frac{1}{2\pi}\left(1+\frac{h_n^2}{2}\cos{2n\phi}\right), \label{psi2}\\
&&\Psi_3(\phi,\phi')\simeq\frac{1}{(2\pi)^2}
\left\{1+\frac{h_n^2}{2}
\left[\cos{2n\phi}+\cos{2n\phi'}\right.\right.\nonumber\\
&&\left.\left.+\cos{2n(\phi-\phi')}\right]\right\}.\label{psi3}
\end{eqnarray}
Consequently,
\begin{eqnarray}
&&B_2[h]\simeq \frac{1}{2}\left({\cal K}_0^{(2)}+\frac{1}{2}{\cal K}_n^{(2)}h_n^2
\right),\label{insert1}\\
&&B_3[h]\simeq \frac{1}{3}\left[{\cal K}_{00}^{(3)}+
\left({\cal K}_{n0}^{(3)}+\frac{1}{2}{\cal K}_{nn}^{(3)}\right)h_n^2\right],
\label{insert2}
\end{eqnarray}
where the following coefficients have been defined:
\begin{eqnarray}
&&{\cal K}_n^{(2)}=\frac{1}{2\pi}\int_0^{2\pi}d\phi {\cal K}^{(2)}(\phi)\cos(2n\phi),\nonumber\\
&&{\cal K}_{nm}^{(3)}=\frac{1}{(2\pi)^2}\int_0^{2\pi}d\phi\int_0^{2\pi}d\phi'
{\cal K}^{(3)}(\phi,\phi')\nonumber\\
&&\times\cos{2(n\phi-m\phi')}. \label{Eq6}
\end{eqnarray}
The Fourier expansion of the ideal free energy Eqn. (\ref{id}) gives
$\varphi_{\rm id}[h]\simeq \log{\eta}-1+h_n^2/4$.
Inserting Eqns. (\ref{insert1}) and (\ref{insert2}) into
(\ref{Eq1}), we obtain the expansion of the total free-energy
per particle as $\varphi=\varphi_{\rm I}+\Delta\varphi$, with
\begin{eqnarray}
&&\varphi_{\rm I}=\log\left(\frac{\eta}{1-\eta}\right)-1+
\left(\frac{{\cal K}_{00}^{(3)}}{3a^2}-1\right)\chi_1(\eta)\nonumber\\
&&-\left(\frac{{\cal K}_0^{(2)}}{2a}-1\right)\chi_2(\eta),\label{la_I}\\
&&\Delta\varphi=\left[\frac{1}{4}+
\frac{1}{3a^2}\left({\cal K}_{n0}^{(3)}+\frac{1}{2}{\cal K}_{nn}^{(3)}
\label{brakets0}
\right)\chi_1(\eta)\right.\nonumber\\
&&\left.-\frac{{\cal K}_n^{(2)}}{4a}\chi_2(\eta)\right]h_n^2,
\label{la_Delta}\\
&&\chi_k(\eta)=\frac{\eta}{1-\eta}+k\log{(1-\eta)},\label{brakets}
\end{eqnarray}
with $\varphi_{\rm I}$ the I phase contribution. $\Delta\varphi$ is different from zero in the
presence of a weak orientational ordering.
All that remains is to calculate the coefficients defined in Eqn. (\ref{Eq6}). As shown in \cite{MR2},
the expression for ${\cal K}_n^{(2)}$, for the particular case of right-angled isosceles triangles, is
\begin{eqnarray}
&&\frac{{\cal K}_n^{(2)}}{2a}-\delta_{n0}=-\frac{1}{\pi(4n^2-1)}\nonumber\\
&&\times \left[2+(-1)^n+
2\sqrt{2}\cos\left(\frac{n\pi}{2}\right)\right],
\end{eqnarray}
with $\delta_{ij}$ the Kronecker-delta.
The coefficients ${\cal K}_{nm}^{(3)}$ cannot be calculated analytically and have been obtained numerically. MC integration was used to evaluate
the spatial kernels in (\ref{Eq3}), while the angular integrals in Eqn. (\ref{Eq6}) were evaluated
using Gaussian-Legendre quadratures.
$2\times 10^6$ random configurations of three particles with fixed
orientations were generated, using the usual protocol: one particle is located at the origin, two
particles are generated at random, both overlapping with the first one,
thus ensuring that the first two Mayer functions are equal to $-1$, and the third Mayer function
connecting the latter is checked. In the case of the angular integrals, a density of
12 points per period were used for the four cases $n=1,\cdots,4$. The estimated error
of the coefficients is less than $0.1\%$.
The minimization of the expanded total free-energy
per particle $\varphi[h]$ with respect to $h_n$ gives the bifurcation condition
$\partial \Delta\varphi/\partial h_n=0$ for $h_n\neq 0$, with $\Delta\varphi$ given by (\ref{brakets0}).
The solutions to this equation are obviously the same as the zeros of the
term enclosed in brackets.
We have solved numerically this equation for the values of packing fractions
$\eta_n$ at bifurcation, considering separately the cases $n=1,2,3$ and $4$.
The results are collected in Table \ref{tabla}. Also included are the results
obtained from the SPT approach, which can be obtained easily by considering the
first two terms of Eqn. (\ref{Eq1}) for the excess free-energy per particle.
In this case the bifurcation points are given analytically by
\begin{eqnarray}
\eta_n^{(\rm SPT)}=\frac{1}{1-{\cal K}_n^{(2)}/a}.
\label{spin}
\end{eqnarray}
\begin{table}
\begin{tabular}{||c|c|c|c|c||}
\hline\hline
$n$ & 1 & 2 & 3 & 4\\ \hline
\ $\eta_n^{(\rm SPT)}$ \ & \ 0.8249 \ & \ $0.9928^*$ \ & \ 0.9821 \ & \ 0.9444 \ \\
\hline
\ $\eta_n^{(\rm B_3)}$ \ & \ 0.7325 \ & \ $0.9794^*$ \ & 0.9328 \ & \ 0.8353 \ \\ \hline
\ $\Delta \eta^{(\rm SPT)}_n$ \ & \ -- \
& \ 0.1679 \ & \ 0.1572 \ & \ 0.1195 \ \\ \hline
\ $\Delta \eta^{(\rm B_3)}_n$ \ & \ -- \ &\ 0.2469 \ &\ 0.2003 \ & \ 0.1062 \ \\
\hline\hline
\end{tabular}
\caption{The bifurcated values of packing fractions $\eta_n^{(\alpha)}$ (with $\alpha=$ SPT or
$B_3$) resulting from SPT and $B_3$ approaches corresponding to bifurcations from I
to N ($n=1)$, T ($n=2$), TR ($n=3$) two numbers means that
O ($n=4$) phases. The differences
$\Delta \eta_n^{(\alpha)}\equiv \eta_n^{(\alpha)}-\eta_1^{(\alpha)}$
for $n\geq 2$ are also shown. An asterisk
indicates that the value does not correspond to a true bifurcation at a
second-order transition; this is because a first-order I-T metastable transition
is found using the SPT theory and a parameterized $h(\phi)$ containing only
one parameter. The same behavior is expected for the $B_3$ approach.}
\label{tabla}
\end{table}
A first conclusion that can be drawn from the results of the table
is that, compared to the second-virial approximation, the third-virial approximation
dramatically decreases the bifurcation packing fractions from the I phase to the orientationaly ordered
phases. In general the third-virial approximation gives a value for the I-N bifurcation,
$\eta_1^{(\rm B_3)}\approx 0.73$, which is quantitatively similar
to that for the I phase coexisting with
the liquid-crystal phase, as predicted by simulations \cite{Dijkstra}.
Note again that simulations predict a very different symmetry (four- or eight- instead of two-fold)
for the latter phase: both, second- and third-virial approaches, predict a stable uniaxial N phase
beyond the I phase.
The failure of DFT to predict the correct symmetry for the liquid-crystal phase could have one or
two of the following explanations: (i) the importance of four and higher-order
particle correlations, quantified through the corresponding virial coefficients,
to correctly predict the stability of the T/O phase
(note that, strictly speaking, the theory generates these virial coefficients from a
density expansion; however, the angular correlations will be poorly represented);
(ii) the presence of strong clustering effects,
which a low-order virial theory would be unable to account for.
In any case, the sequence
$\eta_1^{(\alpha)}<\eta_4^{(\alpha)}<\eta_3^{(\alpha)}$ (see Table \ref{tabla}),
together with the fact that from a value $\eta_2^*\approx 0.955$ a
(highly ordered) metastable T solution is found
(see discussion in Section \ref{minimin}), and taking into account the
corresponding differences
$\Delta \eta_n^{(\alpha)}\equiv \eta_n^{(\alpha)}-\eta_1^{(\alpha)}$ and
$\Delta\eta_2^*=\eta_2^*-\eta_1^{(\rm spt)}\approx 0.13$,
we can conclude that, within the context of DFT theory, the TR, T and O
symmetries
can be discarded as possible
liquid-crystal symmetries for hard right isosceles triangles.
\section{Minimization}
\label{minimin}
A next step beyond the bifurcation analysis consists of analysing the equilibrium solutions
of the extended density functional by minimization.
An important motivation for this calculation is to examine the orientational
distributions and search for the possible existence of secondary peaks in
$h(\phi)$ at $\phi=\{\pi/4,\pi/2,3\pi/4\}$ in the interval $\phi\in[0,\pi]$
(which would point to four- or eightfold orientational symmetry).
Then, we are probing the ability of three-body correlations to capture this symmetry.
First we introduce a parameterization
for the orientational distribution function, which we simply take as a truncated Fourier expansion:
\begin{eqnarray}
h(\phi)=\frac{1}{2\pi}\left(1+\sum_{k=1}^{n_{\rm max}} h_k\cos(2k\phi)\right),
\label{expansion}
\end{eqnarray}
We choose $n_{\rm max}=5$, which will be valid sufficient close to the
bifurcation point.
Substituting (\ref{expansion}) into Eqns. (\ref{psi2}) and (\ref{psi3}), we obtain
\begin{eqnarray}
&&\Psi_2(\phi)=\frac{1}{2\pi}\left(1+\frac{1}{2}\sum_{k=1}^{n_{\rm max}}h_k^2\cos(2k\phi)\right),\\
&&\Psi_3(\phi,\phi')=\frac{1}{(2\pi)^2}\left\{1+\frac{1}{4}
\sum_{(k_1,k_2)\neq(0,0)}^{\text{max}(k_1+k_2)=
n_{\rm max}} h_{k_1}h_{k_2}
\nonumber\right.\\
&&\left. \times h_{k_1+k_2}\left[\cos(2(k_1\phi+k_2\phi'))\right.\right.\nonumber\\
&&\left.\left.+\cos(2(k_2\phi-(k_1+k_2)\phi'))
\right.\right.\nonumber\\
&&\left.\left.+\cos(2(k_2\phi'-(k_1+k_2)\phi))\right]\right\}.
\end{eqnarray}
After substitution into Eqns. (\ref{Eq1}) and (\ref{id}), the total free energy per particle
$\varphi=\varphi_{\rm I} +\Delta \varphi$ is obtained, with $\varphi_{\rm I}$ calculated from (\ref{la_I}).
The excess part over the isotropic contribution, $\Delta \varphi$, is now
\begin{eqnarray}
&&\Delta \varphi=\int_0^{2\pi} d\phi h(\phi) \log {\left[2\pi h(\phi)\right]}+
\frac{1}{6a^2}\sum_{(n,m)\neq (0,0)}^{n+m\le n_{\rm max}}
h_{n}h_{m}\nonumber\\
&&\times h_{n+m} \left[
\frac{{\cal K}^{(3)}_{-n,m}}{2}+{\cal K}^{(3)}_{n,n+m}\right]\chi_1(\eta)\nonumber\\
&&-\frac{1}{4a}\left(\sum_{n\neq 0}^{n_{\rm max}} h_n^2{\cal K}_n^{(2)}\right)\chi_2(\eta).
\label{double}
\end{eqnarray}
It is understood that $h_{n}=1$ for $n=0$ in the double sum of Eqn. (\ref{double}).
As in Section \ref{bifurca} the coefficients ${\cal K}^{(3)}_{nm}$ are evaluated numerically
using MC integration and Gaussian quadrature (note that in this case a total of 71 coefficients
need to be evaluated instead of only 9 in the bifurcation analysis).
The total free energy is then minimized with respect to the Fourier amplitudes
$\{h_k\}$ ($k=1,\cdots,n_{\rm max}$) ($n_{\rm max}=5$) using a Newton-Raphson procedure.
\begin{figure*}
\hspace*{0.8cm}
\includegraphics[width=7.in]{fig2.eps}
\caption{Distribution functions $h(\phi)$ of a N fluid of hard right triangles obtained from the
minimization of: (i) SPT theory (dashed curve), (ii) Extended SPT theory (solid curve).
Three cases of different densities are shown. (a) $\eta=0.8437$ (SPT) and $0.7500$ (Extended SPT).
(b) $\eta=0.8500$ (SPT) and $0.7547$ (Extended SPT).
(c) $\eta=0.8600$ (SPT) and $0.7636$ (Extended SPT). In each case
densities have been chosen with the same relative departure from the respective
bifurcation densities: $3.0\%$ in (a), $4.2\%$ in (b), and $5.5\%$ y (c). In panel (a)
SPT results are plotted for the cases with $n_{\rm max}=5$ and $n_{\rm max}=40$ Fourier
coefficients; both results cannot be distinguished at the scale of the graph. The extended SPT theory
has been calculated with $n_{\rm max}=5$ Fourier coefficients, as explained in the text.}
\label{fig0}
\end{figure*}
Fig. \ref{fig0} shows an example of a minimization of the $B_3$-extended SPT approach
for a packing fraction $\eta=0.75$;
this is not far from the I-N bifurcation point at $\eta_1=0.7325$.
The orientational distribution function $h(\phi)$ is represented, together with the result from
the second-virial SPT theory of Ref. \cite{MR2}. In the latter case $h(\phi)$ was calculated using
$n_{\rm max}=5$ and also $n_{\rm max}=40$, which practically gives the exact result considering
that the last Fourier coefficient, $h_{40}$, obtained from the minimization
has an absolute value of less than $10^{-7}$. The functions of panel (a) (dashed curve),
cannot be distinguished one form the other at the scale of the graph, their mean square
difference being $\epsilon\equiv \sqrt{\int_0^{2\pi}d\phi \left(h(\phi)-\tilde{h}(\phi)\right)^2}=
5.2\times 10^{-4}$. The packing fraction $\eta=0.8437$ was chosen
so as to give the same relative distance $(\eta-\eta_1)/\eta_1$ from the I-N bifurcation point
as the $B_3$-extended theory (solid curve). We can see that the two theories give a qualitatively similar
orientational structure, with the function from the extended theory being more weakly oscillatory because
of the smaller density. We also show the functions obtained from the SPT and $B_3$-extended theories
for higher packing fractions corresponding to the
same relative departures from the respective bifurcation densities.
What is important from these calculations is that neither theory predicts
any trace of secondary peaks and, consequently, of tetratic or octatic correlations, in contrast with the
simulation results.
One may wonder which of the following scenarios takes place at densities above
the I-O bifurcation: (i) The O and N free-energy branches cross each other at some density, or (ii)
the N branch continues to be the lowest one. In order to investigate this point, we minimised the
SPT functional restricting the set of Fourier coefficients $\{h_k\}$ to only those with index $k=4j$
(giving a distribution $h(\phi)$ with perfect O symmetry). A free-energy branch was generated
for a density interval starting at the I-O bifurcation point and up
to densities where the octatic order parameter is $Q_8\simeq 0.97$ (in this situation
the truncated Fourier series still gives correct results).
We also calculated the N branch up to densities for which $Q_2\simeq 0.97$.
As expected, we obtained a ``metastable'' O phase. However, the first scenario above can be
discarded, as the difference between the O and N free-energy branches is huge
(note that the latter
bifurcates at much lower densities); this conclusion results from a simple
extrapolation to the region
where the O phase is metastable. The situation is
even worse in the case of the metastable TR phase, as it bifurcates at higher
packing fraction
(see the table \ref{tabla}). All these results are shown
in Fig. \ref{fig_nueva}(a), where we plot the free-energy
differences
$\Delta\varphi\equiv \varphi_{\rm \alpha}-\varphi_{\rm I}$
($\alpha=$N,O,TR) between $\alpha$ and I phases
calculated from their respective
bifurcation points.
We remark that the T phase bifurcates
from the I phase but within a first-order transition scenario
(see comment in Table \ref{tabla}), unlike the N, O and TR phases.
This point is elucidated not via Fourier-amplitude minimization
(we were unable to obtain a metastable T solution with the proper
restrictions over $\{h_k\}$), but from a simple one-parameter
($\lambda$) minimization of a parameterized orientation distribution function,
\begin{equation}
h(\phi)=\dfrac{e^{\lambda \cos(2n\phi)}}{\pi I_0(\lambda)},\hspace{0.6cm}
\phi\in[0,\pi],
\end{equation}
with $I_0(x)$ the zeroth-order modified Bessel function of the first kind.
The results are shown in Fig. \ref{fig_nueva}(b):
from the bifurcation point (open circle in the figure),
located at a packing fraction $\eta=0.9928$
(see Table \ref{tabla}), an unstable T branch departs towards lower densities,
with a free energy higher than that of the I phase. This branch
terminates at $\eta\approx 0.955$ (open square in the figure), where the T phase
becomes metastable for the first time, with a high order parameter $Q_4$.
From this point a second T branch develops towards higher densities,
$\Delta\varphi$ eventually becoming negative at $\eta\approx 0.968$,
indicating that the T phase is more stable than the I phase.
This is the usual scenario for a first-order phase transition.
When $\eta$ further increases from this value, the T free-energy branch
also crosses the O branch at $\eta\approx 0.988$.
In any case, the free energy of the N phase has by far the lowest free energy,
as can be seen in Fig. \ref{fig_nueva}. The crossing of the T and O branches
at high packing fractions is interesting. It can be understood by invoking
the limit value of the scaled second-virial coefficient (\ref{lab}) as
$\lambda\to\infty$:
\begin{eqnarray}
&&b_2^{(n)}\equiv \lim_{\lambda\to\infty} b_2[h]\nonumber\\
&&=\frac{1}{2an}\left[\frac{A_{\rm excl}(0)+A_{\rm excl}(\pi)}{2}
+\sum_{k=1}^{n-1}A_{\rm excl}\left(\frac{k\pi}{n}\right)\right]-1.
\nonumber\\
\end{eqnarray}
Inserting the known analytic expression for $A_{\rm excl}(\phi)$, we obtain
$b_2^{(1)}<b_2^{(2)}\lesssim b_2^{(4)}<b_2^{(3)}$ for the
N ($n=1$), T ($n=2$), TR ($n=3$) and O ($n=4$) symmetries. This fact explains
the reason for the crossing behavior: the double-averaged excluded area with
respect to $h(\phi)$ for the T symmetry, although similar in magnitude,
is lower than that obtained for the O symmetry. Obviously this occurs only
at very high densities, when the orientational order is almost perfect
and the above asymptotic expression can be justified.
The fact that the O-phase free-energy is lower than that of the T phase
at lower densities implies that the opposite behavior is true when the
orientational distribution function is less sharply peaked.
\begin{figure}
\epsfig{file=fig3a.eps,width=3.in}
\epsfig{file=fig3b.eps,width=3.in}
\caption{Free-energy differences $\Delta\varphi\equiv \varphi_{\alpha}
-\varphi_{I}$, with $\alpha$= N (dot-dashed), O (solid), T (dashed) and
TR (dotted), as a function of packing fraction $\eta$, using the
Fourier-coefficient (a) and one-parameter (b) minimizations.
In (a) we show with a tiny dot-dashed line a simple extrapolation
of the N branch to higher densities. The inset
in (b) show the complete N branch. The filled circles in (a) and (b) show
the bifurcation points at second order transitions while the open
one in (b) labels its first-order counterpart in the T branch. The
open square show the location of the first metastable solution with
T symmetry.}
\label{fig_nueva}
\end{figure}
Even though these calculations were
restricted to the SPT theory, we can be quite confident that the similar scenario results from
the $B_3$ theory (note that the latter theory is difficult to implement at high densities
due to the prohibitively large number of three-body kernels
projected on Fourier space that are necessary to describe a phase with high orientational order).
\begin{figure}
\epsfig{file=fig4.eps,width=3.5in}
\caption{\label{conf}
Possible particle configurations for crystal phases of right triangles.
(a) Square lattice of tetramers. Each tetramer is formed by four triangles. This is
the starting configuration of expansion run indicated by open circles in Fig. \ref{ops}(a).
(b) Square lattice of dimers with ordered orientations. This is the starting configuration
of expansion run shown in Fig. \ref{ops2}. (c) Square lattice of dimers with disordered
orientations. This is the starting configuration of expansion run shown in Fig. \ref{ops1}.}
\end{figure}
\section{Monte Carlo simulations}
\label{MonteCarlo}
Gantapara et al. \cite{Dijkstra} have presented a detailed Monte Carlo simulation
study of the fluid and crystal phases of hard right-angled triangles.
Their isothermal-isobaric simulations on systems of 1600 particles point to the existence of
a liquid-crystal phase between the isotropic fluid and the crystal. Based on the analysis
of order parameters and angular correlation functions, this phase was named
Rhombic (R) and characterised by the $Q_8$ order parameter. Order parameters are defined, in terms of the orientational
distribution function, as
\begin{eqnarray}
Q_k=\int_0^{2\pi}d\phi h(\phi)\cos{(k\phi)}.
\end{eqnarray}
These order parameters characterise the N ($k=2$), T ($k=4$), TR ($k=6$) and O ($k=8$) symmetries.
The term `rhombic' used in \cite{Dijkstra} reflects
the structure of the stable crystal phase, Fig. \ref{conf}(a), which consists of a square lattice of tetramers,
each formed by four triangles in a square configuration \cite{Dijkstra}. Unfortunately, the information
provided by Gantapara et al. does not allow us to ascertain the true symmetry of the liquid-crystal phase.
A pure octatic phase would be characterised by
$Q_8>0$ and $Q_2=Q_4=0$, and by an orientational distribution function $h(\phi)$
with eight equivalent (equal-height) peaks in the interval $[0,2\pi)$, fulfilling
the condition $h(\phi)=h(\phi+\pi/4)$. A focus on the $Q_8$ order parameter
in \cite{Dijkstra} may indicate the occurrence of octatic symmetry in the
intermediate phase between isotropic and crystal phases.
To further investigate
this issue in more detail, and to obtain a test bed for the DFT results, we performed
$NVT$ Monte Carlo simulations on samples of 576 triangles (some selected $NpT$ simulations were
also performed to investigate some specific issues). Although our sample sizes are smaller,
our results complement the work of Gantapara et al. in the sense that they focus on the complete
set of order parameters (rather than only on $Q_8$), and also on the orientational
distribution function.
\begin{figure*}
\epsfig{file=fig5.eps,width=6.5in}
\caption{\label{ops}
Order parameters $Q_k$ (with $k=4$ and $8$) as a function of packing fraction $\eta$, as
obtained from $NVT$ and $NpT$ Monte Carlo simulations on 576 hard right triangles.
Compression runs are represented by squares, while expansion runs are represented with
circles. Panel (a) shows the $Q_4$ order parameter, while (b) shows the $Q_8$
order parameter. In both panels the different runs, indicated in the key, are as follows:
filled circles, NVT expansion from an initial crystalline configuration of tetramers
at $\eta=0.8$ shown in Fig. \ref{conf}(a);
open circles, NVT expansion from an initial crystalline configuration
of tetramers at $\eta=0.98$; filled squares, NVT compression from an initial
crystalline configuration of tetramers at $\eta=0.82$; open squares, NVT compression
from an initial isotropic configuration at $\eta=0.54$; gray squares, NpT
compression from an initial isotropic configuration at low pressure.
The isotropic-liquid-crystal coexistence reported by \cite{Dijkstra} is shown
as a shaded region, bounded by vertical dashed lines corresponding to the
transition densities: a, coexistence density for isotropic
at $0.733$. b, coexistence density for liquid-crystal phase at $0.782$. Also, the vertical line
c is the density of the transition from the
liquid-crystal phase to the crystal phase of tetramers at $0.87$ obtained in \cite{Dijkstra}.
The order parameters $Q_2$ and $Q_6$
are not represented as their values are $<0.1$ for all densities and runs.
Nonvanishing values of $Q_8$ below $\eta=0.733$ in panel (b) might be a finite-size effect.
Regions of stability of the I, T and K phases, according to \cite{Dijkstra},
are indicated with the corresponding labels.
The inset in panel (a) is a zoom of the liquid-crystal--crystal phase transition.}
\end{figure*}
Several $NVT$ runs were performed, following compression or expansion starting from
configurations with different symmetries. All runs comprised $2\times 10^5$ MC steps for equilibration and
$3\times 10^5$ MC steps for averaging. $NpT$ runs were $10$ times longer.
Figs. \ref{ops} shows the order parameters $Q_k$, with $k=4$ and $8$, for some of the runs.
Results for compression and expansion runs are shown by squares and circles,
respectively. Two compression runs in the NVT ensemble were performed. The first (open squares)
started at a low density, $\eta=0.536$, in the isotropic fluid region, and continued up to a
density $\eta=0.812$. Along this run all order parameters are low except $Q_8$,
which shows a steady increase with density. This seems compatible with the results of Gantapara et al.
Since their system size is three
times larger, changes in the order parameter near the isotropic-liquid-crystal
transition are more abrupt in their case. At the end of our compression run values of the order
parameters are $Q_2=0.023\pm 0.003$, $Q_4=0.053\pm 0.004$, $Q_6=0.023\pm 0.006$ and
$Q_8=0.785\pm 0.016$, pointing to octatic symmetry.
The orientation distribution function for a density $\eta=0.802$ is plotted in Fig.
\ref{hphi}, using open squares. The peak heights are not exactly the same, which
may due to statistical fluctuations or imperfect orientational sampling.
To check the compression $NVT$ results, additional $NpT$ simulations were performed (grey symbols
in Fig. \ref{ops}). These results confirm the previous findings on the octatic-like phase
obtained by compressing the system from the isotropic phase.
\begin{figure}
\epsfig{file=fig6.eps,width=3.in,angle=-90}
\caption{\label{hphi}
Orientational distribution function $h(\phi)$ from simulation, for two cases: $\eta=0.802$,
obtained by compression from the isotropic fluid (open squares); and $\eta=0.8$,
obtained by equilibration of a perfectly ordered configuration of tetramers (filled squares).}
\end{figure}
However, the identification of the liquid-crystal phase as an octatic phase is challenged by the
results from our expansion runs. These are indicated in Fig. \ref{ops}
by means of circles. We focus first on the run starting at $\eta=0.8$ (filled circles)
in a (crystalline) configuration of tetramers, Fig. \ref{conf}(a). This density was chosen because
the free energy calculations reported in Ref. \cite{Dijkstra} led the authors to
conclude that the isotropic and liquid-crystal phase coexist along the
density gap $0.733$--$0.782$, indicated in Fig. \ref{ops} by a shaded region bounded
by vertical lines labelled `a' and `b', respectively.
A quasiperfect arrangement of tetramers would give $Q_4\simeq 1\agt Q_8$.
From this perspective, it is no surprise that the largest order parameter in this expansion run is
$Q_4$. The system finally collapses to the isotropic phase below $\eta\simeq 0.73$.
The orientational distribution function along this run (see Fig. \ref{hphi})
confirms that the equilibrium configuration corresponds to a T phase, with
four equivalent peaks in the interval $[0,2\pi)$, fulfilling the condition $h(\phi)=h(\phi+\pi/2)$.
Note that no signs of O symmetry, in the form of local maxima at $\pi/4$, $3\pi/4$, $5\pi/4$ and
$7\pi/4$, are visible. We conclude that the T phase is at least metastable for the
system size used, but our analysis cannot say anything definite as to the true stable phase in
the interval $0.733<\eta<0.782$, either O or T
(detailed free-energy calculations would be necessary to settle this question).
\begin{figure}
\epsfig{file=fig7.eps,width=3.5in}
\caption{\label{ops2}
Order parameters $Q_k$ (with $k=2, 4, 6$ and $8$) as a function of packing fraction $\eta$, as
obtained from $NVT$ Monte Carlo simulations on 576 hard right triangles. A single expansion run
from an initial configuration of ordered dimers
[shown in Fig. \ref{conf}(b)], at $\eta=0.92$ is shown.
Meanings of symbols are explained in the keybox. Vertical dashed lines correspond to the
transition densities reported in \cite{Dijkstra}: a, coexistence density for isotropic.
b, coexistence density for liquid-crystal phase. c: density of the transition from the
liquid-crystal phase to the crystal phase of tetramers. }
\end{figure}
Compression and expansion runs were also conducted to study the
transition between liquid-crystal and crystal phases. In the compression run (filled squares)
we started from the equilibrated point at $\eta=0.8$ of the previous
expansion run (filled circles). The expansion run (open circles) starts
from a perfect crystal of tetramers, Fig. \ref{conf}(a), at $\eta=0.98$ which, as shown
by the simulations of Gantapara et al. \cite{Dijkstra}, is the most stable crystal phase).
The path ends at a density of $\eta=0.86$. These runs indicate
the presence of a first-order phase transition between T and K (crystal) phases,
since the two branches are
not connected smoothly, see inset (a vertical line labelled `c' in Fig. \ref{ops} at density
$\eta=0.87$ indicates the density of the reportedly continuous transition \cite{Dijkstra}).
It is difficult to explain the existence of a strict octatic symmetry for the
liquid-crystal phase of hard right triangles on purely theoretical grounds.
The reason is that this symmetry requires a delicate balance of particle arrangements so as to give
equivalent orientations at multiples of $\pi/4$. We are inclined to think that the global
symmetry of the stable phase should be tetratic,
but the problem persists that simulations indicate two competing liquid-crystal phases,
one phase possessing strong octatic correlations and another phase without such correlations.
In any case, the results reported up to this point indicate that the liquid-crystal phase of hard right triangles is not the
standard uniaxial N phase, as predicted by DFT,
but rather a more symmetric phase with tetratic symmetry and possibly strong octatic correlations.
In order to investigate this result from the point of view of simulation, we performed additional simulations,
following two strategies. First, we conducted expansion runs from the crystal region, starting with particle
configurations different from a square lattice of tetramers. We stress that, according to the
free-energy calculations of Gantapara et al. \cite{Dijkstra},
the crystal of tetramers, Fig. \ref{conf}(a), is slightly more stable than
the crystal of dimers, Fig. \ref{conf}(b). The latter has uniaxial symmetry and can be considered as the crystalline `precursor'
of the uniaxial liquid-crystalline phase predicted by DFT. An obvious question is: could the uniaxial N
phase be at least metastable with respect to the other, more plausible symmetries? Fig. \ref{ops2} presents
an expansion run from a crystal phase made of dimers with identical orientations at density
$\eta=0.92$. The crystal transforms
continuously into a fluid phase with the same uniaxial symmetry (all order parameters are high and in the
order $Q_2>Q_4>Q_6>Q_8$). Along the simulations no indication of even weak tetratic or octatic correlations
could be detected: the orientational distribution function $h(\phi)$ (not shown) only exhibits
peaks at $0$ and $\pi$, with local maxima at other angles being completely absent even at
low densities. This is an indication that the uniaxial
N phase predicted by DFT may be metastable with respect to other more symmetric phases, at
least at the scale of our $10^6$ MC-step simulations.
\begin{figure}
\epsfig{file=fig8.eps,width=3.5in}
\caption{\label{ops1}
Order parameters $Q_k$ (with $k=2, 4, 6$ and $8$) as a function of packing fraction $\eta$, as
obtained from $NVT$ Monte Carlo simulations on 576 hard right triangles. A single expansion run
from an initial configuration of disordered dimers
[shown in Fig. \ref{conf}(c)] at $\eta=0.92$ is shown.
Meanings of symbols are explained in the keybox.
Vertical dashed lines correspond to the
transition densities reported in \cite{Dijkstra}: a, coexistence density for isotropic.
b, coexistence density for liquid-crystal phase. c: density of the transition from the
liquid-crystal phase to the crystal phase of tetramers. }
\end{figure}
A second run was performed on the same system but now from a configuration consisting
of a square lattice of dimers with random orientations, Fig. \ref{conf}(c). The results are shown
in Fig. \ref{ops1}.
This configuration was not considered by Gantapara et al. as a
possible candidate for equilibrium crystal phase, but clearly enjoys a higher entropy than the
uniaxial
crystal of dimers and could compete with the crystal of tetramers (note that the entropy difference per
particle between the crystal of tetramers, Fig. \ref{conf}(a), and the crystal of orientationally
ordered dimers, Fig. \ref{conf}(b), is only $0.011k$ per particle at $\eta=0.91$, the only density
investigated \cite{Dijkstra}; the extra entropy of a crystal of orientationally
disordered dimers, Fig. \ref{conf}(c), could easily overcome the free-energy balance with a
crystal of tetramers in some density range). The initially crystalline configuration
transforms into a fluid tetratic phase at $\eta\simeq 0.86$ (see Fig. \ref{ops1}).
An interesting point is that this
tetratic phase is slightly different from the one obtained by compression of the isotropic phase,
Fig. \ref{ops}.
Clearly a relatively small system of hard right triangles is very prone to developing
metastable phases with a structure sensitively dependent on the initial configuration
and the path followed by the simulation.
\begin{table}
\begin{tabular}{|c|c|c|c|}
\hline\hline
{\bf phase} & {\bf uniaxial} & {\bf tetratic} & {\bf octatic}\\
\hline
$\eta$ & $0.820\pm 0.004$ & $0.822\pm 0.004$ & $0.790\pm 0.003$\\
\hline\hline
\end{tabular}
\caption{\label{table2} Average densities of phases at pressure $\beta pa=13.5$
according to $NpT$ simulations. Each $10^6$ MC step-simulation starts from a
different configuration
(uniaxial, tetratic or octatic) obtained in different $NVT$ runs.}
\end{table}
As a final analysis, we performed $NpT$ simulations at a fixed value of pressure, $\beta pa=
13.5$ (which should be right at the liquid-crystal density interval \cite{Dijkstra}), starting from configurations
with uniaxial, tetratic, and octatic symmetries. The average densities obtained are given in Table \ref{table2}.
Note that $\eta_{\rm octatic}<\eta_{\rm uniaxial}<\eta_{\rm tetratic}$. As a general rule
one expects the stable phase to have the highest density at fixed pressure. Of course,
from a thermodynamic point of view, this is no proof that the T phase is more stable
than the others, but it certainly gives a hint. Interestingly the O phase is considerably less
compact than the T phase, with the uniaxial N phase quite close to the latter.
From the above discussion, we cannot strictly
discard the existence of a purely O phase between
the isotropic fluid and the crystal of a system of hard right-triangle particles. However,
in the region where the O phase could be stable, a competing phase with
clear tetratic symmetry also arises in simulations (starting from the crystal
of tetramers). The T phase is more
plausible than the O phase from the point of view of particle configurations.
Even though the $NpT$ simulations indicate that the T phase might be more stable, a definite answer can only
come from free-energy calculations.
\begin{figure}
\epsfig{file=fig9.eps,width=2.5in}
\caption{\label{clust} Possible two- and four-particle configurations for hard
right triangles. (a) Tetramer. (b) Square dimer. (c) Triangular dimer. (d) Rhomboidal
dimer.}
\end{figure}
The different symmetries
obtained depend very sensitively on the starting configuration and the path, either
compression or expansion, followed in the simulation runs.
Obviously the high packing and the two dimensionality of the system
causes particles to have a low rotational diffusion and to find
difficulties to explore different local symmetries.
An additional effect is the clustering tendencies of particles with right linear sectors.
A recent study relates the formation of
clusters of polyhedral particles with chemical bonding \cite{Glotzer2}. Local particle
configurations vary in type according to the global symmetry of a phase. For hard right
triangles several clusters can be identified, see Fig. \ref{clust}. In the
T phase we expect a high fraction of dimers (two triangles or `monomers' in close
contact along their long or short sides, forming groups with square and triangular shape,
respectively) and tetramers [four close monomers in configurations as shown in
Fig. \ref{clust}(a)], and
a low proportion of dimers forming a rhomboid. The latter are present in the isotropic
and octatic phases, which means that a compression of the isotropic phase will lead to an octatic
texture or frustrated tetratic phase.
In turn, a T phase may also form in the absence of tetramers, giving rise to
different tetratic configurations with a history-dependent structure.
In any case, our simulations seem to rule
out the existence of the uniaxial N phase predicted by the standard SPT-based DFT
or the extended SPT theory including the third virial coefficient.
The intermediate phase between the isotropic and the crystal phases most
likely exhibits strong tetratic and octatic correlations, none of which are detected in the DFT calculations.
However, a final answer will have to wait for a more detailed simulation study including
free-energy calculations, and the stability of a uniaxial N phase in a limited density
interval cannot be ruled out.
\section{Conclusions}
\label{conclusions}
In the present study we have shown that a SPT-based DFT incorporating the exact second and third virial coefficients
is not capable of predicting stable T and O liquid-crystal phases in a fluid of hard right isosceles triangles.
In addition, the orientation distribution function $h(\phi)$ of the only stable phase predicted,
namely the uniaxial phase, has no signature of T or O correlations. However these symmetries were found in the MC
simulations of Ref. \cite{Dijkstra} and in our own simulations, as shown by the orientational distribution function
$h(\phi)$ and the order parameters $Q_k$ ($k=4,8$) in some range of packing fractions.
Our results imply that the stable symmetries of the fluid of hard right triangles are driven by
particle correlations involving more than three particles. In view of the practical impossibility to
implement higher-order correlations in a DFT model, the results of the present article
point to the necessity to formulate some new theoretical description, still within the framework of DFT,
but different from its standard form, to describe the phase behavior of right isosceles triangles.
Indeed virtually all theoretical DFT models constructed so far for the description of uniform phases of two-dimensional
or three-dimensional anisotropic particles are ultimately based on the knowledge of the second virial coefficient.
These models have been successfully applied to the description of liquid-crystal symmetries
for a vast variety of hard particles, and in general they correctly predict the symmetry of these
phases and their stability ranges qualitatively, and even quantitatively in some cases.
The present fluid is, to our knowledge, the first example where the standard theoretical
tool used by the liquid-crystal community since Onsager formulated his theory breaks down.
However, there are different avenues to remedy the deficiencies of the standard
theory, still within the framework of DFT.
We suggest that clustering (or self assembling) of particles may be a fruitful
view to overcome the difficulties posed by an approach based on a virial expansion. Particle shape
alone is enough to identify particularly stable local particle configurations, which can be taken as
entities upon which to formulate a DFT model. This approach would automatically incorporate the
important correlations through the very identification of the possible clusters, the distribution
of which would be obtained in a thermodynamically consistent way, providing a
mechanism to explain the peculiar (four- or eightfold) orientational symmetries
of the liquid-crystal phases of hard right triangles.
For example, the presence of a large amount of dimers and tetramers
forming square-like super-particles and both arranged in a T-like
configurations can lead to monomer axes, depending on their relative fractions, that orient parallel
to four or eight, approximately equivalent, directors. Consequently the angular distribution functions
of monomers could approach the symmetries $h(\phi)=h(\phi+n\pi/4)$ ($n=1,2$).
It is reasonable to appeal to the clustering effect to explain why the T and O phases,
observed in simulations, cannot be stabilized by the usual
implementation of DFT. A more realistic model to account for clustering is to treat the fluid as a polydisperse
mixture of clusters of different sizes and shapes.
An internal energy for clusters, in the line of standard
models for associated fluids, should be introduced.
A study of the relation between phase symmetry and particle clustering, and the formulation of
a cluster model, which we believe is a promising research line, is left for the future.
To end, we would like to comment on two distinct extensions of the problem.
First, our DFT study is restricted only to uniform phases as by construction these
are the liquid-crystal orientational symmetries that the present DFT can predict.
The study of non-uniform phases can be tackled with a theory optimized for
inhomogeneous density profiles. One of the promising versions of DFT
is that constructed from fundamental measure theory
(see for example Ref. \cite{Wittmann}). This theory was formulated for any hard particle geometry.
Even though the theory is computationally expensive, it should be possible to
fix different crystalline lattices with T symmetry, for example the ones
formed by tetramers or by randomly oriented dimers,
and minimize the functional with respect to the density profile. The relative stability
between different crystalline phases or between the fluid and the crystal could then be studied
and compared with the predictions of MC simulations \cite{Dijkstra}.
Finally, one might wonder about the effect of polydispersity in the present system. A
perfect nematic phase of monomers
and also a perfect tetratic phase of tetramers under compression lead to perfect plane tiling
at close packing, with the latter
having greater orientational entropy and consequently lower free energy. If polydispersity in the opening
angle is present, it could affect the stability of these two phases. In particular the balance between the
orientational and configurational entropy can severely be affected in the T phase of tetramers.
The stability of the crystal phase of tetramers should be more affected
by polydispersity because polydisperse monomers cannot fit into
identical tetrameral units to form a crystal. However the effect could be less for the uniaxial crystal formed
by monomers, possibly resulting in a stable uniaxial N phase at lower densities.
\acknowledgements
Financial support under grant FIS2017-86007-C3-1-P
from Ministerio de Econom\'{\i}a, Industria y Competitividad (MINECO) of Spain.
Y. M.-R. acknowledges the support from Grant No. PGC2018-096606-B-I00 (MCIU/AEI/FEDER, UE).
|
\section{Introduction}
\label{sec:intro}
Despite the compelling indirect evidence for new fundamental particles from astrophysical and other observations, no direct discoveries have been confirmed since the identification of the Higgs boson~\cite{ATLAS:2012yve,CMS:2012qbp}. This means that the new physics is either rare, inaccessible, or we are looking in the wrong place for it. This last possibility has motivated a new anomaly detection research program at particle colliders by which search strategies are constructed with less model dependence than previous approaches. Many of these new methods employ modern machine learning to achieve broad sensitivity to unforeseen scenarios~\cite{Collins:2018epr,Collins:2019jip,Andreassen:2020nkr,Nachman:2020lpy,collaboration2020dijet,1815227,Stein:2020rou,Kasieczka:2021tew,Shih:2021kbt,Hallin:2021wme,Farina:2018fyg,Heimel:2018mkt,Roy:2019jae,Blance:2019ibf,Collins:2021nxn,Kahn:2021drv,Amram:2020ykb,Kasieczka:2021xcg,Hajer:2018kqm,DeSimone:2018efk,Mullin:2019mmh,1809.02977,Dillon:2019cqt,Romao:2019dvs,Romao:2020ojy,knapp2020adversarially,Cerri:2018anq,Govorkova:2021utb,Govorkova:2021hqu,1797846,1800445,Amram:2020ykb,Cheng:2020dal,pol2020anomaly,Atkinson:2021nlt,Khosa:2020qrz,Thaprasop:2020mzp,Alexander:2020mbx,aguilarsaavedra2020mass,vanBeekveld:2020txa,Park:2020pak,Faroughy:2020gas,Chakravarti:2021svb,Batson:2021agz,Blance:2021gcs,Bortolato:2021zic,Dillon:2021nxw,Finke:2021sdf,Aarrestad:2021oeb,Caron:2021wmq,Volkovich:2021txe,Ostdiek:2021bem,Fraser:2021lxm,DAgnolo:2018cun,DAgnolo:2019vbw,Dorigo:2021iyy,Aguilar-Saavedra:2017rzt,Mikuni:2020qds} (list from Ref.~\cite{Feickert:2021ajf}).
A complete anomaly detection algorithm is required to have two attributes: it should be sensitive to anomalous events and it should be possible to estimate the rate of Standard Model (SM) events that are labeled as anomalous (false positive rate)~\cite{Nachman:2020lpy}. Complete anomaly detection methods have so far primarily focused on {\it resonant} anomalies, where data sidebands can be used as reference samples to both construct signal-sensitive classifiers and to estimate the SM background~\cite{Collins:2018epr,Collins:2019jip,Andreassen:2020nkr,Nachman:2020lpy,collaboration2020dijet,Amram:2020ykb,1815227,Stein:2020rou,Kahn:2021drv,Kasieczka:2021tew,Hallin:2021wme,Shih:2021kbt}.
Much less well-explored so far has been complete anomaly detection methods for {\it non-resonant} anomalies. One widely studied approach based on unsupervised learning that does not require the new physics to be resonant is the autoencoder~\cite{Farina:2018fyg,Heimel:2018mkt,Roy:2019jae,Cerri:2018anq,Blance:2019ibf,Hajer:2018kqm,DeSimone:2018efk,Mullin:2019mmh,1809.02977,Dillon:2019cqt,Romao:2019dvs,Romao:2020ojy,knapp2020adversarially,1797846,1800445,Amram:2020ykb,Cheng:2020dal,Khosa:2020qrz,Thaprasop:2020mzp,Alexander:2020mbx,aguilarsaavedra2020mass,pol2020anomaly,vanBeekveld:2020txa,Park:2020pak,Faroughy:2020gas,Kasieczka:2021xcg,Chakravarti:2021svb,Batson:2021agz,Blance:2021gcs,Bortolato:2021zic,Collins:2021nxn,Dillon:2021nxw,Finke:2021sdf,Atkinson:2021nlt,Kahn:2021drv,Aarrestad:2021oeb,Caron:2021wmq,Govorkova:2021hqu,Volkovich:2021txe,Govorkova:2021utb,Ostdiek:2021bem,Fraser:2021lxm}.
The idea is to build models for compressing and uncompressing events, trained directly on the (mostly background) data. Events that have a low probability density tend to be poorly reconstructed when compressing and uncompressing compared with events that have a relatively higher probability density. If anomalous events are located in regions of low data probability density, then the reconstruction quality can be used as an anomaly score.
However, autoencoders by themselves are not a complete anomaly detection algorithm - they provide a method for achieving signal sensitivity, but they do not have a natural background estimation component.
In the non-resonant case, one could compare the spectrum of anomalous events with background-only simulations, but this requires an excellent model of the background. Given that we expect the unexpected to occur in regions that are poorly modeled, this is unlikely to be a viable strategy in general.
In this paper, we introduce a new method for detecting non-resonant anomalies, based on autoencoders, that is complete in the sense that it includes both signal sensitivity and simulation-free background estimation. Instead of constructing one autoencoder, we advocate for training two or more autoencoders. The set of autoencoders are trained to be as independent of each other as possible.
While many methods for decorrelating neural networks exist~\cite{Louppe:2016ylz,Dolen:2016kst,Moult:2017okx,Stevens:2013dya,Shimmin:2017mfk,Bradshaw:2019ipy,ATL-PHYS-PUB-2018-014,DiscoFever,Xia:2018kgd,Englert:2018cfo,Wunsch:2019qbo,Rogozhnikov:2014zea,10.1088/2632-2153/ab9023,clavijo2020adversarial,Kasieczka:2020pil,Kitouni:2020xgb,Ghosh:2021hrh} and could be used here, we chose to employ the DisCo decorrelation method first developed in Ref.~\cite{DiscoFever} and explored for simultaneous background estimation in Ref.~\cite{Kasieczka:2020pil}.
Events are labeled as anomalous if the reconstruction quality is poor for all autoencoders. Events labeled as anomalous by one, but not all, of the autoencoders provide the context needed to estimate the Standard Model background in a model independent way, via the ABCD method
An additional benefit of our method is that it can be run equally well online or offline; indeed this forms a second major motivation for our work. Typically, a key assumption is that anomalous events will be saved by the detectors for offline analysis. Due to the immense data rate at the Large Hadron Collider (LHC), it is not possible to save every collision event for offline processing. Instead, a system of triggers are used to save interesting events~\cite{ATLAS:2020esi,CMS:2016ngn}. The definition of interesting is model dependent and therefore the new physics may be thrown away in real time. It is therefore of utmost importance to design model independent strategies for saving anomalous events.
Autoencoders can be run online because they do not require comparing data to a reference sample~\cite{knapp2020adversarially,Cerri:2018anq,Govorkova:2021utb,Govorkova:2021hqu}. However, no autoencoder-based trigger proposal so far has been complete in the sense introduced above. Many conventional triggers are complemented by \textit{support} triggers which provide the context needed for data-driven background estimation offline. Our method provides the first complete anomaly detection strategy in a similar way to these conventional methods. By using two decorrelated autoencoders, we can trigger on potentially anomalous events and then additionally save (at a reduced rate) anti-tagged events in a way that background estimation is possible offline.
This paper is organized as follows. First, we introduce the technique of decorrelated autoencoders in Sec.~\ref{sec:method}. Numerical results with the ADC2021 dataset are presented in Sec.~\ref{sec:results}. By definition, this demonstration highlights an offline application of our approach. Section~\ref{sec:online} provides a discussion about the online-compatibility of our technique for experimental integration online. The paper ends with conclusions and outlook in Sec.~\ref{sec:conclusions}.
\section{Decorrelated Autoencoders}
\label{sec:method}
A vanilla autoencoder is a composition of two functions, an encoder $g$ and a decoder $f$. These two functions are parameterized as neural networks and are optimized to minimize the reconstruction loss:
\begin{align}
\label{eq:auto}
L[f,g]=\sum_i (f(g(x_i))-x_i)^2\,,
\end{align}
where $x\in\mathbb{R}^n$, $g:\mathbb{R}^n\rightarrow\mathbb{R}^m$, and $f:\mathbb{R}^m\rightarrow\mathbb{R}^n$. In order to encourage compression, the latent space dimension is chosen such that $m<n$. A popular variation on this setup is the variational autoencoder~\cite{kingma2014autoencoding,Kingma2019}, whereby the encoding and decoding are probabilistic and the latent space has well-defined statistical properties. The methods proposed here are compatible with variational autoencoders, and while preliminary studies indicate that the results are similar, we leave a full exploration to future work.
Instead of training a single autoencoder as in Eq.~\ref{eq:auto}, we propose to train two (or more) {\it statistically independent} autoencoders at the same time, in order to enable data-driven background estimation. Following \cite{DiscoFever,Kasieczka:2020pil}, we achieve the decorrelation of the autoencoders by including in the training a regularizer term based on the distance correlation (DisCo) measure of statistical dependence. Focusing on the case of two autoencoders $(f_1,g_1)$ and $(f_2,g_2)$ for simplicity, we consider the following loss function:
\begin{align}
\label{eq:autodec}\nonumber
L[f_1,f_2,g_1,g_2]=&\sum_i R_1(x_i)^2+\sum_i R_2(x_i)^2\\
&+\lambda\, \text{DisCo}^2[R_1(X),R_2(X)]\,,
\end{align}
where $R_i(x)=(f_i(g_i(x))-x)^2$, $\lambda > 0$ is a hyperparameter, and DisCo is the distance correlation~\cite{szekely2007, szekely2009, SzeKely:2013:DCT:2486206.2486394,szekely2014}. DisCo is between $0$ and $1$ and is zero if and only if its arguments are independent. The capital $X$ is used in the last term of Eq.~\ref{eq:autodec} to indicate that the distance correlation is computed at the level of a batch of examples $x$, which are realizations of the random variable $X$. Given autoencoders trained via Eq.~\ref{eq:autodec}, we can define counts $N_{\lessgtr,\lessgtr}(\vec{c})=\sum_i\mathbb{I}[R_1(x_i)\lessgtr c_1]\,\mathbb{I}[R_2(x_i)\lessgtr c_2]$, where $\vec{c}=(c_1,c_2)$ are given thresholds and $\mathbb{I}[\cdot]$ is the indicator function that is zero when its argument is false and one otherwise. The signal sensitive region is $N_{>,>}(\vec{c})$ and the other three regions can be used to estimate the background:
\begin{align}
\label{eq:abcd}
N_{>,>}^\text{predicted}(\vec{c})&=\frac{N_{>,<}(\vec{c})N_{<,>}(\vec{c})}{N_{<,<}(\vec{c})}\,.
\end{align}
Equation~\ref{eq:abcd} is known as the ABCD method and the $N_{>,>}(\vec{c})$ is exactly the background in the signal-sensitive region if there are enough events and if the two dimensions are effective at rejecting the background.
\section{Empirical Results}
\label{sec:results}
The performance of the double autoencoder and decorrelation strategy is tested on the ADC2021 dataset, which was created for unsupervised anomaly detection~\cite{40mhz,Govorkova:2021hqu}. In the dataset, proton-proton collisions at the LHC are simulated at center-of-mass energy of 13~TeV. Collision events are required to contain at least one electron ($e$) or muon ($\mu$) with transverse momenta $\mathrm{p_{T}} > 23$~GeV. A set of various Standard Model processes are generated with \textsc{Pythia} 8.240 generator~\cite{Sjostrand:2006za,Sjostrand:2014zea} with detector response modeled by \textsc{Delphes} 3.3.2~\cite{deFavereau:2013fsa,Selvaggi:2014mya,Mertens:2015kba} using the Phase-II CMS detector card.
During the training, 2 million events are used while results are reported using an independent validation set containing 800k SM events.
Four benchmark scenarios containing new physics processes are used to evaluate the performance of the algorithm: a leptoquark (LQ) with 80~GeV mass decaying to a $b$-quark and a $\tau$ lepton, a neutral scalar boson ($A$) of 50~GeV mass decaying to a pair of off-shell $Z$ bosons, which in turn are forced to decay to leptons ($A\rightarrow 4l$), a scalar boson $h^0$ of 60~GeV mass decaying to a pair of $\tau$ leptons ($h^0\rightarrow \tau\tau$), and a charged scalar boson $h^\pm$ with 60~GeV mass, decaying to a $\tau$ lepton and a neutrino ($h^\pm\rightarrow\tau\nu$). In the performance evaluation, each new physics scenario is considered independently, with total amount of events fixed to 0.1\% of the total sample size.
The autoencoders are trained on a sample of pure background events. In practice, this corresponds to the case of training on simulation and testing on data. Differences between data and simulation (which are not modeled or taken into account in the ADC2021 dataset used here) may degrade the autoencoder performance if background data events are not reconstructed as well by the autoencoder as background simulation events. Fortunately, it is well-known from previous studies (see e.g.~\cite{Farina:2018fyg,Heimel:2018mkt,Cerri:2018anq}) that autoencoder training is highly insensitive to low amounts of signal contamination. This means that autoencoders can be trained directly on data with a small amount of signal contamination without a significant change in the learned neural networks. We have explicitly verified this in the case of the decorrelated autoencoders using the $A\rightarrow 4l$ signal.
Each autoencoder architecture is built using deep neural networks containing five fully connected layers. The encoders have 256, 128, 64, 32, and 5 hidden nodes, while the decoder is simply the mirrored version of the encoder. The inputs given to the training are the four-momenta of jets~\cite{Cacciari:2011ma,Cacciari:2005hq} and leptons in ($\mathrm{p_{T}}$, $\eta$, $\phi$, m) coordinates.
Only the first (sorted by $\mathrm{p_{T}}$) four muons, four electrons, and 10 jets in the event are kept with zero padding if fewer objects are present.
The implementation is carried out with \textsc{Tensorflow}~\cite{tensorflow} optimized with the \textsc{Adam}~\cite{adam}. Even though all new physics scenarios considered here contain a mass resonance, no invariant mass information is directly used in the training process. The $\lambda$ parameter from Eq.~\ref{eq:autodec} is fixed to 100 and training batch size fixed to 10k to improve the decorrelation performance. The double autoencoder structure is then trained for a total of 1000 epochs, or stopped if the overall training loss does not improve in an independent testing set for 10 consecutive epochs. The complete model uses 230k trainable weights with a total of 460k floating point operations. The neural network architecture and training procedure were not extensively optimized, due in part to the unsupervised nature of this task.
The performance of each autoencoder for anomaly detection is assessed by using the reconstruction loss as the main discriminator. The
significance improvement characteristic (SIC) curves are built for each new physics scenario shown in Fig.~\ref{fig:roc_sic}. The comparison with a single autoencoder trained without the decorrelation loss is also shown. We also show in Fig.~\ref{fig:roc_sic} the combined performance of both autoencoders, where a ``diagonal" cut that yields the same SM background efficiency for each autoencoder is employed. We see that the signal sensitivity of the combined autoencoders is greater than (in fact roughly double) that of each autoencoder individually, indicating that the decorrelation was successful and each autoencoder learned something independent about the BSM anomalies in question. The SIC distribution for different combinations of thresholds in each autoencoder is shown in App.~\ref{app:sic_curves}.
\begin{figure}[h!]
\centering
\includegraphics[width=0.4\textwidth]{figures/sic_comparison_double.pdf}
\includegraphics[width=0.4\textwidth]{figures/sic_comparison.pdf}
\caption{Top: significance improvement characteristic (SIC) curves of the individual, decorrelated autoencoders, for each new physics benchmark scenario. Bottom: SIC curves for the ``diagonal" combination of both decorrelated autoencoders, compared with that of a single autoencoder (obviously trained without any decorrelation). In both panels, the SIC curves are cut off at lower true positive rates due to low background yields.
}
\label{fig:roc_sic}
\end{figure}
The independence between reconstruction losses is more fully validated by estimating the difference between the background predicted using the ABCD method (Eq.~\ref{eq:abcd}) and the real number of background events in the region of interest. The ratio between the two quantities is shown in Fig.~\ref{fig:closure}. Multiple choices of $\vec{c}$ yielding the same SM efficiency are tested (represented as multiple entries in Fig.~\ref{fig:closure}) for samples containing only SM processes (purple) and mixtures of SM and a new physics process. At lower SM efficiencies, departures from unity (overestimated background~\cite{Kasieczka:2020pil}) are observed in all mixed samples while the highest variation for a sample containing only SM events is 2.5\%, compatible the statistical uncertainty of the sample.
\begin{figure}[h!]
\centering
\includegraphics[width=0.45\textwidth]{figures/closure_comparison.pdf}
\caption{Closure test of the ABCD estimation method for different SM efficiencies and benchmark scenarios. Different selection combinations yielding the same background efficiency are shown as independent entries.}
\label{fig:closure}
\end{figure}
These differences can also be quantified in terms of the signal significance for each benchmark process by comparing the observed and predicted number of background events from the ABCD method. Given $N$ observations in the region of interest with predicted number of background events $B$, the significance is defined as $\frac{N-B}{\sqrt{N}}$ if $N - B >0$ and 0 otherwise. With the initial signal fraction fixed, the total sample significance is around 0.8 prior to the application of the method. As pointed out in Ref.~\cite{Kasieczka:2020pil}, unaccounted contamination from the signal of interest in the ABCD sidebands may result in different significance values when compared to the correct estimation of the background. While this issue can be accounted when performing model specific exclusion limits, we also show in Fig.~\ref{fig:sig} (top) the significance obtained using the ABCD method with and without correcting the number of background events. To avoid fine tuning, the threshold applied to each autoencoder reconstruction loss is the one where both autoencoders have the same SM rejection efficiency.
In all new physics benchmark scenarios, the uncorrected significance for SM efficiencies above 1.5\% is lower than the corrected for SM efficiencies. Nevertheless, all new physics scenarios show significance between 1 to 4 while the SM only sample has a maximum deviation below 1. We have also probed the stability of the method by performing five independent trainings with different random weight initialization. The standard variation of the average significance was below 6\% for all benchmark scenarios tested.
The additional distance correlation loss leads to increased reconstruction loss in the background training sample, resulting in decreased performance compared to a single autoencoder training. This difference is illustrated in Fig.~\ref{fig:sig} (bottom) where the significance is compared with the values obtained from training a single autoencoder with same network architecture. Since the ABCD method is only applicable in the double autoencoder case, no background estimation method is used in the comparison. In all cases, the difference in significance of the double and single autoencoders is less than 30\%. While the single autoencoder consistently outperforms the double autoencoder, the lack of dedicated background estimation might lead to unattainable performances when applied to real particle collisions.
\begin{figure}[h!]
\centering
\includegraphics[width=0.45\textwidth]{figures/significance.pdf}
\includegraphics[width=0.45\textwidth]{figures/sig_comparison.pdf}
\caption{Signal significance for each benchmark scenario (top) when the ABCD method is used to predict the background level (solid lines) compared to the real significance value (dashed lines). In the bottom panel, the comparison of the significance between a single autoencoder (dashed lines) and the double autoencoder (solid lines) is shown. In this case, no background estimation method is used.}
\label{fig:sig}
\end{figure}
\section{Anomaly Detection Online}
\label{sec:online}
The discussion so far has demonstrated that the decorrelated autoencoder protocol is an effective tool for simulation-free, non-resonant anomaly detection. This section briefly describes how this technique is also online compatible.
We envision that in an actual trigger system, we would save all events in the signal sensitive region defined by the two autoencoders and then save a random fraction (`prescale') of events in the three other regions for offline background estimation (similar to existing `support triggers' for certain background processes). The prescale would be set so that the statistical uncertainty on the background prediction is smaller than the statistical uncertainty from events in the signal region. If the SM efficiency in the signal region is $\epsilon$, the trigger rate would scale approximately as $4\epsilon$, including events saved from the background-dominated regions. The autoencoders themselves could be trained directly on data. These data could be from a previous run or from earlier in a given run. We note that this is the first complete online compatible anomaly detection protocol to be proposed - previous proposals have used single autoencoders and do not come with a method for estimating the background.
Moreover, each of our autoencoders is built using only a set of fully connected layers to allow for a memory and time efficient implementation. There have been many recent demonstrations of ultra low-latency implementations of these and related architectures on Field Programable Gate Arrays (FPGAs)~\cite{Duarte:2018ite,Coelho:2020zfu,Aarrestad:2021zos,Heintz:2020soy,Govorkova:2021utb,John:2020sak,Hong:2021snb}. For studies with more computational resources available, the baseline performance of each autoencoder may be enhanced using more complex reconstruction strategies, as studied in \cite{Cheng:2020dal,pol2020anomaly,Atkinson:2021nlt,Collins:2021pld,Orzari:2021suh}.
The ADC2021 community challenge dataset was used in part because it was created for the purpose of developing online methods~\cite{40mhz,Govorkova:2021hqu} as summarized by the challenge title: \textit{Unsupervised New Physics detection at 40 MHz}. However, there are some features of this dataset which limit direct connection to online algorithms. For example, ATLAS and CMS have single lepton triggers that would likely save all of the challenge events for offline processing. Figure~\ref{fig:sig} indicates that our decorrelated autoencoder trigger reduces the bandwidth by nearly two orders of magnitude. This is not necessarily relevant for the lepton-triggered data, but it is a common reduction for dedicated triggers. Another issue is, as we have already noted above, that the ADC2021 dataset does not distinguish between ``data" and ``simulation"; this could be an issue for machine learning methods, which generally require a representative sample for the training data. Expanding the online challenge to other datasets would be interesting for the future.
\section{Conclusions and Outlook}
\label{sec:conclusions}
We have proposed a first complete online-compatible unsupervised non-resonant anomaly detection method that achieves signal sensitivity and can be used to estimate the SM background\footnote{The scripts used to produce the results shown in this work are available at: \url{https://github.com/ViniciusMikuni/DoubleAE}}. Autoencoders, a popular choice of anomaly detection algorithm, are used to identify anomalous events through a reconstruction loss. We advocate for the combination of two or more autoencoders that are trained simultaneously while a distance correlation (DisCo) regularizer term is added to make their reconstruction losses statistically independent. In this strategy, the background from SM events can be estimated with the ABCD method. In the absence of new physics, the method shows a good agreement between the predicted and observed amount of background events. In the presence of new physics, the signal significance varies between 1 and 4 for multiple new physics scenarios with initial contribution amounting to only 0.1\% of all events.
Given that our method is architecture agnostic, it can be readily generalized for other anomaly detection methods whose output is an anomalous score capable of discerning new physics scenarios from background events.
\section*{Acknowledgments}
We thank Barry Dillon and Gregor Kasieczka for feedback on the manuscript. VM and BN are supported by the U.S. Department of Energy (DOE), Office of Science under contract DE-AC02-05CH11231. The work of DS was supported by DOE grant DOE-SC0010008.
|
\section{Introduction}
This work concerns reconstructing sparse solutions of ill-posed problems while encouraging more flexibility than allowed by classical regularization penalties in $\ell^p$ spaces with $p\in (0,\infty)$. Let us first recall some literature concerning $\ell^p$ sparse regularizers. The inspiring paper ~\cite{daubechies2004iteratethresh} pointed out their essential role in promoting sparsity with respect to orthonormal bases for solving inverse problems, in general. The reader is referred to
\cite{claerbout1973robust,taylor1979deconvolution,levy1981reconstruction,santosa1986linear,chen1998basispursuit} for previous works on specific inverse problems and to \cite{tibshirani1996lasso} for statistical approaches (e.g., LASSO). Several theoretical results regarding convergence and error estimates can be found in \cite{lorenz2008reglp,grasmair2008sparseregularization,ramlau2010sparse} for the convex case $p\in[1,2)$, and in \cite{zarzer2009nonconvextikhonov,grasmair2008pleq1,10,bredies2009nonconvexregularization} for the nonconvex case $p\in(0,1)$.
Moreover, it has been shown numerically that $\ell^p$ terms with $p \in (0,1)$, promote sparsity better than the $\ell^1$-norm (see e.g. \cite{10}), for instance in feature selection and compressed sensing \cite{11} (allowing a smaller number of measurements) or in total variation-based
image restoration \cite{5} (providing better edge preservation). There is an increasing interest in this topic not only within signal processing and image restoration, but also within fracture mechanics, optimal control and optimization - see the introduction of \cite{GK1} .
The expansive literature on $\ell^p$ regularization deals with many other fine and interesting aspects, which are however beyond the scope of this work.
The focus here is on flexible sparsity variational regularization that facilitates different penalizations for different coefficients of the regularized solution, and on numerical approaches for the nonconvex regularized problem. Actually, by ``flexible sparsity'' we mean functionals that
regularize different (groups of) indices with a different sparsity
promoting term (see also ~\cite{lorenz2016flexiblesparse}). In general, it is about partitioning the index set $\mathbb{N}$ into
sets $I_{1}$, $I_{2}$,\dots and choosing a specific penalty for $I_k$, e.g., based on a sequence of exponents
$p_{k}>0$. Note that these flexible penalties have been used in various applications for the convex case - see e.g.~\cite{chaari2009minimization,briceno2011proximal,pustelnik2011parallel}.
There are different ways to form a regularization
functional out of these partitions.
First, one can set the $\phi$-functional as
\[
\phi(u) = \sum_{k\in\mathbb{N}} \sum_{j\in I_{k}}w_{j}|u_{j}|^{p_{k}}
\]
with different weights $w_k$. If we denote $u_{I} = (u_{j})_{j\in I}$ and let $\norm{u_I}_{p,w} = (\sum_{j\in I}w_{j}|u_{j}|^{p})^{1/p}$ for weights $w_{j}>0$, we can write this as
\[
\phi(u) = \sum_{k\in\mathbb{N}}\norm{u_{I_{k}}}_{p_{k},w}^{p_{k}}.
\]
A further generalization is considering different functions
$\phi_{k}:[0,\infty)\to [0,\infty)$ which are nondecreasing and satisfy $\phi_{k}(0)=0$, yielding
\begin{equation}\label{eq:phi-func-phi}
\phi(u) = \sum_{k\in\mathbb{N}} \sum_{j\in I_{k}}\phi_{k}(u_{j}).
\end{equation}
Here explicit weights $w_{k}$ are not needed, as they can be incorporated in the functions $\phi_{k}$.
Second, the mixed norms approach adds the different $p$-norms directly and yields
\begin{equation}\label{eq:mixed-norm}
\norm{u}_{p_{k}} = \sum_{k\in\mathbb{N}}\norm{u_{I_{k}}}_{p_{k},w}.
\end{equation}
A notable difference to the $\phi$-functional is that the latter is actually a norm if $p_{k}\geq 1$.
The reader is referred also to the section on group sparsity in \cite{grasmair2011homogeneous}, where the case $p_k=2$ for each $k$ is tackled.
Further generalizations are possible, e.g., one could also consider a functional of the form
$(\sum_{k\in\mathbb{N}}\norm{u_{I_{k}}}_{p_{k},w}^{q})^{1/q}$ (i.e. taking
a $q$-norm of the vector of values of the sub-norms instead of the
$1$-norm) or even combine different groups of sub-norms with
different exponents (e.g., one can do this hierarchically). We do not
pursue these generalizations further.
From a theoretical point of view, we extend the work \cite{lorenz2016flexiblesparse} that involves variable exponents penalties. Thus, we are concerned with existence of minimizers for regularization with nonconvex functions $\phi_k$
$$
\phi(u)=\sum_{k\in\mathbb{N}}\phi_k(\abs{u_k}),
$$
sparsity properties of the minimizers, and convergence to a solution of the original problem in an infinite dimensional setting. As a new challenge, we propose algorithmic approaches for minimizing least squares functions with such general
nonconvex regularization terms, by relying on iterative majorization of the nonconvex part. Thus, we write each nonconvex function $\phi_k$ as a minimum of convex functions by means of the concave conjugate of $\phi_k$. Approaches of this type have been (re)discovered several times. To the best of our knowledge, they concern only nonconvex $\ell^p$ penalties with fixed exponent or are discussed only theoretically for variable penalties. The earliest references we could track down are~\cite{Geman1992,Geman1995,Black1996} - see also \cite{Gorodnitsky1997} using the name iterative reweighting, while more recent rediscoveries are \cite{Chan2014,Lanza2015,37}. We refer also to \cite{Nikolova2005,Allain2006} dealing with convergence of the half-quadratic algorithms.
There are several possibilities to realize the iterative majorization approach. We state here two methods yielding iterative reweighting algorithms, based on concave duality for increasing functions. One of them leads to a sequence of reweighted $\ell^1$-minimization problems (like in~\cite{yu2019iteratively,chen2014convergence}), while the other one leads to a sequence of $\ell^2$-minimization problems, see also~\cite{wipf2010irls}. We investigate convergence of the schemes and monotonicity in the sense of decreasing the objective
function in every iteration. The $\ell^2$-minimization approach is developed into a full
algorithmic framework in Section \ref{sec:optcond}.
Numerical results in two different situations are given, the first one for an academic M matrix example and the second one for a time-dependent optimal control problem. The scheme is proved to be effective and accurate in both situations. In the optimal control example, we carry out a comparison with the fixed penalty case and emphasize the advantages of employing
variable penalties.
A significant reference for the current work is \cite{GK1}, where a monotone scheme for the regularized problem (with fixed $p$) is introduced. This inspired us to propose the algorithm from Section \ref{sec:optcond}. Moreover, in \cite{GK1}, a primal dual active set strategy is studied and tested. The adaptation of this strategy to flexible regularization is out of the scope of the present paper, but will be a subject of future work.
One can mention further numerical schemes for nonconvex regularization with a fixed penalty for all coefficients, such as a generalized gradient projection method \cite{6}, a generalized iterated shrinkage algorithm \cite{52}, and a generalized Krylov subspace method combined with a reweighting technique \cite{Lanza2015}.
In \cite{44}, a functional approach combined with a gradient technique is proposed, and in \cite{36} an alternating direction method of multipliers (ADMM) is studied.
We mention also \cite{XuChXuZh12}, where an iterative half thresholding for fast solution of $L_{1/2}$ regularization is proposed. Finally group sparse optimization problems have been studied in \cite{HuLiMeQiYa17} via an $\ell_{p,q}$ regularization with $p\geq 1$ and $q \in [0,1]$ and numerically solved through a proximal gradient method, and in \cite{BeHa19} through the computation of the proximal mapping and necessary optimality conditions.
Last but not least, a rule on how to choose the flexible functionals is left for future work. We would like to approach this topic
by learning the penalty.
This paper is structured as follows. In Section \ref{sec:2}, we analyse the general nonconvex regularization problem. Section \ref{sec:iterative_majorization} initiates the discussion on iterative majorization approaches for the mentioned problem and presents the necessary concave duality theory. In Section \ref{it_T} we derive a simple $\ell^2$-minimization scheme for flexible sparsity regularization. Section \ref{sec:optcond} is devoted to the theoretical and numerical study of a monotone algorithm that approximates regularized solutions. A reweighted $\ell^1$-minimization approach is proposed and analysed in Section \ref{1-section}.
\section{Variational regularization with flexible functionals}\label{sec:2}
Let
\begin{equation}\label{equation}
Au=y
\end{equation}
denote the operator equation we would like to stably solve via variational regularization, where $A:\ell^2\to Y$ is linear and bounded, $Y$ is a Hilbert space.
Consider the following Tikhonov regularization approach
\begin{equation}\label{reg_nonconvex}
\min_{u\in{\ell^{\phi_k}}}\tfrac12\norm{Au-y}^{2} + \alpha \phi(u),\,\,\,\alpha>0,
\end{equation}
with $\phi$ given by
\begin{equation}\label{phik}
\phi(u)=\sum_{k\in\mathbb{N}}\phi_k(\abs{u_k}),
\end{equation}
where
\[
\ell^{\phi_k} = \{ (u_{k}): \sum_{k}\phi_{k}(|u_{k}|)<\infty\}.
\]
The following assumption will be used throughout this work:
(A1) The functions $\phi_{k}:[0,\infty)\to [0,\infty]$ are nondecreasing, continuous, and satisfy $\phi_k(0)=0$ and $\lim_{t\to\infty}\phi_k(t) = \infty$, $\forall k\in\mathbb{N}$.
In the sequel, we focus on penalties of the form \eqref{phik}
with all $\phi_k$ simultaneously concave. Several examples of functions $\phi$ in this sense that satisfy (A1) are in order:
\begin{enumerate}
\item $\displaystyle{\phi(u)= \sum_{k\in\mathbb{N}} |u_k|^{p_k}}$
\item $\displaystyle{\phi(u)=\sum_{k\in\mathbb{N}} \log(|u_k|^{p_k}+1)}$
\end{enumerate}
with $p_k\in (0,1)$.
Inspired by \cite{akguen_yildirir} and \cite{Zhao12reweighted-l1-minimization}, the following functions are also of interest:
\begin{enumerate}
\item[(3)] $\displaystyle{\phi(u)= \sum_{k\in\mathbb{N}} \log(|u_k|+1)+ |u_k|^{p_k},\,\,\,p_k\in (0,1)}$
\item[(4)] $\displaystyle{\phi(u)= \sum_{k\in\mathbb{N}} (|u_k|+|u_k|^{p_k})^{q_k},\,\,\,p_k, q_k\in (0,1).}$
\item[(5)] $\displaystyle{\phi(u)= \sum_{k\in\mathbb{N}} |u_k|^{p_k}\log(|u_k|+1) ,\,\,\,p_k\in (0,1)}$
\item[(6)] $\displaystyle{\phi(u)= \sum_{k\in\mathbb{N}} |u_k|\left(\log(|u_k|+1)\right)^{p_k} ,\,\,\,p_k\in (0,1)}.$
\end{enumerate}
We will show convergence of minimizers $u_\alpha$ to solutions of \eqref{equation} under additional conditions on $\phi_k$, that will be verified for most of the examples listed above.
Before that, some theoretical background on flexible penalties will be described.
Let us start with a Kadec-Klee (or Radon-Riesz) property for general functions \eqref{phik}.
\begin{proposition}\label{kadec_klee}
Let $\phi_{k}:[0,\infty)\to [0,\infty)$ satisfy (A1). If $\{u^n\}\subset \ell^{\phi_k}$ converges componentwise to $u\in\ell^{\phi_k}$ and converges also in the sense $\phi(u^n)\to \phi(u)$ as $n\to\infty$, then the following convergence holds: $\phi(u^n-u)\to 0$ as $n\to\infty$.
\begin{proof} The statement can be proven using Fatou's Lemma and taking into account that the functions $\phi_k$ are nonnegative - see the the similar proof of Lemma 2 in \cite{grasmair2008sparseregularization}.
\end{proof}
\end{proposition}
The following additional assumptions will be needed in the regularization analysis:
(A2) \[
\phi_k(t)\geq \frac{ct}{t+1},\,\,\,\forall t>0,\,\,\,\forall k\in\mathbb{N},
\]
for some $c>0$. This resembles condition (C3') in \cite{grasmair2010nonconvexsparse} imposed to the single function playing there the role of all these $\phi_k$.
(A3) The set $\{t\geq 0: \phi_k(t)\leq M, \,\forall k\in\mathbb{N}\}$ (where $M>0$) is bounded in the following sense:
There is $L>0$ (not depending on $k$) such that $0\leq t\leq L$, as soon as $\phi_k(t)\leq M$, for any $k\in\mathbb{N}$.
This condition means existence of a uniform bound for the sublevel sets of the functions $\phi_k$.
\begin{remark}\label{examples} Assumptions (A1)-(A3) hold for the following nonconvex functions $\phi_k$ (where (A1) is clearly satisfied for each example):
\begin{enumerate}
\item $\displaystyle{\phi(u)= \sum_{k\in\mathbb{N}} |u_k|^{p_k}}$, with $p_k\in (0,1]$, and denote $p=\inf_kp_k>0.$
Condition (A2) holds with $c=1$, since it amounts to the inequality $t^{p_k}+t^{p_k+1}\geq t$ that is true in both cases $t>1$ and $t\leq 1$.
Regarding (A3): $t^{p_k}\leq M$ for any $k\in\mathbb{N}$ means $t\leq M^{1/{p_k}}\leq M^{1/{p}}$ if $M>1$. Hence, $L=\max\{1, M^{1/{p}}\}$.
\item $\displaystyle{\phi(u)=\sum_{k\in\mathbb{N}} \log(|u_k|^{p_k}+1)}$.
We use the inequality $(a+b)^s\leq a^s+b^s$, for any $a,b>0$ and $s\in [0,1]$. Thus, we have, for any $k\in\mathbb{N}$ and $t>0$:
$\displaystyle{\log(t^{p_k}+1)\geq \log((t+1)^{p_k})=p_k\log(t+1)\geq p\log(t+1)\geq \frac{pt}{t+1}}$, which yields (A2) with $c=p$.
Condition (A3) is verified with $L=\max\{1, (e^M-1)^{1/{p}}\}$.
\item $\displaystyle{\phi(u)= \sum_{k\in\mathbb{N}} \log(|u_k|+1)+ |u_k|^{p_k},\,\,\,p_k\in (0,1)}$.
Here one can proceed by combining the previous two examples, thus yielding (A2) with $c=2$ and $L=\max\{e^M, e^M-1+ M^{1/{p}}\}$.
\item $\displaystyle{\phi(u)= \sum_{k\in\mathbb{N}} (|u_k|+|u_k|^{p_k})^{q_k},\,\,\,p_k, q_k\in (0,1).}$
If $\inf_kq_k=q>0$, then one has
$\displaystyle{ |u_k|^{q_k}\leq(|u_k|+|u_k|^{p_k})^{q_k}\leq |u_k|^{q_k}+|u_k|^{p_kq_k}\leq |u_k|^q+|u_k|^{pq}}$
due to the inequality $(a+b)^s\leq a^s+b^s$ for $a,b>0, s\in(0,1)$. Here one can take $c=1$ and $L=\max\{1, M^{1/{q}}\}+\max\{1, M^{1/{pq}}\}$.
\item $\displaystyle{\phi(u)= \sum_{k\in\mathbb{N}} |u_k|^{p_k}\log(|u_k|+1) ,\,\,\,p_k\in (0,1).}$
\item[(6)] $\displaystyle{\phi(u)= \sum_{k\in\mathbb{N}} |u_k|\left(\log(|u_k|+1)\right)^{p_k} ,\,\,\,p_k\in (0,1)}.$
In the last two examples, the functions $\phi_k$ behave like $t\mapsto t^{p_k+1}$, where now $p_k+1>1$. Minimizers of the corresponding Tikhonov functionals do exist (via coercivity with respect to the $\ell^2$ norm), however the sparsity of the minimizers is not anymore guaranteed.
\end{enumerate}
Note that the popular nonconvex regularizers SCAD \cite{SCAD} and MCP \cite{MCP} are not included in the list above since they do not satisfy the limit condition in (A1).
\end{remark}
The next result will guarantee coercivity of the regularization functional with respect to the $\ell^1$ norm.
\begin{lemma}\label{inclusion}
Let the functions $\phi_k$ verify (A1)-(A3). Then one has $\ell^{\phi_k}\subseteq\ell^1$
in the following sense:
There exists $C>0$ such that $\displaystyle{C\|u\|_1\leq \sum_k \phi_k(|u_k|)}$.
\end{lemma}
\begin{proof}
Let $u\in \ell^{\phi_k}$, that is $\sum_k\phi_k(|u_k|)<\infty$. Hence, there is $M>0$ such that $\phi_k(|u_k|)\leq M$, for any $k\in\mathbb{N}$. Condition (A3) implies existence of an $L>0$ satisfying $|u_k|\leq L$, for all $k\in\mathbb{N}$. By using condition (A2), we have:
\[
\frac{c}{1+L}\|u\|_1\leq \sum_k \frac{c|u_k|}{1+|u_k|}\leq \sum_k \phi_k(|u_k|).
\]
\end{proof}
\begin{comment}
Regarding the coercivity of $\phi$: One can deal with this via
\begin{itemize}
\item the $\ell^2$ topology.
\item the $\ell^1$ topology, in case one has $\ell^1=\ell^{\phi_k}$ (ideally!)
\end{itemize}
\end{comment}
\begin{corollary}
Let the functions $\phi_k$ verify (A1)-(A3). Then $\phi$ is coercive with respect to the $\ell^1$ norm.
\end{corollary}
In the sequel one can see that the minimizers of \eqref{reg_nonconvex} are sparse, whenever they exist.
\begin{proposition}
Let the functions $\phi_k$ satisfy (A1)-(A3). If there is a solution of \eqref{reg_nonconvex}, then it is sparse.
\end{proposition}
\begin{proof}
Let $x$ be a solution of \eqref{reg_nonconvex}. Then one has
\begin{equation}\label{in1}
\frac{1}{2}\|Ax-y\|^2+\alpha \phi(x)\leq \frac{1}{2}\|Au-y\|^2+\alpha \phi(u),
\end{equation}
for any $u\in \ell^{\phi_k}$.
Note that $\sum_n\phi_k(|x_k|)<\infty$ implies $|x_k|\leq L$ for all $k$ and for some $L>0$ (due to (A3)).
Fix
$i\in\mathbb{N}$ and define $u^i=x-x_ie_i$, where $e_i$ is the $i$-th vector of the canonical basis.
By using $u=u^i$ in \eqref{in1}, (A2) and the boundedness of $(x_n)$, we have
\[
\frac{c\alpha |x_i|}{L+1}\leq \alpha\phi_i(|x_i|)\leq \frac{x_{i}^2}{2}\|Ae_i\|^2-x_{i}\langle {Ax-y,Ae_i}\rangle.
\]
One can further write
\[
\frac{c\alpha |x_i|}{L+1}\leq \frac{|x_{i}|^2}{2}\|Ae_i\|^2 +|x_{i}|\,|\langle {Ax-y,Ae_i}\rangle|,
\]
which means that either $x_{i}=0$ or $x_{i}\neq 0$, in which case one can divide by $|x_{i}|$ and obtain
\[
0<\frac{c\alpha }{L+1}\leq \frac{|x_{i}|}{2}\|Ae_i\|^2 +|\langle {Ax-y,Ae_i}\rangle|=\frac{|x_{i}|}{2}\|Ae_i\|^2 +|\langle {A^*(Ax-y),e_i}\rangle|.
\]
If there are infinitely many $i$ with $x_{i}\neq 0$, then taking $i\to\infty$ above yields a contradiction. This is due to the fact that the right-hand side converges to zero as $\|x\|_1<\infty$ cf. Lemma \ref{inclusion}, and $A^*(Ax-y)=(\langle {A^*(Ax-y),e_i}\rangle)\in\ell^2$ since $A^*:Y\to\ell^2$. Therefore, $x$ has only finitely many nonzero components.
\end{proof}
\begin{remark}
In principle, one could also obtain estimates on the sparsity level of a solution $x$ of \eqref{reg_nonconvex} in a similar spirit to \cite[Remark 5.1]{bredies2014nonconvexminimization}, however, this would need a known lower bound on the size of the non-zero entries. In \cite{bredies2014nonconvexminimization}, this is known, since the penalty is $\sum_k|x_k|^p$ for some fixed $p<1$. In our case, our assumptions do not allow to conclude the existence of such a lower bound. What we get is, that every minimizer $x$ fulfills
\begin{align*}
\tfrac12\norm{b}^2 & = \tfrac12\norm{A0-b}^2 + \alpha\phi(0)\\
& \geq \tfrac12\norm{Ax-b}^2 + \alpha\phi(x)\\
& \geq \alpha\sum_{x_k\neq 0}\phi_k(|x_k|).
\end{align*}
This implies that
\begin{align*}
\tfrac12\norm{b}^2\geq \alpha\cdot\inf_{x_k\neq 0}\big[\phi_k(|x_k|)\big]\cdot\#\{k\mid x_k\neq 0\},
\end{align*}
and consequently, the number of nonzero entries in a minimizer $x$ is bounded from above by
\begin{align*}
\frac{\norm{b}^2}{2\alpha\inf_{x_k\neq 0}\big[\phi_k(|x_k|)\big]},
\end{align*}
if the infimum is non-zero. But this can't be guaranteed by our assumptions.
\end{remark}
We discuss next existence of solutions of \eqref{reg_nonconvex}.
\begin{proposition}
\label{exist_uniq} Let $A:\ell^2\rightarrow Y$ be a linear and bounded operator which is also $\ell^1$-weak$^*$-weak sequentially continuous, where $Y$ is a Hilbert space. Assume that the functions $\phi_{k}:[0,\infty)\to [0,\infty)$ verify (A1)-(A3). Then for any $\alpha>0$, the Tikhonov functional \eqref{reg_nonconvex}
has at least one minimizer $u_{\alpha}$ in $\ell^{\phi_k}$.
\end{proposition}
\begin{proof} One can show the statement by usual techniques, taking into account the coercivity of $\phi$ as above and the weak$^*$ lower semicontinuity of the involved functionals (the latter holds due to the componentwise convergence of weak$^*$ convergent subsequences and to the lower semicontinuity of each $\phi_k$).
\end{proof}
A convergence result for the proposed regularization method can be formulated in the following.
\begin{proposition}
Let $A:\ell^2\rightarrow Y$ be a linear and bounded operator which is also $\ell^1$-weak$^*$-weak sequentially continuous, where $Y$ is a Hilbert space. Assume that there is solution of $Au=y$ and that the functions $\phi_{k}:[0,\infty)\to [0,\infty)$ satisfy (A1)-(A3). If $\alpha_n\to 0$ as $n\to\infty$, then there is a subsequence $(u_{\alpha_{n_j}})$ of the sequence of minimizers $(u_{\alpha_{n}})$ of the problems
\[
\min_{u}\tfrac12\norm{Au-y}^{2} + \alpha_n \phi(u),\,\,\,
\]
which converges to a solution $\bar u$ of the operator equation in the sense that $\phi(u_{\alpha_{{n_j}}}-\bar u)\to 0$ as $j\to\infty$ (also in the $\ell^1$-norm).
\end{proposition}\medskip
A similar result holds also in the case that noisy data $y^\delta $ with $\|y-y^\delta\|\leq\delta $ is considered.
\section{Iterative majorization approaches via concave duality}
\label{sec:iterative_majorization}
In the sequel, we focus on numerical methods for the variational nonconvex problem \eqref{reg_nonconvex}.
Some notes on existing literature: \cite{wipf2010irls} analyses general reweighting schemes that lead to algorithms similar to the $\ell^{1}$ and $\ell^{2}$ algorithms here, but does not use the majorization approach. The work \cite{lai2013irls} analyses $\ell^{q}$ minimization with IRLS via $\ell^{1}$ and investigates exact recovery and convergence.
In this section, we consider an approach for minimizing least squares functions with nonconvex regularization terms which is based on iterative majorization of the nonconvex term. The approach has been (re)discovered several times. The first references we know about are \cite{Geman1992,Geman1995,Black1996} and date back to the early 1990s. An apparently independent discovery from the same time can be found in \cite{Gorodnitsky1997} under the name iterative reweighting. Recent rediscoveries are, e.g. \cite{Chan2014,Lanza2015}, while papers that consider convergence of the respective methods are, e.g., \cite{Nikolova2005,Allain2006}.
The idea to minimize a functional of the form
\[
\frac{1}{2}\norm{Ax-y}^{2} + \alpha\sum_{k=1}^{\infty}\phi_{k}(|x_{k}|)
\]
with functions $\phi_k:[0,\infty)\to[0,\infty)$ which fulfil $\phi_k(0)=0$ is as follows: Since the one-dimensional functions $\phi_{k}$ may not be convex, it is of interest to write each function as a minimum of convex functions. There are several possibilities to do this, as one can see below. One can start with using scaled and shifted quadratic functions centered at $0$, i.e.,
\begin{align}\label{eq:def-phik-quad}
\phi_{k}(t) = \inf_{s\geq 0}(st^{2} + \psi_{k}(s))
\end{align}
for some function $\psi_{k}$.
Alternatively, one can also write $\phi_{k}(t)$ as a minimum of scaled and shifted absolute values centered at $0$,
\begin{align}\label{eq:def-phik}
\phi_{k}(t) = \inf_{s\geq 0}(st + \tilde\psi_{k}(s)).
\end{align}
Figure \ref{fig:majorization-of-sqrt} shows the majorization with absolute values and quadratics for the function $\varphi(t)=\sqrt{\abs{t}}$.
\begin{figure}
\centering
\includegraphics{images/Majorization}
\caption{Majorization of the function $\varphi(t) =\sqrt{\abs{t}}$ by shifted and scaled absolute values (left) and quadratics (right).}
\label{fig:majorization-of-sqrt}
\end{figure}
In the first case~\eqref{eq:def-phik-quad}, the minimization problem becomes
\[
\min_{x,s\geq 0}\frac{1}{2}\norm{Ax-y}^{2}+ \sum_{k=1}^{\infty}(s_{k}|x_{k}|^{2} + \psi_{k}(s_{k})).
\]
The minimization over $x$ is now a quadratic problem and the minimization over $s$ decouples over $k$, thus yielding simple one-dimensional minimization problems. Note that the minimization over $s$ and $x$ simultaneously is still difficult, but alternating minimization is comparably simpler.
In the second case~\eqref{eq:def-phik},
one would get
\[
\min_{x,s\geq 0}\frac{1}{2}\norm{Ax-y}^{2}+ \sum_{k=1}^{\infty}(s_{k}|x_{k}| + \tilde\psi_{k}(s_{k})).
\]
The minimization over $x$ is an $\ell^1$-penalized least squares problem, for which a number of methods is available \cite{rao1997focuss,daubechies2003iteratethresh,lorenz2008conv_speed_sparsity,lorenz2008ssn,lorenz2008harditer,figueiredo2007gradproj,figueiredo2009sparsa,vandenberg2007spgl1}, while the minimization over $s$, again, decouples over $k$.
Note that equations~\eqref{eq:def-phik-quad} and~\eqref{eq:def-phik} resemble the definition of the convex conjugate (basically up to sign changes). For example, one can see that $\psi_k$ from~\eqref{eq:def-phik} fulfills $\psi_k(t) = -(-\phi_k)^*(-t)$, where $(-\phi_k)^*$ is the usual convex conjugate.
Since the multiple minus-signs can be confusing, we present in the following a special notion of duality in the case of concave increasing functions.
We did not find a reference that contains all the necessary results regarding the concave conjugate and the superdifferential notions for concave functions. Thus, a presentation of this theory is included for the sake of completeness. It will be largely along the lines of convex duality and hence, most proofs can be omitted.
We will
investigate the following set of functions:
\[
\mathrm{CI} = \{\psi:[0,\infty)\to[-\infty,\infty)\ \mid \psi\ \text{concave and increasing}\}.
\]
For $\psi\in \mathrm{CI}$, define
\[
\maj(\psi) = \{(t,l)\in [0,\infty)\times[-\infty,\infty)\ \mid st+l\geq \psi(s),\,\forall s\geq 0\}.
\]
The pairs $(t,l)\in\maj(\psi)$ parametrize the linear functions on $[0,\infty)$ which majorize $\psi$, where $t$ is its slope and $l$ is the intercept.
Note that this construction is analogous to the situation of the convex conjugate where one shows that a convex and lower semi-continuous function is exactly the supremum of all affine functions below it. This leads to the well known duality theory for convex functions. An adaptation to the situation of concave increasing functions provides us with the following results.
\begin{proposition}
For any subset $M\subset [0,\infty)\times[-\infty,\infty)$, it holds that the function
\[
\varphi(s) = \inf_{(t,l)\in M}st+l
\]
defined on $[0,\infty)$ is in $CI$.
Moreover, every function $\psi\in CI$ can be written as
\[
\psi(s) = \inf_{(t,l)\in\maj(\psi)}st+l.
\]
\end{proposition}
\begin{proof}
Most of the claim follows by analogy to the convex case. The only thing which is missing is the monotonicity, that can be seen as follows: If $s_{1}\leq s_{2}$, then (since $t\geq 0$)
$s_1t+l\leq s_{2}t+l$ for all $(t,l)\in M$. This shows that
$\varphi(s_{1})\leq \varphi(s_{2})$.
\end{proof}
\begin{definition}
For $\psi\in CI$, the concave conjugate is defined by
\[
\psi^{\oplus}(t) = \inf_{s\geq 0}\big(st-\psi(s)\big).
\]
\end{definition}
It is clear that $\psi^{\oplus}\in CI$.
\begin{proposition}\label{prop:double-conjugate}
For all $\psi\in CI$, one has ${\psi^{\oplus}}^{\oplus} = \psi$.
\end{proposition}
We introduce a notion of supergradient for functions in $CI$ (see also \cite{superdiff}):
\begin{definition}
For $\psi\in CI$, we say that $s^{*}\geq 0$ is a supergradient of $\psi$ at $s$, if for all $r\geq 0$ it holds that
\[
\psi(s) + s^{*}(r-s)\geq \psi(r).
\]
We denote the set of all such $s^*$ by ${\hat{\partial}}\psi(s)$ and call it superdifferential.
\end{definition}
Note that ${\hat{\partial}}\psi(s)=-{\partial}(-\psi)(s)$, where $\partial (-\psi)$ is the subdifferential of the convex function $-\psi$.
\begin{theorem}[Fenchel inequality and Fenchel equality]\label{thm:fenchel}
For $\psi\in CI$, one has
\[
\psi(s)+\psi^{\oplus}(s^{*}) \leq ss^{*},\quad \forall s\geq 0, s^*\geq 0.
\]
Moreover, $\psi(s)+\psi^{\oplus}(s^{*}) = ss^{*}$ if and only if $s^{*}\in\hat{\partial}\psi(s)$ (or, equivalently $s\in\hat{\partial}\psi^{\oplus}(s^{*})$).
\end{theorem}
\begin{proof}
The inequality follows from
\[
\psi(s)+\psi^{\oplus}(s^{*}) = \psi(s) + \inf_{r\geq 0}(rs^{*}-\psi(r) )\leq \psi(s) + ss^{*}-\psi(s).
\]
For the equality, note that $s^{*}\in\hat{\partial}\psi(s)$ is by definition equivalent to the inequality
\[
\psi(s) -ss^{*}\geq \psi(r) - s^{*}r,\ \text{for all $r\geq 0$}.
\]
In turn, this is equivalent to
\[
\psi(s) -ss^{*}\geq \sup_{r\geq 0}(\psi(r) - s^{*}r) = -\psi^{\oplus}(s^{*})
\]
which is the same as $\psi(s)+\psi^{\oplus}(s^{*})\geq ss^{*}$. The claim follows with Fenchel's inequality.
\end{proof}
As an example, we take $\psi(s) = s^p$ with $0<p<1$. The concave conjugate is
\begin{align*}
\psi^\oplus(t) = \inf_{s\geq 0}(st-\psi(s)) = \inf_{s\geq 0}(st-s^p).
\end{align*}
The infimum is attained at $s=(t/p)^{1/(p-1)}$ and the value evaluates to
\begin{align*}
\psi^\oplus(t) = \Big(\tfrac{1}{p^{1/(p-1)}} - \tfrac{1}{p^{p/(p-1)}}\Big)t^{p/(p-1)}.
\end{align*}
The special case $p=1/2$, for example, leads to $\psi^\oplus(t) = -1/(4t)$ and hence, by Proposition~\ref{prop:double-conjugate} we have the representation
\begin{align*}
\sqrt{s} = \inf_{t\geq 0}\left(st + \tfrac1{4t}\right).
\end{align*}
\section{Iteratively reweighted quadratic regularization}\label{it_T}
This section uses the concave conjugate to derive flexible majorization approaches which lead to simple minimization algorithms for the regularized problems. A noteworthy property of these methods is their monotonicity, in the sense that they decrease the objective function in every iteration. We consider iterative $\ell^2$-minimization (i.e., quadratic Tikhonov regularization) in an infinite dimensional setting and develop the concept into a full algorithmic framework. Moreover, the resulting monotone algorithm will be tested on two examples, showing its efficiency and accuracy.
To get a series of quadratic problems, we write our objective function as
\[
\min_{x}\norm{Ax-y}^{2}+ \alpha\sum_{k=1}^\infty\psi_k(\abs{x_{k}}^2).
\]
with $\psi_k\in\mathrm{CI}$ for all $k$.
Again, using concave conjugates, we get
\begin{equation}\label{quadr}
\min_{x\geq 0}\Big[\norm{Ax-y}^{2}+\alpha\sum_{k=1}^\infty s_{k}\abs{x_{k}}^2 - \psi_{k}^{\oplus}(s_{k})\Big].
\end{equation}
If we alternatively minimize for $x$ and $s$, we get
\begin{enumerate}
\item Initialize with some $x^{0}$ and $s^{0}$, set $n=0$ and iterate
\item Obtain $x^{n+1}$ by solving $A^*(Ax-y) + \alpha\diag(s^n)x = 0$
\item $s_k^{n+1}\in\hat\partial\psi_k(\abs{x_k^{n+1}}^2)$.
\end{enumerate}
The next lemma shows that the method is indeed of descent type and quantifies the guaranteed descent in each step.
\begin{lemma}\label{lem:descent}
Let $F(x) = \norm{Ax-y}^2 + \alpha\sum_k\psi_k(\abs{x_k}^2)$ and let $x^n$ be the iterates of the above iteration. Then it holds that
\[
F(x^{n+1}) \leq F(x^n) - \norm{Ax^{n+1} - Ax^n}^2 - \alpha\sum_ks_{k}^{n}(x_k^n-x_k^{n+1})^2.
\]
\end{lemma}
\begin{proof}
We write (using $A^*(Ax^{n+1}-y) = -\alpha\diag(s^n)x^{n+1}$)
\begin{align*}
F(x^n)-F(x^{n+1})& = \norm{Ax^n-y}^2 - \norm{Ax^{n+1}-y}^2 + \alpha\sum_k \psi_k(\abs{x_k^n}^2) -\psi_k(\abs{x_k^{n+1}}^2)\\
& =
\norm{Ax^n-Ax^{n+1}}^2 + \scp{x^n-x^{n+1}}{A^T(Ax^{n+1}-y)} + \alpha \sum_k \psi_k(\abs{x_k^n}^2) -\psi_k(\abs{x_k^{n+1}}^2)\\
& = \norm{Ax^n-Ax^{n+1}}^2 + + \alpha \sum_k \psi_k(\abs{x_k^n}^2) -\psi_k(\abs{x_k^{n+1}}^2) - s_k^nx_{k}^{n+1}(x_k^n-x_k^{n+1}).
\end{align*}
Now we use that $\psi_{k}(t^{2})\leq st^{2}-\psi_{k}^{\oplus}(s)$ with $t = x_{k}^{n+1}$ and $s=s_{k}^{n}$ to get
\begin{align*}
F(x^{n})-F(x^{n+1}) & \geq \norm{Ax^{n}-Ax^{n+1}}^{2} + \alpha\sum_{k}\psi_k(\abs{x_{k}^{n}}^{2}) - s_{k}^{n}(x_{k}^{n+1})^{2} + \psi_{k}^{\oplus}(s_{k}^{n}) - s_{k}^{n}x_{k}^{n+1}(x_{k}^{n}-x_{k}^{n+1}).
\end{align*}
By the Fenchel equality (Theorem ~\ref{thm:fenchel}) and since $s_{k}^{n}\in\hat\partial\psi_{k}(\abs{x_{k}^{n}}^{2})$ we have that $\psi_k(\abs{x_{k}^{n}}^{2}) + \psi_{k}^{\oplus}(s_{k}^{n}) = s_{k}^{n}\abs{x_{k}^{n}}^{2}$ and thus
\begin{align*}
F(x^{n})-F(x^{n+1}) & \geq \norm{Ax^{n}-Ax^{n+1}}^{2} + \alpha\sum_{k}s_{k}^{n}\abs{x_{k}^{n}}^{2}- s_{k}^{n}\abs{x_{k}^{n+1}}^{2} - s_{k}^{n}x_{k}^{n+1}(x_{k}^{n}-x_{k}^{n+1})\\
& = \norm{Ax^{n}-Ax^{n+1}}^{2} + \alpha\sum_{k}s_{k}^{n}(x_{k}^{n}-x_{k}^{n+1})^{2}.
\end{align*}
\end{proof}
\begin{lemma}
For the iterates $x^{n}$, $s^{n}$, it holds that
\begin{enumerate}[i)]
\item $(F(x^{n}))$ is convergent,
\item $\displaystyle{\sum_{n=0}^{\infty}\norm{A(x^{n+1}-x^{n})}^{2}\leq F(x^{0})}$,
\item $\displaystyle{\sum_{n=0}^{\infty}\sum_{k=1}^\infty s^{n}_{k}(x^{n+1}_{k}-x^{n}_{k})^{2}\leq
F(x^{0})/\alpha}$,
\item $\displaystyle{\norm{A(x^{n+1}-x^{n})}\to 0}$ for $n\to\infty$, and
\item $\displaystyle{\sum_{k=1}^\infty s^{n}_{k}(x^{n+1}_{k}-x^{n}_{k})^{2}\to 0}$ for $n\to\infty$.
\end{enumerate}
\end{lemma}
\begin{proof}
Summing up the inequalities from Lemma~\ref{lem:descent} for $n=0$ to $m$, we get
\[
F(x^{m+1})\leq F(x^{0}) - \sum_{n=0}^{m}\norm{A(x^{n+1}-x^{n})}^{2} - \alpha\sum_{n=0}^{m}\sum_k s^{n}_{k}(x^{n}_{k}-x^{n+1}_{k})^{2}
\]
from which the claim follows.
\end{proof}
The above lemma does not show convergence of the sequence of iterates yet. This could, in principle, be shown by adapting results from \cite{yu2019iteratively}. We do not pursue this here further, but instead, consider the special case of flexible penalties with varying exponents $p_k$ in the next section.
\section{A monotone algorithm for the regularized problem}\label{sec:optcond}
In this section, the previous approach by iterative $\ell^2$-regularization will be developed into a full algorithmic framework. For simplicity, we will focus on the case where all functions $\phi_k$ are of type $\phi_k(t) = t^{p_k}$ and
will propose an algorithm to solve the minimization problem
\begin{equation}\label{optprob2m}
\min_{x \in \ell^{p_k}}\frac{1}{2}||A x-y||^2+\alpha \sum_{k\in\mathbb{N}}|x_k|^{p_k},
\end{equation}
with
$\displaystyle{\ell^{p_k}:=\{x=(x_k):\, \sum_{k\in\mathbb{N}}|x_k|^{p_k}<\infty\}}$ and $(p_k)\subset (0,1)$. The behavior of the algorithm using other $\phi_k$ functionals will be discussed in the numerics subsection
\subsection{Convergence properties}
Assume that $\displaystyle{\inf_{k\in\mathbb{N}}p_k>0}.
We actually consider
a slightly different version of the quadratic approximation from \eqref{quadr}. That is, we minimize over $x$ the following regularized functional
\begin{equation}
\label{optprobeps2m}
J_\varepsilon(x)=\frac{1}{2}||Ax-y||^2+\alpha \sum_k \Psi_{\varepsilon, p_k}(|x_k|^2),
\end{equation}
where for $\varepsilon>0$, $t \geq 0$ and $p \in (0,1)$,
\begin{equation}\label{psieps}
\Psi_{\varepsilon,p}(t)= \left\{
\begin{array}{ll}
\frac{p}{2}\frac{t}{\varepsilon^{2-p}} \quad &\mbox{for }\,\, 0\leq t \leq \varepsilon^2\\
\noalign{\smallskip}
t^{\frac{p}{2}}-(1-\frac{p}{2})\varepsilon^p \quad & \mbox{ for }\,\, t \geq \varepsilon^2.
\end{array}
\right.\,
\end{equation}
\begin{figure}
\centering
\includegraphics{images/Psieps}
\caption{The function $\Psi_{\varepsilon,p}$ from \eqref{psieps} for different values of $\varepsilon$.}
\label{fig:my_label}
\end{figure}
That is,
\begin{equation}\label{J_e}
J_\varepsilon(x)=\frac{1}{2}||Ax-y||^2+\alpha \sum_{|x_k|\geq \varepsilon}\Big(|x_k|^{p_k} -(1-\tfrac{p_k}2)\varepsilon^{p_k} \Big)+\alpha \sum_{|x_k|\leq \varepsilon}\left( \frac{p_k}{2\varepsilon^{2-p_k}}|x_k|^2\right).
\end{equation}
Note that $\Psi_{\varepsilon,p}$ is well defined, differentiable, and concave. Clearly, $J_{\varepsilon}$ has minimizers in $\ell^2$.
The necessary optimality condition is given by
\begin{equation}
\label{optcondeps2m}
A^*Ax+\frac{\alpha p}{\max(\varepsilon^{2-p},|x|^{2-p})}x=A^*y,
\end{equation}
where the second addend is short for the sequence with component $\frac{\alpha p_k}{\max(\varepsilon^{2-p_k},|x_k|^{2-p_k})}x_k$.
Note that the optimality condition \eqref{optcondeps2m} identifies local minimizers.
In order to solve \eqref{optcondeps2m}, the following iterative procedure is considered:
\begin{equation}\label{iter2m}
A^*Ax^{i+1}+\frac{\alpha p}{\max(\varepsilon^{2-p},|x^{i}|^{2-p})}x^{i+1}=A^*y,\,{i\in\mathbb{N},}
\end{equation}
where nonlinear functions are applied component-wise in the second addend that has components $\frac{\alpha p_k}{\max(\varepsilon^{2-p_k},|x_k^{i}|^{2-p_k})}x_k^{i+1}$. {Note that \eqref{iter2m} is well defined.}
We have the following convergence result.
\begin{theorem}\label{monotdec2m}
For $\varepsilon>0$, let $(x^i)$ be generated by \eqref{iter2m}. Then $(J_{\varepsilon}(x^i))$ is strictly monotonically
decreasing, unless there exists some $i$ such that $x^i = x^{i+1}$, and $x^i$ satisfies the necessary optimality condition \eqref{optcondeps2m}. Moreover, every cluster point of $(x^i)$, of which there exists at least one, is a solution of \eqref{optcondeps2m}.
\end{theorem}
\begin{proof}
\begin{comment}
Since
\begin{equation*}
\end{equation*}
\end{comment}
The proof follows similar arguments to that of Theorem 4.1
in \cite{GK1} and of Lemma \ref{lem:descent}. For the sake of completeness, we sketch the main proof steps. By applying \eqref{iter2m} to $x^{i+1}-x^i$, we get
\begin{multline*}
\frac{1}{2}\|Ax^{i+1}\|^2-\frac{1}{2}\|Ax^{i}\|^2+\frac{1}{2}\|A(x^{i+1}-x^{i})\|^2+\alpha\left\langle \frac{ p}{\max(\varepsilon^{2-p},|x^{i}|^{2-p})}x^{i+1},x^{i+1}-x^{i}\right\rangle\\
=\langle A^*y,x^{i+1}-x^i\rangle.
\end{multline*}
Note that
\begin{equation}\label{eq111}
\left\langle \frac{ p}{\max(\varepsilon^{2-p},|x^{i}|^{2-p})}x^{i+1},x^{i+1}-x^{i}\right\rangle=\frac{1}{2}\sum_{k\in\mathbb{N}}\frac{p_k(|x^{i+1}_k|^2-|x^{i}_k|^2+|x^{i+1}_k-x_k^i|^2)}{\max(\varepsilon^{2-p_k},|x^{i}_k|^{2-p_k})}
\end{equation}
and
\begin{equation}\label{eq112}
\frac{1}{2} \frac{ p_k}{\max(\varepsilon^{2-p_k},|x^{i}_k|^{2-p_k})}(|x^{i+1}_k|^2-|x^{i}_k|^2)=\Psi_{\varepsilon,p_k}'(|x_k^i|^2)(|x^{i+1}_k|^2-|x^{i}_k|^2).
\end{equation}
Since $\Psi_{\varepsilon,p_k}$ is concave, we have
\begin{equation}\label{eq113}
\Psi_{\varepsilon,p_k}(|x^{i+1}_k|^2)-\Psi_{\varepsilon,p_k}(|x^{i}_k|^2)- \frac{1}{2} \frac{ p_k}{\max(\varepsilon^{2-p_k},|x^{i}_k|^{2-p_k})}(|x^{i+1}_k|^2-|x^{i}_k|^2)\leq 0.
\end{equation}
Then, using \eqref{eq111}, \eqref{eq112}, \eqref{eq113}, we get
\begin{equation}\label{eq114}
J_{\varepsilon}(x^{i+1})+\frac{1}{2}\|A(x^{i+1}-x^{i}\|^2+\frac{1}{2}\sum_{k\in\mathbb{N}}\frac{\alpha p_k}{\max(\varepsilon^{2-p_k},|x^{i}_k|^{2-p_k})}|x_k^{i+1}-x_k^i|^2\leq J_{\varepsilon}(x^{i}).
\end{equation}
From \eqref{eq114} and coercivity of $J_{\varepsilon}$ it follows that $(x^i_k) $ is bounded in $\ell^2$ and hence in $\ell^\infty$. This and $\displaystyle{\inf_{k\in\mathbb{N}}p_k>0}$ imply existence of a constant $\kappa>0$ such that
\begin{equation}\label{45}
J_\varepsilon(x^{i+1}) +\frac{1}{2}||A(x^{i+1}-x^i)||^2+\kappa||x^{i+1}-x^{i}||^2 \leq J_\varepsilon(x^i),
\end{equation}
from which we conclude the first part of the theorem.
From \eqref{45}, we deduce that
\begin{equation}\label{46}
\sum_i ||A(x^{i+1}-x^i)||^2+\kappa||x^{i+1}-x^{i}||^2 <\infty.
\end{equation}
Since $(x^i)$ is bounded in $\ell^2$, there exists a subsequence $(x^{i_l})$ and $\bar x \in \ell^2$ such that $(x^{i_l})$ converges weakly to $\bar x$ in $\ell^2$. By \eqref{46} we have that $\displaystyle{\lim_{l\to \infty}x_k^{i_{l+1}}=\lim_{l\to \infty}x_k^{i_{l}}=\bar x_k}$. Testing \eqref{iter2m} with $e_k, k=1, \dots,$ and passing to the limit with respect to $i$ in \eqref{iter2m}, we get that $\bar x$ is a solution to \eqref{optcondeps2m}.
\end{proof}
\begin{remark}
Theorem \ref{monotdec2m} shows that each cluster point is stationary. The extension to global convergence under the Kurdyka-Lojasiewicz (KL) property seems a difficult task. We did not succeed to recover such a convergence, since it is not clear how to prove that $\psi \circ J_\varepsilon$ (where $\psi$ is the KL function) is a Lyapounov function with decreasing rate close to $||x^{i+1}-x^i||$, which is the key point on which the KL argument relies. Take, for example, a (classical) reference for the proof of the convergence under the KL property: Theorem $8$ of paper \cite{A}. In the following inequality due to the KL property,
$$
\psi'(J_\epsilon(x^i))||\partial J_\epsilon(x^i)||\geq 1,
$$
one needs to estimate the lefthand side in terms of $||x^i-x^{i-1}||$.
In our case, we have
$$
\partial J_\epsilon(x^i)=A^*Ax^i-A^*y+\frac{\alpha p}{\max\{\epsilon^{2-p}, |x^i|^{2-p}\}}x^i,
$$
where
$$
A^*Ax^{i}-A^*y=-\frac{\alpha p}{\max\{\epsilon^{2-p}, |x^{i-1}|^{2-p}\}}x^{i}.
$$
Hence,
$$
\partial J_\epsilon(x^i)=\frac{\alpha p}{\max\{\epsilon^{2-p}, |x^i|^{2-p}\}}x^i-\frac{\alpha p}{\max\{\epsilon^{2-p}, |x^{i-1}|^{2-p}\}}x^{i}
$$
should be estimated in terms of $||x^i-x^{i-1}||$ for all $\epsilon>0$, which may not be possible.\\
Moreover we remark that the KL property has been proved for the $\ell^p, p\in [0,1)$ nonconvex penalty (see \cite{KLellp}, section $5$), but in the case of a flexible $\ell^{p_k}, (p_k) \in (0,1)$ penalty like ours, it is not straightforward (at least one would need some further assumptions on the sequence $p_k$, which we did not investigate in detail).
\end{remark}
\begin{remark}
Note that the same result holds true in finite dimension for the problem
\begin{equation}\label{optprob2mlambda}
\min_{x \in \mathbb{R}^d}\frac{1}{2}||A x-y||^2+\alpha \sum_k|(\Lambda x)_k|^{p_k},
\end{equation}
where $\Lambda \in \mathbb{R}^{m\times d}$. However, in order to have existence of a solution $\Lambda$ must satisfy some additional assumptions, e.g. $\mbox{Ker}(A) \cap \mbox{Ker}(\Lambda)=\{0\}$.
The presence of such kind of operator was thoroughly investigated in \cite{GK1}, where the problem is finite dimensional. One main focus of \cite{GK1} was an application in fracture mechanics, where the presence of the operator $\Lambda$ is crucial. In the present paper we are mostly interested in the analysis of the $p_k$-sequence approach, whose setting does not seem to provide any new insight into the fracture mechanics examples considered, e.g., in \cite{GK1}. Indeed, in those kind of examples the regularization term is active just on one component of the solution. However, even though in the present paper we do not focus on the analysis of the case of an operator $\Lambda \neq I$ inside the regularization term, we remark that in the numerical tests such kind of operator could be used.
\end{remark}
The following proposition establishes the convergence of minimizers of \eqref{optprobeps2m} to solutions of \eqref{optprob2m} as $\varepsilon$ goes to zero. The proof uses a technique similar to the one of Proposition 1 of \cite{GK1}.
As it is easily understandable from the above mentioned proof, we just underline that the following result gives an $\ell^2$-convergence of the sequence $(x_\varepsilon)_{\varepsilon>0}$ as $\varepsilon\to 0$ to any cluster point.
\begin{proposition}
Let $(x_\varepsilon)_{\varepsilon>0}$ be solutions to \eqref{optprobeps2m}. Then any cluster point of $(x_\varepsilon)_{\varepsilon>0}$ as $\varepsilon \to 0^{+}$, of which there exists at least one, is a solution of \eqref{optprob2m}.
\end{proposition}
\subsection{Numerical results}\label{numericsDeps}
In the sequel we investigate the performance of the monotone algorithm in practice. For this reason we will work in finite dimension. Our aims are showing the good performance of the $(p_k)$ sequences approach in the monotone scheme, and comparing these results to the ones of \cite{GK1} with $p$ fixed (and $\Lambda=I)$. Thus, we consider two problems that have been investigated also in \cite{GK1}: the first one is an academic example where the operator $A$ is an $M$ matrix, and the second one is a time-dependent optimal control problem.
We have tested two different kinds of penalty, first $\phi_k(t)=t^{p_k}$, and secondly $\phi_k(t)=\log(t^{p_k}+1)$. We describe our results in detail for the first case, and write some remarks on the second case, to underline the main differences.
Our findings show that the $(p_k)$ sequences approach with $p_k \in (0,1)$ proves to be at least as effective as the $p$ fixed one (with $p \in (0,1)$), and in some cases, even more effective. Indeed, the $(p_k)$ sequence approach is more flexible in the sense that it enables to concentrate the "sparsity" constraint where it is more needed. We have exploited this feature in the control example in the way we have chosen the sequence $(p_k)$.
In this example, we are able to localise easily, a priori, an approximate region where the solution is known to be zero. Therefore we "relax" the sparsity constraint by choosing sequences $(p_k)$ which are closer to $1$ in those regions. In the areas where the solution is expected to be nonzero, we let the sequence tend towards the value used in \cite{GK1} in the same situation.
Choosing sequences as described above simplifies the scheme's performance. Consistently with our expectations, the $(p_k)$ sequence approach shows a smaller number of iterations and a lower residue as compared to the approach of \cite{GK1} with $p$ fixed (see Table \ref{tablecontrol1} and Table \ref{tablecontrol2}).
Moreover, the $(p_k)$ sequences approach provides better sparsity as compared to the approach of \cite{GK1} with $p$ fixed (see Table \ref{tablecontrol1} and Table \ref{tablecontrol2}).
The situation is a bit different in the academic $M$ matrix example, since we are not able to easily know a priori information on the behaviour of the solution. However, when testing different types of $(p_k) \subset (0,1)$ sequences, we have observed more sparsity for sequences $p_k \to p$ where $p$ is the value used in \cite{GK1}. For this reason, we will choose this type of sequences when comparing with the results of \cite{GK1}.
The aim of this experiment is to show the good performance of the $(p_k)$ sequence approach in an academic example (with a residue always $O(10^{-9})$), and to show the different behaviour of the solution when changing the sequence $(p_k)\subset (0,1)$.
For convenience of the exposition, we write the algorithm in the following form, see \textbf{Algorithm $1$} for $\phi_k(t)=t^{p_k}$.
A continuation strategy with respect to the parameter $\varepsilon$ is performed. In each of the following example, the initialization and range of $\varepsilon$-values will be described.
The key modifications in the case $\phi_k(t)=\log(t^{p_k}+1)$ are described in Remark \ref{rem:algphilog}.
The algorithm stops when the $\ell^\infty$-norm of the residue of \eqref{optcondeps2m} is $O(10^{-8})$ in the M Matrix example and $O(10^{-15})$ in the time control problem example. At this instance, the $\ell^2$-residue is typically much smaller. Therefore, we find an approximate solution of the $\varepsilon$-regularized optimality condition \eqref{optcondeps2m}.
The system in \eqref{systemalg} is solved through the MATLAB function \textit{mldivive} (that is, the \textit{backslash} command).
The initialization $x^0$ is chosen as the solution of the problem \eqref{optprob2m} where the $\ell^{p_k}$-term is replaced by the $\ell^2$-norm, that is,
\begin{equation}\label{initmoneps}
x^0=(A^*A+2\alpha)^{-1}A^*y.
\end{equation}
Inspired by the findings of \cite{GK1}, we have observed that for some values of $\alpha$ the previous initialization is not suitable (that is, the obtained residue is too big). In order to overcome the problem, we have successfully tested a continuation strategy with respect to increasing $\alpha$-values.
\begin{algorithm}[ht]
\caption{Monotone algorithm + $\varepsilon$-continuation strategy}
\begin{algorithmic}[1]
\STATE Initialize $\varepsilon=\varepsilon^0$, $x^0$ and $i=0$;
\REPEAT
\STATE Solve for $x^{i+1}$
\begin{equation}\label{systemalg}
A^*Ax^{i+1}+\frac{\alpha p}{\max(\varepsilon^{2-p},|x^{i}|^{2-p})} x^{i+1}=A^*y,
\end{equation}
where the second addend is short for the vectors with component $\frac{\alpha p}{\max(\varepsilon^{2-p_k},|x^i_k|^{2-p_k})}x^{i+1}_k$.
\STATE Set $i=i+1$.
\UNTIL{the stopping criterion is fulfilled}.
\STATE Reduce $\varepsilon$ and repeat 2.
\end{algorithmic}
\end{algorithm}
\begin{remark}\label{rem:algphilog}
When $\phi_k(t)=\log(t^{p_k}+1)$, the regularization \eqref{optprobeps2m} becomes
$$
J_\varepsilon(x)=\frac{1}{2}||Ax-y||^2+\alpha \sum_{k\in \mathbb{N}} \log(\Psi_{\varepsilon, p_k}(|x_k|^2)+1),
$$
and equation \eqref{systemalg} written in the following compact form becomes
$$
(A^*A+FM)x^{i+1}=A^*y,
$$
where $F$ is the diagonal matrix with $k$ component $\frac{1}{\Psi_{\varepsilon, p_k}(|x^i_k|^2)+1}$ and $M$ is the diagonal matrix with $k$-component $\frac{\alpha p_k}{\max(\varepsilon^{2-p_k},|x_k^{i}|^{2-p_k})}$.
\end{remark}
Note that, in the exposition of the numerical results, the total number of iterations shown in the tables takes into account the continuation strategy with respect to $\varepsilon$, but it does not take into account the continuation with respect to $\alpha$.
Finally we underline that in all the tests described in the following, the value of the objective functional for each iteration was tested to be monotonically decreasing. Note that this is consistent with the result of Theorem \ref{monotdec2m}.
In the sequel we will adopt the following notation. For $x \in \mathbb{R}^d$ we will denote $|x|_0=\#\{k\, :\, |x_k|> 10^{-10}\},$ $|x|_0^c=\#\{k\, :\, |x_k|\leq 10^{-10}\},$ and by $||x||$ the euclidean norm of $x$.
\subsubsection{M-Matrix example}
Here the monotone scheme will be tested for the M-matrix example. This is a classical academic example that we have chosen to confirm the good performance of our scheme for different values of the sequence $(p_k) \subset (0,1]$. More specifically, we consider
\begin{equation}
\label{optprobM2M}
\min_{x \in \mathbb{R}^{d^2}}\frac{1}{2}||A x-y||^2+\alpha \sum_{k=1}^{d^2}|x_k|^{p_k},
\end{equation}
where $A$ is the backward finite difference gradient
\begin{equation}\label{A}
A=(d+1)\left(\begin{array}{c} G_1\\G_2\end{array}\right),
\end{equation}
with $G_1 \in \mathbb{R}^{d(d+1)\times d^2}, G_2 \in \mathbb{R}^{d(d+1)\times d^2}$ given by
$$
G_1=I \otimes D, \quad G_2=D \otimes I.
$$
Here $I$ is the $d\times d$ identity matrix, $\otimes$ denotes the tensor product and $D\in \mathbb{R}^{(d+1)\times d}$ is given by
\begin{equation}\label{D}
D=\left(\begin{array}{ccccc}
1& 0& 0& \cdots& 0\\ -1& 1& 0& \cdots& 0\\ \vspace{0.2cm}\\ 0& \cdots& 0&-1&1\\0&\cdots&0&0&-1
\end{array}\right).
\end{equation}
Then $A^* A$ is an $M$ matrix coinciding with the $5$-point star discretization on a uniform mesh on a square of the Laplacian with Dirichlet boundary conditions. Note that \eqref{optprobM2M} can be equivalently expressed as
\begin{equation}
\label{optprob2}
\min_{x \in \mathbb{R}^{d^2}}\frac{1}{2}\|A x\|^2-\langle x,f\rangle +\alpha \sum_{k=1}^{d^2}|x_k|^{p_k},
\end{equation}
where $f=A^* y$. If $\alpha=0$, this is the discretized variational form of the elliptic equation
\begin{equation}
\label{elleq}
-\Delta z=f \mbox{ in } \Omega, \quad z=0 \mbox{ on } \partial \Omega.
\end{equation}
For $\alpha>0$, the variational problem \eqref{optprob2} gives a solution with piecewise constant enhancing behaviour.
We choose $f$ as in \cite{GK1}, subsection $3.4$, namely $f$ is a discretization of
$f(x_1,x_2)=10 x_1\mbox{sin}(5x_2) \mbox{cos}(7 x_1)$.
The parameter $\varepsilon$ varies in the same range as in \cite{GK1}, that is, it was initialized with $10^{-1}$ and decreased to $10^{-6}$.
In Table \ref{tableMMatrix} we show the performance of \textbf{Algorithm $1$} for $p\to 0.1$, $h=1/64$ as mesh size and $\alpha$ incrementally increasing by factor of $10$ from $10^{-4}$ to $10$. The sequence $(p_k)$ is approximated by a vector of $N=63^2$ points with $p_0=1.1$, generated by the MATLAB command $(p_k)=0.1+1./P$ with $P=linspace(1, 100, N)$.
In Figure \ref{fig:MMatrix} we show the graphics of the solutions for different values of $\alpha$ for which most changes occur.
One can observe significant differences with respect to different values of $\alpha$. The third row of Table \ref{fig:MMatrix} shows that $|x|^c_0$ increases with $\alpha$, as expected. For example, for $\alpha=1, 10$, we have $|x|^c_0=3969$, or equivalently, $|x|_0=0$, that is, the solution to \eqref{optprob2} is constantly zero. Moreover, we see that $| x|^p_p$ decreases when $\alpha$ increases (see the fourth row).
In the fifth row we show the $\ell^\infty$ norm of the residue, which is $O(10^{-9})$ for all the considered $\alpha$.
Note also that the number of iterations is sensitive with respect to $\alpha$, in particular it increases for $\alpha$ increasing from $10^{-4}$ to $10^{-1}$ and it decreases consistently for $\alpha=1,10$.
The algorithm was also tested for different values of $(p_k) \subset (0,1)$. The results obtained show dependence on $(p_k)$. In particular in Figure \ref{fig:MMatrix2} we show the graphics of the solution for $(p_k)$ a random sequence, that is a vector of $N=63^2$ uniformly distributed random numbers in the interval $(0,1)$, obtained by the command \textit{rand} in MATLAB. Comparing Figure \ref{fig:MMatrix2} B) to Figure \ref{fig:MMatrix} B), we see that for $\alpha=0.1$ the solution is less sparse as compared to the solution for $p_k \to 0.1$.
For any $k$, we say that $k$ is a singular component of the vector $ x$ if $k \in\{k \, :\, |x_k|<\varepsilon\}$. In particular, note that the singular components are the ones where the $\varepsilon$-regularization is most influential. In the sixth row of the tables we show $Sp$, which denotes the number of the singular components of $x$. Note that it coincides with the quantity $|x|_0^c$, which reassures the validity of the $\varepsilon$-strategy.
Moreover, we tested sequences $p_k \to 0$, but no significant improvement was remarked with respect to the above explained results. In part, this seems coherent with the results of \cite{GK1} where $p=0$ was not tested successfully.
We tested the algorithm when $\phi_k(t)=\log(t^{p_k}+1)$, according to the modification shown in Remark \ref{rem:algphilog}. The results are shown in Figure \ref{fig:MMatrixphi} only in the case $\alpha=0.1$ and $\alpha=1$, since no significant changes are observed for smaller values of $\alpha$ with respect to $\alpha=0.1$. Comparing Figure \ref{fig:MMatrixphi} a) with Figure \ref{fig:MMatrix} b) and Figure \ref{fig:MMatrixphi} b) with Figure \ref{fig:MMatrix} c), the solution is less sparse.
Moreover, the number of iterations and the residue are much higher with respect to $\phi_k(t)=t^{p_k}$, respectively of the order $10^3$ and $10^{-1}$.
Finally, we remark that if we modify the initialization \eqref{initmoneps}, the method converges to the same solution with no remarkable modifications in the number of iterations, which is a sign for the global nature of the algorithm.
\begin{table}[ht]
\captionof{table}{$M$-matrix example, $p_k \to0.1$, mesh size $h=\frac{1}{64}$. Results obtained by \textbf{Algorithm $1$}.}
\centering
\begin{tabular}{|l|c|c|c|c|c|c|c|}
\hline\noalign{\smallskip}
{\bf $\alpha$ } &$10^{-4}$& $10^{-3}$ &$10^{-2}$&$10^{-1}$ & $1$&10 \\
\noalign{\smallskip}\hline\noalign{\smallskip}
no. of iterates & 163&785& 2046&1984& 131& 15 \\
{\bf $|x|^c_0$ } &7 &41& 269&2617& 3969&3969\\
{\bf $| x|^p_p$ }& $2*10^3$&$2*10^3$&$ 2* 10^{3}$&$812$&9.6&7.3\\
$\mbox{ Residue }$ & $8*10^{-10}$ & $8*10^{-9}$ & $ 9*10^{-9}$ & $4*10^{-9}$ & $3*10^{-15}$&$10^{-15}$ \\
$Sp$ & 7&41&269& 2617& 3869&3969\\
\noalign{\smallskip}\hline
\end{tabular}
\label{tableMMatrix}
\end{table}
\begin{figure}[ht]
\centering
\subfloat[$\alpha=0.01$]
{
\includegraphics[height=5cm, width=4.5cm]{MmatrixRLG2.jpg}
}
\subfloat[$\alpha=0.1$]
{
\includegraphics[height=5cm, width=4.5cm]{MMatrixRLG3.jpg}
}
\subfloat[$\alpha=1$]
{
\includegraphics[height=5cm, width=4.5cm]{MMatrixRLG4.jpg}
}
\caption{Solution of the M-matrix problem, $p_k\to^+ 0.1,$ mesh size $h=\frac{1}{64}$. Results obtained by \textbf{Algorithm $1$}.}
\label{fig:MMatrix}
\end{figure}
\begin{figure}[ht]
\centering
\subfloat[$\alpha=0.01$]
{
\includegraphics[height=5cm, width=4.5cm]{MMatrixRLGpto02.jpg}
}
\subfloat[$\alpha=0.1$]
{
\includegraphics[height=5cm, width=4.5cm]{MMatrixRLGpto03.jpg}
}
\subfloat[$\alpha=1$]
{
\includegraphics[height=5cm, width=4.5cm]{MMatrixRLGpto04.jpg}
}
\caption{Solution of the M-matrix problem, $p_k \in (0,1]$ random sequence, mesh size $h=\frac{1}{64}$. Results obtained by \textbf{Algorithm $1$}.}
\label{fig:MMatrix2}
\end{figure}
\begin{figure}[ht]
\centering
\subfloat[$\alpha=0.1$]
{
\includegraphics[height=5cm, width=4.5cm]{MMatrixphi33.jpg}
}
\subfloat[$\alpha=1$]
{
\includegraphics[height=5cm, width=4.5cm]{MMatrixphi43.jpg}
}
\caption{Solution of the M-matrix problem, $p_k \subset (0,1], p_k \to 0.1$, mesh size $h=\frac{1}{64}$. Results obtained by \textbf{Algorithm $1$} as modified in Remark \ref{rem:algphilog}.}
\label{fig:MMatrixphi}
\end{figure}
\subsubsection{Time dependent control problem}\label{subsec:control}
The following example, of particular significance for our findings, is taken from \cite{GK1}, subsection $3.2$, to which we refer for further details. For the sake of completeness, we explain in the sequel the setting that we study.
We consider the linear control system
$$
\frac{d}{dt} z(t)=\mathcal{A} z(t)+D u(t), \quad z(0)=0,
$$
that is,
\begin{equation}\label{LCSfinalstate}
z(T)=\int_0^T e^{\mathcal{A}(T-s)} D u(s) ds,
\end{equation}
where the linear closed operator $\mathcal{A}$ generates a $C_0$-semigroup $e^{\mathcal{A}t}$, $t\geq 0$ on the state space $X$. More specifically, we consider the one-dimensional controlled heat equation for $z=z(t,x)$:
\begin{equation}\label{actionscontrol}
z_t=z_{xx}+d_1(x)u_1(t)+d_2(x)u_2(t), \quad x \in (0,1),
\end{equation}
with homogeneous boundary conditions $z(t,0)=z(t,1)=0$ and thus $X=L^2(0,1)$. The differential operator $\mathcal{A}z=z_{xx}$ is discretized in space by the second order finite difference approximation with $n=49$ interior spatial nodes ($\Delta x=\frac{1}{50}$). We use two time dependent controls $\overrightarrow u=(u_1,u_2)$ with corresponding spatial control distributions $d_i$ chosen as step functions:
$$
d_1(x)=\chi_{(.2,.3)}, \quad d_2(x)=\chi_{(.6,.7)}.
$$
The control problem consists in finding the control function $\overrightarrow u$ that steers the state $z(0)=0$ to a neighbourhood of the desired state $z_T$ at the terminal time $T=1$. We discretize the problem in time by the mid-point rule and we define the matrix $A$ as follows
\begin{equation}\label{midpoint}
A \overrightarrow u=\sum_{k=1}^m e^{\mathcal{A}\left(T-t_{k}-\frac{\Delta t}{2}\right)} (D \overrightarrow u)_k \Delta t,
\end{equation}
where $\overrightarrow u=(u_1^1,\cdots, u_1^m,u_2^1,\cdots u_2^m)$ is a discretized control vector whose coordinates represent the values at the mid-point of the intervals $(t_k,t_{k+1})$. Note that in \eqref{midpoint} we denote by $D$ a suitable rearrangement of the matrix $D$ in \eqref{LCSfinalstate} with some abuse of notation. A uniform step-size $\Delta t=\frac{1}{50}$ ($m=50$) is utilized. More specifically, we apply our scheme to the discretized optimal control problem in time and space where $x$ from \eqref{optprob2m} is the discretized control vector $u \in \mathbb{R}^{2m}$ which is mapped by $A$ to the discretized output $z$ at time $1$ by means of \eqref{midpoint} and
$y$ in \eqref{optprob2m} is the discretized target function chosen as the Gaussian distribution $z_T(x)=0.4\,\mbox{exp}(-70(x-.7)^2))$ centered at $x=.7$.
The parameter $\varepsilon$ was initialized with $10^{-3}$ and decreased down to $10^{-8}$.
Since the second control distribution is well within the support of the desired state $z_T$, the authority of this control is expected to be stronger than that of the first one, which is away from the target.
The structure of the example allows us to exploit the flexibility of the $(p_k)$ sequences approach since we know the two regions where the distributions of the control are active. Therefore we can "relax" the sparsity constraint by choosing sequences $(p_k)$ which are closer to $1$ in those regions, whereas in the areas where the solution is expected to be zero, we let the sequence tends towards the value used in \cite{GK1} in the same situation.
More specifically, in Tables \ref{tablecontrol1} and \ref{tablecontrol2} we report the results of our tests for $p=0.5$
and $(p_k)$ approximated by a vector of $N=100$ points with $p_0=0.51$ generated by the MATLAB command $(p_k)=flip(0.5+1./P)$ with $P=linspace(2, 100, N)$, respectively. In both tables, $\alpha$ is incrementally increasing by factor of $10$ from $10^{-2}$ to $1$. We report only the values for the second control $u_2$ since the first control $u_1$ is always zero. Note that the residue is always of the order of $10^{-15}$.
Moreover, the quantity $|u_2|_p^p$ decreases for $\alpha$ increasing, as expected. In Table \ref{tablecontrol1} we see that $|u_2|_0$ is always zero except that for $\alpha=10^{-2}$, whereas in Table \ref{tablecontrol2} for $\alpha=10^{-1}$ it is different from zero. We remark , for smaller values of $\alpha$, we find values of $|u_2|_0$ different from zero (consistent with our expectation) also whit the $(p_k)$ sequence. However, here we choose these values of $\alpha$ since we focus on the comparison between $(p_k)$ and $p$ fixed and for these values of $\alpha$ we found the most interesting results. Consistently with our expectations, for each value of $\alpha$ the number of iterations and the residue are always smaller with the $(p_k)$ sequence than for $p=0.5$ fixed. As an example, we refer to the first rows of Table \ref{tablecontrol1} and Table \ref{tablecontrol2} and we remark that the same holds true for different values of $\alpha$ than those of the above mentioned tables. As mentioned above, we believe that this improvement, as compared to the solution given by the $p$ fixed approach, resides in the fact that with $(p_k)$ sequences we are able to give a "different weight" to each component of the solution, thus simplifying the performance of the algorithm.
Moreover, we found some interesting improvements even with respect to the level of sparsity in the solution. In this regard, we underline that both $|u_2|_0^c$ and $|u_2|^{p_k}$ are smaller for $(p_k)$ (see second and third rows of Table \ref{tablecontrol1}) than when $p=0.5$ is fixed ( see Table \ref{tablecontrol2}).
In the sixth row of the tables we show the number of singular components of the vector $u_2$ at the end of the $\varepsilon$-path following scheme, that is, $Sp:=\#\{ i\,\, |\,\, |(u)_2)_i|<\varepsilon\}$. For all considered values of $\alpha$, we have that $Sp$ is the same as $|u_2|_0^c$, thus confirming the effectiveness of the $\varepsilon$-strategy.
We tested the algorithm when $\phi_k(t)=\log(t^{p_k}+1)$, according to the modification shown in Remark \ref{rem:algphilog}. The results reported in Table \ref{tablecontrol3} show that the algorithm works well enough in this case, too. However, for the tested value of $(p_k)$, the performance is not as good as in the case where $\phi_k(t)=t^{p_k}$. Indeed, the residue is always bigger, the sparsity level of the solution is comparable but slightly smaller, and the number of iterates is slightly higher.
Finally, we remark that, if we change the initialization \eqref{initmoneps}, the method converges to the same solution with no remarkable modifications in the number of iterations.
\begin{table}[tbhp]\caption{Sparsity in a time-dependent control problem, $ (p_k)$, mesh size $h=\frac{1}{50}$. Results obtained by \textbf{Algorithm $1$}.}\label{tablecontrol1}
\centering
\begin{tabular}{|l|c|c|c|c|}
\hline
{\bf $\alpha$ } &$10^{-2}$&$10^{-1}$&$1$\\
\hline
no. of iterates &28&33&19\\
\hline
{\bf $|u_2|_0^c$ } &99&100&100 \\
\hline
{\bf $|u_2|^p_p$ }&6&$10^{-6}$&$10^{-7}$ \\
\hline
$\mbox{ Residue }$ &$10^{-17}$ &$10^{-17}$&$10^{-18}$\\
\hline
$\mbox{Sp}$ &99&100&100\\
\hline
\end{tabular}
\end{table}
\begin{table}[tbhp]\caption{Sparsity in a time-dependent control problem, $ p=0.5$, mesh size $h=\frac{1}{50}$. Results obtained by \textbf{Algorithm $1$}.}\label{tablecontrol2}
\centering
\begin{tabular}{|l|c|c|c|c|}
\hline
{\bf $\alpha$ } &$10^{-2}$&$10^{-1}$&$1$\\
\hline
no. of iterates &35&60&23\\
\hline
{\bf $|u_2|_0^c$ } & 99&99&100 \\
\hline
{\bf $|u_2|^p_p$ }& 6.3&4.7&$10^{-6}$ \\
\hline
$\mbox{ Residue }$ &$10^{-17}$ &$10^{-11}$&$10^{-16}$\\
\hline
$\mbox{Sp}$ &99&100&100\\
\hline
\end{tabular}
\end{table}
\begin{table}[tbhp]\caption{Sparsity in a time-dependent control problem, $ (p_k)$, mesh size $h=\frac{1}{50}$. Results obtained by \textbf{Algorithm $1$} modified as in Remark \ref{rem:algphilog}.}\label{tablecontrol3}
\centering
\begin{tabular}{|l|c|c|c|c|}
\hline
{\bf $\alpha$ } &$10^{-2}$&$10^{-1}$&$1$\\
\hline
no. of iterates &51&51&50\\
\hline
{\bf $|u_2|_0^c$ } &97&98&98 \\
\hline
{\bf $|u_2|^p_p$ }&10&$1.9$&$0.7$ \\
\hline
$\mbox{ Residue }$ &$10^{-1}$ &$10^{-2}$&$10^{-2}$\\
\hline
$\mbox{Sp}$ &97&98&98\\
\hline
\end{tabular}
\end{table}
\section{Iteratively reweighted $\ell^1$-minimization}\label{1-section}
The analysis for the reweighted $\ell^2$-minimization cannot be directly adapted to the $\ell^1$-minimization case, because of the absolute value function singularity. Hence, the setting of this section is finite dimensional. The reader is referred also to \cite{chen2014convergence}, where the fixed exponent penalty case is dealt with.
If the problem under consideration is
\[
\min_{x}\frac{1}{2}\norm{Ax-y}^{2}+ \alpha\sum_{k=1}^N\phi_{k}(\abs{x_{k}})
\]
with $\phi_{k}\in CI$ for all $k$, then we can use the concave conjugates of the $\phi_{k}$ functions to formulate the problem as
\begin{equation}\label{alt1}
\min_{x}\inf_{s_{k}\geq 0}\Big[G(x,s) = \frac{1}{2}\norm{Ax-y}^{2}+\alpha\sum_{k=1}^N (s_{k}\abs{x_{k}} - \phi_{k}^{\oplus}(s_{k}))\Big].
\end{equation}
Since $\phi_{k}$ is concave, $G$ in this problem is convex in $x$ and also convex in all the $s_{k}$ (jointly), but not convex in $x$ and $s_{k}$ jointly. However, we can perform alternating minimization:
\begin{enumerate}
\item Initialize with some $x^{0}$ and $s^{0}$, set $n=0$ and iterate
\item $\displaystyle{x^{n+1}\in\argmin_{x}G(x,s^{n})}$
\item $\displaystyle{s^{n+1}\in\argmin_{s\geq 0} G(x^{n+1},s)}$.
\end{enumerate}
The minimization for $s$ decouples into one-dimensional problems
$\displaystyle{\min_{s}(s\abs{x_{k}} - \phi_{k}^{\oplus}(s))}$ with optimality conditions
$0 \in\abs{x_{k}}+\partial (-\phi_{k}^{\oplus})(s)=\abs{x_{k}}- \hat{\partial}\phi_{k}^{\oplus}(s)$. By
Theorem~\ref{thm:fenchel}, this holds if and only if
$s\in \hat{\partial}\phi_{k}(\abs{x_{k}})$. We obtain the following method consisting of a sequence of convex
$\ell^1$-minimization problems:
Initialize with some $x^{0}$ and $s^{0}$, set $n=0$ and iterate
\begin{align}\label{reweighted_conv_gen}
x^{n+1}\in\argmin_{x}\frac{1}{2} \norm{Ax-y}^2 + \alpha\sum_{k=1}^N s_k^n\abs{x_k}
\end{align}
with
\begin{equation}\label{eps}
s_{k}^{n+1}\in\hat{\partial}\phi_{k}(\abs{x_{k}^{n+1}}+\varepsilon).
\end{equation}
Here, we allow for a shift $\varepsilon\geq 0$, since one may run into problems when $\hat{\partial}\phi_{k}(t)$ grows unbounded near zero (but still include $\varepsilon=0$ to be able to treat the general case).
Thus, we will work with a shifted version of the regularized problem, that is
\begin{equation}\label{shifted}
F(x) = \frac{1}{2}\norm{Ax-y}^{2}+ \alpha\sum_{k=1}^N\phi_{k}(\abs{x_{k}}+\varepsilon),\quad \varepsilon\geq 0 .
\end{equation}
We start the analysis of the method with a proposition on descent:
\begin{proposition}\label{objective_mon}
Assume that the functions $\phi_k$ are in $\mathrm{CI}$ and differentiable on $(0,\infty)$. Let
$(x^{n})$ be generated by the algorithm \eqref{reweighted_conv_gen}-\eqref{eps}.
Then it holds
\begin{equation}\label{decrease}
F(x^{n+1})\leq F(x^n)-\frac{1}{2}\norm{Ax^{n+1}-Ax^n}^2-\alpha D(x^{n+1},x^n),\,\forall n\in\mathbb{N},
\end{equation}
where $D(x^{n+1},x^n)$ is the sum of Bregman distances associated to the convex functions $-\phi_k$, that is
\[\displaystyle{\sum_{k=1}^N D_{-\phi_k}} (|x_k^{n+1}|+\varepsilon,|x_k^n|+\varepsilon) = \sum_{k=1}^{N}\phi_{k}(\abs{x_{k}^{n}}+\varepsilon) - \phi_{k}(\abs{x_{k}^{n+1}}+\varepsilon) - s_{k}^{n}(\abs{x_{k}^{n}}-\abs{x_{k}^{n+1}})
\]
for some $s_{k}^{n}\in\hat\partial\phi_k(|x_k^n|+\varepsilon)$.
\end{proposition}
\begin{proof}
Due to the optimality condition for $x^{n+1}$, there exists $\xi_k^{n+1}\in \partial (|\cdot|)(x^{n+1}_k)$ such that
\begin{equation}\label{optim}
(A^*(Ax^{n+1}-y))_k+\alpha s_{k}^{n}\xi_k^{n+1}=0,\,\, 1\leq k\leq N,
\end{equation}
for some $s_{k}^{n}\in\hat\partial\phi_{k}(\abs{x_{k}^{n}}+\varepsilon)$.
Based on this, one has
\begin{eqnarray}\label{long}
F(x^n)-F(x^{n+1}) &=& \frac{1}{2}\|Ax^n-Ax^{n+1}\|^2+\langle Ax^n-Ax^{n+1},Ax^{n+1}-y\rangle\\
&+& \alpha\sum_{k=1}^N(\phi_k(|x_k^n|+\varepsilon)-\phi_k(|x_k^{n+1}|+\varepsilon))\nonumber\\
&=&\frac{1}{2} \|Ax^n-Ax^{n+1}\|^2+\alpha\sum_{k=1}^N (x_k^{n+1}-x_k^n)s_{k}^{n}\xi_k^{n+1}\nonumber\\
&+& \alpha\sum_{k=1}^N(\varphi_k(|x_k^n|+\varepsilon)-\varphi_k(|x_k^{n+1}|+\varepsilon))\nonumber\\
&\geq & \frac{1}{2}\|Ax^n-Ax^{n+1}\|^2+\alpha\sum_{k=1}^N (|x_k^{n+1}|-|x_k^n|)s_{k}^{n}\nonumber\\
&+& \alpha\sum_{k=1}^N(\varphi_k(|x_k^n|+\varepsilon)-\varphi_k(|x_k^{n+1}|+\varepsilon))\nonumber\\
&=& \frac{1}{2}\|Ax^n-Ax^{n+1}\|^2+\alpha\sum_{k=1}^N D_{-\varphi_k} (|x_k^{n+1}|+\varepsilon,|x_k^n|+\varepsilon).\nonumber
\end{eqnarray}
Note that the last but one inequality followed from properties of subgradients of $|\cdot|$: $\xi_k^{n+1}x_k^n\leq |x_k^n|$ and $\xi_k^{n+1}x_k^{n+1}=|x_k^{n+1}|$.
\end{proof}
The next result shows convergence properties of the sequences generated by \eqref{reweighted_conv_gen}-\eqref{eps}. We recall first a notion that will be used in the analysis, namely sequential consistency of functions. A convex function $g$ defined on a Banach spaces $X$ is called sequentially consistent on its domain if for any two sequences $(x^k)$ and $(y^k)$ such that the first is bounded, one has
\begin{equation}\label{consistent}
D_g(y^k,x^k)\to 0\quad\Rightarrow \|x^k-y^k\|\to 0,\quad\mbox{as}\,\,k\to\infty.
\end{equation}
There are several equivalent definitions for sequential consistency - see, e.g. Theorem 2.10 in \cite{but-res}. Here we will take advantage of the following:
Let $g:\mathbb{R}^N\to(-\infty,\infty]$ be convex, lower semicontinuous, Fr\'{e}chet differentiable
on its domain, such that its derivative is uniformly continuous on bounded sets. Then $g$ is sequentially consistent.
We also recall the notion of stationary point for a locally Lipschitz functional. According to \cite[Chapter 2]{clarke}, if $f:X\to\mathbb{R}$ is a locally Lipschitz function defined on a Banach space $X$, the generalized subdifferential of $f$ at $x\in X$ in the sense of Clarke is defined by
$$\partial f(x)=\{\xi\in X^*: f^\circ(x,v)\geq \langle\xi,v\rangle,\,\forall v\in X\},$$
where $f^\circ(x,v)$ is the Clarke directional derivative
$$f^\circ(x,v)=\limsup_{\substack{y\to x\\ t\searrow 0}}\frac{f(y+tv)-f(y)}{t}\,.$$
A point $x\in X$ is called a stationary point of $f$, if $0\in\partial f(x).$ As expected, any local minimum point of $f$ is also a stationary point of the function.
In the remaining part of this section, we assume that the functions $t\to \phi_k(t+\varepsilon)$ are Lipschitz for any $\varepsilon>0$ (i.e., $\phi_k$ are uniformly Lipschitz on sets away from zero)
Consequently, the mapping $\displaystyle \tilde \phi(x)= \sum_{k=1}^N\phi_{k}(\abs{x_{k}}+\varepsilon)$ is also Lipschitz. Since the least squares term of $F$ is continuously differentiable and the regularizer is Lipschitz, it follows that the Clarke subdifferential of $F$ is additive, that is $\partial F(x)=A^*(Ax-y)+\partial \tilde \phi(x)$, where the latter Clarke sudifferential is given by $\partial\tilde \phi(x)=((\phi_k'(|x_k|+\varepsilon)\xi_k)_k$ with $\xi_k\in \partial (|\cdot|)(x_k)$.
\begin{proposition} Let $\varepsilon>0$. Let the functions $\phi_k$ be in $CI$, coercive and continuously differentiable. Assume that the functions $-\phi_k$ are sequentially consistent on domains that are bounded and away from zero. Then each sequence $(x^n)$ defined by \eqref{reweighted_conv_gen} is bounded and verifies $\displaystyle{\lim_{n\to\infty}(x^{n+1}-x^n)=0}$. Moreover, each accumulation point of $(x^n)$ is a stationary point of the function $F$.
\end{proposition}
\begin{proof} The sequence $(F(x^n))$ is convergent, as it is decreasing due to \eqref{decrease}, and is bounded from below. Thus, it
yields
\begin{equation}\label{zero}
\lim_{n\to\infty} D(x^{n+1},x^n)=0,\quad \lim_{n\to\infty}\|Ax^n-Ax^{n+1}\|=0.
\end{equation}
Consequently,
\begin{equation}\label{zero_i}
\lim_{n\to\infty} D_{-\phi_k} (|x_k^{n+1}|+\varepsilon,|x_k^n|+\varepsilon)=0,\quad \forall k\in\mathbb{N},
\end{equation}
which yields
\begin{equation}\label{limit}
\displaystyle{\lim_{n\to\infty}\left(|x_k^{n+1}|-|x_k^n|\right)=0}
\end{equation}
for each $k$, due to the sequential consistency of $-\phi_k$ for any $k$ on sets that are bounded and away from zero. Here we used the boundedness of $(x^n)$ (that is implied by the boundedness of $(F(x^n))$) and an argument as in the proof of Lemma \ref{inclusion},
\[
\frac{c}{1+L}\|x^n\|_1\leq \sum_{k=1}^N \frac{c|x_k^n|}{1+|x_k^n|}\leq c\sum_{k=1}^N \frac{|x_k^n|+\varepsilon}{1+|x_k^n|+\varepsilon}\leq \sum_{k=1}^N \phi_k(|x_k^n|+\varepsilon)\leq F(x^n).
\]
Now \eqref{limit} and the Lipschitz continuity of $\displaystyle x\mapsto \sum_{k=1}^N\phi_k(|x_k|+\varepsilon)$ combined with the second equality in \eqref{long} lead to
\begin{equation}\label{interm}
\lim_{n\to\infty}\left(x_k^{n+1}-x_k^n\right)\xi_k^{n+1}=0
\end{equation}
for each $k$.
Fix $k\in\{1,\dots,N\}$. One distinguishes two cases. If $x_k^{n+1}\neq 0$ except for finitely many $n\in\mathbb{N}$, then $|\xi_k^{n+1}|=1$, which due to \eqref{interm} implies that
\begin{equation}\label{interm1}
\lim_{n\to\infty}\left(x_k^{n+1}-x_k^n\right)=0.
\end{equation}
In the other case, there is a subsequence $(x_k^{n_j+1})$ which is constant zero. Since
$\displaystyle{\lim_{n\to\infty}\left(|x_k^{n+1}|-|x_k^n|\right)=0}$, one has $x_k^{n_j}\to 0$ and hence, $\displaystyle{\lim_{j\to\infty}(x_k^{n_j+1}-x_k^{n_j})=0}$.
Let $(x^{n_j})$ be a subsequence which converges to $\bar x$ and fix $k\in\{1,\dots,N\}$. By keeping the same notation on a subsequence of the bounded sequence $(\xi^{n_j}_k)$, the closeness of the graph of the subdifferential of $|\cdot|$ (see, e.g. Theorem 24.4 in \cite{rockafellar70}) yields existence of $\xi_k\in\partial (|\cdot|)(\bar x_k)$, such that $\xi^{n_j}_k\to\xi_k$ when $j\to\infty$. Due to \eqref{optim}, it follows that
\begin{equation}\label{optim1}
(A^*(Ax^{n_j}-y))_k+\alpha \phi_k'(|x_k^{n_j-1}-x_k^{n_j}+x_k^{n_j}|+\varepsilon)\xi_k^{n_j}=0,\,1\leq k\leq N.
\end{equation}
By taking limit for $j\to\infty$ in \eqref{optim1} and using the continuity of $\phi'_k$ imply
that $\bar x$ is a {stationary point of $F$}.
\end{proof}
\begin{remark} The functions from Remark \ref{examples} satisfy the sequential consistency property on sets that are bounded and away from zero, since the component functions are twice differentiable with bounded second derivatives, and thus, their derivatives are Lipschitz on sets that are bounded and away from zero.
\end{remark}
\section{Conclusions}
We proposed a general nonconvex approach to reconstruct sparse solutions of ill-posed problems with the aim of enhancing more flexibility than in the classical regularization penalties in $\ell^p$ spaces with $p \in (0,\infty)$. Our analysis touched both theoretical and
numerical aspects. From a theoretical point of view, we studied the convergence of the regularization method and the convergence properties of a couple of majorization techniques, leading
to monotone $\ell^2$-minimization and $\ell^1$-minimization schemes.
Finally, we showed convergence of an algorithm based on $\ell^2$-minimization and tested it in two situations, first for an academic example where the operator is an M matrix, and secondly for a
time-dependent optimal control problem. As shown by numerical results, the procedure is efficient and accurate, highlighting the advantages of using
variable penalties over a fixed penalty.
Further work will concern a different numerical approach, that is, a primal dual active set method for the nonconvex problem. This method is inspired by the one proposed in \cite{GK1} to study nonconvex problems with an $\ell^p$ penalty, $p \in (0,1]$. From a theoretical point of view, the primal dual active set method is interesting in itself since it identifies global minimizers (see Remark 5 of \cite{GK1}). Moreover, the method seems to be more numerically effective than the monotone algorithm in some situations. Last but not least, learning the flexible regularizers is a challenging open problem.
\section{Acknowledgments}
The authors are grateful to Robert Csetnek (University of Vienna) for the discussions on generalized subdifferentials, and to the referee for the interesting comments and suggestions. The work of D.L. has been supported by the ITN-ETN project TraDE-OPT funded by the European Union’s Horizon 2020 research and innovation programme under the Marie Skłodowska-Curie grant agreement No 861137. This work represents only the author’s view and the European Commission is not responsible for any use that may be made of the information it contains. Furthermore, D.L. acknowledges funding from BMBF under grant 05M20MBB and from the DFG under grants LO 1436/9-1 and FOR 3022/1.
\bibliographystyle{plain}
|
\section{Introduction}
Knowledge graph embeddings (KGE) are increasingly deployed in domains with high stake decision making like healthcare and finance \citep{noy2019knowledgegraphs}, where it is critical to identify the potential security vulnerabilities that might cause failure. But the research on adversarial vulnerabilities of KGE models has received little attention.
We study the adversarial vulnerabilities of KGE models through data poisoning attacks.
These attacks craft input perturbations at training time that aim to subvert the learned model's predictions at test time.
Poisoning attacks have been proposed for models that learn from other graph modalities \citep{xu2020advgraphsurvey} but they cannot be applied directly to KGE models. This is because they rely on gradients of all possible entries in a dense adjacency matrix and thus, do not scale to large knowledge graphs with multiple relations.
The main challenge in designing poisoning attacks for KGE models is the large combinatorial search space of candidate perturbations which is of the order of millions for benchmark knowledge graphs with thousands of nodes. Two recent studies \cite{zhang2019kgeattack, pezeshkpour2019criage} attempt to address this problem through random sampling of candidate perturbations \citep{zhang2019kgeattack} or through a vanilla auto-encoder that reconstructs discrete entities and relations from latent space \cite{pezeshkpour2019criage}.
However, random sampling depends on the number of candidates being sampled and the auto-encoder proposed in \citet{pezeshkpour2019criage} is only applicable to multiplicative KGE models.
\captionsetup[figure]{font=small}
\begin{figure}[]
\centering
\begin{subfigure}[htb]{1\columnwidth}
\includegraphics[width=1\columnwidth]{adversarial.jpg}
\end{subfigure}
\caption{Composition based adversarial attack on fraud detection. The knowledge graph consists of two types of entities - Person and BankAccount. The target triple to predict is $(Karl, \mathtt{affiliated\_with}, Joe\_the\_mobster)$. Original KGE model predicts this triple as True. But a malicious attacker adds adversarial triples (in purple) that connect $Karl$ with a non-suspicious person $Bob$ through composition pattern. Now, the KGE model predicts the target triple as False.
}
\label{fig:example}
\end{figure}
In this work, we propose to exploit the inductive abilities of KGE models to craft poisoned examples against the model.
The inductive abilities of KGE models are expressed through different connectivity patterns like symmetry, inversion and composition between relations in the knowledge graph.
We refer to these as \emph{inference patterns}.
We focus on the task of link prediction using KGE models and consider the adversarial goal of \emph{degrading} the predicted rank of \emph{target} missing facts.
To degrade the ranks of target facts, we propose to carefully select a set of decoy facts and exploit the inference patterns to improve performance on this decoy set.
Figure \ref{fig:example} shows an example of the use of composition pattern to degrade KGE model's performance.
We explore a collection of heuristic approaches to select the decoy triples and craft adversarial perturbations that use different inference patterns to improve the model's predictive performance on these decoy triples. Our solution addresses the challenge of large candidate space by breaking down the search space into smaller steps - (i) determining adversarial relations; (ii) determining the decoy entities that most likely violate an inference pattern; and (iii) determining remaining adversarial entities in the inference pattern that are most likely to improve the rank of decoy triples.
We evaluate the proposed attacks on four state-of-art KGE models with varied inductive abilities - DistMult, ComplEx, ConvE and TransE. We use two publicly available benchmark datasets for link prediction - WN18RR and FB15k-237. Comparison against the state-of-art poisoning attacks for KGE models shows that our proposed attacks outperform them in \emph{all} cases. We find that the attacks based on symmetry pattern perform the best and generalize across all model-dataset combinations.
Thus, the main contribution of our research is an effective method to generate data poisoning attacks, which is based on inference patterns captured by KGE models. Through a novel reformulation of the problem of poisoning KGE models, we overcome the existing challenge in the scalability of poisoning attacks for KGE models.
Furthermore, the extent of effectiveness of the attack relying on an inference pattern indicates the KGE model's sensitivity to that pattern.
Thus, our proposed poisoning attacks help in understanding the KGE models.
\section{Problem Formulation}
For a set of entities \( \mathcal{E} \) and a set of relations \( \mathcal{R} \), a knowledge graph is a collection of triples represented as \(
\mathcal{KG} = \{(s,\mathtt{r},o)\, |\, s,o \in \mathcal{E} \, and \, \mathtt{r} \in \mathcal{R} \}
\), where $s, \mathtt{r}, o$ represent the subject, relation and object in a triple.
A Knowledge Graph Embedding (KGE) model encodes entities and relations to a low-dimensional continuous vector space \( \vec{e}_s, \vec{e}_\mathtt{r}, \vec{e}_o \in \mathbb{R}^{k} \) where \( k \) is the embedding dimension. To do so, it uses a scoring function \(f : \mathcal{E} \times \mathcal{R} \times \mathcal{E} \rightarrow \mathbb{R} \) which depends on the entity and relation embeddings to assign a score to each triple \(f_{sro} = f(\vec{e}_s, \vec{e}_\mathtt{r}, \vec{e}_o) \). Table \ref{tab:scoring_functions} shows the scoring functions of state-of-art KGE models studied in this research. The embeddings are learned such that the scores for true (existing) triples in the knowledge graph are higher than the scores for false (non-existing) triples in the knowledge graph.
\begin{table}[]
\centering
\small
\begin{tabular}{|c|c|}
\hline
\textbf{Model} & \textbf{Scoring Function}\\
\hline
\hline
DistMult & $ \langle \vec{e}_s, \vec{e}_\mathtt{r}, \vec{e}_o \rangle$ \\
\hline
ComplEx & $ \Re(\langle \vec{e}_s, \vec{e}_\mathtt{r}, \overline{\vec{e}_o} \rangle)$ \\
\hline
ConvE & $ \langle \sigma(\mathrm{vec}(\sigma([ \overline{\vec{e}_\mathtt{r}}, \overline{\vec{e}_s}] \ast \boldsymbol{\Omega})) \mat{W}), \vec{e}_o \rangle$ \\
\hline
TransE & $-\norm{\vec{e}_s + \vec{e}_\mathtt{r} - \vec{e}_o}$ \\
\hline
\end{tabular}
\caption{Scoring functions \(f_{sro}\) of the KGE models used in this research. For ComplEx, $\vec{e}_s, \vec{e}_\mathtt{r}, \vec{e}_o \in \mathbb{C}^k$; for the remaining models $\vec{e}_s, \vec{e}_\mathtt{r}, \vec{e}_o \in \mathbb{R}^k$. Here, $\langle \cdot \rangle$ denotes the tri-linear dot product; $\sigma$ denotes sigmoid activation function, $\ast$ denotes 2D convolution; $\overline{\ \cdot\ }$ denotes conjugate for complex vectors, and 2D reshaping for real vectors in ConvE model; $\norm{\cdot}$ denotes l-p norm}
\label{tab:scoring_functions}
\end{table}
\textbf{Multiplicative vs Additive Interactions: }
The scoring functions of KGE models exhibit multiplicative or additive interactions \cite{chandrahas2018towards}. The multiplicative models score triples through multiplicative interactions of subject, relation and object embeddings. The scoring function for these models can be expressed as $f_{sro} = \vec{e}_{\mathtt{r}}^\top \mathcal{F}(\vec{e}_s, \vec{e}_o)$ where the function $\mathcal{F}$ measures the compatibility between the subject and object embeddings and varies across different models within this family. DistMult, ComplEx and ConvE have such interactions. On the other hand, additive models score triples through additive interactions of subject, relation and object embeddings. The scoring function for such models can be expressed as $f_{sro} = -\norm{\mat{M}_{\mathtt{r}}^1(\vec{e}_s) + \vec{e}_\mathtt{r} - \mat{M}_{\mathtt{r}}^2(\vec{e}_o)}$ where \( \vec{e}_s, \vec{e}_o \in \mathbb{R}^{k_\mathcal{E}} \), \(\vec{e}_\mathtt{r} \in \mathbb{R}^{k_\mathcal{R}} \) and $\mat{M}_\mathtt{r} \in \mathbb{R}^{k_{\mathcal{E}} \times k_{\mathcal{R}}}$ is the projection matrix from entity space $\mathbb{R}^{k_{\mathcal{E}}}$ to relation space $\mathbb{R}^{k_{\mathcal{R}}}$. TransE has additive interactions.
\textbf{Inductive Capacity of KGE models:}
The general intuition behind the design of the scoring functions of KGE models is to capture logical properties between relations from the observed facts in the knowledge graph. These logical properties or \emph{inference patterns} can then be used to make downstream inferences about entities and relations. For example, the relation $ \mathtt{is\_owned\_by}$ is inverse of the relation $ \mathtt {owns}$, and when the fact $(Account42,\mathtt{is\_owned\_by}, Karl)$ is true, then the fact $(Karl,\mathtt {owns}, Account42)$ is also true and vice versa. A model that can capture inversion pattern can thus predict missing facts about $\mathtt {owns}$ based on observed facts about $\mathtt{is\_owned\_by}$. The most studied inference patterns in the current literature are symmetry, inversion and composition since they occur very frequently in real-world knowledge graphs. In this work, we use these patterns to investigate the adversarial vulnerability of KGE models.
\textbf{Link Prediction:}
Since most of the existing knowledge graphs are incomplete, a standard use case of KGE models is to predict missing triples in the \( \mathcal{KG}\).
This task is evaluated by an entity ranking procedure. Given a test triple $(s,\mathtt{r},o)$, the subject entity is replaced by each entity from \( \mathcal{E} \) in turn. These replacements are referred to as \emph{synthetic negatives}.
The KGE model's scoring function is used to predict scores of these negative triples. The scores are then sorted in descending order and the rank of the correct entity is determined. These steps are repeated for the object entity of the triple.
The state-of-art evaluation metrics for this task are (i) \emph{MR} which is the mean of the predicted ranks, (ii) \emph{MRR} which is the mean of the reciprocals of predicted ranks and (iii) \emph{Hits@n} which count the proportion of correct entities ranked in top-n.
In the filtered setting \citep{bordes2013transe}, negative triples that already exist in the training, validation or test set are filtered out. That is, their scores are ignored while computing the ranks. Depending on the domain of use, either subject or object or both ranks of the test triple are used to determine the model's confidence\footnote{KGE models do not provide model uncertainty estimates.} in predicting a missing link.
\paragraph{Poisoning Attacks on KGE models: }
We study poisoning attacks for the task of link prediction using KGE models. We focus on targeted attacks where the attacker targets a specific set of missing triples instead of the overall model performance. We use the notation $(s,\mathtt{r},o)$ for the \emph{target triple}; in this case, $s,o$ are the \emph{target entities} and $\mathtt{r}$ is the \emph{target relation}. The goal of an adversarial attacker is to degrade the ranks of missing triples which are predicted highly plausible by the model.
The rank of a highly plausible target triple can be degraded by improving the rank of less plausible \emph{decoy triples}. For a target triple $(s,\mathtt{r},o)$, the decoy triple for degrading the rank on object side would be $(s,\mathtt{r},o')$ and the decoy triple for degrading the rank on subject side would be $(s',\mathtt{r},o)$.
Thus, the aim of the adversarial attacker is to select decoy triples from the set of valid synthetic negatives and craft \emph{adversarial edits} to improve their ranks. The attacker does not add the decoy triple itself as an adversarial edit, rather chooses the adversarial edits that would improve the rank of a missing decoy triple through an inference pattern.
\paragraph{Threat Model:}
To ensure reliable vulnerability analysis, we use a white-box attack setting where the attacker has full knowledge of the target KGE model \citep{joseph_nelson_rubinstein_tygar_2019}. They cannot manipulate the model architecture or learned embeddings directly; but only through \emph{addition} of triples to the training data. We focus on adversarial \emph{additions} which are more challenging to design than adversarial deletions for sparse knowledge graphs\footnote{For every target triple, the possible number of adversarial additions in the neighbourhood of each entity are $\mathcal{E} \times \mathcal{R}$. For the benchmark dataset FB15k-237, this is of the order of millions; whereas the \emph{maximum} number of candidates for adversarial deletion are of the order of thousands.}.
As in prior studies \citep{pezeshkpour2019criage, zhang2019kgeattack}, the attacker is restricted to making edits only in the neighbourhood of target entities. They are also restricted to 1 decoy triple for each entity of the target triple.
Furthermore, because of the use of filtered settings for KGE evaluation, the attacker cannot add the decoy triple itself to the training data (which intuitively would be a way to improve the decoy triple's rank).
\section{Poisoning Knowledge Graph Embeddings through Relation Inference Patterns }
Since the inference patterns on the knowledge graph specify a logic property between the relations, they can be expressed as Horn Clauses which is a subset of FOL formulae. For example, a property represented in the form $\forall x,y : (x, \mathtt{owns}, y) \Rightarrow (y, \mathtt{is\_owned\_by}, x)$ means that two entities linked by relation $\mathtt{owns}$ are also likely to be linked by the inverse relation $\mathtt{is\_owned\_by}$. In this expression, the right hand side of the implication $\Rightarrow$ is referred to as the \emph{head} and the left hand side as the \emph{body} of the clause.
Using such expressions, we define the three inference patterns used in our research.
\begin{definition}
The \textbf{symmetry} pattern $\mathcal{P}_s$ is expressed as \(
\forall x,y : (x, \mathtt{r}, y) \Rightarrow (y, \mathtt{r}, x)
\). Here, the relation $\mathtt{r}$ is symmetric relation.
\end{definition}
\begin{definition}
The \textbf{inversion} pattern $\mathcal{P}_i$ is expressed as
\(
\forall x,y : (x, \mathtt{r_i}, y) \Rightarrow (y, \mathtt{r}, x)
\). Here, the relations $\mathtt{r_i}$ and $\mathtt{r}$ are inverse of each other.
\end{definition}
\begin{definition}
The \textbf{composition} pattern $\mathcal{P}_c$ is expressed as
\(
\forall x,y,z : (x, \mathtt{r_1}, z) \wedge (z, \mathtt{r_2}, y) \Rightarrow (x, \mathtt{r}, y)
\). Here, the relation $\mathtt{r}$ is a composition of $\mathtt{r_1}$ and $\mathtt{r_2}$ ; and the $\wedge$ is the conjunction operator from relational logic.
\end{definition}
The mapping \(\mathcal{G}: \mathcal{V \rightarrow \mathcal{E}} \) of variables $\mathcal{V}$ in the above expressions to entities $\mathcal{E}$ is called a grounding.
For example, we can map the logic expression $\forall x,y : (x, \mathtt{owns}, y) \Rightarrow (y, \mathtt{is\_owned\_by}, x)$ to the grounding $(Karl, \mathtt{owns}, Account42) \Rightarrow (Account42, \mathtt{is\_owned\_by}, Karl)$. Thus, a KGE model that captures the inversion pattern will assign a high prediction confidence to the head atom when the body of the clause exists in the graph.
In the above expressions, the decoy triple becomes the head atom and adversarial edits are the triples in the body of the expression.
Since the decoy triple is an object or subject side negative of the target triple, the attacker already knows the relation in the head atom. They now want to determine (i) the adversarial relations in the body of the expression; (ii) the decoy entities which will most likely violate the inference pattern for the chosen relations and; (iii) the remaining entities in the body of the expression which will improve the prediction on the chosen decoy triple. Notice that the attacker needs all three steps for composition pattern only; for inversion pattern, only the first two steps are needed; and for symmetry pattern, only the second step is needed.
Below we describe each step in detail. A computational complexity analysis of all the steps is available in Appendix \ref{apx:complexity_analysis}.
\subsection{Step1: Determine Adversarial Relations}
Expressing the relation patterns as logic expressions is based on relational logic and assumes that the relations are constants. Thus, we use an algebraic approach to determine the relations in the head and body of a clause.
Given the target relation $\mathtt{r}$, we determine the adversarial relations using an algebraic model of inference \citep{yang2015distmult}.
\textbf{Inversion:}
If an atom $(x, \mathtt{r}, y)$ holds true, then for the learned embeddings in multiplicative models, we can assume $\vec{e}_x \circ \vec{e}_{\mathtt{r}} \approx \vec{e}_y$; where $\circ$ denotes the Hadamard (element-wise) product. If the atom $(y, \mathtt{r_i}, x)$ holds true as well, then we can also assume $\vec{e}_y \circ \vec{e}_{\mathtt{r_i}} \approx \vec{e}_x$. Thus, \(\vec{e}_{\mathtt{r}} \circ \vec{e}_{\mathtt{r_i}} \approx \vec{1}\) for inverse relations $\mathtt{r}$ and $\mathtt{r_i}$ when embeddings are learned from multiplicative models. We obtain a similar expression \(\vec{e}_{\mathtt{r}} + \vec{e}_{\mathtt{r_i}} \approx 0\) when embeddings are learned from additive models.
Thus, to determine adversarial relations for \emph{inversion} pattern, we use the pre-trained embeddings to select $\mathtt{r_i}$ that minimizes $\abs{\vec{e}_{\mathtt{r_i}} \vec{e}_{\mathtt{r}}^T - 1}$ for multiplicative models; and $\mathtt{r_i}$ that minimizes $\abs{\vec{e}_{\mathtt{r_i}} + \vec{e}_{\mathtt{r}}}$ for additive models.
\textbf{Composition}: If two atoms $(x, \mathtt{r_1}, y)$ and $(y, \mathtt{r_2}, z)$ hold true, then for multiplicative models, \(\vec{e}_x \circ \vec{e}_{\mathtt{r_1}} \approx \vec{e}_y\) and \(\vec{e}_y \circ \vec{e}_{\mathtt{r_2}} \approx \vec{e}_z \). Therefore, \(\vec{e}_x \circ (\vec{e}_{\mathtt{r_1}} \circ \vec{e}_{\mathtt{r_2}}) \approx \vec{e}_z\). Hence, relation $\mathtt{r}$ is a composition of $\mathtt{r_1}$ and $\mathtt{r_2}$ if $\vec{e}_{\mathtt{r_1}} \circ \vec{e}_{\mathtt{r_2}} \approx \vec{e}_{\mathtt{r}}$. Similarly, for embeddings from additive models, we can model composition as $\vec{e}_{\mathtt{r_1}} + \vec{e}_{\mathtt{r_2}} \approx \vec{e}_{\mathtt{r}}$.
Thus, to determine adversarial relations for \emph{composition} pattern, we use pre-trained embeddings to obtain all possible compositions of ($\mathtt{r_1}, \mathtt{r_2}$). For multiplicative models, we use $\vec{e}_{\mathtt{r_1}} \circ \vec{e}_{\mathtt{r_2}}$ and for additive models we use $\vec{e}_{\mathtt{r_1}} + \vec{e}_{\mathtt{r_2}}$. From these, we choose the relation pair for which the Euclidean distance between the composed relation embeddings and the target relation embedding $\vec{e}_{\mathtt{r}}$ is minimum.
\subsection{Step2: Determine Decoy Entities}
We consider three different heuristic approaches to select the decoy entity - soft truth score, ranks predicted by the KGE model and cosine distance.
\paragraph{Soft Logical Modelling of Inference Patterns}
Once the adversarial relations are determined, we can express the grounding for symmetry, inversion and composition patterns for the decoy triples. We discuss only object side decoy triple for brevity -
\begin{align*}
\mathcal{G}_s: (o', \mathtt{r}, s) &\Rightarrow (s, \mathtt{r}, o') \\
\mathcal{G}_i: (o', \mathtt{r_i},s) &\Rightarrow (s, \mathtt{r}, o') \\
\mathcal{G}_c: (s, \mathtt{r_1}, o'') \wedge (o'', \mathtt{r_2}, o') &\Rightarrow (s, \mathtt{r}, o')
\end{align*}
If the model captures $\mathcal{P}_s$, $\mathcal{P}_i$ or $\mathcal{P}_c$ to assign high rank to the target triple, then the head atom $(s, \mathtt{r}, o')$ of a grounding that violates this pattern is a suitable decoy triple. Adding the body of this grounding to the knowledge graph would improve the model performance on decoy triple through $\mathcal{P}_s$, $\mathcal{P}_i$ or $\mathcal{P}_c$.
To determine the decoy triple this way, we need a measure of the degree to which a grounding satisfies an inference pattern.
We call this measure the \emph{soft truth score} $\phi : \mathcal{G} \rightarrow [0,1]$ - it provides the truth value of a logic expression indicating the degree to which the expression is true. We model the soft truth score of grounded patterns using t-norm based fuzzy logics \citep{hajek1998tnormfuzzylogics}.
The score $f_{sro}$ of an individual atom (i.e. triple) is computed using the KGE model's scoring function. We use the sigmoid function $\sigma(x)=1/(1+\exp(-x))$ to map this score to a continuous truth value in the range $(0,1)$. Hence, the soft truth score for an individual atom is \(\phi(s,\mathtt{r},o) = \sigma(f_{sro})\).
The soft truth score for the grounding of a pattern can then be expressed through logical composition (e.g. $\wedge$ and $\Rightarrow$) of the scores of individual atoms in the grounding. We follow \citep{guo2016kale, guo2018ruge} and define the following compositions for logical conjunction ($\wedge$), disjunction ($\vee$), and negation ($\neg$):
\begin{align*}
\phi(a \wedge b) & = \phi(a) \cdot \phi(b), \\
\phi(a \vee b) & = \phi(a) + \phi(b) - \phi(a) \cdot \phi(b), \\
\phi(\neg a) & = 1 - \phi(a).
\end{align*}
Here, $a$ and $b$ are two logical expressions, which can either be single triples or be constructed by combining triples with logical connectives. If $a$ is a single triple $(s,\mathtt{r},o)$, we have $\phi(a)=\phi(s,\mathtt{r},o)$. Given these compositions, the truth value of any logical expression can be calculated recursively \citep{guo2016kale, guo2018ruge}.
Thus, we obtain the following soft truth scores for the groundings of symmetry, inversion and composition patterns $\mathcal{G}_s$, $\mathcal{G}_i$ and $\mathcal{G}_c$ -
\begin{align*}
\phi(\mathcal{G}_s) &= \phi(o',\mathtt{r},s) \cdot \phi(s,\mathtt{r},o') - \phi(o',\mathtt{r},s) + 1 \\
\phi(\mathcal{G}_i) &= \phi(o',\mathtt{r_i},s) \cdot \phi(s,\mathtt{r},o') - \phi(o',\mathtt{r_i},s) + 1. \\
\phi(\mathcal{G}_c) &= \phi(s,\mathtt{r_1},o'') \cdot \phi(o'',\mathtt{r_2},o') \cdot \phi(s,\mathtt{r},o') \\ &- \phi(s,\mathtt{r_1},o'') \cdot \phi(o'',\mathtt{r_2},o') + 1
\end{align*}
To select the decoy triple $(s,\mathtt{r},o')$ for symmetry and inversion, we score all possible groundings using $\phi(\mathcal{G}_s)$ and $\phi(\mathcal{G}_i)$. The head atom of grounding with minimum score is chosen as decoy triple.
For composition pattern, the soft truth score $\phi(\mathcal{G}_c)$ for candidate decoy triples $(s,\mathtt{r},o')$ contains two entities $(o',o'')$ to be identified. Thus, we use a greedy approach to select the decoy entity $o'$. We use the pre-trained embeddings to group the entities $o''$ into $k$ clusters using K-means clustering and determine a decoy entity with minimum soft truth score for each cluster. We then select the decoy entity $o'$ with minimum score across the $k$ clusters.
\paragraph{KGE Ranks:}
We use the ranking protocol from KGE evaluation to rank the target triple against valid subject and object side negatives $(s',\mathtt{r},o)$ and $(s,\mathtt{r},o')$. For each side, we select the negative triple that is ranked just below the target triple (that is, $negative\_rank = target\_rank + 1$). These are suitable as decoy because their predicted scores are likely not very different from the target triple's score. Thus, the model's prediction confidence for these triples might be effectively manipulated through adversarial additions. This is in contrast to very low ranked triples as decoy; where the model has likely learnt a low score with high confidence.
\paragraph{Cosine Distance:}
A high rank for the target triple $(s,\mathtt{r},o)$ against queries $(s,\mathtt{r},?)$ and $(?,\mathtt{r},o)$ indicates that $\vec{e}_s, \vec{e}_o$ are similar to the embeddings of other subjects and objects related by $\mathtt{r}$ in the training data. Thus, a suitable heuristic for selecting decoy entities $s'$ and $o'$ is to choose ones whose embeddings are dissimilar to $\vec{e}_s, \vec{e}_o$. Since these entities are not likely to occur in the neighbourhood of $o$ and $s$, they will act adversarially to reduce the rank of target triple.
Thus, we select decoy entities $s'$ and $o'$ that have maximum cosine distance from target entities $s$ and $o$ respectively.
\subsection{Step3: Determine Adversarial Entities}
This step is only needed for the composition pattern because the body for this pattern has two adversarial triples.
Given the decoy triple in the head of the composition expression, we select the body of the expression that would maximize the rank of the decoy triple. We use the soft-logical model defined in Step 2 for selecting decoy triples. The soft truth score for composition grounding of decoy triple is given by $\phi(\mathcal{G}_t) = \phi(s,\mathtt{r_1},o'') \cdot \phi(o'',\mathtt{r_2},o') \cdot \phi(s,\mathtt{r},o') - \phi(s,\mathtt{r_1},o'') \cdot \phi(o'',\mathtt{r_2},o') + 1$. We select the entity $o''$ with maximum score because this entity satisfies the composition pattern for the decoy triple and is thus likely to improve the decoy triple's ranks on addition to the knowledge graph.
\begin{table}[]
\centering
\small
\begin{tabular}{c c c c }
\toprule
\textbf{Adversarial Attack Step} & \textbf{Sym} & \textbf{Inv} & \textbf{Com}\\
\midrule
{Determine Adversarial Relations} & n/a & Alg & Alg \\
\midrule
\multirow{3}{*}{\shortstack[l]{Determine Decoy Entities}} & Sft & Sft & Sft \\
& Rnk & Rnk & Rnk \\
& Cos & Cos & Cos \\
\midrule
Determine Adversarial Entities & n/a & n/a & Sft \\
\bottomrule
\end{tabular}
\caption{A summary of heuristic approaches used for different steps of the adversarial attack with symmetry (Sym), inversion (Inv) and composition (Com) pattern. Alg denotes the algebraic model for inference patterns; Sft denotes the soft truth score; Rnk denotes the KGE ranks; and Cos denotes the cosine distance.}
\label{tab:attack_summary}
\end{table}
\section{Evaluation}
The aim of our evaluation is to assess the effectiveness of proposed attacks in \emph{degrading} the predictive performance of KGE models on missing triples that are predicted true.
We use the state-of-art evaluation protocol for data poisoning attacks \citep{xu2020advgraphsurvey}. We train a clean model on the original data; then generate the adversarial edits and add them to the dataset; and finally retrain a new model on this poisoned data. All hyperparameters for training on original and poisoned data remain the same.
We evaluate four models with varying inductive abilities - DistMult, ComplEx, ConvE and TransE; on two publicly available benchmark datasets for link prediction\footnote{https://github.com/TimDettmers/ConvE}- WN18RR and FB15k-237.
We filter out triples from the validation and test set that contain unseen entities.
To assess the attack effectiveness in \emph{degrading} performance on triples predicted as true, we need a set of triples that are predicted as true by the model.
Thus, we select as \emph{target triples}, a subset of the original test set where each triple is ranked $\leq$ 10 by the original model.
Table \ref{tab:data} provides an overview of dataset statistics and the number of target triples selected.
\begin{table}[]
\centering
\footnotesize
\setlength{\tabcolsep}{5pt}
\begin{tabular}{c l ll}
\toprule
\multicolumn{2}{l}{} & \textbf{WN18RR} & \textbf{FB15k-237} \\
\midrule
\multicolumn{2}{l}{Entities} & 40,559 & 14,505 \\
\multicolumn{2}{l}{Relations} & 11 & 237 \\
\multicolumn{2}{l}{Training} & 86,835 & 272,115 \\
\multicolumn{2}{l}{Validation} & 2,824 & 17,526 \\
\multicolumn{2}{l}{Test} & 2,924 & 20,438 \\
\midrule
\multirow{4}{*}{Target}
& DistMult & 1,315 & 3,342 \\
& ComplEx & 1,369 & 3,930 \\
& ConvE & 1,247 & 4,711 \\
& TransE & 1,195 & 5,359 \\
\bottomrule
\end{tabular}
\caption{\small Statistics for the datasets WN18RR and FB15k-237.
We removed triples from the validation and test set that contained unseen entities to ensure that we do not add new entities as adversarial edits.
The numbers above (including the number of entities) reflect this filtering.}
\label{tab:data}
\end{table}
\paragraph{Baselines: }
We compare the proposed methods against the following baselines -
\emph{Random\_n}: Random edits in the neighbourhood of each entity of the target triple.
\emph{Random\_g1}: Global random edits in the knowledge graph which are not restricted to the neighbourhood of entities in the target triple and have 1 edit per decoy triple (like symmetry and inversion).
\emph{Random\_g2}: Global random edits in the knowledge graph which are not restricted to the neighbourhood of entities in the target triple and have 2 edits per decoy triple (like composition).
\emph{Zhang et al.}: Poisoning attack from \cite{zhang2019kgeattack} for edits in the neighbourhood of subject of the target triple. We extend it for both subject and object to match our evaluation protocol. Further implementation details available in Appendix \ref{apx:ijcai_baseline}.
\emph{CRIAGE}: Poisoning attack from \citep{pezeshkpour2019criage}. We use the publicly available implementation and the default attack settings\footnote{https://github.com/pouyapez/criage}. The method was proposed for edits in the neighbourhood of object of the target triple. We extend it for both entities to match our evaluation protocol and to ensure fair evaluation.
\paragraph{Implementation:}
For every attack, we filter out adversarial edit candidates that already exist in the graph.
We also remove duplicate adversarial edits for different targets before adding them to the original dataset.
For Step 2 of the composition attack with ground truth, we use the elbow method to determine the number of clusters for each model-data combination. Further details on KGE model training, computing resources and number of clusters are available in Appendix \ref{apx:implementation}. The source code to reproduce our experiments is available on GitHub\footnote{\url{https://github.com/PeruBhardwaj/InferenceAttack}}.
\subsection{Results}
\label{sec:results}
\begin{table*}
\centering
\small
\setlength{\tabcolsep}{3.5pt}
\begin{tabular}{c l ll ll ll ll }
\toprule
& & \multicolumn{2}{c}{\textbf{DistMult}} & \multicolumn{2}{c}{\textbf{ComplEx}} & \multicolumn{2}{c}{\textbf{ConvE}} & \multicolumn{2}{c}{\textbf{TransE}} \\
\cmidrule(lr){3-4} \cmidrule(lr){5-6} \cmidrule(lr){7-8} \cmidrule(lr){9-10}
& & \multicolumn{1}{l}{\textbf{MRR}} & \multicolumn{1}{l}{\textbf{Hits@1}} & \multicolumn{1}{l}{\textbf{MRR}} & \multicolumn{1}{l}{\textbf{Hits@1}} & \multicolumn{1}{l}{\textbf{MRR}} & \multicolumn{1}{l}{\textbf{Hits@1}} & \multicolumn{1}{l}{\textbf{MRR}} & \multicolumn{1}{l}{\textbf{Hits@1}} \\
\midrule
\textbf{Original} & & 0.90 & 0.85 & 0.89 & 0.84 & 0.92 & 0.89 & 0.36 & 0.03 \\
\midrule
\multirow{5}{*}{\shortstack[l]{\textbf{Baseline} \\ \textbf{Attacks}}}
& Random\_n & 0.86 (-4\%) & 0.83 & 0.84 (-6\%) & 0.80 & 0.90 (-2\%) & 0.88 & 0.28 (-20\%) & 0.01 \\
& Random\_g1 & 0.88 & 0.83 & 0.88 & 0.83 & 0.92 & 0.89 & 0.35 & 0.02 \\
& Random\_g2 & 0.88 & 0.83 & 0.88 & 0.83 & 0.91 & 0.89 & 0.34 & 0.02 \\
\cline{2-10} \\[-7pt]
& Zhang et al. & 0.82 (-8\%) & 0.81 & 0.76 (-14\%) & 0.74 & 0.90 (-2\%) & 0.87 & \textbf{0.24 (-33\%)} & \textbf{0.01} \\
& CRIAGE & 0.87 & 0.84 & - & - & 0.90 & 0.88 & - & - \\
\midrule
\multirow{9}{*}{\shortstack[l]{\textbf{Proposed} \\ \textbf{Attacks}}}
& Sym\_truth & 0.66 & 0.40 & \textbf{0.56 (-33\%)} & \textbf{0.24} & \textbf{0.61 (-34\%)} & \textbf{0.28} & 0.57 & 0.36 \\
& Sym\_rank & 0.61 & 0.32 & \textbf{0.56 (-33\%)} & \textbf{0.24} & 0.62 & 0.31 & 0.25 & 0.02 \\
& Sym\_cos & \textbf{0.57 (-36\%)} & \textbf{0.32} & 0.62 & 0.43 & 0.67 & 0.44 & \textbf{0.24 (-33\%)} & \textbf{0.01} \\
\cline{2-10} \\[-7pt]
& Inv\_truth & 0.87 & 0.83 & 0.86 & 0.80 & 0.90 & 0.87 & 0.34 & 0.03 \\
& Inv\_rank & 0.86 & 0.83 & 0.85 & 0.80 & 0.89 (-4\%) & 0.85 & 0.25 & 0.02 \\
& Inv\_cos & 0.83 (-8\%) & 0.82 & 0.80 (-10\%) & 0.79 & 0.90 & 0.88 & 0.25 (-30\%) & 0.01 \\
\cline{2-10} \\[-7pt]
& Com\_truth & 0.86 & 0.83 & 0.86 & 0.81 & 0.89 & 0.86 & 0.53 (+49\%) & 0.27 \\
& Com\_rank & 0.85 (-5\%) & 0.80 & 0.83 & 0.77 & 0.89 & 0.84 & 0.57 & 0.32 \\
& Com\_cos & 0.86 & 0.77 & 0.82 (-8\%) & 0.70 & 0.88(-4\%) & 0.83 & 0.53 (+49\%) & 0.27 \\
\bottomrule
\end{tabular}
\caption{\small Reduction in MRR and Hits@1 due to different attacks on the \textbf{target split} of WN18RR. First block of rows are the baseline attacks with random edits; second block is state-of-art attacks; remaining are the proposed attacks. For each block, we report the \emph{best} relative percentage difference from original MRR; computed as $(original-poisoned)/original*100$. Lower values indicate better results; best results for each model are in bold. Statistics on the target split are in Table \ref{tab:data}.}
\label{tab:benchmark_mrr_WN18RR}
\end{table*}
Table \ref{tab:benchmark_mrr_WN18RR} and \ref{tab:benchmark_mrr_FB15k-237} show the reduction in MRR and Hits@1 due to different attacks on the WN18RR and FB15k-237 datasets.
We observe that the proposed adversarial attacks outperform the random baselines and the state-of-art poisoning attacks for all KGE models on both datasets.
We see that the attacks based on symmetry inference pattern perform the best across all model-dataset combinations. This indicates the sensitivity of KGE models to symmetry pattern. For DistMult, ComplEx and ConvE, this sensitivity can be explained by the symmetric nature of the scoring functions of these models. That is, the models assign either equal or similar scores to triples that are symmetric opposite of each other. In the case of TransE, the model's sensitivity to symmetry pattern is explained by the translation operation in scoring function. The score of target $(s,\mathtt{r},o)$ is a translation from subject to object embedding through the relation embedding. Symmetry attack adds the adversarial triple $(o',\mathtt{r},s)$ where the relation is same as the target relation, but target subject is the object of adversarial triple. Now, the model learns the embedding of $s$ as a translation from $o'$ through relation $\mathtt{r}$. This adversarially modifies the embedding of $s$ and in turn, the score of $(s,\mathtt{r},o)$.
We see that inversion and composition attacks also perform better than baselines in most cases, but not as good as symmetry. This is particularly true for FB15k-237 where the performance for these patterns is similar to random baselines. For the composition pattern, it is likely that the model has stronger bias for shorter and simpler patterns like symmetry and inversion than for composition. This makes it harder to deceive the model through composition than through symmetry or inverse. Furthermore, FB15k-237 has high connectivity \citep{dettmers2018conve} which means that a KGE model relies on a high number of triples to learn target triples' ranks. Thus, poisoning KGE models for FB15k-237 will likely require more adversarial triples per target triple than that considered in this research.
The inversion pattern is likely ineffective on the benchmark datasets because these datasets do not have any inverse relations \citep{dettmers2018conve,toutanova2015observed}. This implies that our attacks cannot identify the inverse of the target triple's relation in Step 1. We investigate this hypothesis further in Appendix \ref{apx:WN18_analysis}, and evaluate the attacks on WN18 dataset where the inverse relations have not been filtered out. This means that the KGE model can learn the inversion pattern and the inversion attacks can identify the inverse of the target relation. In this setting, we find that the inversion attacks outperform other attacks against ComplEx on WN18, indicating the sensitivity of ComplEx to the inversion pattern when the dataset contains inverse relations.
An exception in the results is the composition pattern on TransE where the model performance improves instead of degrading on the target triples. This is likely due to the model's sensitivity to composition pattern such that adding this pattern improves the performance on all triples, including target triples. To verify this, we checked the change in ranks of decoy triples and found that composition attacks on TransE improve these ranks too. Results for this experiment are available in Appendix \ref{apx:decoy_analysis}. This behaviour of composition also indicates that the selection of adversarial entities in Step 3 of the composition attacks can be improved. It also explains why the increase is more significant for WN18RR than FB15k-237 - WN18RR does not have any composition relations but FB15k-237 does; so adding these to WN18RR shows significant improvement in performance.
We aim to investigate these and more hypotheses about the proposed attacks in future work.
\begin{table*}[]
\centering
\small
\setlength{\tabcolsep}{3.5pt}
\begin{tabular}{c l ll ll ll ll }
\toprule
& & \multicolumn{2}{c}{\textbf{DistMult}} & \multicolumn{2}{c}{\textbf{ComplEx}} & \multicolumn{2}{c}{\textbf{ConvE}} & \multicolumn{2}{c}{\textbf{TransE}} \\
\cmidrule(lr){3-4} \cmidrule(lr){5-6} \cmidrule(lr){7-8} \cmidrule(lr){9-10}
& & \textbf{MRR} & \textbf{Hits@1} & \textbf{MRR} & \textbf{Hits@1} & \textbf{MRR} & \textbf{Hits@1} & \textbf{MRR} & \textbf{Hits@1} \\
\midrule
\textbf{Original} & & 0.61 & 0.38 & 0.61 & 0.45 & 0.61 & 0.45 & 0.63 & 0.48 \\
\midrule
\multirow{5}{*}{\shortstack[l]{\textbf{Baseline} \\ \textbf{Attacks}}}
& Random\_n & 0.54 (-11\%) & 0.40 & 0.54 (-12\%) & 0.40 & 0.56 (-8\%) & 0.41 & 0.60 (-4\%) & 0.45 \\
& Random\_g1 & 0.54 & 0.40 & 0.55 & 0.41 & 0.57 & 0.43 & 0.62 & 0.46 \\
& Random\_g2 & 0.55 & 0.41 & 0.55 & 0.40 & 0.57 & 0.42 & 0.61 & 0.46 \\
\cline{2-10} \\[-7pt]
& Zhang et al. & 0.53 (-13\%) & 0.39 & 0.51 (-16\%) & 0.38 & 0.54 (-11\%) & 0.39 & 0.57 (-10\%) & 0.42 \\
& CRIAGE & 0.54 & 0.41 & - & - & 0.56 & 0.41 & - & - \\
\midrule
\multirow{9}{*}{\shortstack[l]{\textbf{Proposed} \\ \textbf{Attacks}}}
& Sym\_truth & 0.51 & 0.36 & 0.56 & 0.41 & \textbf{0.51} (\textbf{-17\%}) & \textbf{0.34} & 0.62 & 0.48 \\
& Sym\_rank & 0.53 & 0.39 & 0.53 & \textbf{0.38} & 0.55 & 0.38 & \textbf{0.53} (\textbf{-16\%}) & \textbf{0.36} \\
& Sym\_cos & \textbf{0.46} (\textbf{-25\%}) & \textbf{0.31} & \textbf{0.51} (\textbf{-17\%}) & \textbf{0.38}
& 0.52 & 0.37 & 0.55 & 0.40 \\
\cline{2-10} \\[-7pt]
& Inv\_truth & 0.55 & 0.41 & 0.54 & 0.40 & 0.56 & 0.41 & 0.62 & 0.46 \\
& Inv\_rank & 0.56 & 0.43 & 0.55 & 0.40 & 0.55 (-9\%) & 0.40 & 0.58 (-8\%) & 0.42 \\
& Inv\_cos & 0.54 (-11\%) & 0.40 & 0.53 (-14\%) & 0.39 & 0.56 & 0.42 & 0.59 & 0.44 \\
\cline{2-10} \\[-7pt]
& Com\_truth & 0.56 & 0.42 & 0.55 & 0.41 & 0.57 & 0.43 & 0.65 & 0.51 \\
& Com\_rank & 0.56 (-8\%) & 0.42 & 0.55 (-11\%) & 0.40 & 0.56 (-8\%) & 0.41 & 0.69 & 0.48 \\
& Com\_cos & 0.56 (-8\%) & 0.43 & 0.56 & 0.42 & 0.56 & 0.42 & 0.63 (0\%) & 0.49 \\
\bottomrule
\end{tabular}
\caption{\small Reduction in MRR and Hits@1 due to different attacks on the \textbf{target split} of FB15k-237. For each block of rows, we report the \emph{best} relative percentage difference from original MRR; computed as $(original-poisoned)/original*100$. Lower values indicate better results; best results for each model are in bold. Statistics on the target split are in Table \ref{tab:data}.}
\label{tab:benchmark_mrr_FB15k-237}
\end{table*}
\section{Related Work}
KGE models can be categorized into tensor factorization models like DistMult \citep{yang2015distmult} and ComplEx \citep{trouillon2016complex}, neural architectures like ConvE \citep{dettmers2018conve} and translational models like TransE \citep{bordes2013transe}. We refer the reader to \citep{cai2018comprehensive} for a comprehensive survey.
Due to the black-box nature of KGE models, there is an emerging literature on understanding these models.
\citep{pezeshkpour2019criage} and \citep{zhang2019kgeattack} are most closely related to our work as they propose other data poisoning attacks for KGE models.
\citet{minervini2017adversarialsets} and \citet{cai2018kbgan} use adversarial regularization in latent space and adversarial training to improve predictive performance on link prediction. But these adversarial samples are not in the input domain and aim to improve instead of degrade model performance.
Poisoning attacks have also been proposed for models for undirected and single relational graph data like Graph Neural Networks ~\citep{zugner2018nettack, dai2018adversarialgcn} and Network Embedding models \citep{bojchevski2019adversarialnetworkembedding}. A survey of poisoning attacks for graph data is available in \citep{xu2020advgraphsurvey}. But the attacks for these models cannot be applied directly to KGE models because they require gradients of a dense adjacency matrix.
In the literature besides adversarial attacks, \citet{lawrence2020gradientrollback}, \citet{nandwani2020oxkbc} and \citet{zhang2019interaction} generate post-hoc explanations to understand KGE model predictions. \citet{trouillon2019inductive} study the inductive abilities of KGE models as binary relation properties for controlled inference tasks with synthetic datasets. \citet{allen2021interpreting} interpret the structure of knowledge graph embeddings by comparison with word embeddings. On the theoretical side, \citet{wang2018multi} study the expressiveness of various bilinear KGE models and \citet{basulto2018ontologyembedding} study the ability of KGE models to learn hard rules expressed as ontological knowledge.
The soft-logical model of inference patterns in this work is inspired by the literature on injecting logical rules into KGE models. \citet{guo2016kale} and \citet{guo2018ruge} enforce soft logical rules by modelling the triples and rules in a unified framework and jointly learning embeddings from them.
Additionally, our algebraic model of inference patterns, which is used to select adversarial relations, is related to approaches for graph traversal in latent vector space discussed in \citet{yang2015distmult, guu2015traversing, arakelyan2021complexqueryanswering}.
\section{Conclusion}
We propose data poisoning attacks against KGE models based on inference patterns like symmetry, inversion and composition. Our experiments show that the proposed attacks outperform the state-of-art attacks.
Since the attacks rely on relation inference patterns, they can also be used to understand the KGE models. This is because if a KGE model is sensitive to a relation inference pattern, then that pattern should be an effective adversarial attack.
We observe that the attacks based on symmetry pattern generalize across all KGE models which indicates their sensitivity to this pattern.
In the future, we aim to investigate hypotheses about the effect of input graph connectivity and existence of specific inference patterns in datasets.
We note that such investigation of inference pattern attacks will likely be influenced by the choice of datasets. In this paper, we have used benchmark datasets for link prediction. While there are intuitive assumptions about the inference patterns on these datasets, there is no study that formally measures and characterizes the existence of these patterns. This makes it challenging to verify the claims made about the inductive abilities of KGE models, not only by our proposed attacks but also by new KGE models proposed in the literature.
Thus, a promising step in understanding knowledge graph embeddings is to propose datasets and evaluation tasks that test varying degrees of specific inductive abilities.
These will help evaluate new models and serve as a testbed for poisoning attacks. Furthermore, specifications of model performance on datasets with different inference patterns will improve the usability of KGE models in high-stake domains like healthcare and finance.
In addition to understanding model behaviour, the sensitivity of state-of-art KGE models to simple inference patterns indicates that these models can introduce security vulnerabilities in pipelines that use knowledge graph embeddings. Thus, another promising direction for future work is towards mitigating the security vulnerabilities of KGE models. Some preliminary ideas for this research can look into adversarial training; or training an ensemble of different KGE scoring functions; or training an ensemble from subsets of the training dataset. Since our experiments show that state-of-art KGE models are sensitive to symmetry pattern, we call for future research to investigate neural architectures that generalize beyond symmetry even though their predictive performance for link prediction on benchmark datasets might not be the best.
\section*{Acknowledgements}
This research was conducted with the financial support of Accenture Labs and Science Foundation Ireland (SFI) at the ADAPT SFI Research Centre at Trinity College Dublin. The ADAPT SFI Centre for Digital Content Technology is funded by Science Foundation Ireland through the SFI Research Centres Programme and is co-funded under the European Regional Development Fund (ERDF) through Grant No. 13/RC/2106\_P2.
\section*{Broader Impact}
We study the problem of generating data poisoning attacks on KGE models. Data poisoning attacks identify the vulnerabilities in learning algorithms that could be exploited by an adversary to manipulate the model's behaviour \citep{joseph_nelson_rubinstein_tygar_2019, biggio2018wild}. Such manipulation can lead to unintended model behaviour and failure. Identifying these vulnerabilities for KGE models is critical because of their increasing use in domains that need high stakes decision making like heathcare \citep{bendsten2019astrazeneca} and finance \citep{hogan2020knowledgegraphs, noy2019knowledgegraphs}.
In this way, our research is directed towards minimizing the negative consequences of deploying state-of-art KGE models in our society. This honours the ACM code of Ethics of contributing to societal well-being and acknowledging that all people are stakeholders in computing. At the same time, we aim to safeguard the KGE models against potential harm from adversaries and thus honour the ACM code of avoiding harm due to computing systems.
Arguably, because we study vulnerabilities by attacking the KGE models, the proposed attacks can be used by an actual adversary to manipulate the model behaviour of deployed systems. This paradox of an arms race is universal across security research \citep{biggio2018wild}. For our research, we have followed the principle of proactive security as recommended by \citet{joseph_nelson_rubinstein_tygar_2019} and \citet{biggio2018wild}. As opposed to reactive security measures where learning system designers develop countermeasures after the system is attacked, a proactive approach \emph{anticipates} such attacks, simulates them and designs countermeasures before the systems are deployed. Thus, by revealing the vulnerabilities of KGE models, our research provides an opportunity to fix them.
Besides the use case of security, our research can be used in understanding the inductive abilities of KGE models, which are black-box and hard to interpret. We design attacks that rely on the inductive assumptions of a model to be able to deceive that model. Thus, theoretically, the effectiveness of attacks based on one inference pattern over another indicates the model's reliance on one inference pattern over another. However, as we discussed in our paper, realistically, it is challenging to make such claims about the inductive abilities of KGE models because the inference patterns in benchmark datasets are not well defined.
Thus, we would encourage further work to evaluate our proposed attacks by designing benchmark tasks and datasets that measure specific inductive abilities of models. This will not only be useful for evaluating the proposed attacks here, but also for understanding the inductive abilities of existing KGE models. This in turn, can guide the community to design better models. In this direction, we encourage researchers proposing new KGE models to evaluate not only the predictive performance on benchmark datasets, but also the claims made on inductive abilities of these models and their robustness to violations of these implicit assumptions.
\bibliographystyle{acl_natbib}
|
\section{Introduction} \label{sec:introduction}
The identification and macroeconomic impact of monetary policy shocks is a longstanding and open question. The key challenge is that monetary policy responds to both the current and the expected future states of the macroeconomy. Controlling for these unobserved state variables underlies the main difficulty in identifying the causal impact of monetary policy. One way to overcome this challenge, is to measure changes in the prices of financial assets around Central Bank announcements. Under tight time-windows around Central Bank's announcements, it is credible to assume that financial markets react to said announcements only. However, as stressed by \cite{nakamura2018high} and \cite{jarocinski2020deconstructing} among others, these announcements reveal both policy decisions and the central bank's assessment about the present and future economic outlook. Consequently, these high-frequency movements in the prices of financial assets are a response not only to policy decisions but also to the private sector's updated beliefs about the state of the economy. Consider the following example of both policy and private market beliefs around a FOMC meeting.
On September 18, 2007, the FOMC cut the Federal Funds Rate by 50 basis points. As the New York Times reports, ``while an interest rate cut was widely expected, there had been profound uncertainty about whether the Fed would choose a more cautious quarter-point reduction. But the bolder action and an accompanying statement, both approved by a unanimous vote of the central bank’s policy-setting committee, made it clear that the Fed had decided the risks of a recession were too big to ignore." Hence, we see how investors learnt from the aggressive policy decision about the seriously negative prospects facing the economy.
In order to study the causal effects of monetary policy, we would like to decompose high-frequency movements into the part which responds to monetary policy (the monetary component) and the part which responds to information disclosure by the central bank (the information or news component). Studying the effect of the monetary component would allow us to understand the effect of monetary policy, while the information component allows us to understand how central banks influence the economy beyond direct policy tools.
In this paper, we use FOMC statements and newspaper articles to identify and quantify the information component of monetary policy announcements. We do so via state-of-the art text analysis tools in a framework that is both transparent and highly interpretable. Whereas earlier literature has inferred the information component through the introduction of a structural model (see \cite{nakamura2018high}) or imposing restrictions on the moments of high-frequency surprises (see \cite{jarocinski2020deconstructing}), our method does not impose any assumptions on the relationship between the potential shocks and macroeconomic variables. Instead, we use text data to model expectations of the FOMC and the private sector over the Federal Funds Rate (FFR). Our key insight is that the informational component of monetary policy is the part of high frequency movements that is correlated with differences in the expectations of the two.
Our method consists of three stages. First, we use tools in computational linguistics to embed FOMC statements and New York Times articles as numerical vectors. Next, we use the embedded documents to separately model both FOMC and private sector expectations of the FFR via the elastic net, a transparent method for dealing with high dimensional data, and which \cite{giannone2021sparsity} advocates for economic data. Here, we use the expectation of the FOMC to refer to the hypothetical monetary policy rate that would have been chosen in the absence of monetary policy shocks. Stage two allows us to construct the differences in expectation about the FFR which arose due to differences in information alone. Finally, we project high frequency movements of select interest rates around FOMC announcements onto the differences in expectation and interpret the projection as the information component of the high frequency movements.
We apply our framework to a set of identified high-frequency movements which the literature has interpreted as monetary policy shocks.
Following \cite{nakamura2018high} and \cite{jarocinski2020deconstructing} we decompose these high-frequency movements into a monetary component and news component. We find that over our sample period of January 2000 to March 2014, differences in beliefs have predictive power over the high-frequency movement of the Federal Funds Futures around FOMC meetings. We interpret this finding as evidence that FOMC announcements disclose information about the state of the economy.
We show that across these different high-frequency movements, the isolated news component has predictive power over the daily change of short term nominal and real rates around FOMC meetings. In particular, we argue that not controlling for the news component of high-frequency surprises may lead to biases on the impact of monetary policy shocks on domestic financial conditions.
Our contribution to the literature is three-fold. First, we use newspaper text to proxy for private sector information about monetary policy around FOMC meetings. While text analysis has been used to study the information effects of FOMC policy announcements, existing papers have focused on FOMC statements or minutes exclusively (\cite{lucca2009measuring}, \cite{hansen2018transparency}, \cite{benchimol2020communication}, \cite{handlan2020text}). We argue that any informational effects has to operate through differences in expectations, so that both FOMC and private beliefs must be taken into consideration. Hence we augment our analysis of FOMC statements with New York Times articles.
Second, we propose a text analysis framework that is interpretable and highly transparent. Our framework uses the black-box output of text analysis methods to construct conditional expectations of various macroeconomic variables. Studying the effects of these variables -- or their differences -- on high frequency interest rate movements allows us to understand the transmission mechanisms of informational effects. Our method is also highly transparent: the elastic net, which is a form of penalized linear regression, performs explicit regularization using only two tuning parameters.
Third, the empirical applications of our framework shed light on the impact of the informational content of FOMC announcements on both financial and macroeconomic variables. Specifically, the news component of high-frequency monetary policy shocks have a significant impact on short run nominal and real rates, while having little to no effect on longer maturities. Further studying these shocks can provide insights on the specific information advantage the Federal Reserve has over private sector agents.
The paper is organized as follows. Section \ref{sec:data} describes the data. Section \ref{sec:methodology} lays out the text analysis framework in detail. Section \ref{sec:main_results} provides both qualitative and quantitative evidence that our procedure recovers informational content. Section \ref{sec:conclusion} concludes.
\section{Data} \label{sec:data}
This section of the paper describes the two main types of data used: (i) text coming from both the Federal Reserve's FOMC statements and New York Times' articles; (ii) high-frequency financial surprises around FOMC announcements.
The sample period is January 2000 to March 2014. The FOMC statement announcements were sourced from the Federal Reserve's website.\footnote{\url{https://www.federalreserve.gov/monetarypolicy/fomc_historical_year.htm}} The FOMC statements are short summaries of the topics discussed by the committee, released following a meeting. This include the assessment of the current economic situation, and policy decisions regarding the target federal funds rate, discount rate, and since 2008 different types of unconventional monetary policy. In January 2000, the Committee announced that it would issue a statement following each regularly scheduled meeting, regardless of whether there had been a change in monetary policy. We drop unscheduled FOMC meetings from our sample as in \cite{nakamura2018high} and \cite{handlan2020text}. In general, an unscheduled meeting reflects highly unusual economic or political circumstances.
The New York Times' articles were sourced from Northwestern Library, which granted us access up to March 2014.\footnote{We accessed and downloaded these articles on April 2 2021.} Our sample consists of articles published on the day of and day prior to the FOMC meetings. Furthermore we exploit the fact that articles are tagged by subjects and restrict our sample to only those which contain the subject ``federal reserve". These articles typically discuss the current economic situations and contain several private sector projections on both the economy and FOMC policy decisions. Thus, they provide us with a proxy for private sector expectations on the FOMC announcement day.
\newpage
We were able to find at least one such article during the day of and the day before each scheduled FOMC meeting date. This is consistent with the idea that the FOMC meetings are an important event for the financial markets such that they receive news coverage in the lead-up.
The empirical section of the paper uses both high-frequency movements in financial assets and macroeconomic variables for the US. The two high-frequency movements studied are the Fed Funds Futures surprises constructed by \cite{gertler2015monetary} and the policy news shocks constructed by \cite{nakamura2018high}.
We measure the impact of both the isolated monetary and news components using several daily interest rates. The nominal and real Treasury yields are sourced from FRED. Data on macroeconomic aggregates, such as the industrial production index and the consumer price index, are also sourced from FRED.
\section{Methodology} \label{sec:methodology}
This section of the paper details the three-stage process by which we extract the informational component of the high frequency movements around FOMC announcements. The three-stages can be summarized as follows:
\begin{itemize}
\item[Stage 1:] Document Embedding. We use state-of-the-art tools in the fields of natural language processing and machine learning to convert text documents in numeric vectors.
\item[Stage 2:] Constructing Proxy Expectations. We use the vector representations from Stage 1 to predict policy decisions of the FOMC. Under a rational expectation assumption, we interpret our estimated functions as expected policy decisions given the information sets of the private sector and the FOMC respectively.
\item[Stage 3:] Extracting Informational Content. Suppose market movements around FOMC announcements is due to informational effects. Then these movements should be proportional to the difference in expectations. We project high frequency movements onto the difference in expectation to recover their informational component.
\end{itemize}
In the following subsections, we provide further details on their implementation, as well as the assumptions that justify our procedure.
\newpage
Relative to the literature, the main novelty of our approach is in the use of newspaper articles to measure public expectation. Whereas the literature has thus far focused on the FOMC statements alone (see for instance \cite{handlan2020text}, \cite{hansen2016shocking}), we argue below that the information effect should depend on the difference in the expectations of the public and the FOMC, and that neglecting public expectation leads to omitted variable bias. In addition, with the exception of stage 1, which relies on a pre-trained model, the remaining stages involve only simple methods that are highly transparent. Finally, our method of constructing the intermediate expectation term allows us to probe at the mechanism by the information effect operate, as we explore in section \ref{sec:taylorrule}.
\subsection{Stage 1: Document Embedding} \label{subsec:first_stage}
In stage 1, we convert each FOMC statement and NYT article in $768\times1$-vectors using BERT (Bidirectional Encoder Representations from Transformers, \cite{devlin2019bert}). We highlight the challenge of using text data and explain how our BERT overcomes these problems before detailing our procedure.
Text data is difficult to analyze because of their high-dimensional nature. The simplest approach to deal with text is to just treat them as ``bag-of-words". With such an approach, the analyst would count the number of unique words in each article or document and then use the word counts as covariates for linear regressions or as input into other models. However, such an approach would work poorly in our sample, which has over 8,000 unique words in fewer than 150 observations. Whereas earlier papers have made progress by choosing which words to include (see for instance \cite{lucca2009measuring}), such ad-hoc methods not only have high researcher degree-of-freedom but could also leave information on the table.
The high dimensionality problem is compounded by the fact that the meanings of words depend on their context. An FOMC statement that is ``worried about inflation not growth" and one is ``worried about growth not inflation" have identical word counts but express contrary ideas.\footnote{A more subtle example of this ``lexical ambiguity" problem is that ``bank" has completely different meanings in ``river bank" and ``investment bank".} The need to take into account sequential information increases the dimension of text data exponentially. Consider a simple exercise in which we count the number of unique 6-word phrases, so that the example phrases above are counted as distinct entries. In a sample with 8,000 unique words, the possible number of unique phrases could be as high as $8,000^6 = 2.6 \times 10^{23}$. It is therefore largely infeasible to work with counts of phrases alone.
To render text analysis tractable and effective, we turn to BERT -- a state-of-the-art document embedding method that is able to reduce text input into $768\times1$-vectors while preserving sequential information. Its output captures the meaning of text input very well, as demonstrated by their usefulness for natural language processing tasks. For example, models based on BERT excel in question and answering tasks, in which given a question and a passage from Wikipedia containing the answer, the task is to predict the answer text span in the passage. Succeeding in these tasks all require some ``understanding" of the text input, giving us assurance that the BERT representations capture the meaning of words.
The key advantage of using BERT is that it is pre-trained on large amounts of text data (BooksCorpus -- 800M words and English Wikipedia -- 2,500M words). Given only 106 observations, it would have been very difficult to train any reasonably flexible text analysis models. For a similar reason, \cite{handlan2020text} relies on the pre-trained XLNet, a minor variation of BERT, for their analysis.
We note that other papers have employed Latent Dirichlet Allocations (LDA) models on samples of similar sizes [all the Hansen and McMahon papers], even though it is unclear if they are well-behaved given such small samples. In addition, it is known that LDA models have identification issues so that results are driven by the choice of priors even asymptotically (\cite{ke2021text}). Hence, we consider them to be unsuitable for the task at hand.
We initialized pretrained BERT models from an open source implementation provided by Hugging Face.\footnote{Hugging Face is a large open-source community that hosts pre-trained deep learning models, mainly aimed at NLP. Their state-of-the-art transformer models are used by Google, Facebook, Microsoft and AWS} (https://huggingface.co/bert-base-uncased) To use BERT, we first cleaned the input data by removing numbers, dates as well as common stop-words. Stop-words are a set of words which do not contain information about the content of the articles, such as prepositions and conjunctions. Removing these words ``enriches" the information content in text articles. Here, we relied on the stop-word list compiled by the text analysis package \emph{gensim}. Next, the words were tokenized. A token is the most fundamental object in a chosen language dictionary (it could be a word, character, or subword). We used the default tokenizer provided by Hugging Face for use with the bert-base-uncased model for tokenization.
Finally, we split each individual document into windows of 256 tokens, with overlaps of 10 tokens between windows. Each window is tokenized separately, and passed through the BERT model to form individual outputs. The embedding vector of the entire document is then taken to be the mean of the BERT class tokens across all of the windowed outputs.
We note that the cleaning process removes all numbers from the text. Consequently, the actual policy decisions as well as numerical projections about the economy are absent from the documents. This gives us assurance that our method is not simply picking up the policy rates or their forecasts, but rather capturing beliefs about the economy.
\subsection{Stage 2: Constructing Proxy Expectations} \label{subsec:second_stage}
In the second stage, we use the document embeddings to construct proxy expectations for the general public and the FOMC. Our starting point is the equations:
\begin{align*}
\text{FFR}_t & = f^\text{NYT}(X^\text{NYT}_t) + \varepsilon^\text{NYT}_t \\
\text{FFR}_t & = f^\text{FOMC}(X^\text{FOMC}_t) + \varepsilon^\text{FOMC}_t
\end{align*}
where $X^\text{NYT}_t$ and $X^\text{FOMC}_t$ are the embeddings of NYT articles and FOMC statement at time $t$. We further assume that
\begin{equation} \label{eq:news_exogeneity}
\mathbb{E}\left[\varepsilon^\text{NYT}_t \, \big| \, X^\text{NYT}_t\right] = 0 \quad , \quad \mathbb{E}\left[\varepsilon^\text{FOMC}_t \, \big| \, X^\text{FOMC}_t\right] = 0~.
\end{equation}
We interpret the above assumption as follows. Suppose $X^\text{NYT}_t$ captures all information about the economy that is available to the general public at time $t$. Then public expectation about FFR is given by
\begin{equation*}
f^\text{NYT}(X^\text{NYT}_t) = \mathbb{E}\left[\text{FFR}_t \, \big| \, X_t^\text{NYT}\right] ~.
\end{equation*}
The conditional mean assumption
\begin{equation*}
\mathbb{E}\left[\varepsilon^\text{NYT}_t \, \big| \, X^\text{NYT}_t\right] = 0
\end{equation*}
is thus a rational expectation assumption -- the general public has beliefs that are not systematically biased and are correct on average.
We assume the same is true for the FOMC. That is, we assume that $X^\text{FOMC}_t$ contains all the information that the FOMC has about the economy at time $t$. $f^\text{FOMC}(X^\text{FOMC}_t)$ is then the Taylor rule that maps FOMC information into interest rate decisions. $\varepsilon^\text{FOMC}_t$ is monetary policy shock -- implementation error that prevents the FOMC from achieving its desired optimal policy rate. We assume that this monetary policy shock is $0$ on average and mean independent from $X_t^\text{FOMC}$. This is consistent with the current views on monetary policy shock, as discussed in \cite{ramey2016macroeconomic}.
Given our limited sample size, we impose further structure for estimation by assuming that $f^\text{NYT}$ and $f^\text{FOMC}$ take the following linear form:
\begin{align*}
f^\text{NYT}(X_t^\text{NYT}) & = \alpha^\text{NYT} + \beta^\text{NYT}{}'X^\text{NYT}_t \\
f^\text{NYT}(X_t^\text{NYT}) & = \alpha^\text{FOMC} + \beta^\text{FOMC}{}'X^\text{FOMC}_t
\end{align*}
Because we use the pre-trained BERT, $X_t^\text{NYT}, X_t^\text{FOMC} \in \mathbf{R}^{768}$. In order to estimate $\alpha$ and $\beta$ from our sample of size 106, we use the elastic net, defined as:
\begin{equation*}
\left(\hat{\beta}^\text{NYT}, \hat{\alpha}^\text{NYT} \right) = \underset{\beta, \alpha}{\arg\min} \left\lVert \text{FFR} - \alpha - \beta'X^\text{NYT} \right\rVert_2^2 + \gamma\eta \left( \lVert \alpha \rVert_1 + \lVert \beta \rVert_1\right) + \gamma(1-\eta) \left( \lVert \alpha \rVert^2_2 + \lVert \beta \rVert^2_2\right)
\end{equation*}
\begin{equation*}
\left(\hat{\beta}^\text{FOMC}, \hat{\alpha}^\text{FOMC} \right) = \underset{\beta, \alpha}{\arg\min} \left\lVert \text{FFR} - \alpha - \beta'X^\text{FOMC} \right\rVert_2^2 + \gamma\eta \left( \lVert \alpha \rVert_1 + \lVert \beta \rVert_1\right) + \gamma(1-\eta) \left( \lVert \alpha \rVert^2_2 + \lVert \beta \rVert^2_2\right)
\end{equation*}
Here, $\text{FFR}$, $X^\text{NYT}$ and $X^\text{FOMC}$ refer to the stacked matrices of $\text{FFR}_t$, $X_t^\text{NYT}$ and $X_t^\text{FOMC}$ respectively.
The elastic net nests two popular methods of penalized regressions. Suppose $\eta = 1$. Then the elastic net reduces to the LASSO estimator. LASSO performs model selection by setting regression coefficients to zero. The resulting output is typically an interpretable model in which only a few coefficients are non-zero. On the other hand, setting $\eta = 0$ leads to the ridge estimator, a dense model which assigns highly correlated regressors similar coefficients. Since the elastic net is more general, it is able to achieve better prediction accuracy given properly chosen tuning parameters (\cite{zou2005regularization}). Furthermore, it has a Bayesian interpretation as the posterior mode induced by the spike-and-slab prior of \cite{mitchell1988bayesian} and may be more appropriate for economic data (\cite{giannone2021sparsity}).
With our conditional mean assumptions, the elastic net is $L_2$ consistent as long as the space of the covariates is sufficiently rich (\cite{de2009elastic}). Since we are interested in prediction and not model selection, we do not require any sparsity assumption.
Each elastic net requires two tuning parameters $(\lambda, \eta)$. Because these parameters are chosen to maximize $R^2$ in stage 3, we discuss their selection in the next subsection.
\subsection{Stage 3: Extracting Informational Content} \label{subsec:third_stage}
In this subsection, we explain how we use the proxy expectations from stage 2 to decompose high frequency movements of interest rates around FOMC announcements into an information and a monetary component.
Let $\Delta R_t$ denote the high frequency movement of interest rate $R$ around the release of FOMC statements. Suppose $\Delta R_t$ includes reaction by the private sector to information released by the Federal Reserve. This component should be a function to the \emph{difference} in information of the two parties. In particular, if there is no difference in information, this component should be $0$. This motivates the following linear regression:
\begin{equation}\label{eq:stage3}
\Delta R_t = \zeta + \theta \Delta \mathbb{E}[\text{FFR}_t] + \nu_t \quad , \quad \mathbb{E}\left[\nu_t \, | \, \Delta \mathbb{E}[\text{FFR}_t]\right] = 0,
\end{equation}
where
\begin{equation*}
\Delta \mathbb{E}[\text{FFR}_t] := \alpha^\text{FOMC} + \beta^\text{FOMC}{}'X^\text{FOMC}_t - \alpha^\text{NYT} + \beta^\text{NYT}{}'X^\text{NYT}
\end{equation*}
is the difference in the expectations of the private sector and the FOMC based solely on difference in information.
For any choice of tuning parameters $(\lambda, \eta)$ in stage 2, we are able to construct
\begin{equation*}
\widehat{\Delta \mathbb{E}[\text{FFR}_t]} := \hat{\alpha}^\text{FOMC} + \hat{\beta}^\text{FOMC}{}'X^\text{FOMC}_t - \hat{\alpha}^\text{NYT} + \hat{\beta}^\text{NYT}{}'X^\text{NYT}~.
\end{equation*}
We then choose our tuning parameter to maximise $R^2$ in the regression of $\Delta R_t$ on $\widehat{\Delta \mathbb{E}[\text{FFR}_t]}$, where the latter term is a function of $(\lambda, \eta)$.
The regression in equation (\ref{eq:stage3}) has clear meaning. Here, we are projecting the high frequency movement onto the differences in proxy expectation and we interpret the projected component as the part of $\Delta R_t$ that is responding to Federal Reserve information.
Suppose we were able exactly capture news component of $\Delta R_t$. We would then be able to interpret the residual in the above regression as true monetary policy shocks. However, we are cautious that much of the news component may remain due in part to the restrictive linear assumption in equation (\ref{eq:stage3}). On the other hand, given the sample size of 106, it is unclear if more flexible functional forms would simply lead to over-fitting. For this reason, the simple linear regression is our preferred specification for stage 3.
An approach commonly seen in the literature (see for instance \cite{handlan2020text}, \cite{hansen2016shocking}) is to run the regression on the FOMC component only:
\begin{equation}\label{eq:shortregression}
\Delta R_t = \zeta + \theta \cdot f^\text{FOMC}(X^\text{FOMC}_t) + \xi_t \quad
\end{equation}
As we argued above, the informational component of the high frequency movement should depend on the difference in information, and not the level directly. Suppose equation \ref{eq:stage3} is true. We can rewrite equation \ref{eq:shortregression} as
\begin{equation*
\Delta R_t = \zeta + \theta \cdot f^\text{FOMC}(X^\text{FOMC}_t) - \theta \cdot f^\text{NYT}(X^\text{NYT}_t) + \nu_t \quad , \quad \mathbb{E}\left[\nu_t \, | X^\text{FOMC}_t\right] = 0.
\end{equation*}
Here,
\begin{equation*}
\xi_t = - \theta \cdot f^\text{NYT}(X^\text{NYT}_t) + \nu_t
\end{equation*}
is clearly correlated with $f^\text{FOMC}(X^\text{FOMC}_t)$ through $f^\text{NYT}(X^\text{NYT}_t)$, giving rise to omitted variable bias.
Furthermore, equation (\ref{eq:stage3}) is robust in that it can be valid even when the conditions in equation (\ref{eq:news_exogeneity}) is violated. For example, suppose the FOMC has time varying preferences in trading off inflation and output gap, so that a positive shock is more likely when the FOMC preferences have a hawkish realization. This violates the conditional mean assumption in equation (\ref{eq:shortregression}) as long as the statements contains information about its current hawkishness. However, time-varying preferences, or other omitted variables, do not cause a problem for the regression in (\ref{eq:stage3}) as long as: $$\text{Cov}\left(\nu_t, f^\text{FOMC}(X^\text{FOMC}_t)\right) = \text{Cov}\left(\nu_t, f^\text{NYT}(X^\text{NYT}_t)\right)~.$$
Intuitively, our regression is valid as long as these omitted variables skew private sector and FOMC expectations the same way. On the other hand, if equation (\ref{eq:news_exogeneity}) does not hold, \ref{eq:shortregression} will be invalid, even if $X_{t}^\text{FOMC} \perp \!\!\! \perp X_t^\text{NYT}$.
Finally, even if $X_{t}^\text{FOMC} \perp \!\!\! \perp X_t^\text{NYT}$, our specification will still have more power. Including $f^\text{NYT}(X^\text{NYT}_t)$ would lead to more precise estimates by reducing the variance of the residuals. This gels with the idea that our regression is a more direct test of the information hypothesis: if the high frequency movement is driven by agents updating their information set, theory directly predicts that these movement should be proportional to the amount of updating that occurred.
Next, we provide intuitive reasons for choosing tuning parameters that maximize $R^2$ in stage 3. Firstly, the tuning parameters that are chosen are unlikely to overfit the stage 2 elastic nets. To see this, observe that if we overfit stage 2 to the extreme, such that $\hat{\alpha}^\text{FOMC} + \hat{\beta}^\text{FOMC}{}'X^\text{FOMC}_t = \hat{\alpha}^\text{NYT} + \hat{\beta}^\text{NYT}{}'X^\text{NYT}_t = \text{FFR}_t$. Then $\widehat{\Delta \mathbb{E}[\text{FFR}_t]} = 0$ identically so that stage 3 $R^2$ is 0. It is equally unlikely to underfit stage 2. Consider again the extreme case, this time in which $\hat{\beta}^\text{NYT}$ and $\hat{\beta}^\text{FOMC}$ are set to $0$. Now, $\hat{\alpha}^\text{NYT}$ = $\hat{\alpha}^\text{FOMC} = \frac{1}{T} \sum_{t = 1}^T \text{FFR}_t$ so that $\widehat{\Delta \mathbb{E}[\text{FFR}_t]}$ is again $0$. Our procedure guards against both of these problems. Where the method may face issues is in overfitting stage 3. However, we believe that this is less of a concern since stage 3 is a linear regression with only two parameters.
\subsection{Results}
\begin{table}[htbp]
\centering
\caption{Summary Statistics for Stage 3. Robust standard errors are used.}
\begin{tabular}{ccccccc}
& Shock & $R^2$ & $\hat{\theta}$ & S.E. & $t$-stat & $p$-value \\
\midrule
\multirow{3}[2]{*}{Pre-ZLB} & PNS & 0.079 & 0.083 & 0.031 & 2.624 & 0.011 \\
& FFR & 0.038 & 0.051 & 0.044 & 1.152 & 0.254 \\
& FF4 & 0.085 & 0.183 & 0.065 & 2.826 & 0.006 \\
\midrule
\multirow{3}[1]{*}{Full Sample} & PNS & 0.028 & 0.009 & 0.006 & 1.466 & 0.146 \\
& FFR & 0.013 & 0.008 & 0.008 & 0.971 & 0.334 \\
& FF4 & 0.073 & 0.020 & 0.008 & 2.421 & 0.017 \\
\end{tabular}%
\label{tab:stage3}%
\end{table}%
Summary statistics for Stage 3 of our procedure is presented in table \ref{tab:stage3}. We find evidence that high frequency movements around FOMC announcements is in part a response to information from the Federal Reserve. In particular, we note $\hat{\beta}$ is positive for PNS, FFR and FF4. It is also significant at the 5\% level for the PNS and FF4 in the pre-ZLB period, and for FF4 in the full sample. The positive sign implies that when the information set of the FOMC leads it to set higher interest rate than the public was expecting, the market response is positive.
The $R^2$ in stage 3 is relatively low across the board. This could be interpreted as meaning that most of the movement around FOMC announcements is response to pure monetary policy shock. On the other hand, given the our highly restrictive linear model in stage 3, we consider the information effect we found to be a lower bound: with more data and a more flexible stage 3, we might find potentially much larger information effects.
\begin{table}[htbp]
\centering
\caption{Summary Statistics for Stage 2}
\begin{tabular}{cccccccc}
& Shock & $\eta^\text{NYT}$ & $\lambda^\text{NYT}$ & NYT $R^2$ & $\eta^\text{FOMC}$ & $\lambda^\text{FOMC}$ & FOMC $R^2$ \\
\midrule
\multirow{3}[2]{*}{Pre-ZLB} & PNS & 0.550 & 1.150 & 0.122 & 0.250 & 3.275 & 0.082 \\
& FFR & 0.700 & 0.811 & 0.152 & 0.050 & 13.219 & 0.147 \\
& FF4 & 0.400 & 1.630 & 0.068 & 0.400 & 2.310 & 0.060 \\
\midrule
\multirow{3}[1]{*}{Full Sample} & PNS & 1.000 & 0.012 & 0.884 & 1.000 & 0.100 & 0.995 \\
& FFR & 0.650 & 0.050 & 0.872 & 0.800 & 0.142 & 0.974 \\
& FF4 & 0.100 & 0.142 & 0.884 & 1.000 & 0.100 & 0.990 \\
\end{tabular}%
\label{tab:stage2}%
\end{table}%
Table \ref{tab:stage2} presents summary statistics associated with stage 2, including the tuning parameters chosen by our procedure. Larger $\eta$ implies that the chosen model is closer to a LASSO, while smaller $\eta$ implies that the chosen model is closer to a ridge regression. We see evidence for sparsity in PNS and FF4 with the FOMC statement, especially in the full sample period, since the preferred model is the LASSO. This suggests that there are only a few variables which are important for determining the FOMC's desired policy rate. On the other hand, during the pre-ZLB period, the selected models tended towards ridge regressions, suggesting that there are more diverse considerations for determining interest rates prior to the Great Financial Crisis.
We note that $R^2$ is high in the full sample, but much smaller in the the pre-ZLB sample. This is likely because FFR remained largely constant after the Great Financial Crisis. Taking the model at face value, we would infer that monetary policy is easier to predict after post-crisis, and that the variance of monetary policy shocks is smaller. On the other hand, we are cautious that our linear model may be inappropriate for modelling the policy rates, which may be censored at the zero lower bound.
We conclude by noting that our approach is highly transparent. All tuning parameters as well as regression parameters that we chose are contained in tables \ref{tab:stage3} and \ref{tab:stage2}. The only unpresented parameters belong to BERT, which is trained by an external party without reference to any of the variables used in stage 2 and 3.
\subsection{Channels of Information Transmission}\label{sec:taylorrule}
In order to understand how new information from the FOMC translates to market reaction, we consider how differences in expectation of specific macroeconomic variables drive high frequency movements around FOMC announcements. To do that, we extend the analysis thus far to a general economic variable $Y_i$:
\begin{align*}
Y_{i,t} & = f_i^\text{NYT}(X^\text{NYT}_t) + \varepsilon^\text{NYT}_{i,t} \\
Y_{i,t} & = f_i^\text{FOMC}(X^\text{FOMC}_t) + \varepsilon^\text{FOMC}_{i,t}
\end{align*}
where as before we assume that:
\begin{equation*}
\mathbb{E}\left[\varepsilon^\text{NYT}_{i,t} \, \big| \, X^\text{NYT}_t\right] = 0 \quad , \quad \mathbb{E}\left[\varepsilon^\text{FOMC}_{i,t} \, \big| \, X^\text{FOMC}_t\right] = 0~.
\end{equation*}
However, the stage 3 equation of interest is
\begin{equation}\label{eq:stage3_taylor}
\Delta R_t = \zeta + \sum_{i=1}^p \theta_i \cdot \Delta \mathbb{E}[Y_{i,t}] + \nu_t \quad , \quad \mathbb{E}\left[\nu_t \, | \, \Delta \mathbb{E}[Y_t]\right] = 0.
\end{equation}
Suppose $Y_{i,t} = \pi_t$ is inflation at $\pi_t$. In the regression above, we would then interpret $\theta_i$ as the high frequency movement that resulted from differences in inflation expectation. Running regression \ref{eq:stage3_taylor} allows us to understand what type of informational differences is driving the high frequency movement.
In particular, let $Y_{1,t}$ be real GDP in the next nearest quarter and $Y_{2,t}$ be CPI in the next nearest month.\footnote{In particular, we use the ``Consumer Price Index: All Items Excluding Food and Energy''. This choice is motivated by the FOMC targetting a core inflation rate which excludes commodity prices such as Food and Energy which are highly influenced by international markets.} These are the two variables that would be relevant if interest rate was set according to a simple Taylor rule. Results are presented in table \ref{tab:stage3_taylorrule}. Across the board, when the FOMC expects higher inflation than the private sector, the interest rates increase following the announcements. This is in line with the predictions of standard macroeconomic theory, in which the central bank raises rates in response to heightened inflationary pressures regardless of whether the source is demand or supply shocks.
The coefficients on real GDP is mixed. In the pre-ZLB period, the coefficients on real GDP have positive signs across the board. In words, when the FOMC expects higher real GDP than the public, interest rates subsequently increases. In the full sample, however, the sign changes to negative for FFR and FF4. Standard macroeconomic theory does not provide a straightforward answer to how monetary authority should react to GDP fluctuations. For instance, if the central bank receive news of a positive demand shock, then it raises its GDP projections and simultaneously raises interest rates. This gives rise to a positive comovement between expected GDP and monetary rate policy. But, if it receives news of a negative supply shock, a central bank would raise rates even as it lowers GDP projections, leading to negative comovements. Note that the full sample includes episodes of \textit{supply-side} shocks such as tax cuts and/or oil price shocks; and furthermore, we exclude observations from July 2008 to July 2009 inclusive. Hence, it is plausible that in our full sample supply shocks play a large role.
In sum, our method allows us to probe at the mechanisms of information transmission. We find that when the FOMC expects higher inflation than the private sector, interest rates unequivocally increase. However, when the FOMC expects higher real GDP than the private sector, interest rate movements are mixed. These patterns are suggestive of the types of shocks in an economy.
\begin{table}[htbp]
\centering
\caption{Stage 3 Results using real GDP and CPI. Robust standard errors are used.}
\begin{tabular}{clccccc}
& \multicolumn{1}{c}{Shock} & $R^2$ & $\hat{\theta}_\text{RGDP}$ & RGDP $p$-value & $\hat{\theta}_\text{CPI}$ & CPI $p$-value \\
\midrule
\multirow{3}[2]{*}{Pre-ZLB} & PNS & 0.103 & 0.001 & 0.073 & 0.005 & 0.131 \\
& FFR & 0.083 & 0.000 & 0.378 & -0.080 & 0.256 \\
& FF4 & 0.216 & 0.030 & 0.000 & 0.032 & 0.440 \\
\midrule
\multirow{3}[1]{*}{Full Sample} & PNS & 0.061 & 0.000 & 0.548 & 0.003 & 0.174 \\
& FFR & 0.034 & -0.001 & 0.139 & 0.005 & 0.134 \\
& FF4 & 0.067 & -0.002 & 0.051 & 0.008 & 0.055 \\
\end{tabular}%
\label{tab:stage3_taylorrule}%
\end{table}%
\section{Evidence of Informational Content} \label{sec:main_results}
In this section, we provide evidence that the component we isolated using the procedure in Section \ref{sec:methodology} carries informational content. We focus on event study and financial market evidence.\footnote{In Appendix \ref{subsec:svar_evidence} we present additional evidence by estimating the impact of ``Monetary$_t$'' at the aggregate level using a SVAR model. Expectedly, these are not too different from SVARs using the high frequency shocks alone, given the $R^2$ of Stage 3.}
\subsection{Event Study Analysis} \label{subsec:event_analysis}
We study the components of the high-frequency financial surprises isolated in the previous section.
\begin{figure}[ht]
\centering
\caption{Time Series of PNS$_t$ Decomposition. Following \cite{nakamura2018high} we omit observations between July 2008 to July 2009 inclusive.}
\label{fig:TimeSeriesPNS}
\includegraphics[width=12cm,height=9cm]{TimeSeriesPNS.png}
\end{figure}
Given the modest $R^2$ from the decomposition, it is unsurprising that the news component has relatively smaller variance than the monetary component. However, the series is intuitively reasonable. To show this, we zoom in on three FOMC announcements in which the news component was thought to have played an important role ex post.
\begin{itemize}
\item[] \textbf{September 18, 2007}: We return to our opening example on which the FOMC cut its rates by 50 basis points. As our quoted text suggests, the market interpreted the Federal Reserve's fairly aggressive actions and statements to mean that a sharp economic downturn was imminent. Thus it is reassuring that we recorded one of our largest negative news shock on this date.
\item[] \textbf{March 18, 2008}: One of our larger positive news shock was documented on March 18, 2008, even amidst the Great Financial Crisis. On this day, the FOMC cut rates by 75 basis points to 2.25 percent. However, as the New York Times reports, ``though it was one of the biggest one-day rate cuts in decades, investors had been betting heavily that the Fed would cut its key rate a full percentage point in response to strong evidence that a recession has begun and to the deepening crisis on Wall Street." Instead, the Federal Reserve policy reflected concerns about inflation. As the FOMC statement writes, ``inflation has been elevated, and some indicators of inflation expectations have risen". It is therefore reassuring that our measure of news recorded a positive value, reflecting the Federal Reserve's information about inflation.
\item[] \textbf{January 30, 2002}: Another of our largest positive surprises came as the economy was recovering from the bursting of the dot-com bubble and 9-11. At this meeting, the FOMC left interest rates unchanged, which was a departure from its yearlong policy of rate cuts. According to the FOMC, ``Signs that weakness in demand is abating and economic activity is beginning to firm have become more prevalent". As the New York Times reports, ``Investors interpreted [...] the Fed's decision on interest rates as signs that the economy might be stabilizing. Stocks prices moved higher in the afternoon after the Fed's announcement." Here, we have explicit indication that investors interpreted the Federal Reserve's policy as good news, supporting the positive news shock that we recovered.
\end{itemize}
\subsection{Financial Market Evidence} \label{subsec:financial_evidence}
The goal of this subsection of the paper is to identify the effect of the informational content of FOMC announcements on nominal and real interest rates of different maturities.
First, we measure the impact of the informational content of FOMC announcements on nominal and real interest rates. In particular, we estimate the following two empirical equations
\begin{align}
\Delta r_t &= \alpha_0 + \beta \Delta R_t + e_t \label{eq:shock} \\
\Delta r_t &= \alpha_1 + \gamma \text{Monetary}_t + \mu \text{News}_t + \epsilon_t \label{eq:shock_decomposition}
\end{align}
were $\Delta r_t$ is the change in the outcome nominal and/or real interest rate of interest, e.g., the Market Yield on U.S. Treasury Securities at 10-Year Constant Maturity; $\Delta R_t$ is the high-frequency financial surprise around the FOMC announcement; $\text{Monetary}_t$ and $\text{News}_t$ are the decomposition coming from the methodology used in Section \ref{sec:methodology}; $e_t$ and $\epsilon_t$ are error terms; and $\alpha_0$, $\alpha_1$, $\beta$, $\gamma$ and $\mu$ are parameters. The parameter of interests are $\beta$, $\gamma$ and $\mu$. Comparing $\beta$ and $\gamma$ allows us to identify how purging the high-frequency surprises of informational content changes the impact of monetary policy shocks on nominal and real interest rates.
First, we study the impact of the high-frequency Fed Funds Futures on nominal and real rates by estimating Equation \ref{eq:shock} using the high-frequency surprise FF4.
\begin{table}[ht]
\centering
\caption{Response of Interest Rates to Fed Funds Futures}
\label{tab:ff4_rates}
\begin{tabular}{l c c c c}
& (1) & (2) & (3) \\
& FF4$_t$ & $Monetary_t$ & $News_t$ & Observations \\ \hline \hline
& & & & \\
Nominal 3 month & 0.300 & 0.265 & 0.797*** & 106 \\
& (0.218) & (0.234) & (0.248) & \\
Nominal 1 year & 0.486*** & 0.494*** & 0.376 & 106 \\
& (0.128) & (0.123) & (0.316) & \\
Nominal 2 years & 0.571*** & 0.569*** & 0.594 & 106 \\
& (0.209) & (0.210) & (0.509) & \\
Nominal 5 years & 0.408* & 0.240 & 0.432 & 106 \\
& (0.244) & (0.406) & (0.711) & \\
Nominal 10 years & 0.149 & 0.143 & 0.235 & 106 \\
& (0.178) & (0.184) & (0.593) & \\
Nominal 20 years & -0.00137 & 0.000402 & -0.0261 & 106 \\
& (0.135) & (0.138) & (0.547) & \\
TIPS Real 2 years & -0.0473 & -0.0521 & 0.0192 & 74 \\
& (0.131) & (0.139) & (0.530) & \\
TIPS Real 5 years & 0.126 & 0.151 & -0.219 & 82 \\
& (0.191) & (0.175) & (0.824) & \\
TIPS Real 10 years & 0.0124 & 0.0295 & -0.226 & 82 \\
& (0.152) & (0.140) & (0.642) & \\
TIPS Real 20 years & 0.0401 & 0.0355 & 0.0954 & 70 \\
& (0.162) & (0.172) & (0.587) & \\
& & & & \\ \hline \hline
\multicolumn{5}{c}{ Robust Standard Errors in Parentheses} \\
\multicolumn{5}{c}{ *** p$<$0.01, ** p$<$0.05, * p$<$0.1} \\
\end{tabular}
\end{table}
Table \ref{tab:ff4_rates} presents the results of this regressions. Column (1) presents the empirical estimates of coefficient $\beta$ in Equation \ref{eq:shock} for different nominal and real rates. Results seem to show a hump-shaped impact of FF4$_t$ on nominal rates, with no significant impact on the three month nominal rate, a positive impact on nominal rates at the one to 10 year maturities (peaking at 2 years), and no impact on long-term rates. Surprisingly, there does not seem to be a significant impact of FF4$_t$ on the daily change of real interest rates captured by TIPS.\footnote{This result could be explained by different reasons. For example, given the policy change on FOMC statements in the year 2000 and the data availability of text, we have a short sample. This problem worsens for TIPS treasuries as they were introduced later in time.}
Next, we turn to estimating the impact of the different isolated components of FF4$_t$ estimated in Section \ref{sec:methodology}. Columns (2) and (3) present empirical estimates of coefficients $\gamma$ and $\mu$ in Equation \ref{eq:shock_decomposition}, respectively, for different nominal and real interest rates. Interestingly, while the "\textit{Monetary}$_t$" component of FF4$_t$ does not seem to have a significant effect on the 3-month maturity treasury, the "\textit{News}$_t$" component has a significant impact on this short run interest rate. The "\textit{News}$_t$" component does not seem to have significant impact on mid and long-term treasury yields. Consequently, the estimated coefficients of the "\textit{Monetary}$_t$" component is similar to the coefficients presented in Column (1) for FF4$_t$.
Second, we estimate the impact of the "Policy News Shocks" or PNS$_t$ constructed by \cite{nakamura2018high} and its decomposition on nominal and real interest rates. Note, this policy indicator is comprised of changes in multiple nominal interest rates at different maturities spanning the first year of the term structure.\footnote{More specifically, the composite policy indicator is the first principal component of the unanticipated change over the 30-minute windows in the following five interest rates: the Fed funds rate immediately following the FOMC meeting, the expected Fed funds rate immediately following the next FOMC meeting, and expected three-month eurodollar interest rates at horizons of two, three, and four quarters.}
\begin{table}[ht]
\centering
\caption{Response of Interest Rates to Policy News Shocks}
\label{tab:PNS_rates}
\begin{tabular}{l c c c c}
& (1) & (2) & (3) \\
& PNS$_t$ & $Monetary_t$ & $News_t$ & Observations \\ \hline \hline
& & & & \\
Nominal 3 month & 0.670*** & 0.581*** & 1.208*** & 106 \\
& (0.141) & (0.150) & (0.273) & \\
Nominal 1 year & 0.795*** & 0.834*** & 0.558** & 106 \\
& (0.111) & (0.121) & (0.263) & \\
Nominal 2 years & 1.052*** & 1.157*** & 0.420 & 106 \\
& (0.203) & (0.207) & (0.394) & \\
Nominal 5 years & 0.929*** & 1.135*** & -0.310 & 106 \\
& (0.226) & (0.249) & (0.581) & \\
Nominal 10 years & 0.456** & 0.641*** & -0.655 & 106 \\
& (0.184) & (0.202) & (0.516) & \\
Nominal 20 years & 0.225 & 0.333* & -0.421 & 106 \\
& (0.169) & (0.185) & (0.431) & \\
TIPS Real 2 years & 1.057*** & 1.129*** & 0.138 & 74 \\
& (0.239) & (0.241) & (0.560) & \\
TIPS Real 5 years & 0.787*** & 0.944*** & -1.729** & 82 \\
& (0.249) & (0.268) & (0.720) & \\
TIPS Real 10 years & 0.543** & 0.660*** & -1.332* & 82 \\
& (0.210) & (0.224) & (0.673) & \\
TIPS Real 20 years & 0.317 & 0.420* & -0.850 & 70 \\
& (0.210) & (0.231) & (0.643) & \\
& & & & \\ \hline \hline
\multicolumn{5}{c}{ Robust Standard Errors in Parentheses} \\
\multicolumn{5}{c}{ *** p$<$0.01, ** p$<$0.05, * p$<$0.1} \\
\end{tabular}
\end{table}
A priori, this could imply greater variability of PNS$_t$, compared to FF4$_t$. Additionally, in \cite{nakamura2018high} the authors stress that this composite policy indicator captures the effects of “forward
guidance” (FOMC announcements that convey information about future changes in the Fed Funds rate). Column (1) of Table \ref{tab:PNS_rates} presents the empirical estimates of coefficient $\beta$ in Equation \ref{eq:shock}. While results are in line with those for FF4$_t$, the estimates are highly more significant. Again, we find that the empirical estimates for $\beta$ exhibits a hump-shaped impact on nominal interest rate across the term structure. Unlike the results presented in \ref{tab:ff4_rates}, the policy news shock has a significant impact on TIPS yields up to the 10 year maturity rate.
Finally, we turn to estimating the role of the "Monetary$_t$" and "News$_t$" components of Policy New Shocks. Once more, the "News$_t$" component has a large and significant impact on short term nominal interest rates. One way to quantify the relevance of the news component is comparing the estimated coefficients for the "Monetary$_t$" component and those for PNS$_t$. Not controlling for the news component biases up the impact on nominal interest rate between 10 and 5 basis points for short term nominal interest rate, and biases the impact down between 15 and 30 basis points for mid and long term nominal interest rates. In addition, the impact of "Monetary$_t$" on TIPS real rates is greater than the impact of PNS$_t$. Interestingly, the news component which has a large and significant impact on short run nominal interest rate, also seems to have a large but negative impact on long run real rates.
In summary, the paragraphs above studied the impact of high-frequency financial surprises and its decomposition on the daily change of nominal and real rates around FOMC meetings. We find that the news component of these high-frequency surprises has a significant impact on short term nominal interest rates across different specified shocks. We take these results as evidence that differences in the expected policy rate by the FOMC Statements and the New York Times has a systematic and persistent impact on relevant financial rates.
\section{Conclusion} \label{sec:conclusion}
We argue that FOMC announcements contain both information and monetary policy content. We develop a transparent framework to pick out the informational content of high-frequency financial surprises around FOMC announcements by using state-of-the art text analysis on both policy statements and newspapers.
We find that differences in expectations about policy rates between the FOMC and private agents are predictive of high-frequency movements around policy announcements. Furthermore, we show that the isolated informational content has a significant impact on the yields of short-term government bonds. Consequently, ignoring this informational content may bias our understanding of the impact of monetary policy shocks.
We interpret these findings as an existence result. In order to draw reasonable conclusions from a small dataset, we have prioritized simplicity and transparency over more flexible functional forms. Hence, we consider our estimates to be lower bounds for the informational content in FOMC announcements.
\newpage
\section{Introduction} \label{sec:introduction}
The identification and macroeconomic impact of monetary policy shocks is a longstanding and open question. The key challenge is that monetary policy responds to both the current and the expected future states of the macroeconomy. Controlling for these unobserved state variables underlies the main difficulty in identifying the causal impact of monetary policy. One way to overcome this challenge, is to measure changes in the prices of financial assets around Central Bank announcements. Under tight time-windows around Central Bank's announcements, it is credible to assume that financial markets react to said announcements only. However, as stressed by \cite{nakamura2018high} and \cite{jarocinski2020deconstructing} among others, these announcements reveal both policy decisions and the central bank's assessment about the present and future economic outlook. Consequently, these high-frequency movements in the prices of financial assets are a response not only to policy decisions but also to the private sector's updated beliefs about the state of the economy. Consider the following example of both policy and private market beliefs around a FOMC meeting.
On September 18, 2007, the FOMC cut the Federal Funds Rate by 50 basis points. As the New York Times reports, ``while an interest rate cut was widely expected, there had been profound uncertainty about whether the Fed would choose a more cautious quarter-point reduction. But the bolder action and an accompanying statement, both approved by a unanimous vote of the central bank’s policy-setting committee, made it clear that the Fed had decided the risks of a recession were too big to ignore." Hence, we see how investors learnt from the aggressive policy decision about the seriously negative prospects facing the economy.
In order to study the causal effects of monetary policy, we would like to decompose high-frequency movements into the part which responds to monetary policy (the monetary component) and the part which responds to information disclosure by the central bank (the information or news component). Studying the effect of the monetary component would allow us to understand the effect of monetary policy, while the information component allows us to understand how central banks influence the economy beyond direct policy tools.
In this paper, we use FOMC statements and newspaper articles to identify and quantify the information component of monetary policy announcements. We do so via state-of-the art text analysis tools in a framework that is both transparent and highly interpretable. Whereas earlier literature has inferred the information component through the introduction of a structural model (see \cite{nakamura2018high}) or imposing restrictions on the moments of high-frequency surprises (see \cite{jarocinski2020deconstructing}), our method does not impose any assumptions on the relationship between the potential shocks and macroeconomic variables. Instead, we use text data to model expectations of the FOMC and the private sector over the Federal Funds Rate (FFR). Our key insight is that the informational component of monetary policy is the part of high frequency movements that is correlated with differences in the expectations of the two.
Our method consists of three stages. First, we use tools in computational linguistics to embed FOMC statements and New York Times articles as numerical vectors. Next, we use the embedded documents to separately model both FOMC and private sector expectations of the FFR via the elastic net, a transparent method for dealing with high dimensional data, and which \cite{giannone2021sparsity} advocates for economic data. Here, we use the expectation of the FOMC to refer to the hypothetical monetary policy rate that would have been chosen in the absence of monetary policy shocks. Stage two allows us to construct the differences in expectation about the FFR which arose due to differences in information alone. Finally, we project high frequency movements of select interest rates around FOMC announcements onto the differences in expectation and interpret the projection as the information component of the high frequency movements.
We apply our framework to a set of identified high-frequency movements which the literature has interpreted as monetary policy shocks.
Following \cite{nakamura2018high} and \cite{jarocinski2020deconstructing} we decompose these high-frequency movements into a monetary component and news component. We find that over our sample period of January 2000 to March 2014, differences in beliefs have predictive power over the high-frequency movement of the Federal Funds Futures around FOMC meetings. We interpret this finding as evidence that FOMC announcements disclose information about the state of the economy.
We show that across these different high-frequency movements, the isolated news component has predictive power over the daily change of short term nominal and real rates around FOMC meetings. In particular, we argue that not controlling for the news component of high-frequency surprises may lead to biases on the impact of monetary policy shocks on domestic financial conditions.
Our contribution to the literature is three-fold. First, we use newspaper text to proxy for private sector information about monetary policy around FOMC meetings. While text analysis has been used to study the information effects of FOMC policy announcements, existing papers have focused on FOMC statements or minutes exclusively (\cite{lucca2009measuring}, \cite{hansen2018transparency}, \cite{benchimol2020communication}, \cite{handlan2020text}). We argue that any informational effects has to operate through differences in expectations, so that both FOMC and private beliefs must be taken into consideration. Hence we augment our analysis of FOMC statements with New York Times articles.
Second, we propose a text analysis framework that is interpretable and highly transparent. Our framework uses the black-box output of text analysis methods to construct conditional expectations of various macroeconomic variables. Studying the effects of these variables -- or their differences -- on high frequency interest rate movements allows us to understand the transmission mechanisms of informational effects. Our method is also highly transparent: the elastic net, which is a form of penalized linear regression, performs explicit regularization using only two tuning parameters.
Third, the empirical applications of our framework shed light on the impact of the informational content of FOMC announcements on both financial and macroeconomic variables. Specifically, the news component of high-frequency monetary policy shocks have a significant impact on short run nominal and real rates, while having little to no effect on longer maturities. Further studying these shocks can provide insights on the specific information advantage the Federal Reserve has over private sector agents.
The paper is organized as follows. Section \ref{sec:data} describes the data. Section \ref{sec:methodology} lays out the text analysis framework in detail. Section \ref{sec:main_results} provides both qualitative and quantitative evidence that our procedure recovers informational content. Section \ref{sec:conclusion} concludes.
\section{Data} \label{sec:data}
This section of the paper describes the two main types of data used: (i) text coming from both the Federal Reserve's FOMC statements and New York Times' articles; (ii) high-frequency financial surprises around FOMC announcements.
The sample period is January 2000 to March 2014. The FOMC statement announcements were sourced from the Federal Reserve's website.\footnote{\url{https://www.federalreserve.gov/monetarypolicy/fomc_historical_year.htm}} The FOMC statements are short summaries of the topics discussed by the committee, released following a meeting. This include the assessment of the current economic situation, and policy decisions regarding the target federal funds rate, discount rate, and since 2008 different types of unconventional monetary policy. In January 2000, the Committee announced that it would issue a statement following each regularly scheduled meeting, regardless of whether there had been a change in monetary policy. We drop unscheduled FOMC meetings from our sample as in \cite{nakamura2018high} and \cite{handlan2020text}. In general, an unscheduled meeting reflects highly unusual economic or political circumstances.
The New York Times' articles were sourced from Northwestern Library, which granted us access up to March 2014.\footnote{We accessed and downloaded these articles on April 2 2021.} Our sample consists of articles published on the day of and day prior to the FOMC meetings. Furthermore we exploit the fact that articles are tagged by subjects and restrict our sample to only those which contain the subject ``federal reserve". These articles typically discuss the current economic situations and contain several private sector projections on both the economy and FOMC policy decisions. Thus, they provide us with a proxy for private sector expectations on the FOMC announcement day.
\newpage
We were able to find at least one such article during the day of and the day before each scheduled FOMC meeting date. This is consistent with the idea that the FOMC meetings are an important event for the financial markets such that they receive news coverage in the lead-up.
The empirical section of the paper uses both high-frequency movements in financial assets and macroeconomic variables for the US. The two high-frequency movements studied are the Fed Funds Futures surprises constructed by \cite{gertler2015monetary} and the policy news shocks constructed by \cite{nakamura2018high}.
We measure the impact of both the isolated monetary and news components using several daily interest rates. The nominal and real Treasury yields are sourced from FRED. Data on macroeconomic aggregates, such as the industrial production index and the consumer price index, are also sourced from FRED.
\section{Methodology} \label{sec:methodology}
This section of the paper details the three-stage process by which we extract the informational component of the high frequency movements around FOMC announcements. The three-stages can be summarized as follows:
\begin{itemize}
\item[Stage 1:] Document Embedding. We use state-of-the-art tools in the fields of natural language processing and machine learning to convert text documents in numeric vectors.
\item[Stage 2:] Constructing Proxy Expectations. We use the vector representations from Stage 1 to predict policy decisions of the FOMC. Under a rational expectation assumption, we interpret our estimated functions as expected policy decisions given the information sets of the private sector and the FOMC respectively.
\item[Stage 3:] Extracting Informational Content. Suppose market movements around FOMC announcements is due to informational effects. Then these movements should be proportional to the difference in expectations. We project high frequency movements onto the difference in expectation to recover their informational component.
\end{itemize}
In the following subsections, we provide further details on their implementation, as well as the assumptions that justify our procedure.
\newpage
Relative to the literature, the main novelty of our approach is in the use of newspaper articles to measure public expectation. Whereas the literature has thus far focused on the FOMC statements alone (see for instance \cite{handlan2020text}, \cite{hansen2016shocking}), we argue below that the information effect should depend on the difference in the expectations of the public and the FOMC, and that neglecting public expectation leads to omitted variable bias. In addition, with the exception of stage 1, which relies on a pre-trained model, the remaining stages involve only simple methods that are highly transparent. Finally, our method of constructing the intermediate expectation term allows us to probe at the mechanism by the information effect operate, as we explore in section \ref{sec:taylorrule}.
\subsection{Stage 1: Document Embedding} \label{subsec:first_stage}
In stage 1, we convert each FOMC statement and NYT article in $768\times1$-vectors using BERT (Bidirectional Encoder Representations from Transformers, \cite{devlin2019bert}). We highlight the challenge of using text data and explain how our BERT overcomes these problems before detailing our procedure.
Text data is difficult to analyze because of their high-dimensional nature. The simplest approach to deal with text is to just treat them as ``bag-of-words". With such an approach, the analyst would count the number of unique words in each article or document and then use the word counts as covariates for linear regressions or as input into other models. However, such an approach would work poorly in our sample, which has over 8,000 unique words in fewer than 150 observations. Whereas earlier papers have made progress by choosing which words to include (see for instance \cite{lucca2009measuring}), such ad-hoc methods not only have high researcher degree-of-freedom but could also leave information on the table.
The high dimensionality problem is compounded by the fact that the meanings of words depend on their context. An FOMC statement that is ``worried about inflation not growth" and one is ``worried about growth not inflation" have identical word counts but express contrary ideas.\footnote{A more subtle example of this ``lexical ambiguity" problem is that ``bank" has completely different meanings in ``river bank" and ``investment bank".} The need to take into account sequential information increases the dimension of text data exponentially. Consider a simple exercise in which we count the number of unique 6-word phrases, so that the example phrases above are counted as distinct entries. In a sample with 8,000 unique words, the possible number of unique phrases could be as high as $8,000^6 = 2.6 \times 10^{23}$. It is therefore largely infeasible to work with counts of phrases alone.
To render text analysis tractable and effective, we turn to BERT -- a state-of-the-art document embedding method that is able to reduce text input into $768\times1$-vectors while preserving sequential information. Its output captures the meaning of text input very well, as demonstrated by their usefulness for natural language processing tasks. For example, models based on BERT excel in question and answering tasks, in which given a question and a passage from Wikipedia containing the answer, the task is to predict the answer text span in the passage. Succeeding in these tasks all require some ``understanding" of the text input, giving us assurance that the BERT representations capture the meaning of words.
The key advantage of using BERT is that it is pre-trained on large amounts of text data (BooksCorpus -- 800M words and English Wikipedia -- 2,500M words). Given only 106 observations, it would have been very difficult to train any reasonably flexible text analysis models. For a similar reason, \cite{handlan2020text} relies on the pre-trained XLNet, a minor variation of BERT, for their analysis.
We note that other papers have employed Latent Dirichlet Allocations (LDA) models on samples of similar sizes [all the Hansen and McMahon papers], even though it is unclear if they are well-behaved given such small samples. In addition, it is known that LDA models have identification issues so that results are driven by the choice of priors even asymptotically (\cite{ke2021text}). Hence, we consider them to be unsuitable for the task at hand.
We initialized pretrained BERT models from an open source implementation provided by Hugging Face.\footnote{Hugging Face is a large open-source community that hosts pre-trained deep learning models, mainly aimed at NLP. Their state-of-the-art transformer models are used by Google, Facebook, Microsoft and AWS} (https://huggingface.co/bert-base-uncased) To use BERT, we first cleaned the input data by removing numbers, dates as well as common stop-words. Stop-words are a set of words which do not contain information about the content of the articles, such as prepositions and conjunctions. Removing these words ``enriches" the information content in text articles. Here, we relied on the stop-word list compiled by the text analysis package \emph{gensim}. Next, the words were tokenized. A token is the most fundamental object in a chosen language dictionary (it could be a word, character, or subword). We used the default tokenizer provided by Hugging Face for use with the bert-base-uncased model for tokenization.
Finally, we split each individual document into windows of 256 tokens, with overlaps of 10 tokens between windows. Each window is tokenized separately, and passed through the BERT model to form individual outputs. The embedding vector of the entire document is then taken to be the mean of the BERT class tokens across all of the windowed outputs.
We note that the cleaning process removes all numbers from the text. Consequently, the actual policy decisions as well as numerical projections about the economy are absent from the documents. This gives us assurance that our method is not simply picking up the policy rates or their forecasts, but rather capturing beliefs about the economy.
\subsection{Stage 2: Constructing Proxy Expectations} \label{subsec:second_stage}
In the second stage, we use the document embeddings to construct proxy expectations for the general public and the FOMC. Our starting point is the equations:
\begin{align*}
\text{FFR}_t & = f^\text{NYT}(X^\text{NYT}_t) + \varepsilon^\text{NYT}_t \\
\text{FFR}_t & = f^\text{FOMC}(X^\text{FOMC}_t) + \varepsilon^\text{FOMC}_t
\end{align*}
where $X^\text{NYT}_t$ and $X^\text{FOMC}_t$ are the embeddings of NYT articles and FOMC statement at time $t$. We further assume that
\begin{equation} \label{eq:news_exogeneity}
\mathbb{E}\left[\varepsilon^\text{NYT}_t \, \big| \, X^\text{NYT}_t\right] = 0 \quad , \quad \mathbb{E}\left[\varepsilon^\text{FOMC}_t \, \big| \, X^\text{FOMC}_t\right] = 0~.
\end{equation}
We interpret the above assumption as follows. Suppose $X^\text{NYT}_t$ captures all information about the economy that is available to the general public at time $t$. Then public expectation about FFR is given by
\begin{equation*}
f^\text{NYT}(X^\text{NYT}_t) = \mathbb{E}\left[\text{FFR}_t \, \big| \, X_t^\text{NYT}\right] ~.
\end{equation*}
The conditional mean assumption
\begin{equation*}
\mathbb{E}\left[\varepsilon^\text{NYT}_t \, \big| \, X^\text{NYT}_t\right] = 0
\end{equation*}
is thus a rational expectation assumption -- the general public has beliefs that are not systematically biased and are correct on average.
We assume the same is true for the FOMC. That is, we assume that $X^\text{FOMC}_t$ contains all the information that the FOMC has about the economy at time $t$. $f^\text{FOMC}(X^\text{FOMC}_t)$ is then the Taylor rule that maps FOMC information into interest rate decisions. $\varepsilon^\text{FOMC}_t$ is monetary policy shock -- implementation error that prevents the FOMC from achieving its desired optimal policy rate. We assume that this monetary policy shock is $0$ on average and mean independent from $X_t^\text{FOMC}$. This is consistent with the current views on monetary policy shock, as discussed in \cite{ramey2016macroeconomic}.
Given our limited sample size, we impose further structure for estimation by assuming that $f^\text{NYT}$ and $f^\text{FOMC}$ take the following linear form:
\begin{align*}
f^\text{NYT}(X_t^\text{NYT}) & = \alpha^\text{NYT} + \beta^\text{NYT}{}'X^\text{NYT}_t \\
f^\text{NYT}(X_t^\text{NYT}) & = \alpha^\text{FOMC} + \beta^\text{FOMC}{}'X^\text{FOMC}_t
\end{align*}
Because we use the pre-trained BERT, $X_t^\text{NYT}, X_t^\text{FOMC} \in \mathbf{R}^{768}$. In order to estimate $\alpha$ and $\beta$ from our sample of size 106, we use the elastic net, defined as:
\begin{equation*}
\left(\hat{\beta}^\text{NYT}, \hat{\alpha}^\text{NYT} \right) = \underset{\beta, \alpha}{\arg\min} \left\lVert \text{FFR} - \alpha - \beta'X^\text{NYT} \right\rVert_2^2 + \gamma\eta \left( \lVert \alpha \rVert_1 + \lVert \beta \rVert_1\right) + \gamma(1-\eta) \left( \lVert \alpha \rVert^2_2 + \lVert \beta \rVert^2_2\right)
\end{equation*}
\begin{equation*}
\left(\hat{\beta}^\text{FOMC}, \hat{\alpha}^\text{FOMC} \right) = \underset{\beta, \alpha}{\arg\min} \left\lVert \text{FFR} - \alpha - \beta'X^\text{FOMC} \right\rVert_2^2 + \gamma\eta \left( \lVert \alpha \rVert_1 + \lVert \beta \rVert_1\right) + \gamma(1-\eta) \left( \lVert \alpha \rVert^2_2 + \lVert \beta \rVert^2_2\right)
\end{equation*}
Here, $\text{FFR}$, $X^\text{NYT}$ and $X^\text{FOMC}$ refer to the stacked matrices of $\text{FFR}_t$, $X_t^\text{NYT}$ and $X_t^\text{FOMC}$ respectively.
The elastic net nests two popular methods of penalized regressions. Suppose $\eta = 1$. Then the elastic net reduces to the LASSO estimator. LASSO performs model selection by setting regression coefficients to zero. The resulting output is typically an interpretable model in which only a few coefficients are non-zero. On the other hand, setting $\eta = 0$ leads to the ridge estimator, a dense model which assigns highly correlated regressors similar coefficients. Since the elastic net is more general, it is able to achieve better prediction accuracy given properly chosen tuning parameters (\cite{zou2005regularization}). Furthermore, it has a Bayesian interpretation as the posterior mode induced by the spike-and-slab prior of \cite{mitchell1988bayesian} and may be more appropriate for economic data (\cite{giannone2021sparsity}).
With our conditional mean assumptions, the elastic net is $L_2$ consistent as long as the space of the covariates is sufficiently rich (\cite{de2009elastic}). Since we are interested in prediction and not model selection, we do not require any sparsity assumption.
Each elastic net requires two tuning parameters $(\lambda, \eta)$. Because these parameters are chosen to maximize $R^2$ in stage 3, we discuss their selection in the next subsection.
\subsection{Stage 3: Extracting Informational Content} \label{subsec:third_stage}
In this subsection, we explain how we use the proxy expectations from stage 2 to decompose high frequency movements of interest rates around FOMC announcements into an information and a monetary component.
Let $\Delta R_t$ denote the high frequency movement of interest rate $R$ around the release of FOMC statements. Suppose $\Delta R_t$ includes reaction by the private sector to information released by the Federal Reserve. This component should be a function to the \emph{difference} in information of the two parties. In particular, if there is no difference in information, this component should be $0$. This motivates the following linear regression:
\begin{equation}\label{eq:stage3}
\Delta R_t = \zeta + \theta \Delta \mathbb{E}[\text{FFR}_t] + \nu_t \quad , \quad \mathbb{E}\left[\nu_t \, | \, \Delta \mathbb{E}[\text{FFR}_t]\right] = 0,
\end{equation}
where
\begin{equation*}
\Delta \mathbb{E}[\text{FFR}_t] := \alpha^\text{FOMC} + \beta^\text{FOMC}{}'X^\text{FOMC}_t - \alpha^\text{NYT} + \beta^\text{NYT}{}'X^\text{NYT}
\end{equation*}
is the difference in the expectations of the private sector and the FOMC based solely on difference in information.
For any choice of tuning parameters $(\lambda, \eta)$ in stage 2, we are able to construct
\begin{equation*}
\widehat{\Delta \mathbb{E}[\text{FFR}_t]} := \hat{\alpha}^\text{FOMC} + \hat{\beta}^\text{FOMC}{}'X^\text{FOMC}_t - \hat{\alpha}^\text{NYT} + \hat{\beta}^\text{NYT}{}'X^\text{NYT}~.
\end{equation*}
We then choose our tuning parameter to maximise $R^2$ in the regression of $\Delta R_t$ on $\widehat{\Delta \mathbb{E}[\text{FFR}_t]}$, where the latter term is a function of $(\lambda, \eta)$.
The regression in equation (\ref{eq:stage3}) has clear meaning. Here, we are projecting the high frequency movement onto the differences in proxy expectation and we interpret the projected component as the part of $\Delta R_t$ that is responding to Federal Reserve information.
Suppose we were able exactly capture news component of $\Delta R_t$. We would then be able to interpret the residual in the above regression as true monetary policy shocks. However, we are cautious that much of the news component may remain due in part to the restrictive linear assumption in equation (\ref{eq:stage3}). On the other hand, given the sample size of 106, it is unclear if more flexible functional forms would simply lead to over-fitting. For this reason, the simple linear regression is our preferred specification for stage 3.
An approach commonly seen in the literature (see for instance \cite{handlan2020text}, \cite{hansen2016shocking}) is to run the regression on the FOMC component only:
\begin{equation}\label{eq:shortregression}
\Delta R_t = \zeta + \theta \cdot f^\text{FOMC}(X^\text{FOMC}_t) + \xi_t \quad
\end{equation}
As we argued above, the informational component of the high frequency movement should depend on the difference in information, and not the level directly. Suppose equation \ref{eq:stage3} is true. We can rewrite equation \ref{eq:shortregression} as
\begin{equation*
\Delta R_t = \zeta + \theta \cdot f^\text{FOMC}(X^\text{FOMC}_t) - \theta \cdot f^\text{NYT}(X^\text{NYT}_t) + \nu_t \quad , \quad \mathbb{E}\left[\nu_t \, | X^\text{FOMC}_t\right] = 0.
\end{equation*}
Here,
\begin{equation*}
\xi_t = - \theta \cdot f^\text{NYT}(X^\text{NYT}_t) + \nu_t
\end{equation*}
is clearly correlated with $f^\text{FOMC}(X^\text{FOMC}_t)$ through $f^\text{NYT}(X^\text{NYT}_t)$, giving rise to omitted variable bias.
Furthermore, equation (\ref{eq:stage3}) is robust in that it can be valid even when the conditions in equation (\ref{eq:news_exogeneity}) is violated. For example, suppose the FOMC has time varying preferences in trading off inflation and output gap, so that a positive shock is more likely when the FOMC preferences have a hawkish realization. This violates the conditional mean assumption in equation (\ref{eq:shortregression}) as long as the statements contains information about its current hawkishness. However, time-varying preferences, or other omitted variables, do not cause a problem for the regression in (\ref{eq:stage3}) as long as: $$\text{Cov}\left(\nu_t, f^\text{FOMC}(X^\text{FOMC}_t)\right) = \text{Cov}\left(\nu_t, f^\text{NYT}(X^\text{NYT}_t)\right)~.$$
Intuitively, our regression is valid as long as these omitted variables skew private sector and FOMC expectations the same way. On the other hand, if equation (\ref{eq:news_exogeneity}) does not hold, \ref{eq:shortregression} will be invalid, even if $X_{t}^\text{FOMC} \perp \!\!\! \perp X_t^\text{NYT}$.
Finally, even if $X_{t}^\text{FOMC} \perp \!\!\! \perp X_t^\text{NYT}$, our specification will still have more power. Including $f^\text{NYT}(X^\text{NYT}_t)$ would lead to more precise estimates by reducing the variance of the residuals. This gels with the idea that our regression is a more direct test of the information hypothesis: if the high frequency movement is driven by agents updating their information set, theory directly predicts that these movement should be proportional to the amount of updating that occurred.
Next, we provide intuitive reasons for choosing tuning parameters that maximize $R^2$ in stage 3. Firstly, the tuning parameters that are chosen are unlikely to overfit the stage 2 elastic nets. To see this, observe that if we overfit stage 2 to the extreme, such that $\hat{\alpha}^\text{FOMC} + \hat{\beta}^\text{FOMC}{}'X^\text{FOMC}_t = \hat{\alpha}^\text{NYT} + \hat{\beta}^\text{NYT}{}'X^\text{NYT}_t = \text{FFR}_t$. Then $\widehat{\Delta \mathbb{E}[\text{FFR}_t]} = 0$ identically so that stage 3 $R^2$ is 0. It is equally unlikely to underfit stage 2. Consider again the extreme case, this time in which $\hat{\beta}^\text{NYT}$ and $\hat{\beta}^\text{FOMC}$ are set to $0$. Now, $\hat{\alpha}^\text{NYT}$ = $\hat{\alpha}^\text{FOMC} = \frac{1}{T} \sum_{t = 1}^T \text{FFR}_t$ so that $\widehat{\Delta \mathbb{E}[\text{FFR}_t]}$ is again $0$. Our procedure guards against both of these problems. Where the method may face issues is in overfitting stage 3. However, we believe that this is less of a concern since stage 3 is a linear regression with only two parameters.
\subsection{Results}
\begin{table}[htbp]
\centering
\caption{Summary Statistics for Stage 3. Robust standard errors are used.}
\begin{tabular}{ccccccc}
& Shock & $R^2$ & $\hat{\theta}$ & S.E. & $t$-stat & $p$-value \\
\midrule
\multirow{3}[2]{*}{Pre-ZLB} & PNS & 0.079 & 0.083 & 0.031 & 2.624 & 0.011 \\
& FFR & 0.038 & 0.051 & 0.044 & 1.152 & 0.254 \\
& FF4 & 0.085 & 0.183 & 0.065 & 2.826 & 0.006 \\
\midrule
\multirow{3}[1]{*}{Full Sample} & PNS & 0.028 & 0.009 & 0.006 & 1.466 & 0.146 \\
& FFR & 0.013 & 0.008 & 0.008 & 0.971 & 0.334 \\
& FF4 & 0.073 & 0.020 & 0.008 & 2.421 & 0.017 \\
\end{tabular}%
\label{tab:stage3}%
\end{table}%
Summary statistics for Stage 3 of our procedure is presented in table \ref{tab:stage3}. We find evidence that high frequency movements around FOMC announcements is in part a response to information from the Federal Reserve. In particular, we note $\hat{\beta}$ is positive for PNS, FFR and FF4. It is also significant at the 5\% level for the PNS and FF4 in the pre-ZLB period, and for FF4 in the full sample. The positive sign implies that when the information set of the FOMC leads it to set higher interest rate than the public was expecting, the market response is positive.
The $R^2$ in stage 3 is relatively low across the board. This could be interpreted as meaning that most of the movement around FOMC announcements is response to pure monetary policy shock. On the other hand, given the our highly restrictive linear model in stage 3, we consider the information effect we found to be a lower bound: with more data and a more flexible stage 3, we might find potentially much larger information effects.
\begin{table}[htbp]
\centering
\caption{Summary Statistics for Stage 2}
\begin{tabular}{cccccccc}
& Shock & $\eta^\text{NYT}$ & $\lambda^\text{NYT}$ & NYT $R^2$ & $\eta^\text{FOMC}$ & $\lambda^\text{FOMC}$ & FOMC $R^2$ \\
\midrule
\multirow{3}[2]{*}{Pre-ZLB} & PNS & 0.550 & 1.150 & 0.122 & 0.250 & 3.275 & 0.082 \\
& FFR & 0.700 & 0.811 & 0.152 & 0.050 & 13.219 & 0.147 \\
& FF4 & 0.400 & 1.630 & 0.068 & 0.400 & 2.310 & 0.060 \\
\midrule
\multirow{3}[1]{*}{Full Sample} & PNS & 1.000 & 0.012 & 0.884 & 1.000 & 0.100 & 0.995 \\
& FFR & 0.650 & 0.050 & 0.872 & 0.800 & 0.142 & 0.974 \\
& FF4 & 0.100 & 0.142 & 0.884 & 1.000 & 0.100 & 0.990 \\
\end{tabular}%
\label{tab:stage2}%
\end{table}%
Table \ref{tab:stage2} presents summary statistics associated with stage 2, including the tuning parameters chosen by our procedure. Larger $\eta$ implies that the chosen model is closer to a LASSO, while smaller $\eta$ implies that the chosen model is closer to a ridge regression. We see evidence for sparsity in PNS and FF4 with the FOMC statement, especially in the full sample period, since the preferred model is the LASSO. This suggests that there are only a few variables which are important for determining the FOMC's desired policy rate. On the other hand, during the pre-ZLB period, the selected models tended towards ridge regressions, suggesting that there are more diverse considerations for determining interest rates prior to the Great Financial Crisis.
We note that $R^2$ is high in the full sample, but much smaller in the the pre-ZLB sample. This is likely because FFR remained largely constant after the Great Financial Crisis. Taking the model at face value, we would infer that monetary policy is easier to predict after post-crisis, and that the variance of monetary policy shocks is smaller. On the other hand, we are cautious that our linear model may be inappropriate for modelling the policy rates, which may be censored at the zero lower bound.
We conclude by noting that our approach is highly transparent. All tuning parameters as well as regression parameters that we chose are contained in tables \ref{tab:stage3} and \ref{tab:stage2}. The only unpresented parameters belong to BERT, which is trained by an external party without reference to any of the variables used in stage 2 and 3.
\subsection{Channels of Information Transmission}\label{sec:taylorrule}
In order to understand how new information from the FOMC translates to market reaction, we consider how differences in expectation of specific macroeconomic variables drive high frequency movements around FOMC announcements. To do that, we extend the analysis thus far to a general economic variable $Y_i$:
\begin{align*}
Y_{i,t} & = f_i^\text{NYT}(X^\text{NYT}_t) + \varepsilon^\text{NYT}_{i,t} \\
Y_{i,t} & = f_i^\text{FOMC}(X^\text{FOMC}_t) + \varepsilon^\text{FOMC}_{i,t}
\end{align*}
where as before we assume that:
\begin{equation*}
\mathbb{E}\left[\varepsilon^\text{NYT}_{i,t} \, \big| \, X^\text{NYT}_t\right] = 0 \quad , \quad \mathbb{E}\left[\varepsilon^\text{FOMC}_{i,t} \, \big| \, X^\text{FOMC}_t\right] = 0~.
\end{equation*}
However, the stage 3 equation of interest is
\begin{equation}\label{eq:stage3_taylor}
\Delta R_t = \zeta + \sum_{i=1}^p \theta_i \cdot \Delta \mathbb{E}[Y_{i,t}] + \nu_t \quad , \quad \mathbb{E}\left[\nu_t \, | \, \Delta \mathbb{E}[Y_t]\right] = 0.
\end{equation}
Suppose $Y_{i,t} = \pi_t$ is inflation at $\pi_t$. In the regression above, we would then interpret $\theta_i$ as the high frequency movement that resulted from differences in inflation expectation. Running regression \ref{eq:stage3_taylor} allows us to understand what type of informational differences is driving the high frequency movement.
In particular, let $Y_{1,t}$ be real GDP in the next nearest quarter and $Y_{2,t}$ be CPI in the next nearest month.\footnote{In particular, we use the ``Consumer Price Index: All Items Excluding Food and Energy''. This choice is motivated by the FOMC targetting a core inflation rate which excludes commodity prices such as Food and Energy which are highly influenced by international markets.} These are the two variables that would be relevant if interest rate was set according to a simple Taylor rule. Results are presented in table \ref{tab:stage3_taylorrule}. Across the board, when the FOMC expects higher inflation than the private sector, the interest rates increase following the announcements. This is in line with the predictions of standard macroeconomic theory, in which the central bank raises rates in response to heightened inflationary pressures regardless of whether the source is demand or supply shocks.
The coefficients on real GDP is mixed. In the pre-ZLB period, the coefficients on real GDP have positive signs across the board. In words, when the FOMC expects higher real GDP than the public, interest rates subsequently increases. In the full sample, however, the sign changes to negative for FFR and FF4. Standard macroeconomic theory does not provide a straightforward answer to how monetary authority should react to GDP fluctuations. For instance, if the central bank receive news of a positive demand shock, then it raises its GDP projections and simultaneously raises interest rates. This gives rise to a positive comovement between expected GDP and monetary rate policy. But, if it receives news of a negative supply shock, a central bank would raise rates even as it lowers GDP projections, leading to negative comovements. Note that the full sample includes episodes of \textit{supply-side} shocks such as tax cuts and/or oil price shocks; and furthermore, we exclude observations from July 2008 to July 2009 inclusive. Hence, it is plausible that in our full sample supply shocks play a large role.
In sum, our method allows us to probe at the mechanisms of information transmission. We find that when the FOMC expects higher inflation than the private sector, interest rates unequivocally increase. However, when the FOMC expects higher real GDP than the private sector, interest rate movements are mixed. These patterns are suggestive of the types of shocks in an economy.
\begin{table}[htbp]
\centering
\caption{Stage 3 Results using real GDP and CPI. Robust standard errors are used.}
\begin{tabular}{clccccc}
& \multicolumn{1}{c}{Shock} & $R^2$ & $\hat{\theta}_\text{RGDP}$ & RGDP $p$-value & $\hat{\theta}_\text{CPI}$ & CPI $p$-value \\
\midrule
\multirow{3}[2]{*}{Pre-ZLB} & PNS & 0.103 & 0.001 & 0.073 & 0.005 & 0.131 \\
& FFR & 0.083 & 0.000 & 0.378 & -0.080 & 0.256 \\
& FF4 & 0.216 & 0.030 & 0.000 & 0.032 & 0.440 \\
\midrule
\multirow{3}[1]{*}{Full Sample} & PNS & 0.061 & 0.000 & 0.548 & 0.003 & 0.174 \\
& FFR & 0.034 & -0.001 & 0.139 & 0.005 & 0.134 \\
& FF4 & 0.067 & -0.002 & 0.051 & 0.008 & 0.055 \\
\end{tabular}%
\label{tab:stage3_taylorrule}%
\end{table}%
\section{Evidence of Informational Content} \label{sec:main_results}
In this section, we provide evidence that the component we isolated using the procedure in Section \ref{sec:methodology} carries informational content. We focus on event study and financial market evidence.\footnote{In Appendix \ref{subsec:svar_evidence} we present additional evidence by estimating the impact of ``Monetary$_t$'' at the aggregate level using a SVAR model. Expectedly, these are not too different from SVARs using the high frequency shocks alone, given the $R^2$ of Stage 3.}
\subsection{Event Study Analysis} \label{subsec:event_analysis}
We study the components of the high-frequency financial surprises isolated in the previous section.
\begin{figure}[ht]
\centering
\caption{Time Series of PNS$_t$ Decomposition. Following \cite{nakamura2018high} we omit observations between July 2008 to July 2009 inclusive.}
\label{fig:TimeSeriesPNS}
\includegraphics[width=12cm,height=9cm]{TimeSeriesPNS.png}
\end{figure}
Given the modest $R^2$ from the decomposition, it is unsurprising that the news component has relatively smaller variance than the monetary component. However, the series is intuitively reasonable. To show this, we zoom in on three FOMC announcements in which the news component was thought to have played an important role ex post.
\begin{itemize}
\item[] \textbf{September 18, 2007}: We return to our opening example on which the FOMC cut its rates by 50 basis points. As our quoted text suggests, the market interpreted the Federal Reserve's fairly aggressive actions and statements to mean that a sharp economic downturn was imminent. Thus it is reassuring that we recorded one of our largest negative news shock on this date.
\item[] \textbf{March 18, 2008}: One of our larger positive news shock was documented on March 18, 2008, even amidst the Great Financial Crisis. On this day, the FOMC cut rates by 75 basis points to 2.25 percent. However, as the New York Times reports, ``though it was one of the biggest one-day rate cuts in decades, investors had been betting heavily that the Fed would cut its key rate a full percentage point in response to strong evidence that a recession has begun and to the deepening crisis on Wall Street." Instead, the Federal Reserve policy reflected concerns about inflation. As the FOMC statement writes, ``inflation has been elevated, and some indicators of inflation expectations have risen". It is therefore reassuring that our measure of news recorded a positive value, reflecting the Federal Reserve's information about inflation.
\item[] \textbf{January 30, 2002}: Another of our largest positive surprises came as the economy was recovering from the bursting of the dot-com bubble and 9-11. At this meeting, the FOMC left interest rates unchanged, which was a departure from its yearlong policy of rate cuts. According to the FOMC, ``Signs that weakness in demand is abating and economic activity is beginning to firm have become more prevalent". As the New York Times reports, ``Investors interpreted [...] the Fed's decision on interest rates as signs that the economy might be stabilizing. Stocks prices moved higher in the afternoon after the Fed's announcement." Here, we have explicit indication that investors interpreted the Federal Reserve's policy as good news, supporting the positive news shock that we recovered.
\end{itemize}
\subsection{Financial Market Evidence} \label{subsec:financial_evidence}
The goal of this subsection of the paper is to identify the effect of the informational content of FOMC announcements on nominal and real interest rates of different maturities.
First, we measure the impact of the informational content of FOMC announcements on nominal and real interest rates. In particular, we estimate the following two empirical equations
\begin{align}
\Delta r_t &= \alpha_0 + \beta \Delta R_t + e_t \label{eq:shock} \\
\Delta r_t &= \alpha_1 + \gamma \text{Monetary}_t + \mu \text{News}_t + \epsilon_t \label{eq:shock_decomposition}
\end{align}
were $\Delta r_t$ is the change in the outcome nominal and/or real interest rate of interest, e.g., the Market Yield on U.S. Treasury Securities at 10-Year Constant Maturity; $\Delta R_t$ is the high-frequency financial surprise around the FOMC announcement; $\text{Monetary}_t$ and $\text{News}_t$ are the decomposition coming from the methodology used in Section \ref{sec:methodology}; $e_t$ and $\epsilon_t$ are error terms; and $\alpha_0$, $\alpha_1$, $\beta$, $\gamma$ and $\mu$ are parameters. The parameter of interests are $\beta$, $\gamma$ and $\mu$. Comparing $\beta$ and $\gamma$ allows us to identify how purging the high-frequency surprises of informational content changes the impact of monetary policy shocks on nominal and real interest rates.
First, we study the impact of the high-frequency Fed Funds Futures on nominal and real rates by estimating Equation \ref{eq:shock} using the high-frequency surprise FF4.
\begin{table}[ht]
\centering
\caption{Response of Interest Rates to Fed Funds Futures}
\label{tab:ff4_rates}
\begin{tabular}{l c c c c}
& (1) & (2) & (3) \\
& FF4$_t$ & $Monetary_t$ & $News_t$ & Observations \\ \hline \hline
& & & & \\
Nominal 3 month & 0.300 & 0.265 & 0.797*** & 106 \\
& (0.218) & (0.234) & (0.248) & \\
Nominal 1 year & 0.486*** & 0.494*** & 0.376 & 106 \\
& (0.128) & (0.123) & (0.316) & \\
Nominal 2 years & 0.571*** & 0.569*** & 0.594 & 106 \\
& (0.209) & (0.210) & (0.509) & \\
Nominal 5 years & 0.408* & 0.240 & 0.432 & 106 \\
& (0.244) & (0.406) & (0.711) & \\
Nominal 10 years & 0.149 & 0.143 & 0.235 & 106 \\
& (0.178) & (0.184) & (0.593) & \\
Nominal 20 years & -0.00137 & 0.000402 & -0.0261 & 106 \\
& (0.135) & (0.138) & (0.547) & \\
TIPS Real 2 years & -0.0473 & -0.0521 & 0.0192 & 74 \\
& (0.131) & (0.139) & (0.530) & \\
TIPS Real 5 years & 0.126 & 0.151 & -0.219 & 82 \\
& (0.191) & (0.175) & (0.824) & \\
TIPS Real 10 years & 0.0124 & 0.0295 & -0.226 & 82 \\
& (0.152) & (0.140) & (0.642) & \\
TIPS Real 20 years & 0.0401 & 0.0355 & 0.0954 & 70 \\
& (0.162) & (0.172) & (0.587) & \\
& & & & \\ \hline \hline
\multicolumn{5}{c}{ Robust Standard Errors in Parentheses} \\
\multicolumn{5}{c}{ *** p$<$0.01, ** p$<$0.05, * p$<$0.1} \\
\end{tabular}
\end{table}
Table \ref{tab:ff4_rates} presents the results of this regressions. Column (1) presents the empirical estimates of coefficient $\beta$ in Equation \ref{eq:shock} for different nominal and real rates. Results seem to show a hump-shaped impact of FF4$_t$ on nominal rates, with no significant impact on the three month nominal rate, a positive impact on nominal rates at the one to 10 year maturities (peaking at 2 years), and no impact on long-term rates. Surprisingly, there does not seem to be a significant impact of FF4$_t$ on the daily change of real interest rates captured by TIPS.\footnote{This result could be explained by different reasons. For example, given the policy change on FOMC statements in the year 2000 and the data availability of text, we have a short sample. This problem worsens for TIPS treasuries as they were introduced later in time.}
Next, we turn to estimating the impact of the different isolated components of FF4$_t$ estimated in Section \ref{sec:methodology}. Columns (2) and (3) present empirical estimates of coefficients $\gamma$ and $\mu$ in Equation \ref{eq:shock_decomposition}, respectively, for different nominal and real interest rates. Interestingly, while the "\textit{Monetary}$_t$" component of FF4$_t$ does not seem to have a significant effect on the 3-month maturity treasury, the "\textit{News}$_t$" component has a significant impact on this short run interest rate. The "\textit{News}$_t$" component does not seem to have significant impact on mid and long-term treasury yields. Consequently, the estimated coefficients of the "\textit{Monetary}$_t$" component is similar to the coefficients presented in Column (1) for FF4$_t$.
Second, we estimate the impact of the "Policy News Shocks" or PNS$_t$ constructed by \cite{nakamura2018high} and its decomposition on nominal and real interest rates. Note, this policy indicator is comprised of changes in multiple nominal interest rates at different maturities spanning the first year of the term structure.\footnote{More specifically, the composite policy indicator is the first principal component of the unanticipated change over the 30-minute windows in the following five interest rates: the Fed funds rate immediately following the FOMC meeting, the expected Fed funds rate immediately following the next FOMC meeting, and expected three-month eurodollar interest rates at horizons of two, three, and four quarters.}
\begin{table}[ht]
\centering
\caption{Response of Interest Rates to Policy News Shocks}
\label{tab:PNS_rates}
\begin{tabular}{l c c c c}
& (1) & (2) & (3) \\
& PNS$_t$ & $Monetary_t$ & $News_t$ & Observations \\ \hline \hline
& & & & \\
Nominal 3 month & 0.670*** & 0.581*** & 1.208*** & 106 \\
& (0.141) & (0.150) & (0.273) & \\
Nominal 1 year & 0.795*** & 0.834*** & 0.558** & 106 \\
& (0.111) & (0.121) & (0.263) & \\
Nominal 2 years & 1.052*** & 1.157*** & 0.420 & 106 \\
& (0.203) & (0.207) & (0.394) & \\
Nominal 5 years & 0.929*** & 1.135*** & -0.310 & 106 \\
& (0.226) & (0.249) & (0.581) & \\
Nominal 10 years & 0.456** & 0.641*** & -0.655 & 106 \\
& (0.184) & (0.202) & (0.516) & \\
Nominal 20 years & 0.225 & 0.333* & -0.421 & 106 \\
& (0.169) & (0.185) & (0.431) & \\
TIPS Real 2 years & 1.057*** & 1.129*** & 0.138 & 74 \\
& (0.239) & (0.241) & (0.560) & \\
TIPS Real 5 years & 0.787*** & 0.944*** & -1.729** & 82 \\
& (0.249) & (0.268) & (0.720) & \\
TIPS Real 10 years & 0.543** & 0.660*** & -1.332* & 82 \\
& (0.210) & (0.224) & (0.673) & \\
TIPS Real 20 years & 0.317 & 0.420* & -0.850 & 70 \\
& (0.210) & (0.231) & (0.643) & \\
& & & & \\ \hline \hline
\multicolumn{5}{c}{ Robust Standard Errors in Parentheses} \\
\multicolumn{5}{c}{ *** p$<$0.01, ** p$<$0.05, * p$<$0.1} \\
\end{tabular}
\end{table}
A priori, this could imply greater variability of PNS$_t$, compared to FF4$_t$. Additionally, in \cite{nakamura2018high} the authors stress that this composite policy indicator captures the effects of “forward
guidance” (FOMC announcements that convey information about future changes in the Fed Funds rate). Column (1) of Table \ref{tab:PNS_rates} presents the empirical estimates of coefficient $\beta$ in Equation \ref{eq:shock}. While results are in line with those for FF4$_t$, the estimates are highly more significant. Again, we find that the empirical estimates for $\beta$ exhibits a hump-shaped impact on nominal interest rate across the term structure. Unlike the results presented in \ref{tab:ff4_rates}, the policy news shock has a significant impact on TIPS yields up to the 10 year maturity rate.
Finally, we turn to estimating the role of the "Monetary$_t$" and "News$_t$" components of Policy New Shocks. Once more, the "News$_t$" component has a large and significant impact on short term nominal interest rates. One way to quantify the relevance of the news component is comparing the estimated coefficients for the "Monetary$_t$" component and those for PNS$_t$. Not controlling for the news component biases up the impact on nominal interest rate between 10 and 5 basis points for short term nominal interest rate, and biases the impact down between 15 and 30 basis points for mid and long term nominal interest rates. In addition, the impact of "Monetary$_t$" on TIPS real rates is greater than the impact of PNS$_t$. Interestingly, the news component which has a large and significant impact on short run nominal interest rate, also seems to have a large but negative impact on long run real rates.
In summary, the paragraphs above studied the impact of high-frequency financial surprises and its decomposition on the daily change of nominal and real rates around FOMC meetings. We find that the news component of these high-frequency surprises has a significant impact on short term nominal interest rates across different specified shocks. We take these results as evidence that differences in the expected policy rate by the FOMC Statements and the New York Times has a systematic and persistent impact on relevant financial rates.
\section{Conclusion} \label{sec:conclusion}
We argue that FOMC announcements contain both information and monetary policy content. We develop a transparent framework to pick out the informational content of high-frequency financial surprises around FOMC announcements by using state-of-the art text analysis on both policy statements and newspapers.
We find that differences in expectations about policy rates between the FOMC and private agents are predictive of high-frequency movements around policy announcements. Furthermore, we show that the isolated informational content has a significant impact on the yields of short-term government bonds. Consequently, ignoring this informational content may bias our understanding of the impact of monetary policy shocks.
We interpret these findings as an existence result. In order to draw reasonable conclusions from a small dataset, we have prioritized simplicity and transparency over more flexible functional forms. Hence, we consider our estimates to be lower bounds for the informational content in FOMC announcements.
\newpage
|
\section{Introduction}
Recently, cosmological theories and observations have provided interesting information about the universe. Data and measurements from the cosmic microwave background (CMB) show that the matter and energy fluctuations are always unstable on a large scale \cite{2,3}. The leading cause of these cosmic fluctuations is still unknown, and there is no specific explanation. But, like all other phenomena, cosmologists use different scenarios to explain the reasons for these fluctuations, including the inflationary world \cite{6,7,8}. Hence, cosmic inflation is a model for the production of perturbations related to the initial density of the universe, which somehow involves the structure formation. As can be seen from the inflation patterns, the universe has gone through an early period of accelerated expansion to solve the problems in cosmology, such as the horizon, flatness, and monopole problems \cite{10}. This accelerated expansion also led to these quantum fluctuations, and over time, these perturbations intensified under gravity, creating the structure of galaxies and everything in the universe on a large scale \cite{10,12,14}. Researchers have studied different inflation models that the simplest model is as inflation by a slow-roll scalar field \cite{B1, B2}. However, many reasons show that inflationary models may be practical for more than one field. First, in many theories, such as string theory or supersymmetry, and many other areas, we are practically dealing with several fields. Second, using two or more scalar fields, may offer desirable features and have many implications in cosmology. For example, hybrid models that include two scalar fields achieve both inflationary ranges and the area of density fluctuations. These are consistent with observable data and they occur at sub-Planckian scales\cite{17,18,19,20}. However, due to the advantages of studying multi-field inflation models, their analysis is complicated and has particular complexities due to the observable data. When we are faced with two or more scalar fields, perturbations in the relative contribution to the energy density are also possible, along with perturbations in total energy density \cite{22}. These isotropic perturbations may be source of curvature perturbations, and their evolution at the super-horizon scale, which confronts calculations with certain complexities, such as the density power spectrum \cite{26,27,28}. As a result, these fields are associated with several initial conditions that affect the power spectrum \cite{29,31}. The important thing is that the complexity of these multi-field inflation models will show against the observable data. Therefore, we should always consider a complete framework for these models and test multi-field models \cite{31,32,33,38,39,41,42,43}. One of these cases is the use of two-field inflation models. In this article, we want to analyze a new perspective of these two-field inflation models according to specific conditions. For two-field inflation, a number of specific models have been considered already \cite{46,47,48,49,52,53,54,55,56,57,59,60,61,62,63,64,65,66,67}. In general, two-field inflation has been used in many cases and has several implications in cosmology. It is including approximate solutions to the metric perturbations in slow-roll approximation. Also, this scenario include the evolution equation for adiabatic and entropy perturbations for certain models with kinetic corrections. There are also models with specific non-canonical corrections, or some unconventional kinetic corrections \cite{68,70,71,72,73}.\\
Given all the concepts mentioned above, we now want to consider a two-field inflation model concerning swampland conjecture \cite{80}. A conjecture, called weak gravity, has recently been introduced \cite{76,77,78,79,q,w}. According to this conjecture, gravity introduced as the weakest force at high energy limit of theories coupled to the gravity \cite{81,83,84}. There is an area that is consistent with quantum gravity called landscape, but at low energy, the landscape is surrounded by a larger area called swampland that contradicts quantum gravity \cite{85,86,86b}.\\
For inflation models to be compatible with quantum gravity, they must meet two criteria. Some inflation models consistent with these criteria, and many inflation models were inconsistent with them. In general, these two conditions are called swampland distance conjecture, which provides an upper limit for $\Delta\phi$ (variation of scalar fields), and swampland de Sitter (dS) conjecture, which provides a limit for potential slope. In this article, we use the swampland dS conjecture. In general, based on Planck mass $M_{pl}$, these conditions are expressed in the following forms \cite{88,89,90,91,e},
\begin{equation}\label{1}
\frac{\Delta\phi}{M_{pl}}<\mathcal{O}(1),
\end{equation}
and
\begin{equation}\label{2}
M_{pl}\frac{V'}{V}>c,
\end{equation}
where $V$ is a scalar field potential, and $c$ is a positive constant.\\
Given the above concepts, in this paper, we present a two-field inflation model method based on the swampland dS conjecture. Then, we evaluate the compatibility or incompatibility of this model with respect to observable data. This paper is organized as follows. In section 2, we study the two-field inflation model and introduce different types of cosmological parameters. In section 3, according to the swampland dS conjecture and the concepts expressed in this paper, we reobtain all of the parameters in section 2 with new points of view. Then, we investigate the compatibility or incompatibility of this inflation model according to the observable data by plotting some figures related to each of these cosmological parameters. We determine the range of these parameters in the final section before conclusion.
\section{Two-field inflation model}
In this section, we first briefly introduce a two-field inflation model, then evaluate these expressed models with the swampland dS conjecture. In order to examine the inflation models, we first consider the corresponding action and metrics. So, for two scalar fields, the action has following form,
\begin{equation}\label{3}
S=\int d^{4}x\sqrt{-g}\left(\frac{M_{pl}^{2}}{2}R-\frac{1}{2}\partial_{\mu}\phi \partial^{\mu}\phi-\frac{1}{2}\partial_{\mu}\chi \partial^{\mu}\chi-V(\phi,\chi)\right),
\end{equation}
where $R$ and ($\phi,\chi$) are Ricci scalar and two scalar fields, respectively, and $V(\phi,\chi)$ is corresponding potential. The Friedmann-Robertson-Walker (FRW) space-time introduced as following \cite{91b},
\begin{equation}\label{4}
ds^{2}=-dt^{2}+\alpha^{2}(t)\delta_{ij}dx^{i}dx^{j}.
\end{equation}
Concerning the above equation and scale factor $\alpha(t)$, the Friedmann equations for the evolution of $\alpha(t)$ are as follows \cite{92},
\begin{equation}\label{5}
H^{2}=\frac{1}{3M_{pl}^{2}}(\frac{1}{2}\dot{\phi}^{2}+\frac{1}{2}\dot{\chi}^{2}+V(\phi,\chi)),
\end{equation}
and
\begin{equation}\label{6}
-2\dot{H}=\frac{1}{M_{pl}}^{2}(\dot{\phi}^{2}+\dot{\chi}^{2})
\end{equation}
The slow-roll parameters such as $\epsilon$ and $\eta$ are given by,
\begin{equation}\label{7}
\epsilon=\frac{3(\dot{\phi}^{2}+\dot{\chi}^{2})}{\dot{\phi}^{2}+\dot{\chi}^{2}+2V},
\end{equation}
and
\begin{equation}\label{8}
\eta=-\frac{2(\dot{\phi}\ddot{\phi}+\dot{\chi}\ddot{\chi})}{H(\dot{\phi}^{2}+\dot{\chi}^{2})}.
\end{equation}
According to the above concepts and using the Hamilton-Jacobi equation, and also the description of inflation dynamics, we consider the following form of Hubble parameter \cite{92},
\begin{equation}\label{9}
H=H_{0}+H_{1}\phi+H_{2}\chi.
\end{equation}
Also, equations of motion with respect to the mentioned method is as follows,
\begin{equation}\label{10}
\dot{\phi}=-2M_{pl}^{2}\frac{dH}{d\phi},
\end{equation}
and
\begin{equation}\label{11}
\dot{\chi}=-2M_{pl}^{2}\frac{dH}{d\chi}
\end{equation}
Solving differential equations (\ref{10}) and (\ref{11}) one can obtain,
\begin{equation}\label{12}
\phi(t)=-2M_{pl}^{2}H_{1}t+\phi_{0}
\end{equation}
and
\begin{equation}\label{13}
\chi(t)=-2M_{pl}^{2}H_{2}t+\chi_{0}
\end{equation}
Also, the corresponding potential of two scalar fields calculated as following,
\begin{eqnarray}\label{14}
V(\phi,\chi)&=&(3H_{0}^{2}-2M_{pl}^{2}H_{1}^{2}-2M_{pl}^{2}H_{2}^{2})+6H_{1}H_{2}\phi\chi\nonumber\\
&+&6H_{0}H_{1}\phi+6H_{0}H_{2}\chi+3H_{1}^{2}\phi^{2}+3H_{2}^{2}\chi^{2}.
\end{eqnarray}
In order to calculate the scalar spectral index $(n_{s})$ and tensor-to scalar ratio \cite{EPJP}, we can use the power spectrum with respect to $C_{s}k=\alpha H$ with $C_{s}^{2}=1$ and $W_{s}\equiv\frac{(\dot{\phi}^{2}+\dot{\chi}^{2})}{2M_{pl}^{2}H^{2}}$, which yields,
\begin{equation}\label{15}
A_{s}=\frac{H^{2}}{8\pi^{2}W_{s}C{s}^{2}}.
\end{equation}
Therefore, we can obtain,
\begin{equation}\label{16}
n_{s}-1=-2\epsilon-\frac{1}{H}\frac{d}{dt}\ln\epsilon,
\end{equation}
and
\begin{equation}\label{17}
r=16\epsilon.
\end{equation}
We have introduced a two-field inflation model, and we expressed different values for cosmological parameters. In addition to the above, other cosmological parameters such as the number of enfolds, running spectrum index, etc., can be obtained. In next section, we reproduce all these values related to the above concepts. We investigate the two-fields inflation modes with respect to swampland dS conjecture. Finally, we plot some figures to determine the ranges of each of these parameters.
\section{The de Sitter conjecture}
According to all mentioned motivations, and by using a series of direct calculations, we will recalculate the potential and other cosmological parameters. We also specify the range associated with each of the parameters using the swampland dS conjecture. Then, we plot some figures, and determine these ranges. Hence, concerning equations of previous section, the potential can be written as following,
\begin{equation}\label{18}
V=3(H_{0}-2M_{pl}^{2}(H_{1}^{2}+H_{2}^{2})t)^{2}-2M_{pl}^{2}(H_{1}^{2}+H_{2}^{2}).
\end{equation}
Now, according to equation (\ref{2}) and above equation, we obtain,
\begin{equation}\label{19}
\frac{6H_{1}H_{2}}{3(H_{0}-2M_{pl}^{2}(H_{1}^{2}+H_{2}^{2})t)^{2}-2M_{pl}^{2}(H_{1}^{2}+H_{2}^{2})}>c.
\end{equation}
The above equation is the dS conjecture concerning mentioned equations. Also, using the equation (\ref{5}) we can write,
\begin{equation}\label{20}
\epsilon=\frac{4M_{pl}^{2}(H_{1}^{2}+H_{2}^{2})}{2M_{pl}^{2}(H_{0}-2M_{pl}^{2}(H_{1}^{2}+H_{2}^{2})t)^{2}}.
\end{equation}
The scalar spectral index $n_{s}$ and tensor-to-scalar ratio ($r$) regarding the equations (\ref{16}) and (\ref{17}), one can obtain,
\begin{equation}\label{21}
n_{s}-1=-\frac{8M_{pl}^{2}(H_{1}^{2}+H_{2}^{2})}{(H_{0}-2M_{pl}^{2}(H_{1}^{2}+H_{2}^{2})t)^{2}},
\end{equation}
and
\begin{equation}\label{22}
r=16\frac{4M_{pl}^{2}(H_{1}^{2}+H_{2}^{2})}{2M_{pl}^{2}(H_{0}-2M_{pl}^{2}(H_{1}^{2}+H_{2}^{2})t)^{2}}.
\end{equation}
After calculating the above-mentioned values related to cosmological parameters, we will examine the relations proportional to the swampland dS conjecture. It is noteworthy that by inverting equation (\ref{21}), two values are obtained according to the scalar spectrum index, and we will have two relations by placing them in the equation (\ref{19}). Therefore, according to the equation (\ref{21}), by inverting the function and obtaining the relation according to the scalar spectral index $(n_{s})$, and replacing it in the equation (\ref{19}), the equations related to swampland conjecture convert to the following forms,
\begin{equation}\label{23}
-\frac{6M_{pl}^{2}(-1+n_{s})^{2}t^{2}H_{1}H_{2}}{(11+n_{s})(M_{pl}^{2}(-1+n_{s})tH_{0}-2(M_{pl}^{2}+\sqrt{M_{pl}^{4}(1-(-1+n_{s})tH_{0})}))}>c,
\end{equation}
and
\begin{equation}\label{24}
-\frac{6M_{pl}^{2}(-1+n_{s})^{2}t^{2}H_{1}H_{2}}{(11+n_{s})(M_{pl}^{2}(-2+(-1+n_{s})tH_{0})+2\sqrt{M_{pl}^{4}(1+(t-n_{s}t)H_{0})})}>c.
\end{equation}
Now, we perform the same procedure for another parameter, i.e., the tensor-to-scalar ratio. So, according to the equation (\ref{22}), by inverting the function and obtaining the relation in terms of tensor-to-scalar ratio ($r$) and replacement in the equation (\ref{19}), the equations related to swampland conjecture are given as follows. In this part, precisely like the previous part, two different values are obtained,
\begin{equation}\label{25}
-\frac{6(4\sqrt{M_{pl}^{4}(4+rtH_{0})}+M_{pl}^{2}(8+rtH_{0}))H_{1}H_{2}}{M_{pl}^{2}(-48+r)H_{0}^{2}}>c,
\end{equation}
and
\begin{equation}\label{26}
-\frac{6M_{pl}^{2}r^{2}t^{2}H_{1}H_{2}}{(-48+r)(4\sqrt{M_{pl}^{4}(4+rtH_{0})}+M_{pl}^{2}(8+rtH_{0}))}>c.
\end{equation}
After the above calculations, we obtain the range appropriate to each cosmological parameters. Therefore, we determine the range of each of these cosmological parameters by plotting some figures. Of course, with respect to the values of (\ref{16}) and (\ref{17}), the relation between the two parameters of cosmology can be well obtained. Now, according to the above concepts, we plot some figures related to each cosmological parameter, so as you can see in the figures, we have plotted the range of each of the cosmological parameters, such as the scalar spectrum index and the tensor-to-scalar ratio, according to the swampland conditions. The area of each parameter has been compared according to the observable data.\\
Assuming $M_{pl}=1$, we can obtain behavior of the swampland dS conjecture using the equation (\ref{23}) which is represented by Fig. \ref{1}. For $n_{s}<1$, it is decreasing function, while for $n_{s}>1$ it is increasing function, it vanishes for $n_{s}=1$. Also, we show that increasing Hubble parameter increases value of $c$ parameter. In Fig. \ref{1} (a) we vary $H_{2}$, while in Fig. \ref{1} (b) vary $H_{1}$ to see similar result. In Fig. \ref{1} (c) we vary $H_{0}$ and see that is not so important parameter.
\begin{figure}[h!]
\begin{center}
\subfigure[]{
\includegraphics[height=4cm,width=4cm]{fig1.eps}
\label{1a}}
\subfigure[]{
\includegraphics[height=4cm,width=4cm]{fig2.eps}
\label{1b}}
\subfigure[]{
\includegraphics[height=4cm,width=4cm]{fig3.eps}
\label{1c}}
\caption{\small{Behavior of the swampland dS conjecture ($c$) in term of $n_{s}$ in units of constant parameters, consistent to the Eq. (\ref{23}).}}
\label{1}
\end{center}
\end{figure}
Similarly, we obtained behavior of the swampland dS conjecture in term of $n_{s}$ consistent to the equation (\ref{24}) and see that is decreasing function of $n_{s}$ with slow variation (see Fig. \ref{2}). Results are symmetric by variation of $H_{2}$ (Fig. \ref{2} (a)) or $H_{1}$ (Fig. \ref{2} (b)). In Fig. \ref{2} (c) we assumed unit values for $H_{2}$ and $H_{1}$, and vary $H_{0}$.
\begin{figure}[h!]
\begin{center}
\subfigure[]{
\includegraphics[height=4cm,width=4cm]{fig4.eps}
\label{2a}}
\subfigure[]{
\includegraphics[height=4cm,width=4cm]{fig5.eps}
\label{2b}}
\subfigure[]{
\includegraphics[height=4cm,width=4cm]{fig6.eps}
\label{2c}}
\caption{\small{Behavior of the swampland dS conjecture ($c$) in term of $n_{s}$ in units of constant parameters, consistent to the Eq. (\ref{24}).}}
\label{2}
\end{center}
\end{figure}
As shown in the figures, we plotted the values associated with the swampland dS conjecture and the different values obtained for each of the cosmological parameters, such as the scalar spectral index $n_{s}$ in Fig. \ref{1} and Fig. \ref{2} from equations (\ref{23}) and (\ref{24}) as well as the tensor-to-scalar ratio ($r$) in Fig. \ref{3} and Fig. \ref{4} consistent with the change to any of the Hubble parameters. In each plot, except for each mentioned Hubble parameter's changes, we assume other parameters such as $M_{pl}$ as a unit constant positive value. The range associated with these parameters is well determined. As you can see in Fig. \ref{1} and Fig. \ref{2}, the swampland conjecture is well behaved concerning the scalar-spectrum-index for the different values of the Hubble parameter in Fig. \ref{1}, which is derived from the equation (\ref{23}), and shows acceptable values. Similarly, swampland conjectures regarding the tensor-to-scalar ratio as well as the various values of the Hubble parameter per equation (\ref{26}) in Fig. \ref{4} is well defined and shows more acceptable values than by Fig. \ref{3}.\\
\begin{figure}[h!]
\begin{center}
\subfigure[]{
\includegraphics[height=4cm,width=4cm]{fig7.eps}
\label{3a}}
\subfigure[]{
\includegraphics[height=4cm,width=4cm]{fig8.eps}
\label{3b}}
\subfigure[]{
\includegraphics[height=4cm,width=4cm]{fig9.eps}
\label{3c}}
\caption{\small{Behavior of the swampland dS conjecture ($c$) in term of tensor-to-scalar ratio $r$ in units of constant parameters, consistent to the Eq. (\ref{25}).}}
\label{3}
\end{center}
\end{figure}
It is clear from Fig. \ref{4} that $c$ parameter is increasing function of $r$ as well as Hubble parameters.\\
\begin{figure}[h!]
\begin{center}
\subfigure[]{
\includegraphics[height=4cm,width=4cm]{fig10.eps}
\label{4a}}
\subfigure[]{
\includegraphics[height=4cm,width=4cm]{fig11.eps}
\label{4b}}
\subfigure[]{
\includegraphics[height=4cm,width=4cm]{fig12.eps}
\label{4c}}
\caption{\small{Behavior of the swampland dS conjecture ($c$) in term of tensor-to-scalar ratio $r$ in units of constant parameters, consistent to the Eq. (\ref{26}).}}
\label{4}
\end{center}
\end{figure}
Next, by using the equations (\ref{21}) and (\ref{22}), we plot the cosmic parameters $r$ in terms of $n_{s}$ by Fig. \ref{5}. The range of each of these parameters is determined. As you can see in the Fig. \ref{5}, these values are comparable to observable data. In these figures, we examined the change of the two cosmological parameters of the scalar-spectrum-index $(n_{s})$ and the tensor-to-scalar ratio $r$ for differences of the Hubble parameters. We assumed that the other values were positive of a unit order. It is interesting to note that the range and values obtained for single-field inflation models \cite{das,86} are more accurate than the two-field model according to the Swampland conditions and are consistent with observable data.\\
\begin{figure}[h!]
\begin{center}
\subfigure[]{
\includegraphics[height=4cm,width=4cm]{fig16.eps}
\label{6a}}
\subfigure[]{
\includegraphics[height=4cm,width=4cm]{fig17.eps}
\label{6b}}
\subfigure[]{
\includegraphics[height=4cm,width=4cm]{fig18.eps}
\label{6c}}
\caption{\small{The $(r-n_{s})$ plan in units of constant parameters. (a) by variation of $H_{2}$; (b) by variation of $H_{1}$; and (c) by variation of $H_{0}$.}}
\label{6}
\end{center}
\end{figure}
The allowable area for the scalar-spectral-index and the tensor-to-scalar ratio determined according to the Hubble parameter. As mentioned above, the obtained range is somewhat comparable to the observable data \cite{77}. Here this model compared to single-field inflation \cite{86} with the slow-roll condition concerning swampland conjecture, it is less accurate. So, the single-field inflation with swampland conjecture can be more critical in cosmological studies. This paper introduces a two-field inflation model, and we obtained each of the cosmological parameters separately with analytical and numerical analysis. According to the swampland dS conjecture, we examined the inflation model. We determined the range associated with each of the cosmological parameters by plotting some figures consistent with the observable data. In the next works, we will evaluate other implications of this cosmological model according to different conditions.
\section{Conclusions}
Different inflation models according to various conditions such as slow-roll, ultra-slow-roll, constant-roll, and other conditions already investigated in literatures, where many cosmological implications evaluated. In this paper, we investigated a new perspective of the two-field inflation model with respect to the new swampland dS conjecture. Therefore, we studied the two-field inflation model and some cosmological parameters. Then, we obtained the mentioned parameter such as the scalar spectral index $n_{s}$ and tensor-to-scalar ratio ($r$). We applied the mentioned swampland dS conjecture in this paper to our model concerning the observational data like Planck 2018 \cite{1807.06209}.\\
Finally, we compared these values with the observable data by plotting some figures. As mentioned above, the obtained range is somewhat comparable to the observable data. Here, this model compared to single-field inflation with the slow-roll condition concerning swampland conjecture, it is less accurate. So, the single-field inflation with swampland conjecture can be more important in cosmological studies. In general, the concepts expressed in this article can be examined for the multi-field inflation model by using other ideas such as slow-roll and constant-roll conditions. Also, it is interesting to apply two-field scenario to the M-theory compactifications \cite{88}. In the future works, we will discuss these concepts.
|
\section{Introduction}
\label{sec:Intro}
When connecting theory to experiment both classical and quantum mechanics (CM
and QM) must cope with the emergence of randomness and uncertainty. However,
the nature of randomness differs. While inherent in QM (Born's rule), in CM it
dynamically emerges via the instabilities associated with deterministic chaos.
In CM, dynamical systems theory provides the correct set of tools to model and
explain the emergent stochastic dynamics of both Hamiltonian and dissipative systems.
Building on previous results \cite{Anza20a} that leverage the geometric parallels
between classical and quantum state spaces (both symplectic manifolds), in this
manuscript we extend several tools of analysis for the out-of-equilibrium
phenomena of classical systems to the quantum domain. This strengthens the parallels
and provides a novel paradigm for investigating open quantum systems out of equilibrium.
More specifically, following Kolmogorov and Sinai's use of Shannon's information
theory \cite{Shan48a} to quantify degrees of deterministic chaos
\cite{Gelf56a,Kolm56a,Kolm59b,Kolm58,Kolm59,Sina59}, we show that the parallels
go even deeper and lead to new descriptive and quantitative tools. This is done by
leveraging Geometric Quantum Mechanics, an approach to quantum mechanics
based on differential geometry which gets rid of the physical redundancies intrinsic
to the standard, linear algebra, approach.
Quantum mechanics, indeed, is grounded in a formalism in which the states of a discrete
system are vectors in a complex Hilbert space $\mathcal{H}$ of generic finite
dimension $D$. However, it is well-known that such a formulation is redundant
since vectors differing only in normalization and global phase are physically
equivalent. Implementing this equivalence relation leads to the space where
quantum states live: the complex projective Hilbert space $\mathcal{P}(\mathcal{H}) \sim
\mathbb{C}P^{D-1}$. This is the starting point for a differential-geometric
formulation of quantum mechanics called \emph{geometric quantum mechanics}
\cite{STROCCHI1966,Miel68,Kibble1979,Heslot1985,Page87,And90,Gibbons1992,Ashtekar1995,Ashtekar1999,Brody2001,Carinena2007,Chruscinski2006,Marmo2010,Avron2020,Pastorello2015,Pastorello2015a,Pastorello2016,Clemente-Gallardo2013}.
It is important to stress that, while the mathematical formulation differs, the
phenomena addressed are precisely the same as standard quantum mechanics.
Geometric quantum mechanics (GQM) works with probability densities on $\mathcal{P}(\mathcal{H})$.
These are interpreted using ensemble theory, as noninteracting copies of pure
states for the same quantum system, distributed according to some measure. This
leads to the concept of---\emph{geometric quantum state} (GQS)---as an ensemble
of pure states. (For an extensive analysis we recommend Ref. \cite{Brody2001}.)
This is a more fundamental notion of quantum state than the density matrix, as
the latter can be computed from the former, but not vice versa.
Recent work provided a concrete procedure to analytically compute the GQS of an
open, finite-dimensional, quantum system interacting with another one of
arbitrary dimension (both finite and infinite) \cite{Anza20a}. This revealed
the reason why the GQS provides a more accurate description than available with
a density matrix: The GQS retains the details about how a specific ensemble of
pure states emerges from the structure of correlations between the system and
its surroundings.
Starting from this foundation, the following introduces two information-theoretic
concepts to characterize geometric quantum states. The
first is the quantum information dimension. This borrows from Renyi's notion of
the effective dimension of a continuous probability distribution, developed in
the setting of efficiently transmitting continuous variables over noisy
communication channels. Interestingly, for classical variables there are
distributions for which the dimension is not an integer---the realizations of
these are the well-known fractals \cite{Feld12}. It also has an operational
interpretation within communication theory: it is the upper bound on the
lossless compression rate for transmitting GQSs. The second GQS
characterization uses the (related) concept of dimensional geometric entropy.
Accounting for GQS dimension, this entropy quantifies the information a GQS
stores about a quantum system.
The development unfolds as follows. Section \ref{sec:GQM} gives a brief summary
of geometric quantum mechanics and the notion of geometric quantum state.
Section \ref{sec:QID} defines the quantum information dimension, while Section
\ref{sec:GDQE} introduces the dimensional geometric entropy. Sections
\ref{sec:Example1} to \ref{sec:Example4} then analyze several examples,
evaluating these quantities exactly. The first is an open quantum system
interacting with a finite-dimensional environment. The second is an open
quantum system interacting with another with an infinite-dimensional Hilbert
space (continuous degrees of freedom). The third shows how to evaluate these
quantities in a discrete-time chaotic quantum dynamics. The fourth shows how to
evaluate these quantities in the thermodynamic limit for a condensed-matter
system in its ground state. Finally, Section \ref{sec:Discussion} discusses the
results and Section \ref{sec:Conclusions} draws forward-looking conclusions.
\section{Geometric Quantum Mechanics}
\label{sec:GQM}
References
\cite{STROCCHI1966,Miel68,Kibble1979,Heslot1985,Page87,And90,Gibbons1992,Ashtekar1995,Ashtekar1999,Brody2001,Carinena2007,Chruscinski2006,Marmo2010,Avron2020,Pastorello2015,Pastorello2015a,Pastorello2016,Clemente-Gallardo2013}
lay out the mathematical physics of geometric quantum mechanics. Here, we simply
recall the aspects most relevant for our purposes. Throughout, we only address
quantum systems with a Hilbert space $\mathcal{H}$ of finite dimension $D$. In
GQM, the pure states of such systems are points $Z$ in the complex projective
space $\mathcal{P}(\mathcal{H}) \sim \mathbb{C}P^{D-1}$. Given an arbitrary basis
$\left\{\ket{b_{\alpha}}\right\}_{\alpha}$ of $\mathcal{H}$, the pure state
$Z$ has the vector representation:
\begin{align*}
\ket{Z} \coloneqq \sum_{\alpha=0}^{D-1} Z_\alpha \ket{b_\alpha}
~,
\end{align*}
where $Z_\alpha \in \mathbb{C}$ and $Z \sim \lambda Z$, with $\lambda \in
\mathbb{C}$. This space has a rich geometric structure \cite{Bengtsson2017}. In particular,
there is a well-defined metric---the \emph{Fubini-Study metric} $g_{FS}$---and
a related notion of Volume $dV_{FS}$---the Fubini-Study volume element. These
are directly connected by $dV_{FS} = \sqrt{\det g_{FS}}dZ d\overline{Z}$, where
the overbar is the complex conjugate and $dZ$ is the Lebesgue measure. While
a full explication is beyond our current scope, we simply give its explicit form in a
particular coordinate system, specified by $Z_\alpha = \sqrt{p_\alpha}e^{i\phi_\alpha}$: $dV_{FS} =
\prod_{\alpha=1}^{D-1} \left(dp_\alpha d\phi_\alpha / 2\right)$.
On $\mathcal{P}(\mathcal{H})$, one considers ensembles distributed according to a probability
density function or, more generally, a probability measure $\mu$. The simplest example is
the basic definition of the uniform measure: $d\nu_{FS} \coloneqq dV_{FS} /
V_{D-1}$, where the total Fubini-Study volume of $\mathcal{P}(\mathcal{H})$ is $V_{D-1}
= \pi^{D-1} / (D-1)!$. This determines the basic notion of uniform measure
on $\mathcal{P}(\mathcal{H})$. Calling $A$ an element of $\mathcal{P}(\mathcal{H})$'s Borel $\sigma$-algebra and adopting
the De Finetti notation, we have:
\begin{align*}
\nu_{FS}(A) = \frac{1}{V_{D-1}} \int_{A} dV_{FS}
~.
\end{align*}
In general cases, the measure $\mu$ is not uniform and one has:
\begin{align*}
\mu(A) = \int_{A} \mu(d\nu_{FS})
~.
\end{align*}
Looking at the measure-theoretic definition, if $\mu$ is absolutely continuous
with respect to $\nu_{FS}$, then there is a probability density function $q(Z =
z)$ such that:
\begin{align*}
\mu_q(A) = \int_A q(z) d\nu_{FS}
~.
\end{align*}
This is interpreted by saying that $\mu_q(d\nu_{FS}^z) = q(z)d\nu_{FS}^z$ is
the infinitesimal probability of a realization $z$---i.e., a system pure
state---belonging to an infinitesimal volume $d\nu_{FS}^z$ centered at $z \in
\mathcal{P}(\mathcal{H})$. Thus, we can think about the pure state of a quantum system as a
realization $z$ of a random variable $Z$ with sample space $\mathcal{P}(\mathcal{H})$. Here, $z$ is
distributed on $\mathcal{P}(\mathcal{H})$ according to its geometric quantum state $q(z)$ or $\mu$.
Through an abuse of language, we often refer to both the measure $\mu$ and the
density $q$ (when it exists) as geometric quantum states. This is acceptable as
they convey the same kind of information.
\begin{figure}[b]
\centering
\includegraphics[width=.9\columnwidth]{figure_01.png}
\caption{Geometric quantum state $q(Z) = \sum_{k=1}^{2^9} p_k\delta[Z-Z_k]$
on $\mathbb{C}P^1$, with coordinates $(\theta,\phi)$. It is a finite sum
of $2^9$ Dirac measures. Each point is a possible pure state
$\ket{\psi(Z_k)} = \cos \theta_k/2 \ket{0} + \sin \theta_k/2
e^{i\phi_k}\ket{1}$ in which the system can be, with probability $p_k$. The
specific value of each $p_k$ is encoded in a point's color; see legend.
}
\label{fig:GQS}
\end{figure}
Together, the triple $(\mathcal{P}(\mathcal{H}),Z,\mu)$ defines a random variable, in the classical
sense, in which the sample space is continuous and encodes the underlying
quantumness of the physical system we aim to describe. We call this a
\emph{random quantum variable} (RQV). Following standard notation, if $Z$ is
an RQV, we denote a realization with a lowercase corresponding letter $z_j$.
(This is not to be confused with the notation $Z_\alpha$, with a Greek label,
that refers to the $\alpha$-th component of the vector $\ket{Z} \in \mathcal{H}$.
To aid intuition, Fig. \ref{fig:GQS} displays an example of geometric quantum
states of a qubit, using the $(\theta,\phi)$ coordinates to represent $\mathcal{P}(\mathcal{H})$ as
the familiar Bloch sphere.
\section{Quantum Information Dimension}
\label{sec:QID}
One GQM advantage we exploit is that it allows the use of classical measure
theory and this greatly facilitates modeling of quantum behavior. The price
paid is working with an underlying manifold of nontrivial geometry, as just
noted. Intuitively, in this way GQM directly encodes ``quantumness'' in the
underlying sampling space's geometry. Building on this, we now formalize the
dimension of a geometric quantum state, extending Renyi's information theory
\cite{Renyi59} to the quantum domain.
Given $\mathcal{P}(\mathcal{H})$ and a measure $\mu$ on it, we uniformly discretize the manifold
and coarse-grain $\mu$ to obtain a discrete distribution. This is accomplished
as follows, using ``probabilities + phases'' coordinates:
$Z_\alpha = \sqrt{p_\alpha}e^{i\phi_\alpha}$. In this,
$\left\{p_\alpha\right\}_{\alpha=0}^{D-1}$
lives in a probability simplex $\Delta_{D}$, while the phase coordinates
$\left\{\phi_\alpha\right\}_{\alpha=1}^{D-1}$ live on a $(D-1)$-dimensional
torus $\mathbb{T}^{D-1}$.
Given this coordinate set, we directly partition both $\Delta_D$ and
$\mathbb{T}_{D-1}$ separately, still in a uniform fashion. More accurately,
this is a partitioning of the $\sigma$-algebra on $\mathcal{P}(\mathcal{H})$ into a finite and
discrete collection of sets. Since both $\Delta_D$ and $\mathbb{T}_{D-1}$ are
manifolds of real dimension $D-1$, this generates a uniform partition $\left\{
Q(\vec{j},\vec{k})\right\}_{\vec{j},\vec{k}}$ of $\mathcal{P}(\mathcal{H})$ with a total number of
$L^{D-1} \times L^{D-1}$ cells. This means $\bigcup_{j_1,\ldots,j_{D-1}=1}^{L}
\bigcup_{k_1,\ldots,k_{D-1}=1}^{L} Q(\vec{j},\vec{k}) = \mathcal{P}(\mathcal{H})$, with
$Q(\vec{j},\vec{k}) \cap Q(\vec{i},\vec{m}) =
\delta_{\vec{j},\vec{i}}\delta_{\vec{k},\vec{m}}Q(\vec{i},\vec{m})$. Here,
$\vec{j} = (j_1,\ldots,j_{D-1})$ is the multi-index label that runs over the
discretization $\left\{ \Delta_D(\vec{j})\right\}_{\vec{j}}$ of $\Delta_D$,
with $j_\alpha = 1,\ldots, L$ and $\vec{k} = (k_1,\ldots, k_{D-1})$ is the
multi-index label that runs over the discretization $\left\{
\mathbb{T}_{D-1}(\vec{k})\right\}_{\vec{k}}$ of $\mathbb{T}_{D-1}$, with
$k_\alpha = 1,\ldots,L$.
The reason to partition $\mathcal{P}(\mathcal{H})$ using this coordinate system is the resulting
factorization of $d\nu_{FS}$ in ``$d p \times d \phi$'':
\begin{align*}
d\nu_{FS} = dP(\Delta_D) dQ(\mathbb{T}_{D-1})
\end{align*}
where $dP(\Delta_D) = (D-1)! \delta \left(\sum_{\alpha=0}^{D-1}p_\alpha - 1
\right) dp_0 \ldots dp_{D-1}$ is the flat measure on $\Delta_D$ and
$dQ(\mathbb{T}_{D-1})$ is the flat measure on $\mathbb{T}_{D-1}$.
This allows us to separately discretize the probability simplex and the
high-dimensional torus, while also resulting in cells $Q(\vec{j},\vec{k})$
with uniform Fubini-Study volume:
\begin{align}
V_{D-1}(\epsilon) & \coloneqq \nu_{FS}(Q(\vec{j},\vec{k}))
\nonumber \\
& = \frac{1}{L^{2(D-1)}} \nonumber \\
& = \epsilon^{-2(D-1)}
~,
\label{eq:volume}
\end{align}
where $\epsilon$ sets the scale of the partition elements. Note that, despite
the discretization's specific coordinate system, the value of
$V_{D-1}(\epsilon)$ does not depend on the coordinate system used to do the
computation, thanks to the invariance of $d\nu_{FS}$. Therefore, $\left\{
Q(\vec{j},\vec{k})\right\}_{\vec{j},\vec{k}}$ defines a good uniform partition
of $\mathcal{P}(\mathcal{H})$.
Calling $z_{\vec{j},\vec{k}}$ a generic point in $Q(\vec{j},\vec{k})$, we can
now perform the coarse-graining procedure to get a discrete random variable
$Z^\epsilon$, for $Z \in \CP{D-1}$:
\begin{align*}
Z \in Q(\vec{j},\vec{k}) \quad \Rightarrow \quad Z^\epsilon = z_{\vec{j},\vec{k}}
~.
\end{align*}
The probability $p_{\vec{j},\vec{k}}$ of $Z^\epsilon \in \QJK{j}{k}$ is defined via the coarse-graining procedure:
\begin{align}
\mathrm{Pr} \left( Z \in \QJK{j}{k}\right) &= \mu_q(\QJK{j}{k}) \nonumber \\
& = p_{\vec{j},\vec{k}} \nonumber \\
& = q(z_{\vec{j},\vec{k}}) V_{D-1}(\epsilon)
~,
\label{eq:coarse}
\end{align}
where (only) in the last equality we assumed the existence of a density $q(Z)$
and, in that case, $z_{\vec{j},\vec{k}}$ is a point inside $Q(\vec{j},\vec{k})$ defined
implicitly by $q(z_{\vec{j},\vec{k}}) = \mu_q(Q(\vec{j},\vec{k})) / V_{D-1}(\epsilon)$.
For explorations in discretizing continuous random variables see the
lectures in Ref. \cite{Graf00} and relevant developments in Refs.
\cite{Brig19,Pag15}.
Given that $Z^\epsilon$ is a discrete random variable, we can now use Shannon's
discrete functional to evaluate its entropy:
\begin{align}
H\left( Z^\epsilon \right)
= - \sum_{\vec{j},\vec{k}} p_{\vec{j},\vec{k}} \log p_{\vec{j},\vec{k}}
~.
\label{eq:DiscretizedShannon}
\end{align}
Substituting Eq. (\ref{eq:volume}) into Eq. (\ref{eq:coarse}) and calling $\mathfrak{D}$ the \emph{real dimension} of the submanifold of $\mathcal{P}(\mathcal{H})$ on which $q$ has support, then as $\epsilon \to 0$ one has that $H(Z^\epsilon)$ is linear in $-\log \epsilon$ \cite{Renyi59}:
\begin{align*}
H(Z^\epsilon) \sim_{\epsilon \to 0} - \mathfrak{D} \log \epsilon + \mathfrak{h}
~.
\end{align*}
Note that $q$ can also have support on the whole manifold, in which case
$\mathfrak{D} = 2(D-1)$---the real dimension of $\mathcal{P}(\mathcal{H})$. Thus, $\mathfrak{D}$'s
value in this scaling is the \emph{information dimension} while, as we show
shortly, the offset $\mathfrak{H}$ provides a workable definition of
differential entropy for the density $q$ that accounts for the dimension of
$q$'s support.
With this in mind, we are now ready to define a geometric quantum state's
\emph{quantum information dimension}.
\begin{Definition}[Quantum Information Dimension]
Given a finite-dimensional quantum system with state space $\CP{D-1}$ and
geometric quantum state $q$, the latter's quantum information dimension $\mathfrak{D}$ is:
\begin{align}
\mathfrak{D}
\coloneqq \lim _{\epsilon \to 0} \frac{H\left( Z^\epsilon\right)}{- \log \epsilon}\label{eq:D}
~.
\end{align}
\end{Definition}
While alternative definitions are possible, see Ref. \cite{Wu10} and references
therein, they all aim to rigorously grounding the same idea. And, the
alternatives often provide identical values, assuming the validity of some
regularity conditions. The key point is that the essence and the result of the
development do not change under these alternatives. For a detailed explication
of information dimension see Refs. \cite{Renyi59,Wu10,Farm83}.
Note that there are useful theorems for explicitly calculating the information
dimension. We use these in the upcoming sections on example quantum systems.
Before proceeding to the geometric dimensional quantum entropy, though, we
briefly discuss a connection between $\mathfrak{D}$ and analog information
theory, where $\mathfrak{D}$'s classical counterpart has a direct
interpretation.
\subsection*{Information theoretic underpinning of $\mathfrak{D}$}
Quantum information theory takes inspiration from the information theory of
classical discrete sources. However, it is well known that quantum states need
real numbers to be faithfully represented. In fact, they require several complex
numbers or, equivalently, elements of $\mathbb{R}^{2n}$. So, an approach inspired
by information theory is appropriate \cite{Cove91a} if we can identify a
natural extension to situations where the random variables at hand have a
continuous sample space. As such, one can also appeal to analog or continuous
information theory. An example, relevant for our purposes, of a result from
analog information theory is the quasi-lossless compression theorem. Loosely
speaking, this answers the question ``How much can we compress the information
emitted by a continuous source, using continuous variables?''
Rather than giving the full result---for which see Ref. \cite{Wu10}---we simply
discuss the essential point. Consider a continuous source emitting realizations
$\vec{x} \in \mathcal{X}$ of a random variable $X$. We desire to compress its
information. The dimension of $\mathcal{X}$ is arbitrary, but we assume
$\mathcal{X} \subseteq \mathbb{R}^n$, for some $n$.
Compression can be achieved using $(N,K)$-codes---a pair of encoder-decoder
functions. The encoder function $f : \mathcal{X}^N \to \mathcal{Y}^K$ converts
the continuous message into appropriate discrete symbols, belonging to the
space $\mathcal{Y}$. The decoder function $g : \mathcal{Y}^K \to
\mathcal{X}^N$ performs the inverse.
Take the probability of making an error as $\delta = \mathrm{Pr}\left[
g(f(\vec{x})) \neq \vec{x}\right]$. Call $R(\epsilon)$ the infimum of $R \geq
0$ such that the $\left( N ,\lfloor RN\rfloor \right)$ code has $\delta \leq
\epsilon$ error. Assuming a linear form for the encoder and decoder, one
establishes that there is a fundamental limit to the amount of quasi-lossless
(up to $\epsilon$) compression one can reach. This limit is achievable and it
is given by the source's classical information dimension: $R(\epsilon)
\leq \mathfrak{D}(X)$.
Here, with a slight abuse of notation, we use the same symbol $\mathfrak{D}$ to
also identify the classical information dimension. We also stress that this is
only a brief and simplified summary of the comprehensive analysis performed in
Ref. \cite{Wu10}. What is relevant for our purposes is the fact that, despite
its simplicity, it is directly applicable to quantum systems. In particular, it
addresses encoding a quantum source emitting pure states $Z \in \CP{D-1}$ with
a classical continuous distribution given by the geometric quantum state $q$.
Moreover, as quantum states themselves are points on a manifold described by
continuous variables, it can also be applied to the inverse problem of
representing a continuous classical source with quantum states. While this begs
further exploration before making rigorous statements, we believe it hints at
the fact that there is an alternative way, inspired by analog information
theory, of conceptualizing quantum computing.
Before finally moving to dimensional quantum entropy, we highlight a point about
$\mathfrak{D}$. While the understanding based on encoding and communication
theory strengthens the argument for relevance, $\mathfrak{D}$'s general role in
investigating properties of geometric quantum states stands on its own, as it is
independently and rigorously defined.
\section{Dimensional Quantum Entropy}
\label{sec:GDQE}
For a given geometric quantum state, quantum information dimension gives a
notion of effective dimension. It is therefore natural that its value affects
the definition of entropy one assigns to a geometric quantum state.
The standard example comes from comparing discrete and continuous
probability distributions. In the discrete setting there is a unique entropy
definition, given by Shannon's functional:
\begin{align}
H_{\mathrm{discrete}} = - \sum_i p_i \log p_i
~.
\label{eq:DiscreteEntropy}
\end{align}
Its extension to the continuous domain, however, is not unique and
its development and use requires care.
On the one hand, Shannon's original definition of differential entropy for a
continuous variable $X$ with probability distribution $p(x)$ provides a
meaningful and interpretable quantity \cite{Shan48a}:
\begin{align}
H_{\mathrm{continuous}} = - \int p(x) \log p(x) dx
~.
\label{eq:DiffEntropy}
\end{align}
On the other, it is well known that it presents its own challenges and that
alternatives are possible. For example, it is well-known to be sensitive to
rescaling of the measure. When $dx \to k dx$ we have that
$H_{\mathrm{continuous}} \to H_{\mathrm{continuous}} + \log k$.
Thus, when a physical measure is defined up to an overall scale factor, this
quantity is defined up to an overall additive factor. This is an issue that can
often be disregarded as it does not carry physical consequences, in analogy
with the classical notion of energy, defined up to a constant. Practically,
this can be bypassed by fixing the zero point of the entropy to be given by the
uniform distribution, which is realized by using as measure the normalized
volume of the space so that a uniform density simply has constant value equal
to $1$, giving a differential entropy $H_{unif} = \log 1 =0$.
Note, too, that $H_{\mathrm{continuous}}$ can be negative, as $-\log p(x)$ can
be negative when $p(x)$ is a density. This is not a concern, since correctly
interpreting this quantity relies on the asymptotic equipartition property,
which holds for both discrete and continuous random variables, irrespective of
$H_{\mathrm{continuous}}$'s sign; see Ch. 8 of Ref. \cite{Cove91a}.
Moreover and finally, the differential entropy is appropriate only when the
distribution has integer topological dimension. This is not true, for example,
in nonlinear dynamics, in which time-asymptotic statistical states often live
on fractals due, for example, to chaotic behavior. These objects do not have
integer dimension. However, it is possible to define an entropy that takes this
rich phenomenology into account. Again, for the classical result we point to
Refs. \cite{Renyi59,Wu10}. Here, we extend this into the quantum domain as
follows.
\begin{Definition}[Dimensional quantum entropy]
Given a finite-dimensional quantum system with state space $\CP{D-1}$, geometric
quantum state $q$ with quantum information dimension $\mathfrak{D}$, we define
$q$'s dimensional quantum entropy $H_{\mathfrak{D}}\left[q\right]$ as:
\begin{align}
H_{\mathfrak{D}}\left[q\right] \coloneqq \lim_{\epsilon \to 0}
\big( H\left(Z^\epsilon\right) + \mathfrak{D} \log \epsilon\big)
~.
\label{eq:GDQE}
\end{align}
\end{Definition}
Note that this entropy is parametrized by the quantum information dimension. To
provide intuition, consider two simple cases. Shortly after, Secs. \ref{sec:Example1} to \ref{sec:Example4} present detailed example calculations.
First, assuming that $\mathfrak{D} = 0$, we see that $H_{0}\left[Z\right]$ is simply the
continuum limit of $Z^\epsilon$'s entropy $H(Z^\epsilon)$. Second, imagine we
are looking at the uniform distribution over Bloch sphere $\CP{1}$. As this is
an absolutely continuous distribution, it has quantum information dimension
$\mathfrak{D} = 2$ and, therefore, the appropriate notion of entropy should take that
into account.
We also find that when $\mathfrak{D}=2(D-1)$, $H_{\mathfrak{D}=2(D-1)}[Z]$ is equal to the
notion of geometric quantum entropy introduced, as far as we know, by Brody and
Hughston in \cite{Brody2000} and discussed in Refs. \cite{Brody2001,Anza20a,Anza20b,Anza20c}.
In the simple case of a qubit with continuous geometric quantum state $q(Z)$ this is:
\begin{align*}
H_2\left[ Z\right] & = \lim_{\epsilon \to 0}
\left( H(Z^\epsilon) + 2\log \epsilon \right)
\nonumber \\
& =- \int_{\CP{1}}\!\!\! d\nu_{FS}^Z q(Z) \log q(Z) \nonumber \\
& = -\frac{1}{2\pi}\int_{0}^1 \!\!\! dp \int_{0}^{2\pi} \!\!\! \!\!\! d\phi \,\,q(p,\phi) \log q(p,\phi)
~.
\end{align*}
We now discuss two different but related interpretations of $H_{\mathfrak{D}}$. The
first one, of purely information-theoretic nature; the second one, of physical
nature.
\subsubsection*{Information-theoretic interpretation of $H_{\mathfrak{D}}$}
Even in the classical setting, there is no unique definition of entropy for
continuous variables \cite{Rao04,Rao05}. From a resource-theoretic perspective,
one can argue that various definitions address slightly different resources.
Thus, indirectly, their interpretation can be given by identifying appropriate
operational meanings.
In our quantum setting, if $q$ is absolutely continuous, then $\mathfrak{D}= 2(D-1)$
and $H_{\mathfrak{D}}[Z]$ provides the most straightforward definition: the differential
entropy functional, see Ref. \cite{Cove91a}. This is essentially Shannon's
functional
Eq. (\ref{eq:DiscreteEntropy})
adapted to apply to a probability density, in which the sum changes into an integral.
This can be proven directly from its definition in Eq. (\ref{eq:GDQE}), with the assumption that $q$ is
absolutely continuous. Indeed, in this case we have $p_{\vec{j},\vec{k}} =
\mu_q(\QJK{j}{k})=q(Z_{\vec{j},\vec{k}})V_{D-1}(\epsilon)$, for some
$Z_{\vec{j},\vec{k}} \in Q(\vec{j},\vec{k})$. Therefore:
\begin{align}
& H_{2(D-1)}\left[ Z\right] \nonumber \\
& \quad = -\lim_{L \to \infty} \left( \sum_{\vec{j},\vec{k}}p_{\vec{j},\vec{k}} \log p_{\vec{j},\vec{k}} - 2(D-1)\log L \right)\nonumber \\
& \quad = -\lim_{L \to \infty} \frac{1}{L^{2(D-1)}}\sum_{\vec{j},\vec{k}}q(Z_{\vec{j},\vec{k}}) \log q(Z_{\vec{j},\vec{k}}) \nonumber \\
& \quad = -\int_{\CP{D-1}} \!\!\!\!\!\! d\nu_{FS} \,\, q(Z) \log q(Z)
~.
\label{eq:StSpAverage}
\end{align}
While the integral extends to the whole of $\CP{D-1}$, since $\lim_{x \to 0} x
\log x = 0$ only $q(Z)$'s actual support contributes in a nontrivial way. As
with classical continuous variables, the information-theoretic interpretation
of $H_{\mathfrak{D}}[Z]$ hinges on the \emph{asymptotic equipartition property} (AEP)
and on the fact that it characterizes the ``size''---probability decay rate---of
the stochastic process' typical set.
In short, the geometric formalism facilitates importing, \emph{mutatis
mutandis}, the tools of analog information theory (continuous variables)
into the quantum domain. This holds since we can use classical measure theory
to discuss the information-theoretic aspects of quantum states.
The price paid is that the arena where this occurs, which usually is an
arbitrary sample space, is a manifold with geometric rules dictated by quantum
physics. However, from the geometric standpoint, there is nothing special or
uniquely challenging about complex projective spaces. Thus, one can appeal to
standard results, simply by providing the correct setup.
We will argue now in more detail that this holds for the independent and
identically distributed (i.i.d.\xspace) random variables we consider. While somewhat
restrictive, the AEP for i.i.d.\xspace random variables is a fundamental result---one
that lays strong and rigorous foundations for more advanced investigations. For
present purposes, a geometric version of the quantum AEP gives the
information-theoretic interpretation of $H_{\mathfrak{D}}[Z]$.
Results on the classical differential entropy are found in Ref. \cite{Cove91a}.
Here, we provide the proper setup and discuss the results for geometric quantum
states. First, we examine more closely the i.i.d.\xspace~assumption. The projective
space of quantum states of identical systems is not the tensor product of the
projective spaces:
\begin{align*}
\mathcal{P}(\mathcal{H}_D^{\otimes N} ) \neq \mathcal{P}(\mathcal{H}_D)^{\otimes N}
~,
\end{align*}
where $\mathcal{H}_D^{\otimes N}$ is the Hilbert space of $N$ qudits,
$D=\mathrm{dim} ~\mathcal{H}_D$ and $\mathcal{P}(\mathcal{H}_D)^{\otimes N}$
is the manifold of tensor product states of $N$ qudits. This is directly seen
as $\mathcal{P}(\mathcal{H}_D^{\otimes N} ) \sim \mathbb{C}P^{D^N-1}$ while
$\mathcal{P}(\mathcal{H}_D)^{\otimes N} \sim (\mathbb{C}P^{D-1})^{\otimes N}$.
Second, and the key point, the i.i.d.\xspace~assumption guarantees that a geometric
quantum state on $\mathcal{P}(\mathcal{H}_D^{\otimes N} )$ is the product of
$N$ identical geometric quantum states on $\mathcal{P}(\mathcal{H}_D)$. More
precisely, given homogeneous coordinates $Z_{\alpha_1,\ldots,\alpha_N}$ on
$\mathcal{P}(\mathcal{H}_D^{\otimes N} )$, the submanifold of tensor product
states is described by $N$ homogeneous coordinates $\left\{X_{\alpha_i}
\right\}_{i=1}^N$, with $X_{\alpha_i}$ on the $i-$th element $\mathcal{P}(\mathcal{H})$, such that
$Z_{\alpha_1,\ldots,\alpha_N} = \prod_{i=1}^N X_{\alpha_i}$. Together with the
i.i.d.\xspace~assumption, this implies that $q(Z_{\alpha_1,\ldots,\alpha_N}) =
\prod_{i=1}^N q (X_{\alpha_i})$. Geometrically, then, i.i.d.\xspace processes live
on tensor products of the Segre variety embedded in
$\mathcal{P}(\mathcal{H}_D^{\otimes N})$.
In this way, using the tools of classical continuous-variable information
theory, one can easily prove the weak law of large numbers. The details are not
particularly insightful, in that they simply reproduce a particular proof of
the weak law of large numbers, and so are given in App.
\ref{app:QuantumWeakLaw}. In turn, this guarantees that the following geometric
asymptotic equipartition property holds for random quantum variables.
\begin{theorem}[G-AEP for i.i.d.\xspace~quantum processes]
Let $Z_1,\ldots,Z_N$ be a sequence of i.i.d.\xspace~random quantum variables drawn from
$\CP{D-1}$ according to $q(Z)$, then:
\begin{align*}
H_{\mathfrak{D}}[Z] = - \lim_{N \to \infty} \frac{1}{N} \sum_{k=1}^N \log q(z_k)
~.
\end{align*}
\end{theorem}
The limit converges weakly in probability; see the proof in App.
\ref{app:QuantumWeakLaw}. The net result establishes that $H_{\mathfrak{D}}$ is a
well-defined quantum information-theoretic entropy, with clear operational
meaning, directly imported from continuous information theory. Moreover,
this is a tools of practical use as an i.i.d.\xspace sampling of the quantum state space
produces an ergodic process. Hence, state-space averages provide can be
evaluated using sequential time averages and vice versa. Here we do not delve
more on this matter. However, we mention that a deeper and more comprehensive
analysis of the use of geometric quantum mechanics to describe quantum stochastic
processes is possible and will be reported elsewhere.
\begin{comment}
Associated with the AEP, and relevant for the information-theoretic
interpretation of $H_{\mathfrak{D}}[Z]$, is the definition of the \emph{geometric
typical set} as follows:
\begin{align}
A_\epsilon^{N} \coloneqq & \Bigg\{ (Z_1,\ldots,Z_N) \in \mathcal{P}(\mathcal{H}_D)^{\otimes N} : \nonumber \\
& \left. \,\,\, \left\vert - \frac{1}{N}\sum_{k=1}^N \log q(Z_k) - H_{\mathfrak{D}}[Z]\right\vert \leq \epsilon \right\}
\end{align}
Using the appropriate notion of volume, i.e. the Fubini-Study volume, the
properties of the typical set are imported, \emph{mutatis mutandis}, from
classical information theory. Again, see Ref. \cite{Cove91a}.
\begin{theorem}[Geometric Typical Set Properties]
\begin{subequations}
\begin{align}
&\mathrm{Pr}\left( A_\epsilon^N\right) \geq 1 - \epsilon, \mathrm{for} \, N\, \mathrm{sufficiently} \, \mathrm{large}~,\label{eq:TypicalA}\\
&(1-\epsilon)2^{N (H_{\mathfrak{D}}[Z]-\epsilon)} \leq \mathrm{Vol}\left(A^N_\epsilon \right) \leq 2^{N (H_{\mathfrak{D}}[Z]+\epsilon)}~,\label{eq:TypicalB}
\end{align}
\end{subequations}
where, given some set $A \subset \mathcal{P}(\mathcal{H}_D)^{\otimes N}$:
\begin{align*}
\mathrm{Vol} \left( A\right) \coloneqq \int_{\CP{d-1}} \!\!\!\! \!\!\!\! dV_{FS} \ldots \int_{\CP{d-1}} \!\!\!\! \!\!\!\! dV_{FS} \,\, \mathbf{1}_A~,
\end{align*}
and $\mathbf{1}_A$ is the indicator function which is $1$ on $A$ and zeros
otherwise.
\end{theorem}
Together, the AEP and the quantum properties (\ref{eq:TypicalA},
\ref{eq:TypicalB}) provide a clear operational meaning to $H_{\mathfrak{D}}[Z]$: as the
size of the typical set grows exponentially with the number of i.i.d.\xspace~RQV,
$H_{\mathfrak{D}}[Z]$ is the characteristic exponent that evaluates the ratio of
exponential growth. There is also a more practical question whose answer is
provided by $H_{\mathfrak{D}}[Z]$. This is related to its continuous character and its
definition via Eq. (\ref{eq:GDQE}). In agreement with the usual interpretation
of differential entropy, this means that, on average, if we want to describe
$Z$ with $L$-bits accuracy, meaning that we use a digital signal with $2^L$
distinct states, we need $H_{\mathfrak{D}}[Z] + L$ bits.
At this point, it is important to highlight that there is a major difference
with von Neumann entropy, due to the continuous character of $H_{\mathfrak{D}}[Z]$: As
we can have $q(Z) > 1$, we can also have situations in which $H_{\mathfrak{D}}[Z] < 0$.
This is not unphysical or problematic, quite the opposite. By looking at the
connection with the properties of the typical set, as
$\mathrm{Vol}(A_\epsilon^N) \sim 2^{NH_{\mathfrak{D}}[Z]}$, having a negative entropy is
simply the symptom of a probability distribution which has support in a small
region of $\CP{D-1}$.
Further investigations on the relation between $H_{\mathfrak{D}}[Z]$ and the von Neumann
entropy $S_{\mathrm{vN}}(\rho_q)$ of the density matrix $\rho_q$ computed from
$q(Z)$ are ongoing. In particular, with respect to the need for tools to
address measures whose support are fractal sets. This will be reported
elsewhere. Before concluding this section on properties of the geometric
entropy, we discuss one more argument in favor of $H_{\mathfrak{D}}[Z]$. This relates to
its physical interpretation in the context of statistical mechanics.
\end{comment}
\subsubsection*{Physical interpretation of $H_{\mathfrak{D}}$}
Up to this point, $H_{\mathfrak{D}}[Z]$ has only information-theoretic relevance. In
point of fact, though, it is the direct quantum counterpart of Boltzmann's
classical $H$-functional, which has been subject of numerous investigations. In
particular, see Ref. \cite{Uffink07} for a compendium. This is also partly why
we introduce this functional, rather than various alternatives
\cite{Rao04,Rao05}. In the classical setting we have:
\begin{align*}
H_{cl}[p] \coloneqq - \int_{\Gamma}
\frac{d\vec{x} d\vec{p}}{\left( 2 \pi \hbar \right)^{3N}} \,\, f(\vec{x},\vec{p}) \log f(\vec{x},\vec{p})
~,
\end{align*}
where $\Gamma$ is the classical phase space and $d\vec{x} d\vec{p} =
\prod_{\alpha = x}^{y,z}\prod_{i=1}^{N} dx_i^\alpha dp_i^\alpha$ is the volume
measure generated by its symplectic two-form: $\Omega_{cl} = \prod_{i=1}^N
d\vec{x}_i \wedge d\vec{p}_i$.
It is well-known that the classical phase space's symplectic character provides
the correct notion of volume (measure) on the manifold of pure classical states
$(\vec{x},\vec{p})$. This is physically relevant as this notion of volume is heavily
relied upon in classical statistical mechanics. In turn, pure states are Dirac measures
on $\Gamma$. This draws a clear parallel between the classical density of states
$f_{cl}(\vec{x},\vec{p})$ and the geometric quantum state $q(Z)$, especially
when written in canonically conjugate coordinates provided by probabilities and
phases.
The physical relevance, at least in the classical setting, is that the
geometric construction provides the correct mathematical underpinning to define
statistical states in classical mechanics. And, from there, we can build the
tools for statistical mechanics. This line of thinking was spearheaded, to the
best of our knowledge, by Brody and Hughston
\cite{Brody1998,Brody2000,Brody2001,Brody2007} to provide an alternative route
to the use of statistical mechanics for quantum systems. It has also been
developed further to establish a geometric approach to quantum thermodynamics
\cite{Anza20b,Brody2016}.
\section{Examples}
\label{sec:Examples}
This concludes our technical development of the geometric tools for quantum
mechanics. The next four sub-sections show how to compute with them in several
concrete physical cases, using a combination of analytical and numerical
techniques. This first addresses a quantum system in contact with a finite
environment. The second, an electron in two-dimensions. The third, chaotic
dynamics and quantum fractals, including the Baker's and Standard maps. And,
finally, we explore the thermodynamic limit.
\subsection{Case 1: Finite Environment}
\label{sec:Example1}
As a first example, consider a system $S$ that is part of a larger system $SE$
of finite dimension. In this setting $S$ develops correlations with a
finite-dimensional environment $E$. Let $d_E$ and $d_S$ denote the dimensions
of the Hilbert spaces $\mathcal{H}_E$ and $\mathcal{H}_S$ of $E$ and $S$,
respectively. Also, assume the overall system $SE$ to be in a pure state
$\ket{\psi} \in \mathcal{H}_S \otimes \mathcal{H}_E$.
If $\left\{ a_i\right\}_{i=1}^{d_S}$ is a basis of $\mathcal{H}_S$ and
$\left\{ \ket{e_\alpha}\right\}_{\alpha=1}^{d_E}$ a basis within
$\mathcal{H}_E$, we can always write \cite{Anza20a}:
\begin{align}
\ket{\psi} & = \sum_{i=1}^{d_S} \sum_{\alpha=1}^{d_E}
\psi_{i\alpha} \ket{a_i}\ket{e_\alpha} \nonumber \\
& = \sum_{\alpha=1}^{d_E} \sqrt{p_\alpha^E} \ket{\chi_\alpha^S}\ket{e_\alpha}
~.
\label{eq:rev_Schmidt}
\end{align}
Let $\left\{\Pi^E_\alpha = \mathbb{I}_S \otimes \ket{e_\alpha} \!\!
\bra{e_\alpha}\right\}_{\alpha=1}^{d_E}$ be an arbitrary set of projective
measurements on $E$. Then $p_\alpha^E = \bra{\psi} \Pi^E_\alpha \ket{\psi}$ is
the probability of finding the environment in $\ket{e_\alpha}$. And,
$\ket{\chi_\alpha}$ are the system's post-measurement states, upon finding the
environment in state $\ket{e_\alpha}$. This implies that we can always write
the system's reduced density matrix $\rho^S \coloneqq \Tr_E \ket{\psi} \!\!
\bra{\psi}$ as:
\begin{align}
\rho^S = \sum_{\alpha=1}^{d_E} p_\alpha^E \ket{\chi_\alpha^S}\!\!\bra{\chi_\alpha^S}
~.
\label{eq:rho_S}
\end{align}
One can interpret Eq. (\ref{eq:rev_Schmidt}) as a Schmidt-like decomposition in
which the sum runs from $1$ to $d_E$---the dimension of the larger of the two
systems. Note that states $\ket{\chi_\alpha^S}$ do not generally form an
orthogonal set. This environment-induced decomposition of the globally pure
state $\ket{\psi}$ provides a geometric quantum state:
\begin{align}
\mu_S = \sum_{\alpha=1}^{d_E} p_\alpha^E \delta_{{\chi_\alpha}}
~,
\label{eq:q_S}
\end{align}
where $\delta_{\chi_\alpha}$ is the Dirac measure with support on
$\chi_\alpha$---the element of $\mathcal{P}(\mathcal{H})$ corresponding to $\ket{\chi_\alpha}$.
This allows us to extract two general results for when a system interacts with
a finite-dimensional, albeit arbitrarily large, environment.
\begin{theorem}
Given a finite-dimensional quantum system $S$ interacting with another
finite-dimensional quantum environment $E$, $S$'s quantum information dimension
$\mathfrak{D}=0$.
\end{theorem}
This is easily seen from Eq. (\ref{eq:q_S}), which is a finite sum of Dirac
measures, thus having support of dimension zero: a finite number of points.
Hence, for finite environments, its support has always dimension
zero. This also allows us to draw a general result about the dimensional
quantum entropy.
\begin{theorem}\label{theo:3}
Given a finite-dimensional quantum system $S$ interacting with another
finite-dimensional quantum environment $E$, $S$'s dimensional quantum entropy is:
\begin{align*}
H_{0}\left[ q_S\right] = - \sum_{\alpha=1}^{d_E} p_\alpha^E \log p_\alpha^E
~,
\end{align*}
where $p_\alpha^E = \bra{\psi} \Pi^E_\alpha \ket{\psi}$ is the probability of
finding the environment in state $\ket{e_\alpha}$.
\end{theorem}
Two comments are in order. First, the dimensional quantum entropy is invariant
under unitary transformations operating on the system. This is easily seen as
it depends only on $p_\alpha^E$, which has the required behavior.
Second, $H_{0}\left[ q_S\right]$ can (but does not have to) scale with the size
of the environment:
\begin{align*}
H_{0}\left[ q_S\right] \leq \log d_E
~,
\end{align*}
While counterintuitive, this dependence is physically consistent. Indeed, here
we are addressing how the state of a quantum system of size $N_S = \log d_S$
results from its correlations with the state of an environment of size $N_E = \log d_E$.
Since (i) there are $d_E = 2^{N_E}$ distinct environmental states (say, $\ket{e_\alpha}$)
and (ii) via Eqs. (\ref{eq:rev_Schmidt}), (\ref{eq:rho_S}), and (\ref{eq:q_S}) each of them
specifies a pure state $S$, the geometric entropy of $S$ scales, at most,
with the environment's size.
Moreover, we can also extract a lower bound, provided by $\rho^S$'s von Neumann
entropy. Indeed, among all the geometric quantum states with a given $\rho^S$
there is one corresponding to its spectral decomposition $\rho^S = \sum_j
\lambda_j \ket{\lambda_j}\bra{\lambda_j}$. Therefore:
\begin{align*}
H_0[q]\geq S_{vN}\left[ \rho^S(q)\right]
~,
\end{align*}
where we emphasize the dependence of $\rho^S$ on $q$, given by
$[\rho^{S}(q)]_{ij} = \mathbb{E}_q\left[ Z_i \overline{Z}_j\right]$.
The choice of $\ket{e_\alpha}$ reflects physical information about the specific
problem being analyzed. For example, in a thermodynamic setting with
Hamiltonian $H = H_0 + H_{\mathrm{int}}$, with $H_0 = H_S + H_E$, we can choose
$\ket{e_\alpha}$ to be the eigenstates of $H_E$ while $\ket{a_i}$ are the
eigenstates of $H_S$. In this case, if the interaction is weak, we expect the
environment to act as a thermal bath, thus settling on a distribution
$p_\alpha^E$ quite close to a thermal equilibrium distribution $p_\alpha^E
\propto e^{-\beta e_\alpha}$, where $e_\alpha$ is the eigenvalue of $H_E$
corresponding to the eigenvector $\ket{e_\alpha}$: $H_E \ket{e_\alpha} =
e_\alpha \ket{e_\alpha}$.
This first example of calculating the quantum information dimension and
dimensional quantum entropy provides basic intuition about what these
quantities convey about a system's overall behavior resulting from its
correlations with the environment.
\subsection{Case 2: An Electron in a 2D Box}
\label{sec:Example2}
Consider a second class in which a finite quantum system interacts with a
quantum system with continuous variables. A concrete example is an electron
confined to move in a $2D$ rectangular box $\mathcal{R}_{2D} = [x_0,x_1] \times
[y_0,y_1]$ where the position and spin degrees of freedom are assumed to be
entangled. The scenario we have in mind is that of an electron confined to a
certain region in which there is a nonhomogeneous magnetic field $\vec{B}(x,y)$
generating an interaction potential $V(x,y) = (\mu_B g_s / \hbar) \vec{S}
\cdot \vec{B}(x,y)$.
We follow Ref. \cite{Anza20a}'s treatment. Let $\left\{\ket{x,y}\right\}_{x,y}$
be the eigenbasis of the position degrees of freedom and $\left\{
\ket{0},\ket{1}\right\}$ a basis for the spin degree of freedom, Ref.
\cite{Anza20a} showed that a generic state can be written as:
\begin{align*}
\ket{\psi} = \int_{x_0}^{x_1} \!\!\!\!\! dx \int_{y_0}^{y_1} \!\!\!\!\! dy f(x,y) \ket{x,y} \ket{v(x,y)}
~,
\end{align*}
with:
\begin{align*}
\int dx dy \vert f(x,y) \vert^2 & = 1 ~, \\
\ket{v(x,y)} & = \sqrt{p_0(x,y)}e^{i\phi_0(x,y)}\ket{0} \\
& \qquad + \sqrt{p_1(x,y)}e^{i\phi_1(x,y)}\ket{1} ~,\\
p_0 + p_1 & = 1 ~, ~\text{and} \\
(\phi_0,\phi_1) & \sim (0,\phi_1-\phi_0)
~.
\end{align*}
Thus, the spin degree of freedom is described by $f(x,y)$ and $\left\{p_s(x,y),
\phi_s(x,y)\right\}_{s=0,1}$.
\begin{figure*}
\includegraphics[width=\textwidth]{figure_02.pdf}
\caption{Geometric quantum states visited along a single trajectory generated
by the Extended Baker's Map with parameters $\lambda_a=\lambda_b=0.2$,
$\beta=4\pi/10$, and initial condition $(p_0,\phi_0)=(0.32865,0.98886)$.
$N=10^7$ time-steps plotted on the Bloch square $(p,\phi) \in [0,1] \times
[0,2\pi]$. Over time, due to the map's chaotic nature, even a single
trajectory covers a (strange) attractor, with self-similar (fractal)
structure. More specifically, vertically, the attractor has a uniform
structure. Horizontally, it has self-similar, fractal structure,
equivalent to a generalized Cantor set. This is demonstrated, going from
the left panel to the right, via successively magnifying small subsets of
states.
}
\label{fig:BM_Attractor}
\end{figure*}
Examining the operation of partial trace over the position degrees of freedom,
for a generic $\ket{\psi}$, we see that it gives rise to a continuous geometric
quantum state, parametrized by the coordinates $x$ and $y$:
\begin{align*}
\rho^S &= \int dx dy |f(x,y)|^2 \ket{v(x,y)}\bra{v(x,y)}\\
& = \int d\nu_{FS}^{(p,\phi)} q(p,\phi) \ket{p,\phi}\bra{p,\phi}
~,
\end{align*}
where $\ket{p,\phi}= \sqrt{1-p}\ket{0}+\sqrt{p}e^{i\phi}\ket{1}$.
The second equality in the equation above implicitly defines a distribution on the
qubit's projective Hilbert space. The procedure is detailed in in Ref. \cite{Anza20a}.
The following simply summarizes the final result.
Given an operator $\mathcal{O}$, acting only on the Hilbert space of the spin,
we have the following:
\begin{align}
\MV{\mathcal{O}} & = \int_{x_0}^{x_1} \!\!\! dx \int_{y_0}^{y_1} \!\!\!dy
|f(x,y)|^2 \mathcal{O}(v(x,y)) \nonumber \\
& = \int d\nu_{FS}^{(p,\phi)} \, q(p,\phi) \, O(p,\phi)
~,
\label{eq:integral}
\end{align}
where $q$ is a geometric quantum state that depends on $f$ and
$d\nu_{FS}^{(p,\phi)} = dp d\phi / 2\pi$ indicates the uniform Fubini-Study
measure with coordinates $(p,\phi)$. The details of how $q$ depends on $f$ and
on the Fubini-Study metric are not immediately relevant, but can be found in
Ref. \cite{Anza20a}. Here, though, we provide a concrete example to illustrate
computing $\mathfrak{D}[q]$ and $H_{\mathfrak{D}}[q]$.
\begin{figure}
\includegraphics[width=\columnwidth]{figure_03.png}
\caption{Extended Baker's Map information dimension $d_I$: The estimation
incrementally decreases the coarse-graining scale $\epsilon$ and, at each
step, calculates $H(Z^\epsilon)$. Then, excluding some initial points to
avoid saturation, it performs a least-square fit to extract the
$H(Z^\epsilon)$'s growth rate as a function of $\log1/\epsilon$. We
estimate $d_I = 1.31 \pm 0.01$. This is fully consistent with the
analytical prediction of $d_I = 1.31$, plotted in red. See Eq.
(\ref{eq:ID_BM}) and Ref. \cite{Farm83} for the analytical estimate.
}
\label{fig:BM_ID}
\end{figure}
To be concrete, let $p_1(x,y) = \frac{x-x_0}{x_1-x_0}$, $\phi_1(x,y)
= 2\pi\frac{y-y_0}{y_1-y_0}$, and $f(x,y) = \sqrt{G(x,y)}$, where $G(x,y)$ is a
$2D$ Gaussian on $\mathcal{R}_{2D}$:
\begin{align*}
G(x,y) = \left\{ \begin{array}{ll}
\frac{e^{-\frac{1}{2}\left( \frac{x-\mu_x}{\sigma_x}\right)^2}}{\mathcal{N}_x} \frac{e^{-\frac{1}{2}\left( \frac{y-\mu_y}{\sigma_y}\right)^2}}{\mathcal{N}_y}~, & (x,y) \in \mathcal{R}_{2D} \\
& \\
0 & \textrm{otherwise}
\end{array} \right.
~,
\end{align*}
where $(\mu_x,\sigma_x)$ and $(\mu_y,\sigma_y)$ are the average and variance
along the $x$ and $y$ axis, respectively. $\mathcal{N}_x$ and $\mathcal{N}_y$
are normalization factors.
\begin{figure*}
\includegraphics[width=\textwidth]{figure_04.pdf}
\caption{Standard Map on the Bloch Square: Quantum states
$(p,\phi) \in [0,1] \times [0,2\pi]$ iterated at $K=1.15$ over a uniform
grid of initial conditions. (Left) Initial distribution $n=0$: Homogeneous
distribution with $30^2$ points $(p_0^{(j)},\phi_0^{(k)}) = (j/30,2\pi
k/30)$, with $j,k=0,\ldots,29$ running over all initial conditions, each
distinctly colored. (Middle) After only $n=25$ iterations the points begin
to mix and exhibit emergent, according to whether they lie in regions of
periodic, quasiperiodic, or chaotic behavior. (Right) The long-term
distribution, after $n=1000$ iterations. The full range of behavior types
is evident.
}
\label{fig:SM}
\end{figure*}
This constructs a geometric quantum state that is absolutely continuous with
respect to $\nu_{FS}$ and therefore expressible via a probability density
$q(p,\phi)$. With the choices made, we obtain:
\begin{align*}
q(p,\phi) = 2 \pi \frac{\mathrm{exp} \left[{-\frac{1}{2} \left( \frac{p - \mu_p}{\sigma_p}\right)^2}\right]}{\mathcal{N}_p}
\frac{\mathrm{exp}
\left[{-\frac{1}{2}\left( \frac{\phi - \mu_\phi}{\sigma_\phi}\right)^2}\right]}{\mathcal{N}_\phi}
~,
\end{align*}
with $\mathcal{N}_p \coloneqq \int_0^1 dp
e^{-\frac{1}{2}\left(\frac{p-\mu_p}{\sigma_p}\right)^2}$ and $\mathcal{N}_\phi
\coloneqq \int_0^{2\pi} d\phi
e^{-\frac{1}{2}\left(\frac{\phi-\mu_\phi}{\sigma_\phi}\right)^2}$. Moreover,
$\mu_p \coloneqq \frac{\mu_x - x_0}{x_0 - x_1}$, $\sigma_p \coloneqq
\frac{\sigma_x}{x_1 - x_0}$, $\mu_\phi \coloneqq 2\pi \frac{\mu_y - y_0}{y_0 -
y_1}$, and $\sigma_\phi \coloneqq \sigma_y \frac{2\pi}{y_1- y_0}$.
What are the quantum information dimension and the dimensional geometric
entropy of $q(p,\phi)$? Since this is an absolutely continuous density
function, with support on the whole of $\mathcal{P}(\mathcal{H})$, one can directly compute the
limit in Eq.(\ref{eq:D}), obtaining $\mathfrak{D} = 2$. Moreover, $H_2[Z]$
assumes a particularly simple form due to the Gaussian character of $q(p,\phi)$:
\begin{align*}
H_{2}[Z] & = \frac{1}{2}\mathbb{E} \left[ \left(\frac{p-\mu_p}{\sigma_p} \right)^2\right] + \frac{1}{2}\mathbb{E} \left[ \left(\frac{\phi-\mu_\phi}{\sigma_\phi} \right)^2\right] +\\
& \qquad + \log \mathcal{N}_p + \log \mathcal{N}_{\phi} - \log 2\pi \\
& = \log \mathcal{N}_p + \log \mathcal{N}_{\phi} + \log \frac{e}{2\pi}
~.
\end{align*}
Thus, again, $\mathfrak{D}=2$ correctly addresses the dimensionality of the
underlying geometric quantum state and $H_2[Z]$ appropriately quantifies its
entropy.
\subsection{Case 3: Chaotic Dynamics and Quantum Fractals}
\label{sec:Example3}
While the examples above clarify the meaning of, and the technology behind,
information dimension, its strength resides in estimating the dimension of
complex probability distributions, especially those whose support is fractal
\cite{Beck93,Dorf99a,Feld12}. These objects have interesting features, such as
structural self-similarity and spontaneous statistical fluctuations, and often
arise as invariant distributions of the dynamics of complex systems. The
geometric formalism allows us to show how examples imported from the classical
theory of dynamical systems, leading up to fractal invariant sets, are part and
parcel of the phenomenology of quantum systems. In particular, by exploiting
the fact that the Fubini-Study uniform measure on $\mathbb{C}P^1$ in $(p,\phi)$
coordinates is proportional to the Lebesgue measure on the square $[0,1] \times
[0,2\pi]$, we look at two well-known examples of chaotic dynamical systems with
chaotic attractors with fractal support---the Extended Baker's Map
\cite{Farm83} and Chirikov Standard Map \cite{Chir79}. We show how to directly
implement them in quantum systems by leveraging geometric quantum mechanics.
\subsubsection{Baker's Map}
First, we look at the Extended Baker's Map (EBM) that, despite the chaotic
behaviors it generates, can be analytically solved. For a detailed discussion
about its properties, especially those related to the information dimension, we
refer to Ref. \cite{Farm83}.
This map is directly implemented on $\mathbb{C}P^1$ via the following unitary
transformations. Let $B$ denote the Extended Baker's Map, each iteration of $B$
maps a quantum state $(p,\phi) \in \mathbb{C}P^1$ to one and only one quantum
state $\left(p^\prime,\phi^\prime\right)$.
\begin{definition}[Extended Baker's Map]
\begin{align*}
\left(p^\prime,\phi^\prime\right) =
\left\{
\begin{array}{ll}
\left( \lambda_a p,2\pi\frac{\phi}{\beta}\right) & \textrm{if $\phi \leq \beta $}\\
& \\
\left( \frac{1}{2}+\lambda_b p,2\pi\frac{\phi - \beta}{2\pi-\beta}\right) & \textrm{if $\phi > \beta$}
\end{array} \right.
~.
\end{align*}
\end{definition}
Here, we use $\lambda_a \leq \lambda_b \leq \frac{1}{2}$ and $\beta \leq \pi$.
Note that the original extended Baker's Map, as in Ref. \cite{Farm83}, is
defined on the unit square $(x,y)\in [0,1]\times [0,1]$. The above adapts it to
the Bloch square via $(p\to x, \frac{\phi}{2\pi} \to y)$. As a result, $\beta$
is renormalized by a factor $2\pi$ with respect to the one $\alpha$ found in
Ref. \cite{Farm83}: $\alpha \to \beta = 2\pi \alpha$.
Since there is a one-to-one correspondence between points of the Bloch square
and points in $\mathbb{C}P^1$, the action $B[(p,\phi)] =
(p^\prime,\phi^\prime)$ can be implemented on $\mathcal{H}$ as a unitary
transformation, as follows.
First, on the qubit Hilbert space, given any $\ket{\psi}$ there is one and only one orthogonal state $\ket{\psi^{\perp}}$, up to normalization and phase. Thus, a unitary transformation that maps $\ket{\psi}$ onto another $\ket{\phi}$ is directly written as $U = \ket{\phi}\bra{\psi}+\ket{\phi^{\perp}}\bra{\psi^{\perp}}$.
Second, embedding of $\mathbb{C}P^1$ with $(p,\phi)$ coordinates onto the qubit
Hilbert space is given by:
\begin{align*}
(p,\phi) \in \mathbb{C}P^1 \to \ket{p,\phi}
= \sqrt{1-p}\ket{0}+ \sqrt{p}e^{i\phi}\ket{1} \in \mathcal{H}
~.
\end{align*}
With this, given a point $(p,\phi)$, the state orthogonal to
$\ket{p,\phi}$ is simply $\ket{1-p,\phi+\pi}$. This means
$\braket{p,\phi}{1-p,\phi+\pi}= 0$ for all $(p,\phi)$. Hence, this
results in the unitary $U := U(B)$ that implements $B$ on the
Hilbert space:
\begin{align*}
U = \ket{\psi(p^\prime,\phi^\prime)}\bra{\psi(p,\phi)}
+ \ket{\psi^{\perp}(p^\prime,\phi^\prime)} \bra{\psi^{\perp}(p,\phi)}
~.
\end{align*}
As a result, iterates of $U$ directly implement the EBM on $\mathbb{C}P^1$.
Reference \cite{Farm83} gives a detailed discussion on the map's dynamic
properties. Here, we simply recall that, given an arbitrary initial point
$(p_0,\phi_0)$, as a result of the dynamics, the point moves on a subset of the
entire phase space. Thus, the natural measure, that results from the dynamics
over infinite time, is a fractal object. More accurately, the attractor has a
uniform distribution over $\phi$ while it has the structure of an extended
Cantor set with respect to $p$. See Fig. \ref{fig:BM_Attractor} for a plot of
$10^7$ map iterates, illustrating the attractor's self-similar (fractal)
structure.
Moreover, its information dimension $d_I$ is known analytically:
\begin{align}
d_I(\alpha,\lambda_a,\lambda_b) = 1+\frac{-\alpha \log \alpha - (1-\alpha) \log (1-\alpha) }{|\alpha \log \lambda_a + (1-\alpha)\log \lambda_b| }
,
\label{eq:ID_BM}
\end{align}
where $\alpha = \beta/2\pi$. This gives a quantum information dimension
$\mathfrak{D} \approx 1.31$ and it allows us to benchmark the numerical procedure
we used to compute the information dimension, a necessary reference for cases
in which $\mathfrak{D}$ is not known. To extract the dimensional entropy we look at
the estimated zero-point of the curve $H(Z^\epsilon)$ as a function of $-\log \epsilon$.
The linear fit gives $H_{\mathfrak{D}} \approx 0.25 \pm 0.15$. See Fig. \ref{fig:BM_ID}.
\begin{figure}[h]
\centering
\includegraphics[width=\columnwidth]{figure_05.pdf}
\caption{(Left) Periodic orbit with dynamic generated by the Standard Map at
$K=2$ and initial conditions $(p_0,\phi_0)=(0.2,\pi)$. (Right) Numerical
estimation of the information dimension, obtained by extracting the growth
rate of $H(\epsilon)$, shorthand for $H(Z^\epsilon)$, as a function of
$\log(1/\epsilon)$. The estimated value is consistent, up to $2$
significant digits, with the expected $\mathfrak{D}=1$.
}
\label{fig:ID_SM_Periodic}
\end{figure}
\begin{figure}[h]
\centering
\includegraphics[width=\columnwidth]{figure_06.pdf}
\caption{(Left) Chaotic orbit with dynamic generated by the Standard Map at
$K=2$ and initial conditions $(p_0,\phi_0)=(0.1,4\pi/10)$. (Right)
Numerical estimation of the information dimension, as above. Again, the
estimated value is consistent, up to $2$ significant digits, with the
expected one of $\mathfrak{D}=2$.
}
\label{fig:ID_SM_Chaotic}
\end{figure}
\subsection*{Standard Map}
Let's shift attention to the dynamically richer Standard Map (SM). While its
original definition is given on the square of side $2\pi$, it is easily
modified to operate on the Bloch Square $[0,1]\times [0,2\pi]$, i.e.
$\mathbb{C}P^1$ in $(p,\phi)$ coordinates.
\begin{definition}[Standard Map]
\begin{align*}
\begin{array}{l}
p^{'} = p + \frac{K}{2\pi} \sin \phi \\
\\
\phi^{'} = \phi + 2\pi p~,
\end{array}
\end{align*}
\end{definition}
where $p$ is taken modulo $1$, $\phi$ modulo $2\pi$, and $(p_0,\phi_0) \in
[0,1]\times [0,2\pi]$. $K$ is a nonnegative parameter that determines the map's
degree of nonlinearity. Its value is renormalized by $2\pi$ due to the fact that
in its original definition the standard map operates on the unit square $[0,1]\times [0,1]$,
while here we work in $(p,\phi) \in [0,1]\times [0,2\pi]$.
The transformation can be implemented with a set of unitary transformations
$\{S\}$, using the same construction just described in Sec. \ref{sec:Example3}
for the EBM.
For $K=0$ only periodic and quasi-periodic orbits are possible. For $K>0$ the
map generates regions of chaotic behavior and periodic orbits. Increasing $K$,
the extent of periodic orbits decreases, yielding to larger areas of chaotic
behavior. Figure \ref{fig:SM} shows the behavior at $K=1.15$, where both
periodic orbits and chaotic behavior are clearly visible.
As a consequence of the mixed behavior across the state space, the information
dimension of the natural measure, computed over a single trajectory, depends on
the initial condition. If initial conditions lead to chaotic orbits, then we
expect $\mathfrak{D}=2$. While for periodic orbits, we expect $\mathfrak{D}=1$.
We numerically verify this using the same algorithm exploited in the previous
section to estimate EBM's information dimension. Figures
\ref{fig:ID_SM_Periodic} and \ref{fig:ID_SM_Chaotic} plot the results,
consistent with the expected values.
Analogously, for the dimensional entropy there are two different situations,
depending on whether the initial condition leads to periodic or chaotic
behavior. Since in the chaotic case we simply have a 2D integral, here we
look more closely at the second case, in which $\mathfrak{D}=1$, where the following
treatment can be applied.
Referring to Fig. \ref{fig:DQE}, a generic periodic orbit covers a
$1$-dimensional line, which is identified by a generic equation $f(\phi,p)=0$,
whose solutions are parametrized by a curve $\gamma : [0,1] \to \mathbb{C}P^1$,
or a set of them $\left\{ \gamma_i\right\}$, as in the case of Fig.
\ref{fig:ID_SM_Periodic}. In the following, assume that the set of curves
$\gamma_i$ is bijective, so that given a point $Z$ on any curve, there is one
and only one curve the point is part of, thus the functions $\gamma_i(s)$ admit
inverse $\gamma_i^{-1}: \mathbb{C}P^1 \to [0,1]$. While at first this appears
to be a restrictive assumption, one can always use this construction in cases in
which there are overlapping curves, simply by decomposing them into
nonoverlapping subparts.
\begin{figure}
\includegraphics[width=.4\textwidth]{figure_07.png}
\caption{Dimensional quantum entropy of a qubit: Geometric construction for
periodic behavior. The construction gives rise to a natural measure of
dimension $\mathfrak{D}=1$. The measure's characteristics on
$\mathbb{C}P^1$ transfer to a random variable on the unit interval in a way
that does not deform the distribution; That is, it keeps intact the
probability ratios $\frac{\nu_\gamma(\gamma_i)}{\nu_\gamma(\gamma_j)} =
\frac{\mu_\gamma(I_i)}{\mu_\gamma(I_j)} \approx
\frac{n(\gamma_i)}{n(\gamma_j)}$. This holds thanks to the fact that
$\mathfrak{D}=1$ and therefore the support of the measure is a curve,
$\gamma$, on $\mathbb{C}P^1$, that can always be parametrized by
$\gamma(s)$ with $s\in[0,1]$. The green area represents a fictitious
probability density on $\gamma$, mapped onto $[0,1]$.
}
\label{fig:DQE}
\end{figure}
Proceed in this way by analyzing each separately and, since the treatment is
formally the same for each, we simply examine one of them and drop the index
$i$. The function $\gamma$ is the nonvanishing support of the distribution
whose entropy we are evaluating. On $\gamma$, the proper notion of invariant
measure is provided by the Fubini-Study infinitesimal length element:
$dl_{FS}^{\gamma}$, where $dl_{FS}^{\gamma}\coloneqq ||\dot{\gamma}||_{FS} ds$,
where $\dot{\gamma} = (\frac{dp}{ds},\frac{d\phi}{ds})$, $|| v ||_{FS} =
\sqrt{g_{ab}^{FS}v^a v^b}$ is the Fubini-Study norm of a vector $v$ in the
tangent space and $g^{FS}$ is the Fubini-Study metric. Thus, the Fubini-Study
length of the curve $\gamma$ provides a notion of measure on $[0,1]$ that is
invariant under changes of coordinates and by unitary transformations in
$\mathbb{C}P^1$, via $\mu^\gamma_s(ds) \coloneqq dl^\gamma_{FS}=
||\dot{\gamma}(s)||_{FS}ds$. This provides the proper notion of integration on
$[0,1]$ to respect all the necessary invariance properties inherited by the
fact that the points on $\gamma$ belong to $\mathbb{C}P^1$.
Thus, given a measure $\nu^\gamma$ on $\mathbb{C}P^1$ with support on $\gamma$
and density $d\nu^\gamma = \nu(dl^\gamma_{FS}) = f(s) dl^\gamma_{FS}$, the
limit in Eq.(\ref{eq:GDQE}) can be carried out to give:
\begin{align*}
H_1[\nu^\gamma] \coloneqq -\int_0^1 f(s)\log f(s) d\mu_s^\gamma(ds)
~,
\end{align*}
where $f(s)$ is the density or, more appropriately, the Radon-Nikodym
derivative, of $\nu^\gamma$ with respect to $\mu^\gamma$.
For example, one can verify that this procedure gives the expected results in
the case of a uniform distribution. Calling $L[\gamma]$ the Fubini-Study length
of the curve we have that $f_{unif}(s) = \frac{1}{L[\gamma]}$ and entropy $\log
L[\gamma]$. It is worth noting that a most important property of this procedure
is that it facilitates computing the entropy of a $1D$ distribution on $\gamma
\in \mathbb{C}P^1$ by mapping it into the entropy of a continuous density on
$[0,1]$. This amounts to defining $f$ as the continuous density that satisfies
the following consistency constraint: For any arbitrary finite partitioning
$[0,1]=\cup_{i}I_i$, that generate a partition of $\gamma$ into a set of $N$
adjacent curves, $\gamma = \cup_i \gamma_i$, where $\gamma_i = \gamma(I_i)$,
the density $f$ is defined via the following chain of equalities:
\begin{align*}
\nu^\gamma(\gamma_i) & = \lim_{N \to \infty} \frac{n_N(\gamma_i)}{N} \\
& = \int_{I_i} f(s)\mu^{\gamma}(ds) \\
& = \mu^\gamma(I_i)
~,
\end{align*}
for any $i$ and where $n_N(\gamma_i)$ is the number of points belonging to
$\gamma_i$ in a finite (size $N$) sample of the density on $\gamma$. This
construction provides a constructive method to compute analytically $H_1$,
provided one has the form of $\gamma$ and $f(s)$. It also gives a direct way to
numerically estimate $H_1$, via the sampling provided by the dynamics:
$n_N(\gamma_i) / N \approx \nu_\gamma(I_i)$ when $N \gg 1$.
\subsection{Case 4: Thermodynamic Limit}
\label{sec:Example4}
Finally, let's shift to explore dimensions for an overtly physical setting: a
finite quantum system without symmetry that interacts with a finite, but
arbitrarily large, environment. The goal is to infer properties in the
thermodynamic limit. The generic procedure to investigate the thermodynamic
limit in geometric quantum mechanics was established and made explicit in Ref.
\cite{Anza20a}. The following adopts that procedure and investigates the
geometric quantum state of the ground state of an open-boundary $1D$ spin-$1/2$
Heisenberg chain with a broken translational symmetry, a defect, made by
removing the local magnetic field in the last qubit.
Let $\vec{\tau}$ be the system's spin operator and $\vec{\sigma}_j$ the
environment's spin operators. The total Hamiltonian is:
\begin{align*}
H = H_S + H_E(N_E) + H_{\mathrm{int}}(N_E)
~,
\end{align*}
where $N_E$ is the size of the environment, $H_S = \vec{B}\cdot \vec{\tau}$, and:
\begin{align}
& H_E(N) = \sum_{k=1}^{N_E-1} \vec{\sigma}_k \cdot \vec{\sigma_{k+1}}+ \vec{B} \cdot \vec{\sigma}_k~, \nonumber\\
& H_{\mathrm{int}} = \vec{\tau} \cdot \vec{\sigma}_1
~.
\label{eq:Heis}
\end{align}
\begin{figure}[h]
\centering
\includegraphics[width=\columnwidth]{figure_08.jpeg}
\caption{Information dimension of the geometric quantum state of a qubit
interacting with a $1D$ environment with a defective (i.e., nontranslation
invariant) Heisenberg model of progressively increasing size, see Eq.
(\ref{eq:Heis}). The entire system is in its ground state $\ket{GS(N)}$,
where the environment size is $N \in [10,22]$. Each $N$ yields a geometric
quantum state whose information dimension we estimate using the
box-counting algorithm explained in the text, extracting the slope using a
linear fit for $H_N[Z^\epsilon]$. (Inset) A collection of all the data,
together with the linear fits. (Overall) The collection of horizontal lines
displays all $M=13$ estimates and extracts the average and standard
deviation. The result yields $\mathfrak{D}^{(1)}= 0.83 \pm 0.02$. The
shaded area in green, red, and blue correspond to the areas covered by
fluctuations around the average of size $\sigma, 2\sigma$, and $3\sigma$,
respectively, where $\sigma =0.02$ is the standard deviation of the sample
of slopes.
}
\label{fig:ID_LinearFit}
\end{figure}
The defect-bearing Hamiltonian breaks translational symmetry creating a rich
geometric quantum state---one that exhibits self-similarity and fractal
structure. As we will see, the choice is supported by the numerical analysis.
For the latter we used $B_z = 0.5$ and $N \in [10,22]$. At each size $N$ we
used the Lanczos algorithm, available in Python via SciPy, to extract the
ground state $\ket{\mathrm{GS}(N)}$ and calculate the system's associated
geometric quantum state $q_{N}(Z)$. Then, for each environment size, we
estimate the information dimension using the numerical procedure used and
benchmarked in previous sections. This provides $M = 13$ different datasets to
estimate the value of $\mathfrak{D}$.
The estimation was completed in two separate ways, each yielding compatible
results. First, a linear fit was performed by identifying a common region of
linearity for all the $13$ curves $H_N[Z^\epsilon] \propto \mathfrak{D}^{(1)}
(-\log \epsilon)$ analyzed. Then, from the $13$ averages we estimated the
information dimension (and its error) from the average and standard deviation.
The results yield $\mathfrak{D}^{(1)} = 0.83 \pm 0.02$ and are summarized in
Fig. \ref{fig:ID_LinearFit}. Second, we collapsed all the data onto a unique
straight line by removing their estimated vertical offset (setting intercept
equal to $0$). We removed a single outlier, to reduce the error, and checked
that this does not appreciably change the estimation. We then performed linear
regression on the aggregated data points. The results yield $\mathfrak{D}^{(2)}
= 0.84 \pm 0.01$ and are summarized in Fig. \ref{fig:ID_LinearRegression}.
\begin{figure}[h]
\centering
\includegraphics[width=\columnwidth]{figure_09.png}
\caption{Information dimension of the same geometric quantum state, estimated
using aggregated data. To extract the information dimension in the
thermodynamic limit, we do not distinguish between points belonging to
different environment sizes. Aggregating them, we performed linear
regression to extract a prediction, with associated error, of the
curve's slope. The result yields $\mathfrak{D}^{(2)}= 0.84 \pm 0.01$.
}
\label{fig:ID_LinearRegression}
\end{figure}
\begin{figure*}[ht]
\centering
\includegraphics[width=\textwidth]{figure_10.jpeg}
\caption{Support of the geometric quantum state $\ket{GS(N_E)}$ for the ground
state of the Heisenberg defect-Hamiltonian with environment size $N=22$.
The GQS has two separate islands with internal structure that is
self-similar. (Left to Middle to Right panels) Progressively magnifying the
region around each part of the support reveals the distribution's
self-similar support, whose information dimension has been previously
estimated, in the thermodynamic limit, to be $\mathfrak{D} \approx 0.83 \pm
0.02$.
}
\label{fig:GQS_HeisenbergDefect}
\end{figure*}
Altogether, the results support the intuition that the thermodynamic limit is
witness to highly nontrivial geometric quantum states with fractal support.
Increasing the environment's size, the system converges to a self-similar
distribution, with a noninteger information dimension $\mathfrak{D}\approx 0.83
\pm 0.02$. The state set shown in Fig. \ref{fig:GQS_HeisenbergDefect} is
reminiscent of the Cantor set or, more appropriately, one of its
generalizations, e.g., the EBM's invariant distribution in the $x$ direction.
\begin{figure}[ht]
\centering
\includegraphics[width=\linewidth]{figure_11.png}
\caption{Dimensional quantum entropy $H[N_E]$ scaling in the geometric
quantum state $q$ as a function of environment size $N_E$.
The estimated linear growth $H[N_E] \sim 0.66 N_E$ and so the state
has entropy rate of $0.66$.
}
\label{fig:Entropy_HeisenbergDefect}
\end{figure}
Here,
when estimating the dimensional geometric entropy, the
situation differs from the previous examples, as we are not exploring
dynamical properties. Rather, we probe how the system's geometric quantum
state changes as a function of the environment's dimension,
assuming that the entire system is always in the ground state $\ket{GS(N_E)}$ of
a joint Hamiltonian $H_{SE}(N_E)$. Since at each fixed $N_E$ the environment
has finite size, we can directly compute the geometric entropy via its
definition, given explicitly in Thm. \ref{theo:3} and extract its linear
asymptote, giving the geometric entropy rate $h$ in the thermodynamic
limit as:
\begin{align*}
h := \lim_{N_E \to \infty} \frac{H(N_E)}{N_E}
~.
\end{align*}
Thanks to the convergence to linear scaling occurring rather quickly, an
accurate estimate of $h$ is obtained directly from the data. Thus, we have a
reliable estimate up to two significant digits: $h \approx h_{est} = 0.66
\approx 95\% \log 2$. Both data and the results of the linear fit are given in
Fig. \ref{fig:Entropy_HeisenbergDefect}.
This concludes the survey of range of informational properties of geometric
quantum states. Table \ref{tab:summary} summarizes the results. The results
leave several questions and points of discussion, to which we now turn.
After which we draw several conclusions.
\begin{table}[ht]
\small
\begin{tabular}{l|c|c}
& $\mathfrak{D}$ & $H_{\mathfrak{D}}$ \\
\hline
\hline
& & \\
Case 1 (Finite Env)& 0 & $- \sum_{\alpha}p_\alpha \log p_\alpha$ \\
Case 2 ($e^-$ Box)& 2 & $\log \mathcal{N}_p + \log \mathcal{N}_{\phi} + \log \frac{e}{2\pi}$ \\
Case 3 (Baker's Map) & $1.31 \pm 0.01$ & $0.25 \pm 0.15$ \\
Case 3 (SM-Periodic) & 1 & $-\int_0^1 f(s) \log f(s) \mu^\gamma(ds)$\\
Case 3 (SM-Chaotic) & 2 & $- \int_{\!\mathbb{C}P^1} \!d\nu_{FS} q(Z) \log q(Z)$\\
Case 4 (Thermo) & 0.83 & $0.66 N_E$\\
& & \\
\hline
\end{tabular}
\caption{Quantum information dimensions and dimensional quantum entropies for
the geometric quantum states analyzed in Secs.
\ref{sec:Example1}-\ref{sec:Example4}.
}
\label{tab:summary}
\end{table}
\section{Discussion}
\label{sec:Discussion}
It has been over a half century since Kolmogorov and followers showed that
Shannon's information theory \cite{Shan48a} provides essential dynamical
invariants for chaotic physical systems
\cite{Gelf56a,Kolm56a,Kolm59b,Kolm58,Kolm59,Sina59}. Today, practically, we
know that information theory readily applies to physical systems that evolve in
discrete time with either a discrete state space or tractable symbolic dynamics
\cite{Crut01a}. Applying the Shannon entropy functional, this involves
quantities that capture informational features with physical relevance---such
as, a system's randomness and structure. This approach has successfully
described the behavior of both Hamiltonian and dissipative classical systems.
That said, the situation is decidedly less straightforward for physical systems
with an inherently continuous sample space that lack a straightforward symbolic
dynamics. Relying on analog information theory, informational descriptions are
markedly more challenging to define and calculate.
Quantum systems belong to this category, as the space $\mathcal{P}(\mathcal{H})$ of pure states has
continuous nature. Geometric quantum mechanics brings this particular aspect of
quantum systems to the fore, describing their states as probability measures
on $\mathcal{P}(\mathcal{H})$---that is, in terms of geometric quantum states. Thus,
the geometric approach directly leads one to adapt the tools from analog
information theory to the quantum domain.
In this spirit, our development focused on information dimension and
differential entropy, initially proposed by Renyi within the context of analog
information theory. We showed that these tools provide a synthetic view of a
system's geometric quantum state: the information dimension $\mathfrak{D}$
determines the dimensionality of the state's support, while the dimensional
geometric entropy $H_{\mathfrak{D}}$ gives an appropriate differential entropy
for a geometric quantum state with information dimension $\mathfrak{D}$. Once
defined and properly interpreted, we explicitly computed their values in
several examples: a finite-dimensional quantum system interacting with finite
and infinite environments; a qubit evolving with quantum implementations of
nonlinear maps---the Extended Baker's Map and the Standard Map---and finally a
qubit in a progressively larger environment, where we extracted properties in
the thermodynamic limit.
The interest for these investigations is twofold. On the one hand, extending
the tools of dynamical systems theory to the quantum domain is a topic of broad
and long-lived interest. In point of fact, dynamical systems has led to
successful modeling and quantitative understanding of the structures and
behaviors generated by large classes of synthetic and natural systems---from
nonlinear dynamics to the modeling of population dynamics to tackling the
underlying dynamics of information occurring in a computer running classical
algorithms. On the other hand, the phenomenology of open quantum systems, in
equilibrium and far from equilibrium, is a topic of both fundamental and
applied relevance. Indeed, in the past half decade, the rise of the quantum
computing paradigm made concrete several theoretical investigations focused on
the information-theoretic properties embedded in the dynamics of open quantum
systems and the thermodynamic resources necessary for quantum information
processors to run smoothly and efficiently.
We believe the geometric approach is well suited for these goals, for the
following reasons. The notion of geometric quantum state of a system
\cite{Anza20a} encodes not only the statistics of all measurement outcomes one
can perform on the system, as with the density matrix, but also the detailed
structure of the system-environment quantum correlations that determine said
measurement statistics. Hence, determining the information-theoretic properties
of geometric quantum states gives a novel way to understand the phenomenology
of open quantum systems, whose behavior and structure result from exchanging
information-theoretic and energetic resources with an environment. We believe
this will eventually lead to new analytical tools of power sufficient to
deepen our understanding of the phenomenology of open quantum systems, both in
and out of equilibrium.
\section{Conclusions}
\label{sec:Conclusions}
The development's overtly mathematical nature suggests concluding with two
forward-looking comments. First, simple examples of geometric quantum states
yield an integer value for $\mathfrak{D}$. At least in the measure theory of
classical processes, though, it is well-known that this is not typical. There
are very interesting objects that exhibit noninteger information dimension: the
self-similar or Cantor sets, now shorthanded as \emph{fractals}. Indeed, these
structures are critical to the operation of Maxwellian demons \cite{Boyd14b}
and their modern cousins---information engines \cite{Boyd15a}. Comparing classical and quantum domains, it stands to reason that the geometric
quantum formalism provides an interesting arena in which to develop a theory of
quantum fractals. Efforts in this direction are currently ongoing and will
be reported elsewhere. The informational quantities introduced here play a
central role in this endeavor.
Second, while focused here exclusively on $\mathfrak{D}$ and
$H_{\mathfrak{D}}$, it is straightforward to appreciate that the geometric
approach allows for a richer cross-pollination between analog information
theory and quantum information theory. For example, alternative definitions for
core quantities of quantum information theory, based on the geometric approach
and inspired by analog information theory, suggest themselves as parallels of
entropy, relative entropy, mutual information, Kolmogorov-Sinai entropy rate,
excess entropy, bound information, statistical complexity, and many others.
Investigating the relations with their standard quantum counterparts---von
Neumann entropy, quantum mutual information, and the like---presents
interesting challenges. The solutions, we believe, are destined to enrich both
quantum information science and analog information theory.
\section*{Acknowledgments}
\label{sec:acknowledgments}
F.A. thanks Akram Touil, Sebastian Deffner, Marina Radulaski, Davide
Pastorello, and Davide Girolami for various discussions on the quantum
geometric formalism. F.A. and J.P.C. thank David Gier, Dhruva Karkada, Samuel
Loomis, and Ariadna Venegas-Li for helpful discussions, Dhruva Karkada for help
with Fig. \ref{fig:GQS}, and the Telluride Science Research Center for its
hospitality during visits. This material is based upon work supported by, or in
part by, Templeton World Charity Foundation Power of Information Fellowship
TWCF0336, FQXi Grant FQXi-RFP-IPW-1902, and U.S. Army Research Laboratory and
the U. S. Army Research Office under grants W911NF-18-1-0028 and
W911NF-21-1-0048.
|
\section{Introduction} \label{sec:intro}
The goal of the present work is to develop computer assisted arguments for proving the existence
of periodic orbits in state-dependent delay differential equations (SDDEs)
arising as perturbations of ordinary differential equations (ODEs).
After fixing bounds on the norm of the perturbation, the delay, and their derivatives,
our method determines (a) values of the perturbation parameter $\varepsilon$ so that the periodic orbit of the ODE persists
into the SDDE, and (b) provides explicit bounds on the distance,
in an appropriate norm, between the perturbed and unperturbed solutions.
Note that, while an ODE generates a finite
dimensional dynamical system, an SDDE --if it
generates a semi-flow at all-- has phase space as an
infinite dimensional Banach manifold
\cite{MR2019242,MR4261206,MR3642771,MR1345150,MR2084753}.
This makes the perturbation arguments fairly delicate, and one novelty of
our method is that it does not
require to consider the Cauchy problem for the SDDE
Regularity issues are also quite subtle.
For example, if the vector field generating the ODE is real analytic then
any periodic orbits are real analytic as well. Not so for SDDEs. It has been shown
that periodic solutions for SDDEs may be analytic in the neighborhood of
a certain point, and only $C^\infty$ on other portions of the orbit
\cite{MR3229655,MR3748505,MR3992063}. Moreover, it has been
conjectured that this state of affairs is generic.
Regularity questions have practical implications on the set-up of the
perturbative argument -- for example we cannot
employ analytic norms.
To manage these difficulties, we employ
an approach based on the parameterization method
\cite{MR1976079,MR1976080,MR2177465,MR2240743,MR2289544,MR2299977}.
The parameterization method is a functional analytic framework for studying invariant manifolds,
which exploits the fact that recurrent enough solutions often have much nicer properties
than solutions with arbitrary initial conditions. The idea of the method is to
formulate a chart or covering map
for the invariant object as the solution of an invariance equation,
and studying such an equation allows the problem to be attacked using all the
tools of nonlinear analysis and computational mathematics.
A much more thorough description of the parameterization method
with many applications is in
\cite{MR3467671}.
Recently a number of authors have made substantial progress using the
parameterization method to study invariant manifolds in
ill posed problems \cite{MR3900818,MR4066033,MR4120821,MR3749257},
and in particular the method has been used successfully to study
periodic and quasi-periodic solutions of SDDEs
and their attached stable/unstable manifolds
\cite{MR3501842,MR3736145,MR4112213,MR4287353,Per}.
We think of this as an application of the Poincar\'{e} program
in problems where the semi-flow theory is underdeveloped or
otherwise problematic. That is, one builds up an understanding of the
dynamics one invariant object at a time, putting aside the fact that the
dynamics in a full neighborhood of the
invariant sets may or may not make sense at all.
The paper \cite{Per} just cited
develops a-posteriori theorems for state dependent perturbations of periodic orbits
in ODEs. In the present work we implement a procedure sufficient for
verifying the hypotheses of \cite{Per} in concrete examples.
Here we have to balance two competing considerations.
On the one hand, our arguments require a great deal of quantitative
information about the perturbing functions \textit{and}
periodic orbit of the unperturbed system. On the other hand,
we do not want to restrict our attention to ODEs where explicit formulas for
periodic orbits are known analytically. Indeed, our goal is to describe an approach
which works \textit{in principle} for any ODE with an isolated periodic orbit.
In the present work these constraints are simultaneously satisfied using computer-assisted
methods of proof for the ODE. This is a very active area of research, and many viable
options exist for studying periodic solutions. A thorough review of the literature is
a task beyond the scope of this modest introduction, and we refer the interested
reader to the review articles \cite{MR2652784,VANDENBERG_Dynamics,MR4283203}
and books \cite{TUCKER_ValidatedIntroduction,MR3971222}.
In the present work, for reasons that will be elucidated throughout the manuscript,
we employ a computer-assisted method of proof wherein
the ODE is projected into a Banach space of rapidly decaying Chebyshev coefficients.
The truncated problem is solved numerically using Newton's method,
and the existence of a true solution with Chebyshev
series coefficients near our numerical approximation is proven using a
Newton-Kantorovich argument.
The approach is adapted from \cite{AJPJ},
using techniques also from \cite{JPC,Lessard2018}. A readable introduction
to these ideas is found in the first three chapters of \cite{MR3822720}.
Combining the analytical results from \cite{Per} with the mathematically
rigorous computational methods just discussed, we prove results for a number of state-dependent delayed perturbations of a van der Pol equation.
As far as we are aware, the present paper and
\cite{kevinIntegrator} - which deals with the problem of rigorous integration --
are the only two papers thus far in the literature dealing with computer assisted
proofs for SDDEs.
\begin{remark}[SDDE in electrodynamics] \label{rem:electro}
DDEs arise naturally when modeling distributed
systems where communication lags between the various subsystems
cannot be ignored. When the lags themselves depend on the state of the system, we have
SDDEs. Examples are common in biology, control theory, and epidemeology, and
we refer to \cite{MR2457636} for an expansive discussion of relevant
applications.
We single out for further discussion an interesting collection of
problems where the perturbative
arguments developed in the present work could, under suitable modifications,
be very useful.
While the classical Newtonian theory of
$N$-body interactions treats gravitational disturbances as acting instantaneously, it
would be more realistic to incorporate the finite propagation speed of light,
perhaps as a delay.
For example, since the magnitude of the gravitational force due to one body acting on another
is inversely proportional to the square of distance between them, the delay would
depend on the distance divided by the speed of light.
For systems of particles where this fraction is small,
this would result in a
state dependent perturbation of the classical equations of motion.
A much more explicit example is found in the 1949 paper of
Feynmann and Wheeler \cite{MR0032447}, where they put forward a
theory of direct interparticle action for a system of
point charges in electromagnetic interaction,
and discuss its potential advantages.
Subsequent work in this direction is found in Driver's Ph.D. dissertation
on the electromagnetic two body problem \cite{MR2613114}, and in
several of his subsequent works
\cite{MR0151110,MR0146486,MR0151110,MR240421,MR1065250}.
For a theory based on SDDEs to be symmetric under the reversal of time, it is necessary to
include an ``advanced'' delay which restores the symmetry of the equations.
This induces a counter intuitive dependence in the equations of motion on the future state of
the system, in addition to its past
\cite{MR527585,MR553005}. While dependence on future state is
an unusual feature of a physical model, there continues to be much interest
in this approach \cite{MR2639911,MR2912694,MR3389782,MR3456815}.
We mention these works mainly as an
opportunity to stress that the a-posteriori framework employed in the present work
allows for the inclusion of advanced delays, and could in principle be applied to the
electromagnetic theories above.
On the other hand, in their current formulation, the results of \cite{Per} require that
the periodic orbit of the ODE is isolated, and this never occurs in the Hamiltonian ODEs of
classical dynamics. Development of an a-posteriori framework which generalizes the
approach of \cite{Per} to systems with continuous symmetries is the subject
of an upcoming work.
\end{remark}
\begin{remark}[Regularity of the results] \label{rem:regularity}
The arguments in the present work are formulated in
$C^{1+Lip}$ spaces and, though we can bootstrap to obtain more regularity,
we do not obtain bounds on higher derivatives.
Bounds on higher derivatives can be obtained using the methods of
\cite{Per}, but this requires putting more work into the estimates.
\end{remark}
\begin{remark}[Spectral Bases: Chebyshev versus Taylor] \label{rem:ChebIntro}
The computer-assisted proofs in this paper are formulated
using a spectral representation --namely Chebyshev series-- for
the periodic orbit. For the purposes of this paper it is quite valuable to have a
representation of the unperturbed periodic solution as an object in $C^k([0,T], \mathbb{R}^d)$ (with
$k \in \mathbb{N}$,
or $k = \infty$, or even real analytic),
where $T$ is the period and $\mathbb{R}^d$ is the state space for the ODE
(rather than for example a representation of the orbit as a fixed or periodic point
in a Poincar\'e section).
This is because our method requires bounds on the size of the orbit and its derivatives,
and these bounds are easily recovered from the spectral representation.
We note that it is also very natural to study periodic solutions using
Fourier series, and to formulate computer-assisted proofs on Banach spaces
of rapidly decaying Fourier coefficients as in \cite{AJPJ}. However, the
a-posteriori theory of \cite{Per} requires bounds on solutions of
some variational equations associated with the periodic orbit, and solutions
of the variational equation are not periodic. This problem could be addressed
in Fourier space using the Floquet methods developed \cite{MR3436565}.
However, in the present work we found it
expedient to solve the variational equations directly using a Chebyshev scheme.
This requires a Chebyshev representation of the periodic orbit, which is why we
use Chebyshev series throughout.
\end{remark}
The paper is organized as follows. In Section~\ref{sec:Formulation}, we recall some
background from \cite{Per} and introduce a set of polynomial inequalities whose successful
verification leads to the existence of periodic orbits in the perturbed SDDE.
In Section~\ref{sec.alg}, we introduce our main algorithm, which provides a way of computing the coefficients of the polynomials and optimizing their variables so that the polynomial inequalities are satisfied.
In Section~\ref{sec.experi}, we apply our algorithm to prove the existence of periodic orbits in a state-dependent delayed perturbation of the Van der Pol equation. We conclude the paper in Section~\ref{sec:conclusion}.
\section{Formulation of the Problem and the Polynomial Inequalities} \label{sec:Formulation}
In this section, we recall some results from \cite{Per} and describe a set of polynomial inequalities whose successful
verification leads to the existence of the periodic orbits for the perturbed SDDE.
Consider a smooth ODE on $\mathbb{R}^d$
\begin{equation}\label{ode}
\dot x(t)=f(x(t)),
\end{equation}
and assume it has a periodic orbit $\Gamma$, which is parameterized by $K_0\colon \mathbb{T}\to \mathbb{R}^d$ where $\mathbb{T}=\mathbb{R}/\mathbb{Z}$.
In other words, $\Gamma = \{ K_0(t) : t \in [0,1] \}$.
We consider a perturbation involving a state-dependent (forward or backward) delay term, that is
\begin{equation}\label{sdde}
\dot x(t)=f(x(t))+\varepsilon P\Big(x\big(t-r(x(t))\big)\Big),
\end{equation}
where $r:\mathbb{R}^d\to \mathbb{R}$ is not restricted to be positive.
When all but one Floquet multipliers of $\Gamma$ are different from $1$, the result from \cite{Per} ensures
that the perturbed equation \eqref{sdde} also has a periodic orbit for
small enough $\varepsilon$.
In order to find values of $\varepsilon$ so that the periodic orbit
persists, we first summarize the proof in \cite{Per}. The proof there
is based on the parameterization method, first introduced in
\cite{MR2177465,MR1976080,MR1976079}.
The parameterization $K_0\colon \mathbb{T}\to \mathbb{R}^d$ of the periodic orbit for \eqref{ode} with frequency $\omega_0$ satisfies
\begin{equation} \label{po}
\omega_0DK_0(\theta)=f(K_0(\theta)).
\end{equation}
Consider perturbations $\widehat K$ and $\widehat \omega$, of $K_0$ and $\omega_0$ respectively, so that $K \,\stackrel{\mbox{\tiny\textnormal{\raisebox{0ex}[0ex][0ex]{def}}}}{=}\, K_0+\widehat K$
parameterizes the periodic orbit of the perturbed equation
\eqref{sdde} and $\omega \,\stackrel{\mbox{\tiny\textnormal{\raisebox{0ex}[0ex][0ex]{def}}}}{=}\, \omega_0+\widehat \omega$ is the new frequency. Then $\widehat K$ and
$\widehat \omega$ satisfy
\begin{equation}\label{inveq}
\omega_0 D\widehat K(\theta)-Df(K_0(\theta))\widehat K(\theta) =
B^\varepsilon(\theta,\widehat \omega,\widehat K)-\widehat \omega DK_0(\theta),
\end{equation}
where
\begin{align*}
B^\varepsilon(\theta,\widehat \omega,\widehat K)&\,\stackrel{\mbox{\tiny\textnormal{\raisebox{0ex}[0ex][0ex]{def}}}}{=}\, N(\theta, \widehat K)+\varepsilon
P(\widetilde K(\theta))-\widehat \omega D\widehat K(\theta),\\
N(\theta, \widehat K)&\,\stackrel{\mbox{\tiny\textnormal{\raisebox{0ex}[0ex][0ex]{def}}}}{=}\,
f(K_0(\theta)+\widehat K(\theta))-f(K_0(\theta))-Df(K_0(\theta))\widehat K(\theta),
\end{align*}
and the term $\widetilde K$ coming from the delay is given by
\[
\widetilde K(\theta)\,\stackrel{\mbox{\tiny\textnormal{\raisebox{0ex}[0ex][0ex]{def}}}}{=}\, K(\theta - \omega r (K(\theta))).
\]
To solve equation \eqref{inveq}, we consider the variational
equation. For any fixed $\theta_0\in \mathbb{T}$, let $\Phi(\theta;\theta_0)$
be such that
\begin{equation}\label{var}
\omega_0\frac{d}{d\theta}\Phi(\theta;\theta_0) =
Df(K_0(\theta))\Phi(\theta;\theta_0),\quad
\Phi(\theta_0;\theta_0)=Id.
\end{equation}
where $Id$ is the identify matrix in $\mathbb{R}^d$. Then the condition that
$\Gamma$ has all but one Floquet multiplier
different from $1$ is equivalent to the following nondegenerate
assumption \eqref{hypothesisH} on $\Phi(\theta_0+1;\theta_0)$:
\begin{enumerate}
\renewcommand{\theenumi}{H}
\renewcommand{\labelenumi}{(\theenumi)}
\item \label{hypothesisH}\textit{$\Phi(\theta_0+1;\theta_0)$ has a
simple eigenvalue 1 whose eigenspace is generated by
$DK_0(\theta_0)$.}
\end{enumerate}
At the point $K_0(\theta_0)$ on the periodic orbit, the tangent
space has a spectral splitting,
\begin{equation}\label{split}
T_{K_0(\theta_0)}\mathbb{R}^d=E_{\theta_0}\oplus Span\{
DK_0(\theta_0)\}.
\end{equation}
We denote the projections onto $Span\{ DK_0(\theta_0)\}$ and
$E_{\theta_0}$ as $\Pi^{\top}_{\theta_0}$ and $\Pi^{\bot}_{\theta_0}$, respectively.
Now, equation \eqref{inveq} can be solved by a fixed point approach
using the variation of constants method. The fixed point of the
following operator $\Gamma^\varepsilon$ solves \eqref{inveq}, where
\[
\Gamma^\varepsilon(\widehat \omega,\widehat K)=\begin{pmatrix}
\Gamma^\varepsilon_1(\widehat \omega,\widehat K)\\ \Gamma^\varepsilon_2(\widehat \omega,\widehat K)
\end{pmatrix},
\]
with
\begin{align*}
\Gamma^\varepsilon_1(\widehat \omega,\widehat K) & \,\stackrel{\mbox{\tiny\textnormal{\raisebox{0ex}[0ex][0ex]{def}}}}{=}\, \frac{\left\langle\int^{\theta_0+1}_{\theta_0}\Pi^{\top}_{\theta_0}\bigl(\Phi(\theta_0+1;s)B^\varepsilon(s,\widehat \omega,\widehat K)\bigr)\, ds,
DK_0(\theta_0)\right\rangle}{\left| DK_0(\theta_0)\right|^2},
\\
\Gamma^\varepsilon_2(\widehat \omega,\widehat K)(\theta) &\,\stackrel{\mbox{\tiny\textnormal{\raisebox{0ex}[0ex][0ex]{def}}}}{=}\, \Phi(\theta;\theta_0)u_0
+ \frac{1}{\omega_0}\int^{\theta}_{\theta_0}\Phi(\theta;s)\big(B^\varepsilon(s,\widehat \omega,\widehat K)-\Gamma^\varepsilon_1(\widehat \omega,\widehat K)
DK_0(s)\big)\, ds
\end{align*}
where $u_0\in E_{\theta_0}$ satisfies
\[
[Id-\Phi(\theta_0+1;\theta_0)]u_0=\frac{1}{\omega_0}\int^{\theta_0+1}_{\theta_0}\Pi^{\bot}_{\theta_0}\bigl(\Phi(\theta_0+1;s)B^\varepsilon(s,\widehat K,\widehat \omega)
\bigr) \, ds.
\]
Basically, $\Gamma^\varepsilon_2$ provides the updated $\widehat K$ with variation of
constants formula. The initial condition $u_0$ is chosen such that
$\Gamma^\varepsilon_2$ is periodic, and $\Gamma^\varepsilon_1$ makes sure that there
exists such $u_0$.
Let $I_a \,\stackrel{\mbox{\tiny\textnormal{\raisebox{0ex}[0ex][0ex]{def}}}}{=}\, [-a, a]$, and
\begin{equation}\label{space}
\mathcal{E}_\beta \,\stackrel{\mbox{\tiny\textnormal{\raisebox{0ex}[0ex][0ex]{def}}}}{=}\, \Big\{g\colon \mathbb{T}\to\mathbb{R}^d\big|~ \text{$g$ is
$C^{2}$}, ~\Big\|\frac{d^i}{d\theta^i}g(\theta)\Big\|\leq
\beta_i,\ i=0,1,2\Big\},
\end{equation}
where $\| \cdot \|$ is the $C^0$ norm. Note that we could choose any norm in $\mathbb{R}^d$
(e.g. Euclidean norm), but a different choice would lead to a slight change in the estimates.
Define the fixed point operator $\Gamma^\varepsilon$ on
$I_a\times \mathcal{E}_\beta$. For small given $\varepsilon$, if one can show
that: (i) $\Gamma^\varepsilon$ maps $I_a\times \mathcal{E}_\beta$ into itself,
and (ii) $\Gamma^\varepsilon$ is a contraction in $C^0$ distance, then there
is a fixed point $(\widehat \omega^*,\widehat K^*)$ of $\Gamma^\varepsilon$. Notice that
$\widehat \omega^*\in I_a$ and $\widehat K^*$ is $C^{1+\Lip}$, that is differentiable with
Lipschitz derivative, using the fact that the closure of $ \mathcal{E}_\beta$
is a subset of $C^{1+\Lip}$ functions. Therefore, equation
\eqref{inveq} is solved. Indeed, (i) ensures the existence of a fixed
point, and (ii) guarantees the uniqueness of the fixed point. As
demonstrated in \cite{Per}, in order to verify (i) and (ii), it is
sufficient to verify the six following inequalities.
\begin{align*}
|\Gamma^\ep_1(\widehat \omega,\widehat K)|\le a \quad &\Longleftrightarrow \quad Q(\varepsilon, a, \beta_0)\le a\\
\|\Gamma^\ep_2(\widehat \omega,\widehat K)\|\le \beta_0 \quad& \Longleftrightarrow \quad P_0(\varepsilon, a, \beta_0)\le \beta_0\\
\left\|\frac{d}{d\theta}\Gamma^\ep_2(\widehat \omega,\widehat K)\right\|\le \beta_1\quad &\Longleftrightarrow \quad P_1(\varepsilon, a, \beta_0, \beta_1)\le \beta_1\\
\left\|\frac{d^2}{d\theta^2}\Gamma^\ep_2(\widehat \omega,\widehat K)\right\|\le \beta_2 \quad &\Longleftrightarrow \quad P_2(\varepsilon, a, \beta_0, \beta_1,\beta_2)\le \beta_2\\
\Gamma^\ep \text{ is a contraction in } C^0 \quad &\Longleftrightarrow \quad \begin{cases}
\mu_1(\varepsilon, a,\beta_0,\beta_1) < 1\\
\mu_2(\varepsilon, a,\beta_0,\beta_1) < 1
\end{cases}
\end{align*}
As we shall see now in the next section, $Q,~P_0,~P_1,~P_2,~\mu_1$, and $\mu_2$ are polynomials with
computable coefficients which are determined by the unperturbed equation, the
perturbation term $P$, and the delay term $r$.
\subsection{The explicit construction of the polynomials \boldmath$Q$\unboldmath, \boldmath$P_0$\unboldmath, \boldmath$P_1$\unboldmath, \boldmath$P_2$\unboldmath, \boldmath$\mu_1$\unboldmath~and~\boldmath$\mu_2$\unboldmath} \label{sec.polys}
For a given norm $|\cdot|$ on $\mathbb{R}^d$, denote $|A|$ as the operator
norm of matrix $A$. Define constants
\begin{equation}
\label{C1}
\begin{split}
&C_{1,1} \,\stackrel{\mbox{\tiny\textnormal{\raisebox{0ex}[0ex][0ex]{def}}}}{=}\, \|\Phi(\theta_0+1;s)\|=\max_{s\in [\theta_0,\theta_0+1]}\left|\Phi(\theta_0+1;s)\right|,\\
&C_{1,2} \,\stackrel{\mbox{\tiny\textnormal{\raisebox{0ex}[0ex][0ex]{def}}}}{=}\, \|\Phi(\theta;s)\|=\max_{\theta\in [\theta_0,\theta_0+1],~s\in [\theta_0,\theta]}\left|\Phi(\theta;s)\right|,\\
&C_{1,3} \,\stackrel{\mbox{\tiny\textnormal{\raisebox{0ex}[0ex][0ex]{def}}}}{=}\, \|\Phi(\theta;\theta_0)\|=\max_{\theta\in [\theta_0,\theta_0+1]}\left|\Phi(\theta;\theta_0)\right|.
\end{split}
\end{equation}
Let
\begin{align*}
&\left\|\frac{d}{ds}\Phi(\theta_0+1;s)\right\|=\max_{s\in [\theta_0,\theta_0+1]}\left|\frac{d}{ds}\Phi(\theta_0+1;s)\right|,\\
&\left\|\frac{d}{ds}\Phi(\theta;s)\right\|=\max_{\theta\in [\theta_0,\theta_0+1],~s\in [\theta_0,\theta]}\left|\frac{d}{ds}\Phi(\theta;s)\right|,
\end{align*}
and
\begin{equation}
\label{C2}
\begin{split}
&C_{2,1} \,\stackrel{\mbox{\tiny\textnormal{\raisebox{0ex}[0ex][0ex]{def}}}}{=}\, 1+|\Phi(\theta_0+1;\theta_0)|+\left\|\frac{d}{ds}\Phi(\theta_0+1;s)\right\|,\\
&C_{2,2} \,\stackrel{\mbox{\tiny\textnormal{\raisebox{0ex}[0ex][0ex]{def}}}}{=}\, 1+\|\Phi(\theta;\theta_0)\|+\left\|\frac{d}{ds}\Phi(\theta;s)\right\|.
\end{split}
\end{equation}
Recall the spectral splitting in \eqref{split}, let
$\left\|[Id-\Phi(\theta_0+1;\theta_0)]^{-1}\big|_{E_{\theta_0}}\right\|$
be the norm of the operator $[Id-\Phi(\theta_0+1;\theta_0)]^{-1}$
defined in the space $E_{\theta_0}$, let
$\|\Pi^{\top}_{\theta_0}\|=\|\Pi^{\bot}_{\theta_0}\|$ be the norm of the projections
$\Pi^{\top}_{\theta_0}$ and $\Pi^{\bot}_{\theta_0}$, and let
\begin{equation} \label{eq:constant_M}
M \,\stackrel{\mbox{\tiny\textnormal{\raisebox{0ex}[0ex][0ex]{def}}}}{=}\, \frac{\left\|[Id-\Phi(\theta_0+1;\theta_0)]^{-1}\big|_{E_{\theta_0}}\right\|\|\Pi^{\bot}_{\theta_0}\|}{\omega_0}.
\end{equation}
Now we are ready to provide the polynomials explicitly.
\begin{equation}\label{q}
Q(\varepsilon, a, \beta_0)=\frac{\|\Pi^{\top}_{\theta_0}\|}{\big| DK_0(\theta_0)\big|}\Big(\varepsilon C_{1,1}\|P\|+C_{2,1}a\beta_0+\frac{C_{1,1}}{2}\|D^2f\|\beta_0^2\Big),
\end{equation}
where $\|P\|$ is the $C^0$ norm of $P$ in a neighborhood of the
periodic orbit of the unperturbed equation \eqref{ode} (containing
the periodic orbit for the perturbed equation \eqref{sdde}), i.e., a
small neighborhood of size $2\beta_0$ around the unperturbed periodic
orbit, and $\|D^2f\|$ is the supremum of the norm of the bilinear
operator $D^2f$ in the aforementioned neighborhood of the unperturbed
periodic orbit, see more details in section \ref{sec:bound_nbhd}. In
the following, all norms without specification mean the supremum norm.
\begin{align}
\nonumber
P_0(\varepsilon, a, \beta_0) & = \varepsilon \left(M C_{1,3} C_{1,1} +\frac{C_{1,2}}{\omega_0} \right)\|P\|+\frac{C_{1,2}}{\omega_0}\| DK_0\|a + \left(M C_{1,3} C_{2,1} + \frac{C_{2,2}}{\omega_0}\right) a\beta_0 \\
\label{p0}
& \quad +\frac{\|D^2f\|}{2} \left(M C_{1,3} C_{1,1} +\frac{C_{1,2}}{\omega_0} \right)\beta_0^2,
\\
\label{p1}
P_1(\varepsilon, a, \beta_0, \beta_1)&=\frac{1}{\omega_0}\Big(\varepsilon\|P\| +\|DK_0\|a+\|Df\circK_0\|\beta_0 + a\beta_1+\frac{1}{2}\|D^2f\|\beta_0^2\Big),
\\
\label{p2}
P_2(\varepsilon, a, \beta_0, \beta_1,\beta_2)& =\frac{1}{\omega_0}\Big(\|D^2K_0\| a+\|D^2f\circK_0\|\|DK_0\|\beta_0 + \|Df\circK_0\|\beta_1 +\Big\|\frac{d}{d\theta}B^\varepsilon\Big\|\Big),
\end{align}
where
\begin{multline*}
\Big\|\frac{d}{d\theta}B^\varepsilon\Big\|\le \varepsilon\|DP\|(\| D K_0 \| + \beta_1) [1 + (\omega _0 + a) \| Dr \| (\| D K_0 \| + \beta _1)]+a\beta_2\\+\|D^3f\|(\|DK_0\|+\beta_1)\beta_0^2+2\|D^2f\|\beta_0\beta_1.
\end{multline*}
Let
\[
\xi \,\stackrel{\mbox{\tiny\textnormal{\raisebox{0ex}[0ex][0ex]{def}}}}{=}\, \varepsilon\|DP\|+\varepsilon\|DP\|(\|DK_0\|+\beta_1)(\|Dr\|(\omega_0+a)+\|r\|)+\|D^2f\|\beta_0.
\]
Then
\begin{equation}\label{mu1}
\mu_1(\varepsilon, a,\beta_0,\beta_1)=\frac{\|\Pi^{\top}_{\theta_0}\|C_{2,1}}{\big| DK_0(\theta_0)\big|}(\beta_0+a)+\frac{\|\Pi^{\top}_{\theta_0}\|C_{1,1}}{\big| DK_0(\theta_0)\big|}\xi,
\end{equation}
\begin{multline}\label{mu2}
\mu_2(\varepsilon, a,\beta_0,\beta_1)=\left(M C_{1,3}C_{2,1}+\frac{C_{1,2}\|\Pi^{\top}_{\theta_0}\|C_{2,1}\|DK_0\|}{\omega_0\big| DK_0(\theta_0)\big|}+\frac{C_{2,2}}{\omega_0}\right)(\beta_0+a)\\
+\left( M C_{1,3} C_{1,1}+\frac{C_{1,2}\|\Pi^{\top}_{\theta_0}\|C_{1,1}\|DK_0\|}{\omega_0\big| DK_0(\theta_0)\big|}+\frac{C_{1,2}}{\omega_0}\right)\xi.
\end{multline}
Note that similar to $\|P\|$ and $\|D^2 f\|$, $\|r\|$, $\|Dr\|$, $\|DP\|$, and $\|D^3f\|$ are the supremum
norms in the same neighborhood of the unperturbed periodic orbit mentioned before.
In the next section, we introduce an algorithm which provides an efficient and automated way of (a) computing the coefficients of the polynomials $Q$, $P_0$, $P_1$, $P_2$, $\mu_1$ and $\mu_2$; and (b) optimizing their variables $\varepsilon$, $a$, $\beta_0$, $\beta_1$ and $\beta_2$ so that the polynomial inequalities are satisfied, hence providing rigorous and constructive proofs of existence of periodic orbits for the perturbed SDDE.
\section{Algorithm} \label{sec.alg}
In the section, we first introduce our main algorithm. We assume that
the $C^1$ norms of the perturbation $P$ and the forward or backward delay
$r$ are given as inputs.
After that, we provide details for each step in the algorithm. In
particular, the computations of the bounds and the constants
appearing in the inequalities are elaborated. In the algorithm, the
word {\em compute} means that we use computer-assisted proofs to
obtain rigorous enclosures of the quantities we are computing.
\newtheorem{alg}[theorem]{Algorithm}
\begin{alg
\label{alg.delay-cap}
\
\begin{enumerate}
\setlength{\itemsep}{.8em}
\renewcommand*{\theenumi}{\emph{\arabic{enumi}}}
\renewcommand*{\labelenumi}{\theenumi.}
\item [$\star$] \texttt{Input:} The model $\dot x(t)=f(x(t))$ as in
\eqref{ode}, $\|P\|$, $\|DP\|$, $\|r\|$, and $\|Dr\|$ for $P$ and
$r$ in \eqref{sdde}.
\item [$\star$] \texttt{Output:} Positive constants
$\varepsilon,~a,~\beta_0,~\beta_1,~\beta_2$ such that the inequalities
\eqref{q}, \eqref{p0}, \eqref{p1}, \eqref{p2}, \eqref{mu1}, and
\eqref{mu2} are satisfied.
\item Compute a parameterization of the periodic orbit of
\eqref{ode}, $K_0\colon \mathbb{T}\to \mathbb{R}^d$, and the frequency
$\omega_0>0$.
\item For a fixed $\theta _0 \in \mathbb{T}$, compute the
solution of the forward variational equation \eqref{var} and also
the solution of the backward variational equation,
$\Phi(\theta_0;\theta)=\left(\Phi(\theta;\theta_0)\right)^{-1}$,
which verifies
\begin{equation}\label{varb}
\omega_0\frac{d}{d\theta}\Phi(\theta_0;\theta)=-\Phi(\theta_0;\theta)Df(K_0(\theta)),
\quad \Phi(\theta_0;\theta_0)=Id.
\end{equation}
\item Compute the eigenvectors of $\Phi(\theta_0
+1;\theta_0)$.
\item Compute $\|\Pi^{\top}_{\theta_0}\|$, $\|\Pi^{\bot}_{\theta_0}\|$, and $\| [Id
- \Phi(\theta _0 + 1; \theta _0)]^{-1} | _{E _{\theta _0}} \|$.
\item Compute the constants $C_{1,1}, C_{1,2}, C_{1,3}$ given in
\eqref{C1}, $C_{2,1}, C_{2,2}$ given in \eqref{C2}, and $M$ given
in \eqref{eq:constant_M}.
\item Optimize $\varepsilon$ for $a,~\beta_0,~\beta_1,~\beta_2$ in certain
ranges so that the following inequalities are verified.
\begin{equation}\label{ineqs}
\begin{cases}
Q(\varepsilon, a, \beta_0)-a\le 0\\
P_0(\varepsilon, a, \beta_0)-\beta_0\le 0\\
P_1(\varepsilon, a, \beta_0, \beta_1)-\beta_1\le 0 \\
P_2(\varepsilon, a, \beta_0, \beta_1,\beta_2)- \beta_2\le 0\\
\mu_1(\varepsilon, a,\beta_0,\beta_1)-1 < 0\\
\mu_2(\varepsilon, a,\beta_0,\beta_1) -1< 0.
\end{cases}
\end{equation}
\end{enumerate}
\end{alg}
\begin{remark}
Note that Algorithm~\ref{alg.delay-cap} works for a class of
perturbations $P$ and delays $r$, one only needs the $C^1$ norms of
them in a neighborhood of the unperturbed periodic orbit.
\end{remark}
\begin{remark}
Note that if $\varepsilon_0$ verifies the inequalities in \eqref{ineqs} given the constants, so will any $0\le\varepsilon\le\varepsilon_0$.
\end{remark}
\begin{remark}
There are different ways of optimizing, for example, we can also view
$\|P\|$, $\| DP \|$, $\| r \|$, and $\| Dr \|$ as variables of the
polynomials and modify the optimization step in
Algorithm~\ref{alg.delay-cap} to maximize $\|P\|$, $\| DP \|$, $\| r
\|$, and $\| Dr \|$ along with $\varepsilon$. The optimization process
consists, in general, in local searches and thus the initial guesses
play an important role. The choice of the objective function depends
on the goals of the problems, which we will specify in
Section~\ref{sec.experi}.
\end{remark}
\subsection{Computer-Assisted Proofs for the Unperturbed System}
\label{sec.caps-unperturbed}
The first two steps of Algorithm~\ref{alg.delay-cap} require computing
solutions of ODEs, namely a periodic orbit of \eqref{ode} and
solutions of the forward and backward variational equations about the
periodic orbit. These steps are achieved with the tools of rigorously
validated numerics. Using a computer to produce constructive proofs of
existence of solutions of differential equations is by now
well-established, and we refer the interested reader to the survey
papers
\cite{NAKAO_VerifiedPDE,MR2652784,MR1420838,VANDENBERG_Dynamics,GOMEZ_PDESurvey,MR4283203}
and to the books
\cite{MR3971222,TUCKER_ValidatedIntroduction,MR3822720} for more
details. Our approach to compute rigorously the ingredients of Steps 1
and 2 of Algorithm~\ref{alg.delay-cap} uses Chebyshev series expansion
and a Newton-Kantorovich type theorem (i.e. the radii polynomial
approach), as presented in \cite{JPC,AJPJ,MR4292534}. More precisely,
we compute Chebyshev series expansions of the unperturbed periodic
orbit $K _0(\theta)$ of \eqref{po}, the solution $\Phi(\theta;
\theta_0)$ of the forward variational equation \eqref{var}, and the
solution $\Phi(\theta _0;\theta)$ of the backward variational equation
\eqref{varb}. For each of the three problems, a zero-finding problem
is formulated for the Chebyshev coefficients of the solution of the
ODE, which lies in the product of weighed $\ell^1$ spaces that we
denote $\ell_\nu^1$ (for some geometric decay rate $\nu \ge 1$), see
definition in Appendix~\ref{app:l1nu}.
One way to interpret the results is that we have $\bar x$, which is an
approximate finite part of the Chebyshev series, and $\tilde x$, which
is the tail part, such that
\[
x = \bar x + \tilde x,
\]
although we do not know what is exactly $\tilde x$, we have the bound
\[
\qquad \|\tilde x \|_{\ell ^1_\nu} \leq R,
\]
for some explicitly given $R>0$, typically quite small.
In order to use Chebyshev series to represent the solutions, it is
standard to rescale the problem and consider solutions defined on the
interval $[-1,1]$, see Appendix~\ref{app.cheb}. Therefore, we define
the scaling parameter $L$, which is related with the period, the
rescaled periodic orbit $O(s)$, the forward variational flow $F(s)$,
and the backward variational flow $B(s)$ as
\begin{equation*}
\begin{split}
L &= \tfrac{1}{2\omega _0} , \\
O(s) &= K _0\bigl(\tfrac{1}{2} (s + 1) \bigr), \\
F(s) &= \Phi\bigl(\tfrac{1}{2} (s + 1); 0 \bigr), \\
B(s) &= \Phi\bigl(0; \tfrac{1}{2} (s + 1) \bigr),
\end{split}
\end{equation*}
for all $s$ in $[-1,1]$, where we note that we fixed $\theta _0 = 0$. In
particular, for all $-1 \leq s, t \leq 1$,
\[
F(t)B(s) =\Phi\bigl(\tfrac{1}{2} (t + 1); 0 \bigr)\Phi\bigl(0;
\tfrac{1}{2} (s + 1) \bigr)=\Phi\bigl(\tfrac{1}{2}(t+1);
\tfrac{1}{2}(s+1)\bigr),
\]
and with this, we can compute $C_{1,2}$ in \eqref{C1} and $C _{2,2}$
in \eqref{C2}.
To solve \eqref{po}, \eqref{var}, and \eqref{varb}, we look for
Chebyshev series expansions of $O(s)$, $F(s)$ and $B(s)$, and each
solution is computed by applying the radii polynomial approach to a
specific zero-finding problem. Assume that this has been achieved,
we have the numerical approximations and estimations of their
tails, that is
\begin{equation}
\label{eq.LOFB}
\begin{aligned}
L &= \bar L + \tilde L , & |\tilde L | &\leq R, \\
O(s) &= \bar O (s) + \tilde O(s), & \| \tilde O \| &\leq R, \\
F(s) &= \bar F (s) + \tilde F(s), & \| \tilde F \| &\leq R, \\
B(s) &= \bar B (s) + \tilde B(s), & \| \tilde B \| &\leq R,
\end{aligned}
\end{equation}
in certain norms associated with the norm in $\ell ^1 _\nu$.
\subsection{Computation of the bounds}
\label{sec.compu-bounds}
The bounds in the Algorithm~\ref{alg.delay-cap} require to manage the
information from the Section~\ref{sec.caps-unperturbed} and specify
the norms. From now on, we will stick to Euclidean norm on $\mathbb{R}^d$, then, the supremum norms of vector fields are straightforward.
For the derivatives of the vector fields, we are going to
use the Fr\"obenius norm (i.e. Euclidean norm of the
vectorization) as an upper bound of their operator norms.
Note that $O \colon [-1,1] \to \mathbb{R} ^d$ and $F, B \colon [-1,1] \to \mathbb{R}
^{d \times d}$, and the norms appeared in the coefficients of the
polynomials in Algorithm~\ref{alg.delay-cap} are, in essence, the
supremum norms. Since when $\nu=1$, the $\ell ^1 _\nu$ norm is an upper
bound of the supremum norm, we then have
\begin{equation*}
\begin{split}
\| O \| &\leq \sqrt{\sum _{i = 1}^d \| O _i \| _{\ell ^1_1}^2} \leq
\sqrt{\sum _{i = 1}^d \bigl(\| \bar O _i \| _{\ell ^1_1} + R
\bigr)^2} , \\ \| F \| &\leq \sqrt{\sum _{i = 1}^d \sum _{j = 1}^d \|
F_{i,j} \| _{\ell ^1_1}^2} \leq \sqrt{\sum _{i = 1}^d \sum _{j =
1}^d \bigl(\| \bar F_{i,j} \| _{\ell ^1_1} + R \bigr)^2} ,
\end{split}
\end{equation*}
similarly for $B$.
Some of the coefficients in the polynomials involved in
Algorithm~\ref{alg.delay-cap} are now straightforward. However, there
are still a few bounds requiring more computational
effort.
\subsubsection{Bounds on Neighborhoods of the Periodic Orbit}
\label{sec:bound_nbhd}
Some quantities in Algorithm~\ref{alg.delay-cap}, namely $\| D f \|$
appearing in \eqref{p1}, $\| D^2 f\|$ appearing in \eqref{p0},
\eqref{p1} and \eqref{p2}, and $\|D ^3 f\|$ appearing in \eqref{p2},
need to be bounded in a neighborhood of the periodic $K _0$. While
these derivatives are defined everywhere, it is however enough to
consider their bounds in a neighborhood of the unperturbed periodic
orbit.
Therefore, we will rigorously provide a neighborhood enclosing the
unperturbed periodic orbit. Since we use interval arithmetic, the
enclosure will be provided in terms of a hypercube, which admits an
easy computer encoding using interval arithmetic.
Let $O(s)$ be the rigorously proved periodic orbit as in
\eqref{eq.LOFB}. If $O = (O _1, \dotsc, O _d)$ componentwise, then for each $O_ i$, we consider the optimization
problems of minimization and maximization on $s \in [-1,1]$.
Note that to obtain initial approximations of these optimizations, we
consider a non-interval optimization problem with the numerical
approximations $\bar O _i$, then we verify it using, e.g.,
\texttt{verifyconstraintglobalmin} in \textsc{intlab}, see
\cite{Rump2018}. The rigorous verification provides an interval where
the min/max is located, we then evaluate $\bar O _i$ on this
interval. Now taking into account the errors $\tilde O _i$, we are
ready to provide the infimum and the supremum as boundaries of the
hypercube containing the periodic orbit. Here, one can consider some
safety factors to make the enclosure a little bit bigger
although the process described here already provides a rigorous
enclosure. We then enlarge the hypercube by size $\beta_0$ on the upper and lower bounds, which makes sure that the periodic orbit of the perturbed equation \eqref{sdde} lies in the enlarged neighborhood.
The outputs of these bounds will be intervals containing the exact bounds,
to prevent a wrapping effect, the hypercube mesh must be adjusted
until those intervals have a small radius. This adjustment will be
model-dependent and often will be an ad-hoc process.
Once the hypercube is determined, we evaluate upper bounds of $\| D f \|$, $\| D^2
f\|$, and $\|D ^3 f\|$ on the hypercube (possibly with a hypercube
mesh) and return the maximum of these evaluations as the bounds. More precisely, for $f = (f _1, \dotsc, f_d)$, using $\partial$ derivative notation, we have that
on
the hypercube,
\begin{equation*}
\| D f \| \leq \sqrt{\sum _{i,j = 1} ^d \|\partial _i f _j \|_{\ell^1_1}^2}, \qquad
\| D ^2 f \| \leq \sqrt{\sum _{i,j,k = 1} ^d \|\partial _{i,j} f _k\|_{\ell^1_1}^2}, \qquad
\| D ^3 f \| \leq \sqrt{\sum _{i,j,k,l = 1} ^d \|\partial _{i,j,k} f _l\|_{\ell^1_1} ^2},
\end{equation*}
where $\partial _i f _j$ means the partial derivative of $f_j$ with respect to $x_i$, other expressions are similar. Note that for the first inequality above, we used the fact the Fr\"obenius norm is an upper bound of the operator norm of a matrix under Euclidean norm on $\mathbb{R}^d$. The second and third inequalities can be derived from this fact.
\subsubsection{Bounds on the Convolutions} \label{sec:conv}
We have to consider some products involved with the ODE to get several
coefficients for the polynomials in
Algorithm~\ref{alg.delay-cap}. With Chebyshev representations, the
products become convolutions. Since the $\ell ^1_\nu$ space is a
Banach algebra (see Section~\ref{app:l1nu}), we have $\| a \ast b \|
_{\ell ^1_\nu} \leq \| a \| _{\ell ^1_\nu}\| b\| _{\ell ^1_\nu}$ for
all $a,b \in \ell ^1_\nu$, where $\ast: \ell ^1_\nu \times \ell ^1_\nu
\to \ell ^1_\nu$ denotes the discrete convolution.
However, the above inequality is likely to provide overestimated
bounds which will affect the size of $\varepsilon$ in the optimization
step of the inequalities \eqref{ineqs}. To get better results, we
should avoid using the inequality as much as possible.
More precisely, let $a = \bar a + \tilde a$ and $b = \bar b + \tilde
b$ be $\ell ^1_\nu$ elements with exact truncated parts and the tail
parts. If the tail parts are bounded by $R >0$, then
\begin{equation}
\label{eq.conv-bound}
\| a \ast b \| _{\ell ^1_\nu} \leq \| \bar a \ast \bar b \| _{\ell
^1_\nu} + (\|\bar a \| _{\ell ^1_\nu} + \|\bar b \| _{\ell ^1_\nu}
+ R)R.
\end{equation}
Numerically we keep $\bar a \ast \bar b$ and the bound of the tail
$(\|\bar a \| _{\ell ^1_\nu} + \|\bar b \| _{\ell ^1_\nu} +
R)R$. Thus, we can consider a \texttt{class} that encodes the
truncated Chebyshev series and a bound of its tail. In that class we
overload different operations and make elemental operations, such as
sums, products, norms, easily computable.
Note that the bound in \eqref{eq.conv-bound} becomes more complicated
as we increase the number of the convolutions to bound, i.e. for
cubic, quartic, quintic, etc. convolutions. More precisely, we can
keep the numerical parts and let the tail parts be variables of a
polynomial, e.g. $p(s _1, s _2) = (\bar a + s _1)(\bar b + s _2)$, we
then expand everything in monomials and take into account the bounds
of the tails. If the tails are bounded by $R$, then we let $s _1 = s
_2 = R$ in the expansion.
\subsubsection{Bounds on the Derivatives}
Expressions like $C _{2,1}$ in \eqref{C2} can be bounded by using the
ODE systems. That is,
\[
\frac{d}{ds} F(1)B(s) = -F(1) L f(O(s)) B(s).
\]
The norm of the righthand side is now easily computable by
convolutions, taking care of the numerical and tail parts of $F(1)$,
$L$, $O(s)$, and $B(s)$.
Another possible way to bound the norm of the derivative of a function
is to use estimates similar to Cauchy bounds.
\subsubsection{Bounds on Triangle Meshes}
The terms $C _{1,2}$ in \eqref{C1} and $ C _{2,2}$ in \eqref{C2} are
computationally expensive because they require considering a
triangular mesh.
Indeed, the terms $F(t) B(s)$ in $C _{1,2}$ and $-F(t) L f(O(s)) B(s)$
in $C _{2,2}$ can be bounded by taking an interval mesh for the
triangle $-1 \leq s \leq t \leq 1$. That is, for a $m(m+1)/2$ mesh
size, we define the intervals
\begin{align*}
t _k &= -1 + 2 [k-1,k] / m, \\
s _j &= -1 + 2 [j-1,j] / m,
\end{align*}
for integers $k = 1, \dotsc, m$ and $j = 1, \dotsc, k$. We evaluate
the expressions in these intervals (adding the radius $R$), computing
the norms, and returning the maximum. The value of $m$ is chosen in
such a way that the maximum stagnates with respect to larger $m$.
\subsubsection{Bounds on the Projections}
The projections $\Pi _{\theta _0}^\top$ and $\Pi _{\theta _0}^\perp$
have the same norm. To bound them, we first consider the case
$d=2$. The higher dimensional case is similar.
In the two-dimensional case, the monodromy matrix has two eigenpairs
$(\lambda _i, u _i)$, $i=1,2$, without loss of generality, we assume
that $\lambda _1=1$, $\lambda _2 \ne1$. An arbitrary vector $u$ in the
plane is given by $u = a _1 u _1 + a _2
u _2 $ in the basis $\{u _1, u _2\}$. Let $\alpha$ be the smaller angle between these two
eigenvectors. Then by the trigonometric relations (law of sine) (see Figure~\ref{fig.angles})
\begin{equation*}
\frac{\| u \|}{\sin \alpha} = \frac{\| a _1 u _1 \|}{\sin \beta} =
\frac{\| a _2 u _2 \|}{\sin (\alpha - \beta)},
\end{equation*}
then
\[
\| a _1 u _1 \| = \frac{\sin \beta}{\sin \alpha} \| u \| \qquad
\text{and} \qquad \| a _2 u _2 \| = \frac{\sin (\alpha - \beta)}{\sin
\alpha} \| u \|.
\]
As a consequence, the norm of both projections can be bounded as
$\|\Pi _{\theta _0}^\top\|, \|\Pi _{\theta _0}^\perp\| \leq
\frac{1}{\sin \alpha}$. In practice, the angle $\alpha $ can be
computed by the inner product properties, that is
\begin{equation*}
\cos \alpha = \frac{u _1 \cdot u _2}{\| u _1 \| \| u _2 \|}.
\end{equation*}
Note that the eigenvectors $u _1$ and $u _2$ are those of the matrix
$F(1)$, which consists of the numerical part and the tail part. These
eigenvectors need to be verified. We can use for example
\texttt{verifyeig} in \textsc{intlab} for the rigorous verification
(see \cite{Rump2001}).
In the $d$-dimensional case, by assumption \eqref{hypothesisH}, the
monodromy matrix $F(1)$ has a simple eigenvalue $1$ with eigenvector
$u_1$. All the other eigenvectors of eigenvalues $\ne 1$ generate a
hyperplane. Let $\alpha$ be the acute angle between $u_1$ and the
hyperplane, then similar to the $d=2$ case, the norms are bounded by
$\frac{1}{\sin \alpha}$. If we consider a normal vector $\vec n$ to
the hyperplane, the angle $\gamma$ between $u_1$ and $\vec n$
satisfies $\sin \alpha=|\cos \gamma|$, see Figure~\ref{fig.angles}.
\begin{figure}[ht]
\begin{center}
\includegraphics[scale=1]{proj2d.pdf}
\includegraphics[scale=.9]{proj.pdf}
\end{center}
\caption{On the left, the angles in the two-dimensional case $d=2$, a
vector $u$ in the the span of $u _1$ and $u _2$. On the right, the
general case, $u _1$ is the eigenvector of the eigenvalue $1$ and
$\vec n$ is a normal vector of the hyperplane generated by the
other eigenvectors.}
\label{fig.angles}
\end{figure}
\subsubsection{Bounds in the complement of tangential directions}
Now we consider
$\left\|[Id-\Phi(\theta_0+1;\theta_0)]^{-1}\big|_{E_{\theta_0}}\right\|$
in Algorithm~\ref{alg.delay-cap}. Due to the hypothesis \eqref{hypothesisH}, we know that restricted to the
subspace $E _{\theta _0}$ of \eqref{split}, the matrix $Id -
\Phi(\theta_0+1;\theta_0)$ is invertible. In our experiments, see
Section~\ref{sec.experi}, since $E_{\theta_0}$ is 1-dimensional, it is easy to invert $Id -
\Phi(\theta_0+1;\theta_0)$ and compute its norm in $E_{\theta_0}$. In general, one could consider the nonzero singular values of $Id -\Phi(\theta_0+1;\theta_0)$. Since under Euclidean norm, the matrix operator norm is the largest singular value, the norm $\left\|[Id-\Phi(\theta_0+1;\theta_0)]^{-1}\big|_{E_{\theta_0}}\right\|$ we want is the reciprocal of the smallest modulus of the nonzero singular values of $Id -\Phi(\theta_0+1;\theta_0)$.
\begin{comment}
We will discuss general cases below.
If we are in the particular case that all the
eigenvalues of $\Phi(\theta_0+1;\theta_0)$ are, in modulus, less or
equal than $1$, then by Neuman series,
\begin{equation*}
\left\|[Id-\Phi(\theta_0+1;\theta_0)]^{-1}\big|_{E_{\theta_0}}\right\|
\leq \frac{1}{1 - \|\Phi(\theta_0+1;\theta_0)|_{E_{\theta_0}}\|},
\end{equation*}
which is not always easy to rigorous bound.
On the other hand, if the monodromy matrix is diagonalizable, there
are square matrices $U$ and $D$ of the form
\begin{equation*}
D = \left(
\begin{array}{c|c}
\tilde D & 0 \\ \hline
0 & 1
\end{array}
\right) \qquad \text{and} \qquad
U = \left(
\begin{array}{c|c}
\tilde U & U _1
\end{array}
\right),
\end{equation*}
such that $\Phi(\theta_0+1;\theta_0) U = U D$. Therefore,
\begin{equation*}
\left\|[Id-\Phi(\theta_0+1;\theta_0)]^{-1}\big|_{E_{\theta_0}}\right\|
= \| \tilde U (Id - \tilde D)^{-1} \tilde U ^{-1} \|,
\end{equation*}
and the different elements of $\tilde U$ and $\tilde D$ can rigorously
be enclosured by, for instance, the \texttt{verifyeig} function in
\textsc{intlab}.
In the general case, the monodromy matrix may not be diagonalizable,
which means that it has Jordan blocks. However, one
can consider the rigorous enclosure of the complex Schur decomposition
of the monodromy matrix, i.e. $Q T Q ^T$ with $Q$ orthogonal matrix
and $T$ upper triangular. Because of the hypothesis
\eqref{hypothesisH}, we can assume that $T$ takes the form
\begin{equation*}
T = \left(
\begin{array}{c|c}
\tilde T & * \\ \hline
0 & 1
\end{array}
\right),
\end{equation*}
and then
$\left\|[Id-\Phi(\theta_0+1;\theta_0)]^{-1}\big|_{E_{\theta_0}}\right\|
_2 = \|(Id - \tilde T)^{-1}\| _2$. The inverse $Id - \tilde T$ can be
obtained by a finite number of operations. Indeed, if $\tilde T$ is
splited by its diagonal, i.e. $ \tilde T = \tilde D + \tilde R$, then
\begin{multline*}
(Id-\tilde D - \tilde R)^{-1} = ((Id-\tilde D)(Id - (Id-\tilde
D)^{-1}\tilde R))^{-1} \\
= (Id - (Id-\tilde D)^{-1}\tilde R)^{-1} (Id-\tilde D)^{-1} \\
= (Id -(Id-\tilde D)^{-1}\tilde R + \dotsb + (-(Id-\tilde
D)^{-1}\tilde R)^{n-1}) (Id-\tilde D)^{-1} .
\end{multline*}
\end{comment}
\subsection{Solving the Inequalities}
The last step in Algorithm~\ref{alg.delay-cap} consists in optimizing
the inequalities \eqref{ineqs} such that, for instance, the
perturbative parameter $\varepsilon$ is maximized. The variables of
that optimization are $\varepsilon$, $a$, $\beta _0$, $\beta _1$, and
$\beta _2$. All of them must be strictly positive, that is more than
the epsilon machine, e.g. $2^{-52}$. Moreover, the result in
\cite{Per} says that $a$ and $\beta _0$ are of the same order as
$\varepsilon$ asymptotically.
The optimization problem that maximizes the $\varepsilon$ is computed
numerically. That is, we take the upper bounds of the coefficient
intervals of the polynomials in \eqref{ineqs}, we apply the numerical
maximization problem, and then we check the result with the
inequalities having the interval coefficients.
A similar process is applied when we try to optimize the class
of perturbations and forward/backward delays, namely $\| P \|$, $\| DP \|$, $\|r \|$,
and $\|Dr\|$. In this situation, one needs to play more
with the objective function of interest, which depends on the goals of
the proof itself.
To prevent a loss of information, it is convenient to expand the
polynomial in \eqref{ineqs} in monomials. Thus the coefficients in the
numerical search is as sharp as possible. Appendix~\ref{app:monopoly}
shows the explicit monomial expressions for the two
different optimizations proposed here.
In the experiments we use the Optimization Toolbox in Matlab
to get an approximation of the maximizations and \textsc{intlab} to
certify that the validity of the inequalities with the intervals.
\section{The Van der Pol Example}
\label{sec.experi}
For the unperturbed ODE, we consider the Van der Pol equation with
parameter $\mu>0$
\begin{equation}\label{vdp}
\begin{split}
\dot x_1 &= x_2 \\
\dot x_2 &= \mu (1-x_1^2)x_2-x_1.
\end{split}
\end{equation}
Recall that Algorithm \ref{alg.delay-cap} only requires the $C^1$
bounds of $P$ and $r$ on a bounded set. Hence, we will not fix $P$ and
$r$, but rather work with a class of $P$'s and $r$'s.
\subsection{Results}
The main results consists of three theorems, namely
Theorem~\ref{thm.vdp-ars}, Theorem~\ref{thm.vdp-sdde0} and
Theorem~\ref{thm.vdp-sdde}. The first theorem computes and proves the
existence of the periodic orbit, the forward, and the backward
variational flows of the ODE in \eqref{vdp}. It uses the radii
polynomial approach, see Appendix~\ref{app.radii}, which provides and
explicit distance between the numerical approximation and the exact
solution. The second theorem proves the existence of periodic orbits
in state-dependent delay perturbations of \eqref{vdp} (in the form of
\eqref{sdde}) for some values of the perturbation parameter $\varepsilon$, given the norms of the perturbations and the forward or
backward delays. Moreover, we have estimations on the differences
between the frequencies and the periodic orbits before and after the
perturbation. The third theorem establishes that the perturbation and delay terms can be more general.
\begin{theorem} \label{thm.vdp-ars}
Fix a parameter value $\mu \in \{ j/10:j=1,\dots,10\}$. Let $\bar O
\colon [-1,1] \to \mathbb{R} ^2$ be a numerical solution of the periodic
orbit of \eqref{vdp} (passing close to the point $(2,0) \in \mathbb{R}^2$)
consisting of $200$ Chebyshev coefficients (coordinatewise), and
$2\bar L$ be the numerical approximate period. Then the true periodic
orbit $O \colon [-1,1] \to \mathbb{R} ^2$ and its period $2L$ satisfy
\[
\| O - \bar O \|
\leq r _0, \qquad
| L - \bar L | \leq r _0
\]
where $r _0$ depends on $\mu$ and is given in
Table~\ref{tab.vdp-ars}. Moreover, let $\bar F, \bar B \colon [-1,1]
\to \mathbb{R} ^{2\times 2}$ be numerical approximations of the forward and
backward variational flows of $O$ represented by $200$ Chebyshev
coefficients (entrywise). Then the true solutions $F, B \colon [-1,1]
\to \mathbb{R} ^{2\times 2}$ satisfy
\[
\| F - \bar F \|
\leq r _1, \qquad
\| B - \bar B \
\leq r _2
\]
where $r _1$ and $r _2$ are given in Table~\ref{tab.vdp-ars}.
%
\begin{table}[ht]
\centering\tiny {\tt
\begin{tabular}{c||*3c}
$\mu$ & $ r _0$ & $ r _1$ & $ r _2$ \\ \hline \hline
0.1 & 5.573887884260317e-13 & 2.642512865973085e-12 & 2.996819243277690e-12 \\
0.2 & 8.709167228557321e-13 & 5.663812970926017e-12 & 1.015950775601682e-11 \\
0.3 & 1.189397347751034e-12 & 1.077390609220761e-11 & 3.082429090530088e-11 \\
0.4 & 1.557869751645740e-12 & 1.959454762049424e-11 & 9.753694436554251e-11 \\
0.5 & 1.928926778070392e-12 & 3.375851743576695e-11 & 3.280028209406751e-10 \\
0.6 & 2.366661800432010e-12 & 5.761835216827375e-11 & 1.155963389952129e-09 \\
0.7 & 2.865419317705822e-12 & 9.665742902893244e-11 & 4.426491597030142e-09 \\
0.8 & 3.544996952415318e-12 & 1.634487711041929e-10 & 1.894089923170409e-08 \\
0.9 & 4.539421525763888e-12 & 2.781116424781816e-10 & 9.518421901009437e-08 \\
1.0 & 5.835732296028395e-12 & 4.555755576317590e-10 & 5.724940629447956e-07
\end{tabular}
}
\caption{Radii for the computer-assisted proofs for the periodic
orbit and its period ($r_0$), the forward ($r_1$), and the
backward ($r_2$) variational flows for different parameter values
$\mu$ of \eqref{vdp}.}\label{tab.vdp-ars}
\end{table}
\end{theorem}
\begin{remark}
The norm of the difference between $O$ and $\bar O$ in above theorem
should be interpreted as the maximum of the componentwise $\ell
^1_{\nu }$ norm for $\nu=1.01$, where each component of $\bar O$
($200$ Chebyshev coefficients) is viewed as an element in $\ell
^1_{\nu }$ with zero tail. Similar for $ \| F - \bar F \|$ and $\| B
- \bar B \|$.
\end{remark}
\begin{remark}
The number $200$ of Chebyshev coefficients in
Theorem~\ref{thm.vdp-ars} are determined by plotting the coefficients
and truncating before the stagnation in the tail (due to the
precision of the arithmetic as $2^{-52}$ in double-precision.)
\end{remark}
Now we are ready to provide our first persistence result.
\begin{theorem} \label{thm.vdp-sdde0}
Consider a perturbation of the form \eqref{sdde} to equation
\eqref{vdp}, assume that the functions $P$ and $r$ satisfy $\|P\|$,
$\|DP\|$, $\|r\|$, and $\|Dr\|\le 1$, then for values of the
parameter $\mu$ and constants $a$, $ \beta _0$, $ \beta _1$, and $
\beta _2$ in the Table~\ref{tab.abep} below, if $ \varepsilon\le\varepsilon _0$ in
the Table~\ref{tab.abep}, the inequalities in \eqref{ineqs} are
satisfied. Hence, there exists a $C^{1+Lip}$ periodic orbit for the
perturbed system \eqref{sdde}. The differences of the unperturbed and
perturbed periodic orbits and their frequencies lie in the $C^0$
closure of the space $I _a \times \mathcal{E}_\beta$ as in
\eqref{space}.
\begin{table}[ht]
\centering\tiny {\tt
\begin{tabular}{c||*5c}
$\mu$ & $a$ & $ \beta _0$ & $ \beta _1$ & $ \beta _2$ & $ \varepsilon _0$ \\ \hline \hline
0.1 & 3.9289483314e-03 & 1.3814767259e-02 & 1.9351897706e-02 & 3.3489862240 & 2.0479538526e-03 \\
0.2 & 3.9458701226e-03 & 1.2482231722e-02 & 2.4285482147e-02 & 3.3566339504 & 1.8547822608e-03 \\
0.3 & 2.9095469512e-03 & 9.9870189178e-03 & 2.8450975266e-02 & 3.3611368094 & 1.1932985497e-03 \\
0.4 & 1.8632478932e-03 & 7.4544551083e-03 & 4.5960969418e-02 & 3.3707741174 & 6.4266953301e-04 \\
0.5 & 1.0872644592e-03 & 5.2329709138e-03 & 1.0817095356e-01 & 3.4113619375 & 3.1004411778e-04 \\
0.6 & 5.9335608942e-04 & 3.5047699947e-03 & 2.8241308709e-01 & 3.5732825217 & 1.3957190885e-04 \\
0.7 & 3.0480495862e-04 & 2.2543350887e-03 & 4.5110477848e-01 & 3.8463306657 & 5.8261321498e-05 \\
0.8 & 4.7155981698e-05 & 7.9828837072e-04 & 4.6510209568e-01 & 3.9780706528 & 6.9961093824e-06 \\
0.9 & 3.2308284738e-06 & 1.8561771388e-04 & 4.0830621482e-01 & 3.9828277487 & 2.3650493428e-07 \\
1.0 & 2.0172795949e-07 & 4.0905441781e-05 & 3.5946080062e-01 & 3.9846796546 & 9.7963190835e-09
\end{tabular}
}
\caption{Values of $a$ and $\beta$'s, and admissible $\varepsilon_0$ so
that the periodic orbit persists under perturbation.}
\label{tab.abep}
\end{table}
\end{theorem}
\begin{remark}
Theorem~\ref{thm.vdp-sdde0} and bootstrapping techniques lead to
infinite regularity provided that the unperturbed system and the
functions $P$ and $r$ are all smooth.
\end{remark}
The other parameters admits several options of optimization depending
on the aims. Notice that the norm $\| P \|$ can be normalized to $1$
since $P$ is multiplied by the perturbative parameter
$\varepsilon$. Our second perturbative result,
Theorem~\ref{thm.vdp-sdde}, optimizes the threshold $\varepsilon _0$
and also the class of perturbation $\| DP \|$ and of the delay $\| r
\|$, $\| Dr \|$. Then for all $\varepsilon \leq \varepsilon _0$, $\| r
\| \leq c _r$, $\|D r\| \leq c _{Dr}$, and $\| DP \| \leq c _{DP}$
with $\varepsilon_0$, $c _r$, $c _{Dr}$, and $c_{DP}$ in
Table~\ref{tab.vdp-values}, the inequalities in \eqref{ineqs} are
satisfied, and so there is a $C^{1+Lip}$ periodic orbit of the
perturbed system.
\begin{theorem} \label{thm.vdp-sdde}
Let \eqref{vdp} be the unpertubed ODE and consider a perturbation of
the form \eqref{sdde}. Given $a$, $\beta _0$, $\beta _1$, and $\beta
_2$ in Table~\ref{tab.vdp-abetas}. Then for all $\varepsilon \leq
\varepsilon _0$, $\| r \| \leq c _r$, $\|D r\| \leq c _{Dr}$, and $\|
DP \| \leq c _{DP}$ with $\varepsilon _0$, $c _r$, $c _{Dr}$, and $c
_{DP}$ in Table~\ref{tab.vdp-values}, the inequalities in
\eqref{ineqs} are satisfied, and so there exits a $C^{1+Lip}$
periodic orbit for the perturbed system.
\begin{table}[ht]
\centering\tiny{ \tt
\begin{tabular}{c||*4c}
$\mu$ & $ a$ & $ \beta _0$ & $ \beta _1$ & $ \beta _2$ \\ \hline \hline
0.1 & 5.3516317207e-03 & 1.3512619469e-02 & 1.8956002692e+00 & 3.7809698565 \\
0.2 & 4.4313887934e-03 & 1.2622565762e-02 & 1.4920065531e+00 & 3.8602413976 \\
0.3 & 2.8488130419e-03 & 9.6105596050e-03 & 1.1632362043e+00 & 3.9047783796 \\
0.4 & 1.6254775156e-03 & 6.6908933879e-03 & 9.2220210133e-01 & 3.9284935852 \\
0.5 & 8.6242391376e-04 & 4.4107683210e-03 & 7.5126159361e-01 & 3.9475219696 \\
0.6 & 4.1234067787e-04 & 2.6447824466e-03 & 6.3209724862e-01 & 3.9667717362 \\
0.7 & 1.9898994536e-04 & 1.6282965579e-03 & 5.3591862272e-01 & 3.9728038918 \\
0.8 & 2.8152415098e-05 & 5.8068726773e-04 & 4.6148172152e-01 & 3.9758037076 \\
0.9 & 2.3423038382e-06 & 1.5112884117e-04 & 4.0304855658e-01 & 3.9667079412 \\
1.0 & 1.5346224834e-07 & 3.3982751869e-05 & 3.6410084636e-01 & 4.0500566273
\end{tabular}
}
\caption{Values of the space $I _a \times \mathcal{E}_\beta$ in
\eqref{space} for which the perturbed periodic orbit
exists.}\label{tab.vdp-abetas}
\end{table}
\begin{table}[ht]
\centering\tiny{ \tt
\begin{tabular}{c||*4c}
$\mu$ & $ \varepsilon _0$ & $ c _r $ & $ c _{Dr}$ & $c_{DP}$ \\ \hline \hline
0.1 & 5.2514914300e-07 & 1.0525530335e+02 & 9.7470170038e+01 & 1.1923883191e+01 \\
0.2 & 4.8780187758e-07 & 1.0550841730e+02 & 9.7829645680e+01 & 1.2607166526e+01 \\
0.3 & 3.5942433417e-07 & 1.0677903035e+02 & 9.7626077504e+01 & 1.3098287211e+01 \\
0.4 & 2.0629988060e-07 & 1.1537382755e+02 & 9.7823817937e+01 & 1.4572329081e+01 \\
0.5 & 9.4989481038e-08 & 1.3431486893e+02 & 7.9752272527e+01 & 2.1261206986e+01 \\
0.6 & 1.0511326205e-07 & 1.5922648281e+02 & 5.4051960357e+01 & 2.1436738121e+01 \\
0.7 & 9.9701256277e-08 & 1.4962219117e+02 & 5.4940586398e+01 & 1.3762820031e+01 \\
0.8 & 3.4717672547e-08 & 2.8352283832e+02 & 1.2981903915e+02 & 5.0876339676e+00 \\
0.9 & 9.0478966022e-09 & 3.9078461001e+02 & 1.1719345674e+02 & 4.8009874216e+00 \\
1.0 & 1.8649496078e-09 & 3.6114164131e+02 & 1.2261159547e+02 & 4.7222224864e+00
\end{tabular}
}
\caption{Optimized perturbative parameters $\varepsilon _0$, $c _r$,
$c _{Dr}$, and $c _{DP}$.} \label{tab.vdp-values}
\end{table}
\end{theorem}
\begin{remark}
The quantities in Table~\ref{tab.vdp-values} have been obtained by an
optimization process of the inequalities \eqref{ineqs} with the
objective function $\varepsilon ^2 \|r \| \|Dr\| \| DP \|^2$ for each of the
$\mu$ values.
\end{remark}
\subsection{The Zero Finding-problems for the Unperturbed System}
In order to calculate the required coefficients for this example
rigorously, we follow the radii polynomial approach, see
\cite{AJPJ,JPC} and a summary in Appendix \ref{app.radii}, and we use
the \textsc{intlab} package in Matlab, \cite{Ru99a}. We solve a
boundary value problem to get the periodic orbit, and initial value
problems to get the solutions of the forward and backward variational
equations. In the following subsections, each of these problems will
be formulated as a zero-finding problem.
In Algorithm~\ref{alg.delay-cap}, we need to solve for the periodic
orbit $O(s) = (a _1(s), a _2(s))$ and the scaling parameter $L$ that
verifies
\begin{equation}
\label{vdpper}
\frac{d}{d s}
\begin{pmatrix}
a _1 \\ a _2
\end{pmatrix} = L
\begin{pmatrix}
a _2 \\ \mu (1-a_1^2)a_2-a_1
\end{pmatrix} , \quad
O(-1)=O(1) ,\quad\text{and} \quad a _1(-1) = 0,
\end{equation}
the forward system $F(s) =
\left(
\begin{smallmatrix}
v _{11} & v _{12} \\ v _{21} & v _{22}
\end{smallmatrix}
\right)(s)$ that verifies
\begin{equation}
\label{vdpfwd}
\frac{d}{ds}
\begin{pmatrix}
v _{11} & v _{12} \\ v _{21} & v _{22}
\end{pmatrix} =
L\begin{pmatrix}
0 & 1 \\ \vartheta & \varrho
\end{pmatrix}
\begin{pmatrix}
v _{11} & v _{12} \\ v _{21} & v _{22}
\end{pmatrix} \quad \text{and} \quad F(-1) = Id_2,
\end{equation}
with $\vartheta(s) = -2\mu a _1 a _2 - 1 $ and $\varrho(s) = \mu(1 -
a_1^2)$, and the backward system $B(s) = \left(
\begin{smallmatrix}
u _{11} & u _{12} \\ u _{21} & u _{22}
\end{smallmatrix}
\right)(s)$ that verifies
\begin{equation*}
\frac{d}{ds}
\begin{pmatrix}
u _{11} & u _{12} \\ u _{21} & u _{22}
\end{pmatrix} =
-L\begin{pmatrix}
u _{11} & u _{12} \\ u _{21} & u _{22}
\end{pmatrix}
\begin{pmatrix}
0 & 1 \\ \vartheta & \varrho
\end{pmatrix} \quad \text{and} \quad B(-1) = Id_2.
\end{equation*}
We utilize Chebyshev discretization to represent the periodic orbit
and its variational (forward and backward)
flows. Figure~\ref{fig.flows} shows the (numerical) solutions for some
values of the parameter $\mu$. Note that for $\mu \geq 1$ the backward
flow grows rapidly since the periodic orbit attracts strongly as we
can see in its (numerical) real non-trivial eigenvalue $\lambda _\mu$
of the monodromy matrix
\begin{equation*}
\begin{split}
\lambda _{0.5} &= \mathtt{3.917692025927352 \times 10^{-2}}, \\
\lambda _{1} &= \mathtt{8.596950636046152\times 10^{-4}}, \\
\lambda _{1.5} &= \mathtt{6.466756568013210\times 10^{-6}}.
\end{split}
\end{equation*}
\begin{figure}[ht]
\centering
\includegraphics[scale=.41]{flows_3.png}
\caption{On the left $\mu=0.5$, on the middle $\mu =1$, and on the
left $\mu = 1.5$ with $\mu$ parameter in \eqref{vdp}. From top to
bottom the orbit of the periodic orbit passing close to the point
$(2,0)$, its forward and backward variational flows.}
\label{fig.flows}
\end{figure}
\begin{remark}
Note that in \eqref{vdpper} we considered the spatial section $O
_1(-1) = 0$. Thanks to hypothesis \eqref{hypothesisH}, we can solve
the problem successfully.
\end{remark}
\subsubsection{Zero-finding problem for the periodic orbit}
The rigorous proof concerning $L >0$ and the periodic orbit $O \colon
[-1,1] \to \mathbb{R}^2$, with sequence of Chebyshev coefficients
$(a_1, a_2)$, is done in the Banach space $X _\nu \,\stackrel{\mbox{\tiny\textnormal{\raisebox{0ex}[0ex][0ex]{def}}}}{=}\, (\ell
^1_\nu)^2 \times \mathbb{R}$. The space is endowed with the product norm
\begin{equation*}
\| (a_1,a_2,L) \| \,\stackrel{\mbox{\tiny\textnormal{\raisebox{0ex}[0ex][0ex]{def}}}}{=}\, \max \{ \| a_1 \| _{\ell ^1_\nu}, \|
a_2 \| _{\ell ^1_\nu}, |L | \},
\end{equation*}
for $(a_1,a_2,L) \in X _\nu$.
The proof is obtained by solving a boundary value problem, see
Appendix~\ref{app.bvp}, with boundary conditions $a _1(-1) = 0$ (the
phase condition) and $a_i(-1)=a_i(1)$, for $i=1,2$. To define the
equivalent zero-finding map $\mathcal O \colon X _\nu \to X _{\nu'}$,
with $1 < \nu ' \leq \nu $, of \eqref{vdpper}, we first consider
linear operators $M, T \colon \ell ^1_\nu \to \ell ^1_\nu$, and
$\Lambda \colon \ell ^1_\nu \to \ell ^1_{\nu'}$ defined as
\begin{align}
\nonumber
M &\,\stackrel{\mbox{\tiny\textnormal{\raisebox{0ex}[0ex][0ex]{def}}}}{=}\, \begin{pmatrix}
0 & 1 & 0 & 1 & 0& 1 & 0 & \cdots \\
0 & 0 & 0 & 0 & 0 & 0 & \cdots \\
& \ddots & \ddots & \ddots & \ddots & \ddots \\
& \cdots & 0 & 0 & 0 & 0 & 0 & \cdots \\
& & \cdots & \ddots & \ddots & \ddots & \ddots
\end{pmatrix}, \\
\label{eq.T}
T &\,\stackrel{\mbox{\tiny\textnormal{\raisebox{0ex}[0ex][0ex]{def}}}}{=}\, \begin{pmatrix}
0& 0 & 0 & 0 & 0 & 0 & \cdots \\
-1 & 0 & 1 & 0 & 0 & \cdots \\
0 & -1 & 0 & 1 &0 & 0 & \cdots \\
& \ddots & \ddots & \ddots & \ddots & \ddots \\
& \cdots & 0 & -1 & 0 & 1 & \cdots \\
& & \cdots & \ddots & \ddots & \ddots & \ddots
\end{pmatrix},
\intertext{and }
\label{eq.Lambda}
\Lambda &\,\stackrel{\mbox{\tiny\textnormal{\raisebox{0ex}[0ex][0ex]{def}}}}{=}\, \begin{pmatrix}
0 & 0 & 0 & 0 & 0 & 0 & \cdots \\
0 & 2 & 0 & 0 & 0 & \cdots \\
0 & 0 & 4 & 0 & 0 & 0 & \cdots \\
& \ddots & \ddots & \ddots & \ddots & \ddots & \cdots \\
& \cdots & 0 & 0 & 2 k & 0 & \cdots \\
& & \cdots & \ddots & \ddots & \ddots
\end{pmatrix}.
\end{align}
Then the zero-finding map for the periodic orbit is given by $\mathcal
O=(\mathcal O _1, \mathcal O _2, \mathcal O _3)$, where
\begin{equation*}
\begin{aligned}
\mathcal O _i(a_1,a_2,L) & \,\stackrel{\mbox{\tiny\textnormal{\raisebox{0ex}[0ex][0ex]{def}}}}{=}\, (M+\Lambda) a _i +L\cdot T
f_i(a), & i &\in \{1,2\} ,\\ \mathcal O _3(a_1,a_2,L) & \,\stackrel{\mbox{\tiny\textnormal{\raisebox{0ex}[0ex][0ex]{def}}}}{=}\,
(a_1) _0 + 2\sum _{k \geq 1} (a _1)_k (-1)^k,
\end{aligned}
\end{equation*}
with
\begin{equation*}
\begin{split}
f_1(a) &\,\stackrel{\mbox{\tiny\textnormal{\raisebox{0ex}[0ex][0ex]{def}}}}{=}\, a _2, \\
f_2(a) &\,\stackrel{\mbox{\tiny\textnormal{\raisebox{0ex}[0ex][0ex]{def}}}}{=}\, \mu a _2 - \mu (a _1 * a_1 * a _2) - a _1,
\end{split}
\end{equation*}
with $\ast$ being the convolution in $\ell ^1_\nu$, see
Appendix~\ref{app.radii}.
Following the radii polynomial approach, we first compute a numerical
solution $(\bar a_1, \bar a_2,\bar L) \in X_\nu$ such that $\mathcal
O(\bar a_1,\bar a_2,\bar L)\approx0$, that is it is zero up to a given
tolerance, for instance $10^{-12}$.
Then computing the radii polynomial \eqref{eq:radii_polynomial} as in
Lemma \ref{lem:radii}, there is an exact solution $(a_1,a_2,L) \in
X_\nu$ within distance $r _0 > 0$ of the numerical solution such that
$\mathcal O(a_1,a_2,L)=0$.
\subsubsection{Zero-finding problem for the solution of the forward variational equation}
\label{sec:forward}
To compute the solution of \eqref{vdpfwd}, we first assume that one
computed rigorously the values of $L$ and the periodic orbit $ (a_1,
a_2)$ verifying \eqref{vdpper}. We split these information as the sum
of the numerical parts and the tail parts, that is
\begin{equation*}
a_i=\bar a_i+\tilde a_i, \quad \text{for }i \in \{1,2\}, \qquad
L=\bar L+\tilde L.
\end{equation*}
From the proof, we know that $\|\tilde a_i\| _{\ell ^1_\nu}\le r_0$
and $|\tilde L| \le r_0$.
The $\vartheta$ and $\varrho$ in \eqref{vdpfwd} depend on the periodic
orbit, and are represented as Chebyshev series. We split them by their
numerical and tail parts
\[
\vartheta \,\stackrel{\mbox{\tiny\textnormal{\raisebox{0ex}[0ex][0ex]{def}}}}{=}\, \bar \vartheta + \tilde \vartheta, \qquad
\varrho \,\stackrel{\mbox{\tiny\textnormal{\raisebox{0ex}[0ex][0ex]{def}}}}{=}\, \bar \varrho + \tilde \varrho.
\]
Explicitly, the numerical parts are given by
\begin{equation*}
\bar \vartheta _k \,\stackrel{\mbox{\tiny\textnormal{\raisebox{0ex}[0ex][0ex]{def}}}}{=}\, -2 \mu (\bar a _1 \ast \bar a _2)_k -
\delta _{0,k}, \qquad \bar \varrho _k \,\stackrel{\mbox{\tiny\textnormal{\raisebox{0ex}[0ex][0ex]{def}}}}{=}\, \mu \delta _{0,k} -
\mu (\bar a _1 \ast \bar a _1)_k,
\end{equation*}
and the tails contain the crossing terms from the convolutions.
Let $S \colon \ell^1_\nu \to \ell^1_\nu$ be the linear operator
defined by
\begin{equation*}
S \,\stackrel{\mbox{\tiny\textnormal{\raisebox{0ex}[0ex][0ex]{def}}}}{=}\,
\begin{pmatrix}
-1 & 2 & -2 & 2 & -2 & \cdots \\
0 & 0 & 0 & 0 & 0 & \cdots \\
\vdots & \vdots & \vdots & \vdots & \vdots & \ddots \\
0 & 0 & 0 & 0 & 0 & \cdots \\
\end{pmatrix}
\end{equation*}
and let $\ell _\alpha \colon \ell ^1_\nu \to \ell ^1_\nu$ be the
linear operator of convolution multiplication defined, given $\alpha
\in \ell ^1_\nu$, by $\ell _\alpha \colon b \mapsto \alpha \ast b$.
The zero-finding problem equivalent to \eqref{vdpfwd} consists in two
proofs of initial value problems, see Appendix~\ref{app.ivp}. Each of
them corresponds to a column of $F(s)$, and they are given by the map
$\mathcal F _p \colon (\ell ^1_\nu) ^2 \to (\ell ^1_{\nu'})^2$, with
$1 \leq \nu ' < \nu$ defined as
\begin{equation}
\label{eq.zeroopfwd}
\begin{split}
\mathcal F _p (v) &\,\stackrel{\mbox{\tiny\textnormal{\raisebox{0ex}[0ex][0ex]{def}}}}{=}\, \bar {\mathcal F} _p (v) + \tilde
{\mathcal F} _p(v), \\ \bar {\mathcal F} _p(v)
&\,\stackrel{\mbox{\tiny\textnormal{\raisebox{0ex}[0ex][0ex]{def}}}}{=}\, p +
\begin{pmatrix}
S + \Lambda & 0 \\
0 & S + \Lambda
\end{pmatrix} v + \bar L \begin{pmatrix}
0 & T \\ T \ell _{\bar \vartheta} & T \ell _{\bar \varrho}
\end{pmatrix} v, \\
\tilde {\mathcal F} _p(v) &\,\stackrel{\mbox{\tiny\textnormal{\raisebox{0ex}[0ex][0ex]{def}}}}{=}\, \tilde L \begin{pmatrix}
0 & T \\ T \ell _{\bar \vartheta} & T \ell _{\bar \varrho}
\end{pmatrix} v + (\bar L + \tilde L) \begin{pmatrix}
0 & 0 \\ T\ell _{\tilde \vartheta} & T\ell _{\tilde \varrho}
\end{pmatrix} v,
\end{split}
\end{equation}
where the operators $T$ and $\Lambda$ are defined in \eqref{eq.T} and
\eqref{eq.Lambda} respectively.
The element $p\in (\ell^1_\nu)^2$ in \eqref{eq.zeroopfwd} is related
to the initial condition of the chosen column, when $v$ corresponds to
the first column of $F(s)$, then $p$ is the vector of Chebyshev
coefficients of the constant vector
$\left(\begin{smallmatrix}1\\0\end{smallmatrix}\right)$; otherwise,
$p$ is the vector of Chebyshev coefficients of the constant vector
$\left(\begin{smallmatrix}0\\1\end{smallmatrix}\right)$.
Now computing the elements in Lemma~\ref{lem:radii} for the initial
conditions, we prove the existence of the solution $F(s) = \bar F(s) +
\tilde F(s)$ with $\| \tilde F \| = \max _{i,j \in \{1,2\}} \| \tilde
v _{ij}\| _{\ell ^1_\nu} \leq r _1$.
\subsubsection{Zero-finding problem for the solution of the backward variational equation}
Similarly to the forward flow in \eqref{eq.zeroopfwd}, we consider the
map $\mathcal B _p \colon (\ell ^1_\nu) ^2 \to (\ell ^1_{\nu'})^2$
defined by
\begin{equation}
\label{eq.zeroopbwd}
\begin{split}
\mathcal B _p (u) &\,\stackrel{\mbox{\tiny\textnormal{\raisebox{0ex}[0ex][0ex]{def}}}}{=}\, \bar {\mathcal B} _p (u) + \tilde {\mathcal B} _p(u), \\
\bar {\mathcal B} _p(u) &\,\stackrel{\mbox{\tiny\textnormal{\raisebox{0ex}[0ex][0ex]{def}}}}{=}\, p +
\begin{pmatrix}
S + \Lambda & 0 \\
0 & S + \Lambda
\end{pmatrix} u - \bar L \begin{pmatrix}
0 & T \ell _{\bar \vartheta} \\ T & T \ell _{\bar \varrho}
\end{pmatrix} u, \\
\tilde {\mathcal B} _p(u) &\,\stackrel{\mbox{\tiny\textnormal{\raisebox{0ex}[0ex][0ex]{def}}}}{=}\, - \tilde L \begin{pmatrix}
0 & T \ell _{\bar \vartheta} \\ T & T \ell _{\bar \varrho}
\end{pmatrix} u - (\bar L + \tilde L) \begin{pmatrix}
0 & T \ell _{\tilde \vartheta} \\ 0 & T\ell _{\tilde \varrho}
\end{pmatrix} u.
\end{split}
\end{equation}
Now when $p\in (\ell^1_\nu)^2$ in \eqref{eq.zeroopbwd} corresponds to
$\left(\begin{smallmatrix}1\\0\end{smallmatrix}\right)$, we obtain the
first row of $B(s)$; otherwise, when $p$ corresponds to
$\left(\begin{smallmatrix}0\\1\end{smallmatrix}\right)$, we get the
second row.
As in the forward flow case, computing the elements in
Lemma~\ref{lem:radii}, we end up proving the existence of the solution
$B(s) = \bar B(s) + \tilde B(s)$ with $\| \tilde B \| = \max _{i,j \in
\{1,2\}} \| \tilde u _{ij}\| _{\ell ^1_\nu} \leq r _2$.
\subsection{Details of the Computer-Assisted Proofs for the Unperturbed Systems}
We present the results of three proofs. The first one for the periodic orbit and
the period is a nonlinear problem with a cubic term in the case of the
Van der Pol \eqref{vdp}. The two other proofs, \eqref{eq.zeroopfwd}
and \eqref{eq.zeroopbwd}, are linear and they depend on the results
from the first proof.
We used the radii polynomial approach, see Appendix~\ref{app.radii},
for all the three computer-assisted proofs and a common $\nu=1.01$ for
the $\ell ^1_\nu$ space. When we encounter products, like in
$\vartheta$ and $\varrho$, which are convolutions in Chebyshev spaces,
we keep track of all the terms of the numerical and tail parts, see
Section \ref{sec:conv}. In particular, we have to manage quintic
convolution to get the norm of $D^2 K_0$ in the polynomials.
Figure~\ref{fig.ars} shows the final radii values and the
computational times of each proof, in particular, we observed that the
backward variational flow is computationally harder for larger $\mu$
because the periodic orbit becomes more attractive, see
Figure~\ref{fig.flows}.
\begin{figure}[ht]
\centering
\includegraphics[scale=.55]{ars.png}
\caption{Radii of the computer-assisted proofs of the orbit, the
forward, and the backward of \eqref{vdp} for different values of
the parameter $\mu$.}
\label{fig.ars}
\end{figure}
\subsection{Bounds Computation and Optimization Steps}
Following Section~\ref{sec.compu-bounds}, we compute the bounds
required in Algorithm~\ref{alg.delay-cap}. We use different radii of
the computer-assisted proofs, that is, $r _0$ for the period and the
periodic orbit, $r _1$ for the forward flow, and $r _2$ for the
backward flow. Their values depend on the parameter $\mu$ in
\eqref{vdp}. For the triangle meshes we use the size of
$5000$. Overall the computation required around 3 days for each
parameter $\mu$ of the ODE \eqref{vdp}.
We have done two optimization processes. One in
Theorem~\ref{thm.vdp-sdde0} and another one in
Theorem~\ref{thm.vdp-sdde}. In both cases, the variables $a$ and
$\beta = (\beta _0, \beta _1,\beta _2)$ for the space $I _a \times
\mathcal{E} _ \beta$ in \eqref{space} have been restricted to the
domains
\begin{equation*}
\begin{split}
a, \beta _0 &\in (0, 0.1], \\
\beta _1, \beta _2 &\in (0, 5], \\
\varepsilon &\in (0, +\infty),
\end{split}
\end{equation*}
and initial guesses $a = \beta _0 = 10^{-2}$, $\beta _1 = \beta _2 =
0.5$, and $\varepsilon = 10^{-2}$.
In Theorem~\ref{thm.vdp-sdde0} the objective function was just $\varepsilon$
with $\|P\|$,
$\|DP\|$, $\|r\|$, and $\|Dr\|$ equal to $1$. In
Figure~\ref{fig.optim_epsi} we used this procedure to illustrate, for
different values of the parameter $\mu$ in \eqref{vdp}, how sensitive the
numerical threshold $\varepsilon _0$ is when either $\| r \|$, $\| D r \|$, or
$\| D P \|$ ranges in the $x$-axes of the plot and the other two
variables are set to $1$. Thus, from Figure~\ref{fig.optim_epsi}, we
observe that $\| DP \|$ and $\| Dr \|$ have similar, strong effects on
the $\varepsilon_0$. On the other hand, $\|r\|$ presents less
influence in the $\varepsilon_0$. In any case, that sensitivity will
always depend on the model itself and the inputs of the
Algorithm~\ref{alg.delay-cap}.
\begin{figure}[ht]
\centering
\includegraphics[scale=.55]{optim_epsi.png}
\caption{Value of the
$\varepsilon_0$ in \eqref{ineqs}
for the Van der Pol equation \eqref{vdp}, moving either $\| r \|$,
$\| Dr \|$, or $\| DP \|$ and setting the other two quatities equal
to $1$.} \label{fig.optim_epsi}
\end{figure}
In Theorem~\ref{thm.vdp-sdde} the objective function is $\varepsilon ^2 \| r
\| \| Dr \| \| DP \| ^2$ with $\| r \|$, $\| D r \|$, and $\| D P \|$
in the domain $[1,+\infty)$. Figure~\ref{fig.capdelay} shows the
values $a$ and $\beta = (\beta _0,\beta _1,\beta _2 )$ for the space
$I _a \times \mathcal{E} _ \beta$ in \eqref{space} and the numerical
thresholds $\varepsilon _0$, $c _{r}$, $c _{Dr}$, and $c _{DP}$ in
Theorem~\ref{thm.vdp-sdde}. Heuristically, the objective function
was chosen considering that different operands have different
scales. For instance, $\varepsilon$ will be, in general, small and the rest
will be large. Thus, we use multiplication instead of addition. We
are putting $\varepsilon ^2$ and $\|DP \|^2$ because we want them to be
dominant during the optimization. The reason is that the size of
$\varepsilon$ corresponds to the size of the perturbation, and it is not
very interesting if the perturbation term $P$ is close to a constant
vector.
We note that in our result, it is possible that none of $\varepsilon$, $\|r
\|$, $\| Dr \|$, and $\| DP \|$ is maximized, since the optimization
processes are subject to various tolerances, the initial guess, etc.
Our goal is to provide some values of the parameters close to
optimized values so that the inequalities \eqref{ineqs} are satisfied.
Finally, after each of the numerical optimizations, we verify the
inequalities with the interval arithmetic polynomials we got before the
optimization to prove rigorously that they verify the inequalities
\eqref{ineqs}.
\begin{figure}[ht]
\centering
\includegraphics[scale=.55]{capdelay.png}
\caption{Values of the proof optimizing $\varepsilon$, $\| r \|$, $\|
Dr \|$, and $\| DP \|$ for different values of the parameter $\mu$
in \eqref{vdp} in Theorem~\ref{thm.vdp-sdde}.} \label{fig.capdelay}
\end{figure}
\begin{comment}
\subsection{Sensibility on the threshold}
The sensibility of the threshold $\varepsilon_0$ of the perturbative parameter
with respect to the inputs of Algorithm~\ref{alg.delay-cap} depends on
the model. In the Van der Pol case, Figure~\ref{fig.optim_epsi}
shows, for different $\mu$ values in \eqref{vdp}, the approximate
upper bounds of $\varepsilon$ so that \eqref{ineqs} is verified, when one of
$\| r \|$, $\| Dr \|$, and $\| DP \|$ varies with values in the
$x$-axis while the other two are fixed to $1$. Observe that $\| DP
\|$ and $\| Dr \|$ have similar, strong effects on the
$\varepsilon$-threshold. On the other hand, $\|r\|$ presents less influence
in the $\varepsilon$-threshold.
\end{comment}
\section{Conclusion} \label{sec:conclusion}
The theorems established here exploit well-known (but fairly state-of-the-art)
computer-assisted methods of proofs for ODEs,
however the use of these methods in the present work is fairly novel, as
the validated numerical computations are used to verify the hypotheses of
a very singular perturbation theorem. The statement of the theorem
hypothesizes the existence and other more quantitative properties
of an isolated periodic orbit in a nonlinear system of ODEs.
These hypotheses are notoriously difficult to verify in nonlinear ODEs,
especially if the ODEs are far from any perturbative or asymptotic regime.
The validated numerical methods allow us to pass from good numerical
computations, to mathematically rigorous statements about the desired
periodic solution. Moreover, the high order spectral methods used in the
present work provide enough control of the orbits that we can
obtain, a-posteriori bounds for all the constants appearing in the
hypotheses of the perturbation theorem. Indeed, all of this can be made
fairly automatic.
The result given here can be generalized easily to higher dimensional systems of
ODEs, to the case when there are multiple forward or backward delays, either
state-dependent, distributed, or of other types. Of course the polynomials
considered here will need some modifications in other cases,
but the constructions are not fundamentally different.
Using the same framework, but
longer expressions for the polynomials, one should be able to study the case where
small delays are present as in electrodynamics, see \cite{Per}.
Indeed, applying the arguments developed here to relativistic perturbations
of electrodynamics would be a fascinating future project.
Our work used the \textsc{intlab} package, \cite{Ru99a} but other
packages providing rigorous bounds can be applied as well, such as,
\textsc{arb} in \cite{Johansson2017arb} and the CAPD library
\cite{MR4283203}.
\section*{Acknowledgment}
J.G. was supported by the Italian grant MIUR-PRIN 20178CJA2B ``New Frontiers of Celestial
Mechanics: theory and Applications'', the Spanish grant PGC2018-100699-B-I00 (MCIU/AEI/FEDER, UE),
and the Catalan grant 2017 SGR 1374.
J.-P. L. was supported by an NSERC Discovery Grant.
J.Y was supported by the National Science Foundation under Grant No. DMS-1929284 while in residence
at the Institute for Computational and Experimental Research in Mathematics in Providence,
RI, Fall 2021 - Spring 2022.
J.D. Mireles-James was partially supported by the National Science Foundation
Grant DMS 1813501.
|
\section{A TDVP Langevin Equation}
\label{Sec:MPS Langevin Equation}
\begin{figure}[!t]
\includegraphics[width=0.5\textwidth]{Fig1.pdf}
\caption{
{\it Basic Properties of MPS Langevin Evolution:} The TDVP Langevin equation, Eq.(\ref{eq:TDVPLangevin}) describes the evolution of the density matrix through an ensemble of stochastic pure-state trajectories.
Here, the von Neumann entanglement as a function of time for a typical trajectory is shown calculated is shown calculated for different coupling strengths. Here the bond dimension is $D=128$ and we have kept temperature fixed, $T=0.2$, while we vary the noise, $\gamma T$, i.e. coupling $\gamma$. Such plots show three different regimes of behaviour. An initial transient, followed by an approximately linear in $t$ or logarithmic in $t$ growth and finally a saturation at the longest times. This saturation can either be determined by the variational approximation or intrinsically by the interplay of the Hamiltonian and the dissipative bath. It signals whether a lower bond-dimension (hence less computationally intensive) simulation suffices in underpinning the entropy dynamics.}
\label{fig:BasicProperties}
\end{figure}
Langevin equations describe the motion of a system coupled to an environment (or alternatively the motion of slow collective degrees of freedom in an effective bath described by the faster degrees of freedom\cite{zwanzig1960ensemble}) by adding noise and friction terms to the basic equations of motion of the system. If the environmental degrees of freedom are in thermal equilibrium, the friction and noise satisfy a fluctuation-dissipation relation. Applied to quantum systems, the Schr\"odinger equation provides the basic equations of motion. The ensemble of the resulting stochastic Schr\"odinger trajectories recovers the density matrix evolution and is said to be an unravelling of it.
{\it The Langevin equation} over matrix product states studied here, can be written in its Markovian limit as:
\begin{eqnarray}
\langle \partial_i \psi | \partial_j \psi \rangle \dot X_j
&=&
-i \langle \partial_i \psi | \hat H | \psi \rangle
-i \sum_n \langle \partial_i \psi | \hat F_n | \psi \rangle \eta(t)
\nonumber\\
& &
-i \sum_n \gamma \frac{\langle \psi | \hat F_n | \psi \rangle }{dt} \langle \partial_i \psi | \hat F_n | \psi \rangle .
\label{eq:TDVPLangevin}
\end{eqnarray}
The terms on the left-hand side and the first term on the right constitute the conventional time-dependent variational principle (TDVP) equations\cite{HaegemanTDVP}. The second and third terms on the right are, respectively, the noise and friction due to coupling to the environment. $\hat F_n$ are the operators by which the system is coupled to the bath displacement operators. We generally assume these to be spatially local. For spin-half chains they are given by the $x,y $ and $z-$components of the spin operators on each site, each of which couples to a separate bath. The bath is described as a collection of harmonic oscillators and the noise-correlator is determined by the spectrum of oscillators and the temperature of the bath; $\langle \langle \eta(t) \eta(t') \rangle \rangle=2 \gamma T \delta(t-t')$ in the combined classical and Markovian limits.
{\it The TDVP Langevin Equation} is a new approach to unravelling the density matrix evolution of an open system. It combines the study of quantum many-body dynamics on a variational manifold --- here we use matrix product states (MPS)\cite{Orus:2014zl,schollwock2011density,perez2006matrix} --- with the Langevin limit of Keldysh field theories\cite{kamenev2011field,sieberer2016keldysh}.
We combine these approaches by constructing a Keldysh path integral over MPS states\cite{green2016feynman}. To date, MPS techniques have been employed in the study of open systems largely in three ways: by starting with the Lindblad master equation and either describing the density matrix directly as a matrix product operator\cite{verstraete2004matrix,cui2015variational,weimer2019simulation} or else unravelling its evolution over MPS representations of quantum trajectories\cite{daley2009atomic,daley2014quantum,bonnes2014superoperators}; alternatively, tensor network techniques can be employed to describe the bath directly within the state ansatz\cite{strathearn2018efficient}. Of these, the latter is numerically costly and the Lindblad-based approaches formally apply in a different limit\cite{FergusThesis}.
Questions of applicability aside, there are some features of our approach that make it particularly attractive.
It naturally leads to thermal equilibrium, and it can treat the non-Markovian limit. There also exists a natural hydrodynamic limit and Fokker-Plank description. The scaling of the algorithm for each trajectory is the same as the usual TDVP for matrix product states -- moreover, there appears to be a degree of self averaging so that certain quantities are well-approximated by relatively fewer trajectories.
{\it A derivation} of Eq.(\ref{eq:TDVPLangevin}) is given in Appendix A. We develop the Langevin equation from the Keldysh path integral for the time-evolution of the density matrix. The method follows that of Ref.\cite{kamenev2011field} with the modification that the Keldysh path integral is constructed over matrix product states\cite{green2016feynman}. The result adds noise and friction to the time-dependent variational principle constructed over matrix product states\cite{HaegemanTDVP}. A similar construction for any variational class would lead to a similar stochastic equation of motion, which we dub the TDVP-Langevin equation. An alternative, heuristic derivation involves solving the Schr{\"o}dinger equation for the system and bath and substituting the solution of the latter into the equation of motion of the former. Alternative unravellings of the Lindblad equation for the density matrix evolution over MPS\cite{daley2009atomic} apply in different circumstances of relative time and energy scales of the bath and system.
We construct Eq.(\ref{eq:TDVPLangevin}) over matrix product states using conventional methods\cite{HaegemanTDVP}. Integration of this equation is complicated by the friction term. Naively, this requires inversion of a matrix that is proportional to the system size and dimension of the variational manifold. However, recognising that it consists of an outer product of vectors allows an efficient inversion and integration of the equations of motion. Details are given in Appendix B and our code is available at https://github.com/AndrewHallam/Langevin.
{\it Basic properties} of the TDVP Langevin equation are summarised in Fig.\ref{fig:BasicProperties}. A
thermal distribution over the variational manifold is given by a Boltzmann-weighted Haar average over the variational manifold. In the case the of MPS of bond dimension $D$, this average can be performed as a Haar integral over the group $SU(dD)$ (with $d$ the local Hilbert space dimension). Such thermal distributions are fixed points of the Langevin evolution (see Appendix \ref{app:HaarAverage}).
Fig.~\ref{fig:BasicProperties} captures the dynamics of the von Neumann entanglement entropy typical for trajectories over the parameters we consider. In the case shown, temperature $T=0.2$ is kept fixed and friction is increased. Beyond a critical value, $\gamma T\approx 0.1$, entanglement growth is suppressed. The entanglement undergoes a transition from being determined by the variational approximation to becoming intrinsic to the interplay between the Hamiltonian and the dissipative bath. Similar transitions occur when other parameters are kept constant (friction $\gamma$, or noise $\gamma T$). We detail these results in the following section.
\section{Open Evolution of a Rapidly Entangling System}
\label{Sec:Results}
In the absence of coupling to a bath, TDVP equations eventually fail as the entanglement grows beyond that which can be represented on the variational manifold\footnote{TDVP equations for the thermofield purification of the density matrix may escape this fate\cite{hallam2019lyapunov} at least as far as local observations are concerned}. However, just as observed in projective measurements of random circuits, the effects of the environment may restrict the growth of entanglement. {\it In extremis} this might limit entanglement of individual trajectories so that they can be represented on low dimensional variational manifolds. The TDVP Langevin equation will then give a good account of the dynamics at all times, signifying a transition in its classical representability. This is our interpretation of the sequence of results presented in this section.
{\it The Hamiltonian} that we consider is the tilted field Ising model
\begin{equation}
\hat H =
-\sum_i \left[ J \sigma_i^z \sigma_{i+1}^z +h \sigma^z_i+ g \sigma^x_i \right],
\label{eq:Hamiltonian}
\end{equation}
with $J=1$, $g=-1.05$ and $h=0.5$. With these parameters, the Hamiltonian is a far from any integrable point and rapidly thermalising\cite{banuls2011strong,leviatan2017quantum}.
{\it Infinite temperature and vanishing friction}: Fig.\ref{fig:InfiniteT} shows the variation in von Neumann entanglement across the central bond as a function of time for simulations with a range of bond orders and noise strengths. The broad result of these simulations is that the entanglement saturates at long times at a value determined by the bond order of the simulation. This is consistent with an infinite temperature final state with the maximum entanglement supported by the variational manifold. The most interesting aspect of these results is the decreasing rate of early-time entanglement growth with increasing noise strength. Crucially we do not find evidence of an intrinsic saturation of entanglement -- only that dictated by the limitations of the variational approximation.
\begin{figure}
\includegraphics[width=0.47\textwidth]{Fig2.pdf}\caption{
{\it Evolution of Entanglement at Infinite Temperature:} Here we consider the evolution under Eq.(\ref{eq:TDVPLangevin}) with Hamiltonian Eq.(\ref{eq:Hamiltonian}) with $J=1$, $g=-1.05$ and $h=0.5$. The friction coefficient $\gamma=0$ with noise $\gamma T$ finite and initial state $Z$.
a) von Neumann Entropy {\it versus} time at a fixed bond order $D=160$ for different values of $\gamma T$. Without noise, and after an initial transient, the entanglement grows linearly with time before saturating at a constant value less than the maximum determined by the MPS manifold. With non-zero noise, the entanglement growth is linear in time with a growth rate that reduces with increasing $\gamma T$ due to the dephasing effects of the bath. Ultimately, all curves saturate (not shown) to the same value characteristic of an infinite temperature state on the variational manifold.
b) The growth rate of entanglement $\dot{S}$, extracted beyond the initial transient ($t\approx4$), {\it versus} noise.}
\label{fig:InfiniteT}
\end{figure}
{\it Finite temperature and friction}:
Including both noise and friction, we do see such an intrinsic saturation. This is demonstrated in two ways; by considering the saturation of entanglement at long times and by a high fidelity between low- and high-bond order simulations at long times.
\\
\noindent
i. {\it Saturating entanglement}
In order to demonstrate this, we first show in Fig.~\ref{fig:SaturatingS} the long-time average of the von-Neumann entropy. A graph showing the typical time-dependence from which such saturation values are computed is shown in Fig.\ref{fig:BasicProperties}b). For low noise and friction, the saturation is determined by the limitations of the variational manifold. Panels a), b) and c) show that as a function of $\gamma T$ at fixed $\gamma$, then $T$, and $\gamma T$ at fixed $\gamma$, respectively. A threshold is reached for each bond order where it adequately captures the saturation entanglement, thus indicating a transition to increasingly classically simulatable dynamics. The transition can be seen from the point where the trajectories obtained at different bond orders give the same saturation entanglement. From this we can extract a critical $\gamma$ or $\gamma T$ as a function of bond order that we show in each corresponding inset figure.
\\
\noindent
ii. {\it High fidelity as } $t \rightarrow \infty$: We can identify an analogous transition in the fidelity of each trajectory at different bond orders versus a reference trajectory with bond order $D=128$. In this case, we find that beyond a critical combination of $\gamma$ or $\gamma T$, the fidelity of the state at low bond dimension remains close to $1$ for long times. We expand upon this result in Fig.~\ref{fig:DivergentClassicalSimulation}, where we identify a divergent classical simulation time. We note that the fidelity is more sensitive to the time-step as friction is increased -- an issue typical of numerical integration of systems of stochastic differential equations. This makes accessing the critical point of the transition numerically intensive for the parameters and Hamiltonian we consider. The entropy is less sensitive to this.
\begin{figure*}
\includegraphics[trim={6 0 0 0},clip, width=0.32\textwidth]{Fig3a.pdf}
\includegraphics[trim={5 4 5 0},clip, width=0.32\textwidth]{Fig3b.pdf}
\includegraphics[trim={0 0 6 0},clip, width=0.32\textwidth]{Fig3c.pdf}
\caption{
{\it Evolution of Saturation Entanglement at Finite Temperature:} Here we consider the evolution under Eq.(\ref{eq:TDVPLangevin}) with Hamiltonian Eq.(\ref{eq:Hamiltonian}) where $J=1$, $g=-1.05$ and $h=0.5$ and finite $\gamma$ and $T$. In the main figures we show the dependence of the von Neumann entropy as a function of noise and friction: a) {\it versus} $\gamma T$ at fixed $\gamma$, b) {\it versus} $\gamma T$ at fixed $T$, c) {\it versus} $\gamma$ at fixed $\gamma T$.
In each case, at low values of noise and friction, the saturation entanglement $\bar{S}$, is determined by the choice of variational parametrization through the bond order. As the noise and friction are increased, there is a cross-over where the saturation entanglement decreases from this value. Each bond order captures the saturation entanglement for a sufficiently large noise and friction. This is indicated when the entanglement begins to follow the entanglement given by the highest bond order simulation. Where the saturation entanglement of the individual trajectories (distinguished by bond order) dramatically fall to a single curve, so that the lower bond dimension curves coincide with higher bond dimensions, illustrates this transition to situations where the saturation entanglement is intrinsically determined. {\it In each corresponding inset figure}, we have extracted critical dissipation strengths where these transitions occur as a function of bond order.
}
\label{fig:SaturatingS}
\end{figure*}
\begin{figure*}
\includegraphics[trim={0 0.2cm 0 0},clip,width=17.5cm]{Fig4a.pdf}
\includegraphics[trim={0 0 0 0.2cm},clip,width=17.5cm]{Fig4b.pdf}
\newline
\includegraphics[width=17.5cm]{Fig4c.pdf}
\newline
\newline
\includegraphics[width=17.5cm]{Fig4d.pdf}
\caption{
{\it Divergent Classical Simulation Time:} Simulations carried out at a bond order $D$ give a good account of the system evolution up to a time $t^*(D)$. We extract these values versus a reference $D=128$ simulation, which serves as the good account of the system. We do this in two ways, by comparing the difference in von Neumann entanglement entropy between these states and the fidelity with this state. $t^*(D)$ is the time when the simulation with varying bond dimension deviates appreciably from the reference trajectory. The row with panels a), b), c), shows $t^*(D)$ extracted from the entropy, while panels d), e), f) demonstrate this for the fidelity. The fixed variables are split across the columns -- a), d), shows varying $\gamma T$ at fixed $\gamma=0.2$, b), e) $\gamma T$ at fixed $T=0.2$, and c), f) $\gamma$ at fixed $\gamma T=0.25$. Panels g) and h) show typical evolution of entanglement and fidelity with time. This is data from the near-critical point for $T=0.2$, which is for the trajectory with noise $\gamma T = 0.1$. We say that a simulation has failed to provide a good account of the system when the trajectory deviates beyond $\epsilon =0.05$, and the time at which this occurs is $t^*(D)$. In g), this is the point where $\Delta S/S_{D=128} = |S_{D=128}-S_{D}|/S_{D=128} > \epsilon$. Analogously in h), $t^*(D)$ is the time when the fidelity is appreciably different to $1$, i.e. $|\langle\psi_{D=128}(t)|\psi_{D}(t)\rangle|< 1-\epsilon$.
A divergent $t^*(D)$, within either method of extraction, indicates a transition in the classical simulability of the open quantum system.}
\label{fig:DivergentClassicalSimulation}
\end{figure*}
\section{Discussion}
This work introduces a new method to investigate the dynamics of open many-body quantum systems, the TDVP-Langevin equation. We derive this by considering an appropriate limit of the Keldysh path integral constructed over the MPS manifold. Our investigations reveal a phase transition in the applicability of this approach as a function of coupling to the environment -- when the bath temperature and induced friction are sufficiently high, entanglement growth in individual trajectories is suppressed, and a low bond order description works for all time. This is a transition in the classical simulabiity of the open quantum system.
We believe that this transition is related to several other transitions in quantum dynamics that have been observed as a function of coupling the the environment or measurement, including the restriction of entanglement growth in random circuits with projective measurement, the quantum Zeno effect (and perhaps the KT transition in the spin-boson model \cite{leggett1987dynamics,Florens2010quantum,barratt2020dissipative}).
The implications of this result may be far-reaching. In the context of using the TDVP-Langevin equation to simulate open quantum systems, an efficient description for long times is possible for systems in the many-body quantum Zeno phase. Indeed, when a target system is in such a phase, there is no (asymptotic) advantage in using a quantum computer to simulate it. Since many chemical reactions of potential interest for quantum computation occur embedded in a dissipative aqueous environment, this is certainly a point worthy of consideration.
Moreover, viewed from the perspective of a description of the quantum computational device, the transition into the many-body Zeno phase might indicate transitions in the ability to solve quantum problems.
While thresholds of noise for quantum error correction have been identified in the case of gate-based quantum computation, no such thresholds currently exist for adiabatic computation. It is intriguing to speculate that determining whether a putative adiabatic computational device is in its Zeno phase or not might provide similar bounds on performance\cite{crowley2014quantum,barratt2020dissipative}.
We envisage a number of ways in which this work might be developed.
Extending the approach to local observables in closed quantum systems presents some exciting possibilities. In this case the bath would refer to other elements of the system itself and its properties self-consistently determined through the evolution\cite{zwanzig1960ensemble,zwanzig1961memory},
Such a description has the promise of connecting early-time semi-classical descriptions to late-time hydrodynamics and thermalisation. Exploring the Fokker-Planck formulation of the TDVP-Langevin equation would bring a complementary perspective to our analysis\cite{berta2017thermal}.
The accurate description of a quantum system from early to late times is generally not possible because of growing entanglement. However, coupling to the environment can limit this growth and render this achievable. This work has coordinated physical insights from several different perspectives to develop such a numerical scheme. We hope both that the algorithm itself will prove useful and that it will inspire further insights.
\section{Acknowledgements}
We gratefully acknowledge funding from the EPSRC under grants EP/L015242/, EP/S005021/1 and EP/R020612/1.
\bibliographystyle{naturemag}
\section{A TDVP Langevin Equation}
\label{Sec:MPS Langevin Equation}
\begin{figure}[!t]
\includegraphics[width=0.5\textwidth]{Fig1.pdf}
\caption{
{\it Basic Properties of MPS Langevin Evolution:} The TDVP Langevin equation, Eq.(\ref{eq:TDVPLangevin}) describes the evolution of the density matrix through an ensemble of stochastic pure-state trajectories.
Here, the von Neumann entanglement as a function of time for a typical trajectory is shown calculated is shown calculated for different coupling strengths. Here the bond dimension is $D=128$ and we have kept temperature fixed, $T=0.2$, while we vary the noise, $\gamma T$, i.e. coupling $\gamma$. Such plots show three different regimes of behaviour. An initial transient, followed by an approximately linear in $t$ or logarithmic in $t$ growth and finally a saturation at the longest times. This saturation can either be determined by the variational approximation or intrinsically by the interplay of the Hamiltonian and the dissipative bath. It signals whether a lower bond-dimension (hence less computationally intensive) simulation suffices in underpinning the entropy dynamics.}
\label{fig:BasicProperties}
\end{figure}
Langevin equations describe the motion of a system coupled to an environment (or alternatively the motion of slow collective degrees of freedom in an effective bath described by the faster degrees of freedom\cite{zwanzig1960ensemble}) by adding noise and friction terms to the basic equations of motion of the system. If the environmental degrees of freedom are in thermal equilibrium, the friction and noise satisfy a fluctuation-dissipation relation. Applied to quantum systems, the Schr\"odinger equation provides the basic equations of motion. The ensemble of the resulting stochastic Schr\"odinger trajectories recovers the density matrix evolution and is said to be an unravelling of it.
{\it The Langevin equation} over matrix product states studied here, can be written in its Markovian limit as:
\begin{eqnarray}
\langle \partial_i \psi | \partial_j \psi \rangle \dot X_j
&=&
-i \langle \partial_i \psi | \hat H | \psi \rangle
-i \sum_n \langle \partial_i \psi | \hat F_n | \psi \rangle \eta(t)
\nonumber\\
& &
-i \sum_n \gamma \frac{\langle \psi | \hat F_n | \psi \rangle }{dt} \langle \partial_i \psi | \hat F_n | \psi \rangle .
\label{eq:TDVPLangevin}
\end{eqnarray}
The terms on the left-hand side and the first term on the right constitute the conventional time-dependent variational principle (TDVP) equations\cite{HaegemanTDVP}. The second and third terms on the right are, respectively, the noise and friction due to coupling to the environment. $\hat F_n$ are the operators by which the system is coupled to the bath displacement operators. We generally assume these to be spatially local. For spin-half chains they are given by the $x,y $ and $z-$components of the spin operators on each site, each of which couples to a separate bath. The bath is described as a collection of harmonic oscillators and the noise-correlator is determined by the spectrum of oscillators and the temperature of the bath; $\langle \langle \eta(t) \eta(t') \rangle \rangle=2 \gamma T \delta(t-t')$ in the combined classical and Markovian limits.
{\it The TDVP Langevin Equation} is a new approach to unravelling the density matrix evolution of an open system. It combines the study of quantum many-body dynamics on a variational manifold --- here we use matrix product states (MPS)\cite{Orus:2014zl,schollwock2011density,perez2006matrix} --- with the Langevin limit of Keldysh field theories\cite{kamenev2011field,sieberer2016keldysh}.
We combine these approaches by constructing a Keldysh path integral over MPS states\cite{green2016feynman}. To date, MPS techniques have been employed in the study of open systems largely in three ways: by starting with the Lindblad master equation and either describing the density matrix directly as a matrix product operator\cite{verstraete2004matrix,cui2015variational,weimer2019simulation} or else unravelling its evolution over MPS representations of quantum trajectories\cite{daley2009atomic,daley2014quantum,bonnes2014superoperators}; alternatively, tensor network techniques can be employed to describe the bath directly within the state ansatz\cite{strathearn2018efficient}. Of these, the latter is numerically costly and the Lindblad-based approaches formally apply in a different limit\cite{FergusThesis}.
Questions of applicability aside, there are some features of our approach that make it particularly attractive.
It naturally leads to thermal equilibrium, and it can treat the non-Markovian limit. There also exists a natural hydrodynamic limit and Fokker-Plank description. The scaling of the algorithm for each trajectory is the same as the usual TDVP for matrix product states -- moreover, there appears to be a degree of self averaging so that certain quantities are well-approximated by relatively fewer trajectories.
{\it A derivation} of Eq.(\ref{eq:TDVPLangevin}) is given in Appendix A. We develop the Langevin equation from the Keldysh path integral for the time-evolution of the density matrix. The method follows that of Ref.\cite{kamenev2011field} with the modification that the Keldysh path integral is constructed over matrix product states\cite{green2016feynman}. The result adds noise and friction to the time-dependent variational principle constructed over matrix product states\cite{HaegemanTDVP}. A similar construction for any variational class would lead to a similar stochastic equation of motion, which we dub the TDVP-Langevin equation. An alternative, heuristic derivation involves solving the Schr{\"o}dinger equation for the system and bath and substituting the solution of the latter into the equation of motion of the former. Alternative unravellings of the Lindblad equation for the density matrix evolution over MPS\cite{daley2009atomic} apply in different circumstances of relative time and energy scales of the bath and system.
We construct Eq.(\ref{eq:TDVPLangevin}) over matrix product states using conventional methods\cite{HaegemanTDVP}. Integration of this equation is complicated by the friction term. Naively, this requires inversion of a matrix that is proportional to the system size and dimension of the variational manifold. However, recognising that it consists of an outer product of vectors allows an efficient inversion and integration of the equations of motion. Details are given in Appendix B and our code is available at https://github.com/AndrewHallam/Langevin.
{\it Basic properties} of the TDVP Langevin equation are summarised in Fig.\ref{fig:BasicProperties}. A
thermal distribution over the variational manifold is given by a Boltzmann-weighted Haar average over the variational manifold. In the case the of MPS of bond dimension $D$, this average can be performed as a Haar integral over the group $SU(dD)$ (with $d$ the local Hilbert space dimension). Such thermal distributions are fixed points of the Langevin evolution (see Appendix \ref{app:HaarAverage}).
Fig.~\ref{fig:BasicProperties} captures the dynamics of the von Neumann entanglement entropy typical for trajectories over the parameters we consider. In the case shown, temperature $T=0.2$ is kept fixed and friction is increased. Beyond a critical value, $\gamma T\approx 0.1$, entanglement growth is suppressed. The entanglement undergoes a transition from being determined by the variational approximation to becoming intrinsic to the interplay between the Hamiltonian and the dissipative bath. Similar transitions occur when other parameters are kept constant (friction $\gamma$, or noise $\gamma T$). We detail these results in the following section.
\section{Open Evolution of a Rapidly Entangling System}
\label{Sec:Results}
In the absence of coupling to a bath, TDVP equations eventually fail as the entanglement grows beyond that which can be represented on the variational manifold\footnote{TDVP equations for the thermofield purification of the density matrix may escape this fate\cite{hallam2019lyapunov} at least as far as local observations are concerned}. However, just as observed in projective measurements of random circuits, the effects of the environment may restrict the growth of entanglement. {\it In extremis} this might limit entanglement of individual trajectories so that they can be represented on low dimensional variational manifolds. The TDVP Langevin equation will then give a good account of the dynamics at all times, signifying a transition in its classical representability. This is our interpretation of the sequence of results presented in this section.
{\it The Hamiltonian} that we consider is the tilted field Ising model
\begin{equation}
\hat H =
-\sum_i \left[ J \sigma_i^z \sigma_{i+1}^z +h \sigma^z_i+ g \sigma^x_i \right],
\label{eq:Hamiltonian}
\end{equation}
with $J=1$, $g=-1.05$ and $h=0.5$. With these parameters, the Hamiltonian is a far from any integrable point and rapidly thermalising\cite{banuls2011strong,leviatan2017quantum}.
{\it Infinite temperature and vanishing friction}: Fig.\ref{fig:InfiniteT} shows the variation in von Neumann entanglement across the central bond as a function of time for simulations with a range of bond orders and noise strengths. The broad result of these simulations is that the entanglement saturates at long times at a value determined by the bond order of the simulation. This is consistent with an infinite temperature final state with the maximum entanglement supported by the variational manifold. The most interesting aspect of these results is the decreasing rate of early-time entanglement growth with increasing noise strength. Crucially we do not find evidence of an intrinsic saturation of entanglement -- only that dictated by the limitations of the variational approximation.
\begin{figure}
\includegraphics[width=0.47\textwidth]{Fig2.pdf}\caption{
{\it Evolution of Entanglement at Infinite Temperature:} Here we consider the evolution under Eq.(\ref{eq:TDVPLangevin}) with Hamiltonian Eq.(\ref{eq:Hamiltonian}) with $J=1$, $g=-1.05$ and $h=0.5$. The friction coefficient $\gamma=0$ with noise $\gamma T$ finite and initial state $Z$.
a) von Neumann Entropy {\it versus} time at a fixed bond order $D=160$ for different values of $\gamma T$. Without noise, and after an initial transient, the entanglement grows linearly with time before saturating at a constant value less than the maximum determined by the MPS manifold. With non-zero noise, the entanglement growth is linear in time with a growth rate that reduces with increasing $\gamma T$ due to the dephasing effects of the bath. Ultimately, all curves saturate (not shown) to the same value characteristic of an infinite temperature state on the variational manifold.
b) The growth rate of entanglement $\dot{S}$, extracted beyond the initial transient ($t\approx4$), {\it versus} noise.}
\label{fig:InfiniteT}
\end{figure}
{\it Finite temperature and friction}:
Including both noise and friction, we do see such an intrinsic saturation. This is demonstrated in two ways; by considering the saturation of entanglement at long times and by a high fidelity between low- and high-bond order simulations at long times.
\\
\noindent
i. {\it Saturating entanglement}
In order to demonstrate this, we first show in Fig.~\ref{fig:SaturatingS} the long-time average of the von-Neumann entropy. A graph showing the typical time-dependence from which such saturation values are computed is shown in Fig.\ref{fig:BasicProperties}b). For low noise and friction, the saturation is determined by the limitations of the variational manifold. Panels a), b) and c) show that as a function of $\gamma T$ at fixed $\gamma$, then $T$, and $\gamma T$ at fixed $\gamma$, respectively. A threshold is reached for each bond order where it adequately captures the saturation entanglement, thus indicating a transition to increasingly classically simulatable dynamics. The transition can be seen from the point where the trajectories obtained at different bond orders give the same saturation entanglement. From this we can extract a critical $\gamma$ or $\gamma T$ as a function of bond order that we show in each corresponding inset figure.
\\
\noindent
ii. {\it High fidelity as } $t \rightarrow \infty$: We can identify an analogous transition in the fidelity of each trajectory at different bond orders versus a reference trajectory with bond order $D=128$. In this case, we find that beyond a critical combination of $\gamma$ or $\gamma T$, the fidelity of the state at low bond dimension remains close to $1$ for long times. We expand upon this result in Fig.~\ref{fig:DivergentClassicalSimulation}, where we identify a divergent classical simulation time. We note that the fidelity is more sensitive to the time-step as friction is increased -- an issue typical of numerical integration of systems of stochastic differential equations. This makes accessing the critical point of the transition numerically intensive for the parameters and Hamiltonian we consider. The entropy is less sensitive to this.
\begin{figure*}
\includegraphics[trim={6 0 0 0},clip, width=0.32\textwidth]{Fig3a.pdf}
\includegraphics[trim={5 4 5 0},clip, width=0.32\textwidth]{Fig3b.pdf}
\includegraphics[trim={0 0 6 0},clip, width=0.32\textwidth]{Fig3c.pdf}
\caption{
{\it Evolution of Saturation Entanglement at Finite Temperature:} Here we consider the evolution under Eq.(\ref{eq:TDVPLangevin}) with Hamiltonian Eq.(\ref{eq:Hamiltonian}) where $J=1$, $g=-1.05$ and $h=0.5$ and finite $\gamma$ and $T$. In the main figures we show the dependence of the von Neumann entropy as a function of noise and friction: a) {\it versus} $\gamma T$ at fixed $\gamma$, b) {\it versus} $\gamma T$ at fixed $T$, c) {\it versus} $\gamma$ at fixed $\gamma T$.
In each case, at low values of noise and friction, the saturation entanglement $\bar{S}$, is determined by the choice of variational parametrization through the bond order. As the noise and friction are increased, there is a cross-over where the saturation entanglement decreases from this value. Each bond order captures the saturation entanglement for a sufficiently large noise and friction. This is indicated when the entanglement begins to follow the entanglement given by the highest bond order simulation. Where the saturation entanglement of the individual trajectories (distinguished by bond order) dramatically fall to a single curve, so that the lower bond dimension curves coincide with higher bond dimensions, illustrates this transition to situations where the saturation entanglement is intrinsically determined. {\it In each corresponding inset figure}, we have extracted critical dissipation strengths where these transitions occur as a function of bond order.
}
\label{fig:SaturatingS}
\end{figure*}
\begin{figure*}
\includegraphics[trim={0 0.2cm 0 0},clip,width=17.5cm]{Fig4a.pdf}
\includegraphics[trim={0 0 0 0.2cm},clip,width=17.5cm]{Fig4b.pdf}
\newline
\includegraphics[width=17.5cm]{Fig4c.pdf}
\newline
\newline
\includegraphics[width=17.5cm]{Fig4d.pdf}
\caption{
{\it Divergent Classical Simulation Time:} Simulations carried out at a bond order $D$ give a good account of the system evolution up to a time $t^*(D)$. We extract these values versus a reference $D=128$ simulation, which serves as the good account of the system. We do this in two ways, by comparing the difference in von Neumann entanglement entropy between these states and the fidelity with this state. $t^*(D)$ is the time when the simulation with varying bond dimension deviates appreciably from the reference trajectory. The row with panels a), b), c), shows $t^*(D)$ extracted from the entropy, while panels d), e), f) demonstrate this for the fidelity. The fixed variables are split across the columns -- a), d), shows varying $\gamma T$ at fixed $\gamma=0.2$, b), e) $\gamma T$ at fixed $T=0.2$, and c), f) $\gamma$ at fixed $\gamma T=0.25$. Panels g) and h) show typical evolution of entanglement and fidelity with time. This is data from the near-critical point for $T=0.2$, which is for the trajectory with noise $\gamma T = 0.1$. We say that a simulation has failed to provide a good account of the system when the trajectory deviates beyond $\epsilon =0.05$, and the time at which this occurs is $t^*(D)$. In g), this is the point where $\Delta S/S_{D=128} = |S_{D=128}-S_{D}|/S_{D=128} > \epsilon$. Analogously in h), $t^*(D)$ is the time when the fidelity is appreciably different to $1$, i.e. $|\langle\psi_{D=128}(t)|\psi_{D}(t)\rangle|< 1-\epsilon$.
A divergent $t^*(D)$, within either method of extraction, indicates a transition in the classical simulability of the open quantum system.}
\label{fig:DivergentClassicalSimulation}
\end{figure*}
\section{Discussion}
This work introduces a new method to investigate the dynamics of open many-body quantum systems, the TDVP-Langevin equation. We derive this by considering an appropriate limit of the Keldysh path integral constructed over the MPS manifold. Our investigations reveal a phase transition in the applicability of this approach as a function of coupling to the environment -- when the bath temperature and induced friction are sufficiently high, entanglement growth in individual trajectories is suppressed, and a low bond order description works for all time. This is a transition in the classical simulabiity of the open quantum system.
We believe that this transition is related to several other transitions in quantum dynamics that have been observed as a function of coupling the the environment or measurement, including the restriction of entanglement growth in random circuits with projective measurement, the quantum Zeno effect (and perhaps the KT transition in the spin-boson model \cite{leggett1987dynamics,Florens2010quantum,barratt2020dissipative}).
The implications of this result may be far-reaching. In the context of using the TDVP-Langevin equation to simulate open quantum systems, an efficient description for long times is possible for systems in the many-body quantum Zeno phase. Indeed, when a target system is in such a phase, there is no (asymptotic) advantage in using a quantum computer to simulate it. Since many chemical reactions of potential interest for quantum computation occur embedded in a dissipative aqueous environment, this is certainly a point worthy of consideration.
Moreover, viewed from the perspective of a description of the quantum computational device, the transition into the many-body Zeno phase might indicate transitions in the ability to solve quantum problems.
While thresholds of noise for quantum error correction have been identified in the case of gate-based quantum computation, no such thresholds currently exist for adiabatic computation. It is intriguing to speculate that determining whether a putative adiabatic computational device is in its Zeno phase or not might provide similar bounds on performance\cite{crowley2014quantum,barratt2020dissipative}.
We envisage a number of ways in which this work might be developed.
Extending the approach to local observables in closed quantum systems presents some exciting possibilities. In this case the bath would refer to other elements of the system itself and its properties self-consistently determined through the evolution\cite{zwanzig1960ensemble,zwanzig1961memory},
Such a description has the promise of connecting early-time semi-classical descriptions to late-time hydrodynamics and thermalisation. Exploring the Fokker-Planck formulation of the TDVP-Langevin equation would bring a complementary perspective to our analysis\cite{berta2017thermal}.
The accurate description of a quantum system from early to late times is generally not possible because of growing entanglement. However, coupling to the environment can limit this growth and render this achievable. This work has coordinated physical insights from several different perspectives to develop such a numerical scheme. We hope both that the algorithm itself will prove useful and that it will inspire further insights.
\section{Acknowledgements}
We gratefully acknowledge funding from the EPSRC under grants EP/L015242/, EP/S005021/1 and EP/R020612/1.
\bibliographystyle{naturemag}
|
\section{Introduction}
\label{Intro}
Image analysis of digitized histopathological slides can contribute significantly to cancer diagnosis~\cite{irshad2013methods}.
For instance, the diagnosis of cervical cancer and its precancerous stages can be accomplished through assessment of histopathology slides of cervical tissue by pathologists.
An important outcome of the assessment is the
cervical intraepithelial neoplasia (CIN) grade, an essential indicator for abnormality
assessment identified by the abnormal growth of cells on the surface of the cervix. Over the past decade, computer-assisted diagnosis (CAD) algorithms have been developed for histopathology images to complement the opinion of the pathologist for accurate disease detection, diagnosis, and prognosis prediction~\cite{gurcan2009histopathological}. Considering the shortage of pathologists, automatic histopathology image classification systems have great potential
in underdeveloped regions for its low cost and accessibility. Moreover, such a system can help pathologists with diagnosis and potentially mitigate the inter- and intra- pathologist variation.
The supervised training of image recognition systems often requires huge amounts of expert annotated data to reach a high level of accuracy. However, for many practical applications using histopathology images, only small datasets of labeled data are available due to annotation cost and privacy concerns, and the labels are often imbalanced between grades and subtypes.
While traditional data augmentation can increase the amount of training data to some degree, commonly employed random transformations or distortions (such as cropping and flipping) lack flexibility and cannot fill the entire data distribution with missing data samples.
Motivated by the aforementioned difficulties in creating sufficiently large training sets for histopathology image recognition systems, we focus on the problem of expanding training sets with high-quality synthetic examples.
Recently, several works in medical image analysis have leveraged unsupervised learning methods, more specifically, Generative Adversarial Networks (GANs)~\cite{goodfellow2014generative}, to mitigate the effects of small training sets on network training
~\cite{liu2019wasserstein,frid2018gan}. These works show that carefully designed GANs can generate visually appealing synthetic images, but two major issues remain insufficiently investigated for generalized and robust synthetic augmentation: 1)
how to mitigate label ambiguity of generated images; and 2) how to ensure the feature quality of synthetic images used for data augmentation.
In other words, blindly incorporating synthetic samples into the original training set, even if they are visually realistic, is not guaranteed to improve the classification model performance. Synthetic images without quality assurance can potentially adversely alter the data distribution and downgrade model performance. We provide a detailed analysis in Section~\ref{sec:results}.
In this paper, we aim at solving these two issues by designing a novel conditional GAN (cGAN)~\cite{mirza2014conditional} framework, termed as HistoGAN, for high-fidelity histopathology image synthesis, then \textit{selectively} adding synthetic samples generated by HistoGAN to the original training set. Our proposed HistoGAN model consists of multiple progressive generation and refinement modules which gradually generate images with better quality. To encourage the diversity of synthetic images, we incorporate the minibatch discrimination~\cite{salimans2016improved} to reduce the closeness between
examples inside a minibatch. Self attention~\cite{vaswani2017attention} is employed to capture relationships between pixels inside an image. Such relationships contain crucial information about histopathology images, including the density distribution of nuclei and color changes in different locations. Class conditional batch normalization~\cite{de2017modulating} and spectral normalization~\cite{miyato2018spectral} are also utilized to stabilize the adversarial training process and improve the quality of synthetic images. Further, during HistoGAN training, we calculate a smoothed version of Fréchet Inception Distance (FID)~\cite{heusel2017gans} score after each epoch of training so that the trained models can be compared and the model with weights that give rise to the best FID score can be selected. Our proposed HistoGAN consistently generates realistic histopathology image patches on two different datasets, which shows the robustness and generality of the model.
Our proposed selective synthetic augmentation framework consists of two steps. First, we select generated images that can be classified into some class with certainty, by calculating the expectation of predictive entropy of each sample and keeping those samples with relatively low entropy (\textit{i.e.}, high label confidence). Second, we compare the features of real images and synthetic images where the ground truth label of the real images matches the conditional label used to generate the synthetic images, and only select those synthetic images that are sufficiently close to the real-image centroid in feature space. The features of the images are extracted by a feature extractor pre-trained with Monte Carlo dropout (MC-dropout)~\cite{gal2016dropout}. This second step of selection is to ensure that a selected synthetic image indeed belongs to the class that corresponds to the conditional label used to generate it.
The total number of selected samples is determined according to the augmentation ratio $r$ (\textit{i.e.}, the proportion of the number of augmented samples to the number of original training samples). Experimental results show that our proposed HistoGAN model along with selective synthetic augmentation significantly outperforms the baseline ResNet34~\cite{he2016deep} model with traditional augmentation, and also outperforms the synthetic augmentation methods without selection.
To validate the effectiveness and generality of our proposed selective synthetic augmentation framework, we conduct extensive experiments on two histopathology datasets. We first study the 4-class (Normal, CIN 1-3) cervical histopathology image classification problem and evaluate our models on a heterogeneous epithelium image dataset~\cite{xue2019synthetic} with limited and highly unbalanced numbers of patch-level annotations per class label. The second dataset we use is a small subset of the PCam dataset~\cite{veeling2018rotation}, consisting of lymph node histopathology images.
We compare our proposed selective synthetic augmentation method with baseline methods including baseline classification models, models trained with traditional augmentation, and models trained with synthetic augmentation but without quality-assuring selection. Experimental results show that our model achieves significant improvements with $\textbf{6.7}\%$ and $\textbf{2.8}\%$ higher accuracy than baseline classification models on cervical and lymph node datasets, respectively.
The main contributions of this work are as follows:
\begin{itemize}
\setlength{\itemsep}{-0.2em}
\item We design a novel conditional GAN model architecture for synthesizing realistic histopathology image patches.
A smoothed version of FID score is used as a metric to select the best cGAN model during training. With only a limited amount of training data, our GAN model can generate synthetic images with high fidelity and diversity.
\item We propose a selective synthetic augmentation method that actively selects synthetic samples with high confidence of matching to their conditional label and are close to real images in feature space. By only adding selected synthetic samples instead of arbitrary synthetic samples to augment the limited training set, our proposed method can significantly outperform other baseline augmentation methods in improving classification performance.
The proposed selective synthetic augmentation is general and can also be used in conjunction with other augmentation methods.
\item We conduct extensive experiments on both a cervical histopathology dataset and a lymph node histopathology dataset. Compared with baseline models, including our previous state-of-the-art synthetic augmentation model~\cite{xue2019synthetic}, our proposed method improves the augmented classification performance.
\end{itemize}
\section{Related Work}
\subsection{Histopathology Image Classification}
Machine learning, especially deep learning methods have achieved promising results on general histopathology image classification. While whole slide images (WSI) are often with unusually high resolutions, commonly used methods~\cite{hou2016patch, xu2017large, tomita2019attention} alleviate this issue by applying patch-level image classification on cropped image patches or sliding windows rather than the original WSI. Individual classification results on cropped patches are aggregated to infer the final image-level label for the WSI. In such methods, accurate patch-level image classification is fundamental to reach the accuracy level of human pathologists.
In the area of cervical histopathology analysis, existing literature~\cite{chankong2014automatic, guo2016nuclei} have studied various supervised learning methods for nuclei-based cervical cancer classification.
Chankong~\textit{et~al.}~\cite{chankong2014automatic} proposed automatic cervical cancer cell segmentation and classification using fuzzy C-means (FCM) clustering and various types of classifiers.
Guo~\textit{et~al.}~\cite{guo2016nuclei} designed hand-crafted nuclei-based features for fusion-based classification on digitized epithelium histopathology slides with linear discriminant analysis (LDA) and support vector machines (SVM) classifier. While accomplishments have been achieved with fully-supervised learning methods, the training of models require large amounts of expert annotations of cervical histopathology images. Since the annotation process can be expensive, tedious, and time-consuming, it often results in limited or insufficient number of labeled data available for supervised learning models.
\subsection{Conditional Image Synthesis}
Generative adversarial networks (GANs)~\cite{goodfellow2014generative} as an unsupervised learning technique, has enabled a wide variety
of applications including image synthesis, object detection~\cite{li2017perceptual} and image segmentation~\cite{xue2018segan}. Among variants of GANs, conditional GAN (cGAN) generates~\cite{mirza2014conditional, odena2017conditional} more interpretable results with conditional inputs. For instance, images can be generated conditioning on class labels, which enables cGAN to serve as a tool to generate labeled samples for synthetic augmentation. Current state-of-the-art cGAN models often breaks the task into smaller gradual generation or refinement sub-tasks~\cite{zhang2018stackgan++, karras2017progressive}, or employs large scale training~\cite{brock2018large}, which enable them to generate high fidelity images.
In this work, we use our proposed HistoGAN, which is inspired by state-of-the-art cGANs~\cite{zhang2018stackgan++, zhang2019self, brock2018large}, to generate high-fidelity synthetic images to augment classification model training. To improve the quality of synthetic images and stabilize the training process, our model utilizes numerous techniques including minibatch discrimination~\cite{salimans2016improved}, self attention~\cite{vaswani2017attention}, class conditional batch normalization~\cite{de2017modulating}, and spectral normalization~\cite{miyato2018spectral} following prior art. While generating visually appealing histopathology images, HistoGAN serves as an essential prerequisite for the synthetic data augmentation.
\subsection{Synthetic Data Augmentation}
To better utilize training data and reduce over-fitting during the training process, data augmentation has become a common practice for training deep neural networks. The objective of augmentation is to add to the original training set new samples that follow the original data distribution. Therefore, a good augmentation scheme should generate samples that follow the original data distribution but are different from those in the original training set. On the other hand, a bad augmentation scheme can generate samples that deviate from the original data distribution thus can mislead training when added to the training set.
Traditional data augmentation~\cite{wang2017effectiveness} often involves transformations applied directly on original training data, such as cropping, flipping and color jittering. While serving as an implicit regularization, straightforward data augmentation techniques are limited in augmentation diversity. To overcome the limitation of traditional augmentation, several works have been done to improve the effectiveness of data augmentation. Rather than using a pre-defined augmentation policy, Auto Augmentations~\cite{cubuk2019autoaugment, ho2019population} use hyper-parameter searching to automatically find the optimal augmentation policy.
Another popular trend is to generate synthetic images to increase the amount and diversity of original training data, which we denote as \textit{Synthetic Augmentation}. Along this direction, for natural images, Ratner~\textit{et~al.}~\cite{ratner2017learning} learns data transformation with unlabeled data using GANs. GAGAN~\cite{antoniou2017data} and BAGAN~\cite{mariani2018bagan} uses cGANs~\cite{mirza2014conditional} generated samples to augment the standard classifier in the low-data regime. Compared with works done in the natural image domain, issues related to insufficient and imbalanced data are more prominent in the medical image domain. To mitigate these problems, researchers have been working on synthetic augmentation for medical image recognition tasks.
Frid-Adar~\textit{et~al.}~\cite{frid2018gan} proposes to use cGAN generated synthetic CT images to improve the performance of CNN in liver lesion classification. Gupta~\textit{et~al.}~\cite{gupta2019generative} synthesizes lesion images from non-lesion ones using CycleGAN~\cite{zhu2017unpaired}. Bowles~\textit{et~al.}~\cite{bowles2018gan} uses GAN derived synthetic images to augment medical image segmentation models. Zhao~\textit{et~al.}~\cite{zhao2018synthesizing} proposes a GAN model for synthesizing retinal images from small sized samples and uses the synthetic images to improve semantic segmentation performance. Mahapatra~\textit{et~al.}~\cite{mahapatra2018efficient} applies a Bayesian neural network (BNN)~\cite{mackay1992practical} to calculate the informativeness of the synthetic images for improved classification and segmentation results. Zhao~\textit{et~al.}~\cite{zhao2019data} uses transformations of labeled images for one-shot image segmentation. GAN based synthetic augmentation has achieved promising results, but typically blindly adds synthetic samples to the original data. Few consider how to assure the quality of synthetic images or control the augmentation step after image synthesis.
\subsection{Our Previous Work}
In our recent work~\cite{xue2019synthetic}, we propose a feature based filtering mechanism for synthetic augmentation. While improving classification performance, our previous cGAN generated images are not realistic enough and the work lacks rigorous study of its GAN model training and feature extractor training processes.
In this work, we propose an improved GAN model for histopathology image generation, and develop a more general synthetic augmentation framework by reducing the randomness in GAN model and feature extractor training through MC-sampling and FID score based model selection. Our new contributions and differences from previous work
are summarized as follows:
\begin{itemize}
\setlength{\itemsep}{-0.2em}
\item
We design and utilize an improved conditional GAN model architecture, namely HistoGAN, with a self-attention module among other techniques to stabilize the training and improve the quality of synthetic images.
\item
We propose a more general selective synthetic augmentation method which achieves better performances than our previous method.
\item
We conduct more comprehensive experiments including more ablation study and new results on the PCam dataset.
\end{itemize}
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.99\linewidth]{architecture_comparison.pdf}
\end{center}
\caption{Comparison between different training processes. (a) Traditional training pipeline; (b) Conditional GAN augmented training pipeline; (c) Our proposed selective synthetic augmentation with quality assurance. The input to the cGAN are noise vector $z$ and label condition vector $y$.}
\label{fig:arch_comparison}
\end{figure}
\begin{figure*}[ht]
\begin{center}
\includegraphics[width=0.95\linewidth]{method.pdf}
\end{center}
\caption{The architecture of the proposed selective synthetic augmentation algorithm. The $\cup$ symbol indicates that the selected synthetic image set is unioned with the original training set to improve classification model training and test performance. }
\label{fig:architecture}
\end{figure*}
\section{Methodology}
In traditional fully-supervised training methods, the model is trained on training images and the inference is done by feeding the test data to the trained model. In previous GAN-based augmentation works~\cite{frid2018gan,madani2018chest}, a GAN model is first trained to generate some synthetic images based on the training data, then the generated images are added to the original training data as a data augmentation strategy.
However, since the discriminator in GAN only outputs a high level judgement ($0$ or $1$) of the fidelity of generated images, such pipelines cannot guarantee that the generated data contain meaningful features which contribute to improving classification model training. To tackle this issue, we propose a selective synthetic augmentation algorithm to evaluate the quality and fidelity of synthetic images and select only those samples with high-confidence in label correctness and real-image likeness to be added to the training set.
The comparison between different training procedures is illustrated in Fig.~\ref{fig:arch_comparison}.
An overall illustration of our proposed selective synthetic augmentation method can be found in Fig.~\ref{fig:architecture}. We first train a conditional GAN model based on the labeled training images. The optimal model weights is selected based on the smoothed FID score~\cite{heusel2017gans}. A pool of synthetic images are then generated using the selected model. All images are then passed into the image selection module to filter out the ones that fail to contribute sufficient amount of meaningful information. After image selection, a classification model is trained with both original and synthetic training data. Trained classification models can then be used for inference on test data. More details are introduced in the following subsections.
\subsection{HistoGAN Model}
In this section, we introduce our proposed HistoGAN architecture and how to select the best model with highest synthetic image quality from a set of trained models.
\subsubsection{Model Architecture} \label{sec:architecture}
The conventional cGANs~\cite{mirza2014conditional} have an objective function defined as:
\begin{multline}
\min_{\theta_G} \max_{\theta_D} \mathcal{L}_{\text{cGAN}} = \mathbb{E}_{x\sim P_\text{data}}[\log D(x,y)] + \\ \mathbb{E}_{z\sim \mathcal{N}}[\log (1 - D(G(z,y)))] \enspace .\label{Eq:cGAN}
\end{multline}
In the equation above, $x$ represents the real data from an unknown image distribution $P_\text{data}$ and $y$ is the conditional label (\emph{e.g.}, CIN grades). $z$ is a random vector for the generator $G$, drawn from a standard normal distribution $\mathcal{N}(0,1)$. During the training, $G$ and $D$ are alternatively optimized to compete with each other.
\begin{figure*}[ht]
\begin{center}
\includegraphics[width=0.95\linewidth]{histogan_architecture.pdf}
\end{center}
\caption{The architecture of a 3-stage HistoGAN for cervical epithelium synthesis. The number of stages can be adjusted according to the desired final image resolution. Detailed features such as cytoplasm texture and nuclei shapes get progressively refined in synthetic images of higher resolution from stage I to III. The self attention layer is applied after stage I generator where the sketch outline and rough pattern of images are shaping up. Self attention layers are also incorporated in discriminators at all stages to further enforce the consistency of focused local regions more accurately. Conditional batch normalization~\cite{de2017modulating} is used after convolutional layers for flexibly modulating convolutional feature maps.}
\label{fig:net}
\end{figure*}
Since there is no existing cGAN framework specifically designed for histopathology image synthesis, we choose to design a new model, \textit{HistoGAN}, based on previous state-of-the-art conditional GAN models and techniques~\cite{zhang2018stackgan++,brock2018large,zhang2019self}. We aim to generate synthetic images in a coarse-to-fine fashion through multiple stages, where details of images are gradually refined to guarantee the fidelity. The training procedure of HistoGAN is similar to Eq.~\ref{Eq:cGAN}. The generator of the first stage takes a random noise vector and class label as input, and the generator of remaining stages will take the output of the previous stage as input instead of random noise. To
increase diversity among the generated examples and mitigate the issue of mode collapse indicated by the high homogeneity of the synthetic image pool, we incorporate the minibatch discrimination module~\cite{salimans2016improved} into our discriminator.
Following state-of-the-art works in conditional image synthesis~\cite{zhang2019self,brock2018large}, class conditional batch normalization is used in both generators and discriminators to enhance the learning effectiveness of the inter class feature discrepancy. And spectral normalization~\cite{miyato2018spectral} is utilized in discriminators of all stages to further improve model performance.
To better capture the distribution of nucleus density and color changes in histopathology images of different classes, we leverage self attention~\cite{vaswani2017attention,zhang2019self} at early stages of generation and throughout all stages in the discrimination process.
The application of self attention mechanism enables both generator and discriminator to better learn the dependencies between spatial regions by looking at the relationship between one pixel and all other positions in the same image. Similar to~\cite{zhang2019self}, the image features from the previous hidden layer $x$ are first transformed into two feature spaces $q, k$ as query and key in self attention~\cite{vaswani2017attention}
to calculate the attention map. Let $q(x) = W_qx$ and $k(x) =
W_kx$, the attention map over the $i$th location when synthesizing the $j$th region is
\begin{equation}
\alpha_{j, i}=\frac{\exp \left(s_{j i}\right)}{\sum_{i=1}^{N} \exp \left(s_{j i}\right)}, \text { where } s_{j i}=\boldsymbol{q}\left(\boldsymbol{x}_{i}\right)^{T} \boldsymbol{k}\left(\boldsymbol{x}_{j}\right)\enspace.
\label{eq:sa1}
\end{equation}
The output of the self attention of the $j$th region $o_j$ is calculated by applying attention weight over the value $v$ as
\begin{equation}
\boldsymbol{o}_{j}=\sum_{i=1}^{N} \alpha_{j, i} \boldsymbol{v}\left(\boldsymbol{x}_{i}\right), \text { where } \boldsymbol{v}\left(\boldsymbol{x}_{i}\right)=\boldsymbol{W}_{v} \boldsymbol{x}_{i}\enspace.
\label{eq:sa2}
\end{equation}
In all transformation matrices $W_q, W_k$, and $W_v$, weight matrices are implemented as $1\times1$ convolutions. Compared with the StackGAN model implemented in our previous work~\cite{xue2019synthetic}, our HistoGAN generates more realistic image patches which also benefits the following synthetic augmentation step. An example of cGAN result comparison is shown in Fig.~\ref{fig:cervical}.
\subsubsection{Model Selection}\label{model_selection}
During training of the HistoGAN model, the model weights vary from epoch to epoch. A challenge is to determine which model weights gives rise to better synthetic image quality. For natural image synthesis tasks, Inception Score~\cite{salimans2016improved} and Fréchet Inception Distance (FID)~\cite{heusel2017gans} score are two commonly used metrics. The calculation of these two metrics rely on the pre-trained Inception V3~\cite{szegedy2016rethinking} model trained on ImageNet~\cite{deng2009imagenet}. However, since the distribution of natural images and that of medical images such as cervical histopathology images can be quite different, we can not directly use the aforementioned two scores for evaluating our HistoGAN model.
Instead, we follow the calculation of the original FID score while replacing the Inception V3 model pre-trained on ImageNet with a ResNet34~\cite{he2016deep} model pre-trained on the cervical histopathology dataset.
\begin{figure*}[t]
\begin{center}
\includegraphics[width=0.99\linewidth]{smoothfid.pdf}
\end{center}
\caption{FID scores (pre-trained ResNet34) of HistoGAN models saved after different number of epochs of training. Scores are smoothed with varying EMA parameter $\alpha$.}
\label{fig:fid}
\end{figure*}
To compare the trained models after running different numbers of epochs, we save the HistoGAN model after each epoch of training. To estimate the performance of each saved model, we calculate the FID score between the feature vectors of real and generated images extracted from the pre-trained ResNet34 model as follows:
\begin{small}
\begin{multline}
d(x, \tilde{x}) = \left|\left| \mu_{x \sim P_\text{data}}\phi(x) - \mu_{\tilde{x} \sim P_G}\phi(\tilde{x}) \right|\right|_{2}^{2} + \\
\mathbf{Tr} \left (\Sigma_{x \sim P_\text{data}}\phi(x) + \Sigma_{\tilde{x} \sim P_G}\phi(\tilde{x}) - 2 \left (\Sigma_{x \sim P_\text{data}}\phi(x) \Sigma_{\tilde{x} \sim P_G}\phi(\tilde{x}) \right )^{\frac{1}{2}} \right ) \enspace ,
\end{multline}
\end{small}
where $\tilde{x}$ represents synthetic images generated by the saved HistoGAN model being evaluated, and $\phi$ denotes the features extracted from intermediate layers of the pre-trained ResNet34 model. Assume feature vectors follow a multivariate Gaussian distribution, the mean and covariance are estimated for the real and fake data~\cite{borji2019pros} for fréchet distance calculation to measure the visual quality of generated images. Smaller FID scores indicate better visual quality.
Although the FID score itself cannot guarantee agreement with human judgment, trends of FID often provide a reliable estimation of the quality of a GAN model. As we can observe from Fig.~\ref{fig:fid}, due to the instability in GAN training, the FID scores of each saved epoch fluctuate constantly and fail to provide a distinguishable pattern. Based on the unaltered FID scores, one should choose the model saved at epoch 286 or epoch 374. However, one can see that images generated by these chosen models are not satisfactory as in Fig.~\ref{fig:fid}. To get a robust estimation of model quality and mitigate the effect caused by outliers, we apply the Exponential Moving Average (EMA)~\cite{hunter1986exponentially} algorithm to smooth the curve of original FID score. With smoothing, the FID score at time $t$ is:
\begin{equation}
\hat{d} = \left\{\begin{matrix} \text{d}_t,& \enspace t=1
\\ \alpha \hat{d}_{t-1} + (1-\alpha)\text{d}_t,& \enspace t>1 \label{Eq:smoothFID}
\end{matrix}\right.
\end{equation}
We monitor the training process with the smoothed FID. As shown in Fig.~\ref{fig:fid}, different values of $\alpha$ lead to different levels of smoothing in FID and we observed that the chosen model associated with the lowest smoothed FID score has better image quality than the model chosen using the lowest original FID score. In our experiments, we set $\alpha$ to $0.5$ for a medium level of smoothing. One can see that, after smoothing with the EMA algorithm, the minimum in smoothed FID score is reached at epoch 634, which is the GAN model we chose for the follow-on synthetic data augmentation.
\subsection{Image Selection}\label{img_selection}
Given a trained cGAN model, one can sample infinite number of noise-vector inputs from the Gaussian distribution and generate infinite number of synthetic images. While a good cGAN model can generate images that look real, there are no guarantee that those images would be good to be used for augmenting the original training set in visual recognition tasks. In current GAN-based data augmentation methods, with different data augmentation ratio, different number of generated images are added to the training set. However, the effectiveness of such augmentation pipeline is heavily affected by the varying quality of synthetic images as well as the diversity of the images. To reduce the randomness in the synthetic augmentation process and selectively add in new images, we break the whole process into two steps: find samples that can be confidently classified into certain classes thus containing enough diagnosable features; then find samples whose features are within a certain neighborhood of class centroids in the feature space to assure matching between the synthetic image and its assigned label. Such steps are done with a pre-trained feature extractor to calculate centroids for real samples and extract features for fake samples. Considering that a single feature extractor cannot provide robust feature extraction results, we use a feature extractor with Monte Carlo dropout (MC-dropout)~\cite{gal2016dropout} and take the expectation value of multiple samplings to reduce the uncertainty of feature extraction. A depiction of our proposed selective synthetic augmentation algorithm is shown in Fig.~\ref{fig:img_selection} and a detailed description is given in Algorithm~\ref{algorithm}.
The first step of selection is based on label certainty of a sample. In traditional machine learning systems, real samples that lie near the decision boundary are often assumed to contain more important features for classification purposes. However, as we conducted experiments to select good synthetic images, one interesting finding is that selecting the fake samples with more certain labels gives better classification performance than selecting those with less certain labels.
This may be due to the cGAN model being imperfect and conditionally-generated fake examples with less label certainty being more likely to deviate from the real data distribution.
In our algorithm, we evaluate the label certainty of a fake example by calculating the entropy score of its predicted class probabilities.
If the feature extractor is certain that a sample can be classified into a certain class, the entropy score would be low. We rank the entropy scores of all generated images in ascending order and choose the first half of images with lower entropy.
The necessity of this entropy-based selection is proved by experiments on different datasets, which will later be discussed in Section~\ref{Experiments}.
\begin{algorithm}[!t]
\begin{algorithmic}
\caption{Selective Synthetic Augmentation}\label{algorithm}
\STATE \textbf{Input}: a set of trained HistoGAN models \{$G_t$\}, number of classes $\mathcal{C}$, augmentation ratio $r$, number of original training samples $N = \sum_{i=1}^{\mathcal{C}} N_i$.
\STATE \textbf{Output}: selected synthetic samples $\mathcal{X}$ with $|\mathcal{X}|=rN$.
\STATE \textbf{Initialization}: $\mathcal{X}_1 = \emptyset$,
$\hat{t} = \arg \min (\hat{d_t})$,
$G_{\hat{t}}$ generated samples $\mathcal{X}_{0}$ = \{$x_j^i: i \leq \mathcal{C}, j \leq 4rN_i$ \},
entropy $\mathcal{E}^i = \{e_j^i: e_j^i = - \sum p_j^i \log p_j^i, i \leq \mathcal{C}, j \leq 4rN_i \}$. \\
\FOR {$x_j^i \in \mathcal{X}_0$}
\IF {$e_j^i < \text{Median} (\mathcal{E}^i)$}
\STATE $\mathcal{X}_1 = \mathcal{X}_1 \cup \{x_j^i\}$
\ENDIF
\ENDFOR \\
class centroid distance $\mathcal{D}^i = \{d_j^i: d_j^i = D_f (x_j^i, c_i)\}$. \\
\FOR {$x_j^i \in \mathcal{X}_1$}
\STATE $d_j^i = D_f \{x_j^i, c_i\}$
\IF{$d_j^i < \text{Median} (\mathcal{D}^i)$}
\STATE $\mathcal{X} = \mathcal{X} \cup \{x_j^i\}$
\ENDIF
\ENDFOR
\end{algorithmic}
\end{algorithm}
After the entropy selection step, we further select synthetic images based on their distance to class centroids in the feature space. In this second step of selection, all remaining samples that have passed the entropy-based selection will have their feature distances to their class centroids calculated. All distances will be sorted in ascending order and the first half of these samples with smaller distances will be kept.
The motivation behind ranking samples based on their feature distance to class centroids is to help filter out samples whose assigned labels (i.e. the conditional labels used by the cGAN
model to generate them) do not match their classified labels in feature space so that only samples that confidently match with their assigned labels are selected and added to the training set.
In our implementation, instead of using a single run of the feature extractor to extract features, we run the feature extractor multiple times with MC-sampling and then calculate feature distances based on the average feature distance from the multiple runs.
Similar to~\cite{xue2019synthetic}, the feature distance between image $x$ and centroid $c$ is defined as
\begin{equation}
D_{f}(x,c_i) = \frac{1}{K}\sum_{k}\sum_{l} \frac{1}{H_lW_l} \left|\left|\hat{\phi_l^k}(x) - \hat{\phi_l^k}(c_i)\right|\right|_{2}^{2} \enspace ,\label{Eq:feature}
\end{equation}
\noindent where $\hat{\phi_l^k}$ is the unit-normalized activation in
the channel dimension $A_l$ of the $l$th layer of the $k$-th MC-sampling feature extraction network with shape $H_l \times W_l$. We denote the total sampling time as $K$. $D_{f}(x,c_i)$ can be regarded as an estimated cosine distance between sample and $i$-th centroid in the feature space.
The centroid $c$ is calculated as the average feature of all labeled training images in the same class. For class $i$, its centroid $c_i$ is represented by
\begin{equation}
c_i = \left[\frac{1}{N_i}\sum_{j=1}^{N_i}\phi_1(x_j),...,\frac{1}{N_i}\sum_{j=1}^{N_i}\phi_L(x_j)\right] \enspace ,\label{Eq:centroid}
\end{equation}
\noindent where $N_i$ denotes the number of training samples in $i$th class and $x_j$ is the $j$th training sample. Similar to Eq.~\ref{Eq:feature}, $\phi_l$ is the activation extracted from the $l$th layer of the feature extraction network. $L$ is the total number of layers utilized in the feature distance selection. $c_i$ is retained by one time MC-sampling and fixed during the distance calculation.
In conclusion, given augmentation ratio $r$, we first generate $4rN_i$ images for each class $i$, then select $rN_i$ images according to the two-step selection process described above. Regarding the choice of $r$, we provide an
ablation study in Section~\ref{sec:results}.
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.95\linewidth]{image_selection.pdf}
\end{center}
\caption{Illustration of the image selection process. $r$ and $N$ represent the augmentation ratio and the number of original training data. The same feature extractor runs multiple times through MC-dropout for both entropy and class centroid distance calculations to increase robustness.}
\label{fig:img_selection}
\end{figure}
\section{Experiments}\label{Experiments}
\begin{figure*}[!ht]
\begin{center}
\includegraphics[width=0.98\linewidth]{cervical_gan_results.pdf}
\end{center}
\caption{Examples of real images, synthetic images generated from~\cite{xue2019synthetic}, and images generated by our HistoGAN model trained on cervical histopathology dataset before and after selection. Our HistoGAN generates realistic images with clearly better visual quality than those by ~\cite{xue2019synthetic}. Zoom in for better view.}
\label{fig:cervical}
\end{figure*}
\begin{figure*}[!ht]
\begin{center}
\includegraphics[width=0.98\linewidth]{lymph.pdf}
\end{center}
\caption{Examples of real and synthetic images generated by HistoGAN trained on 10\% of PCam dataset.}
\label{fig:pcam}
\end{figure*}
\begin{figure*}[!ht]
\begin{center}
\includegraphics[width=0.98\linewidth]{lymph_full.pdf}
\end{center}
\caption{Examples of real and synthetic images generated by HistoGAN trained on 3\%, 5\%, 10\% and 20\% of PCam dataset. All generated images are chosen from the pool after applying our proposed image selection method. Zoom in for better view.}
\label{fig:pcam_full}
\end{figure*}
\begin{figure*}[!ht]
\begin{center}
\includegraphics[width=0.98\linewidth]{self_attn_map.pdf}
\end{center}
\caption{The attention map extracted from the self attention layer applied after stage I generator as illustrated in Figure \ref{fig:architecture}. The first row shows the synthetic images generated by our proposed HistoGAN model; the second row gives the most attended regions of each image during the GAN training phase by overlaying the attention map on top of the original image. Higher attention scores correspond to the highlighted areas where distinguishing patterns like cell crowding and nuclei distribution are highlighted. It demonstrates the effectiveness of the attention mechanism incorporated in our HistoGAN model.}
\label{fig:self_attn_map}
\end{figure*}
\subsection{Datasets}
The first dataset contains labeled cervical histopathology images collected from a collaborating health sciences center. All images are annotated by the same pathologist. The data processing follows~\cite{xue2019synthetic}, and results in patches with a unified size of $256 \times 128$ pixels. Compared with the dataset used in~\cite{xue2019synthetic}, we include more data for more comprehensive experiments. In total, there are $1,284$ Normal, $410$ CIN1, $481$ CIN2, $472$ CIN3 patches. Examples of the images can be found in the first row of Fig.~\ref{fig:cervical}.
We randomly split the dataset, by patients, into training, validation, and testing sets, with ratio 7:1:2 and keep the ratio of image classes almost the same among different sets. All evaluations and comparisons reported in this section are carried out on the test set.
To further prove the generality of our proposed method, we also conduct experiments on the public PatchCamelyon (PCam) benchmark~\cite{veeling2018rotation}. PCam consists of $327,680$ color patches extracted from histopathologic scans of lymph node sections with unified size of $96 \times 96$ pixels. The PCam dataset is split into 75\%:12.5\%:12.5\% of training, validation, and testing
sets, selected using a hard-negative mining regime. Each image is annotated with a binary label indicating presence of metastatic tissue. To mimic the situation where only a limited amount of training data is available, we use randomly selected $10\%$ of the training set, which has 32,768 patches, to train our proposed HistoGAN model and the baseline classifier. Trained models are evaluated on the full test set.
\begin{figure*}[!ht]
\begin{center}
\includegraphics[width=0.99\linewidth]{histology_tsne_new.png}
\end{center}
\caption{t-SNE of the original and augmented cervical histopathology training set before and after image selection. The augmented training data after selection clearly have more distinguishable features than the ones without selection.}
\label{fig:tsne_cervical}
\end{figure*}
\begin{figure*}[!ht]
\begin{center}
\includegraphics[width=0.99\linewidth]{pcam.png}
\end{center}
\caption{t-SNE of the original and augmented PCam histopathology training set before and after image selection. While data augmentation without image selection increases the number of training samples, the original data distribution is distorted. After image selection, the original data distribution is recovered along with more number of data points. }
\label{fig:tsne_pcam}
\end{figure*}
\subsection{Implementation Details}
\subsubsection{HistoGAN Implementation}
The proposed HistoGAN model is trained in parallel on 4 NVIDIA TITAN Xp GPUs, each with 11G of RAM. We train HistoGAN with WGAN-GP~\cite{gulrajani2017improved} loss on the discriminators at all stages. Based on different sizes of images in the training set, we construct a 3-stage HistoGAN for cervical histopathology images and a 2-stage HistoGAN for the PCam lymph node histopathology images.
The input of the generator at the first stage is the concatenation of random noise and class label (\textit{e.g.}, CIN1-3, Normal) that are first one-hot encoded and then embedded by a transposed convolution layer. The first stage generator consists of $4$ up-sampling blocks with $3\times3$ conv kernels. Each block contains an upsample layer with bilinear interpolation followed by a combination of a convolutional layer with $3\times3$ kernel size. The output then goes through a conditional batch normalization \cite{de2017modulating} layer to modulate convolutional feature maps based on the corresponding assigned labels of the images generated. Blocks of the same architecture but different in and out channels are employed in generators of the next stages respectively, after a set of residual blocks.
Considering the future stages are learning the features from a more granularized level based on the output of the first stage, we employ self attention right after the first stage to facilitate the learning and focus on the desired features that are decisive for classification. Next, together with the real images from the original dataset with the same resolution, synthetic images of each scale are fed into corresponding stages of discriminators.
Inside each discriminator, the main structure contains several down-sampling layers with $4\times4$ conv kernels. Similar to the aforementioned blocks in the generator, class-conditional batch normalization are used after each convolutional layer to embed more class specific information. The down-sampling layers are followed by a $3\times3$ conv layer, a spectral normalization layer, a batch normalization layer, a Leaky ReLU activation layer, a minibatch discrimination~\cite{salimans2016improved} block for preventing mode collapse during GAN training, and a fully connected layer for the final output.
Regarding the hyperparameters, the HistoGAN model used for generating cervical histopathology and PCam images are trained with batch size set to $64$ for the cervical and $256$ for the PCam dataset for $1000$ training epochs with fixed learning rate $2e-4$. The parameter $\delta$ for WGAN-GP loss is set to $50$.
\subsubsection{Model and Image Selection Framework}
In the next step, GAN models at each epoch are saved after the $100^{th}$ epoch for model selection.
For reasons mentioned in Section~\ref{model_selection}, the feature extractor used for FID score calculation is the same as our baseline classifier (ResNet34), followed by EMA-based smoothing to accentuate the pattern of synthetic image quality trend during the GAN training process. The optimal GAN model weights selected for further stages of our purposed sample selection corresponds to the epoch with the lowest adjusted FID score. Next, we generate $4 rN_i$ synthetic images for each class $i$ with the chosen GAN, on which the same feature extractor is run for 5 times in order to extract the predicted probability from the softmax layer for entropy calculation, and also extract feature vectors after each residual block to obtain distance to centroids of ground truth. A dropout layer of rate $0.5$ is inserted before the last residual block right above the fully-connected layer of the feature extractor (ResNet34) for Monte Carlo sampling. Then the generated images are ranked based on the mean of entropy across 5 runs in ascending order, of which half images in each class are kept. The selected pool of synthetic images are further ranked based on the mean of cosine distance to the centroid that corresponds to the assigned label of each image over 5 runs also in ascending order. Similarly, half are filtered out, leaving the rest for the final augmentation.
\subsection{Results Analysis} \label{sec:results}
{
\subsubsection{Evaluation by Expert Pathologists}
To evaluate the quality of images generated by the proposed HistoGAN and validate the effectiveness of the selective synthetic augmentation method, we invited two pathologists to conduct expert evaluation on the cervical histopathology dataset. To prepare for the pathologist evaluation, we randomly chose 100 synthetic images where half of them are before selection and the other half are after selection. These images are then divided into 10 groups. Within each group of 10 images, there are two subgroups of 5 images where one subgroup is from the before-selection set and the other one is from the after-selection set. The 10 groups of images were then presented to the two pathologists who evaluated their quality independently. For each group, a pathologist was asked to choose one subgroup that has better quality, without knowing which subgroup corresponds to the one after selection; if the two subgroups were considered to have similar quality, the pathologist chose a tie. After the pathologists completed their evaluation, we compared their selected subgroups with the ground truth about which subgroups are from the after-selection image set. The comparison result shows that the two pathologists were able to differentiate before-selection subgroups from after-selection subgroups with high consistency: among the 10 groups, they chose the after-selection subgroup as having better quality 7 times, they chose a tie 2 times, and only once they chose the before-selection subgroup as having better quality. This evaluation result demonstrates that our image selection method is highly effective, since the expert pathologists consistently chose the after-selection images as having better quality.
Besides the group-level evaluation of our image selection method, the two pathologists also assessed the quality and realism of the individual synthetic images. They highlighted some realistic characteristics of the synthesized images, such as correct orientation, cell polarity, clear borders, and correct color of the cytoplasm. They also pointed out some unrealistic characteristics that repeatedly appeared in the generated image, such as smudged chromatin, missing nuclear details for large dark nuclei, and incorrect texture of large sheets of keratin. Despite the unrealistic aspects that they saw in the images, the pathologists actually view most of the images as containing meaningful features that make the images diagnosable. We are encouraged by these findings and plan to incorporate such expert knowledge in our future work to further improve our image synthesis model.
}
\subsubsection{Qualitative Evaluation}
The image synthesis results for cervical and lymph node datasets are demonstrated in Fig.~\ref{fig:cervical} and Fig.~\ref{fig:pcam}, respectively. In Fig.~\ref{fig:cervical}, we also show a comparison of synthetic images generated by our previous work \cite{xue2019synthetic} and by our proposed HistoGAN in this work. In both datasets, as we have already achieved promising image generation results, determining whether those samples can be used for data augmentation or not cannot be easily done by human observations. However, the discrepancy between images with and without selection is much more prominent in the feature space. In order to visualize such differences, after training a baseline ResNet34 classifier with the original training data, we use the pre-trained ResNet34 model as the feature extractor to extract features from the last convolutional layer in the ResNet model. We explore the distribution of training samples, including both original images and synthetic images, in the feature space using t-SNE~\cite{maaten2008visualizing}. In Fig.~\ref{fig:tsne_cervical}, without image selection, samples from different classes are
entangled together, introducing obscuring noise that disrupts the data distribution that real data presents. On the contrary, selected images have clearly more distinguishable features and can potentially help with improving the classification model performance. Similar phenomenon is also observed with more noticeable pattern in Fig.~\ref{fig:tsne_pcam}: while data augmentation without image selection increases the number of training samples, the original data distribution is distorted. After image selection, the original data distribution is recovered along with more number of data points.
The self attention mechanism (Sec. \ref{sec:architecture}) is a core improvement of our proposed HistoGAN model in this work as compared to the GAN model used in our previous work \cite{xue2019synthetic}. In order to examine the role of self attention, we visualize the conditional attention maps for images from different classes in Fig. \ref{fig:self_attn_map}. From the figure, one can see that HistoGAN with self attention successfully learns meaningful features by attending to important areas containing patterns most useful in distinguishing images of different disease grades.
\begin{table*}[t]
\begin{center}
\begin{tabular}{|l|c|c|c|c|}
\hline
{} & Accuracy & AUC & Sensitivity & Specificity \\
\hline
Baseline Model~\cite{he2016deep} & 0.754 $\pm$ 0.012 & 0.836 $\pm$ 0.008 & 0.589 $\pm$ 0.017 & 0.892 $\pm$ 0.005 \\
\hline
\enspace + Traditional Augmentation & 0.766 $\pm$ 0.013 & 0.844 $\pm$ 0.009 & 0.623 $\pm$ 0.029 & 0.891 $\pm$ 0.006 \\
\hline
\enspace + GAN Augmentation, r=0.5 & 0.787 $\pm$ 0.005 & 0.858 $\pm$ 0.003 & 0.690 $\pm$ 0.014 & 0.909 $\pm$ 0.003 \\
\hline
\enspace + Single Filtering~\cite{xue2019synthetic}$^{*}$, r=0.5 & 0.808 $\pm$ 0.005 & 0.872 $\pm$ 0.004 & 0.639 $\pm$ 0.015 & 0.912 $\pm$ 0.006 \\
\hline
\enspace + Selective Augmentation, r=0.5 & \textbf{0.821 $\pm$ 0.011} & \textbf{0.881 $\pm$ 0.007} & \textbf{0.671 $\pm$ 0.022} & \textbf{0.917 $\pm$ 0.005} \\
\hline
\end{tabular}
\end{center}
\caption{Classification results of baseline and augmentation models with different settings. Each model is run 5 times for the calculation of all evaluation metrics.
For fair comparison between ~\cite{xue2019synthetic} and our work, we reimplemented~\cite{xue2019synthetic} for it to use the same pool of synthetic images generated by HistoGAN.}
\label{tb:cervical}
\end{table*}
\begin{figure*}[t]
\begin{center}
\includegraphics[width=0.99\linewidth]{augment_ratio.png}
\end{center}
\caption{Classification results of the proposed selective synthetic augmentation with different augmentation ratios on the cervical dataset. $N$ in candidate pool sizes indicates the number of images in the original training dataset. For the same candidate pool size, selected images with different ratios are from the same pool. The error bar represents the standard deviation of classification accuracy from 5 multiple runs of each setting, the middle dot refers to the mean of 5 accuracy scores of the aforementioned multiple runs.}
\label{fig:ratio_comparison}
\end{figure*}
\subsubsection{Quantitative Comparisons}
We report quantitative evaluation scores between all baseline augmentation models and our models including the accuracy, area under the ROC curve (AUC), sensitivity and specificity to provide a comprehensive comparison. All models are run for $5$ rounds with random initialization for fair comparison. The mean and standard deviation results of the $5$ runs are reported.
In Table~\ref{tb:cervical}, we compare quantitative results with different baseline augmentation methods.
We use the same backbone ResNet34 classifier with same hyperparameters setting in all experiments to ensure differences only come from the augmentation mechanisms. Beyond the backbone baseline model ~\cite{he2016deep} without augmentation, we construct a baseline model with traditional augmentation including horizontal flipping and color jittering. Another baseline is GAN augmentation without selection where the training set is expanded by blindly adding GAN-generated images. We also compare the selective augmentation method proposed in this work with our prior work~\cite{xue2019synthetic}.
Since in this work we use HistoGAN, an improved cGAN model that generates better synthetic images (as shown in Fig.~\ref{fig:cervical}) than the cGAN model originally described in~\cite{xue2019synthetic}, we re-implemented~\cite{xue2019synthetic} to also use HistoGAN generated images, for fair comparison of the image selection algorithms.
From Table~\ref{tb:cervical}, one can see that the selective augmentation algorithm brings obvious benefits to all evaluation metrics, and our full model with augmentation ratio $r=0.5$ achieves best performance in all metrics. More specifically,
under $r=0.5$, our image selection method improves the classification result by nearly 2\% compared to the method in our prior work~\cite{xue2019synthetic}. This quantitative result demonstrates that our proposed selection method can better select high-quality images for augmentation than previous work.
To provide further insights on how the choice of the augmentation ratio $r$ affects augmentation performance, we also conduct an ablation study using different values of $r$, on different-sized candidate pools of HistoGAN-generated images. A summary of the ablation study is illustrated in Fig.~\ref{fig:ratio_comparison}. For this study, we generated synthetic image pools of four sizes: $2N$, $4N$, $6N$, and $8N$, where $N$ is the size of the original training set. On these pools, we tested different values of $r$, between $0.4$ and $1.0$. Each test is run for $5$ rounds, and the mean and standard deviation of the $5$ runs are reported. From the results shown in Fig.~\ref{fig:ratio_comparison}, one can see that either too small or too large a value of $r$ compromises the advantage of synthetic augmentation, and the best and most consistent performance gain is achieved at $r=0.5$. This observation is true for all four pools of different sizes. Our explanation for this phenomenon is related to the motivation behind using selective synthetic augmentation: the synthetic images have different levels of quality, and the number of images with good quality and meaningful diverse features generated by a trained GAN model is limited. While our sample selection can provide quality assurance, the total number of diverse, good images that provide complementary information to the existing training set is constrained by the GAN model and more relevantly, by the original labeled training data used to train the GAN model. Therefore, a larger pool of generated images does not always translate to more high-quality images that will be selected by our method, as shown by this ablation study. Once our selection method has chosen those good images generated by the particular GAN model, adding more images such as images that do not improve diversity but may contain artifacts or bad features would indeed add noise to the training set thus degrade performance.
Since our experiments show that the best augmentation performance is achieved at $r=0.5$, we use this value for all ours and other baseline models and all experiments on the PCam dataset.
In Table~\ref{tb:pcam_all}, we use 3\%, 5\%, 10\% and 20\% of the training data in PCam to simulate training sets with limited annotations and evaluate our models on the full testing set. Compared with the cervical dataset, the baseline classification model achieves higher accuracy on the reduced PCam dataset which makes it more difficult to further improve the performance.
However, our model still outperforms all baseline models using training sets of different sizes. For instance, when using 10\% of the entire dataset as training data, the classification accuracy improved by $1\%$ when using HistoGAN generated images for augmentation, without selection. After applying image selection, the accuracy is further improved by another $1.7\%$.
By conducting experiments on two histopathology image datasets and showing improved classification performances, we prove that our proposed HistoGAN model and synthetic augmentation algorithm are general and can be applied to various types of histopathology data.
\begin{table*}[t]
\begin{center}
\begin{tabular}{|l|l|c|c|c|c|}
\hline
{PCam} & Model & Accuracy & AUC & Sensitivity & Specificity \\
\hline
\multirow{2}{*}{ 3 \%}& Baseline & 0.872 $\pm$ 0.0030 &
0.914 $\pm$ 0.0019 &
0.826 $\pm$ 0.0080 &
0.903 $\pm$ 0.0030
\\
&\enspace + Selective Augmentation, r=0.5 & 0.900 $\pm$ 0.0024 &
0.933 $\pm$ 0.0016 &
0.865 $\pm$ 0.0060 &
0.924 $\pm$ 0.0030
\\
\hline
\multirow{2}{*}{ 5 \%}& Baseline & 0.893 $\pm$ 0.0006 &
0.929 $\pm$ 0.0004 &
0.863 $\pm$ 0.0010 &
0.913 $\pm$ 0.0020
\\
& \enspace + Selective Augmentation, r=0.5 & 0.917 $\pm$ 0.0033 &
0.945 $\pm$ 0.0022 &
0.892 $\pm$ 0.0040 &
0.935 $\pm$ 0.0040
\\
\hline
\multirow{2}{*}{10 \%}& Baseline & 0.910 $\pm$ 0.0012 &
0.940 $\pm$ 0.0009 &
0.883 $\pm$ 0.0050 &
0.929 $\pm$ 0.0030
\\
& \enspace + Traditional Augmentation & 0.916 $\pm$ 0.0102 &
0.944 $\pm$ 0.0067 &
0.893 $\pm$ 0.0140 &
0.933 $\pm$ 0.0090
\\
& \enspace + GAN Augmentation, r=0.5 & 0.920 $\pm$ 0.0020 &
0.947 $\pm$ 0.0014 &
0.898 $\pm$ 0.0050 &
0.935 $\pm$ 0.0020
\\
& \enspace + Selective Augmentation, r=0.5 & 0.937 $\pm$ 0.0011 &
0.958 $\pm$ 0.0007 &
0.916 $\pm$ 0.0070 &
0.951 $\pm$ 0.0040
\\
\hline
\multirow{2}{*}{20 \%}& Baseline & 0.932 $\pm$ 0.0014 &
0.955 $\pm$ 0.0010 &
0.909 $\pm$ 0.0080 &
0.948 $\pm$ 0.0040
\\
&+ Selective Augmentation, r=0.5 & 0.948 $\pm$ 0.0003 &
0.965 $\pm$ 0.0005 &
0.931 $\pm$ 0.0040 &
0.960 $\pm$ 0.0020
\\
\hline
\end{tabular}
\end{center}
\caption{
The performance of baseline and augmentation models using 3\%, 5\%, 10\% and 20\% of PCam as the training set. Each model is run 5 times for the calculation of all evaluation metrics. To further prove the effectiveness of our proposed selective augmentation, we compared the performance of our method and other augmentation methods when using 10\% of PCam as the training set. 10\% is chosen for demonstration because in this case the synthetic images show appealing visual quality as we can observe from Figure \ref{fig:pcam_full}, and consistently the classification performance presents improvement by a large margin.}
\label{tb:pcam_all}
\end{table*}
\section{Discussion}
Our proposed selective synthetic augmentation expands the training dataset by selectively adding synthetic images that do not distort the original data distribution, thus providing quality assurance in augmentation. The selected synthetic images are shown to improve the performance of automated image recognition systems with limited amount of manual annotation.
We believe our proposed method is applicable to other histopathology image recognition tasks with insufficient annotated data.
In addition, our proposed image selection algorithm is complementary to existing data augmentation methods, which further indicates the generality of our method.
While our selective synthetic augmentation significantly outperforms all baseline models, partial credits should go to the high-fidelity images generated by our proposed HistoGAN. However, the generated images are still not perfect, especially when viewed by expert pathologists, and we expect to further improve our GAN model with help from clinical experts.
Besides the visual quality of images, the diversity of images also plays a critical role in synthetic augmentation. Since synthetic augmentation is imperative in scenarios with very scarce training samples, combining our pipeline with a GAN model that can learn from limited data~\cite{wang2018transferring,luvcic2019high, noguchi2019image} would further improve the generality of our method. As we provide a solution to assure the synthetic image quality during augmentation, there is still room for improvement in selection mechanisms. More advanced methods for model selection and image selection, such as an end-to-end method and reinforcement learning based method, will be investigated in our future works.
\section{Conclusion}
In this paper, we design a new cGAN model termed HistoGAN for high-fidelity histopathology image synthesis and propose a synthetic augmentation method with quality assurance. By selectively adding realistic samples generated by HistoGAN into the original dataset, our method remarkably boosts the classification performance of baseline models. Experiments on two histopathology image datasets demonstrate the effectiveness and generality of our method.
\section*{Acknowledgments}
This research is supported in part by the Intramural Research
Program of the National Institutes of Health (NIH), National Library
of Medicine, and Lister Hill National Center for Biomedical
Communications.
We gratefully acknowledge the help with expert annotations from Dr. Rosemary Zuna of the University of Oklahoma Health Sciences Center. We also thank Dr. Joe Stanley of Missouri University of Science and Technology for making the cervical histopathology data collection available.
\bibliographystyle{model2-names.bst}\biboptions{numbers}
|
\section{Introduction}
The three-gluon vertex is one of the QCD fundamental Green's functions. This vertex allows the computation of the strong coupling constant and the measurement of a static potential between color charges. Herein we report on an upgrade of the lattice computation of this vertex performed by some of the authors in \cite{duarte2016, proc2016}.
The three-gluon correlation function $G^{a_1 a_2 a_3}_{\mu_1 \mu_2 \mu_3} (p_1, p_2, p_3)$ is given by
\begin{equation}
\langle A^{a_1}_{\mu_1} (p_1) \, A^{a_2}_{\mu_2} (p_2) \, A^{a_3}_{\mu_3} (p_3) \rangle = V \, \delta( p_1 + p_2 + p_3) ~
{G^{a_1 a_2 a_3}_{\mu_1 \mu_2 \mu_3} (p_1, p_2, p_3)}
\end{equation}
and can be written in terms of the gluon propagator $D^{ab}_{\mu\nu}(p^2)$ and the one-particle irreducible (1PI) vertex $\Gamma$ using
\begin{equation}
{G^{a_1a_2a_3}_{\mu_1\mu_2\mu_3} (p_1, p_2, p_3)} = D^{a_1b_1}_{\mu_1\nu_1}(p_1) ~ D^{a_2b_2}_{\mu_2\nu_2}(p_2) ~ D^{a_3b_3}_{\mu_3\nu_3}(p_3)
{\Gamma^{b_1b_2b_3}_{\nu_1\nu_2\nu_3} (p_1, p_2, p_3)} .
\end{equation}
Bose symmetry requires the 1PI vertex to be symmetric under permutations of any pair $(p_i, a_i, \mu_i)$. Given that
\begin{equation}
\Gamma^{a_1 a_2 a_3}_{\mu_1 \mu_2 \mu_3} (p_1, p_2, p_3) = f_{a_1 a_2 a_3} \Gamma_{\mu_1 \mu_2 \mu_3} (p_1, p_2, p_3)
\end{equation}
then the function $\Gamma_{\mu_1 \mu_2 \mu_3} (p_1, p_2, p_3)$ must be antisymmetric under the interchange of any pair $(p_i, \mu_i)$.
A complete description of $\Gamma_{\mu_1 \mu_2 \mu_3} (p_1, p_2, p_3)$ in the continuum requires six Lorentz invariant form factors, two associated to the transverse component $\Gamma^{(t)}$
and the remaining associated to the longitudinal $\Gamma^{(l)}$ \cite{ballchiu}.
\section{Asymmetric momentum configuration}
In this work we consider the computation of the three-gluon vertex in the asymmetric momentum configuration $p_2=0$, as in \cite{alles, duarte2016}. In this case, the correlation function can be written as
\begin{equation}
G_{\mu_1\mu_2\mu_3} (p, 0, -p) = V \frac{N_c(N^2_c-1)}{4} \left[D(p^2)\right]^2 \, D(0) \frac{\Gamma (p^2)}{3} ~ ~ p_{\mu_2} ~T_{\mu_1\mu_3} (p).
\end{equation}
The contraction of the Lorentz $\mu_1$ and $\mu_3$ indices, together with the contraction with the momentum $p_\alpha$, gives
\begin{equation}
G_{\mu \, \alpha \,\mu} (p, 0, -p) \, p_\alpha = V \frac{N_c(N^2_c-1)}{4}
\, \left[D(p^2)\right]^2 \, D(0) ~~\Gamma (p^2) ~~ p^2 .
\end{equation}
From this expression it is possible to extract the form factor $\Gamma (p^2)$. However, a lattice measurement of $\Gamma (p^2)$ requires the computation of the ratio
\begin{equation}
G_{\mu \alpha \mu} (p, 0, -p) p_\alpha / \left[D(p^2)\right]^2 \, D(0)
\end{equation}
and the extraction of $\Gamma (p^2)$ from this ratio will originate large statistical fluctuations at high momenta, where $D(p^2)$ becomes quite small. In fact, assuming Gaussian error propagation, it is possible to show that the statistical error on $\Gamma (p^2)$ behaves as $\Delta \Gamma(p^2) \sim p^4$ in the UV regime \cite{duarte2016}.
\section{Handling of noise, lattice artefacts}
In order to try to deal with the large statistical fluctuations at high momenta, we considered a few strategies \cite{guitese}:
\begin{itemize}
\item explore the ambiguity on the scale setting and perform a binning in the momentum --- all data points in each bin are replaced by a weighted average of the data points;
\item perform a $H(4)$ extrapolation of the lattice data \cite{becirevic1999, soto2009} --- such procedure is based on the remnant $H(4)$ symmetry group
associated with a hypercubic lattice. On the lattice, a scalar quantity $F$ is a function of the $H(4)$ invariants
\begin{displaymath}
p^2 = p^{[2]} = \sum_\mu p^2_\mu , \quad
p^{[4]} = \sum_\mu p^4_\mu , \quad
p^{[6]} = \sum_\mu p^6_\mu , \quad
p^{[8]} = \sum_\mu p^8_\mu ,
\end{displaymath}
i.e. $F_{Lat} = F(p^{[2]}, p^{[4]}, p^{[6]}, p^{[8]})$. The continuum limit will be given by $F(p^{[2]}, 0, 0, 0)$ up to corrections $\mathcal{O}(a^2)$. Having several data points for the same $p^2$ but different $p^{[4]}$, $p^{[6]}$, $p^{[8]}$, an extrapolation of $F_{Lat}$ to the continuum limit can be done, assuming that it can be written as a power series of the H(4) invariants. Note that, in this work, only a linear extrapolation in $p^{[4]}$ is considered.
\end{itemize}
\section{Lattice setup}
In this work we consider the $64^4$ ensemble of 2000 configurations already studied in \cite{duarte2016}, together with a $80^4$ ensemble of 1800 configurations, both generated with the Wilson gauge action at $\beta=6.0$. The rotation to the Landau gauge has been performed using the Fourier accelerated Steepest Descent method \cite{davies} implemented with the help of Chroma \cite{chroma} and PFFT \cite{pfft} libraries. The gluon field is computed using the definition
\begin{equation}
a g_0 A_\mu (x + a \hat{e}_\mu) = \frac{ U_\mu (x) - U^\dagger (x)}{ 2 i g_0}
- \frac{\mbox{Tr} \left[ U_\mu (x) - U^\dagger (x) \right]}{6 i g_0}
\end{equation}
with the momentum space gluon field given by
\begin{equation}
A_\mu (\hat{p}) = \sum_x e^{- i \hat{p} (x + a \hat{e}_\mu) } \, A_\mu (x + a \hat{e}_\mu) \,\,,\,\, \hat{p}_\mu = \frac{2 \, \pi \, n_\mu}{a \, L_\mu}.
\end{equation}
\section{Results}
In Figure \ref{binned} we compare the original and binned data for $\Gamma (p^2)$. The binning of the data suppresses the large statistical errors in the high momentum region and produces a well defined and smooth curve.
\begin{figure}[h]
\vspace{0.55cm}
\centering
\subfigure[$64^4$ lattice.]{ \includegraphics[width=0.42\textwidth]{plots/gamma_64x4.eps} \label{binn64}} \qquad
\subfigure[$80^4$ lattice.]{ \includegraphics[width=0.42\textwidth]{plots/gamma_80x4.eps} \label{binn80}}
\caption{Original and binned data for $\Gamma (p^2)$.}
\label{binned}
\end{figure}
Next, in Figure \ref{binnedoverp2} we compare the binned data for both lattices. The results of the two volumes agree within errors, suggesting that finite volume effects are small.
\begin{figure}[h]
\vspace{0.65cm}
\begin{center}
\includegraphics[width=0.6\textwidth]{plots/gamma_over_p2_compare.eps}
\end{center}
\caption{Comparison of binned data for $\Gamma (p^2)$.}
\label{binnedoverp2}
\end{figure}
In Figure \ref{H4extr} we compare the H(4) extrapolation of the $64^4$ lattice data with the binning of the original data. We observe that the H(4) extrapolation pushes the vertex to higher values in the high momentum regime. Nevertheless, in the infrared region, the extrapolated data is compatible with the original data, for both lattice volumes --- see Figure \ref{H4infra}.
\begin{figure}[h]
\vspace{0.65cm}
\begin{center}
\includegraphics[width=0.6\textwidth]{plots/all_gamma_over_p2_64_H4.eps}
\end{center}
\caption{Results of the H(4) extrapolation of $\Gamma (p^2)$ on the $64^4$ lattice volume.}
\label{H4extr}
\end{figure}
\begin{figure}[h]
\vspace{0.55cm}
\centering
\subfigure[$p^2 \Gamma(p^2)$.]{ \includegraphics[width=0.42\textwidth]{plots/all_gamma.eps} \label{H4infra-p2G}} \qquad
\subfigure[$\Gamma(p^2)$.]{ \includegraphics[width=0.42\textwidth]{plots/all_gamma_over_p2.eps} \label{H4infra-G}}
\caption{Original and H(4) data for both lattice volumes for low momenta.}
\label{H4infra}
\end{figure}
\section{Infrared behaviour of $\Gamma(p^2)$}
No zero crossing of $\Gamma(p^2)$, an indication of ghost dominance in the
infrared, is seen in the lattice data reported here. In order to check for
a change of sign in $\Gamma(p^2)$, in this section we explore the infrared
behaviour of the lattice $\Gamma(p^2)$, using the $80^4$ data for momenta below 1GeV, and fit the data to $\Gamma_1(p^2)=A + Z \ln(p^2)$ and $ \Gamma_2(p^2)=A + Z \ln(p^2+m^2)$. The first one is a typical ansatz considered in recent studies to study the zero crossing, see \cite{guitese} for details, and the second one is a variant of the first one which includes an infrared logarithmic regularizing mass.
In Figure \ref{zerocrossing} we plot the best fits of the lattice data for both fitting functions, obtained through the minimization of $\chi^2/d.o.f.$ .
For $\Gamma_1(p^2)$, we got $\chi^2/d.o.f. = 1.23$ with $A=0.2395(16)$ and $Z=0.0646(21)$. Accordingly the zero crossing occurs at $p_o=157$MeV.
For $\Gamma_2(p^2)$, the parameters take the values $A=0.208(24)$, $Z=0.124(27)$ and $m=0.61(15)$GeV, with a $\chi^2/d.o.f. = 0.95$. As shown in the right plot of Figure \ref{zerocrossing}, in this case there is no zero crossing.
\begin{figure}[h]
\vspace{0.55cm}
\centering
\subfigure[$\Gamma (p^2) = A + Z \ln(p^2)$.]{ \includegraphics[width=0.42\textwidth]{plots/gamma80-fit1.eps} \label{zerocrossing-fit1}} \qquad
\subfigure[$\Gamma (p^2) = A + Z \ln(p^2+m^2)$.]{ \includegraphics[width=0.42\textwidth]{plots/gamma80-fit2.eps} \label{zerocrossing-fit2}}
\caption{Infrared $80^4$ lattice data for $\Gamma(p^2)$ together with some fitting functions. }
\label{zerocrossing}
\end{figure}
\section{Conclusions and outlook}
In this paper we describe an improved calculation of the three gluon vertex on the lattice, for the asymmetric momentum configuration. We use two different lattice volumes $(6.5$ fm$)^4$ and $(8.2$ fm$)^4$, with a common lattice spacing of $a = 0.102$ fm. We show that a H(4) extrapolation of the lattice data pushes the vertex to higher values in UV regime. We proceed with a functional study in the infrared region, considering some functional forms compatible with zero crossing and IR divergence.
Further momentum configurations will be explored in the near future.
\acknowledgments
This work was supported by national funds from FCT
Fundação para a Ciência e a Tecnologia, I. P., within the
Projects UIDB/04564/2020, UIDP/04564/2020, and CERN/FIS-COM/0029/2017.
G. T. R. C. acknowledges financial support from FCT
under Project UIDB/04564/2020, and also from the Generalitat Valenciana
(genT program CIDEGENT/2019/040) and Ministerio de Ciencia e
Innovacion PID2020-113644GB-I00.
P. J. S. acknowledges financial support from FCT
under Contract CEECIND/00488/2017.
This work was granted access to the HPC resources of
the PDC Center for High Performance Computing at the
KTH Royal Institute of Technology, Sweden, made
available within the Distributed European Computing
Initiative by the PRACE-2IP, receiving funding from the
European Communitys Seventh Framework Programme
(FP7/2007-2013) under Grant agreement no. RI-283493.
The use of Lindgren has been provided under DECI-9
project COIMBRALATT. We acknowledge that the results
of this research have been achieved using the PRACE-3IP
project (FP7 RI312763) resource Sisu based in Finland at
CSC. The use of Sisu has been provided under DECI-12
project COIMBRALATT2. We acknowledge the
Laboratory for Advanced Computing at the University of
Coimbra \cite{lca} for providing access to the HPC resource
Navigator. The authors acknowledge Minho Advanced Computing Center
\cite{macc} for providing HPC resources that have contributed to
the research results reported within this paper. This work was
produced with the support of MACC and it was funded by FCT I.P.
under the Advanced Computing Project CPCA/A2/6816/2020, platform Bob.
This work was produced with the support of INCD \cite{incd} funded by FCT and
FEDER under the project 01/SAICT/2016 nº 022153.
|
\section{Introduction}
\label{sec:intro}
In the sparsest cut problem, we are given a graph together with costs and demands on the edges, and our goal is to find a cut that minimizes the ratio between the costs and demands across the cut.
Sparsest cut is among the most fundamental optimization problems that has attracted interests from both computer scientists and mathematicians. Since the problem is $\mathsf{NP}$-hard~\cite{MatulaS90}, the focus has been to study approximation algorithms for the problem.
Over the past four decades, several breakthrough results have eventually culminated in a factor-$\tilde O(\sqrt{\log n})$ approximation in polynomial time~\cite{arora2008euclidean,arora2009expander,leighton1999multicommodity}.
On the lower bound side, the problem is $\mathsf{APX}$-hard~\cite{chuzhoy2009polynomial} and, assuming the Unique Games Conjecture, does not admit any constant-factor approximation in polynomial time~\cite{chawla2006hardness}.
The extensive interest in sparsest cuts stems from both applications and mathematical reasons.
From the point of view of applications, the question of partitioning the universe into two parts while minimizing the ``loss'' across the interface is crucial in any divide-and-conquer approach e.g., in image segmentation.
From a mathematical/geometric viewpoint, the integrality gap of convex relaxations for sparsest cuts is equivalent to the embeddability of any finite metric space (for LP relaxation) and of any negative-type metric (for SDP relaxation)\footnote{A metric $(X,d)$ is said to be \textit{negative type}, if $(X,\sqrt{d})$ embeds isometrically into a Hilbert space.} into $\ell_1$.
Therefore, it is not a surprise that this problem has attracted interest from both computer science and mathematics (geometry, combinatorics, and functional analysis) communities.
The study of sparsest cuts in the low-treewidth regime was initiated in 2010 by Chlamt{\'a}{\v{c}}\xspace, Krauthgamer, and Raghavendra~\cite{ckr10}, who devised a factor-$2^{2^k}$ approximation algorithm (CKR) that runs in time $2^{O(k)} \cdot \operatorname{poly}(n)$, with $k$ being the treewidth of the input graph.
Later, Gupta, Talwar and Witmer~\cite{gtw13} showed how to obtain a factor-$2$ approximation (GTW) with a blown-up run time of $n^{O(k)}$; they further showed that there is no $(2-\varepsilon)$-approximation for any $\varepsilon > 0$ on constant-treewidth graphs, assuming the Unique Games Conjecture.
It remains an intriguing open question whether one can simultaneously achieve the best run time and approximation factor.
In particular, in this paper we address the following question:
\begin{quote}
Does \mbox{\sf Sparsest-Cut}\xspace admit a factor-$2$ approximation that runs in time $2^{O(k)} \cdot \operatorname{poly} (n)$?
\end{quote}
\paragraph{Broader perspectives.}
Given the significance of sparsest cuts, a lot of effort have been invested into understanding when sparsest cut instances are ``easy''.
In trees, optimal sparsest cuts can be found in polynomial time~(see e.g.~\cite{GuptaL19}).
For many other well-known graph classes, finding optimal sparsest cuts is $\mathsf{NP}$-hard, so researchers attempted to find constant-factor approximations in polynomial time.
They have succeeded, over the past two decades, for several classes of graphs, such as outerplanar, $\ell$-outerplanar, bounded-pathwidth and bounded-treewidth graphs~\cite{gtw13,gupta2004cuts,chekuri2006embedding,ckr10,lee2013pathwidth}, as well as planar graphs~\cite{cohen2021quasipolynomial}.
As mentioned earlier, sparsest cuts are not only interesting from the perspective of algorithm design, but also from the perspectives of geometry, probability theory and convex optimization.
Indeed, the famous conjecture of Gupta, Newman, Rabinovich, and Sinclair~\cite{gupta2004cuts} postulates that any minor-free graph metric embeds into $\ell_1$ with a constant distortion, which would imply that all such graphs admit a constant approximation for the sparsest cut problem.
The conjecture has been verified in various graph classes~\cite{lee2013pathwidth,chekuri2006embedding}, but remains open even for bounded-treewidth graph families.
To us, perhaps the most interesting aspect of the treewidth parameter~\cite{ckr10,gtw13} is its connection to the power of hierarchies of increasingly tight convex relaxations (see, for instance, the work by Laurent~\cite{laurent03}).
In this setting, a straightforward (problem-independent!) LP rounding algorithm performs surprisingly well for many ``combinatorial optimization'' problems.
It has been shown to achieve optimal solutions for various fundamental problems in bounded treewidth graphs~\cite{magen2009robust,bienstock2004tree,wainwright2004treewidth} and match the (tight) approximation factors achievable on trees for problems such as {group Steiner tree}~\cite{chalermsook2017beyond,chalermsook2018survivable,garg2000polylogarithmic,halperin2003polylogarithmic}.
In this way, for these aforementioned problems, such a problem-oblivious LP rounding algorithm provides a natural framework to generalize an optimal algorithm on trees to nearly-optimal ones on low (perhaps super-constant) treewidth graphs.
Our work can be seen as trying to develop such understanding in the context of the sparsest cut problem.
\subsection{Our Results}
We present several results that may be seen as an intermediate step towards the optimal result.
Our main technical results are summarized in the following theorem.
\begin{theorem}
\label{thm:main-intro}
For the following functions $t$ and $\alpha$, there are algorithms that run in time $t(k)\cdot \operatorname{poly}(n)$ and achieve approximation factors $\alpha(k)$ for the sparsest cut problem:
\begin{itemize}
\item $t(k) = 2^{O(k)}$ and $\alpha(k) = O(k^2)$.
\item $t(k) = 2^{O(k^2)}$ and $\alpha(k) = O(1)$.
\item For any $\varepsilon > 0$, $t(k) = \exp\paren{O(\frac{k^{1+\varepsilon}}{\varepsilon})}$ and $\alpha(k) = O(1/\varepsilon^2)$.
\end{itemize}
\end{theorem}
Our first result directly improves the approximation factor of $2^{2^k}$ by \chlamtac et al.\@\xspace, while keeping the run time single-exponential in $k$.
Our second result shows that, with only slightly more exponential run time, one can achieve a constant approximation factor.
Compared to Gupta et al., our result has a constant blowup in the approximation factor (but independent of $k$), but has a much better run time ($2^{O(k^2)}$ instead of $n^{O(k)}$); compared to \chlamtac et al.\@\xspace, our result has a much better approximation factor ($O(1)$ instead of $2^{2^k}$), while maintaining nearly the same asymptotic run time.
Finally, our third result gives us an ``approximation scheme'' whose run time exponent converges to a single exponential, while keeping an approximation factor a constant.
We remark that, by plugging in $\varepsilon = \Omega(1/\log k)$, we obtain a factor-$O(\log^2 k)$ approximation in time $k^{O(k)}\cdot\operatorname{poly}(n)$.
\subsection{Overview of Techniques}
Now, we sketch the main ideas used in deriving our results.
We assume certain familiarity with the notions of treewidth and tree decomposition.
Let $G$ be a graph with treewidth $k$ and ${\mathcal T}$ be a tree decomposition of $G$ with a collection of bags $B_t \subseteq V(G)$ for all $t \in V(G)$.
Define the width of ${\mathcal T}$ as $w({\mathcal T}) = \max_{t \in V({\mathcal T})} |B_t| -1$.
The run time of algorithms that deal with the treewidth parameter generally depend on $w({\mathcal T})$, so when designing an algorithm in low-treewidth graphs, one usually starts with a near-optimal tree decomposition in the sense that $w({\mathcal T}) = O(k)$.
To give a concrete example, the CKR algorithm~\cite{ckr10} for sparsest cut runs in time $2^{O(w({\mathcal T}))} \cdot \operatorname{poly}(n)$ and gives approximation factor $2^{2^{w({\mathcal T})}}$ .
Observe that, with slightly higher width $w({\mathcal T}) = O(\log n + \beta(k))$, the CKR algorithm would run in time $2^{\beta(k)} \cdot \operatorname{poly} (n)$.
Our results are obtained via the concept of {\bf combinatorial diameter} of a tree decomposition.
Informally, the combinatorial length between $u$ and $v$ in ${\mathcal T}$ measures the number of ``non-redundant bags'' that lie on the unique path in ${\mathcal T}$ connecting the bags of $u$ and $v$.
We say that the combinatorial diameter $\Delta({\mathcal T})$ of ${\mathcal T}$ is at most $d$ if the combinatorial length of every pair of vertices is at most $d$. Please refer to \Cref{sec:combinatorial diameter} for formal definitions.
Our first key technical observation shows that the approximation factor of the CKR algorithm can be upper bounded in terms of the combinatorial diameter $\min \{ O(\Delta({\mathcal T})^2), 2^{2^{w({\mathcal T})}}\}$. Moreover, in the special case of $\Delta({\mathcal T}) = 1$, the CKR algorithm gives a $2$-approximation, which can be seen by using the arguments of Gupta et al.~\cite{gtw13}.
Therefore, to obtain a fast algorithm with a good approximation factor, it suffices to prove the existence of a tree decomposition with simultaneously low $w({\mathcal T})$ and low $\Delta({\mathcal T})$.
We remark that standard tree decomposition algorithms~\cite{bodlaender2016c} give us $w({\mathcal T}) = O(k)$ and $\Delta({\mathcal T}) = O(\log n)$, so this observation alone does not immediately lead to improved algorithmic results.
However, it allows us to view the results from CKR~\cite{ckr10} and GTW~\cite{gtw13} in the same context: CKR applies the algorithm to the tree decomposition ${\mathcal T}_{CKR}$ with $\Delta({\mathcal T}_{CKR}) = O(\log n)$ and $w({\mathcal T}_{CKR}) = O(k)$, while GTW applies the same algorithm with $\Delta({\mathcal T}_{GTW}) = 1$ and $w({\mathcal T}_{GTW}) = O(k \log n)$.
In other words, the same algorithm is applied to two different ways of decomposing the input graph~$G$ into a tree.
In this paper, we present several new tree decomposition algorithms that optimize the tradeoff between $w({\mathcal T})$ and $\Delta({\mathcal T})$.
Our first algorithm gives a tree decomposition ${\mathcal T}_1$ with $w({\mathcal T}_1) = {O(\log n+k)}$ and $\Delta({\mathcal T}_1) = O(k)$, which leads to a factor-$O(k^2)$ approximation in time $2^{O(k)} \cdot \operatorname{poly}(n)$; this directly improves the approximation factor of CKR while maintaining the same asymptotic run time.
Our second algorithm gives the tree ${\mathcal T}_2$ with $w({\mathcal T}_2) = O(\log n + k^2)$ and $\Delta({\mathcal T}_2) = 4$.
This leads to an algorithm for sparsest cut with run time $2^{O(k^2)} \cdot \operatorname{poly}(n)$ and approximation factor $O(1)$.
Our third algorithm is an approximation scheme which is further parameterized by $\varepsilon >0$.
In particular, for any $\varepsilon>0$, we construct the tree ${\mathcal T}_{3,\varepsilon}$ such that $w({\mathcal T}_{3,\varepsilon}) = O(\log n+ k^{1+\varepsilon}/\varepsilon)$ and $\Delta({\mathcal T}_{3,\varepsilon}) = O(1/\varepsilon)$.
\subsection{Conclusion \& Open Problems}
Our work is an attempt to simultaneously obtain the best run time and approximation factor for sparsest cut in the low-treewidth regime.
Our research question combines the flavors of two very active research areas, namely parameterized complexity and approximation algorithms.
We introduce a new measure of tree decomposition called combinatorial diameter and show various constructions with different tradeoffs between $w({\mathcal T})$ and $\Delta({\mathcal T})$.
We leave the question of getting $2$-approximation in $2^{O(k)} \cdot \operatorname{poly}(n)$ time as the main open problem.
One way to design such an algorithm is to show an existence of a tree decomposition with $w({\mathcal T}) = O(\log n +k)$ and $\Delta({\mathcal T}) = 2$.
An interesting intermediate step would be to show $w({\mathcal T}) = O(\log n +f(k))$ for some function $f$ and $\Delta({\mathcal T}) = 2$, which would imply a fixed-parameter algorithm that yields a $2$-approximation.
Another interesting question is to focus on polynomial-time algorithms and optimize the approximation factor with respect to treewidth.
In particular, is there an $O(\log^{O(1)} k)$ approximation in polynomial time?
This question is open even for the \emph{uniform} sparsest cut problem (unit demand for every vertex pair), for which a fixed-parameter algorithm~\cite{BonsmaBPP2012} but no polynomial-time algorithm is known.
A broader direction that would perhaps complement the study along these lines is to improve our understanding on a natural LP-rounding algorithm on the lift-and-project convex programs in general.
For instance, can we prove a similar tradeoff result for other combinatorial optimization problems in this setting?
One candidate problem is the {\em group Steiner tree} problem, for which a factor-$O(\log^2 n)$ approximation in time $n^{O(k)}$ is known (and the algorithm there is ``the same'' algorithm as used for finding sparsest cuts).
Can we get a factor-$O(\log^2 n)$ approximation in time $2^{O(k)} \cdot \operatorname{poly}(n)$?
\paragraph{Independent Work:} Independent of our work, Cohen-Addad, M\"{o}mke, and Verdugo~\cite{tobias} obtained a $2$-approximation algorithm for sparsest cut in treewidth $k$ graph with running time $2^{2^{O(k)}} \cdot \text{poly}(n)$.
Observe that their result is incomparable with our result: they obtain a better approximation factor, whereas the obtained running time is considerably larger than ours.
Similar to our result, they build on the techniques from \cite{ckr10,gtw13}.
\section{Preliminaries}
\paragraph{Problem Definition} In the \mbox{\sf Sparsest-Cut}\xspace problem (with general demands), the input is a graph $G=(V, E_G)$ with positive edge capacities $\left \{\cp_e \right \}_{e \in E_G}$ and a demand graph $D =(V, E_D)$ (on the same set of vertices) with positive demand values $\left \{\dm_e \right \}_{e \in E_D}$. The aim is to determine
\[
\Phi_{G, D} := \min_{S \subseteq V} \Phi_{G,D}(S), \quad\quad \Phi_{G,D}(S) := \frac{\sum_{e \in E_G(S, V-S)}\cp_e}{\sum_{e \in E_D(S, V-S)} \dm_e}.\]
The value $\Phi_{G,D}(S)$ is called the \emph{sparsity} of the cut $S$.
\paragraph{Tree decomposition}
Let $G=(V, E)$ be a graph. A tree decomposition $({\mathcal T}, \set{B_t}_{t \in V({\mathcal T})})$ of $G$ is a tree ${\mathcal T}$ together with a collection of \emph{bags} $\{B_t\}_{t \in V({\mathcal T})}$, where the bags $B_t \subseteq V(G)$ satisfy the following properties:
\begin{itemize}
\item $V(G) = \bigcup_t B_t$.
\item For any edge $uv \in E(G)$, there is a bag $B_t$ containing both $u$ and $v$.
\item For each vertex $v \in V(G)$, the collection of bags that contain $v$ induces a connected subgraph of ${\mathcal T}$.
\end{itemize}
The treewidth of graph $G$ is defined as the minimum integer $k$ such that there exists a tree decomposition where each bag contains at most $k+1$ vertices.
We generally use $r$ to denote the root of ${\mathcal T}$, and $p\colon V({\mathcal T}) \to V({\mathcal T})$ for the parent of a node with respect to root $r$. %
We sometimes refer to $B_{p(i)}$ as the \emph{parent bag} of $B_i$. %
We denote by ${\mathcal T}_{i \leftrightarrow j}$
the set of nodes on the unique path in tree ${\mathcal T}$ between nodes $i, j \in
V({\mathcal T})$ (possibly $i=j$). %
For a set $X \subseteq V({\mathcal T})$ of bags, we use the shorthand $B(X) = \bigcup_{i \in X} B_i$ (the union of bags for nodes in $X$).
We will treat cuts in a graph as assignments of $\{0,1\}$ to each vertex, and fix some corresponding notation.
\begin{definition}
Let $X$ be some finite set.
An \emph{$X$-assignment} is a map $f\colon X \to \{0,1\}$.
We denote by~$\mathcal{F}[X]$ the set of all $X$-assignments.
For some distribution $\mu$ over $\mathcal{F}[X]$ and set $Y \subseteq X$ we define~$\mu|_Y$ to be the distribution given by
\[
\Pr_{f \sim \mu|_Y}[f = f'] = \Pr_{f\sim \mu}[f|_Y = f'] \quad \forall f' \in \mathcal{F}[Y] \enspace .
\]
\end{definition}
\section{Algorithm and Combinatorial Diameter}
Our approach is based on the new relation between the algorithm of \chlamtac et al.\@\xspace~\cite{ckr10} and our novel notion of ``combinatorial diameter''.
In \Cref{sec:combinatorial diameter}, we present the definition of the combinatorial diameter.
The subsequent sections give the description of \chlamtac et al.\@\xspace and prove the relation to the combinatorial diameter.
\subsection{Our New Concept: Combinatorial Diameter}
\label{sec:combinatorial diameter}
\begin{definition}[Redundant bags]
Fix $s,t \in V({\mathcal T})$.
Let $v \in V({\mathcal T})$ be a bag with exactly two neighbors $u$ and $w$ on the path ${\mathcal T}_{s \leftrightarrow t}$. When $B_v \cap B_w \subseteq B_u$, we say that $v$ is \emph{$(s,t)$-redundant}.
\end{definition}
Intuitively, each node $v$ discarded in the fashion above can be thought of as a subset of $u$, since the vertices $B_v \setminus B_u$ occur only in $B_v$ within ${\mathcal T}_{s \leftrightarrow t}$. %
As a consequence, we can show that they do not affect the rounding behaviour of the CKR algorithm with respect to $s$ and $t$ (therefore ``redundant'').
\begin{definition}[Simplification]
\label{def:simplification}
Let ${\mathcal T}$ be a tree decomposition, and $s,t \in V({\mathcal T})$. We say ${\mathcal T}_{s \leftrightarrow t}$ has \emph{combinatorial length} at most $\ell$ if it can be reduced to a path of length at most $\ell$ by repeatedly applying the following rule:
\begin{itemize}
\item[] Delete an $(s,t)$-redundant node $v$ on path ${\mathcal T}_{s \leftrightarrow t}$, and add the edge~$\set{u,w}$. %
We call this operation \emph{bypassing} $v$.
\end{itemize}
We call any path $P$ generated from ${\mathcal T}_{s\leftrightarrow t}$ in this fashion a \emph{simplification} of ${\mathcal T}_{s \leftrightarrow t}$.
\end{definition}
\begin{definition}[Combinatorial diameter]
The \emph{combinatorial diameter} of ${\mathcal T}$ is defined to be the minimum $\ensuremath{\delta}\xspace$ such that, for all $u,v$, the path ${\mathcal T}_{u \leftrightarrow v}$ has combinatorial length at most $\ensuremath{\delta}\xspace$.
\end{definition}
\subsection{Algorithm Description and Overview}
For completeness, we restate the essential aspects of the algorithm by \chlamtac et al.\@\xspace~\cite{ckr10}.
The algorithm is initially provided a \mbox{\sf Sparsest-Cut}\xspace instance $(G,D,\cp,\dm)$ alongside a tree decomposition ${\mathcal T}$ of $G$ with the width $w({\mathcal T}) = \max_{t} |B_t| - 1$.
The goal is then to compute a cut in $G$ that has low sparsity.
The algorithm starts by computing, for every vertex set $L = B_i \cup \set{s,t}$, consisting of a bag $B_i$ and a pair of vertices $s,t \in V(G)$, a distribution $\mu_L$ over $L$-assignments. %
This collection of distributions for all sets $L$ satisfies the requirement that any two distributions agree on their joint domains, i.e.~$\mu_L|_{L \cap L'} = \mu_{L'}|_{L\cap L'}$ for each pair of sets $L, L'$ with the structure above. %
If we denote $\operatorname{\sf lpcut}\xspace(s,t) = \Pr_{f \sim \mu_{B \cup \{s,t\}}}[f(s) \neq f(t)]$ for any $s,t \in V(G)$, and an arbitrary bag $B$ of~${\mathcal T}$, we can compute the collection of distributions that minimizes
\[
\dfrac{\sum_{\{s,t\} \in E_G}\cp_{\{s,t\}} \cdot \operatorname{\sf lpcut}\xspace(s,t) }{\sum_{\{s,t\} \in E_D}\dm_{\{s,t\}} \cdot \operatorname{\sf lpcut}\xspace(s,t) } \enspace .
\]
Notice that $\operatorname{\sf lpcut}\xspace$ is well-defined by the consistency requirement, since the choice of $B$ does not impact the distribution over $\{s,t\}$-assignments.
For ease of notation, we will refer to the implied distribution over some vertex set $X \subseteq B \cup \{s,t\}$ by $\mu_X$, where formally $\mu_X = \mu_{B \cup \{s,t\}}|_{X}$.
Such a collection of distributions can be computed in time $2^{O(w({\mathcal T}))}\operatorname{poly}(n)$, using Sherali-Adams LP hierarchies, which motivates the function name $\operatorname{\sf lpcut}\xspace$.
It is then rounded to some $V(G)$-assignment $f$ using \autoref{alg:Chlamtac}.
We now recall a number of useful results about the algorithm and the assignment it computes.
Details about the algorithm and the attendant lemmas can be found in the work of \chlamtac et al.\@\xspace~\cite{ckr10}.
Denote by ${\mathcal A}$ the distribution over $V(G)$-assignments produced by the algorithm.
\begin{algorithm}[t]
\SetAlgoLined
\KwData{$G, ({\mathcal T}, \set{B_i}_{i \in V({\mathcal T})}), \{\mu_L\}$}
Start at any bag $B_0$, sample $f|_{B_0}$ from $\mu_{B_0}$\;
We process the bags in non-decreasing order of distance from $B_0$ \;
\ForEach{Bag $B$ with a processed parent bag $B'$}{
Let $B^+ = B\cap B'$ the subset of $B$ on which $f$ is fixed.
Let $B^- := B\setminus B^+$.
Sample $f|_{B^-}$ according to
\[
\Pr[f|_{B^-} = f'] = \Pr_{f^* \sim \mu_B}[f^*|_{B^-} = f'\; \mid \; f^*|_{B^+} =f|_{B^+} ] \quad\forall f' \in \mathcal{F}[B^-]
\]
}
\KwResult{$f$}
\caption{Algorithm \textsc{SC-Round} }
\label{alg:Chlamtac}
\end{algorithm}
\begin{lemma}[\cite{ckr10}, Lemma 3.3]
\label{lem:bagRealisation}
For every bag $B$ the assignment $f|_B$ computed by \autoref{alg:Chlamtac} is distributed according to $\mu_B$, meaning $ \Pr_{ f \sim {\mathcal A}}[f|_B = f'] = \Pr_{f^* \sim \mu_B}[f^* = f']$ for all $f' \in \mathcal{F}[B]$.
\end{lemma}
A direct consequence of this lemma is the fact that any edge $\{s,t\}$ of $G$ is cut by the algorithm with probability $\operatorname{\sf lpcut}\xspace(s,t)$.
In particular, the expected capacity of the rounded cut is therefore
\[
\sum_{\{s,t\} \in E_G}\cp_{\{s,t\}} \cdot \operatorname{\sf lpcut}\xspace(s,t),
\]
which is the value ``predicted'' by the distribution $\mu_L$.
The same property does not hold for the (demand) edges of $D$ since they may not be contained in any bag of ${\mathcal T}$.
Denote by $\operatorname{\sf algcut}\xspace(s,t)$ the probability that the algorithm separates $s$ and $t$, that is, $\operatorname{\sf algcut}\xspace(s,t) = \Pr_{f \sim {\mathcal A}}[f(s) \neq f(t)]$.
We would like to lower bound $\operatorname{\sf algcut}\xspace(s,t) \geq c \; \operatorname{\sf lpcut}\xspace(s,t)$ for all demand edges $\{s,t\}$ and some value $c > 0$.
This would imply that the expected demand of the rounded cut is at least $c\sum_{\{s,t\} \in E_D}\dm_{\{s,t\}} \operatorname{\sf lpcut}\xspace(s,t)$, and having a good expected demand and capacity is sufficient for computing a good solution by the following observation.
\begin{observation}[\cite{ckr10}, Remark 4.3]
\label{obs:derandomisation}
The cut sparsity $\alpha$ predicted by distributions $\set{\mu_L}_L$ is
\[
\alpha := \dfrac{\sum_{\{s,t\} \in E_G}\cp_{\{s,t\}} \cdot \operatorname{\sf lpcut}\xspace(s,t) }{\sum_{\{s,t\} \in E_D}\dm_{\{s,t\}} \cdot \operatorname{\sf lpcut}\xspace(s,t) } \enspace .
\]
Then if $\operatorname{\sf algcut}\xspace(s,t) \geq c \cdot \operatorname{\sf lpcut}\xspace(s,t)$ for all $\{s,t\} \in E_D$ and $\operatorname{\sf algcut}\xspace(s,t) = \operatorname{\sf lpcut}\xspace(s,t)$ for $\{s,t\} \in E_G$, we have
\[
\mathbb{E}_{f \sim {\mathcal A}}\left[\sum_{\{s,t\} \in E_G}\cp_{\{s,t\}} |f(s) - f(t)| - \frac{\alpha}{c}\sum_{\{s,t\} \in E_D}\dm_{\{s,t\}} |f(s) - f(t)|\right] \leq 0 \enspace.
\]
A solution is $c$-approximate if the value in the expectation above is non-positive, and such a solution can either be obtained by repeated rounding or by derandomization
using the method of conditional expectations, without increasing the asymptotic run time.
\end{observation}
This observation implies that the bottleneck to obtaining a good approximation factor is the extent to which our rounding algorithm can approximate the marginal of $\mu_L$ on the individual edges of $D$.
Our main result relates this marginal to the combinatorial diameter of ${\mathcal T}$. It can now be stated as follows:
\begin{theorem}
\label{thm:algo:main}
Let $(G,D,\cp,\dm)$ be an instance of\ \,\mbox{\sf Sparsest-Cut}\xspace, and $({\mathcal T}, \set{B_i}_i)$ a tree decomposition of $G$ with width $w({\mathcal T})$ and combinatorial diameter $\Delta({\mathcal T})$.
Then {\sc SC-ROUND} satisfies $\operatorname{\sf algcut}\xspace(s,t) \geq \Omega\paren{\frac{1}{\Delta({\mathcal T})^2}} \cdot \operatorname{\sf lpcut}\xspace(s,t)$ for every $\{s,t\}\in E_D$. Therefore, we have a factor-$O(\Delta({\mathcal T})^2)$ approximation for sparsest cut with run time $2^{O(w({\mathcal T}))}\cdot \operatorname{poly}(n)$.
\end{theorem}
The rest of this section is devoted to proving this theorem.
\subsection{Step 1: Reduction to Short Path}
In this section, we show that when the combinatorial diameter of the tree decomposition is $\ensuremath{\delta}\xspace = \Delta({\mathcal T})$, the analysis can be reduced to the case of a path decomposition of length $\ensuremath{\delta}\xspace$.
We employ the following lemma to simplify our analysis of the behavior of the algorithm.
\begin{lemma}[\cite{ckr10}, Lemma 3.4]
\label{lem:traverselInv}
The distribution over the assignments $f$ is invariant under any connected traversal of ${\mathcal T}$, i.e. the order in which bags are processed does not matter, as long as they have a previously processed neighbor.
The choice of the first bag $B_0$ also does not impact the distribution.
\end{lemma}
Let $\set{s,t} \in E_D$ be a demand edge. %
If $s$ and $t$ are contained in a common bag, then $\operatorname{\sf algcut}\xspace(s,t) = \operatorname{\sf lpcut}\xspace(s,t)$ by \Cref{lem:bagRealisation} and we are done; therefore, we
assume that there is no bag containing both $s$ and $t$. %
We want to estimate the probability that $s$ and $t$ separated by the algorithm, that is, the probability that $f(s) \neq f(t)$.
The lemma above allows us to reduce to the case in which the algorithm first rounds a bag $B_1$ containing $s$, then rounds bags $B_2, \dots, B_{\ell-1}$ along the path to a bag $B_{\ell}$ containing $t$, and finally $B_{\ell}$.
At this point the algorithm has already assigned $f(s)$ and $f(t)$, so the remaining bags of ${\mathcal T}$ can be rounded in any connected order without impacting the separation probability.
Hence, it is sufficient to characterize the behavior of the rounding algorithm along paths in ${\mathcal T}$.
Let $P$ be the shortest path connecting a bag containing $s$ to a bag containing $t$; %
denote such path by $P = v_1 v_2 \ldots v_{\ell}$ such that $s \in B_{v_1}$ and $t \in B_{v_{\ell}}$. %
By \Cref{lem:traverselInv} we can assume that the algorithm first processes $B_{v_1}$, and then all other bags $B_{v_2}, \dots, B_{v_\ell}$, in
this order.
Observe that, except for $v_1$ and $v_{\ell}$, no other bag of $P$ contains $s$ or $t$.
We repeatedly apply the reduction rule from \Cref{def:simplification} until the resulting path has length at most $\ensuremath{\delta}\xspace$.
The following lemma asserts that the distribution of the algorithm is preserved under this reduction rule.
We slightly abuse the notation and denote by ${\mathcal A}$ the distribution of our algorithm on path $P$ starting from $v_1$.
\begin{lemma}
\label{lem:distpreservereduction}
Let $u,v,w$ be three consecutive internal bags on $P$ with $B_v \cap B_w \subseteq B_u$. Let $P'$ be a simplification of $P$ bypassing $v$ and let ${\mathcal A}'$ be the distribution obtained by running the algorithm on path $P'$, starting on $v_1$.
Then ${\mathcal A}'$ is exactly the same as ${\mathcal A}$ restricted to $B(P')$.
\end{lemma}
\begin{proof}
We can assume, without loss of generality, that $u,v,w$ appear on $P$ in the order of rounding; for otherwise, we apply \Cref{lem:traverselInv} twice: first, to reverse $P$, and preserve the distribution~${\mathcal A}$; then, to undo the reversing of $P'$ caused by the previous application.
We modify the path decomposition $P$ into a (tree) decomposition $\hat{{\mathcal T}}$ as follows: remove bag $v$ and add two new bags $v',v''$ where bag $v'$ is connected to $u$ and $w$ with $B_{v'} = B_u \cap B_v$ and $v''$ is connected to $v'$ with $B_{v''} = B_v$.
This remains a tree decomposition for the vertices in $B(P)$ since vertices in $B_v \setminus B_u$ only occur in the bag $B_{v''}$ (due to our assumption that $B_v \cap B_w \subseteq B_u$).
It is easy to check that run the algorithm {\sc SC-ROUND} on $\hat{{\mathcal T}}$ produces exactly the same distribution as ${\mathcal A}$.
Since $B(P') = B(P) \setminus (B_{v''} \setminus B_{v'} )$, we have that ${\mathcal A}|_{B(P')}$ is the distribution of {\sc SC-ROUND} on the path $\hat{P} = v_1 \ldots u v' w \ldots v_{\ell}$, obtained by removing $v''$ from $\hat{{\mathcal T}}$.
Now since $B_{v'} \subseteq B_u$, the rounding algorithm in fact does not do anything at bag $v'$, so it can be removed without affecting the distribution.
We obtain path $P'$ as a result, and this implies that ${\mathcal A}|_{B(P')}$ is the same distribution as ${\mathcal A}'$.
\end{proof}
This result allows us conduct the rounding analysis on simplifications of paths.
It remains to show that this is beneficial, that is, that the rounding error can be bounded by the length of the path on which we round.
As in the work of \chlamtac et al.\@\xspace~\cite{ckr10}, we use Markov flow graphs to analyze that error.
\subsection{Step 2: Markov Flow Graphs}
\label{sec:markov}
Let $P = v_1, \dots, v_\ell$ be a path with length $\ell$ and $s \in B_{v_1}$, $t\in B_{v_\ell}$. %
We run \autoref{alg:Chlamtac} from $v_1$ to~$v_\ell$ to compute some assignment $f$.
Let ${\mathcal A}$ be the probability distribution of the resulting assignment $f$. %
Recall that $\operatorname{\sf algcut}\xspace(s,t)$ denotes the probability that the algorithm assigns $f(s) \neq f(t)$, and $\operatorname{\sf lpcut}\xspace(s,t)$ is the probability that $s$ and $t$ are separated according to the distributions $\set{\mu_L}_L$, i.e.~$\Pr_{f \sim \mu_{B \cup \{s,t\}}}[f(s) \neq f(t)]$.
In the second step, we analyze the probability of $\operatorname{\sf algcut}\xspace(s,t)$ in terms of $\operatorname{\sf lpcut}\xspace(s,t)$.
This step is encapsulated in the following lemma.
\begin{lemma}
\label{lem:algo:markov}
There exists a directed layered graph $H$ containing nodes $s_0,s_1,t_0,t_1 \in V(H)$ and a weight function $w_H$ on the edges, satisfying the following properties:
\begin{enumerate}
\item For $i = 0,1$, we have that $\Pr_{f \sim {\mathcal A}}[f(s) = i\; \&\; f(t) = 1-i]$ is at least an $\Omega(1/\ell^2)$-fraction of the minimum $(s_i, t_{1-i})$-cut of $H$.\label{lem:algo:markov:i}
\item For $i = 0,1$, the value of a maximum $(s_i, t_{1-i})$-flow in $H$ is at least $\Pr_{f \sim \mu}[f(s) =i\; \&\; f(t) = 1-i]$.\label{lem:algo:markov:ii}
\end{enumerate}
\end{lemma}
\Cref{thm:algo:main} immediately follows from this lemma.
\begin{proof}[Proof of \Cref{thm:algo:main}]
We run the algorithm of \chlamtac et al.\@\xspace to get some $V(G)$-assignment~$f$.
Consider a pair $\set{s,t} \in E_D$.
Using \Cref{lem:traverselInv} and \Cref{lem:distpreservereduction}, we can reduce the analysis to a path $P$ of length at most $\ensuremath{\delta}\xspace$, which is a simplification of a path in ${\mathcal T}$. %
Now, by \Cref{lem:algo:markov} and max-flow-min-cut theorem, we get that
\begin{align*}
\operatorname{\sf algcut}\xspace(s,t)
&= \Pr_{f \sim {\mathcal A}}[f(s) = 0\; \&\; f(t) = 1] + \Pr_{f \sim {\mathcal A}}[f(s) = 1\; \&\; f(t) = 0] \\
&\geq \Omega\paren{\frac{1}{\ensuremath{\delta}\xspace^2}} \paren{\operatorname{mincut}(s_0, t_1) + \operatorname{mincut}(s_1, t_0)} \\
&= \Omega\paren{\frac{1}{\ensuremath{\delta}\xspace^2}} \paren{\operatorname{maxflow}(s_0, t_1) + \operatorname{maxflow}(s_1, t_0)} \\
&\geq \Omega\paren{\frac{1}{\ensuremath{\delta}\xspace^2}} \paren{\Pr_{f \sim \mu}[f(s) =0\; \&\; f(t) = 1] + \Pr_{f \sim \mu}[f(s) =1\; \&\; f(t) = 0]} \\
&= \Omega\paren{\frac{1}{\ensuremath{\delta}\xspace^2}} \operatorname{\sf lpcut}\xspace(s,t) \enspace .
\end{align*}
Therefore, $f$ separates each pair $\set{s,t}$ with probability that is a factor of $O(\ensuremath{\delta}\xspace^2)$ away from $\operatorname{\sf lpcut}\xspace(s,t)$.
Applying Observation \ref{obs:derandomisation} with $c = \Omega(1/\ensuremath{\delta}\xspace^2)$, we can obtain (deterministically) an assignment $f^*$ that is an $O(\ensuremath{\delta}\xspace^2)$-approximation for the \mbox{\sf Sparsest-Cut}\xspace instance.
\end{proof}
The rest of this section is dedicated to proving the \Cref{lem:algo:markov}.
The tools needed for this proof are implicit in the work of \chlamtac et al.\@\xspace~\cite{ckr10}.
We restate them for the sake of completeness and in order to adjust it to our terminology.
The section is organized as follows: first, we describe the construction of our graph $H$, and then we proceed to analyze the values of maximum flow and minimum cut.
We will only analyze the flow and cut for $i= 0$, that is, $(s_0,t_1)$-flow and $(s_0,t_1)$-cut. The other case is analogous.
\paragraph{Construction of Graph $H$:}
Without loss of generality, we can assume that the distributions $\set{\mu_L}_L$ are symmetric in the labels $\{0,1\}$, see \Cref{lem:symmetrization}.
In particular, this gives $\Pr[f(v) = 1] =\linebreak \Pr[f(v) = 0] = 1/2$ for any vertex $v$.
The rounding can be modeled by a simple Markov process. %
Denote by $I_0,\dots, I_\ell$ the sets that are conditioned on in \autoref{alg:Chlamtac}, $I_i = B_{v_i} \cap B_{v_{i+1}}$ for $i \in \set{1,\ldots,
\ell-1}$; we refer to these sets as \emph{conditioning sets}.
For the initial and final sets of the rounding procedure we take $I_0 = \{s\}$, $I_\ell = \{t\}$.
Now we are ready to describe our graph $H$:
\begin{itemize}
\item \textbf{Vertices}: %
Vertices of $H$ are arranged into layers $L_0,\dots,L_\ell$ with $L_i = \mathcal{F}[I_i]$.
Observe that $|L_i| = 2^{|I_i|}$. %
The vertices of $H$ represent the intermediate states the algorithm might
reach.
\item \textbf{Edges}: %
For each $i$, there is a directed edge from every vertex in $L_i$ to every vertex in $L_{i+1}$. %
The weight of the edge $(f_i, f_{i+1})$, for $f_i \in L_i$, $f_{i+1} \in L_{i+1}$, is equal to the probability of joint event,
$w_H(f_i, f_{i+1}) = \Pr[f|_{I_i} = f_i \wedge f|_{I_{i+1}} = f_{i+1}]$.
We remark that the weight is 0 whenever $f_i$ and $f_{i+1}$ are
contradictory, and that probabilities are well defined, as $I_i \cup I_{i+1} \subseteq B_{i+1}$.
\end{itemize}
Observe that the weight of an edge is the probability that both of its endpoints are reached by the algorithm, and hence the probability that the algorithm transitions along that edge.
\newcommand{{\mathcal I}}{{\mathcal I}}
\begin{observation}
Let ${\mathcal I} = \bigcup_{i} I_i$.
The distribution ${\mathcal A}|_{{\mathcal I}}$ can be viewed as the following random walk in~$H$: Pick a random vertex in $L_0$ and start taking a random walk where each edge is taken with probability proportional to its weight. %
Formally, once a node $f_i$ is reached, choose the next node $f_{i+1}$ with probability $w_H(f_i, f_{i+1})/\Pr[f|_{I_i} = f_i]$.
\end{observation}
At this point, we rename ${\mathcal A} := {\mathcal A}|_{{\mathcal I}}$.
Notice that the layer $L_0$ contains two vertices corresponding to the assignment $f(s) = 0$ and $f(s) = 1$, respectively. We denote them by $L_0 = \{s_0, s_1\}$. %
Similarly, $L_{\ell} = \{t_0, t_1\}$. %
Notice further that $\Pr_{f \sim {\mathcal A}}[f(s) = 0, f(t) = 1]$ is exactly the probability that the random walk starts at $s_0 \in L_0$ and ends at $t_1 \in L_{\ell}$.
\paragraph{Maximum $(s_0,t_1)$-Flow:}
We are now ready to show that the value of the maximum $(s_0,t_1)$-flow is at least $\Pr_{f \sim \mu}[f(s) =0, f(t)=1]$.
We define the flow $g\colon E(H) \to \mathbb{R}_{\geq 0}$ as follows, for $i \in \set{1,\ldots,
\ell-1}$, $f_i \in L_i$ and $f_{i+1} \in L_{i+1}$:
\[
g(f_i, f_{i+1}) = \Pr_{f\sim \mu_{B_{v_{i+1}} \cup \{s,t\} }}[ f(s) = 0, f(t) = 1, f|_{I_i} = f_i, f|_{I_{i+1}} = f_{i+1} ] \enspace .
\]
We remark that $g$ is an $s_0$-$t_1$-flow, that is, it satisfies flow conservation at all vertices in $H$ except $s_0,t_1$, and the capacities of
graph $H$ are respected, that is, $g(e) \leq w_H(e)$ for all $e \in E(H)$. %
The value of~$g$ is given by: %
\begin{align*}
\sum_{(s_0,f^*) \in \delta^+(s_0)}g(s_0, f^*) &= \sum_{f^* \in \mathcal{F}[I_1]}\Pr_{f \sim \mu_{B_{v_1} \cup \{s,t\} } }[f(s) = 0, f(t) = 1, f|_{I_1} = f^*]\\
&=\Pr_{f \sim \mu_{B_{v_1} \cup \{s,t\} } }[f(s) = 0, f(t) = 1] \enspace .
\end{align*}
This concludes the proof of Point \ref{lem:algo:markov:ii} of \Cref{lem:algo:markov}.
\paragraph{A Potential Function:}
Before we show a cut with the desired capacity, we need to introduce some notation.
For $i = 0,\ldots, \ell$, let $X_i$ be a random variable indicating the vertex in $L_i$ visited by the random walk (i.e.~picked by the algorithm.
We denote by $\textbf{X}= X_0 X_1 \ldots X_{\ell}$ the path taken in the random walk process. %
We can interchangeably view distribution ${\mathcal A}$ as either the distribution that samples an assignment $f\colon {\mathcal I} \rightarrow \{0,1\}$ or one that samples a (random walk) path $\textbf{X}$.
We define, for every layer $L_i$ and every vertex $v \in L_i$,
\[
A(v):= \Pr_{\textbf{X} \sim {\mathcal A}} [X_0 = s_0 \mid X_i = v] -\frac{1}{2} \enspace .
\]
Intuitively, this function captures the extent to which $v$ has information about the initial state of the Markov process. %
On the one hand, if $A(v)$ is equal to $0$, $v$ knows essentially nothing about $X_0$, the choice of $v$ does not imply anything about $X_0$. %
On the other hand, if $A(v)$ is far from $0$, then we can glean a lot of information about $X_0$ from $v$ being visited; %
in particular, if the probability that $s$ and $t$ are cut is low, we must have $A(t_1) \approx - 1/2$.
To track how $A$ changes from layer to layer, we use the potential function $\phi\colon \{0,\dots,\ell\} \to \mathbb{R}_{\geq 0}$, defined as:
\[
\phi(i) := \operatorname{Var}_{\textbf{X} \sim {\mathcal A}}[A(X_i)] \enspace.
\]
The following lemma by \chlamtac et al.\@\xspace bounds the change in potential in terms of the probability that $X_0 = s_0$ and $X_\ell = t_1$.
\begin{lemma}[\cite{ckr10}, Lemma 5.2]
It holds $\phi(0) - \phi(\ell) \leq 2\Pr[X_0 = s_0 \wedge X_\ell = t_1]$ \enspace .
\end{lemma}
\paragraph{Minimum $(s_0,t_1)$-cut:}
We are now ready to analyze the value of minimum $(s_0,t_1)$-cut in $H$. It suffices to give a lower bound on $\phi(0) - \phi(\ell)$.
This is is possible by the following lemma which is proved implicitly by \chlamtac et al.\@\xspace~\cite{ckr10}.
\begin{lemma}[\cite{ckr10}, Lemma 5.4]
\label{lem:MarkovCut}
Let $C$ be the set of edges $(f_i, f_{i+1})$ in $E(H)$ such that $|A(f_i) - A(f_{i+1})|$ is at least some threshold $\rho > 0$. Then $\sum_{e \in C}w_H(e) \leq (\phi(0) - \phi(\ell)) \cdot 1/\rho^2$.
\end{lemma}
We can apply \Cref{lem:MarkovCut} in the following fashion.
Suppose $A(t_1) \geq 0$. In that case we have $\Pr[X_0 = s_0 \mid X_\ell = t_1] \geq 1/2$, so $s$ and $t$ are cut with probability at least $\frac{1}{2}\operatorname{\sf lpcut}\xspace(s,t)$.
This error is already a small enough, so assume $A(t_1) < 0$. Then $A(s_0) - A(t_1) > 1/2$.
Since every path from $s_0$ to $t_1$ has exactly $\ell$ edges, any such path must contain an edge $(f_i, f_j)$ with $A(f_i) -A(f_j) > 1/ (2\ell)$.
Cutting all such edges therefore separates $s_0$ and $t_1$. Hence, by applying \Cref{lem:MarkovCut}, the minimum $s_0$-$t_1$-cut has size at most
\begin{align*}
O(\ell^2) (\phi(0) - \phi(\ell))
&\leq O(\ell^2) \Pr[X_0 = s_0 \wedge X_\ell = t_1]\\
&=O(\ell^2) \Pr[f(s) = 0 \wedge f(t) = 1] \enspace .
\end{align*}
This concludes the proof of Point \ref{lem:algo:markov:i} of \Cref{lem:algo:markov}. We see that the cutting probability predicted by the distributions is realised by the rounded solution $f$, up to a factor $\Omega(1/\ell^2)$.
This gives an alternative to the analysis given by \chlamtac et al.\@\xspace whose constant depends on the size of the layers of $H$ rather than the number of layers.
While the layer sizes depend only on $k$, the dependence is exponential.
The number of layers is a priori $\log(n)$, which would give a worse approximation guarantee.
However, we will show how to modify a tree decomposition to ensure that $H$ has few layers.
\section{Combinatorially Shallow Tree Decompositions}
In this section, we show how to construct tree decompositions with low combinatorial diameter, thus achieving the approximation results stated in \Cref{thm:main-intro}. %
We start by restricting our consideration to decompositions that are shallow in the traditional sense. %
For a given graph $G$ with treewidth $k$, we consider a tree decomposition $({\mathcal T}, \set{B_i}_{i \in V({\mathcal T})})$ with diameter $\ensuremath{d}\xspace=O(\log n)$ and width $O(k)$~\cite{Bodlaender88}. %
Fix some root $r$ in $V({\mathcal T})$.
Our goal is now to modify ${\mathcal T}$ such that every node has a combinatorially short path to $r$. %
This is a necessary requirement, but perhaps surprisingly it is not sufficient. %
The combinatorial lengths of paths do not necessarily induce a metric on $V({\mathcal T})$%
\footnote{Consider bags $\{ab\}, \{abc\}, \{acd\}, \{ade\}, \{aef\}, \{afg\}, \{a\}$ occuring in that order as a path. The whole path can be reduced to just the endpoints. The subpath $\{ab\}, \{abc\}, \{acd\}, \{ade\}, \{aef\}, \{afg\}$ is irreducible. Thus the distance from $\{ab\}$ to $\{afg\}$ is larger than the sum of the distances from $\{ab\}$ to $\{a\}$ and $\{afg\}$ to $\{a\}$.}, %
and therefore bounding the length to $r$ does not on its own suffice to bound the combinatorial diameter.
We will not show explicitly that the modified structures are in fact tree decompositions.
The proofs are straightforward using \Cref{lem:TreeDecompMonotonicity}.
We introduce three objects, which we call \textbf{bridges}, \textbf{highways}, and \textbf{super-highways}, and show that they can be used to prove the three parts of \Cref{thm:main-intro}.
\subsection{Bridges}
\begin{figure}
\includegraphics[width=\textwidth]{Bridges.pdf}
\caption{Illustration of a path from the root to some node $s$. The square nodes are the synchronization nodes. The bridge from $y$ to its synchronization ancestor is marked with dashes in the first image. The dotted nodes in the second image mark those nodes which can be removed when simplifying the $x$-$s$-path in ${\mathcal T}'$. }
\label{fig:bridges}
\end{figure}
Fix a parameter $\lambda \in \set{1,\dots, \ensuremath{d}\xspace}$. Define $\ell\colon V({\mathcal T}) \to \mathbb{N}_0$ to be the \emph{level} of a node in ${\mathcal T}$, that is, $\ell(v)$ is the number of edges on ${\mathcal T}_{v \leftrightarrow r}$.
\begin{definition}
We call a node a \emph{synchronization node} if its level is a multiple of $\lambda$.
Define also the \emph{synchronization ancestor} $\sigma(v)$ of any node $v$ to be the first node on the path from $v$ to $r$ that is a synchronization node, excluding $v$ itself.
\end{definition}
We can construct a tree decomposition $({\mathcal T}', \set{B'_i}_i)$ by taking ${\mathcal T}'={\mathcal T}$ and setting $B_v' = B({\mathcal T}_{v \leftrightarrow \sigma(v)})$, that is, the new bag is obtained by combining all the bags from $v$ up to its synchronization ancestor. %
This increases the width of the decomposition by a factor of at most $\lambda$. We may view this path connecting $v$ to the synchronization point as a {\bf bridge} crossing over all intermediate nodes in one step.
\begin{lemma}
\label{lem:bridges:diam}
${\mathcal T}'$ has combinatorial diameter $O(\ensuremath{d}\xspace / \lambda)$.
\end{lemma}
\begin{proof}
Fix any two nodes $s,t \in V({\mathcal T}')$ and take $x$ to be their lowest common ancestor in ${\mathcal T}'$.
Then the combinatorial length of ${\mathcal T}'_{s \leftrightarrow t}$ is at most the sum of the combinatorial lengths of ${\mathcal T}'_{s \leftrightarrow x}$ and ${\mathcal T}'_{x \leftrightarrow t}$.%
We remark that triangle inequality holds in this case, because $x$ is on the path from $s$ to $t$. %
Thus, it suffices to show that the combinatorial length of ${\mathcal T}'_{s \leftrightarrow x}$ is $O(\ensuremath{d}\xspace / \lambda)$.
The result follows analogously for ${\mathcal T}'_{x \leftrightarrow t}$.
Using the rules of \Cref{def:simplification}, we can bypass any node that is neither a synchronization node nor $s$ or $x$, since the bag of the unique child (in ${\mathcal T}'_{s \leftrightarrow x}$) of such a node is a superset of its own bag. %
Therefore, the path $\{v \in {\mathcal T}'_{s \leftrightarrow x} | v = s \vee v=x \vee v\text{ is a synchronization node}\}$ is a simplification of~${\mathcal T}'_{s \leftrightarrow x}$.
Since there are at most $\ensuremath{d}\xspace/\lambda$ synchronization nodes on any upward path, the lemma follows.
\end{proof}
This lemma, in conjunction with \Cref{thm:algo:main} and the fact that ${\mathcal T}'$ can be computed in polynomial time from ${\mathcal T}$, yields:
\begin{corollary}
\label{cor:bridge:algo}
For every $\lambda$, there is an algorithm that computes an $O((\log n / \lambda)^2)$-approximation for \mbox{\sf Sparsest-Cut}\xspace instances where $G$ has treewidth at most $k$, in time $2^{O(\lambda k)}\operatorname{poly}(n)$.
Setting $\lambda = \log n /k$ results in an $O(k^2)$-approximation in time $2^{k}\operatorname{poly}(n)$, while setting $\lambda=\log n$ gives an $O(1)$-approximation in time $n^{O(k)}$.
\end{corollary}
\subsection{Highways}
\begin{figure}
\includegraphics[width=\textwidth]{Highways.pdf}
\caption{The dashed nodes in the first image mark the bridge and highway from $y$ to $r$. The other images illustrate the two simplification rounds for the $x$-$s$-path, leaving a path of length $2$.}
\label{fig:highways}
\end{figure}
The idea of extending bags towards the root can be exploited further by adding the vertices in a synchronization bag to all of its descendants.
We may regard this as giving each node a bridge to the next synchronization node, as well as a {\bf highway} along the synchronization nodes towards the root.
This idea leads to the following construction.
Let $({\mathcal T}', \set{B'_i}_i)$ be a modified tree decomposition with ${\mathcal T}'={\mathcal T}$ as before, and
\[
B_v' := B(\{ w \in {\mathcal T}_{v \leftrightarrow r} \mid w \in {\mathcal T}_{v \leftrightarrow \sigma(v)} \vee w \text{ is a synchronization node} \}) \enspace .
\]
The size of these bags is at most $k (\lambda + \ensuremath{d}\xspace / \lambda)$, which for $\lambda = \ensuremath{d}\xspace /k$ gives $\ensuremath{d}\xspace+ k^2 = O(\log n + k^2)$.
Notice that the bag $B_r$ is now contained in any bag $B'_i$, so we have some hope that the combinatorial diameter of $({\mathcal T}', \set{B'_i}_i)$ is low.
Indeed this is true.
\begin{lemma}
\label{lem:highways:diam}
${\mathcal T}'$ has combinatorial diameter at most $3$.
\end{lemma}
\begin{proof}
As before, we split any $s$-$t$-path at $x$, the lowest common ancestor of $s$ and $t$, and consider only the $s$-$x$-path.
Every non-synchronization node $v$ on ${\mathcal T}_{s\leftrightarrow x}$ has a node below it which is either a synchronization node or $s$.
The bag of that node is a superset of $B'_v$, so all non-synchronization nodes except $s$ and $x$ can be bypassed.
Call that reduced path $P$.
Consider the neighbor of $s$ in $P$, which we denote $v$, and assume that $v$ is not the neighbor of $x$ in $P$. %
Then $v$ must be a synchronization node, and its next node in $P$ is $\sigma(v)$. %
Now, the intersection $B'_v \cap B'_{\sigma(v)}$ contains exactly all of the bags of synchronization nodes in ${\mathcal T}_{\sigma(v)\leftrightarrow r}$, and thus, ${B'_v \cap B'_{\sigma(v)} \subseteq B'_s}$. %
This implies that $v$ can be bypassed, and by repeating this process, we can bypass every synchronization node except for the neighbor of $x$.
This gives a possible simplification of ${\mathcal T}_{s \leftrightarrow t}$ as the path $(s, \sigma_s, x, \sigma_t, t)$, where the $\sigma_s$ and $\sigma_t$ are the synchronization nodes below $x$ on the paths to $s$ and $t$, respectively.
There is a further reduction of the whole path, since $B_x'$ is precisely $B_{\sigma_s}' \cap B_{\sigma_t}'$.
This allows us to remove $x$ as well, giving a simplification of length $3$.
\end{proof}
Using the fact that $\ensuremath{d}\xspace \in O(\log n)$, and setting $\lambda = \ensuremath{d}\xspace /k$ gives a fixed-parameter algorithm that yields a constant-factor approximation:
\begin{corollary}
\label{cor:highways:algo}
There exists an algorithm that in time $2^{O(k^2)}\cdot\operatorname{poly}(n)$ computes a factor-$O(1)$ approximation for \mbox{\sf Sparsest-Cut}\xspace instances where $G$ has treewidth at most $k$.
\end{corollary}
\subsection{Super-Highways}
\begin{figure}
\includegraphics[width=\textwidth]{Superhighways.pdf}
\caption{Illustration of an upward path with nodes of layer $-1$ as circles, nodes of layer 0 as diamonds, and nodes of layer $1$ as squares. The root is at some unspecified maximum layer. The dashed nodes in the first image mark the super-highway from $s$ to $r$. The other images illustrate the simplification rounds for the $x$-$s$-path, removing all nodes of some layer in each round, except $x$, $s$, and possibly one node close to $x$.}
\label{fig:superhighways}
\end{figure}
We can think of the previous construction as having two layers, bridges to synchronization nodes and highways along synchronization nodes to the root. The highways need to cover many synchronization nodes, leading to large bags in ${\mathcal T}'$.
To improve on this we introduce a network of {\bf super-highways} of different layers, where each layer covers fewer, more spaced-out synchronization nodes on a root-leaf path.
When we connect a node to the root we can then move up the tree layer by layer with increasing speed, decreasing the size of bags in ${\mathcal T}'$.
This is payed for by the need for an additional node in path simplifications for moving between layers, giving a trade-off between run time and approximation guarantee.
Let $q \in \mathbb{N}$ be a parameter representing the number of layers. %
For a node $v \in {\mathcal T}$, we define the layer of $v$, denoted $\pi(v)$, as
\[
\pi(v) := \max\{-1, \max\{ j\in\{0,\dots, q-1 \} \mid \ell(v) \equiv 0 \mod k^{j/q} \ensuremath{d}\xspace / k\} \} \enspace .
\]
By this definition all synchronization nodes are assigned to some non-negative layer, and all other nodes are on layer $-1$.
We now get a new tree decomposition $({\mathcal T}', \set{B'_i}_i)$ by constructing bags:
\[
B_v' = B(\{ w \in {\mathcal T}_{p(v) \leftrightarrow r} \mid \pi(w) = \max\{ \pi(u) \mid u \in {\mathcal T}_{p(v)\leftrightarrow w} \} \} \cup \{v\})\enspace .
\]
Informally, we start at some node $v$ and move towards $r$ by first taking all nodes of layer $-1$ until we hit a node of layer $0$, then taking only nodes of layer $0$ until we hit layer $1$, and so on.
The nodes at higher layers are spaced further apart. Thus this process ``speeds up'' thereby generating smaller bags.
To be precise, there are $q$ layers and at most $k^{1/q}$ nodes of any one layer in a bag, so ${\mathcal T}'$ has width $O(\ensuremath{d}\xspace + qk^{1+1/q})$.
We now show that $({\mathcal T}', \set{B'_i}_i)$ has combinatorial diameter depending only on $q$.
\begin{lemma}
\label{lem:superhighways:diam}
$({\mathcal T}', \set{B'_i}_i)$ has combinatorial diameter at most $2q+1$.
\end{lemma}
\begin{proof}
As before, we only show that any upward path from $s$ to $x$ has combinatorial length at most~$q+1$.
We need to perform a round of reductions for every layer, with the goal of leaving only $s$, $x$, as well as the first node of at least that layer below $x$.
For layer $-1$, this holds with the same argument as before.
We can now proceed by induction, fixing some layer $i$ and assuming that the $s$-$x$-path $P$ has been reduced to contain only $s$, then nodes of layers $\geq i$, followed by a sequence $(\sigma_{i-1}, \sigma_{i-2}, \ldots, \sigma_0, x)$, where each node $\sigma_j$ is in layer $j$. Here, we assume w.l.o.g.\ that $x$ is at layer $-1$.
Now consider any node $v$ of layer $i$, except the one closest to $x$.
Because its neighbors also have level at least $i$ (or are~$s$), the intersection of their bags can be represented as the union of bags of ${\mathcal T}$ whose layer is at least $i$.
Let~$w$ be the predecessor of $v$ on $s$-$x$-path $P$.
The set $B_w'$ is constructed from some upward path starting at $w$, containing only nodes of non-decreasing layer.
This upward path hits layer~$i$ between~$w$ and~$v$, but not layer $i+1$ since a node of layer $i+1$ would be on $P$ between $w$ and~$v$.
So then~$B_w'$ covers all nodes of layer at least $i$ that $B_v'$ covers, and therefore $v$ can be bypassed, concluding induction.
The simplification of ${\mathcal T}_{s\leftrightarrow x}$ produced in this fashion is a path $(s, \sigma_{q-1}, \dots, \sigma_0, x)$, where $\pi(\sigma_i) = i$.
If we add the same simplification for ${\mathcal T}_{t\leftrightarrow x}$ we get a simplification for ${\mathcal T}_{s\leftrightarrow t}$ that takes the form $(s, \sigma_{q-1}, \dots, \sigma_0, x, \sigma_0',\dots,\sigma_{q-1}',t)$.
As before $x$ can be bypassed since its bag is the intersection of the bags of $\sigma_0$ and $\sigma_0'$.
Thus any $s$-$t$-path in ${\mathcal T}'$ has combinatorial length at most $2q+1$.
\end{proof}
This implies the existence of the following algorithms.
\begin{corollary}
\label{lem:superhighways:algo}
There exists an algorithm that, for any $q\in \mathbb{N}$, computes a factor-$O(q^2)$ approximation for \mbox{\sf Sparsest-Cut}\xspace in time $O(2^{qk^{1+1/q}})\cdot\operatorname{poly}(n)$.
Taking $q = \log k$ gives a factor-$O(\log^2k)$ approximation in time $2^{O(k\log k)}\cdot\operatorname{poly}(n)$.
\end{corollary}
\begin{acks}
Parinya Chalermsook has been supported by European Research Council (ERC) under the
European Union’s Horizon 2020 research and innovation programme (grant
agreement No. 759557) and by Academy of Finland Research Fellowship, under
grant number 310415. %
Joachim Spoerhase has been partially supported by European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme (grant agreement No. 759557). %
Daniel Vaz has been supported by the Alexander von Humboldt Foundation with
funds from the German Federal Ministry of Education and Research (BMBF).”
\end{acks}
\bibliographystyle{abbrv}
|
\section{Introduction}
\label{sec:int}
The magnetic activity of Sun is not identical in two hemispheres---there is always
an asymmetry. This hemispheric asymmetry, also called the north-south asymmetry, has been
observed in the photospheric magnetic field \citep{MK04, McInt13, MK19} as well as in many proxies of the solar activity \citep{MKB17, GC09, norton14, Mordvinovetal20}.
The hemispheric asymmetry is a real feature of the solar cycle and is not an artefact of inaccurate or noisy observations \citep{carbo93}.
\citet{Bell61} found the evidence of hemispheric asymmetry in the number of major flares
and later \citet{Bell62} found a long-term asymmetry in the sunspot area data during Cycles 8--18.
\citet{Swinson86} found a peak in the northern hemispheric solar activity about two years after
sunspot minimum and a 22-years periodicity in the north-south asymmetry. \citet{Verma87} showed that the northern hemisphere is more active during Cycles 19 and 20. \citet{Li09} used group sunspot and sunspot
area data from 1996 to 2007 to show that the solar activity for cycle 23 is dominant in the
southern hemisphere; also see \citet{partha13} who extended this study to some part of cycle 24. A well-known sunspot asymmetry was observed during the Maunder minimum. Most of the sunspots were registered in the southern hemisphere \citep{SN94}.
It is believed that a hydromagnetic dynamo, operating in the solar convection zone, is responsible for the generation and maintenance of the large-scale magnetic field and the cycle of solar activity \citep{Pa55}.
In the current scenario of the solar dynamo \citep{Kar14a, Cha20, Hazra}, a toroidal component of the magnetic field
is largely generated due to the shearing of the poloidal component by the differential rotation. This toroidal field gives back to the poloidal one due to the decay and dispersal of tilted bipolar magnetic regions (BMRs)---
the so-called Babcock--Leighton\ process
and possibly due to the helical convection---the so-called $\alpha$ effect.
Meridional circulation and the small-scale convective flow play the role in transporting the magnetic field from the near-surface layer (the location of Babcock--Leighton\ process)
to the deeper convection zone, where the shearing process is efficient, and thus largely regulates the cycle period.
The turbulent nature of the helical convective flow---the main drive of the dynamo---is expected to make the magnetic field unequal in two hemispheres.
Thus, a hemispheric asymmetry
in the solar magnetic field is unavoidable. Furthermore, the tilts of BMRs, which primarily determines
the poloidal field, has a large scatter around Joy's law \citep{SK12, Wang15, Arlt16, Jha20}.
Thus the tilt scatter makes the poloidal magnetic field irregular and asymmetric \citep{LC17, KM17, Kar20}.
As the poloidal field is the seed for the toroidal field of the next solar cycle,
the asymmetry in the polar field is propagated in the solar cycle \citep{CCJ07}.
Dynamo models have shown that when the turbulent diffusion is sufficiently strong, the coupling between two hemispheres
tries to diminish the asymmetry introduced in the polar field and thus the asymmetry in the solar cycle may not persist for
several cycles \citep{CC06, GC09, Kar10, KM17}.
Dynamo models by including scatter in the BMR tilt \citep{LC17, KM17, KM18} or $\alpha$ term in the poloidal source \citep{OK13, KMB18, HN19} produce hemispheric asymmetry
in the magnetic cycle, which in some parameter regimes, are in agreement with observations. \citet{SC18} have shown that the random excitation of the quadrupole mode of dynamo by the stochastic fluctuations
in the Babcock--Leighton\ process can lead to an asymmetry in the observed magnetic field; also see \citet{Nepo19}.
Thus, all these previous results motivate us to explore whether the solar hemispheric asymmetry is governed by a low-dimensional chaotic process or stochastic process? Is there any long-term memory in the solar cycle asymmetry?
Nonlinear time series analysis is suitable to answer the above questions.
While there exist many such studies for the solar cycle data \citep[e.g.,][]{Ostriakov90, carbo94, jevti01, Letellier06,2009SoPh..260..441S}, only a few such studies are performed in the solar cycle asymmetry.
\citet{carbo93} computed the correlation dimension of the asymmetry of daily sunspot area
during 1874--1989 and did not find any evidence of low-dimensional chaos.
By computing the Higuchi's fractal dimension \citep{Higuchi88} and some nonlinear prediction method for the hemispheric asymmetry of sunspot number
during 1947--1984, \citet{Watari96} concluded
that the sunspot number asymmetry is highly irregular and not deterministic chaos.
In the present work, we shall utilize the maximum available sunspot area data (during 1874--2016) of hemispheric asymmetry and apply multiple nonlinear time series techniques to check the inherent nonlinear properties of the system.
First, we shall compute the correlation dimension ($D_2$) to extract whether the data has any strange attractor in the data and thus this analysis will reveal the existence of any low-dimensional chaos in the underlying system \citep{GP83a}.
Next, we shall compute a fractal dimension using the method given in \citet{Higuchi88} which provides a stable estimate of the fractal dimension when the data is more irregular and non-stationary. Higuchi's dimension will give another independent support of whether the asymmetry data is from a stochastic process or low dimensional chaos. Finally, we shall
compute the Hurst exponent \citep{1969WRR.....5..321M} to check whether the data has any persistent memory or not.
The final conclusion will be presented in Section \Sec{sec:conclusion}.
\section{Observational data}
We use the monthly mean sunspot area data during May 1874 -- September 2016 obtained from the Royal Greenwich Observatory (RGO) \footnote{https://solarscience.msfc.nasa.gov/greenwch.shtml}.
The RGO data have been the only available record of sunspot area separately in two hemispheres for the longer duration and this is routinely used in many studies of the solar activity \citep{Hat15}.
RGO provides the monthly value of the average (over observed day) of the daily sunspot area in the unit of millionths of a hemisphere and it is evenly spaced.
We have also repeated our analyses with the newly available monthly mean hemispheric sunspot number data; \Sec{sec:HemSSN}.
The easy way to measure the asymmetry is to take the difference in the values between two hemispheres \citep{Ballester2004, Chang2007},
i.e. asymmetry,
\begin{equation}
AS=A_N-A_S,
\label{eq:AS}
\end{equation}
where $A_N$ and $A_S$ are the monthly values of the sunspot area in the northern and southern hemispheres, respectively.
We note that during solar maxima, the difference becomes large in comparison to the value during minima and this causes a cyclic pattern in the asymmetry; see \Fig{fig1} top panel.
Furthermore, if a cycle is strong, then the asymmetry is large and vice versa.
Therefore, in the literature \citep{carbo93, Oliver1996, Duchlev2001, GC09, partha13, Priy14},
the asymmetry is also measured by normalizing its strength, namely.
the normalized asymmetry,
\begin{equation}
AS_{Norm}=\frac{A_N-A_S}{A_N+A_S}.
\label{eq:normAS}
\end{equation}
When both $A_N$
and $A_S$ become equal to zero, we set $AS_{Norm}=0$.
We realized that this happens for 15 data points
(which is less than $1\%$ of the total data). However, we
discussed its effect in Section 4 by replacing these points
with interpolated values.
We note that this $AS_{Norm}$ is a different time series
than the $AS$; statistics of the data are different ( \Tab{table1}).
Again this definition of asymmetry is not satisfactory because during the solar minima, when the sunspot area becomes very small and this leads to increase in $AS_{Norm}$; see \Fig{fig1} bottom panel.
It was examined in \citet{Yi1992}, that dividing the difference of the hemispherical sunspot areas by the total sunspot area results in the appearance of a peak in the power spectrum
between 11 and 12 years.
Due to such facts from the literature, where both definitions have been used to measure the solar cycle asymmetry, it becomes necessary to perform analyses of the time series of asymmetry using both the methods that we have discussed above.
\begin{figure}
\centering{\includegraphics[width=.99\columnwidth]{asymmetry5.eps}}
\caption{Time series of the hemispheric asymmetry of monthly average sunspot area
(in unit of millionth of a solar hemisphere)
as measured by $AS$ (top) and $AS_{Norm}$ (bottom).
}
\label{fig1}
\end{figure}
\begin{table}
\caption{
Some statistics of the data. Total number of data points used: 1709. Symbols in second to third columns are as follows: $<S>$, the average of the time series; $rms$, the root mean-squared deviation, and $<{\rm PN}>=\sqrt{<S>}$.
}
\centering
\begin{tabular}{lrrc}
\hline
\hline
Data & $<S>$ & $rms$~ & $<{\rm PN}>/rms$\\
\hline
North & 426.79 & 488.2 & 0.042 \\
South & 409.43 & 470.1 & 0.043 \\
AS & 17.36 & 466.9 & 0.009 \\
AS$_{\rm Norm}$
& 0.01 & 0.6 & 0.225 \\
\hline
\end{tabular}
\label{table1}
\end{table}
\section{Methods}
To identify the nonlinear properties of the hemispheric asymmetry of sunspot area time series, we shall compute three quantities for $AS$ and $AS_{Norm}$, namely, the Correlation dimension, Higuchi's dimension, and Hurst exponent. Below, we discuss briefly how to compute these quantities.
\subsection{Correlation dimension}
First, we shall apply a method of time series analysis to distinguish the random noise in the underlying system from the low dimensional chaos.
This method is to measure a fractal dimension of the strange attractor
in the system which is popularly known as the correlation dimension ($D_2$).
The method for obtaining $D_2$ has been given in \citet{GP83a,GP83b}, also see \citet{takens81,packard80}.
This has also been used in many astrophysical applications \citep[e.g.,][]{schreiber99,misra06,KDM10},
including identifying chaotic dynamics of the solar cycle data \citep{OU90,carbo94,jevti01}.
In this method, we construct a $M$ dimensional phase space using our time series $X(i)$ (where $i= 1, 2, 3 .....N$); $X(i)$ is $AS$ or $AS_{\rm norm}$ in our case.
In this space, any vector has the following form:
\begin{equation}
x_{i} = [X(i), X(i+\tau), ........., X(i+(M-1)\tau)].
\label{eq:vec}
\end{equation}
The time delay, $\tau$ is chosen in such a way that each component becomes independent of each other.
We find that at $\tau = 3$~months, the auto-correlation of the data falls below $1/e$ and thus we
set this value for $\tau$ in our analysis \citep[see][for a detailed discussion on choosing $\tau$]{carbo94}. We checked that our results do not change abruptly
if we increase $\tau$.
Next, the correlation function $C_M(r)$ is given as
\begin{equation}
C_M(r)=\frac{1}{N(N_c-1)}\sum_{i=1}^{N} \sum_{(j,j\neq i)}^{N_c}\Theta(r-|x_{i}-x_{j}|),
\end{equation}
where $x_{i}$ is a reconstructed vector; \Eq{eq:vec}, $\Theta$ is Heaviside function
($\Theta(z)$ = 1 if $z\geq 0$ and 0 if $z < 0$),
$N$ the total number of points and $N_c$ is the number
of centers.
Essentially, $C_M(r)$ gives the number of points that are within a distance $r$ from the centre,
averaged over all the centres.
Then for small $r$, ${D_{2}(M)}$ is given by
\begin{equation}
{D_{2}(M)}= \frac{\rm d log(C_M(r))}{\rm d log(r)}.
\label{eq:slope}
\end{equation}
To compute $C_M(r)$, we divide the entire phase space into $M$ cubes of length
$r$ around a point and count the average number of points. To avoid
the edge effects due to the finite number of data points, we compute
$C_M(r)$ in the range $ r_{min}<r<r_{max}$. Here $r_{min}$ is chosen when
$C_M(r)$ is just greater than one and $r_{max}$ is taken in such a way that all $M$
cubes remain within the embedding space.
For a fixed value of $M$,
$D_{2}$
is computed for different values of $r$ in the linear region of ${\rm log}( C_M( r))$ versus ${\rm log}( r)$ plot using
\Eq{eq:slope}.
The average of all these values will give our final $D_{2}$
and the mean standard deviation over the average value gives the error on $D_{2}$.
The whole calculation is repeated for different values of $M$.
The value of $D_{2}$ should increase initially with
the increase of $M$. However, if the time series is obtained from a low dimensional chaotic system, then
$D_{2}$ tends to saturate above a certain value of $M$. In contrast, for a stochastic system, $D_{2}$ keeps on increasing
with $M$. That is the number of dimensions needed to describe the system in the phase space is infinitely large in a stochastic system. In that case,
$D_{2} \approx M$, for all $M$. Thus, the variation of $D_2$ with $M$ is used to distinguish between the random noise vs low dimensional chaos.
\subsection{Higuchi's dimension}
Previously, many methods for finding stable estimations of the power-law spectral index have been discussed in the literature. Here, we follow the method given in \citet{Higuchi88} to calculate the fractal dimension $D$ of the asymmetry series. This method is helpful in providing a stable estimate of the fractal dimension of the asymmetry time series.
We recall that $X(i)$ is our time series ($i =1, 2, 3,..., N$; $N$ is the total number of observations taken at a regular interval) and thus,
\begin{equation}
X(i): X(1), X(2), X(3), ..., X(N)
\end{equation}
From this, a new time series is constructed in the following manner:
\begin{equation}
X^m_\tau: X(m), X(m+\tau), X(m+2\tau), ..., X\left(m+\left[\frac{N-m}{\tau}\right]\tau\right)
\end{equation}
where $m=1,2,...,\tau$, and $[\:]$ denotes Gauss's notation. The length $L_{m}(\tau)$ of the curve associated to each $X^m_\tau$ is defined as:
\begin{equation}
L_{m}(\tau)=\left\{\left(\sum_{i=1}^{\left[\frac{N-m}{\tau}\right]}|X(m+i\tau)-X(m+(i-1)\tau)|\right)\frac{N-1}{\left[\frac{N-m}{\tau}\right]\tau}\right\}\frac{1}{\tau},
\end{equation}
The average value of the time series length $\langle{L(\tau)}\rangle$ for a given value of $\tau$ is defined as the average of $\tau$ sets of $L_{m}(\tau)$. If $\langle{L(\tau)}\rangle \propto \tau^{-D}$, for the range $\tau_{min} < \tau < \tau_{max}$ then the time series is a fractal and has a dimension $D$ for that range of $\tau$. We find $\langle{L(\tau)}\rangle$ for $\tau=2$ to $\tau=55$ for our analysis, for both $AS$ and for $AS_{Norm}$.
\subsection{Hurst exponent}
Here, we attempt to find the Hurst exponent, $H$, which characterises the persistence of a times series to examine whether the non-periodic variations in the asymmetry time series are a result of a white noise process, an anti-correlated random process, or a correlated random process. We borrow the discussion regarding the Hurst exponent in the context of solar activity data from \citet{1994SoPh..149..395R} and \citet{2009SoPh..260..441S}. We use the $R/S$ method as given in \citet{1969WRR.....5..321M} to obtain the Hurst exponent.
We choose a temporal window $\tau$, where $\tau_{t}>\tau>N$, and $\tau_{t}$ is the Theiler window \citep{1986PhRvA..34.2427T}, to make subsets of time series $X(i)$ as follows:
\begin{equation}
x_i(\tau); X(t_0), X(t_0+1), X(t_0+2), ..., X(t_0+\tau-1),
\end{equation}
where $t_0 = 1, 2, ..., N-\tau+1$. It is important to note that the choice of window used in finding the $R/S$ values for each given $\tau$ is rather important, since non-overlapping windows produce $R/S$ values from comparatively small sample sizes. Lesser number of windows could possibly provide inaccurate values for $H$.
Now we denote the average of these subsets as:
\begin{equation}
\bar{x}(t_0,\tau)=\frac{1}{\tau}\sum_{i=t_0}^{t_0+\tau-1}x_i(\tau).
\end{equation}
Let, $S(t_0,\tau)$ be the standard deviation of $x_i(\tau)$ for the window $\tau$ as follows:
\begin{equation}
S(t_0,\tau)=\sqrt{\frac{1}{\tau-1}\sum_{i=t_0}^{t_0+\tau-1}\left[x^{t_0}_\tau(i)-\bar{x}(t_0,\tau)\right]^2}.
\end{equation}
Now, we define a set of new variables $y_i(t_0,\tau)$, which is the set of cumulative deviations from the mean of $x_i(\tau)$
\begin{equation}
y_i(t_0,\tau)=\sum_{k=t_0}^{t_0+i-1}\left[x^{t_0}_\tau(k)-\bar{x}(t_0,\tau)\right],
\end{equation}
and hence, the range $R$ of $y_i(t_0,\tau)$ is obtained as:
\begin{equation}
R(t_0,\tau)=\max_{1\leq i\leq \tau}y_i(t_0,\tau) - \min_{1\leq i\leq \tau}y_i(t_0,\tau).
\end{equation}
This allows us to define the rescaled range measure $R/S$ as:
\begin{equation}
(R/S)(t_0,\tau)=\frac{R(t_0,\tau)}{S(t_0,\tau)}.
\end{equation}
Calculating the $R/S$ values for each temporal window by moving from $t_0=1$ to $t_0=N-\tau+1$ for window size $\tau$, the rescaled range for $\tau$ is then given as the average of these values
\begin{equation}
(R/S)=\frac{1}{N-\tau+1}\sum_{t_0}(R/S)(t_0,\tau).
\end{equation}
It was observed that the rescaled range for a time window is proportional to $\tau^H$
\begin{equation}
(R/S)_\tau=k\tau^H
\end{equation}
where $k$ is the proportionality constant, and $H$ is the Hurst exponent. To obtain the value of the Hurst exponent, $R/S$ values are plotted for $\tau=11$ to $\tau=1709$ for the Hurst Exponent analysis, for both $AS$ and $AS_{Norm}$.
\begin{figure}
\centering
\includegraphics[width=1.0\columnwidth]{MvsD2.eps}
\caption{
Variation of correlation dimension ($D_2$) with the embedding dimension ($M$) for $AS$ (square points) and $AS_{\rm Norm}$ (asterisks). The dotted line
along the diagonal of the figure indicates the $D2$ variation as expected from an ideal stochastic process. $AS^{Ens.Avg.}$ and $AS_{Norm}^{Ens.Avg.}$ (round points) shows the variation averaged over an ensemble of 1000 AS \& $AS_{Norm}$ time series, respectively (see Section \ref{Sec_Error_estimates}).
}
\label{fig:MvsD2}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=.95\columnwidth]{logLkvslogk8.eps}
\caption{
Variation of the length ($\langle{L(\tau)}\rangle$) with the time interval ($\tau$) for $AS$ and $AS_{Norm}$.
}
\label{fig:higuchi}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=.95\columnwidth]{Hurst_RS.eps}
\caption{
Variation of $R/S$ with time interval ($\tau$) for $AS$ (top) and $AS_{Norm}$ (bottom). $\tau=135$ months and $\tau=271$ months are indicated as (1) and (2).
}
\label{fig:HurstExp}
\end{figure}
A white noise process or a random walk process is defined by a Hurst exponent of $H=0.5$. When the time series has $H>0.5$, it is said to be persistent. Persistence is defined as the tendency for the process to have a memory of the previous step. That is, if there was an increase in the value of the time series, the following step would be more likely to have an increase as well. In such a case, the time series would cover more ``distance'' than a random walk would. Whereas, if $0<H<0.5$, then the time series is said to be anti-persistent. That is, an increase in the value of the time series is more likely to be followed by a decrease, and vice-versa. And in opposition to a persistent case, the time series would cover less ``distance'' than a random walk.
There are other methods to determine the Hurst exponent, and the value that is obtained is sensitive to the method used \citep{weron02}. In order to provide a confidence in the estimate of the Hurst exponent, and to ensure that the result is not method dependent, we shall compute Hurst exponent in two more methods, namely, Detrended Fluctuation Analysis and Periodogram Regression. As these methods are well described in
the literature, we shall describe them briefly in Appendix~A.
\section{Results}
The variation of $D_2$ as function of $M$ for the sunspot area asymmetry is shown in \Fig{fig:MvsD2}. We find that the asymmetry and normalized asymmetry do not show the same behaviour.
Nevertheless, in both cases, $D_2$ increases with an increase of $M$. The lack of saturation in $D_2$ implies that the sunspot area asymmetry is not governed by low-dimensional chaos,
rather it might be driven by a high-dimensional or stochastic process.
This conclusion is in general agreement with \citet{carbo93} who also did not find the evidence of low-dimensional chaos in the asymmetry of sunspot area
data during 1874--1989.
To confirm that the solar cycle asymmetry is really governed by stochastic
or high-dimensional chaos, we observe the value of Higuchi's dimension ($D$).
As seen from \Fig{fig:higuchi},
for $AS$, $D = 1.9125 \pm 0.0001$, and for $AS_{Norm}$, $D = 1.9106 \pm 0.0003$.
We know that when the value of $D$ for a curve is close to $2$, the curve behaves nearly like a surface, i.e., the curve is close to a space-filling curve. Hence, the self-similar nature for time series, i.e., the hallmark of low-dimensional chaos is absent. Therefore, we conclude that the process that generates the hemispheric asymmetry of sunspot area is very likely to be the result of an irregular or stochastic process.
Finally, we explore the memory of these irregular asymmetry data by computing
the Hurst exponent ($H$).
In \Fig{fig:HurstExp}, we show the log-log plots for $R/S$ against $\tau$.
The slope of this curve gives the H value. We, however, see two distinguishable linear scaling regimes, and hence one value of H for all $\tau$ is not adequately representing the data. The previous study for sunspot number cycle also indicted two distinct
regimes \citep{2009SoPh..260..441S}.
For $AS$, in the range: $\tau=11$--110 months, we find $H=0.79$,
while for $\tau=263$--650 months, we obtain $H=0.94$
(\Tab{table2}).
In between these two regimes, there is a small region during 135--271 months (marked by vertical lines in (\Fig{fig:HurstExp}a)) with a weaker slope, which is possibly linked to a period of lower persistence in the trend.
For $AS_{Norm}$ (\Fig{fig:HurstExp}b), in the range $\tau=11$--130 months, we find $H=0.81$, while in the range: $\tau=233$--630 months, we get $H=0.87$. In this case, the change in the slope happens very slowly.
\begin{figure}
\centering
\includegraphics[width=.95\columnwidth]{Hurst_DFA.eps}
\caption{ Variation of $DFA$ with time interval ($\tau$) for $AS$ (top) and $AS_{Norm}$ (bottom). $\tau = 135$ and $\tau = 271$ months are indicated as (1) and (2).
}
\label{fig:HurstExp_DFA}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=.95\columnwidth]{Hurst_PR.eps}
\caption{
Periodogram Geweke-Porter-Hudak (GPH) Method for $AS$ (top) and $AS_{Norm}$ (bottom).
}
\label{fig:HurstExp_PR}
\end{figure}
In the Detrended Fluctuation Analysis (DFA), we get similar results but
the values of $H$ are slightly larger (Figure \ref{fig:HurstExp_DFA}).
For $AS$, in the range: $\tau = 1$--110 months, we find $H = 0.84$, while for $\tau = 263-650$ months, we obtain $H = 1.10$.
Now, as discussed in \cite{BS12, CL17}, we can point out that a value of $H$ above 1 is not impossible. It is a consequence of non-stationarity or a trend not being fully removed from the data. The reliability of DFA as a valid method has been questioned on similar grounds before. But since our R/S analysis still backs up the general result that the latter regime has a higher slope, we can safely negate the effect that this inconsistency may cause.
(\Tab{table2}).
For $AS_{Norm}$, in the range: $\tau = 11-130$ months, we find $H = 0.85$, while in the range: $\tau = 223 - 630$ months, we get $H = 0.91$.
Using the Method of Periodogram Regression (Figure \ref{fig:HurstExp_PR}), for $AS$, we obtain $H (= 0.5+0.21) = 0.71$, while for
$AS_{Norm}$, we obtain $H (= 0.5+0.27) = 0.77$.
Unlike the case in $R/S$ Method and Detrended Fluctuation Analysis, the periodogram (Figure \ref{fig:HurstExp_PR}) does not show two distinct linear scaling regimes. However, the Hurst exponent being larger than 0.5, we can safely conclude that $AS$ and $AS_{Norm}$ time series are persistent in nature and the degree of persistence increases with the time-scale. Larger value of $H$ for
$\tau \gtrapprox $ 22 years ($=264$ months), suggest that
the memory of solar cycle asymmetry persists at least for two cycles.
We have obtained $H$
and $D$
graphically, using the Least Squares method for all of our plots; see e.g., \Fig{fig:higuchi}.
We have also applied Bayesian linear regression to all our power law fits \citep{2004ApJ...609.1134W} using the justification towards a Bayesian method of fitting as opposed to a frequentist method, as explained in \citet{D'Huys2016}. We use the Python module, PyMC3 \citep{Salvatier2016} for this.
We find that the differences in results lie in the range of $10^{-4}$ -- $10^{-3}$, except in the case of of the periodogram method. In that case, for $AS$ and $AS_{Norm}$ we obtain the value of the Hurst exponents as $0.707\pm0.179$ and $0.761\pm0.132$, respectively using Bayesian linear regression, while these values obtained from previous Least Squares method are $0.706$ and $0.772$, respectively.
We recall that while computing $AS_{\rm Norm}$ using \Eq{eq:normAS},
we took $AS_{\rm Norm} = 0$ when both $A_N$ and $A_S$ are zero. Instead of this,
if we replace these points by interpolating the neighbouring points, then
this so-called zero-replacement strategy affects our computed results only marginally. (For example, from this zero-replacement $AS_{\rm Norm}$ data,
the computed value of $D$ is 1.9077 and the values of
$H$ are 0.8110 \& 0.8693 (R/S method), 0.8469 \& 0.9074 (DFA), and 0.7718 (PR). Compare these values with
the corresponding values in \Tab{table2}.)
\begin{table*}
\centering
\caption{
Summary of results obtained for Higuchi's dimension and Hurst exponents.
In the forth and sixth columns, the values are the means and the standard deviations (errors) of the results obtained from an ensemble of 1000 AS \& AS$_{Norm}$ time series, respectively (see text).}
\begin{tabular}{llcccc}
\hline \hline
\multicolumn{1}{l}{Method} & \multicolumn{1}{l}{Window} & \multicolumn{4}{c}{Time Series} \\ \cline{3-6}
& (months) & AS & Ensemble of AS & AS$_{Norm}$ & Ensemble of AS$_{Norm}$ \\ \hline
Higuchi & 2-55 & \multicolumn{1}{l}{1.912} & \multicolumn{1}{c}{1.915$\pm$0.002} & 1.911 & \multicolumn{1}{c}{1.916$\pm$0.002} \\ \hline
\multicolumn{1}{c}{Hurst (R$/$S)} &11-110 & 0.788 & 0.783$\pm$0.008 & 0.811 & \multicolumn{1}{c}{0.813$\pm$0.007} \\
& 263-650 & \multicolumn{1}{l}{0.933} & \multicolumn{1}{c}{0.929$\pm$0.019} &0.871 & \multicolumn{1}{c}{0.879$\pm$0.017} \\ \hline
\multicolumn{1}{c}{Hurst (DFA)} & 11-110 & 0.840 & 0.834$\pm$0.010 & 0.851 & 0.851$\pm$0.008 \\
& 263-650 & \multicolumn{1}{l}{1.102} & \multicolumn{1}{c}{1.096$\pm$0.022} & 0.912 & \multicolumn{1}{c}{0.920$\pm$0.018} \\ \hline
\multicolumn{1}{c}{Hurst (PR)} & K = 28 & 0.706 & 0.704$\pm$0.039 & 0.772 & 0.775$\pm$0.028 \\ \hline
\end{tabular}
\label{table2}
\end{table*}
\section{Error Estimates}
\label{Sec_Error_estimates}
In our study, we used the monthly averaged hemispheric sunspot area as recorded in RGO. Unfortunately, in these data, no error information is given. Therefore we cannot make a direct estimate of error in our computed results. However, using the daily sunspot area
data$^1$,
we can make some estimate of the errors in the following way
using a Bootstrapping technique \citep{bootstrap}.
Let us consider the daily sunspot area data of one month for both the northern and southern hemispheres. We produce 100 resampled datasets with the same size as the number of days for this month by randomly selecting daily pairs of values for north and south, and then computing the corresponding $AS$ and $AS_{Norm}$ for the resamples. We compute the mean for all of these resamples. And then, we compute {\it the mean} ($\mu$) and {\it the standard deviation} ($\sigma$) of the means of all the resampled datasets for a month.
It can be easily seen that this mean is not necessarily the same as what we have used in our earlier analyses.
With these $\mu$ and $\sigma$, we produce an ensemble of 1000 data points (deviate) from a Gaussian distribution and repeat this for all the months to get 1000 time series. Finally, perform our all the nonlinear time series analyses with this ensemble.
Black and magenta filled circles connecting dashed lines in \Fig{fig:MvsD2} show the average $D_2$ behaviour of the ensemble of 1000 $AS$ and $AS_{Norm}$ time series. The error bar represents the $\sigma$ of the computed $D_2$ of 1000 time series. In \Tab{table2}, forth and sixth columns show the mean and error (standard deviation) of $D$ and $H$ from the ensembles.
We clearly see that the mean values of the computed quantities ($D_2$, $D$, and $H$) of the ensemble of 1000 $AS$
\& $AS_{Norm}$
time series are not too far from the ones computed from the original monthly mean time series.
The values of $\sigma$ of the ensemble are also reasonably low, with the quantities being less than 1 standard deviation away from the computed values in most cases.
\section{Discussion and Conclusion}
\label{sec:conclusion}
We have explored some nonlinear properties of the underlying process behind the solar cycle asymmetry using nonlinear time series analysis. For this, we have used the hemispheric monthly sunspot area and number time series, which are the best proxies of the Sun's large-scale magnetic flux available for a longer duration.
Following the literature,
solar cycle asymmetry has been measured in two ways, namely, the hemispheric difference $AS$ and the normalized hemispheric asymmetry $AS_{Norm}$.
We have used three methods of time series analyses to characterise the data.
From the analysis of the correlation dimension $D_2$, we find that the value of $D_2$ does not saturate for higher values of $M$. This indicates that there is no underlying presence of a low-dimensional chaotic attractor that could govern the asymmetry of sunspot area data, in agreement with the conclusion obtained in \citet{carbo93}.
In other words, we can expect that the
asymmetry is likely to be
produced by irregular process.
In our fractal analysis, we see that the value obtained for the Higuchi's fractal dimension ($D$) is close to $2$, which implies that a stochastic process or possibly a high-dimensional chaotic process is the cause of the asymmetry.
In all three methods of computation of Hurst exponent,
we
find the value of Hurst exponent $H$
is above 0.7 for $AS$ data and a little larger
for $AS_{\rm Norm}$.
We find multiple values of $H$ for the same time series. Its value
decreases slightly after about 11 years (one cycle period) and then increases for windows larger than about 22 years. This change in the value of H and thus the persistence is more prominent in $AS$. In general, a memory can be observed for the asymmetry time series and it is larger in long-time scale (beyond 22 years; two cycles).
From our analysis, we conclude that the monthly hemispheric asymmetry of sunspot area is dictated by a stochastic process with some amount of long-memory.
The results from hemispheric sunspot number, which is recently made available
by \citet{Veronig21} during 1874--2020, also shows similar behaviour (during the period May 1874 - September 2016) as
that found in the hemispheric area data; \Sec{sec:HemSSN}.
Stochastic nature of hemispheric asymmetry supports the previous theoretical studies \citep[e.g.,][]{GC09, OK13, KM17, SC18, HN19, Nepo19} which explains the solar cycle asymmetry to be caused by the
irregularity involved in the helical convective flow, and in particular the randomness involved in the Babcock--Leighton process (e.g.,
in the form of tilt of BMRs, emergence rate, meridional flow). Further, the presence of some long-term memory in asymmetry
time series supports the existence of a finite memory of the sun's magnetic field, which is possibly determined by the turbulent diffusion and pumping \citep{CC06, KN12, KM17, KK21}.
\section*{Acknowledgement}
B.B.K. thanks Banibrata Mukhopadhyay, Jayanta Dutta and Vinita Suyal for many discussion on time series analysis and help in writing codes during his PhD time.
He also thanks Bibhuti Kumar Jha and Prasun Dutta for the discussion on error analyses.
Authors thank the anonymous referee who provided us valuable feedback on the earlier versions of this paper.
B.B.K. acknowledges the funding from Department of Science and Technology (SERB/DST), India through the Ramanujan Fellowship (project no SB/S2/RJN-017/2018)
and the support provided by the Alexander von Humboldt Foundation during a part of this project. A.G. acknowledges Kishore Vaigyanik Protsahan Yojana (KVPY) for scholarship.
R.D. acknowledges the DAE Incentive Scheme for Holistic Science Education and Augmentation (DISHA) for scholarship.
\section*{DATA AVAILABILITY}
Sunspot area data used in the present study is obtained from the Royal Greenwich Observatory; http://solarscience.msfc.nasa.gov/greenwch.shtml.
Data of our analyses presented in the article will be shared upon reasonable request to the corresponding author.
\bibliographystyle{mnras}
|
\section{Introduction}
\label{sec:introduction}
The application of \ac{ML} promises far-reaching potentials across industries~\citep{Iansiti.2020}. \ac{ML} has already proven successful in many areas, such as web search or recommender systems in e-commerce, in which a lot of high-quality data exists~\citep{LeCun.2015}. While researchers address \ac{ML}'s growing demand for compute power and use of data with, e.g., distributed \ac{ML} approaches where multiple computing nodes share their resources~\citep{Galakatos2018DistributedML, Kairouz_2019_Advances, Verbraeken2019Survey} and quality issues with data processing, access to data is not only a technical issue. Both traditional \ac{ML} and distributed \ac{ML} approaches assume that their training data is centralized by nature, preventing the applicability of \ac{ML} approaches to domains in which data is sensitive and distributed at the same time. To avoid that \ac{ML} approaches must rely on data to which only a centralized organization or individual has full access, \ac{FL} can aggregate the less sensitive \ac{ML} models that were independently and locally trained by individual clients~\citep{Li_2020_Federated,Yin2021flsurvey}. Consequently, \ac{FL} can enable the use of \ac{ML} applications in domains with strong privacy requirements and contribute to solving the challenge of limited access to sensitive data without invading participating clients' privacy~\citep{Larson_2020_Ethics,yang2019federated}. For instance, researchers at Google aimed to improve next word predictions for mobile devices based on private user data, i.e., the words users are typing~\citep{DBLP:journals/corr/abs-1811-03604}. In the case of autonomous driving, \ac{FL} could reduce the data transmission overhead in vehicular networks while still respecting privacy requirements~\citep{elbir2020federated}. These examples demonstrate \ac{FL}'s capability to avoid the obligation to centralize data. Thus, \ac{FL} approaches improve functionality or even enable new value creation scenarios for \ac{ML}.
Despite these promising applications and developments, \ac{FL} in practice has not yet encountered broad adoption~\citep{Aledhari_2020_Federated}. This can be traced to a variety of design requirements that have not been met simultaneously so far.
For example, \ac{FL} systems require, amongst others, privacy guarantees exceeding \ac{FL}'s privacy by design~\citep{Kaissis2020Secure} as well as high degrees of fairness and integrity~\citep{Aledhari_2020_Federated,Yang_2019_Federated}.
Compared to centralized \ac{ML}, \ac{FL} already ensures a certain level of privacy for participating clients~\citep{Aledhari_2020_Federated}. However, even when replacing centralized data with clients' model updates in a \ac{FL} system, these public model updates can still leak insights on private client data~\citep{Kaissis2020Secure,Hitaj2017Deep, Melis2019Exploiting, Nasr2019Comprehensive, Phong2018Privacy, Zhu_2020_Deep}. While research has addressed this issue with different approaches, there are tradeoffs in terms of performance and integrity, as plausibility checks of the contributed models cannot be made any more when the individual model updates are obfuscated. Second, \ac{FL} systems can be subject to malicious client attacks that try to harm the global model performance by submitting model updates that have been trained on data sets unequal to the client's actual data (or even generated on purpose to harm the quality of the aggregate global model). So-called data-poisoning attacks can reduce model performance by up to 90\,\%~\citep{yang2019federated,Aledhari_2020_Federated}.
Third, the above application examples assume that users contribute their data, computation, and communication resources unconditionally. However, scaling \ac{FL} to broad adoption in practice requires fair and transparent incentive mechanisms to appropriately remunerate clients for their contribution to the global model performance~\citep{Kairouz_2019_Advances,Aledhari_2020_Federated,Yang_2019_Federated,Faltings2017Game,Li_2020_review}. \ac{FL} can be subject to free-riding attacks in which malicious clients fraudulently benefit from the incentive mechanism by submitting model updates that are not based on the respective client’s private data~\citep{Aledhari_2020_Federated, Yang_2019_Federated}. Moreover, in a conventional \ac{FL} setting, clients are forced to trust the central entity to remunerate all clients fairly without being able to check whether this central entity acts truthfully or maliciously.
Satisfying the described privacy, integrity, and fairness requirements in a \ac{FL} system whilst still being scalable requires an interdisciplinary discourse that bundles up a combination of technologies within a \ac{FL} system~\citep{Kairouz_2019_Advances, Li_2020_Federated}.
Even though a vast amount of research in various disciplines already exists, there are, to the best of our knowledge, no systems that jointly deliver privacy, integrity, and fair incentives.
Thus, we explore the following research question:\\
\begin{addmargin}[2em]{2em}
\textit{How can a \ac{FL} system achieve fairness, integrity, and privacy whilst still being practical and scalable?}\\
\end{addmargin}
To answer this research question, we propose a \ac{FL} system that levers blockchain technology, \acf{DP}, and \acp{ZKP}. We thus integrate these emerging technologies to provide a novel and smart \ac{FL}-based architecture with the following properties:
\begin{itemize}
\item \textbf{Fair incentives}: Our proposed \ac{FL} architecture measures the individual contribution to the global model performance per client based on the client's actual parameters (i.e., without the \ac{DP}-noise) and incentivizes each client accordingly. By building our \ac{FL} system based on blockchain, a smart contract enforces the transparent and verifiable distribution of incentives~\citep{Kurtulmus2018Trustless, Mugunthan2020Blockflow}.
\item \textbf{Integrity}: Non-interactive \acp{ZKP} enable clients to validate that fellow clients have truthfully trained their submitted model updates based on private data that they committed to earlier, potentially including a proof of provenance (e.g., from a certified sensor). In doing so, these fellow clients do not have to reveal any of their private data, yet we can guarantee that they do the training and evaluation for their incentive truthfully. Further, we build our \ac{FL} system based on blockchain. In the resulting decentralized setting, there is no trust in a central authority needed regarding censorship and the correct aggregation as well as the availability of the global model. Research has already pointed out that a blockchain can be a suitable replacement for intermediaries in a collaborative process and help achieve standardized communication between participants~\citep{bokolo2021collaboration}. Besides, the blockchain-based design ensures neutrality amongst all clients, the immutability of transactions, and the full transparency of the architecture for all clients.
\item \textbf{Privacy}: To make sure that clients' model updates cannot leak information on patterns within their private data, we leverage \ac{DP} to perturb each clients' model update with Laplacian noise.
\end{itemize}
We discuss and instantiate our architecture for multiple \ac{LR} as \ac{FL} model. For the implementation, we use \acp{SNARK} implemented in \href{https://github.com/iden3/circom}{circom} and \href{https://github.com/iden3/snarkjs}{snarkjs} as well as smart contracts deployed on an \ac{EVM} implemented in \href{https://soliditylang.org/}{Solidity}.
Through implementing and testing the system, we demonstrate that a realization of our architecture can achieve reasonable performance and scalability. Moreover, we gain valuable insights into how the combination of \ac{FL}, \ac{DP}, \acp{ZKP}, and blockchain can be applied to more sophisticated \ac{ML} models beyond \ac{FL}.
By proposing our \ac{FL} system that integrates several emerging technologies in a novel way, we contribute a solution that demonstrates that fairness, integrity, and privacy requirements can be solved in practical settings and thus also improve the real-world applicability of existing \ac{FL} approaches. As we pointed out that using \ac{FL} approaches alone is not practical for some application scenarios, we contribute to overcoming relevant hurdles towards applicability. The developed \ac{FL} system is scalable, ensures integrity, and attracts clients to participate through fair incentives as well as data privacy guarantees.
We structure this paper as follows: In Section~\ref{sec:background}, we briefly present the technical building blocks of our \ac{FL} system (i.e., \ac{FL}, \ac{DP}, \acp{ZKP}, and blockchain technology) and discuss related work on the design of \ac{FL} systems. Afterward, we present the architecture and implementation of our \ac{FL}-system in Section~\ref{sec:architecture} and evaluate it in Section~\ref{sec:evaluation}. Finally, we discuss our results, describe our research's limitations, and outline future research opportunities in Section~\ref{sec:conclusion}.
\subsection{Verifiable Computation and Zero-Knowledge Proofs}
\label{subsec:ZKP}
The notion of \acp{ZKP} was first introduced by~\citet{goldwasser1989knowledge}. \acp{ZKP} are a special form of protocols between a so-called \textit{prover} and a \textit{verifier} in which the prover wants to convince the verifier that she/he knows some value with a specific property (more formally, an element of a language). \acp{ZKP} have the additional property that the prover learns \textit{nothing} beyond this statement.
The most important properties beyond zero-knowledgeness that we build upon in this paper are \textit{completeness} (an honest prover will convince the verifier with high probability if the statement is correct) and \textit{soundness} (any and in particular a malicious prover will convince the verifier of a false statement only with a small probability). By replacing the verifier with a random oracle such as a hash function, a large class of interactive proofs can be transferred to non-interactive proofs (using the so-called Fiat-Shamir heuristic~\citep{fiat1986prove}). As opposed to interactive \acp{ZKP}, prover and verifier do not have to interact with each other in non-interactive \acp{ZKP}. Notably, there are \emph{succinct} \ac{ZKP}, which means that both the proof size as well as the computational complexity of the proof verification is considerably smaller than the complexity of checking the statement by conducting the original computation~\citep{bitansky2013succinct}.
Since the introduction of \ac{ZKP}, there has been a lot of research on them, but practical implementations or even applications remained rare before the beginning of the 2010s. However, starting with~\citet{groth2006perfect}, a period of rapid development of \ac{ZKP} towards practical implementations delivered significant performance improvements, e.g., in~\cite{ben2013snarks}. In recent years, different flavors of non-interactive \acp{ZKP} have emerged, for example, Bulletproofs~\citep{bunz2018bulletproofs}, \acp{SNARK}~\citep{gennaro2013quadratic}, \acp{STARK}~\citep{ben2019scalable}, or hybrid constructions. While they differ in scaling properties and cryptographic assumptions, they all allow creating proofs for the correct execution of a program without displaying all inputs, outputs or intermediate steps. Often, Merkle proofs for the inputs are revealed to commit the prover to the usage of unknown but fixed variables. Domain-specific languages such as \href{https://github.com/iden3/circom}{Circom} or \href{https://github.com/starkware-libs/cairo-lang}{Cairo} allow compiling programs into arithmetic circuits. From these, polynomials are constructed, which in turn can be translated into proving and verification programs through libraries such as \href{https://github.com/scipr-lab/libsnark}{libsnark} or \href{https://github.com/iden3/snarkjs}{snarkjs}. As such generic tools for \ac{ZKP} have significantly matured over the last years, they have increasingly been used in first applications; often associated with blockchains and distributed ledgers, where due to redundant execution, cheap verification without revealing sensitive data is important~\citep{ben2019scalable}. However, the generality to which the correctness of computations can be proved with these frameworks is often limited to prime field operations, complemented by libraries that provide, e.g., circuits for basic cryptographic and arithmetic operations such as hash functions, signature schemes, and comparators.
\subsection{Blockchains and Distributed Ledger Technology}
\label{subsec:DLT}
Blockchain and, more general, distributed ledger technology builds upon peer-to-peer networks in which all data is replicated, shared, and distributed across multiple servers (`nodes')~\citep{butijn2020blockchains}. In blockchains, an append-only structure connects batches of transactions (`blocks') linearly through hash-pointers (`chain') and thus achieves decentralized yet synchronized data management. A so-called consensus mechanism that typically combines cryptographic techniques with economic or social incentives allows deciding which blocks to append as well as the order of transactions within a block~\citep{xiao2020consensus}.
If a majority of the network in a specific metric like hash rate (proof of work), the share of cryptocurrency (proof of stake), or the number of accounts in a permissioned network (voting-based consensus mechanisms) is honest, this guarantees the correct execution of simple payments and programming logic (``smart contracts'') and the practical immutability of the ledger~\citep{xiao2020consensus}. The confidence that a majority of the network behaves as intended without the need to rely on the honesty of a distinguished entity is often referred to as digital trust~\citep{nofer2017blockchain}. Consequently, blockchains allow avoiding dependencies on one or a few distinct entities on digital platforms~\citep{alt2020electronic,fridgen2018cross}. The literature distinguishes between permissionless blockchains (such as those used in cryptocurrencies) where anyone can participate and permissioned blockchains where participation is limited, e.g., to an industry consortium or the public sector~\citep{wust2018you}.
Since the release of the Bitcoin whitepaper~\citep{nakamoto2008peer}, blockchain technology has been used in various applications, e.g., cryptocurrencies, decentralized finance with derivatives and non-fungible tokens, or industry applications.
One early and popular permissionless blockchain that supports a Turing-complete programming language for smart contracts is Ethereum. It provides a decentralized virtual machine environment, namely the \acf{EVM}, for executing smart contracts. Ethereum smart contracts are usually implemented in \href{https://soliditylang.org/}{Solidity}, a high-level programming language with syntax similar to \href{https://www.javascript.com/}{JavaScript}~\citep{Wohrer2018Smart}. Two special properties of \href{https://soliditylang.org/}{Solidity} are the lack of non-deterministic libraries (that would otherwise conflict with the necessarily deterministic design of a blockchain that first orders and then executes transactions) and that the complexity of execution has a price, counted in so-called \emph{gas}. This avoids not only infinite loops but facilitates fair competition for the limited capacity.
However, as blockchains and distributed ledgers exhibit \textit{redundant} storage and computation, they suffer from major challenges. While a high energy consumption is often presumed, in fact, only proof of work blockchains are problematic in this regard~\citep{sedlmeir2020energy}. Two other issues that arise directly from replicated transaction storage and execution are considerably more fundamental: Scalability~\citep{gudgeon2020sok} and privacy~\citep{zhang2019security}. Yet, there are innovative approaches to mitigate these challenges. For scalability, countermeasures range from restricting participation and demanding high computational power, storage, and bandwidth from the participating nodes to sharding and off-chain computations. Off-chain computations are also good for privacy, but lead to the challenge of verification in a system with malicious participants. On the other hand, methods for privacy are technologies like \ac{DP}, \ac{FHE}, or \ac{MPC}~\citep{munilla2021revealing}. \acp{ZKP} can be regarded as a special case of verifiable \ac{MPC} where only one participant contributes private data but the result is verified by all blockchain nodes. Yet, \acp{ZKP} are arguably significantly closer to broad adaption and have been leveraged by many blockchain projects so far, starting with \href{https://z.cash/}{Z-Cash} and now also covering many scalability and privacy projects, many of which are implemented on Ethereum (e.g., \href{https://tornado.cash/}{Tornado-Cash}, \href{https://loopring.org/#/}{Loopring}, \href{https://aztec.network/}{Aztec}, \href{https://www.starkdex.io/}{StarkDEX}).
\subsection{Related Work}
\label{sec:related_work}
Since the term \acl{FL} was introduced by \citet{McMahan2017Communication}, research has focused on improving, amongst others, performance, privacy, integrity, and incentive-mechanisms~\citep{Aledhari_2020_Federated,Yang_2019_Federated}.
To improve the clients' privacy beyond the level that \ac{FL} inherently offers, research came up with three main strategies, namely homomorphic encryption, \ac{MPC}, and DP, which aim to prevent public model updates from leaking private client information. Due to its low complexity and strong privacy guarantee~\citep{Kairouz_2019_Advances}, DP is widely used~\citep{Li_2020_Federated}, even though deploying DP in \ac{ML} leads to a trade-off between maximizing privacy (i.e., adding noise with high variance) and maximizing accuracy. In practice, instead of uploading the actual weights, clients can add DP noise to their weights. For example~\citep{Lu2020Differentially} developed a \ac{FL} system for vehicular networks that combines \ac{DP} with gradient descent to avoid attacks that leak private information from publicly available \ac{ML} model updates. Instead of perturbing model parameters, \ac{DP} noise can also be added to the training data~\citep{Yin2021flsurvey}. However, this approach cannot provide privacy protection since it is not sufficient to make any single record unnoticeable~\citep{domingo2021limits}.
Besides, many works have adopted game-theoretic approaches to motivate clients' participation in a \ac{FL} system and ensure fairness amongst them. As an example, \citet{Khan_2020_Federated} implemented a Stackelberg game to incentivize clients for contributing to training a model and, at the same time, maximize the model's performance. Since clients must trust the central authority to incentivize all clients fairly in a traditional \ac{FL} setting and to provide a decentralized incentive layer for data sharing in general~\citep{jaiman2021user}, researchers have suggested using blockchains and smart contracts for model aggregation (e.g., \citep{Rmanan2020Baffle}) and client remuneration (e.g., \citep{Toyoda2019Mechanism} or \citep{Kang_2019_Incentive}).
\citet{jin2021fl} propose a \ac{FL} architecture that leverages a blockchain for transparency and incentivizing clients and combines it with homomorphic encryption in the aggregation smart contract to prevent leakage from clients' contributions.
However, in all these frameworks, offering incentives to clients also puts the system's integrity at risk as malicious clients may try to fraudulently benefit from the incentive mechanism, e.g., through free-riding attacks~\citep{Aledhari_2020_Federated, Yang_2019_Federated}.
The transparency of smart contracts could help achieve integrity by allowing clients to recalculate the weights that fellow clients submitted. However, even when ignoring the corresponding privacy and scalability challenges, such an approach would lead to a ``verifier's dilemma''~\citep{Luu201Demystifying}, where clients weigh up between accepting the costs of recalculation or trusting other clients. \Acp{ZKP} could offer an efficient solution to the ``verifier's dilemma'' in \ac{FL} and, thus, pave the way to achieving fairness, integrity, and privacy at the same time. Despite \acp{ZKP}' potential for \ac{FL} systems and their increasing adoption (especially in the blockchain domain), it remains an open question how \acp{ZKP} can be used in the context of \ac{FL}~\citep{Kairouz_2019_Advances}.
\citet{Wu2018DIZK} have implemented a \ac{SNARK} that proves the correctness of \ac{LR} parameters by recalculating them. In the case of \ac{LR}, this can be done only using matrix multiplications. However, this approach includes rounding a matrix inverse and, hence, requires further measures to ensure full tamper protection (as we will show in detail in Section~\ref{sec:arch_conarch_parameters}). \citet{Feng2021ZEN} introduced a toolchain to produce verifiable and privacy-preserving \acp{SNARK} that prove correct inference in classification and recognition tasks by taking an existing neural network as input. Also \citet{Zhang2020Zero} as well as \citet{weng2021mystique} implemented \acp{ZKP} that allow verifying whether a particular prediction by a trained \ac{ML} model has been computed truthfully without providing any information about the \ac{ML} model itself. Even though also \citet{weng2021mystique}'s work implements a \ac{ZKP} merely for \ac{ML} model inference, they improved \acp{ZKP}' efficiency to prove the correctness of matrix multiplications and \acp{ZKP}' application to floating point arithmetic, which are both essential ingredients for training \ac{ML} models.
Despite the advancement that these works generate for combining \acp{ZKP} and \ac{ML}, it remains, to the best of our knowledge, still unclear how recent progress in \acp{ZKP}, blockchain technology, and DP can be combined as a technology stack that achieves full privacy, integrity, and fairness in \ac{FL} systems.
\section{Foundations}
\label{sec:background}
\subsection{Federated Learning}
\Acl{FL} describes the concept of training local \ac{ML} models on distributed and private client data without transferring the data beyond the client's reach. After training the local parameters, clients in the \ac{FL} system submit model updates derived from their locally trained parameters to a server that aggregates all local model updates to a global model~\citep{Kairouz_2019_Advances}. The types of potential clients are diverse and range from organizations and mobile devices equipped with sensors to autonomous vehicles. Research has recently applied \ac{FL} systems in several domains including health, the \ac{IoT}, vehicular networks, finance, sales, or smart homes~\citep{Li_2020_Federated, yang2019federated}.
By bringing the computation to the data, \ac{FL} improves clients' privacy~\citep{Li_2020_Federated}. Besides, \ac{FL} can increase the efficiency of existing infrastructure and devices by avoiding the transfer of large data sets to a central server and by utilizing the computational power of edge devices like smartphones or wearables~\citep{Li_2020_review}. In doing so, \ac{FL} is typically associated with the following optimization problem~\citep{Li_2020_Federated}:
\begin{align}
\min\limits_{w_g} F(w_g) = \min\limits_{w_g} \left( \sum\limits_{i=1}^{|I|} b_i F_i(w_g) \right) ~\text{,}
\end{align}
where $|I|$ is the total number of clients, $b_i \geq 0$, $\sum_{|I|} b_i = 1$ denotes the client's relative impact, and $F_i(w)$ is the local objective function. The global, aggregated weight $w_g$ is mostly derived from the local updates $w_i$ using \ac{FedAvg}, an aggregation scheme that computes a weighted average of the local weights $w_i$~\citep{Nilsson_2018_Performance}. When training on \ac{IID} data, \ac{FedAvg} achieves similar results to centralized learning~\citep{Nilsson_2018_Performance}.
\subsection{Local Differential Privacy}
Differential privacy (DP) has been developed by \citet{Dwork2006Calibrating} to allow for analyzing sensitive and private data in a secure way:
Consider a trusted central authority that holds a data set containing sensitive client information. The key idea of DP is to develop a query function on the sensitive data set that returns the true answer plus random noise following a carefully chosen distribution~\citep{Dwork2014Algorithmic}. For example, in a study that asks participants to report a certain personal property, participants report their binary answer by tossing a coin: They respond truthfully if tails and if heads, the participants toss the coin again and report ``yes" if heads and ``no" if tails. In this simple example, the participants' privacy stems from the plausible deniability of any reported value~\citep{Dwork2014Algorithmic}.
The driving force for this privacy guarantee is randomization since the guarantee must hold regardless of all present or even future sources of background information (e.g., from the internet or newspapers). Achieving this requires understanding the input and output space of randomized algorithms. Formally, a randomized algorithm $\mathcal{M}$ with domain $A$ and (discrete) range $B$ is associated with a mapping from $A$ to $\Delta (B)$, the probability simplex over $B$:
\begin{align}
\Delta (B) = \Bigg\{ x \in \mathbb{R}^{|B|}: x_i \geq 0 ~\mathrm{for~all}~ i ~\mathrm{and}~ \sum\limits_{i=1}^{|B|} x_i=1 \Bigg\}.
\end{align}
Then, on input $a \in A$ and given that the probability space is over the coin flips of $\mathcal{M}$, the randomized algorithm $\mathcal{M}$ outputs $\mathcal{M}(a) = b$ with probability $\left( \mathcal{M}(a) \right)_b$ for every $b \in B$~\citep{Dwork2014Algorithmic}.
Before defining DP, note that DP aims to sanitize a query function such that the presence or absence of an individual in the analyzed data set cannot be determined by just observing the output of the query. Now consider two data sets $D$ and $D'$, that are either equal or differ only in the presence or absence of one individual, as histograms. Moreover, a histogram over a universe $\mathcal{X}$ is an object in $\mathbb{N}^{\mathcal{|X|}}$. Then, both data sets $D, D' \in \mathbb{N}^{\mathcal{|X|}}$ are called adjacent if for the $\ell_1$-norm, $\lVert D - D' \rVert_1 \leq 1$ holds. Eventually, \citet{Dwork2006Calibrating} define DP as follows:
\begin{definition*}
\label{def:DP}
A randomized algorithm $\mathcal{M}$ with domain $\mathbb{N}^{\mathcal{|X|}}$ and range $\mathcal{R}$ satisfies $\epsilon$-differential privacy if for every adjacent data sets $D, D' \in \mathbb{N}^{\mathcal{|X|}}$ and any subset $\mathcal{S} \subseteq \mathcal{R}$ we have
\begin{align}
\mathrm{Pr}[ \mathcal{M}(D) \in \mathcal{S} ] \leq e^\epsilon~\mathrm{Pr}[ \mathcal{M}(D') \in \mathcal{S} ],
\end{align}
where $\epsilon>0$ is a privacy parameter.
\end{definition*}
A common way to achieve DP for a numeric function $f$ is to add noise following a Laplacian distribution:
\begin{align}
\label{equation:LapDistribution}
\mathcal{L}(x~|~0,\lambda) &= \frac{1}{2\lambda} \exp{\left( -\frac{|x|}{\lambda} \right).}
\end{align}
When a user wants to learn $f(\mathbf{x}) = \sum_i x_i$ with $\mathbf{x}\in \{0,1\}$, i.e., the total number of 1's in the data set, $\epsilon$-DP can be achieved by adding Laplacian noise~\citep{Dwork2014Algorithmic, Dwork2010Differential}:
\begin{align}
\label{equation:noisysum}
\Tilde{f}(\mathbf{x}) = \sum_i x_i + q, \qquad \mathrm{where}~ q \sim \mathcal{L} \left(0, \frac{\Delta f}{\epsilon} \right).
\end{align}
and $\Delta f$ denotes the sensitivity of $f$, i.e., the maximum difference of $f$ on data sets that differ in only one element~\citep{Dwork2010Differential}.
There are many practical algorithms for providing DP; and in general, they have many characteristics beyond $\varepsilon$ that have an impact on privacy or utility~\citep{garrido2021get}. One special case of DP is \acf{DP}, where the confusion (i.e., random perturbation) is performed locally by clients and not by a central authority~\citep{Dwork2004Privacy}. By doing so, the central authority cannot infer or access the actual client data. According to Definition \ref{def:DP}, anyone accessing $\mathcal{S}$ cannot distinguish whether the true data set is $D$ or $D'$ with confidence (controlled by $\epsilon$ -- the lower $\epsilon$, the higher the privacy and the lower the accuracy and vice versa). Thus, \ac{DP} ensures plausible deniability for the clients~\citep{Nguyen2016Collecting}.
\input{02.2_ZKP}
\input{02.3_Blockchain}
\input{02.4_RelatedWork}
\subsubsection{Weight Proof \texorpdfstring{$\pi^w$}{}}
\label{sec:arch_implementation_weight}
We start by explaining the implementation of $\pi^w$'s circuit, namely \verb|LinRegParams(...)|. First, we define the private and public inputs according to~\eqref{eq:pi^c}. The main template execution requires various variables (cf. listing~\ref{code:weightproof_input}), some of which are also input signals to the circuit for \href{https://github.com/iden3/circom}{circom}-specific reasons. Providing an untruthful value for some of these variables might make adversarial attacks possible. To make sure that these critical input variables are correct, the circuit requires the respective values as input signals to check equality.
Listing~\ref{code:weightproof_main} provides an overview of the implementation of $\pi^w$. The circuit as the basis for the \ac{ZKP} is structured into five parts:
\begin{itemize}
\item \textbf{Step 1} -- Range proofs for $\varepsilon_{\mu}$ and $\varepsilon_{\sigma}$:\\
To ensure that $\mathbf{D}$'s mean $\mu \approx 0$ and variance $\sigma \approx 1$, a certain accuracy is set by $\varepsilon_{\mu}$ and $\varepsilon_{\sigma}$ respectively. Based on both values, \verb|LinRegParams| checks the accuracy of all $\mu$ and $\sigma$ for $X_1, \ldots, X_k$ and $Y$. For example, setting \verb|in_require_meanxn_acc| $= 3$ would require that the absolute value of every $\mu\:\cdot\:n$ is smaller than $\varepsilon_{\mu} = 10^{-3}$ (taking into account the conversion in~\eqref{equation:input_conversion} using $d$). Analogously, \verb|in_require_varxn| sets the upper bound for $\sigma\:\cdot\:n$ via $\varepsilon_{\sigma}$.
\item \textbf{Step 2} -- Check $rt^{\mathbf{D}}$: \\
This step rebuilds $\mathbf{D}$'s Merkle tree with one data point (a prime field element as matrix entry) at each leaf. To improve the system's performance, we do not hash the particular leaves on the lowest level since hashing is costly. This works because the hashing algorithm operates on big numbers that can be sufficiently large to cover any prime field element. After computing the tree, the template ensures that the computed root equals the public input $rt^{\mathbf{D}}$, which will be compared to the commitment specified at registration when calling the smart contract's method \verb|Clients|.
\item \textbf{Steps 3 and 6} -- Range proofs for $\varepsilon_{\mathrm{inverse}}$ and $\varepsilon_{w^{\prime}}$: \\
To verify the upper bound on $\varepsilon_{\mathrm{inverse}}$, the template checks the proximity of every entry in $\left(( \mathbf{X}^{\intercal} \mathbf{X}) \mathbf{Z} - \mathbbm{1} \right)$ to~$0$. For example, setting \verb|in_require_XX_acc| $= 3$ would require that the absolute value of every entry is smaller than $\varepsilon_{\mathrm{inverse}} = 10^{-3}$ (again taking into account the conversion in~\eqref{equation:input_conversion} using $d$). The same applies to \verb|in_require_b_noisy_acc|, $\varepsilon_{w^{\prime}}$, and $w^{\prime} - \tilde{w}^{\prime}$.
\item \textbf{Steps 4 and 5} -- Range proofs for $\vartheta_{\mathbf{Z}}$ and $\vartheta_{\mathbf{X}^{\intercal} Y}$ \\
Both $\vartheta_{\mathbf{Z}}$ and $\vartheta_{\mathbf{X}^{\intercal} Y}$ must be provided as inputs to the template in absolute numbers. The main template first finds the largest (by absolute value) entry in the matrix or vector using the maximum norm:
\begin{align}
\lVert \mathbf{A} \rVert = k\cdot\max\limits_{s,t} |a_{s,t}|
\end{align}
where $\mathbf{A}$ can be either $\mathbf{Z} \in \mathbb{N}^{\left( k+1 \right) \times \left( k+1 \right)}$ or $\mathbf{\mathbf{X}^{\intercal}Y} \in \mathbb{N}^{k+1}$. Then it checks whether \mbox{$0 \leq \lVert\mathbf{A}\rVert \leq \vartheta_{\mathbf{A}}$} holds for both choices of $\mathbf{A}$.
\end{itemize}
Note that along with these six steps, the template performs the computation of $w^{\prime}$ iteratively. The \ac{DP} noise is added in the last step by generating $h_j$ using~\eqref{equation:DP_hash_j}, choosing the respective \ac{DP} noise $q$ from $L$ using~\eqref{equation:DP_L} and~\eqref{equation:DP_p}, as well as adding $q$ as in~\eqref{equation:DP-noise}.
\subsubsection{Cost Proof \texorpdfstring{$\pi^c$}{}}
\label{sec:arch_implementation_cost}
Next, we outline the implementation of $\pi^c$ in our proposed system, for which \verb|LinRegCost(...)| is the main template. $\pi^c$'s private inputs (cf. lines 3 to 10 in Listing~\ref{code:costproof_input}) resemble those of $\pi^w$ except for $w^+$ and $\mathrm{Sign}(w_i)$, which are required to reproduce $c$ as outlined in Section~\ref{sec:arch_conarch_cost}. Note that the inputs of $\mathbf{D}_{\mathrm{test}}$ in lines 12 to 15 of Figure~\ref{code:costproof_input} are declared as \verb|private| signals due to performance reasons (cf. Section~\ref{sec:evaluation} for details), even though they are publicly available. Also $\pi^c$ consists of five major steps to make sure that clients compute and submit their cost $c$ truthfully (cf. Listing~\ref{code:costproof_main}):
\begin{itemize}
\item \textbf{Steps 1 and 2} -- Check $rt^{\mathbf{D}}$ and $rt^{\mathbf{D}_{\mathrm{test}}}$: \\
In addition to checking the Merkle tree root $rt^{\mathbf{D}}$ (as in \verb|LinRegProof(...)|), $\pi^c$ also ensures that clients calculate their $c$ based on the test data set $\mathbf{D}_{\mathrm{test}} = \left(\mathbf{X}_{\mathrm{test}}~Y_{\mathrm{test}}\right)$ by requiring $\tilde{rt}^{\mathbf{D}_{\mathrm{test}}} = rt^{\mathbf{D}_{\mathrm{test}}}$. $\mathbf{D}_{\mathrm{test}}$ is a standardized, publicly available, and central data set that is made available to all clients through a cloud service. Besides, $\mathbf{D}$'s standardization does not need to be checked again, as $\pi^w$ already ensures its standardization and $\pi^c$ verifies that $\tilde{rt}^{\mathbf{D}}$ equals $rt^{\mathbf{D}}$.
\item \textbf{Step 3} -- Range proof for $\varepsilon_w$: \\
Similar to step 5 in \verb|LinRegProof(...)|, we allow controlling $\lVert w - \tilde{w} \rVert$ by performing a range proof using the input $\varepsilon_w$. This means that essentially, the computation of the unperturbed weight $\tilde{w}$ is repeated and, subsequently, its proximity to $w$ is checked analogous to $\pi^w$'s range proof for $\varepsilon_{w^{\prime}}$.
\item \textbf{Steps 4 and 5} -- Check $c$: \\
First, as in \eqref{equation:cost}, step 4 estimates $\hat{Y}_{\mathrm{test}}$. Second, step 5 derives $\tilde{c}$ and ensures that $\tilde{c}$ equals the submitted $c$.
\end{itemize}
\FloatBarrier
\subsubsection{Smart Contract}
\label{sec:arch_implementation_smartcontract}
After introducing \verb|LinRegProof(...)| and \verb|LinRegCost(...)|, this section provides an overview of the governance structure implied by the main smart contract, namely \verb|Clients|.
To reduce \verb|Clients|'s size, we implemented \verb|lib| as a support library. \verb|lib| defines all structs required by \verb|Clients| (cf. Listing~\ref{code:SC_lib}) and repeatedly used methods like, e.g., proof verification calls. We will briefly introduce both major structs \verb|FL_client| and \verb|FL_generic|:
\begin{itemize}
\item \verb|FL_generic|: Contains all global variables for, e.g., defining $k$, $n$, $rt^{\mathbf{D}_{\mathrm{test}}}$, $L$, or the variables to control the range proofs. The client that initially deploys the \verb|Clients| smart contract must instantiate \verb|fl_generic|, which will be the only global instance of the struct.
\item \verb|FL_client|: Upon registering, all clients are assigned to an instance \verb|fl_client| (implemented via a mapping \verb|mapclient| from the particular client's address to the respective instance \verb|fl_client|). \verb|FL_client| includes all client-specific data as, for example, $w^{\prime}_i$, $rt^{\mathbf{D}}_i$, and both proofs.
\end{itemize}
To register, clients call the smart contract's \verb|registerClient(...)| function. As depicted in Listing~\ref{code:SC_main}, calling \verb|registerClient(...)| requires $rt^{\mathbf{D}}_i$ as input, since clients must commit to their data set $\mathbf{D}_i$ upon joining the system. Further, they have to pay the admission fee $B$ defined in \verb|fl_generic| which will be used to distribute the incentive payments later. Moreover, a \verb|clientID|, a mapping \verb|mapID| to connect the clients' addresses with their \verb|clientID|, and the current block hash (at the time of registering) serving as source of public randomness for deriving their \ac{DP} noise $p$ will be set automatically.
Subsequently, clients can upload their $w^{\prime}_i$ by calling the function \verb|uploadBeta(...)| and delivering the following inputs:
\begin{itemize}
\item \verb|t_betaverifier|: Address of the deployed weight verifier contract. The client must deploy this contract before calling \verb|uploadBeta(...)| and provide the respective address such that other clients and \verb|uploadBeta(...)| itself can verify the client's weight proof.
\item \verb|betaproof|: This struct (cf. Listing~\ref{code:SC_lib}) contains the actual proof, which is also the first part of the call data for verifying the weight proof.
\item \verb|beta_noisy_true| is the struct containing the submitted $w^{\prime}_i$.
\end{itemize}
\verb|uploadBeta(...)| will collect the respective public inputs (cf. Listing~\ref{code:weightproof_input}) and, together with the provided \verb|betaproof|, verify the proof $\pi^w$ onchain. Successfully verifying the proof requires correct input data $\mathrm{D}_i$, the correct approximate inverse $\mathrm{Z}$ (accuracy controlled by $d$), and correct computation of $w^{\prime}_i$. If successful, \verb|uploadBeta(...)| saves the submitted \verb|beta| (i.e., the client's $w^{\prime}_i$), sets the client's \verb|betaproof_valid| to \verb|true|, and updates \verb|betaglobal| (i.e., $w_g$). The procedure is analogous (except for updating \verb|betaglobal|) when submitting the cost $c_i$ by calling \verb|uploadCost(...)|, such that we will not further describe the function.
Eventually, to trigger the distribution of incentive payments, only the initial client \verb|t_initialclient| can call the function \verb|Incentivize_clients(...)|. \verb|Incentivize_clients(...)| will calculate the incentive payments $V$ and distribute them to those clients, who have successfully submitted their $w^{\prime}_i$, according to their contribution as described in Section~\ref{sec:arch_conarch_incentives}. However, as this requires trust in the initial client (as only she/he can prompt the incentive payment by calling \verb|Incentivize_clients(...)|), this way of triggering the payoff is probably not suitable in all scenarios. To make the trigger more trustless, the initial client could also specify ways to prompt the payoff. For example, any client could trigger it as soon as a certain threshold of participants is reached, or a certain amount of time (measured in the block number) has passed.
\section{Architecture and Implementation}
\label{sec:architecture}
We organize Section~\ref{sec:architecture} as follows: In Section \ref{sec:arch_conceptual_architecture} we explain how our suggested \ac{FL} system is built up formally. To do so, we divide the \ac{FL} process into four different sections, namely \nameref{sec:arch_conarch_parameters} (Section \ref{sec:arch_conarch_parameters}), \nameref{sec:arch_conarch_aggregation} (Section \ref{sec:arch_conarch_aggregation}), \nameref{sec:arch_conarch_cost} (Section \ref{sec:arch_conarch_cost}), and \nameref{sec:arch_conarch_incentives} (Section \ref{sec:arch_conarch_incentives}). Table~\ref{tab:notation} introduces the notation that we use throughout this section.
\begin{table}[!htb]
\centering
\resizebox{0.6\linewidth}{!}{
\renewcommand{\arraystretch}{0.8}
\setlength{\aboverulesep}{0pt}
\setlength{\belowrulesep}{0pt}
\begin{tabular}{|c|l|}\toprule
\textbf{Symbol} & \textbf{Explanation} \\\midrule
$I$ & Set of indices, $i\in\{1,\ldots,\lvert I\rvert\}$ \\
$u_i$ & Client $i$, $i\in I$ \\
$k$ & Number of features of the linear regression \\
$n$ & Sample size per client of the linear regression (set globally) \\
$d$ & Accuracy of rounded inputs \\
$\mathbf{X}_i$ & $n \times (k+1)$ Matrix of input data including all independent variables \\
$Y_i$ & $n$-dimensional vector of input data containing all dependent variables \\
$\mathbf{Z}_i$ & Approximate inverse of $(\mathbf{X}_i^\intercal \mathbf{X}_i)^{-1}$ \\
$\mathbf{D}_i$ & Private data set of $u_i$ -- an $n \times (k+2)$ matrix: $(X_{i,0}, X_{i,1}\ldots X_{i,k}~Y_i)$ \\
$\mathbf{D}_{\textrm{test}}$ & Public test data set \\
$rt^\mathbf{D}_i$ & Merkle tree root of $\mathbf{D}_i$ \\
$rt^{\mathbf{D}_{\mathrm{test}}}_i$ & Merkle tree root of $\mathbf{D}_{\textrm{test}}$ \\
$l_{\mathrm{train}}$ & Depth of $\mathbf{D}_i$'s Merkle tree (number of levels) \\
$l_{\mathrm{test}}$ & Depth of $\mathbf{D}^{\mathrm{test}}_i$'s Merkle tree (number of levels) \\
$\varepsilon_{\mu}$ & Upper bound for $\mu\:\cdot\:n$ \\
$\varepsilon_{\sigma}$ & Upper bound for $\sigma\:\cdot\:n$ \\
$\varepsilon_{\mathrm{inverse}}$ & Upper bound for $\lVert(\mathbf{X}^\intercal \mathbf{X}) \mathbf{Z} - \mathbbm{1}\rVert$ \\
$\varepsilon_w$ & Upper bound for $\lVert w - \tilde{w} \rVert$ \\
$\varepsilon_{w^{\prime}}$ & Upper bound for $\lVert w^{\prime} - \tilde{w}^\prime \rVert$ \\
$\vartheta_{\mathbf{X}^\intercal Y}$ & Upper bound for $\lVert \mathbf{X}^\intercal Y \rVert$ \\
$\vartheta_{\mathbf{Z}}$ & Upper bound for $\lVert \mathbf{Z} \rVert$ \\
$M$ & Model (corresponds to respective weight vector) \\
$w_{i}$ & Weight vector of $u_i$, corresponding to $M_{i}$ \\
$w_{i}^\prime$ & Weight vector including \ac{DP} noise \\
$\tilde{w}_{i}$ & Recalculated weight vector \\
$\tilde{w}_{i}^\prime$ & Recalculated weight vector including \ac{DP} noise \\
$w_{\mathrm{g}}$ & Weight vector of global model $M_{\mathrm{g}}$ \\
$\varepsilon_{\mathrm{LR}}$ & Error term of \ac{LR} \\
$L$ & Vector including Laplacian \ac{DP} noise \\
$d_{\mathcal{L}}$ & Accuracy of the Laplacian \ac{DP} noise \\
$q$ & \ac{DP} noise \\
$h_j$ & Hash serving as random noise for $q$ \\
$hash\_alg$ & Hash algorithm, i.e., $0$ for `MiMC' or $1$ for `Poseidon' \\
$C$ & Vector of $u_i$'s cost: $(c_1, \dots, c_{|I|})$ \\
$V$ & Vector of $u_i$'s incentive payment: $(v_1, \dots, v_{|I|})$ \\
$B$ & Admission fee payable upon joining the system \\
$\pi(a,b)$ & \ac{ZKP} for statement $a$ with witness $b$ \\
Gen($\pi$) & Construct the \ac{ZKP} $\pi$ (with the proving key) \\
Ver($\pi$) & Verify \ac{ZKP} $\pi$ (with the verification key) \\\midrule
\end{tabular}
}
\caption{Notation used in this paper.}
\label{tab:notation}
\end{table}
\subsection{Conceptual Architecture}
\label{sec:arch_conceptual_architecture}
\subsubsection{Compute and Prove Model Weight}
\label{sec:arch_conarch_parameters}
At the start, participating clients $u_i$ can register at the smart contract \verb|Clients|. To do so, they must commit to the Merkle root $rt^\mathbf{D}_i$ of their private data set $\mathbf{D}_i = (\mathbf{X}_i~Y_i) = (X_{i,0}~X_{i,1}~\ldots~X_{i,k}~Y_i)\in\mathbb{R}^{n \times (k+2)}\sim\mathbb{R}^{n(k+2)}$ with
\begin{align*}
\mathbf{X}_{i} =
\begin{pmatrix}
x_{i,0,1} & x_{i,1,1} & \cdots & x_{i,k,1} \\
x_{i,0,2} & x_{i,1,2} & \cdots & x_{i,k,2} \\
\vdots & \vdots & \ddots & \vdots \\
x_{i,0,n} & x_{i,1,n} & \cdots & x_{i,k,n}
\end{pmatrix}
=
\begin{pmatrix}
1 & x_{i,1,1} & \cdots & x_{i,k,1} \\
1 & x_{i,1,2} & \cdots & x_{i,k,2} \\
\vdots & \vdots & \ddots & \vdots \\
1 & x_{i,1,n} & \cdots & x_{i,k,n}
\end{pmatrix}
\in \mathbb{R}^{n \times (k+1)}
\end{align*}
and
\begin{align*}
Y_i = \left(y_{i,1}, \ldots, y_{i,n} \right)^{\intercal} \in \mathbb{R}^n \textrm{.}
\end{align*}
Prior to computing $rt^\mathbf{D}_i$ and training their \ac{LR} weights, clients must normalize their input data $\mathbf{D}_i$, such that every $X_{i,1},\ldots,X_{i,k},Y_i$ has expectation value \mbox{$\mu=0$} and standard deviation \mbox{$\sigma=1$} (in line with, e.g., \citet{Witten2017Data}). Subsequently, clients train their \ac{LR} weight \mbox{$w_i = (\beta_{i,0}~\beta_{i,1} \ldots \beta_{i,k})^{\intercal}\in\mathbb{R}^k$} by computing
\begin{align}
\label{equation:LinReg_params}
w_i = (\mathbf{X}_i^\intercal \mathbf{X}_i)^{-1} ~ \mathbf{X}_i^\intercal ~ Y_i.
\end{align}
Equation \eqref{equation:LinReg_params} optimizes
\begin{align}
\label{equation:LinReg_RSS}
\min (\text{RSS}) = \min\limits_{w\in\mathbb{R}^k} \left( (Y_i - \mathbf{X}_i w_i)^{\intercal} (Y_i - \mathbf{X}_i w_i) \right)=\lVert\varepsilon_{\mathrm{LR}, i}\rVert_2
\end{align}
where $Y_i = \mathbf{X}_i w_i + \varepsilon_{\mathrm{LR}, i}$ and $\varepsilon_{\mathrm{LR}, i}$ denotes the \ac{LR}'s error term.
Constructing the \ac{ZKP} $\pi^w$ that ensures that $u_i$ computed $w_i$ truthfully (i.e., truly computed the $w_i$ that minimizes the \ac{RSS} using \eqref{equation:LinReg_RSS} based on $u_i$'s originally committed data) requires adapting the protocol to \href{https://github.com/iden3/circom}{circom}'s capabilities that are restricted to prime field operations and, thus, to non-negative integer inputs. To do so, we decide to generally round and subsequently scale all (public and private) \ac{ZKP} inputs $\mathbf{A^{\lvert S \rvert \times \lvert T \rvert}} = (a_{s,t})_{s=1,\ldots,\lvert S \rvert; ~ t=1,\ldots,\lvert T \rvert}$ as in general we cannot assume that $a_{s,t} \in \mathbb{N}_0$. For example, we convert the input $a_{s,t} = 2.5347725$ to $2.53477 \cdot 10^5 = 253477 = \tilde{a}_{s,t}$ in the cae of $d = 5$. Further, to input only positive integers, we introduce a matrix $\mathbf{Sign(A)}$ whose elements indicate the signs of $\mathbf{A}$'s coefficients: $\left(\mathbf{{Sign}(A)}\right)_{s,t}=\mathrm{sign}(a_{s,t})$ where
\begin{align*}
\mathrm{sign}: \mathbb{Z}\to\{-1,1\}, \quad x\mapsto
\begin{cases}
0 & \text{if $x \geq 0$} \\
1 & \text{if $x < 0$}
\end{cases}.
\end{align*}
Then we can write the \ac{ZKP} inputs as
\begin{align}
\label{equation:input_conversion}
\tilde{\mathbf{A}} = 10^{-d}\,\mathbf{{Sign}(A)}\circ \mathbf{\tilde{A}^+}
\qquad \text{where} \quad \tilde{a}^+_{s,t}\in \mathbb{N}_0 \text{.}
\end{align}
Moreover, \mbox{$\lVert \mathbf{\tilde A} - \mathbf{A}\rVert\leq c\cdot 10^{-d}$}, where $c$ depends on the choice of matrix norm. For example. if we use the maximum norm, $c$ can be $0.5$ times the number of columns of $A$.
After this transformation, $u_i$ can compute the \ac{ZKP}. However, as constructing proofs is computationally expensive, designing the \ac{ZKP} generation requires specific consideration and optimizations. For better readability, we will refrain from using client indices $i$ in the following description.
First, to make sure that the client's input data is standardized, we introduce an upper bound for the mean $\mu$. Note that since \href{https://github.com/iden3/circom}{circom} does not allow for non-integer divisions, we apply the upper bound controlling $\mu$ to the sum of all elements in $Y$ and $X_1, \ldots X_k$:
\begin{align}
\label{equation:epsilon_mu}
\begin{split}
\mu_{X_j} \cdot n = \sum\limits_{l=1,\ldots n} x_{j,l} & \leq \varepsilon_{\mu} \\
\mu_Y \cdot n = \sum\limits_{l=1,\ldots n} y_{l} & \leq \varepsilon_{\mu}
\end{split} ~\quad~\forall~j\in\{1,\ldots,k\}.
\end{align}
Similarly, we regulate the data set's variance $\sigma^2$ by applying an upper bound to the squared sums:
\begin{align}
\label{equation:epsilon_sigma}
\begin{split}
\sigma_{X_j}^2 \cdot n = \sum\limits_{l=1,\ldots n}\left(x_{j,l}-\mu\right)^2 & \leq \varepsilon_{\sigma}\\
\sigma_Y^2 \cdot n = \sum\limits_{l=1,\ldots n}\left(y_{l}-\mu\right)^2 & \leq \varepsilon_{\sigma}
\end{split} ~\quad~\forall~j\in\{1,\ldots,k\}.
\end{align}
Hence, $\varepsilon_{\mu}$ and $\varepsilon_{\sigma}$ must be set with respect to $n$.
Recalling \eqref{equation:LinReg_params}, our initial attempt was to compute the inverse $(\mathbf{X}^\intercal \mathbf{X})^{-1}$ within a circuit of the respective \ac{ZKP} by introducing two integers, a denominator and a numerator, per input value. This procedure would allow us to calculate $(\mathbf{X}^\intercal \mathbf{X})^{-1}$ in the circuit (as divisions leading to non-integer values can be replaced by multiplications) without rounding errors (given the input values are precise). However, the approach turned out to have one significant shortcoming: Computing $(\mathbf{X}^\intercal \mathbf{X})^{-1}$ within a circuit using Gaussian elimination quickly results in overflow (i.e., a numerator or denominator can quickly get larger than the size of the prime field). Finding common denominators often requires multiplying all denominators repeatedly. For example, already computing the inverse of $(\mathbf{X}^\intercal \mathbf{X})^{5 \times 5}$ matrices with $d=5$ can cause an overflow (in general, this will depend on the choice of $d$ and $k$, but $2^k\cdot d$ will probably be larger than $78$ in many practical applications). On the other hand, reducing fractions or rounding (using range proofs) before overflow would add further complexity in the Gaussian elimination algorithm. Besides, more complex operations demanded by more sophisticated \ac{ML} algorithms would require to hand-craft a fast numerical solver, so the method does not generalize well. Eventually, this approach would hardly scale to more complicated optimization algorithms.
As matrix inversion inside the circuit or inverting within \href{https://github.com/iden3/circom}{circom}'s prime field is expensive, we decided to follow the approach of~\citet{Wu2018DIZK} to solve the problem by letting clients calculate and provide an inverse $\mathbf{Z} \approx (\mathbf{X}^\intercal \mathbf{X})^{-1}$ \emph{outside the circuit} through a standard solver such as typical matrix inversion tools in \href{https://www.javascript.com/}{Javascript}, \href{https://www.python.org/}{Python}, or \href{https://de.mathworks.com/products/matlab.html}{MATLAB}. We then give this inverse as private input to the circuit (again complying with the above-described conversion to non-negative integer entries). However, due to the finite precision of the matrix inversion, we cannot assume that equality holds. Thus, we apply a slightly different method and only prove the proximity of the calculated inverse to the true inverse. To do so, we check the approximation error of the provided numerical inverse using the following range constraint:
\begin{align}
\label{equation:epsilon_inverse}
\lVert(\mathbf{X}^\intercal \mathbf{X}) \mathbf{Z} - (\mathbf{X}^\intercal \mathbf{X})(\mathbf{X}^\intercal \mathbf{X})^{-1}\rVert = \lVert(\mathbf{X}^\intercal \mathbf{X}) \mathbf{Z} - \mathbbm{1}\rVert \leq \varepsilon_{\mathrm{inverse}} < 1.
\end{align}
As matrix norm, the maximum norm is a convenient choice, i.e., the inequality can be checked index-wise (respecting $k$). Through this bound, we can both control for the effect that the replacement of $(\mathbf{X}^\intercal \mathbf{X})^{-1}$ with its approximation $\mathbf{Z}$ has and ensure that the system remains protected against malicious client attacks.
To do so, we estimate an upper bound for the approximation effects on the distance from $w$ to $\tilde{w}$, with the latter being calculated based on $\mathbf{Z}$, via
\begin{align}
\label{equation:epsilon_w}
\begin{split}
\lVert w - \tilde{w} \rVert
&\leq \lVert (\mathbf{X}^\intercal \mathbf{X})^{-1} \mathbf{X}^\intercal Y - \mathbf{Z} \mathbf{X}^\intercal Y \rVert\\
&\leq \lVert(\mathbf{X}^\intercal \mathbf{X})^{-1}-\mathbf{Z}\rVert \cdot \lVert \mathbf{X}^\intercal Y \rVert \leq \varepsilon_w ,
\end{split}
\end{align}
where we use the submultiplicativity of $\lVert\,\cdot\,\rVert$. Even though all entries of $\mathbf{X}^\intercal$ and $Y$ are normalized, i.e., $X_{1},\ldots,X_{k},Y$ have expectation value $\mu=0$ and the standard deviation $\sigma=1$, $\lVert \mathbf{X}^\intercal Y \rVert$ might still yield large entries. Thus, to control the error $\lVert w - w^\prime \rVert$ between the true weights and the client's approximative result, we desire upper bounds for both $\lVert (\mathbf{X}^\intercal \mathbf{X})^{-1} - \mathbf{Z} \rVert$ and $\lVert \mathbf{X}^\intercal Y \rVert$.
First, we introduce a simple range proof to ensure that $\lVert \mathbf{X}^\intercal Y \rVert$ is bounded:
\begin{align}
\lVert \mathbf{X}^\intercal Y \rVert \leq \vartheta_{\mathbf{X}^\intercal Y}
\quad \mathrm{where}\quad\vartheta_{\mathbf{X}^\intercal Y} \equiv \vartheta_{\mathbf{X}^\intercal Y}\left(k,d\right) \text{.}
\end{align}
In general, the probability that $\lVert \mathbf{X}^\intercal Y \rVert$ is large is very small as the normal distribution decays fast; if by coincidence a single value is big, $d$ might need to be adjusted to increase the accuracy of rounded inputs.
Second, we recall that the true inverse can essentially be replaced by the approximative inverse as long as the approximate inverse is bounded~\citep{Newman1974How}. More precisely, given any matrix norm $\lVert\,\cdot\,\rVert$, a square nonsingular matrix $\mathbf{A}$ and its approximate inverse $\mathbf{Z}$ such that \mbox{$\lVert\mathbf{A} \mathbf{Z} - \mathbbm{1}\rVert \leq \varepsilon_{\mathrm{range}} < 1$}, the bound
\begin{align}
\label{eq:newman_ineq}
\lVert \mathbf{A}^{-1}-\mathbf{Z} \rVert \leq \frac{\lVert \mathbf{Z} \rVert \times \lVert\mathbf{A} \mathbf{Z} - \mathbbm{1} \rVert}{1 - \lVert\mathbf{A}\mathbf{Z} - \mathbbm{1} \rVert}
\end{align}
holds. However, it is not clear that $\mathbf{X}^\intercal \mathbf{X}$ is non-singular, and checking this inside the circuit, e.g., through computing the determinant, would become expensive already for moderate $k$. Fortunately, we can relax the assumption by reconsidering the argument in the derivation of the result in~\citet{Newman1974How}: Provided the candidate for the approximate inverse satisfies $\lVert \mathbf{R}\rVert <1$ for $\mathbf{R}:= \mathbf{A}\mathbf{Z} - \mathbbm{1}$, then the Neumann-series $\mathbbm{1} + \mathbf{R} + \mathbf{R}^2 + \ldots$ converges. Consequently, using the argument in the paper,
\begin{align}
\mathbf{A}\mathbf{Z}\cdot(\mathbbm{1} + \mathbf{R} + \mathbf{R}^2 + \ldots)=(\mathbbm{1}-\mathbf{R})\cdot(\mathbbm{1} + \mathbf{R} + \mathbf{R}^2 + \ldots)=\mathbbm{1}.
\end{align}
It follows that $\mathbf{Z}(\mathbbm{1} + \mathbf{R} + \mathbf{R}^2 + \ldots)$ is a right-inverse of the quadratic matrix $\mathbf{A}$ and hence its unique (left- and right-) inverse, i.e., $\mathbf{A}$ is non-singular. In particular, we learn that if $\mathbf{A}$ was non-singular, we could not find an approximate inverse that satisfies~\eqref{equation:epsilon_inverse}. Applied to $\mathbf{A}:=\mathbf{X}^\intercal \mathbf{X}$, our circuit already implicitly guarantees that $\mathbf{X}^\intercal \mathbf{X}$ is invertible through checking~\eqref{equation:epsilon_inverse}, and~\eqref{eq:newman_ineq} yields
\begin{align}
\label{eq:error_bound_inv}
\lVert (\mathbf{X}^\intercal \mathbf{X})^{-1}-\mathbf{Z} \rVert \leq \frac {\lVert \mathbf{Z} \rVert \times \lVert(\mathbf{X}^\intercal \mathbf{X}) \mathbf{Z} - \mathbbm{1}\rVert} {1 - \lVert(\mathbf{X}^\intercal \mathbf{X}) \mathbf{Z} - \mathbbm{1}\rVert} \leq \vartheta_{\mathbf{Z}} \cdot \varepsilon_{\mathrm{inverse}} ~\text{,}
\end{align}
where $\quad\rVert \mathbf{Z} \lVert ~\leq~ \vartheta_{\mathbf{Z}}$ and $\vartheta_{\mathbf{Z}} \equiv \vartheta_{\mathbf{Z}}\left(k,d\right)$.
Before uploading $w$ to the \ac{EVM} storage, clients must perturb their weight by adding Laplacian \ac{DP}-noise (recall~\eqref{equation:LapDistribution} and~\eqref{equation:noisysum}) to avoid deep leakage~\citep{Zhu_2020_Deep}:
\begin{align}
\label{equation:DP-noise}
w^{\prime} &= w + q, \qquad && q \sim \mathcal{L}(0, \lambda), ~q \in \mathbb{R}^k \\
\label{equation:DP-noise_lambda}
\lambda &= \frac{\Delta}{\epsilon} = \frac{1}{\epsilon} \left( \max \beta_{i,j} - \min \beta_{i,j} \right) \qquad &&i \in \{1,\dots, \lvert I \rvert\}, ~j \in \{0,\dots, k\}.
\end{align}
Note that every entry in $q$ is drawn separately, following $\mathcal{L}(0, \lambda)$. Moreover, since we perturb client weights that will be aggregated, we must set the \ac{DP} security parameter $\epsilon$ globally. As no fellow $\beta_{i,j}$ will be available to any client before uploading the client's own $w$, we must set an estimation for $\Delta$ globally. Recalling that all $w$ are computed based on normalized data, we know that $\beta_0,\dots,\beta_k \in [\text{-}1;1]$ holds. Therefore, we set $\Delta = 2 \cdot 10^d$ as a proxy. Moreover, to generate $\pi^w$, we must ensure that we can reproduce every entry in $q$ while keeping its choice random. Achieving this requires a twofold approach.
First, since \href{https://github.com/iden3/circom}{circom} does not have tools to compute a Laplace-distributed random variable directly, we discretize $\mathcal{L}(0,\lambda)$, introducing a variable $d_{\mathcal{L}} = 10^x$ with $x \in \mathbb{N}$ that defines the interval of a discrete distribution $\mathcal{DL}(0,\lambda)$ of $\mathcal{L}(0,\lambda)$ (i.e., $\mathcal{DL}$'s accuracy). We then construct a vector $L \in \mathbb{R}^{d_{\mathcal{L}}-1}$ whose entries are given by the inverse cumulative distribution function of $\mathcal{DL}(0,\lambda)$:
\begin{align}
\label{equation:DP_L}
L(p) := -\lambda~\text{sgn}\left(p-\frac{1}{2} \right)~\text{ln}\left(1-2\Big|p-\frac{1}{2}\Big|\right), \quad p = \{1,\dots, d_{\mathcal{L}}-1\}.
\end{align}
The finite nature of machines leads to some form of discretization of distributions, which makes DP mechanisms generally vulnerable to attacks~\citep{Mironov2012Significance}. Therefore, there is a research stream dedicated to exploring the effects of discretization on \ac{DP} (e.g., \citeauthor{Balcer2019Differential} (\citeyear{Balcer2019Differential}), \citeauthor{Canonne2021Discrete} (\citeyear{Canonne2021Discrete})). However, in our system, the parameters are currently set in a way that the limitation through $d_{\mathcal{L}}$ will be stricter than that of the precision typically achieved in computer programs.
Second, we draw the underlying randomness $p$ from the random oracle $h_j$ that results from hashing a solid source of entropy like the current block hash of the blockchain and the entry of $Y$ that corresponds to the $\beta_j$ that is being perturbed:
\begin{align}
\label{equation:DP_hash_j}
h_j = H\left(\text{current block hash}~|~y_j \right), \quad j=\{0,\dots,k\}.
\end{align}
So, for example, when perturbing $\beta'_0 = \beta_0 + q_0$, the corresponding entry of $Y$ is $y_0$. The resulting $h_j$ is a 256-bit number. We derive every $p$ by computing
\begin{align}
\label{equation:DP_p}
p = h_j ~\text{mod}~ d_{\mathcal{L}}, \quad j=\{0,\dots,k\} \text{.}
\end{align}
The noise that results from a particular value of $p$, as well as its distribution (which is close to the true Laplace distribution with $\lambda=1$, is illustrated in Figures~\ref{plot:discrete_noise_QQ}~and~\ref{plot:discrete_noise_histogram}).
Essentially, we generate verifiable randomness (similar as used, e.g., in Algorand for electing block producers) deterministically in the circuit by combining entropy from the blockchain and from the training data that was the client previously committed to. This approach ensures that clients cannot influence their \ac{DP}-noise $q$ whilst fellow clients cannot reproduce the $q$ (which would allow them to leak the `unperturbed' $w$) and verify that the noise was produced purely at random: As the client could not predict the entropy taken from the blockchain at the time of committing to $D_i$, and consequently not try different values for $Y$ that yield the desired noise. Additionally, by inputting~$L$, $\pi_w$'s circuits can verify $p$.
\begin{figure}[!htb]
\begin{subfigure}{.5\textwidth}
\centering
\includegraphics[page=1, width=\linewidth, trim=10cm 6cm 10cm 6cm, clip]{Figures/Lagrange-plots.pdf}
\caption{Value of the noise depending on $p$.}
\label{plot:discrete_noise_QQ}
\end{subfigure}
\begin{subfigure}{.5\textwidth}
\centering
\includegraphics[page=2, width=\linewidth, trim=10cm 6cm 10cm 6cm, clip]{Figures/Lagrange-plots.pdf}
\caption{Histogram of the discretized Laplacian noise.}
\label{plot:discrete_noise_histogram}
\end{subfigure}
\caption{Derivation of the discretized Laplacian noise.}
\end{figure}
Now the clients are ready to generate their \acp{ZKP} that allows for verifying the validity of the clients' local computations of $w^{\prime}$. Given the proposed architecture, the \ac{ZKP} $\pi^w(a,b)$ with statement $a$ and witness $b$ that proves that $u$ has truthfully calculated $w^{\prime}$ has the following characteristic:
\begin{align}
\pi^w \left( \left( w^{\prime}, rt^{\mathbf{D}}, \varepsilon_{\mu}, \varepsilon_{\sigma}, \varepsilon_{w^{\prime}}, \varepsilon_{\mathrm{inverse}}, \vartheta_{\mathbf{X}^\intercal Y}, \vartheta_{\mathbf{Z}} \right),
\left( \mathbf{D}, \mathbf{Z} \right) \right).
\end{align}
Besides the above-described range proofs and computation checks, $\pi^w$ ensures a globally set range $\varepsilon_{w^{\prime}}$ for the submitted $w^{\prime}$:
\begin{align*}
\lVert w^{\prime} - \tilde{w}^{\prime} \rVert \leq \varepsilon_{w^{\prime}}.
\end{align*}
Please refer to Section \ref{sec:arch_implementation} for a detailed description of how we implement $\pi^w$.
Using a smart contract and through snarkjs' \verb|export solidityverifier| command, clients can, subsequently to compiling the circuit and generating the proof, upload their $w^{\prime}$ to the \ac{EVM}. The smart contract verifies the proof and, if valid, uploads the client's $w^{\prime}$.
\subsubsection{Model Aggregation}
\label{sec:arch_conarch_aggregation}
For every successfully uploaded $w^{\prime}_i$, the smart contract \verb|Clients| updates $w_{\mathrm{g}}$ by aggregating all weights that have been submitted to the smart contract so far using FedAvg. For a globally constant sample size $n$ in a \ac{LR} setting, the respective aggregation formula equals
\begin{align}
\begin{split}
w_{\mathrm{g}} = \frac{\sum_{i=1}^{\lvert I \rvert} w_i}{\lvert I \rvert} \approx \frac{\sum_{i=1}^{\lvert I \rvert} w^{\prime}_i}{\lvert I \rvert}= w^{\prime}_{\mathrm{g}} \text{,}
\end{split}
\end{align}
depending on $\epsilon$ and $\lvert I \rvert$. Heuristically, the noise is \ac{IID} with expectation value $0$ and finite variance $2\lambda^2$, so by the law of large numbers, the probability that $w_{\mathrm{g}}$ differs significantly from $w^{\prime}_{\mathrm{g}}$ is small (in fact, the deviation is approximately normally distributed with standard deviation $\tfrac{2\lambda^2}{\sqrt{|I|}}$.
\subsubsection{Compute and Prove Model Cost}
\label{sec:arch_conarch_cost}
After submitting and proving their \ac{DP}-weights $w^{\prime}$, clients test their unperturbed $w$ on a global and public data set \mbox{$\mathbf{D}_{\textrm{test}} = (\mathbf{X}_{\mathrm{test}}~Y_{\mathrm{test}}) = (X_{\mathrm{test},0} ~X_{\mathrm{test},1} ~\ldots~X_{\mathrm{test},k}~Y_{\mathrm{test}})\in\mathbb{R}^{n_{\mathrm{test}} \times (k+2)}$} with $X_{\mathrm{test}} = (1, ..., 1)^{\intercal}$, distributed via a public channel (e.g., a cloud provider). Clients have to prove that they computed their model cost $c$ truthfully based on $\mathbf{D}_{\textrm{test}}$ and using their $w$ by constructing a \ac{ZKP} $\pi^c$ with the following characteristic.
\begin{align}
\label{eq:pi^c}
\pi^c \left( \left( c, rt^{\mathbf{D}}, rt^{\mathbf{D}_{\mathrm{test}}}, \varepsilon_w \right), \left( \mathbf{D}, \mathbf{Z}, w \right) \right).
\end{align}
$\pi^c$ ensures that clients used the private input $w$ (i.e., their true and unperturbed weight) to compute $c$ by first reproducing $\tilde{w}$ in a circuit of the \ac{ZKP} and performing the range proof (recall \eqref{equation:epsilon_w})
\begin{align*}
\lVert w - \tilde{w} \rVert \leq \varepsilon_w \quad \text{.}
\end{align*}
Note that the SC \verb|Clients| requires that clients have uploaded a valid $w$ prior to submitting and proving $c$, such that the standardization of $rt^{\mathbf{D}}$ as well as the effects of approximating $\lVert \mathbf{X}^{\intercal} Y \rVert$ through $\lVert \mathbf{Z} \rVert$ are already controlled by $\pi^w$.
Then, as $w$ is a private input to $\pi^c$, we can compute $\tilde{c}$ in a circuit of $\pi^c$ and make sure that
\begin{align}
\label{equation:cost}
\tilde{c} =
\big( Y_{\mathrm{test}} - \hat{Y}_{\mathrm{test}} \big)^{\intercal} \big( Y_{\mathrm{test}} - \hat{Y}_{\mathrm{test}} \big)
= c
\end{align}
with $\hat{Y}_{\mathrm{test}} = \mathbf{X}_{\mathrm{test}}\,\cdot\,w$ holds.
Second, by including $rt^{\mathbf{D}_{\mathrm{test}}}$ as a statement, $\pi^c$ can compute the respective Merkle tree on the test data set $\mathbf{D}^{\mathrm{test}}$ and require that the resulting root $\tilde{rt}^{\mathbf{D}_{\mathrm{test}}}$ equals $rt^{\mathbf{D}_{\mathrm{test}}}$, with the latter being stored in the \ac{EVM}. Since $\mathbf{D}_{\mathrm{test}}$ is publicly available, all clients can easily check its standardization, such that checking $\tilde{rt}^{\mathbf{D}_{\mathrm{test}}} = rt^{\mathbf{D}_{\mathrm{test}}}$ is sufficient. This approach allows the system to guarantee truthful submissions of $c$ without having to store $\mathbf{D}_{\mathrm{test}}$ onchain, which would, depending on $k$ and $n_{\mathrm{test}}$
be costly and challenging. A detailed implementation of $\pi^c$ is again included in Section \ref{sec:arch_implementation}.
\subsubsection{Compute Incentives}
\label{sec:arch_conarch_incentives}
To translate $C = (c_1, \dots, c_{|I|})^{\intercal}$ transparently into each client's incentive payment $v_i$, we choose an efficient approach that can be computed on-chain by the \verb|Clients| smart contract. First, we normalize all submitted and validated $c_i$'s to a vector $C_{\mathrm{norm}} \in \mathbb{R}^{|I|}$ that has expectation value \mbox{$\mu=0$} and the standard deviation \mbox{$\sigma=1$}. We multiply all entries in $C_{\mathrm{norm}}$ with $(-1)$ as lower $c_i$ implies higher model performance and, subsequently, replace negative entries with $0$ (i.e., ensure that all $c_{\mathrm{norm,}i}$ that are higher than the mean of all $c_i$ are not incentivized). Next, we scale the data such that $\sum\limits_{i=0}^{|I|} c_{\mathrm{norm},i} = 1$. Upon entry, clients have to pay an admission fee $B$, such that we can distribute the incentive payments $V = (v_1, \dots, v_{|I|})^{\intercal} \in \mathbb{R}^{+}$ as follows:
\begin{align}
V = B \cdot |I_{\mathrm{valid}}| \cdot C_{\mathrm{norm}}
\end{align}
where $|I_{\mathrm{valid}}|$ is the number of clients with valid $w^{\prime}_i$ and $c_i$.
The payments are distributed by the function \verb|reward_clients()| which can only be executed by the client who has initially deployed the \verb|Clients| smart contract. Moreover, by using \href{https://soliditylang.org/}{Solidity}'s \verb|mapping|, every client (i.e., address) can only register once as the data would be overwritten when joining multiple times and a new \verb|clientID| would be set. We summarize the system outlined above in Table \ref{tab:architecture}.
\begin{table}[!htb]
\centering
\scriptsize
\renewcommand{\arraystretch}{1.5}
\begin{tabularx}{\textwidth}{|c|p{7cm}|X|}\toprule
\textbf{Step} & \textbf{Client \texorpdfstring{$u_{i}$}{ui}} & \textbf{Smart Contract} \\\midrule
\midrule
\multicolumn{3}{|c|}{Training phase}\\
\midrule
1 & \textbf{Join system:} Join the system by fetching the model $M_i$ (i.e., number of features $k$ and the sample size $n$), hashing $u_i$'s database, writing $rt^\mathbf{D}_i$ into storage, and paying the admission fee $B$ & \\
2 & \textbf{Train model:} Train $M_i$ by computing $w_i$ & \\
3 & \textbf{Compute and upload \ac{DP} weight $w^{\prime}_i$:} Compute and upload differentially private (\ac{DP}) weight $w^{\prime}_i$ & \\
4 & \textbf{Compute and submit weight proof $\pi^w_i$:} \newline Gen$\left( \pi^w_i \right) \space \rightarrow \space \pi^w_i$ and upload $\pi^w_i$. $w^{\prime}_i$ will only be uploaded into storage if $\pi^w_i$ is valid & \\
5 & & \textbf{Verify all \ac{DP} weights:} \newline Ver$\left(\pi^w_i\right)~\forall~i\in\{1,\ldots,\lvert I\rvert\}$. Write every $w^{\prime}_i$ into storage if $\pi^w_i$ has been validated \\
\midrule
\multicolumn{3}{|c|}{Testing phase}\\
\midrule
6 & & \textbf{Global model aggregation:} Aggregate all $w'_i$ to $w^{\prime}_{\mathrm{g}} \approx w_{\mathrm{g}}$ and write $w^{\prime}_{\mathrm{g}}$ into storage\\
7 & \textbf{Compute $u_i$'s cost $\mathrm{c}_i$:} Fetch the public test data set and evaluate own model performance (\ac{RSS}) using $u_i$'s true weight $w_i$. Submit own accuracy as $u_i$'s cost $c_i$ & \\
8 & \textbf{Compute and upload cost proof $\pi^c_i$:} \newline Gen$ \left( \pi^c_i \right) \space \rightarrow \space \pi^c_i$ and upload $\pi^c_i$. $c_i$ will only be uploaded into storage if $\pi^c_i$ is valid & \\
9 & & \textbf{Verify all cost proofs:\newline} Ver$\left(\pi^c_i\right)~\forall~i\in\{1,\ldots,\lvert I\rvert\}$ \\
10 & & \textbf{Compute incentives:} Compute and distribute incentive payments $V$ according to $C$ \\
\midrule
\end{tabularx}
\caption{Steps of our protocol for private, fair, and honest \ac{FL}.}
\label{tab:architecture}
\end{table}
\subsection{Implementation}
\label{sec:arch_implementation}
In this section, we describe the implementation of our architecture. Please note that, following \href{https://github.com/iden3/circom}{circom} terminology and in contrast to Section~\ref{sec:arch_conceptual_architecture}, we will now refer to single `circuits' that are callable with input and output signals as `templates', whereas we refer to a set of `templates' that make up a \ac{ZKP} as `circuits'. `Signals' (\verb|signal|) are the basis for defining constraints (using \verb|===|) and can only be assigned a value once (using \verb|<--| or \verb|-->|). Both \verb|<==| and \verb|==>| declare a constraint and assign a signal's value at the same time. On the other hand, \href{https://github.com/iden3/circom}{circom} also includes conventional variables (\verb|var|). As we use snarkjs and hence a framework for proof creation and verification based on \acp{SNARK}, a so-called trusted setup is required initially. It consists of two phases, one of which is independent of the circuit (`powers of tau'\footnote{The `powers of tau' ceremony, also referred to as `phase~1 trusted setup', is a circuit-agnostic \ac{MPC} ceremony where multiple independent parties collaboratively construct common parameters from their secret random values. The parameters allow to obtain a proving and a verification key in a later stage. Note that the \ac{SNARK} protocol's integrity guarantee (``soundness'') is compromised if all parties' random values are exposed. It is, however, important to note that while information about the private inputs to the \ac{MPC} would allow to create fake proofs and hence to violate the integrity of our artifact, the privacy of the participants' data would still be ensured even in this case~\citep{wilcox2021snark}.}) and that we could just import, as well as one that is circuit-specific. For implementation purposes, we conducted the trusted setup alone. However, for practical applications, a group of trusted parties would need to conduct an \ac{MPC} such that the participants in the \ac{FL} system would be confident that at least one group member deleted their input to the \ac{MPC}.
We will include several, truncated source code excerpts throughout the following section. Please find the whole project on \href{https://github.com/timon131/ma_webstorm_v4}{GitHub}\footnote{\url{github.com/timon131/ma_webstorm_v4}}. Note that some templates are taken from \href{https://www.iden3.io/}{iden3}'s \href{https://github.com/iden3/circomlib}{circomlib}.
\input{04.3.1_Weight}
\input{04.3.2_Cost}
\input{04.3.3_SC}
\section{Evaluation}
\label{sec:evaluation}
To evaluate our system, we ran several tests with different parameters. $k$ and $n$ are mainly influencing the complexity of our system, since they determine the number of Merkle tree leaves and, thus, the number of hashes that we compute in a \ac{SNARK}. Note that more specifically, $(k+1)\,\cdot\,n$ for $\pi^w$ and $(k+1)\,\cdot\,(n+n_{\mathrm{test}})$ for $\pi^c$ determines the number of leaves since we leave out $X_{0,i} = (1,\ldots,1)$ and instead only hash $X_{1,i},\ldots,X_{k,i}$ as well as $Y_i$. Besides the number of leaves, also the choice of the hashing function can have a significant impact on the system performance. For example, when switching from `Poseidon'~\citep{grassi2021poseidon} to `MiMC'~\citep{albrecht2016mimc} hashing, the number of constraints of our system roughly increases by a factor $4$. Since this factor remains approximately the same when increasing circuit complexity (i.e., the number of constraints), we tested our artifact only with `Poseidon' hashes on an AWS instance (Ubuntu 20.04, 16 virtual CPU cores, and 64 GB RAM). Instead of \href{https://github.com/iden3/snarkjs}{snarkjs}' default web assembly compiler, we used a native compiler\footnote{\href{https://github.com/Fluidex/snarkit}{snarkit by Fluidex}} (C\texttt{++}) that can handle larger circuits.
Table \ref{tab:evaluation_pi-w_partial} provides an excerpt of our test results (Tables~\ref{tab:evaluation_pi-w_full}~and~\ref{tab:evaluation_pi-c} in the appendix contain the entire test data for $\pi^w$ and $\pi^c$ respectively). Note that we used \href{https://github.com/iden3/rapidsnark}{rapidsnark} to generate the proofs (cf. ``proof gen duration''). Besides, due to their similar implementation (cf. Section~\ref{sec:arch_implementation}), the cost proof's performance test results do not differ significantly from the weight proof's results. We will, therefore, mainly focus on the weight proof's results:
\begin{table}[!htb]
\centering
\scriptsize
\setlength{\aboverulesep}{0pt}
\setlength{\belowrulesep}{0pt}
\renewcommand{\arraystretch}{1.5}
\resizebox{\columnwidth}{!}{
\begin{tabular}{|r|r|r||R{1.25cm}|R{1.25cm}|R{1.4cm}|R{1.6cm}|R{1.25cm}|R{1.6cm}|R{1.25cm}|}\toprule
\textbf{$k$} & \textbf{$n$} & \textbf{$(k+1)\,\cdot\,n$} & \textbf{\# constraints} & \textbf{key gen duration (s)} & \textbf{proof gen duration (s)} & \textbf{verification duration (s)} & \textbf{circuit size (MB)} & \textbf{verification key size (kB)} & \textbf{proof size (bytes)} \\
\midrule
4 & 100 & 500 & 89,819 & 177 & 8.641 & 0.236 & 46 & 25 & 708 \\
4 & 250 & 1,250 & 208,162 & 337 & 19.410 & 0.237 & 102 & 25 & 709 \\
4 & 500 & 2,500 & 404,962 & 659 & 36.868 & 0.237 & 198 & 25 & 708 \\
4 & 750 & 3,750 & 601,520 & 1089 & 57.042 & 0.237 & 310 & 25 & 705 \\
4 & 1,000 & 5,000 & 798,659 & 1360 & 74.945 & 0.232 & 391 & 25 & 705 \\
4 & 1,500 & 7,500 & 1,190,620 & 2571 & 120.736 & 0.263 & 614 & 25 & 705 \\
\midrule
\end{tabular}
}
\caption{Evaluation of essential parameters for building and verifying the $\pi^w$ circuit depending on $k$ and $n$.}
\label{tab:evaluation_pi-w_partial}
\end{table}
Before analyzing our system's performance, we describe an essential optimization that prior tests suggested: We found that the recomputations of the Merkle trees are the artifact's dominating operations in terms of constraints. Therefore, we optimized the computation of $rt^{\mathbf{D}}$ and $rt^{\mathbf{D}_{\mathrm{test}}}$ by hashing six instead of two data points at the leaf level at once and hard-coding `empty' leaf hashes to avoid additional hashing in those Merkle trees that are not entirely filled with data points on the leaf level (depending on $k$, $n$, and $n_{\mathrm{test}}$, not every leaf necessarily represents a data point as the bottom level must -- given our Merkle tree optimization -- always contain $2^{\,\mathrm{depth}}\,\cdot\,3$ leaves). This optimization reduced the circuit complexity by roughly 50\%. Then, computing the circuit-specific trusted setup for a \ac{LR} on a training data set with $k=4$ features and sample size $n=1,000$ per client in a \ac{FL} setting takes roughly $23$ minutes each for $\pi^w$ (cf. Table~\ref{tab:evaluation_pi-w_partial}) and $\pi^c$ (cf. Table~\ref{tab:evaluation_pi-c} in the appendix). Note that, as outlined above, the circuit-specific trusted setup must only be computed once and can then be used by all clients for a particular \ac{FL} learning task.
Using the proving key from the trusted setup, every client must spend roughly $2.5$ minutes on computing both proofs that allow fellow clients to verify their submitted weight $w$ and cost $c$ (for $k=4$ and $n=1,000$). In practical \ac{FL} applications, the proof generation likely runs on less sophisticated machines than our AWS instance. We, therefore, also tested the proof duration on a single CPU core (also for combinations other than $k=4$ and $n=1,000$). The results reveal that the witness generation duration was not sensitive to this limitation, whereas the duration for generating the proof using \href{https://github.com/iden3/rapidsnark}{rapidsnark} increased seven times from $4.3$ to $29.4$ seconds and the proof verification duration four times from $0.24$ to $0.97$ seconds. In total, the proof duration grew $1.4$ times from around $75$ to $103$ seconds. Moreover, the RAM required to generate and validate proofs remains below 4 GB and agnostic to the number of constraints.
Looking at the performance of the system's smart contract on the \ac{EVM}, we find gas cost of around $2.2$~million for all $\pi^w$-related on-chain operations and $920,000$ for all $\pi^c$-related operations. This translates into a total of roughly USD~$1270$ for all on-chain operations (USD~$900$ for $\pi^w$ operations and USD~$370$ for $\pi^c$ operations) per client, given a gas price of 100~gwei and a rate of 1~ETH~=~USD~4,000. While in absolute terms this is prohibitively expensive, it is only $21\%$ of the current public Ethereum target block capacity as well as $10.5\%$ of the block limit. Note that the $\pi^c$ operations are significantly cheaper than the $\pi^w$ operations, mainly since the latter requires the vector $L$ as public input and updates $w_{\mathrm{g}}$. As long as operations on public blockchains are that costly, using a permissioned blockchain like Quorum can allow not only to reduce costs but also allowing for considerably higher throughput~\citep{sedlmeir2021benchmarking}.
To assess our system's scalability, we focus on analyzing the impacts of increasing $k$, $n$, and $n_{\mathrm{test}}$ on the circuit-specific trusted setup, the proof generation and verification, as well as the gas cost of the smart contract. Both Figure~\ref{plot:scalability_pi-w} and Table~\ref{tab:evaluation_pi-w_partial} show results that are in line with the expected scaling properties of \acp{SNARK}: The computation of the circuit-specific trusted setups (i.e., `key gen duration') as well as the time it takes to generate the proofs (i.e., `proof duration') scale linearly with \mbox{$(k+1)\,\cdot\,n$} or \mbox{$(k+1)\,\cdot\,(n+n_{\mathrm{test}})$} (cf. Table~\ref{tab:evaluation_pi-c} in the appendix). Moreover, verification duration, verification key size, and proof size do not significantly change when increasing the circuit complexity. Thus, increasing $k$, $n$, or $n_{\mathrm{test}}$ will have only minor effects on the system's overall computation effort, since the proof duration might grow for an individual client whilst the redundant on-chain proof storage and verification cost will remain constant. Even when translating our results to circuits as huge as those in the \href{https://github.com/hermeznetwork}{Hermez project} (the project also used \href{https://github.com/iden3/circom}{circom} with \href{https://github.com/iden3/snarkjs}{snarkjs} and manages to compute a proof for around $100$~million constraints in few minutes on a server with 64 virtual cores and 1 TB RAM~\citep{hermez2021rapidsnark}), we find promising scaling performance: Given the test results, we expect that $\pi^w$ and $\pi^c$ with \mbox{$(k+1)\,\cdot\,n = 600,000$} data points and \mbox{$n_{\mathrm{test}} = 0.1\,\cdot\,n = 6,000$} would result in roughly $100$ million constraints and hence be feasible to prove in a few minutes to hours per proof (depending on the hardware used; in our setup roughly $2.5$ hours with 16~CPU cores or approximately~$3.5$ hours with one core) whilst handling a huge sample size (e.g.,~$k=5$ and $n=100,000$). In this case, the one-time trusted setups would take a few hours to a few days for each participant of the trusted setup (which we assume will not necessarily be conducted by the clients in practice, but a small to medium-sized group or maybe also research institutions, as could be observed for the trusted setup used in Z-Cash, zkSync, etc.). Besides, we expect the verification duration, verification key size, and proof size to remain constant in this scenario.
\begin{figure}[!htb]
\begin{subfigure}{.5\textwidth}
\centering
\includegraphics[page=1, width=\linewidth, trim=10.9cm 5.5cm 10.9cm 5.5cm, clip]{Figures/Performance-plots_v3.pdf}
\caption{Scaling of the circuit complexity (i.e., the number of constraints) with $(k+1)\,\cdot\,n$.}
\label{plot:scalability_pi-w_constraints}
\end{subfigure}
\begin{subfigure}{.5\textwidth}
\centering
\includegraphics[page=5, width=\linewidth, trim=10.9cm 5.5cm 10.9cm 5.5cm, clip]{Figures/Performance-plots_v3.pdf}
\caption{Scaling of the verification duration, verification key size, and proof size with the number of constraints.}
\label{plot:scalability_pi-w_constant}
\end{subfigure}
\caption{$\pi^w$ scalability analysis (client perspective).}
\label{plot:scalability_pi-w}
\end{figure}
Also, the gas cost for the \ac{EVM} operations is independent of the proof complexity. The only data that must be stored on-chain to verify a proof are the proof itself, the verification key, and the public inputs. Both the proof size as well as the verification key size are constant for every $k$, $n$, and $n_{\mathrm{test}}$. The public inputs' volume only increases when either the number of features $k$ or the \ac{DP} noise discretization interval $d_{\mathcal{L}}$ grows, since this leads to an increase in the number of entries in the weight vector $w^{\prime}$ or the \ac{DP} noise vector $L$. Thus, raising $w^{\prime}$ or $L$ increases the payload size, which in turn leads to a rise in gas cost. However, for moderate $k$ and $d_{\mathcal{L}}$, the total gas cost remains small since the respective proof verification itself is responsible for the major part of the gas cost. Therefore, both proofs can be stored as well as verified cheaply on-chain also for large circuits and the particular gas cost remains, as mentioned above, approximately at $2.2$~million for all $\pi^w$-related on-chain operations and $920,000$ for all $\pi^c$-related operations.
Eventually, a clients' effort does not grow with the size of other clients' underlying data sets or the number of participating clients. Further, the number of on-chain transactions only grows linearly with the number of clients and is independent of the size of the clients' training data as well as the central test data. These results promise high potential also for more sophisticated \ac{ML} applications that require higher data volumes.
Eventually, we refrain from testing the accuracy of our \ac{FL} system. The only difference regarding accuracy between our and a plain \ac{FL} system is the \ac{DP} noise that we add to the clients’ weights; and since aggregation is an average over all local weights plus noise, this difference is just an average of \ac{IID} Laplacian-distributed random variables, which has mean $0$ and standard deviation proportional to $\frac{\varepsilon}{\sqrt{\lvert I\rvert}}$. The value of $\varepsilon$ and $\lvert I\rvert$ hence completely determine the change in accuracy of our approach compared to the literature.
\section{Discussion and Conclusion}
\label{sec:conclusion}
Our paper aims at answering the research question of how a \ac{FL} system can achieve fairness, integrity, and privacy simultaneously whilst still being practical and scalable. After identifying a business need for these requirements and arguing that related work so far has not proposed an architecture that satisfies them, we described our proposed system in Section~\ref{sec:architecture} based on a combination of blockchain, \ac{ZKP}, and \ac{DP}. Our conceptual discussion of the architecture and the experiments that we describe in Section~\ref{sec:evaluation} suggest that our implementation indeed offers a practical solution to confidential, fair, and tamper-resistant \ac{FL} that achieves reasonable performance and scalability properties for \acp{LR}. Next to suggesting a pathway to effectively combining \ac{FL} with blockchain, \ac{ZKP}, and \ac{DP}, we develop a system that allows clients to verifying that other clients truthfully trained their local model and received a fair compensation for participating in the case of
\ac{LR}. Thus, we go beyond existing research that has suggested \acp{ZKP} for verifying \ac{ML} model inference based on already trained models.
However, our research is not without limitations and reveals potential for future research that we will outline in this section to conclude our paper. First, even though our suggested system does currently only support \acp{LR}, we tried to design it as generic as possible to allow for adapting the system to other classes of \ac{ML} protocols. Specifically, we move computationally intensive parts (in our case, the inversion of $\mathbf{X}^\intercal\mathbf{X}$) outside the circuit and only check certain properties of the result (in our case, that the approximate inverse that must be provided as private input to the \ac{ZKP} is indeed close to the true inverse, cf. Section~\ref{sec:arch_conarch_parameters}) rather than recomputing the result. For the case of \ac{LR}, we derived an error bound (see~\eqref{eq:error_bound_inv}) that allowed us to significantly reduce the complexity of the respective circuit while maintaining full tamper resistance. The approach of only proving specific properties that the local weights need to satisfy to indicate the integrity of training may be generalizable to further classes of \ac{ML} models, also considering the recent advancement in using \acp{ZKP} for frequent operations in \ac{ML} (e.g., \citep{weng2021mystique}). Since training \ac{ML} models often involves solving a convex optimization problem where optimality can be checked locally (e.g., all partial derivatives in the allowed directions are $0$)~\citep{bubeck2014convex}, we encourage future research to adapt the suggested architecture to more sophisticated classes of \ac{ML} models beyond \ac{LR}. Furthermore, the aggregation of weights in federated learning is often linear~\citep{Nilsson_2018_Performance}, so our concept of perturbing the local weights with verifiable noise is applicable beyond multiple \acp{LR}: Assuming that the training data is \ac{IID} (a common assumption in \ac{FL} applications), the local weights are also \ac{IID}. Moreover, as we construct the clients' \ac{DP} noise from a random oracle, the noise is also \ac{IID} and has mean $0$ and finite variance by construction. Since the weights and the noise applied to them are by construction independent, by the linearity of the averaging algorithm FedAvg, the error in the global model is a weighted average of local noise and the random variable (noise times weight) is \ac{IID} with expectation value $0$ and finite variance. By the law of large numbers, the error in the aggregated global model hence converges to $0$. Thus, for a large number of clients, the error term in the aggregate model as introduced by \ac{DP} is typically small. It follows that our approach of using \ac{DP} (which has been proposed by several other scholars but is particularly relevant in our system because we can prove not only the correct training but also the correct addition of noise) extends to more sophisticated \ac{ML} models beyond \ac{LR}.
To further improve the performance and practicality of our system, we aim at implementing the \acp{ZKP} via \acp{STARK} for improved proof creation performance, post-quantum security, and eliminating the need for a trusted setup in the future. Further, the system's scalability would further benefit from a recursive verification mechanism that reduces the complexity of verifying weight and cost proofs. This could be implemented by building on batching techniques~\citep{gailly2021snarkpack} or recursive proofs~\citep{bowe2020zexe,chiesa2020fractal} and would facilitate scalability by hierarchical aggregation (e.g., only $\log(\lvert I\rvert)$ verification steps would be required on-chain). Besides, even though recalculating $w$ in $\pi^c$ is not the proof's main complexity driver, we see optimization potential there: One could commit a hash of the weight (without noise) and some random salt in the smart contract (proving that the hash was computed like that), so for computing the costs we only need to prove that we used the pre-image of this hash (without salt) for the computation. This can further reduce the cost proof's number of constraints. Next to optimizations of our own code and the performance improvements in libraries that support the generation of \ac{ZKP} (e.g., \acp{STARK} or \href{https://github.com/iden3/circom}{circom} 2.0), we are also confident that hardware acceleration for faster \ac{ZKP}-related operations and particularly proving will be available soon, as research in this area is already conducted, e.g., by projects that build on Ethereum and that leverage \acp{ZKP}~\citep{matterlabs2020acceleration}. This may allow getting even shorter proof times also on devices that are computationally more restricted than a Laptop.
Moreover, we acknowledge that there are still some attacks on integrity: When clients know the learning task (i.e., in our case, the parameters stored on \verb|fl_generic|) prior to committing their data when joining the system, they could attack the system by manipulating their data set before committing to it. However, reverse engineering the data in order to get a desired (malicious) result for the weights is likely more effort than just contributing arbitrarily chosen weights. Moreover, we expect many use cases for our proposed system to be built on sensor data (e.g., from vehicular networks or health applications). Given this, and the availability of certified sensors (e.g., by means of a crypto-chip on the sensor and a certificate of the manufacturer), as are emerging, for example, in Germany's Smart Meter rollout~\citep{djamali2021asset}, the \ac{ZKP}-based approach could handle this issue by including a proof of authenticity (i.e., a proof that the data was signed by a private key that is bound to a certificate that was in turn signed by a trusted, publicly known entity) for the sensor input data when committing to it. This would be easy to integrate at the costs of an additional signature verification per Merkle tree leaf (around 5,000 constraints per leaf for a Schnorr signature). This still does not protect against physical manipulation (imagine putting a temperature sensor to a place where it is not supposed to be), but may offer a reasonable degree of trust in data provenance in many practical scenarios.
Our current incentive mechanism relies on the existence of a central and public data set as described in Section~\ref{sec:arch_conarch_incentives}. Since we acknowledge that this hypothesis is not always feasible in practice, we will work on developing other, effective incentive mechanisms. In doing so, we intend to ensure the mechanisms' fairness by incorporating the research of~\citet{Shapley1953Value}. In doing so, our construction allows us to perform the local evaluation of each client on their actual weights with and without noise, so we get new degrees of freedom for proving the correct evaluation and deriving fair incentives.
Lastly, to ensure the system's applicability in practice, we plan a twofold approach. First, as we are currently training the test model on the \href{https://www.kaggle.com/camnugent/california-housing-prices}{California Housing Prices} data set, we will also run accuracy tests to assess how our conceptual approach translates into data-related performance in practice. Second, we will evaluate the perspective of enterprises on the new features that the combination of \ac{FL} with blockchain technology, \ac{DP}, and \ac{ZKP} offers.
We believe that the convergence of emerging technologies like \ac{ML} and blockchains in the context of data generated by the IoT, as \citet{guggenberger2021emerging} or \citet{singh2020blockiotintelligence} suggest, combined with privacy-enhancing technologies in the context of sensor data sharing or derived models in data markets~\citep{munilla2021revealing}, has the potential to facilitate many new use cases as well as business models and can inspire the field of computer science. As all these aspects are relevant for \ac{FL}, we are expecting interesting results that further extend our design to more complex models in the future.
\section*{Abbreviations}
\acrodefplural{SHARK}[SHARKs]{succinct hybrid arguments of knowledge}
\acrodefplural{SNARK}[SNARKs]{succinct non-interactive arguments of knowledge}
\acrodefplural{STARK}[STARKs]{scalable transparent arguments of knowledge}
\begin{acronym}[SHARK]
\setlength{\itemsep}{-0.1cm}
\acro{DL}{distributed machine learning}
\acro{DP}[LDP]{local differential privacy}
\acro{_DP}[DP]{differential privacy}
\acro{EVM}[EVM]{Ethereum virtual machine}
\acro{FedAvg}[FedAvg]{federated averaging}
\acro{FL}[FL]{federated machine learning}
\acro{FHE}[FHE]{fully homomorphic encryption}
\acro{GDPR}[GDPR]{general data protection regulation}
\acro{HFL}[HFL]{horizontal federated learning}
\acro{IID}[i.i.d.]{independent and identically distributed}
\acro{IoT}[IoT]{internet of things}
\acrodef{LOO}[LOO]{leave-one-out}
\acro{LR}[LR]{linear regression}
\acro{ML}[ML]{machine learning}
\acro{MPC}[MPC]{multiparty computation}
\acro{PKI}[PKI]{public key infrastructure}
\acro{RSS}[RSS]{residual sum of squares}
\acro{SNARK}[SNARK]{succinct non-interactive argument of knowledge}
\acroplural{SNARK}[SNARKs]{succinct non-interactive arguments of knowledge}
\acro{STARK}[STARK]{scalable transparent argument of knowledge}
\acro{ZKP}[ZKP]{zero-knowledge proof}
\end{acronym}
\section*{Appendix}
\subsection{Selected Code Snippets}
\label{sec:appendix}
\input{Listings/weightproof_input}
\input{Listings/weightproof_main}
\input{Listings/costproof_input}
\input{Listings/costproof_main}
\input{Listings/sc_lib}
\input{Listings/sc_main}
\pagebreak
\subsection{Performance of the \texorpdfstring{$\pi^w$}{} circuit}
\begin{table}[!htb]
\centering
\scriptsize
\setlength{\aboverulesep}{0pt}
\setlength{\belowrulesep}{0pt}
\renewcommand{\arraystretch}{1.5}
\resizebox{\columnwidth}{!}{
\begin{tabular}{|r|r|r||R{1.25cm}|R{1.25cm}|R{1.4cm}|R{1.6cm}|R{1.25cm}|R{1.6cm}|R{1.25cm}|R{1.25cm}|}\toprule
\textbf{$k$} & \textbf{$n$} & \textbf{$(k+1)\,\cdot\,n$} & \textbf{\# constraints} & \textbf{key gen duration (s)} & \textbf{proof gen duration (s)} & \textbf{verification duration (s)} & \textbf{circuit size (MB)} & \textbf{verification key size (kB)} & \textbf{proof size (bytes)} \\
\midrule
4 & 20 & 100 & 26,894 & 84 & 2.981 & 0.238 & 15 & 25 & 707 \\
4 & 30 & 150 & 34,340 & 93 & 3.619 & 0.239 & 20 & 25 & 710 \\
4 & 40 & 200 & 42,834 & 113 & 4.530 & 0.245 & 23 & 25 & 707 \\
4 & 50 & 250 & 50,272 & 117 & 5.091 & 0.239 & 26 & 25 & 705 \\
4 & 100 & 500 & 89,819 & 177 & 8.641 & 0.236 & 46 & 25 & 708 \\
4 & 150 & 750 & 128,362 & 213 & 12.080 & 0.238 & 62 & 25 & 706 \\
4 & 200 & 1,000 & 168,619 & 294 & 15.996 & 0.241 & 86 & 25 & 706 \\
4 & 250 & 1,250 & 208,162 & 337 & 19.410 & 0.237 & 102 & 25 & 709 \\
4 & 300 & 1,500 & 246,658 & 372 & 22.400 & 0.241 & 118 & 25 & 706 \\
4 & 350 & 1,750 & 286,670 & 508 & 27.322 & 0.241 & 150 & 25 & 707 \\
4 & 400 & 2,000 & 325,975 & 558 & 31.217 & 0.241 & 166 & 25 & 707 \\
4 & 450 & 2,250 & 364,709 & 627 & 33.409 & 0.238 & 182 & 25 & 707 \\
4 & 500 & 2,500 & 404,962 & 659 & 36.868 & 0.237 & 198 & 25 & 708 \\
4 & 550 & 2,750 & 444,030 & 713 & 40.447 & 0.242 & 241 & 25 & 708 \\
4 & 600 & 3,000 & 482,522 & 803 & 43.499 & 0.246 & 230 & 25 & 707 \\
4 & 650 & 3,250 & 522,775 & 990 & 49.498 & 0.242 & 246 & 25 & 707 \\
4 & 700 & 3,500 & 562,317 & 1070 & 54.096 & 0.236 & 294 & 25 & 704 \\
4 & 750 & 3,750 & 601,520 & 1089 & 57.042 & 0.237 & 310 & 25 & 705 \\
4 & 800 & 4,000 & 640,825 & 1148 & 59.630 & 0.236 & 326 & 25 & 710 \\
4 & 900 & 4,500 & 718,864 & 1246 & 70.394 & 0.234 & 358 & 25 & 706 \\
4 & 1,000 & 5,000 & 798,659 & 1360 & 74.945 & 0.232 & 391 & 25 & 705 \\
4 & 1,500 & 7,500 & 1,190,620 & 2571 & 120.736 & 0.263 & 614 & 25 & 705 \\
\midrule
\end{tabular}
}
\caption{Evaluation of essential parameters for building and verifying the $\pi^w$ circuit depending on $k$ and $n$.}
\label{tab:evaluation_pi-w_full}
\end{table}
\pagebreak
\subsection{Performance of the \texorpdfstring{$\pi^c$}{} circuit}
\begin{table}[!htb]
\centering
\scriptsize
\setlength{\aboverulesep}{0pt}
\setlength{\belowrulesep}{0pt}
\renewcommand{\arraystretch}{1.5}
\resizebox{\columnwidth}{!}{
\begin{tabular}{|r|r|r|r||R{1.25cm}|R{1.25cm}|R{1.4cm}|R{1.6cm}|R{1.25cm}|R{1.6cm}|R{1.25cm}|R{1.25cm}|}\toprule
\textbf{$k$} & \textbf{$n$} & \textbf{$n_{\mathrm{test}}$} & \parbox[c]{1.4cm}{\textbf{$(k+1)\,\cdot\, (n+n_{\mathrm{test}})$}} & \textbf{\# constraints} & \textbf{key gen duration (s)} & \textbf{proof gen duration (s)} & \textbf{verification duration (s)} & \textbf{circuit size (MB)} & \textbf{verification key size (kB)} & \textbf{proof size (bytes)} \\
\midrule
4 & 20 & 10 & 150 & 20,959 & 82 & 3.005 & 0.243 & 11 & 4.2 & 707 \\
4 & 30 & 10 & 200 & 26,940 & 90 & 3.472 & 0.237 & 13 & 4.2 & 706 \\
4 & 40 & 10 & 250 & 33,971 & 108 & 4.439 & 0.241 & 18 & 4.2 & 704 \\
4 & 50 & 10 & 300 & 39,952 & 114 & 4.951 & 0.250 & 21 & 4.2 & 710 \\
4 & 100 & 10 & 550 & 72,194 & 174 & 8.152 & 0.244 & 38 & 4.2 & 705 \\
4 & 150 & 15 & 825 & 105,940 & 206 & 11.108 & 0.243 & 51 & 4.2 & 708 \\
4 & 200 & 20 & 1,100 & 141,684 & 293 & 15.439 & 0.243 & 74 & 4.2 & 707 \\
4 & 250 & 25 & 1,375 & 176,240 & 339 & 18.568 & 0.247 & 88 & 4.2 & 704 \\
4 & 500 & 50 & 2,750 & 349,881 & 659 & 36.949 & 0.244 & 174 & 4.2 & 708 \\
4 & 1,000 & 100 & 5,500 & 697,265 & 1380 & 73.696 & 0.243 & 346 & 4.2 & 705 \\
\midrule
\end{tabular}
}
\caption{Evaluation of essential parameters for building and verifying the $\pi^c$ circuit depending on $k$, $n$, and $n_{\mathrm{test}}$.}
\label{tab:evaluation_pi-c}
\end{table}
\section*{References}
\bibliographystyle{unsrtnat}
|
\section{Introduction} \label{sect_intro_chapgit}
In this paper, we use geometric invariant theory (GIT) to construct moduli spaces $\mathcal{M}^N_{1,n}$ of linear self-maps of $\mathbb{P}^N$ with $n$ marked points, considered up to projective equivalence.
\begin{defn}
Let $N, n \in \mathbb N.$ A \emph{linear map on $\mathbb{P}^N$ with $n$ marked points}, or \emph{marked linear map}, is a tuple $(T, v_1, \hdots, v_n),$ where $T \colon \mathbb{P}^N \dashrightarrow \mathbb{P}^N$ is the projectivization of a linear map and $v_1, \hdots, v_n \in \mathbb{P}^N$.
\end{defn}
Marked linear maps arise in algebraic dynamics, combinatorial algebraic geometry, and integrable systems; see Sections \ref{sect_alg_dyn}-\ref{sect_integrable_systems} for these connections. Our primary motivation is the problem of understanding moduli of algebraic dynamical systems \cite{MR4007163, MR4366860, MR4132597, MR2741188, MR2884382}. Doyle-Silverman constructed moduli spaces $\mathcal{M}_{d,n}^N$ of degree $d \geq 2$ dynamical systems on $\mathbb{P}^N$ with $n$ marked points \cite{MR4132597}. These moduli spaces are used to study dynamical systems with marked periodic orbits, but little is known about their structure. We construct these moduli spaces in the linear case $d = 1$.
We find that, although linear maps are the simplest algebraic dynamical systems, marking points leads to a remarkably subtle moduli space. We show that $\mathcal{M}^N_{1,n}$ exists, is rational, and admits a dynamically meaningful compactification $\bar{\mathcal{M}}^N_{1,n}$ with a delicate combinatorial structure. In this way, our moduli problem is loosely analogous to the widely-studied moduli space of curves of genus $0$ with $n$ marked points.
We work over a fixed algebraically closed field $k$. Fix integers $N, n \in \mathbb N$. Let $\Mat_{N+1}$ be the space of linear endomorphisms of the vector space $k^{N+1}$, and let $\mathbb{P}(\Mat_{N+1}) = \mathbb{P}^{N^2 + 2 N}$ be its projectivization. In coordinates, the elements of $\mathbb{P}(\Mat_{N+1})$ are $(N+1) \times (N+1)$ nonzero matrices, up to scale. We consider the variety
$$\mathcal{X}_{N,n} \colonequals \mathbb{P}(\Mat_{N+1}) \times (\mathbb{P}^N)^n,$$
equipped with the following action of $\SL_{N+1}$: for each $A \in \SL_{N+1}$,
\begin{equation} \label{eq_acn}
A \cdot (T,v_1, \hdots, v_n) \colonequals (ATA^{-1}, Av_1, \hdots, Av_n).
\end{equation}
The GIT quotient of $\mathcal{X}_{N,n}$ by the action \eqref{eq_acn} is defined relative to a choice of very ample, $\SL_{N+1}$-linearized invertible sheaf $\mathcal{L}$. Note that any very ample invertible sheaf $\mathcal{L}$ has a unique $\SL_{N+1}$-linearization. We write $(\mathcal{X}_{N,n} \GITQ \SL_{N+1})(\mathcal{L})$ for the GIT quotient of $\mathcal{X}_{N,n}$ by the action \eqref{eq_acn} relative to $\mathcal{L}$.
Our main objects of study are the GIT quotients for various sheaves $\mathcal{L}$,
$$\mathcal{M}_{1, n}^N(\mathcal{L}) \colonequals (\mathcal{X}_{N,n} \GITQ \SL_{N+1})(\mathcal{L}).$$
The price of working with GIT is that some geometric objects are so degenerate that they cannot be represented in a quotient variety. Points of $\mathcal{X}_{N,n}$ are classified as \emph{unstable}, \emph{semistable}, or \emph{stable} relative to the sheaf $\mathcal{L}$. Stable points have the desirable property that their orbits are in bijection with the points of a geometric quotient variety. The semistable points are certain degenerations of families of stable points, and they can be used to construct a natural compactification of the stable quotient.
Generally, a geometric object in $\mathbb{P}^N$ is GIT unstable if it has too high an order of contact with some flag in $\mathbb{P}^N$. The classic example is due to Mumford \cite{MR1304906}. A configuration $(v_1, \hdots, v_n) \in (\mathbb{P}^N)^n$ of $n$ points in $\mathbb{P}^N$ is semistable relative to the diagonal action of $\SL_{N+1}$ and the sheaf $\mathcal{O}(1, \hdots, 1)$ if, for each nontrivial proper subspace $H \subsetneq k^{N+1}$,
\begin{equation} \label{eq_mumford_template}
\# \{i \colon v_i \in H \} \leq \frac{n}{N+1}(\dim H).
\end{equation}
Our main theorem characterizes the GIT stable and semistable points of $\mathcal{X}_{N,n}$, in the spirit of Mumford's example. We show that the GIT stability of a point $(T, v_1, \hdots, v_n) \in \mathcal{X}_{N,n}$ can be determined by counting the points $v_i$ lying on certain flags that play a special dynamical role relative to $T$. To state the theorem, we define an invariant that encapsulates the dynamics of $T$ on a given flag.
\begin{defn}
\label{defn:flaginkN1}
A \emph{flag~$\mathcal{H}$ in $k^{N+1}$} is a nested sequence of linear subspaces
\[
\mathcal{H} :
0 \subsetneq H_1 \subsetneq \hdots \subsetneq H_\gamma \subsetneq k^{N+1}.
\]
For notational convenience, we extend the list of subspaces in~$\mathcal{H}$ by setting
\[
H_0 = \{0\}, \quad H_{\gamma+1} = k^{N+1}.
\]
Given a flag $\mathcal{H}$ and $T \in \PP(\Mat_{N+1})$, the \emph{Hessenberg function of $\mathcal{H}$ relative to $T$} is defined by
\begin{align*}
h_{T,\mathcal{H}} : \; & \{0, \hdots, \gamma + 1\} \to \{0, \hdots \gamma + 1\}, \\
& i \mapsto \min \{j : TH_i \subseteq H_j\}.
\end{align*}
Given a flag $\mathcal{H}$ and a point $T\in\PP(\Mat_{N+1})$, we say that~$\mathcal{H}$ is of \emph{Type~I,~II, or~III relative to~$T$} if it has the following properties, although we note that~$\mathcal{H}$ need not satisfy any of these conditions\textup:
\begin{itemize}
\item \textup{Type I:} $\gamma = 1$, and $\mathcal{H} = (H_1)$ satisfies either
\[
\Bigl( 0 \neq TH_1 \subseteq H_1 \Bigr)
\qquad\text{or}\qquad
\Bigl(TH_1 \subseteq H_1 \quad \text{and} \quad T(k^{N+1}) \not\subseteq H_1
\Bigr).
\]
\item \textup{Type II:} The Hessenberg function $h_{T, \mathcal{H}}$ satisfies, for all $t$ in the range $1 \leq t \leq \gamma,$
$$h_{T, \mathcal{H}}(t) = t + 1.$$
\item \textup{Type III:} The Hessenberg function $h_{T, \mathcal{H}}$ satisfies, for all $t$ in the range $1 \leq t \leq \gamma + 1$,
$$h_{T, \mathcal{H}}(t) = t - 1.$$
\end{itemize}
\end{defn}
Intuitively speaking, Type II flags are unfurled by $T$ in $\gamma + 1$ steps, and Type III flags are furled by $T$ in $\gamma + 1$ steps. A map admits a Type III flag if and only if it is nilpotent. Type I flags are simply those with $\gamma = 1$ that are not Type II or Type III.
Our main theorem shows that GIT stability of a marked linear map can be checked on these three types of flags.
\begin{thm}
\label{thm_main}
Let $k$ be an algebraically closed field, and let $N, n, q \in \mathbb N$. We consider the variety
\[
\mathcal{X}_{N,n} = \PP(\Mat_{N+1}) \times (\mathbb{P}^N)^n
\]
equipped with the action \eqref{eq_acn} of $\SL_{N+1}$, the sheaf $\mathcal{L} = \mathcal{O}(q, 1, \hdots, 1)$, and the unique $\SL_{N+1}$-linearization.
\begin{itemize}
\item[\textup{(a)}]
Let $T\in\PP(\Mat_{N+1})$, and let $\mathcal{H}$ be a flag in $k^{N+1}$. Then $\mathcal{H}$ belongs to at most one of the three types I, II, or III relative to~$T$ described in Definition~\ref{defn:flaginkN1}.
\item[\textup{(b)}]
A point $(T,v_1, \hdots, v_n) \in \mathcal{X}_{N,n}$ is GIT semistable if and only if, for every $1\le\gamma\le N$ and every flag $\mathcal{H}=(H_1,\ldots,H_\gamma)$ in $k^{N+1}$ that is of Type~I,~II, or~III relative to~$T$ as described in Definition~\ref{defn:flaginkN1}, we have
\begin{equation} \label{eq_main}
\sum_{j=1}^\gamma \# \{i \colon v_i \in H_j\} \leq \frac{n}{N+1} \sum_{j=1}^\gamma \dim H_j +
\begin{cases}
0 &\text{if $\mathcal{H}$ is Type I relative to $T$,} \\
q &\text{if $\mathcal{H}$ is Type II relative to $T$,} \\
-q &\text{if $\mathcal{H}$ is Type III relative to $T$.} \\
\end{cases}
\end{equation}
\item[\textup{(c)}]
A point $(T,v_1, \hdots, v_n) \in \mathcal{X}_{N,n}$ is GIT stable if and only if, for every flag $\mathcal{H}$, the inequality \eqref{eq_main} in \textup{(b)} is strict.
\end{itemize}
\end{thm}
We deduce Theorem \ref{thm_main} as a special case of Theorem \ref{thm_main_any_sheaf}, which describes GIT stability and semistability relative to any very ample invertible sheaf $\mathcal{L} = \mathcal{O}(q, m_1, \hdots, m_n)$, where $q, m_1, \hdots, m_n \in \mathbb N$.
\begin{example} \label{ex_one_marked_point}
When the number of marked points is $n = 1$, the complicated criteria of Theorem \ref{thm_main} condense as follows. The stable and semistable loci are equal. The pair $(T,v)$ is stable if $T$ is non-nilpotent and $v$ is a cyclic vector for $T$ up to scaling. The stable quotient is simply $\mathbb{P}^N$, recovering the Mumford-Suominen moduli space of linear maps equipped with a cyclic vector \cite{MR0437531}. See Section \ref{sect_one_marked_point} for details.
\end{example}
\begin{example} \label{ex_dim_one}
On $\mathbb{P}^1$, complete flags are in 1-1 correspondence with points. If $T$ is non-nilpotent, then fixed points correspond to Type I flags and all other points correspond to Type II flags. If $T$ is nilpotent, then its indeterminacy point corresponds to a Type III flag and all other points correspond to Type II flags. Thus, Theorem \ref{thm_main} reduces to the following semistability criterion. If $T$ is non-nilpotent, then up to $n/2 + q$ points may coincide at a general point of $\mathbb{P}^1$, and up to $n/2$ points may coincide at each fixed point of $T$. If $T$ is nilpotent, then up to $n/2 + q$ points may coincide at a general point of $\mathbb{P}^1$, and up to $n/2 - q$ points may coincide at the indeterminacy point of $T$. Thus fixed points and indeterminacy points behave somewhat like marked points.
\end{example}
\begin{example} \label{ex_id}
For any $N, n, q \in \mathbb N$, if $T$ is the identity map, then there are no Type II or III flags, and each linear subvariety of $\mathbb{P}^N$ defines a Type I flag. Thus Theorem \ref{thm_main} specializes to Mumford's stability test \eqref{eq_mumford_template}.
\end{example}
\begin{example} \label{ex_large_q}
For fixed choices of $N, n\in \mathbb N$ in Theorem \ref{thm_main}, if $q$ is sufficiently large, then the Type II condition of \eqref{eq_main} is vacuously true, and the Type III condition of \eqref{eq_main} is vacuously false. Thus, if $q$ is sufficiently large, a marked linear map $(T, v)$ is stable relative to $\mathcal{O}(q,1,\hdots,1)$ if and only if $T$ is non-nilpotent and $v$ meets the Mumford stability test \eqref{eq_mumford_template} for each $T$-invariant proper subspace $H \subsetneq k^{N+1}$.
\end{example}
These examples suggest the following interpretation of Theorem \ref{thm_main}: the more that the map $T$ moves around the marked points, the better for GIT stability; the more marked points that eventually enter the kernel of $T$, the worse.
Our work is a starting point for describing the structure of $\mathcal{M}_{1,n}^N$ in more detail. Mumford and Suominen showed that the stable locus for the conjugation action of $\SL_{N+1}$ on $\PP(\Mat_{N+1})$ is empty \cite{MR0437531}. Marking points and using an appropriate sheaf resolves this problem.
\begin{cor} \label{cor_stable_nonempty}
Let $n \geq 1, N \geq 1$, and $q \geq n$. The stable locus in $\mathcal{X}_{N,n}$, relative to the sheaf $\mathcal{L} = \mathcal{O}(q, 1, \hdots, 1)$ with its unique linearization, is nonempty.
\end{cor}
The moduli space $\mathcal{M}^1_d$ of unmarked degree $d \geq 2$ dynamical systems on $\mathbb{P}^1$ is rational \cite{MR2741188}. A central open problem in algebraic dynamics is to determine the rationality of the analogous moduli spaces $\mathcal{M}_d^N$ of dynamical systems on $\mathbb{P}^N$ with $N > 1$. In previous work, we showed that, in the planar case ($N = 2$), if $n \geq 4$, the moduli space $\mathcal{M}_{1,n}^2$ is birational to $\mathbb{P}^{2n}$, by working in explicit coordinates \cite[Theorem 3.2]{weinreich2021}. This generalizes to any dimension:
\begin{thm} \label{thm_rational}
For any $n, N \in \mathbb N$ and any very ample invertible sheaf $\mathcal{L}$, if the stable locus of $\mathcal{X}_{N,n}$ relative to $\mathcal{L}$ is nonempty, the variety $\mathcal{M}_{1,n}^N(\mathcal{L})$ is rational of dimension $nN$.
\end{thm}
\begin{question} We suggest some directions for future research into these moduli spaces.
\begin{enumerate}
\item Does Theorem \ref{thm_main} generalize to other Lie types?
\item What is the isomorphism type of the moduli space $\mathcal{M}_{1,2}^1 (\mathcal{L})$ of 2-marked linear maps on $\mathbb{P}^1$?
\item What is the variation of the isomorphism type of $\mathcal{M}_{1,n}^N (\mathcal{L})$ with $\mathcal{L}$?
\item What is the dimension of the GIT boundary? What are its irreducible components, particularly when $N = 1$?
\item For which $n, N, q \in \mathbb N$ and sheaves $\mathcal{L}$ are the stable and semistable loci equal?
\end{enumerate}
\end{question}
We now sketch the proof of our main theorem, then discuss motivation from three areas: algebraic dynamics, combinatorial algebraic geometry, and integrable systems.
\subsection{Sketch of the proof of Theorem \ref{thm_main}}
The proof of Theorem \ref{thm_main} is via polyhedral combinatorics. There is a standard method in GIT for checking stability of a point. To any maximal torus in $\SL_{N+1}$ and point $p \in \mathcal{X}_{N,n}$, we associate a convex polytope called the \emph{weight polytope} or \emph{state polytope}. A point $p$ is GIT stable if and only if its weight polytope contains the origin.
We find that the weight polytopes arising from the conjugation action of $\SL_{N+1}$ on $\PP(\Mat_{N+1})$ are \emph{root polytopes of $A$ type} \cite{MR4310906}. The vertices of these polytopes are root vectors of the $A_N$ lattice. However, the weight polytopes that arise from the combined action of $\SL_{N+1}$ on $\mathcal{X}_{N,n}$ are rather complicated. To deal with this issue, we introduce a variant of the weight polytope called the \emph{corner polyhedron} (Definition \ref{def_corner}). The corner polyhedron is the Minkowski sum of the standard weight polytope and the nonnegative orthant. The idea is that, since the Weyl group of the $A_N$ lattice is the symmetric group $\mathcal{S}_N$, every 1-parameter subgroup in $\SL_{N+1}$ can be put in a descending form, and the corner polyhedron describes stability relative to descending 1-parameter subgroups. This allows us to frame complicated systems of inequalities, like those in \cite{MR4132597}, in the intuitive language of convex geometry.
We show that the action of $\SL_{N+1}$ on the factor $(\mathbb{P}^N)^n$ has the effect of translating the corner polyhedron for the factor $\PP(\Mat_{N+1})$. Thus we understand the corner polyhedra for the combined action by computing all facets of the corner polyhedra of type $A_N$. Each facet imposes a necessary condition for GIT stability.
Our results on the vertices and facets of corner polyhedra of type $A_N$ are of independent combinatorial interest; they are contained in Proposition \ref{prop_vertices} and \ref{prop_facets}. For similar results about root polytopes, see \cite{MR4310906}. We also introduce a combinatorial gadget called the \emph{matrix profile} (Definition \ref{def_profile}) which classifies corner polyhedra.
Finally, in Section \ref{sect_proof}, we finish the proof. We show that the many necessary conditions for GIT stability imposed by the facets of corner polyhedra can in fact be checked on just Type I, II, and III flags.
\subsection{Motivation from dynamical moduli spaces} \label{sect_alg_dyn}
A \emph{dynamical system of degree $d$} on $\mathbb{P}^N$ is a morphism $\mathbb{P}^N \to \mathbb{P}^N$ of degree $d$. Since conjugating a morphism by a projective transformation produces a morphism with the same dynamics, the \emph{moduli space $\mathcal{M}_d^N$ of degree $d$ dynamical systems on $\mathbb{P}^N$} is defined as the space of morphisms $\mathbb{P}^N \to \mathbb{P}^N$ up to the action of $\SL_{N+1}$ by conjugation. These moduli spaces were constructed with GIT by Silverman \cite{MR1635900} in degree $d \geq 2$ and dimension $N = 1$, and by Levy and Petsche-Szpiro-Tepper in degree $d \geq 2$ and dimension $N > 1$ \cite{MR2741188, MR2567424}. The analogous moduli space in degree $d = 1$ is impossible to construct because the GIT stable locus is empty \cite{MR0437531}.
The moduli space $\mathcal{M}^N_d$ admits a GIT semistable compactification $\bar{\mathcal{M}}^N_d$ that includes some morphisms of lower degree and rational maps; these degenerate objects are represented in the GIT boundary of the quotient, that is, $\bar{\mathcal{M}}^N_d \smallsetminus \mathcal{M}^N_d$. One might speculate that these rational maps are somehow dynamically special, although we note that the natural iteration maps between these moduli spaces have indeterminacy \cite{MR1764925}.
Now we discuss the reasons for considering marked points. The $\SL_{N+1}$-action \eqref{eq_acn}, defining projective equivalence of marked linear maps, preserves dynamical relationships among the marked points. That is, if we write
$$(T', v'_1, \hdots, v'_n) = A \cdot (T, v_1, \hdots, v_n),$$
then $A$ takes $T$-orbits to $T'$-orbits. Further, the orbit relations of the marked points are preserved; for instance, if for some $i, j$ we have $T(v_i) = v_j$, then also $T'(v'_i) = v'_j$. Any set of relations of this kind defines a subvariety of $\mathcal{M}_{1,n}^N$.
These notions were formalized by Doyle-Silverman as the theory of \emph{portrait moduli spaces} \cite{MR4132597}. They define the \emph{moduli space $\mathcal{M}_{d}^N(\mathcal{P})$ of degree $d \geq 2$ dynamical systems on $\mathbb{P}^N$ with portrait $\mathcal{P}$}. A \emph{portrait} is a weighted digraph that prescribes dynamical relations on the marked points. Our moduli space $\mathcal{M}_{1,n}^N$ is the special case of this construction where $d = 1$ and $\mathcal{P}$ has $n$ vertices and no edges.
Doyle-Silverman obtained a partial description of GIT stability in degree $d \geq 2$ \cite[Theorem 8.1]{MR4132597}; our Theorem \ref{thm_main} is a complete description in degree $d = 1$. In principle, our techniques would give an exact characterization of GIT stability in degree $d \geq 2$, but the weight polyhedra will be much more complicated.
\subsection{Motivation from combinatorial algebraic geometry} \label{sect_dynamics_on_flags}
Definition \ref{defn:flaginkN1} associates a Hessenberg function $h_{T, \mathcal{H}}$ to each flag $\mathcal{H}$ of length $\gamma + 1$ and projective linear map $T$. This function is a nondecreasing self-map of $\{0, \hdots, \gamma + 1\}$. The Hessenberg function captures the dynamics induced on the flag $\mathcal{H}$ by $T$, in the following sense: lifting $T$ to an endomorphism of $k^{N+1}$, there is a semiconjugacy from $T$ to $h_{T, \mathcal{H}}$ as set-maps, defined by
\begin{align*}
&k^{N+1} \to \{0, \hdots, \gamma + 1\},\\
&v \mapsto \min \{j \colon v \in H_j \}.
\end{align*}
Hessenberg functions of flags arise in combinatorial algebraic geometry. Flags in $k^{N+1}$ are parametrized by \emph{flag varieties}, and each flag variety is stratified into \emph{Hessenberg varieties} via the associated Hessenberg functions relative to a fixed endomorphism $T$. The dynamics of $T$ tend to be reflected in the structure of the corresponding Hessenberg varieties. There is a substantial literature exploring the geometry of Hessenberg varieties in the complete flag variety, especially their dimension, singularities, and cohomology \cite{MR1043857, MR4391520, MR2275912}. Our work uses the notion of Hessenberg functions for incomplete flags; the corresponding generalized Hessenberg varieties are the subject of the recent paper \cite{incomplete_hessenberg}.
Theorem \ref{thm_main} shows that, whereas one normally expects each flag in $\mathbb{P}^N$ to contribute a necessary condition for GIT stability, for our moduli problem, stability can be checked only on a few Hessenberg strata. This is a significant reduction; whereas the complete flag variety on $\mathbb{P}^N$ has dimension $\binom{N + 1}{2}$, the Hessenberg variety of complete Type II flags has dimension at most $N$, and the Hessenberg variety of complete Type III flags has dimension at most $0$.
\subsection{Motivation from integrable systems} \label{sect_integrable_systems}
The moduli space $\mathcal{M}^N_{1,n}$ of $n$-marked linear maps on $\mathbb{P}^N$ is the domain of a wide class of integrable systems called \emph{generalized pentagram maps} \cite{MR3253683, izosimov, MR3118623, MR3093293, MR2679816, MR2434454, MR3161305}. In this context, the space $\mathcal{M}^N_{1,n}$ is usually studied as a real or complex manifold that parametrizes \emph{twisted polygons}.
\begin{defn} \label{defn_tw_poly}
Let $N, n \geq 1$ be integers. A \emph{twisted $n$-gon} $v$ in $\mathbb{P}^N$ is a bi-infinite sequence $(v_i)_{i \in \mathbb Z}$ of points in $\mathbb{P}^N$, called \emph{vertices}, such that there exists a unique invertible projective transformation $T$, the \emph{monodromy}, with the property that, for all $i \in \mathbb Z$,
$$Tv_i = v_{i + n}.$$
\end{defn}
A generic twisted $n$-gon is determined by its first $n$ vertices $v_1, \hdots, v_n$ and the monodromy $T$. This embeds the set of sufficiently generic twisted $n$-gons in the variety $\mathcal{X}_{N,n}$. Then the $\SL_{N+1}$-action \eqref{eq_acn} describes the effect of changing of coordinates on $\mathbb{P}^N$, so the moduli space $\mathcal{M}^N_{1,n}$ parametrizes projective equivalence classes of twisted $n$-gons in $\mathbb{P}^N$ and some degenerations.
Theorem \ref{thm_main} shows that the moduli space $\mathcal{M}^N_{1,n}$ can be equipped with the structure of an algebraic variety over any algebraically closed field, and provides a compactification $\bar{\mathcal{M}}^N_{1,n}$ in the category of varieties.
The moduli space of twisted $n$-gons is of interest because any projectively natural operation induces a rational self-map of $\mathcal{M}^N_{1,n}$ with an invariant fibration. For instance, the pentagram map \cite{MR1181089, MR2434454} is a dynamical system $\mathcal{M}^2_{1,n} \dashrightarrow \mathcal{M}^2_{1,n}$. Fixing $n \geq 4$ and cyclically ordering the vertices, we send each sufficiently generic twisted polygon $(v_i)$ to $(v'_i)$, where $$v'_i = \overline{v_{i - 2} v_{i + 1}} \cap \overline{v_{i - 1} v_{i + 2}}.$$
The map is projectively natural, so it defines a rational self-map of $\mathcal{M}^2_{1,n}$. Note that the $\SL_3$-equivalence class of the monodromy $T$ is an invariant of the pentagram map, providing two algebraically independent conserved quantities. In fact, the pentagram map is a discrete algebraic completely integrable system: it has an iterate which is birational to a translation on a family of abelian varieties \cite{MR3161305, weinreich2021}.
Other maps on $\mathcal{M}^N_{1,n}$ that are known or believed to be integrable include generalized pentagram maps \cite{izosimov, MR3118623, MR3356734}, cross-ratio dynamics \cite{crossratdynamics}, and maps associated to moves on triple-crossing diagrams \cite{affolter2021crossratio}. More generally, the following recipe produces discrete dynamical systems on $\mathcal{M}^N_{1,n}$. Choose an integer $n' \leq n$, and let $\phi \colon (\mathbb{P}^N)^{n'} \dashrightarrow \mathbb{P}^N$ be any rational map which respects projective transformations of $\mathbb{P}^N$. Then we get a (partially defined) self-map of the set of sequences $\mathbb Z \to \mathbb{P}^N$, defined by the rule $(v_i) \mapsto (v'_i)$, where
$$v'_i = \phi(v_{i}, \hdots, v_{i+n'-1}).$$
Again there is an induced map $\Phi$ on $\mathcal{M}^N_{1,n}$ that admits at least $N$ independent integrals. Silverman, in a talk at the 2021 Joint Mathematics Meetings, proposed studying the integrability of $\Phi$ when $\phi$ is chosen to be the Cayley-Bacharach map $(\mathbb{P}^2)^8 \dashrightarrow \mathbb{P}^2$.
Thus we have the following curious observation: the moduli space $\mathcal{M}^N_{1,n}$ both \emph{is} a moduli space of dynamical systems with marked points, and itself admits many interesting rational dynamical systems.
\subsection{Outline}
Section \ref{sec_prelim} contains standard preliminaries from geometric invariant theory. New results begin in Section \ref{sect_polyhedra} with the definition of the corner polyhedron, the classification of the corner polyhedra of type $A_N$, and a menagerie of examples. In Section \ref{sect_proof}, we show that, instead of checking stability on every flag, one can just check Type I, II, and III flags, and we prove Theorem \ref{thm_main}, Corollary \ref{cor_stable_nonempty}, and Theorem \ref{thm_rational}. Section \ref{sect_one_marked_point} illustrates the main results in the special case $n = 1$ of one marked point.
\subsection{Acknowledgments}
The author thanks his advisor, Joseph Silverman, for many thought-provoking discussions and for a careful reading of this manuscript. Further thanks to Niklas Affolter, Madeline Brandt, Brendan Hassett, Han-Bom Moon, Richard Schwartz, and Mariel Supina for helpful conversations. The author was supported by an NSF Graduate Research Fellowship. An earlier version of this work appeared
as Chapter 4 of the author's Ph.D. thesis \cite{weinreich_thesis}.
\section{Preliminaries} \label{sec_prelim}
In this section, we recall some standard concepts from geometric invariant theory: categorical quotients, geometric quotients, good quotients, $G$-linearized sheaves, GIT stability, Hilbert-Mumford numerical invariants, destabilizing 1-parameter subgroups, weight sets, and the weight polytope. For a detailed development, see \cite[Chapter 6--9]{MR2004511}.
\begin{defn}
Let an algebraic group $G$ act algebraically on a variety $V$, and let $\alpha: G \times V \to V$ denote the action. A \emph{categorical quotient} is a pair $(V', \chi)$ consisting of a variety $V'$ and a $G$-invariant morphism $V \to V'$, such that for each variety $V''$, every $G$-invariant morphism $V \to V''$ factors through $\chi$ uniquely. We typically suppress the map $\chi$ from the notation.
A \emph{geometric quotient} is a categorical quotient $(V', \chi)$ with the additional property that if $v, v' \in V$ have the same image in $V'$, then they share a $G$-orbit.
A categorical or geometric quotient $(V', \chi)$ is called \emph{good} if it satisfies the additional properties:
\begin{enumerate}
\item For any open subset of $U$ of $V'$, the homomorphism $\chi^*: \mathcal{O}(U) \to \mathcal{O}(\chi^{-1}(U))$ is an isomorphism onto the subring $\mathcal{O}(\chi^{-1}(U))^G$ of $G$-invariant sections.
\item If $W$ is a closed $G$-invariant subset of $V$, then $\chi(W)$ is a closed subset of $V'$.
\item If $W_1, W_2$ are closed $G$-invariant subsets of $V$ with empty intersection, then their images in $V'$ have empty intersection.
\end{enumerate}
Let $V$ be a projective variety acted on by a geometrically reductive group $G$. Let $\mathcal{L}$ be an invertible sheaf on $V$, and equip $\mathcal{L}$ with a $G$-linearization. If $\mathcal{L}$ is very ample, this data is more or less equivalent to a choice of embedding of $V$ into a projective space so that $G$ acts via a linear group. An extension of the $G$-action on $V$ to a linear action on that projective space is called a \emph{$G$-linearization}. A \emph{$G$-linearized sheaf} is a very ample invertible sheaf together with a choice of $G$-linearization.
\end{defn}
\begin{prop} \label{prop_sl_linearizations}
For any very ample invertible sheaf $\mathcal{L}$ on an irreducible normal projective variety $V$ acted on by $\SL_{N+1}$, there exists a unique $\SL_{N+1}$-linearization of $\mathcal{L}$.
\end{prop}
Proposition \ref{prop_sl_linearizations} follows from the tools of \cite[Chapter 7]{MR2004511} and is an exercise in that chapter.
\begin{remark}
Our work fits into this setup, since $V = \mathcal{X}_{N,n}$ is projective and $G = \SL_{N+1}$ is geometrically reductive over any algebraically closed field. The very ample invertible sheaves up to isomorphism are classified by tuples $(q, m_1, \hdots, m_n) \in \mathbb N^{n+1}$; they are of the form $\mathcal{O}(q, m_1, \hdots, m_n)$. Thus these tuples also classify $\SL_{N+1}$-linearized sheaves on $\mathcal{X}_{N,n}$.
\end{remark}
\begin{defn} \label{defn_stab}
Let $\mathcal{L}$ be a $G$-linearized sheaf on $V$. A point $v \in V$ is \emph{semistable} with respect to $\mathcal{L}$ if, for some $m \in \mathbb N$, the tensor power $\mathcal{L}^{\otimes m}$
admits a global section $s \in \Gamma(V, \mathcal{L}^{\otimes m})^G$ such that the set
$$ V_s \colonequals \{ y \in V \colon s(y) \neq 0 \}$$
is affine and contains $v$.
A point $v \in V$ is \emph{stable} if it is semistable and the set $V_s$ may be chosen such that the stabilizer of $v$ in $G$ is finite and all orbits of $G$ in $V_s$ are closed. A point $v \in V$ is \emph{unstable} if it is not semistable.
\end{defn}
Denote the sets of stable, semistable, and unstable points with respect to $\mathcal{L}$ (with its $G$-linearization) by
$$V\suptxt{s}(\mathcal{L}),\; V\suptxt{ss}(\mathcal{L}), \;V\suptxt{us}(\mathcal{L}).$$
In general, any of these sets could be empty.
Note that there is some variation in the literature regarding these definitions. In our notation, ``not stable'' and ``unstable'' mean different things. One could also consider the closely related concept of polystability, but we do not pursue this.
The following theorem shows that the semistable quotient is a compactification of the stable quotient.
\begin{thm} \label{thm_stabquo}
Let $V$ be a variety with an algebraic action of a geometrically reductive group $G$. Let $\mathcal{L}$ be a $G$-linearized sheaf.
\begin{enumerate}
\item
There exists a categorical quotient $V\suptxt{ss}(\mathcal{L}) / \! / G$, and it is quasi-projective. If $V$ is projective, then $V\suptxt{ss}(\mathcal{L}) / \! / G$ is projective.
\item
There exists a geometric quotient $ V\suptxt{s}(\mathcal{L}) / G$.
\item
The stable locus $V\suptxt{s}(\mathcal{L})$ is a Zariski open subset of the semistable locus $V\suptxt{ss}(\mathcal{L})$, which is in turn an open subset of $V$. The geometric quotient $V\suptxt{s}(\mathcal{L}) / G$ is an open subset of the categorical quotient $V\suptxt{ss}(\mathcal{L}) / \! / G.$
\end{enumerate}
\end{thm}
For a proof, see \cite[Theorem 8.1 and Proposition 8.1]{MR2004511}.
Suppose that the $G$-linearized sheaf $\mathcal{L}$ embeds $V$ in $\mathbb{P}^D$.
Then $G$ acts via a linear representation in $\SL_{D+1}$. Let $\G\subtxt{m}$ denote the multiplicative group scheme over $k$. Let $\ell \colon \G\subtxt{m} \to \SL_{D+1}$ be a 1-parameter subgroup. Every 1-parameter subgroup of $\SL_{D+1}$ is diagonal in some basis of $\mathcal{L}$. Such a basis is called a \emph{diagonalizing basis for} $\ell$. Choose a diagonalizing basis for $\ell$. For some $\lambda_1, \hdots, \lambda_{D+1} \in \mathbb Z$, the matrix of $\ell(\tau)$ is
\[
\ell(\tau) = \begin{bmatrix}
\tau^{\lambda_1} \\
& \tau^{\lambda_2} \\
& & \ddots \\
& & & \tau^{\lambda_{D+1}}
\end{bmatrix}.
\]
Suppose that $v \in V$, and in the diagonalizing basis for $\ell$, the homogeneous coordinates of $v$ are $v = [X_1 : \hdots : X_{D+1}]$. The \emph{numerical invariant} of $\ell$ at $v$ (with respect to $\mathcal{L}$ with the chosen linearization) is
$$\mu(v, \ell) \colonequals \min_{1 \leq i \leq D+1} \{\lambda_i \colon X_i \neq 0 \}.$$
This quantity is independent of choice of diagonalizing basis for $\ell$. Note that the sign on $\mu$ varies in the literature.
Now we can state the \emph{Hilbert-Mumford numerical criterion}.
\begin{thm}[Mumford] \label{thm_numcrit}
With notation as above,
$$V\suptxt{ss} = \{v \colon \mu(v, \ell) \leq 0 \text{ for all } \ell\},$$
$$V\suptxt{s} = \{v \colon \mu(v, \ell) < 0 \text{ for all } \ell \}.$$
\end{thm}
For a proof, see \cite[Theorem 9.1]{MR2004511}.
A \emph{de-semistabilizing 1-parameter subgroup} for $v \in V$ is a 1-parameter subgroup $\ell$ for which $\mu(v, \ell) \not\leq 0$.
A \emph{destabilizing 1-parameter subgroup} for $v \in V$ is a one-parameter subgroup $\ell$ for which $\mu(v, \ell) \not< 0$. By Theorem \ref{thm_numcrit}, semistability of $v \in V$ is equivalent to having no de-semistabilizing 1-parameter subgroups. Similarly, stability is equivalent to having no destabilizing 1-parameter subgroups.
We set more notation:
\begin{itemize}
\item $\Theta$, a maximal torus in $\SL_{N+1}$;
\item $W_\mathbb Z$, the character lattice of $\Theta$;
\item $W$, the $\mathbb R$-vector space $W_\mathbb Z \otimes \mathbb R$;
\item $R_\mathbb Z$, the lattice of 1-parameter subgroups of $\Theta$;
\item $R$, the $\mathbb R$-vector space $R_\mathbb Z \otimes \mathbb R$;
\item $\langle \cdot, \cdot \rangle: W \times R \to \mathbb R$ is the pairing defined as follows. For any $\chi \in W_\mathbb Z$ and $\ell \in R_\mathbb Z$, let $\langle \chi, \ell \rangle$ be the integer $\lambda$ such that the composed map $\chi \circ \ell : \G\subtxt{m} \to \G\subtxt{m}$ is $\tau \mapsto \tau^\lambda$. Extend linearly.
\end{itemize}
A choice of isomorphism $\Theta \cong \G\subtxt{m}^N$ induces an identification $W_\mathbb Z \cong \mathbb Z^{N+1}$. We also get an embedding $r \colon R_\mathbb Z \hookrightarrow \mathbb Z^{N+1}$ onto the hyperplane in $\mathbb Z^{N+1}$ where the coordinates sum to 0. Note that $R_\mathbb Z$ is isomorphic to the $A_N$ lattice rather than a free abelian group. With these identifications, the inner product $\langle \cdot, \cdot \rangle$ is computed by dot product in $\mathbb R^{N+1}$.
The $\SL_{N+1}$-linearization of $\mathcal{L}$ induces a representation $\Theta \hookrightarrow \GL_{D+1}$. Any such representation splits completely, because $\Theta$ is linearly reductive. Therefore we can choose a diagonalizing basis $X_1, \hdots, X_{D+1}$ for the action of $\Theta$ on $\mathcal{L}$. Each basis element $X_i$ has an associated character $\chi_i \in W_\mathbb Z$.
The \emph{weight} of $X_i$, denoted $\wt(X_i)$, is the image of the character $\chi_i$ in $\mathbb Z^{N+1}$. The \emph{weight set} $\wt(v) \subset W_\mathbb Z$ of $v \in V$, relative to the maximal torus $\Theta$ and the choice of identification $\Theta \cong \G\subtxt{m}^N$, is the set
$$\wt(v) = \{ \wt(X_i) \colon X_i(v) \neq 0, \text{ for each } i = 1, \hdots, D + 1 \}.$$
The \emph{weight polytope} of $v$, denoted $\Pi(v)$, is the convex hull of $\wt(X_i)$ in $\mathbb R^{N+1}$. It is a convex polytope. We denote its interior by $\Pi(v)^\circ$.
We further define
$$V\suptxt{ss}_\Theta = \{v \colon \mu(v, \ell) \leq 0 \text{ for all } \ell \subseteq \Theta \},$$
$$V\suptxt{s}_\Theta = \{v \colon \mu(v, \ell) < 0 \text{ for all } \ell \subseteq \Theta \}.$$
The weight polytope of a point $v$ depends on the identification $\Theta \cong \G\subtxt{m}^N$, but the weight set viewed in $W$ is basis-independent, so the statements $0 \in P(v)$ and $0 \in P(v)^\circ$ are basis-independent. In fact, we have the following consequence of Proposition \ref{thm_numcrit}:
\begin{prop}
Given a maximal torus $\Theta$ of $\SL_{N+1}$ and an identification $\Theta \cong \G\subtxt{m}^N$, and a point $v \in V$, the following are equivalent:
\begin{enumerate}
\item $v \in V\suptxt{ss}_\Theta$,
\item $0 \in \Pi(v)$ for some diagonalizing basis of $\Theta$.
\end{enumerate}
Similarly, the following are equivalent:
\begin{enumerate}
\setcounter{enumi}{2}
\item $v \in V\suptxt{s}_\Theta$,
\item $0 \in \Pi(v)^\circ$ for some diagonalizing basis of $\Theta$.
\end{enumerate}
\end{prop}
For a proof, see \cite[Theorem 9.2]{MR2004511}.
Suppose that $\SL_{N+1}$ acts on a projective variety $V$ equipped with sheaf $\mathcal{L}$. Let $B$ be a basis of $k^{N+1}$. The basis $B$ has an associated maximal torus $\Theta$ in $\SL_{N+1}$ containing all the 1-parameter subgroups of $\SL_{N+1}$ that are diagonal in $B$. A basis $B'$ for $\mathcal{L}$ is called \emph{$B$-aligned} if the representation of $\Theta$ in $\GL_{D+1}$ is diagonal in $B'$.
The next two lemmas are standard; they describe how weight polytopes relate to the Segre and Veronese maps.
\begin{defn}
Given two subsets $\Pi_1$, $\Pi_2$ of a $k$-vector space, the \emph{Minkowski sum} of $\Pi_1$ and $\Pi_2$ is the set
$$\Pi_1 + \Pi_2 \colonequals \{ \pi_1 + \pi_2 \colon \pi_1 \in \Pi_1, \, \pi_2 \in \Pi_2 \}.$$
Given any scalar $m \in k$ and subset $\Pi$ of a $k$-vector space, define
$$m\Pi \colonequals \{ m \pi \colon \pi \in \Pi \}.$$
\end{defn}
\begin{lemma} \label{lem_segre}
Suppose that $\SL_{N+1}$ acts on two projective varieties $V_1, V_2$ via the linearized sheaves $\mathcal{L}_1, \mathcal{L}_2$.
We consider the product action of $\SL_{N+1}$ on $V_1 \times V_2$.
Suppose that $B$ is a basis of $k^{N+1}$ and $B_1, B_2$ are $B$-aligned bases of $\mathcal{L}_1, \mathcal{L}_2$.
There is a $B$-aligned basis $B_1 \times B_2$ formed of all products of the form $b_1 b_2$, where $b_1 \in B_1$ and $b_2 \in B_2$.
The weight polytope of $(p_1, p_2)$ with respect to $B_1 \times B_2$ is the Minkowski sum
$$\Pi(p_1, p_2) = \Pi(p_1) + \Pi(p_2).$$
\end{lemma}
\begin{lemma} \label{lem_veronese}
Suppose that $\SL_{N+1}$ acts on a projective variety $\mathcal{X}$ via the sheaf $\mathcal{L}$. For any $m \in \mathbb N$, we consider the induced action of $\SL_{N+1}$ on $\mathcal{X}$ with the tensor power sheaf $\mathcal{L}^{\otimes m}$. Let $B$ be a basis of $k^{N+1}$, and Suppose that $B'$ is a $B$-aligned basis of $\mathcal{L}$. There is a $B$-aligned basis $B''$ on $\mathcal{L}^{\otimes m}$ consisting of all $m$-fold products of elements of $B'$.
Let $\Pi(p)$ denote the weight polytope of $p$ for $\mathcal{L}$, and let $\Pi^{(m)}(p)$ denote the weight polytope of $p$ for $\mathcal{L}^{\otimes m}$. Then
$$\Pi^{(m)}(p) = m\Pi(p).$$
\end{lemma}
The proofs are direct applications of the definitions of the Segre and Veronese embeddings; we omit them.
\section{Corner polyhedra and stability} \label{sect_polyhedra}
In this section, we introduce our key technical tool, the corner polyhedron. We use corner polyhedra instead of weight polytopes to simplify the analysis of GIT stability. Our first step is to derive the stability-detecting property of corner polyhedra from the corresponding property of weight polytopes. The corner polyhedron stability test is given in Proposition \ref{prop_corner_numerical}.
We recall some terminology. The setting is any $\mathbb R$-vector space. Unless stated otherwise, half-spaces are assumed to be closed.
\begin{itemize}
\item A \emph{convex polyhedron} is an intersection of finitely many half-spaces. All polyhedra we consider are convex, so we may just write \emph{polyhedron}.
\item A \emph{convex polyhedral cone} is an intersection of finitely many half-spaces, all of which contain the origin at the boundary.
\item A \emph{polytope} is a bounded polyhedron.
\item A \emph{supporting halfspace} of a polyhedron is a halfspace containing the polyhedron and containing some point on its boundary.
\item A \emph{face} of a polyhedron is any set which is the intersection of that polyhedron with a supporting halfspace.
\item For any $D \in \mathbb N$, a \emph{facet} of a $D$-dimensional polyhedron is a face of dimension $D - 1$.
\item Each facet has a unique \emph{associated supporting halfspace}.
\item A \emph{vertex} is a 0-dimensional face.
\item A polyhedron is \emph{full-dimensional} if it is a polyhedron with topological dimension equal to the ambient vector space.
\item A polyhedron is \emph{pointed} if it has at least one vertex.
\item A polytope is \emph{integral} with respect to a lattice if its vertices are lattice points.
\item A polyhedron is \emph{rational} with respect to a chosen basis if the associated supporting halfspace of each facet is defined over $\mathbb Q$ when written in that basis.
\end{itemize}
We use the following basic principles of convex polyhedra throughout. For any set $\zeta$ in an $\mathbb R$-vector space, let $\Conv(\zeta)$ denote its convex hull. We use $+$ to denote Minkowski sum.
\begin{itemize}
\item The convex hull of a finite set is a polytope.
\item A convex polyhedron is equal to the intersection of the associated supporting halfspaces of its facets.
\item Given convex polyhedra $\Pi_1, \Pi_2$, the Minkowski sum $\Pi_1 + \Pi_2$ is a polyhedron. Each face of $\Pi_1 + \Pi_2$ is the Minkowski sum of some face of $\Pi_1$ and some face of $\Pi_2$.
\item Every convex polyhedron can be expressed as the Minkowski sum of a convex polyhedral cone and a convex polytope.
\item Any functional on a polytope achieves its minimum at some vertex.
\item An integral polytope is also rational as a polyhedron.
\item The Minkowski sum of rational polyhedra is rational.
\end{itemize}
For any $N \in \mathbb N$, let $\mathbb R^N_+$ denote the \emph{nonnegative orthant}, i.e., the set of points in $\mathbb R^N$ with all nonnegative coordinates.
The \emph{corner hull} of a set $\zeta \subseteq \mathbb R^N$, denoted $\Corner(\zeta)$, is defined by
$$\Corner(\zeta) = \Conv(\zeta) + \mathbb R^N_+,$$
where $\mathbb R^N_+$ denotes the nonnegative orthant.
We repeatedly use the following properties of the corner hull:
\begin{enumerate}
\item The corner hull of a finite set is a finite, full-dimensional, pointed, unbounded, convex polyhedron. The vertices of the corner hull of a set $\zeta$ are themselves elements of $\zeta$.
\item For any sets $\zeta, \zeta' \subset \mathbb R^N$,
$$\Corner(\zeta + \zeta') = \Corner(\zeta) + \Corner(\zeta').$$
\item For any real number $m > 0$, and any set $\zeta \subset \mathbb R^N$,
$$\Corner(m \zeta) = m \Corner (\zeta).$$
\end{enumerate}
Our next lemma is a version of the hyperplane separation theorem, a central tool in convex geometry, for the corner hull. It shows that membership in the corner hull can be tested by functionals with positive coordinates in the dual space.
\begin{lemma}[Positive hyperplane separation] \label{lem_hyperplane_sep}
Let $(\mathbb R^N)^\vee_+$ denote the nonnegative orthant in the dual space to $\mathbb R^N$, with respect to the dual basis. Let $(\mathbb Z^N)^\vee_+$ denote the set of standard lattice points within it.
\begin{enumerate}
\item
For any finite set $\zeta \subset \mathbb R^N$,
we have $0 \in \Corner( \zeta )$
if and only if there is no $r \in (\mathbb R^N)^\vee_+ \smallsetminus \{0\}$ such that, for all $w \in \zeta$,
$$r(w) > 0.$$
\item
For any finite set of lattice points $\zeta \subset \mathbb Z^N$,
we have $0 \in \Corner( \zeta )$
if and only if there is no lattice point $r \in (\mathbb Z^N)^\vee_+ \smallsetminus \{0\}$ such that, for all $w \in \zeta$,
$$r(w) > 0.$$
\item
For any finite set $\zeta \subset \mathbb R^N$, we have
$0 \in \Corner( \zeta )^\circ$
if and only if there is no $r \in (\mathbb R^N)^\vee_+ \smallsetminus \{0\}$ such that, for all $w \in \zeta$,
$$r(w) \geq 0.$$
\item
For any finite set of lattice points $\zeta \subset \mathbb Z^N$, we have
$0 \in \Corner( \zeta )^\circ$
if and only if there is no lattice point $r \in (\mathbb Z^N)^\vee_+ \smallsetminus \{0\}$ such that, for all $w \in \zeta$,
$$r(w) \geq 0.$$
\end{enumerate}
\end{lemma}
\begin{proof}
\par(1)\enspace Suppose that $0 \in \Corner(\zeta)$. Then there exists some point $w_0 \in \Conv(\zeta)$ such that, for all $r \in (\mathbb R^N)^\vee_+,$
$$0 \geq r(w_0).$$
Functionals on $\Conv(\zeta)$ achieve their minima on an element of $\zeta$. So, for each $r \in (\mathbb R^N)^\vee_+$, there exists some $w \in \zeta$, depending on $r$, such that $r(w_0) \geq r(w)$. Thus
$$0 \geq r(w).$$
For the other direction, suppose that $0 \not\in \Corner(\zeta)$. There exists some shortest-length vector $w_0$ in $\Corner(\zeta)$. By the hyperplane separation theorem, the functional $r$ dual to $w_0$ takes only positive values on $\Corner(\zeta)$. In particular we know that $r$ takes only positive values on $\zeta$. The coordinates of $w_0$ are non-negative, since otherwise the negative coordinates could be replaced by 0 to obtain a shorter-length vector in $\Corner(\zeta)$. The coordinates of $w_0$ are rational, since the points of $\zeta$ are lattice points. Thus some positive integer multiple of $r$ is in $(\mathbb R^N)^\vee_+ \smallsetminus \{0\}.$
\par(2)\enspace
One direction follows immediately from (1). For the other direction, note that in the proof of (1), since $\Corner(\zeta)$ is the sum of an integral polytope and a rational polyhedral cone, all its facets are rational. Thus the shortest vector $w_0$ has rational coordinates, so a sufficiently high multiple of its dual is the desired lattice point $r$.
\par(3)\enspace We have
$$\Corner(\zeta)^\circ = (O^+)^\circ + \Conv(\zeta).$$
Thus if $0 \in \Corner(\zeta)^\circ$, then there exists some point $w_0 \in \Conv(\zeta)$ such that, for all $r \in (\mathbb R^N)_+^\vee \smallsetminus \{0\}$,
$$0 > r(w_0).$$
Following the same argument as in (1), for each $r$ there exists some $w \in \zeta$ such that $r(w_0) \geq r(w)$, so $0 > r(w)$.
For the other direction, we already proved it in (1), except for the case that $0$ is on the boundary of $\Corner(\zeta)$. Then $0$ belongs to a facet of $\Corner(\zeta)$. Take $r$ to be the functional defining that facet's associated supporting halfspace.
\par(4)\enspace Similar to (2) and (3); omitted.
\end{proof}
It turns out that, because we are working with $\SL_{N+1}$-actions, computing in the space $\mathbb R^{N+1}$ is redundant. All our computations happen in the hyperplane of vectors with coordinates that sum to 0. The next definitions set up a preferred basis for this hyperplane and its dual space.
Let $N \in \mathbb N$. We consider the lattice $\mathbb Z^{N+1}$ embedded in $W = \mathbb R^{N+1}$. We denote the standard basis vectors by $e_1, \hdots, e_{N+1}$. We denote the dual space of $\mathbb R^{N+1}$ by $R = (\mathbb R^{N+1})^\vee$. The space $R$ is spanned by the dual coordinates to the standard basis vectors, denoted $x_1, \hdots, x_{N+1}$.
Let $\bm{1} = (1, \hdots, 1) \in W$. Let $\bar{W}$ be the hyperplane orthogonal to $\bm{1}$. The elements of $W$ that belong to $\bar{W}$ are vectors with coordinates that sum to 0.
Let $\bar{R} = (\bar{W})^\vee$. We consider the basis $s_1, \hdots, s_N$ of $\bar{R}$ defined by
$$s_i = \sum_{j=1}^i x_j.$$
Let $\bar{R}^+$ denote the nonnegative $\mathbb R$-span of $s_1, \hdots, s_N.$
We work with this particular basis $s_1, \hdots, s_N$ because it has the following convenient property. There is a restriction map $R \to \bar{R}$ defined by restricting functionals from $W$ to $\bar{W}$. For any $r \in R$, we denote its restriction by $\bar{r}$. If $r = (r_1, \hdots, r_{N+1})$, we have $r_1 \geq \hdots \geq r_{D+1}$ if and only if $\bar{r} \in \bar{R}^+$. Further, if $r(\bm{1}) = 0$, then we have $\bar{r} = 0$ if and only if $r_1 = \hdots = r_{D+1} = 0.$
Let $O^+$ denote the nonnegative orthant in $\bar{W}$ relative to the dual basis $s_1, \hdots, s_N$. Its interior $(O^+)^\circ$ is the strictly positive orthant.
The \emph{corner hull} of a set $\zeta \subseteq \bar{W}$, denoted $\Corner(\zeta)$, is defined as the corner hull of the image of $\zeta$ in $\mathbb R^N$ via the identification $\bar{W} \cong \mathbb R^N$ from the preferred basis $s_1, \hdots, s_N$. That is,
$$\Corner(\zeta) = \Conv(\zeta) + O^+.$$
\begin{defn} \label{def_corner_poly}
Consider a projective variety $V$ equipped with a $\SL_{N+1}$-linearized sheaf. Let $B$ be a basis of $k^{N+1}$. Let $W \cong \mathbb R^{N+1}$ be the vector space over the character lattice corresponding to basis $B$, and let $\bar{W}$ be the hyperplane of points of $Z$ with coordinates that sum to 0. Let $\Pi(v) \subset W$ denote the weight polytope of $v$ with respect to this basis. Let $\bar{\Pi}(v)$ denote the projection of $\Pi(v)$ to $\bar{W}$. The \emph{corner polyhedron} of $v \in V$, denoted $\Pi_+(v)$, is the corner hull
$$\Corner(\bar{\Pi}(v)) \subseteq \mathbb R^N.$$
We call the vertices of $\Pi_+(v)$ the \emph{control points}. The set of control points is denoted $\Ctrl(v)$.
\end{defn}
We emphasize that the weight polytope $\Pi(v)$ is a subset of the vector space $W$, which has dimension $N + 1$, whereas the corner polyhedron $\Pi_+(v)$ is a subset of the subspace $\bar{W}$, which has dimension $N$.
Let $B$ be a basis, with vectors $b_1, \hdots, b_{N+1}$. We emphasize that $B$ is ordered. A 1-parameter subgroup of $\SL_{N+1}$ is said to be \emph{along $B$} if, written in basis $B$, it is of the form $\diag [\tau^{r_1}, \tau^{r_2}, \hdots, \tau^{r_{N+1}}]$, where $r_1 \geq \hdots \geq r_{N+1}$. Every $1$-parameter subgroup of $\SL_{N+1}$ is along some basis $B$ because we can permute basis vectors to get a new basis.
Let $V$ be a variety acted on by $\SL_{N+1}$, equipped with a linearized sheaf $\mathcal{L}$. Given an ordered $B$, an element $v \in V$ is called \emph{$B$-stable} if there is no destabilizing 1-parameter subgroup of $\SL_{N+1}$ along $B$. Similarly, a point $v$ is \emph{$B$-semistable} if there is no de-semistabilizing 1-parameter subgroup along $B$.
Now we interpret the Hilbert-Mumford criterion in these terms. The idea to interpret the Hilbert-Mumford criterion with the weight polytope is standard; see \cite[Chapter 9]{MR2004511}. The original contribution in Proposition \ref{prop_corner_numerical} is that the corner hull takes advantage of the symmetry of $\SL_{N+1}$.
\begin{prop} \label{prop_corner_numerical}
Suppose $\SL_{N+1}$ acts on a projective variety $V$ via the sheaf $\mathcal{L}$. Let $B$ be a basis of $k^{N+1}$. Then $v \in V$ is $B$-semistable if and only if
$$ 0 \in \Pi_+(v).$$
The point $v \in V$ is $B$-stable if and only if
$$ 0 \in \Pi_+(v)^\circ.$$
\end{prop}
\begin{proof}
One checks $B$-semistability by computing the numerical invariant $\mu(v, \ell)$ for every 1-parameter subgroup $\ell$ along $B$. We have
$$\mu(v, \ell) = \min_{w \in \wt(v)} r(\ell)(w).$$
For any $w \in W$, let $\bar{w}$ denote the projection of $w$ to $\bar{W}$. For any $\ell$, the vector $r(\ell)$ is orthogonal to $(1, \hdots, 1)$. So, for any $\ell$ and $w$, we have
$$r(\ell)(w) = r(\ell)(\bar{w}) = \bar{r}(\ell)(\bar{w}).$$
As $\ell$ ranges over $1$-parameter subgroups along $B$, the vector $\bar{r}(\ell)$ ranges over nonzero lattice points in $\bar{R}^+$.
Thus $v$ is $B$-semistable if and only if, for all nonzero lattice points $\bar{r} \in \bar{R}^+$, there exists $w \in \wt(v)$ such that
$$ \bar{r} ( \bar{w} ) \leq 0.$$
By the choice of basis $s_1, \hdots, s_N$ on $\bar{R}$ and Lemma \ref{lem_hyperplane_sep} (2), this is equivalent to
$$0 \in \Pi_+(v).$$
The $B$-stability claim follows from the same argument, but applying Lemma \ref{lem_hyperplane_sep} (4) instead.
\end{proof}
\subsection{First application: configurations of points} \label{sec_first_application}
In this section, we show how to analyze the diagonal action of $\SL_{N+1}$ on $(\mathbb{P}^N)^n$ with the corner polyhedron; here, no linear map $T$ is involved. These results are essential for our proof of Theorem \ref{thm_main}, and we recover a classical result of Mumford as Theorem \ref{thm_classic_git} along the way.
For any $m \in \mathbb N$, let $[m] \subset \mathbb N$ denote the set $\{1, \hdots, m\}$.
We first compute the corner polyhedron associated to a given basis for a single point in $\mathbb{P}^N$; it is just a translate of the nonnegative orthant.
\begin{prop} \label{prop_one_point}
Consider the action of $\SL_{N+1}$ on $\mathbb{P}^N$, with $\mathcal{L} = \mathcal{O}(m)$ for some $m \geq 1$. Choose a basis $B$ of $k^{N+1}$, let $X_1, \hdots, X_{N+1}$ be induced coordinates on $\mathbb{P}^N$, and let $B'$ be the induced $B$-aligned basis from the Segre map. Given a point $v \in \mathbb{P}^N$, Let $i$ be the largest index such that $X_i(v) \neq 0$. Then
$$\Ctrl(v, \mathcal{L}) = \left\{ m \cdot \left( -\frac{1}{N+1}, \hdots, -\frac{i - 1}{N+1}, 1 - \frac{i}{N+1}, \hdots, 1 - \frac{N}{N+1} \right) \right\}.$$
\end{prop}
\begin{proof}
Assume $m = 1$, since then the general case follows from Lemma \ref{lem_veronese}.
Let $\ell$ be a 1-parameter subgroup along a basis $B$, with $r(\ell) = (r_1, \hdots, r_{N+1})$. Let $X_1, \hdots, X_{N+1}$ be generators of $\mathcal{O}(1)$ obtained from $B$. Then, for all $i \in [N+1]$, the weight of $X_i$ is the standard basis vector
$$\wt(X_i) = e_i.$$
Therefore, we have
\[
s_j(\wt(X_i)) =
\begin{cases}
0 & \text{ if } j < i, \\
1 & \text{ otherwise.} \\
\end{cases}
\]
Since the sum of the entries of $\wt(X_i)$ is 1, we calculate
$$\overline{\wt}(X_i) = \wt(X_i) - \frac{1}{N+1} \bm{1}.$$
We have $s_j(\bm{1}) = j$. This gives us
\[
s_j(\overline{\wt}(X_i)) =
\begin{cases}
-\frac{j}{N+1} & \text{ if } j < i, \\
1 - \frac{j}{N+1} & \text{ otherwise.} \\
\end{cases}
\]
Thus, for all $(i, i', j)$ such that $i < i'$, we have
$$s_j(\overline{\wt}(X_{i})) \leq s_j(\overline{\wt}(X_i')).$$
It follows that every element of $\overline{\wt}(v)$ belongs to the nonnegative orthant at $\overline{\wt}(X_i)$, where $i$ is the largest index for which $X_i(v) \neq 0$.
\end{proof}
Let $\bm{m} = (m_1, \hdots, m_n)$ be an $n$-tuple of positive integer weights. Each such vector determines a very ample sheaf $\mathcal{O}(\bm{m})$ on $(\mathbb{P}^N)^n$. For present and later use, we define a vector
\begin{equation} \label{eq_config_vector}
\eta(v) := \left(
\sum_{i=1}^n \frac{-j m_i}{N+1} + \sum_{i \in [n]: \; X_{j+1}(v_i) = \hdots = X_{N+1}(v_i) = 0} m_i
\right)_{j \in [N]}.
\end{equation}
We now extend the previous proposition to ordered configurations of points in $\mathbb{P}^N$.
\begin{prop} \label{prop_tuple}
Consider the diagonal action of $\SL_{N+1}$ on $(\mathbb{P}^N)^n$, with $\mathcal{L} = \mathcal{O}(\bm{m})$. Choose a basis $B$ of $k^{N+1}$, and let $B'$ be the induced $B$-aligned basis from the Segre map. Then, for each $v = (v_1, \hdots, v_n) \in (\mathbb{P}^N)^n.$
\[
\Ctrl(v, \mathcal{L}) = \{ \eta(v) \}.
\]
\end{prop}
\begin{proof}
The case $n = 1$ is the result of Proposition \ref{prop_one_point}. By Lemma \ref{lem_segre}, the set $\Ctrl(v, \mathcal{L})$ consists of just one vector, and we can compute it as the sum, for $1 \leq i \leq n$, of the control vector of $v$ relative to $\mathcal{O}(m_i)$. The result follows.
\end{proof}
As a corollary, we obtain Mumford's classical description of GIT stability of projective configurations, which serves as the template for Theorem \ref{thm_main}. We note that the proof we give here is essentially the proof that can be found in \cite[Chapter 11]{MR2004511}, translated into the language of convex geometry.
\begin{thm}[Mumford, \cite{MR1304906}] \label{thm_classic_git}
Consider the diagonal action of $\SL_{N+1}$ on $(\mathbb{P}^N)^n$, with $\mathcal{L} = \mathcal{O}(\bm{m})$. Then $v = (v_1, \hdots, v_n)$ is semistable if and only if, for every nonzero proper subspace $H \subset k^{N+1}$, we have
\begin{equation} \label{eq_mum_classic}
\sum_{i \in [n] \colon v_i \in H} m_i \leq \frac{\dim H}{N + 1} \left( \sum_{i=1}^n m_i \right).
\end{equation}
The configuration $v$ is stable if and only if the inequality \eqref{eq_mum_classic} is strict.
\end{thm}
\begin{proof}
By Proposition \ref{prop_corner_numerical}, the point $v$ is semistable if and only if, in each basis $B$ of $k^{N+1}$, there is a $B$-aligned basis $B'$ of $\mathcal{L}$ such that $0 \in \Pi_+(v)$. For any given basis $B$, we let $B'$ be the natural one coming from the Segre and Veronese embeddings (Proposition \ref{prop_tuple}). By Proposition \ref{prop_tuple}, a point $v$ is $B$-semistable if and only if each entry of $\eta(v)$ is nonpositive. So, for each $j \in [N]$, the $j$-th coordinate of of $\eta(v)$ imposes the condition
\[
\sum_{i \in [n]: \; X_{j+1}(v_i) = \hdots = X_{N+1}(v_i) = 0} m_i \leq \sum_{i=1}^n \frac{j m_i}{N+1}.
\]
Let $H_j = \Span(X_1, \hdots, X_j)$, viewed as a subspace of $k^{N+1}$. Then $j = \dim H_j$, and the condition $X_{j+1}(v_i) = \hdots = X_{N+1}(v_i) = 0$ is equivalent to $v \in H_j$. Therefore, $B$-semistability is equivalent to satisfying \eqref{eq_mum_classic} for the sequence of subspaces $H_1, \hdots, H_N$ coming from basis $B$. Every nontrivial proper subspace appears this way for some basis $B$, so the result follows.
For stability, the same argument works using strict inequalities.
\end{proof}
\subsection{Corner polyhedra of type \texorpdfstring{$A_N$}{A N}} Having understood stability of configurations of marked points, we turn to the complementary problem: describing corner polyhedra of unmarked linear maps.
The main result of this subsection, Proposition \ref{prop_vertices}, calculates the vertices of the corner polyhedron associated to any matrix $M$. We use this description to classify corner polyhedra (Corollary \ref{cor_bijection}) in terms of a combinatorial gadget called the \emph{profile} of a matrix (Definition \ref{def_profile}). We encourage the reader to consult Section \ref{sect_examples} for examples to see the general theory in action.
Consider the action of $\SL_{N+1}$ by conjugation on $\PP(\Mat_{N+1})$. Let $B$ be some basis of $k^{N+1}$. We get an induced basis of the sheaf $\mathcal{O}(1)$ on $\PP(\Mat_{N+1})$ by considering the $(N + 1)^2$ matrix coordinates in basis $B$; we denote these coordinates by $a_{ij}$, where $i, j \in [N+1].$ Any 1-parameter subgroup $\ell$ that is diagonal in $B$ has an associated vector $r(\ell) = (r_1, \hdots, r_{N+1})$, and we have
$\ell(\tau) \cdot a_{ij} = \tau^{r_i - r_j}$. It follows that, with the choice of sheaf $\mathcal{L} = \mathcal{O}(1)$,
$$\wt(a_{ij}) = e_i - e_j.$$
Notice that these weights $\wt(a_{ij})$ are roots of the $A_N$ lattice.
\begin{defn}
A \emph{root polytope of type $A_N$} is the convex hull of a nonempty set of vectors of the form $e_i - e_j$, where $i, j \in [N+1]$. Note that we allow $i = j$. The \emph{full root polytope of type $A_N$} is the root polytope defined with all possible vectors $e_i - e_j$.
\end{defn}
The combinatorics of root polytopes is known in detail; see \cite{MR2801233, MR1697418} and especially \cite{MR4310906}. Full root polytopes are orbit polytopes, since their vertices are the orbit of $(1,-1,0,\hdots,0)$ by the symmetric group $S_{N+1}$.
We introduce the analogous definition for corner polyhedra.
\begin{defn} \label{def_corner}
A \emph{corner polyhedron of type $A_N$} is the corner hull $\Corner(\zeta)$ of a finite, nonempty set $\zeta$ of vectors of the form $e_i - e_j$, where $i, j \in [N+1]$. Note that we allow $i = j.$
\end{defn}
By Definition \ref{def_corner_poly}, the corner polyhedra of type $A_N$ are exactly the polyhedra of the form $\Pi_+(M)$, for some $M \in \mathbb{P}(\Mat_{N+1})$.
To compute the vertices and faces of corner polyhedra, it is convenient to work in the hyperplane of vectors in $\mathbb R^{N+1}$ with coordinates that sum to 0, for which we have a preferred basis. We recall our notation:
\begin{itemize}
\item $W$, the $\mathbb R$-vector space over the character lattice in basis $B$, identified with $\mathbb R^{N+1}$ by the basis $e_1, \hdots, e_{N+1}$.
\item $\bar{W} \subset W$, the hyperplane orthogonal to $\bm{1}$.
\item $R$, the dual space to $W$, identified with $\mathbb R^{N+1}$ by the basis $x_1, \hdots, x_{N+1}$.
\item $s_i$, for each $i \in [N]$, defined as the functional $x_1 + \hdots + x_i$.
\item For any $w \in W$, we write $\bar{w}$ for the projection of $w$ to $\bar{W}$, identified with $\mathbb R^N$ by the dual basis to $s_1, \hdots, s_N$.
\item $\bar{R}$, the dual space of $\bar{W}$, identified with $\mathbb R^N$ by the basis $s_1, \hdots, s_N \in \bar{R}$.
\item $\Pi_+(M)$, the corner polyhedron associated to $M$.
\end{itemize}
Our next lemma calculates weights in this preferred basis. Note that the trichotomy that appears here (\emph{on} vs. \emph{below} vs. \emph{above} the main diagonal) persists through the rest of our argument, and ultimately leads to the definitions of Type I, II, and III flags in Theorem \ref{thm_main}.
\begin{lemma} \label{lem_root_weights}
Let $a_{ij}$ be an entry coordinate of $\Mat_{N+1}$. Then $\overline{\wt}(a_{ij})$ depends on the $a_{ij}$ location relative to the main diagonal, as follows:
\begin{enumerate}
\item If $i = j$, then $\overline{\wt}(a_{ij}) = 0$.
\item If $i > j$, then
$$\overline{\wt}(a_{ij}) = (0, \hdots, 0, -1, \hdots, -1, 0, \hdots, 0),$$
where the $-1$-block runs from $j$ to $i - 1$ inclusive.
\item If $i < j$, then
$$\overline{\wt}(a_{ij}) = (0, \hdots, 0, 1, \hdots, 1, 0, \hdots, 0),$$ where the 1-block runs from $i$ to $j - 1$ inclusive.
\end{enumerate}
\end{lemma}
\begin{proof}
First, notice that $\wt(a_{ij}) = e_i - e_j$ is in $\bar{W}$ already, so $\overline{\wt}(a_{ij}) = \wt(a_{ij})$. The result follows from evaluating the functionals $s_1, \hdots, s_N$ on $\overline{\wt}{(a_{ij})}$.
\end{proof}
We now discuss the simplest examples in dimension $N = 1, 2, 3$. We include a catalogue of more elaborate examples in Section \ref{sect_examples}.
\begin{example}
Let $M \in \Mat_{N+1}$ be a matrix with all entries nonzero. Then the weight polytope $\Pi(M)$ is the full root polytope of type $A_N$.
\begin{itemize}
\item If $N = 1$, then $\Pi(M)$, viewed as a subset of $W = \mathbb R^2$, is the line segment from $(-1,1)$ to $(1,-1)$. In the preferred basis $s_1$ on $\bar{W}$, the polytope $\Pi(M)$ is the interval $[-1, 1]$. The corner polyhedron $\Pi_+(M)$ is the ray $[-1, \infty)$.
\item If $N = 2$, then $\Pi(M)$ is a regular hexagon in $W = \mathbb R^3$. Its vertices in $W$ are
\[
(-1,0,1), (-1,1,0), (0,-1,1), (1,0,-1), (1, -1, 0), (0,1,-1).
\]
Note that $(0,0,0) \in \wt(M)$ is not a vertex of $\Pi(M)$ because it is in the interior. In the basis $s_1, s_2$ on $\bar{W}$, the vertices of $\Pi(M)$ are
\[
(-1,-1), (-1,0), (0,-1), (1,1), (1,0), (0,1).
\]
We see from this description that the corner polyhedron $\Pi_+(M)$ is the nonnegative orthant translated by $(-1,-1)$.
\item If $N = 3$, then $\Pi(T)$ is a cuboctahedron. In $W$, its vertices are the 12 vectors with entries $0, 0, -1, 1$ in some order. In $\bar{W}$, its vertices are
\begin{align*}
&(-1,-1,-1), (-1,-1,0),(0,-1,-1), (-1,0,0), (0,-1,0), (0,0,-1), \\
&(1,1,1), (1,1,0),(0,1,1), (1,0,0), (0,1,0), (0,0,1).
\end{align*}
Thus $\Pi_+(M)$ is the nonnegative orthant translated by $(-1,-1,-1)$.
\end{itemize}
\end{example}
As the examples show, the corner polyhedron $\Pi_+(M)$ is generally much simpler than the weight polytope $\Pi(M)$.
We define a partial order on the coordinates of $\PP(\Mat_{N+1})$ that will help us identify the vertices of the corner polyhedron.
\begin{defn} \label{def_outweighing}
Given $(i,j), (i',j') \in [N+1]^2$, we say $(i, j)$ \emph{outweighs} $(i', j')$ if $\overline{\wt}(a_{i'j'}) \in \overline{\wt}(a_{ij}) + O^+$ and $\overline{\wt}(a_{i'j'}) \neq \overline{\wt}(a_{ij})$. Equivalently, entry $(i, j)$ outweighs $(i', j')$ if $\overline{\wt}(a_{i'j'}) \neq \overline{\wt}(a_{ij})$ and, for each functional $s = s_1, \hdots, s_N$, we have
$$s(\overline{\wt}(a_{ij})) \leq s(\overline{\wt}(a_{i'j'})).$$
The outweighing relation is a strict partial order on $[N+1]^2$. That is, outweighing is irreflexive, transitive and antisymmetric.
A \emph{minimal element} with respect to outweighing is an entry that is not outweighed by any other entries. To describe $\Pi_+(M)$, we only need to know the nonzero entries $(i,j)$ of $M$ that are minimal with respect to outweighing, among all nonzero entries of $M$.
\end{defn}
Outweighing can be described explicitly by cases, depending on the location of the entries $(i,j), (i',j')$ relative to the main diagonal.
\begin{lemma} \label{lem_outweighing}
Let $(i,j), (i',j') \in [N+1]^2.$
\begin{enumerate}
\item If $i > j$ and $i' > j'$, then $(i, j)$ outweighs $(i', j')$ if and only if $(i, j) \neq (i',j')$ and $i \geq i'$ and $j \leq j'$.
\item If $i > j$ and $i' \leq j'$, then $(i, j)$ outweighs $(i', j')$.
\item If $i < j$ and $i' < j'$, then $(i,j)$ outweighs $(i', j')$ if and only if $(i, j) \neq (i', j')$ and $i \geq i'$ and $j \leq j'.$
\end{enumerate}
\end{lemma}
\begin{proof}
These results follow immediately from Lemma \ref{lem_root_weights}.
\end{proof}
\begin{example}
Suppose that $N = 3$, so $M \in \Mat_{4}$.
\begin{itemize}
\item The entry $(4,2)$ outweighs every entry on or above the main diagonal, and also outweighs $(4,3)$ and $(3,2)$.
\item The entry $(2,4)$ only outweighs $(3,4)$ and $(4,4)$.
\item The entry $(2,2)$ outweighs every entry above the main diagonal.
\end{itemize}
\end{example}
To keep track of minimal elements (Definition \ref{def_outweighing}), we introduce a combinatorial gadget called the \emph{profile} of a matrix. The matrix profile provides the key link between corner polyhedra and the special Hessenberg functions that appear in Theorem \ref{thm_main}. We show in Corollary \ref{cor_bijection} that the profile of a matrix exactly describes its corner polyhedron.
\begin{defn} \label{def_profile}
A \emph{path} in an $(N+1) \times (N+1)$ matrix is a word of length $2N + 2$ in the letters $\bm{\downarrow}$ and $\bm{\rightarrow}$, consisting of $N + 1$ instances of each letter. We identify each path with the corresponding sequence of down-directed and right-directed line segments in the gridlines between entries. A \emph{Dyck path} is a path that does not cross the main diagonal, but may meet it. A \emph{lower Dyck path} is a Dyck path beginning with $\bm{\downarrow}$, and an \emph{upper Dyck path} is a Dyck path beginning with $\bm{\rightarrow}$. Note that the first letter determines whether the Dyck path is below or above the main diagonal.
A path \emph{contains} a pair $(i,j) \in [N+1]^2$ if $(i,j)$ is above the path.
The \emph{profile} of a matrix $M$, denoted $\Prof(M)$, is defined as follows.
\begin{enumerate}
\item Suppose that $M$ is not strictly-upper-triangular. Then $\Prof(M)$ is defined to be the upper Dyck path in the grid lines of $M$ that contains the fewest entries among all upper Dyck paths that contain all nonvanishing entries.
\item Suppose that $M$ is strictly upper-triangular. Then $\Prof(M)$ is defined as the lower Dyck path that contains the fewest entries among all lower Dyck paths that contain all nonvanishing entries.
\end{enumerate}
We call $\Prof(0)$ the \emph{trivial profile}; we treat it as an exceptional case.
\end{defn}
\begin{remark}
For strictly upper-triangular matrices $M$, the profile $\Prof(M)$ is exactly the data of the Hessenberg function (Definition \ref{defn:flaginkN1}) associated to $M$ with respect to the standard complete flag. For all other $M$, the profile of $M$ is the data of the pointwise maximum of the identity function and that Hessenberg function.
\end{remark}
\begin{defn} \label{def_pivotal_entry}
A \emph{pivotal entry} of a matrix $M$ is a pair $(i,j)$, where $i, j \in [N+1]$, such that the path $\Prof(M)$ follows the closest gridlines to the left of $(i,j)$ and below $(i,j)$. In other words, the pivotal entries correspond to appearances of the subword $\bm{\downarrow} \bm{\rightarrow}$ in $\Prof(M)$.
\end{defn}
\begin{lemma} \label{lem_minimal_entries}
Let $M \in \Mat_{N+1}$ be a nonzero matrix. Let $E(M)$ denote the set of nonvanishing entries $(i, j) \in [N+1]^2$ of $M$ that are minimal with respect to outweighing (Definition \ref{def_outweighing}) among all nonvanishing entries of $M$. Then $E(M)$ is a subset of the set of pivotal entries of $M$, as follows:
\begin{enumerate}
\item If $M$ is upper-triangular but not strictly-upper-triangular, then $E(M)$ is the set of nonvanishing diagonal entries of $M$.
\item If $M$ is not upper-triangular, then $E(M)$ is the set of pivotal entries $(i,j)$ of $M$ such that $i > j$.
\item If $M$ is strictly upper-triangular, then $E(M)$ is the set of pivotal entries of $M$.
\end{enumerate}
\end{lemma}
\begin{proof}
\par(1)\enspace Since $M$ is upper-triangular but not strictly-upper-triangular, it has a nonvanishing diagonal entry, so by Lemma \ref{lem_outweighing}, no entry $(i, j)$ with $i < j$ is minimal. Since $M$ is upper-triangular, by Lemma \ref{lem_outweighing}, all the diagonal nonvanishing entries are minimal.
\par(2) \enspace Since $M$ is not upper-triangular, it has a nonvanishing entry below the main diagonal. By Lemma \ref{lem_outweighing}, that entry outweighs every entry on or above the main diagonal. So every element of $E(M)$ is below the main diagonal. Consider a nonvanishing entry $(i',j')$ below the main diagonal. The entry $(i', j')$ is outweighed by some entry $(i, j)$ if and only if $(i',j')$ is not pivotal.
\par(3) \enspace Consider an entry $(i',j')$ above the main diagonal. Since $M$ is strictly upper-triangular, by Lemma \ref{lem_outweighing}, the entry $(i',j')$ is outweighed by some entry $(i,j)$ if and only if $(i',j')$ is not pivotal.
\end{proof}
The set of pivotal entries of $M$ has a natural ordering, namely the order in which the pivotal entries appear along the path from upper-left to bottom-right.
Recall that the set $\Ctrl(M)$ of \emph{control points} is the set of vertices of the corner polyhedron $\Pi_+(M)$. The next proposition is the key result of this section. It shows that control points essentially correspond to pivotal entries of $M$.
\begin{prop} \label{prop_vertices}
For any nonzero matrix $M \in \Mat_{N+1}$, there is a canonical ordering on the elements of $\Ctrl(M)$.
\begin{itemize}
\item If $M$ is upper triangular but not strictly-upper-triangular, then $\Ctrl(M) = \{0\}$, so the ordering is the trivial one.
\item For any other nonzero $M$, the off-diagonal pivotal entries of $M$ are in bijection with $\Ctrl(M)$ by the map
$$(i,j) \mapsto \overline{\wt}(a_{ij}).$$
The canonical ordering on $\Ctrl(M)$ can be defined in any of the following equivalent ways:
\begin{itemize}
\item Take the ordering inherited from the natural ordering on the pivotal entries;
\item Take the ordering by the index of the first nonzero coordinate;
\item Take the ordering by the index of the last nonzero coordinate.
\end{itemize}
\end{itemize}
\end{prop}
\begin{proof}
Let $E(M)$ be as defined in Lemma \ref{lem_minimal_entries}. Let
$$\overline{\wt}(E(M)) = \{ \overline{\wt}(a_{ij}) \colon (i,j) \in E(M) \}.$$
We consider the same three cases.
\par(1)\enspace Suppose $M$ is upper triangular but not strictly-upper-triangular. Then it follows immediately from Lemma \ref{lem_root_weights} that $\Ctrl(M) = \{0\}$.
\par(2)\enspace Suppose $M$ is not of the form of case (1). We claim that $\Ctrl(M) = \overline{\wt}(E(M))$. The containment $\Ctrl(M) \subseteq \overline{\wt}(E(M))$ follows from the definition of $E(M)$, since if some entry $(i,j)$ is outweighed by a nonvanishing entry $(i', j')$, then $\overline{\wt}(a_{ij}) \in \overline{\wt}(a_{i' j'}) + O^+$ and $\overline{\wt}(a_{ij}) \neq \overline{\wt}(a_{i'j'})$, so $\overline{\wt}(a_{ij})$ is not a vertex.
By Lemma \ref{lem_minimal_entries}, there is an induced ordering on $E(M)$ from the canonical ordering on the pivotal entries. Since $\overline{\wt}$ is injective on $E(M)$, we also obtain an ordering on $\overline{\wt}(E(M))$, and by restriction, on $\Ctrl(M)$. This is the canonical ordering in the theorem statement. The canonical ordering on the set of pivotal entries can equivalently be described as ordering by increasing row or increasing column in $M$. Combining this observation with the explicit description of the weights in Lemma \ref{lem_root_weights} gives us the three equivalent definitions of the ordering on $\Ctrl(M)$.
We claim $\Ctrl(M) \supseteq \overline{\wt}(E(M))$. To prove this, we subdivide into two cases.
\par(2a)\enspace Assume $M$ is non-upper-triangular. Then $\Prof(M)$ is a lower Dyck path.
Given $w \in \overline{\wt}(E(M))$, we must show that $w$ is a vertex of $\Pi_+(M)$. This is true if there are $N$ independent linear functionals $r \in \bar{R}^+$ which achieve their minimum on $\Pi_+(M)$ at $w$. Say $\kappa = \# \overline{\wt}(E(M))$. Using the canonical ordering, let $w_c$ be the $c$-th element of $\overline{\wt}(E(M))$. It is of the form
$$(0,\hdots,0,-1,\hdots,-1,0,\hdots,0),$$
where, for some $t_c, u_c \in [N]$ with $t_c \leq u_c$, the $-1$'s run from index $t_c$ to $u_c$ inclusive. It follows from the canonical ordering that, for any $c', c'' \in [\kappa]$ with $c' < c''$, we have $t_{c'} < t_{c''}$ and $u_{c'} < u_{c''}$. Therefore, the following $N$ functionals on $\overline{\wt}(E(M))$ are minimized at $w_c$.
\begin{itemize}
\item For all $i$ in the range $1 \leq i < t_c$, the functional
$s_i + s_{u_{c}}$ on $\overline{\wt}(E(M))$ achieves its minimum value $-1$ at $w_c$.
\item For all $i$ in the range $t_c \leq i \leq u_c$, the functional
$s_i + s_{u_c}$
on $\overline{\wt}(E(M))$ achieves its minimum value $-2$ at $w_c$.
\item For all $i$ in the range $u_c < i \leq N$, the functional
$s_{t_c} + s_i$
on $\overline{\wt}(E(M))$ achieves its minimum value $-1$ at $w_c$.
\end{itemize}
The $N$ functionals named above are easily seen to be independent, so $w_c$ is a vertex of the corner hull $\Pi_+(M)$ of $\overline{\wt}(E(M))$.
\par(2b)\enspace Assume $M$ is strictly upper-triangular. Then $\Prof(M)$ is an upper Dyck path. As in (2a), we claim $\Ctrl(M) \supseteq \overline{\wt}(E(M))$. Given $w \in \overline{\wt}(E(M))$, we seek $N$ independent linear functionals $r \in \bar{R}^+$ which achieve their minimum on $\Pi_+(M)$ at $w$. Say $\kappa = \# \overline{\wt}(E(M))$. Using the canonical ordering, let $w_c$ be the $c$-th element of $\overline{\wt}(E(M))$. It is of the form
$$(0,\hdots,0,1,\hdots,1,0,\hdots,0),$$
where for some $t_c, u_c \in [N]$ with $t_c \leq u_c$, the $1$'s run from index $t_c$ to $u_c$ inclusive. It follows from the canonical ordering that, for any $c', c'' \in [\kappa]$ with $c' < c''$, we have $t_{c'} < t_{c''}$ and $u_{c'} < u_{c''}$. Then the following $N$ functionals are independent and simultaneously minimized at $w_c$.
\begin{itemize}
\item For all $1 \leq i < t_c$ and $i > u_c$, the functional $s_i$ on $\overline{\wt}(E(M))$ achieves its minimum value $0$ at $w_c$.
\item For all $i$ in the range $t_c \leq u_c$, the functional
$$s_1 + \hdots + s_{t_c - 1} + s_i + s_{u_c + 1} + \hdots + s_\kappa$$
achieves its minimum value 1 at $w_c$.
\end{itemize}
\end{proof}
Finally, we classify corner polyhedra of type $A_N$.
\begin{cor} \label{cor_bijection}
Two nonzero matrices in $\Mat_{N+1}$ have the same corner polyhedron if and only if they have the same profile.
\end{cor}
\begin{proof}
Let $M_1, M_2 \in \Mat_{N+1}$. By the basic properties of the corner hull operation, we have $\Pi_+(M_1) = \Pi_+(M_2)$ if and only if $\Ctrl(M_1) = \Ctrl(M_2)$. By Lemma \ref{prop_vertices}, two matrices $M_1$ and $M_2$ satisfy $\Ctrl(M_1) = \Ctrl(M_2)$ if and only if $\overline{\wt}(E(M_1)) = \overline{\wt}(E(M_2))$. The set $\overline{\wt}(E(M))$ is determined by the off-diagonal pivotal entries of $M$, which is determined by $\Prof(M)$.
\end{proof}
\begin{cor} \label{cor_catalan}
Let $C_N$ be the $N$-th Catalan number. The number of corner polyhedra of type $A_N$ is $2 C_{N+1} - 1$.
\end{cor}
\begin{proof}
The corner polyhedra of type $A_N$ all arise as corner polyhedra associated to matrices by Corollary \ref{cor_bijection}, and so it suffices to count the set
$$\{ \Prof(M) \colon 0 \neq M \in \Mat_{N+1}\}$$
instead. Any lower Dyck path and nontrivial upper Dyck path may appear as a matrix profile. By definition, the Catalan numbers count Dyck paths, so there are $C_{N+1}$ lower Dyck paths and $C_{N+1} - 1$ nontrivial upper Dyck paths.
\end{proof}
For another connection between Catalan numbers and $A_N$ root polytopes, see \cite[Theorem 8]{MR1697418}.
\subsection{Facets of corner polyhedra of type \texorpdfstring{$A_N$}{A N}} \label{sec_facets}
We have calculated the vertices of $\Pi_+(M)$ for any nonzero matrix $M \in \Mat_{N+1}$ (Proposition \ref{prop_vertices}). In this section, we calculate the facets of $\Pi_+(M)$ (Proposition \ref{prop_facets}). These results are the key input for proving Theorem \ref{thm_main}; the facets of $\Pi_+(M)$ roughly correspond to Type I, II, and III flags relative to the linear map $M$.
We work with halfspaces in $\bar{W}$, which are of the form $\{ w \in \bar{W} \colon s(w) \geq c \}$ for some $s \in \bar{R}^+, c \in \mathbb R$. For convenience, we write halfspaces as $\{ s \geq c \}$.
For any nonempty $I \subseteq [N]$, we define a functional
$$s_I := \sum_{i \in I} s_i.$$
\begin{defn}
The \emph{standard flag in $k^{N+1}$} is the sequence
$$0 = S_0 \subsetneq S_1 \subsetneq \hdots \subsetneq S_{N+1} = k^{N+1}$$ of subspaces of $k^{N+1}$, where for each $\rho \in [N]$, the subspace $S_\rho$ is the span of the first $\rho$ standard basis vectors.
The \emph{Hessenberg function} of $M$, denoted $h_M$, is the Hessenberg function of the linear map defined by $M$ on $k^{N+1}$ relative to the standard flag:
\begin{align*}
h_{M} : \; & \{0, \hdots, N + 1\} \to \{0, \hdots N + 1\}, \\
& i \mapsto \min \{j : MS_i \subseteq S_j\}.
\end{align*}
Each nonempty subset
$$I = \{i_1, \hdots, i_\gamma\} \subseteq [N]$$
determines a subflag
$$0 = S_{i_0} \subsetneq S_{i_1} \subsetneq \hdots \subsetneq S_{i_\gamma} \subsetneq S_{i_{\gamma + 1}} = k^{N+1}$$
of the standard flag. Let $h_{M, I}$ denote the Hessenberg function of $M$ with respect to that subflag. Explicitly,
\begin{align*}
h_{M, I} : \; & \{0, \hdots, \gamma + 1 \} \to \{ 0, \hdots, \gamma + 1 \}, \\
& t \mapsto \min \{t' : MS_{i_t} \subseteq S_{i_{t'}}\}.
\end{align*}
\end{defn}
The following proposition is the main result of this section.
\begin{prop} \label{prop_facets}
Let $M \in \Mat_{N + 1}$ be a nonzero matrix. Let $0 = S_0 \subset S_1 \subset \hdots \subset S_{N+1}$ be the standard flag. The facets of the corner polyhedron $\Pi_+(M)$ are of four kinds, called (F-1A), (F-1B), (F-2A), and (F-2B), as follows. For each nonempty subset $$I = \{i_1, \hdots, i_\gamma\} \subseteq [N],$$
where by convention we let $\gamma = \#I$ and
$$0 = i_0 < i_1 < \hdots < i_\gamma < i_{\gamma + 1} = N + 1,$$
there is at most one facet of $\Pi_+(M)$.
If $M$ is not strictly-upper-triangular:
\begin{itemize}
\item (F-1A): If
\begin{equation*}
\label{eq_1A}
\gamma = 1 \quad \text{and} \quad M S_{i_1} \subseteq S_{i_1},
\tag{\ref{prop_facets}.1A}
\end{equation*}
there is a facet
$$\{ s_I \geq 0 \}.$$
\item (F-1B): If
\begin{equation*} \label{eq_1B}
h_{M, I}(t) = t + 1 \quad (1 \leq t \leq \gamma),
\tag{\ref{prop_facets}.1B}
\end{equation*}
there is a facet
$$ \left\{ s_I \geq -1 \right\}. $$
\end{itemize}
If $M$ is strictly upper-triangular:
\begin{itemize}
\item (F-2A): If
\begin{equation*} \label{eq_2A}
\gamma = 1 \quad \text{ and } \quad (MS_{N+1} \not\subseteq S_{i_1} \quad \text{or} \quad MS_{i_1} \neq 0),
\tag{\ref{prop_facets}.2A}
\end{equation*}
there is a facet
$$\{ s_I \geq 0 \}.$$
\item (F-2B): If
\begin{equation*} \label{eq_2B}
h_{M, I}(t) = t - 1 \quad (1 \leq t \leq \gamma + 1),
\tag{\ref{prop_facets}.2B}
\end{equation*}
there is a facet
$$\{ s_I \geq 1\}.$$
\end{itemize}
\end{prop}
The proof of Proposition \ref{prop_facets} is involved. We first prove intermediate results, Lemmas \ref{lem_facets_lower_ctrl} and \ref{lem_facets_upper_ctrl}, that compute facets of $\Pi_+(M)$ via the following combinatorial gadget.
\begin{defn}
For any nonzero matrix $M \in \Mat_{N+1}$, we view $\Ctrl(M)$ as a matrix, the \emph{control matrix} of $M$, by ordering its elements according to the canonical ordering of Proposition \ref{prop_vertices}. Let $\kappa = \# \Ctrl(M)$. Then $\Ctrl(M)$ is a $N \times \kappa$ matrix.
For examples, see Section \ref{sect_examples}.
\end{defn}
The form of $\Ctrl(M)$ is one of three kinds.
\begin{enumerate}
\item If $M$ is upper-triangular, but not strictly so, then $\Ctrl(M) = 0.$
\item If $M$ is not upper-triangular, then the entries of $\Ctrl(M)$ are each 0 or $-1$. For each $c \in [\kappa]$, say the $c$-th off-diagonal pivotal entry of $M$ is $(t_c, u_c - 1)$. Then column $c$ of $\Ctrl(M)$ consists of a single block of $-1$'s running from $t_c$ to $u_c$ inclusive, and has all other entries 0. By Proposition \ref{prop_vertices}, if $c', c'' \in [\kappa]$ satisfy $c' < c''$, then $t_{c'} < t_{c''}$ and $u_{c'} < u_{c''}$.
\item If $M$ is strictly upper-triangular, then the entries of $\Ctrl(M)$ are each $0$ or $1$. For each $c \in [\kappa]$, say the $c$-th off-diagonal pivotal entry of $M$ is $(u_c, t_c - 1)$; note the difference from the previous case. Then column $c$ of $\Ctrl(M)$ consists of a single block of $1$'s running from $t_c$ to $u_c$ inclusive, and has all other entries 0. By Proposition \ref{prop_vertices}, if $c', c'' \in [\kappa]$ satisfy $c' < c''$, then $t_{c'} < t_{c''}$ and $u_{c'} < u_{c''}$.
\end{enumerate}
In order to understand facets from the control matrix, we look at certain submatrices, which themselves correspond to lower-dimensional corner polyhedra. We set this argument up as follows.
Let $y_1, \hdots, y_N \in \bar{W}$ be the dual vectors to $s_1, \hdots, s_N$. Given any subset $I \subseteq [N]$, let $\bar{W}_I$ be the subspace of $\bar{W}$ generated by the $y_i, i \in I$. If $i' \not\in I$, then $s_{i'}$ is identically 0 on $\bar{W}_I$. Let $\gamma = \#I$. Taking $s_i$ for each $i \in I$ as coordinates identifies $\bar{W}_I$ with $\mathbb R^{\gamma}$. Define $\Pi_+^I(M)$ to be the projection of $\Pi_+(M)$ onto $\bar{W}_I$.
Observe that $\Pi_+^I(M)$ is itself a corner polyhedron in $\bar{W}_I$ with this choice of basis. We denote the set of control points of $\Pi_+^I(M)$ in $\bar{W}_I$ by ${\Verts}_+^I (M)$. It follows that there is a canonical ordering on ${\Verts}_+^I (M)$, so writing $\kappa' = \# {\Verts}_+^I(M)$, the set ${\Verts}_+^I(M)$ can be identified with a $\gamma \times \kappa'$ matrix.
The matrix ${\Verts}_+^I(M)$ can be obtained from $\Ctrl(M)$, as follows. Take the $\gamma \times \kappa$ minor of $\Ctrl(M)$ consisting of the rows with indices in $I$. If there are two columns $w_1, w_2$ in the minor with distinct indices such that $w_1 \in w_2 + O_+$ in $\bar{W}_I$, then remove $w_1$. Repeat the previous step until no such pairs exist, and the result is ${\Verts}_+^I(M)$.
In Lemma \ref{lem_facets_lower_ctrl}, we show how the facets of $\Pi_+(M)$ may be read off the control matrix for certain matrices, an intermediate step in proving Proposition \ref{prop_facets}.
\begin{lemma} \label{lem_facets_lower_ctrl}
Let $M \in \Mat_{N+1}$ be not strictly-upper-triangular. The facets of the corner polyhedron $\Pi_+(M)$ are as follows.
\begin{itemize}
\item (F-1A): For each all-zero row $\rho$ of $\Ctrl(M)$, there is a facet
$$\{ s_\rho \geq 0\}.$$
\item (F-1B): For each nonempty subset $I \subseteq [N]$ such that ${\Verts}_+^I(M) = -1$, there is a facet
$$\{ s_I \geq -1\}.$$
\end{itemize}
\end{lemma}
\begin{proof}
First, we check that these halfspaces are facets of $\Pi_+(M)$.
If we have an all-zero row $\rho$ of $\Ctrl(M)$, then writing $I = \{\rho\}$, we find ${\Verts}_+^I(M) = 0 \in \Mat_1$. Thus $\Pi_+^I(M)$ has one facet $F$, defined by $\{s_\rho \geq 0\}$. We can see that $\{ s_\rho \geq 0\}$ is a supporting halfspace of $\Pi_+(M)$. To check that it defines a facet, we show
$$\dim (\Pi_+(M) \cap \{ s_\rho = 0 \}) \geq N - 1.$$
To see this, notice that the functional $s_\rho$ is identically $0$ on the convex cone $F'$ generated by $y_i, i \not\in I$. Since $F$ and $F'$ are orthogonal, we have $\dim F + F' = N - 1$, and
$$F + F' \subseteq \Pi_+(M) \cap \{ s_\rho = 0\}.$$
Suppose that a nonempty subset $I \subseteq [N]$ satisfies ${\Verts}_+^I(M) = -1 \in \Mat_\gamma$. Then $\{s_I = -1\}$ contains $\gamma$ vertices of ${\Verts}_+^I(M)$, and $\Pi^I_+(M) \subset \{s_I \geq -1\}$, so $\{s_I \geq -1\}$ defines a facet $F$ of $\Pi_+^I(M)$. Thus $\{s_I \geq -1\}$ is a supporting halfspace of $\Pi_+(M)$. To check that it defines a facet, we show
$$\dim(\Pi_+(M) \cap \{s_I = -1\}) \geq N - 1.$$
To see this, notice that the functional $s_I$ is identically 0 on the convex cone $F'$ generated by $y_i, i \not\in I$. Since $F$ and $F'$ are orthogonal, we have $$\dim F + F' = \gamma - 1 + (N - \gamma) = N - 1,$$
and $$F + F' \subseteq \Pi_+(M) \cap \{s_I = -1\}.$$
It remains to show that every facet of $\Pi_+(M)$ is of one of these two forms. Suppose that $\{ r > c \}$ is a facet, where $r \in \bar{R}^+$ is a nonzero functional and $c \in \mathbb R$. Let $I \subseteq [N]$ be the set of indices $i \in [N]$ such that $r(y_i) \neq 0$. Since $r \neq 0$, the set $I$ is nonempty. Projecting to $\bar{W}_I$, the halfspace $\{ r > c \}$ defines a facet $F$ of $\Pi_+^I(M)$. We claim that $\Pi_+^I(M)$ has $\#I$ vertices. Since ${\Verts}_+^I(M)$ has $\#I$ rows, by the canonical ordering, it has at most $\#I$ columns, so $\Pi_+^I(M)$ has at most $\#I$ vertices. On the other hand, the facet $F$ is the Minkowski sum of some face $F''$ of $O^+$ in $\bar{W}_I$ and a face of $\Conv({\Verts}_+^I(M))$. But $F''$ is necessarily the origin, by construction of the set $I$. It follows that $F$ is a facet of the convex polytope $\Conv({\Verts}_+^I(M))$. Since $\dim F = \#I - 1$, at least $\#I$ elements of ${\Verts}_+^I(M)$ lie on $F$, supplying the claimed number of vertices of $\Pi_+^I(M)$.
Since ${\Verts}_+^I(M)$ is a square matrix, it follows from the canonical ordering that either ${\Verts}_+^I(M) = 0$ or ${\Verts}_+^I(M) = -1 \in \Mat_\gamma$, where $\gamma = \# I$. If ${\Verts}_+^I(M) = 0$, then $I$ is some singleton $\{\rho\}$, and this determines the facet; we conclude $\{r \geq c\} = \{s_\rho \geq 0\}$. If ${\Verts}_+^I(M) = -1$, then the facet containing the necessary $\gamma$ vertices is given explicitly by $\{r \geq c\} = \{s_i \geq -1\}$.
\end{proof}
In Lemma \ref{lem_facets_upper_ctrl}, we show how the facets of $\Pi_+(M)$ may be read off the control matrix for strictly-upper-triangular matrices, another step in proving Proposition \ref{prop_facets}.
\begin{lemma} \label{lem_facets_upper_ctrl}
Let $M \in \Mat_{N+1}$ be strictly upper-triangular and nonzero. The facets of the corner polyhedron $\Pi_+(M)$ are as follows.
\begin{itemize}
\item (F-2A): For each $\rho \in [N]$ such that ${\Verts}_+^I(M) = 0 \in \Mat_1$, there is a facet
$$\{s_\rho \geq 0\}.$$
\item (F-2B): For each nonempty subset $I \subseteq [N]$ such that ${\Verts}_+^I(M) = 1 \in \Mat_{\#I}$, there is a facet
$$\{s_I \geq 1 \}.$$
\end{itemize}
\end{lemma}
\begin{proof}
The argument resembles the proof of Lemma \ref{lem_facets_lower_ctrl}. First, we check that these halfspaces are facets of $\Pi_+(M)$.
Suppose that $I = \{\rho\}$ satisfies ${\Verts}_+^I(M) = 0 \in \Mat_1$. By the proof of Lemma \ref{lem_facets_lower_ctrl}, the supporting halfspace $\{s_\rho \geq 0 \}$ defines a facet of $\Pi_+(M)$.
Suppose that a nonempty subset $I \subseteq [N]$ satisfies ${\Verts}_+^I(M) = 1 \in \Mat_\gamma$, where $\gamma = \#I$. Then $\{s_I = 1\}$ contains all $\gamma$ vertices of $\Pi_+^I(M)$, so again by the proof of Lemma \ref{lem_facets_lower_ctrl}, there is a facet of $\Pi_+(M)$ defined by $\{s_I \geq 1\}$.
It remains to show that every facet of $\Pi_+(M)$ is of one of these two forms. By the proof of Lemma \ref{lem_facets_lower_ctrl}, if $\{r > c\}$ defines a facet of $\Pi_+(M)$, and $I = \{i \in [N] \colon r(y_i) \neq 0\}$, then ${\Verts}_+^I(M)$ is a square matrix. It follows from the canonical ordering that either ${\Verts}_+^I(M) = 0$ or ${\Verts}_+^I(M) = 1 \in \Mat_\gamma$, where $\gamma = \#I$. If ${\Verts}_+^I(M) = 0$, then $I$ is some singleton $\{\rho\}$, and this determines the facet; we conclude $\{r \geq c\} = \{s_\rho \geq 0\}$. If ${\Verts}_+^I(M) = 1$, then the facet containing the necessary $\gamma$ vertices is given explicitly by $\{r \geq c\} = \{s_I \geq 1\}$.
\end{proof}
Finally, we prove the main result of this section.
\begin{proof}[Proof of Proposition \ref{prop_facets}]
By Lemma \ref{lem_facets_lower_ctrl} and Lemma \ref{lem_facets_upper_ctrl}, it suffices to prove the equivalence of the conditions for each facet type.
Recall the definition of $E(M)$ from Lemma \ref{lem_minimal_entries}. Observe that, for any not strictly-upper-triangular $M$, and for any $\rho \in [N]$ and $c \in [\kappa]$,
\[
\Ctrl(M)_{\rho c} =
\begin{cases}
-1, & \text{the } c \text{-th element of } E(M) \text{ is in rectangle } [\rho + 1 , N + 1] \times [1, \rho],\\
0, & \text{otherwise}.
\end{cases}
\]
Similarly, for any strictly-upper-triangular $M$,
\[
\Ctrl(M)_{\rho c} =
\begin{cases}
1, & \text{the } c \text{-th element of } E(M) \text{ is in rectangle } [1, \rho] \times [\rho + 1, N + 1],\\
0, & \text{otherwise}.
\end{cases}
\]
\par(F-1A)\enspace We must show that, for a given nonempty subset $I \subseteq [N]$, the condition \eqref{eq_1A} is equivalent to ${\Verts}_+^I(M) = 0 \in \Mat_1$. Since $\gamma = 1$ is equivalent to ${\Verts}_+^I$ having exactly one row, we must show that under this condition,
$${\Verts}_+^I(M) = [0] \Leftrightarrow MS_{i_1} \subseteq S_{i_1} .$$
Suppose that ${\Verts}_+^I(M) = [0]$. By the canonical ordering, row $i_1$ of $\Ctrl_M$ is all zeroes. By the observation, there are no pivotal entries of $M$ in rectangle $[i_1 + 1, N + 1] \times [1, i_1]$. Then, all entries of $M$ in that rectangle must be 0. It follows that $MS_{i_1} \subseteq S_{i_1}$.
For the reverse implication, suppose that $MS_{i_1} \subseteq S_{i_1}$. All entries of $M$ in the rectangle $[i_1 + 1, N + 1] \times [1, i_1]$ are zero, so in particular, there are no pivotal entries of $M$ in that rectangle. Then, by the observation, we deduce that $\Ctrl(M)$ has an all-0 row $i_1$, so ${\Verts}_+^I(M) = [0]$.
\par(F-1B)\enspace For each nonempty $I \subseteq [N]$, we must prove the equivalence
$${\Verts}_+^I(M) = -1 \Leftrightarrow \eqref{eq_1B}.$$
If $I$ satisfies \eqref{eq_1B}, we claim there at least $\#I$ elements of ${\Verts}_+^I(M)$. To each $i \in I$, we can associate an element of $E(M)$, as follows. In column $i$ of $M$, let $(i', i)$ be the bottom-most nonzero entry; then consider the left-most nonzero entry in row $i'$; call this entry $\iota(i)$. The entry $(i', j')$ is in $E(M)$ since the path $\Prof(M)$ borders it to the left and below. By \eqref{eq_1B}, the map $\iota \colon I \to E(M)$ is injective. We may compute the entries of ${\Verts}_+^I(M)$ using the observation to directly calculate that ${\Verts}_+^I(M) = -1$.
Suppose that $I$ does not satisfy \eqref{eq_1B}. Then there is some $i \in I$ such that $S_i = MS_i$ or $MS_i \not\subseteq S_{i'}$, where $i'$ is the next element in $I$. In the first case, the rectangle $[i+1, N + 1] \times [1, i]$ of $M$ is all zero. By the observation, row $i$ of ${\Verts}_+^I(M)$ is all 0, so ${\Verts}_+^I(M) \neq -1$. In the second case, there exists a nonzero entry of $M$ in $[i' + 1, N + 1] \times [ 1, i]$, so there is a pivotal entry in that rectangle. By the observation, the corresponding column of $\Ctrl(M)$ has a $-1$ in row $i$ and $i'$, so ${\Verts}_+^I(M) \neq -1$.
\par(F-2A)\enspace We must show that, for nonzero strictly-upper-triangular $M$,
$${\Verts}_+^I(M) = [0] \Leftrightarrow \eqref{eq_2A}.$$
Suppose that ${\Verts}_+^I(M) \neq [0]$. By the canonical ordering, row $i_1$ of $\Ctrl(M)$ is all 1's. By the observation, every pivotal entry of $M$ is in the rectangle $[1, i_1] \times [i_1 + 1, N + 1]$. Then every nonzero entry of $M$ is in that rectangle, so \eqref{eq_2A} does not hold. For the reverse implication, if \eqref{eq_2A} does not hold, then every entry of $M$ is in that rectangle, so every pivotal entry of $M$ is in that rectangle, so ${\Verts}_+^I(M) = [1].$
\par(F-2B)\enspace We must show that, for nonzero strictly-upper-triangular $M$, for each nonempty $I \subseteq [N]$,
$${\Verts}_+^I(M) = 1 \Leftrightarrow \eqref{eq_2B}.$$
Suppose that ${\Verts}_+^I(M) = 1$. Then, using the observation, in the notation of Proposition \ref{prop_facets}, for each $0 \leq t \leq \#I$, the rectangle $[i_t + 1, N + 1] \times [1, i_{t+1}]$ of $M$ has no pivotal entries, and for each $1 \leq t \leq \#I$, there is a pivotal entry of $M$ in $[i_{t - 1} + 1, i_t] \times [1, i_{t+1}]$. It follows that the rectangle $[i_t + 1, N + 1] \times [1, i_{t+1}]$ is all 0's and there is a nonzero entry in $[i_{t - 1} + 1, i_t] \times [1, i_{t+1}]$, implying \eqref{eq_2B}.
Conversely, suppose that \eqref{eq_2B} holds. For each $0 \leq t \leq \#I$, the rectangle $[i_t + 1, N + 1] \times [1, i_{t+1}]$ is all 0's, and for each $1 \leq t \leq \#I$, there is a nonzero entry in $[i_{t - 1} + 1, i_t] \times [1, i_{t+1}]$. It follows that the map $\iota$ of (F-1B) is injective, so there are at least $\#I$ elements of $\Ctrl (M)$. Using the observation, we calculate explicitly that ${\Verts}_+^I(M) = 1.$
\end{proof}
\subsection{Examples} \label{sect_examples}
In the following examples, we fix some $N \geq 1$ and consider some $(N + 1) \times (N + 1)$ matrix $M$. We mark nonzero entries by $\bm{\ast}$, and entries with no conditions on them by $\ast$.
\begin{example} \label{ex_generic}
Suppose that entry $(N+1, 1)$ of $M$ is nonzero. Then the profile $\Prof(M)$ is the lower Dyck path
\[
\left[
\begin{array}{ccc}
\multicolumn{1}{|c}{*} & \hdots & \ast \\
\multicolumn{1}{|c}{\vdots} & \ddots & \vdots \\
\multicolumn{1}{|c}{\bm{*}} & \hdots & \ast \\ \cline{1-3}
\end{array}
\right].
\]
By Proposition \ref{prop_vertices}, the control matrix $\Ctrl(M)$ is the $N \times 1$ matrix with every entry equal to $-1$. Thus the facets of $\Pi_+(M)$ are, for each $i \in [N]$,
$$\{ s_i \geq -1 \}.$$
This is obvious because $\Ctrl(M)$ has just one column, but we can also see it as a special case of Proposition \ref{prop_facets} in which the facets are all of type (F-1B).
This is the least degenerate matrix $M$, and the corner polyhedron is as simple as can be. But note that, even though a generic matrix is described by this example, a generic linear transformation $T$ is not, since there is always some basis in which $T$ is of a different form.
\end{example}
\begin{example} \label{ex_most_degen}
Suppose that the only nonzero entry of $M$ is $(1, N + 1)$. Then the profile $\Prof(M)$ is the upper Dyck path
\[
\left[
\begin{array}{cccc}
\cline{1-3}
0 & \hdots & 0 & \multicolumn{1}{|c}{\bm{*}} \\ \cline{4-4}
0 & \ddots & 0 & \multicolumn{1}{c|}{0} \\
\vdots & \ddots & \ddots & \multicolumn{1}{c|}{\vdots} \\
0 & \hdots & 0 & \multicolumn{1}{c|}{0} \\
\end{array}
\right].
\]
We can think of this matrix $M$ as the worst-case scenario for stability. By Proposition \ref{prop_vertices}, the control matrix $\Ctrl(M)$ is the $N \times 1$ matrix with every entry equal to 1. Since there is only one column of $\Ctrl(M)$, the corner polyhedron $\Pi_+(M)$ has facets $\{s_i \geq 1\}$, for each $i \in [N]$. As a special case of Proposition \ref{prop_facets}, the facets are all of type (F-2B).
\end{example}
\begin{example} \label{ex_main_stair}
Suppose that $M$ is upper-triangular and entry $(1,1)$ is nonzero. This example includes the case $M = 1$. Since there is a nonzero entry on the main diagonal, the profile $\Prof(M)$ is a lower Dyck path, specifically
\[
\left[
\begin{array} {cccc}
\multicolumn{1}{|c}{\bm{*}} & \ast & \hdots & \ast \\ \cline{1-1}
0 & \multicolumn{1}{|c}{*} & \hdots & \ast \\ \cline{2-2}
0 & 0 & \ddots & \ast \\
0 & 0 & \hdots & \multicolumn{1}{|c}{*} \\ \cline{4-4}
\end{array}
\right].
\]
By Proposition \ref{prop_vertices}, the control matrix $\Ctrl(M)$ is the $N \times 1$ matrix with every entry equal to 0. Thus $\Pi_+(M)$ is just the nonnegative orthant. The corner polyhedron $\Pi_+(M)$ has facets $\{s_i \geq 0\}$, for each $i \in [N]$. As a special case of Proposition \ref{prop_facets}, all the facets are of type (F-1A).
When $N = 2$, we can draw $\Pi_+(M)$ in the plane:
\begin{center}
\begin{tikzpicture}[line cap=round,line join=round,x=0.4cm,y=0.4cm]
\clip(-2,-2) rectangle (2.5,2.5);
\draw [line width=1pt] (-2,0)-- (3,0);
\draw [line width=1pt] (0,-2)-- (0,3);
\fill[line width=2pt,fill=black,fill opacity=0.1] (0,0) -- (3,0) -- (3,3) -- (0,3) -- cycle;
\draw [line width=2pt] (0,0)-- (3,0);
\draw [line width=2pt] (3,0)-- (3,3);
\draw [line width=2pt] (3,3)-- (0,3);
\draw [line width=2pt] (0,3)-- (0,0);
\begin{scriptsize}
\draw [fill=black] (0,0) circle (2.5pt);
\end{scriptsize}
\end{tikzpicture}
\end{center}
\end{example}
Now we get to some more interesting examples.
\begin{example} \label{ex_lower_stair}
Suppose that $M$ has entry $(i + 1, i)$ nonzero for each $1 \leq i < N + 1$, and the entries $(i, j)$ with $j > i + 1$ are all 0. Then the profile of $M$ is the lower Dyck path shown:
\[
\small
\left[
\begin{array}{ccccccc}
\multicolumn{1}{|c}{*} & \ast & \hdots & \ast & \ast & \ast \\
\multicolumn{1}{|c}{\bm{*}} & \ast & \hdots & \ast & \ast & \ast\\ \cline{1-1}
0 & \multicolumn{1}{|c}{\bm{*}} & \hdots & \ast & \ast & \ast \\ \cline{2-2}
\vdots & \vdots & \ddots & \vdots & \vdots & \vdots \\
0 & 0 & \hdots & \multicolumn{1}{|c}{\bm{*}} & \ast & \ast \\ \cline{4-4}
0 & 0 & \hdots & 0 & \multicolumn{1}{|c}{\bm{*}} & \ast \\ \cline{5-6}
\end{array}
\right]
\]
By Proposition \ref{prop_vertices}, we have $\Ctrl(M) = -1 \in \Mat_N$. Thus there are $N$ vertices of $\Pi_+(M)$. By Proposition \ref{prop_facets}, the facets of $\Pi_+(M)$ are, for each nonempty subset $I \subseteq [N]$, given by
$$\left\{ \sum_{i \in I} s_i \geq -1 \right\}.$$ Therefore there are $2^N - 1$ facets, all of type (F-1B).
When $N = 2$, we can draw $\Pi_+(M)$ in the plane:
\begin{center}
\begin{tikzpicture}[line cap=round,line join=round,x=0.4cm,y=0.4cm]
\clip(-2,-2) rectangle (2.5,2.5);
\draw [line width=1pt] (-2,0)-- (3,0);
\draw [line width=1pt] (0,-2)-- (0,3);
\fill[line width=2pt,fill=black,fill opacity=0.1] (-1,0) -- (0,-1) -- (3,-1) -- (3,3) -- (-1,3) -- cycle;
\draw [line width=2pt] (-1,0)-- (0,-1);
\draw [line width=2pt] (0,-1)-- (3,-1);
\draw [line width=2pt] (3,-1)-- (3,3);
\draw [line width=2pt] (3,3)-- (-1,3);
\draw [line width=2pt] (-1,3)-- (-1,0);
\begin{scriptsize}
\draw [fill=black] (-1,0) circle (2.5pt);
\draw [fill=black] (0,-1) circle (2.5pt);
\end{scriptsize}
\end{tikzpicture}
\end{center}
\end{example}
\begin{example} \label{ex_upper_stair}
Suppose that $M$ has entry $(i, i + 1)$ nonzero for each $1 \leq i < N + 1$, and all entries on or below the main diagonal are equal to $0$. Then $\Prof(M)$ is the upper Dyck path shown:
\[
\small
\left[
\begin{array}{cccccc}
\cline{1-1}
0 & \multicolumn{1}{|c}{\bm{*}} & \ast & \hdots & \ast & \ast \\ \cline{2-2}
0 & 0 & \multicolumn{1}{|c}{\bm{*}} & \hdots & \ast & \ast \\ \cline{3-3}
\vdots & \vdots & \ddots & \ddots & \vdots & \vdots \\
0 & 0 & 0 & \ddots & \multicolumn{1}{|c}{\bm{*}} & \ast \\ \cline{5-5}
0 & 0 & 0 & \hdots & 0 & \multicolumn{1}{|c}{\bm{*}} \\ \cline{6-6}
0 & 0 & 0 & \hdots & 0 & \multicolumn{1}{c|}{0}
\end{array}
\right]
\]
By Proposition \ref{prop_vertices}, we have $\Ctrl(M) = 1 \in \Mat_N$. Thus there are $N$ vertices of $\Pi_+(M)$. By Proposition \ref{prop_facets}, there are $N + 1$ facets: a Type (F-2A) facet $\{s_i \geq 0\}$ for each $i \in [N]$, and a Type (F-2B) facet $\{\sum_{i \in I} s_i \geq 1\}$ corresponding to $I = [N]$. Comparing with Example \ref{ex_lower_stair}, we see that the seemingly dual cases $\Ctrl(M) = 1$ and $\Ctrl(M) = -1$ produce quite different corner polyhedra.
When $N = 2$, we can draw $\Pi_+(M)$ in the plane:
\begin{center}
\begin{tikzpicture}[line cap=round,line join=round,x=0.4cm,y=0.4cm]
\clip(-2,-2) rectangle (2.5,2.5);
\draw [line width=1pt] (-2,0)-- (3,0);
\draw [line width=1pt] (0,-2)-- (0,3);
\fill[line width=2pt,fill=black,fill opacity=0.1] (0,1) -- (1,0) -- (3,0) -- (3,3) -- (0,3) -- cycle;
\draw [line width=2pt] (0,1)-- (1,0);
\draw [line width=2pt] (1,0)-- (3,0);
\draw [line width=2pt] (3,0)-- (3,3);
\draw [line width=2pt] (3,3)-- (0,3);
\draw [line width=2pt] (0,3)-- (0,1);
\begin{scriptsize}
\draw [fill=black] (0,1) circle (2.5pt);
\draw [fill=black] (1,0) circle (2.5pt);
\end{scriptsize}
\end{tikzpicture}
\end{center}
\end{example}
\begin{example}
Let $N = 7$. We present a more elaborate example of matrix with a lower Dyck path as its profile:
\[
M =
\tiny
\left[
\begin{array}{cccccccc}
\multicolumn{1}{|c}{*} & \ast & \ast & \ast & \ast & \ast & \ast & \ast \\
\multicolumn{1}{|c}{*} & \ast & \ast & \ast & \ast & \ast & \ast & \ast \\
\multicolumn{1}{|c}{\bm{*}} & \ast & \ast & \ast & \ast & \ast & \ast & \ast \\ \cline{1-1}
0 & \multicolumn{1}{|c}{*} & \ast & \ast & \ast & \ast & \ast & \ast \\
0 & \multicolumn{1}{|c}{\bm{*}} & \ast & \ast & \ast & \ast & \ast & \ast \\ \cline{2-5}
0 & 0 & 0 & 0 & 0 & \multicolumn{1}{|c}{0} & \ast & \ast \\ \cline{6-6}
0 & 0 & 0 & 0 & 0 & 0 & \multicolumn{1}{|c}{*} & \ast \\
0 & 0 & 0 & 0 & 0 & 0 & \multicolumn{1}{|c}{\bm{*}} & \ast \\ \cline{7-8}
\end{array}
\right].
\]
We emphasize that entry $(6,6)$ is 0, but that the profile is a lower Dyck path by definition, thus must go below entry $(6,6)$. By Proposition \ref{prop_vertices},
\[
\Ctrl(M) =
\tiny
\begin{bmatrix}
-1 & 0 & 0 \\
-1 & -1 & 0 \\
0 & -1 & 0 \\
0 & -1 & 0 \\
0 & 0 & 0 \\
0 & 0 & 0 \\
0 & 0 & -1 \\
\end{bmatrix}.
\]
By Proposition \ref{prop_facets}, there are Type (F-1A) facets
$$\{s_5 \geq 0\}, \quad \{s_6 \geq 0\},$$
and there are Type (F-1B) facets $\{ s_I \geq -1 \}$ for each
\begin{align*}
I \in \{
& \{ 1 \}, \{ 2 \}, \{ 3 \}, \{ 4 \}, \{ 7 \}, \\
& \{1, 3\}, \{1, 4\}, \{1, 7\}, \\
& \{1, 3, 7\}, \{1, 4, 7\} \}.
\end{align*}
\end{example}
\begin{example}
Consider
\[
M =
\tiny
\left[
\begin{array}{ccccc}
\cline{1-2}
0 & 0 & \multicolumn{1}{|c}{\bm{*}} & \ast & \ast \\ \cline{3-3}
0 & 0 & 0 & \multicolumn{1}{|c}{\bm{*}} & \ast \\ \cline{4-5}
0 & 0 & 0 & 0 & \multicolumn{1}{c|}{0} \\
0 & 0 & 0 & 0 & \multicolumn{1}{c|}{0} \\
0 & 0 & 0 & 0 & \multicolumn{1}{c|}{0} \\
\end{array}
\right].
\]
By Proposition \ref{prop_vertices},
\[
\tiny
\Ctrl(M) =
\begin{bmatrix}
1 & 0 \\
1 & 1 \\
0 & 1 \\
0 & 0 \\
\end{bmatrix}.
\]
By Proposition \ref{prop_facets}, the Type (F-2A) facets are $\{s_i \geq 0\}$ for each $i \in \{1, 3, 4\}$, and the Type (F-2B) facets are $\{ s_I \geq 1 \}$ for each
$I \in \{ \{2\}, \{1, 3\} \}.$
\end{example}
\section{Linear Maps with Marked Points} \label{sect_proof}
In this section, we describe GIT stability for marked linear maps relative to any sheaf. We get Theorem \ref{thm_main} and the other results of the introduction as corollaries.
To describe GIT stability for marked linear maps, we plug the facet computation of Proposition \ref{prop_facets} into the Hilbert-Mumford criterion for corner polyhedra, Proposition \ref{prop_corner_numerical}. Then we show that these stability tests can be checked on just Type I, II, and III flags, by studying the ways to embed these flags in complete flags. The underlying idea is that, any time a facet of a corner polyhedron imposes a stability test, we can find the same condition imposed by some Type I, II, or III flag.
This technique should be compared to Mumford's stability test for point configurations, Theorem \ref{thm_classic_git}. In that proof, it was unnecessary to test stability with every complete flag, since the facet conditions were very simple and could be checked on individual subspaces. Our method here is similar, but we have to consider all four kinds of facets described in Section \ref{sec_facets}. The main difficulty is in setting up the necessary terminology to keep track of the various flag and facet types.
We recall our setting. Let $N, n \in \mathbb N$. We define $\Mat_{N+1} \cong \mathbb{A}^{N^2 + 2N + 1}$ to be the space of linear endomorphisms $k^{N+1} \to k^{N+1}$. We write $\PP(\Mat_{N+1})$ for the projectivization of the nonzero elements of $\Mat_{N+1}$. Let $\mathcal{X}_{N,n} = (\mathbb{P}^N)^n \times \PP(\Mat_{N+1})$. We write its elements as
$$(T,v) = (T, v_1, \hdots, v_n).$$
Note that $\PP(\Mat_{N+1})$ is not a group, but it nevertheless admits the conjugation action of $\SL_{N+1}$.
We study the action \eqref{eq_acn} of $\SL_{N+1}$ on $\mathcal{X}_{N,n}$, which we restate for convenience:
$$A \cdot (T,v) = (ATA^{-1}, Av_1, \hdots, Av_n).$$
A \emph{flag} $\mathcal{H}$ in $k^{N+1}$ is a sequence, of some length $\gamma \in [N]$, of linear subspaces
$$0 \subsetneq H_1 \subsetneq \hdots \subsetneq H_\gamma \subset k^{N+1}.$$
A flag is \emph{complete} if $\gamma = N + 1$ and $H_\gamma = k^{N+1}$. Equivalently, a flag is complete if, for all $i \in [\gamma]$, we have $\dim V_i = i$.
A \emph{completion} of a flag $\mathcal{H}$ is a complete flag $\mathcal{H}'$ such that every subspace appearing in $\mathcal{H}$ also appears in $\mathcal{H}'$. A \emph{completed flag pair} $(\mathcal{H}, \mathcal{H}')$ is a pair of flags such that $\mathcal{H}'$ is a completion of $\mathcal{H}$.
Each basis $b_1, \hdots, b_{N+1}$ of $k^{N+1}$ has an \emph{associated flag}: for each $i = 1, \hdots, N + 1$, let
$$H_i = \Span(b_1, \hdots, b_i).$$
If $\mathcal{H}$ is a complete flag, a \emph{flag basis} is any basis which has $\mathcal{H}$ as its associated flag. Every complete flag has a flag basis.
Fix $T \in \PP(\Mat_{N+1})$. We define some special types (P-1A), (P-1B), (P-2A), (P-2B) of completed flag pair, with respect to $T$. Note that some completed flag pairs do not belong to any of these types.
\begin{itemize}
\item A completed flag pair
$(\mathcal{H}, \mathcal{H}')$ is \emph{Type} (P-1A) if:
\begin{itemize}
\item $\mathcal{H}$ consists of just one space $H$, and
\item In any associated basis of $\mathcal{H}'$, the matrix $M$ of $T$ is not strictly-upper-triangular, and
\item the matrix $M$ has property \eqref{eq_1A}, where $i_1 \in [N]$ is the value such that $H'_{i_1} = H$.
\end{itemize}
\item A completed flag pair $(\mathcal{H}, \mathcal{H}')$ is of \emph{Type} (P-1B) if:
\begin{itemize}
\item In any associated basis of $\mathcal{H}'$, the matrix $M$ of $T$ is not strictly-upper-triangular, and
\item The matrix $M$ has property \eqref{eq_1B}, where $I$ is the set of indices determined by the inclusion of $\mathcal{H}$ in $\mathcal{H}'$.
\end{itemize}
\item A completed flag pair $(\mathcal{H},\mathcal{H'})$ is \emph{Type} (P-2A) if:
\begin{itemize}
\item $\mathcal{H}$ consists of just one space $H$, and
\item The matrix $M$ of $T$ in any associated basis of $\mathcal{H}'$ is strictly upper triangular, and
\item The matrix $M$ has property \eqref{eq_2A}, where $i_1 \in [N]$ is the value such that $H'_{i_1} = H$.
\end{itemize}
\item A completed flag pair $(\mathcal{H}, \mathcal{H}')$ is of \emph{Type} (P-2B) if:
\begin{itemize}
\item In any associated basis of $\mathcal{H}'$, the matrix $M$ of $T$ is not strictly-upper-triangular, and
\item The matrix $M$ has property \eqref{eq_2B}, where $I$ is the set of indices determined by the inclusion of $\mathcal{H}$ in $\mathcal{H}'$.
\end{itemize}
\end{itemize}
We recall the three types I, II, III of a flag $\mathcal{H}$ relative to $T$ that were defined in Theorem \ref{thm_main}. Given a flag $\mathcal{H}$ and $T \in \PP(\Mat_{N+1})$, the \emph{Hessenberg function of $\mathcal{H}$ relative to $T$} is defined by
\begin{align*}
h_{T,\mathcal{H}} : \; & \{0, \hdots, \gamma + 1\} \to \{0, \hdots \gamma + 1\}, \\
& i \mapsto \min \{j : TH_i \subseteq H_j\}.
\end{align*}
Given a flag $\mathcal{H}$ and a point $T\in\PP(\Mat_{N+1})$, we say that~$\mathcal{H}$ is of \emph{Type~I,~II, or~III relative to~$T$} if it has the following properties, if any:
\begin{itemize}
\item \textup{Type I:} $\gamma = 1$, and $\mathcal{H} = (H_1)$ satisfies either
\[
\Bigl( 0 \neq TH_1 \subseteq H_1 \Bigr)
\qquad\text{or}\qquad
\Bigl(TH_1 \subseteq H_1 \quad \text{and} \quad T(k^{N+1}) \not\subseteq H_1
\Bigr).
\]
\item \textup{Type II:} The Hessenberg function $h_{T, \mathcal{H}}$ satisfies, for all $t$ in the range $1 \leq t \leq \gamma,$
$$h_{T, \mathcal{H}}(t) = t + 1.$$
\item \textup{Type III:} The Hessenberg function $h_{T, \mathcal{H}}$ satisfies, for all $t$ in the range $1 \leq t \leq \gamma + 1$,
$$h_{T, \mathcal{H}}(t) = t - 1.$$
\end{itemize}
\begin{example}
We give some examples of Type I, II, and III flags.
\begin{itemize}
\item For any $T \in \PP(\Mat_{N+1})$, every flag consisting of just one subspace $H$ is of Type I, II, or III.
\item Any non-nilpotent $T \in \PP(\Mat_{N+1})$ has flags of Type I, taking $\mathcal{H} = H_1$ to be the eigenspace of any nonzero eigenvalue.
\item If $T = 1$, then every flag is Type I, and we recover the condition \eqref{eq_mumford_template} of Mumford's example.
\item Suppose $T \neq 1$ and $T$ is invertible. Then there is some vector $v$ such that $v \neq T(v) \neq 0$. Let $H_1 = \Span(v)$ and $H_2 = \Span(v, T(v))$. The flag
$H_1 \subsetneq H_2$ is Type II.
\item Suppose $T$ has a cyclic vector; that is, there exists $v \in k^{N+1}$ such that $v, T(v), \hdots, T^{N}(v)$ form a basis. For each $i = 1, \hdots, N$, let
$$H_i = \Span(v, T(v), \hdots, T^{i-1}(v)).$$
The flag $H_1 \subsetneq \hdots \subsetneq H_{N}$ is Type II.
\item Suppose the Jordan form of $T$ has all 0's on the diagonal and all 1's on the first off-diagonal. Then there exists $v \in k^{N+1}$, such that for all $0 \leq i < N$, we have $T^i(v) \neq 0$, and $T^N(v) = 0$. Let
$$H_i = \Span(v, T(v), \hdots, T^{i-1}(v)).$$
The flag $H_1 \subsetneq \hdots \subsetneq H_{N}$ is Type III.
\item Type III flags exist only if $T$ is nilpotent. But a nilpotent map $T$ may still have flags of Type I and II.
\end{itemize}
\end{example}
\begin{remark}
One can also define Type I, II, and III flags without reference to Hessenberg functions, as follows.
\begin{itemize}
\item \textup{Type I:} $\gamma = 1$, and $\mathcal{H} = (H_1)$ satisfies either
\[
\Bigl( 0 \neq TH_1 \subseteq H_1 \Bigr)
\qquad\text{or}\qquad
\Bigl(TH_1 \subseteq H_1 \quad \text{and} \quad T(k^{N+1}) \not\subseteq H_1
\Bigr).
\]
\item \textup{Type II:} For all $t$ in the range $1 \leq t \leq \gamma$, we have
$$TH_t \not\subseteq H_t \quad \text{and} \quad TH_t \subseteq H_{t+1}.$$
\item \textup{Type III:} With the convention that $H_{-1} = \emptyset$, for all $t$ in the range $0 \leq t \leq \gamma$, we have
$$TH_{t+1} \subseteq H_t \quad \text{and} \quad TH_{t+1} \not\subseteq H_{t-1} .$$
\end{itemize}
\end{remark}
For any point $v \in \mathbb{P}^N$ and flag $\mathcal{H}$, we define
$$\Omega(v, \mathcal{H}) := \sum_{j \in [\gamma]} \left(
\sum_{i=1}^n \frac{-j m_i}{N+1} + \sum_{i \in [n]: \; X_{j+1}(v_i) = \hdots = X_{N+1}(v_i) = 0} m_i
\right).$$
It arises in our calculations as follows. Recall the notation $\eta(v)$ from \eqref{eq_config_vector}. If $\mathcal{H'}$ is any completion of $\mathcal{H}$, and the spaces in $\mathcal{H}$ have indices corresponding to $I \subseteq [N]$ in $\mathcal{H}'$, we have
\begin{equation} \label{eq_omega_functional}
\Omega(v, \mathcal{H}) = s_I( \eta(v) ).
\end{equation}
\begin{remark}
The function $\Omega(v, \mathcal{H})$ measures how much the tuple $v$ intersects the flag $\mathcal{H}$. The more intersections there are between points of $v$ and subspaces in $\mathcal{H}$, the larger the value of $\Omega(v, \mathcal{H})$. Notice that the weighted count $\Omega(v, \mathcal{H})$ is unrelated to the space $\PP(\Mat_{N+1})$; it is solely a function of the $n$-tuple $v$ and the flag $\mathcal{H}$. To give an idea of how the quantity $\Omega(v, \mathcal{H})$ behaves, observe that the following are equivalent:
\begin{enumerate}
\item The tuple $v \in (\mathbb{P}^N)^n$ is semistable relative to $\mathcal{O}(\bm{m}).$
\item For every linear subspace $H$, we have $\Omega(v, H) \leq 0.$
\item For every flag $\mathcal{H}$, we have $\Omega(v, \mathcal{H}) \leq 0.$
\end{enumerate}
The equivalence of (1) and (2) is just the result of Theorem \ref{thm_classic_git}. The equivalence of (2) and (3) follows from the fact that, for any flag $\mathcal{H}$, the sum $\Omega(v, \mathcal{H})$ is computed over the individual subspaces making up $\mathcal{H}$.
\end{remark}
Finally we prove our main result, which describes GIT stability of marked linear maps for any sheaf. We prove Theorem \ref{thm_main} as a special case.
\begin{thm} \label{thm_main_any_sheaf}
Consider $\mathcal{X}_{N,n}$ with the action of $\SL_{N+1}$ and sheaf $\mathcal{L} = \mathcal{O}(q, \bm{m})$. For each flag $\mathcal{H}$ in $k^{N+1}$ of Type I, II, or III, define
\[
c(T, \mathcal{H}) =
\begin{cases}
0, & \mathcal{H} \text{ is Type I}, \\
1, & \mathcal{H} \text{ is Type II}, \\
-1, & \mathcal{H} \text{ is Type III}. \\
\end{cases}
\]
Then $(T, v) \in \mathcal{X}_{N,n}$ is GIT semistable relative to $\mathcal{L}$ if and only if, for every Type I, II, or III flag $\mathcal{H}$, we have
\begin{equation} \label{eq_flag_solo}
\Omega(v, \mathcal{H}) \leq q c(T, \mathcal{H}).
\end{equation}
A point $(T, v) \in \mathcal{X}_{N,n}$ is GIT stable relative to $\mathcal{L}$ if and only if, for every Type I, II, or III flag $\mathcal{H}$, the inequality \eqref{eq_flag_solo} is strict.
\end{thm}
\begin{proof}
Given a basis $B$ of $k^{N+1}$, let $\Pi_+(T,v)$ denote the corner polyhedron of $(T,v)$. By Proposition \ref{prop_corner_numerical}, the point $(T,v)$ is $B$-semistable if and only if, in the $B$-aligned basis $B'$ of $\mathcal{L}$ obtained by Segre and Veronese maps, we have
$$0 \in \Pi_+(T,v).$$
Let $M$ be the matrix of $T$ in basis $B$, and let $\Pi_+(M)$ be the associated corner polyhedron of Type $A_N$, which is by definition the corner polyhedron of $M$ relative to $\mathcal{O}(1)$. Let $\Pi_+(v)$ be the corner polyhedron of $v$ relative to $\mathcal{O}(\bm{m})$. By Lemma \ref{lem_segre} and \ref{lem_veronese},
$$ \Pi_+(T,v) = \Pi_+(v) + q \Pi_+(M).$$
By Proposition \ref{prop_tuple}, the corner polyhedron $\Pi_+(v)$ has just one vertex $\eta(v)$. Thus $B$-semistability is equivalent to
$$0 \in q \Pi_+(M) + \eta(v).$$
Equivalently,
$$-\eta(v) \in q \Pi_+(M).$$
By Proposition \ref{prop_facets}, the polyhedron $\Pi_+(M)$ is an intersection of halfspaces of the form $\{s_I \geq c\}$, for various nonempty sets $I \subseteq [N]$ and values $c \in \{0, 1, -1\}.$
Suppose that $(T,v)$ is semistable relative to $\mathcal{L}$. Then, by the preceding discussion, $(T,v)$ satisfies some list of conditions of the form
$$ \Omega(v, \mathcal{H}) \leq c_0(T, \mathcal{H})q, $$
where $\mathcal{H}$ is a flag and $c_0(T, \mathcal{H}) \in \{0, 1, -1\}$.
We show that the conditions described in the theorem statement appear in this way.
Let $\mathcal{H}$ be Type I. Let the single element of $\mathcal{H}$ be $H$. We claim there exists a flag completion $\mathcal{H}'$ of $\mathcal{H}$ such that the completed flag pair $(\mathcal{H}, \mathcal{H}')$ is of Type (P-1A). Indeed, let $\mathcal{H}''$ be any flag completion of $\mathcal{H}$, with some associated basis $B''$. Since $0 \neq TH_1$ or $T(k^{N+1}) \subsetneq H$, there is a permutation $B'$ of the basis $B''$ that fixes $H$, such that the matrix of $T$ is not strictly-upper-triangular. Let $\mathcal{H}'$ be the flag associated to $B'$. Since $TH_1 \subseteq H_1$, the pair $(\mathcal{H}, \mathcal{H}')$ is Type (P-1A). Let $\rho \in [N]$ be the value such that $H = \mathcal{H}'_\rho$. By Proposition \ref{prop_facets}, in the basis $B'$, there is a facet of $\Pi_+(M)$ with supporting halfspace $\{s_\rho \geq 0 \}$. Thus $s_\rho(z)(-\eta) \geq 0$, so by \eqref{eq_omega_functional}, we have $\Omega(v, H) \leq 0$.
Let $\mathcal{H}$ be Type II. There exists a flag completion $\mathcal{H}'$ of $\mathcal{H}$ such that the completed flag pair $(\mathcal{H}, \mathcal{H}')$ is of Type (P-1B). Let $I \subseteq [N]$ be the subset of indices corresponding to the inclusion of $\mathcal{H}$ in $\mathcal{H}'$. By Proposition \ref{prop_facets}, writing $M$ for the matrix of $T$ in basis $B$, there is a facet of $\Pi_+(M)$ with supporting halfspace $\{s_I \geq -1 \}$. Thus $s_I(z)(-\eta) \geq -q$, so by \eqref{eq_omega_functional}, we have $\Omega(v, \mathcal{H}) \leq q$.
Let $\mathcal{H}$ be Type III. There exists a flag completion $\mathcal{H}'$ of $\mathcal{H}$ such that the completed flag pair $(\mathcal{H}, \mathcal{H}')$ is of Type (P-2B); namely, let $\mathcal{H}'$ be any completion of
$$T^{-1}(0) \subsetneq T^{-1}(H_1) \hdots \subsetneq T^{-1}(H_{\gamma - 1}).$$ Let $I \subseteq [N]$ be the subset of indices corresponding to the inclusion of $\mathcal{H}$ in $\mathcal{H}'$. By Proposition \ref{prop_facets}, writing $M$ for the matrix of $T$ in basis $B$, there is a facet of $\Pi_+(M)$ with supporting halfspace $\{ s_I \geq 1 \}$. Thus $\Omega(v, \mathcal{H}) \leq -q.$
We have proved one direction of the theorem; now we prove the other.
For every complete flag $\mathcal{H}'$, and every subflag $\mathcal{H}$ of $\mathcal{H}'$ such that the completed flag pair $(\mathcal{H}, \mathcal{H}')$ is Type (P-1A), (P-1B), (P-2A), or (P-2B), we define a quantity $c(T, \mathcal{H}, \mathcal{H}')$ by
\[
c(T, \mathcal{H}, \mathcal{H'}) =
\begin{cases}
0, & (\mathcal{H}, \mathcal{H'}) \text{ is Type (P-1A) for } T, \\
-1, & (\mathcal{H}, \mathcal{H'}) \text{ is Type (P-1B) for } T, \\
0, & (\mathcal{H}, \mathcal{H'}) \text{ is Type (P-2A) for } T, \\
1, & (\mathcal{H}, \mathcal{H'}) \text{ is Type (P-2B) for } T. \\
\end{cases}
\]
We claim that, if \eqref{eq_flag_solo} holds
for all flags $\mathcal{H}$ of Type I, II, and III, then for all completed flag pairs $(\mathcal{H}, \mathcal{H}')$ of Type (P-1A), (P-1B), (P-2A), or (P-2B),
\begin{equation} \label{eq_flag_pairs}
\Omega(v, \mathcal{H}) \leq c(T, \mathcal{H}, \mathcal{H}') q.
\end{equation}
Indeed, assuming \eqref{eq_flag_pairs}, by Proposition \ref{prop_facets}, for every basis $B$ of $k^{N+1}$, writing $M$ for the matrix of $T$, we have $-\eta(v) \in q\Pi_+(M)$ by \eqref{eq_omega_functional}. Hence $0 \in \Pi_+(T,v)$ for every basis $B$, proving semistability.
Finally, we check that \eqref{eq_flag_solo} implies \eqref{eq_flag_pairs}:
\begin{itemize}
\item If $(\mathcal{H}, \mathcal{H}')$ is Type (P-1A) or (P-2A), then $\mathcal{H}$ is Type I and $c(T, \mathcal{H}, \mathcal{H}') = c(T, \mathcal{H})$.
\item
If $(\mathcal{H}, \mathcal{H}')$ is Type (P-1B), then $\mathcal{H}$ is Type II and $c(T, \mathcal{H}, \mathcal{H'}) = c(T, \mathcal{H})$.
\item
If $(\mathcal{H}, \mathcal{H})'$ is Type (P-2B), then $\mathcal{H}$ is Type III and $c(T, \mathcal{H}, \mathcal{H}') = c(T, \mathcal{H})$.
\end{itemize}
For stability, the same argument works using strict inequalities.
\end{proof}
Our main theorem follows:
\begin{proof}[Proof of Theorem \ref{thm_main}]
In Theorem \ref{thm_main_any_sheaf}, take $m_1 = \hdots = m_n = 1$.
\end{proof}
Finally, we prove two useful facts about the structure of the quotient: it is nonempty (Corollary \ref{cor_stable_nonempty}) and rational (Theorem \ref{thm_rational}).
\begin{proof}[Proof of Corollary \ref{cor_stable_nonempty}]
Let $T \neq 1$ be invertible with distinct eigenvalues. Then there exists a point $v$ that is not in any $T$-invariant proper linear subspace. Let $v_1 = \hdots = v_n = v$. Let us check that the stability conditions in Theorem \ref{thm_main} hold. Since $T$ is invertible, it is not nilpotent, so there are no Type III flags relative to $T$. The conditions for the Type II flags are met since $q \geq n$. The Type I flags are simply the invariant linear subspaces for $T$, but these do not contain $v$, so the Type I conditions are all met.
\end{proof}
\begin{proof}[Proof of Theorem \ref{thm_rational}]
The existence and dimension count follow immediately from Corollary \ref{cor_stable_nonempty}. The claim is independent of $\mathcal{L}$ because varying the sheaf $\mathcal{L}$ does not change the birational type of the quotient. Let $\mathcal{S}_{N+1}$ denote the symmetric group on $N + 1$ letters. Given a generic point $(T, v_1, \hdots, v_n) \in \mathcal{X}_{N,n}$, consider the $N + 1$ fixed points of $T$ as an element of $(\mathbb{P}^N)^n / \mathcal{S}_{N+1}$. There exists a projective transformation that diagonalizes $T$ to some matrix $\diag[\lambda_1, \hdots, \lambda_{N+1}]$ of eigenvalues and takes $v_1$ to $[1: \hdots : 1]$. For each $i$ in the range $1 < i \leq n$, let $v'_i$ be the image of $v_i$. Since there is no ordering on the eigenvectors, a projective transformation with this property is unique up to the action of $\mathcal{S}_{N+1}$. Thus the multivalued map
$$(T, v_1, \hdots, v_n) \dashrightarrow ([\lambda_1 : \hdots : \lambda_{N+1}], v'_2, \hdots, v'_n)$$
descends to a well-defined rational map
$$\mathcal{X}_{N,n} \GITQ \SL_{N+1} \dashrightarrow (\mathbb{P}^N)^{n} / \mathcal{S}_{N+1},$$
where $\mathcal{S}_{N+1}$ acts diagonally on each factor by permuting coordinates. Then, to prove rationality, it suffices to show that each factor $\mathbb{P}^N/ \mathcal{S}_{N+1}$ is rational. The theorem on symmetric functions states that $\mathbb{A}^{N+1} / \mathcal{S}_{N+1} \cong \mathbb{A}^{N+1}$, so $\mathbb{P}^N / \mathcal{S}_{N+1}$ is a weighted projective space, and these are rational.
\end{proof}
\section{Application: one marked point} \label{sect_one_marked_point}
In this section, we work out a special case of Theorem \ref{thm_main}, described in Example \ref{ex_one_marked_point}. We show that, when there is one marked point, our results specialize to a moduli space of Mumford-Suominen \cite{MR0437531}.
Let the dimension $N \geq 1$, let the number of marked points be $n = 1$, and let $q = 1$. With these choices, Theorem \ref{thm_main} describes stability for linear self-maps of $\mathbb{P}^N$ marked with a single point, embedded in projective space via the Segre embedding $\mathbb{P}^{N^2 + 2N} \times \mathbb{P}^N \hookrightarrow \mathbb{P}^{(N+1)^3 - 1}$. We show that, with these parameters, the stable quotient is isomorphic to $\mathbb{P}^N$ (Proposition \ref{prop_quotient_is_PN}). Note that any choice of $q \geq 1$ gives the same result; we choose $q = 1$ for concreteness.
First, we collect the most basic consequences of Theorem \ref{thm_main}. Theorem \ref{thm_main} describes (semi)stability by checking the inequality \eqref{eq_main} for each flag of Type I, II, or III. With these parameters, the right side of \eqref{eq_main} is not an integer, and the left side is an integer, so the inequality \eqref{eq_main} is always strict. Thus
$$\mathcal{X}_{N,1}(1,1)\suptxt{s} = \mathcal{X}_{N,1}(1,1)\suptxt{ss}.$$
Further, according to Corollary \ref{cor_stable_nonempty} and Theorem \ref{thm_rational}, the stable locus is nonempty and the (semi)stable quotient is a rational projective variety of dimension $N$.
To say more about the structure of the quotient, we need the more detailed description of stability from Theorem \ref{thm_main}. With these parameters, Theorem \ref{thm_main} condenses to a simple linear-algebraic condition.
\begin{lemma} \label{lem_one_marked_point}
A marked linear map $(T,v) \in \mathcal{X}_{N,1}$ is stable relative to $\mathcal{O}(1,1)$ if and only if $v, T(v), \hdots, T^N(v)$ form a basis of $k^{N+1}$ and $T^{N+1}(v) \neq 0.$
\end{lemma}
\begin{proof}
First, observe that if $T$ has any Type III flags, then inspecting \eqref{eq_main}, the pair $(T,v)$ is unstable. Every nilpotent map admits a Type III flag, so we conclude that if $T$ is nilpotent, then $(T,v)$ is unstable. Second, if $v$ lies on a Type I flag of $T$, then $(T,v)$ is unstable. The Type II flags impose no further conditions because, for any Type II flag, the left side of \eqref{eq_main} is less than the right side of \eqref{eq_main}. So by Theorem \ref{thm_main}, we have that $(T,v)$ is stable if and only if $T$ is non-nilpotent and $v$ lies on no Type I flag for $T$.
Now we claim that, if $(T,v)$ is stable, then the points $v, T(v), T^2(v), \hdots, T^{N-1}(v), T^N(v)$ are a basis of $k^{N+1}$. Indeed, if not, then there is a $T$-invariant proper subspace containing $v$, thus a Type I flag containing $v$, given by
$$\Span \bigl( T^m (v): 0 \leq m \leq N - 1 \bigr).$$
Conversely, suppose that a pair $(T,v)$ satisfies the condition that $v, \hdots, T^N(v)$ form a basis of $k^{N+1}$. Then the only $T$-invariant subspace containing $v$ is the whole space $k^{N+1}$, and thus $v$ is not contained in any Type I flag. Note further that, for any such $(T,v)$, nilpotence is equivalent to $T^{N+1}(v) = 0.$
\end{proof}
The simple characterization of stability in Lemma \ref{lem_one_marked_point} allows us to explicitly describe the structure of the stable quotient.
\begin{prop} \label{prop_quotient_is_PN}
We have
$$ (\mathcal{M}_{1,1}^N(1,1))\suptxt{s} \cong \mathbb{P}^N.$$
\end{prop}
\begin{proof}
We exhibit an explicit isomorphism $(\mathcal{M}_{1,1}^N)\suptxt{s} \to \mathbb{P}^N$. Let $(T,v)$ be a stable marked linear map relative to $\mathcal{O}(1,1).$ By Lemma \ref{lem_one_marked_point}, there is a unique choice of coordinates on $\mathbb{P}^N$ such that $v = [1: 0 : \hdots : 0]$ and, for some point
$$[\alpha_1 : \alpha_2 : \hdots : \alpha_{N+1}] \in \mathbb{P}^N,$$
the projective matrix of $T$ is of the form
\[
\begin{bmatrix}
0 & & & & \alpha_1 \\
1 & 0 & & & \alpha_2 \\
& 1 & \ddots & & \vdots \\
& & \ddots & 0 & \alpha_N \\
& & & 1 & \alpha_{N+1} \\
\end{bmatrix}.
\]
We note that the characteristic polynomial of this matrix is the monic polynomial with coefficients $\alpha_1, \hdots, \alpha_{N+1}$.
The map
\begin{align*}
\alpha : (\mathcal{M}_{1,1}^N)\suptxt{s} &\to \mathbb{P}^N,\\
(T,v) &\mapsto [\alpha_1 \colon \hdots \colon \alpha_{N+1}]
\end{align*}
is algebraic and has an inverse, since for any point $[\alpha_1 \colon \hdots \colon \alpha_{N+1}]$, we can write down the matrix $T$ above and set $v = [1 \colon 0 \colon \hdots \colon 0]$. Hence $\alpha$ is an isomorphism.
\end{proof}
\begin{remark}
A vector $v$ such that $v, T(v), \hdots, T^N(v)$ forms a basis of $k^{N+1}$ is called a \emph{cyclic vector} of $T$. Thus $\mathcal{M}_{1,1}^N$ may be described as the moduli space of non-nilpotent linear maps on projective space equipped with a cyclic vector. Mumford and Suominen studied this moduli space, or rather, its affine version, in \cite{MR0437531}. They observe that the stable locus for the moduli problem of unmarked linear maps is empty, but that this deficiency can be addressed by including the data of a cyclic vector. Lemma \ref{lem_one_marked_point} shows that this choice of extra data arises as a natural consequence of the stability theory of marked linear maps.
\end{remark}
\bibliographystyle{plain}
|
\section*{Methods}
\subsection*{Oscillator relaxation and dephasing in the displaced frame}
With photon loss at a rate $\kappa$, the oscillator's density matrix evolves according to the quantum master equation in Lindblad form
\begin{align}
\partial_t \rho &= -i \left[H,\, \rho\right] + \kappa \mathcal{D}\left[a\right]\rho, \\
H &= H_0 + \varepsilon a^\dagger + h.c.,
\end{align}
were $\mathcal{D}[L] = L\rho L^\dagger - \left(1/2\right)\left\{L^\dagger L, \rho\right\}$, $H_0$ is the oscillator's Hamiltonian, and we have included a time-dependent oscillator drive $\varepsilon(t)$. Here, we take $\hbar = 1$.
Evolution of the density matrix in a time-dependent displaced frame $\tilde{\rho} = D^\dagger \left(\alpha\right) \rho D\left(\alpha\right)$ is given by the equivalent master equation
\begin{align}
\partial_t \tilde{\rho} &= -i\left[\tilde{H}, \tilde{\rho} \right] + \kappa \mathcal{D}\left[a + \alpha\right]\tilde{\rho} \\
\tilde{H} &= D^\dagger \left(\alpha\right)H_0 D\left(\alpha\right) + \left( -i \partial_t \alpha + \varepsilon \right)a^\dagger + h.c.
\end{align}
In particular, the displaced frame Lindbladian can be recast as
\begin{equation}
\kappa \mathcal{D}\left[a + \alpha\right]\tilde{\rho} = \kappa \mathcal{D}\left[a\right]\tilde{\rho} - i \left[i\frac{\kappa}{2} \left(\alpha^* a - \alpha a^\dagger\right), \tilde{\rho}\right],
\end{equation}
corresponding to photon loss at a rate $\kappa$, and a Hermitian re-centering force at a rate $\frac{\kappa}{2}|\alpha|$. This deterministic force can be lumped into the effective displaced-frame Hamiltonian, giving
\begin{align}
\partial_t \tilde{\rho} &= -i\left[\tilde{\tilde{H}}, \tilde{\rho} \right] + \kappa \mathcal{D}\left[a\right]\tilde{\rho} \\
\tilde{\tilde{H}} &= D^\dagger \left(\alpha\right)H_0 D\left(\alpha\right) + \left( -i \partial_t \alpha - i \frac{\kappa}{2} \alpha + \varepsilon \right)a^\dagger + h.c.
\end{align}
Given a desired $\alpha(t)$, $\varepsilon(t)$ can be chosen such that the term in parentheses is zero, satisfying the classical Langevin equation for $\alpha(t)$ given in the main text and counteracting the re-centering force. With this choice of drive, the deterministic evolution is accounted for, and relaxation in the displaced frame is not enhanced compared to relaxation at the origin of phase space. The classical drive equation can also be modified to account for all linear terms in $\tilde{H}$, including those caused by nonlinear terms in $H_0$ (see Supplementary Information section 4).
White-noise oscillator dephasing is given by the master equation $\partial_t \rho = 2\kappa_\phi\mathcal{D}\left[a^\dagger a\right]\rho$. Defining the superoperator $\mathcal{S}\left[X,Y\right]\rho = X \rho Y^\dagger - \left\{Y^\dagger X, \rho \right\}$, oscillator dephasing is transformed in the displaced frame to
\begin{equation}
\begin{aligned}
&\partial_t \tilde{\rho} = 2\kappa_\phi \mathcal{D}[(a^\dagger + \alpha^*)(a + \alpha)]\tilde{\rho} \\
&=2\kappa_\phi \left\{ \mathcal{D}[a^\dagger a]\tilde{\rho} + |\alpha|^2 \left(\mathcal{D}[a]\tilde{\rho} + \mathcal{D}[a^\dagger]\tilde{\rho}\right)\right. \\
&+ \alpha^{2}\mathcal{S}\left[a^\dagger, a\right]\tilde{\rho} + \alpha^{*2}\mathcal{S}\left[a, a^\dagger\right]\tilde{\rho} \\
&+ \alpha \left(\mathcal{S}\left[a^\dagger a, a\right]\tilde{\rho} + \mathcal{S}\left[a^\dagger, a^\dagger a\right]\tilde{\rho}\right)\\
&\left.+ \alpha^* \left(\mathcal{S}\left[a^\dagger a, a^\dagger\right]\tilde{\rho} + \mathcal{S}\left[a, a^\dagger a\right]\tilde{\rho}\right)\right\}.
\end{aligned}
\end{equation}
In the displaced frame, the noise is dominated by diffusion-like terms at rate $2\kappa_\phi |\alpha|^2$, and unlike the relaxation case, there is no deterministic part that can be counteracted with a simple displacement. However, this master equation is only valid in the Markovian regime, and typically the spectral density of oscillator frequency fluctuations non-white due to effects such as two-level-system defects \cite{niepceStabilitySuperconductingResonators2021}. In the colored noise case, it is possible that part of the enhanced dephasing noise could be echoed away using symmetric pulse constructions \cite{ballSoftwareToolsQuantum2021}.
\subsection*{Universality of ECD Control}
Universal control of the oscillator is the ability to perform arbitrary unitary transformations which are generated by Hamiltonians polynomial in $q =(1/\sqrt{2})(a^\dagger + a)$ and $p = (i/\sqrt{2})(a^\dagger - a)$ \cite{lloydQuantumComputationContinuous1999b, braunsteinQuantumInformationContinuous2005}. Here, we extend this definition to universal control of the oscillator and qubit, which is the ability to perform arbitrary unitary transformations which are generated by linear combinations of Hamiltonians of the form $q^j p^k \sigma_i$ where $j,k$ are non-negative integers and $\sigma_i \in \left\{I, \sigma_x, \sigma_y, \sigma_z\right\}$.
Given a set of generating Hamiltonians $\left\{A,B\right\}$, the two identities
\begin{align}
e^{-iA \delta t}e^{-iB \delta t}e^{iA \delta t}e^{iB \delta t} &= e^{\left[A,B\right]\delta t^2} + O(\delta t^3), \\
e^{iA \delta t/2}e^{iB \delta t / 2}e^{iB \delta t/2}e^{iA \delta t} &= e^{i(A + B)\delta t} + O(\delta t^3),
\end{align}
can be used to generate the action of the Hamiltonian $-i\left[A,B\right]$ and the Hamiltonian $A + B$ in the limit $\delta t \rightarrow 0$ \cite{braunsteinQuantumInformationContinuous2005}. By repeated application of the identities above, we can generate evolution which is any superposition of nested commutators of the original set of generators \cite{dalessandroIntroductionQuantumControl2007}.
Starting with the set of generators for $\text{ECD}(\beta)$ and $R_\varphi(\theta)$, $\left\{q \sigma_z, p\sigma_z, \sigma_x, \sigma_y\right\}$,
commutators such as $\left[q \sigma_z, \sigma_x \right] \propto q \sigma_y$ and $\left[\sigma_x, \sigma_y \right] \propto \sigma_z$ can be used to expand the set to $\left\{\sigma_i, q \sigma_i, p \sigma_i\right\}$ where $i \in \left\{x,y,z\right\}$. This shows that effectively, by rotating the qubit between conditional displacements, the ECD gate set can create more general Rabi type interactions between the oscillator and qubit, where qubit-mediated nonlinear gates have been proposed \cite{parkQubitmediatedDeterministicNonlinear2017,parkDeterministicNonlinearPhase2018}.
By using commutators similar to $\left[q\sigma_x, q\sigma_y\right] \propto q^2 \sigma_z$, our set can further be expanded to all quadratic polynomials of $q\sigma_i$ and $p\sigma_i$. This process can be iterated in order to generate any $q^j p^k \sigma_i$ product, where $i \in \left\{x,y,z\right\}$. Terms which do not contain a Pauli operator such as $q^j p^k$ can be generated from commutators such as $\left[q^{j+1}p^{k}\sigma_z, p \sigma_z \right] \propto q^j p^k$. With this, the full Lie algebra for polynomial operators on the qubit and oscillator Hilbert space is generated.
\subsection*{Speed limit of control}
The maximum interaction rate between the oscillator and qubit will be limited by the maximum displacement in the oscillator before higher-order nonlinear effects begin to invalidate the dispersive approximation. From \cite{blaisCircuitQuantumElectrodynamics2021} the critical oscillator photon number for the $j^\text{th}$ transmon state is
\begin{align}
n_\text{crit}^j = \frac{1}{2j + 1}\left(\frac{|\Delta - jE_C|^2}{4 g^2} - j\right)
\end{align}
where $\Delta$ is the transmon-oscillator detuning, $g$ is the linear transmon-oscillator coupling rate, and $E_C$ is the charging energy of the transmon. For our experimental parameters, $n_\text{crit}^g \approx 2740$ and $n_\text{crit}^e \approx 910$. These bounds are not strict, however they provide a guiding principle for the maximum photon number before higher order effects become important (see Supplementary Information Section 3.)
With this, the maximum conditional displacement rate is $g_\text{eff}^\text{max} = \alpha_0^\text{max}\chi \approx \sqrt{n_\text{crit}^e}\chi$ using the critical photon number for the first excited state of the transmon. From perturbation theory, the transmon-oscillator dispersive coupling is $\chi \approx \left(2 g^2 E_C\right)/\left(\Delta (\Delta- E_C)\right)$ and the transmon anharmonicity is $K \approx E_C$ \cite{blaisCircuitQuantumElectrodynamics2021}. In the regime $\Delta \gg E_C$ we can approximate $\Delta - E_C \approx \Delta$ and combine the above expressions to find
\begin{align}
g_\text{eff}^\text{max} \approx \sqrt{\frac{\chi K}{6}}
\end{align}
We note other experiments using sideband three-wave-mixing interactions are similarly limited by a bound $\propto \sqrt{\chi K}$ \cite{pechalMicrowaveControlledGenerationShaped2014,touzardGatedConditionalDisplacement2019,rosenblumCNOTGateMultiphoton2018b}. This suggests that at a fixed dispersive shift, increasing transmon anharmonicity could lead to faster interaction rates, giving a path forward for engineering higher-fidelity gates with enhanced effective three-wave interactions.
\section*{Acknowledgments}
We thank Nicholas Frattini, Rodrigo Cortiñas, Christa Flühmann, and Xu Xiao for helpful discussions on oscillator control and driven nonlinear systems. We are grateful to Jacob Curtis and Billy Kalfus for technical assistance. We thank Ioannis Tsioutsios and Luigi Frunzio for device fabrication assistance. We thank Max Hays, Ben Brock, James Teoh, Chris Wang, Aniket Maiti, Philippe Campagne-Ibarcq, Steven Touzard, and Serge Rosenblum for helpful feedback. We thank the Yale Center for Research Computing for technical support and high performance computing resources. This research was sponsored by the Army Research Office (ARO) under grant number W911NF-18-1-0212, W911NF-16-1-0349, W911NF-18-1-0020, and by the Air Force Office of Scientific Research under grant number FA9550-19-1-0399. The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of the Army Research Office (ARO), or the U.S. Government. The U.S. Government is authorized to reproduce and distribute reprints for Government purposes notwithstanding any copyright notation herein.
\section*{Author contributions}
A.E., S.S.E., M.H.D. and R.J.S. developed the large displacement control method. A.E., S.R.J. V.S., and A.Z.D. implemented the numerical ECD parameter optimization. A.E., V.S., and A.Z.D. conducted the measurements. A.E., B.R.,V.S., and S.M.G. developed the theory. J.V. and A.E. performed numerical analysis of the strongly driven nonlinear oscillator. A.E. and M.H.D. wrote the manuscript with feedback from all authors.
\section*{Competing interests}
R.J.S. and M.H.D. are founders and R.J.S. is a shareholder of Quantum Circuits, Inc.
\let\oldaddcontentsline\addcontentslin
\renewcommand{\addcontentsline}[3]{
\putbib[ECD]
\let\addcontentsline\oldaddcontentslin
\end{bibunit}
\setcounter{equation}{0}
\setcounter{figure}{0}
\setcounter{table}{0}
\setcounter{section}{0}
\makeatletter
\renewcommand{\theequation}{S\arabic{equation}}
\renewcommand{\thefigure}{S\arabic{figure}}
\renewcommand{\thetable}{S\arabic{table}}
\renewcommand{\thesection}{S\arabic{section}}
\crefname{table}{Table}{Tables}
\Crefname{table}{Table}{Tables}
\crefname{figure}{Fig.}{figs.}
\Crefname{figure}{Fig.}{Figures}
\crefname{section}{section}{sections}
\Crefname{section}{Section}{Sections}
\crefname{subsection}{section}{sections}
\Crefname{subsection}{Section}{Sections}
\title{Supplementary Information: Fast Universal Control of an Oscillator with Weak Dispersive Coupling to a Qubit}
\clearpage
\maketitle
{
\onecolumngrid
\hypersetup{linkcolor=black}
\tableofcontents
}
\begin{bibunit}[naturemag]
\section{Experimental design}
\label{sec:exp}
The sample consists of two coaxial microwave cavities machined out of aluminum 6061 alloy anchored at the base stage of a dilution refrigerator operating at $\SI{20}{mK}$. The lower-frequency cavity is used as a high-Q storage oscillator, while the other is overcoupled to a transmission line and used for readout of a fixed-frequency transmon qubit bridging the two cavities. The transmon includes a double-angle-evaporated $\mathrm{Al/AlO_x/Al}$ Josephson-junction fabricated on a sapphire substrate. An FPGA system is used to control the transmon and cavity with a DAC sampling rate of $\SI{1}{GS/s}$. The package and transmon used here is the same as was used in \cite{campagne-ibarcqQuantumErrorCorrection2020}, with device parameters that have aged since that publication. We refer the reader to \cite{campagne-ibarcqQuantumErrorCorrection2020} and associated supplementary material for more details, as well as the wiring diagram, for which the only major difference here is the lowering of amplification power and addition of room-temperature microwave switches on the storage and readout line for better noise properties.
Single-shot readout is performed using a SNAIL parametric amplifier operating with 20 dB of gain in phase-preserving mode \cite{frattiniOptimizingNonlinearityDissipation2018}. We use a square readout pulse of length $\SI{100}{ns}$ and acquire signal for $\SI{824}{ns}$. With additional FPGA delays, the total readout time is $\SI{1.176}{us}$, leading to a readout fidelity greater than $98\%$ as inferred by the measured average contrast of thresholded Rabi fringes. A measurement based feedback routine is used to prepare the transmon in $\ket{g}$ before each experimental iteration.
In equilibrium, the transmon's excited state population is $n_\text{th,q} \sim 0.0092$, corresponding to a temperature of $\sim\SI{68}{mK}$. We rely on a wait time longer than $5T_{1,c}$ between each experiment for the cavity to relax to near-equilibrium. As a conservative estimate, we assume the cavity mode is at the same temperature as the transmon when estimating error sources in \cref{sec:error budget}, corresponding to a cavity excited state population of $n_\text{th,c} \sim 0.025$ before the start of each experiment.
\section{System Hamiltonian and Parameters}
Given the range of displacements used in this work, our system is well described by the effective Hamiltonian \cite{niggBlackBoxSuperconductingCircuit2012, minevEnergyparticipationQuantizationJosephson2021a, blaisCircuitQuantumElectrodynamics2021}
\begin{align}
\label{eq:system Hamiltonian}
&\frac{H}{\hbar} = \Delta a^\dagger a - \chi a^\dagger a q^\dagger q - \chi'a^{\dagger2} a^2q^\dagger q -K_c a^{\dagger2} a^2 - K_q q^{\dagger2} q^2 + \varepsilon^*(t) a + \Omega^*(t)q + h.c.
\end{align}
where $a$ ($q$) are bosonic annihilation operators for the hybridized oscillator-like (transmon-like) modes. $\Omega(t)$ and $\varepsilon(t)$ are complex-valued drives generated by IQ modulation, and we have ignored terms rotating at twice the drive frequencies. $H$ is written in the co-rotating frame of the qubit and cavity drives, and in this work we use $\Delta = \chi/2$ when performing ECD gates, representing a cavity drive at frequency $(\omega_g + \omega_e)/2$, where $\omega_g$ ($\omega_e$) is the cavity frequency with the transmon in the ground (excited) state. Under this choice, the Hamiltonian given in the main text corresponds to projecting \cref{eq:system Hamiltonian} onto the ground and excited state transmon manifold using the mapping $\sigma_z = 1 - 2q^\dagger q$ and only keeping the dispersive interaction term. Hamiltonian parameters, as well as measured decoherence rates and mode frequencies, are given in \cref{tab:system parameters}, and measurement techniques for these some of these values are described in \cref{sec:calibration}.
\begin{table}[h]
\begin{tabular}{l@{\hskip 0.25in}l@{\hskip 0.25in}l}
\textbf{parameter} & \textbf{value} \\
\hline
transmon g-e transition frequency & $\omega_{ge} = 2\pi \times \SI{6.65}{GHz}$ \\
transmon anharmonicity & $K = 2K_q = 2\pi \times \SI{193}{MHz}$ \\
transmon Ramsey coherence & $T_{2R,q} = \SI{30}{us}$ \\
transmon echo coherence & $T_{2E,q} = \SI{65}{us}$ \\
bare transmon relaxation& $ T_{1,q} = \SI{50}{us}$ \\
transmon relaxation with $\bar{n}_\text{cav} = 900$ & $ \tilde{T}_{1,q} = \SI{30}{us}$ \\
transmon equilibrium population & $ n_\text{th} = 0.0092$ \\
readout frequency & $\omega_r = 2\pi \times \SI{8.22}{GHz}$ \\
readout dispersive shift & $\chi_{qr} = 2\pi \times\SI{0.96}{MHz}$ \\
readout relaxation rate & $\kappa_r = \SI{1.7}{MHz}$ \\
storage cavity frequency & $\omega_c = 2\pi \times \SI{5.26}{GHz}$\\
storage dispersive shift & $\chi = 2\pi \times \SI{32.8}{kHz}$ \\
storage second-order dispersive shift & $\chi' = 2\pi \times \SI{1.5}{Hz}$\\
storage cavity Kerr & $2K_c = 2\pi \times \SI{1}{Hz}$\\
storage cavity relaxation & $T_{1,c} = \SI{436}{us}$ \\
storage cavity Ramsey coherence & $T_{2R,c} < 2 T_{1,c} = \SI{872}{us}$
\end{tabular}
\caption{\label{tab:system parameters} \textbf{Measured system parameters and loss rates}. Measurement of the dispersive shift, the second order dispersive shift, and Kerr is described in \cref{sec:hamiltonian measurement}. The Cavity relaxation time $T_{1,c}$ is measured by preparing a coherent state $\alpha_0 = 3.6$ and measuring $\braket{\hat{a}^\dagger \hat{a}}(t)$ using time-dependent transmon spectroscopy. The limit on the cavity Ramsey coherence time $T_{2,c}$ is inferred from the cavity relaxation time.}
\end{table}
\section{System Characterization and Calibration}
\label{sec:calibration}
In this section, we outline techniques to characterize an oscillator and qubit coupled with $\chi/2\pi$ on the order of or smaller than qubit decoherence rates. In this regime, methods for calibration of control and Hamiltonian parameters which rely on large number-splitting are inefficient. For these calibrations, we rely on semiclassical phase-space trajectories as derived in \cref{sec:trajectories}. In \cref{sec:out and back}, we describe the \textit{out-and-back} method, which uses large displacements of the oscillator mode to realize a measurement of Hamiltonian parameters. Finally in \cref{sec:geometric phase} we describe a simple geometric phase measurement which is used to calibrate the oscillator drive strength $|\varepsilon|$.
\subsection{Semiclassical phase space trajectories}
\label{sec:trajectories}
Starting from $H$ in \cref{eq:system Hamiltonian}, we perform a time-dependent displaced frame transformation using the unitary $U = D^\dag(\alpha(t)) = \exp\left\{\alpha^*(t) a - \alpha(t) a^\dagger\right\}$. This modifies the state according to $\tilde{\rho}(t) = D^\dag(\alpha(t)) \rho(t) D(\alpha(t))$, and the Hamiltonian according to
$H \rightarrow \tilde{H} = UH U^\dagger + \left(i\hbar\right) \left(\partial_t U\right) U^\dagger = D^\dagger \left(\alpha(t)\right) H D\left(\alpha(t)\right) + \left(i\hbar\right) \left(a \partial_t \alpha^*(t) - a^\dagger \partial_t \alpha(t)\right)$, giving
\begin{equation}
\label{eq:displaced H}
\begin{aligned}
\frac{\tilde{H}}{\hbar} &= \Delta a^\dagger a - (\chi + 4\chi'|\alpha|^2) a^\dagger a q^\dagger q - \chi' a^{\dagger2} a^2 q^\dagger q - K_c a^{\dagger2} a^2 - K_q q^{\dagger 2}q^2 - (\chi + 2 |\alpha|^2 \chi')(\alpha^*a + \alpha a^\dagger)q^\dagger q & \\
&- (\chi|\alpha|^2 + \chi' |\alpha|^4)q^\dagger q - 4K_c|\alpha|^2 a^\dagger a + \left(\Delta \alpha^* -2K_c|\alpha|^2\alpha^* + i (\partial_t \alpha^*) + i \frac{\kappa}{2} \alpha^* + \varepsilon^* \right)a + h.c. & \\
&- K_c\left(2\alpha a^{\dagger 2}a + \alpha^2 a^{\dagger2} + h.c.\right) - \chi'\left(2\alpha a^{\dagger 2}a + \alpha^2 a^{\dagger2} + h.c.\right)q^\dagger q+ \Omega^*(t)q + h.c.
\end{aligned}
\end{equation}
We have also included the deterministic part of oscillator relaxation at a rate $\kappa/2$ as described in the methods section of the main text.
\subsubsection{Simulating in the displaced frame}
\label{sec:simulation displaced frame}
Given an oscillator drive $\varepsilon(t)$, it is numerically advantageous to cancel the classical part of the oscillator's phase-space trajectory by picking a time-dependent frame $\alpha(t)$ which cancels the displacement term in $\tilde{H}$. This is done by solving
\begin{equation}
\label{eq:cm trajectory}
\begin{aligned}
\partial_t \alpha(t) &= -i \Delta \alpha(t) + 2i K_c |\alpha(t)|^2 \alpha(t) - \frac{\kappa}{2} \alpha(t) - i \varepsilon(t) \\
\alpha(0) &= 0
\end{aligned}
\end{equation}
for $\alpha(t)$. Once $\alpha(t)$ is known, $\tilde{H}$ without the linear terms in $a$ and $a^\dagger$ is used to efficiently simulate a cavity and transmon evolution in the displaced frame using a truncated Hilbert space. This displaced frame Hamiltonian is used for master equation simulations in \cref{sec:error budget}.
\subsubsection{Semiclassical trajectories}
During a period where the qubit populations stay constant ($\Omega(t) =0$), we can instead determine the oscillator's phase-space trajectories conditioned on the qubit's ground or excited state. This semiclassical approximation is done by replacing $q^\dagger q$ in $\tilde{H}$ with $\left\{0,1\right\}$ for transmon states $\left\{\ket{g},\ket{e}\right\}$. This replacement gives the two Hamiltonian sectors
\begin{equation}
\begin{aligned}
\frac{\tilde{H}_g}{\hbar} &= (\Delta - 4K_c|\alpha|^2) a^\dagger a - K_c a^{\dagger2} a^2 + \left(\Delta \alpha^* -2K_c|\alpha|^2\alpha^* + i (\partial_t \alpha^*) + i \frac{\kappa}{2} \alpha^* + \varepsilon^* \right)a + h.c. \\
&- K_c\left(2\alpha a^{\dagger 2}a + \alpha^2 a^{\dagger2} + h.c.\right), \\
\frac{\tilde{H}_e}{\hbar} &= (\Delta - \chi - 4\chi'|\alpha|^2 - 4K_c|\alpha|^2) a^\dagger a - (\chi' + K_c) a^{\dagger2} a^2 \\
&+ \left(\Delta \alpha^* -2K_c|\alpha|^2\alpha^* + i (\partial_t \alpha^*) + i \frac{\kappa}{2} \alpha^* + \varepsilon^* - (\chi + 2 |\alpha|^2 \chi')\alpha^* \right)a + h.c. \\
&- (K_c + \chi')\left(2\alpha a^{\dagger 2}a + \alpha^2 a^{\dagger2} + h.c.\right)
\end{aligned}
\end{equation}
describing the dynamics of the driven oscillator when the transmon is in the ground or excited state. Similar to the displaced-frame simulation, the linear part of these Hamiltonians can be individually cancelled, resulting in the two equations
\begin{equation}
\label{eq:nonlinear trajectories}
\begin{aligned}
\partial_t \alpha_g(t) &= -i \Delta \alpha_g(t) + 2i K_c |\alpha_g(t)|^2 \alpha_g(t) - \frac{\kappa}{2} \alpha_g(t) - i \varepsilon(t)\\
\partial_t \alpha_e(t) &=-i \Delta \alpha_e(t) + 2i K_c |\alpha_e(t)|^2 \alpha_e(t) - \frac{\kappa}{2} \alpha_e(t) - i \varepsilon(t) + i(\chi + 2\chi'|\alpha_e(t)|^2)\alpha_e(t)
\end{aligned}
\end{equation}
which can be used to calculate the semiclassical trajectories for the ground or excited states during periods when $\Omega(t) = 0$. In the case of a conditional displacement, after each $\pi$ pulse, the Hamiltonians are swapped, and the result from the previous part of the trajectory is used to seed the next initial value problem. In our simulations, we solve these nonlinear initial value problems using a central-difference method with trajectories sampled at $\SI{1}{ns}$. These trajectories are used in \cref{sec:hamiltonian measurement} for Hamiltonian parameter calibration and in \cref{sec:ECD optimization} for optimization of the cavity and qubit drives to produce ECD gates.
\subsection{Measurement of Hamiltonian parameters using out-and-back}
\label{sec:hamiltonian measurement}
\label{sec:out and back}
Although many of the Hamiltonian terms are small relative to the rate of transmon decoherence, they can be estimated in experiment using large displacements to enhance their effective strength. Here, we make use of this enhancement through the measurement sequence shown in \cref{fig:out and back a}. We first prepare the qubit in the ground or excited state then displace the oscillator out by $\alpha_0$. After a time $t$, the oscillator is displaced back by $-e^{i\phi} \alpha_0$, where $\phi$ is swept. The second displacement serves as an attempt to displace the oscillator's state back to the origin of phase-space. If the attempt is successful, $\phi$ encodes the oscillator's coherent-state phase accumulation at a displacement $\alpha_0$ after a time $t$. A narrow-bandwidth $\pi$ pulse ($\sigma = \SI{200}{ns}$) is then used as a probe, only flipping the transmon's state measured by $m_2$ if the oscillator's state is close to the origin of phase space. Finally, we postselect the results of $m_2$ on the condition $m_1 = \ket{\psi_i}$, where $\ket{\psi_i}$ is the initial transmon state, in order to remove the influence of transmon relaxation or heating.
In \cref{fig:out and back c}, we show the results of this experiment with $t = \SI{1}{us}$ and $N = 5$ repetitions of out and back, used to enhance the sensitivity. In initial state $\ket{e}$, the signal is lost above $\braket{n}\gtrsim 2500$ photons, and in initial state $\ket{g}$, the signal is lost above $\braket{n} \gtrsim 7500$ photons. These values represent the oscillator photon numbers at which the qubit is excited outside of the $\ket{g},\ket{e}$ manifold due to higher-order nonlinear transitions, a process sometimes referred to as \textit{bright-stating} that has been observed in previous experiments using readout resonators \cite{reedHighFidelityReadoutCircuit2010a, sankMeasurementInducedStateTransitions2016, lescanneEscapeDrivenQuantum2019}. Such an effect could be suppressed by using an inductive shunt, proving a path forward to engineering faster gates \cite{verneyStructuralInstabilityDriven2019a}. \cref{fig:out and back} also indicates the critical photon number at which the dispersive approximation begins to fail calculated using the expressions in the methods section of the main text, $n_\text{crit}^g \approx 2740$ and $n_\text{crit}^e \approx 910$. This serves as a guiding principle for the speed limit of control.
\begin{figure*}[!ht]
\begin{center}
\vspace{-2\baselineskip
\phantomsubfloat{\label{fig:out and back a}}
\phantomsubfloat{\label{fig:out and back b}}
\phantomsubfloat{\label{fig:out and back c}}
\phantomsubfloat{\label{fig:out and back d}}
\includegraphics{out_and_back}
\caption{\label{fig:out and back} \textbf{Out-and-back measurement} \textbf{a.} Experimental sequence to measure the phase accumulation of a coherent state with radius $\alpha_0$ after a time $t$ when the transmon is in the ground or excited state. \textbf{b.} Measurement result with initial transmon state $\ket{g}$ (top panel) and $\ket{e}$ (bottom panel) with a fixed wait time of $t=\SI{1}{us}$ and $N=5$ repetitions post-selected on measurement $m_1$ matching the prepared state. \textbf{c.} Measured cavity frequency dispersion (open circles) compared to numerical diagonalization of \cref{eq:coupled H} (solid lines). The open circles is a fit of different experimental data than shown in panel (b), with higher resolution out to $\alpha_0^2 = 3000$. \textbf{d} Measured transmon relaxation as a function of $\braket{n}$ found by analyzing $m_1$ measurement results with initial transmon state $\ket{e}$ and sweeping $t$. $\phi$ is chosen close to values which successfully displace the cavity state back to the origin of phase space. Measurement results shown for three consecutive days with each experiment averaging for around 1 hour.}
\end{center}
\end{figure*}
At each $\braket{n}$ below the bright-stating point, the relative oscillator frequency when the transmon is in the ground or excited state is obtained by fitting $\braket{\sigma_z}$ to a Gaussian function and dividing its mean phase accumulation by the wait time $t$. The resulting relative frequency dispersion as a function of average cavity photon number $\braket{n} = \alpha_0^2$ for transmon states $\ket{g}$ and $\ket{e}$ is shown in \cref{fig:out and back c} by the open circles. Note that the finite duration of the displacement pulses will also influence the result, and as a secondary check, the experiment can be repeated at different wait times $t$. Clearly the dispersion is not constant, representing the need to include higher-order nonlinear terms beyond the dispersive coupling in our effective Hamiltonian, as is done in \cref{eq:system Hamiltonian}.
To extract the effective Hamiltonian parameters, we fit the measured dispersion (open circles in \cref{fig:out and back}) to that expected from $H$ in \cref{eq:system Hamiltonian}. In particular, the semiclassical trajectories in \cref{eq:nonlinear trajectories} which govern the evolution of coherent states predict the effective cavity rotation frequencies to be
\begin{align}
\Delta_{g} &= \Delta - 2K_c\alpha_0^2\\
\Delta_{e} &= \Delta - \chi - (2K_c + 2\chi')\alpha_0^2
\end{align}
as a function of the average number of photons in the cavity $\alpha_0^2$ when the transmon is in the ground and excited states respectively. By fitting the sum and difference of two dispersion curves to the sum and difference of these linear functions, we can extract the four unknown Hamiltonian parameters $\chi$, $\chi'$, $K$, and $\Delta$. The dispersion fits well to a linear function in the range of interest for control, for photon numbers up to $\alpha_0^2 < 2000$, with results given in \cref{tab:system parameters}. For experiments using echoed conditional displacements, we use this experiment to calibrate the cavity drive frequency such that $\Delta = \chi/2$.
To further confirm our model, we compare the measured dispersion to a numerical diagonalization of the coupled transmon-oscillator Hamiltonian
\begin{equation}
\label{eq:coupled H}
\frac{H}{\hbar} = 4 E_c (\hat{N} - N_g)^2 - E_J \cos\left(\hat{\varphi}\right) + \omega_{c}^{\text{bare}} a^\dagger a + g (\hat{N} - N_g)\left(a + a^\dagger \right)
\end{equation}
where $\hat{N}$ is the Cooper-pair number operator, $\hat{\varphi}$ is the conjugate Josephson phase, $N_g$ is the offset charge in units of $2e$, and $\hat{a}$ is the bare cavity mode annihilation operator \cite{kochChargeinsensitiveQubitDesign2007}. For photon numbers up to $\braket{n} \approx 1000$, we have confirmed that the oscillator's dispersion for qubit states $\ket{g}$ and $\ket{e}$ does not depend on offset charge, so we set $N_g = 0$. With this choice, we use second-order perturbation theory to find the bare Hamiltonian parameters which fit the measured hybridized mode frequencies, transmon anharmonicity, and dispersion at low $\braket{n} \approx 0$, resulting in the bare parameters $g/2\pi = \SI{9.12}{MHz}$, $E_j/2\pi = \SI{32.33}{GHz}$, $E_c/2\pi = \SI{181}{MHz}$ and $\omega_{c}^{\text{bare}} = \SI{5.26}{GHz}$. Using these parameters, we numerically diagonalize \cref{eq:coupled H} in the basis of transmon eigenstates with a Hilbert space of 2800 oscillator states and 12 transmon states. The resulting dispersion is shown by the solid lines in \cref{fig:out and back c}, which have excellent agreement with the measured dispersion. Also, the diagonalization predicts a breakdown of our quantum number assignment algorithm for transmon state $\ket{e}$ at 2500 photons in the oscillator, matching the measured breakdown in experiment. Finally, we note that higher-order perturbation theory can also be used to predict an analytic effective Hamiltonian at large photon numbers in the cavity, as shown in \cite{venkatramanStaticEffectiveHamiltonian2022}.
We use a similar out-and-back measurement to probe the transmon relaxation and heating rate while the cavity is displaced to a large coherent state, since a reduction in transmon lifetime has been observed when displacing readout resonators \cite{minevCatchReverseQuantum2019}. For this, we use the out-and-back sequence in \cref{fig:out and back a}, except we sweep $t$, and fix $\phi(t)$ close to phases that displace the oscillator's state back to the origin of phase-space at each $\braket{n}$ given the measured dispersion. In this case, we focus on the result of $m_1$. We find that, up to $2000$ oscillator photons, there is no appreciable heating out of $\ket{g}$ when displacing the cavity state, indicating that the dressed dephasing rate is small \cite{boissonneaultDispersiveRegimeCircuit2009}. However, when preparing the transmon in $\ket{e}$, we measure that the transmon's relaxation rate shows a dependence on cavity photon number. In \cref{fig:out and back d}, we plot the measured probability of the transmon remaining in $\ket{e}$ after a wait time $t$ up to $\SI{4}{us}$ when displacing the cavity to $\braket{n} = \alpha_0^2$, with experiments run on three consecutive days. We suspect this time-dependent $\tilde{T}_{1,q}$ vs $\bar{n}_\text{cav}$ effect is caused by fluctuating two-level-systems (TLS) which come into resonance with the transmon as it is stark-shifted by cavity photons \cite{klimovFluctuationsEnergyRelaxationTimes2018, burnettDecoherenceBenchmarkingSuperconducting2019,carrollDynamicsSuperconductingQubit2021}. Although the $T_1$ vs $\bar{n}$ changes with time, we find an average value of $\tilde{T}_{1,q} \approx \SI{30}{us}$ for $\bar{n} = 900$ over the data plotted in \cref{fig:out and back d}, and we use this value when performing master equation simulations in \cref{sec:error budget}.
\subsection{Oscillator drive strength calibration using geometric phase}
\label{sec:geometric phase}
In this section, we discuss a simple experiment which can be used to calibrate a linear oscillator drive amplitude in the weak-$\chi$ regime. Starting with the qubit prepared in $\ket{\psi_i} = \frac{1}{\sqrt{2}}\left(\ket{g} + \ket{e}\right)$, we construct an oscillator drive sequence which encloses an area in phase space for both $\ket{g}$ and $\ket{e}$ trajectories. By disentangling the qubit and oscillator at the end of the sequence, the qubit will be left in the state $\ket{\psi_f} = \frac{1}{\sqrt{2}}\left(\ket{g} + e^{i\phi}\ket{e}\right)$ where $\phi$ encodes the enclosed area \cite{chaturvediBerryPhaseCoherent1987,vacantiGeometricphaseBackactionMesoscopic2012,pechalGeometricPhaseNonadiabatic2012a,songContinuousvariableGeometricPhase2017}. Given an arbitrary displacement pulse shape $g(t)$ with a length $t_p$, a simple pulse sequence that accomplishes this is
\begin{equation}
\varepsilon(t) = \varepsilon_0\left[g(t) - r g(t - (t_p + t_w)) - r g(t - (2t_p + t_w)) + g(t - (2t_p + 2t_w))\right]
\end{equation}
as shown in \cref{fig:geometric phase a} with a phase-space trajectory shown in \cref{fig:geometric phase b}. This drive is similar to a conditional displacement without a qubit $\pi$ pulse, and the goal here is to calibrate the pulse scale $\varepsilon_0$.
To analyze this sequence, we note that for low photon numbers, the Hamiltonian is well described by only the dispersive term. With this approximation, the semiclassical trajectories in \cref{eq:nonlinear trajectories} can be solved with initial value $\alpha_{_{e}^{g}}(0) = 0$ giving
\begin{equation}
\label{eq:dispersive ivp sol}
\alpha_{_{e}^{g}}(t) = e^{-\frac{1}{2}\left(\pm i\chi + \kappa\right)t} \left(e^{\frac{1}{2}\left(\pm i\chi + \kappa\right)t_0}\alpha_{_{e}^{g}}(t
_0) - i \int_{t_0}^t e^{\frac{1}{2}\left(\pm i\chi + \kappa\right)\tau}\varepsilon(\tau)d\tau\right).
\end{equation}
Subsituting $\varepsilon(t)$ into \cref{eq:dispersive ivp sol}, we solve for the ratio of the middle-two pulses, $r$, such that the condition $\alpha_{_e^g} = 0$ is satisfied at the end of the entire sequence, and the qubit and oscillator are disentangled. Remarkably, this ratio is independent of the shape of the displacement pulse $g(t)$, and is found to be
\begin{equation}
\label{eq:geometric pulse ratio}
r = \frac{1 + e^{\frac{1}{2}\left(\pm i\chi + \kappa\right)(3t_p + 2t_w)} }{e^{\frac{1}{2}\left(\pm i\chi + \kappa\right)(t_p + t_w)} + e^{\frac{1}{2}\left(\pm i\chi + \kappa\right)(2t_p + t_w)} }
\end{equation}
where $t_p$ is the length of the displacement pulses $g(t)$ and $t_w$ is the wait time. By Taylor expanding this ratio in orders of $\kappa$, we find in the limit $(t_p + t_w) << 1/\kappa$,
\begin{equation}
r = \cos\left(\frac{\chi}{4}(3 t_p + 2t_w)\right)\sec\left(\frac{\chi}{4} t_p\right)
\end{equation}
independent of the qubit state. In the high-Q oscillator limit, this sequence disentangles the oscillator and qubit and can be used to measure the geometric area enclosed by the trajectory.
Using \cref{eq:dispersive ivp sol}, we numerically integrate the sequence with displacement pulses $g(t)$ chosen as truncated Gaussians with $\sigma = \SI{11}{ns}$ and $t_p = \SI{44}{ns}$ to find the phase-space trajectories for $\alpha_{_{e}^{g}}$, including the maximum phase-space radius $\alpha_0$, and the associated geometric phase difference. By measuring $\braket{\sigma_x}$ and $\braket{\sigma_y}$ while sweeping the drive scale, this phase difference is fit to the experiment, allowing a calibration of $\varepsilon_0$ in terms of DAC amplitude. In \cref{fig:geometric phase c} we show an example of the measured phase accumulation as $\varepsilon_0$ is linearly scaled ($t_w = \SI{200}{ns}$), along with the phase predicted using the integrated geometric area. If desired, nonlinear effects can be included by using \cref{eq:nonlinear trajectories} when calculating the trajectories, however in this case it is no longer guaranteed that $r$ given by \cref{eq:geometric pulse ratio} will exactly disentangle the oscillator and qubit. Using this, we find a maximum drive amplitude of $|\varepsilon|_\text{max}/2\pi \approx \SI{400}{MHz}$ before saturating our room-temperature amplification chain.
\begin{figure*}[ht]
\begin{center}
\vspace{-2\baselineskip
\phantomsubfloat{\label{fig:geometric phase a}}
\phantomsubfloat{\label{fig:geometric phase b}}
\phantomsubfloat{\label{fig:geometric phase c}}
\includegraphics{geometric_phase}
\caption{\label{fig:geometric phase} \textbf{Geometric phase measurement} \textbf{a.} Geometric phase measurement drive sequence and drive parameterization. \textbf{b.} Phase space trajectories for $\alpha_g$ and $\alpha_e$. \textbf{c} Measured $\braket{\sigma_x}$ and $\braket{\sigma_y}$ compared to expected phase found by solving the semiclassical trajectories in \cref{eq:nonlinear trajectories} (solid lines) as a function of phase space radius $\alpha_0$.}
\end{center}
\end{figure*}
\section{The Echoed Conditional Displacement Gate}
In \cref{sec:ideal gate} we derive the echoed conditional displacement unitary assuming an ideal dispersive Hamiltonian. Next, in \cref{sec:ECD optimization}, we optimize the ECD gate considering drive constraints and higher-order nonlinearities.
\subsection{Derivation of the ECD gate}
\label{sec:ideal gate}
Here, $\varepsilon(t)$ is a complex-valued function representing the envelope of an I-Q modulated drive with a carrier frequency $(\omega_g + \omega_e)/2$, where $\omega_g$ ($\omega_e$) is the oscillator's frequency when the qubit is in the ground (excited) state. In the co-rotating frame at the drive frequency and qubit frequency, the ideal Hamiltonian is
\begin{equation}
\frac{H}{\hbar} = \chi a^\dagger a \frac{\sigma_z}{2} + \varepsilon^*(t) a + \varepsilon(t) a^\dagger
\end{equation}
where we have neglected terms rotating at twice the drive frequency.
The echoed conditional displacement gate consists of two driving steps with a qubit $\pi$ pulse between, here assumed instantaneous. In this case, the general solution to the time-dependent Schrödinger equation $i\hbar\partial_t U = H U$ is
\begin{equation}
U = \mathcal{T} e^{-\frac{i}{\hbar} \int_{t_1}^{T} H(\tau) d\tau}\sigma_x \mathcal{T}e^{-\frac{i}{\hbar} \int_{0}^{t_1} H(\tau) d\tau}
\end{equation}
where $t_1$ is the time of the $\pi$ pulse, $T$ is the total time of the gate, and $\mathcal{T}$ is the time-ordering operator.
To represent the action of the $\pi$ pulse flipping the sign of $\sigma_z$ between the two trajectories, we instead modify the dispersive Hamiltonian to include a function $z(t) = \pm 1$ which represents the sign of $\sigma_z$, giving
\begin{equation}
\frac{H}{\hbar} = \chi a^\dagger a \frac{\sigma_z z(t)}{2} + \varepsilon^*(t) a + \varepsilon(t) a^\dagger.
\end{equation}
Using this modified Hamiltonian, we take as an ansatz for the solution of the Schrödinger equation
\begin{equation}
U = e^{i \theta \frac{\sigma_z}{2}}e^{a^\dagger (\gamma + \delta \sigma_z) - a(\gamma^* + \delta^* \sigma_z)}e^{i\phi a^\dagger a \sigma_z}
\end{equation}
where $\theta$ represents an ancilla qubit phase, $\gamma$ and $\delta$ represent a displacement and conditional displacement, $\phi$ represents a qubit state-dependent rotation of the oscillator, and these variables are time-dependent.
Ignoring a global phase, the Schrödinger equation gives
\begin{align}
\partial_t \theta &= - 2 \text{Re}\left[\delta \varepsilon^*\right] \\
\partial_t \gamma &= -i\frac{\chi}{2} z(t)\delta - i\varepsilon \\
\partial_t \delta &= -i \frac{\chi}{2}z(t) \gamma \\
\partial_t \phi &= -\frac{\chi}{2}z(t).
\end{align}
These equations can be solved, giving
\begin{align}
\label{eq:sol start}
\theta(t) &= -2 \int_{0}^{t} d\tau \text{Re}\left[\varepsilon^*(\tau)\delta(\tau)\right] \\
\gamma(t) &= -i \int_{0}^{t} d\tau \cos\left[\phi(\tau) - \phi(t)\right]\varepsilon(\tau) \\
\delta(t) &= -\int_{0}^{t} d\tau \sin\left[\phi(\tau) - \phi(t)\right]\varepsilon(\tau) \\
\phi(t) &= -\frac{\chi}{2} \int_{0}^{t} d\tau z(\tau).
\label{eq:sol end}
\end{align}
The state-dependent rotation of the oscillator can be canceled by setting $\phi(T) = 0$. For the echoed conditional displacements in this work, this is done by applying a single qubit $\pi$ pulse at time $T/2$. If desired, more qubit echos can be included, subject to the condition $\int_{0}^{T} d\tau z(\tau) =0$.
Using the Baker-Campbell-Hausdorff formula, the conditional displacement and displacement can be separated, giving the overall unitary
\begin{equation}
\label{eq:cd_unitary}
U = \sigma_xe^{i \theta' \frac{\sigma_z}{2}} D(\lambda) CD(\beta)
\end{equation}
corresponding to a conditional displacement $CD(\beta) = D(\beta/2)\ket{g}\bra{g} + D(-\beta/2)\ket{e}\bra{e}$, a displacement, and an additional qubit phase with the parameters $\beta = 2\delta(T)$, $\lambda = \gamma(T)$, and $\theta' = \theta(T) + 2 \text{Im}\left[\gamma(T)\delta(T)\right]$. We have also explicitly included a $\sigma_x$ operator to represent the action of the single $\pi$ pulse (here $\text{ECD}\left(\beta\right) = \sigma_x CD(\beta)$).
To realize an ECD gate, we aim to null the qubit phase and oscillator displacement. In the limit of instantaneous displacements and motivated by the geometry of rotating phase-space, this can be perfectly achieved by choosing the drive
\begin{align}
\label{eq:perfect drive}
\varepsilon(t) = \alpha\left[\delta(t) - 2\delta(t - T/2)\cos\left(\chi T /4\right) + \delta(t - T)\cos\left(\chi T/2 \right)\right]
\end{align}
where $\delta(t)$ is the Dirac delta function. This drive corresponds to the displacement sequence described in Fig.\ 1b of the main text with $\beta = 2 i \alpha e^{i\phi} \sin\left(\chi T/2\right)$ and $|\alpha| = \alpha_0$.
The drive in \cref{eq:perfect drive} cannot be realized in experiment due to bandwidth and amplitude limits of a realistic microwave drive.
Also, effects such as photon loss, higher-order nonlinearities, and the finite duration of the qubit $\pi$ pulse are not taken into account in \Crefrange{eq:sol start}{eq:sol end}.
To realize a high-fidelity ECD gate in the presence of these effects, we optimize $\varepsilon(t)$ using semiclassical trajectories as described in the next \cref{sec:ECD optimization}.
\subsection{Optimization of the ECD gate}
\label{sec:ECD optimization}
In our experiment, the dynamics can slightly differ from those described by \Crefrange{eq:sol start}{eq:sol end} due to the second-order dispersive shift $\chi'$ and the oscillator Kerr $K_c$ which become relevant at large phase-space displacements. These effects can be studied by examining the displaced-frame Hamiltonian in \cref{eq:displaced H}. In $\tilde{H}$, nonlinear terms proportional to $\chi'$ or $K_c$ can generally cause distortions to the state. However, simulations indicate that these terms do not significantly decrease the fidelity of ECD control protocols given our system parameters in \cref{tab:system parameters}. This is partially because the deleterious effect of terms which are proportional to $\text{sign}(\alpha)$ or $\text{sign}(\sigma_z)$ are significantly reduced due to the phase-space echo $\alpha \rightarrow -\alpha$ and qubit echo $\ket{g} \leftrightarrow \ket{e}$ at time $T/2$ during the ECD gate, which cancels part of their on-average effect to the state distortion in the same way the qubit-state dependent oscillator rotation is canceled during the ideal ECD gate.
With this in mind, we optimize ECD gates using semiclassical trajectories (\cref{sec:trajectories}) which account for the linear displaced-frame terms (proportional to $a$), including those caused by the second-order disperse shift, Kerr, and photon loss. We assume a form of the unitary still given by \cref{eq:cd_unitary}, with the values of $\beta$, $\lambda$ and $\theta'$ calculated using the trajectories $\alpha_{_e^g}(t)$ for the ground and excited qubit states. We note that $\text{ECD}(\beta)$ can be generated with any $\alpha_0$ or $\chi$ as long as $\chi T < 2\pi$ and the qubit rotation pulse bandwidth is sufficiently large compared to $\chi$. In particular, a large $\alpha_0$ is not required.
To construct each ECD gate, we start by imposing the drive to be of the form shown in \cref{fig:CD drives a}, which replaces the Dirac $\delta$-functions in \cref{eq:perfect drive} with fixed-length Gaussian waveforms, chosen in our experiment with a standard deviation of $\sigma = \SI{11}{ns}$ and a total length of $4\sigma = \SI{44}{ns}$. This simplification is chosen so that the drive strength required to realize the large displacements used in this work remain in the linear regime of our room temperature amplification chain, and so the displacements take the exact same form as those used to calibrate the drive amplitude using the geometric phase measurement shown in \cref{sec:geometric phase}. The amplitude ratio of the second, third, and fourth Gaussian to the first are given by $r_2$, $r_3$, and $r_4$, and the wait time between the displacements is given by $t_w$.
Using a simple optimization strategy, we find the values of $\left\{\varepsilon_0, r_2,r_3, r_4 \right\}$ which realize a target conditional displacement $\beta$ with intermediate phase-space radius $\alpha_0$ in the shortest time $t_w$. Starting with a large guess time $t_w$, the parameters are optimized with a Nelder-Mead method using the cost function
\begin{align}
\text{C}=
\left|\alpha_g(T/2) + \alpha_e(T/2)\right|^2 + \left|\alpha_g(T) + \alpha_e(T)\right|^2 + \left(\frac{\left|\alpha_g(T/4)+ \alpha_e(T/4)\right|}{2} - \alpha_0\right)^2 + \left(\frac{\left|\alpha_g(3T/4)+ \alpha_e(3T/4)\right|}{2} - \alpha_0\right)^2
\end{align}
where $\alpha_{_e^g}(t)$ is calculated using \cref{eq:nonlinear trajectories} including the second-order dispersive shift, Kerr, and photon loss.
This cost function minimizes the final and midpoint net displacement, and ensures an intermediate phase space radius of $\alpha_0$ for the first and third displacements. Once $\left\{\varepsilon_0, r_2,r_3, r_4 \right\}$ have converged at a given $t_w$, $\beta$ is calculated as $\alpha_e(T) - \alpha_g(T)$, and $t_w$ is stepped down until the target $\beta$ is realized. If $t_w$ reaches 0, and the target $\beta$ has not been reached, then $\alpha_0$ is reduced until the target $\beta$ is realized. We note that shorter pulses could instead be used in this small-$\beta$ case, however in our proof of principle example, we keep the displacement duration fixed such that the pulses do not occupy a larger bandwidth than those used in the calibration.
The form of the resulting conditional displacement strongly depends on $\beta$ and the choice of $\alpha_0$. In \cref{fig:CD drives b} and \cref{fig:CD drives c} we illustrate the result of this optimization using our system parameters for $\beta = 1$ in two different regimes, $\alpha_0 = 10$ and $\alpha_0 = 50$. In the first case, a majority of the conditional displacement is accumulated during the wait times $t_w$. In the second case, $t_w$ is reduced to $0$, resulting in $\alpha_0$ being further lowered to $\approx 45$ to realize the gate. In this regime, the conditional displacement is accumulated during the driving periods, and increasing the target $\alpha_0$ does not result in a faster gate after optimization. This is the reason for the drive-constraint limit shown in Fig. 2c of the main text.
\begin{figure*}[!ht]
\begin{center}
\vspace{-2\baselineskip
\phantomsubfloat{\label{fig:CD drives a}}
\phantomsubfloat{\label{fig:CD drives b}}
\phantomsubfloat{\label{fig:CD drives c}}
\includegraphics{cd_geometry}
\caption{\label{fig:CD drives} \textbf{Echoed conditional displacement} \textbf{a.} Parameterization of the drive $\varepsilon(t)$ for the echoed conditional displacement. \textbf{b. and c.} Echoeded conditional displacement gate for $\beta = 1$ optimized with target $\alpha_0=10$ (\textbf{b}) and $\alpha_0 = 50$ (c). Top panels show the resulting drives, and bottom panels show the semi-classical phase space trajectories at a large aspect ratio.}
\end{center}
\end{figure*}
A full ECD sequence specified by the parameters $\left\{\vec\beta,\vec\phi, \vec\theta \right\}$ is compiled into drives $\varepsilon(t)$ and $\Omega(t)$ by first optimizing the drives for each ECD gate given a target $\alpha_0$. These are interleaved with qubit rotations, which are performed by modifying the phase and amplitude of fixed-length Gaussian pulses with $\sigma = \SI{6}{ns}$ and total length $4\sigma = \SI{24}{ns}$ independently calibrated in experiment. We note that pulse shaping techniques such as derivative reduction by adiabatic gate (DRAG) could also be incorporated to realize shorter transmon pulses \cite{motzoiSimplePulsesElimination2009,chenMeasuringSuppressingQuantum2016a}. Finally, the phase of each qubit pulse is updated to account for the additional phases associated with each ECD gate ($\theta'$) calculated by \cref{eq:sol start}. This correction is done by keeping track of the qubit frame given all preceding ECD gates and updating the phase $\varphi$ of each qubit rotation gate accordingly. As an example, in \cref{fig:GKP pulse} we show the compiled ECD pulse sequence used to prepare $\ket{+Z}_\text{GKP}$ in the cavity. Here, we use $\alpha_0 = 30$, however some ECD gates are performed at a smaller $\alpha_0' < \alpha_0$ resulting from the finite displacement pulse duration and the ECD optimization procedure described above.
\begin{figure*}[!ht]
\begin{center}
\includegraphics{gkp_pulse}
\caption{\label{fig:GKP pulse} \textbf{GKP state preparation ECD pulse sequence} used to prepare $\ket{+Z}_\text{GKP}$ in the cavity starting from vacuum using a circuit depth $N=9$ (here, $\Delta_\text{target} = 0.306$ as described in \cref{sec:GKP state analysis} and the ECD control parameters $\left\{\vec\beta,\vec\phi, \vec\theta \right\}$ are found using the procedure described in \cref{sec:ECD optimization}). $\varepsilon(t)$ is the cavity drive and $\Omega(t)$ is the transmon drive. Red and blue colors denote the real and imaginary parts of these drives, respectively.}
\end{center}
\end{figure*}
Finally, we note that the ECD gate set is designed to be well suited in the weak-$\chi$ regime since it requires fast unselective qubit rotations, an important operation that can become challenging at large dispersive shifts. With independent experiments on a different sample not presented in this work, we have confirmed the validity of the ECD approach for $\chi/2\pi \sim \SI{200}{kHz}$. To realize faster gates using a larger dispersive shift on the order of $\chi/2\pi \gtrsim \SI{1}{MHz}$, the gate set could be modified to take the partially selective nature of the finite-bandwidth qubit rotations into account, or GRAPE based techniques could be incorporated.
\section{Characteristic function tomography}
\label{sec:tomography}
\subsection{Measurement and Post-Processing}
The tomographic sequence used to measure the characteristic function after each ECD sequence is shown in \cref{fig:tomography a}. As shown in previous works (\cite{fluhmannDirectCharacteristicFunctionTomography2020, fluhmannEncodingQubitTrappedion2019, campagne-ibarcqQuantumErrorCorrection2020}), the oscillator's characteristic function defined as $\mathcal{C}(\beta) = \text{Tr}\left(D(\beta)\rho\right)$ can be measured using a conditional displacement embedded within a qubit ramsey sequence resulting in $\braket{\sigma_x - i\sigma_y} = \braket{D(\beta)}$ before the second $\pi/2$ pulse. By varying the phase of the second $\pi/2$ pulse, we can measure either the real or imaginary part of $\mathcal{C}(\beta)$ by measuring either $\braket{\sigma_x}$ or $\braket{\sigma_y}$ respectively. We also include a first measurement $m_1$ to disentangle the qubit and oscillator before the tomography, and postselect the results of the characteristic function ($m_2$) on $m_1 = \ket{g}$. We note that the pulses in this work are designed to realize state preparation sequences of the form $\ket{0}\ket{g} \rightarrow \ket{\psi}\ket{g}$ which disentangle the oscillator and qubit after the pulse. However, due to decoherence during the pulse, there is small residual entanglement between the oscillator and qubit, hence the need for $m_1$, with probabilities of $m_1 = \ket{g}$ given in the main text.
\begin{figure*}[h]
\begin{center}
\vspace{-2\baselineskip
\phantomsubfloat{\label{fig:tomography a}}
\phantomsubfloat{\label{fig:tomography b}}
\phantomsubfloat{\label{fig:tomography c}}
\includegraphics{tomography}
\caption{\label{fig:tomography}\textbf{Tomography} \textbf{a.} Characteristic function measurement sequence \textbf{b.} Measured qubit phase accumulation (open circles) after the \textit{cat-and-back} sequence (depicted in the inset). Data matches the phase predicted by \cref{eq:sol end} (solid line). \textbf{c.} Example of characteristic function post-processing using the measured GKP $\ket{+Z}$ state.}
\end{center}
\end{figure*}
To simplify tomography, the ECD gate used to measure the characteristic function is realized using the optimization method in \cref{sec:ECD optimization} for $\beta=1.0$ using $\alpha_0=2.0$. The amplitude and phase of this ECD gate is swept in experiment to sample the characteristic function. This low $\alpha_0$ is chosen so the second-order dispersive shift and Kerr have a negligible effect on the tomography, and the applied gate is well described by \cref{eq:cd_unitary}. From this, the additional qubit phase $\theta'$ can be easily predicted and corrected in post-processing.
To verify the validity of \cref{eq:sol start} in predicting this phase, we measure the phase $\theta'$ by using a \textit{cat-and-back} ECD sequence depicted in the inset of \cref{fig:tomography b}: We prepare the qubit in $\ket{+x} = \frac{1}{\sqrt{2}}\left(\ket{g} + \ket{e}\right)$, then apply two conditional displacements: $\text{ECD}(\beta), \text{ECD}(-\beta)$, with a qubit $\pi$ pulse between, after which the qubit phase is measured (by measuring $\braket{\sigma_x}$ and $\braket{\sigma_y}$). Here, the $\text{ECD}$ gates are the same as those used for characteristic function tomography ($\alpha_0 = 2$ at $\beta=1$, and amplitude is swept). The resulting phase $\theta'$ in \cref{fig:tomography b} shows excellent agreement with the prediction from \cref{eq:sol start}.
With this, the post-processing of the tomographic $\braket{\sigma_x}$ and $\braket{\sigma_y}$ is depicted in \cref{fig:tomography c}, using the measured data for the $\ket{+Z}_\text{GKP}$ state as an example. In experiment, we alternate between $\pm \pi/2$ for the first $\pi/2$ pulse to symmetrize the transmon's $T_1$ error channel during readout. Since the characteristic function obeys the property $\mathcal{C} (-\beta) = \mathcal{C}^*(\beta)$, we only measure half of the real and imaginary parts, then mirror about the $\text{Re}(\beta) = 0$ axis. The characteristic function is found by applying a phase correction $\mathcal{C}(\beta) \rightarrow e^{i|\beta|^2\theta'_0}\mathcal{C}(\beta)$ where $\theta'_0$ is the phase predicted by \cref{eq:sol start} for $\beta = 1$ using the ECD pulse optimized for tomography ($\alpha_0 = 2$). We note that this phase is slightly different than the phase found using the cat-and-back experiment described above, which is only verifying the validity of \cref{eq:sol start}. Finally, the data is scaled such that $\mathcal{C}(0) = 1$, effectively accounting for qubit decoherence ($T_{2E}$) during the tomography.
\subsection{Density matrix reconstruction}
\label{sec:density matrix recon}
To estimate the fidelity and purity of the oscillator state in experiment, we employ density matrix reconstruction using maximum likelihood estimation. For this, we use the measured real and imaginary parts of the characteristic functions, taken with 1,280 averages per point and use a numerical, iterative, convex optimization algorithm. For all demonstrations, the measured imaginary parts are close to zero, as is expected for states with symmetric Wigner functions. Any small deviations away from zero in the imaginary part are captured by the reconstruction.
For the Fock states, binomial states, and GKP states, the reconstruction is done in the Fock basis. For the squeezed states, reconstruction is performed in the squeezed-Fock basis with a basis squeezing equal to the target squeezing. The reconstruction Hilbert space size is swept, and a Hilbert space trunction is chosen such that increasing or decreasing the truncation does not change the fidelity or purity within the quoted error bars. For the binomial,GKP, and squeezed states, some states display a small phase-space rotation in the tomography. For these, a small inverse rotation is applied to the reconstructed density matrix. The maximum change in fidelity from this rotation is $1(\%)$ for the $\ket{-Z}_\text{GKP}$ state.
\subsection{Effective squeezing measurement}
To find the effective squeezing of prepared states, the reconstructed density matrices are used to calculate the position quadrature probability distributions, $P(x) = \text{Tr}\left(\rho \ket{x}\bra{x}\right)$, with results shown in \cref{fig:squeezing a}, including the measured cavity equilibrium thermal state for comparison. Here, a small rotation is applied to the reconstructed density matrices before calculating the probability distributions to align the squeezed quadrature. Also shown is the purity ($p = \text{Tr}\left(\rho^2\right)$) of each reconstructed density matrix. These distributions are fit to Gaussian functions to extract the squeezing in each state quoted in the main text. We compare these results to a calculation of the Fisher information using the full reconstructed probability distributions,
\begin{align}
I_c = 2\int dx\, \left(\partial_x \log P(x)\right)^2 P(x).
\end{align}
The Fisher information is a measure of the ability to sense small displacements along the position quadrature using the state with respect to homodyne detection \cite{parisQuantumStateEstimation2004, hastrupUnconditionalPreparationSqueezed2021}. For ideal squeezed states, $I_{c} = 2/\braket{\Delta x^2}$. Although the calculated purity of the squeezed states decreases with larger target squeezing, the states can still be used to sense small displacements, since only the probability distribution $P(x) = \text{Tr}\left(\rho \ket{x}\bra{x}\right)$ enters into the Fisher information. In \cref{fig:squeezing b}, we compare the calculated $I_C$ for each state to the target value. Finally, we also compare to $I_c = 2/\sigma_x^2$, where $\sigma_x^2$ is the variance of the Gaussian fit, indicating that the squeezed quadrature distributions are well approximated by a Gaussian, even though the full tomographies show some non-Gaussian features.
\begin{figure*}[!h]
\begin{center}
\vspace{-2\baselineskip
\phantomsubfloat{\label{fig:squeezing a}}
\phantomsubfloat{\label{fig:squeezing b}}
\includegraphics{squeezing}
\caption{\label{fig:squeezing} \textbf{Squeezed state distributions} \textbf{a.} Reconstructed position quadrature probability distributions and squeezing values found by a fit to Gaussian functions. Also shown are purities calculated as $p = \text{Tr} (\rho^2)$. \textbf{b.} Calculated Fisher information from the probability distributions (orange), compared to that expected from a Gaussian fit (blue).}
\end{center}
\end{figure*}
\subsection{Binomial code analysis}
The binomial \textit{kitten code} is the smallest binomial code for which the loss of a single photon is correctable \cite{michaelNewClassQuantum2016,huQuantumErrorCorrection2019, grimsmoQuantumComputingRotationsymmetric2020, gertlerProtectingBosonicQubit2021,burkhartErrorDetectedStateTransfer2021}, with logical states given by $\ket{+Z} = \left(\ket{0} + \ket{4}\right) / \sqrt{2}$ and $\ket{-Z} = \ket{2}$. In \cref{tab:Binomial expectation values} we give the estimated expectation values of the logical Pauli operators of the prepared states found numerically using the reconstructed density matrices. We also quantify how errors in the prepared states could be corrected by ideal error correction, as some errors during the state preparation are in principle correctable. In particular, the correctable code space includes the normalized error states $\ket{+Z}_\text{e} = a\ket{+Z}/ |\braket{+Z|a^\dagger a |+Z}|^2 = \ket{3}$ and $\ket{-Z}_\text{e} = a\ket{-Z}/ |\braket{-Z|a^\dagger a |-Z}|^2 = \ket{1}$. From these, we can define the logical operators corresponding to applying ideal error correction then performing a logical measurement, given by
\begin{equation}
\begin{aligned}
&X_c = \ket{+Z}\bra{-Z} + \ket{+Z}_\text{e}\bra{-Z}_\text{e} + \ket{-Z}\bra{+Z} + \ket{-Z}_e\bra{+Z}_e \\
&Y_c = i\ket{-Z}\bra{+Z} + i\ket{-Z}_e\bra{+Z}_e - i\ket{+Z}\bra{-Z} - i\ket{+Z}_e\bra{-Z}_e\\
&Z_c = \ket{+Z}\bra{+Z} + \ket{+Z}_\text{e}\bra{+Z}_\text{e} - \ket{-Z}\bra{-Z} - \ket{-Z}_e\bra{-Z}_e \\
&I_c = \ket{+Z}\bra{+Z} + \ket{+Z}_\text{e}\bra{+Z}_\text{e} + \ket{-Z}\bra{-Z} + \ket{-Z}_e\bra{-Z}_e.
\end{aligned}
\end{equation}
The expectation values of these operators quantify the logical information encoded in the prepared states after ideal error correction assuming a photon loss error channel. We calculate these expectation values using the reconstructed density matrices with results shown in \cref{tab:Binomial expectation values}. If these prepared states were to be used in an error correction setting, the error decoding model should instead be adapted to fit the actual errors encountered during state preparation.
\begin{table}[!h]
\begin{tabular}{c|cccc|cccc}
State& $\braket{I}$ & $\braket{X}$ & $\braket{Y} $ & $\braket{Z}$ & $\braket{I_c}$ & $\braket{X_c}$ & $\braket{Y_c}$ & $\braket{Z_c}$\\
\hline
$\ket{+X}_\text{bin}$ & 0.90 & 0.98 & 0.08 & -0.05 & 0.95 & 0.92 & 0.09 & -0.08 \\
$\ket{+Y}_\text{bin}$ & 0.92 & -0.06 & 0.90 & -0.10 & 0.98 & -0.06 & 0.96 & -0.13 \\
$\ket{+Z}_\text{bin}$ & 0.92 & 0.03 & -0.05 & 0.92 & 0.97 & 0.04 & -0.04 & 0.96\\
$\ket{-Z}_\text{bin}$ & 0.93 & -0.01 & -0.10 & -0.93 & 0.99 & -0.02 & -0.10 & -0.98\\
\end{tabular}
\caption{\label{tab:Binomial expectation values} \textbf{Binomial code Pauli expectation values} found numerically using the reconstructed density matrices.}
\end{table}
\subsection{GKP code analysis}
\label{sec:GKP state analysis}
The finite energy square GKP code stabilizers and logical Pauli operators are defined as \cite{gottesmanEncodingQubitOscillator2001,menicucciFaultTolerantMeasurementBasedQuantum2014,nohQuantumCapacityBounds2019,royerStabilizationFiniteEnergyGottesmanKitaevPreskill2020,terhalScalableBosonicQuantum2020,grimsmoQuantumErrorCorrection2021}
\begin{align}
&S_{q,\Delta} = e^{-\Delta^2 a^\dagger a} D\left(i\sqrt{2\pi}\right) e^{\Delta^2 a^\dagger a} &S_{p,\Delta} = e^{-\Delta^2 a^\dagger a} D\left(\sqrt{2\pi}\right) e^{\Delta^2 a^\dagger a} \\
&X_{\Delta} = e^{-\Delta^2 a^\dagger a} D\left(\sqrt{\pi/2}\right) e^{\Delta^2 a^\dagger a} &Z_{\Delta} = e^{-\Delta^2 a^\dagger a} D\left(i\sqrt{\pi/2}\right) e^{\Delta^2 a^\dagger a}
\end{align}
with $Y_\Delta = iX_\Delta Z_\Delta $. The target GKP states are found numerically by letting $\ket{+Z_\Delta}$ be the groundstate of a fictitious Hamiltonian $H = -S_{q,\Delta} - S_{p,\Delta} - Z_\Delta$, then by applying the appropriate finite energy logical operators and normalizing. Here, we use a target state squeezing of $\Delta_\text{target} = 0.306$ (10.3 dB). To estimate the effective squeezing of the prepared GKP states in experiment, we find the value of $\Delta$ that maximizes the expectation value of the projector onto the finite-energy code space $P_\Delta = \ket{+Z_\Delta}\bra{+Z_\Delta} + \ket{-Z_\Delta}\bra{-Z_\Delta}$ using the reconstructed density matrices. We find a squeezing of $\Delta_\text{exp} = 0.35$ (9.1 dB) for all prepared states within the precision of the reconstruction. In addition to the fidelities given in the main text, we quantify the quality of the prepared states here by numerically estimating the expectation values of the finite-energy Pauli operators and stabilizers with results given in \cref{tab:GKP expectation values}.
In addition, we quantify the prepared GKP states by the logical expectation values that would result from an ideal homodyne detection. In \cref{fig:GKP marginals}, we plot the reconstructed marginal probability distributions for the prepared GKP states along the generalized quadrature coordinate $x_\theta = \left(e^{i\theta }a^\dagger + e^{-i\theta}a\right)/\sqrt{2}$ with $\theta = \left\{0,\pi/2, \pi/4\right\}$ ($x_0 = x$ and $x_{\frac{\pi}{2}} = p$). From these, we can define the corresponding homodyne expectation values ($X_H$, $Y_H$, and $Z_H$) resulting from integrating the probability distributions and assigning a logical value. These are given by the total probability of finding a homodyne measurement result in sectors closest to $x_0 \mod 2\sqrt{\pi} = 0$ or $x_0 \mod 2\sqrt{\pi} = \sqrt{\pi}$ for $Z_H = \pm 1$, with analogous definitions for $X_H$ and $Y_H$ \cite{gottesmanEncodingQubitOscillator2001}. The results are given in \cref{tab:GKP expectation values}.
\begin{figure*}[ht]
\begin{center}
\includegraphics{GKP_marginals}
\caption{\label{fig:GKP marginals} \textbf{Reconstructed GKP state marginal probability distributions.} Blue and white bins represent the integration zones corresponding to assigning the associated logical Pauli operator $\pm 1$ respectively.}
\end{center}
\end{figure*}
\begin{table}[!h]
\begin{tabular}{c|ccccc|ccc}
State& $\text{Re}\left(\braket{S_{q,\Delta}}\right)$ & $\text{Re}\left(\braket{S_{p,\Delta}}\right)$ & $\text{Re}\left(\braket{X_\Delta}\right)$ & $\text{Re}\left(\braket{Y_\Delta}\right) $ & $\text{Re}\left(\braket{Z_\Delta}\right)$ & $\braket{X_\text{H}}$ & $\braket{Y_\text{H}}$ & $\braket{Z_\text{H}}$\\
\hline
$\ket{+Z}_\text{GKP}$ & 0.75 & 0.88 & 0.01 & -0.02 & 0.94 & 0.00 & -0.01 & 0.94\\
$\ket{+Y}_\text{GKP}$ & 0.78 & 0.78 &0.02&0.87&0.05 &0.01 & 0.87 & 0.06 \\
$\ket{-Z}_\text{GKP}$ & 0.81 & 0.71 &0.02 & 0.03 & -0.85 & 0.02 & 0.02 & -0.91 \\
\end{tabular}
\caption{\label{tab:GKP expectation values} \textbf{GKP code stabilizer and Pauli expectation values}. For the finite energy stabilizers and Pauli operators, we use $\Delta = \Delta_\text{exp} = 0.35$.}
\end{table}
\section{Sources of Infidelity}
In this section, we use simulations to estimate the sources of infidelity for the Fock state, binomial state, and GKP state preparation examples shown in the main text.
\label{sec:error budget}
\subsection{Decoherence-Free Error Budget}
First, we estimate how accurately our pulse compilation procedure, described in \cref{sec:ECD optimization}, can realize ideal ECD control unitaries, $U_\text{ECD}$ in \cref{eq:ECD unitary}, especially in the presence of Kerr and the second-order dispersive shift. In the open red circles of \cref{fig:lossless error budget}, we show simulated oscillator state preparation fidelities of the ideal ECD unitaries, defined as $\mathcal{F}_g = |\braket{\psi_\text{target}|\psi_g}|^2$
where $\ket{\psi_g}$ is the oscillator's state after postselecting the qubit in $\ket{g}$.
Next, we use the pulse compilation procedure described in \cref{sec:ECD optimization} with our system parameters, except we set $K_s = 0$ and $\chi' = 0$ to construct oscillator and qubit pulses which realize the ECD sequences without these higher order nonlinearities.
These pulses are simulated with $K_s = 0$ and $\chi' = 0$, and the resulting fidelities are shown by the blue triangles in \cref{fig:lossless error budget}.
These fidelites are close to the ideal unitary fidelities, demonstrating our ability to realize ECD sequences with the ideal dispersive Hamiltonian. These pulses are then simulated using \cref{eq:displaced H} with all nonlinear terms included using the measured system parameters, and the results are shown by the green crosses in \cref{fig:lossless error budget}. These infidelities are significantly higher, demonstrating the need to account for higher order nonlinaerities in the pulse construction.
Finally, we include the measured values of $K_c$ and $\chi'$ in the pulse construction, following the procedure outlined in \cref{sec:ECD optimization} which corrects for the linear contributions of Kerr and the second-order dispersive shift in the displaced frame. The resulting pulses are simulated with the full Hamiltonian, resulting in the infidelities given by purple diamonds in \cref{fig:lossless error budget}. These infidelities are close to the pulses optimized and simulated with $K=0$ and $\chi'=0$, indicating that correcting for the linear contributions of these terms is enough to significantly reduce their deleterious impact on the overall gate fidelity.
\begin{figure*}[!ht]
\begin{center}
\includegraphics{error_budget_lossless}
\caption{\label{fig:lossless error budget} \textbf{Decoherence free error budget} Simulated fidelity of constructed ECD sequences under different conditions. Red circles: result from unitary ECD parameter optimization. Blue triangles: pulses constructed and simulated with $K = \chi' = 0$. Green crosses: pulses optimized with $K = \chi' = 0$ and simulated with measured $K$ and $\chi'$ (nonzero). Purple diamonds: pulses optimized and simulated with measured $K$ and $\chi'$ (nonzero)f.}
\end{center}
\end{figure*}
\subsection{Impact of decoherence}
Next, we study the impact of transmon and cavity decoherence on the state preparation fidelity. In particular, we perform master equation simulations in the time-dependent displaced frame. Under a unitary frame transformation $\rho \rightarrow \tilde{\rho} = U\rho U^\dagger$, the master equation $\partial_t \rho = -(i/\hbar) \left[H,\rho\right] + \sum_i \mathcal{D}[L_i]\rho$ becomes $\partial_t \tilde{\rho} = -(i/\hbar) \left[\tilde{H},\rho\right] + \sum_i \mathcal{D}[U L_i U^\dagger]\tilde{\rho}$ with $\tilde{H} = UH U^\dagger + i\hbar \left(\partial_t U\right)$. Using the time-dependent displaced frame in \cref{sec:trajectories}, we evolve the joint transmon-cavity density matrix according to
\begin{equation}
\begin{aligned}
\label{eq:master equation}
&\partial_t \tilde{\rho} = -\frac{i}{\hbar}\left[\tilde{H}(t), \tilde{\rho}\right] + \gamma_\downarrow \mathcal{D}[q]\tilde{\rho} + \gamma_\uparrow \mathcal{D}[q^\dagger] \tilde{\rho} + 2\gamma_\phi \mathcal{D}[q^\dagger q]\tilde{\rho} \\
&+ \kappa_\downarrow \mathcal{D}[a]\tilde{\rho} + \kappa_\uparrow \mathcal{D}[a^\dagger + \alpha^*(t)]\tilde{\rho} + 2 \kappa_\phi \mathcal{D}[(a^\dagger + \alpha^*(t))(a + \alpha(t))]\tilde{\rho}
\end{aligned}
\end{equation}
where $\tilde{H}(t)$ is the displaced frame Hamiltonian in \cref{eq:displaced H} and $\alpha(t)$ is the nonlinear response to the drive given by solving \cref{eq:cm trajectory}. By simulating in the displaced frame which tracks the classical trajectory of the state's center-of-probability in phase space, we reduce the Hilbert space truncation required to accurately capture the dynamics. This is especially important considering our pulses drive the oscillator to photon numbers of $\sim 10^3$.
\begin{figure*}[!ht]
\begin{center}
\includegraphics{error_budget}
\caption{\label{fig:error budget} \textbf{Measured fidelities, purities, and simulated error budget} For each state preparation example, black dots indicate infidelities found from maximum likelihood reconstruction using the measured characteristic functions after postselecting the transmon in $\ket{g}$. Also shown is the purity of the reconstructed states $p = \text{Tr}\left(\rho^2\right)$. For each state, we include the baseline infidelity (labeled \textit{decoherence-free}) which is equivalent to the purple diamonds in \cref{fig:lossless error budget}. On top of this baseline, contributions to the total simulated infidelity using the measured decoherence rates are shown by the colored bars. Here, \textit{other} (bars not visible) includes contributions from transmon dephasing, transmon heating, and cavity heating at the quoted rates. Also included is simulated contribution due to intrinsic cavity dephasing at a rate $\kappa_\phi/2\pi = \left(\SI{150}{ms}\right)^{-1} \approx \SI{1}{Hz}$.}
\end{center}
\end{figure*}
We first simulate the impact of transmon relaxation, heating, and dephasing, as well as cavity relaxation and heating, given the measured rates in \cref{tab:system parameters}. When simulating transmon relaxation, we use the averaged measured relaxation rate with $\braket{n}_\text{cav} = 900$ photons in the cavity $\tilde{T}_{1,q}\approx \SI{30}{us}$, however we note this changes with time as shown in \cref{fig:out and back}, and the contribution from this error channel is also expected to change depending on the particular $\tilde{T}_{1,q}$ at the time of averaging. In addition, we approximate the qubit dephasing rate to be $\gamma_\phi = \gamma_{2E} - \gamma_1/2$, which assumes the echoing of low-frequency noise during the ECD gates results in an effective white noise dephasing.
The results are summarized by the colored bars in \cref{fig:error budget}, where each contribution to the error budget is simulated by only including a single decoherence mechanism. Also included is infidelity associated to an initial cavity thermal state ($n_\text{th} = 0.025$) since we do not employ active cavity cooling before each experiment. Independent simulations have verified that adding the infidelities of individual error channels is a good predictor of the total infidelity when simulating with all error channels combined. Out of these error channels, transmon relaxation has the biggest overall impact on the infidelity as ancilla relaxation during the conditional displacements can result in large oscillator displacements \cite{campagne-ibarcqQuantumErrorCorrection2020}.
\subsection{Discussion}
As shown in \cref{fig:error budget}, these decoherence mechanisms alone under-predict the infidelity found in experiment. Possible additional sources of infidelity include unknown transfer functions \cite{gustavssonImprovingQuantumGate2013,rolRestlessTuneupHighFidelity2017a,jergerSituCharacterizationQubit2019}, drifts in parameters, additional cavity heating due to the strong drives, and cavity dephasing. Out of these mechanisms, we simulate the effect of cavity dephasing using the displaced-frame dephasing term in \cref{eq:master equation}.
In this experiment, we do not have a direct measurement of the small intrinsic cavity dephasing rate, and such a measurement is an ongoing topic of investigation. We instead use master equation simulations to study the impact of cavity dephasing. A previous work on the same physical sample has also used simulations to bound the dephasing rate to $\kappa_\phi \lesssim \SI{1}{Hz}$ \cite{campagne-ibarcqQuantumErrorCorrection2020} and here we find a similar result. In \cref{fig:error budget} we include a contribution to the infidelity by simulating the pulses with a cavity dephasing rate of $\kappa_\phi/2\pi = \left(\SI{150}{ms}\right)^{-1} \approx \SI{1}{Hz}$, shown by the light grey bars. For some state preparation experiments, such as Fock $\ket{5}$ and GKP $\ket{+Z}$, this rate roughly matches the measured infidelity. This is evidence that the dephasing rate is likely smaller than $\SI{1}{Hz}$, matching the result from \cite{campagne-ibarcqQuantumErrorCorrection2020}. However, for other experiments, the error bars still under predict the fidelity, indicating other unknown mechanisms.
\section{Optimization of ECD circuit parameters}
\label{sec:numerical optimization}
For state preparation using ECD control, the quantum control problem we aim to solve is
\begin{align}
\label{eq:ECD unitary}
&U_\text{ECD} = D(\beta_{N+1}/2)R_{\varphi_{N+1}}\left(\theta_{N+1}\right)\text{ECD}(\beta_N)R_{\varphi_N}\left(\theta_N\right)...\text{ECD}(\beta_2)R_{\varphi_2}\left(\theta_2\right)\text{ECD}(\beta_1)R_{\varphi_1}\left(\theta_1\right) \\
\label{eq:full fidelty}
&\mathcal{F} = |\braket{\psi_t|U_\text{ECD}|\psi_i}|^2\\
&\left\{\vec{\beta}, \vec{\varphi}, \vec{\theta} \right\} = \argmax_{\left\{\vec{\beta}, \vec{\varphi}, \vec{\theta} \right\}} (\mathcal{F})
\end{align}
with initial state $\ket{\psi_i}$ and target state $\ket{\psi_t}$.
The circuit depth $N$ should be chosen such that $\mathcal{F}$ at its maximum is above an acceptable value with experimental considerations in mind. Although we focus on state preparation in this work, the optimization method described can also be used to realize a general unitary $U_\text{target}$ by replacing the Fidelity function with $\mathcal{F} = \left| \frac{1}{\text{Tr}(P)} \text{Tr}\left(PU_\text{target}^\dag U_\text{ECD}\right)\right|^2$, where $P$ is a projector onto a possible subspace of interest \cite{ballSoftwareToolsQuantum2021}. In all protocols presented in this work we include a final qubit rotation $R_{\varphi_{N+1}}\left(\theta_{N+1}\right)$ and displacement $D(\beta_{N+1}/2)$ after the last ECD gate in the optimizer. Often, the optimizer converges to protocols with $\beta_{N+1} = 0$. These gates are implemented quickly with respect to typical ECD gates and are not included in the quoted circuit depths $N$ which only counts the total number of ECD gates.
Here, we realize a multi-start method to solve this non-convex problem by optimizing $B$ independent variations of $U_\text{ECD}$ in a parallel manner.
Denoting the $j^{\text{th}}$ variation of the ECD unitary as $U_{\text{ECD},j}$ with circuit parameters $\left\{\vec{\beta}_j, \vec{\varphi}_j, \vec{\theta}_j \right\}$ and fidelity $\mathcal{F}_j = |\braket{\psi_t|U_{\text{ECD},j}|\psi_i}|^2$, we perform gradient descent on all $4BN$ real parameters using the total cost function $\text{C} = \sum_{j = 1}^B \log ( 1 - \mathcal{F}_j)$. Since the cost function is a simple sum of independent logarithmic cost functions, gradient descent of $\text{C}$ realizes independent gradient descent of each circuit realization in parallel. The optimization is stopped when any $\mathcal{F}_j$ reaches the target fidelity, and the parameters from the $j^\text{th}$ circuit are selected.
We realize gradient descent of the cost function using \textit{Adam} \cite{kingmaAdamMethodStochastic2017} implemented in TensorFlow. To construct this cost function and its gradient, we represent the batch of circuit parameters by the tensors $\bm{\beta}$, $\bm{\varphi}$ and $\bm{\theta}$ of dimensions $B \times N$ such that $\bm{\beta}_{ji}$ is the $i^\text{th}$ parameter appearing in circuit $j$.
Given a tensor-product structure of $\mathcal{H} = \mathbb{C}^2 \otimes \mathbb{C}^{N_o}$ where $N_o$ is the truncation of the oscillator's Hilbert space, the ECD unitaries and cost function are represented in the block-matrix form
\begin{align}
&U_{\text{ECD}, j} = \bm{b}_{jN}...\bm{b}_{j2}\bm{b}_{j1} \\
&\bm{b}_{ji} = \text{ECD}(\bm{\beta}_{ji})R_{\bm{\varphi}_{ji}}(\bm{\theta}_{ji}) =
\begin{pmatrix}
D^\dag(\bm{B}_{ji})e^{i\bm{\Phi}_{ji}}\sin\bm{\Theta}_{ji} & D^\dag(\bm{B}_{ji})\cos\bm{\Theta}_{ji} \\
D(\bm{B}_{ji})\cos\bm{\Theta}_{ji} & -D(\bm{B}_{ji})e^{-i\bm{\Phi}_{ji}}\sin\bm{\Theta}_{ji}
\end{pmatrix} \\
&\text{C} = \sum_{j=1}^{N} \log\left(1 -|\bra{\psi_t} \bm{b}_{jN}...\bm{b}_{j2}\bm{b}_{j1}\ket{\psi_i}|^2\right)
\end{align}
with reduced parameters $\bm{B} = \frac{\bm{\beta}}{2}$, $\bm{\Theta} = \frac{\bm{\theta}}{2}$ and $\bm{\Phi} = \bm{\varphi} - \frac{\pi}{2}$, and $D$ is the displacement operator defined on the oscillator's Hilbert space truncated to $N_o$. To construct each block operator $\bm{b}_{ji}$, we first compute the displacement operators. For this, we use the batched displacement operator implementation in \cite{sivakModelFreeQuantumControl2021}, which uses pre-diagonalization of the truncated position and momentum operators to efficiently construct the displacement matrices. With this, we compute all $B\times N$ displacement opertors $D(\bm{B})$ in parallel, then build the block matrices $b_{ji}$ block-by-block while reusing computed functions to minimize the total number of computations. Once these blocks are computed, the cost function is implemented by contracting along the $i$ index, taking the logarithm, then contracting along the $j$ index. To compute the gradient of the cost function with respect to $\bm{\beta}$, $\bm{\varphi}$, and $\bm{\theta}$, we use TensorFlow to realize reverse-accumulation automatic differentiation.
\begin{figure*}[!ht]
\begin{center}
\includegraphics{optimization}
\caption{\label{fig:optimization} \textbf{ECD circuit optimization} Examples of optimization curves for Fock 5 state preparation. Each line represents a circuit fidelity $F_j$, and 500 randomly initialized circuits are optimized in parallel. In the case of small circuit depth $N=3$ (left panel), the best infidelity reached is $\sim 0.3$. With a larger circuit depth $N=9$ (right panel), one circuit out of the 500 reaches an infidelity of $\sim 0.003$, demonstrating the need for multi-start optimization.}
\end{center}
\end{figure*}
In \cref{fig:optimization} we show an example by plotting the infidelities of each circuit in the batch as a function of optimization epoch for Fock $\ket{5}$ state preparation ($\ket{0}\ket{g} \rightarrow \ket{5}\ket{g}$) using two different circuit depths, $N=3$ and $N=9$. For these examples, we use a batch size of $B=500$ circuits, with each epoch representing $100$ steps of gradient descent using Adam with a learning rate of 0.001 carried out using an Nvidia Tesla v100 GPU.
In \cref{fig:GKP pulse}, we show an example of the pulse parameters found using this procedure for state preparation of the GKP $\ket{+Z}$ logical state. The magnitude of the echoed conditional displacements found in this example is typical of most demonstrations in this work - the largest $|\beta|$ found is $|\beta| \approx 2.75$. We note the optimization procedure does not include a constraint on $|\beta|$, but generally the scale of $|\beta|$s found is set by the phase-space extent of the target state.
As shown in this example, the optimizer often converges to pulses with interpretable values of qubit rotation angles and phases: values in that pulse are close to $\pi/4$, $\pi/2$, etc, and this is a common feature for the demonstrations in this work.
\section{Optimization of logical gates on a finite energy GKP code}
In this section, we focus on the implementation of logical gates using ECD control. In particular, we demonstrate numerical optimization of the phase and T gates for a finite-energy square GKP encoding.
For these gates, the target state maps $\left\{\ket{\psi_{i}}\right\} \rightarrow \left\{\ket{\psi_{t}}\right\}$ acting on the finite energy logical subspace are given by
\begin{equation}
\label{eq:GKP state maps}
\begin{aligned}
&S: \{\ket{+Z}_\Delta \ket{g},\ket{-Z}_\Delta\ket{g}\} \rightarrow \{\ket{+Z}_\Delta\ket{g},e^{i\pi/2}\ket{-Z}_\Delta\ket{g}\} \\
&T: \{\ket{+Z}_\Delta\ket{g},\ket{-Z}_\Delta\ket{g}\} \rightarrow \{\ket{+Z}_\Delta\ket{g},e^{i\pi/4}\ket{-Z}_\Delta\ket{g}\}
\end{aligned}
\end{equation}
where we have also included the condition that the ancilla qubit starts and ends in $\ket{g}$.
To optimize these logical gates, we modify the cost function in \cref{sec:ECD optimization} to be
\begin{align}
&C = -\sum_j \text{Re}\left(\bra{\psi_{t,j}} U_\text{ECD}\ket{\psi_{i,j}}\right)
\end{align}
where the sum is carried out over a logical state map, such as the state map for the $S$ and $T$ gates given in \cref{eq:GKP state maps}. Here, we only optimize the gate over the logical subspace, and in future work, the optimization could be modified to focus on error transparent gates \cite{maErrortransparentOperationsLogical2020}.
To quantify the quality of the optimized logical gates, we numerically calculate their average fidelity \cite{nielsenSimpleFormulaAverage2002}, defined as
\begin{align}
\overline{\mathcal{F}} = \frac{1}{6}\text{Tr}\left(R^T[\mathcal{U}_\text{target}] R[\mathcal{E}]\right) + \frac{1}{3}
\end{align}
where $R_{ij}[\mathcal{E}] = \frac{1}{2} \text{Tr}\left(\sigma_i \mathcal{E}\left[\sigma_j\right]\right)$ is the Pauli transfer matrix (PTM) associated to a quantum channel $\mathcal{E}$. Here, we define the finite energy logical Pauli operators using the numerically computed logical states as described in \cref{sec:GKP state analysis} to be $X_\Delta = \left(\ket{+Z}_\Delta \bra{-Z}_\Delta + \ket{-Z}_\Delta \bra{+Z}_\Delta \right)\ket{g}\bra{g}$ and analogous definitions for $I_\Delta$, $Y_\Delta$, and $Z_\Delta$. For these operations, the target unitary channel is defined as $\mathcal{U}_\text{target}\left[\rho\right] = U_\text{target} \rho U^\dagger_\text{target}$ and the applied channel is $\mathcal{E}\left[\rho\right] = U_\text{ECD} \rho U^\dagger_\text{ECD}$, where $U_\text{ECD}$ is the result of the optimization.
The optimization results for the finite energy $S$ and $T$ gates are shown in \cref{fig:GKP gates} at three different squeezing values $\Delta = \left\{ 0.25, 0.31, 0.35 \right\}$. Remarkably, these gates can be performed with a low circuit depth, only requiring $N=3$ ECD gates to reach a channel fidelity $\overline{\mathcal{F}} \approx 0.99$ for the T gate and $N=4$ for the S gate. These results indicate that the ECD gate set is especially well-suited for control over the GKP code, however we note that these gate implementations are not fault tolerant with respect to ancilla qubit errors.
\begin{figure*}[!ht]
\begin{center}
\includegraphics{GKP_gates}
\caption{\label{fig:GKP gates} \textbf{Optimized finite energy S and T average fidelity.} Plotted is the average fidelity for optimized ECD circuits at different values of $\Delta$ and circuit depth $N$. S gate results are shown with circles, T gate results are indicated with triangles, and the color indicates the squeezing. Also shown (insets) are the Pauli transfer matrices for the ECD S and T gates optimized at $N=3, \Delta=0.25$. Differences in Fidelity at different values of $\Delta$ are likely caused by the small finite overlap of the finite-energy logical states.}
\end{center}
\end{figure*}
\section{Optimization of GRAPE and SNAP pulses}
In this section, we outline numerical methods used to generate the GRAPE and SNAP points in Fig. 2b. For these pulses, we optimize the fidelity of the quantum control problem $\ket{0}\ket{g} \rightarrow \ket{n}\ket{g}$ for the Fock states $n = 1...5$.
\subsection{Optimization of GRAPE}
We use the methods described in \cite{heeresImplementingUniversalGate2017} to optimize GRAPE pulses for the cavity and qubit using our measured $\chi$. The driven dispersive Hamiltonian is used with piecewise-constant pulses sampled every $\SI{33}{ns}$ and an oscillator Hilbert space truncation of $N = 50$. To find the numerical quantum speed limit (QSL) associated with each Fock state preparation, we sweep the length of the pulse and pick the shortest pulse with an optimized Fidelity $\mathcal{F} > 0.99$. We also employ typical bandwidth and amplitude constraints when optimizing these pulses. We note that the cavity drive amplitudes used in our ECD gates are over an order-of-magnitude larger than typical cavity drive amplitudes used in optimized GRAPE pulses \cite{heeresImplementingUniversalGate2017}. Our results confirm that GRAPE pulses optimized in the usual way take a time greater than $2\pi/\chi$ as also observed in many state-of-the art bosonic control experiments \cite{ofekExtendingLifetimeQuantum2016a,heeresImplementingUniversalGate2017,axlineOndemandQuantumState2018,huQuantumErrorCorrection2019,wangHeisenberglimitedSinglemodeQuantum2019a,wangEfficientMultiphotonSampling2020a,gertlerProtectingBosonicQubit2021,burkhartErrorDetectedStateTransfer2021, curtisSingleshotNumberresolvedDetection2021a}.
\subsection{Optimization of SNAP}
The SNAP-displacement control sequence is parameterized as $D^\dagger(\alpha) {\rm SNAP}(\varphi) D(\alpha)$ as in \cite{foselEfficientCavityControl2020}. For each target Fock state $|n\rangle$ with $n=1...5$ and circuit depth $T=1...5$, we optimize the parameters of this control sequence with reinforcement learning \cite{sivakModelFreeQuantumControl2021}. For each configuration $(n, T)$, we repeat the training 10 times with different random initial seeds. The results show that to achieve fidelity $>99\%$, the circuit depth has to be $T\ge2$ for Fock states $n=1...3$, and $T\ge3$ for $n=4...5$. The best performing protocols are compared against ECD control in the main text. For SNAP times in the main text, we assume a gate time of $2\pi /\chi$, however SNAP is typically implemented with longer gate times \cite{heeresCavityStateManipulation2015}.
\section{Code Availability}
Code for the two-step optimization procedure - ECD parameter optimization (\cref{sec:ECD optimization}) and ECD pulse construction (\cref{sec:numerical optimization}) - is available at \url{https://github.com/alec-eickbusch/ECD_control}.
\putbib[ECD]
\end{bibunit}
\makeatother
\end{document} |
\section{Introduction}
The first flight of the Micro-X sounding rocket \cite{Wikus:2010} occurred in July 2018, marking the first time Transition Edge Sensors (TES) were operated in space~\cite{Adams:2019, Adams:2021}. A failure of the Attitude Control System (ACS) caused Micro-X to tumble in space for its full five minute observation. While no scientific data was obtained, the rotation of the rocket through Earth’s magnetic field offers a unique opportunity to study how TESs and their SQUID-based readout systems respond in space. We present a model to describe the system’s response to external magnetic fields based on the flown NIST MUX06a SQUIDs, a three-stage time-division multiplexing (TDM) system \cite{Stiehl:2011, Korte:2003}.
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=0.45\linewidth,keepaspectratio]{fig_squid_circuit.jpg}
\caption{Circuit Diagram for the Micro-X detectors and MUX06a SQUID readout. Each TES pixel is periodically read out by a SQ1, which is amplified by a SQ2 and SA unique to a column of 16 pixels. The four column-specific tuning parameters are the TES bias current, the SQ2 bias current, the SQ2 flux bias, and the SA flux bias. The SQ2 flux bias changes where the output of SQ1 maps onto the input of the SQ2 V-$\Phi$. The superconducting summing coil couples individual SQ1s to the column-specific SQ2, and is the primary point of magnetic susceptibility in the circuit. External magnetic fields that couple to the system act as an additional SQ2 flux bias. Republished from~\cite{Wikus:2009}.}
\label{fig:squid_circuit}
\end{center}
\end{figure}
The Micro-X TES array~\cite{Eckart:2009} is read out through a three-stage SQUID system in a flux-locked loop, shown in Fig.~\ref{fig:squid_circuit}. A characteristic property of SQUIDs is their periodic voltage response (V) to magnetic flux ($\Phi$), or their V-$\Phi$ relation, with a period equal to the magnetic flux quantum $\Phi_0$. The 128 TES pixels are divided into two independent science chains (X and Y), arranged in four columns (A--D) of 16 pixels each. Each pixel is read out by a first stage SQUID (SQ1) unique to the pixel, and then periodically read out by a second stage SQUID (SQ2) and a SQUID array (SA), both unique to the column. This SQUID system has several properties that the model aims to reflect. These include: asymmetric V-$\Phi$ relationships that have a steep, high-gain side, and a more gradual, low-gain side; SQ1s that are strongly coupled to the SQ2 input such that the high-gain domain of the SQ2 is fully covered by the output of SQ1; SAs that have an approximately linear response over the output range of the SQ2s; and a magnetic response that is dominated by the effective area of the SQ2 summing coil used to couple the SQ1s to the SQ2 in each column~\cite{Stiehl:2011}. Additionally, the system is tuned to utilize the high-gain side of the SQ2s.
\section{Three-Stage SQUID Model}
\label{sec:model}
To reflect the asymmetric V-$\Phi$ response we will model each SQUID stage as a truncated sawtooth wave summation, Eq.~(\ref{eqn:sawtooth}). Increasing the number of terms in the summation, n, increases how skewed the V-$\Phi$ will be. The input to each SQUID stage will be modeled in units of $\Phi_0$, and the output voltage will be normalized between $\pm1$. The three major model parameters are the number of SQ1 and SQ2 sawtooth terms, $n_1$ and $n_2$, and the $scale$ that maps the output of SQ1 onto the input of SQ2. The model parameters are fit to the real SQUIDs using composite V-$\Phi$ data. Typical values and an example of data fit to the model output are shown in Fig.~\ref{fig:squid_fit}.
\begin{equation}
\sum_{k=1}^{n} \frac{{2n \choose n-k}}{{2n \choose n}} \frac{\sin{kx}}{k}
\label{eqn:sawtooth}
\end{equation}
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=0.65\linewidth,keepaspectratio]{fig_squid_fit.jpg}
\caption{Model fit to a composite V-$\Phi$ for YC01, a pixel characteristic of the array, tuned on the high-gain size of SQ2. Fit parameters $n_1$=2, $n_2$=7, $scale$=1.2. (Color figure online.)}
\label{fig:squid_fit}
\end{center}
\end{figure}
The structure of the model is shown in Fig.~\ref{fig:3stage_model}. For a given flux input the SQUID response can be traced through a SQ1 V-$\Phi$ (\textit{bottom left}), the SQ2 \mbox{V-$\Phi$} (\textit{bottom right}), and the SA \mbox{V-$\Phi$} (\textit{top}) to produce the normalized voltage output. When operating in a flux-locked loop, the voltage output is used by the PI controller to adjust the flux input, which establishes a locked baseline value. By adjusting the relationship between the modeled SQUID stages we can simulate changing the SQUID tuning parameters described in Fig.~\ref{fig:squid_circuit}. The most important tuning parameter is the SQ2 flux bias, which determines where the output of SQ1 maps onto the input of SQ2. The large effective area of the SQ2 summing coil dominates the coupling of the SQUIDs to external magnetic fields. This means the effect of external magnetic fields on the SQUIDs can be modeled as a change in the SQ2 flux bias. DC offsets to the SQ2 flux input produced in this way effectively shift the SQ1 V-$\Phi$ up or down, represented by the various SQ1 V-$\Phi$s in Fig.~\ref{fig:3stage_model}. If the change in the SQ2 flux bias is sufficiently large, then there may not exist a stable SQ1 flux input for the flux-locked loop, as shown by the dotted SQ1 curve. When this happens the SQUIDs ``unlock" and cannot read out that TES. If the SQ2 flux bias continues to change, then the system may relock on the low-gain side of the SQ2. When this happens, the locked baseline will also jump to the low-gain side of the SQ1 to preserve the PI relationship between the SQ1 flux input and the normalized voltage output.
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=0.55\linewidth,keepaspectratio]{fig_3stage_model.jpg}
\caption{Model relationship between the three SQUID stages. SQ1 V-$\Phi$s tuned on the high-gain side of the SQ2 are shown in \textit{solid purple} and their corresponding locked baseline values in \textit{solid grey}. Those tuned on the low-gain side of the SQ2 are shown in \textit{dashed purple} and \textit{dashed grey}. A SQ1 without a stable lock point is shown in \textit{dotted purple}. The projection of a properly tuned SQ1 through the SQ2 and SA is highlighted in \textit{grey}. The model \textit{scale} parameter is shown where the SQ1 projects onto the SQ2 V-$\Phi$ \textit{bottom right}. (Color figure online.)}
\label{fig:3stage_model}
\end{center}
\end{figure}
Using this model, a composite V-$\Phi$ can be calculated for different SQ2 flux biases, along with any possible locked baseline values. Example composite V-$\Phi$ relationships are detailed in Fig.~\ref{fig:param_space}a, and the corresponding phase space in Fig.~\ref{fig:param_space}b. For a nominally tuned SQUID, as the SQ2 flux bias changes from zero the baseline follows the red curve through phase space with an initially linear response of 1.4e-2~$\Phi_{0}$~per~mG through the SQ2 summing loop, assuming a SQ2 effective area of 468.5~$\mu m^2$~\cite{Stiehl:2011}. This response increases after a 0.12~$\Phi_0$ baseline shift, and unlocks in the highlighted region after a 0.17~$\Phi_0$ baseline shift. Similar behaviour is expected when the SQUIDs relock on the low-gain side of the SQ2, with a few notable differences. The baseline will relock about 0.5~$\Phi_0$ away from where it unlocked as the PI loop forces the lock point to the low-gain side of SQ1. Additionally, the locked baseline response is greater and in the opposite direction on the SQ2 low-gain side as it is on the high-gain side. This is shown clearly in the projected locked baseline values in Fig.~\ref{fig:3stage_model}, where equal steps in SQ2 flux input produce small decreases in baseline on the high-gain side of the SQ2, and larger increases in baseline on the low-gain side. The linear baseline response on the SQ2 low-gain side is 2.3e-2~$\Phi_0$~per~mG throgh the SQ2 summing loop.
\begin{figure}[htbp]
\begin{center}
\begin{subfigure}
\centering
\includegraphics[width=0.4\linewidth,keepaspectratio]{fig_sq2_compound.jpg}
\end{subfigure}
\begin{subfigure}
\centering
\includegraphics[width=0.4\linewidth,keepaspectratio]{fig_sq2_lockpoints.jpg}
\end{subfigure}
\caption{(a), \textit{left}: Modeled composite V-$\Phi$ curves and locked baseline values at various SQ2 flux inputs. (b), \textit{Right}: Parameter space showing the relationship between SQ2 flux input and locked baseline values. Values highlighted in \textit{grey} have no stable lock points. Lock points from the example V-$\Phi$ curves (\textit{left}) marked in \textit{red circles}. (Color figure online.)}
\label{fig:param_space}
\end{center}
\end{figure}
\section{Observed Flight Effects}
This model can explain many of the anomalous system effects observed in flight due to the instrument tumbling in Earth's magnetic field.
\subsection{SQUID Unlocking}
During the five-minute observation period many of the SQUIDs unlocked for approximately two minutes. Additionally, the locked SQ1 baseline appears to track with the external magnetic field in the negative rocket x-direction. The baseline correlation with magnetic fields in this direction specifically is due to the intricacies of the detector's superconducting magnetic shielding, and will be discussed in more detail in an upcoming instrument paper reviewing the first flight of Micro-X. An example of this response is shown in Fig.~\ref{fig:obs_unlocking} and is explained by several features of the model. The marked pre-flight baseline value should be considered well tuned on the high-gain side of the SQ2, and any changes in the baseline due to Earth's magnetic field should be referenced from this point. The difference between the pre-flight baseline and the baseline at the start of observation is due to the different orientation of the rocket in Earth's magnetic field. As the locked baseline drifted during observation, the SQUIDs unlocked when the correlated magnetic field value is at a maximum. Additionally, the baseline response is observed to shift more rapidly before and after the SQUIDs unlocking - a clear prediction of the modeled SQUID relationships. The baseline response is observed to increase after drifting 0.12~$\Phi_0$, and unlock after drifting 0.16~$\Phi_0$. This is consistent with the model fit to these SQUIDs, and can be seen in Fig.~\ref{fig:param_space}b.
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=0.65\linewidth,keepaspectratio]{fig_obs_unlocking.jpg}
\caption{Example locked baseline response during observation (\textit{left axis}) and magnetic field in the negative rocket x-direction (\textit{right axis}). Time is referenced from launch. This SQUID was unlocked between 212~s and 298~s. The \textit{purple triangle} marks the pixel's baseline when it was well tuned before flight. (Color figure online.)}
\label{fig:obs_unlocking}
\end{center}
\end{figure}
\subsection{Baseline Oscillation in Powered Flight}
During powered flight, the rocket spins about the thrust (z) axis at approximately 4~Hz for stability. The external magnetic field in the rocket x-direction produces an oscillation on the locked feedback baseline as shown for pixels XB03 and YB13 in Fig.~\ref{fig:powered_flight}. Note that the detectors are tuned and operate at 75~mK, but are kept at 300~mK during powered flight. Therefore, the ADR magnet current and the SQ2 flux input are arbitrary at this time. Some pixels continued to read out normally (XB03), while others either unlocked or relocked on the low-gain side of their SQ2 (YB13). The locked baseline of pixel XB03 tracks the magnetic field in the positive rocket x-direction at 1.1e-4~$\Phi_{0}$~per~mG, while pixel YB13 tracks at 3.3e-4~$\Phi_{0}$~per~mG.\footnote{In addition to the baseline response to external magnetic fields inverting between the high-gain and low-gain sides of the SQ2, the response also inverts between detector sides X and Y. This was verified via post-flight testing with a Helmholtz coil, and is likely due to the spatial layout of the MUX06a chips and bending of the field inside the superconducting shielding. This will be discussed in more detail in an upcoming instrument paper reviewing the first flight of Micro-X.} The model predicts that the baseline response should be 1.6 times greater on the SQ2 low-gain side than on the high-gain side. The observed ratio of baseline responses is therefore higher than would be expected if XB03 and YB13 were in the linear regions of the SQ2 high- and low-gain side, respectively. However, all of column XB remained well behaved during powered flight and was most likely in the linear region, while column YB was in the non-linear region as evidenced by pixel YB00 in Fig.~\ref{fig:powered_flight}. A larger ratio of responses is therefore expected.
As the SQ2 flux bias changes, if a SQ1 is sufficiently coupled to the SQ2, it might not unlock and instead jump directly between the two sides of the SQ2. In doing so, the baseline response will change in magnitude and reverse direction, and the size of the baseline jumps will be approximately 0.5~$\Phi_0$ as discussed in Section~\ref{sec:model}. Both of these features are clearly visible in the baseline of pixel YB00.
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=0.65\linewidth,keepaspectratio]{fig_powered_flight.jpg}
\caption{Example locked baseline responses during powered flight (\textit{purple}) and magnetic field in the rocket x-direction (\textit{grey}). Pixel XB03 remained locked on the high-gain side of the SQ2, while pixel YB13 relocked on the low-gain side. Pixel YB00 jumps between lock points on the high-gain and low-gain sides of SQ2 as the rocket spins. Time is referenced from launch. (Color figure online.)}
\label{fig:powered_flight}
\end{center}
\end{figure}
\section{Conclusions}
Modeling the relationships between the three MUX06a SQUID stages helps explain many of the behaviours observed in the first Micro-X flight. The primary source of magnetic susceptibility in the system was the superconducting summing coil that connects SQ1 to SQ2, increasing the effective area of the SQ2 stage. Treating the external magnetic field as an additional SQ2 flux input models many of the effects observed in flight: the pixels can become unlocked if the SQ1 output is shifted to a state where no stable lock point exists; before pixels unlock, the locked baseline has an increased sensitivity to the external field; and pixels that lock on the low-gain side of SQ2 have an increased and inverted baseline response.
For the upcoming second Micro-X flight in 2022 the NIST MUX06a SQUIDs have been replaced with the less magnetically susceptible NIST MUX18b SQUIDs~\cite{Reintsema:2019}. See the upcoming instrument paper for a full review of the Micro-x first flight, and~\cite{Manzagol:2021} for details on progress towards the 2022 reflight. The datasets generated during and/or analyzed during the current study are available from the corresponding author on reasonable request.
\begin{acknowledgements}
This work was supported by NASA Grants 80NSSC20K0430 and 80NSSC21K1856. Part of this work was performed under the auspices of the U.S. Department of Energy by Lawrence Livermore National Laboratory under Contract DE-AC52-07NA27344.
\end{acknowledgements}
\pagebreak
|
\section{Introduction}
On the one hand, given a graph $G=(V,E)$, the problem of finding a subset of disjoint edges (matching) of maximum size is tractable, while on the other hand, its generalization to hyper-graphs, even when the edge length is $3$, is intractable (that is, it is NP-hard). The general problem is known as the \textsf{Hyper-graph Matching} problem. The hyper-graph $H=(W,F)$ can be equivalently viewed as a set system $({\mathcal U},{\mathcal S})$,
where the universe (or the ground set) ${\mathcal U}$ corresponds to the vertex set $W$ and ${\mathcal S}$ corresponds to the collection of hyperedges $F$. Then, finding a maximum matching in $H$ is equivalent to finding maximum number of pairwise disjoint sets (packing) in ${\mathcal S}$. Thus, the \textsf{Hyper-graph Matching} problem is also known as the \mbox{\sf Set Packing}\xspace problem, which is a fundamental problem in combinatorial optimization with numerous applications~\cite{vemuganti1998applications}. While this problem captures many classical combinatorial problems such as maximum independent set (or maximum clique), $k$-dimensional matching and also, some graph packing problems~\cite{chataigner2009approximation,hassin2006approximation}, this generalization also makes it intractable in several regimes.
One computational regime in which \mbox{\sf Set Packing}\xspace has been explored extenstively is approximation algorithms.
Since \mbox{\sf Set Packing}\xspace generalizes the maximum independent set problem~\cite{ausiello1980structure}, it inherits the inapproximability of the latter problem~\cite{10.5555/874062.875499}. This immediately implies that the trivial approximation of picking simply one set in the packing is roughly the best to hope for. Furthermore, approximations in terms of $|{\mathcal U}|$ are also not hopeful since the result also implies inapproximability bound of $|{\mathcal U}|^{1/2 -\epsilon}$, which is matched by~\cite{halldorsson2000independent}. To combat these intractabilities, various restrictions of \mbox{\sf Set Packing}\xspace have been studied. Particularly, a restriction where the size of the sets in ${\mathcal S}$ is bounded by some integer $k$, which is known as $k$\mbox{\sf -Set Packing}\xspace, is also a well-studied problem. However, $k$\mbox{\sf -Set Packing}\xspace captures the independent set problem in bounded degree graphs, which again is a notoriously hard problem to approximate beyond the ``trivial'' bound~\cite{austrin2009inapproximability,bansal2018lovasz}. While \cite{hazan2006complexity} improves the lower bound for $k$\mbox{\sf -Set Packing}\xspace to $\Omega(k/\ln k)$, the best known approximation is $(k+1+\epsilon)/3$ \cite{chan2012linear,cygan2013improved}, yielding a logarithmic gap between the bounds.
Besides approximation algorithms, \mbox{\sf Set Packing}\xspace has also been studied from the parameterized complexity perspectives (with the standard parameter on the size of an optimal packing solution).
This problem is known as the {\sf Parameterized Set Packing}\xspace or \mbox{\sf PSP}\xspace problem, in which, given an instance $({\mathcal U},{\mathcal S},r)$, we need to decide whether there exists a packing of size $r$. Unfortunately, even \mbox{\sf PSP}\xspace remains intractable and is, actually, \textsf{W[1]}\xspace-complete \cite{Downey:2012:PC:2464827}. In fact, \textit{Exponential Time Hypothesis} (ETH) implies that the trivial \textit{enumerative algorithm} running in $O^{*}(|{\mathcal S}|^r)$ time that finds an $r$-packing, is asymptotically our best hope~\cite{10.5555/2815661}. The algorithmic outlook for \mbox{\sf PSP}\xspace gets even worse due to \cite{chalermsook2017gap} which also rules out $o(r)$-\textsf{FPT}\xspace-approximation algorithm assuming the \textit{Gap Exponential Time Hypothesis} (Gap-ETH) \cite{manurangsi2016birthday,dinur2016mildly}.
Informally, it rules out non-trivial approximations for \mbox{\sf Set Packing}\xspace even when the optimal packing is small. Assuming a weaker hypothesis of $\textsf{FPT}\xspace = \textsf{W[1]}\xspace$, very recently ~\cite{s2021polynomial} showed that there is no \textsf{FPT}\xspace algorithm for \mbox{\sf PSP}\xspace problem that finds a packing of size $r/r^{1/H(r)}$, for any increasing function $H(\cdot)$, when given a promise that there is an $r$-packing in the instance.
Thus, the flurry of these negative results make it likely that \mbox{\sf Set Packing}\xspace problem is intractable in all computational regimes.
In this paper, we consider a restriction of \mbox{\sf Set Packing}\xspace in the parameterized setting.
We consider \mbox{\sf PSP}\xspace on \textit{compact} instances. We say that an instance $({\mathcal U}, {\mathcal S}, r)$ of $\mbox{\sf PSP}\xspace$ is compact if $|{\mathcal U}| = f(r) \cdot \Theta( \textsf{poly}\xspace(\log |{\mathcal S}|))$, for some function $f(r) \ge r$, that is, the universe is relatively small compared to the number of sets\footnote{There are in fact two ways to define compactness of instances: either compact ${\mathcal U}$ or compact ${\mathcal S}$. For compact ${\mathcal S}$, observe that when $|{\mathcal S}| = f(r) \cdot\Theta( \textsf{poly}\xspace(\log |{\mathcal U}|))$, then the \textit{enumerative algorithm} running in time $O^{*}(|{\mathcal S}|^r) = O^*(f(r)^r \log ^r |{\mathcal U}|)$ is already fixed parameter tractable~\cite{cai2001subexponential}. Thus, the interesting case is when the universe is compact, which is the case we will be focusing on.}.
Besides the algorithmic motivation, compact instances have recently been used as an ``intermediate step'' to prove FPT inapproximability results of the classical (non-compact) problems (see, e.g., \cite{bhattacharyya_et_al:LIPIcs:2016:6362,lin:LIPIcs:2019:10657} where the compact instances were used in proving FPT-inapproximability of the $k$-EvenSet and Dominating Set).
We hope that studying \mbox{\sf Compact \psp}\xspace would lead to some ideas that would be useful in proving total FPT inapproximability of \mbox{\sf PSP}\xspace (that is, to weaken the Gap-ETH assumption used in~\cite{chalermsook2017gap}).
%
\subsection{Our Results}
Our main result is the following dichotomy of {\sf Parameterized Set Packing}\xspace.
\begin{theorem}[Dichotomy] \label{thm:dich}
The following dichotomy holds for \mbox{\sf PSP}\xspace:
\begin{itemize}
\item[$\circ$] If $|{\mathcal U}| = f(r)\cdot o(\log |{\mathcal S}|)$, for any $f$, then \mbox{\sf PSP}\xspace is in \textsf{FPT}\xspace.
\item[$\circ$] \mbox{\sf PSP}\xspace remains {{\wone}-hard}\xspace even when $|{\mathcal U}| = r \cdot \Theta(\log |{\mathcal S}|)$.
\end{itemize}
\end{theorem}
The algorithmic result follows from well-known dynamic programming based algorithms~\cite{10.1137/070683933,10.5555/2815661} that run in time $O^*(2^{|{\mathcal U}|})$, and observing that the run time is fixed parameter tractable when $|{\mathcal U}| = f(r)o(\log |{\mathcal S}|)$. Parameterized tractability follows since $O(2^{O(f(r) \cdot s(n))})$ is fixed parameter tractable \cite{cai2001subexponential} for any $s(n)=o(\log n)$ which is non-decreasing and unbounded. However, for completeness, in Appendix~\ref{ss:dialgo} we present a simple dynamic program algorithm based on finding longest path in a DAG running in time $O^*(2^{|{\mathcal U}|})$.
The main contribution of our work is the \textsf{W[1]}\xspace-hardness of \mbox{\sf PSP}\xspace even when $|{\mathcal U}| = r \cdot \Theta(\log |{\mathcal S}|)$. Towards this, we show an \textsf{FPT}\xspace-reduction from a {{\wone}-hard}\xspace problem, called \mbox{\sf Subgraph Isomorphism}\xspace, to \mbox{\sf Compact \psp}\xspace. In fact, our hardness result can be strengthened assuming a stronger complexity assumption \cite{marx2007can,10.5555/2815661}.
\begin{theorem}\label{thm:enuhard}
\mbox{\sf Compact \psp}\xspace requires time $|{\mathcal S}|^{\Omega \left(r/\log r\right)}$ even when $|{\mathcal U}| = r \cdot \Theta(\log |{\mathcal S}|)$, unless ETH fails.
\end{theorem}
First, we want to remark that the universe size in the above theorem is tight (up-to $\log r$ factor) since having an ETH-hard instance with $|{\mathcal U}| = o(r/\log r) \cdot \Theta(\log |{\mathcal S}|)$ contradicts ETH.
The result of Theorem~\ref{thm:enuhard} follows from the hardness result of \mbox{\sf Subgraph Isomorphism}\xspace due to \cite{marx2007can}, and from the fact that the hardness reduction of Theorem~\ref{thm:dich} from \mbox{\sf Subgraph Isomorphism}\xspace is parameter preserving up to a multiplicative constant. Note that since \mbox{\sf PSP}\xspace can be trivially solved by enumeration in time $O^*\left(|S|^r\right)$, the above result says that, even for the compact instances this is essentially our best hope, up to a $\log$ factor in the exponent. Consequently, the dichotomy theorem coupled with Theorem~\ref{thm:enuhard} says that, as soon as instances get asymptotically smaller, not only can we beat the enumerative algorithm, but actually obtain an \textsf{FPT}\xspace algorithm.
\iffalse
Interestingly, our reduction from \mbox{\sf Subgraph Isomorphism}\xspace to \mbox{\sf Compact \psp}\xspace can be easily modified to obtain a reduction from $k$\mbox{\sf -Clique}\xspace to \mbox{\sf Compact \psp}\xspace with $r=\Theta(k^2)$.
Hence, combining this reduction with \textsf{FPT}\xspace-inapproximibility results~\cite{DBLP:conf/stoc/Lin21,lin2021lower,s2021polynomial} of $k$\mbox{\sf -Clique}\xspace imply the same \textsf{FPT}\xspace-inapproximibility results for \mbox{\sf Compact \psp}\xspace. The present best bound for $k$\mbox{\sf -Clique}\xspace, due to~\cite{s2021polynomial}, imply the following result for \mbox{\sf Compact \psp}\xspace.
\begin{theorem}
For any increasing computable function $H(r)\le r$, there is no \textsf{FPT}\xspace algorithm that finds packing of size $r/r^{1/H(r)}$ of the \mbox{\sf Compact \psp}\xspace problem even when $|{\mathcal U}| = r \cdot \Theta(\log |{\mathcal S}|)$, unless $\textsf{FPT}\xspace = \textsf{W[1]}\xspace$.
\end{theorem}
\fi
\paragraph{\textbf{Other results.}} \label{ss:other}
We note that our construction framework can be extended to other combinatorial problems.
First, we consider the {\sf Exact $r$-Covering}\xspace problem: given a set system $({\mathcal U},{\mathcal S})$ and an integer $r$, decide if there is a cover of size at most $r$ which is also a packing. \cite{10.1007/978-3-642-40450-4_57} show {\sf Compact \ecr}\xspace is {{\wone}-hard}\xspace, when $|{\mathcal U}| = r^2\cdot \Theta(\log |{\mathcal S}|)$. Our construction yields better universe size.
\begin{theorem} \label{thm:cecpw}
{\sf Compact \ecr}\xspace is {{\wone}-hard}\xspace, even when $|{\mathcal U}|=r\cdot \Theta(\log |{\mathcal S}|)$.
\end{theorem}
\iffalse
Assuming ETH, the result of [Jones at al.] imply that {\sf Compact \ecr}\xspace require time $|{\mathcal S}|^{\Omega(r/\log r)}$, when $|{\mathcal U}| = r^2 \cdot \Theta(\log |{\mathcal S}|)$. Our construction yields tighter universe size (up-to $\log r$ factor) since having an ETH-hard instance with $|{\mathcal U}| = o(r/\log r) \cdot \Theta(\log |{\mathcal S}|)$ contradicts ETH.
\begin{theorem} \label{thm:cecpeth}
{\sf Compact \ecr}\xspace require time $|{\mathcal S}|^{\Omega(r/\log r)}$ even when $|{\mathcal U}|=r\cdot \Theta(\log |{\mathcal S}|)$, unless ETH fails.
\end{theorem}
We obtain a similar dichotomy theorem for the {\sf Exact $r$-Covering}\xspace problem\footnote{{\sf Exact $r$-Covering}\xspace is known to be {{\wone}-hard}\xspace when $|{\mathcal U}| = \Theta \left( |{\mathcal S}| \right)$ from a reduction from the \textsf{Perfect Code} problem \cite{downey1995fixed, 10.5555/2568438}.}.
\textcolor{red}{already implied by Jones at al. So just talk about improving the universe for hardness.}
\begin{theorem} \label{thm:dichex}
The following dichotomy holds:
\begin{itemize}
\item[$\circ$] If $|{\mathcal U}| = f(r)\cdot o(\log |{\mathcal S}|)$, for any $f$, then {\sf Exact $r$-Covering}\xspace is in \textsf{FPT}\xspace.
\item[$\circ$] {\sf Exact $r$-Covering}\xspace remains {{\wone}-hard}\xspace even when $|{\mathcal U}| = r \cdot \Theta( \log |{\mathcal S}|)$.
\end{itemize}
\end{theorem}
\fi
Next we consider the $r$\mbox{\sf -VectorSum}\xspace problem. Given a collection ${\mathcal C}$ of $N$ vectors in $\mathbb{F}^d_2$, and a target vector $\vec{b} \in \mathbb{F}^d_2$, $r$\mbox{\sf -VectorSum}\xspace asks if there are at most $r$ vectors in ${\mathcal C}$ that sum to $\vec{b}$. The corresponding compact problem, \mbox{\sf Compact $r$-VectorSum}\xspace, is defined when $d = f(r) \cdot \Theta(\textsf{poly}\xspace(\log N))$, for some $f(r)\ge r$.
The present known hardness, due to \cite{bhattacharyya_et_al:LIPIcs:2016:6362}, shows that \mbox{\sf Compact $r$-VectorSum}\xspace requires time $N^{\Omega(\sqrt{r})}$ under ETH. We improve their result matching the trivial upper bound, up to a $\log$ factor in the exponent.
\begin{theorem}\label{thm:ckvs}
\mbox{\sf Compact $r$-VectorSum}\xspace requires time $N^{\Omega(r/\log r)}$, even when $d=r \cdot \Theta(\log N)$, unless ETH fails..
\end{theorem}
\subsection{Our contributions and comparison to existing works} \label{ss:comp}
In this section, we compare our contribution with existing works to highlight its significance.
To our best knowledge, the compact version of combinatorial problems has not previously been formalized and investigated. However, several existing reductions already imply the hardness of compact version. Here, we review and compare the related results.
\subsubsection{Our Contribution.}
As far as we know, there are no results showing \textsf{W[1]}\xspace-hardness \mbox{\sf Compact \psp}\xspace, and hence the corresponding dichotomy (Theorem~\ref{thm:dich}). A key contribution of this paper is to show such result for \mbox{\sf Compact \psp}\xspace. Additionally, our \textsf{W[1]}\xspace-hardness holds even when $|{\mathcal U}| = r \cdot \Theta(\log |S|)$. On the way, we also show an almost tight ETH-hardness result (up-to $\log r$ factor) for \mbox{\sf Compact \psp}\xspace, with tight (up-to $\log r$ factor) universe size of $|{\mathcal U}|= r\cdot \Theta(\log |{\mathcal S}|)$.
Interestingly, we show both of these results with a single \textsf{FPT}\xspace reduction. In addition, our framework can easily be extended to improve the hardness result of {\sf Compact \ecr}\xspace and \mbox{\sf Compact $r$-VectorSum}\xspace.
In the following points, we survey some known hardness results for {\sf Set $r$-Covering}\xspace in the compact regime. The reason for discussing these results is that, although being for {\sf Compact \scr}\xspace, they imply the hardness for {\sf Compact \ecr}\xspace as the intended cover in these instances is always a partition. The limitation of these reductions is that they can not be extended to \mbox{\sf PSP}\xspace. For example, if we consider the Soundness (No case) proof in these reductions, then we note that they rely crucially on the fact that no candidate solution can cover the entire universe. In fact, it is quite easy to find $r$ sets that are mutually disjoint but do not form a cover in the No case.
Therefore, these constructions do not correspond to \mbox{\sf Compact \psp}\xspace. The heart of our construction lies in ensuring that, in the No case, any $r$ sets intersect. To this end, we construct a combinatorial gadget called Compatible Intersecting Set System (ISS). This gadget is a pair of set systems $({\mathcal A},{\mathcal B})$ over a universe $U$ that guarantees two properties: First, every pair of sets within each set system intersects, and second, for any set $a \in A$, there exists $b\in B$ such that $a$ intersects every set in $B$ except $b$. Further, we present a simple greedy algorithm that finds such compatible ISS pair $({\mathcal A},{\mathcal B})$ over a universe of size $N$, each having roughly $2^{O(N)}$ sets. Note that this gadget, we use to build our compact hard instance, also has a ``compact" universe - in the sense - that the universe size is logarithmic in the number of sets in the set system.
\begin{enumerate}
\item \textbf{\cite{10.1007/978-3-642-40450-4_57}} shows that {\sf Compact \ecr}\xspace is {{\wone}-hard}\xspace\footnote{Actually, they show hardness for {\sf Compact \scr}\xspace. But a closer inspection of their construction shows that the intended set cover is also a packing.}. As mentioned before, the construction fails to produce an instance of \mbox{\sf PSP}\xspace.
Further, their hard instance has universe $|{\mathcal U}| = r^2 \cdot \Theta(\log |{\mathcal S}|)$, in contrast to Theorem~\ref{thm:cecpw}.
To achieve \textsf{W[1]}\xspace-hardness, they show an \textsf{FPT}\xspace reduction from Partitioned-\mbox{\sf Subgraph Isomorphism}\xspace.
The high level idea of the construction is similar to ours: First assign each vertex of $G$ a logarithm length binary vector. Then, create two kinds of sets: $V$-sets that capture the mapping of the vertices and $E$-sets that capture the mapping of edges. The idea is to use the pattern vector to create these sets so that there is an isomorphic copy of $G$ to $H$ if and only if there are $|V(H)|$ many $V$-sets and $|E(H)|$ many $E$-sets covering the universe.
\item \textbf{\cite{10.1145/3188745.3188896}} shows {\sf Compact \scr}\xspace is {{\wone}-hard}\xspace using a reduction from $k$\mbox{\sf -Clique}\xspace to {\sf Set $r$-Covering}\xspace with $r=\Theta(k^2)$ and $|{\mathcal U}| = r^{3/2} \cdot \Theta(\log |{\mathcal S}|)$. However, their reduction does not give tight ETH-hardness. In contrast, \cite{puatracscu2010possibility} shows tight ETH hardness for {\sf Compact \scr}\xspace, in the sense that they show that {\sf Compact \scr}\xspace requires time $|{\mathcal S}|^{\Omega\left(r\right)}$.
\end{enumerate}
\subsubsection{Gap-ETH and \mbox{\sf Compact \psp}\xspace.} We point out that, using a stronger assumption of Gap-ETH, it is easy to obtain a tight lower bound of $|{\mathcal S}|^{\Omega(r)}$ time for \mbox{\sf Compact \psp}\xspace, when $|{\mathcal U}|=r \cdot \Theta(\log |{\mathcal S}|)$. However, note that the assumptions in our results (Theorem~\ref{thm:dich} and Theorem~\ref{thm:enuhard}) are much weaker than Gap-ETH.
To this end, we start with the result of \cite{doi:10.1137/1.9781611975994.5} that shows \textsf{FPT}\xspace inapproximibility for Max $r$-Coverage, assuming Gap-ETH. Specifically, it shows that, given an instance ${\mathcal I} = ({\mathcal U},{\mathcal S},r)$ of Max $r$-Coverage, there is no $|{\mathcal S}|^{o(r)}$ time algorithm that decides if there exists $r$ sets that cover ${\mathcal U}$ or every $r$-sets cover at most $(1-1/e)|{\mathcal U}|$ elements.
We sketch the idea here. A crucial property of the construction (which was missing in the above reductions) is that, for any collection $T \subseteq {\mathcal S}$ of size $r$, $T$ covers ${\mathcal U}$ if and only if the sets in $T$ are disjoint. We construct an instance ${\mathcal I}' = ({\mathcal U}',{\mathcal S},r)$ of \mbox{\sf Compact \psp}\xspace as follows. Sample $100r \log |{\mathcal S}|$ elements from ${\mathcal U}$ and add them to ${\mathcal U}'$. Now, in the No case, since any $r$-sets $T$ of ${\mathcal S}$ cover at most $(1-1/e)|{\mathcal U}|$ elements, it follows that the probability that $T$ covers ${\mathcal U}'$ is at most $e^{-\frac{100r}{e} \log |{\mathcal S}|}$. Hence, taking union bound over all $r$-sets of ${\mathcal S}$, we have that, with probability at least $(1-n^{\Omega(-r)})$, no $r$-sets cover ${\mathcal U}'$.
\subsubsection{Other related results.}
\begin{itemize}
\item {{\sf Exact $r$-Covering}\xspace: } As mentioned before, the reduction of \cite{puatracscu2010possibility} for {\sf Compact \scr}\xspace, can be easily modified to achieve tight ETH hardness of {\sf Compact \ecr}\xspace. The idea is to start the reduction from \mbox{\sc $1$-in-$3$-SAT}\xspace, instead from \mbox{\sc $3$-SAT}\xspace, which produces an instance of {\sf Compact \ecr}\xspace.
\item $r$\mbox{\sf -VectorSum}\xspace: The reduction of \cite{bhattacharyya_et_al:LIPIcs:2016:6362} shows that \mbox{\sf Compact $r$-VectorSum}\xspace requires time $N^{\Omega(\sqrt{k})}$ under ETH.
\end{itemize}
\subsection{Overview of Techniques} \label{ss:tech}
In this section, we sketch the proof of the main results.
To prove the hardness part of our Theorem~\ref{thm:dich}, we present a reduction from \mbox{\sf Subgraph Isomorphism}\xspace, which asks, given a graph $G$ on $n$ vertices and another graph $H$ with $k$ edges, if there is a subgraph of $G$ isomorphic (not necessarily induced) to $H$, with parameter $k$. The reduction produces an instance ${\mathcal I} = ({\mathcal U},{\mathcal S},r)$ of \mbox{\sf PSP}\xspace in \textsf{FPT}\xspace time such that $r = \Theta(k)$ and $|{\mathcal U}| = \Theta(r \log |{\mathcal S}|)$.
We remark that the classical reduction given in \cite{Downey:2012:PC:2464827} also has parameter $r = \Theta(k)$, but $|{\mathcal U}|$ is linear in the size of $G$, which is the size of $|{\mathcal S}|$. Below, we present a reduction that is, in fact, incorrect but illustrates some of the main ideas.
Our reduction constructs the instance ${\mathcal I} = ({\mathcal U},{\mathcal S},r)$ of \mbox{\sf PSP}\xspace using a special set system gadget -- which we call the Intersecting Set System (ISS) gadget. A set system ${\mathcal A}=(U_A,S_A)$, with $M$ sets over $N$ elements, is called an $(M,N)$-Intersecting set system, if every pair $s^i, s^j \in S_A$ intersects. We show how to efficiently constructs an $(M,N)$-ISS ${\mathcal A}=(U_A,S_A)$ with $M > 2^{N/4}$, for even $N$. Let $U_A = \{1,2,\cdots,N\}$, then the set $S_A$ consists of all the $N/2$ sized complement free\footnote{$S_A$ is complement free if $s \in S_A \implies \bar{s} \notin S_A$.} subsets of $U_A$. Then, note that $M = |S_A| = \frac{1}{2}{N \choose {N/2}} > 2^{N/4}$. Further, $s^i \in S_A$ intersects every $s^j \in S_A$ because $s^j \ne \bar{s}^i$ and $|s^j|=N/2$.
We are now ready to present the reduction.
Let the given instance of \mbox{\sf Subgraph Isomorphism}\xspace be ${\mathcal J} = (G=(V_G,E_G), H=(V_H,E_H),k)$. Let $\ell := |V_H|$, $n:= |V_G|$ and $m:= |E_G|$. Further, let $V(G) = \{1,\cdots,n\}$. Also, let $d(v)$ be the degree of $v \in V_H$. Note that $\ell \le 2k$, since isomorphic sub-graph in $G$ to $H$ is not necessarily induced.
Let ${\mathcal A}=(U_A,S_A)$ be the $(M,N)$-ISS gadget specified above with $N=10\log n$, and hence $M \ge n$.
We view the sets in this ISS as $S_A=\{s^\alpha\}_{\alpha \in V(G)}$.
We construct an instance ${\mathcal I} = ({\mathcal U},{\mathcal S},r)$ of \mbox{\sf Compact \psp}\xspace as follows.
For every $v \in V_H$, and $w \in N_H(v)$, let ${\mathcal A}_{v,w}=(U_{v,w}, S_{v,w})$ be a (distinct) copy of our ISS ${\mathcal A}$ (that is, the universes $\{U_{v,w}\}_{w \in N_H(v)}$ are disjoint). Denote by ${\mathcal A}_v := \{{\mathcal A}_{v,w}\}_{w \in N(v)}$.
Let $U_v:= \cup_{w \in N(v)} U_{v,w}$. The universe ${\mathcal U}$ in ${\mathcal I}$ is just the union of all the universes $U_v$, i.e.,
$
{\mathcal U} := \cup_{v\in V_H} U_v.
$
Now, we will construct sets in ${\mathcal S}$. There are two types of sets, that we call $V$-sets and $E$-sets.
For every $\alpha \in V(G)$ and $v \in V(H)$, add the set $S_{\alpha \mapsto v}$ to ${\mathcal S}$ such that
$
S_{\alpha \mapsto v} := \cup_{w \in N(v)} s^\alpha_{v,w}.
$
These sets are referred to as $V$-sets.
For each edge $(\alpha,\beta) \in E_G$ and each edge $(v,w) \in E_H$, add a set $S_{(\alpha,\beta) \mapsto (v,w)}$ to ${\mathcal S}$ such that
$
S_{(\alpha,\beta) \mapsto (v,w)} := \bar{s}^\alpha_{v,w} \cup \bar{s}^\beta_{w,v}.
$
These sets are called $E$-sets.
Finally, setting the parameter $r=\ell + k$, concludes the construction of \mbox{\sf PSP}\xspace instance ${\mathcal I} = ({\mathcal U},{\mathcal S},r)$. First, note that for the base ISS gadget ${\mathcal A} = (U_A,S_A)$, we have that $|U_A| = \Theta(\log n)$. Hence, $|{\mathcal U}| = \sum_{i \in [\ell]} \sum_{j \in [d(v_i)]} |U_A| = \Theta(k \log n)$, where as $|{\mathcal S}| = \Theta(mk+n\ell) = O(n^2k)$. Thus, we have that $|{\mathcal U}| = \Theta(r \log |{\mathcal S}|)$, since $r =\Theta(k)$. So, ${\mathcal I}$ is, indeed, an instance of \mbox{\sf Compact \psp}\xspace.
To illustrate the main ideas, we analyze the completeness and discuss how the soundness fails.
In the completeness case, let us assume that there exists an injection $\phi: V_H \rightarrow V_G$ which specifies the isomorphic subgraph in $G$.
The set packing solution $T \subseteq {\mathcal S}$ can choose the following sets:
\[\{S_{\phi(v) \mapsto v}\}_{v \in V_H} \mbox{ and } \{S_{(\phi(v), \phi(w)) \mapsto (v,w)} \}_{(v,w) \in E(H)} \]
Notice that we have chosen $\ell+k$ sets from ${\mathcal S}$. Now, we verify that these sets are disjoint. To this end, we will show that when $T$ is restricted to any universe $U_{v,w}$, for $v \in V_H$ and $w \in N(v)$, forms a packing. Let $T\mid_{U_{v,w}}$ be the restriction\footnote{Formally, $T\mid_{U_{v,w}} := \{t \cap U_{v,w} : t \in T\}$.} of $T$ on $U_{v,w}$. Then note that $T\mid_{U_{v,w}}$ is a packing since
$
T\mid_{U_{v,w}} = \{ s^{\phi(v)}_{v,w},\ \bar{s}^{\phi(v)}_{v,w} \}
$
For soundness, consider a ``non-cheating'' solution $T \subseteq {\mathcal S}$ such that $|T|= r$, where there is at most one $V$-set from each vertex of $G$. Let $T_V$ and $T_E$ denote the $V$-sets and $E$-sets of $T$ respectively. Also, assume that $|T_V| = \ell$ and $|T_E| = k$. Let $V_H = \{v_1,\cdots,v_\ell\}$. Relabel the sets in $T_V$ as $T_V=\{T^i_V : \exists S^\alpha_{u\mapsto v_i} \in T_V, \text{ for some } v_i \in V_H \}$.
\begin{proposition}
$T$ covers ${\mathcal U}$.
\end{proposition}
\begin{proof}
First note that $|{\mathcal U}| = \sum_{v \in V_H} |U_v| = \sum_{v \in V_H} \sum_{w \in N(v)} |U_{v,w}| = 2kN$. Since $T^i_V$ contains $d(v_i)N/2$ many elements of ${\mathcal U}$, we have
$
|\bigcup_{i \in [\ell]} T^i_V| = \frac{N}{2} \sum_{i \in [\ell]} d(v_i) = kN
$
as $T_V$ is a packing and $|T_V|=\ell$ due to assumption. On the other hand, any $t \in T_E$ contains $N$ elements of ${\mathcal U}$, and hence
$
|\bigcup_{t \in T_E} t| = kN,
$
as $T_E$ is a packing and $|T_E|=k$ due to assumption. Since $T$ is a packing
$
|\bigcup_{t \in T} t| = 2kN = |{\mathcal U}|.
$\qed
\end{proof}
Now consider $V' := \{\alpha'_1, \cdots, \alpha'_\ell\} \subseteq V_G$, where $\alpha'_i = \alpha$ such that $T_V^i = S^\alpha_{u\mapsto v_i}$. We claim that $G[V']$ is isomorphic to $H$ with isomorphism $\phi:V_H \rightarrow V_G$ such that $\phi(v_i) = \alpha'_i$. To this end, it is sufficient to show that $(v_i,v_j) \in E_H \implies (\phi(v_i),\phi(v_j)) \in E_{G[V']}$. Note that since ${\mathcal A}_{v_i,v_j}$ is an ISS, $T_V\mid_{U_{v_i,v_j}} = T^i_V\mid_{U_{v_i,v_j}} = s^{\alpha'_i}_{v_i,v_j}$. Further, as $T$ is a packing covering ${\mathcal U}$, we have that $T_E\mid_{U_{v_i,v_j}} = \bar{s}^{{\alpha}'_i}_{v_i,v_j}$.
Similarly, since $T_V\mid_{U_{v_j,v_i}} = T^j_V\mid_{U_{v_j,v_i}} = s^{\alpha'_j}_{v_j,v_i}$, it must be that $T_E\mid_{U_{v_j,v_i}} = \bar{s}^{{\alpha}'_j}_{v_j,v_i}$. But this implies that $S_{(\alpha'_i,\alpha'_j) \mapsto (v_i,v_j)} \in T_E$. This means that $(\alpha'_i,\alpha'_j) = (\phi(v_i),\phi(v_j)) \in E_{G[V']}$.
In order to make the soundness work in the general setting, we need to take care of the following.
\begin{enumerate}
\item \textit{Multiple $V$-sets or $E$-sets in optimal packing.}
We want that an optimal packing in ${\mathcal I}$ should have at most one set from each vertex, and at most one set from each edge of $G$ since multiple sets from a vertex (or an edge) would correspond to that vertex (edge) in $G$ being isomorphic to multiple vertices (edges) in $H$.
\item \textit{Sizes of $V$-sets and $E$-sets in optimal packing.} We want that optimal packing has exactly $\ell$ $V$-sets and $k$ $E$-sets.
\item \textit{Mismatch between $V$-sets and $E$-sets in optimal packing.}
Since only $V$-sets contain sets from an ISS, we would like to enforce the condition that the sets used in constructing $E$-sets also form an ISS, which would allow us to recover the isomorphic subgraph corresponding to a packing.
\end{enumerate}
In Section \ref{ss:dihardness}, we show how to modify this construction to fix the above issues.
\subsection{Open problems}
Theorem~\ref{thm:enuhard} is very close to being optimal. So an interesting problem is that to show \mbox{\sf Compact \psp}\xspace also needs time $|{\mathcal S}|^{\Omega(r)}$ time even when $|{\mathcal U}| = r \cdot \Theta(\log |{\mathcal S}|)$, assuming ETH. Note that, in Section~\ref{ss:comp}, we discussed that such lower bound exists, but instead assuming Gap-ETH.
Another question is \textsf{FPT}\xspace approximating \mbox{\sf Compact \psp}\xspace: Given a promise that there is an $r$-packing, is it possible to find a packing of size $o(r)$ in \textsf{FPT}\xspace time? Note that for general \mbox{\sf PSP}\xspace problem, there is no $o(r)$ \textsf{FPT}\xspace-approximation, assuming Gap-ETH. However, as mentioned before, recent results~\cite{DBLP:conf/stoc/Lin21,lin2021lower,s2021polynomial} use a weaker assumption of $\textsf{W[1]}\xspace \ne \textsf{FPT}\xspace$ for showing \textsf{FPT}\xspace-inapproximibility within a polynomial factor in $r$ for \mbox{\sf PSP}\xspace. Showing such hardness of approximation for \mbox{\sf Compact \psp}\xspace is another open problem.
\iffalse
\subsection{Further Related Results}
To our knowledge, the compact version of combinatorial problems has not been formalized and studied before. But several existing reductions already imply hardness of compact version of the problem. Here we review these related results.
\begin{itemize}
\item {\sf Set $r$-Covering}\xspace: \cite{10.1145/3188745.3188896} shows {\sf Compact \scr}\xspace is {{\wone}-hard}\xspace; however, their reduction does not give tight ETH-hardness. On the other hand,
\cite{puatracscu2010possibility} shows tight ETH hardness for {\sf Compact \scr}\xspace, in the sense that they show that {\sf Compact \scr}\xspace requires time $|{\mathcal S}|^{\Omega\left(r\right)}$.
\item {{\sf Exact $r$-Covering}\xspace: } We observe that the reduction of \cite{puatracscu2010possibility} for {\sf Compact \scr}\xspace, can be easily modified to achieve tight ETH hardness of {\sf Compact \ecr}\xspace. The idea is to start the reduction from \mbox{\sc $1$-in-$3$-SAT}\xspace, instead from \mbox{\sc $3$-SAT}\xspace, which produces an instance of {\sf Compact \ecr}\xspace.
\item $r$\mbox{\sf -VectorSum}\xspace: The reduction of \cite{bhattacharyya_et_al:LIPIcs:2016:6362} shows that \mbox{\sf Compact $r$-VectorSum}\xspace requires time $N^{\Omega(\sqrt{k})}$ under ETH.
\end{itemize}
\fi
\paragraph{Organization.}
We start with the preliminaries in Section~\ref{sec:prel}. The proof of Theorem~\ref{thm:dich} is presented in Section~\ref{sec:mainpf}. The proofs of Theorem~\ref{thm:cecpw} and Theorem~\ref{thm:ckvs} are presented in Appendix~\ref{sec:cecr} and Appendix~\ref{sec:ckvs} respectively.
\section{Preliminaries} \label{sec:prel}
\subsection{Notations}
For $q \in \mathbb{N}$, denote by $[q]$, the set $\{1,\cdots, q\}$. For a finite set $[q]$ and $i \in [q]$, we overload $`+'$ operator and denote by $i+1$ as the (cyclic) successor of $i$ in $[q]$. Thus, the successor of $q$ is $1$ in $[q]$. Similarly, denote by $i-1$ as the (cyclic) predecessor of $i$ in $[q]$. Thus, the predecessor of $1$ is $q$ in the set $[q]$.
All the $\log$s are in base $2$.
For a graph $G=(V,E)$ and a vertex $v \in V$, denote by $N(v)$, the set of vertices adjacent to $v$. Further, $d(v)$ denotes the degree of $v$, i.e., $d(v) := |N(v)|$.
For a finite universe $U$ and $s \subseteq U$, denote by $\bar{s}$ as the complement of $s$ under $U$, i.e., $\bar{s} := U \setminus s$. Similarly, for a family of sets $S=\{s_1,\cdots,s_M\}$ over $U$, we denote by $\cmp(S):=\{\bar{s}_1,\cdots,\bar{s}_M\}$.
Further, for a subset $s \subseteq U$ and a sub-universe $U' \subseteq U$, denote by $s \mid_{U'}$ as the restriction of $s$ on sub-universe $U'$, i.e., $s \mid_{U'} := s \cap U'$. Similarly, for a family of sets $S=\{s_1,\cdots,s_M\}$ over $U$, denote by $S\mid_{U'}$ as the restriction of every set of $S$ on $U'$, i.e., $S\mid_{U'}:= \{s_1\mid_{U'},\cdots,s_M\mid_{U'}\}$. For a set system $A = (U_A,S_A)$, we denote the complement set system $\bar{A} := (U_A,\cmp(S_A))$.
\subsection{Parameterized Complexity} \label{ss:pcmpty}
The parameterized complexity theory concerns computational aspects of languages $(L,\kappa) \subseteq \Sigma^* \times \mathbb{N}$. Here $\kappa \in \mathbb{N}$ is called the parameter. Thus a parameterized problem is a classical problem together with a parameter $\kappa$. As an example consider the following classical NP-complete problem.
\noindent\fbox{%
\parbox{\textwidth}{%
\mbox{\sc Clique}\xspace\\
\textit{Instance: } A graph $G$ and $k \in \mathbb{N}$\\
\textit{Problem: } Decide if $G$ has a clique of size $k$
}%
}
\\
Now consider a parameterized version of \mbox{\sc Clique}\xspace defined by $\kappa := k$.
\noindent\fbox{%
\parbox{\textwidth}{%
$k$\mbox{\sf -Clique}\xspace\\
\textit{Instance: } A graph $G$ and $k \in \mathbb{N}$\\
\textit{Parameter: } $k$\\
\textit{Problem: } Decide if $G$ has a clique of size $k$
}%
}
When the parameter $\kappa$ represents the size of solution, then it is called \textit{natural} parameter.
\begin{definition} [Fixed Parameter Tractable]
A parameterized problem $(L,\kappa)$ is called \textit{fixed parameter tractable} if there is an algorithm $A$, a constant $c$ and a computable function $f: \mathbb{N} \rightarrow \mathbb{N}$ such that on all inputs $y=(x,k)$, $A$ decides whether $x$ is in $L$ and runs in time at most $f(k)\cdot |x|^c$.
\end{definition}
The complexity class \textsf{FPT}\xspace is the set of all fixed parameter tractable problems.
In this paper, we consider parameterized problems with natural parameter i.e. $\kappa$ represents the size of solution.
Once we define the class \textsf{FPT}\xspace, the next natural thing is to define \textit{parameterized reduction} or \textit{\textsf{FPT}\xspace-reduction} with the intention that such a reduction from parameterized problem $Q$ to another parameterized problem $Q'$ allows converting an \textsf{FPT}\xspace algorithm of $Q'$ to an \textsf{FPT}\xspace algorithm of $Q$.
\begin{definition} [\textsf{FPT}\xspace-reduction]
An \textsf{FPT}\xspace-reduction from $Q \subseteq \Sigma^* \times \mathbb{N}$ to $Q' \subseteq \Sigma^* \times \mathbb{N}$ is an algorithm $R$ mapping from $\Sigma^* \times \mathbb{N} $ to $\Sigma^* \times \mathbb{N}$ such that for all $y=(x,k), R(y) \in Q'$ if and only if $y \in Q$, and for some computable function $f$ and a constant $c$, $R(y)$ runs in time $f(k)\cdot |x|^c$ and $R(y)=(x',k')$, where $k'\le g(k)$ for some computable function $g$.
\end{definition}
An extensive treatment of the subject can be found in \cite{Downey:2012:PC:2464827,Flum:2006:PCT:1121738}.
\subsection{Problem definitions} \label{ss:pdef}
\begin{definition}[\mbox{\sf Set Packing}\xspace]
Given a collection of sets $\mathcal{S} =\{S_1,\dots, S_m\}$ over an universe $\mathcal{U} = \{e_1,\dots,e_n\}$, the \mbox{\sf Set Packing}\xspace problem asks to find maximum number of sets in ${\mathcal S}$ that are pairwise disjoint. An instance of \mbox{\sf Set Packing}\xspace is denoted as $ ({\mathcal U},{\mathcal S})$.
\end{definition}
\begin{definition}[\mbox{\sf PSP}\xspace]
Given a collection of sets ${\mathcal S} =\{S_1,\dots, S_m\}$ over an universe ${\mathcal U} = \{e_1,\dots,e_n\}$ and an integer $r$, the \mbox{\sf PSP}\xspace problem asks if there is a collection of sets $\mathcal{S}' \subseteq \mathcal{S}$ such that $|\mathcal{S}'| \ge r$ and, $S_i \cap S_j = \emptyset$ for every $S_i \ne S_j \in \mathcal{S'}$.
An instance of \mbox{\sf PSP}\xspace is denoted as $ ({\mathcal U},{\mathcal S},r)$.
\end{definition}
\mbox{\sf Compact \psp}\xspace is defined when the instances have $|{\mathcal U}|= f(r) \cdot \Theta(\textsf{poly}\xspace(\log |{\mathcal S}|))$, for some function $f(r)\ge r$.
Given two graphs $G=(V_G,E_G)$ and $H=(V_H,E_H)$, a homomorphism from $H$ to $G$ is a map $\phi: V_H \rightarrow V_G$ such that if $(v_i,v_j) \in E_H$ then $(\phi(v_i), \phi(v_j)) \in E_G$.
\begin{definition}[\mbox{\sf Subgraph Isomorphism}\xspace]
Given a graph $G=(V_G,E_G)$ and a smaller graph $H=(V_H,E_H)$ with $|E_H|= k$, the \mbox{\sf Subgraph Isomorphism}\xspace problem asks if there is an injective homomorphism from $H$ to $G$.
An instance of \mbox{\sf Subgraph Isomorphism}\xspace is denoted as $(G=(V_G,E_G), H=(V_H,E_H),k)$
\end{definition}
That is, it asks if there is a subgraph of $G$ that is isomorphic (not necessary induced) to $H$.
The parameterized version of \mbox{\sf Subgraph Isomorphism}\xspace has parameter $\kappa = |E_H| = k$.
Without loss of generality, we can assume $H$ has no isolated vertices.
\begin{observation} \label{obs:verinH}
For the \mbox{\sf Subgraph Isomorphism}\xspace problem, we have that $|V_H| \le 2k$, as $|E_H|= k$. Further, every vertex of $H$ has degree at most $k$.
\end{observation}
\section{Dichotomy of \mbox{\sf PSP}\xspace} \label{sec:mainpf}
In this section we prove the hardness part of the dichotomy theorem (Theorem~\ref{thm:dich}). First in Section~\ref{ss:iss}, we identify the gadget and its associated properties that would be useful for the reduction.
Then, in Section~\ref{ss:dihardness}, using this gadget
we show an \textsf{FPT}\xspace-reduction from \mbox{\sf Subgraph Isomorphism}\xspace to \mbox{\sf Compact \psp}\xspace.
\subsection{Compatible Intersecting Set System pair} \label{ss:iss}
A set system ${\mathcal A} = (U_A,S_A)$ is called an $(M,N)$-Intersecting set system (ISS), if it contains $M$ sets over $N$ elements such that every pair $s,t \in S_A$ intersects.
\begin{definition} [Compatible ISS pair]
Given two ISS ${\mathcal A}=(U,S_A)$ and ${\mathcal B}=(U,S_B)$, we say that $({\mathcal A},{\mathcal B})$ is a compatible ISS pair if there exists a bijection $f: S_A \mapsto S_B$ such that
\begin{itemize}
\item(Complement disjointness) $\forall s \in S_A$, we have $s \cap f(s) = \emptyset$, and
\item(Complement exchange) $\forall s \in S_A$, the set system ${\mathcal A}_s =(U,(S_A \setminus \{s\}) \cup \{f(s)\})$ is an ISS.
\end{itemize}
\end{definition}
For a compatible ISS pair $({\mathcal A},{\mathcal B})$, we say $f(s)\in S_B$ is a complement of $s \in S_A$.
\begin{observation} \label{obs:compiss}
We observe the following symmetric properties of a compatible ISS pair $({\mathcal A},{\mathcal B})$.
\begin{enumerate}
\item Since $f$ is a bijection, $|S_A| = |S_B|$.
\item $\forall s' \in S_B$, the set system ${\mathcal B}_{s'} =(U,(S_B \setminus \{s'\}) \cup \{f^{-1}(s')\})$ is an ISS.
\item Since $f$ is a bijection, we have that $f^{-1}(s') \in S_A$ is a complement of $s' \in S_B$. Hence, we say $s \in S_A$ and $s' \in S_A$ are complements of each other, if $s'= f(s)$.
\end{enumerate}
\end{observation}
The following lemma efficiently computes a compatible $(M,N)$-ISS pair with $M$ exponential in $N$.
\begin{lemma}\label{lm:spset}
For even $N \ge 2$, we can compute a compatible $(M,N)$-ISS pair $({\mathcal A},{\mathcal B})$ with $M \ge 2^{N/2-1}$ in time polynomial in $M$ and $N$.
\end{lemma}
\begin{proof}
Given even $N \ge 2$, we construct two set systems ${\mathcal A} =(U,S_A)$ and ${\mathcal B}=(U,S_B)$ greedily as follows. First set $U = [N]$, and unmark all the subsets $s \subseteq U$ of size $\frac{N}{2}$. Then, for every subset $s \subseteq U, |s|=\frac{N}{2}$ that is unmarked, add $s$ to $S_A$ and $\bar{s}$ to $S_B$, and mark both $s$ and $\bar{s}$.
Note that $|S_A|= |S_B|$. First, we claim that both ${\mathcal A}$ and ${\mathcal B}$ are $(M,N)$-ISS. To verify ${\mathcal A}$ is an ISS, observe that any $s, t \in S_A$ intersects since $|s| = |t| = N/2$ and $t \ne \bar{s}$. Similarly, ${\mathcal B}$ is an ISS, as any $\bar{s}, \bar{t} \in S_B$ intersects since $|\bar{s}| = |\bar{t}| = N/2$ and $\bar{t} \ne s$. Next, to lower bound $M$, note that the number of subsets of $[N]$ having exactly half the elements is ${N \choose {N/2}}$. Hence, picking only complement free such subsets, we have that $M \ge \frac{1}{2} {N \choose {N/2}} \ge 2^{N/2-1}$. Thus, the total time to construct $({\mathcal A},{\mathcal B})$ is $O(2^{2N})$, which is polynomial in $M$.
Finally, the following proposition shows that $({\mathcal A},{\mathcal B})$ is a compatible ISS pair.
\begin{proposition}
For the ISS pair $({\mathcal A},{\mathcal B})$, consider the bijection $f: S_A \mapsto S_B$ with $f(s) = \bar{s}$, for $s \in S_A$. Then, for $s \in S_A$, we have that ${\mathcal A}_s = (U,S_A \setminus s \cup f(s))$ is an ISS.
\end{proposition}
\begin{proof}
We will prove that for $s \in S_A$, $f(s) =\bar{s}$ intersects with every set in $S_A \setminus s$. Fix any ${t} \in {S_A} \setminus s$. Then, since $|\bar{s}| = |\bar{t}| = \frac{N}{2}$ and $\bar{t} \ne \bar{s}$, we have $\bar{s} \setminus \bar{t} \ne \emptyset$. But $\bar{s} \cap {t} = \bar{s} \setminus \bar{t} \ne \emptyset$.\qed
\end{proof}
\qed
\end{proof
\subsection{Hardness of \mbox{\sf Compact \psp}\xspace} \label{ss:dihardness}
Our hardness result follows from the following reduction since \mbox{\sf Subgraph Isomorphism}\xspace is hard for \textsf{W[1]}\xspace.
\begin{theorem} \label{thm:sgired}
There is an \textsf{FPT}\xspace-reduction that, for every instance ${\mathcal I} = (G=(V_G,E_G),H=(V_H,E_H),k)$ of \mbox{\sf Subgraph Isomorphism}\xspace with $|V_G|=n$ and $|E_G|=m$, computes $\mu = O(k!)$ instances ${\mathcal J}_p=({\mathcal U}_p,{\mathcal S}_p,r), p \in [\mu]$ of \mbox{\sf PSP}\xspace with the following properties:
\begin{itemize}
\item [$\circ$] $|{\mathcal U}_p|= \Theta(k \log n)$
\item[$\circ$] $|{\mathcal S}_p|= \Theta(n^2k + mk)$
\item[$\circ$] $r=\Theta(k)$
\end{itemize}
such that there is a subgraph of $G$ isomorphic to $H$ if and only if there is an $r$-packing in at least one of the instances ${\mathcal J}_p$.
\end{theorem}
Observe that every instance ${\mathcal J}_p$ is an instance of \mbox{\sf Compact \psp}\xspace since $|{\mathcal U}_p| = \Theta(k \log n)=r \Theta \left( \log |{\mathcal S}_p| \right)$ \footnote{In fact, $|{\mathcal U}_p| = \Theta \left( r\log \left( |{\mathcal S}_p|/r \right) \right)$ }. The construction follows the approach outlined in Section~\ref{ss:tech}.
\begin{proof}
Let $V_G = \{1,\cdots,n\}$. Let $({\mathcal A},\bar{{\mathcal A}})$ be the compatible $(M,N)$-ISS pair given by Lemma \ref{lm:spset}, for $N = 2\lceil \log (n+1) \rceil +2$. We call the ISS ${\mathcal A}= (U_A,S_A)$ as the base ISS gadget.
Further, assume an arbitrary ordering on $S_A= \{s^1,\cdots,s^M\}$. Since $M \ge 2^{N/2-1} = 2^{\lceil\log (n+1)\rceil} > n$, every $\alpha \in V_G$ can also be identified by set $s^\alpha \in S_A$ corresponding to the index $\alpha \in [M]$.
For each ordering $p: V_H \rightarrow [\ell]$, create an instance ${\mathcal J}_p = ({\mathcal U}_p, {\mathcal S}_p,r)$ of \mbox{\sf Compact \psp}\xspace as follows. Rename the vertices of $V_H$ as $\{v_1,\cdots,v_\ell \}$ such that $v_i := v \in V_H$ such that $p(v)=i$.
For each $v_i \in V_H$, create a collection ${\mathcal C}_{v_i}$ of $d(v_i)+1$ many different copies of base ISS gadget as follows:
\[
{\mathcal C}_{v_i} :=\{{\mathcal A}_{v_i,0}, \{{\mathcal A}_{v_i,w}\}_{w \in N(v_i)} \}
\]
where, ${\mathcal A}_{v_i,0} = (U_{v_i,0},S_{v_i,0}) \text{ and } {\mathcal A}_{v_i,w} = (U_{v_i,w},S_{v_i,w}).$
For brevity, we let ${\mathcal A}_{v_i} :=\{{\mathcal A}_{v_i,w}\}_{w \in N(v_i)}$ such that ${\mathcal A}_{v_i} =(U_{v_i}, S_{v_i})$, where $U_{v_i} := \cup_{w \in N(v)} U_{v_i,w}$ and $S_{v_i} := \times_{w\in N(v)} S_{v_i,w}$.
Thus, we can write ${\mathcal C}_{v_i}$ concisely as ${\mathcal C}_{v_i} = \{{\mathcal A}_{v_i,0}, {\mathcal A}_{v_i}\}$.
For each ${\mathcal C}_{v_i}$, consider the union of universes of both ISS, i.e, $U_{{\mathcal C}_i} := U_{v_i,0} \cup U_{v_i}$. Now, we define the universe ${\mathcal U}_p$ of ${\mathcal J}_p$ as the union of all the universes $U_{{\mathcal C}_i}, i \in [\ell]$, i.e.,
\[
{\mathcal U}_p := \bigcup_{i \in [\ell]} U_{{\mathcal C}_i}.
\]
Next, we construct the collection ${\mathcal S}_p$ of ${\mathcal J}_p$. The sets in ${\mathcal S}_p$ will be union of $V$-sets and $E$-sets, corresponding to the vertices and the edges of $G$ respectively. For $\alpha \in V_G$ and $v \in V_H$, we define the set $S^\alpha_v := \cup_{w \in N(v)} s^\alpha_{v,w}$. Note that $S^\alpha_v \in S_v$ of ISS ${\mathcal A}_v$.
\paragraph*{$V$-sets:} For each $\alpha \in V_G$, for each $v_i \in \{v_1,\cdots, v_{\ell-1}\}$, and for each $\beta > \alpha \in V_G$, add a set $S_{\alpha \mapsto v_i, \beta}$ to ${\mathcal S}_p$ such that
\[
S_{\alpha \mapsto v_i, \beta} := {s}^{\alpha}_{v_i,0} \; \bigcup \; S^\alpha_{v_i} \; \bigcup \; \bar{s}^{\beta}_{v_{i+1},0}
\]
Further, for each $\alpha \in V_G$, and for each $\beta< \alpha \in V_G$, add a set $S_{\alpha \mapsto v_\ell, \beta}$ to ${\mathcal S}_p$ such that
\[
S_{\alpha \mapsto v_\ell, u'} := {s}^{\alpha}_{v_\ell,0} \; \bigcup \; S^\alpha_{v_\ell} \; \bigcup \; \bar{s}^{\beta}_{v_{1},0}
\]
\paragraph*{$E$-sets:} For each edge $(\alpha,\beta) \in E_G$ and each edge $(v_i,v_j) \in E_H$, add a set $S_{(\alpha,\beta) \mapsto (v_i,v_j)}$ to ${\mathcal S}_p$ such that
\[
S_{(\alpha,\beta) \mapsto (v_i,v_j)} := \bar{s}^\alpha_{v_i,v_j} \cup \bar{s}^{\beta}_{v_j,v_i}
\]
\paragraph*{Parameter} Set $r := k + \ell$.
This concludes the construction. Before we prove its correctness, we note the size of the constructed instance ${\mathcal J}_p$. First, note that $r= \Theta(k)$, since $\ell \le 2k$. Then, we have
\[
|{\mathcal U}_p| = \sum_{i=1}^{\ell} |U_{{\mathcal C}_i}| = \sum_{i=1}^{\ell} (d(v_i) + 1) N = (2k + \ell)N = \Theta(k\log n),
\]
and
$
|{\mathcal S}_p| \le n\cdot \ell \cdot n + mk = \Theta(n^2k).
$
\paragraph*{\textbf{Yes case:}} Suppose there is a subgraph $G'=(V_{G'},E_{G'})$ of $G$ that is isomorphic to $H$ with isomorphism $\phi:V_{H} \rightarrow V_{G'}$. Let $V_{G'}=\{\alpha_{1},\alpha_{2},\cdots, \alpha_{\ell}\} \subseteq [n]$ such that $\alpha_1<\alpha_{2} <\cdots<\alpha_{\ell}$.
Relabel the vertices of $H$ as $\{v_1,\cdots,v_\ell\}$, where $v_i := \phi^{-1}(\alpha_{i}), i \in [\ell]$.
Now, consider the ordering $p$ of $V_H$ such that $p(v_i):=i$, for $i \in [\ell]$, and fix the corresponding instance ${\mathcal J}_p=({\mathcal U}_p,{\mathcal S}_p,r)$. Consider the following collection of $V$-sets and $E$-sets,
\[
T_V := \bigcup_{i \in [\ell]} S_{\alpha_{i} \mapsto v_i, \alpha_{i+1}} \qquad T_E := \bigcup_{(v_i,v_{j}) \in E_H} S_{(\alpha_{i}, \alpha_{j}) \mapsto (v_i,v_{j})}
\]
Let $T:=T_V \cup T_E$. Note that, $T_V \subseteq {\mathcal S}_p$ due to construction, and $T_E \subseteq {\mathcal S}_p$ due to $\phi$, hence $T \subseteq {\mathcal S}_p$. Further, $|T| = |T_V| + |T_E| = \ell + k=r$, as required. Now, we claim that $T$ forms a packing in ${\mathcal J}_p$.
Towards this goal, note that it is sufficient to show that the sets in $T\mid_{U_{{\mathcal C}_i}}$ are mutually disjoint, for all $i \in [\ell]$. To this end, it is sufficient to show that both $T \mid_{U_{v_i,0}}$ and $T \mid_{U_{v_i}}$ are packing, for all $i \in [\ell]$. Fix $i \in [\ell]$ and consider the following case analysis.
\begin{enumerate}
\item \textit{$T \mid_{U_{v_i,0}}:$} Note that we only need to show $T_V \mid_{U_{v_i,0}}$ is a packing since $T_E \mid_{U_{v_i,0}} = \emptyset$ by construction. Notice that $T_V \mid_{U_{v_i,0}}$ is a packing since,
\begin{align*}
T_V \mid_{U_{v_i,0}} =
\begin{cases}
\{S_{\alpha_{i-1} \mapsto {v_{i-1}}, \alpha_i}\mid_{U_{v_i,0}}, S_{\alpha_{i} \mapsto v_i, \alpha_{i+1}}\mid_{U_{v_i,0}}\} = \{\bar{s}^{\alpha_i}_{v_i,0}, s^{\alpha_i}_{v_i,0}\} & \text{if }i \ne 1 \\
\{S_{\alpha_{\ell} \mapsto {v_{\ell}}, \alpha_1}\mid_{U_{v_1,0}}, S_{\alpha_{1} \mapsto v_1, \alpha_{2}}\mid_{U_{v_1,0}}\}=\{\bar{s}^{\alpha_1}_{v_1,0}, s^{\alpha_1}_{v_1,0}\} & \text{if }i = 1.
\end{cases}
\end{align*}
\item \textit{$T\mid_{U_{v_i}}:$} It is sufficient to show that $T \mid_{U_{v_i,v_j}}$ is a packing, $\forall v_j \in N(v_i)$. But this is true as,
\begin{align*}
T \mid_{U_{v_i,v_j}} &= \{T_V \mid_{U_{v_i,v_j}}, T_E \mid_{U_{v_i,v_j}}\} \\
&= \{ S_{\alpha_i\mapsto v_i, \alpha_{i+1}}\mid_{U_{v_i,v_j}}, S_{(\alpha_i,\alpha_j) \mapsto (v_i,v_j)}\mid_{U_{v_i,v_j}} \}\\
&= \{s^{\alpha_i}_{v_i,v_j},\bar{s}^{\alpha_i}_{v_i,v_j} \}.
\end{align*}
\end{enumerate}
\paragraph*{\textbf{No case:}}
Suppose there is an $r$-packing $T \subseteq {\mathcal S}_p$ for some instance ${\mathcal J}_p, p\in[\mu]$, then we show that there is a subgraph $G_T$ of $G$ that is isomorphic to $H$. First note that, $p \in [\mu]$ gives a labeling $\{v_1,\cdots, v_\ell\}$ of $V_H$ such that $v_i=p^{-1}(i)$, for $i\in [\ell]$.
Next, partition $T$ into $T_V$ and $T_E$, where $T_V$ contains $V$-sets of $T$, and $T_E$ contains $E$-sets of $T$. This can be easily done since $t \in T$ is a $V$-set if and only if $t\mid_{U_{v_i,0}} = s^{\alpha}_{v_i,0}$, for some $\alpha \in V_G, v_i \in V_H$ and $s^{\alpha}_{v_i,0} \in {\mathcal A}_{v_i,0}$. We define the following universes that will be helpful for the proof.
\[
U_0 := \{U_{v_i,0}\}_{v_i \in V_H} \quad\text{and}\quad U_1 := \{U_{v_i}\}_{v_i \in V_H}.
\]
We claim the following.
\begin{lemma}\label{lm:setsizes}
$|T_V| = \ell$ and $|T_E| = k$.
\end{lemma}
\begin{proof}
Note that for $t \in T_V$, we have $t\mid_{U_0} = \{s^{\alpha}_{v_i,0}, \bar{s}^{\beta}_{v_{i+1},0}\}$, for some $\alpha,\beta \in V_G$ and $v_i, v_{i+1} \in V_H$. Thus, from Lemma~\ref{lm:spset}, it follows that $|t\mid_{U_0}| = N$. Since $|U_0| = \ell N$, we have $|T_V| \le \ell$.
For bounding $|T_E|$, consider $t\in T_E$, and note that $t \mid_{U_1} = \{\bar{s}^\alpha_{v_i,v_j},\bar{s}^{\beta}_{v_j,v_i}\}$, for some $(\alpha,\beta) \in E_G$ and $(v_i,v_j) \in E_H$.
As $\bar{{\mathcal A}}$ is an ISS, we have that $U_1$ is also the union of universes of $2k$ many base ISS $\{\bar{{\mathcal A}}_{v_i}\}_{i \in [\ell]}$. Since each $t\in T_E$ contains sets from two such ISS, it follows that $|T_E| \le k$.
Finally, since $|T| = r = \ell + k$, this means $|T_V| = \ell$ and $|T_E| = k$. \qed
\end{proof}
Let $T_V = \{T^1_{V},\cdots, T^\ell_{V}\}$ such that for $i \in [\ell]$, $T^i_V := t \in T_V$ such that $t\mid_{U_0} \ni s^\alpha_{v_i,0}$, for some $\alpha \in V_G$.
\begin{lemma} \label{cl:tcovu}
$T$ covers the whole universe ${\mathcal U}_p$.
\end{lemma}
\begin{proof}
Since ${\mathcal U}_p = \{U_0,U_1\}$, we will show that $T\mid_{U_0}$ and $T\mid_{U_1}$ covers $U_0$ and $U_1$ respectively. For $U_0$, note that $T\mid_{U_0} = T_V\mid_{U_0}$ by construction. For $T^i_V\in T_V$, as noted in the proof of above lemma, $|T^i_V\mid_{U_0}| = N$. Since the sets in $T_V$ are mutually disjoint, we have that $|\cup_{i \in [\ell]} T^i_V \mid_{U_0}| = \ell N = |U_0|$, as desired.
Next, for $U_1$, note that $|U_1| = 2kN$. Consider $T^i_V\in T_V$ and notice $|T^i_V\mid_{U_1}|=\frac{N}{2}d(v_i)$ since $T^i_V \mid_{U_1} = S^\alpha_{v_i}$, for some $\alpha \in V_G$. Since the sets in $T_V$ are mutually disjoint, we have
\[
|\bigcup_{i \in [\ell]} T^i_V \mid_{U_1}|= \sum_{i=1}^{\ell} |T^i_V\mid_{U_1}| = \sum_{i=1}^{\ell} \frac{N}{2} d(v_i) = kN
\]
Now consider $t = S_{(\alpha,\beta) \mapsto (v_i,v_j)} \in T_E$, for some $(\alpha,\beta) \in E_G$ and $(v_i,v_j) \in E_H$. Then, $t\mid_{U_1} = \{\bar{s}^\alpha_{v_i,v_j}, \bar{s}^{\beta}_{v_j,v_i}\}$. Hence, $|t\mid_{U_1}| = N$. Since the sets in $T_E$ are mutually disjoint, we have
\[
|\bigcup_{t \in T_E} t \mid_{U_1}| = \sum_{t \in T_E} |t\mid_{U_1}| = kN
\]
Since $T$ is a packing, we have
\[
|\bigcup_{\tau \in T} \tau \mid_{U_1}| = |\bigcup_{i \in [\ell]} T^i_V \mid_{U_1}| + |\bigcup_{t \in T_E} t \mid_{U_1}| = 2kN = |U_1|
\]
completing the proof.
\qed
\end{proof
Consider the set $V_{T} := \{\alpha : T^i_V \mid_{U_{v_i,0}} = s^{\alpha}_{v_i,0}, i \in [\ell]\} \subseteq V_G$. Let us denote the vertices in $V_{T}$ as $\{\alpha_1,\cdots,\alpha_\ell\}$, where $\alpha_i:=\alpha$ such that $ T^i_V \mid_{U_{v_i,0}} = s^{\alpha}_{v_i,0}$, for $i \in [\ell]$.
Following proof asserts that $\alpha_1 < \cdots < \alpha_\ell$, which is sufficient to conclude that $V_T$ contains non-duplicate vertices.
\begin{lemma}\label{lm:ethuvertex}
For each vertex $\alpha \in V_G$, there is at most one $V$-set $S_{\alpha \mapsto v_i, \beta}$ in $T_V$, for some $v_i \in V_H$ and $\beta \in V_G$.
\end{lemma}
\begin{proof}
We show $\alpha_{1}<\cdots<\alpha_{\ell}$, for $V_T$, which is sufficient. For any $i \in [\ell-1]$, we have that $T^i_V = S_{\alpha_i \mapsto v_i, \beta}$, for some $\beta > \alpha_i$. This means $T^i_V\mid_{U_{v_{i+1},0}}= \bar{s}^{\beta}_{v_{i+1},0}$. From Lemma~\ref{cl:tcovu}, since $T$ covers ${\mathcal U}_p$, and in particular $T$ covers $U_{v_{i+1},0}$, it should be that $T^{i+1}_V \mid_{U_{v_{i+1},0}} = s^{\beta}_{v_{i+1},0}$ as $({\mathcal A}_{v_i+1,0}, \bar{{\mathcal A}}_{v_i+1,0})$ is a compatible ISS pair. Hence, $\alpha_{i+1} =\beta > \alpha_i$.\qed
\end{proof
We have a similar claim for $E$-sets.
\begin{lemma} \label{lm:ethuedge}
For every edge $(\alpha,\beta) \in E_G$, there is at most one $E$-set $S_{(\alpha,\beta) \mapsto (v_i,v_j)}$ in $T_E$, for some $(v_i,v_j) \in E_H$.
\end{lemma}
\begin{proof}
Suppose, there are two different sets $S_{(\alpha,\beta) \mapsto (v_i,v_j)}, S_{(\alpha,\beta) \mapsto (v'_i,v'_j)} \in T_E$, then we will show that there are two sets $S_{\alpha \mapsto v_i, \gamma}, S_{\alpha \mapsto v'_i, \delta} \in T_V$, for some $\gamma,\delta \in V_G$, contradicting Lemma~\ref{lm:ethuvertex}. Since, $S_{(\alpha,\beta) \mapsto (v_i,v_j)}, S_{(\alpha,\beta) \mapsto (v'_i,v'_j)} \in T_E$, it holds that $T_E\mid_{U_{v_i,v_j}} = \bar{s}^\alpha_{v_i,v_j}$, and $T_E\mid_{U_{v'_i,v'_j}}=\bar{s}^\alpha_{v'_i,v'_j}$. As $T$ covers ${\mathcal U}_p$, in particular, $T$ covers $U_{v_i,v_j}$, it must be that $T_V\mid_{U_{v_i,v_j}} = s^\alpha_{v_i,v_j}$ as $({\mathcal A}_{v_i+1,0}, \bar{{\mathcal A}}_{v_i+1,0})$ is a compatible ISS pair. Similarly, as $T$ covers $U_{v'_i,v'_j}$, it must be that $T_V\mid_{U_{v'_i,v'_j}} = s^\alpha_{v'_i,v'_j}$ as $({\mathcal A}_{v_i+1,0}, \bar{{\mathcal A}}_{v_i+1,0})$ is a compatible ISS pair.
This implies that $T_V \mid_{U_{v_i}} = S^\alpha_{v_i}$ and $T_V \mid_{U_{v'_i}} = S^\alpha_{v'_i}$. Thus, $S_{\alpha \mapsto v_i,\gamma}, S_{\alpha \mapsto v'_i,\delta} \in T_V$, for some $\gamma,\delta \in V_G$.\qed
\end{proof
Let $G_T := G[V_T] =(V_T,E_T)$, be the induced subgraph of $G$ on $V_T$. The following lemma says that $G_T$ is, indeed, isomorphic to $H$ completing the proof of No case.
\begin{lemma} \label{cl:ethiso}
There exists an isomorphism (homomorphism) $\phi : V_H \rightarrow V_{T}$ given by $\phi(v_i) = \alpha_{i}$. Hence, $G_T= G[V_T]$ is isomorphic to $H$.
\end{lemma}
\begin{proof}
We need to show that for any $(v_i,v_j) \in E_H$, it holds that $(\phi(v_i),\phi(v_j))= (\alpha_i,\alpha_j) \in E_T$.
Consider the sub-universes $U_{v_i,v_j}$ and $U_{v_j,v_i}$, and note that $T^i_V \mid_{U_{v_i,v_j}}= s^{\alpha_i}_{v_i,v_j}$.
As $T$ covers $U_{v_i,v_j}$, it holds that $T_E \mid_{U_{v_i,v_j}} = \bar{s}^{\alpha_i}_{v_i,v_j}$ since $({\mathcal A}_{v_i,v_j}, \bar{{\mathcal A}}_{v_i,v_j})$ is a compatible ISS pair. Hence $S_{(\alpha_i,\beta) \mapsto (v_i,v_j)} \in T_E$, for some $(\alpha_i,\beta) \in E_G$. This implies that $T_E \mid_{U_{v_j,v_i}} = \bar{s}^{\beta}_{v_j,v_i}$.
As $T$ is covers $U_{v_j,v_i}$ and $T^j_v \mid_{U_{v_j,v_i}}= s^{\alpha_j}_{v_j,v_i}$, it must be the case that $\beta = \alpha_j$ since $({\mathcal A}_{v_j,v_i}, \bar{{\mathcal A}}_{v_j,v_i})$ is a compatible ISS pair.
\qed
\end{proof
\qed
\end{proof
\textbf{Acknowledgment: } I thank Parinya Chalermsook for the informative discussions about the results of the paper, and for providing guidance on writing this paper. I also thank anonymous reviewers for their helpful suggestions.
\bibliographystyle{plain}
|
\section{Introduction}
The chaotic nature of generic quantum many-body systems has triggered interest across many fields of physics, ranging from nuclear physics \cite{wigner1993characteristic} and condensed matter theory (see e.g. \cite{beenakker1997random}) to quantum gravity \cite{Kazakov:1985ea}. This converging interest can be explained by one of the most fascinating concepts in physics: universality. It is perhaps in quantum chaos that lies the strongest form of universality: the hamiltonian of any generic (i.e. non-integrable) quantum system behaves approximately like a random matrix \cite{bohigas1984characterization}. To be precise, it is the statistics of nearby energy levels that obey universal distributions. The most common probe of energy level statistics is the spectral form factor
\begin{equation}
Z(\beta+i t) Z(\beta -it) = \sum_{n,m} e^{-\beta (E_n+E_m)} e^{it(E_n-E_m)} \,
\end{equation}
The behaviour of the spectral form factor is universal in chaotic systems (see for example \cite{Cotler:2016fpe,Altland:2020ccq} for a recent discussion). It decreases at early times due to the destructive nature of the interference caused by the phases. At sufficiently late times, which depends on the system under investigation, this decay gives way to erratic oscillations. However, the mean signal of this erratic curve still follows a universal pattern: it rises linearly (this is known as the \textit{ramp}) before saturating on a \textit{plateau}. This behaviour has a physical meaning: it is related to the repulsion of nearby eigenvalues, and ultimately comes from the fact that the spectrum of the Hamiltonian is discrete \cite{Altland:2017eao, Cotler:2016fpe}.
Another useful probe of chaos is the auto-correlation function
\begin{equation} \label{autocor}
\braket{O(t) O(0)}_{\beta} = \sum_{n,m} |O_{nm}|^2 e^{-\beta E_n} e^{it(E_n-E_m)} \,
\end{equation}
where $O_{mn}$ are the matrix elements of the operator $O$ in energy eigenstates. The behaviour of this quantity is also universal, at least for an appropriate choice of operator $O$ \cite{Altland:2021rqn}. The time-scales can, however, be different than for the spectral form factor. Both the spectral form factor and the auto-correlation function are universal probes of chaos that can be defined in arbitrary quantum systems. There may also exist other probes of chaos that are special to certain classes of quantum systems. The goal of this paper is to explore such probes in the context of conformal fields theories (CFTs).
Conformal symmetry organizes the dynamics of the theory in a convenient manner. The spectrum of the Hamiltonian is given by the spectrum of local operators $\Delta_i$, and the rest of the dynamical data is encoded in the OPE coefficients $c_{ijk}$. At the local level, a CFT is fully specified by this data, and quantum chaos must therefore be reformulated in terms of the data $\{\Delta_i, c_{ijk}\}$. There are various quantities that can be built from this dynamical data and that would encode properties of quantum chaos. Many quantities of this type have not been studied from the point of view of RMT universality and in this work, we initiate a program to do so. Consider the quantity\footnote{In $d=2$, quantities like this can have a nice path integral interpretation on a higher genus surface. In $d>2$, this may no longer be the case, but they can be defined nonetheless.}
\begin{equation}
\mathcal{Z}(\beta_1, \cdots,\beta_n) \equiv \sum_{{\Delta_i}} f\left[c \right] e^{-\sum_{i}\beta_i \Delta_i} \,,
\end{equation}
where $f\left[c\right]$ is a function that depends on at least $2n/3$ OPE coefficients. An example with $n=6$ would be $c_{ijk}c_{ijl}^*c_{mnl}c_{mnk}^*$. We can now build a \textit{generalized spectral form factor}
\begin{equation}
F(t_1, \cdots, t_2) \equiv |\mathcal{Z}(\beta_1+it_1, \cdots,\beta_n+it_n)|^2 \,.
\end{equation}
In this paper, we present a general framework to embed the dynamical data of the CFT in a quantum mechanical setting, where we can apply random matrix theory.
\begin{figure}[t]
\centering
\begin{tikzpicture}
\draw[thin, blue] (7.6,0.57) .. controls (7.3,0.7) and (6.18,0.35) .. node (b) [above] {$\tau_3$} (6.18,0.06)
\draw[thin, blue!50] (7.6,0.57) .. controls (7.8,0.3) and (6.3,-0.13) .. (6.18,0.06)
\draw[thin, gray!50] (5.67,0) .. controls (5.67,-0.3) and (2.35,-0.3) .. (2.35,0)
\draw[thin, gray!90] (5.67,0) .. controls (5.67,0.3) and (2.35,0.3) .. node (g) [below left] {$\tau_2$} (2.35,0)
\draw[thin, red!85] (2,0.08) .. controls (1.75,0.13) and (1.75,0.95) .. node (r) [left] {$\tau_1$} (2,1)
\draw[thin, red!35] (2,0.08) .. controls (2.25,0.13) and (2.25,0.95) .. (2,1)
\draw[thick] (0,0) .. controls (0,0.5) and (1,1) .. (2,1) .. controls (3,1) and (3,0.5) .. (4,0.5) .. controls (5,0.5) and (5,1) .. (6,1) .. controls (7,1) and (8,0.5) .. (8,0);
\draw[thick] (0,0) .. controls (0,-0.5) and (1,-1) .. (2,-1) .. controls (3,-1) and (3,-0.5) .. (4,-0.5) .. controls (5,-0.5) and (5,-1) .. (6,-1) .. controls (7,-1) and (8,-0.5) .. (8,0);
\draw[thick] (1.5,0.05) .. controls (2,-0.1) and (2,-0.1) .. (2.5,0.05) ;
\draw[thick] (1.65,0.001) .. controls (2,0.1) and (2,0.1) .. (2.35,0.001) ;
\draw[thick] (5.5,0.05) .. controls (6,-0.1) and (6,-0.1) .. (6.5,0.05) ;
\draw[thick] (5.65,0.002) .. controls (6,0.1) and (6,0.1) .. (6.35,0.002) ;
\end{tikzpicture}
\caption{Genus-2 surface with three different cycles related to the three moduli $\tau_i$.}
\label{fig.g2s}
\end{figure}
While we expect that this framework will work in full generality, on practical grounds, we focus on a simple probe of heavy OPE coefficients that is inspired by the genus-2 partition function in two-dimensional CFTs. This is defined a
\begin{equation} \label{genus2pf}
Z(\tau_1,\tau_2,\tau_3) = \sum_{O_1,O_2,O_3} |c_{123}|^2 q_1^{\Delta_1} q_2^{\Delta_1} q_3^{\Delta_3} \,, \qquad q_i=e^{2\pi i \tau_i} \,.
\end{equation}
The sum over triplets of operators is a sum over \textit{all} operators of the CFT. One could naturally reduce the sums to be only over primaries by defining appropriate conformal blocks. For simplicity, we will not do so in this paper as we work in a toy-model which neglects the contribution of descendants. We comment on this further below.
For concreteness, it is useful to view \rref{genus2pf} in the context of two-dimensional CFTs, but the calculations we perform are more general. For $d=2$, the $\tau_i$ are moduli of the genus-2 surface and are related to its period matrix, see Fig. \ref{fig.g2s}. This quantity is then given by a tripled sum over energy eigenstates, weighted by the OPE coefficients squared (see \cite{Belin:2017nze,Cardy:2017qhl,Cho:2017fzo,Collier:2019weq} for various studies of the genus-2 partition function). In CFTs, $O_{mn}=c_{Onm}$, such that one might think that this quantity is similar to the auto-correlation function: a Boltzmann sum weighted by OPE coefficients. However, there is a crucial difference. The expression \rref{genus2pf} does not correspond to the expectation value of a fixed operator; instead, it is like a correlation function where the probe operators are summed over. Moreover, in the small $\tau_i$ limit, the sum is dominated by OPE coefficients with three heavy operators, such that it probes different physics than \rref{autocor}.
We now introduce the associated \textit{genus-2 spectral form factor} as a new probe of quantum chaos. It is defined as
\begin{equation} \label{genus2sffintro}
F(t_1,t_2,t_3)= Z(\tau_1+it_1,\tau_2+it_2,\tau_3+it_3) Z(\tau_1+it_1,\tau_2+it_2,\tau_3+it_3)^* \,.
\end{equation}
This quantity is sensitive to both the spectral statistics (it involves six densities of states) and the statistics of OPE coefficients. One would like to study the behaviour of this quantity in chaotic (i.e. non-integrable) CFTs.
Unfortunately, an explicit computation of this quantity appears to be difficult. It is not known how to solve non-integrable CFTs analytically and the conformal bootstrap is not presently powerful enough to efficiently probe a quantity like \rref{genus2sffintro} numerically. Instead, we investigate the implications of random matrix universality on the genus-2 spectral form factor. We propose an alternative description of the genus-2 spectral form factor that we expect to be valid at sufficiently late times and proceed to study it using insights from random matrix theory. Therefore, we can regard our random-matrix analysis as a toy-model that closely captures the relevant late time physics.
\begin{figure}[th]
\centering
\begin{tikzpicture}
\draw[thick,->] (0,0.5) node(orgn){} -- (0,5) node[anchor=south east] {$\ln|Z|^2$};
\draw[thick,->] (0,0.5) -- (6.0,0.5) node[anchor=north west] {$\ln t$};
\fill[fill=blue!20] (1,1)--(2.5,2.6)--(2.5,2.3);
\fill[fill=blue!20] (2.5,2.3) rectangle (5.0,2.6);
\draw [blue!25,
declare function={
fn1(\t,\b)= 1 + ln(1+ \t + 0.3*\t*(sin(\b*\t)) + 0.3*\t*(sin(\b*random(0.5,2)*\t)));
fnx(\x) = 1+ln(1+\x);
}]
plot [domain=0:1.5, samples=144, smooth] ({fnx(\x)},{fn1(\x,2500)});
\draw [blue!25,
declare function={
fn1(\t,\b)= 1 + ln(1+ \t + 0.3*\t*(sin(\b*\t)) + 0.3*\t*(sin(\b*random(0.5,2)*\t)));
fnx(\x) = 1+ln(1+\x);
}]
plot [domain=1.5:3.3, samples=144, smooth] ({fnx(\x)},{fn1(\x,2000)});
\draw [blue!25,
declare function={
fn2(\t,\b)= 2.5 + ln(1 + 0.3*(sin(\b*\t)) + 0.3*(sin(\b*random(0.5,2)*\t)));
fnx2(\x) = 2.5+ln(1+\x);
}]
plot [domain=0:1, samples=74, smooth] ({fnx2(\x)},{fn2(\x,5000)});
\draw [blue!25,
declare function={
fn2(\t,\b)= 2.5 + ln(1 + 0.3*(sin(\b*\t)) + 0.3*(sin(\b*random(0.5,2)*\t)));
fnx2(\x) = 2.5+ln(1+\x);
}]
plot [domain=1:1.6, samples=74, smooth] ({fnx2(\x)},{fn2(\x,5000)});
\draw [blue!25,
declare function={
fn2(\t,\b)= 2.5 + ln(1 + 0.3*(sin(\b*\t)) + 0.3*(sin(\b*random(0.5,2)*\t)));
fnx2(\x) = 3.3+ln(1+\x);
}]
plot [domain=0:1, samples=74, smooth] ({fnx2(\x)},{fn2(\x,5000)});
\draw [blue!25,
declare function={
fn2(\t,\b)= 2.5 + ln(1 + 0.3*(sin(\b*\t)) + 0.3*(sin(\b*random(0.5,2)*\t)));
fnx2(\x) = 3.7+ln(1+\x);
}]
plot [domain=0:1, samples=74, smooth] ({fnx2(\x)},{fn2(\x,5000)});
\draw [blue!25,
declare function={
fn2(\t,\b)= 2.5 + ln(1 + 0.3*(sin(\b*\t)) + 0.3*(sin(\b*random(0.5,2)*\t)));
fnx2(\x) = 4.1+ln(1+\x);
}]
plot [domain=0:1, samples=74, smooth] ({fnx2(\x)},{fn2(\x,5000)});
\draw [blue!25,
declare function={
fn2(\t,\b)= 2.5 + ln(1 + 0.3*(sin(\b*\t)) + 0.3*(sin(\b*random(0.5,2)*\t)));
fnx2(\x) = 4.5+ln(1+\x);
}]
plot [domain=0:1, samples=74, smooth] ({fnx2(\x)},{fn2(\x,5000)});
\draw [blue!25,
declare function={
fn1(\t,\b)= 1 + ln(1+ \t + 0.3*\t*(sin(\b*\t)) + 0.3*\t*(sin(\b*random(0.5,2)*\t)));
fnx(\x) = 1+ln(1+\x);
}]
plot [domain=0:1.5, samples=144, smooth] ({fnx(\x)},{fn1(\x,1863)});
\draw [blue!25,
declare function={
fn1(\t,\b)= 1 + ln(1+ \t + 0.3*\t*(sin(\b*\t)) + 0.3*\t*(sin(\b*random(0.5,2)*\t)));
fnx(\x) = 1+ln(1+\x);
}]
plot [domain=1.5:3.3, samples=144, smooth] ({fnx(\x)},{fn1(\x,1732)});
\draw [blue!25,
declare function={
fn2(\t,\b)= 2.5 + ln(1 + 0.3*(sin(\b*\t)) + 0.3*(sin(\b*random(0.5,2)*\t)));
fnx2(\x) = 2.5+ln(1+\x);
}]
plot [domain=0:1, samples=74, smooth] ({fnx2(\x)},{fn2(\x,3282)});
\draw [blue!25,
declare function={
fn2(\t,\b)= 2.5 + ln(1 + 0.3*(sin(\b*\t)) + 0.3*(sin(\b*random(0.5,2)*\t)));
fnx2(\x) = 2.5+ln(1+\x);
}]
plot [domain=1:1.6, samples=74, smooth] ({fnx2(\x)},{fn2(\x,2957)});
\draw [blue!25,
declare function={
fn2(\t,\b)= 2.5 + ln(1 + 0.3*(sin(\b*\t)) + 0.3*(sin(\b*random(0.5,2)*\t)));
fnx2(\x) = 3.3+ln(1+\x);
}]
plot [domain=0:1, samples=74, smooth] ({fnx2(\x)},{fn2(\x,4829)});
\draw [blue!25,
declare function={
fn2(\t,\b)= 2.5 + ln(1 + 0.3*(sin(\b*\t)) + 0.3*(sin(\b*random(0.5,2)*\t)));
fnx2(\x) = 3.7+ln(1+\x);
}]
plot [domain=0:1, samples=74, smooth] ({fnx2(\x)},{fn2(\x,2839)});
\draw [blue!25,
declare function={
fn2(\t,\b)= 2.5 + ln(1 + 0.3*(sin(\b*\t)) + 0.3*(sin(\b*random(0.5,2)*\t)));
fnx2(\x) = 4.1+ln(1+\x);
}]
plot [domain=0:1, samples=74, smooth] ({fnx2(\x)},{fn2(\x,2930)});
\draw [blue!25,
declare function={
fn2(\t,\b)= 2.5 + ln(1 + 0.3*(sin(\b*\t)) + 0.3*(sin(\b*random(0.5,2)*\t)));
fnx2(\x) = 4.5+ln(1+\x);
}]
plot [domain=0:1, samples=74, smooth] ({fnx2(\x)},{fn2(\x,4837)});
\draw[line width = 1pt, red] (0,4.5) node(strt){} ..controls (0.5,4.5) and (0,1) .. (1,1);
\draw[line width = 1pt, blue] (1,1).. controls (1.2,1) and (1.5,1.5) .. (2.5,2.5) -- (5.2,2.5);
\end{tikzpicture}
\caption{The genus-2 spectral form factor as a function of time. After an initial decay, the signal displays a ramp and plateau at sufficiently late times. The ramp and the plateau are the average of the noisy signal. The $\sigma$-model captures the moments of this noisy signal.}
\label{fig.g2t}
\end{figure}
\subsection{A probe for the genus-2 partition function}
In this paper, we define an operator,
\begin{align} \label{Odefintro}
\langle i \, j \, k | \mathbb O | i' \, j' \, k'\rangle = c^*_{ijk} c_{i'j'k'} \,,
\end{align}
whose matrix elements encapsulate the OPE coefficients of the theory under investigation. $\ket{i},\ket{j},\ket{k}$ are energy eigenstates such that this operator can be thought of as a linear operator acting on a tripled Hilbert space. The genus-two partition function is then defined as\footnote{Here, we take a particular slice of the genus-2 moduli space where the three moduli are the same.}
\begin{equation}\label{eq.Genus-2.Partition.Function}
Z(\tau_1=\tau_2=\tau_3=i \beta) = {\rm Tr} \left[ \mathbb O e^{-\beta(H_1+H_2+H_3)} \right]\,,
\end{equation}
which is a ``thermal" one-point function on this tripled Hilbert space. Having turned the problem into an effective quantum-mechanical one, we use insights from operator statistics and random matrix universality to probe the genus-2 spectral form factor.
In chaotic quantum systems, aspects of operator statistics are captured by the Eigenstate Thermalization Hypothesis (ETH) \cite{deutsch1991quantum,srednicki1994chaos}. For simple operators $O$, the matrix elements have the general structure
\begin{equation}
\braket{n|O|m}= \delta_{mn} f(\bar{E}) + g(\bar{E},\delta E) e^{-S(\bar{E})/2} R_{mn} \,,
\end{equation}
where $\bar{E}$ and $\delta E$ are the mean energy and energy difference, and $f$ and $g$ are two smooth functions related to the microcanonical one- and two-point functions. The coefficients $R_{mn}$ are fixed in any given theory, but their statistical distribution is approximately gaussian random. For this reason, we call them pseudo-random variables. The ETH ansatz has not been proven, even though it can be verified numerically, for example in the holographically relevant case of the SYK model \cite{Sonner:2017hxc}. Analytical results pointing to the validity of ETH in this model have furthermore been established in \cite{Nayak:2019khe, Nayak:2019evx}. ETH has also been studied in higher dimensions, \cite{Lashkari:2016vgj, Dymarsky:2018ccu,Delacretaz:2020nit}, in particular in 2-dimensional CFTs where Virasoro symmetry introduces additional subtleties \cite{Fitzpatrick:2015zha, Basu:2017kzo, Romero-Bermudez:2018dim, Brehm:2018ipf, Hikida:2018khg, Guo:2018pvi, Maloney:2018yrz, Dymarsky:2018lhf, Dymarsky:2019etq, Datta:2019jeo, Besken:2019bsu}.
More generally, there is a symmetry-breaking argument (causal symmetry breaking) that describes the fine-grained structure of the energy spectrum as well as the associated eigenstates of a chaotic many-body system \cite{Altland:2020ccq}. The resulting effective field theory dictates both spectral and operator correlations (see Section \ref{sec.basics}). For the statistics of operator correlations in chaotic systems, in addition to random-matrix like correlations of energy eigenvalues, Haar unitary averages over the eigenstates contribute equally importantly \cite{Altland:2021rqn}. This lends credibility to a typicality argument in favor of the ETH \cite{deutsch1991quantum,srednicki1994chaos}: if we replace an energy eigenstate with a Haar-random state in a microcanonical window, a version of ETH follows. The jump to ETH is to assume that the energy eigenstates are as good as Haar-random states in a chaotic system since the Hamiltonian is close to a random matrix.\footnote{It is worthwhile to note that ETH also describes the physics of thermalization, which occurs at much earlier time scales than the random matrix phase. }
For chaotic CFTs, a generalization of ETH called the OPE Randomness Hypothesis (ORH) was proposed in \cite{Belin:2020hea}. It proposes to treat the OPE coefficients involving heavy (i.e. $\Delta\to\infty$) operators as pseudo-random variables. This includes the ETH type OPE coefficients $O_{mn}=c_{Omn}$ for some light operator $O$, but also makes a prediction for the structure of other types of OPE coefficients, in particular the OPE coefficients of three heavy operators. One of the aims of this paper is to provide evidence for the ORH using Haar-typicality and random matrix universality.
An important motivation behind the present work is also to illuminate our understanding of the AdS/CFT correspondence. In the dual gravitational theory, the semi-classical limit was long thought to only capture the early time decaying behaviour of physical observables like the auto-correlation, which ultimately leads to information loss \cite{Maldacena:2001kr}. Recent work for the nAdS${}_2$/nCFT${}_1$ correspondence between JT gravity and its matrix model dual demonstrated that the contribution of certain gravitational configurations known as Euclidean wormholes are capable of reproducing the linearly growing ramp-like behaviour \cite{Saad:2018bqo, Saad:2019lba}. A universal description in \cite{Altland:2020ccq} shows that the Euclidean wormholes capture the moments of the signals at sufficiently late times and hence reproduce the ramp-plateau behaviour. In higher dimensions, a detailed understanding of this picture is still being developed (see \cite{Cotler:2020ugk,Belin:2020hea} for gravity in AdS$_3$). In particular, it is not known what gravitational configurations reproduce the plateau behaviour. From \cite{Altland:2021rqn}, we have strong indications that the spacetime topology of these configurations should be the same as that of the wormhole solutions that explain the ramp.
From a wormhole perspective, the 2d genus-2 spectral form factor is an ideal observable to sharpen our understanding of holography. On the gravitational side, the relevant wormhole is called the genus-2 wormhole, stretching between two asymptotic genus-2 surfaces. Unlike the geometries relevant to compute the standard spectral form factor, the genus-2 wormhole is a true saddle of the gravitational equations of motion, and it can be perturbatively stable when embedded in certain top-down theories \cite{Maldacena:2004rf}. Precisely matching the expressions that we derive in this paper with the gravitational contribution of the genus-2 wormhole requires explicit knowledge of the statistical distribution of heavy-heavy-heavy OPE coefficients. Asymptotic formulas for these OPE coefficients are known, either for the Gaussian part \cite{Cardy:2017qhl,Collier:2019weq} or for the non-gaussian corrections \cite{Belin:2021ryy}. However, one can only trust that these asymptotic formulas correctly encode the statistical distribution of the OPE coefficients if one can argue that they are pseudo-random variables. The asymptotic formulas are blind to his fact, and one of the main goals of this paper is to explain to what extent the OPE coefficients have a random nature to them.
\subsection{Summary of Results}
The two main results we present in this paper are as follows. First, we provide evidence for the ORH based on random matrix theory. We show that the operator \rref{Odefintro} satisfies the following statistics. Its mean is given by\footnote{The overline notation $\overline{\phantom{a}\cdot\phantom{a}}$ means that we have effectively evaluated the matrix elements in random matrix theory. Thus, the mean and average we obtain are valid in the ergodic limit, so these results should hold for nearby matrix elements (defined by the Thouless time). Whether or not this extends to the full microcanonical window is a more difficult question that cannot be probed from ergodicity. Similar statements apply to the ETH (see for example \cite{Dymarsky:2018ccu,Richter:2020bkf}).}
\begin{align}
\label{eq.signal-same.intro}
\overline{\langle i_i i_2 i_3|\mathbb O |i'_1 i'_2 i'_3\rangle} &= \mathbb T{\rm r} \mathbb{O} ~ \frac1{D_1D_2D_3} ~ \delta_{i_1,i_1'}\delta_{i_2,i_2'}\delta_{i_3,i_3'} ~,
\end{align}
and its variance is
\begin{eqnarray}\label{eq.variance.intro}
\overline{\langle i_i i_2 i_3|\mathbb O |i'_1 i'_2 i'_3\rangle^2}- \overline{\langle i_i i_2 i_3|\mathbb O |i'_1 i'_2 i'_3\rangle}^2 ~ &\approx& ~ \left(\frac{1}{D_1 D_2 D_3}\right)^2 \Big[\ ~\mathbb T{\rm r} \mathbb{O} ~ \mathbb T{\rm r} \mathbb{O} \\
&\hspace{-5.75cm}+&\hspace{-3cm} \mathbb T{\rm r}_{\rm sunset} \mathbb{O}^2\Big(\delta_{i_1i'_1}\delta_{i_2i'_2} + \delta_{i_1i'_1}\delta_{i_3i'_3} + \delta_{i_3i'_3}\delta_{i_2i'_2} ~+ \delta_{i_1i'_1} + \delta_{i_2i'_2} + \delta_{i_3i'_3}\Big)\Big]~ \,. \notag
\end{eqnarray}
Here, $D_i$ are the dimensions of the three (microcanonical) Hilbert spaces which we have taken to be distinct for the three energy windows $i_1,i_2,i_3$,\footnote{More complicated expressions where all three windows are the same can be found in Section \ref{singlewindsec}.} and $\mathbb T{\rm r} \mathbb{O}$ and $\mathbb T{\rm r}_{\rm sunset} \mathbb{O}^2$ represent microcanonical traces. The sunset trace is a particular contraction given by $c_{ijk}c^{*}_{ijl}c_{mnl}c_{mnk}^{*}$.
The mean and variance of the matrix elements of $ \mathbb{O}$ are compatible with the OPE coefficients $c_{ijk}$ being random and approximately Gaussian random variables, as advocated by the OPE Randomness Hypothesis. For the distribution to be approximately Gaussian, $\mathbb T{\rm r}_{\rm sunset} \mathbb{O}^2$ must be exponentially suppressed compared to $(\mathbb T{\rm r} \mathbb{O})^2$. Here, this suppression is not built-in and must be taken as an additional assumption.\footnote{A similar assumption enters in ETH, where the microcanonical one- and two-point functions are assumed to be of the same order.} Note however that in two-dimensional CFTs, the sunset contraction has indeed be shown to be exponentially suppressed \cite{Belin:2021ryy}.
It is tempting to assemble the mean and variance of $ \mathbb{O}$ into a statistical formula with random variables as in ETH, which we do in \rref{resultdiffwindow2}. There are however crucial points to be kept in mind. The most important caveat is that the diagonal and random components of $c_{ijk}c_{lmn}^*$ are the same size. This is in fact expected from writing down a statistical ansatz for a composite structure such as $c_{ijk}c_{lmn}^*$. The composite nature of this object gives its variance a size comparable to its mean, which is expected from cross-Wick contractions that appear in the calculation of the variance. This is very similar to correlation functions of multi-trace operators in a generalized free theory, which contain connected components even though the underlying theory is Gaussian.
Our second result is the late-time behaviour of the genus-2 spectral form factor. It is most easily written in energy space, in terms of the resolvent (see \rref{eq.new-resolvent}). We find that the genus-2 spectral form factor resolvent is given by
\begin{equation}
R(s)= \frac{2\pi^2\rho^2(E)}{D^6} \left(\pi \delta(s) + 1 - \frac{\sin^2(s)}{s^2}\right) {\rm Tr} (\mathbb O)^2 ~
\end{equation}
where $s$ is the energy difference in units of mean level spacing of the tripled Hilbert space. This is the genus-2 version of the standard sine-kernel. The sin kernel is well-known for the standard spectral form factor, and predicts both a ramp and a plateau. We see that the behaviour is similar for the genus-2 spectral form factor. Here, it appears that spectral correlations are the dominant effect and the operator statistics do not really play a prominent role (this is why no sunset trace appears). Note however that the Thouless time, i.e. the time at which we can trust the random matrix behaviour may depend on the observable, and thus the correlations of the OPE coefficients can be important in determining the Thouless time for the genus-2 spectral form factor. Note that in any case, the ramp time which is of the order of inverse mean level spacing, is much larger here because of the enlarged dimensionality of the tripled Hilbert space.
The main assumption that we make in this paper is that random matrix universality applies to conformal field theories, i.e. that CFTs have a finite Thouless time. While this statement is expected to hold in arbitrary quantum systems (in fact at the Thouless time, even the locality of the quantum field theory has been washed away), it has not been proven and the Thouless time needs to be calculated theory by theory. It remains one of the big open problems to understand quantum chaos in QFTs and CFTs, and goes beyond the scope of this work.
This paper is organized as follows. In section \ref{sec.basics}, we review the EFT of quantum chaos and introduce the quantum mechanical model for the OPE coefficients in terms of a linear operator on a tripled Hilbert space. In section \ref{sec.ope}, we study Haar-averages over unitaries and determine the statistics of OPE coefficients making use of the linear operator $\mathbb O$. In section \ref{sec.g2}, we use the EFT of quantum chaos to study the genus-2 spectral form factor. We end with a conclusion and a discussion of open questions in section \ref{sec.discussion}.
\section{Our Setup and the EFT of Quantum Chaos}
\label{sec.basics}
In this section, we start with a review of the effective field theory that describes the ergodic regime of chaotic quantum systems \cite{Altland:2020ccq}. The quantum ergodic regime of a physical theory is defined as the one in which the behaviour of physical observables is indistinguishable from that in a random matrix theory (RMT).\footnote{The exact nature of which RMT itself depends on the time-reversal symmetry and the fermionic symmetries of the physical theory. This is classified in a 10-fold way by Altland and Zirnbauer, \cite{PhysRevB.55.1142}, building on the pioneering work of Wigner, Dyson and Mehta \cite{wigner1993characteristic,dyson1962statistical,mehta-book}.} More accurately, one should say that the behaviour is indistinguishable from a typical representative of the random matrix ensemble. Hence, it is possible to model a physical system using a random matrix ensemble in this limit. We then introduce a new formalism that is well suited to study the behaviour of OPE coefficients and the genus-2 partition function in the quantum ergodic limit. We will describe how to translate heavy-heavy-heavy OPE coefficients in terms of a standard quantum mechanical system. In subsequent sections, we use this formalism to probe the statistical behaviour of OPE coefficients and the genus-2 spectral form factor.
\subsection{The effective field theory of quantum chaos}
\label{sec.review}
Here, we give a brief overview of our main tool for studying the spectral probes in quantum theories: a symmetry-based effective field theory of quantum chaos described in \cite{Altland:2020ccq}. We refer the reader to \cite{Altland:2021rqn} for a more detailed exposition.
As we have emphasized a few times by now, chaos manifests itself in the fine structure of energy levels, and does so in a remarkably universal fashion \cite{wigner1993characteristic, dyson1962statistical, altland1997nonstandard}. One natural way to understand this universality is by identifying a symmetry and its associated breaking pattern, which gives rise to chaotic spectral correlations. Furthermore, in order to fully resolve the fine-grained structure of the energy spectrum and associated eigenfunctions, it turns out to be essential to work in a framework that is capable of controlling both perturbative (in $e^S$) {\it and} non-perturbative (in $e^S$) contributions to our quantities of interest. It is thus fortuitous that both these aims can be achieved by considering ratios of so-called spectral determinants, the first one of interest to the present discussion being
\begin{equation}\label{eq.Z4GenFunctional}
{\cal Z}^{(4)} \left( z_1\,\ldots ,z_4 \right) = \frac{{\rm Det} \left(z_1-H \right) {\rm Det} \left(z_2-H \right)}{{\rm Det} \left(z_3-H \right){\rm Det} \left(z_4-H \right)}
\end{equation}
where $H$ is the Hamiltonian of the chaotic quantum system in question, and $z_i$ are four different energies. As we will see in more detail below, the framework we are about to lay out here allows us to evaluate the spectral determinants above in terms of an effective field theory that becomes semi-classical for large $D = e^S$. Within this EFT we can then understand non-perturbative (in $e^S$) physics via the sum over saddle points, while perturbative (in $e^S$) physics comes from the perturbation theory around each individual saddle.
The utility of quantities like ${\cal Z}^{(4)}(\hat z)$ is that they allow us to generate insertions of the spectral density $\rho(z)$, via the relation $\mp i \pi \rho(z) = {\rm Im} {\rm Tr} \frac{1}{z-H}$, where
\begin{equation}
{\rm Tr} \frac{1}{z-H} = \frac{\partial_z {\rm Det}(z-H)}{{\rm Det}(z'-H)} \Biggr|_{z=z'}\,.
\end{equation}
This opens the door to study chaotic correlations of the spectral density. The main idea of the EFT of quantum chaos is to develop a field theory of this ratio of spectral determinants. It may be useful to note that this is a field theory that lives on the space of Hamiltonians, and its modes parametrize how close a given Hamiltonian is to being an RMT Hamiltonian. The crucial point, however, is the realization that such a theory necessarily possesses a continuous symmetry that reflects the fact that the four copies of the system Hamiltonian that enter \rref{eq.Z4GenFunctional} are by design identical. In order to make this symmetry apparent, as is described in detail in \cite{Altland:2020ccq} it is convenient to exponentiate the determinants via Gaussian integrals over bosonic/fermionic auxiliary variables (for determinants in denominator/numerator, respectively). The exact symmetry breaking pattern differs from one Altland-Zirnbauer symmetry class to another, but in the simplest case (the unitary class), it corresponds to a breaking of
\begin{equation}\label{eq.SymmetryBreaking}
G={\rm U}(2,2 | 2) \longrightarrow K={\rm U}(1|1)\times {\rm U}(1|1)\,,
\end{equation}
and the EFT of quantum chaos takes the form of a non-linear sigma model with target space $G/K$. The precise sense in which one should understand this symmetry breaking is via a mean-field solution of the field theory obtained by exponentiating our ratio of determinants
\begin{equation}\label{eq.DensityEFT}
\int [dQ] e^{-S[Q]} :=\left\langle \frac{{\rm Det} \left(z_1-H \right) {\rm Det} \left(z_2-H \right)}{{\rm Det} \left(z_3-H \right){\rm Det} \left(z_4-H \right)} \right\rangle_{\Delta E \,\, {\rm or}\,\, P(H),
\ldots}\,,
\end{equation}
where $Q$ is a collective degree of freedom parametrizing the Goldstone manifold \rref{eq.SymmetryBreaking} and which is defined in detail in \cite{Altland:2020ccq}. We indicated with angle brackets that the determinants are to be averaged over small (microcanonical) energy windows, or alternatively over a probability distribution of the Hamiltonian, or any other generic averaging procedures. The main point here is that the resulting physics is dictated by our symmetry breaking principle and does therefore not depend on the averaging procedure. It is then interesting to note that such a (slightly averaged) version of a spectral determinant can be defined for an \emph{individual} quantum system, for example via averaging over some small microcanonical energy window. In other words, from the EFT point of view, an individual quantum system gives rise to exactly the same IR EFT of quantum chaos as a full random-matrix ensemble.
This is, of course, nothing but a technical way of restating random-matrix universality: {\it an individual quantum chaotic system has spectral correlations that are well described by random matrix theory.}\footnote{Some work where the emergence of the ensembles has been discussed are: \cite{Anninos:2016szt, Balasubramanian:2020lux,Verlinde:2021kgt, Verlinde:2021jwu}.} This fundamental fact leaves its imprint on a variety of physical quantities. Admittedly, the EFT has its limitations and doesn't provide a description of arbitrary fine-grained observables. For this reason, while at late times, the EFT captures the moments of the noisy signal, it remains an open question to understand how far one can push this description to capture the entire signal. Moreover, the EFT also breaks down at low temperatures when the thermal sum is dominated by the contribution of the low energy states (those near the ground state).
In this paper, we are primarily concerned with chaotic conformal field theories, and in particular, the statistical distribution of their OPE coefficients. We are thus interested in applying the EFT approach outlined above to the operator \rref{eq.OPEoperator}. The main novelty is that we find it convenient to consider not just the energy density itself, but an ``operator-weighted" energy density, that is
\begin{equation}\label{eq.OperatorWeightedDensity}
\rho_{ O} (E) := {\rm Im} {\rm Tr} \left[ \frac{1}{E-H} O \right]~.
\end{equation}
It may already be obvious to the reader at this point that such a quantity gives us a way to define spectral sums weighted by operator matrix elements. In particular, with our purpose-built operator $\mathbb O$, the spectral sums are weighted by OPE coefficients: we precisely get the kind of quantities introduced in \rref{eq.Genus-2.Partition.Function}.
Furthermore, this is but a small extension of the ideas we just reviewed, and proceeds as follows. We start again by noting that
\begin{equation}
{\rm {Tr}} \left( \frac{1}{z-H} O \right) =\frac{\partial}{\partial h} \frac{{\rm Det}(z-H+ h\, O)}{{\rm Det}(z-H + h' O)} \Biggr|_{h=h'=0}\,.
\end{equation}
From this we can then proceed as above and define a field theory of sourced spectral determinants, starting from
\begin{equation}\label{eq.SourcedEFT}
\int [dQ] e^{-S[Q, h]} = \left\langle \frac{{\rm Det} \left(z_1-H + h_1 O \right) {\rm Det} \left(z_2-H + h_2 O \right)}{{\rm Det} \left(z_3-H + h_3 O \right){\rm Det} \left(z_4-H+ h_4 O \right)} \right\rangle_{\Delta E\,\,\, {\rm or}\,\, P(H), }
\end{equation}
where we have denoted all four possible sources by adding the parameter $h$ to the effective action. An important difference between the effective actions \eqref{eq.DensityEFT} and \eqref{eq.SourcedEFT} is the contribution of Haar average over unitary matrices. The equation \eqref{eq.DensityEFT} is invariant under unitary transformations of the Hamiltonian, $H\to U H U^{-1}$, and therefore only depends on the eigenvalues of the Hamiltonian. However, the insertion of the sources $O$ can break the unitary invariance giving rise to non-trivial contributions due to the Haar integrations. Haar averages are also central to the study of correlation functions in ensembles of typical states \cite{Pollack:2020gfa, Freivogel:2021ivu}.
In the following section, we represent the operator $\mathbb{O}$ in a fixed basis, and relative unitary rotations between this fixed basis (in which $\mathbb{O}$ is represented) and the eigenbasis of the Hamiltonian are averaged over using the Haar measure. However, not all observables depend on the Haar average over the unitaries. In \autoref{sec.ope}, we study specific matrix elements of these operators, akin to ETH. In this case the Haar average over the unitary matrices are crucial and play a dominant role in the behaviour of the operators under investigation. In \autoref{sec.g2}, we study the genus-2 spectral form factor. This observable is defined as a trace over the states of the Hilbert space and is insensitive to the Haar average over the unitaries. Both of these observables are important in their own way. The results of \autoref{sec.ope} show that the ORH should be interpreted in a way similar to ETH as discussed in \cite{DAlessio:2015qtq}. The results of \autoref{sec.g2} highlight that the underlying mechanism that determines the genus-2 spectral form factor is the same as the one giving rise to the ORH.
We now note that as before the limit of large Hilbert space dimension $D\gg 1$ allows us to evaluate the integral \rref{eq.SourcedEFT} by saddle point, each saddle giving rise to a perturbative expansion controlled by $D^{-1}=e^{-S}$. Furthermore, the physics of the relevant saddle points is once more controlled by the same causal symmetry breaking structure, so that we are again able to capture the universal chaotic EFT in each sector by a non-linear sigma model with target space $G/K$, as in \rref{eq.SymmetryBreaking} above. Note again that this is true for the simplest of the ten RMT ensembles, namely the unitary class, while the symmetry breaking and thus the sigma-model manifold for the remaining nine take a slightly different form.
\subsection{A tripled Hilbert space}
\label{sec.trip-Hil}
Having provided a pedagogical overview of the tools that we use for our purpose of probing quantum chaos, we move on to develop the formalism that is most useful to the study of CFTs. We start with defining a Hilbert space formed by taking three copies of our original CFT Hilbert space and considering the tensor-product Hilbert space. The advantage of this will become clear in a moment. We refer to it as the tripled Hilbert space,
\begin{equation}
\mathcal H^{\otimes3} = \mathcal H \otimes \mathcal H \otimes \mathcal H~.
\end{equation}
The Hamiltonian defined on the tripled Hilbert space is,
\begin{equation}\label{eq.tripHam}
H_{\textrm{full}} = H \otimes \mathds 1 \otimes \mathds 1 + \varkappa_1 \mathds 1 \otimes H \otimes \mathds 1 + \varkappa_2 \mathds 1 \otimes \mathds 1 \otimes H~,
\end{equation}
where, $H$ is the original CFT Hamiltonian. The parameters $\varkappa_i$ are introduced to be able to probe three different moduli in the genus-2 partition function \rref{genus2pf}. In this paper, we are mostly interested in the slice of moduli space where all three moduli are equal and thus often consider $\varkappa_i=1$. In appendix \ref{appA}, we discuss the spectrum of $H_{\rm full}$ and its relationship with the Hamiltonian of the original theory, $H$. We now consider a linear operator acting on this Hilbert space that has the property that one obtains the product of two CFT OPE coefficients when projected onto the energy eigenbasis,
\begin{align} \label{eq.OPEoperator}
\langle l \, m \, n| \mathbb O | i \, j \, k\rangle = c^*_{lmn}c_{ijk} \,.
\end{align}
This should be viewed as a formal definition of the operator, which we can do even without knowing the exact value of the heavy-heavy-heavy OPE coefficients. Nevertheless as we will see, ergodicity and RMT dictates the statistics of the coefficients of this operator $ \mathbb O$. This parallels nicely the situation for ETH, where the microscopic matrix elements are not known but one can derive their statistics from ergodicity, which essentially renders the simple operators Haar-random operators.
Here, and in all subsequent discussions in this paper, we are using the bold notation for traces in the tripled Hilbert space to distinguish it from the trace in the standard Hilbert space.
In fact, we have now translated our CFT question into a usual quantum mechanics setup, so there is no longer really any reason to distinguish our setup with standard questions related to ETH, up to some small differences that we now mention.
The two main differences are related to the structure of the Hilbert space and of the Hamiltonian. Let us focus on a microcanonical window centered at energy $E$.
The first point to note is that the size of the Hilbert space is given by
\begin{equation}
D^3 \,,\qquad D=e^{S(E)}=e^{2\pi \sqrt{\frac{c}{3}E}} \,.
\end{equation}
We have used the Cardy formula for the entropy \cite{cardyformula}. It is important \textit{not} to take the energy on the tripled Hilbert space which is $3E$, and think that the dimension of the Hilbert space is then $e^{S(3E)}$. In other words, one should be careful to use the Cardy formula on the original CFT Hilbert space and not on the tripled one.
The next difference is that in usual RMT, we would average over the Hamiltonian on the full Hilbert space. This is a random matrix of dimension $D^3$. Here, we see that in order to connect to a CFT calculation the relevant averaging should be done on a matrix of size $D$, which is then assembled into a bigger matrix following \rref{eq.tripHam}.
Finally, we would like to emphasize again that we have dropped the contribution of descendants, and are only keeping primary operators. One should thus view our setup as a toy-model for CFTs, which captures some but not all of the dynamics. Based on the sparseness of descendants with respect to primaries in CFTs with $d>2$, one expects our toy-model to accurately reflect the behaviour in those CFTs. In $d=2$, the presence of Virasoro symmetry induces extra subtleties, which deserve to be studied in more detail, but at large $c$, primaries should again dominate the dynamics. Moreover, we have not carefully imposed crossing symmetry of all heavy operators.
Keeping these issues in mind, we are now ready to proceed and analyze the consequences of ergodicity for the statistics of heavy operators in a CFT. This can be done thanks to an effective theory that we have reviewed in section \ref{sec.review}.
\subsection{The EFT description of random OPE coefficients}
\label{sec.eft.ope}
Let us now apply the ideas we have reviewed in section \ref{sec.review} to the formalism of section \ref{sec.trip-Hil}. For the operator that encapsulates the OPE coefficients, \eqref{eq.OPEoperator}, we study an observable that we call operator resolvent,
\begin{align}
\label{eq.new-resolvent}
R(\omega) &= \sum_{i,j} \langle i|\mathbb O | i \rangle \langle j|\mathbb O | j \rangle \delta(E_i-E_j-\omega)\nonumber \\
&= \int\!\!dE_1dE_2~\rho_{\mathbb O}(E_1) \rho_{\mathbb O}(E_2) \delta(E_1-E_2-\omega)~,\\[10pt]
\mathbb T{\rm r} [\mathbb O] \mathbb T{\rm r}[\mathbb O] &= \int \!\! d\omega R(\omega)~.
\end{align}
This is a variation of the observable that was studied in \cite{Altland:2021rqn}.
Rewriting the terms containing $\rho_{\mathbb O}(E)$ that appear in the integrand more carefully, we have,
\begin{align}\label{eq.op2pt}
\mathbb T{\rm r} [\mathbb O] \mathbb T{\rm r}[\mathbb O] &= \int \!\!dE_1\,dE_2 \ \rho_{\mathbb O}(E_1)\rho_{\mathbb O}(E_2) \nonumber \\
&= \frac1{2\pi^2}\,\mathfrak{Re}\int\!\!dE_1dE_2\,\Big(\mathbb T{\rm r}\left[G^+\!(E_1)\mathbb O\right] \mathbb T{\rm r}\left[G^-\!(E_2)\mathbb O\right] \nonumber\\
&\hspace{4cm}- \mathbb T{\rm r}\left[G^+\!(E_1)\mathbb O\right] \mathbb T{\rm r}\left[G^+\!(E_2)\mathbb O\right]\Big)\nonumber \\
&=: \frac1{2\pi^2} \int dEd\omega \, \Big(R^\pm(E,\omega) - R^{\ddag}(E,\omega)\Big)~.
\end{align}
The retarded/advanced Green's functions appearing above are defined by the identity,
\begin{equation}
G^\pm(E) = \frac1{E\pm i \varepsilon - H_{\rm full}}~.
\end{equation}
The resolvents that are labeled by the causality superscripts are defined by the equivalence of the second and the third line on the RHS of \eqref{eq.op2pt}. These are distinguished from the resolvent refined in \eqref{eq.new-resolvent} by the presence of superscripts as well as different number of arguments. The interesting physics\footnote{More specifically, the $\sigma$-model that governs that late time effective field theory of quantum chaos arises solely from this term. In order to get the full correct answer, including disconnected pieces the other contribution needs of course to be added as well.} arises from the first term, $R^\pm$, in \rref{eq.op2pt} above and we focus our interest on this object from now on. The above quantity is useful to compute a variety of physical observables. The static expression, as it appears in \eqref{eq.op2pt}, computes the variance of the expectation value of the OPE operator. When the integral over the mean energy, $E$, is restricted to a smaller microcanonical energy window, it reproduces the statistics of certain specific OPE coefficients instead. Furthermore, appropriate integral transforms with respect to the energy arguments of the expression gives rise to the genus-2 spectral form factor,
\begin{align}\label{eq.path.resolvents}
\left|Z[\tau_1,\tau_2,\tau_3]\right|^2 &= \frac1{2\pi^2}\int \!\!dE\, d\omega \, e^{-2\beta E+it\omega} \, \Big(R^\pm(E,\omega) - R^{\ddag}\!(E,\omega)\Big) \\[5pt]
&\text{where, } \tau_1=\tau_2=\tau_3=t+i\beta \nonumber~.
\end{align}
It now becomes apparent that the parameters, $\varkappa_i$, are related to the different cycles, $\tau_i$, of the genus-2 surface, see \autoref{fig.g2s}. The observables we are studying here are simply the ratio of determinants as described above \eqref{eq.SourcedEFT},
\begin{align}\label{eq.pathintwrite}
R^\pm &= -\mathfrak{Re}~ \partial_{h_+}\partial_{h_-} \left. \left\langle \frac{{\rm Det} \left(z^+_1-H_{\rm full} \right) {\rm Det} \left(z^-_2-H_{\rm full} \right)}{{\rm Det} \left(z^+_3-H_{\rm full} + h_+ \mathbb O \right){\rm Det} \left(z^-_4-H_{\rm full}+ h_- \mathbb O \right)} \right\rangle_{\rm avg.} \right|_{h_\pm=0}
\end{align}
In principle, we have not used any averaging prescription at this point and all the expressions are simply an exact rewriting. It is at this stage, once we represent the observables as determinant operators that we perform some averaging prescription as relevant to the problem.
This ratio of determinants can be expressed in a path integral representation over certain graded fields, $\Psi, \bar\Psi$,
\begin{align}\label{eq.OpDet.path}
\left\langle \frac{{\rm Det} \left(z^+_1-H_{\rm full} \right) {\rm Det} \left(z^-_2-H_{\rm full} \right)}{{\rm Det} \left(z^+_3-H_{\rm full} + h_+ \mathbb O \right){\rm Det} \left(z^-_4-H_{\rm full}+ h_- \mathbb O \right)} \right\rangle_{\rm avg.}& \nonumber\\
&\hspace{-3cm}= \left\langle\! \int \!\! D\bar \Psi D\Psi \ \exp\left[i\bar \Psi \left( \hat z-\!H_{\rm full}\!-\hat h \right) \Psi\right]\right\rangle_{\rm avg.}~.
\end{align}
In the above path integral, we have introduced a component of $\Psi$ fields for each determinant insertion. The determinant insertions in the numerator correspond to fermionic variables while those in the denominator correspond to bosonic variables. Hence the $\Psi$ field has a graded $U(1,1|2)$ structure.\footnote{Understanding the pseudo-Riemannian structure is slightly subtle and arises from having convergent bosonic integrals. See \cite{Altland:2021rqn} for detailed discussion.} Thereby, we have a further enlarged Hilbert space corresponding to the direct product of retarded/advanced (RA), graded boson/fermion (bf) and the tripled Hilbert space. The matrix, $\hat z$, corresponds to energies at which the observables are being probed and $\hat h$ are the source terms,
\begin{eqnarray}
\hat z &=&\underbrace{ \bigg(E\,\mathds 1 + \left(\frac\omega2 + i \varepsilon \right) \sigma_3\bigg)^{\rm RA}}_{\mathrm{advanced/retarded}} \quad\otimes
\underbrace{\quad\mathds 1^{\rm bf}\quad}_{\mathrm{boson/fermion}} \quad\otimes \underbrace{\quad \mathds 1^{\otimes3}\quad}_{\mathrm{tripled}\,\,\mathrm{Hilbert}\,\,\mathrm{space}} ~,\\
\hat h &=& \begin{pmatrix} h_+&0\\0&0\end{pmatrix} \otimes P_b \otimes \mathbb O + \begin{pmatrix} 0&0\\0&h_-\end{pmatrix} \otimes P_b \otimes \mathbb O^\dagger~.
\end{eqnarray}
In the first line we have made the tensor product and graded structure of our theory completely explicit. The source terms in the second line are written with this structure in mind. The projector, $P_b$, projects onto the bosonic sector in which the sources are introduced in \eqref{eq.pathintwrite}.
\subsection{Validity of the EFT and Upshot of the Analysis}
To summarize, the EFT provides a way to probe the ergodic limit of chaotic quantum systems. This should also apply to chaotic CFTs: there exists a quantum ergodic regime in which the physical Hamiltonian can be {\it modelled} by an ensemble over Hamiltonians, $H$, chosen from the appropriate Altland-Zirnbauer symmetry class (unitary in the context of the present work). A requirement for the EFT to hold is that it has a finite regime of applicability. In terms of time scales, this is often captured by the Thouless time: the time after which RMT universality kicks in. It is both theory and observable dependent. The main assumption we are making in this paper is that the Thouless time for the thermal trace of the linear operator $\mathbb{O}$ is finite (and parametrically smaller than the Heisenberg time given by the inverse mean-level spacing). As long as this is true, we can apply our EFT and the framework presented above holds. To really show that the Thouless time is finite, one should carefully integrate out the massive modes as is akin to effective theories and show that the EFT becomes a theory of random Hamiltonians which implements Haar averages. This has been carried out in the SYK model \cite{Altland:2017eao,Altland:2020ccq,Altland:2021rqn}, but a general derivation applicable to arbitrary chaotic systems (including CFTs) is beyond the scope of this work.
We consider two main applications of this EFT. First, we saw that it implements Haar-averages between operator bases and the eigenbases of the Hamiltonian. This procedure should be done within some microcanonical window. In \autoref{sec.ope}, we explicitly preform these Haar averages on $\mathbb{O}$ to study the statistics of OPE coefficients. Since we have a tripled Hilbert space structure, we really have three microcanonical windows which can be taken to be the same or different. Our analysis produces statistics that match the expectations of the OPE Randomness Hypothesis in the ergodic regime. One should remember that while the Haar averages act on the whole microcanonical window, one should only really trust correlations of matrix elements that are close enough in energy space.\footnote{This does not constrain the three microcanonical windows to be close to one another, but simply says that the indices $i,i'$ in $C_{ijk}C^*_{i'j'k'}$ should be close in energy space. The indices $i,j$ are allowed to lie in different microcanonical windows that are very far away in energy space.} As stated above, as long as the Thouless time is finite, the results we present are on solid ground.
The second application is to calculate the genus-2 spectral form factor in the ergodic limit. In \autoref{sec.g2}, we make direct use of the EFT to do so. We find a ramp and a plateau.
\section{Statistics of OPE coefficients}
\label{sec.ope}
As we saw above in section \ref{sec.review}, the EFT of quantum chaos impacts matrix elements of some operator $O$ in a particular way: it essentially implements a Haar average over unitaries between the eigenbasis of the Hamiltonian and that of $O$. Note that this is the logic behind the ETH ansatz in the first place: simple operators are indistinguishable from a typical operator chosen from a Gaussian distribution of operators in a microcanonical window. By performing unitary averages, one can essentially derive the ETH ansatz (see for example \cite{DAlessio:2016rwt}). The role of the EFT is to \textit{derive} that such unitary averages capture the right physics. In this section, we study the implication of these statements for our linear operator $\mathbb{O}$ on the tripled Hilbert space.
Recall, from \eqref{eq.OPEoperator}, that we can rewrite the product of OPE coefficients as specific elements of the operator $\mathbb O$,\footnote{In a CFT, OPE coefficients have the conjugation property $c_{OOO}=c^*_{O^{\dagger}O^{\dagger}O^{\dagger}}$. This structure must be preserved by the Haar averages, and could influence the RMT universality class of CFTs. We will not be careful about this fact and simply apply unitary averages, but we discuss this point further in the discussion section. }
\begin{align}\label{eq.csqO}
c^*_{lmn}c_{ijk} &= \bra{l \, m \, n} \mathbb O \ket{i \, j \, k}\,,
\end{align}
where we have lightened the notation a little, denoting energy eigenstates states $\ket{i \, j \, k}$. As we see below, it is important to distinguish whether the triplet of states $\ket{i},\ket{j},\ket{k}$ are in the same microcanonical window or not. We discuss both cases, and start with the simpler case where all three states lie in different microcanonical windows.
\subsection{Different microcanonical ensembles}
To start, note that the EFT of quantum chaos states that the energy eigenstates have a random, uncorrelated projection along some fixed vector(s) in the Hilbert space \cite{mehta-book}. Let us fix some basis for the tripled Hilbert space and the particular choice of basis we make is not important since the energy eigenstates are randomized. We understand the states $\ket{j_{1,2,3}}$ to be vectors in this fixed basis, to distinguish with the eigenbasis of the Hamiltonian which we label by $\ket{i_{1,2,3}}$. In this basis, we can write the operator as
\begin{equation}
\mathbb{O} =\sum_{j_1,j_2,j_3,j_1',j_2',j_3'} \Omega_{j_1j_2j_3}\Omega_{j_1'j_2'j_3'}^* \ket{j_1j_2j_3} \bra{j_1'j_2'j_3'} \,,
\end{equation}
We now write the matrix elements of $\mathbb{O}$ in energy eigenstates as
\begin{align}
\label{eq.diff-micro}
\langle i_i i_2 i_3|\mathbb O |i'_1 i'_2 i'_3\rangle &=\sum_{j_1,j_2,j_3,j_1',j_2',j_3'}\Omega_{j_1j_2j_3}\Omega_{j_1'j_2'j_3'} \langle i_i i_2 i_3 |j_1j_2j_3\rangle \langle j_1'j_2'j_3' |i'_1 i'_2 i'_3\rangle \nonumber\\
&= \sum_{j_1,j_2,j_3,j_1',j_2',j_3'}\Omega_{j_1j_2j_3}\Omega_{j_1'j_2'j_3'} U_{i_1j_1}^{(1)} U_{i_2j_2}^{(2)} U_{i_3j_3}^{(3)} U_{j_1' i'_1}^{(1)\dagger} U_{j_2' i'_2}^{(2)\dagger} U_{j_3' i'_3}^{(3)\dagger}
\end{align}
The $U^{(k)}$ are unitaries that implement a change of basis between the fixed basis $\ket{j_k}$ and the eigenstate of the Hamiltonian, $\ket{i_k}$, within a microcanonical window. Note that we have three \textit{different} unitaries $U^{(k)}$, because we are taking the three energy $i_1$, $i_2$, $i_3$ to lie in three different microcanonical windows.
\subsubsection*{Statistical mean of operators}
Ergodicity now implies that the unitaries are random, which means we can integrate over them with the Haar measure. Performing the Haar average over the product of the unitaries gives \cite{Collins2003, collins2006integration},
\begin{align}
\label{eq.signal-diff}
\overline{\langle i_i i_2 i_3|\mathbb O |i'_1 i'_2 i'_3\rangle} &=\sum_{j_1,j_2,j_3,j_1',j_2',j_3'}\Omega_{j_1j_2j_3}\Omega_{j_1'j_2'j_3'} \int \prod_{k=1}^{3} \left[dU^{(k)}\right] U_{i_1j_1}^{(1)} U_{i_2j_2}^{(2)} U_{i_3j_3}^{(3)} U_{j_1' i'_1}^{(1)\dagger} U_{j_2' i'_2}^{(2)\dagger} U_{j_3' i'_3}^{(3)\dagger} \nonumber \\
&= \sum_{j_1,j_2,j_3}| \Omega_{j_1,j_2,j_3}|^2 ~~\prod_{m=1}^3 {\rm Wg}(D_i,1,\mathds 1) \delta_{i_mi'_m}
\end{align}
The first argument of the Weingarten functions, Wg, is the dimensionality of the unitary operators. The dimensionality of the unitary operators is the same as that of the microcanonical Hilbert space and are called $D_i$s.
The second argument is $n$, the number of unitary operators inserted in the Haar integral (not counting the daggered-unitary insertions). The third argument is an element of the permutation group, $S_n$. $\mathds 1\in S_n$ corresponds to the identity element that doesn't permute any element in the set. In this case, we simply have
\begin{equation}
{\rm Wg}(D_i,1,\mathds 1)=\frac{1}{D_i} \,.
\end{equation}
There are several remarkable features that have happened, once the Haar average has been done. First, note that the coefficients are now purely diagonal, the off-diagonal elements having been forced to zero upon doing the Haar average. Second, the only dependence on $\mathbb{O}$ that is left is the term $ \sum_{j_1,j_2,j_3}| \Omega_{j_1,j_2,j_3}|^2$, but this is simply the trace of the operator in the corresponding microcanonical windows.\footnote{We have a microcanonical trace because the indices, $j_1,j_2,j_3$, only take values corresponding to the states that lie in the relevant microcanonical window.} All dependence on the individual $\Omega_{j_1,j_2,j_3}$ or on the choice of basis $\ket{j_{1,2,3}}$ has been washed out. Since the trace is basis independent, it is most convenient to write it in the energy basis and we thus have that the coefficient of the Kronecker delta is
\begin{equation}
\frac{1}{D_1 D_2 D_3 }\mathbb T{\rm r} \mathbb{O} = \frac{1}{D_1 D_2 D_3 }\sum_{ijk} |c_{ijk}|^2 \,,
\end{equation}
which is the averaged sum over OPE coefficients squared. This is \textit{coarse-grained} information which we have access to without knowing the detailed microscopics. In ETH, this would correspond to the diagonal smooth function $f(\bar{E})$.
Just like in ETH, this function of the mean energies is not fixed from first principles, and can be theory dependent. For CFTs in general dimensions, it is currently unknown how to constrain the heavy-heavy-heavy OPE coefficients (this is explored in \cite{6ptasymptotics}). However, for $d=2$, Virasoro symmetry fixes the averaged OPE coefficients. This is a result of modular invariance of the genus-2 partition function. Explicit expressions for the averaged OPE coefficients can be found in \cite{Cardy:2017qhl,Collier:2019weq}.
It is worth mentioning that the OPE coefficients are invariant under the permutation of the three indices, and so the statistics should be built in a way that preserves this symmetry. This is a minor technical step that we have not taken into account in the presentation so far, for the sake of clarity. It would be straightforward to do so, but we leave this procedure as being done implicitly to avoid cluttering expressions.
\subsubsection*{Statistical variance of operators}
Having obtained the mean of the matrix elements of $\mathbb{O}$, we now proceed to study their statistical variance. We thus wish to compute
\begin{equation}
\overline{\langle i_i i_2 i_3|\mathbb O |i'_1 i'_2 i'_3\rangle^2}- \overline{\langle i_i i_2 i_3|\mathbb O |i'_1 i'_2 i'_3\rangle}^2~,
\end{equation}
which can be done in a similar fashion by an appropriate integration over the unitary matrices using the Haar measure. We give the explicit and detailed computation in appendix \ref{app.variance}. The leading order contribution to the variance is given by
\begin{eqnarray}\label{eq.variance}
\overline{\langle i_i i_2 i_3|\mathbb O |i'_1 i'_2 i'_3\rangle^2}- \overline{\langle i_i i_2 i_3|\mathbb O |i'_1 i'_2 i'_3\rangle}^2 ~ &\approx& ~ \left(\frac{1}{D_1 D_2 D_3}\right)^2 \Big[\ ~\mathbb T{\rm r} \mathbb{O} ~ \mathbb T{\rm r} \mathbb{O} \\
&\hspace{-5.75cm}+&\hspace{-3cm} \mathbb T{\rm r}_{\rm sunset} \mathbb{O}^2\Big(\delta_{i_1i'_1}\delta_{i_2i'_2} + \delta_{i_1i'_1}\delta_{i_3i'_3} + \delta_{i_3i'_3}\delta_{i_2i'_2} ~+ \delta_{i_1i'_1} + \delta_{i_2i'_2} + \delta_{i_3i'_3}\Big)\Big]~ \,, \notag
\end{eqnarray}
where we have defined the sunset trace contraction as
\begin{equation}\label{eq.sunset}
\mathbb T{\rm r}_{\rm sunset} \mathbb{O}^2=c_{j_1j_2j_3}c^*_{j_1j_2k_3} ~ c_{k_1k_2k_3}c^*_{k_1k_2j_3} \,.
\end{equation}
A diagram of this contraction is depicted in \autoref{fig.sunset}.
\begin{figure}[t]
\centering
\begin{tikzpicture}
\draw[line width = 1pt] (0,0) node(st1) {$*\quad$} ..controls (1,1) and (3,1) .. (4,0) ..controls (3,-1) and (1,-1) .. cycle;
\draw[line width = 1pt] (0,2) ..controls (1,3) and (3,3) .. (4,2) node(st2) {$\quad*$} ..controls (3,1) and (1,1) .. cycle;
\draw[line width = 1pt] (0,0) -- (0,2);
\draw[line width = 1pt] (4,0) -- (4,2);
\end{tikzpicture}
\caption{Sunset diagrams are the diagrammatic representation of the contractions depicted in \eqref{eq.sunset}. Each vertex of the diagram represents an insertion of a OPE coefficient. The ones corresponding to the insertions of $c^*$s are marked with a star, while the unmarked vertices are the $c$ insertions. Each edge in the diagram corresponds to the index contractions between OPE coefficients.}
\label{fig.sunset}
\end{figure}
This particular sum of OPE coefficients appears in the sunset channel decomposition of a genus-3 partition function, hence the name we have chosen for the trace.
It is important to note that the mean is as large as the variance (and for that matter all the subsequent moments of the operator $\mathbb O$). Hence, we can not apply the central limit theorem and write the operator $\mathbb O$ using random variables which are approximately Gaussian. If we wish to write an ansatz for $\mathbb O$ which resembles ETH, we can write
\begin{align} \label{resultdiffwindow2}
c^*_{i_1i_2i_3} c_{i'_1i'_2i'_3} \approx \frac{\mathbb T{\rm r} \mathbb{O}}{D_1 D_2 D_3} (\delta_{i_1,i_1'}\delta_{i_2,i_2'}\delta_{i_3,i_3'}+R_{i_1,i_2,i_3,i_1',i_2',i_3'}) \nonumber \\ + \frac{\sqrt{ \mathbb T{\rm r}_{\rm sunset} \mathbb{O}^2}}{D_1 D_2 D_3}\textrm{Sym} &\left[\delta_{i_1,i_1'} R_{i_2,i_3,i_2',i_3'}+\delta_{i_1,i_1'}\delta_{i_2,i_2'} R_{i_3,i_3'} \right] \,,
\end{align}
where the $\textrm{Sym}$ stands for $S_3$ symmetrizing between the indices 1, 2 and
3, and simultaneously acting on 1', 2' and 3' in the same way. We must however be very careful with what we mean by such an equation.
Contrary to ETH, the variables $R$ above are \textit{not} approximately Gaussian random variables. They are statistical variables that have zero mean and unit variance, but they have non-trivial higher moments that are of the same size as the first and second moments. The reason for this is that we are trying to write random variables directly for the operator $\mathbb O$, which is really a composite tensor made out of two $c_{ijk}$. It is the $c_{ijk}$ that are approximately Gaussian random, and the non-trivial random matrix in the diagonal $R_{i_1,i_2,i_3,i_1',i_2',i_3'}$, which is the same size as the mean, comes from the fact that there is a cross Wick contraction in the variance of $\mathbb O$ which involves four $c_{ijk}$. One can make an intuitive parallel with large $N$ factorization in large $N$ gauges theories. At infinite $N$, the theory is a generalized free theory and all correlation functions are obtained from Wick contractions. However, connected correlation function of multi-trace operators are non-trivial. This does not mean that the underlying theory is not Gaussian, a free theory is obviously Gaussian, but simply that we are considering composite operators. The same logic applies to \rref{resultdiffwindow2}.
To sum up, it is worth emphasizing what \rref{resultdiffwindow2} really means. The microscopic definition of the operator $\mathbb{O}$ was given in \rref{eq.csqO} and seems in tension with the result \eqref{resultdiffwindow2}. One should view these results in the same way as one views ETH: there is a microscopic definition of the matrix elements, and there is the ETH ansatz which is true in a statistical sense. It should be understood as statistically true, once sampled over enough individual coefficients. While the general structure of our result follows closely that of matrix elements in the ETH, there are also important distinctions to take into account, in particular the composite nature of the operator $\mathbb O$. We discuss them in section \ref{ETHcomp}, but we first give the general expression when all microcanonical windows are the same.
\subsection{A single microcanonical ensemble \label{singlewindsec} }
In this section we analyse the behaviour of the operator elements when the external states belong to the same microcanonical window of energies. In this case, all unitaries that enter in \rref{eq.signal-diff} are the same. Once integrated using the Haar measure, they produce new structure because more Wick contractions are allowed. Let us start once more by calculating the mean value of operators.
\subsubsection*{Statistical mean of operators}
We have
\begin{align}
\label{eq.same-micro}
\langle i_i i_2 i_3|\mathbb O |i'_1 i'_2 i'_3\rangle &=\sum_{j_1,j_2,j_3,j_1',j_2',j_3'}\Omega_{j_1j_2j_3}\Omega_{j_1'j_2'j_3'} U_{i_1j_1} U_{i_2j_2} U_{i_3j_3} U_{j_1' i'_1}^{\dagger} U_{j_2' i'_2}^{\dagger} U_{j_3' i'_3}^{\dagger}
\end{align}
The Haar integration over $U$ is more complicated but the final result is
\begin{align}
\label{eq.signal-same}
\overline{\langle i_i i_2 i_3|\mathbb O |i'_1 i'_2 i'_3\rangle} &= \mathbb T{\rm r} \mathbb{O} ~ \Big( {\rm Wg}(D_1,3,\mathds 1) + 3{\rm Wg}(D_1,3,2\cdot1) + 2{\rm Wg}(D_1,3, 3) \Big) \nonumber \\
& \qquad \bigg(\delta_{EE'} + \delta_{i_1i'_2}\delta_{i_2i'_3}\delta_{i_3i'_1} + \delta_{i_1i'_3}\delta_{i_2i'_1}\delta_{i_3i'_2} \nonumber \\
&\qquad \qquad + \delta_{i_1i'_1}\delta_{i_3i'_2}\delta_{i_2i'_3} + \delta_{i_1i'_3}\delta_{i_2i'_2}\delta_{i_3i'_1} + \delta_{i_1i'_2}\delta_{i_2i'_1}\delta_{i_3i'_3} \bigg) \nonumber \\
&= \mathbb T{\rm r} \mathbb{O} ~ \frac1{D_1(D_1+1)(D_1+2)} ~ [\delta_{EE'}]_{S_3} ~.
\end{align}
Note that in this case we have both diagonal as well as ``off-diagonal" terms. However, the off-diagonal terms are essentially the permutations of three indices labelling the OPE coefficients, so they are rather trivial. To avoid clutter, we denote the $S_3$ symmetrised terms in the primed coordinates by a subscript, $[\delta_{EE'}]_{S_3}$.
To leading order in the large $D_1$ limit, the coefficient of the delta function is simply $\mathbb T{\rm r}\mathbb{O}/D_1^3$ which agrees with \eqref{eq.signal-diff}, once we have set the three windows to be the same size. This makes sense, we wouldn't expect there to be a big different between the same and distinct microcanonical windows. There are only two small differences: first, we see that the indices are symmetrized, which is natural in this case. Second, we see that there are further exponentially suppressed corrections to the leading coefficient.
\subsubsection*{Statistical variance of operators}
Once again, to be able to describe the random nature of the matrix elements of $\mathbb O$, we need to compute their variance. In the present scenario, we find
\begin{align}\label{eq.variance.same}
\overline{\langle i_i i_2 i_3|\mathbb O |i'_1 i'_2 i'_3\rangle^2}- \overline{\langle i_i i_2 i_3|\mathbb O |i'_1 i'_2 i'_3\rangle}^2&\nonumber\\
&\hspace{-3.25cm}\approx {\rm Wg}(D_1,6,\mathds 1)\bigg[ ~ \mathbb T{\rm r}\mathbb O \mathbb T{\rm r}\mathbb O \left( 1+\delta_{i_1i_2} + \delta_{i_2i_3} + \delta_{i_3i_1} + \delta_{i_1i_2} \delta_{i_2i_3} \delta_{i_3i_1} \right) \times (\cdots)' \nonumber \\
&\hspace{-0.5cm}+ ~ \mathbb T{\rm r}_{\rm sunset}[\mathbb O^2] \Big(\delta_{i_1i'_1}\delta_{i_2i'_2} + \cdots 647 \text{ terms } \cdots \Big) \bigg]~.
\end{align}
The term $(\cdots)'$ represents the same parenthesis that precedes it, but with primed indices. The 647 other terms correspond to all possible combinations of delta functions and symmetrizations thereof. The gory details are given in appendix \ref{haarsame}.
Therefore the statistical properties of the heavy-heavy-heavy OPE coefficients as encapsulated by the operator, $\mathbb O$, can be effectively given written as,
\begin{align} \label{Oonewind}
c^*_{i_1i_2i_3} c_{i'_1i'_2i'_3} &\approx \frac{\mathbb T{\rm r} \mathbb{O}}{D_1^3} (\textrm{Sym}_{S_3}[\delta_{i_1,i_1'}\delta_{i_2,i_2'}\delta_{i_3,i_3'}]+R^s_{i_1,i_2,i_3,i_1',i_2',i_3'} \nonumber \\
&\quad +\frac{\sqrt{ \mathbb T{\rm r}_{\rm sunset} \mathbb{O}^2}}{D_1^3}\textrm{Sym}_{\textrm{full}} \left[\delta_{i_1,i_1'} R_{i_2,i_3,i_2',i_3'} +\delta_{i_1,i_1'}\delta_{i_2,i_2'} R_{i_3,i_3'} \right] \,,
\end{align}
where $R^s$ is now a symmetric random tensor and $\textrm{Sym}_{\textrm{full}}$ is now fully symmetrizing in all indices using the symmetric group $S_6$. Once again, one must be careful with the interpretation of this formula, the statistical variables $R$ are not approximately Gaussian random, but their mean vanishes and they have unit variance.
We see that this is a more sophisticated symmetrized version of \rref{resultdiffwindow2}, but other than that it shares the same general structure. This gives us a measure of the variance of the OPE coefficients as conjectured in \cite{Belin:2020hea}. We now move to discussing the difference and similarities between $\mathbb{O}$ and an operator in ETH.
\subsection{Comparison with conventional ETH \label{ETHcomp}}
The framework we have developed to study the heavy-heavy-heavy OPE coefficients naturally shares many features with conventional ETH where the objects of study are the matrix elements of operators. This follows from the fact that we have defined a linear operator on a tripled Hilbert space. Nevertheless, there are also a few differences that we would like to underline.
The first difference we have come across is the fact that the diagonal and off-diagonal matrix elements of $\mathbb{O}$ are of the same size. This should be contrasted with ETH, where the diagonal contribution given by the microcanonical one-point function is expected to be much larger.\footnote{Note that this separation between diagonal and off-diagonal parts is not universal and in particular in 2d CFTs, the diagonal term is kinematically exponentially suppressed because one-point functions of primary operators vanish on the cylinder.} The reason for this is that the operator $\mathbb{O}$ is not the most general operator that acts on a Hilbert space of size $D^3$,\footnote{In fact, it is a projector. Had we picked the fixed basis to be the eigenbasis of $\mathbb{O}$, we would have had a single non-zero eigenvalue, which shows the operator is very constrained.} but rather is a composite operator made out of two tensors $c_{ijk}$. It is thus specified by $D^3$ numbers, rather than a generic operator which would require $D^6$ terms. Comparing to ETH, this would be similar to defining a vector $v$ such that
\begin{equation}
R_{mn}=v_m v_n^* \,,
\end{equation}
Clearly, the most general matrix is not of this type. Now, if $v$ effectively becomes averaged over, the diagonal and off-diagonal elements are on different footing since the diagonal elements are real and positive numbers. Something similar is happening for us with the operator $\mathbb{O}$.
The fact that the off-diagonal matrix elements are of the same size as the diagonal ones should however be seen as a feature, not a bug. The ergodic nature of the OPE coefficients is consistent with the observed behaviour of the operator $\mathbb O$, that can be regarded as a composite operator in terms of the OPE coefficients $c_{ijk}$. The Gaussian nature of the $c_{ijk}$ imply ``large" higher moments for $\mathbb{O}$. This is why the variance was the same order as the mean, because there is a non-trivial Wick contraction in the variance of $\mathbb{O}$, which contains four $c_{ijk}$. The same is true for higher moments, that all have the same size as the mean. We would like to emphasize that this is not inconsistent with the $c_{ijk}$ being Gaussian, but rather happens precisely because they are!
A second difference with ETH comes from the size of the true non-Gaussianities, which in our case are measured by $\sqrt{ \mathbb T{\rm r}_{\rm sunset} \mathbb{O}^2}$. In ETH, one finds that the variance is suppressed by $e^{-S/2}$ as long as the microcanonical two-point function is of the same order as the microcanonical one-point function. Here, the condition for the true non-Gaussianities to be small is that
\begin{equation}
\sqrt{ \mathbb T{\rm r}_{\rm sunset} \mathbb{O}^2} \ll \mathbb T{\rm r} \mathbb{O} \,.
\end{equation}
This can be viewed as the condition for the OPE coefficients to be approximately Gaussian random variables. It is currently unknown if this is true in all CFTs, but it has been proven in $d=2$ \cite{Belin:2021ryy}. It is however important to underline that it is \textit{not} an output of our analysis, but rather has to be assumed as an extra condition.
Finally, we would like to emphasize that we did not do a $D^3$ unitary average over the basis of the Hamiltonian, which is what would be required if we had naively applied standard ETH to $\mathbb O$ with a random Hamiltonian of size $D^3$. If we would have done this larger unitary average, $\mathbb O$ would have behavd as
\begin{align}
\langle E|\mathbb O| E'\rangle \approx \frac{\mathbb T{\rm r} \mathbb O}{D^3} \Big(\delta_{EE'} + R_{EE'}\Big)~,
\end{align}
which misses the contributions of the sunset diagrams. As we now see, observables like the genus-2 spectral form factor do not receive contributions from the Haar-average over the unitary operators, and thus do not actually perceive the triple tensor-product structure of the Hilbert space. We discuss this in further in the following section, as we now turn to the study of the genus-2 spectral form factor using the EFT of quantum chaos.
\section{EFT for the Genus-2 Spectral Form Factor}
\label{sec.g2}
In this section we change gears and compute the genus-2 spectral form factor, \eqref{genus2sffintro},
\begin{align} \label{genus2sffintro.sec4}
F(t_1,t_2,t_3) &= |Z(\tau_1+it_1,\tau_2+it_2,\tau_3+it_3) |^2 \,,\\
&Z(\tau_1,\tau_2,\tau_3) = \sum_{O_1,O_2,O_3} |C_{123}|^2 q_1^{\Delta_1} q_2^{\Delta_1} q_3^{\Delta_3} \,, \qquad q_i=e^{2\pi i \tau_i} \,.
\label{eq.g2.sum.explct}
\end{align}
As defined, the genus-2 spectral form factor involves a sum over all the operators of the CFT. Because of the state-operator correspondence, this amounts to a sum over the entire Hilbert space of the theory. Nevertheless, at sufficiently high ``temperatures" (i.e. for small moduli) the dominant contribution arises from the high energy part of the spectrum. For holographic CFTs, as long as one is above the Hawking-Page temperature, this would probe the chaotic sector of the spectrum. Alternatively, one could restrict the sum in \eqref{eq.g2.sum.explct} to a microcanonical window within this high energy spectrum. This microcanonical genus-2 spectral form factor would capture the chaotic physics that we are interested in equally well. For simplicity, we consider the sum over the entire Hilbert space of our theory, while pointing out the differences with a microcanonical version of the spectral form factor when relevant.\footnote{Note that we are considering CFTs, so formally the dimension of the full Hilbert space $D=\infty$. In the end, we study a fixed microcanonical window of size $D_1$ which in practice corresponds to setting $D=D_1$. We keep $D$ general is what follows, see the discussion after \rref{eq.HHA}.}
In section \ref{sec.eft.ope}, we discussed how one can write the genus-2 spectral form factor as a correlation function of a linear operator that acts on the tripled Hilbert space. This facilitates a rewriting of the genus-2 partition function as the path integral
\begin{align}\label{eq.resolvent.path}
R^\pm &= - \mathfrak{Re} \left[\partial_{h_+}\partial_{h_-} Z[\hat h]\right], \nonumber \\
Z[\hat h] &= \left\langle \int \!\! D\bar \Psi D\Psi \ \exp\left[i\bar \Psi \cdot \left( \hat z-H_{\rm full}-\hat h \right)\cdot \Psi\right] \right\rangle_{\rm avg.}~.
\end{align}
An important difference from the observables studied in the preceding sections is that the genus-2 spectral form factor involves traces over either the entire Hilbert space or a part of it - an appropriate microcanonical window. The cyclicity of traces implies that the unitaries that rotate the physical states of the Hilbert space cancel against their complex conjugates. Consequently, the genus-2 spectral form factor only depends on the statistics of the eigenvalues in the ergodic limit but not on that of the eigenstates. As we shall see below, this implies that in the quantum ergodic limit, the information about the product nature of the underlying Hilbert space is not relevant. Although, when working with microcanonical windows one would need to restrict the auxilliary graded vectors introduced in the above equation, as well as the corresponding unitary rotations to only the relevant sector of the Hilbert space.
We now bring the formalism described in section \ref{sec.review} to bear on the genus-2 spectral form factor. This involves writing down the effective field theory of chaos based on the tripled Hilbert space we made copious use of in this work. The EFT of quantum chaos is a theory on the space of Hamiltonians, whose dynamics captures the approach of a generic chaotic Hamiltonian to the predictions of random-matrix theory in the ergodic limit (see e.g. \cite{Altland:2020ccq,Altland:2021rqn}).
In order to proceed, let us write the Hamiltonian, $H$, in a complete basis of hermitian matrices, $X_a$, in $D$-dimensions, which we then adapt to the tripled Hilbert space
\begin{align}
\mathbb X_a &\equiv X_a\otimes \mathds 1 \otimes \mathds 1 + \varkappa_1 \mathds 1 \otimes X_a \otimes \mathds 1 + \varkappa_2 \mathds 1 \otimes \mathds 1 \otimes X_a,
\end{align}
This basis is labelled by the parameter $a$ which takes values, $a = 0,1,2,\ldots , D^2-1$ and $X_0 = \mathds 1$. We can then expand the tripled Hamiltonian in this basis as\footnote{In appendix \ref{appA}, we discuss the spectrum of $H_{\textrm{full}}$ and its relationship with the Hamiltonian of the original theory, H.}
\begin{align}
H_{\rm full} &= a^a \mathbb X_a
\end{align}
The ergodic limit is characterised by the indistinguishability of the physical Hamiltonian from a typical representative of the corresponding RMT ensemble. How a given chaotic Hamiltonian approaches this limit is a highly interesting subject, \cite{Altland:2020ccq,Altland:2021rqn}, but for the present purposes we may simply assume that the theory has such an ergodic limit. The strength of our formalism is that it is well-suited to study the non-universal modes that lead to deviation from the ergodic limit at times earlier than the {\it Thouless/ergodic time}. These non-universal modes, referred to as massive modes, arise from the fact that the matrix elements of a physical Hamiltonian are highly correlated. Outside the ergodic limit, this is reflected in correlations between the $a^a$ parameters. However, in the quantum ergodic limit, random-matrix universality implies that the individual elements of the Hamiltonian, $H_{ij}$, are chosen from a Gaussian distribution. This translates to performing a Gaussian disorder average over the $a^a$ parameters with a probability distribution given by,
\begin{equation}
P(a^a) = \exp\left[-\frac1{2\bar a} \left(a^a\right)^2\right]~,
\end{equation}
for some fixed $\bar a$. Performing the integral over these parameters gives us the expression
\begin{equation}
Z[\hat h] = \!\! \int \!\!\!D\bar \Psi D\Psi \, \exp\left[i\bar \Psi \cdot \left( \hat z-\hat h \right)\cdot \Psi - \frac {\bar a}2 \sum_a\left(\bar \Psi \mathbb X_a \Psi \right)^2\right]~.
\end{equation}
whose last term is quartic in the graded fields $\Psi$. This quartic term can be treated by a Hubbard-Stratanovich transformation to rewrite the path integral in terms of the new fields, $A^a$,
\begin{align}
Z[\hat h] &= \int DA^a D\bar \Psi D\Psi \ \exp\left[i\bar \Psi \cdot \left( \hat z-\hat h \right)\cdot \Psi +i \sum_a \mathbb S\mathbb T{\rm r}\left[A^a \left(\mathbb X_a \Psi \bar \Psi \right)\right] \right. \nonumber \\
&\hspace{5cm} \left. - \frac 1{2\bar a} \sum_a \mathbb S\mathbb T{\rm r} \left[ (A^a)^2\right]\right]\,.
\end{align}
The resulting action is quadratic in the $\Psi$ fields and which can therefore be integrated exactly, so that finally we arrive at the expression
\begin{align}
Z[\hat h] &= \int DA^a \exp\left[ -\frac1{2 \bar a}\sum_a\mathbb S\mathbb T{\rm r}\left[(A^a)^2\right] - \mathbb S\mathbb T{\rm r}\ln\left[ \hat z - \hat h + A^a\mathbb X_a \right] \right]\,.
\end{align}
In the above and in some of the following expressions we are not taking care of the normalisation factors which can be easily fixed by requiring that the final expression for the path integral should be equal to 1 when all the sources are switched off. This is true for the simple reason that the determinants that appear in the numerator and the denominator of \eqref{eq.pathintwrite} are identical in the absence of sources.
The final expression above is amenable to saddle point analysis when the size of the Hilbert space is large. The saddle point equations are
\begin{equation}\label{eq.full.saddle}
A^a = - \bar a \left( \hat z + A^a\mathbb X_a \right)^{-1}\, \mathbb X_a~.
\end{equation}
Note that we have dropped the source term in the above saddle equation. Following a standard treatment, the sources are treated perturbatively and the operators are assumed to not backreact on the solutions of the saddle equations.\footnote{This is an assumption. We comment on this point further in the discussion section.} At this stage, we consider the homogeneous Hilbert space ansatz according to which the saddle point solution of the above path integral is homogeneous in the complete $D^3$ dimensional Hilbert space,
\begin{equation}\label{eq.HHA}
A^a = \hat y \,\delta^a_0 \, \mathds 1^{\otimes 3}~.
\end{equation}
The physical motivation and the validity of this ansatz is discussed in great detail in \cite{Altland:2020ccq}. When working with a microcanonical window instead of the entire Hilbert space, the homogeneous Hilbert space ansatz is restricted only to the relevant part of the Hilbert space in this equation. This implies that the identity matrix, $\mathds 1^{\otimes 3}$, is not an identity matrix in the entire Hilbert space but only a subsector of it. The factors of Hilbert space dimensions that appear in the following discussion will therefore get modified accordingly to reflect the dimensionality of the microcanonical Hilbert space (effectively, this means taking $D=D_1$ from now on).
The path integral simplifies on taking the ansatz, \eqref{eq.HHA}, also making the validity of the saddle point approximation more apparent in the large $D$ limit,
\begin{align}\label{eq.homogeneous-action}
Z[\hat h] & =\int \!\!D\hat y \exp\Bigg[ \!\!-\frac{D^3}{2 \bar a}\sum_a\mathbb S\mathbb T{\rm r}\left[(\hat y)^2\right] - D^3 \, \mathbb S\mathbb T{\rm r}\ln\left[ \hat z + \gamma \hat y \right] -\mathbb S\mathbb T{\rm r}\ln\left[\mathds 1 + \left(\hat z + \gamma \hat y\right)^{-1} \hat h\right]\!\!\Bigg]~,\\
&\hspace{5cm}\gamma := 1+\varkappa_1+\varkappa_2~.
\end{align}
Note that we have chosen to write the term containing the sources seprately. Because we evaluate the path integral at $\hat h=0$, this term is expanded in powers of $\hat h$ starting at linear order. In the $D\gg1$ limit, the solutions of the saddle point equation, \eqref{eq.full.saddle}, simplifies to take the form
\begin{align}
\hat y &= - \bar a \left( \hat z + \hat y \right)^{-1}~,\\
\Rightarrow \hat y &= -\frac E{2 \gamma} + i \Lambda \sqrt{\bar a - \frac{E^2}{4\gamma^2}},
\end{align}
where the matrix $\Lambda$ determines the different choices of saddle, as
\begin{align}
\Lambda_0 &= \sigma_3^{\rm RA} \otimes \mathds 1^{\rm bf} \label{eq.leading-saddle}\\
\Lambda_{\rm AA} &= \sigma_3^{\rm RA} \otimes \sigma_3^{\rm bf} \label{eq.subleading-saddle}
\end{align}
Following standard convention, we refer to these as the \emph{standard saddle} and the \emph{Andreev-Altschuler saddle}, respectively. The standard saddle point corresponds to the following density of states in the tripled Hilbert space,
\begin{equation}\label{eq.TripledDensityOfStates}
\rho(E) = \frac{D^3}{\pi \sqrt{\bar a} \gamma} \sqrt{1 - \frac {E^2}{4 \bar a \gamma^2}}, \quad \Delta(E) = \frac1{\rho(E)}~.
\end{equation}
We have also defined the mean level spacing, $\Delta(E)$ as the inverse of the density of states, $\rho(E)$. One way of seeing \rref{eq.TripledDensityOfStates} is to use \rref{eq.OperatorWeightedDensity} - \rref{eq.SourcedEFT} to compute
\begin{equation}\label{eq.1ptdense}
\rho(E) := \rho_{ \mathds 1}(E) = \frac{1}{\pi} {\rm Im}\,\partial_{h_1} Z[\hat h] \Bigr|_{\hat h =0}\,.
\end{equation}
Notice that as long as the energy arguments that appear in $\hat z$ are all the same, the path integral \eqref{eq.homogeneous-action} is symmetric under the $U(1,1|2)$ rotations in the graded spaces,
\begin{equation}
\Psi \to g \cdot \Psi , \quad g\in U(1,1|2)~.
\end{equation}
which translates to following transformation of the Hubbard-Stratonovic field, $\hat y$,
\begin{equation}
\hat y \to g \cdot \hat y \cdot g^\dagger~.
\end{equation}
Both the saddle points break this $U(1,1|2)$ symmetry spontaneously to $U(1|1)\times U(1|1)$. In fact, when $\omega=0$, the entire coset manifold, $U(1,1|2)/U(1|1)\times U(1|1)$, is a space of solutions of the saddle point solutions. We parametrise this manifold as a `rotation' of the standard saddle point,
\begin{align}\label{eq.pions2}
T \Lambda_0 T^{-1} \in \frac{U(1,1|2)}{U(1|1)\times U(1|1)}.
\end{align}
The Andreev-Altschuler saddle point lies on this manifold and corresponds to,
\begin{equation}
T = T_0 = \mathds 1^{RA} \otimes P_b + \sigma_1^{RA} \otimes P_f~.
\end{equation}
The fact that saddle point solutions break the symmetry spontaneously can be argued as follows.
The sign of the regulator, $\varepsilon$, that differentiates between the advanced and the retarded sector leads to this spontaneous symmetry breaking. In a fashion the reader might be familiar from the treatment of advanced and retarded correlation functions in elementary QFT, the regulator displaces the poles in the integration plane of the eigenvalues of $\hat y$ by a small amount along the imaginary direction. The integration contour runs along the real axis for the bosonic components of the graded matrix, $\hat y$. The presence of the infinitesimally displaced pole restricts the deformation of this integration contour to only one of the two choices for the saddle point solutions. This choice is different for different causalities, hence leading to the spontaneous symmetry breaking. The fermionic sector has no such restrictions because of the lack of poles in the integration plane. For a more detailed technical discussion, we refer the reader to \cite{Altland:2021rqn, Haake-book}.
The precise nature of the structure of symmetry breaking relies on the observable under study. For example, in the computation of the density of states in \eqref{eq.1ptdense}, $\rho(E)$ involves only one of the two causality sectors. Therefore the symmetry breaking is irrelevant in this case. Similarly, for the computation of higher point functions, we need to introduce additional copies of the Hilbert space corresponding to additional insertion of the determinant operators. In this case too, the symmetry breaking will be different. Such observables are not the subject of current work.
The $U(1,1|2)$ symmetry is explicitly broken when $\omega\neq0$, which corresponds to different values of $z_i$s. In this case, we do not have a full coset manifold worth of saddle points solutions. However, the standard saddle and the Andreev-Altschuler saddle continue to be solutions to the saddle point equations. The additional action-cost associated with the points of the coset manifold can be computed in the limit $s = \pi \omega/\Delta \ll 1$. We treat $s = \pi \omega/\Delta$ as a perturbative parameter and expand the action, \eqref{eq.homogeneous-action}, as a power series in this parameter. Under such a treatment, various points on the coset-manifold can therefore be regarded as \emph{pseudo-Goldstone bosons}.
It is useful to parametrise the points on the coset manifold in terms of \emph{pion fields}, $W$,
\begin{align}\label{eq.pions}
Q = T \Lambda_0 T^{-1} \in \frac{U(1,1|2)}{U(1|1)\times U(1|1)}, \quad T = e^{-W}~,\quad
W = -\begin{pmatrix}0&B\\ \tilde B&0\end{pmatrix}~.
\end{align}
On substituting the above parametrisation of the coset manifold into the action, \eqref{eq.homogeneous-action}, and performing the perturbative expansion in small $\omega$ one obtains the leading order term,
\begin{equation}\label{eq.quark-mass}
2i \pi \frac\omega{ \Delta(E) } \ \mathbb S\mathbb T{\rm r}\left[B \tilde B\right]~.
\end{equation}
The above term can be compared to the ``quark mass'' for the pions in the chiral perturbation theory. Following a top down approach, for complex Hamiltonians one can write down an effective action on the coset manifold in increasing orders of $\omega$ as well as $B,\tilde B$,
\begin{align}
\int \!\!dQ\, e^{-S[Q;\omega]}, \qquad Q \in \frac{U(1,1|2)}{U(1|1)\times U(1|1)}~,
\end{align}
in the same spirit as the chiral perturbation theory.
The integration on the coset manifold can be performed exactly \cite{Haake-book}. However, because the integration is one-loop exact, \cite{Haake-book, Efetov-book}, it is sufficient to perform the integration perturbatively in the pion fields around the individual saddle points. It is for this reason we have expanded the action in $B,\tilde B$ fields and kept only the leading quadratic term in above equation.
To compute the resolvent, \eqref{eq.resolvent.path}, we need to take the derivative of the path integral, \eqref{eq.homogeneous-action}, with respect to the sources, $h_\pm$. This amounts to expanding the last term in this action that depends on the source to quadratic order in $\hat h$. Taking derivatives of this term with respect to the sources introduces certain pre-exponential terms in the $\sigma$-model path integral,
\begin{equation}\label{eq.pre-exp}\begin{aligned}
\left( \frac1{\bar a\gamma^2} + \frac{4\pi^2 \rho^2(E)}{D^6} \mathbb S\mathbb T{\rm r}\left[ B \tilde B P_b\right] \mathbb S\mathbb T{\rm r}\left[ \tilde B B P_b\right]\right) \mathbb T{\rm r}\mathbb O \, \mathbb T{\rm r} \mathbb O^\dagger\\
- \frac{4\pi^2 \rho^2(E)}{D^6} \mathbb S\mathbb T{\rm r}\left[ B P_b \tilde B P_b \right] \mathbb T{\rm r}[\mathbb O \mathbb O^\dagger] + \cdots~.
\end{aligned}\end{equation}
Here, we have explicitly written only the terms that involve non-trivial contraction of $B, \tilde B$ fields. There are other terms that evaluate to zero and aren't shown above. From \eqref{eq.quark-mass} one can deduce the following contraction rules between the pion fields:
\begin{align}
\label{eq.BBtprop1}
\mathbb S\mathbb T{\rm r}[{B}\cdot X] \, \mathbb S\mathbb T{\rm r}[ { \tilde B} \cdot Y] &= \frac1{2\pi \rho(E)} \frac i{\omega+i0^+} \mathbb S\mathbb T{\rm r}[X Y]~,\\
\label{eq.BBtprop2}
\mathbb S\mathbb T{\rm r}[\source{B} \cdot X \cdot \target{\tilde B}\cdot Y] &= \frac1{2\pi \rho(E)} \frac i{\omega+i0^+} \mathbb S\mathbb T{\rm r}[X] \mathbb S\mathbb T{\rm r}[ Y]~.
\end{align}
Using these contraction rules, one can solve \eqref{eq.pre-exp} to get the following answer for the resolvent,
\begin{equation}\label{eq.std.saddle.ans}
R(\omega) = \left(\frac1{\bar a \gamma^2} - \frac1{D^6 \omega^2}\right) {\rm Tr} \mathbb O \, {\rm Tr} \mathbb O^\dagger~.
\end{equation}
Note that contractions between the $B, \tilde B$ that have a common index is zero. For example, one does not receive a contribution from contracting the fields within the same $\mathbb S\mathbb T{\rm r}$ terms. This is because it corresponds to $X=\mathds 1$ in \eqref{eq.BBtprop2} and $\mathbb S\mathbb T{\rm r} \mathds 1=0$.
As we remarked above, for small $s$ when the symmetry breaking is small, one needs to perform the integration over the entire coset manifold. However, one-loop exactness of the path integral implies that the exact answer is the sum over one-loop contributions around the individual saddle points. The result in \eqref{eq.std.saddle.ans} is the contribution around the standard saddle point. Therefore, we need to add to the above answer the contribution of the Andreev-Altschuler saddle point as well. Additionally, for the complete answer we also need to add the contribution of the $R^{\ddag}$, \eqref{eq.path.resolvents}. The computation of these terms closely follows equivalent computations presented in \cite{Altland:2021rqn}, and therefore we do not discuss these details here. Adding in all these contributions, one gets the full answer for the resolvent in the ergodic limit,
\begin{equation}
R(s)= \frac{2\pi^2\rho^2(E)}{D^6} \left(\pi \delta(s) + 1 - \frac{\sin^2(s)}{s^2}\right) {\rm Tr} \mathbb O \, {\rm Tr} \mathbb O^\dagger ~
\end{equation}
In this expression, note that the leading disconnected contribution corresponding to `1' is enhanced by a factor of $D^3$ with respect to the asymptotic contribution of the connected piece given by the $\delta(s)$. This happens because $s = \pi \rho(E) \omega \sim D^3 \omega$.
Let us consider only the connected contribution,
\begin{equation}\label{eq.connected-only}
\frac{2\pi^2\rho^2(E)}{D^6} \left(\pi \delta(s) - \frac{\sin^2(s)}{s^2}\right) {\rm Tr} \mathbb O \, {\rm Tr} \mathbb O^\dagger ~.
\end{equation}
This term corresponds to the ramp-plateau behaviour depicted in \autoref{fig.g2t} after performing the Fourier transform to the time domain,
\begin{equation}
\int({\rm connected\, term}) e^{i\omega t} = \left( 2 + t \, \frac{\Delta(E)}{\pi} + \left(2- t\,\frac{\Delta(E)}{\pi} \right)\, \text{sgn}\left(t\, \frac{\Delta(E)}{\pi} -2\right) \right)~.
\end{equation}
This term is zero at $t=0$. The disconnected term gives,
\begin{equation}
\int({\rm disconnected\, term}) e^{i\omega t} = \frac{2\pi^2\rho^2(E)}{D^6} {\rm Tr} \mathbb O \, {\rm Tr} \mathbb O^\dagger \, \delta(t)~.
\end{equation}
The above integration gives rise to a Dirac-delta function only when the integration over $\omega$ is unbounded. In physical examples, the range of $\omega$ integration is fixed by the width of the spectrum ($\sim \sqrt{\bar a} \gamma$) and therefore the width of the {\it regulated}-delta function is controlled by $1/{(\sqrt{\bar a} \gamma)}$. Importantly, the value of the genus-2 partition spectral form factor is non-zero at $t=0$ (it is simply given by the square of the genus-2 partition function).
From a gravitational point of view, there are two types of geometry relevant for the genus-2 spectral form factor. Handlebody geometries \cite{Krasnov:2000zq}, which correspond to a bulk filling of a genus-2 surface. For the square of the genus-2 partition function, one can thus have two disconnected handlebodies. The second type of geometry is the genus-2 wormhole, see \cite{Maldacena:2004rf,Belin:2020hea}, which connects two genus-2 boundaries. One interesting aspect of the genus-2 wormhole is that it is a true saddle-point of the gravitational equations of motion, even at $t=0$. This should be contrasted with the double-trumpet or the double-cone \cite{Saad:2018bqo,Saad:2019lba,Cotler:2020ugk,Cotler:2020lxj} relevant for the usual spectral form factor. As we increase time, the genus-2 wormhole eventually dominates over the two disconnected handlebodies. It is not known when this happens exactly, but the fact that the genus-2 wormhole is a true saddle may play a role. In the current treatment, we have considered the homogeneous Hilbert space ansatz, which studies the theory only in the ergodic regime. At earlier times, it becomes crucial that the Hamiltonian is fixed. This can be captured by studying the contributions of the massive modes like in \cite{Altland:2021rqn}. It would be interesting to understand whether the fact that the genus-2 wormhole is a true saddle even at early times is important. We will address this in future work.
\section{Discussion}
\label{sec.discussion}
In this paper, we have analyzed the statistics of heavy-heavy-heavy OPE coefficients from the point of view of ergodicity and random matrix theory. To transform the problem into that of ordinary quantum mechanics, we defined a linearized operator on the tripled Hilbert space of the CFT whose matrix elements are given by the product of two OPE coefficients. We then proceeded to analyze the properties of this operator from a random matrix theory point of view. We argue that the behaviour of this operator in the ergodic limit can be explained by the statistics of the OPE coefficients, a conjecture originally proposed in \cite{Belin:2020hea}. Using the effective theory of quantum chaos, we also studied the genus-2 spectral form factor and found that it should have a ramp and plateau.
\subsection{Open questions}
We conclude the paper with some open questions.
\subsubsection*{From CFT to QM}
Part of the goal of this paper was to translate a CFT question which does not manifestly have a nice counterpart in quantum mechanics, into a framework that can be tackled strictly within quantum mechanics using random matrix theory. To do so, we defined a linear operator in a tripled Hilbert space. The next step was to assume that this operator is a random operator on the tripled Hilbert space. From the effective theory of quantum chaos, this amounts to assuming that the insertion of the operator source does not cause a large backreaction on the saddle-point. At this stage, this is an assumption, and it would be very interesting to understand whether this assumption is true. Also note that the framework we have presented would work for other tensors made out of more OPE coefficients, and it would be interesting to see whether the validity of the saddle-point assumption depends on the particular choice of tensor we made (i.e. on the particular combination of OPE coefficients).
This ties to a more profound understanding of what a heavy CFT operator is. An alternative route to that studied in this paper would be to study the expectation value of complicated (i.e. extensive) operators in high energy states \a`{a} la ETH. This could presumably be done numerically, although might quickly become difficult. The major problem that remains, even with the numerical methods, is to understand what the right basis of operators is. In CFTs, the state-operator correspondence tells us that local operators and energy eigenstates are one and the same. We don't see such an obvious choice in quantum mechanics. In a chaotic spin-chain, what is the ``right" extensive operator to pick? Should one consider randomized extensive operators?
Perhaps the most interesting theory to study in this regard is the SYK model. The SYK model can offer a bridge between a notion of complicated operator in QM and the CFT language, since the theory is conformal in the IR. A complicated operator would be one that is built from an $\mathcal{O}(1)$ fraction of the $N$ fermions, and the question becomes understanding what that maps to in the IR, in terms of the spectrum of local operators. This could be tackled using a combination of numerical and analytics techniques following up on \cite{Sonner:2017hxc,Nayak:2019khe,Nayak:2019evx}.
Another point to mention is that CFTs always have more symmetries than generic quantum systems. One should thus organize the operators according to representations of the conformal group, and statements like the ETH should be always understood as applied to primary operators. This is particularly relevant in 2d CFTs, because of Virasoro symmetry. We do not expect the results to exhibit qualitative differences once this has been taken into account, but it be would be interesting to understand how this affects the EFT of quantum chaos.
\subsubsection*{Minimal models and the genus-2 spectral form factor}
Another interesting avenue to explore would be to study the genus-2 spectral form factor in rational CFTs that can be solved. The most natural example is the minimal models. For the usual spectral form factor, this was studied in \cite{Benjamin:2018kre}. Note that the minimal models are completely solved, so there is in no conceptual obstruction in computing the genus-2 spectral form factor. There are however some practical problems, coming from the fact that the genus-2 conformal blocks are not known in closed form, contrary to Virasoro characters. For plotting the genus-2 spectral form factor, this should not be too much of a problem since one can get the blocks numerically.
It would be interesting to check whether the minimal models at large level do display a ramp and plateau for the genus-2 spectral form factor. Even more interestingly, one could check if the dip time of the usual spectral form factor and that of its genus-2 counterpart are the same or not. In this paper, since we used only ergodicity arguments, we don't have an understanding of the Thouless time for the genus-2 spectral form factor. The Thouless time is both theory \textit{and} observable dependent, and therefore it is natural to think that the Thouless time could be quite different for the genus-2 spectral form factor.
\subsubsection*{The Symmetry Class of a CFT}
In section \ref{sec.ope}, we computed the mean and variance of the matrix elements of $\mathbb{O}$. We saw that the variance of the matrix elements involved a ${\rm {Tr}} \mathbb{O}^2$ term. The trace here is given by a particular cyclic contraction of the indices, and matches the decomposition of a genus-3 partition function in the sunset channel \cite{Belin:2021ryy}. However, we have obtained a single type of trace for $\mathbb{O}^2$ and there are in fact 5 possible trace structures involving four OPE coefficients (corresponding to the 5 OPE decompositions of a genus-3 surface), two examples of which are
\begin{eqnarray}
{\rm {Tr}}_{\textrm{skyline}}c^4&=& c_{abc} c_{ade}^* c_{fbe} c_{fdc}^* \\
{\rm {Tr}}_{\textrm{comb}}c^4&=&c_{aad}c_{bbe}^*c_{ccf} c_{def}^* \,.
\end{eqnarray}
To correctly account for the statistics of the heavy-heavy-heavy OPE coefficients, we need all five of these structures \cite{Belin:2021ryy}. It is easy to see that the reason that we did not find all five structures lies in the way we applied unitary Haar averages to the operator $\mathbb{O}$.
One could imagine a more refined analysis in which we are more careful about the discrete symmetries of a CFT and the operator
under consideration. Generic CFT's have real and complex fields, and under complex conjugation of operators the OPE coefficients are also complex conjugated. This particular symmetry is not implemented in the unitary averaging we considered so
far, and taking it into account will presumably lead to a different type of averaging, a different type of RMT and and a different sigma model corresponding to a different symmetry class. It is our expectation that such a more precise treatment will lead to some quantitative changes in our result, in particular that a richer structure of random matrices and index structures will appear, but that the qualitative conclusions remain unaltered. Moreover, we expect that the leading order scaling with $D$ remains the same, and that we still
have a version of a ramp and a plateau in the relevant generalized form factors. We leave the interesting question which symmetry classes are realized in CFTs, with and without
operator statistics, and how these impact randomness and chaos, to a future study.
\subsubsection*{The OPE coefficients $c_{LLH}$}
There are essentially three types of OPE coefficients involving heavy operators, depending on whether one, two or three of the operators are heavy. As explained in the introduction, the operators $c_{LHH}$ are easiest to understand since the ETH ansatz makes a prediction for them. In this paper, we have developed a tool to study the OPE coefficients $c_{HHH}$. The final OPE coefficients to understand are thus $c_{LLH}$. Asymptotic formulas for such coefficients are also known \cite{Pappadopulo:2012jk,Das:2017cnv,Mukhametzhanov:2018zja,Collier:2019weq,6ptasymptotics} and it is natural to ask what random matrix theory has to say about them.
While our framework could in principle be used to study such operators as well, it seems important to underline a physical difference between these objects and the ones studied in this paper: there is no spectral form factor that one can define only with $c_{LLH}$ coefficients. In fact, the most Lorentzian probe of these OPE coefficients that one can cook up is simply the out-of-time-ordered four-point function on the plane. This is also a probe of quantum chaos, but of a quite different nature.\footnote{In fact, it really probes scrambling more than chaos. Even if the two are often related, they are not exactly the same thing. For example, the rate of scrambling is theory dependent while the eigenvalue statistics is much more universal. This suggests that how random the $c_{LLH}$ truly are could also be theory dependent. A similar observation was made in \cite{Caron-Huot:2020ouj}.} Since this object maps to a thermal correlator, but on Rindler space (a non-compact space), there is no discrete eigenspectrum to probe and hence no ramp and plateau. It would be interesting to understand better how random matrix theory constrains these OPE coefficients.
\section*{Acknowledgements}
We are happy to thank Alexander Altland, Daniel Jafferis, Diego Liska for fruitful discussions. The presentation of certain results contained in this paper was improved following a comment posted by Brehm, Das and Datta. JdB is supported by the European Research Council
under the European Unions Seventh Framework Programme (FP7/2007-2013), ERC Grant
agreement ADG 834878.
JS thanks ENS Paris for hospitality during the final stages of this work.
This work has been partially supported by the SNF through Project Grants 200020 182513, as well as the NCCR 51NF40-141869 The Mathematics of Physics (SwissMAP).
|
\section{Introduction}
Disorder averaging is a well-defined procedure in any parametric family of quantum field theories (QFTs). Operationally, one
fixes the parameters of the theory, computes correlation functions for this choice, and then at the end, performs a
further average over a \emph{classical} probability distribution for these parameters. This is clearly a useful tool
for gaining access to ``typical'' behavior in various systems with a high degree of complexity (see, e.g., \cite{Anderson:1958vr,Sachdev:1992fk,Parisi:1992tk}).
It has also appeared in the context of holography (see, e.g., \cite{Hartnoll:2014cua, Aharony:2015aea, kitaev, Maldacena:2016hyu,Afkhami-Jeddi:2020ezh,Maloney:2020nni,Benjamin:2021wzr}) as well as other areas within high energy theory \cite{Rothstein:2012hk,Green:2014xqa,Craig:2017ppp, Balasubramanian:2020lux}.
Indeed, recent bottom-up considerations suggest that combining the principles of holography with effective field theory
in the context of the Euclidean gravitational path integral naturally results in the appearance of disorder averaging in the conformal field theory (CFT) on the boundary (see, e.g., \cite{Maldacena:2016hyu,Maldacena:2017axo,Penington:2019kki,Saad:2019lba,Stanford:2019vob,Marolf:2020xie} as well as \cite{Giddings:1989bq,Giddings:2020yes,Cottrell:2018ash,Sonner:2017hxc}).
One particularly surprising aspect of these considerations is that, at present, they have resisted an embedding in
string theory.\footnote{Other examples include the case of ``double holography'' \cite{Almheiri:2019hni}, which involves coupling a large $N$ gauge theory to gravity. For massless gravity, this is in sharp contradiction with all known string constructions and Swampland considerations \cite{Ooguri:2006in,Heckman:2019bzm}, but for a massive graviton, it might be possible \cite{Geng:2020qvw,Uhlemann:2021nhu,Geng:2021hlu,Raju:2021lwh}.} It is not hard to see that finding such a completion might be difficult, because the presence of an explicit classical distribution would seem to require treating gravity as an open system.
The hope, then, is that constraints imposed from requiring a UV completion can be sufficiently decoupled from long distance effects in the putative gravity dual. That being said, it is not entirely obvious that this is really the case. For example, in \cite{McNamara:2020uza} it was argued that the Swampland cobordism conjecture (see \cite{McNamara:2019rup}) precludes the existence of an AdS/CFT correspondence with ensemble averaging for $D > 2$ boundary systems, and in the case of $D \le 2$, the corresponding systems should be viewed as UV completed in a higher-dimensional system. A closely related point is that for many of these considerations, it is actually quite important that the couplings have no position dependence at all. This is required to have a consistent interpretation in terms of the creation of baby universes.\footnote{This can be seen as a consequence of Gauss's law for translation invariance; an emitted baby universe cannot carry energy or momentum because it has no boundary. We thank H.~Maxfield for helpful correspondence on this point.}
Our aim in this paper will be to engineer ensemble averaged QFTs embedded in string theory.
Let us state at the outset that aesthetically, the construction we present is a rather contrived UV completion.
That being said, it provides us with a general framework for testing the claim that disorder averaging in the context of
holography can truly be decoupled from stringy considerations.
The main idea behind our construction is to produce QFTs decoupled from gravity via open strings
localized on a small patch of the compactification geometry. To get a statistical ensemble, we simply consider multiple stacks of branes
at different locations in the transverse extra dimensions. In particular, by varying the profile of non-normalizable modes
in these directions, we can realize different low energy effective field theories with \emph{identical} field content but with
different values of the physical parameters. Given a set of $\ensuremath K$ stacks with couplings $\lambda_\ensuremath k$ for $\ensuremath k = 1, \dotsc, \ensuremath K$, we can consider a special class of operators $O_\ensuremath k$ for $\ensuremath k = 1, \dotsc, \ensuremath K$ given by ``tracing'' over all the different stacks:
\begin{equation}
\ensuremath\mathbb{O} \equiv O_1 + \dotsb + O_\ensuremath K\,.
\end{equation}
The key point is that the correlation functions of these $\ensuremath\mathbb{O}$s factorize to leading order:
\begin{equation}\label{eq:Ofactor}
\langle \ensuremath\mathbb{O}(x) \ensuremath\mathbb{O}'(y) \rangle \approx \sum_{1 \le \ensuremath k \le \ensuremath K} \langle O_\ensuremath k(x) O'_\ensuremath k(y) \rangle\,,
\end{equation}
which in turn leads to an averaging over couplings. At short distances this approximation breaks down because we become sensitive to massive excitations that have been integrated out to reach the effective field theory in the first place.
Depending on the number density of stacks with a given value of $\lambda_\ensuremath k$, it is clear that this is building up a ``binned'' version of an ensemble average. Provided we can engineer a suitable internal profile for the couplings and populate the stacks at distinct values of the couplings, it would appear that we can produce a large class of probability distributions $p(\lambda)$ for ensemble averaging. Note also that taking $\ensuremath K \gg 1$ provides a general way to get a good approximation by the binned distribution of its idealized smooth counterpart.
A particularly important special case is provided by requiring each
stack to realize a conventional holographic dual with the same value
of the bulk cosmological constant. In this case, we observe that the correlators
for the operators $\ensuremath\mathbb{O}(x)$ reconstruct a \emph{single}
anti-de Sitter space (AdS) throat region. In this context, the appearance of a disorder average means that asymptotically,
we do not restrict ourselves to a fixed number of boundary components, but allow this to fluctuate, much as in \cite{Marolf:2020xie}.
Provided we only consider a number of boundary components much smaller than $\ensuremath K$, we thus make contact with the proposal of Marolf and Maxfield \cite{Marolf:2020xie}, and in the limit where $\ensuremath K$ is very large, this provides an adequate way to build up an ensemble average and its holographic dual.
However, with an actual construction in hand, we can also identify two general ways in which our UV completion breaks down.
First of all, we clearly have a large number of sequestered stacks of branes,
so if we proceed to higher energies, we should observe additional contributions beyond the
approximate factorization appearing in \cref{eq:Ofactor}. This is not altogether surprising, but already points to the fact
that the UV completion does place an intrinsic limitation on the sorts of correlation functions we can consider.
Perhaps more surprisingly, there is \emph{another} way in which the approximation can break down, and it is something that
occurs even if we restrict to observables deep in the infrared. This is due to the fact that our UV completion implicitly makes reference to a fixed $\ensuremath K$, and we can actually distinguish between our binned approximation and a smooth distribution after sampling $n_*$ times (see \cref{eq:nstar} later on for the precise definition). In particular, if we consider any bulk object characterized
by a density matrix with order $n_*$ or more entries (as, for example, we would need to
discuss in constructing the Page curve of a macroscopic black hole \cite{Page:1993wv}, see, e.g., \cite{Page:2013dx}),
then our putative ensemble average has been pushed beyond its regime of validity.
To make these considerations precise, we present a number of examples illustrating
how to generate ensemble averaging for appropriate string-based constructions. Curiously enough, the case where we can maintain
the most control is for $D = 4$ superconformal field theories (SCFTs) with an ensemble average over marginal coupling constants.
We illustrate this both in terms of compactifications of 6D little string theories (LSTs), brane box configurations, and
D3-brane probes of orbifold singularities. In this case, the extra dimensions transverse to the brane stacks
provide enough flexibility to produce a nearly arbitrary probability distribution with support on
the moduli space of marginal couplings. As another class of examples, we consider various $D = 2$ SCFTs obtained in a similar fashion.
In this case, we find that for examples where we can reliably extract an $\AdS_3$ dual description, it is often simplest to consider the fibration of a Calabi--Yau $n$-fold over a subvariety of its moduli space, resulting in a higher-dimensional (non-compact) Calabi--Yau geometry (see \cref{sec:FIBRATION}). The limitation of this sort of construction is then that our ensemble averages are necessarily restricted to a particular subset of moduli. Similar considerations hold for $D = 1$ ``SCFTs'' of the sort that appear in the construction of 4D black holes obtained from type II strings on Calabi--Yau threefold backgrounds. Here, we again get an ensemble average, as associated with $\AdS_2$ vacua. It is worth pointing out that at no point do we truly get a 2D gravitational theory; rather we get an $\AdS_2 \times S^2 \times X_6$ background.
Our method meets with less success in the case of $\AdS_4$, $\AdS_6$, and $\AdS_7$ vacua, as associated with 3D, 5D, and 6D SCFTs. In the case of 3D SCFTs, this may just be a failure of imagination / stamina. In the case of $D > 4$ SCFTs, we face the fact that there are no marginal deformations available that preserve supersymmetry \cite{Louis:2015mka, Cordova:2016xhm}.
The rest of this paper is organized as follows. In \cref{sec:GENERAL}, we present some more details on the general idea of our construction, and its regime of validity. We present a holographic interpretation of this construction in \cref{sec:HOLO}. \Cref{sec:EXAMPLES} presents explicit string-based examples. We present a brief discussion in \cref{sec:DISC}.
Some additional technical details are deferred to \cref{sec:FIBRATION}.
\section{Engineering an Ensemble}\label{sec:GENERAL}
In this \namecref{sec:GENERAL}, we present the main idea of generating a QFT ensemble average in string theory constructions,\footnote{For the reader interested in learning more about string theory, the authors recommend~\cite{Zwiebach:2004tj, Ibanez:2012zz}. It is a fascinating subject.} giving explicit examples later in \cref{sec:EXAMPLES}. We are specifically interested in the case where the coupling constants of the QFT are truly constant in the sense that they have no spacetime dependence. This is the case that has been of primary interest in recent holographic investigations, and also turns out to be the most challenging one to engineer in the context of string constructions. See reference \cite{Balasubramanian:2020lux} for how to get an ensemble average in a gravitational system but with spacetime-dependent couplings.
Our goal will be to mimic the effects of disorder averaging in a QFT via a string construction. More precisely, we suppose that our QFT depends on a set of couplings $\{\lambda\} \equiv \lambda$, and that we have a smooth classical distribution $p_\text{smooth}(\lambda)$. Given operators $O^{(1)}, \dotsc, O^{(m)}$ of the QFT, the disorder averaged correlator is obtained by evaluating the correlation function with respect to a fixed value of $\lambda$, and then performing a further averaging with respect to $p_\text{smooth}(\lambda)$:
\begin{equation}
\overline{\langle O^{(1)} \dotsm O^{(m)} \rangle} \equiv \int \@ifnextchar^{\DIFF}{\DIFF^{}}{\lambda} \; p_\text{smooth}(\lambda) \langle O^{(1)} \dotsm O^{(m)} \rangle\,.
\end{equation}
Our plan will be to engineer an ensemble of QFTs that are decoupled from gravity. Loosely speaking, this is arranged by taking a limit in which we try to retain some localized degrees of freedom while decoupling the gravitational degrees of freedom.\footnote{For example, open strings attached to a stack of D-branes.} This in turn means that some degrees of freedom are non-normalizable, and these descend to coupling constants $\{\lambda\}$ of the QFT sector. An important comment here is that even though these degrees of freedom are constants in $D < 10$ spacetime dimensions, they are still dynamical in the full $D = 10$ (resp., $D = 11$) spacetime associated with string theory (resp., M-theory).\footnote{Indeed, in reference \cite{Balasubramanian:2020lux} it was noted that in a string compactification with multiple QFT sectors coupled to gravity, performing a partial trace over all but one such QFT sector results in a system characterized by a mixed state for position-dependent coupling constants. A subtlety with this approach is that the process of decoupling gravity also tends to force the previously obtained mixed state back into a pure state, and we will ultimately need to take a different tack to generate a position-independent ensemble average.}
\begin{figure}[t!]
\centering
\includegraphics[width=15.5cm]{brane_sample.pdf}
\caption{Copies of the same QFT sector separated from one another in the extra dimensions. QFTs are engineered on the worldvolume of branes (colored grey). The extra-dimensional geometry is colored blue.}
\label{fig:QFT sample in string theory}
\end{figure}
Now, precisely because we are dealing with a non-compact extra-dimensional geometry, we are free to consider multiple copies of the same QFT sector, just separated from one another in the extra dimensions. This stringy setup is roughly illustrated in \cref{fig:QFT sample in string theory}.
The parameters on each local sector are then specified by the internal profile of the non-normalizable modes in the vicinity of each local model. Introducing $\ensuremath K$ such local sectors, we can label the corresponding couplings as $\lambda_{\ensuremath k}$ for $\ensuremath k = 1, \dotsc, \ensuremath K$, where $\lambda_\ensuremath k \equiv \{\lambda_\ensuremath k\}$ denotes the full set of couplings in each local model. Since the field content is assumed to be identical, we also have the same set of operators $\{O_\ensuremath k\}$ for each local model. The different sectors are decoupled only in an approximate sense because we can also consider degrees of freedom that stretch from one sector to the other (e.g., open strings in D-brane models), and this introduces a UV cutoff $\Lambda_\text{UV}$ for this approximate factorization of the Hilbert space of states:
\begin{equation}
\ensuremath\mathbb{H} \approx \ensuremath\mathcal{H}_1 \otimes \dotsb \otimes \ensuremath\mathcal{H}_\ensuremath K\,.
\end{equation}
In the effective action for the full system, the action breaks up into a set of distinct contributions that then mix via higher-dimension operators with suppression scale $\Lambda_\text{UV}$:
\begin{equation}
\ensuremath\mathbb{S} = S_1 + \dotsb + S_\ensuremath K + S_\text{mix}\,.
\end{equation}
Given a specific operator of a single sector, there is a natural subset of operators obtained
by forming a sum over all such sectors:
\begin{equation}\label{eq:otown}
\ensuremath\mathbb{O} \equiv O_1 + \dotsb + O_\ensuremath K\,.
\end{equation}
Normalized correlation functions for $n$ such operators are then evaluated as:
\begin{equation}
\langle \ensuremath\mathbb{O}^{(1)} \dotsm \ensuremath\mathbb{O}^{(n)} \rangle_\text{normalized} \equiv \frac{1}{\ensuremath K} \langle \ensuremath\mathbb{O}^{(1)} \dotsm \ensuremath\mathbb{O}^{(n)} \rangle_{\ensuremath\mathbb{H}}\,,
\end{equation}
where on the righthand side the correlation function is evaluated with respect to the ground state of the full system. The pre-factor of $1 / \ensuremath K$ can be understood as the statement that we just require a well-behaved large $\ensuremath K$ limit. Another way to understand the same requirement is that we are just measuring all correlation functions in units of the traced identity operator:
\begin{equation}
\ensuremath\mathbb{I} = \bm{\mathrm{id}}_1 + \dotsb + \bm{\mathrm{id}}_\ensuremath K\,.
\end{equation}
Correlation functions for the $\ensuremath\mathbb{O}$ operators close to leading order. To see why, we note that:
\begin{equation}
\langle \ensuremath\mathbb{O}^{(1)} \dotsm \ensuremath\mathbb{O}^{(n)} \rangle_\text{normalized} \approx \frac{1}{\ensuremath K} \sum_{1 \le \ensuremath k \le \ensuremath K} \langle O_\ensuremath k^{(1)} \dotsm O_\ensuremath k^{(n)} \rangle_{\ensuremath\mathcal{H}_\ensuremath k}\,.
\end{equation}
where to leading order the different factors $\ensuremath\mathcal{H}_i$ are decoupled from one another. There can also be cross-terms between
the stacks, but these are all subleading contributions. Consider first connected correlators. In this case, sequestering the
stacks suppresses such contributions. For disconnected correlators (for example $\langle O_k O_k \rangle \langle O_l O_l \rangle$ with $k \ne l$), we further note that such contributions are kinematically suppressed relative to their connected counterparts
(as is clear by passing to momentum space).\footnote{We thank H.~Maxfield for correspondence on this point.}
In this sense it is a consistent truncation.
Our main claim is that this can be used to build up a discretized approximation to a disorder averaging by a continuous distribution $p_\text{smooth}(\lambda)$. To see why, we observe that each expectation value over $\ensuremath\mathcal{H}_\ensuremath k$ makes reference to the couplings
$\{\lambda_\ensuremath k\}$ of that sector. If we have a total of $\ensuremath K(\lambda)$ sectors with a particular set of couplings, then the probability associated with this choice is:
\begin{equation}\label{eq:discDist}
p_\text{disc}(\lambda) = \frac{\ensuremath K(\lambda)}{\ensuremath K}\,.
\end{equation}
As it stands, this is a discrete probability distribution, but it is important to note that in any actual string construction, there is an intrinsic ``spread'', as associated with the overall tension of a brane/region of localization in the internal geometry.\footnote{For example, there is a minimal length scale $\ell_\text{min} \sim \left(\frac{1}{T_p}\right)^{\frac{1}{p + 1}}$ that a D$p$-brane of tension $T_p \sim (g_s\ell_\text{st}^{p + 1})^{-1}$ can probe (see, e.g., \cite{Shenker:1995xq}).} For this reason, it is more appropriate to view our construction as building up a continuous distribution, but one that has been suitably ``binned''. More precisely, introducing an indicator function $I_{\lambda', \varepsilon_{\lambda'}}(\lambda)$, which has unit area and has support on a small region of size $\varepsilon_{\lambda'}$ centered around $\lambda'$, the probability of drawing $\lambda$ builds up a histogram comprised of small bins of size $\varepsilon_{\lambda'}$:
\begin{equation}\label{eq:binnedDist}
p_\text{bin}(\lambda) = \sum_{\lambda'} I_{\lambda' , \varepsilon_{\lambda'}}({\lambda}) \frac{\ensuremath K(\lambda')}{\ensuremath K}\,.
\end{equation}
The values of the $\varepsilon_{\lambda'}$ depend on the UV cutoff $\Lambda_\text{UV}$, as well as specific details of the model and construction.
The approximation just developed enables us to closely mimic the disorder average we would get for a smooth probability distribution, but there are clearly some limitations. One of these is already apparent from the general setup: in the correlation functions for the $\ensuremath\mathbb{O}$ operators, we observe that factorization will begin to break down when any correlators probe a short distance scale of size $\Lambda_\text{UV}$. Indeed, this is just the statement that we have a UV cutoff. Provided we work at long distances close to an infrared fixed point, we can hope to neglect such contributions.
There is also an entropic breakdown as associated with sampling the distribution $p_\text{bin}(\lambda)$ a large number of times. To see the issue, we consider two sorts of observers, a ``daemonic observer'' who has access to the full set of operators $O_i$, and another ``ignorant observer'' who is confined to making do with just the $\ensuremath\mathbb{O}$ operators.\footnote{This is related to the broader question of how well a low energy observer can ever hope to reconstruct a given UV completion, see, e.g., \cite{Heckman:2013kza,
Balasubramanian:2014bfa, Heckman:2016wte, Heckman:2016jud, Fowler:2020rkl, Balasubramanian:2020lux}
as well as \cite{Balasubramanian:1996bn, mehta2014exact, Erdmenger:2020vmo, Halverson:2020trp,
Stout:2021ubb, Erdmenger:2021sot, Erbin:2021kqf, Fowler:2021oje}.} For the daemonic observer, they can, after performing $n$ measurements associated with correlations between the $O_i$ and the $\ensuremath\mathbb{O}$ operators, extract $n$ independent and identically distributed (iid) draws from $p_\text{bin}$. The discrepancy between $p_\text{smooth}(\lambda)$ and $p_\text{bin}(\lambda)$ is captured by the relative entropy/Kullback--Leibler divergence, which reflects the amount of information we would gain upon switching from the binned distribution $p_\text{bin}(\lambda)$ to the ``true'' (although, from our perspective, unobtainable) distribution $p_\text{smooth}$:
\begin{equation}\label{eq:nstar}
D_\text{KL}(p_\text{smooth} \mid\mid p_\text{bin}) = \int \@ifnextchar^{\DIFF}{\DIFF^{}}{\lambda} \; p_\text{smooth}(\lambda) \log\frac{p_\text{smooth}(\lambda)}{p_\text{bin}(\lambda)} \approx c \varepsilon^2 \equiv \frac{1}{n_*}\,,
\end{equation}
where $c$ an order one constant dependent on the details of the distribution, and $\varepsilon$ is a representative value of the
size of our histogram bins. We can view $\varepsilon^{-1}$ as telling us the total number of distinct histogram bins. An observer who samples the distribution order $n \sim n_*$ times will be able to detect the difference between the string construction and the smooth ``idealization''. Let us note that typically $\varepsilon^{-1} \ll \ensuremath K$, since we necessarily need to group multiple values of the couplings in a single histogram bin to get an adequate approximation of $p_\text{smooth}(\lambda)$ in the first place.
Similar considerations hold for the ``ignorant observer,'' but their strategy for inferring the existence of a distribution over couplings is somewhat different. In this case, the point is to sample a number of different correlation functions over just the $\ensuremath\mathbb{O}$s, and in so doing infer the higher moments of the probability distribution $p_{\text{bin}}(\lambda)$. Here, distinguishability is governed by the number of moments of the distribution they are able to extract.
While the details of a particular model will dictate the specific operators to consider,
we can always consider the partition function in Euclidean signature, as obtained by placing
our QFT on some $D$-dimensional background. For a fixed value of the couplings $\lambda$, we have,
in each of our local sector path integrals:
\begin{equation}
Z_\lambda = \int \@ifnextchar^{\BIGDIFF}{\BIGDIFF^{}}{\phi} \; e^{-S_\lambda[\phi]}\,,
\end{equation}
in the standard notation. Observe that in our stringy construction,
the quantity $\overline{Z}$ is obtained from just performing a sum over each individual local sector:
\begin{equation}\label{eq:Zover}
\overline{Z} = \frac{1}{\ensuremath K} \sum_\ensuremath K \int \@ifnextchar^{\BIGDIFF}{\BIGDIFF^{}}{\phi_\ensuremath k} \; e^{-S_{\lambda_\ensuremath k}[\phi_\ensuremath k]}\,.
\end{equation}
Once we engineer $p_\text{bin}(\lambda)$, we also implicitly get a probability distribution over just $Z$. To see why, we can similarly consider the quantities $Z_{\ensuremath k}, Z_{\ensuremath k}^2, \dotsc$, and compute the corresponding moments for the partition function.
In terms of our $\ensuremath\mathbb{O}$ operator formalism, we construct the corresponding operators as follows.
First of all, we build a copied unnormalized thermal density matrix:
\begin{equation}
\ensuremath\mathbb{P}^{(1)} \equiv \rho_1 + \dotsb + \rho_\ensuremath K\,,
\end{equation}
with $\rho_\ensuremath k \equiv \exp(-\beta H_\ensuremath k) \otimes \rho_{\text{GND}, \ensuremath k_\perp}$. Here,
$H_\ensuremath k$ is the Hamiltonian on the $\ensuremath k^\text{th}$ stack and we explicitly tensor by $\rho_{\text{GND}, \ensuremath k_\perp}$,
the ground state associated with all the other Hilbert space factors, since as far as the evaluation of the partition function is
concerned, an observer there has no access to any other states.\footnote{It is of course tempting to
also consider a different class of operators defined as $\ensuremath\mathcal{O} = \sum_\ensuremath k O_\ensuremath k \otimes \rho_{\text{GND}, \ensuremath k_\perp}$,
which share many of the same properties as the $\ensuremath\mathbb{O}$ operators. One issue is that it is physically rather awkward to
create an excitation in one stack, and simultaneously enforce a projection onto the ground state of the other stacks.
As already mentioned, however, this is quite appropriate in constructing a corresponding partition function.}
Upon evaluating $\ensuremath K^{-1} \Tr\ensuremath\mathbb{P}^{(1)}$, we then get
just $\overline{Z}$, as in \cref{eq:Zover}. To get the higher order terms such as $Z_\ensuremath k^m$,
we need a corresponding operator acting on a Hilbert space of states, so we
consider $m$ replicas of $\ensuremath\mathcal{H}_\ensuremath k$, namely the $m$-fold tensor product $\ensuremath\mathcal{H}_\ensuremath k^{\otimes m}$.
With respect to this, we introduce the replica density matrix
for the $\ensuremath k^\text{th}$ stack:
\begin{equation}
\rho^{(m, \text{rep})}_\ensuremath k \equiv \rho^{(1)}_\ensuremath k \rho^{(2)}_\ensuremath k \dotsm \rho^{(m - 1)}_\ensuremath k \rho^{(m)}_\ensuremath k\,,
\end{equation}
and then the corresponding copied thermal density matrix including all the replicas is:
\begin{equation}
\ensuremath\mathbb{P}^{(m)} \equiv \rho^{(m, \text{rep})}_1 + \dotsb + \rho^{(m, \text{rep})}_\ensuremath K\,.
\end{equation}
Upon evaluating $\ensuremath K^{-1} \Tr \ensuremath\mathbb{P}^{(m)}$, we then get
just $\overline{Z^m}$. Continuing in this fashion, we can clearly
treat the partition function itself as a random variable,
and evaluate its moments.
\section{Holographic Interpretation}\label{sec:HOLO}
Let us now specialize to the case where each local model is a conformal field theory with a semi-classical AdS gravity dual with the same bulk value of the cosmological constant. In this case, we clearly obtain a large number of AdS throats in which bulk fields $\Phi^\text{bulk}_\ensuremath k $ have a boundary condition set by the particular values of the couplings on the boundary:
\begin{equation}
\Phi^\text{bulk}_\ensuremath k \to \lambda_\ensuremath k\,.
\end{equation}
Our aim will be to understand the sense in which the construction just presented can be interpreted in terms of a \emph{single} AdS throat region. In the process, we will make contact with the baby universe interpretation of ensemble averaging proposed in \cite{Marolf:2020xie}, but one that respects the Swampland constraints of \cite{McNamara:2019rup}.
Our approach to this question will be to focus on the operator subsector defined by the $\ensuremath\mathbb{O}$s of our QFT (now a CFT) sector. Along these lines, consider a local operator $\ensuremath\mathbb{O}(x)$:
\begin{equation}
\ensuremath\mathbb{O}(x) = O_1(x) + \dotsb + O_\ensuremath K(x)\,,
\end{equation}
where each summand has the same field content on its respective stack.
By the standard rules of \cite{Gubser:1998bc}, we know that for each $O_\ensuremath k(x)$, we can (in principle) write down a corresponding bulk field profile in an $\AdS_{(\ensuremath k)}$ geometry. Said differently, each bulk quantity ``casts a shadow'' corresponding to a specific dual in the CFT. On the other hand, precisely because the $O_\ensuremath k$ are built from the same fields, we see that even though we are dealing with a large number of AdS throats, the collective motion described by $\ensuremath\mathbb{O}(x)$ only accesses a single AdS. Said differently, because the connected correlators for the $\ensuremath\mathbb{O}$ operators close, according to the standard AdS/CFT dictionary \cite{Gubser:1998bc}, they reconstruct a single AdS throat.
Another way to arrive at the same conclusion is to consider the geometric entanglement entropy for the ground state between a ball $B$ and its complement $B^c$. Again, if we were initially dealing with a single local sector of our construction, we would simply introduce the pure state $\rho_\ensuremath k = \ket{0}_\ensuremath k \, _\ensuremath k\bra{0}$. As is well known, this has a gravity dual description in terms of a ``minimal area surface'' homologous to $B$, and its ``area'' tracks with the entanglement entropy \cite{Ryu:2006bv,Hubeny:2007xt}. In the present setting, the ground state is given by the tensor product:
\begin{equation}
\rho_\text{GND} = \rho_1 \otimes \dotsb \otimes \rho_\ensuremath K\,.
\end{equation}
In the limit where the additional throats are sequestered from each other, the partial trace collapses to a single ``diagonal'' contribution. Again, the interpretation is that for this set of states, we are building up a single bulk ``minimal area surface''.
Summarizing the discussion so far, we have seen that even though our stringy geometry is building up a large number of AdS throats,
the closed subsector defined by the $\ensuremath\mathbb{O}$ operators only reconstructs a single throat, and this is the one that produces an ensemble-averaged CFT. In other words, only a single bulk AdS geometry is needed to match to the subsector associated with the $\ensuremath\mathbb{O}$s. See Figure \ref{multi to single throat} for an illustrative depiction.
\begin{figure}
\centering
\includegraphics[width=12cm]{multi_to_single.pdf}
\caption{On the left, the string theory construction builds a large number of $D$-dimensional CFTs, each of which has its own dual AdS throat. The closed subset of operators defined by the $\ensuremath\mathbb{O}$s, however, only reconstructs a single AdS throat, which is dual to an ensemble-averaged CFT, shown on the right.}
\label{multi to single throat}
\end{figure}
In the UV complete realization in terms of multiple AdS throats, one might of course ask whether there could be wormhole configurations that join these individual throats, perhaps via some generalization of the construction presented in \cite{Balasubramanian:2020ffd}. The general point is that our construction only mimics an ensemble average provided we have sufficient statistics. For this reason, it is natural to expect that the dominant contribution from saddle points of the Euclidean path integral instead comes from wormhole configurations that have fractionated, i.e., they join many boundaries.
Of course, the notion of ensemble averaging in AdS/CFT has recently been a topic of much interest, and so it is natural to ask how the present description fits with this. To this end, we next turn to a brief summary of the proposal of Marolf and Maxfield (MM) \cite{Marolf:2020xie} in terms of baby universes, and then explain why it can be a valid approximation compatible with the considerations of McNamara and Vafa (McV) \cite{McNamara:2019rup}.
\subsection{Baby Universe Disintegration}
To frame the discussion to follow, we first provide a brief summary of the MM proposal for how to capture the effects of ensemble averaging in AdS/CFT from the perspective of the gravity dual. Following \cite{Marolf:2020xie}, consider an AdS gravity theory with a set of fields denoted as $\Phi$ (including the metric), with boundary conditions labelled by $J$: $\Phi \sim J$. Note that the boundary can have more than one component generically. The Euclidean gravitational path integral defined by an asymptotic boundary with $n$ connected components is then
\begin{equation}\label{eq:def of gravity path integral}
\langle Z[J_1] \dotsm Z[J_n] \rangle \equiv \int_{\Phi \sim J} \@ifnextchar^{\BIGDIFF}{\BIGDIFF^{}}{\Phi} \; e^{-S_\text{grav}[\Phi]}\,,
\end{equation}
where $J_1, \dotsc, J_n$ correspond to different components of the asymptotic boundary. As a point of notation, let us emphasize that here and throughout this subsection, $J_a$ really are just the boundary couplings, but to emphasize that they are \emph{not} associated with a particular set of stacks in our UV completion, we shall write $J_a$, with $a$ having no relation to the indexing of all the $\lambda_\ensuremath k$.
The path integral defined in \cref{eq:def of gravity path integral} cannot generically be factorized into those of disconnected boundaries \cite{Witten:1999xp,Maldacena:2004rf}, e.g., with $n=2$
\begin{equation}
\langle Z[J_1] Z[J_2] \rangle \ne \langle Z[J_1] \rangle \langle Z[J_2] \rangle\,.
\end{equation}
This is because of the presence of Euclidean wormholes corresponding to the bulk manifold whose connected component includes the two boundaries. See \cref{fig:factorization}. From the bulk perspective, this non-factorization comes from the dynamical interactions of two boundaries connected by Euclidean wormholes. From the boundary-CFT point of view, nevertheless, this non-factorization should be rather interpreted as the ensemble average over a classical probability distribution as follows.
\begin{figure}[t!]
\centering
\includegraphics[width=9cm]{factorization.pdf}
\caption{The presence of Euclidean wormholes results in non-factorization of the gravitational path integral.}
\label{fig:factorization}
\end{figure}
By cutting open the above integral carefully so that the intermediate slice intersects no asymptotically AdS boundaries, one defines the baby universe Hilbert space $\ensuremath\mathcal{H}_\text{BU}$ for the complete set of intermediate states separating ``past'' and ``future''. The set of boundary conditions $\{J_1, \dotsc, J_n\}$ is then associated to a state
\begin{equation}
\ket{Z[J_1] \dotsm Z[J_n]} \in \ensuremath\mathcal{H}_\text{BU}\,.
\end{equation}
One special state is the Hartle--Hawking state with no boundary. Its norm gives rise to the ``cosmological partition function''
\begin{equation}
\braket{\text{HH}}{\text{HH}} =\int_\text{no boundary} \@ifnextchar^{\BIGDIFF}{\BIGDIFF^{}}{\Phi} \; e^{-S_\text{grav}[\Phi]}\,.
\end{equation}
Next, introduce operators $\widehat{Z[J]}$ on $\ensuremath\mathcal{H}_\text{BU}$ for any boundary condition $J$ so that
\begin{equation}
\widehat{Z[J]} \ket{Z[J_1] \dotsm Z[J_n]} = \ket{Z[J] Z[J_1] \dotsm Z[J_n]}\,.
\end{equation}
The eigenstates of $\widehat{Z[J]}$ then form a basis of $\ensuremath\mathcal{H}_\text{BU}$, and are defined by
\begin{equation}
\widehat{Z[J]} \ket{\alpha} = Z_\alpha[J] \ket{\alpha} \quad \forall J\,,
\end{equation}
with $\braket{\alpha'}{\alpha} = \delta_{\alpha' \alpha}$.
Note that any boundary condition can be derived from the corresponding operators acting on Hartle--Hawking state as
\begin{equation}
\ket{Z[J_1] \dotsm Z[J_n]} = \widehat{Z[J_1]} \dotsm \widehat{Z[J_n]} \ket{\text{HH}}\,.
\end{equation}
The gravitational path integral in \cref{eq:def of gravity path integral} can then be expressed as
\begin{equation}
\begin{aligned}
\langle Z[J_1] \dotsm Z[J_n] \rangle &= \braopket{\text{HH}}{\widehat{Z[J_1]} \dotsm \widehat{Z[J_n]}}{\text{HH}} \\
&= \sum_{\alpha_0, \dotsc, \alpha_n} \braket{\text{HH}}{\alpha_0} \braopket{\alpha_0}{Z[J_1]}{\alpha_1} \dotsm \braopket{\alpha_{n - 1}}{Z[J_n]}{\alpha_n} \braket{\alpha_{n}}{\text{HH}} \\
&=\braket{\text{HH}}{\text{HH}} \sum_\alpha p_\alpha Z_\alpha[J_1] \dotsm Z_\alpha[J_n]\,,
\end{aligned}
\end{equation}
where $p_\alpha$ is the probability for each $\alpha$ state, computed by $p_\alpha = \frac{\abs{\braket{\text{HH}}{\alpha}}^2}{\braket{\text{HH}}{\text{HH}}}$.
The appearance of the $\ket{\alpha}$ states is rather disturbing, especially in the context of string theory where we have no evidence at all for such tunable parameters. Indeed, in string constructions coupled to gravity, all known examples of coupling constants descend from dynamical moduli. This point was significantly sharpened in \cite{McNamara:2019rup} where they showed that in even more general terms, the cobordism hypothesis of the Swampland program is enough to require the baby universe Hilbert space $\ensuremath\mathcal{H}_\text{BU}$ to be one-dimensional, namely we can indeed speak of fixing the boundary values of the coupling constants, just as we would in ``standard'' AdS/CFT. From this perspective, ensemble averaging really has no general meaning in a UV complete framework such as string theory.
We now argue that in spite of appearances, our embedding in string theory provides a way to make contact with both proposals.
First of all, the very fact that we have an ensemble average means that any putative gravity dual will likely have to match on to the
characterization provided by the MM picture. Indeed, assuming we can work out the ``standard'' AdS dual for a single throat in our construction, we already know the bulk field content, and thus in principle can discuss the computation of the Hartle--Hawking wave function for this theory.
Note that the probability distribution $p_\text{bin}(\lambda)$ we engineer
in string theory does not directly correspond to the probability $p_\alpha$
for $\alpha$ states, though the two notions are clearly implicitly related in some way.
From the perspective of the bulk gravitational system, it is tempting
to say that the specific details of the path integral dictate a particular ``preferred choice'' for $p_\text{smooth}(\lambda)$
(see, e.g., \cite{Saad:2019lba, Stanford:2019vob, Marolf:2020xie}), although even this relies on having enough data in the form of a specific set of $\alpha$ states, and a specific choice of bulk gravitational action with which to construct the Hartle--Hawking state in the first place. Indeed, turning the discussion around, there seems to be little constraint on what sort of $p_\text{bin}(\lambda)$s we can end up generating,
and so we leave it as an interesting question to determine precisely how to fill in this entry of the AdS/CFT correspondence.
From the perspective of the present construction, we take this to mean that there are ambiguities in specifying the Euclidean path integral
for quantum gravity, and resolving these ambiguities in different ways can result in different choices for the ensemble average in the boundary theory.
Once we accept the existence of the $\ket{Z[J_1] \dotsm Z[J_n]}$ states, the appearance of the baby universe states $\ket{\alpha}$ would appear to follow. At this point, however, we recall that our ensemble average picture can break down, both at high energies, but also entropically whenever $n$ gets sufficiently large. Probing either regime of validity shatters the illusion and we can no longer work in terms of a single AdS throat region. For example, when the number of boundary components $n$ becomes sufficiently large, (i.e., of order $n_*$ of \cref{eq:discDist}), then we have already seen there can be an entropic breakdown. Observe that when $n$ is really of order $\ensuremath K$, we can even resolve the individual AdS throats, and so in this limit, we just recover the standard AdS/CFT dictionary, with a single one-dimensional baby universe Hilbert space for each throat, much as in \cite{McNamara:2020uza}.
Of course, one of the main reasons to seek out an AdS/CFT interpretation of our ensemble average system is its potential use in
studying aspects of quantum gravity in anti-de Sitter space. Along these lines, it is also natural to ask about whether the approximation is reliable enough to provide access to the microscopic details of a black hole. To arrange this, we follow the procedure in \cite{Maldacena:1997re, Witten:1998zw} and consider the CFT on the background $S^1 \times S^{D - 1}$. The size of the thermal circle sets a corresponding temperature (and thus size) for an AdS--Schwarzschild geometry, but one can of course consider more elaborate configurations with various chemical potentials switched on. Now, suppose we are interested in probing the $n_\text{micro}$ microstates of this black hole, perhaps as captured by Hawking radiation quanta. To construct a quantity such as the Page curve, a boundary observer will need to sample order $n^2_\text{micro}$ times \cite{Page:1993wv}. However, if $n_\text{micro}$ exceeds $\sqrt{n_*} \sim \varepsilon^{-1}$ of \cref{eq:nstar}, then the ensemble interpretation becomes problematic. In light of this, it is unclear (at least to us) how we can use this setup to learn about the Page curve of a macroscopic black hole.
It is what it is.
\section{Examples} \label{sec:EXAMPLES}
Having demonstrated the main ideas behind ensemble averaging in the context of string theory, we now turn to explicit examples.
Our aim here is not to be exhaustive, but rather to showcase a few different methods, including their advantages and disadvantages.
For starters, we will focus on supersymmetric quantum field theories since these are the ones over which we have maximal control. We shall also primarily focus on superconformal field theories (SCFTs), since these have a chance (in a suitable large-$N$ limit) of having a holographic dual. Within this setting, we identify two general methods for building an ensemble, one that we refer to as various brane box models (and their dual incarnations), and another based on building up new compactification geometries by fibering an existing geometry over a non-compact subvariety of its moduli space.
One way to build a large class of QFTs is with D-branes that end on NS5-branes, so-called ``brane box'' models. Arranging these NS5-branes in various repeating patterns such as $d$-dimensional generalizations of a cube, we can produce a rich set of possible SCFTs. For our purposes, the important point is that for $D \le 4$ systems, the relative positions between the NS5-branes descend to non-normalizable parameters of the resulting quantum field theory. We get an ensemble average by repeating this construction in some of the directions transverse to the original $d$ dimensions used to make a single instance of the SCFT. For example, if we attempt to engineer a $D$-dimensional QFT using spacetime-filling branes in the geometry $\ensuremath\mathbb{R}^{D - 1, 1} \times \ensuremath\mathbb{R}^d \times Y_{10 - D - d}$, then the entire configuration sits at a single point of the transverse $Y$ geometry, which in many cases of interest is just $\ensuremath\mathbb{R}^{10 - D - d}$. Moving to a different point of $Y$, we can then arrange for a different choice of couplings. By a chain of T-dualities, these constructions can also be related to the worldvolume theory of branes probing singularities, and these in turn can often be generated by appropriate compactifications of 6D SCFTs and little string theories (LSTs). An important feature of this method of construction is that precisely because we can tune the moduli in the transverse $Y$ directions, we can use this to engineer an essentially arbitrary probability distribution over the couplings of the model. The disadvantage of this approach is that in some cases it is difficult to guarantee that we generate a theory with a candidate AdS dual.
Another way to generate examples consists of taking a system of branes wrapped on subspaces of a Calabi--Yau $d$-fold $X_{2 d}$ (a $2 d$-real-dimensional space). More precisely, we assume that the geometry takes the form $\ensuremath\mathbb{R}^{D - 1, 1} \times Y_{2 m} \times X_{2 d}$, where $Y$ is an $m$-complex-dimensional geometry transverse to the branes. In particular, our branes sit at a particular point of $Y$ and the parameters of the QFT descend from the geometric moduli of $X$. To get an ensemble average, we consider $\ensuremath\mathcal{M}_X$, the moduli space of $X$. Observe that we can consider the total space $\ensuremath\mathcal{X}$ as defined by $X \to \ensuremath\mathcal{X} \to B$, where $B \subset \ensuremath\mathcal{M}_X$. Cutting out an $m$-complex-dimensional subspace $B$ of $\ensuremath\mathcal{M}_X$ that has no singular fibers then generates a non-compact Calabi--Yau of real dimension $2 m + 2 d$.\footnote{We thank T.~Pantev for discussions on this point.} We can also entertain more general fibrations, possibly with singular fibers, and we present some explicit examples of precisely this sort in \cref{sec:FIBRATION}. An important advantage of this approach is that such brane constructions in Calabi--Yau compactifications often come with readily defined AdS duals. A drawback of this approach is that it does not, in general, allow us to engineer an arbitrary ensemble average. This is simply because the best we can do is to sweep out a probability distribution with support on an $m$-complex-dimensional subspace of the full moduli space.
In the remainder of this \namecref{sec:EXAMPLES}, we turn to some particular examples, illustrating
the pitfalls, the possibilities, the perils, and the promise of generating ensemble averages.
\subsection{$D = 4$}
We now engineer some ensembles of $D = 4$ SCFTs. We begin by constructing an ensemble for $\ensuremath\mathcal{N} = 4$ Super Yang--Mills theory, and then turn to examples with lower supersymmetry.
\subsubsection{Warmup: Approximating $\ensuremath\mathcal{N} = 4$ SYM}
Perhaps the simplest case to consider is that of type IIB string theory on the background $\ensuremath\mathbb{R}^{3, 1} \times \ensuremath\mathbb{C}^3$ with a stack of $N_\text{c}$ spacetime-filling D3-branes sitting at a point of $\ensuremath\mathbb{C}^3$. At low energies, the open string degrees of freedom realize $\ensuremath\mathcal{N} = 4$ Super Yang--Mills theory with gauge group $\U(N_\text{c})$. The value of the complexified gauge coupling $\tau$ is controlled by the background value of the type IIB axio-dilaton:
\begin{equation}
\tau = C_{0} + i \exp(-\phi)\,.
\end{equation}
In type IIB supergravity, this is characterized by the extremal brane solution with constant axio-dilaton profile, localized source for the self-dual five-form flux, and metric (see, e.g., \cite{Horowitz:1991cd}):
\begin{equation}
\@ifnextchar^{\DIFF}{\DIFF^{}}{s}^2 = H^{-1 / 2} \@ifnextchar^{\DIFF}{\DIFF^{}}{s}_{\ensuremath\mathbb{R}^{3, 1}}^2 + H^{1/2} \@ifnextchar^{\DIFF}{\DIFF^{}}{s}_{\ensuremath\mathbb{C}^3}^2\,, \quad \text{with } H = 1 + \frac{4 \pi g_\text{s} N_\text{c} \alpha'^2}{r^4}\,,
\end{equation}
where $r$ is the distance from the D3-brane stack. As is well known, in the near-horizon limit, this produces an $\AdS_5 \times S^5$ geometry with $N_\text{c}$ units of self-dual five-form flux threading the two factors \cite{Maldacena:1997re}. The AdS radius and $S^5$ radius $L$ are correlated, and related to the open string parameters as $L^4 = 2 g_\text{YM}^2 N_\text{c} \alpha'^4$. We are, of course, free to consider moving these D3-brane stacks to separate points in $\ensuremath\mathbb{C}^3$. In all these local sectors, the value of the axio-dilaton is always the same, and we generate a rather trivial probability distribution of values for the axio-dilaton. The supergravity approximation is the same, the only change being the harmonic function $H$, which is now given by
\begin{equation}
H = 1 + \sum_\ensuremath k \frac{4 \pi g_\text{s} N_\text{c} \alpha'^2}{\abs{\V{r} - \V{r}_\ensuremath k}^4}\,,
\end{equation}
with $\V{r}_k$ the position of the $k$th stack.
To get a more general class of distributions, we now introduce an additional source as specified by a stack of D7-branes that sits at a point of the middle factor in $\ensuremath\mathbb{R}^{3,1} \times \ensuremath\mathbb{C}_\perp \times \ensuremath\mathbb{C}^2$, and fills the remaining eight directions.
The brane configuration is as follows:
\vspace{2em}
\begin{center}
\begin{tabularx}{0.5\textwidth}{c|*{10}{>{\centering\arraybackslash $}X<{$}}}
& \tikzmark{r31L} 0 & 1 & 2 & 3 \tikzmark{r31R} & \tikzmark{cPerpL} 4 & 5 \tikzmark{cPerpR} & \tikzmark{c2L} 6 & 7 & 8 & 9 \tikzmark{c2R} \\ \hline
D3 & \times & \times & \times & \times & & & & & & \\
D7 & \times & \times & \times & \times & & & \times & \times & \times & \times
\end{tabularx}
%
\begin{tikzpicture}[overlay, remember picture]
\draw [decorate,decoration={brace,amplitude=10pt,raise=4pt}] (r31L.west) --node[above=14pt]{$\ensuremath\mathbb{R}^{3, 1}$} (r31R.east);
\draw [decorate,decoration={brace,amplitude=10pt,raise=4pt}] (cPerpL.west) --node[above=14pt]{$\ensuremath\mathbb{C}_\perp$} (cPerpR.east);
\draw [decorate,decoration={brace,amplitude=10pt,raise=4pt}] (c2L.west) --node[above=14pt]{$\ensuremath\mathbb{C}^2$} (c2R.east);
\end{tikzpicture}
\end{center}
Doing so produces a position-dependent profile for the axio-dilaton, but also breaks half the supersymmetry in the system. Indeed,
if we now have a stack of D3-branes located at distinct points of the geometry, then each can experience a different value of the axio-dilaton $\tau_\ensuremath k$, and the low-energy effective action on each stack is of the form
\begin{equation}
S_\ensuremath k = S_{\ensuremath\mathcal{N} = 4}(\tau_\ensuremath k) + S_{\ensuremath\mathcal{N} = 2}\,,
\end{equation}
where the contribution from explicit $\ensuremath\mathcal{N} = 2$ breaking terms is captured by a collection of higher-dimension operators. The precise form of these contributions can be worked out by noting that this D3/D7 system is just engineering a 4D $\ensuremath\mathcal{N} = 2$ gauge theory. There are two distance scales that control the strength of these higher-dimension operators. One is the relative separation between the D3-branes,
\begin{equation}
\Lambda_{\ensuremath k_1, \ensuremath k_2} \equiv \frac{\dist(\DThree_{\ensuremath k_1}, \DThree_{\ensuremath k_2})}{\alpha'}\,,
\end{equation}
as measured in the full $\ensuremath\mathbb{C}^3$ factor transverse to all the D3-branes,
and the other is the relative separation between the D3-branes and the D7-branes,
\begin{equation}
\Lambda_{\ensuremath k, \DSeven} \equiv \frac{\dist(\DThree_\ensuremath k, \DSeven)}{\alpha'}\,,
\end{equation}
as measured in the $\ensuremath\mathbb{C}_\perp$ factor transverse to the D7-branes. Provided we only ask questions at low energies compared with these cutoffs, we get an adequate approximation to ensemble averaging in $\ensuremath\mathcal{N} = 4$ SYM, but one that has an $\ensuremath\mathcal{N} = 2$ UV completion.
Let us now turn to the class of ensembles we can actually engineer in this setting. First of all, the whole point of introducing a stack of D7-branes is that we can thus generate a position-dependent axio-dilaton. More broadly, this and more general choices of non-perturbative bound states of 7-branes can be understood in F-theory \cite{Vafa:1996xn, Morrison:1996na, Morrison:1996pp} by considering a non-compact elliptically fibered K3 surface with minimal Weierstrass model:
\begin{equation}
y^2 = x^3 + f(z) x + g(z)\,,
\end{equation}
where in the present setting, $f(z)$ and $g(z)$ are treated as polynomials in the holomorphic coordinate $z$ of the $\ensuremath\mathbb{C}_\perp$ factor. The possible values of the axio-dilaton are implicitly encoded in the $\SL(2, \ensuremath\mathbb{Z})$-invariant $j$-function:
\begin{equation}
j = 1728 \frac{4 f^3}{4 f^3 + 27 g^2}\,,
\end{equation}
which has the weak-coupling expansion in $q = \exp(2 \pi i \tau)$ given by
\begin{equation}
j = q^{-1} + 744 + 196884 q + \dotsb\,.
\end{equation}
For example, $j = \infty$ corresponds to weak coupling at $\tau = i \infty$, and $j = 1728$ corresponds to $\tau = i$, while $j = 0$ corresponds to $\tau = \exp(2 \pi i / 6)$. In the case of a single stack of $M$ D7-branes sitting at $z = 0$, we just have
\begin{equation}
j_{\DSeven} = z^{-M}\,.
\end{equation}
To get a particular value of the axio-dilaton (or, more precisely, its $j$-invariant), we simply consider D3-branes at the desired value of $z$. Note that this intrinsically comes with some limitations, because to populate the distribution near the $\tau = i \infty$ region of moduli space, we necessarily must move close to the stack of D7-branes, which in turn lowers the UV scale $\Lambda_{\ensuremath k, \DSeven}$ in our effective field theory. The other issue we face is how to sequester the D3-brane stacks from one another. This is less problematic, because even if they sit at the same point of $\ensuremath\mathbb{C}_\perp$, we are free to move them away from each other in the $\ensuremath\mathbb{C}^2$ factor. We thus conclude that the UV cutoff is set by $\Lambda_{\ensuremath k, \DSeven}$, and this in turn depends on what sort of distribution we wish to engineer.
Consider next the supergravity background generated by our D3/D7 system. Since we are dealing with D3-brane probes of an F-theory geometry, the main change is that the metric on the $\ensuremath\mathbb{C}_\perp$ factor is controlled by that of the non-compact elliptically fibered K3 space $\KThree \to \ensuremath\mathbb{C}_\perp$. In particular, we observe that in the near-horizon limit for each local stack, we indeed get a collection of individual AdS throats, but of different sizes as dictated by the local profile of the $\tau_\ensuremath k$. Note also that the subleading $\ensuremath\mathcal{N} = 2$ breaking terms amount to a deviation away from a pure $\AdS_5 \times S^5$ geometry.
Precisely because the value of the bulk cosmological constant is different for each local sector, there is no sense in which we can give a holographic interpretation in terms of a single AdS geometry. This is in accord with the fact that consistency of the MM proposal requires the baby universe Hilbert space interpretation to be trivial (i.e., one-dimensional) in this special case \cite{Marolf:2020xie, McNamara:2020uza}.
\subsubsection{Quiver Gauge Theory Ensemble}
We now proceed to engineer an ensemble average of quiver gauge theories that enjoys a holographic dual. The model we consider consists of $N_\text{c}$ D3-branes probing the orbifold singularity $\ensuremath\mathbb{C}^2 / \ensuremath\mathbb{Z}_M$ with group action $(u, v) \mapsto (\xi u, \xi^{-1} v)$ on the holomorphic coordinates, where $\xi$ is a primitive $M$th root of unity. As is well known from \cite{Douglas:1996sw, Lawrence:1998ja}, the worldvolume theory for this model results in a 4D $\ensuremath\mathcal{N} = 2$ SCFT described by a quiver gauge theory with gauge groups arranged in a circular ring, joined by hypermultiplets in bifundamental representations (see \cref{fig:N=2 quiver}). The model also comes with a collection of marginal parameters, captured by the holomorphic couplings $(\tau^{(1)}, \dotsc, \tau^{(M)}) \equiv \V{\tau}$. These are encoded in the choice of closed string moduli.
\begin{figure}[t!]
\centering
\includegraphics[width=4cm]{N=2quiver.pdf}
\caption{Quiver diagram of the 4D $\ensuremath\mathcal{N} = 2$ SCFT obtained from $N_\text{c}$ D3-branes probing $\ensuremath\mathbb{C}^2 / \ensuremath\mathbb{Z}_M$, for $M = 4$.
Each node represents an $\SU(N_\text{c})$ gauge group, and links between them denote bifundamental hypermultiplets.}
\label{fig:N=2 quiver}
\end{figure}
To better understand the sense in which these closed string moduli are tunable, it is helpful to consider some dual realizations of the same low-energy effective field theory. One way to proceed is to observe that there is a 6D little string theory obtained via F-theory from a configuration of collapsing $-2$-curves arranged in a circular ring (see, e.g., \cite{Bhardwaj:2015oru}). Wrapping $N_\text{c}$ D7-branes on each $-2$-curve (a Kodaira $I_{N_\text{c}}$ fiber) then results in a 6D quiver gauge theory. Compactifying on a further $T^2$ then produces the desired 4D $\ensuremath\mathcal{N} = 2$ SCFT.
In this realization, the gauge couplings descend from the complexified K{\"a}hler volume of $T^2 \times \Sigma_m$, for $m = 1, \dotsc, M$ labelling the different $-2$ curves. In particular, each of these is a tunable complexified gauge coupling in the 4D field theory.
The tensor branch of the 6D little string theory can also be realized from a configuration of $M$ NS5-branes arranged in a circular ring with $N_\text{c}$ D6-branes suspended in between each neighboring pair. In this picture, the relative distance between each NS5-brane sets the value of 6D gauge coupling, and further compactification on a $T^2$ again results in the same 4D gauge theory.
To be more explicit, let us consider T-dualizing the $T^2$ wrapped by D6-branes; the 4D gauge theory is then realized alternatively by the following D4/NS5 system:
\begin{equation*}
\begin{array}{c|*{10}{c}}
& 0 & 1 & 2 & 3 & 4 & 5 & 6 & 7 & 8 & 9 \\ \hline
N_\text{c} \text{ D4s} & \times & \times & \times & \times & & & \times & & & \\
M \text{ NS5s} & \times & \times & \times & \times & \times & \times & & & &
\end{array}
\end{equation*}
See \cref{fig:D4-NS5} for an illustration of this brane configuration in the 4, 5, and 6 directions. Observe also that a T-duality on direction 6 directly connects this construction to that of $N_\text{c}$ D3-branes probing the same $\ensuremath\mathbb{C}^2 / \ensuremath\mathbb{Z}_M$ singularity. In all these cases, then, we have a geometric characterization of the resulting moduli.
\begin{figure}[t!]
\centering
\includegraphics[width=11cm]{D4_NS5.pdf}
\caption{D4/NS5 system leading to the 4D quiver gauge theory. Each D4-brane segment corresponds to a gauge group, whose coupling is set by the relative distance in the 6 direction between the two boundary NS5-branes as $\frac{1}{g_i^2}=\frac{x_{i + 1}^6 - x_i^6}{g_\text{s} \sqrt{\alpha'}}$. The effective 4D SCFT on the D4-brane worldvolume is the same quiver gauge theory shown in \cref{fig:N=2 quiver}.}
\label{fig:D4-NS5}
\end{figure}
Let us now turn to the construction of an ensemble average. Returning to the picture of D3-branes probing an orbifold, observe that we can generate a large ensemble of local singularities via the hypersurface in $\ensuremath\mathbb{C}^3$ given by:
\begin{equation}
y^2 = x^2 + \prod_\ensuremath k (z - z_\ensuremath k)^{M_\ensuremath k}\,.
\end{equation}
Near $x = y = (z - z_\ensuremath k) = 0$, this generates a $\ensuremath\mathbb{C}^2 / \ensuremath\mathbb{Z}_{M_\ensuremath k}$ singularity, so if we take $M_\ensuremath k = M$ for all $\ensuremath k$, and place $N_\text{c}$ D3-branes near each singularity, we produce an ensemble of such systems. Observe that because the closed string moduli of each singularity are decoupled, we can arrange a nearly arbitrary ensemble. The same is clear from working with any of the other local pictures. Not also that there is a natural UV completion available in terms of a stack of $\ensuremath K N_\text{c}$ D3-branes probing a $\ensuremath\mathbb{C}^2 / \ensuremath\mathbb{Z}_{\ensuremath K M}$ singularity.
In the large-$N_\text{c}$ limit, we can also arrange for a holographic dual description. To get the same size AdS, we just require that the value of the string theory dilaton (in the D3-brane probe picture) is the same for each stack. In the quiver gauge theory, this is controlled by the particular combination of parameters
\begin{equation}
\tau_\text{IIB} = \tau^{(1)} + \dotsb + \tau^{(M)}
\end{equation}
at each local model, which we take to be equal for each value of $\ensuremath k = 1, \dotsc, \ensuremath K$. In this description, the remaining $M - 1$ degrees of freedom are tunable marginal couplings over which our ensemble runs.
Clearly, one can consider far more elaborate examples, but the essential point is that at least for $D = 4$ SCFTs, we can engineer a wide variety of ensemble averages.
\subsection{$D = 3$}
Let us briefly comment on the construction of $D = 3$ SCFTs, and issues with engineering ensemble averages in this setting. First of all, we can just take a $D = 4$ SCFT and compactify it on a circle. In many cases, this can also result in a 3D SCFT. From the perspective of the examples just considered, we could alternatively start from a 6D LST or 6D SCFT and compactify it on a three-manifold with negative sectional curvature. The primary challenge from this perspective is how to build a class of compactification geometries where we can vary the marginal couplings.
\footnote{A general class of 3D SCFTs can be obtained by compactification of 6D $\ensuremath\mathcal{N} = (2,0)$ SCFTs on a three-manifold, which are known as the $T[M_3]$ theories (see, e.g., \cite{Dimofte:2011ju,Gadde:2013sca}). When $M_3 = \Sigma \times S^1$, where $\Sigma$ is a Riemann surface, $T[M_3]$ admits marginal deformations. However, the holographic dual of $T[M_3]$ in this case is not $\AdS_4$ gravity, but rather a ``gravitational domain wall" separating two AdS regions (see, e.g., \cite{Gukov:2015qea}). Even if one focuses on the CFT side, the top-down approach to building an ensemble average of these 3D SCFTs is not clear to us. We thank S.~Gukov for discussions on this point.}
To this general point, one might attempt to realize examples using M-theory on an eight-manifold with a large number of singularities. The theory of $N_\text{c}$ M2-branes probing any one singularity will produce an $\AdS_4 \times X_7$ background, and in principle the geometric moduli of $X_7$ can be traced to corresponding deformations in the 3D SCFT. One can consider, for instance, an ensemble average over the Chern--Simons levels of the ABJ(M) model \cite{Aharony:2008ug, Aharony:2008gk} by introducing $\ensuremath K$ copies of $N_\text{c}$ M2-branes probing $\ensuremath\mathbb{C}^4 / \ensuremath\mathbb{Z}_{M_\ensuremath k}$ singularities. The difficulty we now face is that the radius of the dual $\AdS_4$ scales as $L^2 \propto N_\text{c} / M_\ensuremath k$ with other parameters fixed. On the other hand, if we attempt to average over the Chern--Simons couplings, we must then simultaneously adjust the value of $N_\text{c}$. It would be interesting to build an explicit ensemble average in this case, but this would seem to require first obtaining a better understanding of how the geometry of a given background descends to marginal parameters of
the field theory.\footnote{See references \cite{Gauntlett:2005jb, Imeroni:2008cr, Bobev:2021gza, Bobev:2021yya} for some examples of how these marginal couplings are realized in various $\AdS_4$ gravity duals. However, the geometric or brane realization counterparts of these deformations are not yet known, so it is unclear to us how to build an ensemble with random couplings from a top-down approach.
We thank N.~Bobev and A.~Tomasiello for helpful correspondence on this point.}
\subsection{$D = 2$}
Let us now turn to the construction of $D = 2$ SCFTs, and, when appropriate, their $\AdS_3$ duals.
We begin with an example of a brane box model where we have a great deal of freedom in generating
an ensemble average. The downside, however, is that the existence of an $\AdS_3$ dual is not always apparent.
The other general method we consider yields an $\AdS_3 \times S^3 \times X_4$ dual, but at the expense of a less-flexible class of possible probability distributions.
\subsubsection{Brane Box Examples}
Similarly to 4D SCFTs, in addition to branes probing geometries, one can also engineer 2D theories using intersecting branes. Let us now consider 2D $\ensuremath\mathcal{N} = (0, 2)$ theories engineered by the brane box model \cite{Garcia-Compean:1998sla,Franco:2015tya}.
The brane configuration consists of $M$ D4-branes, $\ensuremath m$ NS5-branes, $\ensuremath m'$ NS5$'$-branes, and $\ensuremath m''$ NS5$''$-branes, as follows:
\begin{equation*}
\begin{array}{c|*{10}{c}}
& 0 & 1 & 2 & 3 & 4 & 5 & 6 & 7 & 8 & 9 \\ \hline
\text{D4} & \times & \times & \times & & \times & & \times & & & \\
\text{NS5} & \times & \times & \times & \times & \times & \times & & & & \\
\text{NS5}' & \times & \times & \times & \times & & & \times & \times & & \\
\text{NS5}'' & \times & \times & & & \times & \times & \times & \times & &
\end{array}
\end{equation*}
where all branes sit on the same position in the $8, 9$-plane. The D4-branes are finite in the 2, 4, and 6 directions as a $T^3$, and bounded by three types of NS5-branes. For each pair of parallel NS5-branes, the brane configuration in directions 2, 4, and 6 is a 3D box filled by D4-branes. With $\ensuremath m$ NS5-branes, $\ensuremath m'$ NS5$'$-branes, and $\ensuremath m''$ NS5$''$-branes, we have a 3D grid containing $\ensuremath m \cdot \ensuremath m' \cdot \ensuremath m''$ boxes, as illustrated in \cref{fig:brane box}.
\begin{figure}[t!]
\centering
\includegraphics[width=10cm]{brane_box.pdf}
\caption{Brane box model on $T^3$ with $\ensuremath m = 2, \ensuremath m' = 3$ and $\ensuremath m'' = 2$. Every box bounded by NS5-, NS5$'$-, and NS5$''$-branes is filled by $M$ D4-branes, corresponding to one $\U(M)$ gauge group in the 2D effective field theory.}
\label{fig:brane box}
\end{figure}
The effective field theory on D4-branes in the non-compact 0 and 1 directions is a 2D $\ensuremath\mathcal{N} = (0, 2)$ gauge theory with $\U(M)^{\ensuremath m \cdot \ensuremath m' \cdot \ensuremath m''}$ gauge symmetry. Each $\U(M)$ gauge group corresponds to one brane box on the $T^3$ comprising the 2, 4, and 6 directions. Performing T-duality on this $T^3$, the brane configuration becomes D1-branes probing a $\ensuremath\mathbb{C}^4 / (\ensuremath\mathbb{Z}_\ensuremath m \times \ensuremath\mathbb{Z}_{\ensuremath m'} \times \ensuremath\mathbb{Z}_{\ensuremath m''})$ singularity, with orbifold group action:
\begin{equation}
\begin{aligned}
(z_1, z_2, z_3, z_4) &\to (z_1, z_2, e^{\frac{2 \pi i}{\ensuremath m}} z_3, e^{\frac{2 \pi i}{\ensuremath m}} z_4)\,, \\
(z_1, z_2, z_3, z_4) &\to (z_1, e^{\frac{2 \pi i}{\ensuremath m'}} z_2, z_3, e^{\frac{2 \pi i}{\ensuremath m'}} z_4)\,, \\
(z_1, z_2, z_3, z_4) &\to (e^{\frac{2 \pi i}{\ensuremath m''}} z_1, z_2, z_3, e^{\frac{2 \pi i}{\ensuremath m''}} z_4)\,,
\end{aligned}
\end{equation}
which indeed preserves $\ensuremath\mathcal{N} = (0, 2)$ supersymmetry on the 2D worldvolume of the D1-branes.
Due to the non-vanishing elliptic genera of these 2D $\ensuremath\mathcal{N} = (0, 2)$ theories computed in \cite{Franco:2017cjj}, it is natural to conjecture that these theories flow to SCFTs in the IR.\footnote{To our knowledge, the explicit SCFTs at the IR fixed point and the AdS gravity duals for these $\ensuremath\mathcal{N} = (0,2)$ theories have not yet been constructed, which by itself is a problem deserving further investigation.} Here, we will assume the existence of the SCFT in the IR and build the ensemble average.\footnote{We thank N.~Benjamin, S.~Franco, and S.~Gukov for helpful discussions on this point.}
One of the marginal couplings in 2D is the Fayet--Iliopoulos (FI) term for the $\U(1)$ factor of each $\U(N)$ gauge group. The corresponding FI parameters are encoded in the positions of the NS5-branes in several directions. Namely, for a given $\U(M)$ gauge group associated with the $(\ensuremath m, \ensuremath m', \ensuremath m'')$-th brane box, its FI parameter is given in terms of the separations of the NS5-, NS5$'$-, and NS5$''$-branes in directions 7, 5, and 3, respectively (see, e.g., \cite{Garcia-Compean:1998sla}):
\begin{equation}
r_{\ensuremath m, \ensuremath m', \ensuremath m''} = \frac{(x^7_{\ensuremath m + 1} - x^7_\ensuremath m) + (x^5_{\ensuremath m' + 1} - x^5_{\ensuremath m'}) + (x^3_{\ensuremath m'' + 1} - x^3_{\ensuremath m''})}{\sqrt{\alpha'}}.
\end{equation}
We can build an ensemble of brane box models by repeating this construction at different points in the $8, 9$-plane. Indeed, tuning the relative positions of the NS5-, NS5$'$-, and NS5$''$-branes in each instance of the construction, we can engineer an ensemble of 2D $\ensuremath\mathcal{N} = (0, 2)$ SCFTs in which the FI-parameters are drawn from a probability distribution.
A drawback of this approach is that although this provides us with a way to generate ensemble averaging in 2D SCFTs, the existence of a putative AdS dual is somewhat unclear. We now turn to examples for which we understand the holographic dual more clearly.
\subsubsection{Holographic Example}
To generate some examples with a holographic dual, we consider type IIB string theory on the background $\ensuremath\mathbb{R}_\text{time} \times \ensuremath\mathbb{C}^2 \times S^1 \times \KThree$, with $N_5$ D5-branes wrapped on $S^1 \times \KThree$ and $N_1$ D1-branes wrapped on the $S^1$ factor. We keep the D1/D5 system coincident at the same point of $\ensuremath\mathbb{C}^2$. This engineers a 2D SCFT with $\ensuremath\mathcal{N} = (4, 4)$ supersymmetry on the spacetime $\ensuremath\mathbb{R}_\text{time} \times S^1$. The brane configuration is as follows:
\vspace{2em}
\begin{center}
\begin{tabularx}{0.6\textwidth}{c|*{10}{>{\centering\arraybackslash $}X<{$}}}
& \tikzmark{rTimeL} 0 \tikzmark{rTimeR} & \tikzmark{s1L} 1 \tikzmark{s1R} & \tikzmark{c2L} 2 & 3 & 4 & 5 \tikzmark{c2R} & \tikzmark{k3L} 6 & 7 & 8 & 9 \tikzmark{k3R} \\ \hline
D1 & \times & \times & & & & & & & & \\
D5 & \times & \times & & & & & \times & \times & \times & \times
\end{tabularx}
%
\begin{tikzpicture}[overlay, remember picture]
\draw [decorate,decoration={brace,amplitude=10pt,raise=4pt}] (rTimeL.west) --node[above=14pt]{$\ensuremath\mathbb{R}_\text{time}$} (rTimeR.east);
\draw [decorate,decoration={brace,amplitude=10pt,raise=4pt}] (s1L.west) --node[above=14pt]{$S^1$} (s1R.east);
\draw [decorate,decoration={brace,amplitude=10pt,raise=4pt}] (c2L.west) --node[above=14pt]{$\ensuremath\mathbb{C}^2$} (c2R.east);
\draw [decorate,decoration={brace,amplitude=10pt,raise=4pt}] (k3L.west) --node[above=14pt]{$\KThree$} (k3R.east);
\end{tikzpicture}
\end{center}
The corresponding supergravity solution describes an extremal black string, and in the near-horizon limit, it is given by the geometry $\AdS_3 \times S^3 \times \KThree$, which figures prominently in the $\AdS_3 / \CFT_2$ correspondence \cite{Maldacena:1997re} (see also \cite{Brown:1986nw}). The size of the AdS radius in Planck units is proportional to $N_1 N_5$, which is in turn proportional to the central charge of the 2D SCFT $c = 6 N_1 N_5$. For our purposes, the important point is that this 2D SCFT comes with a set of marginal couplings controlled by the metric moduli of the K3 surface, and geometrically, these can be understood (via duality) as Narain moduli of the sort that appear in \cite{Maloney:2020nni,Afkhami-Jeddi:2020ezh,Benjamin:2021wzr}.
To generate an ensemble, we begin by observing that $\ensuremath\mathbb{C}^2 = \ensuremath\mathbb{C}_\text{seq} \times \ensuremath\mathbb{C}_\text{base}$ can be written as a product of two complex lines. We shall use the $\ensuremath\mathbb{C}_\text{seq}$ factor to sequester the different QFTs from one another, and will use the $\ensuremath\mathbb{C}_\text{base}$ factor to build a K3-fibered Calabi--Yau threefold $X_3 \to \ensuremath\mathbb{C}_\text{base}$ with fiber given by our K3 surface. Populating a different number of local sectors at each value of the moduli, we can again construct a probability distribution in a one-parameter subspace of the K3 moduli space. Let us also note that, much as in our discussion of the D3/D7 system, approaching a singular point in the moduli space of K3s limits the regime of validity of our approximation.
\subsection{$D = 1$}
Finally, let us turn to the case of one-dimensional quantum mechanical systems and their approximate $\AdS_2$ duals. Along these lines, we observe that type IIA strings on $\ensuremath\mathbb{R}_\text{time} \times \ensuremath\mathbb{R}^3 \times X$, with $X$ a Calabi--Yau threefold, produces a 4D $\ensuremath\mathcal{N} = 2$ supergravity background. Wrapping D0-, D2-, D4-, and D6-branes on holomorphic cycles of $X$ can, for suitable charges, result in a 4D black hole solution with near-horizon limit $\AdS_2 \times S^2$, with radius set by the background value of the charges (see, e.g., \cite{Simons:2004nm}). The brane configuration is shown below:
\vspace{2em}
\begin{center}
\begin{tabularx}{0.6\textwidth}{c|*{10}{>{\centering\arraybackslash $}X<{$}}}
& \tikzmark{rTimeL} 0 \tikzmark{rTimeR} & \tikzmark{r3L} 1 & 2 & 3 \tikzmark{r3R} & \tikzmark{xL} 4 & 5 & 6 & 7 & 8 & 9 \tikzmark{xR} \\ \hline
D0 & \times & & & & & & & & & \\
D2 & \times & & & & [ & \multicolumn{4}{c}{\text{2 of 6 directions}} & ] \\
D4 & \times & & & & [ & \multicolumn{4}{c}{\text{4 of 6 directions}} & ] \\
D6 & \times & & & & \times & \times & \times & \times & \times & \times
\end{tabularx}
%
\begin{tikzpicture}[overlay, remember picture]
\draw [decorate,decoration={brace,amplitude=10pt,raise=4pt}] (rTimeL.west) --node[above=14pt]{$\ensuremath\mathbb{R}_\text{time}$} (rTimeR.east);
\draw [decorate,decoration={brace,amplitude=10pt,raise=4pt}] (r3L.west) --node[above=14pt]{$\ensuremath\mathbb{R}^3$} (r3R.east);
\draw [decorate,decoration={brace,amplitude=10pt,raise=4pt}] (xL.west) --node[above=14pt]{$X$} (xR.east);
\end{tikzpicture}
\end{center}
Here, the bracketed directions denote the subset of directions containing the cycles wrapped by D2- and D4-branes, but do not specify the exact directions those cycles fill out. In the 4D theory, the D0- and D2-brane charges give rise to an electric charge vector $\V{Q}$, while the D4- and D6-brane charges give rise to a magnetic charge vector $\V{P}$. In general, it is a challenging problem to construct multi-center black hole solutions, but in the special case where all the centers have the \emph{same} charge $(\V{Q}_\ensuremath k, \V{P}_\ensuremath k) = (\V{Q}, \V{P})$, the configurations are mutually BPS, and in particular can be separated from one another in arbitrary directions of the $\ensuremath\mathbb{R}^{3}$ factor. For our purposes, we shall find it convenient to write $\ensuremath\mathbb{R}^3 = \ensuremath\mathbb{R}_\text{seq} \times \ensuremath\mathbb{C}_\text{base}$.
The worldvolume theory of this configuration is also challenging to describe, but at least at weak string coupling (away from the supergravity limit), there is a corresponding quiver quantum mechanics description available (see, e.g., \cite{Denef:2002ru}).
In the type IIA description, the superpotential couplings of this model are controlled by the complex structure moduli of $X$, and in the mirror type IIB description, they are captured by the K{\"a}hler moduli. An important open problem is to explain the sense in which this quiver quantum mechanics actually ``flows'' to a 1D SCFT.\footnote{Of course, in one dimension we face the issue that the CFT condition of a traceless stress tensor would appear to trivialize the theory altogether. So, we must already be prepared to work in terms of an approximate notion of conformal invariance where we introduce an explicit IR cutoff from the start, much as in \cite{Maldacena:2016hyu}.}
To generate an ensemble average in the 1D system, we return to the decomposition $\ensuremath\mathbb{R}^3 = \ensuremath\mathbb{R}_\text{seq} \times \ensuremath\mathbb{C}_\text{base}$. Since we can retain a supersymmetric multi-center configuration no matter how we separate the branes in the $\ensuremath\mathbb{R}^3$ direction, we use the $\ensuremath\mathbb{R}_\text{seq}$ factor to sequester the black holes, and treat $\ensuremath\mathbb{C}_\text{base}$ as the base of a non-compact Calabi--Yau fourfold $X_4 \to \ensuremath\mathbb{C}_\text{base}$, with fiber given by $X$ at a specified value of the moduli. Much as in our 2D holographic example, this provides us with an ensemble average over a one-parameter subspace of the moduli space.\footnote{The disorder averaged quiver quantum mechanics has been studied in \cite{Anninos:2016szt}, where the probability distribution is chosen to be Gaussian. Our construction in this section can be regarded as why/how the ensemble arise in the quiver quantum mechanics.}
An important subtlety with this construction is that as we vary the Calabi--Yau moduli, we may be forced to deal with a further fragmentation of a single center solution into a multi-center solution (the supergravity analog of wall-crossing), as studied for example in \cite{Denef:2001xn,Denef:2007vg}. For our present purposes, this introduces some additional complications into the $\CFT_1$ description and any putative $\AdS_2$ dual. One can view this as either imposing additional restrictions on what sorts of distributions we can engineer, or alternatively, as an opportunity to come to grips with multiple large $\AdS_2$ throats right from the start.\footnote{We thank N.~Bobev for comments.} As an additional comment, let us note that if we move away from the truncated subsector of operators, we get a large number of $\AdS_2 \times S^2$ throats, much as in the stringy baby universe construction of reference \cite{Dijkgraaf:2005bp}.
\subsection{$D>4$}
For SCFTs in $D = 5$ or $6$,\footnote{Note that $D = 6$ is the highest spacetime dimension where SCFTs exist \cite{Nahm:1977tg}.} it was demonstrated in \cite{Louis:2015mka,Cordova:2016xhm} that there is no supersymmetry-preserving marginal deformation. It would be interesting to address this issue in the context of the MM picture, since it implicitly imposes restrictions on the sort of boundary conditions for the bulk gravitational path integral.
\section{Discussion}\label{sec:DISC}
In this paper, we have presented a string-based method for engineering ensemble averaged QFTs. The main idea in our construction is to first build multiple copies of the same QFT, but in which the non-normalizable modes of the background geometry vary in the transverse directions. Truncating to the subsector of operators that are distributed over all the local sectors, we have shown that this can be used to build up an approximation to ensemble averaging with a UV completion. The construction can also be used to realize models with an approximate AdS dual. There is an intrinsic regime of validity that can be detected just by sampling macroscopic objects a large number of times. This is important because it cuts to the core issue of whether holography and effective field theory can actually be combined without reference to any putative UV completion. Our (admittedly contrived) construction serves as a counterexample to the claim that a UV completion is ``not necessary''.
In light of these considerations, it would be interesting to revisit the computation of a Page curve for a macroscopic black hole (see, e.g., \cite{Penington:2019npb, Almheiri:2019psf, Almheiri:2019hni}). For some recent critiques of these calculations, see, e.g., \cite{Geng:2020qvw,Geng:2021hlu,Raju:2021lwh}.
It is also natural to ask whether we could directly couple our ensemble of QFTs back to gravity. At least in the context of string theory, this is deeply problematic because in all known constructions, there are sharp upper bounds on the total number of extra sectors we can introduce. Moreover, once we recouple to gravity, all the moduli will once again become dynamical and must be stabilized in some way.
Aside from these intrinsic limitations of our method of construction, it is also interesting to ask about the extent to which we can generate more general classes of probability distributions. For example, in our examples based on fibering a Calabi--Yau manifold over its moduli space, we are automatically restricted in the sorts of distributions we can build. It would be helpful to pinpoint whether there is a deeper reason for such constraints.
\newpage
\section*{Acknowledgments}
We thank V.~Balasubramanian, N.~Benjamin, N.~Bobev, S.~Franco,
S.~Gukov, H.~Maxfield, J.~McNamara, T.~Pantev, D.~Pei, M.~Porrati,
A.~Tomasiello, and C.~Vafa for helpful discussions and correspondence.
We thank N.~Bobev, J.~Magan, and E.~Shaghoulian for helpful comments
on an earlier draft. JJH and XY thank the 2021 Simons
summer workshop at the Simons Center for Geometry and Physics for
kind hospitality during part of this work. XY thanks the UPenn theory group
for hospitality during this work. XY also thanks the Caltech Particle Theory Group
for hospitality during part of this work. The work of JJH and APT is supported by the DOE (HEP)
Award DE-SC0013528.
|
\section{Introduction}
The interaction of a single light mode with a single or multiple quantum emitters (QEs) is a fundamental model that is successfully applied to numerous systems.
Semiconductor quantum dots (QDs) in a microcavity are a prominent realization of such quantum light sources with a variety of different applications ranging from QD lasers \cite{chow2014emi,czerniuk2017pic} involving QDs in laterally extended cavities formed by Bragg mirrors down to single QE structures for applications in the field of quantum information technology \cite{michler2017qua}.
Color centers of different types placed in a microcavity, such as nitrogen-vacancy \cite{englund2010det,beha2012dia,fehler2019eff,janitz2020cav} or silicon-vacancy centers \cite{lee2012opt, fehler2020pur} in diamond, defect states in hexagonal boron nitride \cite{proscia2020mic,froch2020cou}, or organic molecules \cite{pino2018ten} constitute another class of QEs coupled to a single light mode, which are currently extensively studied.
Increasing possibilities to create deterministic photon emitter structures, e.g., by strain-patterning of $2$D semiconductors \cite{branny2017det,palacios2017lar,kern2016nan}, by lithographic positioning of nanodiamonds \cite{schrinner2020int} or by irradiation with a narrow helium beam \cite{klein2019sit} also pave the way for studying systems with well-defined finite numbers of photon emitters.
Various types of cavities have been realized to host QEs, such as photonic crystal cavities or micropillar cavities \cite{lodahl2015int}.
While structures with many emitters are often treated in terms of a semiclassical model, in which the emitters are described by quantum mechanical few-level systems interacting with a classical electromagnetic field \cite{gehrig2002mes,cartar2017sel,jurgens2020sem}, a single emitter in a cavity is usually described in terms of quantum optics by the Jaynes-Cummings (JC) model \cite{jaynes1963com,shore1993the,groll2020fou}.
Such a quantum optical treatment is necessary because of the strong correlations between the quantum states of the emitter and the photon field inside the cavity.
Between these limiting cases there is an interesting class of systems represented by a finite number of emitters $N$ in a cavity.
Based on the experience in the limiting cases, one would expect that for small $N$ a quantum optical treatment is necessary while for increasing $N$ the dynamical behavior approaches the one obtained from a semiclassical treatment.
In this paper we address this question by analyzing the optically induced dynamics of systems with $1$ up to $60$ QEs in a cavity on a quantum optical level and comparing the results with a semiclassical treatment.
We demonstrate that the spectra obtained by the quantum optical model approach the semiclassical result for increasing number of emitters, however for the numbers studied here they are still much broader.
Furthermore, we show that quantum effects play a crucial role for the photon states as revealed by a Wigner function analysis.
In a previous study we have performed semiclassical simulations of a planar ensemble of QEs, modeled as two-level systems (TLSs), in a one-dimensional photonic cavity driven by a short external laser pulse coupled in through one of the mirrors \cite{jurgens2020sem}.
We have shown that there exists a sharp transition in this semiclassical undamped model, associated with characteristic changes in the spectrum of the light field when varying the pulse amplitudes.
For small amplitudes exciton-polariton-like spectra occur, while Rabi oscillations emerge for large amplitudes.
We have furthermore shown that a sharp transition remains also for an ensemble of QEs with a Gaussian energy distribution.
In the semiclassical limit the dynamics depend on the density of QEs, represented by a continuous number, which is related to the fact that also the light field is a continuous classical variable.
In a quantum optical treatment the light field is described in terms of Fock states involving discrete numbers of photons, which can be absorbed or emitted by the QEs, and the number $N$ of QEs becomes a relevant quantity.
Because of the tensor product structure of the Hilbert space of composite quantum systems, the complexity increases exponentially with increasing $N$, which provides strong limitations on the analytical and numerical feasibility.
Therefore, additional restrictions of the model are necessary, which can be: (i) a limitation to very small numbers $N$ \cite{youssef2010som,albert2013mic,droenner2017col}, (ii) a limitation to small excitation intensities (i.e., small numbers of photons compared to the number of QEs) \cite{tsyplyatyev2009dyn,tsyplyatyev2010cla},or (iii) assuming ensembles consisting of $N$ QEs with identical \cite{tavis1968exa} or a small number of distinct \cite{dhar2018var} frequencies.
Here we will rely on the assumption of an ensemble with identical frequencies, which is further motivated by our finding in Ref.~\cite{jurgens2020sem} that the general features observed when moving from the low-excitation to the high-excitation regime are preserved also for a QE ensemble of non-vanishing width.
In this work we compare the spectra and the dynamics of the semiclassical and quantum optical solutions of $N$ QEs coupled to a single mode of the electric field after a pulsed excitation of the cavity.
The laser pulse creates a coherent field inside the cavity \cite{groll2020fou}, considered as an initial condition for the electric field, which then initiates the dynamics in the coupled QE-light system.
The paper is organized as follows.
In Sec.~\ref{sec.model} we introduce the model and derive the equations of motion, showing that the semiclassical limit is obtained as the mean-field approximation of the full equations of motion.
Then the exact solution of the full quantum system with $N$ QEs, i.e., the Tavis-Cummings (TC) model, is briefly reviewed and applied to the present initial conditions.
Section~\ref{sec.comparison} provides a detailed comparison of the spectral and temporal characteristics in the semiclassical and the quantum optical model for QE numbers between $N=1$ and $60$.
In Sec.~\ref{sec.wigner} the Wigner function is introduced and used to obtain information on the full quantum state of the light field during the initial dynamics.
Section~\ref{sec.collapse} extends this discussion to later times by analyzing the quantum state during the collapse and revival of the field amplitude. While up to this section we concentrate on a Hamiltonian model, in Sec.~\ref{sec.dissipation} we extend the model by including dissipation.
Finally, in Sec.~\ref{sec.conclusions} we briefly summarize our results.
\section{The Tavis-Cummings Model}\label{sec.model}
We consider an ensemble of $N$ QEs treated as TLSs with transition energies $\hbar \highindex{\omega_x}{k}$, coupled with strengths $\hbar \highindex{g}{k}$ to a single light field mode with energy $\hbar \omega_0$ and annihilation (creation) operator $\hat{a}$ $(\hat{a}^{\dagger})$.
This leads to the Hamiltonian of the TC model \cite{tavis1968exa,krimer2019cri,zens2019cri}
\begin{align}
\hat{H} &= \hbar \omega_0 \hat{a}^\dagger \hat{a} + \frac{\hbar}{2} \sum_{k=1}^N \highindex{\omega_x}{k} \highindex{\hat{\sigma}_3}{k} \nonumber \\
&\quad + \hbar \sum_{k=1}^N \highindex{g}{k} \left( \hat{a} \highindex{\hat{\sigma}_+}{k} + \hat{a}^\dagger \highindex{\hat{\sigma}_-}{k} \right), \label{eq:H_ens}
\end{align}
where $\highindex{\hat{\sigma}_\pm}{k}=\frac{1}{2}\left(\highindex{\hat{\sigma}_1}{k}\pm i \highindex{\hat{\sigma}_2}{k}\right)$, and $\{\highindex{\hat{\sigma}_j}{k}, j=1,2,3\}$ denote the Pauli matrices acting on QE $k$.
They are directly related to the inversion ($\highindex{\hat{\sigma}_3}{k}$) and to the polarization ($\highindex{\hat{\sigma}_\pm}{k}$) of the QEs.
In general, the transition frequencies and coupling strengths may be different for each QE \cite{zens2019cri,krimer2019cri,eastham2009qua,dhar2018var}.
Note that the Hamiltonian in \cref{eq:H_ens} involves the standard rotating wave approximation (RWA); the Hamiltonian for the QE-light coupling without RWA is called Dicke Hamiltonian \cite{dicke1954coh,bastarrachea2014com1,bastarrachea2014com2}.
Depending on the system parameters, the counter rotating terms can have an influence on the ground state of the system \cite{bastarrachea2014com1,bastarrachea2014com2} or its dynamics \cite{seke1995squ,seke1997eff,crescente2020ult}.
For the parameters studied here, however, the RWA can be assumed to be well justified, as is also confirmed by our calculations within the semiclassical model \cite{jurgens2020sem}.
\begin{figure}
\centering
\includegraphics[width=1\columnwidth]{fig1.pdf}
\caption{Sketch of the system. $N$ QEs are coupled to a single light mode inside a cavity. The cavity field is excited using a short external laser pulse.}
\label{fig1}
\end{figure}
The system is excited using a short laser pulse coupled in through one of the mirrors, such that the cavity mode is initialized in a coherent state \cite{groll2020fou}, as schematically shown in \cref{fig1}.
The generated light field then couples to the QEs and starts the internal dynamics.
As has been seen in the semiclassical case in Ref.~\cite{jurgens2020sem}, for a pulse which is shorter than the characteristic time scale of the dynamics, the excitation can be very well approximated by taking the excited cavity mode as an initial condition and disregarding the details of the injection process.
We will follow this scheme also here.
Therefore, the driving laser field is not included in the Hamiltonian given in \cref{eq:H_ens}.
This is different from studies investigating the case of a constant driving \cite{zens2019cri,krimer2019cri}.
The influence of initial Fock- and coherent states are discussed in Ref.~\cite{seke1989col}, where collapse and revival phenomena of different origin are found.
Other excitation schemes like chirped excitation of the QEs \cite{eastham2009qua} or different initial conditions \cite{keeling2009qua} have also been studied.
\subsection{Equations of motion and semiclassical limit}
To clearly see the relation between the quantum optical and the semiclassical model we start by deriving the equations of motion for the expectation values of the operators $\hat{a}$, $\highindex{\hat{\sigma}_3}{k}$ and $\highindex{\hat{\sigma}_-}{k}$, i.e., the coherent field amplitude, as well as the inversion and the polarization of QE $k$, respectively.
Using Ehrenfest's theorem in the rotating frame with $\hat{a} = \hat{\tilde{a}} e^{-i\omega_0 t}$, $\highindex{\hat{\sigma}_-}{k} = \highindex{\hat{\tilde{\sigma}}_-}{k} e^{-i\omega_0 t}$, the equations of motion read
\begin{subequations}
\begin{align}
\dv{t} \erw{\hat{\tilde{a}}} &= -i \sum_{k=1}^N \highindex{g}{k} \erw{\highindex{\hat{\tilde{\sigma}}_-}{k}} , \\
\dv{t} \erw{\highindex{\hat{\sigma}_3}{k}} &= 2 i \highindex{g}{k} \left[ \erw{\hat{\tilde{a}}^\dagger \highindex{\hat{\tilde{\sigma}}_-}{k}} - \erw{\hat{\tilde{a}} \highindex{\hat{\tilde{\sigma}}_+}{k}} \right] , \\
\dv{t} \erw{\highindex{\hat{\tilde{\sigma}}_-}{k}} &= -i \left(\highindex{\omega_x}{k}-\omega_0\right) \erw{\highindex{\hat{\tilde{\sigma}}_-}{k}} + i \highindex{g}{k} \erw{\hat{\tilde{a}} \highindex{\hat{\sigma}_3}{k}}.
\end{align}
\end{subequations}
Obviously, this set of equations of motion is not closed, instead on the right hand side it involves expectation values of products of a field and a QE operator.
These terms describe the influence of quantum correlations between light and matter on the dynamics.
By separating these quantities via $\erw{\hat{x} \hat{y}} = \erw{\hat{x}} \erw{\hat{y}} + \korr{\hat{x} \hat{y}}$ into mean field contributions and correlations, where $\hat{x}$ and $\hat{y}$ denote any two operators, and neglecting the correlations we arrive at
\begin{subequations}
\begin{align}
\dv{t} \erw{\hat{\tilde{a}}} &= -i \sum_{k=1}^N \highindex{g}{k} \erw{\highindex{\hat{\tilde{\sigma}}_-}{k}} , \\
\dv{t} \erw{\highindex{\hat{\sigma}_3}{k}} &= 2 i \highindex{g}{k} \left[ \erw{\hat{\tilde{a}}^\dagger} \erw{\highindex{\hat{\tilde{\sigma}}_-}{k}} - \erw{\hat{\tilde{a}}} \erw{\highindex{\hat{\tilde{\sigma}}_+}{k}} \right] , \\
\dv{t} \erw{\highindex{\hat{\tilde{\sigma}}_-}{k}} &= -i \left(\highindex{\omega_x}{k}-\omega_0\right) \erw{\highindex{\hat{\tilde{\sigma}}_-}{k}} + i \highindex{g}{k} \erw{\hat{\tilde{a}}} \erw{\highindex{\hat{\sigma}_3}{k}}.
\end{align}
\end{subequations}
The dynamics of $\erw{\hat{\tilde{a}}}$, $\erw{\highindex{\hat{\sigma}_3}{k}}$ and $\erw{\highindex{\hat{\tilde{\sigma}}_-}{k}}$ only depend on the exciton frequency of the QE $\highindex{\omega_x}{k}$ and its coupling constant $\highindex{g}{k}$.
Assuming that $\highindex{g}{k}$ only depends on its frequency $\highindex{\omega_x}{k}$, i.e., $\highindex{g}{k} = g\left(\highindex{\omega_x}{k}\right)$, it is possible to characterize the system using the continuous variable $\omega_x$ instead of the index $k$ of each QE assigning $\erw{\hat{\sigma}_3} (\omega_x) = \erw{\highindex{\hat{\sigma}_3}{k}}$ and $\erw{\hat{\sigma}_\pm} (\omega_x) = \erw{\highindex{\hat{\sigma}_\pm}{k}}$ for $\omega_x = \highindex{\omega_x}{k}$.
Now we rescale the dynamical variables and the time, according to
\begin{subequations} \label{eq:transformation}
\begin{align}
\alpha &= \frac{1}{\sqrt{N}} \erw{\hat{\tilde{a}}}, \\
\beta(\omega_x) &= \erw{\hat{\sigma}_3}(\omega_x), \\
\gamma(\omega_x) &= -i \erw{\hat{\tilde{\sigma}}_-} (\omega_x), \\
\tau &= G \sqrt{N} t, \label{eq:trafo_tau}
\end{align}
\label{eq:trafo}%
\end{subequations}
where $G = \frac{1}{N} \sum_{k=1}^N \highindex{g}{k}$ is the mean coupling strength.
The variable $\alpha$, which is the photon amplitude normalized to the square root of $N$, will be important throughout the paper.
$\tau$ denotes the scaled time.
This leads to the equations of motion
\begin{subequations}
\begin{align}
\dv{\tau} \alpha(\tau) &= \int \dd{\omega_x} \rho_{\rm QE}(\omega_x) \gamma(\omega_x,\tau) ,\\
\dv{\tau} \beta(\omega_x,\tau) &= -2\frac{g(\omega_x)}{G} \left[ \alpha^*(\tau) \gamma(\omega_x,\tau) + \alpha(\tau) \gamma^*(\omega_x,\tau) \right] ,\\
\dv{\tau} \gamma(\omega_x,\tau) &= -i \frac{\omega_x-\omega_0}{G\sqrt{N}} \gamma(\omega_x,\tau) + \frac{g(\omega_x)}{G} \alpha(\tau) \beta(\omega_x,\tau),
\end{align}
\label{eq:eom_semiclassical}%
\end{subequations}
where $\rho_{\rm QE} = \frac{1}{GN} \sum_{k=1}^N \highindex{g}{k} \delta(\omega_x - \highindex{\omega_x}{k})$ is the effective spectral QE distribution.
Comparing this result with the equations obtained from a semiclassical model for a QE ensemble in a Bragg mirror cavity [Eqs.~(34) in Ref.~\cite{jurgens2020sem}], we find that they are in complete agreement when setting $G = M \sqrt{\frac{u_c(z_0) v_c^*(z_0) \omega_0}{2\hbar A n(z_0)^2 \epsilon_0}}$, where $M$ is the mean value of the dipole moments, $A$ is a normalization area, $n(z_0)$ is the refractive index at the position of the QEs, $u_c(z_0)$ is the eigenmode of the wave equation at the position of the QE layer $z_0$ and $v_c(z_0)=-\frac{d^2}{dz^2}u_c(z_0)$ is the eigenmode of the adjoint wave equation.
Note that in Ref.~\cite{jurgens2020sem} furthermore a frequency-independent coupling matrix element $g(\omega_x)=G$ has been assumed.
Like in the semiclassical case studied in Ref.~\cite{jurgens2020sem}, here we are interested in the dynamics of the system after exciting the cavity with a short laser pulse.
In practice, we will assume that the pulse creates a coherent state inside the cavity, such that the system can initially be described by the coherent amplitude $\xi = \erw{a}(0) = \alpha_0\sqrt{N}$ of the field \cite{groll2020fou}, where $\left|\alpha_0\right|^2$ is the mean initial photon number per QE, with all QEs being in their respective ground state.
This corresponds to the initial conditions $\alpha(\tau=0)=\alpha_0$, $\beta(\omega_x,\tau=0)=-1$ and $\gamma(\omega_x,\tau=0)=0$.
Pulsed excitations, where instead of the cavity an ensemble of QEs is directly pumped, have been studied, e.g., in Ref.~\cite{eastham2009qua}.
The semiclassical (or mean field) equations shown in Eqs.~(\ref{eq:eom_semiclassical}) can be solved numerically for arbitrary ensemble shapes $\rho_{\rm QE}(\omega_x) $ using standard Runge-Kutta methods \cite{jurgens2020sem}.
In contrast to the semiclassical case, a complete quantum optical solution of the Hamiltonian $\hat{H}$ of \cref{eq:H_ens} for arbitrary spectral distributions and numbers $N$ of QEs is difficult because of the exponential increase of Hilbert space dimensionality with increasing $N$.
The quantum system with arbitrary frequency distributions has been investigated in Refs.~\cite{tsyplyatyev2009dyn,tsyplyatyev2010cla} where, however, only weak excitations, i.e., small numbers of photons compared to the number of QEs, have been studied.
Here we are interested in the comparison between semiclassical and quantum optical behavior from low excitations up to the Rabi oscillation regime, the latter being characterized by photon numbers larger than $N$.
Motivated by the findings in Ref.~\cite{jurgens2020sem} that the general features of the spectra are preserved also in the presence of spectral broadening, we restrict our study to the case of identical QEs with $\highindex{\omega_x}{k} = \omega_x$, $\highindex{g}{k} = g=G$ and $\omega_x = \omega_0$ being resonant to the cavity mode.
Then the model reduces to the standard TC model, for which a full solution is available.
The TC and Dicke models include a normal- to super-radiant phase-transition in the semiclassical limit ($N \rightarrow \infty$) \cite{bastarrachea2014com1,bastarrachea2014com2,castanos2009coh,emary2003qua,zou2013qua}.
This phase transition is hard to realize experimentally, therefore also coherently driven systems in the presence of damping have been studied \cite{krimer2019cri,zens2019cri,munoz2019sym} and phase transitions and hysteresis effects were found depending on the amplitude of the driving field \cite{krimer2019cri,zens2019cri,zou2013qua}.
It was shown that quantum effects are negligible outside a critical region of the coherent amplitude, when the number of two-level systems $N$ is large \cite{zens2019cri}.
Far from this critical order parameter the system can then be described by its semiclassical limit $N \rightarrow \infty$.
In contrast to these equilibrium phase transitions or phase transitions between stationary non-equilibrium states, the transition discussed in Ref.~\cite{jurgens2020sem} occurs between different types of time dependent states.
Beside these ground state and steady-state behaviors, Dicke and TC models also show interesting dynamical quantum features like squeezing \cite{retamal1997sqe,hassan1993per,genes2003spi,seke1995squ,seke1997eff,ramon1998col}, collapse and revival phenomena \cite{agarwal2012tav,ramon1998col,meunier2006ent,jarvis2009dyn} or superradiance and subradiance \cite{temnov2005sup,gegg2018sup}.
We will come back to some of these features below.
\subsection{Derivation of the quantized model}
For identical QEs the Hamiltonian of \cref{eq:H_ens} agrees with the standard TC model \cite{tavis1968exa} and can be written in an angular momentum representation
\begin{align}
\hat{H} = \hbar \omega_0 \hat{a}^\dagger \hat{a} + \hbar \omega_x \hat{J}_3 + \hbar g \left( \hat{a} \hat{J}_+ + \hat{a}^\dagger \hat{J}_- \right) \label{eq:H}
\end{align}
with total angular momentum operators
\begin{align}
\hat{J}_3 = \frac{1}{2} \sum_{k=1}^N \highindex{\hat{\sigma}_3}{k}, &&
\hat{J}_\pm = \sum_{k=1}^N \highindex{\hat{\sigma}_\pm}{k}. \nonumber
\end{align}
The total Hilbert space $\mathcal{H}$ is a tensor product of the Hilbert space of the photons $\mathcal{H}_\te{Phot}$ and the Hilbert space of the QEs $\mathcal{H}_\te{QE}$.
The Fock states with $n$ photons $\ket{n}$ constitute a basis of $\mathcal{H}_\te{Phot}$, whereas a basis of $\mathcal{H}_\te{QE}$ is given by the angular momentum states $\ket{j, m}$ characterized by the total angular momentum quantum number $j$ and the eigenvalue $m$ of the third component $\hat{J}_3$.
The Hamiltonian preserves the quantum number $j$ and, assuming all QEs to be in the ground state at $t=0$, the quantum number $j$ is given by half of the total number of QEs $j=N/2$.
The quantum number $m= -j$ then corresponds to the state with zero excitations and $m=j$ corresponds to the case of fully inverted QEs.
In the basis of the product states $\ket{n}\ket{j, m}$ the Hamiltonian is tridiagonal and the subspaces with constant $n + m + j$ decouple.
Here we classify the subspaces by the maximum number of photons $n$, i.e., the number of photons for the case that all QEs are in their respective ground state, implying $m=-j$.
These subspaces can be diagonalized either analytically \cite{tavis1968exa} or numerically leading to the eigenbasis $\ket{j, n; l}$ with
\begin{align}
\ket{j, n; l} &= \sum_{k=0}^{\min(N, n)} c_{k, l}^{(N,n)} \ket{n-k}\ket{j, -j+k} , \\
\hat{H}\ket{j, n; l} &= E_l^{(N,n)} \ket{j, n; l}.
\end{align}
In each subspace there are $\min(N, n)+1$ states and the state vector $\ket{\highindex{\Psi}{N}(t)} \in \mathcal{H}$ of the full system has the time dependence
\begin{align}
\ket{\highindex{\Psi}{N}(t)} = e^{-\frac{i}{\hbar}\hat{H}t} \ket{\highindex{\Psi}{N}(0)}. \label{eq:time_state}
\end{align}
As explained above (and confirmed in the semiclassical model \cite{jurgens2020sem}) we assume that the exciting pulse has no direct influence on the QEs inside the cavity but instantaneously creates a coherent photon state $\ket{\xi} = e^{-\frac{\abs{\xi}^2}{2}} \sum_n \frac{\xi^n}{\sqrt{n!}} \ket{n}$ at time $t=0$ \cite{groll2020fou}.
The coherent amplitude $\xi = \erw{a}(0)$ is the initial condition for the strength of the light field inside the cavity and $\abs{\xi}^2$ corresponds to the initial mean number of photons in the system.
The initial state of the full system is therefore given by the product of the state $\ket{j, -j}$ reflecting the fact that all QEs are in their respective ground state and a coherent state $\ket{\xi}$ for the light field, i.e., $\ket{\highindex{\Psi}{N}(0)} = \ket{\xi} \ket{j, -j}$.
The subsequent time evolution of the state is then given by
\begin{align}
\ket{\highindex{\Psi}{N}(t)} = e^{-\frac{\abs{\xi}^2}{2}} \sum_{n, l} \frac{\xi^n}{\sqrt{n!}} c_{0, l}^{*(N,n)} e^{-\frac{i}{\hbar}\highindex{E_l}{N,n}t}\ket{j, n; l}. \label{eq:time_evolution_state}
\end{align}
Using the transformation from \cref{eq:transformation}, the expectation value $\highindex{\alpha}{N}(t) = \erw{\hat{\alpha}} = \erw{\hat{\tilde{a}}} / \sqrt{N}$ in the rotating frame reads
\begin{align}
\highindex{\alpha}{N}(t) &= \frac{e^{-\abs{\xi}^2}}{\sqrt{N}} \sum_{n, l, l'}
\frac{\xi^n}{\sqrt{n!}} \frac{\xi^{*(n-1)}}{\sqrt{(n-1)!}}\ccoeff{c}{0, l}{N,n} \coeff{c}{0, l'}{N,n-1} \nonumber \\
& \times e^{-\frac{i}{\hbar} \left( E_l^{(N,n)} - E_{l'}^{(N,n-1)} - \hbar \omega_0 \right)t} B_{l,l'}^{(N, n)}, \label{eq:erw_x_N}
\end{align}
where
\begin{align}
B_{l,l'}^{(N, n)} &= \sum_{k=0}^{\min(N, n-1)} \ccoeff{c}{k, l'}{N,n-1} \coeff{c}{k, l}{N,n} \sqrt{n-k}. \nonumber
\end{align}
The semiclassical limit, calculated using \cref{eq:eom_semiclassical} with $\rho(\omega_x) = \delta(\omega_x)$, corresponds to the limit $N \rightarrow \infty$ (similar to Ref.~\cite{zens2019cri}) and will therefore be denoted by $\highindex{\alpha}{\infty}(t)$.
For small values $\xi \ll 1$ only the Fock states with $n = 0$ and $n = 1$ contribute to \cref{eq:erw_x_N} leading to
\begin{align}
\highindex{\alpha}{N}(t) &\approx \xi \frac{e^{-\abs{\xi}^2}}{\sqrt{N}} \sum_{l=0}^{1} \sum_{l'=0}^0 \ccoeff{c}{0,l}{N,1} \coeff{c}{0,l'}{N,0} \nonumber \\
&\times e^{-\frac{i}{\hbar} \left( E_l^{(N,1)} - E_{l'}^{(N,0)} - \hbar \omega_0 \right) t} B_{l,l'}^{(N, 1)}.
\end{align}
The subspaces with $n=0$ and $n=1$ photons can be diagonalized analytically similar to the JC model \cite{haroche2006exp}, leading to two frequency contributions at
\begin{align}
\frac{1}{\hbar}\left(E_l^{(N,1)} - E_{0}^{(N,0)} - \hbar \omega_0 \right) = (-1)^l g \sqrt{N} \label{eq:splitting}
\end{align}
for $l \in \{0, 1\}$ whose difference scales with $\sqrt{N}$.
These findings are in good agreement with Refs.~\cite{cummings1983exa,tsyplyatyev2009dyn,tsyplyatyev2010cla}, because in these cases the number of excited QEs is small compared to their total number.
It also fits very well to the semiclassical Dicke model \cite{bakemeier2013dyn} in the limit of weak couplings where the RWA holds.
The limit of small initial values $\highindex{\alpha}{N}(0) = \alpha_0 = \frac{\xi}{\sqrt{N}} \ll 1$ corresponds to the formation of an exciton-polariton.
The splitting depends on the number of QEs, identical to the semiclassical case \cite{jurgens2020sem}.
\section{Comparison of the spectra and the dynamics}\label{sec.comparison}
\subsection{Spectra}
In \cref{fig2} we show the spectrum of the light field, i.e., the Fourier transform of $\alpha(\tau)$, depending on the initial coherent amplitude $\alpha_0$ for different numbers $N$ of QEs.
The spectrum $\tilde{f}(\omega_\tau)$ of a function $f(\tau)$ is defined by
\begin{align}
\tilde{f}(\omega_\tau) = \Re \left(\int \dd{\tau} f(\tau) e^{-i\omega_\tau \tau - \epsilon \abs{\tau}}\right) \label{eq:fourier}
\end{align}
where $\omega_\tau$ is the frequency corresponding to the dimensionless time $\tau$ defined in \cref{eq:trafo_tau} and $\epsilon = 0.01\, \te{meV} / \hbar G \sqrt{N}$ has been introduced as a damping parameter to broaden the spectral distribution for a better visibility.
Let us start by briefly reviewing the result obtained in the semiclassical limit ($N = \infty$) for a $\delta$-shaped ensemble of QEs (see Ref.~\cite{jurgens2020sem}).
As shown in \cref{fig2} (a), we find exciton-polariton dynamics for initial values $\alpha_0 < 1$, which results in two spectral peaks at frequencies $\omega_\tau = \pm \omega_{\alpha_0}$ in the spectrum of the light field amplitude. For small initial values $\alpha_0 \ll 1$ the polariton frequency is given by $\omega_{\alpha_0}=1$, with increasing $\alpha_0$ it decreases. In this regime the occupation of the QEs never reaches one.
For initial values $\alpha_0 > 1$ three peaks at $\omega_\tau = 0, \pm \omega_{\alpha_0}$ are found. This reflects Rabi oscillations with the Rabi frequency approaching $\omega_{\alpha_0} = \sqrt{2(2\alpha_0^2 - 1)}$ for $\alpha_0 \gg 1$. The peak at the frequency zero reflects the fact that the field amplitude never vanishes in this regime.
At $\alpha_0 = 1$ a sharp transition between these regimes is observed.
We will call this initial value the transition point.
This transition can be explained by noting that \cref{eq:eom_semiclassical} can be transformed into the equation of motion for a classical particle moving in a double well potential which depends on the initial conditions of the dynamics \cite{jurgens2020sem}. Similar transitions to such three peak structures have also been found in the case of damped and driven systems \cite{munoz2019sym}.
\begin{figure}
\centering
\includegraphics[width=1\columnwidth]{fig2.pdf}
\caption{Spectrum of $\highindex{\alpha}{N}$ for different initial values $\alpha_0$. (a) Semiclassical limit determined via \cref{eq:eom_semiclassical} together with \cref{eq:fourier}. (b)-(d) Quantum optical treatment determined via \cref{eq:erw_x_N} together with \cref{eq:fourier} for the cases (b) $N=40$, (c) $N = 20$ and (d) $N=1$. The coupling strength is chosen, such that $\hbar G \sqrt{N} = const.$, independent of $N$.}
\label{fig2}
\end{figure}
Spectra obtained from the TC model for $N = 40, 20, 1$ are plotted in \cref{fig2} (b)-(d) again as functions of the initial value $\alpha_0$. For comparability we have set $\hbar G \sqrt{N} = const.$ in all calculations.
We apply this scaling of the coupling constant, such that the polariton splitting is independent of the number of QEs and can be easily compared to the semiclassical model.
For $N=1$ [\cref{fig2} (d)] we see the spectrum of the JC model.
Here the purely discrete nature of the JC ladder can be observed leading to energetically well separated transitions between the ladder rungs for the relatively low number of photons (note that $\alpha_0=2$ corresponds to an initial mean photon number of four).
At very small values $\alpha_0 \ll 1$ only the lowest rung of the JC ladder contributes to the dynamics leading to the same two polariton frequencies as in the semiclassical case, as discussed before.
For increasing $\alpha_0$, however, additional spectral lines appear.
In contrast to the semiclassical model shown in panel (a), the sharp transition at $\alpha_0 = 1$ cannot be observed and no continuous shift of the transition lines as a function of $\alpha_0$ is seen.
Also for initial values $\alpha_0 > 1$, where the semiclassical model reveals Rabi oscillations, the JC spectrum in this parameter range (\cref{fig2} (d)) looks very different from the semiclassical spectrum in (a).
If $\alpha_0$ is increased further, the JC ladder splittings of neighboring rungs approach each other and neighboring spectral lines get closer leading to essentially three lines reflecting the physics of the Mollow triplet \cite{mollow1969pow,valle2010mol} and thus approaching the same structure as in the semiclassical case.
This is shown in \cref{fig3}, where the spectrum obtained from the JC model is plotted for initial values up to $\alpha_0 = 5$ (corresponding to 25 photons on average).
\begin{figure}
\centering
\includegraphics[width=1\columnwidth]{fig3.pdf}
\caption{Spectrum of $\highindex{\alpha}{1}$ (i.e., $N=1$) for initial values between $\alpha_0 = 0$ and $\alpha_0 = 5$.}
\label{fig3}
\end{figure}
In Figs.~\ref{fig2} (b) and (c) the spectra are shown for QE numbers $N=40$ and $N=20$, respectively.
For small initial values $\alpha_0 \ll 1$ the exciton-polariton splitting with peaks at $\omega_\tau= \pm 1$ can be seen independent of the number of QEs.
For higher initial values the number of eigenstates contributing to the spectra at a given value $\alpha_0$ strongly increases with increasing $N$, resulting in a decrease of the energy separation of neighboring spectral peaks.
For $N=20$ [\cref{fig2} (c)] the general structure of the semiclassical spectrum can already be recognized but the discrete structure of the quantized system is still well resolved.
At $N=40$ [\cref{fig2} (b)] the broadened $\delta$-peaks overlap and the general structure of the semiclassical case, shown in panel (a), is now clearly visible.
Compared to the semiclassical model (a), the transition at $\alpha_0 = 1$ is not as sharp and the discrete structure of the individual rungs can still be seen around $\alpha_0 = 1$.
In the semiclassical model the transition point $\alpha_0 = 1$ corresponds to an infinitely slow evolution of the system into an unstable fixed point.
When this fixed point is reached all QEs are excited and there are no more photons left inside the cavity \cite{jurgens2020sem}.
Since quantum effects, in particular spontaneous emission, are not included in the semiclassical model, the system stays in its unstable fixed point.
In the region around $\alpha_0 = 1$ quantum effects play an important role, leading to the broadened transition seen in \cref{fig2}.
These findings reflect the strong influence of the QE-photon correlations at $\alpha_0 \approx 1$ similar to Ref.~\cite{zens2019cri}.
Although the sharp transition at $\alpha_0=1$ is not present in the quantum model, many features of the semiclassical limit can already be seen for relatively small $N\approx 40$.
In the following we will compare the dynamics obtained from the semiclassical and the quantum optical calculations to get a better understanding of the differences between these two approaches.
\subsection{Dynamics}
\begin{figure}
\centering
\includegraphics[width=1\columnwidth]{fig4.pdf}
\caption{Dynamics of $\highindex{\alpha}{N}$ for different initial conditions (a) $\alpha_0 = 0.1$, (b) $\alpha_0 = 0.5$, (c) $\alpha_0 = 1.1$ and (d) $\alpha_0 = 2$ and different numbers of QEs [$N=20$ (blue line), $N=40$ (red line), $N=60$ (green line), $N=\infty$ (grey line)]. The oscillation periods in the case $N=\infty$ are found to be $\tau_{0.1}\approx 6.28$, $\tau_{0.5}\approx 6.7$, $\tau_{1.1}\approx 4.1$, and $\tau_{2.0} \approx 1.68$.}
\label{fig4}
\end{figure}
The dynamics of $\highindex{\alpha}{N}$ are shown in \cref{fig4} for different initial values $\alpha_0$ and different $N$ as a function of the dimension-less time $\tau$ [\cref{eq:trafo_tau}].
Note that while in general $\highindex{\alpha}{N}$ and $\highindex{\alpha}{\infty}$ are complex variables, in the case of resonance ($\omega_x = \omega_0$) between the light field and the QE transitions as well as for real initial conditions, as considered here, they are real quantities.
This also explains the symmetry with respect to $\omega=0$ in all the spectra in \cref{fig2,fig3}.
In the semiclassical model (grey line) the coherent field amplitude exhibits undamped oscillations for all values of $\alpha_0 \ne 1$ with an oscillation period given by $\tau_{\alpha_0}=2\pi/\omega_{\alpha_0}$. For the lowest and highest values of $\alpha_0$ the oscillation periods essentially agree with the polaritonic limit $\tau_{0.1}\approx 2\pi$ and the Rabi frequency limit $\tau_{2.0} \approx 2\pi / \sqrt{2(2\alpha_0^2 - 1)} \approx 1.68$. For the intermediate values we find $\tau_{0.5}\approx 6.7$ and $\tau_{1.1}\approx 4.1$.
For small values [$\alpha_0=0.1$ in \cref{fig4} (a)] the lines of the quantum optical (colored lines) and semiclassical solutions cannot be distinguished, thereby showing an almost perfect agreement.
For larger initial values [$\alpha_0=0.5$ in \cref{fig4} (b)] $\highindex{\alpha}{N}$ also agrees with the semiclassical model for small $\tau$.
However, for times $\tau \gtrsim 10$ the deviations between the quantum calculations and the semiclassical model become more pronounced; furthermore they are stronger for smaller numbers of QEs, as can be seen by comparing the case of $N=20$ (blue line) with the cases $N=40$ (red line) and $N=60$ (green line).
Interestingly, we find that the frequencies of the oscillations fit very well, even if the amplitudes increasingly differ as a function of time, which explains the qualitatively good agreement of the spectra [\cref{fig2} (a)-(c)] in terms of the positions of the peaks.
At initial values close to $\alpha_0=1$, where in the semiclassical model the transition occurs [\cref{fig4} (c), $\alpha_0=1.1$], the dynamics differ strongly already after the first minimum of the coherent amplitude, because the correlations have a strong impact in this parameter regime, as has been shown in Ref.~\cite{zens2019cri} in the damped and driven system.
For still larger initial values [$\alpha_0=2$ in \cref{fig4} (d)], we see that the dynamics agree for the first few oscillations and then increasingly deviate.
Again, the deviations start earlier and are stronger for smaller QE numbers $N$.
The strong decay seen in \cref{fig4} (d) reflects the collapse part of the collapse and revival phenomenon in the JC and TC model \cite{ramon1998col,agarwal2012tav,meunier2006ent,jarvis2009dyn}.
Thus we can conclude from \cref{fig4} that for larger $N$ (green curves) the coherent amplitude agrees with the semiclassical model on longer time scales than for smaller $N$ (red and blue curves).
Apart from the dynamical transition around $\alpha_0 = 1$, the frequency in general matches very well, resulting in the good agreement of the position of the resonances within the spectra, even if the amplitudes differ strongly for larger times.
The large broadening of the peak structure seen in \cref{fig2} (b) and (c) is caused by the collapse of the coherent amplitude; indeed we find that the spectra are less broadened for higher numbers of QEs, where the collapse occurs at later times.
A larger collapse time therefore leads to sharper peaks in the spectrum and a better matching of the dynamics between the semiclassical and quantum optical approach.
\begin{figure}[t]
\centering
\includegraphics[width=0.9\columnwidth]{fig5.pdf}
\caption{Dynamics of the mean photon number per QE $\bar{n}_{\te{Phot}}^{(N)} = \erw{\hat{\alpha}^\dagger \hat{\alpha}}$ for (a) $\alpha_0 = 0.5$, (b) and (c) $\alpha_0 = 2$ for different numbers of QEs $N$. Panel (c) shows the same results as in (b) for $N=20$, $40$, $60$ on a longer time scale.}
\label{fig5}
\end{figure}
While in the semiclassical model, due to the coherence of the light field, the field intensity is strictly related to the field amplitude, this does not hold in the quantum optical model anymore.
In fact, the collapse and revival phenomenon in JC and TC models is associated with the loss and build-up of coherence of the light field, respectively.
To get a deeper understanding of these phenomena in our system we investigate the dynamics of the mean number of photons per QE
\begin{align}
\overline{n}_\te{Phot}^{(N)}(t) = \bra{\highindex{\Psi}{N}(t)}\hat{\alpha}^\dagger \hat{\alpha}\ket{\highindex{\Psi}{N}(t)}.
\end{align}
$\overline{n}_\te{Phot}^{(N)} (t)$ is calculated analog to \cref{eq:erw_x_N}.
The total number of photons in the system is then given by $N \cdot \overline{n}_\te{Phot}^{(N)} $.
Initially the light field is in a coherent state with $\overline{n}_\te{Phot}^{(N)}(0) = \abs{\alpha_0}^2$.
In the semiclassical model the system remains in a coherent state, such that at any time the relation $\overline{n}_\te{Phot}^{(\infty)}(t) = \abs{\highindex{\alpha}{\infty}(t)}^2$ holds.
The dynamics of the mean photon number per QE for different $\alpha_0$ and $N$ is shown in \cref{fig5}, where the line colors have the same meaning as in \cref{fig4}.
For small initial values ($\alpha_0 = 0.5$) the temporal evolution of the mean photon number is shown in \cref{fig5} (a).
In this case there are initially on average $0.25$ photons per QE in the system, i.e., there are less photons than QEs.
Every QE can thus absorb and re-emit up to $0.25$ photon on average.
In the semiclassical model this happens periodically and $\overline{n}_\te{Phot}^{(\infty)}$ continues to oscillate between $0.25$ and zero.
In the quantum optical model the oscillations decay and, as the coherent amplitude in \cref{fig4} (b) relaxes to zero, the mean photon number relaxes to approximately half the initial value.
The dynamics of the mean photon number above the transition point (here $\alpha_0 = 2$) are shown in \cref{fig5} (b) and the respective long time behavior is shown in \cref{fig5} (c).
Now, initially the mean photon number per QE is four.
Each QE absorbs and re-emits up to one photon leading to the oscillations of $\overline{n}_\te{Phot}$ between $4$ and $3$.
This corresponds to modulations between $4N$ and $3N$ photons in total.
Again, the oscillations persist in the semiclassical model while in the quantum optical case at longer times a quasi-stationary value of $\overline{n}_\te{Phot}^{(N)} = 3.5$ is reached, indicating that every QE has absorbed on average half a photon.
In the long time behavior [see \cref{fig5} (c)] besides the just described collapse also the corresponding revival can be seen \cite{agarwal2012tav,jarvis2009dyn} (note that for $N=60$ the revival is beyond the plotted times).
With the given initial conditions the revival time increases with the number of QEs, as seen in \cref{fig5} (c).
We will come back to the phenomenon of collapse and revival in Sec.~\ref{sec.collapse}.
\section{Wigner functions}\label{sec.wigner}
The different time evolutions of the squared coherent field amplitude and of the mean photon number are clear indications that the quantum state of the light field strongly deviates from a coherent state.
In the following we will analyze these deviations in more detail.
The complete quantum state of a single mode light field can be best visualized in a phase space representation by calculating the corresponding Husimi $Q$ function or the Wigner function.
We focus here on the Wigner function due to the easy identification of non-classical features in terms of negativities of this function.
The Wigner function of a coherent state is an isotropic two-dimensional Gaussian.
Deviations from the coherent state are thus directly reflected in deformations of the Wigner function from this shape.
The Wigner function $\wigner{W}{N}{\lambda, t}$ is a function of the complex phase space variable $\lambda = U + i\Pi$ and defined by \cite{haroche2006exp}
\begin{align}
\wigner{W}{N}{\lambda, t} = \frac{2}{\pi} \Tr[ D(-\lambda) \hat{\rho}^{(N)}(t) D(\lambda) \mathcal{P} ]
\end{align}
with the density operator $\hat{\rho}^{(N)}(t) = \kb{\Psi^{(N)}(t)}{\Psi^{(N)}(t)}$, the displacement operator $D(\lambda) = \exp(\lambda \hat{a}^\dagger - \lambda^* \hat{a})$, the parity operator $\mathcal{P} = \exp(i\pi \hat{a}^\dagger \hat{a})$, and the trace $\Tr[\dots]$.
The superscript $(N)$ denotes the number of QEs in the system.
Using the eigenstates $\ket{j, n; l}$ and the state $\ket{\Psi^{(N)}(t)}$ defined in \cref{eq:time_evolution_state}, the Wigner function can be written as
\begin{align}
\wigner{W}{N}{\lambda, t} = \sum_{\substack{n, n' \\l, l'}} \rho^{(N,n, n')}_{l, l'}(t) \, w^{(N,n, n')}_{l, l'}(\lambda) \nonumber
\end{align}
with
\begin{align}
w^{(N,n, n')}_{l, l'}(\lambda) &= \frac{2}{\pi} \Tr[D(-\lambda) \kb{j, n; l}{j, n'; l'} D(\lambda) \mathcal{P}] , \nonumber \\
\rho^{(N, n, n')}_{l, l'}(t) &= \exp(-\abs{\xi}^2) \frac{\xi^n}{\sqrt{n!}} \frac{\xi^{*n'}}{\sqrt{n'!}} \ccoeff{c}{0, l}{N,n} \coeff{c}{0, l'}{N, n'} \nonumber \\
&\times \exp[-\frac{i}{\hbar} \left( E_l^{(N,n)} - E_{l'}^{(N,n')} \right) t ]. \nonumber
\end{align}
As is shown in \cref{app:wigner}, by using the properties of the Fock states, a closed expression for the Wigner function can be obtained.
To relate the Wigner function to our results, we express it in the transformed frame defined by \cref{eq:transformation}.
By introducing the coordinates in the transformed frame $U_\alpha = U/\sqrt{N}$ and $\Pi_\alpha = \Pi/\sqrt{N}$ and the transformed Wigner function
\begin{align}
&\wigner{\widetilde{W}}{N}{U_\alpha, \Pi_\alpha, \tau} \nonumber\\
&\quad = N \, \wigner{W}{N}{\sqrt{N} (U_\alpha + i\Pi_\alpha) e^{-i\omega_0 \frac{\tau}{g\sqrt{N}}}, \frac{\tau}{g\sqrt{N}}},
\end{align}
the field amplitude is obtained by
\begin{align}
\highindex{\alpha}{N} &= \erw{\hat{\alpha}}^{(N)} \\
&=\int \dd{U_\alpha} \dd{\Pi_\alpha} \left( U_\alpha + i\Pi_\alpha \right) \wigner{\widetilde{W}}{N}{U_\alpha, \Pi_\alpha, \tau}, \label{eq:erw_wigner} \nonumber
\end{align}
with $\hat{\alpha} = \hat{\tilde{a}} / \sqrt{N}$.
The derivation can be found in \cref{app:wigner_transformation}.
The quadrature operators in the transformed frame are given by
\begin{align}
\hat{u}_\alpha = \frac{1}{2} \left( \hat{\alpha} + \hat{\alpha}^\dagger \right), \hspace{1cm}
\hat{\pi}_\alpha = \frac{1}{2i} \left( \hat{\alpha} - \hat{\alpha}^\dagger \right). \nonumber
\end{align}
As already mentioned above, in all the cases studied here $\highindex{\alpha}{N}$ is a real quantity leading to
\begin{subequations}
\begin{align}
\erw{\hat{u}_\alpha}^{(N)} &= \highindex{\alpha}{N}, \\
\erw{\hat{\pi}_\alpha}^{(N)} &= 0.
\end{align}
\end{subequations}
This reflects the fact that the transformed Wigner function is here always symmetric in $\Pi_\alpha$.
The variances of the Wigner function are given by
\begin{align}
&\left(\Delta u_{\alpha}^{(N)}\right)^2 = \erw{\hat{u}_\alpha^2}^{(N)} - \left(\erw{\hat{u}_\alpha}^{(N)}\right)^2, \nonumber \\
&\left(\Delta \pi_{\alpha}^{(N)}\right)^2 = \erw{\hat{\pi}_\alpha^2}^{(N)} - \left(\erw{\hat{\pi}_\alpha}^{(N)}\right)^2, \nonumber
\end{align}
where $\Delta u_{\alpha}^{(N)}$ and $\Delta \pi_{\alpha}^{(N)}$ are the widths in the $U_\alpha$ and $\Pi_\alpha$ direction, respectively.
At $\tau = 0$ the Wigner function is given by a Gaussian distribution shifted in $U_\alpha$ direction by the amount $\alpha_0$ and with widths $\Delta u_{\alpha, \te{coh.}}^{(N)} = \Delta \pi_{\alpha, \te{coh.}}^{(N)} = (4N)^{-1/2}$ in the scaled frame.
Deviations from this value indicate nonclassical states, because they are directly related to the incoherent part of the photon occupation $\erw{\alpha^\dagger \alpha}^{(N)}_{\mathrm{incoh.}}$ according to
\begin{subequations}\label{eq:fluctuations}
\begin{align}
\left(\Delta_u^{(N)}\right)^2 &= \left( \Delta u_\alpha^{(N)} \right)^2 - \left( \Delta u_{\alpha, \te{coh}}^{(N)} \right)^2, \\
\left(\Delta_\pi^{(N)}\right)^2 &= \left( \Delta \pi_\alpha^{(N)} \right)^2 - \left( \Delta \pi_{\alpha, \te{coh}}^{(N)} \right)^2, \\
\erw{\hat{\alpha}^\dagger \hat{\alpha}}_{\mathrm{incoh.}}^{(N)} &= \erw{\hat{\alpha}^\dagger \hat{\alpha}}^{(N)} - \erw{\hat{\alpha}^\dagger}^{(N)} \erw{\hat{\alpha}}^{(N)} \nonumber \\
&= \left( \Delta_u^{(N)} \right)^2 + \left( \Delta_\pi^{(N)} \right)^2\, .
\end{align}
\end{subequations}
\noindent
We will refer to $\Delta_u^{(N)}$ and $\Delta_\pi^{(N)}$ as excess fluctuations.
We now discuss the Wigner function $\wigner{\widetilde{W}}{N}{U_\alpha, \Pi_\alpha, \tau}$ at the extrema and turning points of $\highindex{\alpha}{N}(\tau)$ during the first period.
The Wigner function at these times is shown for $N=20$ QEs in \cref{fig6} for the three initial values $\alpha_0 = 0.1$ (left column), $\alpha_0 = 0.5$ (central column), and $\alpha_0 = 2.0$ (right column).
The times, increasing from top to bottom, are given in units of the intrinsic oscillation period $\tau_{\alpha_0}$, which for the present cases have been found to be $\tau_{0.1}\approx 6.28$, $\tau_{0.5}\approx 6.7$, and $\tau_{2.0} \approx 1.68$ (see \cref{fig4}).
Note that the color scale is chosen such that saturation is reached for smaller values in the negative than positive direction to easier identify negativities in the Wigner function, which are clear signatures of quantum behavior.
\begin{figure}
\centering
\includegraphics[width=1\columnwidth]{{fig6.pdf}}
\caption{Wigner functions $\wigner{\widetilde{W}}{N}{U_\alpha, \Pi_\alpha, t}/N$ for initial values $\alpha_0 = 0.1$ (left column), $\alpha_0 = 0.5$ (central column) and $\alpha_0 = 2$ (right column) for $N=20$ QEs at different times during the first oscillation period $0 \leq \tau \leq \tau_{\alpha_0}$ (increasing from top to bottom). The oscillation periods are $\tau_{0.1}\approx 6.28$, $\tau_{0.5}\approx 6.7$, and $\tau_{2.0} \approx 1.68$ (see \cref{fig4}). Note that the axes of the right column are scaled differently.}
\label{fig6}
\end{figure}
For initial values $\alpha_0 \leq 1$ (first two columns) the Wigner function oscillates between $+\alpha_0$ and $-\alpha_0$ on the $\Pi_\alpha = 0$ line in $U_\alpha$-direction.
Note that the untransformed Wigner function $\wigner{W}{N}{\lambda, t}$ rotates with the frequency $\omega_0$ around the origin, while the transformed function $\wigner{\widetilde{W}}{N}{U_\alpha, \Pi_\alpha, \tau}$ is defined in this rotating frame and therefore oscillates along the $U_\alpha$ axis.
For $\alpha_0 = 0.1$ (left column) the Wigner function keeps its Gaussian shape during the first period and it essentially returns to the initial state after a complete oscillation.
This is in agreement with the previous finding that the system behaves similar to the semiclassical one.
For $\alpha_0 = 0.5$ (central column) the shape of the Wigner function changes and after a complete oscillation ($\tau = \tau_{0.5}$, bottom panel) we already see clear deviations from the initial state associated with very small negative contributions.
In addition to the negative parts, the distribution broadens mainly in the $\Pi_\alpha$ direction, while the width in the $U_\alpha$ direction stays approximately the same as in the initial state.
Using \cref{eq:erw_wigner}, one can see that the broadening in the $\Pi_\alpha$ direction has no impact on the real part of the expectation value.
Therefore, the mean value $\erw{u_\alpha}^{(N)}(\tau)$ does not deviate strongly from the coherent case at the same time, which results in the good agreement with the semiclassical model, as seen in \cref{fig4}.
Despite this similarity of the mean value with the coherent amplitude, we see clear changes in the quantum state.
For $\alpha_0 = 2$ (right column), i.e., above the transition point, the dynamics during the first period are restricted to positive $U_\alpha$, reflecting the fact that now $\highindex{\alpha}{N}$ remains positive (see \cref{fig4}).
Again, after the first oscillation the Wigner function is stretched in the $\Pi_\alpha$ direction compared to the initial state.
Similar to the case with $\alpha_0 = 0.5$, the distribution in the $U_\alpha$ direction does not differ significantly, leading to the good agreement of the dynamics of the mean value $\erw{\hat{u}_\alpha}^{(N)}(\tau)$ during the first oscillation period, as seen in \cref{fig4}.
\begin{figure}
\centering
\includegraphics[width=1\columnwidth]{{fig7.pdf}}
\caption{Wigner functions $\wigner{\widetilde{W}}{N}{U_\alpha, \Pi_\alpha, t} / N$ for initial values $\alpha_0 = 0.1$ (left column), $\alpha_0 = 0.5$ (central column) and $\alpha_0 = 2$ (right column) for $N=20$ QEs at different times during the second oscillation period $\tau_{\alpha_0} \leq \tau \leq 2\tau_{\alpha_0}$ (increasing from top to bottom). The oscillation periods are $\tau_{0.1}\approx 6.28$, $\tau_{0.5}\approx 6.7$, and $\tau_{2.0} \approx 1.68$ (see \cref{fig4}). Note that the axes of the right column are scaled differently.}
\label{fig7}
\end{figure}
The Wigner functions during the second oscillation period of \cref{fig4} are shown in \cref{fig7} for the same initial values.
Note that we changed the color bar scaling for $\alpha_0 = 2$ (right column) for better visibility.
For small initial values (left column) the same behavior as before continues.
This underlines the interpretation, that the semiclassical model is a good approximation for small initial values of the light field.
In this case the quantum state of the light remains to a large degree in a coherent state.
For larger amplitudes, but still below the transition ($\alpha_0 = 0.5$, central column), the deformations of the Wigner function further increase compared to the first period (central column, \cref{fig6}).
At the times $\tau=1.25\tau_{0.5}$ and $\tau=1.75\tau_{0.5}$ we now also observe a stretching along the $U_\alpha$ direction while for $\tau=1.5\tau_{0.5}$ and $\tau=2\tau_{0.5}$ the stretching occurs mainly in the $\Pi_\alpha$ direction.
At the latter times we also observe the development of a ringlike structure, similar to negativities appearing in the Wigner function of higher Fock states \cite{schleich2011qua}.
Here, however, the rings are not centered around the origin.
These times ($\tau = 1.5\tau_{0.5}$ and $\tau = 2\tau_{0.5}$) correspond to local extrema in the $\highindex{\alpha}{N}(t)$ dynamics, where the strongest deviations between quantum and semiclassical treatment have been observed in \cref{fig4}.
The same trend can be seen for $\alpha_0 = 2$ (right column), where the Wigner function exhibits a significant spread in the $\Pi_\alpha$ direction and negative parts at certain times.
Interestingly, here the negative parts are most prominent at times $\tau = 1.25\tau_{2.0}$ and $\tau=1.5\tau_{2.0}$, where $\highindex{\alpha}{N}$ is in good agreement with the semiclassical model (see \cref{fig4}).
\begin{figure}
\centering
\includegraphics[width=1\columnwidth]{{fig8.pdf}}
\caption{Excess fluctuations $\Delta_u^{(N)}$ (solid) and $\Delta_\pi^{(N)}$ (dashed) of the Wigner function for $N=20$ (blue lines), $N=60$ (red lines) and $N=\infty$ (black lines).}
\label{fig8}
\end{figure}
The color plots of the Wigner functions in \cref{fig6,fig7} provide a good overview of the qualitative changes of the quantum state of the light field compared to the coherent state in the semiclassical model.
However, it is often not so easy to extract quantitative information directly from these plots.
Therefore, in \cref{fig8} we investigate the temporal evolution of the excess fluctuations $\Delta_u^{(N)}$ (solid lines) in the $U_\alpha$ direction and $\Delta_\pi^{(N)}$ (dashed lines) in the $\Pi_\alpha$ direction, as defined in \cref{eq:fluctuations} (a) and (b) for $N=20$ (blue lines) and $N=60$ (red lines).
In the semiclassical limit the light field is in a coherent state, leading to $\Delta_u^{(\infty)} = \Delta_\pi^{(\infty)} = 0$ (black solid lines).
For finite numbers $N$ and for a small initial condition $\alpha_0 = 0.1$ [\cref{fig8} (a)] the fluctuations oscillate with a very small amplitude around the semiclassical value, indicating that already at this initial condition there are some deviations from the coherent state.
However, they are very small and therefore not visible in the color plots in \cref{fig6,fig7}.
For larger initial conditions as seen in \cref{fig8} (b) and (c), the excess fluctuations keep oscillating with a significant amplitude and additionally experience a steady increase.
While in the case of $\alpha_0 = 0.5$ [\cref{fig8} (b)] the increase is similar for $\Delta_u^{(N)}$ and $\Delta_\pi^{(N)}$, for $\alpha_0 = 2$ [\cref{fig8} (c)] $\Delta_\pi^{(N)}$ grows much faster than $\Delta_u^{(N)}$, consistent with the finding of a large broadening in $\Pi_\alpha$ direction in \cref{fig7}.
We also see that the excess fluctuations during the first oscillation periods drop slightly below zero indicating squeezing \cite{retamal1997sqe,hassan1993per}.
While for $\alpha_0=0.1$ and $\alpha_0 = 0.5$ squeezing occurs both in the $U_\alpha$ and the $\Pi_\alpha$ direction, for $\alpha_0 = 2$ the increase of $\Delta_\pi^{(N)}$ is so strong that only $\Delta_u^{(N)}$ exhibits squeezing.
The formation of squeezed states on small time scales in the TC model has also been investigated for different initial states in Refs.~\cite{retamal1997sqe,seke1995squ,seke1997eff}.
Squeezing has also been observed in the resonance fluorescence of a single QD \cite{schulte2015qua}.
When comparing the fluctuations for different numbers of QEs, we find that the deviations from the value of the coherent state decrease with increasing $N$.
This is in line with the findings that the correlations vanish for large $N$ \cite{zens2019cri}, however it also shows that the values up to $N=60$ studied here are still far from this limiting case.
\section{Collapse and revival}\label{sec.collapse}
In the results discussed until now we have mainly focused on the initial part of the dynamics, in which the coherent amplitude decays and the oscillating mean photon number relaxes to a time-independent value.
As is well known for the JC and TC models, and as has already been seen in \cref{fig5}(c), this collapse of the oscillatory dynamics is accompanied by a revival of the coherent amplitude and of the oscillations in the mean photon number at a revival time $\tau_R$.
For a fixed and sufficiently large initial value $\alpha_0 = \xi / \sqrt{N}$ and with $\hbar G \sqrt{N}= \te{const.}$ the revival time $\tau_{R,\te{analytical}}=2\pi \alpha_0 N$ (compare Ref.~\cite{meunier2006ent}) increases linearly with increasing $N$.
Note that by choosing $\xi = \te{const.}$ and $g=\te{const.}$ the revival time would be independent of the number of QEs.
For the JC model it has been shown that in the limit of large mean photon numbers the field and the QE states factorize at half-integer and integer multiples of the revival time \cite{gea-banacloche1990col,gea-banacloche1991ato}.
While at integer multiples the field is again in a coherent state, for half-integer multiples it is in a superposition of two coherent states, i.e., in a Schr\"odinger cat state.
It has been shown that in the case of a single QD in a cavity this can be used for the preparation of a cat state \cite{cosacchi2021sch}.
These results have been generalized to the TC model \cite{meunier2006ent,jarvis2009dyn}.
There, however, a complete factorization occurs only for a specific class of initial states, which does not comprise the initial state in our case, consisting of all QEs in their respective ground state.
Nevertheless, as will be seen below, the field exhibits interesting quantum states at specific times.
\begin{figure}
\centering
\includegraphics[width=1\columnwidth]{{fig9.pdf}}
\caption{Wigner functions $\wigner{\widetilde{W}}{N}{U_\alpha, \Pi_\alpha, t} / N$ for the initial value $\alpha_0 = 2$ for $N=20$ (left column) and $N=21$ (right column) at different times given in units of the revival time $\tau_R$.}
\label{fig9}
\end{figure}
In \cref{fig9} we have plotted the Wigner function for $N=20$ (left column) and $N=21$ (right column) QEs and for the initial condition $\alpha_0 = 2$ at four different times: $\tau=\tau_R/10$, $\tau=\tau_R/3$, $\tau=\tau_R/2$, and $\tau=\tau_R$ from top to bottom, where $\tau_R$ is the revival time extracted from our numerical calculations.
For a better interpretation of these results \cref{fig10} shows the temporal evolution of the coherent amplitude over a time scale including the first two revivals [similar to \cref{fig4}(d)].
At all times the overall structure of the Wigner function does not resemble a single Gaussian distribution anymore.
Instead, typically multiple localized structures (red areas) are seen with large interference-like patterns (red and blue stripes) in between.
Already at $\tau_R/10$ (upper panel in \cref{fig9}) the light field shows pronounced quantum features indicated by large negative contributions of the Wigner function.
When increasing the time from $\tau_R/10$ over $\tau_R/3$ to $\tau_R/2$ one observes that the number of localized (red) features reduces.
At $\tau_R/M$ there are exacly $M$ of those features visible until at $\tau_R$ (lower panel) only one single feature remains which is concentrated in a rather small region of the phase space.
There is however one striking difference between the case $N=20$ and $N=21$ despite the small change in the QE number.
At $\tau = \tau_R$ the Wigner function is strongly concentrated around $(U_\alpha, \Pi_\alpha) = (2,0)$ for $N=20$ and $(U_\alpha, \Pi_\alpha) = (-2,0)$ for $N=21$.
The resulting coherent amplitude in the frame rotating with the frequency of the light mode occurs with the same sign as the initial amplitude in the case of even $N$, while for odd $N$ it appears with opposite sign (see also \cref{fig10} and Ref.~\cite{jarvis2009dyn}).
This difference can be traced back to the fact that in the angular momentum representation [\cref{eq:H}] the TC model for even $N$ refers to a bosonic system while for odd $N$ it describes a fermionic system.
\begin{figure}
\centering
\includegraphics[width=1\columnwidth]{{fig10.pdf}}
\caption{Dynamics of $\highindex{\alpha}{N}$ for $\alpha_0 = 2$ and $N=20$ (blue) and $N=21$ (red) on a time scale including the first and second revival.}
\label{fig10}
\end{figure}
To understand the shape of the Wigner functions for different $\tau$ and $N$ it is instructive to refer to the limiting case of high mean photon numbers.
In this regime it has been shown, first for the JC model \cite{gea-banacloche1990col,gea-banacloche1991ato} and later for the TC model \cite{meunier2006ent,jarvis2009dyn}, that the state of the complete system can be written in the form
\begin{align}
\ket{\highindex{\Psi}{N}(t)}= \sum_{k=-N/2}^{N/2} \beta_k(t) \ket{D_k(t)} \otimes \ket{\Phi_k(t)}\ ,
\end{align}
where $\beta_k(t)$ are time-dependent coefficients, $\{\ket{D_k(t)}\}$ is a set of $N+1$ orthogonal states of the $N$ QEs defined, e.g., in Ref.~\cite{meunier2006ent}, and $\ket{\Phi_k(t)}$ are the so-called Gea-Banacloche states \cite{meunier2006ent,jarvis2009dyn}
\begin{align}\label{eq:gea}
\ket{\Phi_k(t)}=\ket{\xi e^{i2\pi k t/t_R}}\ ,
\end{align}
which are $N+1$ coherent states with mean photon number $|\xi|^2$.
In our Wigner functions defined in the scaled phase space $(U_\alpha, \Pi_\alpha)$ they correspond to $N+1$ Gaussians rotating with different frequencies on a circle with radius $\alpha_0$, located at time $\tau$ at $(U_\alpha, \Pi_\alpha)= (\alpha_0 \cos(2\pi k \tau/\tau_R), \alpha_0 \sin(2\pi k \tau/\tau_R))$.
At $\tau = 0$ all Gaussians overlap at $(U_\alpha, \Pi_\alpha) = (\alpha_0, 0)$. At the beginning of the dynamics the states with $k>0$ move in the positive $\Pi_\alpha$ direction, those with $k<0$ move in the negative $\Pi_\alpha$ direction.
States with different $|k|$ move with different velocities leading to the smearing out of the initial coherent state, as seen in \cref{fig6,fig7} (right column) and \cref{fig8} (c), where the width $\Delta_\pi^{(N)}$ increases much faster than $\Delta_u^{(N)}$.
As discussed in Ref.~\cite{meunier2006ent} this separation of the Gaussian distributions is the reason for the collapse of the Rabi oscillations.
As seen in \cref{fig4} (d) the collapse also leads to the deviation between the quantum and semiclassical model.
From Eq.~\eqref{eq:gea} we deduce that at the revival time $\tau=\tau_R$ all Gaussians overlap again, for even numbers $N$ (i.e., in the bosonic case) at $(U_\alpha, \Pi_\alpha) = (\alpha_0, 0)$ and for odd $N$ (i.e., in the fermionic case) at $(U_\alpha, \Pi_\alpha) = (-\alpha_0, 0)$ (see also Ref.~\cite{jarvis2009dyn}).
This is indeed what we see in the bottom panels of \cref{fig9}, except for the fact that the Wigner function, although very localized compared to the other depicted times, is significantly smeared out compared to a coherent state.
This is related to the fact that in our case with on average four photons per QE, we are still far from the limit of large photon numbers.
From Eq.~\eqref{eq:gea} we also find that at the times $\tau=\tau_R/M$ with $1 \leq M \leq N$ some of the Gaussians overlap, such that we observe $M$ isolated localizations of the Wigner function located at equal distances along the circle.
This is clearly seen in \cref{fig9} for the values $M=10, 3, 2, 1$ (from top to bottom).
These partial overlaps lead to small revivals of the Rabi oscillations at intermediate times \cite{meunier2006ent,jarvis2009dyn}, which are not visible in the photon amplitude $\alpha^{(N)}$ [\cref{fig4}], but very weakly in the inversion and the photon number [see \cref{fig5} (c) at $\tau \approx 120$ (blue curve), $\tau \approx 240$ (red curve) and $\tau \approx 360$ (green curve)].
Interestingly, between the maxima of the Wigner function there are pronounced negative contributions, which is a clear indication of non-classical behavior.
In particular, for $M=2$ the Wigner function has the qualitative shape of a Schr\"odinger cat state consisting of two maxima and an interference pattern in between.
This holds despite the fact that according to the general theory we do not expect a factorization into QE and field states \cite{jarvis2009dyn}, except for the case $N=1$, i.e., the JC model \cite{gea-banacloche1990col,gea-banacloche1991ato}.
Also for higher values of $M$, here shown for $M=3$ and $M=10$, we observe increasingly complex interference patterns between the maxima indicating coherences between the corresponding quasi-coherent states.
\section{Impact of dissipation}\label{sec.dissipation}
\noindent
So far we only discussed the pure Hamiltonian dynamics of the system.
In real systems the interaction of the system with external baths plays an important role for the internal dynamics of the quantum state and leads to dissipative effects.
JC \cite{gea-banacloche1993jay,puri1986col} and TC \cite{meunier2006ent,munoz2019sym,dhar2018var,lemini2018bou} models with dissipation, often in the presence of an external driving, have been widely studied.
In this section we study the impact of dissipation on the temporal and spectral properties as well as on the Wigner functions for our system.
We assume that the lifetime of the TLS is much longer than the lifetime of photons in the resonator mode, such that we only take into account the decay of the photon mode (e.g., due to photon losses through the mirrors of the cavity) with a rate $\Gamma$ \cite{meunier2006ent,munoz2019sym}.
To analyze the changes induced by the damping we solve the Master equation in Lindblad form with the dissipation rate $\Gamma$, which in the frame scaled with the dimensionless dissipation rate $\widetilde{\Gamma} = \Gamma N / (G \sqrt{N})$ is given by
\begin{align}
\dv{\tau} \hat{\rho}^{(N)} &= -\frac{i}{\hbar G \sqrt{N}} \left[ \hat{H}, \hat{\rho}^{(N)} \right] \nonumber \\
&+ \widetilde{\Gamma} \left[ \hat{\alpha} \hat{\rho}^{(N)} \hat{\alpha}^\dagger - \frac{1}{2} \left( \hat{\rho}^{(N)} \hat{\alpha}^\dagger \hat{\alpha} + \hat{\alpha}^\dagger \hat{\alpha} \hat{\rho}^{(N)} \right)\right],
\end{align}
use the same initial condition as above
\begin{align}
\hat{\rho}^{(N)}(0) &= \ket{\Psi^{(N)}(0)} \bra{\Psi^{(N)}(0)}, \nonumber
\end{align}
calculate the expectation value
\begin{align}
\alpha^{(N)}(\tau) = \Tr[\hat{\rho}^{(N)}(\tau) \hat{\alpha}],\nonumber
\end{align}
and the corresponding Wigner functions in the scaled frame $\wigner{\widetilde{W}}{N}{U_\alpha, \Pi_\alpha, \tau}$.
\begin{figure}
\centering
\includegraphics[width=1\columnwidth]{{fig11.pdf}}
\caption{(a) Dynamics of $\highindex{\alpha}{N}$ for $N=20$ with various damping rates $\widetilde{\Gamma}$ and (b) corresponding spectra. The inset shows an enlargement of the region of the right side peaks.}
\label{fig11}
\end{figure}
Figure~\ref{fig11} shows the dynamics of the coherent amplitude $\highindex{\alpha}{N}(\tau)$ in (a) and the corresponding spectra in (b) for various damping rates $\widetilde{\Gamma}$ and $N=20$.
In \cref{fig11} (a) we clearly see that the damping has a remarkable influence on the strength and the time of the revival.
The revival time reduces for increasing $\widetilde{\Gamma}$ (as was similarly observed in the damped JC model \cite{puri1986col}), which is not surprising since the revival time depends on the average number of photons and the dissipation is associated with the loss of photons.
We also observe that the amplitude of the weak oscillations during the revival is significantly reduced even for relatively small dissipation.
As seen in the spectra in \cref{fig11} (b), this damping leads to a broadening of the individual lines constituting the side peaks, such that already for $\widetilde{\Gamma} \approx 0.01$ smooth side peaks emerge, while in the central peak the individual lines are still visible.
If the damping increases further and reaches values larger than the inverse collapse time, the revival vanishes and the initial decay of the coherent amplitude becomes faster.
This leads to a strong broadening also of the central line in the spectrum (not shown).
\begin{figure}
\centering
\includegraphics[width=1\columnwidth]{{fig12.pdf}}
\caption{Transformed Wigner functions $\wigner{\widetilde{W}}{N}{U_\alpha, \Pi_\alpha, \tau} / N$ at the time $\tau_R / 10$ for different damping rates $\widetilde{\Gamma}$. The value of $\tau_R$ refers to the revival time of the undamped system. The black circles correspond to the trajectory on which the initial Gaussian would move for $\widetilde{\Gamma} = 0$.}
\label{fig12}
\end{figure}
In \cref{fig12} the Wigner functions $\wigner{\widetilde{W}}{N}{U_\alpha, \Pi_\alpha, \tau_R/10}$ at the time $\tau = \tau_R/10$ are plotted for different damping rates $\widetilde{\Gamma}$, where $\tau_R$ refers to the revival time of the undamped system. The black circle marks the trajectory of the initial Gaussian without any coupling or dissipation.
By comparing the position of the peaks and the black circle one can see, that the entire distribution shrinks towards the center of the phase space reflecting the loss of photons from the system.
For increasing $\widetilde{\Gamma}$ we clearly see that the interferences in the middle of the circle vanish first, while the ones on the circumference of the circle between neighboring peaks remain present for relatively large $\widetilde{\Gamma} > 0.1$. Indeed, as has been shown in Ref.~\cite{meunier2006ent}, the decoherence functional depends on the phase difference $\Delta \phi_{k, l} = \phi_k - \phi_l$ between the Gea-Banacloche states $\ket{\Phi_k(\tau)} = \ket{\xi e^{i\phi_k}}$ with $\phi_k = 2\pi k \tau /\tau_R$ when neglecting the loss of photons.
Due to the dissipation large values of $\Delta \phi_{k, l}$ are averaged out on shorter time scales than small $\Delta \phi_{k, l}$.
In \cref{fig12} the coherences on the circumference are induced by neighboring peaks (small $\Delta \phi_{k, l}$) and the ones in the center by opposing lines (large $\Delta \phi_{k, l}$), which explains why the coherences in the center are damped away earlier than the ones on the circumference.
\section{Conclusions}\label{sec.conclusions}
We have presented a detailed comparison of the spectral and temporal characteristics obtained from a semiclassical and a fully quantum optical treatment of an ensemble of QEs in a cavity, in which a coherent initial state is created by excitation with an ultrashort laser pulse.
We have found that the spectra of the quantum system show the same features as the semiclassical system even for relatively small numbers of QEs, however, they exhibit a broadening which decreases with increasing number of QEs.
The collective width seen in the spectra can be traced back to the finite collapse time observed in the dynamics of the quantum system.
Because of this collapse, which is absent in the semiclassical case, the coherent field amplitude matches only on small time scales.
During this collapse time, however, the frequencies agree very well leading to the similarity of the spectra.
Adding the effect of dissipation, we explicitly showed that with increasing damping rate the side peaks in the Rabi regime and for higher rates also the central peak are broadened until eventually the revival vanishes.
By investigating the temporal evolution of the Wigner function for the cavity mode at small times, it was possible to show that a good agreement in the expectation values is in general not associated with a quasi-classical behavior of the quantum state.
In the Wigner functions we found squeezing at short times and strong broadenings, in particular in the $\Pi_\alpha$ direction, on time scales where the temporal field characteristics of the semiclassical and the quantum model still agree very well.
The deviations from a classical state thus seem to have only a minor influence on the expectation values on these time scales.
Nonetheless we have found that the semiclassical model describes the complete quantum system very well for small initial field values, i.e., small amplitudes of the exciting laser pulse.
We then have studied the Wigner function at different times during the collapse and revival of the coherent field amplitude.
At the revival time $\tau_R$ the Wigner function is indeed concentrated in the phase space region of a corresponding coherent state, the shapes, however, for the values of the mean photon numbers studied here deviate strongly from the circular shape, expected for a coherent state.
At the times $\tau_R/M$ for $1 \leq M \leq N$ the Wigner function exhibits the structure of Schr\"odinger cat states consisting of a superposition of $M$ quasi-coherent states distributed on a circle around the origin in phase space.
By additionally accounting for dissipation, we showed that for increasing dissipation rates the coherences in the center of the circle are destroyed first, while the ones between neighboring states are much more robust.
\section*{Acknowledgments}
DW thanks NAWA for financial support within the ULAM program (No. PPN/ULM/2019/00064).
|
\section{Introduction}\label{introduction section}
Throughout this paper, $M$ is either a compact 3-manifold or a compact manifold with a link removed.
We assume we are in the piecewise-linear category, and all manifolds may have boundary unless specified to be closed. When a manifold is said to be hyperbolic, then there are no spherical or projective plane boundaries, and there is a hyperbolic metric on the manifold obtained by removing torus boundaries and Klein bottle boundaries, with finite hyperbolic volume.
When there are additional higher genus boundaries, for hyperbolicity, we further assume that they are realized as totally geodesic surfaces in the hyperbolic metric on the manifold. This is called tg-hyperbolicity. We do not assume that surfaces in $M$ are connected.
When a 3-manifold possesses a hyperbolic metric, we can use the geometric invariants from hyperbolicity to understand the 3-manifold. In the case that the 3-manifold has a hyperbolic metric of finite hyperbolic volume, the Mostow-Prasad Rigidity Theorem says that this metric, and hence its volume, is a topological invariant.
A great amount of research has been devoted to approximating the volume of particular families of hyperbolic 3-manifolds. Much of this work has focused on link complements in the 3-sphere $S^3$.
For instance, in \cite{Lackenby}, it was proved that if $L$ is an alternating link in $S^3$ that is hyperbolic, then \[v_{\text{oct}} (t(D) -2)/2 < \vol(S^3\setminus L )< 10 v_{\text{tet}} (t(D) -1)\] where $t(D)$ is the number of twist regions in a twist-reduced alternating diagram of $L$, $v_{\text{oct}}$ is the volume of an ideal regular octahedron, approximately 3.6638, and $v_{\text{tet}}$ is the volume of an ideal regular tetrahedron, approximately 1.0149. Additional lower bounds for hyperbolic alternating link volumes appear in \cite{CKP2}.
If we restrict to hyperbolic Montesinos links, lower bounds for volumes that depend on the combinatorics of the link diagram are provided in \cite{FP} and \cite{FKP}. Other categories of links with lower bounds include prime link diagrams with twist-reduced diagrams of twist number at least two and all twist regions containing at least seven crossings \cite{FKP2}, homogeneously adequate links \cite{FKP3}, and positive braids \cite{FKP, Giambrone}.
However, in all cases, obtaining accurate lower bounds on volume has proved elusive. In this paper, given a link in $S^3$ that can be appropriately decomposed along surfaces into tangles, we define certain types of hyperbolicity for the tangles which, when satisfied, provide an associated volume for each tangle.
Using results from \cite{Agol-Storm-Thurston}, as generalized in \cite{CFW}, we can then show that links obtained from the tangles are themselves hyperbolic.
Further, we obtain surprisingly accurate lower bounds on the volumes of the links from the sum of the volumes of the constituent tangles. However, unlike the previous lower bounds that depend only on the combinatorics of the diagrams, this method does assume that the volumes of the constituent tangles are known. We provide examples of the various volumes of tangles in the last section.
These results for links in $S^3$ are particular cases of a much more general theorem that shows that if there is an appropriate decomposition of a 3-manifold into pieces, each of which is hyperbolic in the appropriate way, the manifold is hyperbolic with volume at least as large as the sum of the associated volumes of the pieces, versions of which we prove in Sections 2 and 3. In the rest of the introduction, we focus on the application to links in $S^3$.
For our purposes, a tangle is a properly embedded disjoint union of $m$ arcs and potentially some additional circles in a ball.
These are sometimes called $2m$-tangles in the literature.
The first situation applies to a link $L$ in the open thickened torus $T \times (0,1)$, which we can think of as the complement of the Hopf link in $S^3$.
Suppose $L$ decomposes into $n$ tangles $(\cT_i)_{i=1}^n$, arranged in a cycle that follows the core of $T \times (0,1)$ as pictured in Figure \ref{Torus cycle figure}(a). There may be an arbitrary nonzero number of connecting strands between any two adjacent tangles.
\begin{figure}
\centering
\includegraphics[width=.6\textwidth]{Figures/torus_cycles_3d_patterned.pdf}
\caption{Part (a) shows a link decomposing into a cycle of tangles in $T \times (0,1)$.
Each node is replaced with a tangle.
Part (b) shows surfaces that decompose the link into tangles.
}
\label{Torus cycle figure}
\end{figure}
We may construct a surface $\Sigma \subset M \setminus L$ by taking the disjoint union of a collection of multiply-punctured annuli, each corresponding to a meridianal annulus between adjacent tangles as shown in Figure \ref{Torus cycle figure}(b).
Cutting the manifold open along $\Sigma$ yields the disjoint union of the individual tangles in thickened cylinders ending at the ``boundary'' meridianal annuli.
Since there is a homeomorphism of $T \times (0,1)$ taking meridians to $(p,q)$-curves, we could have just as easily cut $L$ along a collection of annuli which bound $(p,q)$ curves on each component of $T \times \partial I$.
We refer to such a tangle as a \emph{thickened-cylinder tangle}, and we associate with it a link in a thickened torus called its \emph{double} or \emph{2-replicant} $\Dthick{2}{\cT_i}$, formed by doubling across the image of $\Sigma$.
If this link is hyperbolic in $T \times (0,1)$, we say that $\cT_i$ is \emph{2-hyperbolic}, and define the 2-volume $\volsolid{2}{\cT_i}$ to be half of the volume of this double.
We prove that if $L \subset T \times (0,1)$ is a link in a thickened torus such that there exist surfaces separating $L$ into thickened-cylinder tangles $\cT_i$ and each $\cT_i$ is 2-hyperbolic,
then $L$ is hyperbolic, and its volume satisfies
\[
\vol(L) \geq \sum_i \volsolid{2}{\cT_i}.
\]
If a given thickened-cylinder tangle can be further decomposed into a cycle, we may ask about hyperbolicity of the individual tangles in cubes, called \emph{cubical tangles}.
An example of this is illustrated in Figure \ref{thickened torus example}.
A cubical tangle $\cT_{ij}$ doubles to a thickened-cylinder tangle.
We say that a cubical tangle is \emph{$(2,2)$-hyperbolic} if the double of its double $\Dthick{(2,2)}{\cT_{ij}}$ is a hyperbolic link, in which case we define the \emph{$(2,2)$-volume}
\[
\volthick{(2,2)}{\cT_{ij}} := \frac{1}{4} \vol\prn{\Dthick{(2,2)}{\cT_{ij}}}
\]
In fact, given a decomposition of a fixed thickened-cylinder tangle $\cT_i$ into cubical tangles $\cT_{ij}$, the double of the thickened-cylinder tangle can be realized as a cycle of thickened-cylinder tangles, each of which is doubled from some $\cT_{ij}$.
Then, if $\cT_i$ is a thickened-cylinder tangle, and there are surfaces separating $\cT_i$ into a cycle $\cT_{ij}$ of cubical tangles and each $\cT_{ij}$ is $(2,2)$-hyperbolic,
then, $\cT_i$ is $2$-hyperbolic, with volume satisfying
\[
\volthick{2}{\cT_i} \geq \sum_j \volthick{(2,2)}{\cT_{ij}}.
\]
Moreover, if a link $L$ in a thickened torus decomposes into cubical tangles such that each $\cT_{ij}$ is $(2,2)$-hyperbolic, then $L$ is hyperbolic with volume satisfying
\[
\vol(L) \geq \sum_{ij} \volthick{(2,2)}{\cT_{ij}}.
\]
In a future paper \cite{selectalternating}, we will give a large class of $2$-hyperbolic thickened-cylinder tangles, called \emph{select alternating thickened cylinder tangles} and a class of $(2,2)$-hyperbolic cubical tangles called \emph{select alternating cubical tangles}.
As a corollary to the second of those, it is shown that all rational tangles in a cube with one endpoint on each edge of the square, other than the integer tangles 0, $\pm 1$, and $\infty$, are (2,2)-hyperbolic cubical tangles.
\begin{figure}[htbp]
\centering
\includegraphics[scale=.15]{Figures/thickened_torus_example.png}
\caption{A link $L$ in $T \times I$ and a decomposition of $T \times I$ into a thickened cylinder and two cubes corresponding to the tangles. One of the cubes is shaded.}
\label{thickened torus example}
\end{figure}
\begin{example} Consider the link $L$ in $T \times (0,1)$ shown in Figure \ref{thickened torus example}. We can separate $L$ along a surface into a thickened-cylinder tangle $\cT_1$ on the left, and two cubical tangles $\cT_{21}$ and $\cT_{22}$ on the right. By \cite{selectalternating}, we know $\cT_1$ is 2-hyperbolic, while $\cT_{21}$ and $\cT_{22}$ are both $(2,2)$-hyperbolic. Thus, by Theorem \ref{Torus theorem} and Theorem \ref{Square tangles theorem}, $\mbox{vol}^2_{T \times I}(\cT_1) + \mbox{vol}^{(2,2)}_{T \times I}(\cT_{21}) + \mbox{vol}^{(2,2)}_{T \times I}(\cT_{22})$ gives a lower bound of 34.7259 on the volume of $L$. The actual volume is 37.8023.
\end{example}
The next result applies to a link $L$ in $S^3$ constructed as a cycle of $2n$ tangles, which we call a {\it bracelet link}. We define the length of the bracelet link to be $2n$.
For the case $n=2$, this is pictured in Figure \ref{braceletlink}. Each tangle is connected to the subsequent tangle in the cycle by two or more strands. For example, a Montesinos link is an example of a bracelet link, where the individual tangles are rational tangles and there are exactly two strands between adjacent tangles (absorbing the additional twist allowed into one of the tangles). If the number of tangles used in the construction of the Montesinos link is odd, we can treat one adjacent pair of tangles as a single tangle so that the number of tangles is even.
\begin{figure}[htbp]
\centering
\includegraphics[scale=1.2]{Figures/braceletlink.pdf}
\caption{A bracelet link of length 4.}
\label{braceletlink}
\end{figure}
Given a bracelet link, we cut the 3-sphere into ``flying" \emph{saucers} $S_1, \dots, S_{2n}$ obtained by taking $n$ spheres all sharing the circle $\alpha$ in $S^3$ corresponding to the vertical axis at the center of the link and each cutting through the strands connecting two adjacent tangles. The result is $2n$ tangles, each in a unique saucer, the boundary of which is two bowed disks sharing the central circle. The endpoints of the tangle lie on the two disks. We call the $i$th saucer $S_i$ and its tangle $\cT_i$, as in Figure \ref{saucer}. For convenience, in certain situations we let $\cT_i$ represent both the tangle and its saucer. We call the two punctured disks that make up the boundary of the saucer the \emph{faces} of the saucer.
Let $\ell_i$ be the number of stands connecting $\cT_i$ and $\cT_{i+1}$, and
$\ell_{0}$ the number of strands connecting $\cT_{2n}$ with $\cT_1$. Note that $\ell_i \geq 2$.
\begin{figure}[htbp]
\centering
\includegraphics[scale=.7]{Figures/saucer.pdf}
\caption{A saucer and tangle.}
\label{saucer}
\end{figure}
We define the \emph{$2n$-replicant} $\Dsphere{2n}{\cT_i}$ of $\cT_i$ to be the bracelet link formed by the cycle $(\cT_i,\cT_i^R,\cT_i, \dots,\cT_i^R)$ of $2n$ tangles where $\cT_i^R$ is the ``reflected tangle'' obtained by reflecting the saucer for $\cT_i$ through the face containing the endpoints to one side of $\cT_i$ and perpendicular to the strands.
So $\cT_i$ has $\ell_{i-1}$ endpoints on the first face and $\ell_i$ on the second, whereas $\cT_i^R$ has $\ell_i$ on the first face and $\ell_{i-1}$ on the second. See Figure \ref{braceletreflection}.
\begin{figure}[htbp]
\centering
\includegraphics[scale=1.2]{Figures/braceletreflection.pdf}
\caption{The $2n$-replica $\Dsphere{2n}{\cT_i}$ of $\cT_i$.}
\label{braceletreflection}
\end{figure}
We say that $\cT_i$ is \emph{$2n$-hyperbolic} if $\Dsphere{2n}{\cT_i}$ is hyperbolic, in which case we define
\[
\volsphere{2n}{\cT_i} := \frac{\vol\prn{\Dsphere{2n}{\cT_i}}}{2n}.
\]
We prove that if $L$ is a bracelet link made of a cycle $(\cT_i)_{i=1}^{2n}$ of $2n$ saucer tangles such that each $\cT_i$ is $2n$-hyperbolic, then $L$ is hyperbolic and the volumes satisfy
\[
\vol(L) \geq \sum_i \volsphere{2n}{\cT_i}.
\]
\begin{figure}[htbp]
\centering
\includegraphics[scale=.5]{Figures/pretzellink.pdf}
\caption{A bracelet link.}
\label{pretzellink}
\end{figure}
\begin{example}
To see an explicit example, consider the length six bracelet link projection of Figure \ref{pretzellink}. The SnapPy computer program \cite{SnapPy} yields a volume of approximately 32.9819.
The link is prime, alternating and twist-reduced with twist number 6, so Lackenby's lower bound on volume applies and yields $2v_{\text{oct}} \approx 7.3276.$ The lower bound that comes from \cite{FP} and \cite{FKP} is $3v_{\text{oct}} \approx 10.9914.$ Other known bounds are lower than these. To apply our theorem, we split the link into six saucer tangles along the union of three spheres all intersecting at the central circle, five of the tangles of which are vertical integer 4 tangles and one of which is a vertical integer 5 tangle. Summing $\volsphere{6}{\cT_i}$ for $i = 1, \dots, 6$, we obtain a lower bound of 32.7858.
The theorem also implies that if we alter this bracelet link by replacing some of the tangles by their reflections, so that the resulting link is non-alternating, it will still be hyperbolic and the same lower bound on volume applies. All of the volumes for this collection of links are lower than the volume of the alternating bracelet link with which we started and the least volume is realized for the links where half of the tangles have been reflected. The corresponding volume for these is approximately 32.7926. Thus, in this case, the lower bound is within .01 of the actual volume.
\end{example}
In certain situations, we can further decompose the saucer tangles. A tangle in a tetrahedron such that all endpoints are on the cross-sectional square as in Figure \ref{tetrahedraltangle}, with at least one endpoint on each edge of the square and none at the corners, is called a {\it tetrahedral tangle.}
A saucer tangle can sometimes be decomposed into tetrahedral tangles. In particular, as in Figure \ref{toruslattice}, certain links form an $(m,n)$-rectangular lattice of tangles on an unknotted torus $T$. In this case, the 3-sphere can be decomposed into $mn$ tetrahedra, each containing a tetrahedral tangle. For each tetrahedron, one edge lies on the core curve of the solid torus to the inside of $T$ and is shared by $n$ other tetrahedra, and one edge lies on the core curve of the outside solid torus, and is shared by $m$ tetrahedra. The remaining four edges of each tetrahedron are each shared by four others. In particular, on the standardly embedded torus, there are $m$ columns and $n$ rows of tangles.
\begin{figure}[htbp]
\centering
\includegraphics[scale=.7]{Figures/toruslattice.pdf}
\caption{A link given as a torus lattice in $S^3$.}
\label{toruslattice}
\end{figure}
\begin{figure}[htbp]
\begin{subfigure}[b]{0.2\textwidth}
\includegraphics[width=\textwidth]{Figures/saucertet.pdf}
\caption{Decomposing a saucer into tetrahedra.}
\label{fig:f1}
\end{subfigure}
\begin{subfigure}[b]{0.2\textwidth}
\includegraphics[width=\textwidth]{Figures/tetrahedraltangle.pdf}
\caption{A tetrahedral tangle.}
\label{tetrahedraltangle}
\end{subfigure}
\caption{Decomposing saucer tangles into tetrahedral tangles.}
\label{suacertet2}
\end{figure}
Given a tetrahedral tangle $\cT_i$, we can form the $(2m, 2n)$-replicant, denoted $D^{(2m, 2n)}(\cT_i)$ by creating a $2m \times 2n$ torus lattice link with copies of $\cT_i$ replacing each vertex, such that two connected tangles are reflections of one another.
We say that a tetrahedral tangle $\cT_i$ is \emph{$(2m, 2n)$-hyperbolic} if $\Dsphere{(2m,2n)}{\cT_i}$ is hyperbolic, in which case we define
\[
\volsphere{(2m,2n)}{\cT_i} := \frac{\vol\prn{\Dsphere{2n}{\cT_i}}}{4mn}.
\]
We show that if $L$ is a $2m \times 2n$ torus lattice link made of a collection $\{\cT_{i,j}\}_{i=1}^{2n}$ of $4mn$ tetrahedral tangles such that each $\cT_{i,j}$ is $(2m,2n)$-hyperbolic, then, $L$ is hyperbolic and the volumes satisfy
\[
\vol(L) \geq \sum_{i,j} \volsphere{(2m,2n)}{\cT_{i,j}}.\]
In fact, it is not necessary that every saucer tangle decompose into tetrahedral tangles. If even a subset of them do, we obtain a similar result.
\begin{example} \label{tettangleexample} The tetrahedral tangle $\cT$ in its projection to the cross-sectional square in Figure \ref{tetrahedralbigon}(a) is $(2m,2n)$-hyperbolic for all $m,n \geq 1$; this can be checked for $m = n = 1$ explicitly, and extended to arbitrary $(2m,2n)$ by Theorem \ref{Saucer hyperbolicity theorem}.
Thus any $(2m, 2n)$-torus lattice link in $S^3$ constructed from copies of just this tangle is hyperbolic and has volume at least $4mn \volsphere{(2m,2n)}{\cT}$. For example, since $\volsphere{(2,2)}{\cT} = 3.1322306$, any $(2,2)$-torus lattice link in $S^3$ constructed by placing this tangle or its reflections over a vertical or horizontal line into the four slots in the (2,2)-lattice has volume at least 12.5289226. The lower bound on volume is realized by the link in Figure \ref{tetrahedralbigon}(b) obtained by placing the tangles in the slots so each is a reflection of the others across the separating surfaces, making those surfaces totally geodesic. This is in fact the $(2,2)$-replicant of the tangle.
However, caution must be exercised. If the tangle is rotated $90^\circ$ before being placed in the same tetrahedron or if its crossings are switched, it is no longer $(2,2)$-hyperbolic. And if that tangle is included as an option in the (2,2)-lattice, the link may no longer be hyperbolic and even if it is, the volume of the link is no longer necessarily bounded below by 12.5289226.
If on the other hand, we considered this tangle as a tangle in a cube, it and its reflections and all of their rotations are (2,2)-hyperbolic in the thickened torus.
\end{example}
\begin{figure}[htbp]
\centering
\includegraphics[scale=.7]{Figures/rationaltanglelink.pdf}
\caption{A 2-tangle and a (2,2)-lattice link generated by copies of it.}
\label{tetrahedralbigon}
\end{figure}
Similar results for tangles that we have described in the thickened torus and in the 3-sphere also apply to links in an open solid torus, which we can think of as the complement of a single unknotted component, as well as to links in $S^2 \times S^1$.
The proofs of the main result and generalizations rely on the following theorems.
The first is well known \cite[Thm 1.10.15]{Klingenberg}.
\begin{theorem}\label{fixed surface lemma}
Let $M$ be a Riemannian manifold and let $f:M \rightarrow M$ be a nontrivial isometry.
Then, the subset $\Fix(f)$ is a union of embedded totally geodesic submanifolds.
\end{theorem}
The second is due to Agol, Storm, and Thurston in \cite[Thm. 9.1]{Agol-Storm-Thurston}, as generalized by Calegari, Freedman and Walker in \cite[Thm. 5.5]{CFW}.
\begin{theorem}\label{Cutting and pasting theorem}
Let $\overline{M}$ be a compact manifold with interior $M$ a hyperbolic 3-manifold of finite-volume.
Let $\Sigma$ be a properly embedded two-sided totally geodesic surface, let $\psi:\Sigma \rightarrow \Sigma$ be a diffeomorphism, and let $M'$ be the manifold formed by surgering along $\Sigma$ and gluing the resulting pieces together via $\psi$.
Then, $M'$ is a hyperbolic 3-manifold of finite volume, satisfying
\[
\vol(M') \geq \vol(M).
\]
Equality is attained if and only if $\Sigma$ is totally geodesic in $M'$.
\end{theorem}
We describe the idea of the proof of our main result for bracelet links. We begin with a copy of the $2n$-replicant of each tangle making up the bracelet link. Each of the decomposition surfaces in these link complements are totally geodesic by Theorem \ref{fixed surface lemma}. Then by repeated cutting and regluing, always along totally geodesic surfaces, we obtain $2n$ copies of the original link complement. Then Theorem \ref{Cutting and pasting theorem} implies that the resulting set of copies must have total volume greater than the sum of the volume of the replicants, and therefore the original link complement has volume greater than the sum of the volumes of the individual tangles. This turns a geometric question into a combinatorial question. We answer that question in a very general setting in Section 2, where the main theorem is stated and proved.
In Section 3, we generalize the main theorem in various ways.
In Section 4, we apply these results to prove the theorems listed above for the various link complements in $S^3$.
In Section 5, we consider the determination of appropriate hyperbolicity for various kinds of tangles.
In particular, we prove that if $\cT$ is a $2n$-hyperbolic saucer tangle for some $n \geq 1$,
then, $\cT$ is $2m$-hyperbolic for all $m \geq n$.
Thus, once we know a given tangle is $2n$-hyperbolic, and so it can be used to prove hyperbolicity and lower bound volume of a $2n$-bracelet link, it can also be used in any bracelet link of greater length.
A corollary to this result shows that if a tetrahedral tangle is $(2m,2n)$-hyperbolic, then it is $(2r,2s)$-hyperbolic for all $r \geq m$ and $s \geq n$.
In Section 6, we consider explicit examples of saucer tangles. We use results of \cite{Volz}, \cite{Wu} and Theorem \ref{Saucer hyperbolicity theorem} to determine $2n$-hyperbolicity for all arborescent tangles.
We go on to prove certain solid-cylinder tangles are 2-hyperbolic by relating them to $2m$-hyperbolic saucer tangles having alternating projections.
Both the cyclic graph on $2n$ vertices and the torus lattice graph on $mn$ vertices are examples of graphs in $S^3$ such that the replacement of the vertices with tangles and the edges with strands connecting the tangles yields a link for which a version of Theorem \ref{Main theorem separating} holds. That is to say, if all of the constituent tangles of a given link $L$ are hyperbolic in the appropriate way, then they each have an associated volume and the link is hyperbolic with volume at least as large as the sum of the volumes of the tangles. We call such a graph a \emph{tangle reflection graph}. In Section 8, we extend beyond cyclic graphs and torus lattice graphs to a variety of other tangle reflection graphs.
In Section 9, we provide tables of various types of volumes for some tangles.
\section{The main theorem for a separating starburst}\label{Main theorem separating section}
\subsection{Definitions and statement of the theorem}
In this section we prove that if given an appropriately defined collection of surfaces in a 3-manifold that together intersect in a 1-manifold, such that the pieces they decompose the manifold into are appropriately hyperbolic, then the manifold is hyperbolic with volume at least as large as the sum of the appropriately defined volumes of the pieces.
In the case that the manifold $N$ has boundary components of genus greater than one, we double the manifold across these boundaries. The resulting manifold $N'$ has only torus boundary, and the collection of surfaces doubles to a collection of surfaces in the new manifold. Then we can work with that manifold. The manifold $N$ is tg-hyperbolic if and only if the manifold $N'$ is hyperbolic. Thus, from now on we will work with manifolds with no higher genus boundary, keeping in mind the results for those manifolds imply results for those that have such boundary components.
We motivate the more general results by returning to the example of a bracelet link $L \subset S^3$ that decomposes into a cycle of $2m$ saucer tangles $(\cT_i)_{i=1}^{2m}.$
There is a collection $\bS = (\Sigma_i)_{i=1}^m$ of surfaces in $M = S^3 \setminus L$ that separates $M$ into $2m$ pieces that may each be viewed as the complement of a particular tangle $\cT_i$ in a saucer.
Local to the locus where these surfaces intersect, these resemble the product of a circle with Figure \ref{newstarburstfigure}(a).
We cut along $\bS$ and analyze the resulting pieces, generalizing to the following situation.
\begin{definition}
Let $M$ be a compact orientable 3-manifold, possibly with torus boundaries, and let $\bS = \cbr{\Sigma_1,\dots,\Sigma_m}$ be a collection of $m$ two-sided surfaces properly embedded in $M$.
We denote by $M \dbs \bS$ the manifold $M \dbs (\Sigma_1 \cup \cdots \cup \Sigma_m)$ obtained by removing the interior of a regular neighborhood of the union of the surfaces.
We assume that any point in the intersection set of two of the surfaces is in fact in the intersection set of all of the surfaces, which is to say, the intersection set $E = \Sigma_i \cap \Sigma_j = \bigcap _{k = 1}^m \Sigma_k$ for any $i \neq j$. We further assume that $E$ is a compact 1-manifold properly embedded in $M$.
We suppose that there is a regular neighborhood $U$ of $E$ such that the induced pair $(U,U \cap \cup_{k=1}^m \Sigma_k)$ is diffeomorphic to $(D \times E, S_m \times E)$, where $D$ is an open disk and $S_m$ is the union of $m$ open line segments $L_i$ in $D$, in cyclic order, each passing directly through a single intersection point at the center, as in Figure \ref{newstarburstfigure}(a).
We further require that the diffeomorphism takes $U \cap \Sigma_i$ to $L_i \times E$ in cyclic order.
Suppose further that $M \dbs \bS$ can be expressed as the disjoint union of $2m$ 3-manifolds, denoted $M_1, \dots M_{2m}$, each intersecting the boundaries of the regular neighborhoods of two of the surfaces $\Sigma_j$ with indices differing by one modulo $2m$, as in Figure \ref{newstarburstfigure}(b).
Then we say that $(M, \bS)$ is an {\bf $m$-sheeted separating starburst}.
\end{definition}
For convenience we allow the case of $E = \emptyset$ in the definition.
\begin{figure}
\centering
\includegraphics[scale=0.6]{Figures/starburst1.pdf}
\caption{The starburst $S_m$ for $m = 3$ appears in (a). The cross section of the neighborhood $U$, after the removal of the regular neighborhood of the union of the surfaces appears in (b).
}
\label{newstarburstfigure}
\end{figure}
In the case of a bracelet link $N = S^3 \setminus L$, the set $\bS$ consists of $m$ punctured spheres, all intersecting in a single circle.
The manifold $N \dbs \bS$ is a disjoint union of $2m$ connected 3-manifolds $N_i$, each given by the complement of a tangle in a saucer.
We seek to associate with each tangle complement $N_i$ a notion of hyperbolicity, and we do so by bootstrapping the definition of hyperbolicity defined for links.
As defined in the introduction, in the case of a saucer tangle, there is a well-defined notion of a \emph{$2m$-replicant} formed by gluing together a cycle of $2m$ copies of $\cT_i$ along compatible endpoints.
We say that $\cT_i$ is \emph{$2m$-hyperbolic} if its $2m$-replicant is a hyperbolic link.
This gluing occurs by specifying surfaces $V^1_i,V^2_i \subset \partial N_i$ such that $V^1_i \cap V^2_i = E$, given by the \emph{counter-clockwise pointing} and \emph{clockwise pointing} sides of the subsurface of $\partial N_i$ corresponding with $\bS$, and gluing copies of $V^j_i$ in different copies of $N_i$ together in a cycle.
In analogy with the case of tangles, we may define the $2m$-replicant of this data as follows.
\begin{definition}
A \emph{piece} is a pair $(P,(V^1,V^2))$ where $P$ is a 3-manifold with distinguished surfaces $V^1,V^2 \subset \partial P$ such that $V^1 \cap V^2$ is a 1-manifold (possibly disconnected and possibly with boundary).
We may define the $2m$-replicant of a piece $D^{2m}(P,\prn{V^1,V^2})$ by reflecting $2m$ times over $V^1$ or $V^2$ as follows.
\begin{equation}\label{2n replication equation}
D^{2m}(P,\prn{V^1,V^2}) := \frac{\coprod_{i=1}^{2m} P}{(V^1,2i) \sim (V^1,2i+1) \hspace{10pt} \text{ and } \hspace{10pt} (V^2,2i-1) \sim (V^2,2i)}
\end{equation}
Here, arithmetic is performed modulo $2m$.
We say that $(P,\prn{V^1,V^2})$ is \emph{$2m$-hyperbolic} if $D^{2m}(P,\prn{V^1,V^2})$ is hyperbolic, in which case we define
\[
\vol^{2m}(P,\prn{V^1,V^2}) := \frac{1}{2m} \vol(D^{2m}(P,\prn{V^1,V^2})).
\]
In the case that the surfaces $(V^i)$ are clear, we will often simply refer to such a piece as $P$, and write $D^{2m}(P)$ and $\vol^{2m}(P)$ for the replicant and volume.
\end{definition}
\begin{remark}
The order of $V^1$ and $V^2$ can be interchanged without changing the diffeomorphism class of the replicant.
However, we use the syntax of ordered pairs for reasons that will be clear in the later \emph{$\ell$-piece} definition.
\end{remark}
We let $H$ represent the surface that is the closure of the complement of $V^1 \cup V^2$ in $\partial P$. Then $H$ corresponds to boundary components in $D^{2m}(P,\prn{V^1,V^2})$. When $D^{2m}(P, (V^1, V^2))$ has boundary components of genus one, keep in mind that we remove that boundary component when determining the hyperbolicity of $D^{2m}(P, (V^1, V^2))$.
For an $m$-sheeted separating starburst $(M, \bS)$, the manifold $M \dbs \bS$ is a disjoint union of $2m$ pieces, each with two possibly disconnected surfaces on its boundary corresponding to two consecutive surfaces in the starburst. We denote the $i$th piece and its two distinguished surfaces on its boundary by $(M_i,(V^1_i, V^2_i))$ where gluing $V^2_i$ in the boundary of the $M_i$ to $V^1_{i+1}$ in the boundary of $M_i$, for all $i$ cyclically returns us to the manifold $M$.
We will prove the following result, which demonstrates that one can use $2m$-hyperbolicity of suitably defined pieces of a manifold to prove hyperbolicity and also to obtain a lower bound on the volume of the manifold. This version of the theorem applies to bracelet links in the 3-sphere.
\begin{theorem}\label{Main theorem separating}
Suppose $(M,\bS)$ is an $m$-sheeted separating starburst in a 3-manifold $M$ with all boundary components tori, and $(M_i,(V^1_i,V^2_i))_{i=1}^{2m}$ are the associated pieces. If
$E$ does not intersect $\partial M$ and if all $M_i$ are $2m$-hyperbolic, then $M$ is hyperbolic, with volume satisfying
\[
\vol(M) \geq \sum_{i=1}^{2m} \vol^{2m}(M_i).
\]
Equality is attained if and only if each surface $\Sigma_i \subset M$ is totally geodesic.
\end{theorem}
\begin{warning}
The notation $V^j_i$ is suppressed in the volume bound because the surfaces are chosen canonically, but in general, differing choice of surfaces $V^j_i$ may yield different geometry, which may not obey the volume bound.
\end{warning}
In the case that $E = \emptyset$, so the surfaces $\{\Sigma_i\}$ do not intersect one another, the result is an immediate corollary of Theorem \ref{Cutting and pasting theorem}. For each $i$, we form $D^{2m}\prn{M_i}$, and then, cutting and pasting along the $2m$ totally geodesic surfaces in each of the $D^{2m}\prn{M_i}$, we can form $2m$ copies of $M$, with total volume at least as large as the total volume of the $2m$-replicants. Dividing by $2m$ yields the result.
This is the case for thickened-cylinder tangles in the thickened torus, for instance.
We henceforth assume $E \neq \emptyset$.
In this case, we prove Theorem \ref{Main theorem separating} in Subsection \ref{Separating subsection}.
Then, we consider a generalization of this to \emph{separating starbursts in a piece} in Subsection \ref{Separating inductive subsection}.
A further generalization holds in the case that $M \dbs \bS$ does not necessarily decompose as the disjoint union of $\cbr{M_i}$,
$E$ may intersect $\partial M$, and $M$ is not necessarily orientable.
These will be considered in Section \ref{Main theorem nonseparating section}.
\subsection{Proof of the main theorem for a separating starburst in a manifold}\label{Separating subsection}
For the duration of this subsection, we fix the index $m$, starburst $(M,\bS),$ and resulting pieces $(M_i,(V^1_i,V^2_i))$ as in Theorem \ref{Main theorem separating}, and we assume that each $(M_i,(V^1_i,V^2_i))$ is $2m$-hyperbolic.
It will help to refer to Example \ref{Algebra example} while following the proof.
We suppress the notation $V^j_i$ most of the time.
Note that both $M$ and the $2m$-replicants $D^{2m}(M_i,(V^1_i,V^2_i))$ are formed by gluing together $2m$ elements of $\cbr{M_i}_{i=1}^{2m}$ in a cycle such that all such gluings are given by canonical isometries $V^j_i \simeq V^j_i$ or $V^2_i \simeq V^1_{i+1}$, with arithmetic performed modulo $2m$.
Furthermore, $D^{2m}(M_i) = D^{2m}(M_i,(V^1_i,V^2_i))$ possesses $m$ totally geodesic surfaces. Choosing one of them, $F$, the manifold $D^{2m}(M_i) \dbs F$ is composed of two connected components, where each connected component is formed by gluing together $m$ many copies of $M_i$.
Forming a manifold that is composed of a cycle of pieces, where $m$-many copies of $M_i$ are connected, followed by $m$-many copies of $M_j$, for any $1 \leq i,j \leq 2m$, Theorem \ref{Cutting and pasting theorem} implies that it is hyperbolic with volume at least $\frac{1}{2} \prn{\vol^{2m}(M_i) + \vol^{2m}(M_j)}$.
Our strategy will use this logic to confer properties between different manifolds formed by cycles of $2m$ elements of $\cbr{M_i}$.
We use vector spaces to represent the possible configurations, which we now define.
\begin{construction}\label{Cyclic word construction}
Let $\widetilde W$ be the set of cyclic words of length $2m$ from the alphabet $\cbr{T_1, T_1^R,\dots,T_{2m},T_{2m}^R}$ such that the words are expressed as a product
\[
t_1 \cdots t_{2m}
\]
where for all $i$, the product $t_it_{i+1}$ is a word from the following set
\[
\cbr{T_jT_{j+1}}_{j=1}^m \cup \cbr{T_j^R T_j}_{j=1}^m \cup \cbr{T_j T_j^R}_{j=1}^m \cup \cbr{T_{j+1}^R T_j^R}_{j=1}^m
\]
with arithmetic done modulo $2m$;
these correspond precisely with the manifolds that can be formed formally as cycles of $2m$ elements of $\cbr{M_i}$ along the canonical diffeomorphisms described in the preceding discussion.
We accordingly call the letters $T_i^R$ \emph{reflected}.
There is a map from $\widetilde W$ into the cyclic words on alphabet $\cbr{T_i}$ made by replacing $T_i^R$ with $T_i$.
This map is not injective in the case $2m = 2$;
in this case, the elements $T_1T_2$ and $T_1^RT_2^R$ constitute the only pair of distinct cyclic words that map to the same word without reflections.
We argue that the map is injective when $2m \geq 4$.
Indeed, given the image of a valid word $w$ containing multiple letters, whether a letter is reflected is determined by its exponent and whether the following letter has one-larger or one-smaller index than the preceding letter.
Since these are distinct indices mod $2m$, no two distinct words have the same image containing multiple letters.
Noting that $(T_iT_i^R)^m = (T_i^RT_i)^m$ as cyclic words, this yields injectivity.
We refer to the image of this map as $W$.
For an arbitrary word $w = t_1\cdots t_{2m} \in W$, define $w^R := t_{2m}\cdots t_1$.
Define the subset $\Wknoti i \subset W$ of \emph{palindromic words on $i$ letters} to be those words expressed as a product $t_1^{a_1} \cdots t_i^{a_i}t_i^{a_i} \cdots t_1^{a_1}$, where each $a_i \geq 0$.
This defines a filtration on $W$:
\[
\Wknoti 1 \subset \cdots \subset \Wknoti m \subset W.
\]
Define the $\QQ$-vector space $\Mknoti i$ generated by $\Wknoti i$, and define $U$ to be generated by $W$.
\end{construction}
As in the above discussion, the set $\widetilde W$ corresponds naturally with a collection of 3-manifolds.
In fact, $T_1T_2$ and $T_1^RT_2^R$ correspond with the same 3-manifold up to homeomorphism by reflecting one to obtain the other.
Hence for all $w \in W$, we may unambiguously define the 3-manifold $\cS(w)$ to be the 3-manifold formed by gluing together a cycle of $2m$ many pieces from $\cbr{\prn{M_i,\bS}}$ such that the associated indices correspond with the indices of the letters in $w$.
Further, we may realize the free commutative monoid $\ZZ_{\geq 0}[W] \subset U$ via 3-manifolds by defining
\[
\cS(w_1 + \cdots + w_k) := \cS(w_1) \sqcup \cdots \sqcup \cS(w_k).
\]
Using this, we may define volume. By taking hyperbolic volumes (with non-hyperbolic manifolds having volume 0), there is a map
\[
\vol:W \rightarrow \RR_{\geq 0}
\hspace{50pt}
\vol(w) := \vol(\cS(w)).
\]
This extends linearly to a map $\vol:U \rightarrow \RR$.
Further, whenever $\cS(w_1)$ and $\cS(w_2)$ are hyperbolic, we have $\vol(w_1 + w_2) = \vol(\cS(w_1) \sqcup \cS(w_2)) = \vol(w_1) + \vol(w_2)$.
By the preceding discussion, we have $\cS(T_1T_2\cdots T_{2m}) = M$, so that
\begin{equation}\label{Volume of manifold}
\vol(T_1T_2 \cdots T_{2m}) = \vol(M).
\end{equation}
Additionally, we have $\cS(T_i^{2m}) = D^{2m}(M_i,(V^1_i,V^2_i))$ for all $i$.
Hence the linear transformation $\vol:\Mknoti 1 \rightarrow \RR$ is simply the map
\begin{equation}\label{Evaluation morphism}
\vol : \Mknoti 1 \rightarrow \RR \hspace{50pt} \text{ defined on basis by } \hspace{50pt} T_i^{2m} \mapsto 2m \cdot \vol^{2m}(M_i)
\end{equation}
We prove our theorem by using Theorem \ref{fixed surface lemma} and Theorem \ref{Cutting and pasting theorem} to relate the volumes of different elements of $U$.
These relations will be realized via the following quotient map
\[
\pi: U \twoheadrightarrow V := U / \Span\cbr{2a_1a_2 - a_1a_1^R - a_2a_2^R \mid \operatorname{len}(a_1) = m}
\]
which equates a cyclic word with the average of the ``doubled'' halves made when splitting the word into sub-words of length $m$.
We will see in the following proposition that these relations reduce cyclic words to sums of words made up of a single letter.
\begin{proposition}\label{Reducability proposition}
The restriction $\pi|_{\Mknoti 1}:\Mknoti i \rightarrow V$ is an isomorphism.
In particular, we have
\[
2m \cdot \vol(w) \geq \sum_{i=1}^{2m} q_i(w) \cdot \vol(T_i^{2m}).
\]
for all $w \in W$, where $q_i(w)$ denotes the number of appearances of the letter $T_i$ in $w$.
\end{proposition}
\begin{figure}[htpb]
\centering
\includegraphics[width=.8\textwidth]{Figures/pizza.pdf}
\caption{Some examples of relations laid out by $\pi$ in the case $m = 5$.the roles of $T_1$ and $T_2$ can be reversed.
}
\label{Pizza figure}
\end{figure}
First we work through an example.
\begin{example}\label{Algebra example}
Fix $2m = 10$.
Let $v_i \in \Wknoti 2$ be the monotonic cyclic word $T_1^i T_2^{m-i} T_2^{m-i} T_1^i$ and define $x_j := T_j^{2m}$.
Then, by Figure \ref{Pizza figure} there is a cycle of words $(v_4,v_3,v_1,v_2)$ such that each element of this cycle is expressed as an $\QQ$-linear combination in $V$ of the next in the cycle and an element of $\Wknoti 1$.
In particular, repeatedly applying the relations depicted, we obtain
\[
v_4 = \frac{1}{2}(x_1 + v_3) = \frac{1}{2}(x_1 + \frac{1}{2} (x_1 + v_1)) = \frac{1}{2}(x_1 + \frac{1}{2} (x_1 + \frac{1}{2}(x_2 + v_2))) = \frac{1}{2}(x_1 + \frac{1}{2} (x_1 + \frac{1}{2}(x_2 + \frac{1}{2}(x_2 + v_4))))
\]
This in turn is equivalent to the linear relation $v_4 = \frac{4}{5}x_1 + \frac{1}{5}x_2$, so $\pi(v_4)$ is generated by $\pi(\Wknoti 1)$.
In particular, each of the moves outlined in Figure \ref{Pizza figure} are implemented by Theorem \ref{Cutting and pasting theorem}, so that we have
\[
\vol(v_4) \geq \vol\prn{\frac{1}{2}(x_1 + \frac{1}{2} (x_1 + \frac{1}{2}(x_2 + \frac{1}{2}(x_2 + v_4)))) }
\]
and hence $10\vol(v_4) \geq 8\vol(T_1^{10}) + 2\vol(T_2^{10})$, verifying Proposition \ref{Reducability proposition} in this case.
\end{example}
The following lemma establishes that the map $\pi$ corresponds with relations on volume.
\begin{lemma}\label{Volume relation lemma separating}
Suppose that $w = a_1a_2 \in W$ where $\operatorname{len}(a_1) = m$, and define $w_j := a_ja_j^R$.
Then,
\begin{equation}\label{Volume relation equation}
2\vol(w) \geq \vol\prn{w_1 + w_2}.
\end{equation}
Further, we have $2q_i(w) = q_i(w_1) + q_i(w_2)$ for all $i$, where $q_i(w)$ denotes the number of appearances of the letter $T_i$ in $w$.
\end{lemma}
\begin{proof}
We prove Equation \eqref{Volume relation equation} purely topologically, and note that the second statement is clear.
Since $w_j = a_ja_j^R$, there is a diffeomorphism of $\cS(w_j)$ which sends the piece corresponding with a letter in $a_j$ to the corresponding letter in $a_j^R$;
this diffeomorphism fixes a surface $F_j \subset \cS(w_j)$ such that $\cS(w_j) \dbs F_j$ is the disjoint union of two copies of a manifold formed by gluing pieces of $\cbr{M_i}$ together corresponding with the word $a_j$.
Furthermore, we may glue together one such manifold for each $w_1$ and $w_2$ to form $\cS(w)$;
this outlines a totally geodesic surface $F_1 \sqcup F_2 \subset \cS(w_1 + w_2)$ such that cutting and gluing along $F_1 \sqcup F_2$ yields $\cS(2w)$.
Hence equation \eqref{Volume relation equation} follows from Theorem \ref{Cutting and pasting theorem}.
\end{proof}
Proposition \ref{Reducability proposition} will be proved primarily via the following lemma.
\begin{lemma}\label{Pizza inductive lemma}
Fix $i \geq 2$.
For every $w \in \Wknoti i$, there is a relation of the form $2w = w' + w''$ where $w' \in \Wknoti i$ and $w'' \in \Wknoti{i-1}$.
\end{lemma}
\begin{proof}[Proof of Lemma \ref{Pizza inductive lemma}] Let $2n$ be the length of $w$.
Write
\begin{equation}\label{The cycle notation}
w = t_1^{a_1} \cdots t_i^{a_i}t_i^{a_i} \cdots t_1^{a_1} = t_1^{2a_1}t_2^{a_2} \cdots t_i^{a_i}t_i^{a_i} \cdots t_2^{a_2}.
\end{equation}
We may replace $w$ with $w^R$, so suppose without loss of generality that $a_1 \leq a_i$.
Group the leftmost $n$ terms in the right-hand side of \eqref{The cycle notation} into a word $x$, and write $w = xy$.
Note that, since $a_1 \leq a_i$, we have
\[
x = t_1^{2a_1} \cdots t_{i-1}^{a_{i-1}} t_{i}^{b}
\]
for some $b \leq a_i$.
Hence $xx^R \in \Wknoti i$.
Further, note that we have $2a_1 \leq n$, so that
\[
y = t_i^{2a_i - b} \cdots t_2^{a_2}
\]
implying that $yy^R \in \Wknoti{i-1}$.
Setting $w' := xx^R$ and $w'' := yy^R$ yields the desired relation.
\end{proof}
\begin{proof}[Proof of Proposition \ref{Reducability proposition}]
Note that there are no nontrivial relations imposed on $\Mknoti 1$;
hence $\pi|_{\Mknoti 1}$ is injective.
It suffices to prove that $\pi(\Wknoti 1)$ generates $V$.
First we will prove inductively that $\pi(\Mknoti 1)$ contains $\pi(\Wknoti i)$ for each $i$, with base case $i = 1$ trivially verified.
For the case $i \geq 2$ and element $w_1 \in \Wknoti i$, repeated application of Lemma \ref{Pizza inductive lemma} yields a sequence
\[
(w_1,w_2,w_3,\dots)
\]
where $w_j \in \Wknoti i$ for all $j$ and there exists a relation of the form $\pi(w_j) = \frac{\pi(w) + \pi(w_{j+1})}{2}$ where $w \in W_{i-1} \subset \pi(\Mknoti 1)$.
Since $W_i$ is finite, there exist two indices $j < k$ such that $w_j = c \cdot w_k$ for some $c > 1$;
then, there is some $u' \in \Mknoti{i-1}$ such that $\pi(w_j) = \frac{\pi(u')}{c - 1}$ by construction.
The inductive hypothesis then guarantees that $\pi(w_j) \in \pi(\Mknoti 1)$, completing the inductive step.
In particular, the above construction yields the inequality $\vol \pi(w_j) \geq \frac{\vol \pi(u')}{c-1}$, and induction yields that $\vol \pi(w_j)$ is at least the volume of an element of $\Mknoti 1$ with coordinates summing to 1, and hence $\vol \pi(w_j) \geq 2m \cdot \sum_i q_i(w_j) \vol(T_i^{2m})$, completing the inductive step.
Using this, for an arbitrary word $w \in W$, we may pick an arbitrary relation $\pi(w) = \pi(w_1) + \pi(w_2)$, and note that $w_1,w_2 \in \Wknoti m$.
By the above statement, $\pi(w) \in \pi(\Mknoti m) = \pi(\Mknoti 1)$ and $2\vol(w) \geq \vol(w_1) + \vol(w_2) \geq 2m \sum_i q_i(w) \vol(T_i^{2m})$, as desired.
\end{proof}
\begin{proof}[Proof of Theorem \ref{Main theorem separating}]
For hyperbolicity and the volume bound, we simply combine Theorem \ref{Cutting and pasting theorem}, Proposition \ref{Reducability proposition} and Equation \eqref{Evaluation morphism}.
Together these yield hyperbolicity and
\begin{align*}
2m \cdot \vol(M)
&= 2m \cdot \vol(T_1 \cdots T_{2m})\\
&\geq \sum_{i=1}^{2m} \vol^{2m}(M_i).
\end{align*}
We now characterize equality of this bound.
Suppose there is some $\Sigma \in \bS$ that is not totally geodesic.
Then Lemma \ref{Volume relation lemma separating} and Theorem \ref{Cutting and pasting theorem} together imply that
\begin{align*}
2m \cdot \vol(M)
&= 2m \cdot \vol(w)\\
&> m \cdot \prn{\vol(w_1) + \vol(w_2)}\\ &\geq \sum_i \vol^{2m}(M_i)
\end{align*}
where $\Sigma$ splits $w$ into $w_1w_2$.
Suppose conversely that each $\Sigma \in \bS$ is totally geodesic.
By repeatedly cutting along totally geodesic surfaces, then gluing along isometric totally geodesic surfaces, we may directly construct $\coprod_{i=1}^{2m} D^{2m}(M_i)$ from $\coprod_{i=1}^{2m} M$.
By Theorem \ref{Cutting and pasting theorem}, this implies that
\[
\vol(M) = \sum_i \vol^{2m}(M^i)
\]
as desired.
\end{proof}
\subsection{Proof of the main theorem for a separating starburst in an \texorpdfstring{$\ell$}{l}-piece}\label{Separating inductive subsection}
The example from Section \ref{introduction section}, where a given saucer tangle can be decomposed into tetrahedral tangles as in Figure \ref{suacertet2} motivates the following definition.
\begin{definition}
Let $P$ be a 3-manifold with distinguished surfaces $V^1,V^2,\dots,V^{2\ell} \subset \partial P$ such that $V^{i} \cap V^{j}$ is a 1-manifold (possibly with boundary, or possibly empty) for each distinct $1 \leq i,j \leq 2\ell$.
We call such data an \emph{$\ell$-piece}, and we call a 3-manifold with no distinguished surfaces a \emph{0-piece}.
\end{definition}
For an $\ell$-piece, define $R = \overline{\partial P \setminus \cup V_i}$. Given an $\ell$-piece $(P,(V_i)_{i=1}^{2\ell}))$, some index $1 \leq k \leq \ell$, and some positive integer $m$, we may form the 3-manifold $D^{2m}(P,(V^{2k - 1},V^{2k}))$.
For any $j \neq 2k-1,2k$, the points $D^{2m}(V^j,(V^{2k - 1},V^{2k}))$ form a surface contained in $\partial D^{2m}(P,(V^{2k - 1},V^{2k}))$, and the intersections of these surfaces remain 1-manifolds.
Hence the data $\prn{D^{2m}(P,(V^{2k - 1},V^{2k})),\prn{D^{2m}(V^{j},(V^{2k - 1},V^{2k}))}_{j \neq 2k-1,2k}}$ constitute an $(\ell-1)$-piece.
That is, one can $2m$-replicate over one of the pairs of surfaces in an $\ell$-piece, which yields an $(\ell-1)$-piece.
Iteration of this construction yields the notion of an $\bm$-replicant, which we define now
\begin{definition}
We define the $\bm = (2m_1,\dots,2m_\ell)$-replicant of an $\ell$-piece $D^{\bm}(P,\prn{V^1,V^2,\dots,V^{2\ell}})$ inductively as
\[
D^{(2m_1,\dots,2m_\ell)}(P,\prn{V^j}_{j=1}^{2\ell}) := D^{(2m_2,\dots,2m_\ell)}\prn{D^{2m}(P,(V^{2k - 1},V^{2k})),\prn{D^{2m}(V^{j},(V^{2k - 1},V^{2k}))}_{j=3}^{2\ell}}
\]
that is, the $(2m_1,\dots,2m_\ell)$-replicant is the $(2m_2,\dots,2m_{\ell})$-replicant of the $(\ell-1)$-piece resulting from doubling everything across $V^{1}$ and $V^{2}$.
We say that $(P,\prn{V^j})$ is \emph{$\bm$-hyperbolic} if $D^{\bm}(P,\prn{V^j})$ is hyperbolic, in which case we define
\[
\vol^{\bm}(P,\prn{V^j}) := \frac{1}{\prod_{j=1}^\ell 2m_j}\vol(D^{\bm}(P,\prn{V^j}).
\]
Similar to before, if the tuple $\prn{V^j}$ is clear, we often simply refer to the piece as $P$, and refer to its $\bm$-replicant as $D^\bm(P)$ and its $\bm$-volume as $\vol^\bm(P)$.
\end{definition}
The above definition appears to involve a choice of permutation on $\ell$ letters defining the order over which the pairs are ``doubled over.''
However, the following result shows that every such choice yields the same manifold.
We say that an \emph{isomorphism of pieces} is a diffeomorphism of their underlying manifolds which restricts to a diffeomorphism of each surface.
\begin{proposition}\label{Replication commuting proposition}
Let $(P,(V^j))$ be an $\ell$-piece.
For any $1 \leq k \leq \ell$, define the notation
\[
D^{2m}_k(P,(V^j)) := \prn{D^{2m}(P,(V^{2k - 1},V^{2k})),\prn{D^{2m}(V^{j},(V^{2k - 1},V^{2k}))}_{j \neq 2k-1,2k}}.
\]
Then, for any distinct $1 \leq k,k' \leq \ell$ and any positive integers $m$ and $m'$, there is an isomorphism of pieces
\begin{equation}\label{Double replication equation}
D^{2m}_{\widetilde k} D^{2m'}_{k'} (P,(V^j)) \simeq D^{2m'}_{\widetilde k'} D^{2m}_{k} (P,(V^j))
\end{equation}
where $\widetilde k = k + 1$ if $k > k'$ and $k$ otherwise and $\widetilde k' = k' + 1$ if $ k' > k$ and $k'$ otherwise.
In particular, for any $\bm = (2m_1,\dots,2m_\ell)$ and $1 \leq k \leq 2\ell$, we have
\begin{equation}\label{Replication commuting equation}
D^{\bm}\prn{P,(V^j)_{j=1}^{2\ell}} = D^{(2m_1,\dots,\widehat{2m_k},\dots,2m_\ell)}\prn{D^{2m_k}(P,(V^{2k - 1},V^{2k})),\prn{D^{2m_k}(V^{j},(V^{2k - 1},V^{2k}))}_{j \neq 2k-1,2k}}
\end{equation}
where $\widehat .$ denotes exclusion from a list.
\end{proposition}
\begin{proof}
Note that both sides of \eqref{Double replication equation} are expressed as a quotient of the disjoint union of $4mm'$ copies of $P$.
For the left hand side, we first quotient together subspaces of $2m'$ copies of $P$ along $V^{2k' - 1}$ and $V^{2k'}$, then we quotient together such subspaces along the images of $V^{2k-1}$ and $V^{2k}$.
The right hand side is analogous, with $(m',k')$ and $(m,k)$ interchanged.
This may be expressed in both cases as a quotient of the disjoint union of $4mm'$ copies of $P$, arranged in a ``grid,'' where a point is either identified with nothing, identified with a point on the east or west, or identified with a point on the north or south (with combinations of the last two allowed).
These relations are the same in both cases, so the resulting manifolds are diffeomorphic;
furthermore, this diffeomorphism clearly carries the piece surfaces to each other, so the resulting pieces are isomorphic.
We prove Equation \eqref{Replication commuting equation} via a generalization. Each permutation $\pi \in S_\ell$ defines a manifold $D^{\bm}_\pi(P,(V^j))$, where one replicates over the $\pi(1)$ surface first, the $\pi(2)$ second etc.
Equation \eqref{Double replication equation} yields a diffeomorphism $D^{\bm}_\pi(P,(V^j)) \simeq D^{\bm}_{\tau \pi}(P,(V^j))$ for any simple transposition $\tau$.
Since simple transpositions generate $S_\ell$, this implies that $D^{\bm}_\pi(P,(V^j)) \simeq D^{\bm}_{\rho}(P,(V^j))$, for any $\pi,\rho \in S_\ell$, which generalizes Equation \eqref{Replication commuting equation}.
\end{proof}
By similar logic to the $\ell = 1$ case, an $m$-sheeted separating starburst in an $(\ell-1)$-piece $\prn{M,\prn{V^j}_{j=1}^{2(\ell-1)}}$ yields a cycle of $2m$ many $\ell$-pieces $\prn{M_i,\prn{V_i^j}_{j=1}^{2\ell}}$. We define $H$ to be the closure of $\partial M \setminus \cup_{j = 1}^{2(\ell-1)} V^j$. We denote the closure of the complement of $V^1_i \cup V^2_i$ in $\partial M_i$ by $H_i$. Note that $H_i \subset \partial M$.
We generalize Theorem \ref{Main theorem separating} directly to $\ell$-pieces in the following theorem.
\begin{theorem} \label{Main theorem separating inductive}
Let $(M,\bS)$ be an $m$-sheeted separating starburst with all boundary components tori. Suppose that $\prn{M,\prn{V^j}_{j=1}^{2(\ell-1)}}$ is an $(\ell-1)$-piece, and suppose the associated $\ell$-pieces $(M_i,\prn{V_i^j}_{j=1}^{2\ell})$ are each $(2m_1,\dots,2m_\ell)$-hyperbolic and $V^{2j -1}_i \cap V^{2j}_i$ does not intersect $H_i$ for all $j$.
Then, $(M,\prn{V_i}_{i=1}^{2(\ell-1)})$ is $(2m_1,\dots,2m_{\ell-1})$-hyperbolic with volume satisfying
\[
\vol^{(2m_1,\dots,2m_{\ell-1})}\prn{M} \geq \sum_{i=1}^{2m_\ell} \vol^{(2m_1,\dots,2m_\ell)}\prn{M_i}.
\]
This equality is attained if and only if each $\Sigma \in \bS$ is totally geodesic in $M$.
\end{theorem}
\begin{proof}
For brevity, we will write $D^{m_1}(V^j)$ to replace $D^{m_1}(V^j,(V^1,V^2))$.
We prove this theorem inductively. In the case $\ell = 1$,
the result is given by Theorem \ref{Main theorem separating}.
Suppose inductively that Theorem \ref{Main theorem separating inductive} is satisfied for starbursts in a $k$-piece for $k \leq \ell - 2$.
The definition
\[
D^{(2m_1,\dots,2m_\ell)}(P,\prn{V^1,\dots,V^{2\ell}}) = D^{(2m_2,\dots,2m_{\ell})}\prn{D^{2m_1}(P,\prn{V^{1},V^{2}}),\prn{D^{m_1}(V_3),\dots,D^{m_1}(V^{2\ell})}}
\]
and Proposition \ref{Replication commuting proposition} together imply that the $(2m_1,\dots,2m_\ell)$-hyperbolicity of each $\ell$-piece is equivalent to $(2m_2,\dots,2m_\ell)$-hyperbolicity of pieces corresponding with a separating starburst $D^{2m_1}(\bS)$ in the $(\ell-2)$-piece $\prn{D^{2m_1}(M),(D^{m_1}(V_3),\dots,D^{m_1}(V^{2(\ell-1)}))}$.
Our inductive hypothesis then implies that $\prn{D^{2m_1}(M),(D^{m_1}(V_3),\dots,D^{m_1}(V^{2\ell})}$ is $(2m_2,\dots,2m_\ell)$-hyperbolic, or equivalently, that $(M,(V^1,\dots,V^{2\ell}))$ is $\bm$-hyperbolic.
A similar argument yields volume bounds:
\begin{align*}
2m_1 \cdot \vol^{(2m_1,\dots,2m_{\ell-1})}(M,\prn{V^j}_{j=1}^{2(\ell-1)}) &= \vol^{(2m_2,\dots,2m_{\ell-1})}\prn{D^{2m_1}(M,\prn{V^{1},V^{2}}),\prn{V^j}_{j=3}^{2(\ell-1)}}\\
&\geq \sum_i \vol^{(2m_2,\dots,2m_{\ell})}\prn{D^{2m_1}(M_i,\prn{V_i^{1},V_i^{2}}),\prn{D^{m_1}(V_i^j)}_{j=3}^{2\ell}}\\
&= 2m_1 \cdot \sum_i \vol^{(2m_1,\dots,2m_\ell)}(M_i,\prn{V_i^j}_{j=1}^{2\ell})).
\end{align*}
In fact, this equality is attained if and only if equality is attained in the middle inequality above.
By the inductive hypothesis, this is true if and only if each doubled surface in $\bS$ is totally geodesic in $D^{2m_1}(M,(V^1,V^2))$.
This is true if and only if each surface in $\bS$ is totally geodesic in $M$, as desired.
\end{proof}
\section{The main theorem in the general case}\label{Main theorem nonseparating section}
\subsection{Definitions}
In Section \ref{Main theorem separating section}, we defined suitable collections of surfaces, called \emph{$m$-sheeted separating starbursts}, which separate a 3-manifold $M$ into a disjoint union of $2m$ pieces such that hyperbolicity of the pieces implies hyperbolicity of $M$.
In fact, we may summarize this with one datum rather than $2m$:
defining the disjoint union of $\ell$-pieces by
\[
(P,(V^j)) \sqcup (P',(V'^j)) := (P \sqcup P', (V^j \sqcup V'^j)).
\]
For any $\bm \in 2\ZZ^{\ell}_{>0}$, there is a diffeomorphism
\[
D^{\bm}\prn{(P,(V^j)) \sqcup (P',(V'_j))} \simeq D^{\bm}\prn{P,(V^j)} \sqcup D^{\bm}\prn{P',(V^j)}.
\]
In particular, since we have $(M \dbs \bS,(V^1,V^2)) = \coprod_{i=1}^{2m} (M_i,(V_i^1,V_i^2))$ as in Construction \ref{Piece construction}, we may summarize Theorem \ref{Main theorem separating} by stating that $2m$-hyperbolicity of $(M \dbs \bS,(V^1,V^2))$ implies hyperbolicity of $M$, with volume satisfying
\[
\vol(M) \geq \vol^{2m}(M \dbs \bS).
\]
In the case of a starburst in an $(\ell-1)$-piece, we further have that $(2m_1,\dots,2m_\ell)$-hyperbolicity of the $\ell$-piece $\prn{M \dbs \bS, (V^j)_{i=1}^{2\ell}}$ implies $(2m_1,\dots,2m_{\ell-1})$-hyperbolicity of $\prn{M,(V^j)_{i1}^{2(\ell-1)}}$, with volumes satisfying
\[
\vol^{(2m_1,\dots,2m_{\ell-1})}\prn{M} \geq \vol^{(2m_1,\dots,2m_\ell)}\prn{M \dbs \bS}.
\]
This formulation is suggestive;
the contruction of $(M \dbs \bS, (V^j))$ doesn't appear to depend on the fact that $M \dbs \bS$ separates into $M_i$ in an essential way.
To expand on this, we make the following definition, which relaxes orientability of $M$ and the separation of $M$ into a cycle of pieces.
\begin{definition}\label{Starburst definition}
Let $M$ be a 3-manifold and let $\bS = \cbr{\Sigma_1,\dots,\Sigma_m}$ be a set of properly embedded two-sided surfaces in $M$.
Let $E := \bigcup_{i\neq j} \Sigma_i \cap \Sigma_j$ be the locus at which at least two surfaces intersect, and suppose that $E = \Sigma_i \cap \Sigma_j$ for each $i \neq j$.
Let $\rho:\bigcup_i \Sigma_i \dbs E \rightarrow \cbr{+,-}$ be a ``two-coloration'' of the connected components of $\Sigma_i \dbs E$ for each $i$.
We say that the data $(M,\bS,\rho)$ constitute an \emph{n-sheeted starburst in $M$} (often, simply a \emph{starburst}) if the following conditions are met:
\begin{enumerate}
\item the intersection locus $E$ is a compact one-manifold (possibly disconnected and possibly with boundary) properly embedded in $M$;
\item \label{Diffeomorphism condition} there is a neighborhood $E \subset U \subset M$ and a diffeomorphism of pairs $(U,U \cap\prn{\bigcup_{i=1}^m \Sigma_i}) \simeq (D \times E, \bigcup_{i=1}^m L_i \times E)$ which restricts to a diffeomorphism $\Sigma_i \cap U \simeq L_i \times E$ for each $i$, with notation defined in Figure \ref{newstarburstfigure} and \ref{newstarburstfigure 2
;
\item \label{Coloration condition} any two connected components of $\Sigma_i \dbs E$ who neighbor the same points of $E$ in $M$ have opposing colorations via $\rho$;
\item \label{Compatible direction condition} for each connected component of $E$, given the induced orientations of normal bundles and labels of the lines $L_i \subset S_m$, all orientations of normal bundles on a $+$-labelled component of $L_i \setminus E$ share a direction (either clockwise or counterclockwise).
\end{enumerate}
\end{definition}
\begin{remark}
We may formally define "all $+$-labelled components of $L_i \setminus E$ share direction" several equivalent ways;
one convenient one is to note that the subspace of points $d(\mathbf{L})$ of $\partial(D \dbs \bigcup_i L_i)$ corresponding with $\bigcup_i L_i$ have $2n$ connected components, and each surface together with an orientation of its normal bundle determines two connected components of $d(\mathbf{L})$ having either positive orientation and $+$ label or negative orientation and $-$ label, as in Construction \ref{Piece construction}.
One may replace condition \eqref{Compatible direction condition} with the requirement that this correspondence identifies all $2n$ components of $d(\mathbf{L})$, or equivalently, than no $+$ and $-$ labelled components of half-lines identify the same component.
\end{remark}
\begin{figure}
\centering
\includegraphics[scale=0.6]{starburst2.pdf}
\caption{Continuation of Figure \ref{newstarburstfigure}, now using the orientation and labelling data.
Part (c) includes arrows denoting the orientation, and labels denoting $\rho$.
Part (d) shows the parts of surfaces near $E$ comprising $V^1_i$.
}
\label{newstarburstfigure 2}
\end{figure}
For this paragraph, let $(M,\bS)$ be a separating starburst.
Then, by definition, $(M,\bS)$ satisfies the first two conditions for a starburst.
The portion $d(\Sigma_i) \setminus d(E) \subset \partial M$ corresponding with $\Sigma_i$ can be expressed as the union of $V^1_i$, $V^2_{i-1}$, $V^1_{i+n}$, and $V^2_{i + n - 1}$, with the first pair and last pair of surfaces corresponding with a common component of $\Sigma_i \dbs E$.
The parts $V^2_i$ all point \emph{clockwise} local to $E$, so we may choose the labeling $\rho|_{\Sigma_i}:\Sigma_i \dbs E \rightarrow \cbr{+,-}$ so that all of $V^2_i$ corresponds with either positive orientation and $+$-labelling or negative orientation and $-$-labelling;
we may further choose labelling so that $V^1_i$ corresponds with either positive orientation and $-$-labeling or negative orientation and $+$-labelling.
Conditions \eqref{Coloration condition} and \eqref{Compatible direction condition} then follow from the fact that $\bS$ separates $M$ into $M_i$.
Hence separating starbursts are themselves starbursts, and we may reverse-engineer the notion of \emph{clockwise} or \emph{clockwise side} from orientations and labelling.
We will use this to extend the piece structure from separating starbursts to all starbursts.
Henceforth let $(M,\bS)$ be an $m$-sheeted starburst.
We now explicitly construct surfaces rendering $M \dbs \bS$ to be a piece, generalizing the piece structure on each $M_i$ in the separating case.
In the same sense that the outer boundary of the saucer containing a tangle is separated by copies of $E$ into a \emph{clockwise} and \emph{counterclockwise side}, we define the piece structure by using the 2-coloration and orientation of the normal bundles of $\Sigma_i$ to choose surfaces which, local to the points corresponding with $E$, appear to distinguish the \emph{clockwise} and \emph{counterclockwise side} as well.
\begin{construction}\label{Piece construction}
For $\Sigma_i \in \bS$, let $d(\Sigma_i)$ be the union of the points in $\partial (M \dbs \bS)$ which correspond with $\Sigma_i$, and let $d(\bS)$ be the union of the $d(\Sigma_i)$. Let $d(E)$ be defined similarly for $E$.
Every point of $d(\Sigma_i) \setminus d(E)$ corresponds with a point in $\Sigma_i \setminus E$ and for a point in $\Sigma_i \setminus E$, there are two points in $d(\Sigma_i) \setminus d(E)$ which correspond to it.
One comes about as a boundary point on the positively oriented side of a normal bundle on $\Sigma_i$ and one comes about from the negatively oriented side, so we may accordingly call a point in $d(\Sigma_i) \setminus d(E)$ positive or negative.
By nature of the orientation, if two points in $d(\Sigma_i) \setminus d(E)$ reside in the same connected component, they must either both be positive or both be negative.
Hence we may refer to a positively-oriented component of $d(\Sigma_i) \setminus d(E)$ as one whose points are positive away from $d(E)$, and a negatively-oriented component otherwise.
Similarly, every component of $d(\Sigma) \setminus d(E)$ has points corresponding with a unique component of $\Sigma_i \dbs E$, so there is a two-coloration $\bigcup_i d(\Sigma_i) \setminus d(E) \rightarrow \cbr{+,-}$ induced by $\rho$, and we will also call this $\rho$.
We say that a component of $d(\bS)\setminus d(E)$ (that is, a component of some $d(\Sigma_i) \setminus d(E)$) as \emph{counter-clockwise-pointing} if it is either $+$-labeled and positively-oriented or $-$-labelled and negatively-oriented, and \emph{clockwise-pointing} otherwise.
Let $V_i^2$ be the union of $E$ and the clockwise-pointing components of $d(\Sigma_i) \setminus d(E)$, and let $V_i^1$ be the union of $E$ and the counterclockwise-pointing components of $d(\Sigma_i) \setminus d(E)$.
Let $V^j := \bigcup_i V_i^j$ be the union of the clockwise-pointing or counterclockwise-pointing components.
We visualize this via Figure \ref{newstarburstfigure 2}.
By condition \eqref{Compatible direction condition}, each component of $d(E)$ separates a clockwise-pointing component and a counterclockwise-pointing component;
hence $V^1 \cap V^2 = d(E) = \bigsqcup_{i=1}^{2m} E$, which is a 1-manifold.
\end{construction}
\begin{remark}
In the case that every connected component of $\Sigma_i$ intersects $E$, note that every allowable choice of orientations of normal bundles and labellings identifies the same piece structure on $M \dbs \bS$;
in general, varying orientations and labellings allows one to construct up to $2^C$ many isomorphism classes of piece structures, where $C$ is the number of connected components of $\bigcup_i \Sigma_i$ not intersecting $E$.
\end{remark}
This construction yields a piece structure $(M \dbs \bS, (V^1, V^2))$.
Hence there is a notion of replication and $2m$-hyperbolicity of a starburst, and we may ask whether generalizations of Theorem \ref{Main theorem separating} applies in this case. But we would like to allow the intersection set $E$ of the surfaces to intersect the boundary of the manifold. This is the topic of the next subsection.
\subsection{Boundary compatibility and the torus boundary case}
In Theorem \ref{Main theorem separating}, we assumed that the manifold $M$ had only torus boundary and $E$ did not intersect the boundary. We shortly generalize these conditions. In the case that $E$ intersects a torus boundary of $M$, let $G$ be the graph on $T$ formed by the intersection of $\bigcup_i \partial \Sigma_i$ with $T$. We prove the following result.
\begin{lemma}\label{Eintersectstorusboundary}
If $(M, \bS)$ is an $m$-sheeted separating starburst, $E$ intersects a torus boundary $T$ of $M$, and the $2m$-replicants of the corresponding pieces are all hyperbolic, then $G$ is connected, $m$ = 2, and all of the complementary regions to $G$ on $T$ are squares.
\end{lemma}
\begin{proof}
Note that every surface $\Sigma_i$ must touch every vertex of the graph. Also, hyperbolicity of the replicants means that the replicant cannot have spherical boundaries. This would, for instance occur, if any complementary region of $G$ on $T$ is a bigon.
We first consider the simple closed curves that make up $\Sigma_i \cap T$ for a fixed value of $i$. None can be trivial curves on $T$ as follows. If there were trivial curves, take an innermost such. Let $D$ be the disk it bounds on $T$. So $D$ has no other curves from $\Sigma _i$ inside it. Therefore it has no vertices in its interior, since every vertex must touch a curve from $\Sigma_i$. If $D$ has vertices on its boundary, there are curves from each of the other surfaces that cross it. But that there are no vertices inside $D$ implies that we can take an outermost arc from the other surfaces inside $D$ and it forms a bigon with an arc on the boundary of $D$, which will generate a sphere in the boundary of the corresponding replicant, a contradiction to its hyperbolicity.
On the other hand, suppose there are no vertices on $\partial D$. Then when we form $D^{2m} (M_i)$ for the piece containing $D$ on its boundary, we again obtain a sphere boundary, contradicting hyperbolicity of the replicants.
Hence, for all $i$, every curve in $\Sigma _i \cap T$ must be nontrivial on $T$. So for each fixed $i$, the curves in $\Sigma_i \cap T$ must all be parallel $(p_i, q_i)$-curves. Further, the only options for complementary regions to $G$ are disks and annuli.
If there is an annulus in $M_i \cap T$, take its boundary to be meridians for convenience. Then for all $i$ and all of the curves in $\Sigma_i \cap T$, $q_i = 0$. That is to say, for all $i$, all simple closed curves in $\Sigma_i\cap T$ are meridians.
Consider the first two surfaces $\Sigma_1$ and $\Sigma_2$. The fact that all curves in their intersections with $T$ are meridians implies that any two that cross must do so an even number of times. Further, for some pair of curves $\gamma_1$ in $\partial \Sigma_1$ and $\gamma_2$ in $\partial \Sigma_2$, there must be an arc $\alpha_1$ on $\gamma_1$ and an arc $\alpha_2$ on $\gamma_2$ such that they form a bigon with sides that are not crossed by any other curve from $\partial \Sigma_1$ or $\partial \Sigma_2$. Additional arcs from other surfaces could enter the bigon through one of the vertices. But since there can be no vertices inside the bigon as there are no curves from $\partial \Sigma_1$ or $\partial \Sigma_2$ inside it, any such arc must just exit out the opposite vertex. Thus, they decompose the bigon into thinner bigons, which again lift to spheres in the corresponding replicants, contradicting their hyperbolicity.
So there are no annuli.
We now show that $(p_i, q_i) \neq (p_j, q_j)$ for $i \neq j$. The proof is the same argument as above. Namely if curves from $\partial \Sigma_1$ and $\partial \Sigma_2$ on $T$ are parallel, they must cross, and then we can find an innermost bigon between the two types of curves and there can be no other boundary curves of $\Sigma_1$ or $\Sigma_2$ inside the bigon formed. So only curves from one vertex to the other can appear in the bigon, which makes a thinner bigon, a contradiction.
So now if we consider a $(p_1, q_1)$-curve $\gamma_1$ and a $(p_2, q_2)$-curve $\gamma_2$ from two different surfaces, they must cross. But in fact they must cross a minimal amount of times, namely $|p_1 q_2 - p_2 q_1|$ times, as if not, there will be a bigon formed between them with no other curves from these two surfaces in the bigon and the same argument shows there is a bigon.
So we have collections of parallel curves for each surface. We can always choose one set of them to be thought of as meridians. Then if there are only two surfaces, they decompose the torus into squares and $m = 2$. Then all replicants have torus boundaries corresponding to $T$.
If there are three surfaces, the first two surface boundaries decompose the torus into squares. The third surface boundaries must intersect the first two only at their intersections. Since they cannot generate bigons, they must cut diagonally across the squares.
Note that if there were a fourth surface, it would also have to have its boundary curves on $T$ intersecting only at the existing vertices and this would force bigons. Thus, $m = 2$ or 3.
In the case $m =3$, the torus is decomposed into triangles by the graph $G$. However, there can be no odd cycles in $G$. This is because the cyclic order of pieces around a vertex must be preserved. Hence, for any edge in $G$, the cyclic orders must be clockwise around one and counterclockwise around the other. In particular, this means $G$ must be a bipartite graph, with all even-length cycles.
\end{proof}
In particular, this lemma ensures that if $E$ intersects a torus boundary of $M$, then if the replicants are hyperbolic, as we assume in Theorem \ref{Main theorem separating}, the boundaries of the replicants of the corresponding pieces are also tori.
We would like to generalize Theorems \ref{Main theorem separating} and \ref{Main theorem separating inductive} to manifolds with higher genus boundary. Such a manifold is said to be tg-hyperbolic if when all torus boundaries are removed, it has a hyperbolic metric such that its higher genus boundaries are totally geodesic. This is equivalent to the fact that if the manifold is doubled across its higher genus boundaries, and the torus boundaries are again removed, the resulting manifold is hyperbolic. A tg-hyperbolic manifold has a well-defined finite hyperbolic volume.
But in order to generalize to this situation, in the case that $E$ intersects the higher genus boundary components, we need to know that the $2m$-replicants of the corresponding pieces that intersect those boundaries also have corresponding genus greater than 1. Hence we have the following definition.
\begin{definition}Any connected sub-surface $Q_i$ of the boundary of $M_i$ that lies in the boundary of $M$ will correspond to boundary-components of $D^{2m}(M_i,(V^1_i, V^2_i))$. We define a piece $(M_i,(V^1_i, V^2_i))$ to be {\bf boundary-compatible} with the manifold $M$ from which it comes if for any connected components of $\partial M_i \subset \partial M$, the corresponding boundary components of $D^{2m}(M_i,(V^1_i, V^2_i))$ have genus one if and only if the corresponding boundary components of $M$ have genus one. Note that all of the boundary-components of $D^{2m}(M_i,(V^1_i, V^2_i))$ coming from a particular connected component $Q_i$ of $\partial M_i$ will have the same genus.
\end{definition}
\subsection{Statement of the General Theorems}
We are now ready to state the generalization of Theorem \ref{Main theorem separating}.
We prove the following.
\begin{theorem}\label{Main theorem}
Suppose $(M,\bS,\rho)$ is an $m$-sheeted starburst in a 3-manifold $M$, and $(M \dbs \bS,(V^1,V^2))$ is the associated piece.
If $(M \dbs \bS, (V^1,V^2))$ are $2m$-hyperbolic and boundary-compatible with $M$, then $M$ is hyperbolic, with volume satisfying
\[
\vol(M) \geq \vol^{2m}(M \dbs \bS).
\]
Equality is attained if and only if each surface $\Sigma_i \subset M$ is totally geodesic.
\end{theorem}
Note that Lemma \ref{Eintersectstorusboundary} implies that if $m > 2$, it must be the case that $E$ intersects only higher genus components of $\partial M$ to obtain hyperbolicity of the pieces.
Similarly to the separating case, a starburst in an $(\ell-1)$-piece $(M,(V^j)_{i=1}^{2(\ell-1)})$ yields an $\ell$-piece $(M \dbs \bS, (V^j)_{i=1}^{2\ell})$, and there is an associated notion of $\bm$-hyperbolicity.
We will prove the following generalization of Theorem \ref{Main theorem separating inductive} in this case.
\begin{theorem}\label{Main theorem inductive}
Suppose that $\prn{M,\prn{V^j}_{j=1}^{2(\ell-1)}}$ is an $(\ell-1)$-piece, suppose $(M,\bS,\rho)$ is an $m$-sheeted, boundary-compatible starburst, and suppose the associated $\ell$-piece $(M \dbs \bS,\prn{V^j}_{j=1}^{2\ell})$ is $(2m_1,\dots,2m_\ell)$-hyperbolic.
Then, $(M,\prn{V^j}_{j=1}^{2(\ell-1)})$ is $(2m_1,\dots,2m_{\ell-1})$-hyperbolic with volume satisfying
\[
\vol^{(2m_1,\dots,2m_{\ell-1})}\prn{M} \geq \vol^{(2m_1,\dots,2m_\ell)}\prn{M \dbs \bS}.
\]
This equality is attained if and only if each $\Sigma \in \bS$ is totally geodesic in $M$.
\end{theorem}
In Subsection \ref{Non-orientable subsection}, we reduce Theorem \ref{Main theorem} to the case that $M$ is orientable.
Subsection \ref{Non-orientable subsection} can be skipped by the reader only interested in working with orientable 3-manifolds.
Then, in Subsection \ref{secondproofsubsection} we give an alternative proof of Theorem \ref{Main theorem separating} which directly uses the piece structure on $M \dbs \bS$, rather than individual $M_i$.
We use this as motivation to prove Theorems \ref{Main theorem} and \ref{Main theorem inductive} in Subsection \ref{Nonseparating subsection}.
\subsection{Reduction of Theorem \ref{Main theorem} to the orientable case}\label{Non-orientable subsection}
In general, we denote the orientable double cover of $M$ as $M^*$, with map $\pi_M:M^* \rightarrow M$. In the case $M$ is already orientable, $M^*$ represents the double cover obtained by taking two copies of
$M$.
\begin{proposition}\label{Orientable double cover proposition}
Suppose $(M,\bS)$ is an $m$-sheeted starburst.
\begin{enumerate}
\item Let $\Sigma_i^* := \pi_{M}^{-1}(\Sigma_i)$, let $E^* := \pi_M^{-1}(E)$, and let $\bS^* := \cbr{\Sigma^*_i}$.
Define labels $\rho^*:\bigcup_i \Sigma^*_i \dbs E^* \rightarrow \cbr{+,-}$ by the composition
\[
\bigcup_i \Sigma^*_i \dbs E^* \xrightarrow \pi \bigcup_i \Sigma_i \dbs E \xrightarrow{\rho} \cbr{+,-}.
\]
Then, $(M^*,\bS^*,\cbr{\rho^*})$ constitute an $m$-sheeted starburst.
\item Cutting and $2m$-replication commute with taking the orientable double cover;
in particular, letting $(M^* \dbs \bS^*,(V^{1*},V^{2*}))$ be the piece on the orientable double cover, we have
\[
D^{2m}(M \dbs \bS,(V^1,V^2))^* = D^{2m}(M^* \dbs \bS^*,(V^{1*},V^{2*})).
\]
\end{enumerate}
\end{proposition}
For part (1), our argument will extend naturally to say that, for an $m$-sheeted starburst $(M,\bS,\cbr{\rho})$ and a finite covering map $\pi:M' \rightarrow M$, the set of preimages $\pi^{-1}(\bS) := \cbr{\pi^{-1}(\Sigma_i) \mid \Sigma_i \in \bS}$ endowed with labels given by composition
\[
\begin{tikzcd}
\rho \pi : \bigcup_i \pi^{-1}(\Sigma_i) \dbs \pi^{-1}(E) \longrightarrow[r,"\pi"]
& \bigcup_i \Sigma_i \dbs E \longrightarrow[r,"\rho"]
& \cbr{+,-}
\end{tikzcd}
\]
constitute an $m$-sheeted starburst in $M'$.
Verifying this will consist mostly of noting that each condition in Definition \ref{Starburst definition} can be verified locally, sometimes verifying compatibility of local data between certain pairs of points.
\begin{proof}[Proof of Proposition \ref{Orientable double cover proposition}]
\textit{Part (1).}
Each $\Sigma_i^*$ is a properly embedded surface in $M^*$.
A trivialization of the normal bundle of $\Sigma_i$ lifts to a trivialization of the normal bundle of $\Sigma_i^*$, so each $\Sigma_i^*$ is a two-sided surface.
The intersection locus $E^* = \bigcup_{i \neq j} \Sigma_i^* \cap \Sigma_j^*$ satisfies $E^* = \pi^{-1}(\Sigma_i \cap \Sigma_j) = \Sigma_i^* \cap \Sigma_j^*$ for any $i \neq j$.
Furthermore, $E^*$ is the preimage of a 1-manifold $E$ under a covering map, so it is a 1-manifold.
The existence of a diffeomorphism $(U,U \cap (\bigcup_i \Sigma_i)) \simeq (D \times E, S_m \times E)$ respecting the cyclic order is equivalent to existence, for each point $p \in E$, of a neighborhood $U_p$ of $p$ and a diffeomorphism $(U_p,U_p \cap (\bigcup_i \Sigma_i)) \simeq (D \times I, S_m \times I)$ respecting the cyclic order of $\bS$ and $S_m$ for each $p$.
For each $p^* \in E^*$, composing with a local diffeomorphism $U_{p^*} \xrightarrow{\sim} \pi U_{p^*}$ of a neighborhood $U_{p^*}$ of $p^*$ yields such a diffeomorphism.
Similarly, compatibility of the direction induced by orientations of the normal bundles of $\Sigma_i$ and by $\rho$ can be verified local to each point of $E$;
since the orientation and labeling are compatible with a local diffeomorphism $\pi$, this compatibility on $(M^*,\bS^*,\cbr{\rho^*})$ follows from compatibility on $(M,\bS,\cbr{\rho^*})$.
Hence we have defined an $m$-sheeted starburst in $M^*$.
\textit{Part (2).}
We may characterize the cover $(M^* \dbs \bS^*) \rightarrow M \dbs \bS$ on an open set $(M^* \setminus \bS^*) \rightarrow M \setminus \bS$.
The latter is the restriction of the orientable double cover of $M$ onto an open set, and hence an immersed submanifold of codimension 0.
Hence it is the orientable double cover, and we have $(M^* \dbs \bS^*) \simeq (M \dbs \bS)^*$.
The orientable double cover is compatible with disjoint unions, so the induced cover on $\coprod_{i=1}^{2m} (M \dbs \bS)$ is the orientable double cover.
The covering translations permute fibers, and the surfaces $V^{j*}$ are given by the preimages of $V^{j*}$, with the identifications defining $D^{2m}(M^* \dbs \bS^*,(V^{1*},V^{2*}))$ defined fiberwise;
hence the orientable double cover on $\coprod_{i=1}^{2m}(M \dbs \bS)$ descends to the orientable double cover on $D^{2m}(M \dbs \bS, (V^{1*},V^{2*}))$, i.e. the cover $D^{2m}(M^* \dbs \bS^*,(V^{1*},V^{2*})) \rightarrow D^{2m}(M \dbs \bS, (V^1,V^2))$ is isomorphic to the orientable double cover, and in particular we have an diffeomorphism of their total spaces, as we wanted to show.
\end{proof}
By this proposition, we are able to reduce to the orientable case by passing to the orientable double cover. Hyperbolicity of $D^{2m}(M \dbs \bS,(V^1,V^2))$ implies hyperbolicity of $D^{2m}(M^* \dbs \bS^*,(V^{1*},V^{2*}))$, which implies hyperbolicity of $M^*$, which implies hyperbolicity of $M$.
Additionally, we have
\[
\vol(M) = \frac{1}{2}\vol(M^*) \geq \frac{1}{2}\vol^{2m}(M^* \dbs \bS^*,(V^{1*},V^{2*})) = \vol^{2m}(M \dbs \bS, (V^1,V^2)).
\]
with equality attained if and only if each element of $\bS^*$ is totally geodesic in $M^*$, which occurs if and only if each element of $\bS$ is totally geodesic in $M$.
\subsection{A second proof of Theorem \ref{Main theorem separating} (the main theorem for a separating starburst)}\label{secondproofsubsection}
For the duration of this subsection and Subsection \ref{Nonseparating subsection} we fix the notation $\bm = (2m_1,\dots,2m_\ell)$, $(M,\bS)$, and $(V^j)_{i=1}^{2\ell}$ as in Theorem \ref{Main theorem}, and we assume that $\prn{M \dbs \bS,(V^j)_{i=1}^{2\ell}}$ is $\bm$-hyperbolic.
This subsection focuses mainly on giving an analog of the proof given for Theorem \ref{Main theorem separating} without constructions that essentially rely on the separating condition.
This will provide motivation for the proof of Theorem \ref{Main theorem} in Subsection \ref{Nonseparating subsection}. However, the proof of Theorem \ref{Main theorem} is independent of this subsection.
Recall the set $W$ of cyclic words on alphabet $\cbr{T_i}$ parameterizing the bracelet links formed by the tangles $\cT_i$ of a generic bracelet link on $2m$ tangles, as defined in Construction \ref{Cyclic word construction}.
We had previously defined a set of manifolds $\cbr{\cS(w) \mid w \in W}$ as the topological foundation of our strategy.
However, the manifold $\cS(w)$ consisted of individual pieces $M_i$, and hence it cannot directly generalize to $M \dbs \bS$.
Intuitively, we need to \emph{fill in} the remaining $(M_j)_{i \neq j}$ corresponding with each copy of $M_i$ in $\cS(w)$;
in the separating case, this can be done simply by defining the manifolds
\begin{equation}\label{Symmetric manifold equation}
\cSc\prn{T_{i_1} \cdots T_{i_{2m}}} := \coprod_{j=1}^{2m} \cS\prn{T_{i_1 + j} \cdots T_{i_{2m} + j}}
\end{equation}
and extending this to $\ZZ_{\geq 0}[W]$ by $\cSc(\sum_i w_i) := \coprod_i \cSc(w_i)$.
Our manifolds are grounded to $M$ and $D^{2m}(M \dbs \bS,(V^1,V^2))$ by the following identifications:
\begin{align*}
\cSc\prn{T_1 \cdots T_{2m}} &= 2m \cdot \cS\prn{T_1 \cdots T_{2m}}\\
&= \coprod_{i=1}^{2m} M\\
\cSc(T_1^{2m})
&= \coprod_{i=1}^{2m} \cS(T_i^{2m})\\
&= \coprod_{i=1}^{2m} D^{2m}(M_i,(V^1_i,V^2_i))\\
&= D^{2m}(M \dbs \bS, (V^1,V^2)).
\end{align*}
Furthermore, there is a natural notion of volume
\[
\volc(w) := \vol \cSc(w)
\]
with the volume of non-hyperbolic manifolds equal to 0.
It suffices to prove that $\volc\prn{T_1 \cdots T_{2m}} \geq \volc\prn{T_1^{2m}}$.
In fact, the relation
\[
2\volc(a_1a_2) \geq \volc(a_1a_1^R) + \volc(a_2a_2^R)
\]
follows from the same relation for the volume of each summand in \eqref{Symmetric manifold equation}.
This together with the algebra of Subsection \ref{Separating subsection} immediately imply Theorem \ref{Main theorem separating}.
\subsection{Proof of the main theorem in the general case}\label{Nonseparating subsection}
Using the results of Subsection \ref{Non-orientable subsection}, we henceforth assume that $M$ is orientable.
By boundary compatibility, we may double across higher genus boundary, which takes the original starburst to another starburst in the resulting manifold, and henceforth assume that $M$ has torus boundary.
As exemplified in Subsection \ref{secondproofsubsection}, we may prove Theorem \ref{Main theorem} via the following strategy:
\begin{enumerate}
\item First, we define a set of manifolds $\cbr{\cSc(w) \mid w \in W}$ such that $\cSc(T_i^{2m}) = D^{2m}(M \dbs \bS, (V^1, V^2))$ and $\cSc(T_1 \cdots T_{2m})$ is the total space of a $2m$-fold cover of $M$.
\item Next, we prove the relation $2\vol \cSc(a_1a_2) \geq \vol \cSc(a_1a_1^R) + \vol \cSc(a_2a_2^R)$ holds in this case.
\item Last, we use the algebraic results given in Subsection \ref{Separating subsection} to prove the theorem.
\end{enumerate}
We recall some notation for the purpose of defining $\widetilde S(w)$.
Recall that the set $d(\Sigma_i) \subset \partial M \dbs \bS$ consists of the points corresponding with $\Sigma_i$;
for each $i$ we write $V_i^2$ for the clockwise side of the points of $d(\Sigma_i)$ and $V^{i}_1$ for the counter-clockwise side of the points of $d(\Sigma_{i+1})$.
We have $V^j = \bigsqcup_i V^j_i.$
In this language, we can describe the $2m$-replicant as
\[
D^{2m}(M \dbs \bS, (V^1,V^2)) = \frac{\coprod_{k=1}^{2m} (M \dbs \bS)}{\prn{V_i^{1},2k} \sim \prn{V_i^{2},2k + 1} \hspace{10pt} \text{and} \hspace{10pt} \prn{V_i^{2},2k - 1} \sim \prn{V_i^{1},2k}}
\]
We will extend this to arbitrary elements of $\widetilde W$ in the following construction.
\begin{construction}
Setting $\tilde w = t_1 \cdots t_{2m}$, and fixing (arbitrarily) a choice of first letter, define
\begin{align*}
R^1(k,\tilde w) &:= \begin{cases}
1 & \text{if}\hspace{.1cm} t_k = T_j \text{ for some } j\\
2 & \text{if}\hspace{.1cm} t_k = T_j^R \text{ for some } j\\
\end{cases}\\
R^2(k,\tilde w) &:= \begin{cases}
2 & \text{if}\hspace{.1cm} t_k = T_j \text{ for some } j\\
1 & \text{if}\hspace{.1cm} t_k = T_j^R \text{ for some } j\\
\end{cases}\\
\text{if} \hspace{10pt} t_k \in \cbr{T_j,T_j^R} \hspace{10pt} \text{and} \hspace{10pt} t_{k+1} \in \cbr{T_{j'},T_{j'}^R} \hspace{60pt} \Delta(k,\tilde w) &:= j' - j
\end{align*}
Using this language, we may define the following:
\[
\cSc(\tilde w) := \frac{\coprod_{k=1}^{2m} (M \dbs \bS)}{\prn{V_i^{R^1(2k,\tilde w)},2k} \sim \prn{V_{i + \Delta(2k,\tilde w)}^{R^2(2k+1,\tilde w)},2k + 1} \hspace{10pt} \text{and} \hspace{10pt} \prn{V_i^{R^2(2k - 1,\tilde w)},2k - 1} \sim \prn{V_{i + \Delta(2k - 1,\tilde w)}^{R^1(2k,\tilde w)},2k}}
\]
Note that there exists a diffeomorphism $\cSc(T_1T_2) \simeq \cSc(T_1^RT_2^R)$, and hence this descends to a construction of a $W$-indexed collection of manifolds.
\end{construction}
This generalizes the construction of $\cSc(w)$ in the separating case;
for instance, in that case the identification
\[
\prn{V_i^{R_1(2k,w)},2k} \sim \prn{V_{i+\Delta(2k,w)}^{R_2(2k+1,w)},2k+1}
\]
represents the gluing of the piece at index $2k$ to the piece at index $2k + 1$ within the submanifold $\cS(t_{j_1 + i} \cdots t_{j_{2m} + i}) \subset \cSc(w)$.
The reader may verify that, for each tuple $(i,j,k)$, there is a unique tuple $(i',j',k')$ such that there exists an identification $(V_i^j,k) \sim (V_{i'}^{j'},k')$ in $\cSc(w)$.
Now that we've defined $\widetilde S(w)$, we verify its relationship with $(M,\bS)$ in the following lemma.
\begin{lemma}\label{Pieces and manifold lemma}
Let $(M,\bS,\cbr{\rho})$ be an $m$-sheeted starburst, and let $\cSc(w)$ be the corresponding word for each $w \in W$.
\begin{enumerate}
\item For each $j$ there is a diffeomorphism
\[
\cSc\prn{T_j^{2m}} \simeq D^{2m}(M \dbs \bS)
\]
so that $\cSc\prn{T_j^{2m}}$ is hyperbolic if and only if $M \dbs \bS$ is $2m$-hyperbolic, in which case we have
\[
\volc\prn{T_j^{2m}} = 2m \cdot \vol^{2m}(M \dbs \bS).
\]
\item There is a $2m$-fold covering map
\[
\cSc\prn{T_1 \cdots T_{2m}} \rightarrow M
\]
so that $\cSc\prn{T_1 \cdots T_{2m}}$ is hyperbolic if and only if $M$ is hyperbolic, in which case we have
\[
\volc\prn{T_1 \cdots T_{2m}} = 2m \cdot \vol(M).
\]
\end{enumerate}
\end{lemma}
\begin{proof}
Part (1) follows by noting that
\[
\cSc(T_j^{2m}) = \frac{\coprod_{k=1}^{2m} (M \dbs \bS)}{\prn{V_i^{1},2k} \sim \prn{V_i^{2},2k + 1} \hspace{10pt} \text{and} \hspace{10pt} \prn{V_i^{2},2k -1} \sim \prn{V_i^{1},2k}} = D^{2m}(M \dbs \bS,(V^1,V^2))
\]
For part (2), note that
\[
\cSc(T_1 \cdots T_{2m}) = \frac{\coprod_{k=1}^{2m} (M \dbs \bS)}{\prn{V_i^{2},k} \sim \prn{V^{i + 1}_{1},k+1}}
\]
has points either given by a point in the interior of $M \dbs \bS$, by an equivalence class of two points $(p,k)$ and $(p',k')$ where $p$ and $p'$ are points of $V_i^2$ and $V_{i+1}^1$ who correspond with the same point of $\Sigma_{i+1}$, or by an equivalence class of $2m$ points who correspond with the same point of $E$.
This correspondence outlines a map
\[
\cSc(T_1 \cdots T_{2m}) \rightarrow M.
\]
We will exhibit this map as a covering map, with local trivialization at points in the interior of $M \dbs \bS$ by an $M$-neighborhood which doesn't intersect any element of $\bS$.
For points on some $\Sigma_i \setminus E$, we may choose the trivializing neighborhood to be the union of two ``half-balls'' in two different copies of $M \dbs \bS$ in $\cSc(T_1 \cdots T_{2m})$.
For points on $E$, we may choose the trivializing neighborhood to be the union of $2m$ ``slices'' in the $2m$ copies of $M \dbs \bS$ in $\cSc(T_1 \cdots T_{2m})$.
Note that the fiber of each point $p \in M \dbs \bS$ is given by $\cbr{(p,k) \mid 1 \leq k \leq 2m}$.
Hence the fibers of this covering map have size $2m$, as desired.
\end{proof}
As outlined in our strategy, we now prove the following proposition, implementing the relations used in the proof of Theorem \ref{Main theorem separating}.
\begin{lemma}\label{Volume relation lemma}
Suppose that $w = a_1a_2 \in W$ where $\operatorname{len}(a_1) = m$, and define $w_j := a_ja_j^R$.
If $\cSc(w_1)$ an $\cSc(w_2)$ are hyperbolic, then $\cSc(w)$ is hyperbolic, with volume satisfying
\[
2\volc(w) \geq \volc\prn{w_1 + w_2}.
\]
\end{lemma}
\begin{proof}
Assume without loss of generality that $\cSc(w_1)$ was constructed fixing the first letter of $a_1$ as the first letter of the word.
Then, there is a diffeomorphism of $\coprod_{k=1}^{2m} M \dbs \bS$ sending $\prn{M \dbs \bS,k} \rightarrow \prn{M \dbs \bS, 2m - k}$.
For a general word $v$ this descends to a diffeormophism $\cSc(v) \rightarrow \cSc(v^R)$;
in the case of a palindromic word, this descends to a diffeomorphism of $\cSc(v)$.
In particular, the fixed point set of this diffeomorphism is the surface
\[
F_j := \bigcup_k \prn{V_1^{R_1(1,w)} \cup V_m^{R_2(m,w)},k} \subset \cSc(w_j).
\]
By Theorem \ref{fixed surface lemma}, $F_j$ is totally geodesic;
since $a_1a_2 \in W$, the surfaces $F_1$ and $F_2$ are diffeomorphic.
Hence we may form the 3-manifold $X$ by gluing together the $\cSc(a_ja_j^R) \dbs F_j$ along a diffeomorphism of the points corresponding to $F_j$ induced by the diffeomorphism $F_1 \simeq F_2$.
Then, by Theorem \ref{Cutting and pasting theorem}, $X$ is hyperbolic, with volume satisfying
\[
\vol(X) \geq \volc(w_1 + w_2).
\]
To prove the lemma, it suffices to supply a double cover $p:X \rightarrow \cSc(w)$.
This may be defined by noting that there exists a map $\cSc(w_1) \dbs F_1 \sqcup \cSc(w_2) \dbs F_2 \rightarrow \cSc(w)$ identifying the points corresponding with word $a_j$ in the domain with the same points corresponding with $a_j$ in the codomain;
two points in the domain are identified by the diffeomorphism of $F_1$ and $F_2$ only if they are mapped to the same point of $\cSc(w)$, so this descends to a map $p:X \rightarrow \cSc(w)$.
We verify that $p$ is a covering map.
For a point contained within the points corresponding with subword $a_j$ and away from the image of $F_j$ in $\cSc(w)$, we may choose a trivializing neighborhood which is contained in the points corresponding with $a_j$.
For a point on the image of $F_j$, we may choose the neighborhood composed of the union of ``half-balls'' in the points corresponding with each $a_1$ and $a_2$.
Hence this is a covering map.
It is clear that this map has fibers of size 2, and hence $\cSc(w)$ is hyperbolic with volume satisfying
\[
2 \cdot \volc(w) = \vol(X) \geq \volc(w_1 + w_2),
\]
as desired.
\end{proof}
\begin{proof}[Proof of Theorem \ref{Main theorem}]
Given Lemmas \ref{Pieces and manifold lemma} and \ref{Volume relation lemma}, the proof of Proposition \ref{Reducability proposition} yields hyperbolicity of $M$ and the inequality
\begin{align*}
2m \cdot \vol(M)
&= \volc(T_1 \cdots T_{2m})\\
&\geq \frac{1}{2m} \sum_{i=1}^m \volc(T_i)\\
&= 2m \cdot \vol^{2m}(M \dbs \bS).
\end{align*}
We now characterize when this equality is attained.
Let $a_1 := T_1 \cdots T_{m}$ and let $a_2 := T_{m+1} \cdots T_{2m}$.
Then, in the notation of Lemma \ref{Volume relation equation}, the image of the surfaces $F^1_j \subset \cSc(a_ja_j^R)$ in $\cSc(T_1 \cdots T_{2m})$ is given by the preimage of $\Sigma_1$ under the covering map $\cSc(T_1 \cdots T_{2m}) \rightarrow M$.
Call this surface $F^1$, and note that $F^1$ is totally geodesic if and only if $\Sigma_1$ is totally geodesic;
this generalizes immediately to $\Sigma_j$, and let the associated surface be $F^j$.
Suppose first that some $\Sigma_j$ is not totally geodesic.
Without loss of generality, we may fix $j = 1$.
Then, note that $F^1 \subset \cSc(T_1 \cdots T_{2m})$ is not totally geodesic, so its preimage in the total space of the double cover $X \rightarrow \cSc(T_1 \cdots T_{2m})$ is not totally geodesic either.
Then, since $X$ is formed by cutting and pasting $\cSc(a_ja_j^R)$ along diffeomorphisms of totally geodesic surfaces, Theorem \ref{Cutting and pasting theorem} implies that $\vol(X) > \volc(a_1a_1^R) + \volc(a_2a_2^R)$.
This in turn implies
\begin{align*}
4m \cdot \vol(M)
&= 2 \cdot \volc(T_1 \cdots T_{2m})\\
&= \vol(X)\\
&> \volc(a_1a_1^R) + \volc(a_2a_2^R)\\
&\geq 4m \cdot \vol^{\cdot} (M \dbs \bS),
\end{align*}
as desired.
Suppose conversely that each $\Sigma_i$ is totally geodesic;
then, each of the surfaces $F^j \subset \cSc(T_1 \cdots T_{2m})$ are totally geodesic, and hence each of the preimages of these surfaces in $X$ are totally geodesic.
We may repeatedly cut along totally geodesic surfaces to yield $\coprod_{k=1}^{2m} (M \dbs \bS)$, and glue along isometric totally geodesic surfaces to yield $\cSc(T_i) = D^{2m}(M \dbs \bS, (V^1,V^2))$, directly yielding the equality
\[
2m \cdot \vol(M)
= \volc(T_1 \cdots T_{2m})
= \vol^{2m}(M \dbs \bS).
\]
\end{proof}
Having proven our theorem in the case of a starburst in a manifold, we may extend our theorem to starbursts in $\ell$-pieces identically to our extension from Theorem \ref{Main theorem separating} to Theorem \ref{Main theorem separating inductive}, with references to Theorem \ref{Main theorem separating} replaced by Theorem \ref{Main theorem}, and omitting the word ``separating.''
\subsection{Composition of pieces}
In the following short section, we state and prove a corollary to Theorem \ref{Main theorem} about \emph{combining hyperbolic pieces}.
We will use this to prove hyperbolicity of composition of tangles in Sections \ref{Knot theory section} and \ref{hyperbolicitytangles}.
\begin{corollary}\label{General composition corollary}
Let $\cbr{(P_i, (V_i^1,V_i^2))}$ be a set of $n$ pieces, each $2mn$-hyperbolic, equipped with diffeomorphisms $\phi_i:V_i^2 \xrightarrow\sim V_{i+1}^1$ for each $1 \leq i \leq {n-1}$.
Then, the piece
\[
(P,(V_1^1, V_n^2)) \hspace{50pt} \text{where} \hspace{50pt} P = \frac{\coprod_i P_i}{V_i^2 \sim_{\phi_i} V_{i+1}^2}
\]
is $2m$-hyperbolic, with volume satisfying
\[
\vol^{2m}(P) \geq \frac{1}{n} \sum_{i=1}^n \vol^{2mn}(P_i).
\]
\end{corollary}
\begin{proof}
Note that $D^{2m}(P,(V_1^1,V_n^2))$ contains, by tracking the images of $V_i^j$, an $mn$-sheeted starburst $\bS$ which separates the manifold into $m$-namy copies of $P_i$ for each $i$.
Hence Theorem \ref{Main theorem} applied to the replicant $D^{2m}(P,(V_1^1,V_n^2))$ yields that the manifold is hyperbolic, with volume bound as described in the corollary.
\end{proof}
\section{Applications to knot theory} \label{Knot theory section}
Given a manifold $M$ and a surface $\Sigma \subset \partial M$, we define a \emph{tangle in $(M,\Sigma)$} to be an embedded compact 1-manifold in $M$ with boundary contained in $\Sigma$.
Given an $\ell$-piece $(P,(V^j))$, we define a \emph{tangle in $(P,(V^j))$} to be a tangle in $(P,\bigcup_n V_{2n-1} \cap V_{2n})$.
A link may be viewed as a tangle with no arcs.
Given a tangle in an $\ell$-piece $(M,\prn{V_i})$ and a starburst $(M,\bS,\rho)$ in the $\ell$-piece, there is an associated tangle in the $(\ell+1)$-piece on $M \dbs \bS$.
Further, if we have an $\ell$-piece that is a disjoint union $(M \sqcup M',\cbr{V_i \sqcup V'_i})$ containing a tangle $\cT$, then $\cT$ decomposes into the disjoint union of tangles in $(M,\cbr{V_i})$ and $(M',\cbr{V'_i})$.
We explore various applications of Theorem \ref{Main theorem} to the theory of links that are constructed out of tangles.
In Subsection \ref{thickenedtorussection} we explore hyperbolic links within the thickened torus, which can be thought of as the complement of the Hopf link in $S^3$. We decompose the links into thickened-cylinder tangles, and further decompose those into cubical tangles.
In Subsection \ref{solidtorussection}, we consider links in an open solid torus, which can be thought of as the complement of a trivial component in $S^3$. There, we decompose links into wedge tangles.
Following this, in Subsection \ref{Bracelet subsection} we explore hyperbolic links in the 3-sphere which decompose into a cycle of saucer tangles.
Later, in Theorem \ref{Saucer hyperbolicity theorem}, we show that $2n$-hyperbolicity of saucer tangles implies $2m$-hyperbolicity for any $m \geq n$, and hence a cycle of $n \geq 2m$ saucer tangles which are each $2m$-hyperbolic glue together to yield a hyperbolic link.
We decompose saucer tangles further into tetrahedral tangles, and prove the corresponding results there.
In Subsection \ref{spheretimescirclesubsection}, we consider torus lattice links in $S^2 \times S^1$ and decompose them into thickened bigon tangles, proving similar results in this context as well.
\subsection{Links within the thickened torus}\label{thickenedtorussection}
Let $T := S^1 \times S^1$ be the torus.
We define two different classes of tangles arising from the study of links in $T \times I$.
\begin{definition}\label{Thickened torus tangle definition}
Let $C := S^1 \times I \times I$ be the thickened cylinder and let $R_0 := S^1 \times I \times \cbr{0}$ and $R_1 := S^1 \times I \times \cbr{1}$ be the annuli at heights $\{ 0 \}$ and $\{ 1 \}$.
A \emph{thickened-cylinder tangle} $\cT_i$ is a tangle in $(C,R_0 \cup R_1)$, with endpoints in the two annuli at the ends of the thickened cylinder.
By reflecting across $R_0$ and $R_1$, a thickened-cylinder tangle $\cT_i$ determines a link $\Dthick{2}{\cT_i} \subset S^1 \times S^1 \times I$ in a thickened torus.
We say that $\cT_i$ is \emph{2-hyperbolic} if $\Dthick{2}{\cT_i}$ is a hyperbolic link, in which case we define
\[
\volthick{2}{\cT_i} := \frac{\vol\prn{\Dthick{2}{\cT_i}}}{2}.
\]
Let $Q = I^3$ be the cube.
Define the \emph{top and bottom faces} $R_{0i} := I \times \cbr{i} \times I$ and the \emph{left and right faces} $R_{1i} := I \times I \times \cbr{i}$.
A \emph{cubical tangle} $\cT_{ij}$ is a tangle in $(Q,I \times \partial(I^2) - I \times \partial I \times \partial I)$.
By reflecting in two opposite faces of the cube to obtain a second copy glued to the first, a cubical tangle $\cT_{ij}$ determines a thickened-cylinder tangle. Via another reflection operation, this in turn determines a link $\Dthick{(2,2)}{\cT_{ij}} \subset S^1 \times S^1 \times I$ in a thickened torus.
We say that $\cT_{ij}$ is \emph{$(2,2)$-hyperbolic} if $\Dthick{(2,2)}{\cT_{ij}}$ is a hyperbolic link, in which case we define
\[
\volthick{(2,2)}{\cT_{ij}} := \frac{\vol\prn{\Dthick{(2,2)}{\cT_{ij}}}}{4}.
\]
\end{definition}
See Figure \ref{cubicaltangle} for cubical tangle $\cT$ that when doubled, yields a thickened-cylinder tangle, which when doubled again generates $D^{(2,2)}(\cT)$.
\begin{figure}[htpb]
\centering
\includegraphics[width=.9\textwidth]{Figures/cubicaltangle2.pdf}
\caption{ A cubical tangle generates a thickened-cylinder tangle which in turn generates a link in a thickened torus
}
\label{cubicaltangle}
\end{figure}
The following theorems summarize the situations of Figures \ref{Torus cycle figure} and \ref{thickened torus example}, and is an easy corollary of Theorem \ref{Main theorem}.
\begin{theorem}\label{Torus theorem}
Let $L \subset T \times I$ be a link in a thickened torus such that there exists disjoint surfaces separating $L$ into thickened-cylinder tangles $\cT_i$ and each $\cT_i$ is 2-hyperbolic.
Then $L$ is hyperbolic, and its volume satisfies
\[
\vol(L) \geq \sum_i \volthick{2}{\cT_i}.
\]
\qed
\end{theorem}
\begin{theorem} \label{Square tangles theorem}
Let $\cT_i$ be a thickened-cylinder tangle, and suppose there are disjoint surfaces separating $\cT_i$ into a cycle $\cT_{ij}$ of cubical tangles and each $\cT_{ij}$ is $(2,2)$-hyperbolic.
Then, $\cT_i$ is $2$-hyperbolic within, with volume satisfying
\[
\volthick{2}{\cT_i} \geq \sum_j \volthick{(2,2)}{\cT_{ij}}.
\]
If a link $L$ in a thickened torus decomposes into cubical tangles such that each $\cT_{ij}$ is $(2,2)$-hyperbolic, then $L$ is hyperbolic with volume satisfying
\[
\vol(L) \geq \sum_{ij} \volthick{(2,2)}{\cT_{ij}}.
\]
\qed
\end{theorem}
Such surfaces as in Theorem \ref{Torus theorem} can be found by taking a sequence of disjoint punctured annuli tracing out a $(p,q)$-curve on each boundary torus of $T \times I$.
Since there is a homeomorphism of $T \times I$ taking such a punctured annulus to one bounding meridians, Theorem \ref{Torus theorem} may be pictured as \emph{cutting the thickened torus along meridians} as in Figure \ref{Torus cycle figure}.
Further, Theorem \ref{Square tangles theorem} may be pictured by cutting a thickened cylinder into cubes, as in Figure \ref{thickened torus example}.
Note that diffeomorphisms of pieces lift to diffeomorphisms of their $2n$-replicants.
Hence Mostow-Prasad rigidity implies that we may replace any piece with a diffeomorphic piece while maintaining the same information regarding hyperbolicity and volume.
Using this, we may replace each thickened-cylinder tangle with one whose endpoints are ``standardized'' such that whenever thickened-cylinder tangle $\cT_i$ has the same number of endpoints at its top as $\cT_j$ has at its bottom, there is an evident gluing of the tangles
\[
(C,\cT_i \circ \cT_j) := (C,\cT_i) \cup_{R_1 \sim R_0} (C,\cT_j)
\]
which attaches the first copy of $R_1$ to the second copy of $R_0$.
Through an analogous process, we may define two ways of gluing cubical tangles:
\emph{top-to-bottom}, or \emph{left-to-right}, written
\[
(Q,\cT_{ij} \circ_a \cT_{kl}) := (Q,\cT_{ij}) \cup_{R_{a0} \sim R_{a1}} (Q,\cT_{kl}))
\]
whenever the numbers of endpoints are compatible.
The following establishes a relationship between hyperbolicity and composition, and it follows as a clear application of Theorem \ref{Main theorem inductive}.
\begin{proposition}\label{Tangle composition proposition}
Suppose $\cT_i,\cT_j$ are $2$-hyperbolic thickened-cylinder tangles, and suppose that that $\cT_i \circ \cT_j$ is defined.
Then, $\cT_i \circ \cT_j$ is $2$-hyperbolic with volume satisfying
\[
\volthick{2}{\cT_i \circ \cT_j} \geq \volthick{2}{\cT_i} + \volthick{2}{\cT_j}
\]
Suppose instead that $\cT_{ij},\cT_{kl}$ are $(2,2)$-hyperbolic cubical tangles and suppose that $\cT_{ij} \circ_a \cT_{kl}$ is defined.
Then, $\cT_{ij} \circ_a \cT_{kl}$ is $(2,2)$-hyperbolic in a thickened torus, with volume satisfying
\[
\volthick{(2,2)}{\cT_{ij} \circ_a \cT_{kl}} \geq
\volthick{(2,2)}{\cT_{ij}}
+ \volthick{(2,2)}{\cT_{kl}}
\]
\qed
\end{proposition}
In a subsequent paper \cite{selectalternating}, we will provide many examples of 2-hyperbolic thickened-cylinder tangles.
\subsection{Links within a solid torus} \label{solidtorussection}
Let $V := D \times S^1$ be the solid torus, where $D$ is the 2-dimensional disk.
Unlike the thickened torus case, cutting links $L \subset V$ via starbursts that intersect the boundary $\partial V = T$ in meridians and longitudes yields tangles in different ambient spaces.
In particular, cutting along a starburst with longitude boundaries yields a cycle of tangles living in \emph{cyclical wedges}, and cutting along a surface bounded by disjoint meridians yields a cycle of tangles living in solid cylinders.
Each of these may be further cut into square tangles that live in wedges.
We give analogous definitions to the thickened torus case in order to find lower bounds on volumes of links $L \subset V$.
\begin{definition}\label{Solid torus tangle definition}
A \emph{solid-cylinder tangle} $\cT_i$ is a tangle in $(D \times I, D \times \partial I)$.
In a similar manner to thickened-cylinder tangles, doubling a solid-cylinder tangle $\cT_i$ over $D \times \partial I$ determines a link in a solid torus, written
\[
(V,\Dsolid{2}{\cT_i}) := (D \times I,\cT_i) \cup_{D \times \partial I} (D \times \cT_i).
\]
We say that $\cT_i$ is \emph{$2$-hyperbolic} if $\Dsolid{2}{\cT_i}$ is hyperbolic, in which case we define
\[
\volsolid{2}{\cT_i} := \frac{\vol\prn{\Dsolid{2}{\cT_i}}}{2}.
\]
Let $P$ be the solid triangle, and let $F \subset P$ be the union of two edges of $P$.
A \emph{cyclic wedge tangle} $\cT_i$ is a tangle in $(S^1 \times P, S^1 \times F)$.
Via the reflection $r:P \rightarrow P$ sending one edge of $F$ to another, a cyclic wedge tangle $\cT_i$ determines a reflected cyclic wedge tangle $\cT_i^R$.
Then, gluing adjacent tangles together along the length $2n$ cycle sequence $(\cT_i,\cT_i^R,\dots,\cT_i^R)$ determines a link $\Dsolid{2n}{\cT_i} \subset V$, called the \emph{$2n$-replicant}.
We say that $\cT_i$ is \emph{$2n$-hyperbolic} if $\Dsolid{2n}{\cT_i}$ is hyperbolic, in which case we define the volume
\[
\volsolid{2n}{\cT_i} := \frac{\vol\prn{\Dsolid{2n}{\cT_i}}}{2n}.
\]
A \emph{wedge tangle} $\cT_{ij}$ is a tangle in $(I \times P, I \times F \cup \partial I \times P)$.
Similarly, $r$ determines a reflected tangle $\cT^R_{ij}$, the sequence $(\cT_{ij},\cT_{ij}^R,\dots,\cT_{ij}^R)$ determines a solid cylindrical tangle, and doubling determines a cyclic wedge tangle.
Fixing the number $2n$, the $2n$-replicant of the cyclic wedge tangle is the same as the $2$-replicant of the solid cylindrical tangle, so we may define the \emph{$(2n,2)$-replicant} $\Dsolid{(2n,2)}{\cT_i}$ to be the resulting link from either replication.
We say that $\cT_{ij}$ is \emph{$(2n,2)$-hyperbolic} if $\Dsolid{(2n,2)}{\cT_{ij}}$ is hyperbolic, in which case we define
\[
\volsolid{(2n,2)}{\cT_{ij}} := \frac{\vol\prn{\Dsolid{(2n,2)}{\cT_{ij}}}}{4n}.
\]
\end{definition}
See Figure \ref{wedgetangle} for a wedge tangle $\cT$ which, when doubled one way, yields a solid-cylinder tangle, and when doubled another way, yields a cyclic wedge tangle. Each of those, when doubled, both generate $D^{(2,2)}(\cT)$.
The fact that doubling over the two sets of endpoints commute with each other is a special case of Proposition \ref{Replication commuting proposition}.
\begin{figure}[htpb]
\centering
\includegraphics[width=.7\textwidth]{Figures/wedgetangle.pdf}
\caption{A wedge tangle generates either a solid-cylinder tangle or a cyclic wedge tangle, each of which generates the same link in the solid torus.}
\label{wedgetangle}
\end{figure}
Using this notation, we have a similar result to Theorems \ref{Torus theorem} and \ref{Square tangles theorem}, which we condense into a single theorem which follows immediately from Theorem \ref{Main theorem inductive}.
\begin{theorem}\label{Solid torus theorem}
Let $L \subset V$ be a link in a solid torus.
\begin{enumerate}[label=\upshape{(\roman*)},ref={\thetheorem(\roman*)}]
\item
Suppose there exists a surface separating $L$ into solid-cylinder tangles $\cT_i$ such that each $\cT_i$ is $2$-hyperbolic.
Then $L$ is hyperbolic, and its volume satisfies
\[
\vol(M) \geq \sum_i \volsolid{2}{\cT_i}.
\]
\item Suppose there exists a starburst separating $L$ into $2n$ cyclic wedge tangles $\cT_i$ such that each $\cT_i$ is $2n$-hyperbolic.
Then $L$ is hyperbolic, and its volume satisfies
\[
\vol(M) \geq \sum_i \volsolid{2n}{\cT_i}.
\]
\item
Suppose there is a starburst separating a solid-cylinder tangle $\cT_{i}$ into a cycle of $2n$ wedge tangles $\cT_{ij}$ such that each wedge tangle is $(2n,2)$-hyperbolic.
Then, $\cT_i$ is $2$-hyperbolic in the solid torus, with volume satisfying
\[
\volsolid{2}{\cT_i} \geq \sum_j \volsolid{(2n,2)}{\cT_{ij}}.
\]
\item
Suppose there is a starburst separating a cyclic wedge tangle $\cT_{i}$ into a cycle of wedge tangles $\cT_{ij}$ such that each wedge tangle is $(2n,2)$-hyperbolic in a solid torus.
Then, $\cT_i$ is $2n$-hyperbolic in the solid torus, with volume satisfying
\[ \volsolid{2n}{\cT_i} \geq \sum_j \volsolid{(2n,2)}{\cT_{ij}}.
\]
\end{enumerate}
\qed
\end{theorem}
We describe an analog to Proposition \ref{Tangle composition proposition};
the class of 2-hyperbolic solid-cylinder tangles is closed under composition by Theorem \ref{Main theorem}.
Further, Corollary \ref{General composition corollary} implies that the composition of $n$ cyclic-wedge tangles which are each $2mn$-hyperbolic is $2m$-hyperbolic in the solid torus.
Similar statements hold for wedge tangles, depending on which direction the composition follows.
There is a strong relationship between solid-cylinder tangles and thickened-cylinder tangles.
If a solid-cylinder tangle $\cT$ has an unknotted strand with endpoints on both components of $D \times \partial I$, then the complement of that strand determines a thickened-cylinder tangle.
The solid-cylinder tangle may be reconstructed from this thickened-cylinder tangle.
The complement of the double of the solid-cylinder tangle is homeomorphic to the complement of the associated thickened-cylinder tangle, so \cite{selectalternating} will yield a large class of examples of $2$-hyperbolic solid-cylinder tangles.
\subsection{Bracelet links within the sphere} \label{Bracelet subsection}
\begin{definition}\label{Bracelet tangle definition}
A \emph{saucer tangle} is a tangle in the piece defined in Figure \ref{saucer} which has at least two endpoints on each face and composition of saucer tangles is defined by attaching their endpoints, as in the thickened-cylinder case. As defined in the introduction, if $\cT_i$ is a saucer tangle, then the length-$2n$ sequence $(\cT_i,\cT_i^R,\dots,\cT_i^R)$ determines a bracelet link $\Dsphere{2n}{\cT_i} \subset S^3$ in the 3-sphere, called it $2n$-replicant.
We say that $\cT_i$ is \emph{$2n$-hyperbolic} if $\Dsphere{2n}{\cT_i}$ is a hyperbolic link, in which case we define the volume
\[
\volsphere{2n}{\cT_i} := \frac{\vol\prn{\Dsphere{2n}{\cT_i}}}{2n}.
\]
\end{definition}
\begin{remark}
We make the restriction that saucer tangles have two endpoints on each face in service of Theorem \ref{Saucer hyperbolicity theorem}.
This does not fundamentally change the class of $2m$-hyperbolic saucer tangles relative to the class of tangles without such a restriction;
in the case $m \geq 2$, any tangle with $\leq 1$ endpoints on one face will have $2m$-replicant which is not a prime link, and hence is not hyperbolic, and in the case of a $2m=2$ hyperbolic tangle, we can choose a piece structure which has the same $2$-replicant and has at least 2 endpoints on each face.
\end{remark}
Unlike Theorems \ref{Torus theorem}, \ref{Square tangles theorem}, and \ref{Solid torus theorem}, the intersection of the surfaces making up the evident starburst is closed in this case;
hence we may directly conclude the following theorem from Theorem \ref{Main theorem separating}.
\begin{theorem}\label{Bracelet theorem}
If $L$ is a bracelet link made of a cycle $(\cT_i)_{i=1}^{2n}$ of $2n$ saucer tangles such that each $\cT_i$ is $2n$-hyperbolic, then $L$ is hyperbolic and the volumes satisfy
\[
\vol(L) \geq \sum_i \volsphere{2n}{\cT_i}.
\]
\qed
\end{theorem}
Hyperbolicity of saucer tangles will be discussed further in Section \ref{hyperbolicitytangles}, including composition of tangles and the relationship between $2m$-hyperbolicity and $2m'$-hyperbolicity for $m,m'$ distinct.
\subsection{Links in \texorpdfstring{$S^2 \times S^1$}{S2 x S3}} \label{spheretimescirclesubsection}
When we glue two solid tori together along their boundary, meridian to meridian, we obtain $S^2 \times S^1$. Given a torus lattice link $L$ projected onto the separating torus, we can glue a triangular prism from each solid torus together along a square face from each on the separating torus to obtain a thickened bigon. See Figure \ref{thickened bigon square tangle}. Thus we can decompose $L$ into thickened bigon tangles.By Theorem \ref{Main theorem separating}, a similar theorem to our previous cases holds.
\begin{figure}[htbp]
\centering
\includegraphics[scale=.4]{Figures/bigontangle.pdf}
\caption{A square tangle in a thickened bigon.}
\label{thickened bigon square tangle}
\end{figure}
\begin{theorem}\label{S^2 times S^1 theorem}
Let $L$ be an $(2m, 2n)$-torus lattice link in $S^2 \times S^1$, consisting of $4mn$ thickened bigon tangles $\{ \cT_{ij}\}$. If each tangle $\cT_{ij}$ is $(2,2n)$-hyperbolic, then $L$ is hyperbolic, and its volume satisfies
\[
\vol(M) \geq \sum_{i,j} \volsolid{(2,2n)}{\cT_i}.
\]
\end{theorem}
\section{Various hyperbolicities for tangles} \label{hyperbolicitytangles}
Define a saucer tangle to be a {\it braid} if all strings of the tangle move monotonically left to right from the left face to the right face, intersecting each disk in the saucer parallel to the faces exactly once. We note that if a tangle is $2n$-hyperbolic, then its composition with a braid is still $2n$-hyperbolic with the same $2n$-volume. This is immediate because the composition of an $\ell$-string braid with its reflection yields a trivial braid of $\ell$ strings. Hence, the $2n$-replicant of a tangle $T$ composed with a braid yields the $2n$-replicant of $T$ alone.
\begin{theorem}\label{Saucer hyperbolicity theorem}
Suppose $\cT$ is a $2n$-hyperbolic tangle in a saucer for some $n \geq 1$.
Then, $\cT$ is $2m$-hyperbolic for all $m \geq n$.
\end{theorem}
\begin{proof}
It suffices to prove that $\cT$ is $2(n+1)$-hyperbolic, which is by definition showing $D^{2(n+1)}(\cT)$ is hyperbolic. Let $L'$ be the link that is given by the $2(n+1)$-replicant of $\cT$. So we want to show $S^3 \setminus L'$ is hyperbolic.
We are assuming $D^{2n}(\cT)$ is hyperbolic, which means that the link $L$ in $S^3$ that is the $2n$-replicant of $\cT$ is hyperbolic. Define $\bS$ to be the collection of surfaces that decompose $D^{2n}(\cT)$ into homeomorphic copies of $\cT$, and let $E$ be the intersection set of the surfaces. Note that in the case $n=1$, $E = \emptyset$.
Because $D^{2n}(\cT)$ is hyperbolic and reflection in any of the surfaces is a symmetry of the manifold, Theorem \ref{fixed surface lemma} implies that each surface is totally geodesic, which immediately implies it is incompressible in $D^{2n}(\cT)$.
Let $\bS'$ be the collection of surfaces that decompose $D^{2(n+1)}(\cT)$ into homeomorphic copies of the saucer tangle $\cT$. Let $E'$ be their intersection set. Note that $E' \neq \emptyset$ for all $n$.
We begin the proof by showing that each surface in $\bS'$ is incompressible.
Then, we will use this to prove that $L'$ is nontrivial, and use this to prove that essential disks induce essential spheres.
Then, we again use incompressibility to prove that essential spheres induce essential spheres in $D^{2n}(\cT)$, so $2n$-hyperbolicity precludes essential disks and spheres in $D^{2(n+1)}(\cT)$.
Then, we prove that $L'$ is a prime link, precluding certain essential annuli.
Given a torus in $D^{2(n+1)}(\cT)$, we use this together with a strategy of restricting to pieces and replicating in order to prove that there are no essential tori in $D^{2(n+1)}(\cT)$.
Lastly, we use the properties of $L'$ already proved to show that $L'$ is not Seifert fibered, and use \cite{Hatcher, BM} to conclude that $D^{2(n+1)}(\cT)$ has no essential annuli, implying that it is hyperbolic.
\sauceremph{Each surface in $\bS'$ is incompressible.}
Suppose for contradiction that $\Sigma \in \bS'$ is compressible.
Choose a compressing disk $D$ for $\Sigma$ that has the least number of intersections with $E'$, and then the least number of simple closed curve intersections with $\bS'$ among such disks.
The intersection graph of $D$ with $\bS'$ on $D$ will be a collection of arcs with both endpoints on the boundary and a collection of simple closed curves.
If there is a simple closed intersection curve with a surface $\Sigma'$, we replace the disk $D$ with an innermost such disk on $D$, which does not intersect any other surfaces if it exists. This contradicts the minimality of the intersection set for $D$, so there are no simple closed curves of intersection.
If there are arcs of intersection on $D$, choose an outermost arc of intersection on $D$ cutting a bigonal disk $D'$ from $D$. Then $D'$ is a bigonal disk in a particular saucer $S_i$ with its edges in the faces of $S_i$. If both edges are in the same face of $S_i$, then when replicated in $D^{2n}(\cT)$, it yields a set of spheres, each of which must bound balls in $S^3 \setminus L$. Thus we could have isotoped $D$ to eliminate intersections in $D^{2(n+1)}(\cT)$, a contradiction to minimality.
If the two edges of $D'$ are on distinct faces of $S_i$, replicating $D'$ yields a sphere in $D^{2n}(\cT)$, which must be inessential.
Hence it bounds a ball. Thus $D'$ must cut a ball from the saucer tangle in $S_i$. This means that $D$ could have been isotoped to eliminate the intersection curve in $D^{2(n+1)}(\cT)$, contradicting minimality. This proves that the surfaces in $\bS'$ are all incompressible.
\sauceremph{The link $L'$ is nontrivial.}
Suppose first that the link $L'$ is trivial. Then $L'$ consists of a single component bounding a disk $D$. By incompressibility of the surfaces in $\bS'$, we can remove all simple closed curves of intersection with any of the surfaces with $D$. For any surface $\Sigma$ in $\bS'$, consider an outermost intersection arc in $D \cap \Sigma$ on $D$.
It cuts a disk $D'$ from $D$, the boundary of a regular neighborhood of which is a compressing disk for $\Sigma$, since $\Sigma$ has at least four punctures by $L'$.
Thus, $D$ and therefore the link $L'$ cannot intersect any of the surfaces in $\bS'$, again a contradiction to our construction. Therefore $L'$ cannot be a trivial knot.
\sauceremph{The manifold $D^{2(n+1)}(\cT)$ has no essential disks or spheres.}
Note that for a nontrivial link, the boundary of a regular neighborhood of an essential disk in its complement is an essential sphere.
Hence we can prove that there are no essential disks in $D^{2(n+1)}(\cT)$ by proving that there are no essential spheres.
Suppose now that there is an essential sphere
$S$ in $D^{2(n+1)}(\cT)$. Because of the incompressibility of the surfaces in $\bS'$, we can isotope the sphere to miss each in turn until it misses all of the surfaces. Hence, it now resides in a single saucer. Thus, there is a copy $S'$ that exists in $D^{2n}(\cT)$ (in fact $2n$ copies). But in $D^{2n}(\cT)$, $S'$ must bound a ball. Hence it bounds a ball in the single saucer. Thus, $S$ bounds a ball in $D^{2(n+1)}(\cT)$, a contradiction to the essentiality of the sphere.
\sauceremph{The link $L'$ is prime.}
Before moving on to essential tori, we first prove that the link $L'$ in $S^3$ is prime. Suppose not, and let $S$ be the essential (incompressible, boundary-incompressible) twice punctured sphere in the complement of $L'$. Consider its intersections with the surfaces in $\bS'$, and assume we have chosen $S$ to have a minimum number of intersection curves. If it intersects one of the surfaces $\Sigma$ in a closed curve that is trivial on the punctured sphere, then we can isotope to remove the intersection. If it intersects one of the surfaces in a closed curve that is nontrivial on the punctured sphere and that avoids the axis of the starburst on $\Sigma$, and all intersection curves avoid the axis, then taking an innermost such on $S$ yields a disk $D$ that is once-punctured on $S$ and that is properly embedded in a single saucer. This reflects to a twice-punctured sphere $S'$ in $D^{2n}(\cT)$, which must then bound an unknotted arc to one side in $L$ since $L$ must be prime.
If $n \geq 2$, then there are multiple such twice-punctured spheres so it must be that they contain a trivial arc to the inside.
However, then $D$ cuts a ball from the saucer in which it lies which contains an unknotted arc in $D^{2(n+1)}(\cT)$, and we can isotope $D$ to remove the intersection, contradicting minimality of the number of intersection curves.
If $n = 1$, it could be that the unknotted arc is to the outside of $S'$. But then the opposite face of the saucer contains only one endpoint of the tangle, a contradiction to our assumption that the number of endpoints of the tangles in a face is at least two.
Finally, suppose that $S$ intersects a surface $\Sigma$ in a curve nontrivial on $S$ that intersects the axis $E'$ of the starburst on $\Sigma$. Then $S$ intersects all of the surfaces in $\bS'$. The set of intersection arcs forms a $2(n+1)$-valent graph on $S$ with an even number of vertices. Choose a component of this graph that is innermost on $S$, in the sense that all of its complementary regions but one do not contain further components of the graph. Dropping the other components of the graph, we now have a connected $2(n+1)$-valent graph on $S$ with an even number of vertices.
In fact, there can be no odd cycles in $G$. This is because the cyclic order of pieces around a vertex must be preserved. Hence, for any edge in $G$, the cyclic orders must be clockwise around one endpoint and counterclockwise around the other. In particular, this means $G$ must be a bipartite graph, with all even-length cycles.
For such a graph on a sphere, define $f_{2n}$ to be the number of complementary faces with $2n$ edges. Then if $v$, $e$ and $f$ are the number of vertices, edges and faces, by Euler characteristic, we know $v-e+f = 2$, where $e = (n+1) v$, so $-\frac{n}{n+1}e + f = 2$. However, $$f = f_2 + f_4 + f_6 +\dots$$ and $$e = \frac{2f_2 + 4f_4 + 6f_6 + \dots}{2}$$ Plugging these into the Euler equation and algebra yields:
$$\frac{f_2}{n+1} +\frac{1-n}{n+1}f_4 + \frac{1-2n}{n+1}f_6 +\dots = 2$$
Thus, $$f_2 \geq 2(n+1) \geq 4$$
At most two of the at least four bigonal faces can be punctured, since $S$ has only two punctures, so there is a bigonal disk face $D'$ that is not punctured. Then by the same argument as before (replicating in $D^{2n}(\cT)$ to obtain a sphere which must bound a ball), we can isotope $S$ in $D^{2(n+1)}(\cT)$ to lower the number of intersections, a contradiction to our choice of $S$. Thus $L'$ is prime in $S^3$.
\sauceremph{The manifold $D^{2(n+1)}(\cT)$ has no essential tori.}
Suppose now that there is an essential torus $T$ in $D^{2(n+1)}(\cT)$. By the incompressibility of the surfaces in $\bS'$, we can isotope $T$ to eliminate any trivial intersection curves on the torus with any of the surfaces in $\bS'$.
We first consider the case that the torus does not intersect the axis of the starburst. Then $T$ intersects the collection of surfaces $\bS'$ in a finite collection of parallel curves on $T$. Assume we have chosen $T$ to have a minimal number of these intersections. These curves divide $T$ into a finite collection of annuli $\{A_i\}$ which are properly embedded in $D^{2(n+1)}(\cT) \dbs \bS'$. If there is an annulus $A_i$ in a given saucer $S_i$ with its boundaries on the same copy of a surface $\Sigma$ from $\bS'$, then it replicates in
$D^{2n}(\cT)$ to a collection of tori, each made up of two copies of the annulus, and each of which cannot be essential. Let $T'$ be one of these tori. Then it is either boundary-parallel or it is compressible. If it is boundary-parallel, the annulus $A_i$ is boundary-parallel to a segment of a link component in the saucer $S_i$. Then meridianally compressing $T$ in $D^{2(n+1)}(\cT)$ either yields an essential annulus, contradicting primeness of $L'$ or it shows that $T$ was in fact boundary-parallel, also a contradiction.
If $T'$ is compressible, then there is a disk $D$ that realizes the compression with boundary a longitude on $T'$. The disk $D$ can be isotoped to intersect a single saucer in a disk with boundary one arc a nontrivial arc on the annulus $A_i$ and the other arc on the boundary of the saucer. Hence, we can isotope a nontrivial arc on the annulus $A_i$ into the neighboring saucer, leaving a disk from $T$ still in $S_i$. By incompressibility, we can eliminate this intersection, a contradiction to minimality of the intersection curves.
Thus, all annuli $A_i$ have their two boundaries on two different faces of each saucer. Since a pair of adjacent saucers and tangles are homeomorphic via reflection, we can consider $A_i$ and $A_{i+1}^R$ in the same saucer $S_i$. Note that both $A_i$ and $A_{i+1}^R$ share one boundary component on the face $F_1$ of $\partial S_i$.
We now isotope $A_i$ and $A_{i+1}^R$ into general position and consider intersection curves between them. These can take the form of arcs or simple closed curves. If there is a simple closed curve intersection, then if it is trivial on both annuli, by taking innermost such curves, the disks cut from them form a sphere in $S_i$. But as $S_i$ appears in $D^{2n}(\cT)$, this sphere must bound a ball. Hence we can isotope to remove the intersection curve. If one of the two curves is trivial, this contradicts the incompressibility of the torus.
We consider the case that both are nontrivial shortly.
\begin{figure}[htbp]
\centering
\includegraphics[scale=0.7]{Figures/intersectingannuli.pdf}
\caption{The two annuli $A_i$ and $A_{i+1}^R$ can intersect in various ways.}
\label{intersectingannuli}
\end{figure}
Suppose there is an intersection arc beginning and ending on their shared boundary, as in Figure \ref{intersectingannuli}(a). Then it must cut a disk from each and again, we can form a sphere that allows us to isotope the intersection away.
Suppose there is an intersection curve that runs from the shared boundary on $F_1$ to the opposite boundary on each of $A_i$ and $A_{i+1}^R$ on $F_2$, as in Figure \ref{intersectingannuli}(b). Then the opposite boundaries must cross each other on $F_2$, so there must be an even number of such curves. Then a pair of disks on the annuli cut from two adjacent vertical arcs on each share their bottom edge and their two vertical edges. Together they form a disk with boundary in $F_2$. When we take the neighboring saucer, this disk doubles to a sphere, which appears in these two copies of the saucer and hence in $D^{2n}(\cT)$.
But it must bound a ball in $D^{2n}(\cT)$ and hence in $D^{2(n+1)}(\cT)$. Thus, we can isotope the annuli to eliminate these two intersection curves.
So there are now no intersection arcs that end on the shared boundary of the two annuli. Suppose that there is a simple closed curve of intersection. As shown above, it must be nontrivial on both annuli, as in Figure \ref{intersectingannuli}(c). Taking the two annuli it cuts from $A_i$ and $A_{i+1}^R$ that share their shared boundary, we can construct a torus living in $S_i$, with a longitude on the boundary of $S_i$. Since $S_i$ appears in $D^{2n}(\cT)$, this torus $T'$ must either compress or be boundary parallel there. Because $A_i$ and $A_{i+1}^R$ are incompressible, $T'$ cannot compress to the outside. If it compresses to the inside, then it compresses to a sphere $S' \subset S_i$, which must bound a ball; hence $T'$ bounds a solid torus in $S_i$, and we can isotope $A_{i+1}^R$ to $A_i$, eliminating the intersection curve.
If $T'$ is boundary parallel, then there is a link component that forms the core curve for the solid torus bounded by $T'$. But this link component wraps once around both annuli. When we reflect $S_i$ in $D^{2n}(\cT)$, this link component is reflected to a second copy that is parallel to the first. That is, they will form the boundary of an annulus that follows either $A_i$ and its reflection or $A_{i+1}$ and its reflection, contradicting the hyperbolicity of $D^{2n}(\cT)$. Thus there are no simple closed curves of intersection.
The final possibility for intersection curves are arcs of intersection that end on the other unshared boundaries of the two annuli, as in Figure \ref{intersectingannuli}(d). These must cut disks from the two annuli. Together these two disks form a disk with boundary on $F_2$. By incompressibilty, we know that this disk together with the disk it cuts from $F_2$ bounds a ball, and therefore we can isotope the two annuli to eliminate the intersection.
Thus, we now have no intersections curves. The union of the two annuli forms a new annulus, the reflected double over $F_2$ of which is a torus $T''$. That torus exists in $D^{2n}(\cT)$ and therefore it must either be boundary parallel or compressible there. If it is boundary parallel there, there is a single link component inside it that forms the core curve for the solid torus bounded by $T''$. But this contradicts the fact $T''$ was obtained by reflection doubling. If $T''$ is compressible then it must bound a solid torus with no components inside it. Therefore $A_{i+1}^R$ is isotopic to $A_i$.
Thus we see that the original torus $T$ can be isotoped so that all of the $A_i$ are the same in each $S_i$. That is, the torus is made up of $A_1, A_1^R, A_1, \dots, A_1^R$, in the respective copies of $S_1$ and $S_1^R$, making up $D^{2(n+1)}({\cT})$.
Consider the corresponding torus $T'''$ in $D^{2n}(\cT)$. It consists of $2n$ copies of $A_1$. It must be either boundary-parallel or compressible. if it is boundary-parallel the solid cylinder bounded by $A_1$ contains a core curve that is an arc in a link component. But then $T$ is also boundary-parallel in $D^{2(n+1)}(\cT)$.
If $T'''$ is compressible in $D^{2n}(\cT)$, the compression must be to the outside. Thus, there is a compressing disk $D$ which we can choose to minimize the intersection curves with $\bS$ and such that its boundary is a longitude on $T'''$. By incompressibility of the surfaces, we can isotope it so that it has a single arc of intersection with each surface $\Sigma_i$ and all of those arcs meet at the intersection of the surfaces. Then take $D'$ to be its intersection with saucer $S_1$. If we then take the union of the copies of $D'$ and its reflections, we obtain a compressing disk for $T$ in $D^{2(n+1)}(\cT)$. Thus there are no essential tori that avoid $E$ in $D^{2(n+1)}(\cT)$.
Suppose now that there is an essential torus in $D^{2(n+1)}(\cT)$ that intersects the axis $E$ of the starburst. We take the essential torus with the least number of intersections with $E$, and then with the least number of curves of intersection with $\bS'$. Since $E$ is the only place that any two of the surfaces intersect and all the surfaces intersect there, the collection of intersection curves of $T$ with $\bS'$ forms a graph $G$ on $T$ with an even number of vertices, each of valency $2(n+1) \geq 4$. Since $T$ intersects each surface in a disjoint set of curves, each surface $\Sigma_i$ intersects $T$ in a parallel set of $(p_i, q_i)$-curves.
If there is a component of $G$ that lies in either a disk or annulus on $T$, then as in the case of an essential sphere that intersects $E'$, there is a complementary region to $G$ that is a bigon, and this generates a sphere in $D^{2n}(\cT)$ that bounds a ball, and that then allows us to isotope $T$ to reduce the number of intersections with $E'$ in $D^{2(n+1)}(\cT)$, a contradiction.
So $G$ must consist of a single component such that all of its complementary regions are disks.
We use a similar argument in the proof of Lemma \ref{Eintersectstorusboundary}. As mentioned previously, $G$ cannot have any bigonal complementary regions. This forces $n= 1$ or $n=2$. In the second case, there are three surfaces in $\bS'$ and and all complementary regions on $T$ are triangles. However, as mentioned previously, there can be no odd cycles in $G$, as cyclic order must be preserved between different vertices.
So $n = 1$ and there are two surfaces in $\bS'$, each a punctured sphere. $T$ intersects each one in a collection of curves. Taking an innermost such curve on the sphere, we know it bounds a disk in $S^3$ intersecting $T$ only in its boundary. So it must either be a meridian or a longitude. Since the two sets of curves cannot be parallel, as otherwise they would generate a bigon, one set consists of meridians and the other of longitudes, and $T$ is an unknotted torus in $S^3$. If we take any square $Q$ complementary to $G$ on $T$, it is properly embedded in $\cT$ and $\partial Q$ crosses $E$ four times. It doubles to a sphere living in $D^{2n}(\cT) = D(\cT)$, which by hyperbolicity, bounds a ball. So to one side of $Q$ in $\cT$, there are no link arcs. Therefore, we can isotope a properly embedded arc in $Q$ to $E$. We can use this fact to isotope $T$ in $D^4(\cT)$ to lower the number of intersections with $E$, a contradiction to our choice of $T$.
\sauceremph{The manifold $D^{2(n+1)}(\cT)$ has no essential annuli.}
We now consider an essential annulus $A$ in $D^{2(n+1)}(\cT)$.
Lemma 1.16 of \cite{Hatcher}, shows that if a manifold is connected, orientable, irreducible and atoroidal, which we have shown our link complement to be, then the existence of an incompressible boundary-incompressible annulus implies the manifold is Seifert fibered. In \cite{BM}, a complete determination of link complements in $S^3$ that are Seifert fibered is given. In particular, the only options that are nontrivial, prime, and that have no essential tori are a Hopf link, or a $(p,q)$-torus knot $K$, with annulus on the defining torus, and possibly with one or both of the core curves of the solid tori to either side of the torus, denoted $\alpha_1$ and $\alpha_2$, included as well.
There are no spheres in $S^3$ that become incompressible punctured spheres in the Hopf link complement. So $L'$ is a torus knot with possibly two additional components as described.
If neither core curve is present, so we are in a torus knot complement, then there are no incompressible punctured spheres to play the role of the surfaces in $\bS'$ (\cite {Tsau}). In the case that $\alpha_1$ or $\alpha_2$ are present, the $(p,q)$-torus knot $K$ lies on a torus $T$ defined by the annulus $A$, with a solid torus or a cored solid torus to each side.
Since the fundamental groups of a solid torus and cored solid torus are $\mathbf{Z}$ or $\mathbf{Z} \times \mathbf{Z}$, and a planar surface has fundamental group a free product of $\mathbf{Z}$'s, the only incompressible sub-surfaces of our punctured sphere properly embedded in either one are a disk, a once-punctured disk whose puncture corresponds with a puncture of $\Sigma$, and an annulus whose boundaries are not punctures of $\Sigma$. But note that the boundaries of any of these surfaces can intersect the punctures on $\Sigma$ since they may cross the $(p,q)$-torus knot on $T$. Thus, an incompressible punctured sphere $\Sigma \subset D^{2(n+1)}(\cT)$ can only intersect each side in a disjoint union of disks, once-punctured disks, or annuli.
In the case of the annulus, both boundaries are on $T$, and the annulus is boundary-parallel to $T$.
If there is a disk $D$ that has boundary trivial on $T$, then it cuts a ball $B$ from that side of $T$. If there are no further components of $\Sigma \setminus (\Sigma \cap T)$ inside the ball, then either $\partial D$ intersects $L'$ or it does not. If not, we can isotope $\Sigma$ to decrease the number of intersections with $T$. If it does, then we can isotope an arc of $L'$ to $D$ while fixing it endpoints, implying $\Sigma$ is boundary-compressible and hence compressible.
If there is another component of $\Sigma \setminus (\Sigma \cap T)$ in $B$, we can take an innermost such. It cannot be a punctured disk. If it is a disk, the argument just given eliminates it. If it is an annulus $A'$, that annulus must have boundaries trivial on $T$, and that annulus cuts a solid torus $V$ off of that side of $T$. If $\partial A'$ does not intersect $L'$, we can isotope $A'$ to eliminate the intersections with $T$. If it does intersect $L'$ then there is an arc of $L'$ that can be isotoped fixing its endpoints into $A'$, once again contradicting incompressibility of $\Sigma$.
A similar argument eliminates all annuli in $\Sigma \setminus (\Sigma \cap T)$ with boundary trivial in $T$ by taking an innermost such and applying the argument already given.
By taking an innermost curve of intersection of $\Sigma \cap T$ on $\Sigma$, there is a disk or punctured disk to one side of $T$, and it must have meridianal boundary in the solid torus or cored solid torus to that side of $T$.
Hence, to the other side, there must be an annulus $A'$ with boundaries that are longitudinal in the solid torus or cored solid torus to that side. Such an annulus splits the solid torus or cored solid tori into two solid tori, at most one of which can contain the possibly missing core. To that side, take an innermost such annulus $A''$. That annulus is boundary-parallel to $T$, and if its boundary misses $L'$, we can eliminate the intersections. If its boundaries intersect $L'$, then we can isotope an arc of $L'$ into $A'$, contradicting incompressibility of $\Sigma$.
\begin{comment}Let $\Sigma$ be one of the incompressible punctured spheres in $\bS'$ and let $\Sigma'$ be the corresponding unpunctured sphere in $S^3$. Then after isotopy, $A$ cannot intersect $\Sigma'$ in a trivial curve. This implies that $T$ intersects $\Sigma'$ in a collection of nontrivial parallel curves on $T$. At least two of these curves must be innermost on the sphere $\Sigma'$, and therefore those two must correspond to either a pair of meridianal disks to the inside of $T$ or a pair of longitudinal disks to the outside of $T$. Thus, intersection curves in $\Sigma' \cap T$ must either be all meridians or all longitudes. In either case, in the two solid tori to either side of $T$ in $S^3$, an orientable incompressible properly embedded planar surface that is a component of $\Sigma' \setminus (\Sigma' \cap T)$ must either be a disk or an annulus that is boundary-parallel since these are the only incompressible 2-sided surfaces in a solid torus. And there must be at least one such annulus on one of the two sides of $T$. Choose such an annulus $A'$ that is outermost in the solid torus bounded by $T$ on that side. If one of the two curves $\alpha_i$ is present, and $A'$ is punctured by $\alpha_i$, then it must be punctured an even number of times and there is an arc of $\alpha_i$ that can be isotoped relative its boundary into $A'$. This means that $\Sigma$ is boundary-compressible and hence compressible, a contradiction. Otherwise, a boundary component of $A'$ intersects the torus knot at least $q$ times if its boundary is a meridian and $p$ times if its boundary is a longitude. But in either case, there will be an arc of the torus knot that is isotopic relative its endpoints into the annulus. This also implies that the sphere $\Sigma$ is boundary-compressible and therefore compressible, a contradiction.
{\color{magenta} I find this whole argument on torus links very confusing and hard to picture/draw. Could you possibly walk me through it at some point soon?}
{\color{blue} I made a few changes to try to clarify, but happy to talk about it.}
\end{comment}
\end{proof}
\begin{remark}
Though Theorem \ref{Saucer hyperbolicity theorem} may extend to other special classes of pieces, it is not true for all pieces.
For instance, the restriction that saucer tangles have at least two endpoints on each face is necessary. If we allowed a saucer tangle that had only one endpoint on a face, we could still choose the tangle so that it is 2-hyperbolic. But it will never be $2m$-hyperbolic for any $m > 1$ as it will not yield a prime knot or link.
Another example where Theorem \ref{Saucer hyperbolicity theorem} fails to extend comes from Lemma \ref{Eintersectstorusboundary}.
Given a $4$-hyperbolic piece $P$ whose $4$-replicant has an intersection locus intersecting a torus boundary component, its $2m$-replicant also has an intersection locus intersecting a torus boundary component.
Hence Lemma \ref{Eintersectstorusboundary} implies that $P$ is not $2m$-hyperbolic for any $m > 2$.
\begin{comment}
Any 2-sheeted starburst yielding hyperbolic pieces and having an intersection locus that intersects a torus boundary component yields a piece whose $2m$-replicant is an $m$-sheeted starburst with intersection locus intersecting a torus boundary.
Hence such a piece is $2m$-hyperbolic for $m = 2$ and no other values of $m$.
For example, note that any 2-piece can be given a new piece structure that generates a 4-replicant diffeomorphic to the $(2,2)$-replicant of the 2-piece structure.
Hence by Lemma \ref{Eintersectstorusboundary}, every $(2,2)$-hyperbolic cubical tangle yields an example of a piece which is $2m$-hyperbolic for $m = 2$ and no other values of $m$.
\end{comment}
\end{remark}
\begin{comment}
\begin{proof}
It suffices to prove that $\cT$ is $2(n+1)$-hyperbolic, which is by definition showing $D^{2(n+1)}(\cT)$ is hyperbolic. Let $L'$ be the link that is given by the $2(n+1)$-replicant of $\cT$. So we want to show $S^3 \setminus L'$ is hyperbolic.
We are assuming $D^{2n}(\cT)$ is hyperbolic, which means that the link $L$ in $S^3$ that is the $2n$-replicant of $\cT$ is hyperbolic. Define $\bS$ to be the collection of surfaces that decompose $D^{2n}(\cT)$ into homeomorphic copies of $\cT$, and let $E$ be the intersection set of the surfaces. Note that in the case $n=1$, $E = \emptyset$.
Because $D^{2n}(\cT)$ is hyperbolic and reflection in any of the surfaces is a symmetry of the manifold, Theorem \ref{fixed surface lemma} implies that each surface is totally geodesic, which immediately implies it is incompressible in $D^{2n}(\cT)$.
Let $\bS'$ be the collection of surfaces that decompose $D^{2(n+1)}(\cT)$ into homeomorphic copies of the saucer tangle $\cT$. Let $E'$ be their intersection set. Note that $E' \neq \emptyset$ for all $n$. We first show each surface $\Sigma$ in $\bS'$ is incompressible.
If not, choose a compressing disk $D$ for $\Sigma$ that has the least number of intersections with $E$ and then the least number of simple closed curve intersections. The intersection graph of $D$ with $\bS'$ on $D$ will be a collection of arcs with both endpoints on the boundary and a collection of simple closed curves.
If there is a simple closed intersection curve with a surface $\Sigma'$, we replace the disk $D$ with an innermost such disk on $D$, which does not intersect any other surfaces if it exists. This contradicts the minimality of the intersection set for $D$ so there are no simple closed curves of intersection.
If there are arcs of intersection on $D$, choose an outermost arc of intersection on $D$ cutting a bigonal disk $D'$ from $D$. Then $D'$ is a bigonal disk with its edges in the faces of a particular saucer $S_i$. If both edges are in the same face of $S_i$, then when replicated in $D^{2n}(\cT)$, it yields a set of spheres, each of which must bound balls in $S^3 \setminus L$. Thus we could have isotoped $D$ to eliminate intersections in $D^{2(n+1)}(\cT)$, a contradiction to minimality.
If the two edges of $D'$ are on distinct faces of $S_i$, replicating $D'$ in $D^{2n}(\cT)$, we obtain a sphere which must be inessential. Hence it bounds a ball. Thus $D'$ must cut a ball from the saucer tangle in $S_i$. This means that $D$ could have been isotoped to eliminate the intersection curve in $D^{2(n+1)}(\cT)$, contradicting minimality. This proves that the surfaces in $\bS'$ are all incompressible.
In order to show that the $2(n+1)$-replicant of $\cT$ is hyperbolic, we show that it contains no essential disks, spheres, annuli or tori. Note that it is enough to show that the link $L'$ is not a trivial knot and there are no essential spheres, tori or annuli, since for a nontrivial link, the boundary of a regular neighborhood of an essential disk is an essential sphere.
Suppose first that the link $L'$ is trivial. Then $L'$ consists of a single component bounding a disk $D$. By incompressibility of the surfaces in $\bS'$, we can remove all simple closed curves of intersection with any of the surfaces with $D$. For any surface $\Sigma$ in $\bS$, consider an outermost intersection arc in $D \cap \Sigma$ on $D$. It cuts a disk $D'$ from $D$, the boundary of a regular neighborhood of which is a compressing disk for $\Sigma$, since $\Sigma$ has at least four punctures by $L'$. Thus, $D$ and therefore the link $L'$ cannot intersect any of the surfaces in $\Sigma'$, again a contradiction to our construction. Therefore $L'$ cannot be a trivial knot.
Suppose now that there is an essential sphere
$S$ in $D^{2(n+1)}(\cT)$. Because of the incompressibility of the surfaces in $\bS'$, we can isotope the sphere to miss each in turn until it misses all of the surfaces. Hence, it now resides in a single saucer. Thus, there is a copy $S'$ that exists in $D^{2n}(\cT)$ (in fact $2n$ copies). But in $D^{2n}(\cT)$, $S'$ must bound a ball. Hence it bounds a ball in the single saucer. Thus, $S$ bounds a ball in $D^{2(n+1)}(\cT)$, a contradiction to the essentiality of the sphere.
Before moving on to essential tori, we first prove that the link $L'$ in $S^3$ is prime. Suppose not. Then let $S$ be the essential (incompressible, boundary-incompressible) twice punctured sphere in the complement of $L'$. Consider its intersections with the surfaces in $\bS'$, and assume we have chosen $S$ to have a minimum number of intersection curves. If it intersects one of the surfaces $\Sigma$ in a closed curve that is trivial on the punctured sphere, then we can isotope to remove the intersection. If it intersects one of the surfaces in a closed curve that is nontrivial on the punctured sphere and that avoids the axis of the starburst on $\Sigma$, and all intersection curves avoid the axis, then taking an innermost such on $S$ yields a disk $D$ that is once-punctured on $S$ and that is properly embedded in a single saucer. This reflects to a twice-punctured sphere $S'$ in $D^{2n}(\cT)$, which must then bound an unknotted arc to one side in $L$ since $L$ must be prime.
If $n \geq 2$, then there are multiple such twice-punctured spheres so it must be that they contain a trivial arc to the inside.
However, then $D$ cuts a ball from the saucer in which it lies that contains an unknotted arc in $D^{2(n+1)}(\cT)$, and we can isotope $D$ to remove the intersection, contradicting minimality of the number of intersection curves.
If $n = 1$, it could be that the unknotted arc is to the outside of $S'$. But then the opposite face of the saucer contains only one endpoint of the tangle, a contradiction to our assumption that the number of endpoints of the tangles in a face is at least two.
Finally, suppose that $S$ intersects a surface $\Sigma$ in a curve nontrivial on $S$ that intersects the axis $E'$ of the starburst on $\Sigma$. Then $S$ intersects all of the surfaces in $\bS'$. The set of intersection arcs forms a $2(n+1)$-valent graph on $S$ with an even number of vertices. Choose a component of this graph that is innermost on $S$, in the sense that all of its complementary regions but one do not contain further components of the graph. Dropping the one complementary region, the others must be disk faces that together make up a disk $D$. If all faces have two edges, then because $2(n+1) \geq 4$, there are at least three bigonal faces. If any faces have more than two edges, then each of the regions complementary to the face on $D$ must contain a bigonal face that is outermost. Thus, in either case, there are at least three bigonal faces. At most two of them can be punctured, since $S$ has only two punctures, so there is a bigonal disk $D'$ that is not punctured. Then by the same argument as before (replicating in $D^{2n}(\cT)$ to obtain a sphere which must bound a ball), we can isotope $S$ in $D^{2(n+1)}(\cT)$ to lower the number of intersections. a contradiction to our choice of $S$. Thus $L'$ is prime in $S^3$.
Suppose now that there is an essential torus $T$ in $D^{2(n+1)}(\cT)$. By the incompressibility of the surfaces in $\bS'$, we can isotope $T$ to eliminate any trivial intersection curves on the torus with any of the surfaces in $\bS'$.
We first consider the case that the torus does not intersect the axis of the starburst. Then $T$ intersects the collection of surfaces $\bS'$ in a finite collection of parallel curves on $T$. Assume we have chosen $T$ to have a minimal number of these intersections. These curves divide $T$ into a finite collection of annuli $\{A_i\}$ which are properly embedded in $D^{2(n+1)}(\cT) \dbs \bS'$. If there is an annulus $A_i$ in a given saucer $S_i$ with its boundaries on the same copy of a surface $\Sigma$ from $\bS'$, then it replicates in
$D^{2n}(\cT)$ to a collection of tori, each made up of two copies of the annulus, and each of which cannot be essential. Let $T'$ be one of these tori. Then it is either boundary-parallel or it is compressible. If it is boundary-parallel, the annulus $A_i$ is boundary-parallel to a segment of a link component in the saucer $S_i$. Then meridianally compressing $T$ in $D^{2(n+1)}(\cT)$ either yields an essential annulus, contradicting primeness of $L'$ or it shows that $T$ was in fact boundary-parallel, also a contradiction.
If $T'$ is compressible, then there is a disk $D$ that realizes the compression with boundary a longitude on $T'$. The disk $D$ can be isotoped to intersect a single saucer in a disk with boundary one arc a nontrivial arc on the annulus $A_i$ and the other arc on the boundary of the saucer. Hence, we can isotope a nontrivial arc on the annulus $A_i$ into the neighboring saucer, leaving a disk from $T$ still in $S_i$. By incompressibility, we can eliminate this intersection, a contradiction to minimality of the intersection curves.
Thus, all annuli $A_i$ have their two boundaries on two different faces of each saucer. Since a pair of adjacent saucers and tangles are homeomorphic via reflection, we can consider $A_i$ and $A_{i+1}^R$ in the same saucer $S_i$. Note that both $A_i$ and $A_{i+1}^R$ share one boundary component on the face $F_1$ of $\partial S_i$.
We now isotope $A_i$ and $A_{i+1}^R$ into general position and consider intersection curves between them. These can take the form of arcs or simple closed curves. If there is a simple closed curve intersection, then if it is trivial on both annuli, by taking innermost such curves, the disks cut from them form a sphere in $S_i$. But as $S_i$ appears in $D^{2n}(\cT)$, this sphere must bound a ball. Hence we can isotope to remove the intersection curve. If one of the two curves is trivial, this contradicts the incompressibility of the torus.
We consider the case that both are nontrivial shortly.
\begin{figure}[htbp]
\centering
\includegraphics[scale=0.7]{Figures/intersectingannuli.pdf}
\caption{The two annuli $A_i$ and $A_{i+1}^R$ can intersect in various ways.}
\label{intersectingannuli}
\end{figure}
Suppose there is an intersection arc beginning and ending on their shared boundary, as in Figure \ref{intersectingannuli}(a). Then it must cut a disk from each and again, we can form a sphere that allows us to isotope the intersection away.
Suppose there is an intersection curve that runs from the shared boundary on $F_1$ to the opposite boundary on each of $A_i$ and $A_{i+1}^R$ on $F_2$, as in Figure \ref{intersectingannuli}(b). Then the opposite boundaries must cross each other on $F_2$, so there must be an even number of such curves. Then a pair of disks on the annuli cut from two adjacent vertical arcs on each share their bottom edge and their two vertical edges. Together they form a disk with boundary in $F_2$. When we take the neighboring saucer, this disk doubles to a sphere, which appears in these two copies of the saucer and hence in $D^{2n}(\cT)$.
But it must bound a ball in $D^{2n}(\cT)$ and hence in $D^{2(n+1)}(\cT)$. Thus, we can isotope the annuli to eliminate these two intersection curves.
So there are now no intersection arcs that end on the shared boundary of the two annuli. Suppose that there is a simple closed curve of intersection. As shown above, it must be nontrivial on both annuli, as in Figure \ref{intersectingannuli}(c). Taking the two annuli it cuts from $A_i$ and $A_{i+1}^R$ that share their shared boundary, we can construct a torus living in $S_i$, with longitude on the boundary of $S_i$. Since $S_i$ appears in $D^{2n}(\cT)$, this torus $T'$ must either compress or be boundary parallel there. Because $A_i$ and $A_{i+1}^R$ are incompressible, $T'$ cannot compress to the outside. If it compresses to the inside, then $T'$ bounds a solid torus in $S_i$, and we can isotope $A_{i+1}^R$ to $A_i$, eliminating the intersection curve.
If $T'$ is boundary parallel, then there is a link component that forms the core curve for the solid torus bounded by $T'$. But this link component wraps once around both annuli. When we reflect $S_i$ in $D^{2n}(\cT)$, this link component is reflected to a second copy that is parallel to the first. That is, they will form the boundary of an annulus that follows either $A_i$ and its reflection or $A_{i+1}$ and its reflection, contradicting the hyperbolicity of $D^{2n}(\cT)$. Thus there are no simple closed curves of intersection.
The final possibility for intersection curves are arcs of intersection that end on the other unshared boundaries of the two annuli, as in Figure \ref{intersectingannuli}(d). These must cut disks from the two annuli. Together these two disks form a disk with boundary on $F_2$. By incompressibilty, we know that this disk together with the disk it cuts from $F_2$ bounds a ball, and therefore we can isotope the two annuli to eliminate the intersection.
Thus, we now have no intersections curves. The union of the two annuli forms an new annulus, the reflected double over $F_2$ of which is a torus $T''$. That torus exists in $D^{2n}(\cT)$ and therefore it must either be boundary parallel or compressible there. If it is boundary parallel there, there is a single link component inside it that forms the core curve for the solid torus bounded by $T''$. But this contradicts the fact $T''$ was obtained by reflection doubling. If $T''$ is compressible then it must bound a solid torus with no components inside it. Therefore $A_{i+1}^R$ is isotopic to $A_i$.
Thus we see that the original torus $T$ can be isotoped so that all of the $A_i$ are the same in each $S_i$. That is, the torus is made up of $A_1, A_1^R, A_1, \dots, A_1^R$, in the respective copies of $S_1$ and $S_1^R$, making up $D^{2(n+1)}({\cT})$.
Consider the corresponding torus $T'''$ in $D^{2n}(\cT)$. It consists of $2n$ copies of $A_1$. It must be either boundary-parallel or compressible. if it is boundary-parallel the solid cylinder bounded by $A_1$ contains a core curve that is an arc in a link component. But then $T$ is also boundary-parallel in $D^{2(n+1)}(\cT)$.
If $T'''$ is compressible in $D^{2n}(\cT)$, the compression must be to the outside. Thus, there is a compressing disk $D$ which we can choose to minimize the intersection curves with $\bS$ and such that its boundary is a longitude on $T'''$. By incompressibility of the surfaces, we can isotope it so that it has a single arc of intersection with each surface $\Sigma_i$ and all of those arcs meet at the intersection of the surfaces. Then take $D'$ to be its intersection with saucer $S_1$. If we then take the union of the copies of $D'$ and its reflections, we obtain a compressing disk for $T$ in $D^{2(n+1)}(\cT)$. Thus there are no essential tori that avoid $E$ in $D^{2(n+1)}(\cT)$.
Suppose now that there is an essential torus in $D^{2(n+1)}(\cT)$ that intersects the axis $E$ of the starburst. We take the essential torus with the least number of intersections with $E$, and then with the least number of curves of intersection with $\bS'$. Since $E$ is the only place that any two of the surfaces intersect and all the surfaces intersect there, the collection of intersection curves of $T$ with $\bS'$ forms a graph $G$ on $T$ with an even number of vertices, each of valency $2(n+1) \geq 4$. Since $T$ intersects each surface in a disjoint set of curves, each surface $\Sigma_i$ intersects $T$ in a parallel set of $(p_i, q_i)$-curves.
If there is a component of $G$ that lies in either a disk or annulus on $T$, then as in the case of an essential sphere that intersects $E'$, there is a complementary region to $G$ that is a bigon, and this generates a sphere in $D^{2n}(\cT)$ that bounds a ball, and that then allows us to isotope $T$ to reduce the number of intersections with $E'$ in $D^{2(n+1)}(\cT)$, a contradiction.
So $G$ must consist of a single component such that all of its complementary regions are disks.
We use a similar argument to the proof of Lemma \ref{Eintersectstorusboundary}. As mentioned previously, $G$ cannot have any bigonal complementary regions. This forces $n= 1$ or $n=2$. In the second case, there are three surfaces in $\bS'$ and and all complementary regions on $T$ are triangles. However, there can be no odd cycles in $G$. This is because the cyclic order of pieces around a vertex must be preserved. Hence, for any edge in $G$, the cyclic orders must be clockwise around one endpoint and counterclockwise around the other. In particular, this means $G$ must be a bipartite graph, with all even-length cycles.
So $n = 1$ and there are two surfaces in $\bS'$, each a punctured sphere. $T$ intersects each one in a collection of curves. Taking an innermost such curve on the sphere, we know it bounds a disk in $S^3$ intersecting $T$ only in its boundary. So it must either be a meridian or a longitude. Since the two sets of curves cannot be parallel as again that would generate a bigon, one set consists of meridians and the other of longitudes, and $T$ is an unknotted torus in $S^3$. If we take any square $Q$ complementary to $G$ on $T$, it is properly embedded in $\cT$ and $\partial Q$ crosses $E$ four times. It doubles to a sphere living in $D^2n(\cT) = D(\cT)$, which by hyperbolicity, bounds a ball. So to one side of $Q$ in $\cT$, there are no link arcs. Therefore, we can isotope a properly embedded arc in $Q$ T $E$. We can use this fact to isotope $T$ in $D^4{\cT}$ to lower the number of intersections with $E$, a contradiction to our choice of $T$.
We now consider an essential annulus $A$.
Lemma 1.16 of \cite{Hatcher}, shows that if a manifold is connected, orientable, irreducible and atoroidal, which we have shown our link complement to be, then the existence of an incompressible boundary-incompressible annulus implies the manifold is Seifert fibered. In \cite{BM}, a complete determination of link complements in $S^3$ that are Seifert fibered is given. In particular, the only options that are nontrivial, prime and that have no essential tori are a Hopf link, or a $(p,q)$-torus knot $K$, with annulus on the defining torus, and possibly with one or both of the core curves of the solid tori to either side of the torus, denoted $\alpha_1$ and $\alpha_2$, included as well.
The Hopf link complement does not contain any incompressible punctured spheres (where here, we do not include punctured disks as punctured spheres). So $L'$ is a torus knot with possibly two additional components as described.
If neither core curve is present, so we are in a torus knot complement, then there are no incompressible punctured spheres to play the role of the surfaces in $\bS'$ (\cite {Tsau}). In the case that $\alpha_1$ and/or $\alpha_2$ are present, the $(p,q)$-torus knot $K$ lies on a torus $T$ defined by the annulus $A$, with a solid torus or a cored solid torus to each side.
Let $\Sigma$ be one of the incompressible punctured spheres in $\bS'$. Then after isotopy, $A$ cannot intersect $\Sigma$ in a trivial curve. This implies that $T$ intersects $\Sigma$ in a collection of nontrivial parallel curves on $T$. At least two of these curves must be innermost on the sphere $\Sigma$, and therefore those two must correspond to either a pair of meridianal disks to the inside of $T$ or a pair of longitudinal disks to the outside of $T$. Thus, intersection curves in $\Sigma \cap T$ must either be all meridians or all longitudes. In either case, in the two solid tori to either side of $T$ in $S^3$, an orientable incompressible properly embedded planar surface that is a component of $\Sigma \setminus (\Sigma \cap T)$ must either be a disk or an annulus that is boundary-parallel. And there must be at least one such annulus. Choose such an annulus that is outermost in the solid torus. A boundary component of that annulus intersects the torus knot at least $q$ times if its boundary is a meridian and $p$ times if its boundary is a longitude. But in either case, there will be an arc of the torus knot that is isotopic relative its endpoints into the annulus. This implies that the sphere $\Sigma$ is not boundary-incompressible or for that matter, incompressible, a contradiction.
\end{proof}
\end{comment}
We gain the following useful corollary.
\begin{corollary}\label{Saucer composition corollary}
Suppose $\{\cT_j\}$ is a collection of $n$ $2m$-hyperbolic saucer tangles such that the composition $\cT_i \circ \cT_{i+1}$ is defined for all $i$.
Then, $\cT_1 \circ \cT_2 \circ \cdots \circ \cT_n$ is $2m$-hyperbolic.
\end{corollary}
\begin{proof}
Theorem \ref{Saucer hyperbolicity theorem} implies that each $\cT_j$ is $2mn$-hyperbolic, hence Corollary \ref{General composition corollary} implies that the composition is $2m$-hyperbolic.
\end{proof}
The analogy to tangles in $T \times I$ carries to the following theorems.
\begin{theorem}\label{Bracelet theorem2}
Suppose $L$ is a bracelet link made of a cycle $(\cT_i)_{i=1}^{n}$ of $n \geq 2m$ saucer tangles such that each $\cT_i$ is $2m$-hyperbolic.
Then, $L$ is hyperbolic.
If $n = 2m$, then the volume satisfies
\[
\vol(L) \geq \sum_i \volsphere{2n}{\cT_i}.
\]\qed
\end{theorem}
\begin{theorem}\label{tetrahedral hyperbolicity theorem}
Suppose $\cT$ is a tetrahedral tangle that is $(2m, 2n)$-hyperbolic for some $m, n \geq 1$.
Then, $\cT$ is $(2r, 2s)$-hyperbolic for all $r \geq m$ and $s \geq n$.
\end{theorem}
\begin{proof} The $(2m, 2n)$-replicant of $\cT$ can be decomposed into $2m$ saucer tangles, each consisting of $n$ copies of $\cT$ and $n$ copies of its reflection. But then the fact the $(2m, 2n)$-replicant of $\cT$ is hyperbolic is equivalent to the fact the $2n$-replicant of the saucer tangle is hyperbolic. But then we know the $2s$-replicant of the saucer tangle is hyperbolic by the theorem, so $\cT$ is $(2m, 2s)$-hyperbolic. We can repeat the process, only this time decomposing into $2s$ saucer tangles, each consisting of $m$ copies of $\cT$ and $m$ copies of its reflection. Again the theorem implies that this saucer tangle is $2r$-hyperbolic, implying that $\cT$ is $(2r, 2s)$-hyperbolic.
\end{proof}
\section{Explicit Tangles} \label{explicittangles}
In the case of bracelet links, $2$-hyperbolicity of some classes of saucer tangles have been previously studied. We consider one such class here.
\begin{figure}[htbp]
\centering
\includegraphics[width=.8\textwidth]{Figures/rational_examples.pdf}
\caption{Pictured are some particular examples of arborescent tangles.
The integer tangles are defined analogously to 3, with tangle $m$ having $m$ crossings.
The tangle $Q_m$ has $m$ ``loops.''
}
\label{Arborescent examples figure}
\end{figure}
\begin{figure}[htbp]
\centering
\includegraphics[width=.8\textwidth]{Figures/rational_operations.pdf}
\caption{Pictured are the four operations defined by \cite{Conway} to generate arborescent tangles.
We realize arborescent tangles as saucer tangles with the endpoints on the far left or far right, and hence composition of arborescent tangles corresponds with the sum operation.}
\label{Arborescent operations figure}
\end{figure}
\subsection{Arborescent tangles}\label{Arborescent tangles subsection}
Rational tangles, as defined by Conway in \cite{Conway} are examples of saucer tangles. We can think of a rational tangle as constructed by starting with either the 0 tangle corresponding to two horizontal strands or the $\infty$ tangle corresponding to two vertical strands.Then we alternately twist the right two endpoints and the the bottom two endpoints to create a link in the 3-ball. Note that when we start, there is a disk $D$ known as the \emph{defining disk} of the rational tangle, that is properly embedded in the ball and that separates the two strands. It continues to exist as we twist the endpoints, but its boundary becomes more complicated on the boundary of the ball.
Given a rational tangle in a ball, we can choose a disk $D'$ on the boundary of the ball that contains two of the endpoints of the tangle, called a \emph{composition disk}. Then given two such rational tangles with choices of disks on the boundary, we can compose the tangles by gluing the one disk to the other.
A Montesinos tangle is obtained by composing rational tangles in a row, left to right, our composition disks either being the leftside disk or the rightside disk. Arborescent tangle are defined to include rational tangles and the compositions of arborescent tangles along any composition disks.
We use previous results to determine the $2n$-hyperbolicity of all arborescent tangles.
Let $Q_m$ be the arborescent tangles defined in Figure \ref{Arborescent examples figure}.
We say that an arborescent tangle $\cT$ \emph{contains $Q_m$} if there is a 4-times punctured sphere in $D \times I \setminus \cT$, called a \emph{Conway sphere}, cutting $\cT$ into two tangles, one of which is $Q_m$.
We say that a link contains $Q_m$ if it possesses such a Conway sphere.
The following Theorem was proved in \cite[Thm. 3.22]{Volz}.
It is related to a theorem originally in an unpublished manuscript by Bonahon and Siebenmann (see \cite{Bonahon}), also proved in \cite{Volz} and \cite{Futer} by essentially different means.
\begin{theorem}[\cite{Volz}]\label{Arborescent theorem}
Suppose $\cT$ is an arborescent tangle.
Then, $\cT$ is $2$-hyperbolic if and only if none of the following hold:
\begin{enumerate}
\item $\cT$ is a rational tangle.
\item $\cT = Q_m \circ \cT'$ or $\cT = \cT' \circ Q_m$ for some $m \geq 1$ and some arborescent tangle $\cT'$.
\item $\cT$ contains $Q_m$ for some $m \geq 2$.
\end{enumerate}\qed
\end{theorem}
Note that a rational tangle $\cT$ can never be 2-hyperbolic since the defining disk doubles to a separating sphere for the resulting link.
We will use this theorem and Corollary \ref{General composition corollary} to determine $2n$-hyperbolicity.
In order to state the result, we define a saucer tangle $\cT$ to be \emph{principally $2n$-hyperbolic} if it is $2n$ hyperbolic and either $n=1$ or $\cT$ is not $2(n-1)$-hyperbolic.
We say that a saucer tangle is \emph{entirely non-hyperbolic} if it is not $2n$-hyperbolic for any $n$.
\begin{corollary}\label{Arborescent corollary}
Suppose $\cT$ is an arborescent tangle.
Then, $\cT$ is entirely non-hyperbolic if and only if one of the following hold:
\begin{enumerate}
\item $\cT$ is an integer tangle.
\item $\cT = Q_m \circ \cT'$ or $\cT = \cT' \circ Q_m$ for some $m \geq 1$.
\item $\cT$ contains $Q_m$ for some $m \geq 2$.
\end{enumerate}
Further, $\cT$ is principally 2-hyperbolic if and only if none of these hold and $\cT$ is non-rational.
Moreover, $\cT$ is principally 4-hyperbolic if and only if it is a non-integer rational tangle other than $Q_{\text{clasp}}.$
Last, $\cT$ is principally 6-hyperbolic if and only if it is equal to $Q_{\text{clasp}}.$
\end{corollary}
We use the fact (see \cite{Kauffman_classification}), that a sum of two rational tangles is rational if and only if one of the two is an integer tangle.
\begin{proof}[Proof of Corollary \ref{Arborescent corollary}]
We first assume that $\cT$ is an arborescent tangle satisfying one of (1)-(3), and we show that $\cT \circ \cT^R \cdots \widetilde \cT$ satisfies one of the conditions of Theorem \eqref{Arborescent theorem}, where $\widetilde \cT$ stands for either $\cT$ or $\cT^R$ depending on the parity.
Suppose that $\cT$ is an integer tangle.
Then, by the preceding discussion, we see inductively that
\[
\cT \circ \cT^R \cdots \widetilde \cT = \cT \circ \prn{\cT^R \cdots \widetilde \cT}
\]
is a rational tangle (in fact an integer tangle).
Hence Theorem \ref{Arborescent theorem} implies that no such tangles are 2-hyperbolic, so \ref{General composition corollary} implies that $\cT$ is entirely non-hyperbolic.
Suppose instead that $\cT = Q_m \circ \cT'$ or $ \cT' \circ Q_m$ for some $m \geq 1$.
Then, $\cT^R \circ \cT \cdots \widetilde \cT$ contains $Q_{2m}$, so we may assume that condition (3) holds.
In fact, if condition (3) holds for $\cT$, then it holds for $\cT \circ \cT^R \cdots \widetilde \cT$, so Theorem \ref{Arborescent theorem} and Corollary \ref{General composition corollary} together imply that $\cT$ is entirely non-hyperbolic.
The corollary for principal 2-hyperbolicity is a restatement of Theorem \ref{Arborescent theorem}.
For 4-hyperbolicity, note that $\cT$ is not an integer tangle, so $\cT^R$ is also not an integer tangle.
Hence $\cT \circ \cT^R$ is not a rational tangle.
Since $Q_m$ contains more than two components, any rational tangle satisfies conditions (2) and (3).
Further, conditions (2), (3) together imply that $\cT$ neither contains $Q_m$ for $m \geq 2$ nor attains equality with $Q_m \circ T$ or $T \circ Q_m$ for $m \geq 1$.
Hence $\cT \circ \cT^R$ is 2-hyperbolic and $\cT$ is not 2-hyperbolic.
Hence $\cT$ is principally 4-hyperbolic.
The corollary is well-known for $Q_{\text{clasp}}$ (see for instance \cite{NR}).
\end{proof}
\subsection{Alternating tangles} \label{alternatingtangles}
We now explore a class of 2-hyperbolic solid cylindrical tangles related to alternating saucer tangles.
\begin{definition}
A saucer tangle $\cT$ is an \emph{alternating tangle} if it possesses a projection to a cross-sectional bigon of the saucer such that the crossings alternate along every strand.
\end{definition}
Since they are topologically indistinguishable, a saucer tangle $\cT$ determines a canonical solid-cylinder tangle $\cT_{C}$.
We show the following:
\begin{proposition}\label{Hyperbolic alternating tangle proposition}
Suppose $\cT$ is a $2m$-hyperbolic alternating saucer tangle which intersects each end at two points.
Then, $\cT_{C}$ is a 2-hyperbolic solid-cylinder tangle with volume satisfying
\[
\volsolid{2}{\cT_C} > \volsphere{2m}{\cT}.
\]
\end{proposition}
We use the following construction.
\begin{definition}
A saucer tangle $\cT$ determines another saucer tangle $\widehat \cT$ with a projection identical to that of $\cT$ except that all crossings are switched between under-crossings and over-crossings.
Using this, the length $2m$ cycle $\prn{\cT,\widehat \cT^R,\dots,\widehat \cT^R}$ determines an alternating bracelet link $D^{2m}_{\text{mod}}(\cT)$, called the \emph{modified $2m$-replicant}.
\end{definition}
If $\cT$ is $2m$-hyperbolic, then $\widehat \cT$ is as well, and by Theorem \ref{Bracelet theorem}, the link$D^{2m}_{\text{mod}}\prn{\cT}$ is also hyperbolic, with volume satisfying
\[
\vol\prn{D^{2m}_{\text{mod}}{\cT}} > 2m \volsphere{2m}{\cT}
\]
\begin{proof}[Proof of Proposition \ref{Hyperbolic alternating tangle proposition}]
The link $D^{2m}_{\text{mod}}(\cT)$ is alternating by construction. Since it is hyperbolic, it must be nonsplit, prime and it must not be a torus link.
Using this, we may construct a \emph{generalized augmented alternating link} by adding in a single unknotted component along the central axis of the bracelet link $D^{2m}_{\text{mod}}(\cT)$.
By \cite{Adams_generalized_augmented}, this link is hyperbolic, and by \cite[Theorem 3.4(c)]{Thurston_kleinian} , its volume is greater than the volume of $D^{2m}_{\text{mod}}(\cT)$.
In fact, the complement of this link is homeomorphic to that of the link in the open solid torus determined by $\prn{\cT_C,\widehat \cT^R_C,\dots,\widehat \cT^R_C}$.
Call this link $D^{2m}_{\text{mod}}(\cT_C)$.
There is a collection of disjoint twice-punctured meridianal disks in the open solid torus which separate the copies of $\cT_C$ and $\widehat \cT^R_C$ in $D^{2m}_{\text{mod}}(\cT_C)$ from the rest of the link.
By \cite{Adams_thrice_punctured}, these twice-punctured disks are totally geodesic, and cutting and then regluing along any subset of them will yield another hyperbolic manifold. In particular, $\cT_C$ is 2-hyperbolic.
Since $\volsolid{2}{\cT_C} = \volsolid{2}{\widehat \cT_C^R}$, Theorem \ref{Solid torus theorem} implies that $\vol\prn {D^{2m}_{\text{mod}}(\cT_C)} = 2m \volsolid{2}{\cT_C}$.
We may summarize these inequalities as follows:
\begin{align*}
\volsphere{2m}{\cT}
&= \frac{\vol\prn{D^{2m}(\cT)}}{2m} \\
&\leq \frac{\vol\prn{D^{2m}_{\text{mod}}(\cT)}}{2m}\\
& < \frac{\vol\prn{D^{2m}_{\text{mod}}(\cT_C)}}{2m}\\
&= \volsolid{2}{\cT_C}.
\end{align*}
\end{proof}
Note that because the twice-punctured meridianal disks are totally geodesic, the link $J$ obtained by taking a single copy of $\cT$, and gluing its left endpoints to its right endpoints (generating a ``bracelet link'' of length 1) and then removing a single unknotted component corresponding to the axis of the bracelet link is hyperbolic and its volume is equal to $\volsolid{2}{\cT_C}$. A corollary applies to Montesinos tangles.
\begin{corollary}
Let $\cT_C$ be a non-integer Montesinos tangle in a solid cylinder.
Then $\cT_C$ is 2-hyperbolic.
\end{corollary}
\begin{proof}
Let $\cT$ be the corresponding Montesinos tangle in a saucer.
Then, $\cT$ decomposes into a composition of non-integer rational tangles (by including an integer tangle on the tail end of the preceding rational tangle if necessary).
By Corollary \ref{Arborescent corollary} these are $2m$-hyperbolic for some $m$, and by \cite{Kauffman_classification} they are alternating.
Since $\cT_C \circ \cT'_C = (\cT \circ \cT')_C$, this implies that $\cT_C$ decomposes into $2$-hyperbolic solid cylindrical tangles.
Then, Theorem \ref{Main theorem inductive} implies that $\cT_C$ is 2-hyperbolic, as desired.
\end{proof}
\section{Application to Other Tangle Reflection Graphs}\label{tanglereflectiongraphs}
Up to now, we have used the cyclic graph on $2n$ vertices or the $(2n, 2m)$-torus lattice graph to create links. We defined the replicants of a given tangle $\cT$ in terms of the link obtained by alternately reflecting the tangle $\cT$ around the graph.
We now generalize the construction. Let $G$ be a finite connected $r$-valent bipartite graph with bipartition of the vertices into two sets $U$ and $V$ such that $G$ is embedded in an orientable 3-manifold $M$. Let $\bS$ a collection of surfaces in $M$, each intersecting $G$ transversely at a subset of the edges such that for each such surface $\Sigma$, there is an orientation-reversing involution of $M$, which we call a reflection, with fixed point set $\Sigma$. Further, assume that for each pair of vertices sharing an edge, there is a distinguished such reflection that switches the two vertices, sends $G$ back to itself, switches $U$ and $V$ and sends the collection $\bS$ back to itself.
Further assume that in the finite group $\mathcal{G}$ generated by the collection of reflections, the only element that fixes a vertex is the trivial element.
Then we call $(M, G,\bS)$ a {\it reflection graph} of $M$.
The group $\mathcal{G}$ is fully determined as the group of isometries of $M$ generated by reflections about $\bS$.
Let $L$ be a link formed by inserting tangles $\cT_1, \cT_2, \dots, \cT_{2n}$ into the $2n$ vertices of $G$ such that the nonzero number of outgoing strands along each edge match up from the two vertices. For each such tangle $\cT$, we can form the $G$-replicant $D^G(\cT)$ of the tangle by removing all tangles except for it and then using the reflections to reflect $\cT$ across each outgoing edge. Then we repeat the process with all subsequent edges.
To see that the definition of the replicant is well-defined, we need to show that there are $r$ distinct edge classes at each vertex under the action of $\mathcal{G}$. Suppose not, and that there is a nontrivial element $\mu$ of $\mathcal{G}$ that identified two edges that share a vertex $y$. Then since $\mu$ cannot fix $y$, it must send the other endpoint of the first edge, call it $x$ to $y$ and $y$ to the other endpoint of the second edge, call it $z$.
But because $y$ and $z$ share an edge, there must be a reflection $\lambda$ that sends $y$ to $z$ and $z$ to $y$. But then $\lambda \circ \mu(y) = y$. hence, $\lambda \circ \mu$ must be the identity so $\mu = \lambda^{-1} = \lambda$. But $\lambda$ sends $z$ to $y$ whereas $\mu$ sends $x$ to $y$, a contradiction.
The fact there are exactly $r$ equivalence classes of edges follows by reflecting to get from one vertex to an adjacent one. Then the $r$ equivalence classes with representatives at the first vertex will generate representatives of the same classes at the next vertex. Repeating for adjacent vertices will eventually make this true at all vertices.
Since the graph is bipartite, each cycle is even and this construction is well-defined. If the complement of the $G$-replicant of $\cT$ is hyperbolic, we say $\cT$ is a $G$-hyperbolic tangle with volume $\vol(D^G(\cT))/ |G|$.
If we have a combinatorial proof that a collection of $2n$ copies of the $G$-replicants for the tangles can be cut and reglued, always along totally geodesic surfaces, in order to obtain $2n$ copies of $M \setminus L$, then we know that the analog of Theorem \ref{Bracelet theorem} applies. That is to say that under these circumstances, if each tangle we insert at the vertices of the graph is $G$-hyperbolic, then the resulting link is hyperbolic with volume greater than the sum of $1/2n$ times the volumes of the $G$-replicas. Under these circumstances, we say that $(M, G, \bS)$ is a {\it tangle reflection graph} for $M$.
\begin{theorem} \label{tangle reflection graph} Let $(M, G, \bS)$ be a tangle reflection graph with $2n$ vertices in a compact orientable 3-manifold $M$ with possible torus boundaries.
If there is a orientation-reversing symmetry $\phi$ of $M$ with fixed point set an orientable surface $S$, such that $G$ lies in $S$ and $\phi$ preserves each surface in $\bS$, then there exists a new tangle reflection graph $(M, G', \bS')$ for $M$, where $G'$ is the product of $G$ with the graph $P_1$ consisting of one edge and two vertices and $\bS'$ is the collection $\bS$ together with the surface $S$.
\end{theorem}
\begin{proof}
Push $G$ off $S$ in a choice of normal direction, and call that copy of $G$ the upper copy. Take a second copy of $G$ that is the reflection of $G$ by $\phi$, and call that the lower copy. Then we connect the corresponding vertices of the two copies of $G$ by edges that are perpendicular to $S$, which we call {\it vertical edges}. This yields an embedding of $G'$ in $M$ such that $(M, G', \bS')$ is a reflection graph.
It remains to show that $(M, G', \bS')$ is a tangle reflection graph. Let $L'$ be a link formed from $(G', \bS')$, with tangles $\cT_1, \dots \cT_{2n}$ in the upper copy of $G$ and tangles $\cT_{2n+1} , \dots, \cT_{4n}$ in the lower copy of $G$.
We form two new links $L''$ and $L'''$ from $(G',\bS')$ such that the first consists of $\cT_1, \dots \cT_{2n}$ in the upper copy of $G$, and then their reflections across the vertical edges in the lower copy of $G$. The second consists of $\cT_{2n+1} , \dots, \cT_{4n}$ in the lower copy of $G$ and then their reflections across the vertical edges in the upper copy of $G$. Treating each pair of tangles along a vertical edge as a single tangle, we know that because $(G, \bS)$ is a tangle reflection graph, we can form the $G$-replicants of each of the now $2n$ tangles in $L''$ and then by cutting along totally geodesic surfaces and regluing, we can create $2n $ copies of $L''$ in $M$. Similarly, from the $2n$-replicants of the tangles in $L'''$ we can create $2n$ copies of $L'''$ in $M$. In each case, the sum of the volumes of the $2n$ copies of the complements of the links have volumes at least as great as the sum of the volumes of the $G$-replicants. But each of these links has a reflection surface such that reflecting across it switches the two copies of each $\cT_i$ within it. So we can cut the manifolds open along the totally geodesic surfaces and reglue to obtain $4n$ copies of $L$ in $M$. Hence the theorem follows.
\end{proof}
The following examples are immediate from the theorem.
Our discussion of bracelet links showed that a cycle graph $C_{2n}$ with $2n$ vertices is a tangle reflection graph in $S^3$.
That graph can be isotoped to live on a sphere $S$ that is the fixed-point set for a reflection symmetry of the 3-sphere. Therefore Theorem \ref{tangle reflection graph} implies that we obtain a prism graph with $4n$ vertices that is a tangle reflection graph. Note that each vertex in the graph is trivalent so our tangles have three bundles of strands exiting.
We can reapply Theorem \ref{tangle reflection graph} to the prism graph, since it also can be isotoped to live on a sphere.
Thus, we generate a hyperprism graph of $8n$ vertices, as in Figure \ref{hyperprism}, which must also be a tangle reflection graph. Of course, this last example is also a $(2n, 4)$-torus lattice graph, so we already knew that it was a tangle reflection graph.
\begin{figure}[htbp]
\centering
\includegraphics[width=.7\textwidth]{Figures/tanglereflectiongraphs.pdf}
\caption{That the cycle graph is a tangle reflection graph in $S^3$ implies the prism graph is also, which implies the hyperprism graph is as well.}
\label{hyperprism}
\end{figure}
Unlike what happened with the cycle graph, in these graphs, we can have any positive number of strands running along an edge, including a single strand.
Another example is generated by the $(2m, 2n)$-torus lattice graph, denoted $G$, which we know to be a tangle reflection graph. If we take that graph to lie in either the intermediate torus in $T \times (0,1)$ or to lie in the torus splitting $S^2 \times S^1$ into two solid tori, then there is a reflection of the manifold that preserves the torus and therefore the graph. Hence, Theorem \ref{tangle reflection graph} implies that the graph obtained as a product of $G$ with $P_1$ is also a tangle reflection graph in those manifolds.
As further potential tangle reflection graphs, we can look to the 1-skeleta of the Archimedean solids.
Three of these solids are relevant to this construction since the only cycles in the 1-skeleta have even length, namely the truncated octahedron, truncated cuboctahedron, and truncated icosidodecahedron.
We observe that in $S^3$, the graphs determined by the $1$-skeleta of these three solids are reflection graphs. We expect that volume bounds similar to those proven for simpler configurations would hold for links in $S^3$ with tangles at the vertices of these Archimedean solids, and connecting strands along the edges. However, the combinatorics quickly becomes difficult.
But note that if these can be shown to be tangle reflection graphs, then Theorem \ref{tangle reflection graph} implies that their products with $P_1$ would also be tangle reflection graphs.
\section{Volume Calculations} \label{volumecalculations}
In this section we give tables of volumes of tangles living in various objects, all calculated using the SnapPy computer program \cite{SnapPy}.
A \emph{rational square tangle} is defined to be a rational tangle projected onto a square, having one endpoint at the midpoint of each edge of the square. Table \ref{rational square tangle} gives the volumes of some such tangles in a variety of contexts. The first column specifies the object in which the rational square tangle is embedded. The second column specifies the manifold $M$. The third column specifies the $m$ for the $(2, 2m)$-replicant we perform on the tangle when finding its volume. The subsequent columns correspond to particular rational tangles given in Conway notation, where the first four are integer tangles.
Note that on the standardly embedded torus, $(2, 2m)$-replicant means $2$ columns and $2m$ rows. We do not give the $(2m,2)$-replicant volumes as in $T \times I$, $\mbox{vol}_{T \times I}^{(2m, 2n)}(\cT)= \mbox{vol}_{T \times I}^{(2n, 2m)}(\cT)$. In $D^2 \times S^1$, $\mbox{vol}_{D^2 \times S^1}^{(2m, 2)}(\cT) = m \mbox{vol}_{D^2 \times S^1}^{(2, 2)}(\cT)$. And in $S^3$, $\mbox{vol}_{S^3}^{(2m, 2n)}(\cT)$ gives the same volume as $\mbox{vol}_{S^3}^{(2n, 2m)}(\cT')$, where $\cT'$ is obtained from $\cT$ by rotating $90^\circ$ and then changing all crossings. In the case of the tangles listed, this generates a tangle that is the reflection of the initial tangle, and hence the volumes are equal.
As a note, caution should be exercised. For instance, in $S^3$ and $D^2 \times S^1$, both the integer 2-tangle and the integer 3-tangle are not $(2m,2)$-hyperbolic for any positive integer $m$ for one choice of crossings in the tangle and are $(2m,2)$-hyperbolic for all positive $m$ for the other choice of crossings. See Example \ref{tettangleexample} for the case where the 2-tangle is $(2m,2)$-hyperbolic for all positive $m$. In $S^2 \times S^1$, the integer 2-tangle and 3-tangle are not $(2m,2)$-hyperbolic for any choice of crossings.
\begin{table}[htbp]
\centering
\caption{$\mbox{vol}_M^{(2, 2m)}$ of rational square tangles}
\begin{tabular}{|c|c|c||r|r|r|r|r|}
\hline
Embedded in a & Link is in $M$ & $m$ & \multicolumn{1}{|c|}{2} & \multicolumn{1}{|c|}{3} & \multicolumn{1}{|c|}{4} & \multicolumn{1}{|c|}{5} & \multicolumn{1}{|c|}{2 1} \\ \hline \hline
cube & $T \times I$ & 1 & 4.3692852 & 6.26446133 & 7.37277416 & 8.05714498 & 7.48167784 \\ \hline
\multirow{3}{*}{triangular prism} & \multirow{3}{*}{$D^2 \times S^1$} & 1 & 3.66386238 & 5.07470803 & 5.82605812 & 6.26446133 & 6.57223395 \\
& & 2 & 4.20148819 & 5.97374285 & 6.98517306 & 7.59952104 & 7.26467498 \\
& & 3 & 4.2953372 & 6.92143228 & 7.20028164 & 7.85266168 & 7.38599241 \\ \hline
\multirow{3}{*}{tetrahedron} & \multirow{3}{*}{$S^3$} & 1 & 3.13223067 & 4.3062076 & 4.90649786 & 5.24442274 & 5.81283664 \\
& & 2 & 3.538345513 & 4.89348601 & 5.60919297 & 6.02345588 & 6.39075686 \\
& & 3 & 3.60861463 & 5.53828532 & 5.73062363 & 6.1583781 & 6.4921938 \\ \hline
\end{tabular}
\label{rational square tangle}
\end{table}
We may define a \emph{rational thickened-cylinder tangle} to be obtained by having one endpoint of a rational tangle at the midpoint of each side of a square, thickening the square and then gluing the north and south pair of opposite sides of the thickened square. See Table \ref{integer cylindrical tangles} for $\mbox{vol}_{T \times I}^2$ of example integer thickened-cylinder tangles (c.f. Definition \ref{Thickened torus tangle definition}). Note that when we have rational thickened-cylinder tangles, since the tangles only intersect the left and right boundaries of the thickened cylinder once, these left and right boundaries become once-punctured annuli. Therefore they are totally geodesic in the hyperbolic structure of the resulting link complement in the thickened torus. Hence, the 2-volumes of the constituent tangles add to give the volume of the resulting link exactly.
\begin{table}[htbp]
\centering
\caption{$\mbox{vol}_{T \times I}^2$ of integer cylindrical tangles}
\begin{tabular}{|c|c|c|c|}
\hline
2 & 3 & 4 & 5 \\ \hline \hline
5.33348957 & 7.32772475 & 8.35550215 & 8.92931782 \\ \hline
\end{tabular}
\label{integer cylindrical tangles}
\end{table}
We may also define a \emph{reciprocal integer saucer tangle} to be an integer tangle with NE and NW strands going out one side of a square cross section of a saucer, and SE and SW strands going out the other.
Note that $D^{2m}(p/q) = D^{2m}(-p/q)$, so it suffices to consider these for nonnegative integers.
The reciprocal saucer tangle $1/0 = \infty$ and $1/1 = 1$ are entirely non-hyperbolic.
Further, Corollary \ref{Arborescent corollary} implies that $1/2 = Q_{\text{clasp}}$ is principally $6$-hyperbolic and $1/n$ is principally $4$-hyperbolic for all $n > 2$.
Table \ref{reciprocal integer saucer tangle} gives the beginning of a table of $\mbox{vol}^{2m}_{S^3}$ for integer reciprocal saucer tangles.
Note that $\mbox{vol}^{2m}_{S^3}(\cT)$ of the reciprocal integer saucer tangle $1/2$ corresponds to $\frac{1}{2n}$ the volume of the minimally twisted $2m$-chain link.
\begin{table}[htbp]
\centering
\caption{$\mbox{vol}^{2m}_{S^3}$ of reciprocal integer saucer tangles}
\begin{tabular}{|c||r|r|r|r|}
\hline
$m$ & \multicolumn{1}{|c|}{1/2} & \multicolumn{1}{|c|}{1/3} & \multicolumn{1}{|c|}{1/4} & \multicolumn{1}{|c|}{1/5} \\ \hline \hline
2 & 0 & 3.13223067 & 4.30620760 & 4.90649786 \\
3 & 2.44257492 & 4.45025692 & 5.38411452 & 5.86524434 \\
4 & 3.01152301 & 4.85098130 & 5.72360375 & 6.17292141\\
5 & 3.25515403 & 5.28582982 & 5.87549250 & 6.31134186\\
\hline
\end{tabular}
\label{reciprocal integer saucer tangle}
\end{table}
We note that for a reciprocal integer saucer tangle, by Proposition \ref{Hyperbolic alternating tangle proposition}, $\vol\prn{D^{2m}(\cT)}$ is less than the volume $\volsolid{2}{\cT_C}$. In fact, as $m$ increases, the volumes will approach the upper bound from below. This follows because $\volsolid{2}{\cT_C} = \volsolid{2m}{\cT_C}$ since $D^{2m}(\cT_C)$ is an $m$-fold cyclic cover of $D^2(\cT_C)$. And $D^{2m}(\cT)$ comes from surgery on $D^{2m}(\cT_C)$, with surgery coefficients increasing with $m$.
Hence, in Table \ref{limitvolumes}, we include some of these limit volumes as well. So each column in Table \ref{reciprocal integer saucer tangle} has values approaching the corresponding entry in Table \ref{limitvolumes}. Note that the limit volumes in Table \ref{limitvolumes} themselves limit to 7.32772474 from below, which is the volume of the Borromean rings. So all $2m$-replicants of reciprocal integer tangles have volume less than 7.32772474.
\begin{table}[htbp]
\centering
\caption{$\volsolid{2}{\cT_C}$ for integer reciprocal tangles}
\begin{tabular}{|c|c|c|c|}
\hline
1/2 & 1/3 & 1/4 & 1/5 \\ \hline
3.66386237 & 5.33489567 & 6.13813879 & 6.5517432897 \\ \hline
\end{tabular}
\label{limitvolumes}
\end{table}
\printbibliography
\end{document}
|
\section{Introduction}
The $k$-center problem is one of the most fundamental problems in clustering. The input to the $k$-center problem consists of a finite metric space $(X, d)$, where $X$ is a set of $n$ points, and $d: X \times X \to \mathbb{R}^+$ is the associated distance function satisfying triangle inequality. We are also given a parameter $k$, where $1 \le k \le n$. A solution to the $k$-center problem consists of a set $C \subseteq X$ of size at most $k$, and the cost of this solution is $\max_{p \in X} d(p, C)$, i.e., the maximum distance of a point to its nearest center in $C$. Alternatively, a solution can be thought of as a set of $k$ balls of radius $\max_{p \in X} d(p, C)$, centered around points in $C$, that covers the entire set of points $X$. The goal is to find a solution of smallest radius. We say that a solution $C'$ is an $\alpha$-approximation, if the cost of $C'$ is at most $\alpha$ times the optimal radius. Several $2$-approximations are known for the $k$-center problem \cite{hochbaumS1985best,hochbaumM1985approximation}. A simple reduction from the Minimum Dominating Set problem shows that the $k$-center problem is \textsf{NP}-hard. In fact, the same reduction also shows that it is \textsf{NP}-hard to get a $(2-\epsilon)$-approximation for any $\epsilon > 0$.
Several generalizations of the vanilla $k$-center problem have been considered in the literature, given its fundamental nature in the domain of clustering and approximation algorithms. One natural generalization is the \emph{Robust $k$-center} or \emph{$k$-center with outliers} problem, where we are additionally given a parameter $m$, and the goal is to find a solution that covers at least $m$ points of $X$. Note that the remaining at most $n-m$ points can be thought of as outliers with respect to the clustering computed. Charikar et al.\ \cite{charikar2001algorithms}, who introduced this problem, showed that a simple greedy algorithm gives a $3$-approximation for the problem. Subsequently, the approximation guarantee was improved by \cite{CGK20,harris2017lottery}, who gave a $2$-approximation, which is optimal in light of the aforementioned $(2-\epsilon)$-hardness result.
The focus of our paper is the Non-Uniform $k$-Center (NU$k$C\xspace), which was introduced by Chakrabarty et al.\ \cite{CGK20}. A formal definition follows.
\begin{definition}[$t$-NU$k$C\xspace] \label{def:nukc}
The input is an instance $\mathcal{I} = ((X, d), (k_1, k_2, \ldots, k_t), (r_1, r_1, \ldots, r_t))$, where $r_1 \ge r_2 \ge \ldots r_t \geq 0$, and the $k_i$ are positive integers. The goal is to find sets $C_i \subseteq X$ for $1 \le i \le t$, such that $|C_i| \le k_i$, and the union of balls of radius $\alpha r_i$ around the centers in $C_i$, over $1 \le i \le t$, covers the entire set of points $X$. The objective is to minimize the value of the dilation factor $\alpha$.
\end{definition}
In the Robust $t$-NU$k$C\xspace problem, we are required to cover at least $m$ points of $X$ using such a solution. We note that the special case of (Robust) $t$-NU$k$C\xspace with $t = 1$ corresponds to the (Robust) $k$-center problem. Chakrabarty et al.\ \cite{CGK20} gave a bicriteria approximation for $t$-NU$k$C\xspace for arbitrary $t$, i.e., they give a solution containing $O(k_i)$ balls of radius $O(r_i)$ for $1 \le i \le t$. They also give a $(1+\sqrt 5)$-approximation for $2$-NU$k$C\xspace. Furthermore, they conjectured that there exists a polynomial time $O(1)$-approximation for $t$-NU$k$C\xspace for constant $t$. Subsequently, Chakrabarty and Negahbani \cite{ChakrabartyN21} made some progress by giving a $10$-approximation for Robust $2$-NU$k$C\xspace. Very recently, Jia et al.\ \cite{jia2021towards} showed an approximate equivalence between $(t+1)$-NU$k$C\xspace and Robust $t$-NU$k$C\xspace, thereby observing that the previous result of \cite{ChakrabartyN21} readily implies a $23$-approximation for $3$-NU$k$C\xspace. We note that the techniques from Inamdar and Varadarajan \cite{inamdar2020capacitated} implicitly give an $O(1)$-approximation for $t$-NU$k$C\xspace for any $t\ge 1$, in $k^{O(k)} \cdot n^{O(1)}$ time, where $k = \sum_t k_t$. That is, one gets an \emph{FPT approximation}. Finally, we also note that Bandyapadhyay \cite{Bandyapadhyay20nukc} gave an exact algorithm for perturbation resilient instances of NU$k$C\xspace in polynomial time.
Another related variant of $k$-center is the \emph{Colorful $k$-center} problem. Here, the set of points $X$ is partitioned into $\ell$ \emph{color classes}, $X_1 \cup X_2 \cup \ldots \cup X_\ell$. Each color class $X_j$ has a coverage requirement $m_j$, and the goal is to find a set of $k$ balls of smallest radius that satisfy the coverage requirements of all the color classes. Note that this is a generalization of Robust $k$-center to multiple types of coverage constraints. Bandyapadhyay et al.\ \cite{Bandyapadhyay0P19} introduced this problem, and gave a \emph{pseudo-approximation}, i.e., their algorithm returns an $2$-approximate solution using at most $k+\ell-1$ centers. Furthermore, they managed to improve this to a true $O(1)$-approximation in the Euclidean plane for constant number of color classes. Subsequently, Jia et al.\ \cite{JiaSS20} and Anegg et al.\ \cite{AneggAKZ20} independently gave (true) $3$ and $4$-approximations respectively for the Colorful $k$-center (with constant $\ell$) in arbitrary metrics.
\vspace{0.4cm}\paragraph{Our Results and Techniques.}
Our main result is an $O(1)$-approximation for $4$-NU$k$C\xspace. We obtain this result via a sequence of reductions; some of these reductions are from prior work while some are developed here and constitute our main contribution. Along the way, we combine various tools and techniques from the aforementioned literature of Robust, Colorful, and Non-Uniform versions of $k$-center.
First, we reduce the $4$-NU$k$C\xspace problem to the Robust $3$-NU$k$C\xspace problem, following Jia et al.\ \cite{jia2021towards}. Next, we reduce the Robust $3$-NU$k$C\xspace to \emph{well-separated} Robust $3$-NU$k$C\xspace, by adapting the approach of Chakrabarty and Negahbani \cite{ChakrabartyN21}.\footnote{In this discussion, ``reduction'' refers to a polynomial time (possibly Turing) reduction from problem $A$ to problem $B$, such that (i) a feasible instance of $A$ yields (possibly polynomially many) instance(s) of $B$, and (ii) a constant approximation for $B$ implies a constant approximation for $A$.} In a well-separated instance, we are given a set of potential centers for the balls of radius $r_1$, such that the distance between any two of these centers is at least $c \cdot r_1$, for a parameter $c \geq 2$.
Before describing how to solve Well-Separated Robust $3$-NU$k$C\xspace, we give a sequence of reductions, which constitute the technical core of our paper. First, we show that any instance of Robust $t$-NU$k$C\xspace can be transformed to an instance of ``Colorful'' $(t-1)$-NU$k$C\xspace, where we want to cover certain number of red and blue points using the specified number of balls of $t-1$ distinct radii. Thus, this reduction reduces the number of radii classes from $t$ to $t-1$ at the expense of increasing the number of coverage constraints from $1$ to $2$. In our next reduction, we show that Colorful $(t-1)$-NU$k$C\xspace can be reduced to Colorful $(t-1)$-NU$k$C\xspace with an additional ``self-coverage'' property, i.e., the radius $r_{t-1}$ can be assumed to be $0$. Just like the aforementioned reduction from \cite{jia2021towards}, these two reductions are generic, and hold for any value of $t \ge 2$. These reductions crucially appeal to the classical greedy algorithm and its analysis from Charikar et al.\ \cite{charikar2001algorithms}, which is a tool that has been not been exploited in the NU$k$C\xspace literature thus far. We believe that these connections between Colorful and Robust versions of NU$k$C\xspace are interesting in their own right, and may be helpful toward obtaining a true $O(1)$-approximation for $t$-NU$k$C\xspace for fixed $t$.
We apply these two new reductions to transform Well-Separated Robust $3$-NU$k$C\xspace to \emph{Well-Separated Colorful} $2$-NU$k$C\xspace, with $r_{2} = 0$. The latter problem can be solved in polynomial time using dynamic programming in a straightforward way. Since each of our reductions preserves the approximation factor up to a constant, this implies an $O(1)$-approximation for $4$-NU$k$C\xspace.
Our overall algorithm for $4$-NU$k$C\xspace is combinatorial, except for the step where we reduce Robust $3$-NU$k$C\xspace to Well-Separated Robust $3$-NU$k$C\xspace using the round-or-cut approach of \cite{ChakrabartyN21}. Thus, we avoid an additional ``inner loop'' of round-or-cut that is employed in recent work \cite{ChakrabartyN21,jia2021towards}.\footnote{A by-product of one of our reductions is a purely combinatorial approximation algorithm for colorful $k$-center, in contrast with the LP-based approaches in \cite{Bandyapadhyay0P19,AneggAKZ20,JiaSS20}.}
\subsection{Main Algorithm for $4$-NU$k$C\xspace} \label{sec:main-algo}
Let $\mathcal{I} = ((X, d), (k_1, \ldots, k_4), (r_1, \ldots, r_4))$ be the given instance of $4$-NU$k$C\xspace, which we assume is feasible. First, using the reduction Section \ref{sec:t-nukc-to-robust}, we reduce it to an instance $\mathcal{I}' = ((X, d), (\mathbbm{1}, m) (r'_1, r'_2, r'_3), \allowbreak (k_1, k_2, k_3))$ of Robust $3$-NU$k$C\xspace. Recall that Lemma \ref{lem:tplus1-to-t} implies that $\mathcal{I}'$ is feasible, and furthermore an $O(1)$-approximation for $\mathcal{I}'$ implies an $O(1)$-approximation for $\mathcal{I}$.
Next, we use the round-or-cut framework methodology from \cite{ChakrabartyN21} on the instance $\mathcal{I}'$, as described in Section \ref{sec:t-to-robust}. Essentially, this is a Turing reduction from Robust $3$-NU$k$C\xspace to (polynomially many instances of) \emph{Well-Separated} Robust $3$-NU$k$C\xspace. In a well-separated instance, we are given a set of potential centers for the balls of radius $r'_1$, such that the distance between any two potential centers is at least $3r'_1$. At a high level, this reduction uses the ellipsoid algorithm, and each iteration of ellipsoid algorithm returns a candidate LP solution such that, (1) it can be rounded to obtain an $O(1)$-approximate solution for $\mathcal{I}'$, or (2) One can obtain polynomially many instances of \emph{well-separated} Robust $3$-NU$k$C\xspace, at least one of which is feasible, or (3) If none of the obtained instances is feasible, then one an obtain a hyperplane separating the LP solution from the integer hull of coverages.
\paragraph{Solving a Well-Separated Instance.} For the sake of simplicity let $\mathcal{J}$ be one of the instances of Well-Separated Robust $3$-NU$k$C\xspace, along with a well-separated set $Y$ that is a candidate set for the centers of balls of radius $r'_1$. Furthermore, let us assume that $\mathcal{J}$ is feasible. First, the reduction in Section \ref{sec:robust-colorful}, given the instance $\mathcal{J}$, produces $O(n)$ instances $\mathcal{J}(\ell)$ of Colorful $2$-NU$k$C\xspace, such that at least one of the instances is feasible. Then, we apply the reduction from Section \ref{sec:self-coverage} on each of these instances to ensure the \emph{self-coverage} property, i.e., we obtain an instance $\mathcal{J}'(\ell)$ of Colorful $2$-NU$k$C\xspace with $r''_1 = c_2 r'_2 + c_3 r'_3$, and $r''_2 = 0$. Finally, assuming that the resulting instance $\mathcal{J}'(\ell)$ is feasible, it is possible to find a feasible solution using dynamic programming, using the algorithm from Section \ref{sec:dp}. This algorithm supposes that the instance is \emph{Well-Separated} w.r.t.\ a smaller separation factor of $2$. We argue in the next paragraph that this property holds in each each of the instances $\mathcal{J}'(\ell)$.
In order to show that the set $Y$ well-separated w.r.t. the new top level radius $r''_1$, we need to show that $3r'_1 \ge 2r''_1$, i.e., $r'_1 \ge c_2 r'_2 + c_3 r'_3 \ge \beta \cdot r'_2$ for some sufficiently large constant $\beta$. This assumption is without loss of generality, since, if two consecutive radii classes are within a $\beta$ factor, it is possible to combine them into a single radius class, at the expense of an $O(\beta)$ factor in the approximation guarantee.
Assuming the instance $\mathcal{J}$ is feasible, a feasible solution to an instance $\mathcal{J}'(\ell)$ can be mapped back to an $O(1)$-approximate solution to $\mathcal{J}$, and then to $\mathcal{I}$, since each reduction preserves the approximation guarantee up to an $O(1)$ factor.
\begin{theorem} \label{theorem:main}
There exists a polynomial time $O(1)$-approximation algorithm for $4$-NU$k$C\xspace.
\end{theorem}
We have overviewed how the various sections of the paper come together in deriving Theorem \ref{theorem:main}. Before proceeding to these sections, we describe a greedy clustering procedure that we need.
\subsection{Greedy Clustering}
Assume we are given (i) a metric space $(X, d)$, where $X$ is finite, (ii) a radius $r \geq 0$, (iii) an expansion parameter $\gamma \geq 1$, (iv) a subset $Y \subseteq X$ and a weight function $\omega: Y \to \mathbb{Z}^+$. The weight $\omega(y)$ can be thought of as the multiplicity of $y \in Y$, or how many points are co-located at $y$. We describe a greedy clustering procedure, from Charikar et al. \cite{charikar2001algorithms}, that is used to partition the point set $Y$ into clusters, each of which is contained in a ball of radius $\gamma r$. This clustering procedure, together with its properties, is a crucial ingredient of our approach.
\begin{algorithm}
\caption{\textsc{GreedyClustering}($Y, X, r \ge 0, \gamma \ge 1, \omega: Y \to \mathbb{Z}^+$)} \label{alg:greedyclustering}
\begin{algorithmic}[1]
\Statex We require that $Y \subseteq X$
\State Let $U \gets Y$, $M \gets \emptyset$
\While{$U \neq \emptyset$}
\State $p = \arg\max_{q \in X}\omega(U \cap B(q, r))$ \label{line:argmax}
\State $C(p) \coloneqq U \cap B(p, \gamma r)$; $wt(p) \coloneqq \omega(C(p))$
\State $U \gets U \setminus C(p)$
\State $M \gets M \cup \{p\}$ \Comment{{\small\texttt{We will refer to $p$ as a \emph{mega-point} with cluster $C(p)$ of weight $w(p)$}}}
\EndWhile
\State \Return $(M, \{C(p)\}_{p \in M}, \{wt(p)\}_{p \in M})$
\end{algorithmic}
\end{algorithm}
In line~\ref{line:argmax}, we only consider $q \in X$ such that $U \cap B(q,r) \neq \emptyset$. Notice that it is possible that $\omega(U \cap B(q, r)) = 0$ if $\omega(y) = 0$ for each $y \in U$. Furthermore, notice that we do not require that $q \in U$ for it to be an eligible point in line~\ref{line:argmax}.
We summarize some of the key properties of this algorithm in the following observations.
\begin{observation}
\label{greedy:observe}
\begin{enumerate}
\item For any $p \in M$, $C(p) \subseteq B(p, \gamma r)$,
\item Point $y \in Y$ belongs to the cluster $C(p)$, such that $p$ is the \emph{first} among all $q \in M$ satisfying $d(y, q) \le \gamma r$.
\item The sets $\{C(p)\}_{p \in M}$ partition $Y$, which implies that
\item $\sum_{p \in M} wt(p) = \omega(Y)$, where $\omega(Z) = \sum_{z \in Z} \omega(z)$ for any $Z \subseteq Y$.
\item If $p_i$ and $p_j$ are the points added to $M$ in iterations $i \leq j$, then $wt(p_i) \geq wt(p_j)$.
\item For any two distinct $p, q \in M$, $d(p, q) > (\gamma -1) r$.
\end{enumerate}
\end{observation}
\begin{proof}
The first five properties are immediate from the description of the algorithm. Now, we prove the sixth property. Suppose for contradiction that there exist $p, q \in M$ with $d(p, q) \le (\gamma -1) r$, and without loss of generality, $p$ was added to $M$ before $q$. Then, note that at the end of this iteration, $B(q, r) \cap U = \emptyset$. Therefore, $q$ will subsequently never be a candidate for being added to $M$ in line~\ref{line:argmax}.
\end{proof}
A key property of this greedy clustering, established by Charikar et al.~\cite{charikar2001algorithms}, is that for any $k \geq 1$ balls of radius $r$, the weight of the points in the first $k$ clusters is at least as large as the weight of the points covered by the $k$ balls.
\begin{lemma}
\label{charikar-lemma}
Suppose that the parameter $\gamma$ used in Algorithm~\ref{alg:greedyclustering} is at least $3$. Let $\mathcal{B}$ be any collecion of $k \geq 1$ balls of radius $r$, each centered at a point in $X$. Let $M'$ consist of the first $k'$ points of $M$ chosen by the algorithm, where $k' = \min \{k, |M|\}$. We have
\[ \sum_{p \in M'} wt(p) = \omega \left( \bigcup_{p \in M'} C(p) \right) \geq \omega \left( Y \cap \bigcup_{B \in \mathcal{B}} B \right). \]
\end{lemma}
The equality follows from the definition of $wt(p)$ and the fact that the clusters partition $Y$, as stated in Obervation~\ref{greedy:observe}.
\section{Preliminaries}
\subsection{Problem Definitions}
In the following, we set up the basic notation and define the problems we will consider in the paper. We consider a finite metric space $(X, d)$, where $X$ is a finite set of (usually $n$) points, and $d$ is a distance function satisfying triangle inequality. If $Y$ is a subset of $X$, then by slightly abusing the notation, we use $(Y, d)$ to denote the metric space where the distance function $d$ is restricted to the points of $Y$. Let $p \in X$, $Y \subseteq X$, and $r \ge 0$. Then, we use $d(p, Y) \coloneqq \min_{y \in Y} d(p, y)$, and denote by $B(p, r)$ the \emph{ball} of radius $r$ centered at $p$, i.e., $B(p, r) \coloneqq \{ q \in X : d(p, q) \le r \}$. We say that a ball $B(p, r)$ \emph{covers} a point $q$ iff $q \in B(p, r)$; a set of balls $\mathcal{B}$ (resp.\ a tuple of sets of balls $(\mathcal{B}_1, \mathcal{B}_2, \ldots, \mathcal{B}_t)$) covers $q$ if there exists a ball in $\mathcal{B}$ that covers $q$ (resp.\ $\bigcup_{1 \le i \le t} \mathcal{B}_i$ that covers $q$). Analogously, a set of points $Y \subseteq X$ is covered iff every point in $Y$ is covered. For a function $f: S \to \mathbb{R}^+$ or $f: S \to \natural$, and $R \subseteq S$, we define $f(R) \coloneqq \sum_{r \in R} f(r)$.
\begin{definition}[Decision Version of $t$-NU$k$C\xspace]
\ \\The input is an instance $\mathcal{I} = ((X, d), (k_1, k_2, \ldots, k_t), (r_1, r_2, \ldots, r_t))$, where $r_1 \ge r_2 \ge \ldots r_t \geq 0$, and each $k_i$ is a non-negative integer. The goal is to determine whether there exists a solution $(\mathcal{B}_1, \mathcal{B}_2, \ldots, \mathcal{B}_t)$, where for each $1 \le i \le t$, $\mathcal{B}_i$ is a set with at most $k_i$ balls of radius $r_i$, that covers the entire set of points $X$. Such a solution is called a \emph{feasible solution}, and if the instance $\mathcal{I}$ has a feasible solution, then $\mathcal{I}$ is said to be \emph{feasible}.
\\An algorithm is said to be an $\alpha$-approximation algorithm (with $\alpha \ge 1$), if given a feasible instance $\mathcal{I}$, it returns a solution $(\mathcal{B}_1, \mathcal{B}_2, \ldots, \mathcal{B}_t)$, where for each $1 \le i \le t$, $\mathcal{B}_i$ is a collection of at most $k_i$ balls of radius $\alpha r_i$, such that the solution covers $X$.
\end{definition}
Next, we define the robust version of $t$-NU$k$C\xspace.
\begin{definition}[Decision Version of Robust $t$-NU$k$C\xspace]
\ \\The input is an instance $\mathcal{I} = ((X, d), (\omega, m), (k_1, k_2, \ldots, k_t), (r_1, r_2, \ldots, r_t))$. The setup is the same as in $t$-NU$k$C\xspace, except for the following: $\omega: X \to \mathbb{Z}^+$ is a weight function, and $1 \le m \le w(X)$ is a parameter. The goal is to determine whether there exists a feasible solution, i.e., $(\mathcal{B}_1, \mathcal{B}_2, \ldots, \mathcal{B}_t)$ of appropriate sizes and radii (as defined above), such that the total weight of the points covered is at least $m$. An $\alpha$-approximate solution covers points of weight at least $m$ while using at most $k_i$ balls of radius $\alpha r_i$ for each $1 \le i \le t$.
\end{definition}
We will frequently consider the \emph{unweighted} version of Robust $t$-NU$k$C\xspace, i.e., where the weight of every point in $X$ is unit. Let $\mathbbm{1}$ denote this unit weight function. Now we define the Colorful $t$-NU$k$C\xspace problem, which generalizes Robust $t$-NU$k$C\xspace.
\begin{definition}[Decision Version of Colorful $t$-NU$k$C\xspace]
\ \\The input is an instance $\mathcal{I}= ((X, d), (w_r, w_b, m_r, m_b) , (k_1, k_2, \ldots, k_t), (r_1, r_2, \ldots, r_t))$. The setup is similar as in Robust $t$-NU$k$C\xspace, except that we have \emph{two} weight functions $w_r, w_b: X \to \mathbb{R}^+$ (corresponding to \emph{red} and \emph{blue} weight respectively). A feasible solution covers a set of points with red weight at least $m_r$, and blue weight at least $m_b$. The notion of approximation is the same as above.
\end{definition}
We note that the preceding definition naturally extends to an arbitrary number $\chi \ge 2$ of colors (i.e., $\chi$ different weight functions over $X$). However, we will not need that level of generality in this paper.
\subsection{From $(t+1)$-NUkC to $t$-Robust-NUkC} \label{subsec:t-nukc-to-robust}
In this section, we show an approximate equivalence of $t+1$-NUkC and unweighted $t$-Robust-NUkC. In particular, we show the following lemma.
\begin{lemma}\label{lem:tplus1-to-t}\
\begin{enumerate}
\item Suppose there exists an $\alpha$-approximation algorithm for $t+1$-NUkC. Then, there exists an $\alpha$-approximation algorithm for unweighted $t$-Robust NUkC.
\item Suppose there exists a $\beta$-approximation algorithm for unweighted $t$-Robust-NUkC. Then there exists a $3\beta+2$-approximation algorithm for $t+1$-NUkC.
\end{enumerate}
\end{lemma}
\begin{proof}
Note that the first claim is trivial, since an instance of $t$-Robust-NUkC is a special case of NUkC, as follows. Let $\mathcal{I} = ((X, d), (\mathbbm{1}, m), (r_1, r_2, \ldots, r_t), (k_1, k_2, \ldots, k_t))$ be an instance of unweighted $t$-Robust-NUkC, where $m$ is the coverage requirement. Then, observe that it is equivalent to the instance $\mathcal{I}' = ((X, d), (r_1, r_2, \ldots, r_t, 0), (k_1, k_2, \ldots, k_t, n-m))$ of $t+1$-NUkC. An $\alpha$-approximate solution to $\mathcal{I}'$ immediately gives an $\alpha$-approximate solution to $\mathcal{I}$. We now proceed to the second claim.
Consider an instance $\mathcal{I} = ((X, d), (r_1, r_2, \ldots, r_t, r_{t+1}), (k_1, k_2, \ldots, k_t, k_{t+1}))$ of $(t+1)$-NUkC. Note that we have to cover all points of $X$ in the instance $\mathcal{I}$. First, we compute a $2r_{t+1}$-net $Y$ of $X$. That is compute $Y \subseteq X$ with the following properties: (i) $d(u, v) > 2r_{t+1}$ for any $u, v \in Y$, and (ii) for any $u \in X \setminus Y$, there exists a $v \in Y$ such that $d(u, v) \le 2r_{t+1}$. Let $\varphi: X \to Y$ be a mapping that assigns every point in $X$ to its nearest point in $Y$ (breaking ties arbitrarily). Our reduction constructs the instance $\mathcal{I}' = ((Y, d), (\mathbbm{1}, |Y|-k_{t+1}), (k_1, k_2, \ldots, k_t), (r'_1, r'_2, \ldots, r'_t))$ of $t$-Robust-NUkC with at most $k_{t+1}$ outliers, where $r'_i = r_i + 2r_t$ for $1 \le i \le t$.
We now argue that if $\mathcal{I}$ is feasible, then so is $\mathcal{I}'$.
Fix a solution $(\mathcal{B}_i)_{i \in [t+1]}$ for the original instance $\mathcal{I}$, where $\mathcal{B}_i$ is a set of at most $k_i$ balls of radius $r_i$. Let $Y' \subseteq Y$ be the set of points in $Y$ covered by $(\mathcal{B}_i)_{i \in [t]}$, the balls of the $t$ largest radii types. For each ball $B(c_i, r_i) \in \mathcal{B}_i$, we add $B(\varphi(c_i), r'_i)$ to obtain the set $\mathcal{B}'_i$ of balls; recall $r'_i = r_i + 2r_t$. Note that the resulting solution $(\mathcal{B}'_i)_{i \in [t]}$ covers the set of points $Y'$. Now, let $Y'' = Y \setminus Y'$ be the set of points covered by $\mathcal{B}_{t+1}$, the balls of radius $r_{t+1}$. The distance between any two points of $Y$, and thus $Y''$, is greater than $2r_{t+1}$. Therefore, a ball of radius radius $r_{t+1}$ covers at most one point of $Y''$, which implies that $|Y''| \le |\mathcal{B}_{t+1}| \le k_{t+1}$. Thus $(\mathcal{B}'_i)_{i \in [t]}$ is a feasible solution for instance $\mathcal{I}'$, with the points in $Y''$ being the set of outliers of size at most $k_{t+1}$.
We now argue that from a $\beta$-approximate solution to $\mathcal{I}'$, we can efficiently construct a $(3\beta + 2)$-approximate solution to $\mathcal{I}$. Fix a solution $(\mathcal{B}'_i)_{i \in [t]}$ for the instance $\mathcal{I}'$ that covers at least $|Y|-k_{t+1}$ points of $Y$, where $\mathcal{B}'_i$ consists of $k_i$ balls of radius $\beta r'_{i}$, for $1 \le i \le t$. To obtain a solution for the original instance $\mathcal{I}$, we proceed as follows. We expand the radius of every ball in $\mathcal{B}'_i$ by an additive factor of $2r_{t+1}$ to obtain $\mathcal{B}_i$. Note that the resulting radius for each ball in $\mathcal{B}_i$ is $\beta r_i + 2\beta r_{t+1} + 2r_{t+1} \le (3\beta+2) \cdot r_{i}$. Note that if a ball in solution $(\mathcal{B}'_i)_{i \in [t]}$ covers $y \in Y$, then the additively expanded version of the ball covers every point $x \in \varphi^{-1}(y)$. For every outlier point $y \in Y$ not covered by $(\mathcal{B}'_i)_{i \in [t]}$, we add a ball of radius $2r_{t+1}$ centered at $y$ to $\mathcal{B}_{t+1}$; this ball covers all points $x \in \varphi^{-1}(y)$. As the number of outliers is at most $k_{t+1}$, we have $|\mathcal{B}_{t+1}| \leq k_{t+1}$. The resulting solution $(\mathcal{B}_i)_{i \in [t+1]}$ covers all the points of $X$, and has approximation guarantee $3\beta+2$.
\end{proof}
\section{Definitions, Main Result, and Greedy Clustering}
\subsection{Problem Definitions}
In the following, we set up the basic notation and define the problems we will consider in the paper. We consider a finite metric space $(X, d)$, where $X$ is a finite set of (usually $n$) points, and $d$ is a distance function satisfying triangle inequality. If $Y$ is a subset of $X$, then by slightly abusing the notation, we use $(Y, d)$ to denote the metric space where the distance function $d$ is restricted to the points of $Y$. Let $p \in X$, $Y \subseteq X$, and $r \ge 0$. Then, we use $d(p, Y) \coloneqq \min_{y \in Y} d(p, y)$, and denote by $B(p, r)$ the \emph{ball} of radius $r$ centered at $p$, i.e., $B(p, r) \coloneqq \{ q \in X : d(p, q) \le r \}$. We say that a ball $B(p, r)$ \emph{covers} a point $q$ iff $q \in B(p, r)$; a set of balls $\mathcal{B}$ (resp.\ a tuple of sets of balls $(\mathcal{B}_1, \mathcal{B}_2, \ldots, \mathcal{B}_t)$) covers $q$ if there exists a ball in $\mathcal{B}$ that covers $q$ (resp.\ $\bigcup_{1 \le i \le t} \mathcal{B}_i$ that covers $q$). Analogously, a set of points $Y \subseteq X$ is covered iff every point in $Y$ is covered. For a function $f: S \to \mathbb{R}^+$ or $f: S \to \natural$, and $R \subseteq S$, we define $f(R) \coloneqq \sum_{r \in R} f(r)$.
\begin{definition}[Decision Version of $t$-NU$k$C\xspace]
\ \\The input is an instance $\mathcal{I} = ((X, d), (k_1, k_2, \ldots, k_t), (r_1, r_2, \ldots, r_t))$, where $r_1 \ge r_2 \ge \ldots r_t \geq 0$, and each $k_i$ is a non-negative integer. The goal is to determine whether there exists a solution $(\mathcal{B}_1, \mathcal{B}_2, \ldots, \mathcal{B}_t)$, where for each $1 \le i \le t$, $\mathcal{B}_i$ is a set with at most $k_i$ balls of radius $r_i$, that covers the entire set of points $X$. Such a solution is called a \emph{feasible solution}, and if the instance $\mathcal{I}$ has a feasible solution, then $\mathcal{I}$ is said to be \emph{feasible}.
\\An algorithm is said to be an $\alpha$-approximation algorithm (with $\alpha \ge 1$), if given a feasible instance $\mathcal{I}$, it returns a solution $(\mathcal{B}_1, \mathcal{B}_2, \ldots, \mathcal{B}_t)$, where for each $1 \le i \le t$, $\mathcal{B}_i$ is a collection of at most $k_i$ balls of radius $\alpha r_i$, such that the solution covers $X$.
\end{definition}
Next, we define the robust version of $t$-NU$k$C\xspace.
\begin{definition}[Decision Version of Robust $t$-NU$k$C\xspace]
\ \\The input is an instance $\mathcal{I} = ((X, d), (\omega, m), (k_1, k_2, \ldots, k_t), (r_1, r_2, \ldots, r_t))$. The setup is the same as in $t$-NU$k$C\xspace, except for the following: $\omega: X \to \mathbb{Z}^+$ is a weight function, and $1 \le m \le \omega(X)$ is a parameter. The goal is to determine whether there exists a feasible solution, i.e., $(\mathcal{B}_1, \mathcal{B}_2, \ldots, \mathcal{B}_t)$ of appropriate sizes and radii (as defined above), such that the total weight of the points covered is at least $m$. An $\alpha$-approximate solution covers points of weight at least $m$ while using at most $k_i$ balls of radius $\alpha r_i$ for each $1 \le i \le t$.
\end{definition}
We will frequently consider the \emph{unweighted} version of Robust $t$-NU$k$C\xspace, i.e., where the weight of every point in $X$ is unit. Let $\mathbbm{1}$ denote this unit weight function. Now we define the Colorful $t$-NU$k$C\xspace problem, which generalizes Robust $t$-NU$k$C\xspace.
\begin{definition}[Decision Version of Colorful $t$-NU$k$C\xspace]
\ \\The input is an instance $\mathcal{I}= ((X, d), (\omega_r, \omega_b, m_r, m_b) , (k_1, k_2, \ldots, k_t), (r_1, r_2, \ldots, r_t))$. The setup is similar as in Robust $t$-NU$k$C\xspace, except that we have \emph{two} weight functions $\omega_r, \omega_b: X \to \mathbb{Z}^+$ (corresponding to \emph{red} and \emph{blue} weight respectively). A feasible solution covers a set of points with red weight at least $m_r$, and blue weight at least $m_b$. The notion of approximation is the same as above.
\end{definition}
We note that the preceding definition naturally extends to an arbitrary number $\chi \ge 2$ of colors (i.e., $\chi$ different weight functions over $X$). However, we will not need that level of generality in this paper.
\section{From Robust $t$-NU$k$C\xspace to Colorful $(t-1)$-NU$k$C\xspace} \label{sec:robust-colorful}
Let $\mathcal{I} = ((X, d), (\omega, m) , (k_1, k_2, \ldots, k_t), (r_1, r_2, \ldots, r_t))$ be an instance of Robust $t$-NU$k$C\xspace. The reduction to Colorful $(t-1)$-NuKC consists of two phases. In the first phase, we use Algorithm~\ref{alg:greedyclustering} to reduce the instance $\mathcal{I}$ to an instance $\mathcal{I}'$ focused on the cluster centers output by the greedy algorithm. A key property of this reduction is that we may set $r_t = 0$ in the instance $\mathcal{I}'$ -- each ball at level $t$ is allowed to cover at most one point.
In the second phase, we transform $\mathcal{I}'$ to $O(n)$ instances of Colorful $(t-1)$-NU$k$C\xspace. Assuming there exists a feasible solution for $\mathcal{I}'$, at least one of the instances $\mathcal{I}''$ of Colorful $(t-1)$-NU$k$C\xspace has a feasible solution, and any approximate solution to $\mathcal{I}''$ can be used to obtain an approximate solution to $\mathcal{I}'$ (and thus to $\mathcal{I}$).
\paragraph{Phase 1.}
Let $\mathcal{I} = ((X, d), (\omega, m), (k_1, k_2, \ldots, k_t), (r_1, r_2, \ldots, r_t))$ be an instance of Robust $t$-NU$k$C\xspace. We call the algorithm \textsc{GreedyClustering}$(X, X, r_t, 3, \omega)$, and obtain a set of points $M$ with the corresponding clusters $C(p)$ for $p \in M$. The greedy algorithm also returns a weight $wt(p) = \omega(C(p))$ for each $p \in M$. Let us number the points of $M$ as $p_i$, where $i$ is the iteration in which $p_i$ was added to the set $M$ by algorithm \textsc{GreedyClustering}$(X, X, r_t, 3, \omega)$. This gives an ordering $\sigma = \langle p_1, p_2, \ldots, p_{|M|} \rangle $ of the points in $M$. Note that $wt(p_i) \geq wt(p_j)$ for $i \geq j$.
We define a weight function $\lambda: X \rightarrow \mathbb{Z}^+$. Let $\lambda(p) = wt(p)$ for $p \in M$ and $\lambda(p) = 0$ for $p \in X \setminus M$. Note that for $p \in M$, $\lambda(p) = wt(p) = \omega(C(p))$. Thus, for each $p \in M$, we are moving the weight from points in cluster $C(p)$ to the cluster center $p$. Clearly, $\omega(X) = \lambda(X)$.
The output of Phase 1 is the instance $\mathcal{I}' = ((X, d), (\lambda, m), (k_1, k_2, \ldots, k_t), (r'_1, r'_2, \ldots, r'_{t-1}, 0))$ of $t$-Robust-NuKC, where $r'_i = r_i + 3r_t$.
Note that in the instance $\mathcal{I}'$, we have $r'_t = 0$, whereas the other radii in $\mathcal{I}$ have been increased by an additive factor of $3 r_t$. The following claim relates instances $\mathcal{I}$ and $\mathcal{I}'$.
\begin{lemma}
\label{lemma:phase1}
(a) If instance $\mathcal{I}$ has a feasible solution, then so does the instance $\mathcal{I}'$. (b) Given a solution $(\mathcal{B}'_i)_{i \in [t]}$ for $\mathcal{I}'$ that uses at most $k_i$ balls of radius $\alpha r'_i$ for every $i \in [t]$, we can obtain a solution $(\mathcal{B}_i)_{i \in [t]}$ for $\mathcal{I}$ that uses at most $k_i$ balls of radius at most $\alpha r'_i + 3r_t \leq \alpha r_i + (3 \alpha + 3) r_t$ for $1 \le i \le t$.
\end{lemma}
\begin{proof}
We begin with part (b). For each ball in $B(p,r)$ that is part of the solution $(\mathcal{B}'_i)_{i \in [t]}$, we replace it with the ball $B(p, r + 3r_t)$ to obtain a solution $(\mathcal{B}_i)_{i \in [t]}$ for $\mathcal{I}$. That is, we expand each ball by an additive $3 r_t$. If $B(p,r)$ covers $q \in M$, then $B(p, r + 3r_t)$ covers $C(q)$, and $\lambda(q) = \omega(C(q))$. Let $M' \subseteq M$ denote the points covered by $(\mathcal{B}'_i)_{i \in [t]}$. The weight of the points covered by $(\mathcal{B}_i)_{i \in [t]}$ is at least
\[ \sum_{p \in M'} \omega(C(p)) = \sum_{p \in M'} \lambda(p) \geq m. \]
We now establish (a). Fix a feasible solution $(\mathcal{B}_i)_{i \in [t]}$ to $\mathcal{I}$ that covers $\omega$-weight at least $m$, where $\mathcal{B}_i$ is a set of at most $k_i$ balls of radius $r_i$, for $i \in [t]$. Let $M_1 \subseteq M$ be the set of points $p$ such that some point in $C(p)$ is covered by a ball in $\mathcal{B}_1, \mathcal{B}_2, \ldots, \mathcal{B}_{t-1}$.
Now let $M_2 = M \setminus M_1$ be the set of points $p$, such that any point in $C(p)$ is either covered by a ball from $\mathcal{B}_t$, or is an outlier. Let $X_i \coloneqq \bigcup_{p \in M_i} C(p)$ for $i = 1, 2$. Note that $X = X_1 \sqcup X_2$.
Note that in the sequence $\sigma = \langle p_1, p_2, \ldots, p_{|M|} \rangle$, the points of $M_1$ and $M_2$ may appear in an interleaved fashion. Let $p_{i_1}, p_{i_2}, \ldots, p_{i_{|M_2|}}$ be the subsequence restricted to the points in $M_2$. In the following lemma, we argue that the first $k_t$ points in this subsequence are sufficient to replace the balls in $\mathcal{B}_t$. Let $k' = \min \{|\mathcal{B}_t|, |M_2|\} \leq k_t$.
\begin{lemma} \label{lem:greedy-ordering}
There exists a subset $M_2^+ \subseteq M_2$ of size at most $k'$ such that $\sum_{p \in M_2^+} wt(p) \ge \omega \left(X_2 \cap \bigcup_{\ensuremath{\mathsf{Ball}} \in \mathcal{B}_t} \ensuremath{\mathsf{Ball}} \right).$
\end{lemma}
\begin{proof}
Let $M_2^+ = \{p_{i_1}, p_{i_2}, \ldots, p_{i_{k'}}\}$. That is, $M_2^+$ consists of the first $k'$ points of $M_2$ picked by the greedy algorithm. Recall that $M_2^+ \subseteq M_2$, and thus for $p_{i_j} \in M_2^+$, it holds that $C(p_{i_j}) \subseteq X_2$.
Now imagine calling the algorithm \textsc{GreedyClustering}$(X_2, X, r_t, 3, \omega)$. Observe that in the iteration $1 \le j \le |M_2|$, this algorithm will select point $p_{i_j}$ (as defined above) in Line 3, and the corresponding cluster and its weight will be $C(p_{i_j})$ and $wt(p_{i_j})$ -- exactly as in the execution of \textsc{GreedyClustering}$(X, X, r_t, 3, \omega)$. That is, the algorithm \textsc{GreedyClustering}$(X_2, X, r_t, 3, \omega)$ will output $M_2$ and the clusters $C(p)$ for each $p \in M_2$.
Now, $\mathcal{B}_t$ consists of a set of $|\mathcal{B}_t|$ balls of radius $r_t$. The lemma now follows from Lemma~\ref{charikar-lemma} applied to \textsc{GreedyClustering}$(X_2, X, r_t, 3, \omega)$.
\end{proof}
Using Lemma \ref{lem:greedy-ordering}, we now construct a solution to instance $\mathcal{I}'$. Fix index $1 \leq i \leq t-1$, and $\mathcal{B}'_i$ denote the set of balls obtained by expanding each ball in $\mathcal{B}_i$ by an additive $3 r_t$. Note that each ball in $\mathcal{B}'_i$ has radius $r'_i = r_i + 3 r_t$. For every point $p \in M_2^+$, we add a ball of radius $0$ around it and let $\mathcal{B}'_t$ be the resulting set of balls. Note that $|\mathcal{B}'_t| = |M_2^+| \leq k' \leq k_t$.
By definition, for each point $p \in M_1$, there is a ball in $(\mathcal{B}_i)_{i \in [t-1]}$ that intersects cluster $C(p)$, whose points are at distance at most $3 r_t$ from $p$. It follows that the balls in $(\mathcal{B}'_i)_{i \in [t-1]}$ cover each point in $M_1$.
Using Lemma \ref{lem:greedy-ordering}, the coverage of $(\mathcal{B}'_i)_{i \in [t]}$ in instance $\mathcal{I}'$ is at least
\[ \sum_{p \in M_1} wt(p) + \sum_{p \in M_2^+} wt(p) \geq \omega(X_1) + \omega\left(X_2 \cap \bigcup_{\ensuremath{\mathsf{Ball}} \in \mathcal{B}_t} \ensuremath{\mathsf{Ball}} \right) \geq m. \]
The final inequality follows because any point covered by solution $(\mathcal{B}_i)_{i \in [t]}$ for $\mathcal{I}$ either belongs to $X_1$ or to $X_2 \cap \bigcup_{\ensuremath{\mathsf{Ball}} \in \mathcal{B}_t} \ensuremath{\mathsf{Ball}}$. Thus, we have shown that $\mathcal{I}'$ has a feasible solution.
\end{proof}
\paragraph{Phase 2.}
Now we describe the second phase of the algorithm. We have the instance $\mathcal{I}' = ((X, d), (\lambda, m), (k_1, k_2, \ldots, k_t), (r'_1, r'_2, \ldots, r'_{t-1}, 0))$ of Robust $t$-NU$k$C\xspace that is output by Phase 1. Phase 2 takes $\mathcal{I}'$ as input and generates an instance $\mathcal{I}(\ell)$, for each $0 \leq \ell \leq |X|$, of the Colorful $(t-1)$-NU$k$C\xspace problem. Note that the number of generated instances is $|X| + 1 = O(n)$. If $\mathcal{I}'$ is feasible, at least one of these $|X| + 1$ instances will be feasible.
Let $\sigma = \langle p_1, p_2, \ldots, p_{|X|} \rangle $ be an ordering of the points in $X$ by non-increasing $\lambda$. That is, $\lambda(p_i) \geq \lambda(p_j)$ for $i \leq j$.
Fix an index $0 \leq \ell \leq |X|$. We now describe the instance $\mathcal{I}(\ell)$ of colorful $(t-1)$-NU$k$C\xspace. Let $R = \{ p_1, p_2, \ldots, p_{\ell}\}$ denote the set of {\em red} points, and $B = \{ p_{\ell +1}, p_{\ell + 2}, \ldots, p_{|X|} \}$ denote the set of {\em blue} points. For each $p \in B$, define its blue weight as $\omega_b(p) \coloneqq \lambda(p)$; for each $p \in R$, define its blue weight as $\omega_b(p) \coloneqq 0$. Define the blue coverage $m_b$ for instance $\mathcal{I}(\ell)$ as $m_b \coloneqq m - \lambda(R)$. We define the red weight function $\omega_r$ in a slightly different manner. For each red point $p \in R$, let its red weight $\omega_r(p) \coloneqq 1$; for each $p \in B$, let red weight $\omega_r(p) \coloneqq 0$. Let $m_r \coloneqq \sum_{p \in R} \omega_r(p) - k_t = |R| - k_t$ denote the red coverage for instance $\mathcal{I}(\ell)$. Note that $\omega_r$ is supported on $R$ and $\omega_b$ on $B$. Let $\mathcal{I}(\ell) \coloneqq ((X, d), (\omega_r, \omega_b, m_r, m_b), (k_1, k_2, \ldots, k_{t-1}), (r'_1, r'_2, \ldots, r'_{t-1}) )$ denote the resulting instance of Colorful $(t-1)$-NU$k$C\xspace problem. Recall that a solution to this instance is required to cover red weight that adds up to at least $m_r$, and blue weight that adds up to at least $m_b$. (In instance $\mathcal{I}(\ell)$, the point sets $R$ and $B$, the red and blue weights, and total coverage requirements $m_r$ and $m_b$ all depend on the index $\ell$. This dependence is not made explicit in the notation, so as to keep it simple.)
We now relate the instance $\mathcal{I}'$ to the instances $\mathcal{I}(\ell)$, for $0 \leq \ell \leq |X|$.
\begin{lemma}
\label{lemma:phase2}
(a) If the instance $\mathcal{I}' = ((X, d), (\lambda, m), (k_1, k_2, \ldots, k_t), (r'_1, r'_2, \ldots, r'_{t-1}, 0) )$ is feasible, then there exists an $0 \leq \ell^* \leq |X|$ such that instance $\mathcal{I}(\ell^*)$ is feasible. \\
(b) Let $\mathcal{I}(\ell) = ((X, d), (\omega_r, \omega_b, m_r, m_b), (k_1, k_2, \ldots, k_{t-1}), (r'_1, r'_2, \ldots, r'_{t-1}) )$ be a generated instance of Colorful $(t-1)$-NuKc, and suppose $(\mathcal{B}''_i)_{i \in [t-1]}$ is a solution to this instance such that $\mathcal{B}''_i$ contains at most $k_i$ balls of radius $\alpha r'_i$ for $1 \le i \le t-1$, and covers red weight at least $m_r$ and blue weight at least $m_b$. Then, we can efficiently obtain a solution to the instance $\mathcal{I}'$ that uses at most $k_i$ balls of radius $\alpha r'_i$ for $1 \le i \le t -1$, and at most $k_t$ balls of radius $0$.
\end{lemma}
\begin{proof}
We first show part (b). In instance $\mathcal{I}(\ell)$, the red weight $\omega_r(p) = 1$ for each $p \in R$, so the solution $(\mathcal{B}''_i)_{i \in [t-1]}$ covers at least $m_r = \sum_{p \in R} \omega_r(p) - k_t = |R| - k_t$ red points. So the number of red points that are not covered is at most $k_t$. Construct $\mathcal{B}'_t$ by adding a ball of radius $0$ at each uncovered point in $R$. Thus, $|\mathcal{B}'_t| \leq k_t$.
Let $\mathcal{B}'_i = \mathcal{B}''_i$ for each $1 \leq i \leq t-1$. Now, we argue that the solution $(\mathcal{B}'_i)_{i \in [t]}$ covers weight at least $m$ in instance $\mathcal{I}'$. Note that this solution covers all points in $R$, and a subset $C \subseteq B$ such that $\omega_b(C) \geq m_b = m - \lambda(R)$. Thus the coverage for $\mathcal{I}'$ is at least
\[ \lambda(R) + \lambda(C) = \lambda(R) + \omega_b(C) \geq \lambda(R) + m_b = m. \]
We now turn to part (a). Fix a feasible solution $(\mathcal{B}'_i)_{i \in [t]}$ to $\mathcal{I}'$. Let $M_1 \subseteq X$ denote the subset consisting of each point covered by a ball in $\mathcal{B}'_i$, for $1 \leq i \leq t-1$. Let $M_2 = X \setminus M_1$. Each point in $M_2$ is either an outlier or is covered by a ball in $\mathcal{B}'_t$.
Note that in the sequence $\sigma = \langle p_1, p_2, \ldots, p_{|X|} \rangle$, the points of $M_1$ and $M_2$ may appear in an interleaved fashion. Let $p_{i_1}, p_{i_2}, \ldots, p_{i_{|M_2|}}$ be the subsequence restricted to the points in $M_2$. Let $k' = \min \{k_t, |M_2|\}$, and let $M_2^+ = \{ p_{i_1}, p_{i_2}, \ldots, p_{i_{k'}} \}$. A key observation is that $\lambda(M_2^+)$ is at least as large as the total weight of the points in $M_2$ covered by balls in $\mathcal{B}'_t$. This is because each ball in $\mathcal{B}'_t$ has radius $0$ and can cover only one point in $M_2$; and the maximum coverage using such balls is obtained by placing them at the points in $M_2$ with the highest weights, i.e, $M_2^+$. Without loss of generality, we assume that $\mathcal{B}'_t$ consists of balls of radius $0$ placed at each point in $M_2^+$.
Now, let the index $\ell^* \coloneqq i_{k'}$. We now argue that the instance $\mathcal{I}(\ell^*)$ of colorful $(t-1)$-NU$k$C\xspace is feasible. In particular, we argue that $(\mathcal{B}'_i)_{i \in [t-1]}$ is a solution. Consider the set $R = \{p_1, p_2, \ldots, p_{\ell^*}\}$ of red points in $\mathcal{I}(\ell^*)$. Each point in $R$ is either in $M_1$ or in $M_2^+$, and is therefore covered by $(\mathcal{B}'_i)_{i \in [t]}$. It follows that $(\mathcal{B}'_i)_{i \in [t-1]}$ covers at least $|R| - |\mathcal{B}'_t| \geq |R| - |k_t| = m_r$ points of $R$. In other words, the red weight in $\mathcal{I}(\ell^*)$ covered by $(\mathcal{B}'_i)_{i \in [t-1]}$ is at least $m_r$.
Now consider the set $B = \{p_{\ell +1}, p_{\ell + 2}, \ldots, p_{|X|}\}$ of blue points in $\mathcal{I}(\ell^*)$. Let $C \subseteq B$ denote the blue points covered by solution $(\mathcal{B}'_i)_{i \in [t]}$. As $(\mathcal{B}'_i)_{i \in [t]}$ covers points with weight at least $m$ in instance $\mathcal{I}'$, we have $\lambda(R) + \lambda(C) \geq m$; thus, $\lambda(C) \geq m - \lambda(R) = m_b$. However, the balls in $\mathcal{B}'_t$ do not cover any point in $B$. We conclude that the balls in $(\mathcal{B}'_i)_{i \in [t-1]}$ cover all points in $C$. For any $p \in B$, we have $\lambda(p) = \omega_b(p)$. It follows that the blue weight in $\mathcal{I}(\ell^*)$ covered by $(\mathcal{B}'_i)_{i \in [t-1]}$ is at least $\omega_b(C) = \lambda(C) \geq m_b$. This concludes the proof of part (a).
\end{proof}
Combining Lemmas \ref{lemma:phase1} and \ref{lemma:phase2} from Phases 1 and 2, we obtain the following reduction from robust $t$-NuKC to colorful $(t-1)$-NuKC.
\begin{theorem} \label{thm:t-to-t-1}
There is a polynomial-time algorithm that, given an instance $\mathcal{I} = ((X, d), (\omega, m), \allowbreak (k_1, \ldots, k_t), (r_1, \ldots, r_t) )$ of Robust $t$-NU$k$C\xspace, outputs a collection of $O(n)$ instances of Colorful $(t-1)$-NU$k$C\xspace with the following properties: (a) If $\mathcal{I}$ is feasible, then at least one of the instances $\mathcal{I}(\ell) = ((X,d), \allowbreak (\omega_r, \omega_b, m_r, m_b), (k_1, \ldots, k_{t-1}), (r'_1, \ldots, r'_{t-1}) )$ of Colorful $(t-1)$-NU$k$C\xspace is feasible; (b) given an $\alpha$-approximate solution to some instance $\mathcal{I}(\ell)$, we can efficiently construct a solution to $\mathcal{I}$ that uses at most $k_i$ balls of radius at most $\alpha r_i + (3\alpha + 3)r_t$.
\end{theorem}
\begin{remark} In part (a), the feasible solution for $\mathcal{I}(\ell)$ that is constructed from the feasible solution for $\mathcal{I}$ has the following useful property: for any $\ensuremath{\mathsf{Ball}}$ of radius $r'_i = r_i + 3 r_t$ in the feasible solution for $\mathcal{I}(\ell)$, the center of $\ensuremath{\mathsf{Ball}}$ is also the center of some ball of radius $r_i$ in the feasible solution for $\mathcal{I}$.
\end{remark}
\section{From $(t+1)$-NUkC to Robust $t$-NUkC} \label{sec:t-nukc-to-robust}
In this section, we show an approximate equivalence of $t+1$-NUkC and Robust $t$-NUkC. Note that Jia et al.\ \cite{jia2021towards} recently showed a very similar result. However, our proof is slightly different from theirs, and we describe it here for the sake of completeness.
\begin{lemma}\label{lem:tplus1-to-t}\
\begin{enumerate}
\item Suppose there exists an $\alpha$-approximation algorithm for $(t+1)$-NUkC. Then, there exists an $\alpha$-approximation algorithm for unweighted Robust $t$-NUkC.
\item Suppose there exists a $\beta$-approximation algorithm for unweighted Robust $t$-NUkC. Then there exists a $3\beta+2$-approximation algorithm for $(t+1)$-NUkC.
\end{enumerate}
\end{lemma}
\begin{proof}
Note that the first claim is trivial, since an instance of Robust $t$-NUkC is a special case of NUkC, as follows. Let $\mathcal{I} = ((X, d), (\mathbbm{1}, m), (r_1, r_2, \ldots, r_t), (k_1, k_2, \ldots, k_t))$ be an instance of unweighted $t$-Robust-NUkC, where $m$ is the coverage requirement. Then, observe that it is equivalent to the instance $\mathcal{I}' = ((X, d), (r_1, r_2, \ldots, r_t, 0), (k_1, k_2, \ldots, k_t, n-m))$ of $t+1$-NUkC. An $\alpha$-approximate solution to $\mathcal{I}'$ immediately gives an $\alpha$-approximate solution to $\mathcal{I}$. We now proceed to the second claim.
Consider an instance $\mathcal{I} = ((X, d), (r_1, r_2, \ldots, r_t, r_{t+1}), (k_1, k_2, \ldots, k_t, k_{t+1}))$ of $(t+1)$-NUkC. Note that we have to cover all points of $X$ in the instance $\mathcal{I}$. First, we compute a $2r_{t+1}$-net $Y$ of $X$. That is compute $Y \subseteq X$ with the following properties: (i) $d(u, v) > 2r_{t+1}$ for any $u, v \in Y$, and (ii) for any $u \in X \setminus Y$, there exists a $v \in Y$ such that $d(u, v) \le 2r_{t+1}$. Let $\varphi: X \to Y$ be a mapping that assigns every point in $X$ to its nearest point in $Y$ (breaking ties arbitrarily). Our reduction constructs the instance $\mathcal{I}' = ((Y, d), (\mathbbm{1}, |Y|-k_{t+1}), (k_1, k_2, \ldots, k_t), (r'_1, r'_2, \ldots, r'_t))$ of $t$-Robust-NUkC with at most $k_{t+1}$ outliers, where $r'_i = r_i + 2r_t$ for $1 \le i \le t$.
We now argue that if $\mathcal{I}$ is feasible, then so is $\mathcal{I}'$.
Fix a solution $(\mathcal{B}_i)_{i \in [t+1]}$ for the original instance $\mathcal{I}$, where $\mathcal{B}_i$ is a set of at most $k_i$ balls of radius $r_i$. Let $Y' \subseteq Y$ be the set of points in $Y$ covered by $(\mathcal{B}_i)_{i \in [t]}$, the balls of the $t$ largest radii types. For each ball $B(c_i, r_i) \in \mathcal{B}_i$, we add $B(\varphi(c_i), r'_i)$ to obtain the set $\mathcal{B}'_i$ of balls; recall $r'_i = r_i + 2r_t$. Note that the resulting solution $(\mathcal{B}'_i)_{i \in [t]}$ covers the set of points $Y'$. Now, let $Y'' = Y \setminus Y'$ be the set of points covered by $\mathcal{B}_{t+1}$, the balls of radius $r_{t+1}$. The distance between any two points of $Y$, and thus $Y''$, is greater than $2r_{t+1}$. Therefore, a ball of radius radius $r_{t+1}$ covers at most one point of $Y''$, which implies that $|Y''| \le |\mathcal{B}_{t+1}| \le k_{t+1}$. Thus $(\mathcal{B}'_i)_{i \in [t]}$ is a feasible solution for instance $\mathcal{I}'$, with the points in $Y''$ being the set of outliers of size at most $k_{t+1}$.
We now argue that from a $\beta$-approximate solution to $\mathcal{I}'$, we can efficiently construct a $(3\beta + 2)$-approximate solution to $\mathcal{I}$. Fix a solution $(\mathcal{B}'_i)_{i \in [t]}$ for the instance $\mathcal{I}'$ that covers at least $|Y|-k_{t+1}$ points of $Y$, where $\mathcal{B}'_i$ consists of $k_i$ balls of radius $\beta r'_{i}$, for $1 \le i \le t$. To obtain a solution for the original instance $\mathcal{I}$, we proceed as follows. We expand the radius of every ball in $\mathcal{B}'_i$ by an additive factor of $2r_{t+1}$ to obtain $\mathcal{B}_i$. Note that the resulting radius for each ball in $\mathcal{B}_i$ is $\beta r_i + 2\beta r_{t+1} + 2r_{t+1} \le (3\beta+2) \cdot r_{i}$. Note that if a ball in solution $(\mathcal{B}'_i)_{i \in [t]}$ covers $y \in Y$, then the additively expanded version of the ball covers every point $x \in \varphi^{-1}(y)$. For every outlier point $y \in Y$ not covered by $(\mathcal{B}'_i)_{i \in [t]}$, we add a ball of radius $2r_{t+1}$ centered at $y$ to $\mathcal{B}_{t+1}$; this ball covers all points $x \in \varphi^{-1}(y)$. As the number of outliers is at most $k_{t+1}$, we have $|\mathcal{B}_{t+1}| \leq k_{t+1}$. The resulting solution $(\mathcal{B}_i)_{i \in [t+1]}$ covers all the points of $X$, and has approximation guarantee $3\beta+2$.
\end{proof}
\section{Ensuring Self-Coverage in Colorful $2$-NU$k$C\xspace} \label{sec:self-coverage}
We assume that we are given as input a Colorful $2$-NU$k$C\xspace instance $\mathcal{I} = ((X, d), (\omega_r, \omega_b, m_r, m_b), \allowbreak (k_1, k_2), (r_1, r_2) )$. Recall that $\omega_r: X \rightarrow \mathbb{Z}^+$ (resp. $\omega_b: X \rightarrow \mathbb{Z}^+$) is the red (resp. blue) weight function. The task in Colorful $2$-NU$k$C\xspace is to find a solution $(\mathcal{B}_1, \mathcal{B}_2)$ such that (1) $|\mathcal{B}_i| \le k_i$ for $i = 1, 2$, and (2) the point set $Y \subseteq X$ covered by the solution satisfies $\omega_r(Y) \geq m_r$ and $\omega_b(Y) \geq m_b$, (i.e., the solution covers points with total red weight at least $m_r$, and blue weight at least $m_b$.) In this section, we show that $\mathcal{I}$ can be reduced to an instance of Colorful $2$-NU$k$C\xspace with $r_2 = 0$. The fact that each ball of radius $r_2$ can only cover its center in the target instance is what we mean by the term {\em self-coverage}. This reduction actually generalizes to Colorful $t$-NU$k$C\xspace, but we address the case $t = 2$ to keep the notation simpler.
Our reduction proceeds in two phases. In Phase 1, we construct an intermediate instance where we can ensure blue self-coverage. Then in Phase 2, we modify the intermediate instance so as to obtain red self-coverage as well.
\paragraph{Phase 1.} In this step, we call the greedy clustering algorithm using the {\em blue weight function} $\omega_b$. In particular, we call \textsc{GreedyClustering}($X, X, r_2, 3, \omega_b$) (See Algorithm \ref{alg:greedyclustering}). This algorithm returns a set of points $M \subseteq X$, where every $p \in M$ has a cluster $C(p)$ and weight $wt(p)$ such that (1) $\{C(p)\}_{p \in M}$ is a partition of $X$; (2) for any $p \in M$, $wt(p) = \omega_b(C(p))$, the blue weight of the cluster, and (3) $d(q, p) \le 3r_2$ for any $q \in C(p)$. Furthermore, the greedy algorithm naturally defines an ordering $\sigma = \langle p_1, p_2, \ldots, p_{|M|} \rangle$ of $M$ -- this is the order in which the points were added to $M$.
We define a new weight function $\lambda_b: X \rightarrow \mathbb{Z}^+$ as follows: $\lambda_b(p) \coloneqq wt(p)$ if $p \in M$ and $\lambda_b(p) \coloneqq 0$ if $p \in X \setminus M$. Note that for $p \in M$, we have $wt(p) = \omega_b(C(p))$. So the new weight function $\lambda_b$ is obtained from $\omega_b$ by moving weight from each cluster $C(p)$ to the cluster center $p$.
Phase 1 outputs the intermediate instance $\mathcal{I}' = ((X, d), (\omega_r, \lambda_b, m_r, m_b), (k_1, k_2), (r'_1, r'_2) )$ of Colorful $2$-NU$k$C\xspace, where $r'_1 = r_1 + 6 r_2$ and $r'_2 = 5 r_2$. A solution $(\mathcal{B}'_1, \mathcal{B}'_2)$ for $\mathcal{I}'$ is said to be {\em structured} if it has the following properties.
\begin{enumerate}
\item It is a solution to $\mathcal{I}'$ viewed as an instance of Colorful $2$-NU$k$C\xspace.
\item Let $Y \subseteq X$, the set of points {\em self-covered} by solution $(\mathcal{B}'_1, \mathcal{B}'_2)$, consist of points $p \in X$ such that either (a) $p$ is covered by $\mathcal{B}'_1$, or (b) $p$ is the center of some ball in $\mathcal{B}'_2$. We require that
\[ \lambda_b(Y) \geq m_b. \]
\end{enumerate}
Thus, a structured solution covers red weight in the usual way; for blue weight, a ball in $\mathcal{B}'_2$ can only contribute blue coverage for its center,
The following lemma relates instances $\mathcal{I}$ and $\mathcal{I}'$.
\begin{lemma}
\label{lemma:phase1:colorful}
(a) If instance $\mathcal{I}$ has a feasible solution, then the instance $\mathcal{I}'$ has a feasible solution that is also structured. (b) Given a solution $(\mathcal{B}'_1, \mathcal{B}'_2)$ for $\mathcal{I}'$ that uses at most $k_i$ balls of radius $\alpha r'_i$ for every $i \in \{1, 2\}$, we can obtain a solution $(\mathcal{B}_1, \mathcal{B}_2)$ for $\mathcal{I}$ that uses at most $k_i$ balls of radius $\alpha r'_i + 3r_2 \leq \alpha r_i + (6 \alpha + 3) r_2$ for $i \in \{1,2\}$.
\end{lemma}
Part (b) is straightforward as the red weights are unchanged in going from $\mathcal{I}$ to $\mathcal{I}'$, and the blue weights are moved by at most $3 r_2$. (Note that we don't require in part (b) that the solution to $\mathcal{I}'$ be structured.)
In the rest of this section, we establish (a). Fix a feasible solution $(\mathcal{B}_1, \mathcal{B}_2)$ to $\mathcal{I}$. Thus, (1) $|\mathcal{B}_i| \le k_i$ for $i = 1, 2$, and (2) the point set $Y \subseteq X$ covered by the solution satisfies $\omega_r(Y) \geq m_r$ and $\omega_b(Y) \geq m_b$, (i.e., the solution covers points with total red weight at least $m_r$, and blue weight at least $m_b$.)
Let $M_1 \subseteq M$ be the set of points $p$ such that some point in $C(p)$ is covered by a ball in $\mathcal{B}_1$. Now let $M_2 = M \setminus M_1$ be the set of points $p$ such that any point in $C(p)$ is either covered by a ball from $\mathcal{B}_2$, or is an outlier. Let $X_i \coloneqq \bigcup_{p \in M_i} C(p)$ for $i = 1, 2$. Note that $X = X_1 \sqcup X_2$.
We construct a solution $(\mathcal{B}'_1, \mathcal{B}'_2)$ for instance $\mathcal{I}'$ as follows. The set $\mathcal{B}'_1$ is obtained by expanding each ball in $\mathcal{B}_1$ by an additive factor of $6 r_2$. Thus, the balls in $\mathcal{B}'_1$ cover $X_1$. As in the proof of Lemma~\ref{lemma:phase1}, we construct a subset $N \subseteq M_2$ of size at most $|\mathcal{B}_2|$. We let $\mathcal{B}'_2$ consist of the balls of radius $r'_2 = 5 r_2$, each centered at a point in $N$. The set $N$ will have the following properties:
\begin{eqnarray}
\omega_r( X_2 \cap \bigcup_{\ensuremath{\mathsf{Ball}} \in \mathcal{B}'_2} \ensuremath{\mathsf{Ball}} ) & \geq & \omega_r(X_2 \cap \bigcup_{\ensuremath{\mathsf{Ball}} \in \mathcal{B}_2} \ensuremath{\mathsf{Ball}} ) \label{eq:red} \\
\sum_{p \in N} wt(p) & \geq & \omega_b(X_2 \cap \bigcup_{\ensuremath{\mathsf{Ball}} \in \mathcal{B}_2} \ensuremath{\mathsf{Ball}}) \label{eq:blue}
\end{eqnarray}
It is easy to verify that these two guarantees imply that $(\mathcal{B}'_1, \mathcal{B}'_2)$ is a structured, feasible solution to $\mathcal{I}'$:
The red weight covered by $(\mathcal{B}'_1,\mathcal{B}'_2)$ is at least
\[ \omega_r(X_1) + \omega_r( X_2 \cap \bigcup_{\ensuremath{\mathsf{Ball}} \in \mathcal{B}'_2} \ensuremath{\mathsf{Ball}} ) \geq \omega_r(X_1) + \omega_r( X_2 \cap \bigcup_{\ensuremath{\mathsf{Ball}} \in \mathcal{B}_2} \ensuremath{\mathsf{Ball}} ) \geq m_r.\]
The set $M_1 \cup N$ is self-covered by $(\mathcal{B}'_1, \mathcal{B}'_2)$. We have
\[ \lambda_b(M_1) + \lambda_b(N) = \omega_b(X_1) + \sum_{p \in N} wt(p) \geq \omega_b(X_1) + \omega_b(X_2 \cap \bigcup_{\ensuremath{\mathsf{Ball}} \in \mathcal{B}_2} \ensuremath{\mathsf{Ball}}) \geq m_b. \]
We now describe the construction of $N$ and establish properties (\ref{eq:red}) and (\ref{eq:blue}). At a high level, this is similar to what we did for $M_2^+$ in Lemma~\ref{lemma:phase1}; but it is more involved as we need to ensure that both properties hold.
\begin{algorithm}
\caption{\textsc{Mapping Procedure}$(\widehat{M}, \sigma, \widehat{\mathcal{B}}, \{C(p)\}_{p \in \widehat{M}})$} \label{alg:mapping}
\begin{algorithmic}[1]
\State Index the points of $\widehat{M}$ as $q_1, q_2, \ldots$ according to the ordering $\sigma$
\State For every $\ensuremath{\mathsf{Ball}} \in \widehat{\mathcal{B}}$, $\varphi(\ensuremath{\mathsf{Ball}}) \coloneqq q_i$, where $q_i \in \widehat{M}$ is the \emph{first} point $q$ s.t. $\ensuremath{\mathsf{Ball}} \cap C(q) \neq \emptyset$
\State $\ell = 0$; $\mathcal{T} \gets \emptyset$
\While{there exists a $\ensuremath{\mathsf{Ball}} \in \widehat{\mathcal{B}}$ that does not belong to any $D_{j}$ with $j \le \ell$}
\State $\ell \gets \ell+1$
\State $q_i \in \widehat{M} \setminus \bigcup_{j = 1}^{\ell-1} N_\ell$ be the first point $q$ with $|\varphi^{-1}(q)| > 0$
\State \texttt{pending} $ \gets |\varphi^{-1}(q_i)| - 1$
\State $N_\ell \gets \{q_i\}$, $D_\ell \gets \varphi^{-1}(q_i)$
\While{$\texttt{pending} > 0$ \textbf{ and } $i+1 \le |\widehat{M}|$}
\State $i \gets i+1$
\State $\texttt{pending} \gets \texttt{pending} + |\varphi^{-1}(q_i)| - 1$
\State $N_\ell \gets N_\ell \cup \{q_i\}$, $D_\ell \gets D_\ell \cup \varphi^{-1}(q_i)$
\EndWhile
\State Add $(N_\ell, D_\ell)$ to $\mathcal{T}$
\EndWhile
\State Return $\mathcal{T}$
\end{algorithmic}
\end{algorithm}
Let $\widehat{\mathcal{B}}_2 = \{\ensuremath{\mathsf{Ball}} \in \mathcal{B}_2 \ | \ \ensuremath{\mathsf{Ball}} \cap X_2 \neq \emptyset \}$.
The set $N$ is obtained via \textsc{Mapping Procedure}, given in Algorithm \ref{alg:mapping}. In particular, we invoke \textsc{Mapping Procedure}($M_2, \sigma, \widehat{\mathcal{B}}_2, \{C(p\}_{p \in M_2}$). We describe Algorithm \ref{alg:mapping} at a high level. First, we map every ball in $\widehat{\mathcal{B}}_2$ to the \emph{first} (according to $\sigma$) point $q$ in $M_2$ whose cluster $C(q)$ has a non-empty intersection with the ball -- this is the definition of $\varphi$. Now, some points $q \in M_2$ may get mapped by more than one ball. Then, we create a ``grouping procedure'' that creates pairs $(N_\ell, D_\ell)$ as follows. We start from the \emph{first} (according to $\sigma$) point $q_i$ that is mapped by at least one ball. We add $q_i$ to $N_\ell$, and the balls that were mapped to $q_i$ to the set $D_\ell$. Now, if $|\varphi^{-1}(q_i)| > 1$, then we aim to find $|\varphi^{-1}(q_i)| -1$ additional points after $q_i$ to be added to $N_\ell$. Furthermore, it is important in the analysis that these points be \emph{consecutive} according to $\sigma_{|M_2}$. The variable \texttt{pending} keeps track of how many additional distinct points need to be added to $N_\ell$ to match the number of distinct balls in $D_\ell$ at the current time. Thus, if $|\varphi^{-1}(q_i)| > 1$, we add $q_{i+1}$ to $N_\ell$ as well. At this stage, it may happen that $\varphi^{-1}(q_{i+1}) \neq \emptyset$. Then, we add $\varphi^{-1}(q_{i+1})$ to $D_\ell$, and update the variable \texttt{pending} appropriately. If the variable \texttt{pending} becomes $0$, then $|N_\ell| = |D_\ell|$, at which point the inner while loop terminates. By construction, the points added to $N_\ell$ form a contiguous sub-sequence of $\sigma_{|M_2}$. We add the pair $(N_\ell, D_\ell)$ to $\mathcal{T}$. At this point, if there still exists a ball of $\widehat{\mathcal{B}}_2$ that does not belong to any $D_j$ with $j \le \ell$, we start the construction of the next pair $(N_{\ell+1}, D_{\ell+1})$. Note that in all but the last iteration of the outer while loop, it holds that $|N_\ell| = |D_\ell|$. However, in the last iteration $t$, the loop may terminate with $|N_t| \le |D_t|$.
The invocation of \textsc{Mapping Procedure}($M_2, \sigma, \widehat{\mathcal{B}}_2, \{C(p\}_{p \in M_2}$) returns
$\mathcal{T} = \{(N_1, D_1), (N_2, D_2), \allowbreak \ldots, (N_t, D_t)\}$. In the following observation, we summarize a few key properties of this collection of pairs.
\begin{observation} \label{obs:mapping-properties}
$\mathcal{T} = \{(N_1, D_1), (N_2, D_2), \ldots, (N_t, D_t)\}$ satisfies the following properties.
\begin{enumerate}
\item For each $1 \leq \ell \leq t$, we have $\emptyset \neq N_\ell \subseteq M_2$; Furthermore, the points of $N_\ell$ form a contiguous subsequence of $M_2$ ordered according to $\sigma$.
The sets $N_1, N_2, \ldots, N_t $ are pairwise disjoint.
\item For each $1 \leq \ell \leq t$, we have $\emptyset \neq D_\ell \subseteq \widehat{\mathcal{B}}_2$. The sets $D_1, D_2, \ldots, D_t$ form a partition of $\widehat{\mathcal{B}}_2$.
\item $|N_\ell| = |D_\ell|$ for $\ell < t$, and $|N_t| \le |D_t|$.
\end{enumerate}
\end{observation}
Now we prove the following key lemma.
\begin{lemma} \label{lem:blue-greedy}
For any $1 \leq \ell \leq t$, the following properties hold.
\begin{enumerate}[label=(\Alph*)]
\item For any ball $B(c, r_2) \in D_\ell$, there exists a $q \in N_\ell$ such that $B(c, r_2) \subseteq B(q, 5r_2)$. \label{prop:one}
\item $\displaystyle \omega_b\lr{ X_2 \cap \bigcup_{B(c, r_2) \in D_\ell} B(c, r_2)} \le \sum_{p \in N_\ell} wt(p) $. \label{prop:two}
\end{enumerate}
\end{lemma}
\begin{proof}
For any $\ensuremath{\mathsf{Ball}} = B(c, r_2) \in D_\ell$, $q_i = \varphi(\ensuremath{\mathsf{Ball}}) \in N_\ell$. By the definition of $q_i$, it holds that $C(q_i) \cap \ensuremath{\mathsf{Ball}} \neq \emptyset$. Therefore, for any point $p \in \ensuremath{\mathsf{Ball}}$, it holds that $d(p, q_i) \le d(p, c) + d(c, p') + d(p', q_i) \le r_2 + r_2 + 3r_2 = 5r_2$, where $p' \in C(q_i) \cap \ensuremath{\mathsf{Ball}}$. This proves property \ref{prop:one}.
Let $\mathcal{X}_\ell \coloneqq X_2 \cap \lr{\lr{\bigcup_{q \in N_\ell} C(q)} \cup \lr{ \bigcup_{\ensuremath{\mathsf{Ball}} \in D_\ell} \ensuremath{\mathsf{Ball}} } }$. That is, $\mathcal{X}_\ell$ denotes the set of those points in $X_2$ that belong to the clusters of all the points in $N_\ell$, as well as those in the balls in $D_\ell$. Now, imagine calling \textsc{GreedyClustering}$(\mathcal{X}_\ell, X, r_2, 3, \omega_b)$. As in the proof of Lemma \ref{lem:greedy-ordering}, the main observation is that the set of clusters computed in the first $|N_\ell |$ iterations is exactly $\{C(q)\}_{q \in N_\ell}$. Thus, property \ref{prop:two} in the lemma follows from
Lemma~\ref{charikar-lemma} applied to \textsc{GreedyClustering}$(\mathcal{X}_\ell, X, r_2, 3, \omega_b)$.
\end{proof}
We now set $N = \bigcup_{1 \leq \ell \leq t} N_{\ell}$. Note that
\[ |N| = \sum_{\ell} |N_\ell | \leq \sum_{\ell} |D_\ell | = |\widehat{\mathcal{B}}_2| \leq |\mathcal{B}_2|. \]
Recall that for instance $\mathcal{I}'$, we set $\mathcal{B}'_2 = \{ B(q, 5r_2) \ | \ q \in N \}.$ We now argue that $N$ satisfies properties (\ref{eq:red}) and (\ref{eq:blue}).
By Property \ref{prop:one} of Lemma \ref{lem:blue-greedy}, we have that for any $\ensuremath{\mathsf{Ball}} \in \mathcal{B}_2$, there is a $\ensuremath{\mathsf{Ball}}' \in \mathcal{B}'_2$ such that $X_2 \cap \ensuremath{\mathsf{Ball}} \subseteq X_2 \cap \ensuremath{\mathsf{Ball}}'$. Thus, $\lr{X_2 \cap \bigcup_{\ensuremath{\mathsf{Ball}} \in \mathcal{B}_2} \ensuremath{\mathsf{Ball}}} \subseteq \lr{X_2 \cap \bigcup_{\ensuremath{\mathsf{Ball}} \in \mathcal{B}'_2} \ensuremath{\mathsf{Ball}}}$, which implies property (\ref{eq:red}).
Using Property \ref{prop:two} of Lemma \ref{lem:blue-greedy}, we have
\[\sum_{p \in N} wt(p) = \sum_{\ell} \sum_{p \in N_\ell} wt(p) \geq \sum_{\ell} \omega_b \left(X_2 \cap \bigcup_{\ensuremath{\mathsf{Ball}} \in D_\ell} \ensuremath{\mathsf{Ball}} \right) \geq \omega_b \left(X_2 \cap \bigcup_{\ensuremath{\mathsf{Ball}} \in \mathcal{B}_2} \ensuremath{\mathsf{Ball}} \right),\]
which is property (\ref{eq:red}).
\textbf{Phase 2.}
Phase 1 outputs an instance $\mathcal{I}' = ((X, d), (\lambda_r, \lambda_b, m_r, m_b), (k_1, k_2), (r'_1, r'_2) )$ of Colorful $2$-NU$k$C\xspace. In Phase 2, we transform this into an instance $\mathcal{I}'' = ((X,d), (\chi_r, \chi_b, m_r, m_b), (k_1, k_2), (r''_1, 0) )$ of Colorful $2$-NU$k$C\xspace where the radius at the second level is $0$.
In this step, we call the greedy clustering algorithm (Algorithm \ref{alg:greedyclustering}) using the {\em red weight function} $\lambda_r$. In particular, we will call \textsc{GreedyClustering}($X, X, r'_2, 3, \lambda_r$). This algorithm returns a set of points $M \subseteq X$, where every $p \in M$ has a cluster $C(p)$ and weight $wt(p)$ such that (1) $\{C(p)\}_{p \in M}$ is a partition of $X$, (2) For any $p \in M$, $wt(p) = \lambda_r(C(p))$, the red weight of the cluster, and (3) $d(q, p) \le 3r'_2$ for any $q \in C(p)$. Furthermore, the greedy algorithm naturally defines an ordering $\sigma = \langle p_1, p_2, \ldots, p_{|M|} \rangle$ of $M$ -- this is the order in which the points were added to $M$.
We define the red weight function $\chi_r$ for $\mathcal{I}''$ as follows: $\chi_r(p) \coloneqq \lambda_r(C(p))$ for $p \in M$, and $\chi_r(p) \coloneqq 0$ for $p \in X \setminus M$.
We define a $\phi: X \to M $ as follows: $\phi(p)$ is the first point in $M$ (according to $\sigma$) such that $B(p, r'_2) \cap C(p) \neq \emptyset$. Note that $\phi(p)$ exists and $d(p, \phi(p)) \leq 4 r'_2$. We define the blue weight function $\chi_b$ for $\mathcal{I}''$ as follows: $\chi_b(p) \coloneqq \sum_{q \in \phi^{-1}(p)} \lambda_b(q)$ for $p \in M$, and $\chi_b(p) \coloneqq 0$ for $p \in X \setminus M$.
Finally, we let $r''_1 = r'_1 + 4r'_2$, and obtain the instance $\mathcal{I}'' = ((X,d), (\chi_r, \chi_b, m_r, m_b), (k_1, k_2), (r''_1, 0) )$ of Colorful $2$-NU$k$C\xspace. The following lemma relates instances $\mathcal{I}'$ and $\mathcal{I}''$.
\begin{lemma}
\label{lemma:phase2:colorful}
(a) If instance $\mathcal{I}'$ has a feasible solution that is structured, then the instance $\mathcal{I}''$ has a feasible solution. (b) Given a solution $(\mathcal{B}''_1, \mathcal{B}''_2)$ for $\mathcal{I}'$ that uses at most $k_i$ balls of radius $\alpha r''_i$ for each $i \in \{1, 2\}$, we can obtain a solution $(\mathcal{B}'_1, \mathcal{B}'_2)$ for $\mathcal{I}'$ that uses at most $k_i$ balls of radius $\alpha r''_i + 4r'_2 \leq \alpha r'_i + (4 \alpha + 4) r'_2$ for $i \in \{1,2\}$.
\end{lemma}
Again, part (b) follows from the fact that in constructing $\mathcal{I}''$ from $\mathcal{I}'$, we move weights by a distance of at most $4 r'_2$. Note that we do not claim that the solution to $\mathcal{I}'$ constructed in part (b) is structured.
In the rest of this section, we establish part (a). Fix a feasible solution $(\mathcal{B}'_1, \mathcal{B}'_2)$ for $\mathcal{I}'$ that is also structured. Our construction of a feasible solution for $\mathcal{I}'$ is analogous to what we did in Phase 1.
Let $M_1 \subseteq M$ be the set of points $p$ such that there exists some point $x$ satisfying (i) $x$ is covered by a ball in $\mathcal{B}'_1$, and (ii) $d(x,p) \leq 4 r'_1$. Note that $M_1$ includes any $p \in M$ such $C(p)$ contains a point covered by a ball in $\mathcal{B}'_1$. Now let $M_2 = M \setminus M_1$; note that for $p \in M_2$, any point in $C(p)$ is either covered by a ball from $\mathcal{B}'_2$, or is an outlier. Let $X_i \coloneqq \bigcup_{p \in M_i} C(p)$ for $i = 1, 2$. Note that $X = X_1 \sqcup X_2$.
Let $\widehat{\mathcal{B}}'_2 = \{ \ensuremath{\mathsf{Ball}} \in \mathcal{B}'_2 \ | \ \ensuremath{\mathsf{Ball}} \cap X_2 \neq \emptyset \}.$ We invoke \textsc{Mapping Procedure}($M_2, \sigma, \widehat{\mathcal{B}}'_2, \{C(p\}_{p \in M_2}$) and $\mathcal{T} = \{(N_1, D_1), (N_2, D_2), \ldots, (N_t, D_t)\}$. We let $N = \bigcup_{1 \leq \ell \leq t} N_{\ell}$.
As in phase 1, we have that $|N| \leq |\widehat{\mathcal{B}}'_2| \leq | \mathcal{B}'_2 |$. The set $N$ also satisfies the following property, which is the analog of Property \ref{eq:blue}.
\begin{equation}
\sum_{p \in N} wt(p) \geq \lambda_r(X_2 \cap \bigcup_{\ensuremath{\mathsf{Ball}} \in \mathcal{B}'_2} \ensuremath{\mathsf{Ball}}) \label{eq:blue-red}
\end{equation}
We now construct a solution $(\mathcal{B}''_1, \mathcal{B}''_2)$ for $\mathcal{I}''$. The set $\mathcal{B}''_1$ is obtained by expanding each ball in
$\mathcal{B}'_1$ by an additive $4 r'_2$; each ball in $\mathcal{B}''_1$ has radius $r''_1$. Note that by definition of $M_1$, the balls in $\mathcal{B}''_1$ cover $M_1$. The set $\mathcal{B}''_2$ is obtained by including in it a ball of radius $0$ at each point in $N$. Note that $|\mathcal{B}''_2| = |N| \leq |\mathcal{B}'_2|$.
We now argue that $(\mathcal{B}''_1, \mathcal{B}''_2)$ provides adequate coverage. Red coverage is analogous to blue coverage in phase 1, using property \ref{eq:blue-red}:
\[ \chi_r(M_1) + \chi_r(N) = \lambda_r(X_1) + \sum_{p \in N} wt(p) \geq \lambda_r(X_1) + \lambda_r(X_2 \cap \bigcup_{\ensuremath{\mathsf{Ball}} \in \mathcal{B}'_2} \ensuremath{\mathsf{Ball}}) \geq m_r. \]
For blue coverage, let $Y \subseteq X$ denote the set of points {\em self-covered} by the structured, feasible solution $(\mathcal{B}'_1, \mathcal{B}'_2)$ with $\lambda_b(Y) \geq m_b$. We argue that for each $y \in Y$, we have $\phi(y) \in M_1 \cup N$. If $y$ is covered by a ball in $\mathcal{B}'_1$, then as $d(y, \phi(y)) \leq 4 r'_2$, we conclude that $\phi(y) \in M_1$ using the definition of $M_1$. Otherwise, $y$ is the center of some ball in $B(y,r'_2) \in \mathcal{B}'_2$. Assume $\phi(y) \not\in M_1$. Then by the definition of $\phi$, $\phi(y)$ is the first point $p \in M_2$ such that $B(y, r'_2)$ intersects $C(p)$. But this means $\phi(y)$ is the same as $\varphi( B(y, r'_2) )$ computed in \textsc{Mapping Procedure}($M_2, \sigma, \widehat{\mathcal{B}}'_2, \{C(p\}_{p \in M_2}$). Thus, $B(y, r'_2) \in D_{\ell}$ and $\phi(y) \in N_{\ell}$ for some pair
$(N_{\ell}, D_{\ell})$ in $\mathcal{T}$. We conclude $\phi(y) \in N = \bigcup_{\ell} N_\ell$.
Thus, the blue coverage of $(\mathcal{B}''_1, \mathcal{B}''_2)$ is at least
\[\chi_b(M_1) + \chi_b(N) \geq \sum_{p \in M_1 \cup N} \phi^{-1}(p) \geq \sum_{y \in Y} \lambda_b(y) \geq m_b.\]
This completes the proof of Lemma~\ref{lemma:phase2:colorful} and concludes our description of Phase 2. Combining Phase 1 and Phase 2, we conclude with the main result of this section.
\begin{theorem}
\label{theorem:self-coverage}
There is a polynomial-time algorithm that transforms a Colorful $2$-NU$k$C\xspace instance $\mathcal{I} = ((X, d), (\omega_r, \omega_b, m_r, m_b), (k_1, k_2), (r_1, r_2) )$ into an instance $\mathcal{I}''= ((X,d), (\chi_r, \chi_b, m_r, m_b), (k_1, k_2), (r''_1, 0) )$ of Colorful $2$-NU$k$C\xspace with $r''_1 = r_1 + 26 r_2$, and has the following properties: (a) If $\mathcal{I}$ has a feasible solution, then so does $\mathcal{I}''$; (b) Given an $\alpha$-approximate solution to $\mathcal{I}''$, we can construct, in polynomial time, a $c \cdot \alpha$-approximate solution to $\mathcal{I}$, where $c > 0$ is an absolute constant.
\end{theorem}
\begin{remark} In part (a), the feasible solution $(\mathcal{B}''_1, \mathcal{B}''_2)$ to $\mathcal{I}'$ that is constructed from feasible solution $(\mathcal{B}_1, \mathcal{B}_2)$ to $\mathcal{I}$ has the following useful property: for any $\ensuremath{\mathsf{Ball}} \in \mathcal{B}''_1$, the center of $\ensuremath{\mathsf{Ball}}$ is also the center of some ball in $\mathcal{B}'_1$.
\end{remark}
\iffalse
\textbf{Remark.} Consider an instance of Colorful $k$-Center problem with $\chi \ge 2$ colors, where we want to find a set of $k$ balls of the smallest possible radius that covers, the number of points covered of every color is at least the coverage requirement of that color. The algorithm described in this section gives an $O(\chi)$-approximation for this problem in $n^{O(\chi)}$ time, as follows. Let $r = r_2$ denote the optimal radius for this problem. Then, we set $k_1 = 0$, and $r_1 > r$. Then, we perform the greedy clustering w.r.t.\ the $i$-th color class in the $i$-th iteration, which returns a set of mega-points of color $i$. Then, $w_i(p)$ denotes the weight of the points in the cluster of such a mega-point $p$. Then, for $1 \le j \le i-1$, we define $w_j(p)$ as the weight of ``nearby'' points of color $j$, where ``nearby'' is defined to be the appropriate subset of the $O((i-1)r)$-neighborhood around $p$. This is similar to the Phase 2 of the analysis above. At the end of iteration $\chi$, we have a collection of mega-points $p$ with weights $w_1(p), \ldots, w_\chi(p)$, where the weights come from $O(\chi r)$-neighborhood around $p$. Then, the task is reduced to find a subset of $k$ points, such that the total $i$-th weight is at least the $i$-th coverage requirement. This can be found in time $n^{O(\chi)}$ time using Dynamic Programming. We omit the details.
\fi
\section{Setup for Robust $t$-NU$k$C\xspace} \label{sec:setup}
Let $\mathcal{I} = ((X, d), (\mathbbm{1}, m) (k_1, \ldots, k_t), (r_1, \ldots, r_t))$ be an instance of Robust $t$-NUkC. First we state the natural LP relaxation for $\mathcal{I}$. Recall that the goal is to cover at least $m$ points.
\begin{align*}
\sum_{v \in X} \mathsf{cov}(v) &\ge m
\\\sum_{u \in X} x_{i, u} &\le k_i &\forall 1 \le i \le t
\\\mathsf{cov}_i(v) &= \sum_{u \in B(v, r_i)} x_{i, u} &\forall 1 \le i \le t, \forall v \in X
\\\mathsf{cov}(v) &= \min \LR{\sum_{i = 1}^t \mathsf{cov}_i(v), 1} &\forall v \in X
\\x_{i, u} &\ge 0 &\forall 1 \le i \le t, \forall u \in X.
\end{align*}
Let $\mathscr{F}$ denote the set of all tuples of subsets $(S_1, \ldots, S_t)$, where $|S_i| \le k_i$ for $1 \le i \le t$. For $v \in X$, and $1 \le i \le t$, we say that $(S_1, \ldots, S_t) \in \mathscr{F}$ covers $v$ with radius $r_i$, if $d(v, S_i) \le r_i$. Let $\mathscr{F}_i(v) \subseteq \mathscr{F}$ denote the subset of solutions that cover $v$ with radius $r_i$ -- where, the sets $\mathscr{F}_i(v)$ of solutions are assumed to be disjoint by including a solution in $\mathscr{F}_i(v)$ of the smallest index $i$, if it appears in multiple such sets.
If the instance $\mathcal{I}$ is feasible, then the integer hull of the coverages, $\Poly^{\I}_{\cov}$ as given below, must be non-empty.
\begin{align*}
\Poly^{\I}_{\cov}:
\\ \sum_{v \in X} \sum_{i \in [t]} \mathsf{cov}_i(v) &\ge m
\\ \sum_{S \in \mathscr{F}_i(v)} z_S &= \mathsf{cov}_i(v) &\forall i \in [t], \forall v \in X
\\ \sum_{S \in \mathscr{F}} z_S &= 1
\\ z_S &\ge 0 &\forall S \in \mathscr{F}
\end{align*}
Next, we give a few definitions from \cite{ChakrabartyN21}, generalized to arbitrary $t \ge 2$, for the sake of completeness. These definitions are used in the round-or-cut framework that reduces an instance of Robust $t$-NU$k$C\xspace to Well-Separated Robust $t$-NU$k$C\xspace, as described in Section \ref{sec:t-to-robust}.
\paragraph{$t$-Firefighter Problem.}
The input is a collection of height-$t$ trees, where $L_1$ is the set of roots, and for any $v \in L_i$ with $i \ge 1$, $a_j(v)$ represents the ancestor of $v$ that belongs to $L_j$, where $1 \le j \le i$ ($a_i(v) = v$).
Furthermore, let $w: L_t \to \mathbb{N}$ be a weight function on the leaves. For a root $u \in L_1$, we use $\mathsf{Leaf}(u)$ to denote the set of leaves, i.e., nodes in $L_t$ in the tree rooted at $u$.
Note that the $\{ \mathsf{Leaf}(u): u \in L_1 \}$ partitions $L_t$. Thus, $((L_1, \ldots, L_t), (a_1, a_2, \ldots, a_t), \mathsf{Leaf}, w)$ completely describes the structure of the tree, where $a_i(v): \bigcup_{i \le j \le t} L_j \to L_i$ is an ancestor function as defined above. Now we define the $t$-FF problem.
\begin{definition}[$t$-FF Problem] \label{def:t-ff}
Given height-$t$ trees $( \mathcal{T} = (L_1, \ldots, L_t), (a_1, \ldots, a_t), \mathsf{Leaf}, w)$, along with budgets $(k_1, \ldots, k_t)$, we say that $T = (T_1, \ldots, T_t)$, with $T_i \subseteq L_i$ is a feasible solution, if $|T_i| \le k_i$ for $1 \le i \le t$. Let $\mathcal{C}(T) = \{ v \in L_t : a_i(v) \in T_i \text{ for some $1 \le i \le t$} \}$ be the set of leaves covered by the solution. Then, the objective is to find a feasible solution maximizing the weight of the leaves covered. This instance is represented as $\mathcal{I} = (\mathcal{T} = ((L_1, \ldots, L_t), (a_1, \ldots, a_t), \mathsf{Leaf}, w), (k_1, \ldots, k_t))$,
\end{definition}
\begin{definition}[The solution $y$] \label{def:soln-y}
Given $\mathsf{cov}$, and a collection $\mathcal{T}$ of rooted trees, let $L_1$ denote the set of roots, and let $L_i$, $i > 1$ denote the set of vertices at $j$-th level. Furthermore, for any node $v \in L_i$ with $i > 1$, let $a_j(v)$ denote the ancestor of $v$ that belongs to $L_j$, where $1 \le j < i$. Then, the solution $y$ is defined as follows.
$$y(v) = \begin{cases}
\mathsf{cov}_1(v) & \text{ if } v \in L_1
\\ \min \LR{ \mathsf{cov}_i(v), 1 - \sum_{j < i} \mathsf{cov}_j(a_j(v)) } & \text{ if } v \in L_i, i > 1
\end{cases}
$$
\end{definition}
\begin{definition}[The Sparse LP] \label{def:sparse-lp}
\begin{align*}
\max \sum_{v \in L_t} w(v) Y(v) &
\\\sum_{u \in L_1} y_u &\le k_1 - t
\\\sum_{u \in L_i} y_u &\le k_i \qquad \forall 2 \le i \le t
\\Y(v) \coloneqq y_v &+ \sum_{i = 1}^{t-1} y_{a_i(v)} \qquad \forall v \in L_{t}
\end{align*}
\end{definition}
We now describe two subroutines that are used in the Reduction from Robust $t$-NU$k$C\xspace to Well-Separated Robust $t$-NU$k$C\xspace. We use the same notation and convention as in \cite{ChakrabartyN21}. These two algorithms (Algorithm \ref{alg:hs} and Algorithm \ref{alg:cgk}) are named after Hochbaum, and Shmoys \cite{hochbaumS1985best}; and Chakrabarty, Goyal, and Krishnaswamy \cite{CGK20}, respectively.
\begin{algorithm}
\caption{HS$(\text{Metric space} (X, d), r \ge 0, \text{ assignment } \mathsf{cov}: X \to \mathbb{R}^+)$} \label{alg:hs}
\begin{algorithmic}[1]
\State $R \gets 0$
\While{$U \neq \emptyset$}
\State $u \gets \arg\max_{v \in U} \mathsf{cov}(v) $
\State $R \gets R \cup \{u\}$
\State $\mathsf{Child}(u) \gets \{ v \in U : d(u, v) \le r \}$
\State $U \gets U \setminus \mathsf{Child}(u)$
\EndWhile
\State \Return $R, \{ \mathsf{Child}(u) : u \in R \}$.
\end{algorithmic}
\end{algorithm}
\begin{algorithm}
\caption{CGK} \label{alg:cgk}
\begin{algorithmic}[1]
\Statex \textbf{Input: } Robust $t$-NU$k$C\xspace instance $\mathcal{I} = ((X, d), (\omega, m), (r_1, \ldots, r_t), (k_1, \ldots, k_t))$,
\Statex \qquad\qquad $(\alpha_1, \ldots, \alpha_t)$, where $\alpha_i > 0$ for $1 \le i \le t$,
\Statex \qquad\qquad $\mathsf{cov} = (\mathsf{cov}_1, \ldots, \mathsf{cov}_t)$, where each $\mathsf{cov}_i: X \to \mathbb{R}^+$
\For{$i = t$ \textbf{downto} $1$}
\State $(L_i, \LR{\mathsf{Child}_i(v) : v \in L_i}) \gets $ HS$((X, d), \alpha_i r_i, \mathsf{cov}'_i \coloneqq \sum_{j = 1}^i \mathsf{cov}_j)$
\EndFor
\State \emph{Construct and Return a $t$-FF instance using $\{L_i, \mathsf{Child}_{i} \}_{1 \le i \le t}$ as described below.}
\end{algorithmic}
\end{algorithm}
We construct the $t$-FF instance based on the sets $L_i$'s constructed, as follows. Consider some $1 \le i \le t-1$, and some $u \in L_i$. Then, for every $v \in \mathsf{Child}_i(u)$, we make $v$ a child of $u$ in a tree $T$. Note that $L_1$ is the set of roots of the trees constructed in this way. Then, we define $\mathsf{Leaf}(u) = \{ v \in L_t : \text{ $v$ is a leaf in the tree rooted at $v$ } \}$, and let $a_i: L_t \to \bigcup_{i \le j \le t} L_j$ be the ancestor function as defined above. Finally, for every $u \in L_t$, let $w(u) = |\mathsf{Child}_t(u)|$. Then, we return the $t$-FF instance $\mathcal{I} = (\mathcal{T} = ((L_1, \ldots, L_t), (a_1, \ldots, a_t), \mathsf{Leaf}, w), (k_1, \ldots, k_t))$.
\section{From Robust $t$-NU$k$C\xspace to Well-Separated Robust $t$-NU$k$C\xspace} \label{sec:t-to-robust}
In this section, we use the round-or-cut framework of \cite{ChakrabartyN21} to give a Turing reduction from Robust $t$-NU$k$C\xspace to (polynomially many instances of) \emph{Well-Separated} Robust $t$-NU$k$C\xspace. Furthermore, $c$-approximation for a feasible instance of the latter problem will imply an $O(c)$-approximation for the original instance of Robust $t$-NU$k$C\xspace.
\paragraph{Round-or-Cut Framework.} Let $\mathcal{I}= ((X, d), (\mathbbm{1}, m), (k_1, k_2, \ldots, k_t), (r_1, r_2, \ldots, r_t))$ be the given instance of Robust $t$-NU$k$C\xspace (we assume that we are working with unit-weight instance, where we want to cover at least $m$ points of $X$). We adopt the round-or-cut framework of \cite{ChakrabartyN21} (also \cite{chakrabarty2018generalized}) to separate an LP solution from the integer hull of coverages (see Section \ref{sec:setup} in the appendix for the definitions thereof). Even though \cite{ChakrabartyN21} discuss this for $t = 2$, it easily generalizes to arbitrary $t \ge 2$. Thus, we only sketch the high level idea.
Let $\mathsf{cov} = (\mathsf{cov}_1, \mathsf{cov}_2, \ldots, \mathsf{cov}_t: \forall v \in X)$ be a candidate solution returned by the ellipsoid algorithm. First, we check whether $\mathsf{cov}(X) \ge m$, and report as the separating hyperplane if this does not hold. Now, we call CGK Algorithm (see Section \ref{sec:setup}) with $\alpha_1 = 6$, and $\alpha_i = 2$ for all $2 \le i \le t$ to get a $t$-FF instance $(\mathcal{T} = ((L_1, \ldots, L_t), (a_1, \ldots, a_t), \mathsf{Leaf}, w), (k_1, \ldots, k_t))$. Here, for any $i \in [t]$, any distinct $p, q \in L_i$ satisfy that $d(p, q) > 3r_i$. Then, we let $\{y_v: v \in \bigcup_{i} L_i\}$ be the solution as defined in Section \ref{sec:setup}, see Definition \ref{def:soln-y}. Now we check if $\mathsf{cov}_i(L_i) \le k_i$ for $i \in [t]$, and report if any of these $t$ inequalities is not satisfied. Finally, the algorithm checks the value of $y(L_1)$, and branches into the following two cases.
In the first case, if $y(L_1) \le k_1 - t$, then as argued by \cite{ChakrabartyN21}, it can be shown that a sparse LP that is related to the $t$-FF problem (see Definitions \ref{def:t-ff} and \ref{def:sparse-lp}) admits an \emph{almost-integral} solution. That is, a basic feasible solution to the sparse LP contains at most $t$ strictly fractional variables. By rounding up all such variables to $1$, one can obtain an $O(1)$-approximation for the original instance $\mathcal{I}$. Note that here we need the assumption that the ratio between the values of consecutive radii is at least $\beta$ -- otherwise we can merge the two consecutive radii classes into a single class.
In the second case, $y(L_1) > k_1 - t$. In this case, we use a generalization of an argument from \cite{ChakrabartyN21} as follows. We enumerate every subset $Q \subseteq X$ of size at most $t-1$, and add a ball of radius $r_1$ around each point in $Q$. Let $X'$ be the set of points covered by balls of radius $r_1$ around $Q$. Then, we modify the weight of the points of $X'$ to be $0$, and let $\mathbbm{1}_{X\setminus X'}$ be the resulting weight function. Let $\mathcal{I}(Q) = ((X, d), (\mathbbm{1}_{X \setminus X'}, m-|X'|), (2r_1, r_2, \ldots, r_t), (k_1 - |Q|, k_2, \ldots, k_t))$ be the resulting residual instance of \emph{Well-Separated} $t$-NU$k$C\xspace, where the \emph{well-separatedness} property imposes that the $2r_1$ centers must be chosen from $Y \coloneqq L_1 \setminus Q$ -- note that the distance between any two distinct points in $L_1$, and thus $Y$, is at least $6r_1 = 3 \cdot 2r_1$, i.e., the set $Y$ is well-separated w.r.t. the new radius $r_1$. An argument from \cite{ChakrabartyN21} implies that if $\mathcal{I}$ is feasible, then either (a) at least one of the well-separated instances $\mathcal{I}(Q)$ is feasible for some $Q \subseteq X$ of size at most $t-1$, or (b) the hyperplane $y(L_1) \le k_1 -t$ separates the LP solution $\mathsf{cov}$ from the integer hull of coverages. Furthermore, an argument from \cite{ChakrabartyN21} implies that a constant approximation to any of the instances implies a constant approximation to $\mathcal{I}$.
Note that the ellipsoid algorithm terminates in polynomially many iterations, and each iteration produces at most $n^{t}$ instances of \emph{Well-Separated} Robust $t$-NU$k$C\xspace. Thus, we get the following theorem.
\begin{theorem} \label{theorem:robust-to-ws}
Suppose there exists an algorithm that, given an instance $\mathcal{J}$ of \emph{Well-Separated} Robust $t$-NU$k$C\xspace, in time $f(n, t)$, either finds an $\alpha$-approximation to $\mathcal{J}$, or correctly determines that $\mathcal{J}$ is not feasible. Then, there exists an algorithm to obtain an $c \cdot \alpha$-approximation for any instance of Robust $t$-NU$k$C\xspace, running in time $n^{O(t)} \cdot f(n, t)$.
\end{theorem}
\iffalse
\paragraph{Algorithm.}
Suppose we have an instance $\mathcal{I} = ((X, d), (k_1, k_2, k_3), (r_1, r_2, r_3), m)$ of Robust $3$-NUkC problem. First, without loss of generality, we may assume that for some large constant $\beta \ge 1$, it holds that $r_1 \ge \beta r_2 \ge \beta r_3$. Suppose not, and say $r_{1} < \beta r_2$. Then, an $O(1)$-approximation for the $2$-NUkC instance $((X, d), (k_1 + k_2, k_3), (r_1, r_3), m)$ --- which can be obtained by the results of CN --- implies an $O(\beta)$-approximation for $\mathcal{I}$. A similar argument implies that $r_2 \ge \beta r_3$. We will fix the value of $\beta$ later.
As in CN, we apply the round-on-cut methodology on $\Poly^{\I}_{\cov}$, stated earlier. Given a $\hat{\mathsf{cov}} = (\hat{\mathsf{cov}}_i(v): i \in [3], v \in X)$, we want to either find an $O(1)$-approximation for $\mathcal{I}$, or find a hyperplane separating from $\Poly^{\I}_{\cov}$.
First, we check whether $\mathsf{cov}(X) \ge m$, and report as the separating hyperplane if this does not hold. Now, we call CGK Algorithm to get a $3$-FF instance $\mathcal{T} = ((L_1, L_2, L_3, \mathsf{Leaf}, w), (k_1, k_2, k_3))$. Let $\{y_v: v \in L_1 \cup L_2 \cup L_3\}$ be the solution as described in preliminaries. Now we check if $\mathsf{cov}_i(L_i) \le k_i$ for $i \in [3]$, and report if any of these three inequalities is not satisfied. Finally, the algorithm checks the value of $y(L_1)$, and branches into the following two cases.
\textbf{Case 1.} $y(L_1) \le k_1 - 3$.
In this case, we want to obtain a feasible solution to the instance $\mathcal{I}$ by looking at a simpler LP solution. In particular, consider the following LP.
\begin{align*}
\text{max} \sum_{v \in L_3} w(v) \cdot Y(v) & \quad
\\\text{s.t. } \qquad \quad\ \ \sum_{u \in L_1} y_u &\le k_1 -3
\\ \sum_{u \in L_2} y_u &\le k_2
\\ \sum_{u \in L_3} y_u &\le k_3
\\ Y(u) &= y_u + y_{p(u)} + y_{p(p(u))} &\forall u \in L_3
\end{align*}
Since $y(L_1) \le k_1 -3$, there is a feasible solution to the previous LP with objective value at least $m$. Now, let $(y'(v): v \in L_1 \cup L_2 \cup L_3)$ be a solution, and let $T_1 = \{ v \in L_1 : y'_v \ge 0 \}$. Now, $y'(T_1) = y'(L_1) \le k_1 -3$. Note that this LP has $|L_3|$ variables (essentially, as $Y(v)$ is only a dummy variable), and $|L_3|+3$ constraints. Therefore, there are at most $3$ variables that are strictly fractional. Therefore, $|T_1| \le k_1$. Our goal is to pick subsets $T_2 \subseteq L_2, T_3 \subseteq L_3$, such that the solution $(T_1, T_2, T_3)$ covers the maximum weight of leaves. To this end, let $U_1$ be the set of leaves not covered by $T_1$, and let $T_2 \subseteq L_2$ denote a subset of size at most $k_2$ that covers the maximum weight of leaves from $U_1$ -- this can be found by sorting the vertices of $L_2$ by non-decreasing weight of leaves covered, and picking the first $k_2$ vertices according to this order. Similarly, let $U_2$ denote the set of leaves not covered by $T_1 \cup T_2$, and let $T_3 \subseteq L_3$ denote a subset of size at most $k_3$ that covers the maximum weight of leaves in $U_2$. By combining the ``greedy'' manner in which the subsets $T_2, T_3$ were chosen, along with the exchanging arguments as in pipage rounding, it can be shown that the solution $(T_1, T_2, T_3)$ has weight at least $m$. Then, we obtain a feasible solution for the instance $\mathcal{I}$ with $O(1)$-approximate guarantee.
\textbf{Case 2. } $y(L_1) > k_1 -3$.
In this case, we will either find an $O(1)$-approximation, or prove that the hyperplane $\hat{\mathsf{cov}}_1(L_1) \le k_1 -3$ separates $\hat{\mathsf{cov}}$ from $\Poly^{\I}_{\cov}$. Similar to CN, we will prove the following lemma.
\begin{lemma} \label{lem:well-separated}
Let $(\mathsf{cov}_1, \mathsf{cov}_2, \mathsf{cov}_3) \in \Poly^{\I}_{\cov}$ be fractional coverages, and suppose there exists a subset $Y \subseteq X$ with $d(u, v) > 5 r_1$ for all $u, v \in Y$.
Let $\mathcal{Q} = \{Q \subseteq X : |Q| \le 2 \text{ and } \forall q \in Q: d(q, Y) > r_1 \}$. For every $Q \in \mathcal{Q}$, define an Well-Separated $3$-NUkC instance $\mathcal{I}_Q \coloneqq ((X \setminus \bigcup_{q \in Q} B(q, r_1)), (k_1-|Q|, k_2, k_3), (2r_1, r_2, r_3), Y, m - |\bigcup_{q \in Q} B(q, r_1)|)$.
Either $\mathsf{cov}_1(Y) \le k_1 -3$, or at least one of the $O(n^2)$ instances $\mathcal{I}_S: S \in \S$ is feasible.
\end{lemma}
\begin{proof}
We will assume that $\mathsf{cov}_1(Y) > k_1 - 3$, and prove that one of the instances of the form $\mathcal{I}_Q$ are feasible. Since $d(u, v) > 5 r_1$ for any $u, v \in Y$. Therefore, each of the instances is, in fact, well-separated w.r.t.\ the set $Y$.
Since $\mathsf{cov}_1(Y) > k_1 -3$, and $\mathsf{cov} \in \Poly^{\I}_{\cov}$, there exists some $S = (S_1, S_2, S_3) \in \mathscr{F}$ such that $S_1$ covers strictly more than $k_1 - 3$ points of $Y$. Fix such an $S = (S_1, S_2, S_3)$. Let $A \coloneqq \{ a \in S_1 : d(a, Y) > r_1 \}$, and $B \coloneqq \{ b \in S_1 : d(b, Y) \le r_1 \} = S_1 \setminus A$. Since any distinct points in $Y$ are more than $5 r_1 > 2r_1$ apart, no center in $A$ can serve two distinct points of $Y$. Thus, $|A| > k_1 -3$, which implies $|B| < 3$, i.e., $|B| \le 2$. Therefore, $B \in \mathcal{Q}$ -- note that $B$ may be equal to $\emptyset$, which also belongs to $\mathcal{Q}$.
Let $\mathcal{I}_B$ be the Well-Separated $3$-NUkC instance as defined in the statement of the lemma. We will show that $\mathcal{I}_B$ is feasible by constructing a solution $S'$. Note that the coverage requirement of $\mathcal{I}_B$ is $m' \coloneqq m - |\bigcup_{q \in B} B(q, r_1)|$, which is equal to the number of points covered by the solution $(A, S_2, S_3)$ in $\mathcal{I}$. Thus, we need to argue that the solution $S'$ covers at least $m'$ points in $\mathcal{I}_B$. To this end, let $S'_1 \coloneqq \{ f \in Y : d(f, A) \le r_1 \}$, $S'_2 \coloneqq S_2$, $S'_3 \coloneqq S_3$, and let $S' = (S'_1, S'_2, S'_3)$. Consider $f \in Y$, and the corresponding $a \in A$ such that $d(f, a) \le r_1$. Then, $B(a, r_1) \subseteq B(f, 2r_1)$. Furthermore, $|S'_1| = |A|$. Therefore, the solution $S'$ is feasible for the instance $\mathcal{I}_B$.
\end{proof}
{\Large \red{I thought the argument of Chakrabarty et al.\ (cases 1 and 2 and Lemma \ref{lem:well-separated} above) did not generalize for $t \ge 3$, but that is not true. Even though they do not state it as such, it should generalize for arbitrary $t \ge 3$ with suitable modifications. I guess it is up to us whether to redo this analysis. But if we do, I guess it is better to state it for general $t$ for the sake of completeness, otherwise it doesn't make much sense to re-prove for a specific value of $t$ (i.e., 3).}}
Now, we use Lemma \ref{lem:well-separated} to obtain either an $O(1)$-approximation for $\mathcal{I}$, or report a separating hyperplane. Given a $\hat{\mathsf{cov}}$ with $y(L_1) > k_1 - 3$, we create $O(n^2)$ instances $\mathcal{I}_Q$ as defined in the statement of Lemma \ref{lem:well-separated}, by letting $Y \gets L_1$. Note that $L_1$ is well-separated by construction. Now, Lemma \ref{lem:well-separated} implies that either one of the instances $\mathcal{I}_Q$ is feasible, or $\hat{\mathsf{cov}}_1(Y) \le k_1 -3$ is a separating hyperplane.
Fix an instance $\mathcal{I}_Q$. The main idea is to reduce $\mathcal{I}_Q$ to $O(n^3)$ instances of (Well-Separated) Colorful $2$-NUkC problem with self-coverage (C$2$-NUkC-SC), by chaining the reductions from Section 2 and 3. We will crucially need the well-separated property (given by the set $Y$) to solve the latter problem. Therefore, we will need to ``keep track'' of the points of $Y$ as they are clustered into mega-points as a result of greedy clustering used at various places. We will observe that, the weight of any mega-point comes from its $c_3 r_3$ (resp. $c_2 r_2$) neighborhood during the reduction of Section 2 (resp. Section 3). However, the points of $Y$ are at least $5r_1$ apart, and we may assume wlog $r_1 \gg r_2 \ge r_3$. Therefore, the mega-points ``representing'' the points of $Y$ in the final instance of C$2$-NUkC-SC, will be more than $5 r_1 - \max(c_2 + c_3) \cdot (r_2 + r_3) \ge 4 r_1$ apart. Theorems 2 and 3 imply that, assuming $\mathcal{I}_Q$ is feasible, at least one of the $O(n)$ instances of C$2$-NUkC-SC will be feasible. Furthermore, an approximate solution to the C$2$-NUkC-SC instance, will imply an $O(1)$-approximation for $\mathcal{I}_Q$ by translating this solution back.
First, we use the reduction from Section 2 to reduce the instance $\mathcal{I}_Q$ to obtain $O(n^3)$ instances of Colorful $2$-NUkC using Theorem \ref{thm:t-to-t-1}. First, if $\mathcal{I}_Q$ is feasible, then one of the $O(n^3)$ instances of Colorful $2$-NUkC is feasible. Using the properties of the construction, it follows that any red/blue point receives its red/blue weight from within its $3r_3$ neighborhood. Thus, if distinct points $y_1, y_2 \in Y$ contribute to the red/blue weights of distinct points $p, q \in \ensuremath{\mathsf{Red}} \cup \ensuremath{\mathsf{Blue}}$, then $d(p, q) > d(y_1, y_2) - 6r_3 = 5 r_1 - 6r_3$. Furthermore, in each of the Colorful $2$-NUkC instances, the radii incur an additive $O(r_3)$ factor -- $r'_i = r_i + 9r_3$ for $i = 1, 2$. Now, each of these Colorful $2$-NUkC instances is transformed into an instance of C$2$-NUkC-SC by two rounds of greedy clustering at the $O(r_2)$ level. Thus, in the final instance, a red mega-point receives its coverage from its $23r'_2 = 23(r_2 + 9r_3) = 23 r_2 + 207r_3$ neighborhood. Again, the distance between two red mega-points that have a point of $Y$ contributing to their weight, is at least $5 r_1 - (46r_2 - 414r_3)$. However, the radius $r''_1 = r'_1 + 54r'_2 = r_1 + 54(r_2 + 9r_3) = r_1 + 54r_2 + 486r_3$, which implies that, $5 r_1 - 46r_3 - 414r_2 = 5 r''_1 - 316r_2 - 2844r_3$. Thus, by assuming that $r_1 > 3160r_2$, the distance between the corresponding red mega-points is greater than $4r''_1$, which proves the well-separated property. Finally, an $O(1)$-approximation to the C$2$-NUkC-SC instance that is feasible, will imply an $O(1)$-approximation to the instance $\mathcal{I}_Q$, which in turn, implies an $O(1)$-approximation to the original instance $\mathcal{I}$.
On the other hand, if each of the $O(n)$ instances of C$2$-NUkC-SC turn out to be infeasible, then we conclude that the original instance $\mathcal{I}_Q$ was infeasible. Finally, if each of the $O(n^2)$ instances $\mathcal{I}_Q$ are infeasible, then Lemma \ref{lem:well-separated} implies that $\hat{\mathsf{cov}}_1(Y) \le k_1 -3$ is a separating hyperplane, which can be reported to the Ellipsoid algorithm.
\fi
\section{Solving Well-Separated Colorful $2$-NUkC} \label{sec:dp}
We assume that we are given a \emph{well-separated} instance $\mathcal{I} = ((X, d), (\omega_r, \omega_b, m_r, m_b) (k_1, k_2), (r_1, 0))$ of Colorful $2$-NU$k$C\xspace. The \emph{well-separatedness} of the instance comes with the following additional input and restriction -- we are given an additional set $Y \subseteq X$ as an input. The set $Y$ is \emph{well-separated}, i.e., for any $u, v \in Y$, $d(u, v) > 2 r_1$. The additional restriction is that, the set of centers of balls of radius $r_1$ must be chosen from the set $Y$. We sketch how to solve such an instance optimally in polynomial time using dynamic programming.
Let $z \coloneqq |Y|$, and let $Y = \{y_1, y_2, \ldots, y_z\}$. For $1 \le i \le z$, let $X_i \coloneqq B(y_i, r_1) \cap X$, and let $X_{z+1} \coloneqq X \setminus \lr{\bigcup_{1 \le i \le z} X_i}$. Note that $\{X_i\}_{1 \le i \le z+1}$ is a partition of $X$.
For any $X' \subseteq X$ and non-negative integers $k, n_r, n_b $, let $F(X', k, n_r, n_b)$ be \textbf{true} if there exists a subset $X'' \subseteq X'$ of size at most $k$, and (red, blue) weight at least $(n_r, n_b)$; and \textbf{false} otherwise.\footnote{We use \emph{$X''$ has (red, blue) weight at least $(n_r, n_b)$} as shorthand for $\omega_r(X'') \geq n_r$ and $\omega_b(X'') \geq n_b$.} For a particular subset $X'$, the value of $F(X', k, n_r, n_b)$ can be found in polynomial time using dynamic programming, since the values $k, n_r, n_b$ are at most $n$.
For $(1, 0, 0, 0, 0) \le (i, k'_1, k'_2, n_r, n_b) \le (z+1, k_1, k_2, m_r, m_b)$, let $G(i, k'_1, k'_2, n_r, n_b)$ be \textbf{true} if it is possible to obtain (red, blue) coverage of at least $(n_r, n_b)$ from the set of points $\bigcup_{1 \le j \le i} X_j$, using at most $k'_1$ balls of radius $r_1$ and $k'_2$ balls of radius $0$; and \textbf{false} otherwise. Note that if $G(i-1, k'_1, k'_2, n_r, n_b) = \textbf{true}$, then $G(i, k'_1, k'_2, n_r, n_b)$ is trivially \textbf{true}. Otherwise, suppose some points in $X_i$ are covered. We consider two possibilities: either (A) $X_i$ is covered using a ball of radius $r_1$ (note that for $i \le z$ this is possible by definition; for $i = z+1$ we omit this case), and the remaining (red, blue) coverage comes from $\bigcup_{1 \le j \le i-1} X_j$, or (B) We use some $1 \le t \le \min\{k'_2, |X_i|\}$ balls of radius $0$ to achieve the (red, blue) coverage of $(n'_r, n'_b)$ from within $X_i$, and the remaining (red, blue) coverage comes from $\bigcup_{1 \le j \le i-1} X_j$. Note that in case (B), for a fixed guess of $(t, n'_r, n'_b)$, the subproblem for $X_i$ corresponds to $F(X_i, t, n'_r, n'_b)$ as defined in the previous paragraph, and can be solved in polynomial time. It is straightforward to convert this recursive argument to compute $G(z+1, k_1, k_2, m_r, m_b)$ into a dynamic programming algorithm that also finds a feasible solution, and it can be implemented in polynomial time. We omit the details.
|
\section{Introduction}
\label{sec:introduction}
Set theory is usually traced back to Cantor who used sets in an informal way giving rise to what is called \emph{naive set theory}. Nowadays we usually refer to axiomatic set theory which was formulated by Zermelo and Fraenkel and which is called Zermelo-Fraenkel Set Theory or short ZFC. When saying Set Theory
\footnote{I am capitalising both \textbf{Set Theory} and \textbf{Type Theory} when referring to specific theories with this name not in the sense of a set theory or a type theory.}
we mean ZFC.
Type Theory was introduced by Per Martin-L\"of \cite{martin1975intuitionistic} and there are several incarnations. The early \emph{Extensional Type Theory} (ETT) \cite{martin1984intuitionistic} gave way to \emph{Intensional Type Theory} (ITT) \cite{nordstrom1990programming} but recently, heavily influenced by \mbox{Voevodsky} using concepts from Homotopy Theory, Homotopy Type Theory (HoTT) \cite{hottbook} was developed.
\footnote{Since not everybody wants to read the whole book, introductory presentations of HoTT include
\cite{altenkirch2019naive,ahrens2019univalent,grayson2018introduction}.}
When saying Type Theory we mean HoTT.
In Set Theory all mathematical objects are viewed as sets and we write $a \in A$ to mean that $a$ is an element of the set $A$. Since in modern Set Theory we don't use \emph{urelements}, $a$ is a set again, a process that only stops at the empty set $\{\}$. The same element can occur in different sets and two sets are equal iff they have the same elements (axiom of extensionality).
Set Theory
is based on classical logic
\footnote{There are alternatives like CZF and IZF that are based on intuitionistic logic, see \cite{sep-set-theory-constructive}}
, formally it is presented in the framework of first order predicate logic. Set Theory can be viewed as an alternative to higher order logic where the quantification over predicates is replaced by quantifying over sets, corresponding to predicates. There are alternative set theories which use intuitionistic predicate logic as the framework but they all have in common that we first fix the logic as a framework and then formulate the set theory within this logic.
Type Theory is based on the idea that all mathematical objects belong to a type and can only be understood as elements of a given type. This corresponds to mathematical practice where we conceive of a statement quantifying over all natural numbers as only talking about those while in Set Theory this is represented by quantifying over all sets and singling out those which are elements of a particular one, the set of natural numbers. In Type Theory an element can only be an element of one type
\footnote{I am here presenting a pure form of Type Theory. They are extensions which allow subtyping, which is usually an attempt to recover features of Set Theory. I prefer to view these sort of developments as notational conveniences as opposed to foundational features.}
hence we cannot talk about elements in isolation which means we cannot reason about representations. This discipline enables a high degree of extensionality which is realised by Voevodsky's univalence axiom entailing that isomorphic structures are equal.
Types have elements which we write as $a : A$ meaning $a$ is an element of type $A$. While superficially similar to the set-theoretic $a \in A$ it is fundamentally different. Saying $a : A$ is part of the static structure of a mathematical text, only texts which use types consistently make sense. In particular we cannot ask whether $a : A$ because this is an aspect of our text not the subject we are talking about. We say that $a : A$ is a judgement not a proposition.
The formal presentation of Type Theory also requires another judgement which is called definitional or judgemental equality. E.g. when we define $n : \mathbb{N}$ as $n :\equiv 3$ we also know that $(1,2,3) : \mathbb{N}^n$ but to see this we need to exploit the fact that $n \equiv 3$ statically. Again $a \equiv b$ is a judgement which is a static property of a mathematical text not a predicate. Judgemental equality also reflects the computational character of Type Theory, e.g. we know that all (closed) natural number expressions can be reduced to a numeral, e.g. $3 + 4 \equiv 7$. Type Theory also introduces \emph{propositional equality} but it makes sense to distinguish the static judgemental equality from the usual propositional equality. Indeed being equal by definition is a common phrase in mathematics.
Mathematicians often consider ZFC as the only foundation of mathematics, and frequently don't actually want to think much about foundations. We argue here that modern Type Theory, i.e. HoTT, is a preferable and should be considered as an alternative. We will look at different aspects of type-theoretic and set-theoretic foundations and compare the two: we start with the different approaches to logic, we consider the question of representations and in particular representation independence, which leads to the univalence principle which changes the way we understand equality. Then we discuss the relation between the two central aspects where Type Theory and Set Theory are different: structuralism and constructivity and make the novel observation that the two are connected. We conclude with a personal reflection which also includes a short summary of the basic idea of the type theoretic approach.
\section{Related work}
\label{sec:related-work}
In \cite{maddy} the author develops a set of criteria for possible foundations and compares Set Theory, category theory and Homotopy Type Theory using these criteria. First of all it seems rather misleading to include category theory as a possible foundation because while it provides \emph{essential guidance} (one of the criteria) it requires another foundation to explain what basic constructions are intuitively permissible.
\footnote{Indeed in \cite{maddy}, Maddy complains: \emph{I don’t know what Voevodsky finds lacking in category-theoretic
foundations -- perhaps that it fails to provide a Generous Arena?}}
Indeed categorical guidance can be misleading since it would suggest that symmetric cartesian categories would be a good concept but you need to go beyond category theory to see that only preorders satisfy this condition. However, category theory, fits very well with HoTT, indeed the universal properties have unique solutions in HoTT, while they are only unique up to isomorphisms. Maddy in \cite{maddy}, footnote 28, admits that she is \emph{not sure what these thinkers take to be wrong with ZFC}, where the thinkers seem to refer to Voevodsky and others. It is the purpose of this paper to explain what we think is wrong with ZFC and can be addressed by using HoTT. Maddy also argues that she thinks that HoTT would need to adopt the axiom of choice to be able to encode the classical theory of sets
\footnote{In loc.cit, p.29: \emph{But to get even to ETCS, we have to add the axiom of choice,} where ETCS stands for \emph{The Elementary Theory of the Category of Sets}}
--- which is in my experience
a common reaction of philosophers who are only too willing to give in to intimidation by classical mathematicians and let them define the yardsticks of mathematics. Her conclusion is that there is no problem with set theory and that the only advantage HoTT would have to offer is proof checking. While it is doubtful that this even correct (there are proof systems based on classical mathematics), it is also misleading. What we try to argue in this paper is that the fundamental point of HoTT is conceptual, the fact that it lends itself to implementation is a welcome additional benefit.
Our views are similar to \cite{awodey2014structuralism} who explains why HoTT is essential for structural mathematics. The present paper is different in that we emphasise also the constructive aspects of HoTT and relate them to its structural properties.
\section{The role of logic}
\label{sec:role-logic}
Type Theory does not rely on predicate logic as a framework but introduces logic using the propositions as types explanation --- in this sense logic is an emergent aspect of Type Theory not a prerequisite. In constructive thinking we acknowledge that mathematical constructions are taking place in our heads and that we communicate them using shared intuitions. From this point of view, it is not the notion of truth which is fundamental but the notion of evidence or proof. A proposition is given by saying what we need as evidence to accept that the proposition holds. In the context of Type Theory we can do this by assigning to every proposition the type of evidence for this proposition. This is also called the Curry-Howard equivalence but this is based on the assumption that we already have a logic and a Type Theory and then observe a formal relationship between the two, while I \emph{define} logic by the translation.
\footnote{The propositions as types explanation is also related to the Brouwer-Heyting-Kolmogorov (BHK) semantics which introduces untyped realisers of propositions.}
The translation of propositional connectives into type theoretic connectives is rather straightforward, for example we translate implication $P \implies Q$ as the function type $P \to Q$ between the associated types of evidence which I denote by the same letters here
\footnote{To be more precise we can introduce an operation $\mathrm{Prf}$ which assign to every proposition $P$ te type of its proofs $\mathbf{Prf}(P)$ and then we stipulate that $\mathbf{Prf}(P \implies Q) \equiv \mathbf{Prf}(P) \to \mathbf{Prf}(Q)$.}
. The idea is that evidence for $P$ implies $Q$ is a function which maps evidence for $P$ into evidence of $Q$ --- function types are a primitive concept of Type Theory
\footnote{We explain a function as \emph{black box} where we can input elements of the domain and which outputs elements of the codomain. While we have no access to the mechanism (hence the box is black) the only way to actually construct a function is via effective means. However, the theory works as well if we believe in some sort of magic, like that everything is decidable (the principle of the excluded middle). Formally, the properties of functions are represented by the laws of $\lambda$-calculus and the principle of functional extensionality.}.
Similar we can translate $\forall x:A.P(x)$ by the dependent function type
\footnote{The dependent function type is a generalisation of the usual function type where the codomain can vary over the domain. This corresponds to the cartesian product of a family of sets in Set Theory, hence the use of $\Pi$}
denoted as $\Pi x:A.P(x)$, the idea is that evidence for a universally quantified statement is a (dependent) function that assigns to every element $a : A$ evidence for $P(a)$. Negation $\neg P$ we interpret as saying that $P$ is impossible means $P$ implies falsum ($P \implies \mathrm{False}$) where falsum ($\mathrm{False}$) is interpreted as the empty type.
The translation of disjunction and existential quantification depends on what exactly we mean by a proposition. Here there has been a historic change mainly since the introduction of HoTT. In ITT we identified types and propositions which leads to some strange artefacts. For example we would denote a subset $\{ x : A \mid P(x) \}$ as the type of dependent pairs $(a,p)$ where $a : A$ and $p : P(a)$ (this is written as $\Sigma x:A.P(x)$)
\footnote{The justification for this notation is similar as before, in Set Theory the type of dependent pairs corresponds to an infinite sum hence $\Sigma$.}
. Now the embedding from the subset into $A$ is given by the first projection $\pi_1 : \{ x : A \mid P(x) \} \to A$
\footnote{Here $\pi_1 : \Sigma x:A.B(x) \to A$ is defined as $\pi_1(a,b) \equiv a$ and $\pi_2 : \Pi p : \Sigma x:A.B(x). B(\pi_1(p))$ with $\pi_2(a,b)\equiv b$.}.
However, it turns out that this is not in general an injection since we can have different proofs $p$ for the same element $a$. Also the naive translation of existential statements uses again dependent pairs, i.e. evidence for $\exists x:A.P(x)$ is given by a pair $(a,p)$ with $a : A$ and $p : P(a)$, we write this type as $\Sigma x:A.P(x)$ as before. However, this interpretation of existence gives rise to an apparent proof of the axiom of choice: if we know that for all $x : A$ there exists a $y : B$ such that a certain relation $R(x,y)$ holds then we can extract the witness and its justification from the proof. More precisely we can define
\begin{align*}
&\mathrm{ac} : (\forall x : A.\exists y:B . R(x,y)) \to \exists f : A \to B . \forall x:A.R(x,f(x))\\
&\mathrm{ac}(f) = (\lambda a. \pi_1(f(a)), \lambda a.\pi_2(f(x)))
\end{align*}
However, this \emph{type theoretic axiom of choice} hardly corresponds to the axiom of choice as it is used in Set Theory. Indeed, it is not an axiom but just a derivable fact.
In HoTT we say a proposition is a type with at most one element. The idea here is that a proposition should not contain any additional information, its only purpose is to assert something. As a consequence of univalence we obtain that two propositions are equal if they are logically equivalent --- this is called propositional extensionality. The embedding from a subset to a set ($\pi_1 : \{ x : A \mid P(x) \} \to A $) is now an injection because there is at most one proof that the proposition $P(a)$ holds. To translate disjunction and existence we use propositional truncation, that is to any type $A$ we introduce a proposition $||A||$ which corresponds to the proposition that $A$ is inhabited.
\footnote{Formally, $||A||$ can be defined as a higher inductive type with two constructors $|-| : A \to ||A||$ and $\mathrm{is-prop} : \forall x y : ||A||. x= y$.}
That is if there is an element $a:A$ we obtain an element of $||A||$ but we cannot distinguish them anymore. Now we can say that a proof of existence, i.e. evidence for $\exists x:A.P(x)$ is the propositional truncation of the naive translation, i.e. $||\Sigma x:A.P(x)||$. Given this interpretation the translation of
\[ (\forall x : A.\exists y:B . R(x,y)) \to \exists f : A \to B . \forall x:A.R(x,f(x))\]
is not provable since we extract an element from a propositional truncation which is clearly wrong because here all elements have been identified. And indeed this formulation of the axiom of choice has the power of the classical axiom, in particular we can derive the excluded middle using Diaconescu's construction.
By the principle of the excluded middle
\footnote{The term \emph{excluded middle} doesn't actually reflect its meaning. \emph{Excluded} clearly is a negative statement, hence we should render it is saying \emph{It cannot be that a proposition is neither true nor false}, which can be translated is $\neg\neg(P \vee \neg P)$, which \emph{is} constructively provable. A better term for the universal assertion of $P\vee\neg P$, would be \emph{universal decidability} or \emph{magicae nigrae} in latin.}
we mean
\[ \forall P : \mathbf{Prop} . P \vee \neg P \]
Here I write $\mathbf{Prop}$ for all \emph{propositional types}, i.e. types with at most one element and $P \vee Q$ is the translation of disjunction using propositional truncation, ie. $|| P + Q||$
\footnote{Here $A+B$ is the coproduct or binary sum in Type Theory corresponding to disjoint union in Set Theory. It can be defined as $A+B = \Sigma b : \mathrm{Bool}.P(B)$ with $P(\mathrm{true}) \equiv A$ and $P(\mathrm{false})\equiv B$.}
. Actually in this particular example we don't need to truncate because it can never happen that both $P$ and $\neg P$ hold. Clearly there is no function which assigns to every proposition either a proof of it or a proof of its negation. We don't even need to appeal to the undecidability of our logic but observe that we cannot look \emph{into} a proposition due to propositional extensionality. Hence excluded middle seems to suggest that we can decide any proposition without even looking at it!
However, we can justify excluded middle by a different translation of the connectives and a different explanation what is a proposition. That is we can interpret $P \vee_\mathrm{class} Q$ as the statement that not both are false $\neg(\neg P \wedge \neg Q)$ and similar we can explain existence $\exists_\mathrm{class} x:A.P(x)$ by saying it is not the case that $P$ is always false, i.e. $\neg (\forall x:A.\neg P(x))$. Note that we don't need any truncation because negated statements are always propositions. Moreover the translations are valid equivalences in classical logic. However, to obtain the correct behaviour of the connectives, in particular to justify reasoning by cases and the elimination rule of the existential, we need to limit propositions to the negative ones, i.e. to the ones for which proof by contradiction holds, that is we define $\mathbf{Prop}_{\mathrm class} = \{ P : \mathbf{Prop} \mid \neg \neg P \to P\}$. So for example if we know $P\vee Q$ and $P \to R$ and $Q \to R$ we can conclude $\neg \neg R$ and if $R$ is negative we also know $R$. Similar reasoning justifies the rule of existential elimination. Moreover the remaining connectives preserve the property of being negative. The negative translation I have described here is called the G\"odel-Gentzen translation, first described by G\"odel \cite{godel1933intuitionistischen}.
We may think that there is no difference between propositions as types with at most one element and negative propositions because in neither case can we extract information. However, while the axiom of choice is unprovable in both cases, the axiom of unique choice is provable for the type theoretic interpretation of propositions, that is
\begin{align*}
&\mathrm{auc} : (\forall x : A.\exists! y:B . R(x,y)) \to \exists f : A \to B . \forall x:A.R(x,f(x))\\
&\mathrm{auc}(f) = (\lambda a. \pi_1(f(a)), \lambda a.\pi_2(f(x)))
\end{align*}
where $\exists! x:A.P(x)$ means \emph{unique existence} which can be defined as
\[\exists x:A.P(x)\wedge \forall y:A.P(y) \implies x=y.\]
This can be proven in Type Theory because the statement of unique existence is already a proposition anyway and hence we can extract the witness. However, using the negative translation we cannot justify unique choice.
\section{Representation}
\label{sec:representation}
To illustrate the difference between Set Theory and Type Theory let's look at the representation of numbers, starting with the natural numbers. In Set Theory natural numbers are encoded using only sets, the most common encoding is due to von Neumann:
\begin{align*}
0 & = \{\} \\
1 & = \{ 0 \}\\
2 & = \{ 0 , 1 \}\\
\vdots & \quad \vdots \\
n & = \{ 0 , 1, \dots , n-1 \}
\end{align*}
However, there are many alternative ways, e.g. the following is due to Zermelo:
\begin{align*}
0 & = \{\} \\
1 & = \{ 0 \}\\
2 & = \{ 1 \}\\
\vdots & \quad \vdots \\
n & = \{ n-1 \}
\end{align*}
So for example the number 3 is represented as $\{\{\},\{\{\}\},\{\{\},\{\{\}\}\}$ using von Neumann's encoding and as $\{\{\{\}\}\}$ in Zermelo's. The axiom of infinity uses one particular representation to introduce the existence of the infinite set of natural numbers, e.g. von Neumann's, and using the axiom of replacement we can deduce that the alternative representation, e.g. Zermelo's also forms a set.
We can formulate properties that distinguish the two representations, for example in von Neumann's encoding we have that $n \subseteq n+1$ which fails in Zermelo's, e.g. $1 \not\subseteq 2$ if we use Zermelo's encoding. This statement is not really a statement about numbers but about their encoding, we say it is not structural. Mathematicians would normally avoid non-structural properties, because they entail that results are may not be transferable between different representations of the same concept. However, frequently non-structural properties are exploited to prove structural properties and then it is not clear whether the result is transferable.
In Type Theory natural numbers are an example of an inductive type, defined by the constructors. Following Peano we say that natural numbers are generated from $0 : \mathbb{N}$ and $\mathrm{suc} : \mathbb{N} \to \mathbb{N}$ (the successor, $\mathrm{suc}(n)$ means $n+1$). To define functions out of the natural numbers we use an elimination principle that allows us to perform dependent recursion, that is given a family of types indexed by the natural numbers $M : \mathbb{N} \to \mathbf{Type}$ and an element $z : M(0)$ and $s : \Pi n:\mathbb{N} . M(n) \to M(\mathrm{suc}(n))$ we can define a function $f : \Pi n:\mathbb{N}.M(n)$ with the definitional equalities $f(0)\equiv m(0)$ and $f(\mathrm{suc}(m)) \equiv s(n,f(n))$. The elimination principle allows us to define functions like addition but also to derive properties, e.g. that addition forms a commutative monoid exploiting the propositions as types explanation.
The definition of the natural numbers precisely identifies the structural properties of the type: it is equivalent to stating that $\mathbb{N}$ is an initial algebra, a natural number object, in the language of category theory. And because we cannot talk about elements in isolation it is not possible to even state non-structural properties of the natural numbers. Indeed, we cannot distinguish different representations, for example using binary numbers instead.
Both in Set Theory and in Type Theory we can go on and construct more elaborate sets/types of numbers, e.g. $\mathbb{Z},\mathbb{Q},\mathbb{R},\mathbb{C}$. In Set Theory we can talk about subsets, e.g. we may want to stipulate that $\mathbb{N} \subseteq \mathbb{Z}$ but it turns out that this doesn't hold for the most natural choice of encodings.
\footnote{E.g. we may represent integers as equivalence classes of pairs of natural numbers with the same difference, i.e. $(a,b) = (c,d)$ if $a+d = c+b$.}.
Yes, we can start with the biggest number class we want to represent and then define the other ones as subsets via comprehension.
In Type Theory the notion of a subset doesn't make sense in general for types. We can observe that we have a canonical embedding function between the different number types, e.g. $\mathbb{N} \to \mathbb{Z}$ and we can agree notationally that we can omit this function when it is clear from the context that we use a natural number when an integer was required.
While we don't have a notion of subset on types, we can talk about subsets of a given type $A$ which are just propositionally valued functions $P : A \to \mathbf{Prop}$. Given any two of such subsets we can define $P \subseteq Q$ as $\forall x:A.P(x) \to Q(x)$ in Type Theory. This means we can actually play the same trick as in Set Theory and define our number classes as subsets of the largest number class we want to consider and we have indeed the subset relations we may expect.
Hence Type Theory allows us to do basically the same things as Set Theory as far as numbers are concerned (modulo the question of constructivity) but in a more disciplined fashion limiting the statements we can express and prove to purely structural ones.
One may think that Type Theory is somehow less minimalistic, while in Set Theory we only need $\{\dots\}$ to represent all mathematical concepts there seems to be quite a menagerie of type constructors: $\Pi$-types, $\Sigma$-types, equality types and so on. However, to be fair we should compare the axioms of Set Theory with the basic constructions of Type Theory. And indeed while in applied Type Theory we may introduce a lot of apparently unrelated types we can often reduce them to a basic collection. So for example most inductive constructions can be reduced to the type of well-founded trees or $W$-types. The reduction of more advanced type constructors such as Higher Inductive-Inductive types (HIITs) are subject of current work, similar to the question whether advanced set theoretic constructions can be reduced to some axioms.
\section{Univalence and equality}
\label{sec:univalence-equality}
Type Theory doesn't allow us to make statements about representations because we cannot talk about elements in isolation. This means that we cannot observe intensional properties of our constructions. This already applies to Intensional Type Theory, so for example we cannot observe any difference between two functions which are pointwise equal. However, on the other hand in ITT we cannot prove them to be equal either. This is a consequence of the way the equality type is treated in ITT. That is for any element $a : A$ there is a canonical element of the equality type $\mathrm{refl}(a) : a = a$. Now consider $f,g : \mathbb{N} \to \mathbb{N}$ with
\begin{align*}
f(x) & = x + 0 \\
g(x) & = 0 + x
\end{align*}
Using induction it is easy to show that $f(x) = g(x)$ for all $x:\mathbb{N}$. However, the only prove of $f = g$ would need to be canonical (since we have no other assumptions) i.e. $\mathrm{refl}(h)$ but this would require that $h \equiv f \equiv g$ and $f$ and $g$ are clearly not definitionally equal.
Hence in ITT while we cannot distinguish extensionally equal functions we do not identify them either. This seems to be a rather inconvenient incompleteness of ITT, which is overcome by HoTT. In HoTT a type doesn't just have elements but also it comes with an equality type that tells what is the type of evidence that two objects are equal. In the case of the function type this evidence is equivalent to functional extensionality. Another example is propositional extensionality: two propositions are equal if and only if they are logically equivalent. By incorporating these principles Type Theory is as extensional as Set Theory but it turns out it is much better.
This is a consequence of the fact that we cannot observe the representations of objects. What does this mean for types: what can we observe about a type? Now we can observe the cardinality of a type --- e.g. whether it has two elements or whether it is countably infinite and so on --- but nothing else. Hence extensionally two types with the same number of elements should be considered equal. This may appear counterintuitive on the first glance because it entails for example that $\mathbb{Z} = \mathbb{N}$. But then can you write a property of a type that distinguished $\mathbb{Z}$ from $\mathbb{N}$? You may say that addition behaves differently for the two but you just have the type and no operation on them. Indeed, if we talk about the algebraic structure, e.g. we consider them as commutative monoids, then they are not equivalent because we can distinguish them.
It is a consequence of the univalence principle that two isomorphic types are equal
\footnote{The actual statement of univalence is slightly different and uses the notion of an \emph{equivalence} which refines isomorphism in a setting where equality is not always propositional. A function $f : A \to B$ is an equivalence if for any $b : B$ there is a unique $a :A$ \textbf{and} $p : f(a) = b$. Since equivalence is reflexive there is an obvious function from equality of two types to their equivalence. Univalence states that this function is an equivalence. }
where by isomorphic I mean that there a functions in both directions which when composed are equal to the identity. So basically two types are equal if there is a one-to-one correspondence between their elements. This also applies to algebraic structures like the aforementioned commutative monoids. While the monoids associated to $\mathbb{Z}$ and $\mathbb{N}$ are not isomorphic we cannot distinguish different representations of the same algebraic structure, e.g. using unary or binary numbers to represent $\mathbb{N}$. And indeed it is a consequence of univalence that isomorphic algebraic structures are equal.
This reflects mathematical practice to view isomorphic structures as equal. However, this is certainly not supported by Set Theory which can distinguish isomorphic structures. Yes, indeed all structural properties are preserved but what exactly are those? In HoTT all properties are structural, hence the problem disappears.
There is a price to pay for this elegance. We usually expect that equality is a proposition but this ceases to hold when we say that evidence for equality of types is isomorphism because there is usually more than one isomorphism between isomorphic types.
\footnote{One can force equality to be a proposition by truncating it. However, this means that we loose important properties such as the ability to always replace equals by equals.}
The conclusion is that equality of structures in general is not a proposition but a structure itself.
What is equality actually? I said that every type comes with a notion of equality, with an equality type but what are the properties of equality types. If equality were just a proposition then it would be easy: equality is just an equivalence relation (i.e. reflexive, symmetric and transitive) and every function has to preserve it. This view is called the setoid interpretation and it is a useful easy case of the actual story.
\footnote{Setoids have been used to justify extensional principles, see \cite{altenkirch1999extensional,altenkirch2019setoid}.}
But since we have types and we can iterate the construction of equality types, i.e. we get a whole infinite tower of equalities, the answer is a bit more involved. The first step up from equivalence relations are groupoids, these are categories where every morphism corresponding to an equality type has an inverse. Or more naively, it is the combination of an equivalence relation and a group.
\footnote{A one object groupoid is a group while a groupoid where every homset is propositional is an equivalence relation.}
However, we need to iterate this process any number of times arriving at weak infinity groupoids. They are called weak because all the properties don't hold strictly but only upto higher equality. This is where homotopy theory comes in helpful because these structures have been investigated and they can be defined as simplicial sets with the Kan property or more exactly Kan fibrations for dependent types (see \cite{nlab:kan_fibration}). Actually in the context of Type Theory it turns out that cubical sets are technically better behaved.
\footnote{In \cite{cohen2016cubical} cubical sets are used to provide a constructive explanation of the univalence axiom. This semantics gives rise to cubical Type Theory,
which is now implemented in cubical agda.\cite{vezzosi2021cubical}. Cubical Type Theory can be understood as a refinement of HoTT. }
\section{Structuralism and constructivity}
\label{sec:struct-constr}
We may think that the constructive nature of Type Theory which comes from the propositions as types approach to logic is rather orthogonal to its support of structural mathematics which stems from the fact that elementhood is treated statically. Indeed, we can add classical principles like the unrestricted axiom of choice (and hence excluded middle) to HoTT and still have univalence. So are these two aspects orthogonal? Can we just have univalence but don't embrace constructive Mathematics?
While this is possible in principle, I would argue that there is indeed a connection between these two aspects. Univalence forces us to revisit the notion of equality and accept that equality isn't always a proposition. It is this step which at the same time renders many applications of choice unnecessary. Choice in conventional mathematics becomes necessary because we only talk about propositions, and hence choices need to be hidden in propositions. When we later need to refer to the choices we have just hidden, we need to use the axiom of choice. In Type Theory we can indeed exploit the fact that our constructions may be types and not propositions and the choices are explicit.
\footnote{There is another case where HoTT can avoid the use of choice, which comes from using Higher Inductive Types. This is used in the definitions of the Cauchy Reals in the HoTT book, \cite{hottbook} section 11.3, and for example in the definition of the partiality monad. \cite{altenkirch2017partiality}.}
A simple example for this phenomena is the way categories are treated in HoTT. Categories are a generalisation of preorders but indeed we prefer to deal with partial orders where the equivalence introduced by the order coincides with equality. Can we do the same for categories? Indeed in HoTT we can introduce the notion of a \emph{univalent category}
\footnote{In the terminology of the HoTT book, ordinary categories are called precategories and univalent categories are called categories.}
, where isomorphism and equality of objects are equivalent (i.e. isomorphic). And indeed most categories that arise from semantic considerations -- like the category of sets, or categories of algebraic structure -- are univalent. In contrast the counterpart to univalent categories in a classical setting are \emph{skeletal categories}, which are very rare and their constructions only succeeds by heavy use of the axiom of choice.
\section{Conclusions}
\label{sec:conclusions}
Axiomatic Set Theory was developed in the 1930ies and then it constituted considerable progress in making the foundations of mathematics precise. However, much has happened since and it is time to reconsider the foundations of mathematics, in particular in light of a shift of major applications of mathematics from natural science to computer science and related disciplines.
From the viewpoint of computer science, it seems almost preposterous to think that all mathematical concepts can be adequately represented as sets, which correspond to a type of trees where we ignore order and multiplicity of subtrees. This encoding seems to be not very different than the insight that all structures in computing can be represented as a sequence of 0s and 1s. Indeed, it seems to me that set theory is something like the machine language of mathematics, but it doesn't share the redeeming feature that in computing we \emph{need} to translate everything to machine language. Mathematics which is a construction of our mind doesn't need a machine language it should rather reflect our basic intuitions.
Set Theory enables us to ask stupid questions, like \emph{Are the integers a subset of the rationals? }or \emph{Is 2 a subset of 3}? Mathematicians tell me that they just don't ask these questions so they don't really need a better language. This reminds me, who worked as a programmer in the 1980ies, of the attitude of some of my colleagues who didn't see the need for high-level languages because they knew how to do the right thing on the machine level.
The basic tenants of Type Theory are this: mathematical concepts are represented as types, where every type comes with a static notion of elementhood and an equality type. The main property of the equality type is that objects are equal if they represent the same mathematical object, in particular we view objects as equal if they satisfy the same properties -- this is just paraphrasing Leibniz. A consequence of this is that equalities aren't always propositions because they may be me more than one way that two objects can be equal. Logic emerges in Type Theory, via the propositions as types explanation, which replaces the philosophically difficult concept of truth with the notion of evidence which can be expressed as a type.
As far as constructivity is concerned, I would like to quote Frank Pfenning who recently said in a lecture \emph{I have lost the ability to understand classical logic}. I share the same problem, specifically with the boolean explanation of logic: it seems to me that we cannot \emph{assign} a boolean to every formula in arithmetic. Now I understand that \emph{assign} here seems to refer to a \emph{classical function}. However, and this may be because I am computer scientist, it seems to me a strange abuse of the word assignment if you have one but cannot tell me what is the output for a given input.
It seems to me that the classical approach can only succeed by an intensive process of brain-washing, replacing naive concepts, like function, which rather elaborate replacements. It seems that many mathematicians think that this process is necessary for them to continue their leading edge research in modern mathematics. That may well be so, but maybe we shouldn't give up the hope that there is the possibility for a more na\"ive form of mathematics which favours explainability over latest fashion research.
\bibliographystyle{alpha}
|
\section{Introduction}\label{sec:intro}
Symmetries have long been a foundational concept and tool in physics. In particular, continuous symmetries are those described by transformations that vary continuously as a function of some parameterization, mathematically modeled by Lie groups.
There is a vast range of different continuous symmetry groups
that may naturally arise in physical scenarios, which are associated with corresponding conservation laws as dictated by the celebrated Noether's theorem \cite{noether2005invariant}. In quantum mechanics, two basic but important examples are $U(1)$ and $SU(2)$ symmetry groups, respectively associated with a conserved charge (particle number, energy) and spin polarization (isospin) conservation.
A phenomenon that has drawn great recent interest in quantum information and physics is that continuous symmetries place fundamental limitations on the accuracy of quantum error correction (QEC)~\cite{hayden2017error,eastin2009restrictions}, which was introduced as a technique to protect quantum computation~\cite{shor1995scheme,nielsen2002quantum,gottesman2010introduction,lidar2013quantum} and recently found to play fundamental roles in many areas in physics such as quantum gravity \cite{almheiri2015bulk,pastawski2015holographic} and condensed matter physics \cite{kitaev2003fault,zeng2019quantum,brandao2019quantum,PhysRevLett.120.200503}. More specifically, the Eastin--Knill theorem \cite{eastin2009restrictions} indicates that if a (finite-dimensional) code implements any continuous group of logical gates transversally (such codes are dubbed \emph{covariant codes}), then it cannot exactly correct local errors. Moreover, there has been a series of recent works that further investigate approximate QEC by such codes, providing bounds on their accuracy as well as explicit constructions~\cite{faist2019continuous,woods2019continuous,wang2019quasi,kubica2020using,zhou2020new,yang2020covariant,wang2021theory,tajima2021symmetry,KongLiu21:random}. These results have significant implications to fault-tolerant quantum computation due to the desirable transversality feature~\cite{shor1996fault,nielsen2002quantum,gottesman2010introduction,lidar2013quantum}. Remarkably, covariant codes have also found intriguing connections to several important physical scenarios, in particular, quantum reference frames \cite{hayden2017error,woods2019continuous}, quantum many-body systems \cite{brandao2019quantum,PhysRevLett.120.200503,wang2019quasi}, and AdS/CFT correspondence \cite{harlow2018symmetries,harlow2018constraints,kohler2019toy,faist2019continuous,woods2019continuous}.
When symmetries arise in theoretical studies, they are usually assumed to be exactly respected by default. Indeed, the existing results on covariant codes~\cite{hayden2017error,faist2019continuous,woods2019continuous,wang2019quasi,kubica2020using,zhou2020new,yang2020covariant,wang2021theory,tajima2021symmetry,KongLiu21:random} are mostly concerned with the accuracy of error correction under exact symmetries. However, it is often important or even necessary to consider cases where the symmetries or conservation laws are only approximate in physical scenarios. For instance, the noise and imperfections in quantum many-body systems and other symmetry breaking mechanisms may cause deviations from the exact symmetry conditions and conservation laws. In particle physics, it is also well known that many fundamental symmetries are only approximate \cite{Witten2018}. More notably,
for quantum gravity, it is commonly believed that exact global symmetries are fundamentally forbidden \cite{Misner1957,Giddings1988,PhysRevD.52.912,Arkani_Hamed_2007,BanksSeiberg11} (justified in more concrete terms in AdS/CFT \cite{harlow2018constraints,harlow2018symmetries}).
Alas, our understanding of approximate symmetries, especially on a quantitative level, is very limited, raising the need for a systematic study of symmetry violation measures. In particular, for QEC and associated problems in physics,
it is imperative to understand quantitatively how symmetry violation is induced from QEC accuracy.
The goal of this work is to establish a rigorous and comprehensive theory of approximate continuous symmetry measures in quantum channels and codes, and in particular, the interplay between them and QEC accuracy, which allows us to formally understand symmetry violation in QEC codes.
More specifically, we introduce three different meaningful measures of the degree of symmetry violation respectively in terms of group-global and group-local covariance violation, and charge conservation violation, based on which quantitative notions of approximately covariant codes are defined correspondingly.
{We establish a series of trade-off bounds between the QEC inaccuracy and the above approximate symmetry measures, employing techniques from approximate QEC~\cite{knill1997theory,beny2010general}, quantum metrology~\cite{giovannetti2011advances,kolodynski2013efficient,zhou2020theory} and quantum resource theory~\cite{chitambar2019quantum,marvian2020coherence,FangLiu19:nogo,fang2020no}.}
In particular,
our results induce various lower bounds on different forms of symmetry violation for the commonly studied exact QEC codes, which imply fine-grained restrictions on the transversally implementable logical gates, improving the Eastin--Knill theorem.
To exemplify the general theory, we present two explicit families of approximately covariant codes that nearly saturate certain lower bounds.
In the end, we provide a blueprint for several potential applications to quantum gravity and condensed matter physics.
The main goal of this Letter is to elucidate the intuitions behind our approaches and report the key results. Interested readers may refer to the companion paper \cite{SM} for details, additional results, and more in-depth discussions.
\section{Characterizing approximate symmetries in quantum channels and codes}
We first discuss the quantitative characterization of symmetry violation in quantum dynamics from a general standpoint. Let $G$ be a compact Lie group corresponding to the continuous symmetry of interest. Denote by ${\mathcal{E}}_{B\leftarrow A}$ a quantum channel from system $A$ to system $B$.
The channel exactly respects symmetry $G$ if it is covariant with respect to the group actions, i.e.,
$
{\mathcal{E}}_{B\leftarrow A}\circ {\mathcal{U}}_{A,g} = {\mathcal{U}}_{B,g}\circ {\mathcal{E}}_{B\leftarrow A}
$
or equivalently ${\mathcal{U}}_{B,g}^\dagger \circ {\mathcal{E}}_{B\leftarrow A} \circ {\mathcal{U}}_{A,g} = {\mathcal{E}}_{B\leftarrow A}$ for all $g\in G$,
where we use ${\mathcal{U}}(\cdot) := U(\cdot )U^\dagger$ to denote the channel action of unitary $U$, and $U_g$ is given by some unitary representation of $G$ (on the appropriate system).
To characterize the deviation from the exact symmetry, we may consider the mismatch between the two sides of the covariance condition.
Then an intuitive group-global measure, is the maximum mismatch as given by some channel distance $D$:
\begin{align}
\delta_\group := \max_{g\in G} D({\mathcal{E}}_{B\leftarrow A}\circ {\mathcal{U}}_{A,g}, {\mathcal{U}}_{B,g}\circ {\mathcal{E}}_{B\leftarrow A}).
\end{align}
Note that we will not explicitly write down the arguments of the measures as long as they are unambiguous.
Another meaningful notion is the group-local symmetry violation around a certain point $g_0$ in the group at which the symmetry condition holds, i.e., ${\mathcal{E}}_{{B\leftarrow A}}\circ{\mathcal{U}}_{A,g_0} = {\mathcal{U}}_{B,g_0}\circ {\mathcal{E}}_{{B\leftarrow A}}$. We are interested in the local geometry of the mismatch around this point.
Let the symmetry actions be parametrized by $\bm{\theta} = \{\theta_k\in\mathbb{R}\}$ via some unitary representation which gives $U_g = e^{i \bm{J} \cdot \bm{\theta}}$ where $\bm{J} = \{J_k\}$ are infinitesimal generators of $G$.
\begin{equation}
\label{eq:point-general}
\delta_{\mathrm{P}}:=\\ \sqrt{2 \nabla^2 D({\mathcal{E}}_{B\leftarrow A}\circ {\mathcal{U}}_{A,g}, {\mathcal{U}}_{B,g}\circ {\mathcal{E}}_{B\leftarrow A})^2|_{g=g_0}},
\end{equation}
(the square root and the coefficient $\sqrt{2}$ in the definition are chosen to simplify calculations later), where $(\nabla^2 \mathscr{P})_{ij} = \partial^2 \mathscr{P}/\partial\theta_i\partial\theta_j$ is the Hessian matrix of function $\mathscr{P}$ \footnote{The general existence and properties of the Hessian matrix are left for study elsewhere.}.
Finally, it is natural to consider the deviation from conservation laws. Specifically, each $J_k$ is associated with a charge, and we can quantify the variation of the charge for input $\rho$ by
$
\delta_{{\mathrm{C}},k}(\rho) := \abs{\mathrm{Tr} J_{k,B}{\mathcal{E}}_{B\leftarrow A}(\rho) - \mathrm{Tr} J_{k,A}\rho}
$,
where $J_{k,A}$ and $J_{k,B}$ are the appropriate generators on systems $A$ and $B$ so the trace gives the expectation value of the charge.
Note that in the QEC context, one usually considers isometric encoding channels ${\mathcal{E}}$, for which symmetries imply corresponding conservation laws, but this is not true for general quantum channels~\cite{cirstoiu2020robustness}---$\delta_{\mathrm{C}}$ are not necessarily zero for covariant ${\mathcal{E}}$. Also note that $\delta_{\mathrm{P}}$ and $\delta_{\mathrm{C}}$ only depend on the local geometry of the symmetry group,
for which reason we shall refer to both as local symmetry measures.
Now we apply the above discussion to our QEC setting in more concrete terms. Consider the channel purified distance~\cite{gilchrist2005distance,tomamichel2015quantum,LiuWinter19} defined by $P(\Phi_1,\Phi_2) := \max_{\rho} P((\Phi_1\otimes{\mathbbm{1}})(\rho),(\Phi_2\otimes{\mathbbm{1}})(\rho))$, where $P(\rho,\sigma) := \sqrt{1 - f(\rho,\sigma)^2}$ is the purified distance between quantum states with fidelity $f(\rho,\sigma) := \mathrm{Tr}(\sqrt{\rho^{1/2}\sigma\rho^{1/2}})$ \footnote{In this work we extensively use the purified distance and channel purified distance, which are well-behaved distance measures and commonly used in the quantum information literature. Furthermore, the choice of purified distance directly relates the local covariance violation to the well known quantum Fisher information (QFI)~\cite{helstrom1976quantum,holevo1982probabilistic,hubner1992explicit,sommers2003bures}. In principle, one may also consider other metrics. In \cite{SM}, we also discuss the situations where one uses the diamond distance~\cite{watrous2018theory}, another standard channel distance measure.
}.
A QEC code is defined by an encoding channel ${\mathcal{E}}_{{S\leftarrow L}}$ that maps a logical system $L$ to a physical system $S$, and $S$ is subject to a noise channel ${\mathcal{N}}_S$. Ideally, for an exact QEC code, we can find a recovery channel ${\mathcal{R}}_{{L\leftarrow S}}$ that achieves ${\mathcal{R}}_{{L\leftarrow S}}\circ {\mathcal{N}}_{S}\circ{\mathcal{E}}_{{S\leftarrow L}} = {\mathbbm{1}}_L$ where ${\mathbbm{1}}_L$ is the logical identity channel, indicating that the noise effects are perfectly recovered by the QEC procedure.
For cases where QEC is only done approximately, we characterize the optimal inaccuracy by the \emph{QEC inaccuracy}
\begin{align}
\epsilon &:= \min_{{\mathcal{R}}_{{L\leftarrow S}}} P({\mathcal{R}}_{{L\leftarrow S}}\circ {\mathcal{N}}_{S}\circ{\mathcal{E}}_{{S\leftarrow L}}, {\mathbbm{1}}_L).
\end{align}
From here on, we shall base our discussion on $U(1)$ symmetry, which is of fundamental importance in itself and sufficient to reveal the key phenomena. Consider the family of logical gates $U_{L,\theta} = e^{-iH_L\theta}, \theta \in {\mathbb{R}}$ implemented by physical gates $U_{S,\theta} = e^{-iH_S\theta}, \theta \in {\mathbb{R}}$, which are $U(1)$ symmetry actions respectively generated by non-trivial Hamiltonians (charge observables) $H_L$ and $H_S$.
Let ${\mathcal{N}}_{S}(\cdot) = \sum_i K_{S,i}(\cdot)K_{S,i}^\dagger$ where $K_{S,i}$ are Kraus operators. A sufficient condition for the non-existence of exactly covariant QEC codes is $H_S \in {\rm span}\{K_{S,i}^\dagger K_{S,j},\,\forall i,j\}$, which we will refer to as the ``Hamiltonian-in-Kraus-span'' (HKS) condition~\cite{zhou2020new}. A key scenario obeying the HKS condition that we will revisit is when $H_S$ is 1-local (corresponding to transversal $U_{S,\theta}$) and ${\mathcal{N}}_S$ is a single-erasure noise where one subsystem chosen uniformly at random is erased.
When the HKS condition fails, examples of exactly covariant QEC codes exist, e.g., when ${\mathcal{N}}_{S} = {\mathbbm{1}}$ (noiseless dynamics), when $H_S$ is a Pauli-X operator and ${\mathcal{N}}_S$ is dephasing noise~\cite{kessler2014quantum,arrad2014increasing}, and when $H_S$ is 2-local and ${\mathcal{N}}_S$ is single-erasure noise~\cite{gottesman2016quantum}.
Therefore, we will assume the HKS condition holds true through out this Letter and study the trade-off between approximate QEC and approximate covariance (see \figref{fig:setting}).
\begin{figure}[bt]
\centering
\includegraphics[width=0.38\textwidth]{setting.pdf}
\caption{\label{fig:setting}
{We study the trade-off between QEC inaccuracy (the deviation of the QEC procedure from the logical identity channel) and symmetry violation (the deviation of the encoding map from being covariant with respect to symmetry actions).}
}
\end{figure}
The formal notions of approximately covariant codes that we shall consider are as follows.
i)~\emph{Group-global covariance violation}: \begin{align}
\delta_\group := \max_\theta P({\mathcal{U}}_{S,\theta}\circ{\mathcal{E}}_{{S\leftarrow L}}, {\mathcal{E}}_{{S\leftarrow L}}\circ{\mathcal{U}}_{L,\theta}).
\end{align}
$\delta_\group = 0$ if and only if the code is exactly covariant.
ii)~\emph{Group-local~\footnote{In the following, we shall always refer to ``group-global'' and ``group-local'' as ``global'' and ``local'' for simplicity, which should not be confused with the spatially geometrical notions of ``global'' and ``local'' often used in physics.} (point) covariance violation} in the vicinity of a certain $\theta_0$ where the code is exactly covariant: We may assume $\theta_0 = 0$ without loss of generality because if not, we can always redefine the encoding channel to be ${\mathcal{U}}_{S,\theta_0}\circ {\mathcal{E}}_{{S\leftarrow L}}$ and the discussion follows. The quantum Fisher information (QFI)~\cite{helstrom1976quantum,holevo1982probabilistic,hubner1992explicit,sommers2003bures} $F(\rho_{\theta}) := 2\partial^2 P(\rho_{\theta},\rho_{\theta'})^2/\partial \theta'^2|_{\theta' = \theta}$ is a standard quantifier of the amount of information $\rho_\theta$ carries about $\theta$ locally~\cite{braunstein1994statistical,hubner1992explicit}. Letting $D$ be the channel purified distance in \eqref{eq:point-general} leads to the local covariance violation at the point $\theta = 0$:
\begin{equation}
\delta_{\mathrm{P}} := \sqrt{F({\mathcal{U}}_{S,\theta}\circ{\mathcal{E}}_{{S\leftarrow L}}\circ{\mathcal{U}}_{L,\theta}^\dagger)|_{\theta = 0}},
\end{equation}
where the channel QFI $F(\cdot)$ is given by $F(\Phi_\theta) = \max_{\rho} F((\Phi_\theta\otimes{\mathbbm{1}})(\rho))$~\cite{fujiwara2008fibre}.
iii)~\emph{Charge conservation violation}: Recall that the logical and physical charge observables (generators) are respectively Hamiltonians $H_L$ and $H_S$.
As mentioned, isometric covariant channels are always charge conserving, i.e., $H_L - \nu {\mathbbm{1}} = {\mathcal{E}}_{{L\leftarrow S}}^\dagger(H_S)$ (up to some constant offset $\nu$, which does not affect the $U(1)$ group representations) where ${\mathcal{E}}^\dagger_{{L\leftarrow S}}$ is the dual of the encoding channel~\cite{faist2019continuous}.
Hence, we consider
\begin{align}
\delta_{\mathrm{C}} := \Delta\big(H_L - {\mathcal{E}}_{{L\leftarrow S}}^\dagger(H_S)\big),
\end{align}
where $\Delta(\cdot)$ denotes the difference between the maximum and minimum eigenvalues of $(\cdot)$. Note that $\delta_{\mathrm{C}} = 2\min_{\nu\in{\mathbb{R}}} \max_\rho| \mathrm{Tr}(H_S{\mathcal{E}}_{{S\leftarrow L}}(\rho))-\mathrm{Tr} ((H_L-\nu {\mathbbm{1}})\rho)|$, where we allow a constant offset on the definitions of charges.
\section{Symmetry vs. QEC}
We now introduce our main results on the trade-off between these approximate symmetry measures and the QEC inaccuracy (see \cite{SM} for detailed proofs and extended results).
We first consider the global measure $\delta_\group$, which is of primary interest. The following two trade-off bounds can be established:
\begin{theorem}
\label{thm:global-1}
$\delta_\group \gtrsim \sqrt{({\Delta H_L - 2\epsilon {\mathfrak{J}}})/{\Delta H_S}}$, when ${\mathcal{E}}_{{S\leftarrow L}}$ is isometric.
\end{theorem}
\begin{theorem}
\label{thm:global-2}
$\epsilon + \delta_\group \gtrsim {\Delta H_L}/{\sqrt{4{\mathfrak{F}}}}$.
\end{theorem}
\noindent Here ${\mathfrak{J}} := \min_{h \in {\mathbb{H}}} \Delta h$ where ${\mathbb{H}}$ is the subset of Hermitian matrices $h$ such that $H_S = \sum_{ij} h_{ij} K_{S,i}^\dagger K_{S,j}$, and ${\mathfrak{F}}$ is the regularized channel QFI of the noisy physical gate ${\mathcal{N}}_{S}\circ{\mathcal{U}}_{S,\theta}$~\footnote{Note that the regularized QFI of a channel $\Phi_\theta$ is defined by ${F^\infty}(\Phi_\theta) = \lim_{N\rightarrow\infty} F(\Phi_\theta^{\otimes N})/N$~\cite{kolodynski2013efficient,zhou2020theory}.}. Both ${\mathfrak{J}}$ and ${\mathfrak{F}}$ are positive functions of $H_S$ and ${\mathcal{N}}_S$ computable using semidefinite programming~\cite{SM}.
Note that by ``$x \gtrsim y$'' we mean $x \geq \ell(y)$ for some function $\ell(y) = y + O(y^2)$.
\thmref{thm:global-1} and \thmref{thm:global-2} both give non-trivial lower bounds on $\epsilon$ for sufficiently small $\delta_\group$ and vice versa, demonstrating the robustness of the incompatibility between QEC and continuous symmetries. The two bounds can behave differently and complement each other in different settings. {For example, consider quantum codes in an $n$-partite system with 1-local $H_S$ such that $\Delta H_S = O(n)$ and $\Delta H_L = O(1)$, and single-erasure noise
}. We have ${\mathfrak{J}} = O(n)$, so \thmref{thm:global-1} gives a $\Omega(1/n)$ lower bound on $\epsilon$ for sufficiently small $\delta$, and a $\Omega(1/\sqrt{n})$ lower bound on $\delta$ for sufficiently small $\epsilon$. On the other hand, ${\mathfrak{F}} = O(n^2)$, so \thmref{thm:global-2} only gives a $\Omega(1/\sqrt{n})$ lower bound on $\delta$. For stronger noise such that ${\mathfrak{F}} = O(n)$, \thmref{thm:global-2} can be comparable to the first one or even outperform it, as discussed in~\cite{SM}.
We now explain the ideas behind the proofs. \thmref{thm:global-1} is derived from a notion that we call \emph{charge fluctuation} defined by $\chi := \braket{0_L|{\mathcal{E}}^\dagger_{{L\leftarrow S}}(H_S)|0_L} - \braket{1_L|{\mathcal{E}}^\dagger_{{L\leftarrow S}}(H_S)|1_L}$, where $\ket{0_L}$ and $\ket{1_L}$ are respectively eigenstates of $H_L$ corresponding to the largest and the smallest eigenvalues. For exact QEC codes, $\chi = 0$ because of the HKS condition and the Knill--Laflamme condition~\cite{knill1997theory} which indicates that $\Pi K_{S,i}^\dagger K_{S,j} \Pi \propto \Pi$ for all $i,j$ where $\Pi$ is the projection onto the code subspace.
Meanwhile, for exactly covariant codes, $\chi = \Delta H_L$, because ${\mathcal{E}}^\dagger_{{L\leftarrow S}}(H_S) = H_L - \nu {\mathbbm{1}}$ for some $\nu \in {\mathbb{R}}$~\cite{faist2019continuous}.
Then by relating $\delta_\group$ and $\epsilon$ respectively to $\abs{\Delta H_L-\chi}$ and $\abs{\chi}$, a trade-off can be established. To be specific, we have $\delta_\group \gtrsim \sqrt{\abs{\Delta H_L - \chi}/\Delta H_S}$ from the definition of $\delta_\group$, and $\abs{\chi} \leq 2\epsilon {\mathfrak{J}}$ by using the approximate Knill--Laflamme condition~\cite{beny2010general} (note that ${\mathfrak{J}}$ is closely related to the HKS condition).
\thmref{thm:global-2} comes from another notion that we call \emph{gate implementation error}, defined by $\gamma := \min_{{\mathcal{R}}_{{L\leftarrow S}}} \max_\theta P({\mathcal{R}}_{{L\leftarrow S}}\circ {\mathcal{N}}_{S}\circ{\mathcal{U}}_{S,\theta}\circ{\mathcal{E}}_{{S\leftarrow L}}, {\mathcal{U}}_{L,\theta})$, which quantifies the error of implementing an ideal set of logical gates ${\mathcal{U}}_{L,\theta}$ using the error-corrected noisy gates ${\mathcal{R}}_{{L\leftarrow S}}\circ{\mathcal{N}}_{S}\circ{\mathcal{U}}_{S,\theta}\circ{\mathcal{E}}_{{S\leftarrow L}}$. It can be shown that $\gamma \leq \delta_\group + \epsilon$~\cite{SM}, putting QEC accuracy and symmetry on the same footing. {The result then follows from ${\mathfrak{F}} \gtrsim (\Delta H_L)^2/(4\gamma^2)$, derived using the monotonicity of the regularized QFI~\cite{kolodynski2013efficient,zhou2020theory} (${\mathfrak{F}}$ necessarily upper bounds the regularized QFI of the error-corrected noisy gate which is approximately $(\Delta H_L)^2/(4\gamma^2)$)}.
Besides the above quantum metrology technique, we also derive another version of \thmref{thm:global-2} using quantum resource theory which can additionally induce trade-off on the average-case behavior over different input states~\cite{SM}.
When setting $\delta_\group = 0$, \thmref{thm:global-1} and \thmref{thm:global-2} recover previous results on exactly covariant codes~\cite{faist2019continuous,woods2019continuous,kubica2020using,zhou2020new,yang2020covariant}. Furthermore, our study here applies to general codes, in particular providing symmetry bounds on exact QEC codes.
A remarkable application of our symmetry bounds is the following result
on the optimal set of transversal logical gates for general QEC codes -- a key problem in fault tolerance -- refining the Eastin--Knill theorem:
\begin{corollary}
\label{col:gate}
Suppose an $n$-qudit QEC code with distance at least $2$ admits a transversal implementation $V_S = \bigotimes_{l=1}^n e^{-i2\pi T_{S_l}/D}$ of the logical gate $V_L = e^{-i2\pi T_L/D}$ where $D$ is a positive integer and $T_{L,S}$ are Hermitian operators with integer eigenvalues~\footnote{The size of the logical system can in principle be arbitrary here.}.
It holds that $D = O\big(\max\big\{{(\sum_{l=1}^n \Delta T_{S_l})^{3/2}}/{\Delta T_L^{1/2}},\Delta T_L\big\}\big)$.
\end{corollary}
\noindent
A simple and common case is $T=Z$, where larger $D$ corresponds to phase rotations with higher precision.
For stabilizer codes, \corollaryref{col:gate} implies that $\tilde{V}_S = \bigotimes_{l=1}^n e^{-i 2\pi a_l Z_l/D}$ where $a_l$ is an integer and $D$ is a power of two (which is the most general form of transversal logical gates up to local Clifford equivalences~\cite{zeng2011transversality,anderson2016classification}), can only implement logical gates up to the \mbox{$O(\log n)$-th} level of the Clifford hierarchy when $a_l = O(\mathrm{poly}(n))$~\cite{SM}.
Note that several similar results~\cite{zeng2011transversality,bravyi2013classification,pastawski2015fault,anderson2016classification,jochym2018disjointness} were known for stabilizer codes, but our \corollaryref{col:gate} holds generally for arbitrary QEC codes.
A key insight behind \corollaryref{col:gate} is that the gate precision is quantitatively associated with the covariance violation with respect to the Hamiltonians $H_S = \sum_{\ell=1}^n T_{S_\ell}$ and $H_L = T_L$ that is bounded by our results above.
As for the local symmetry measures, we can also establish the following trade-off bounds \footnote{Note that \thmref{thm:local} holds true for arbitrary Hermitian operators $H_L$ and $H_S$, which do not necessarily share a common period like generators of $U(1)$ representations.}:
\begin{theorem}
\label{thm:local}
$\delta_{\mathrm{P}} + 2\epsilon{\mathfrak{J}} \geq \Delta H_L$ and $\delta_{{\mathrm{C}}} + 2\epsilon{\mathfrak{J}} \geq \Delta H_L$.
\end{theorem}
\noindent That is, $\delta_{\mathrm{P}}$ and $\delta_{\mathrm{C}}$ are lower bounded by a constant for small $\epsilon$ unlike $\delta_\group$.
However, also note that $\delta_{\mathrm{P}}$ and $\delta_{\mathrm{C}}$ may naturally be superconstant (e.g.,~consider the trivial encoding ${\mathcal{E}}_{{S\leftarrow L}} = {\mathbbm{1}}$, for which we usually have $\delta_{\mathrm{P}}=\delta_{\mathrm{C}}=\Delta(H_S-H_L) = \Theta(n)$), indicating that the constant or even sublinear scaling requires non-trivial code structures.
Note that \thmref{thm:global-1} and \thmref{thm:local} can be further improved by replacing ${\mathfrak{J}}$ with another quantity which is in general code-dependent but usually upper bounded by a smaller number (see \cite{SM}). Remarks on possible extensions of our definitions and results to non-compact groups and infinite-dimensional codes can also be found in \cite{SM}.
\section{Concrete examples}
We now present two explicit codes that demonstrate important approximate covariance features (details in~\cite{SM}).
\begin{enumerate}[(i)]
\item Quantum Reed--Muller code. We consider $[[n = 2^t -1,1,3]]$ ($t \geq 3$) quantum Reed--Muller codes which
{exactly corrects single-erasure noise~\cite{steane1999quantum,SM}.
They} admit a transversal implementation $\bigotimes_l \big(e^{i\pi Z_l/2^{t-1}}\big)$ of the logical operator $e^{-i\pi Z_L/2^{t-1}}$.
\item Modified thermodynamic code. We also construct a parametrized family of codes that exhibits the complete QEC-symmetry trade-off by modifying the thermodynamic code~\cite{brandao2019quantum,faist2019continuous}.
Our code is a function of a continuously tunable parameter $q \in [0,1]$, which is exactly covariant on the $q = 0$ end and exactly error-correcting on the $q = 1$ end.
\end{enumerate}
Consider single-erasure noise.
For (i),
we find that $\delta_\group \simeq \sqrt{4/n}$ for large $n$ (``$\simeq$'' indicates equivalence up to the leading order), saturating the lower bound $\simeq\sqrt{1/n}$ up to a constant factor.
For (ii), we find that $\delta_\group \simeq \sqrt{4q/n}$ and $\epsilon \simeq (1-q)m/2n$, both saturating the scaling of their lower bounds; when $q$ increases (decreases), the code becomes less (more) symmetric but more (less) accurate.
As for the local symmetry measures, in both examples, $\delta_{\mathrm{C}}$ saturates our lower bound up to the leading order, while $\delta_{\mathrm{P}}$ increases with $n$ {and has a gap with its lower bound}~\cite{SM}.
These two codes not only exemplify interesting constructions of approximately covariant codes, but also indicate the near-optimality of many of our bounds presented above.
\section{Potential physical applications}
We would also like to point out a few important areas in physics where our theory is potentially useful.
First, we expect our theory to provide new quantitative insights into the symmetry problem in quantum gravity, which is widely studied and of special importance to quantum gravity (see e.g.,~\cite{Misner1957,Giddings1988,PhysRevD.52.912,BanksSeiberg11,Arkani_Hamed_2007,harlow2018symmetries,Palti2019}), through the following lenses:
i) Holography and AdS/CFT correspondence: AdS/CFT is known to have fundamental connections with QEC \cite{almheiri2015bulk,pastawski2015holographic} and indeed, the no-global-symmetry arguments of Harlow and Ooguri \cite{harlow2018constraints,harlow2018symmetries} is underpinned by insights from QEC. In particular, for the continuous symmetry case, the situation becomes largely equivalent to Eastin--Knill (due to the transversality deduced from inherent properties of AdS/CFT \cite{almheiri2015bulk,PhysRevLett.117.021601,harlow2018constraints,harlow2018symmetries,faist2019continuous}),
indicating that our theory may induce quantitative statements about the breaking of global symmetries in AdS/CFT. ii) Black hole evaporation: A standard no-global-symmetry argument is based on certain inconsistencies between black hole evaporation and charge conservation \cite{BanksSeiberg11} (the ``weak gravity'' conjectures \cite{Arkani_Hamed_2007} are closely relevant). It could be interesting to apply our results to the evaporation process of charged black holes through e.g.,~the Hayden--Preskill model \cite{hayden2007}, which can be understood in terms of QEC (see also \cite{Yoshida:softmode,liu2020,Nakata20,tajima2021symmetry} which discuss Hayden--Preskill with symmetries). To summarize, in these scenarios, our theory potentially indicates interesting quantitative bounds on symmetry-violating effects (operators, terms, modes, etc.).
Note that there are some recent field or string theory calculations on approximate symmetries in quantum gravity \cite{Fichet2020,Hsin21,Chen2021} and it could be intriguing to draw comparisons with our quantum information results.
Moreover, QEC and symmetries naturally arise together in various areas in many-body physics, such as topological phases of matter \cite{Bravyi2010,Michalakis2013,Kitaev2003,Yoshida15:color,Yoshida2017,zeng2019quantum,RobertsBartlett20,wang2021comparative,moessner_moore_2021} and information scrambling \cite{PhysRevX.8.031057,PhysRevX.8.031058,Rakovszky19:renyi,Znidaric2020,Huang20,Yoshida:softmode,tajima2021symmetry,KongLiu21:random,Huang22}, where our theory of their competition may find applications. In particular, given the close relationship between QEC and entanglement, we expect our theory to also shed light on the interplay between symmetries and entanglement, which is of clear importance to the physics of many-body quantum systems.
\section{Summary and outlook}
In this work, we developed a systematic quantitative theory of the interplay between continuous symmetries and QEC by introducing several notions of approximately covariant codes based on both global and local symmetry violation and analyzing their QEC properties.
A key message is that the degree of symmetry (in multiple senses) and optimal QEC accuracy of a code are concurrently limited by trade-off relations between them, which may have numerous interesting implications in quantum computation and physics.
We point out a few directions that are worth further study. {First,
it would be interesting to further understand whether the two trade-off relations between global symmetry and QEC, which exhibit different behaviours under different noise models, can be unified.}
Also, a natural future task is to extend our study to more general symmetry groups such as $SU(d)$ with multiple generators and parameters, for which more involved representation theory machinery is expected to be useful~\cite{faist2019continuous,KongLiu21:random}.
Discrete symmetries are also broadly important and worth further exploring---although their incompatibility with QEC does not appear as fundamental as continuous symmetries~\cite{hayden2017error}, we do know interesting cases where it stems from simple additional constraints (e.g., AdS/CFT codes \cite{harlow2018constraints,harlow2018symmetries,faist2019continuous}). It would be interesting to understand the case of discrete symmetries in more general terms.
Finally, in the last section we pointed to a few relevant physical problems in the hope of enriching the interaction between quantum information and physics, and there could be more. It would be worthwhile to further consolidate these connections in physics languages.
\vspace*{0.2in}
\begin{acknowledgments}
We thank Daniel Gottesman, Liang Jiang, Dong-Sheng Wang, Beni Yoshida for relevant discussion and feedback.
Z.-W.L. is supported by Perimeter Institute for Theoretical Physics.
Research at Perimeter Institute is supported in part by the Government of Canada through the Department of Innovation, Science and Economic Development Canada and by the Province of Ontario through the Ministry of Colleges and Universities.
S.Z. acknowledges funding provided by the Institute for Quantum Information and Matter, an NSF Physics Frontiers Center (NSF Grant PHY-1733907). S.Z. also acknowledges support through the University of Chicago from ARO (W911NF-18-1-0020, W911NF-18-1-0212), ARO MURI (W911NF-16-1-0349), AFOSR MURI (FA9550-19-1-0399, FA9550-21-1-0209), DoE Q-NEXT, NSF (EFMA-1640959, OMA-1936118, EEC-1941583), NTT Research, and the Packard Foundation (2013-39273), where part of this work was done.
\end{acknowledgments}
\bibliographystyle{aps}
|
\section{Introduction}
\label{introduction}
\begin{figure*}
\centering
\includegraphics[width=1.0\linewidth]{finalPlot_r5to60_dT1_2_3_4_5_6_7_8_9_10_im.png}
\caption{The 126\arcsec$\times$126\arcsec thumbnail image (left) shows (248370)~2005~QN$_{173}${} (green dashed arrow) at center with a tail (white arrows) oriented towards 5 o'clock. This 89~s $z$-band exposure captured with the DECam is the only image in which we could unambiguously identify activity. We conducted wedge photometry (right) that shows the tail orientation is $251\fdg3\pm1\fdg4$ (blue star), in close agreement with the $251\fdg6$ antisolar angle (yellow $\odot$) and the $251\fdg7$ antimotion vector (red $v$) as computed by JPL Horizons. The plot shows counts radially outward from the the object center at (0,0).}
\label{fig:wedgephot}
\end{figure*}
Active asteroids are objects that are dynamically asteroidal but that display comet-like activity such as a tail or coma \citep{hsiehActiveAsteroidsMystery2006}. Activity may be caused by mechanisms unrelated to volatiles (e.g., impact, rotational disruption) or by sublimation as is typically the case with comets. Sublimation driven active objects provide key insights into the present-day volatile distribution in our solar system, as well as clues about the origins of those volatiles and how they arrived on Earth \citep{hsiehPopulationCometsMain2006}. These objects have been persistently difficult to study because of the small numbers detected to date: fewer than 30 active asteroids, of which roughly half are thought to exhibit sublimation driven activity; see \citealt{chandlerSAFARISearchingAsteroids2018} for a summary.
When the aforementioned sublimation driven activity is connected with a main-belt asteroid, the object is classified as a main-belt comet (MBC). MBCs are often characterized by activity near perihelion and the absence of activity elsewhere in the orbit \citep{hsiehMainbeltCometsPanSTARRS12015,agarwalBinaryMainbeltComet2017,hsieh2016ReactivationsMainbelt2018}, suggesting that the primary activity mechanism is sublimation of volatiles such as water ice \citep{snodgrassMainBeltComets2017}.
By contrast, stochastic events like impacts may result in comet-like activity but, in such cases, the appearance of activity is expected to cease once the material dissipates. Roughly 60\% of known active asteroids have been observed to display activity during only a single apparition \citep{chandlerSAFARISearchingAsteroids2018}.
Asteroid (7968), now comet 133P/Elst-Pizarro, was the first active main-belt asteroid to be discovered. While it was unclear at the time whether the activity was sublimation driven \citep{boehnhardt1996_133p,boehnhardt1998_133p} or due to a one-time event \citep{tothImpactgeneratedActivityPeriod2000}, subsequent apparitions showing activity indicated sublimation was the cause \citep{hsiehStrangeCase133P2004,hsiehReturnActivityMainbelt2010}. This example illustrates the importance of detecting additional activity epochs
Asteroid (248370)~2005~QN$_{173}${} is a 3.2$\pm$0.4~km diameter \citep{hsiehPhysicalCharacterizationMainBelt2021} outer main-belt asteroid ($a$=3.075~au, $e$=0.226, $i$=$0\fdg067$) that has a 5.37~yr orbit that ranges from a perihelion distance of $q$=2.374~au to an aphelion distance of $Q$=3.761~au. The object first drew particular attention when it was reported as active on 2021 July 9 \citep{fitzsimmons2483702005QN1732021}. Subsequently, Zwicky Transient Facility (ZTF) data were used to help constrain the activity onset to between 2020 July 10 and 2021 June 11 \citep{kelley2483702005QN2021}.
We set out to locate archival astronomical images of (248370)~2005~QN$_{173}${} in order to characterize prior activity. We made use of solar system object thumbnails (small image cutouts like Figure \ref{fig:wedgephot}) derived from publicly available archival data. We previously demonstrated how our data sources, such as the Dark Energy Camera (DECam), are well suited to discovering and characterizing active objects \citep{chandlerSAFARISearchingAsteroids2018,chandlerSixYearsSustained2019,chandlerCometaryActivityDiscovered2020a}.
Here we report activity of (248370)~2005~QN$_{173}${} on 2016 July 22 \citep{chandler2483702005QN2021}, an apparition prior to the 2021 outburst. We describe the process by which the activity was identified and examine the implications of this discovery.
\section{Second Activity Epoch}
\label{sec:secondActivityEpoch}
In order to find an additional activity epoch for (248370)~2005~QN$_{173}${}, we searched, assessed, and analyzed publicly available archival image data, building upon the methods of \cite{chandlerSAFARISearchingAsteroids2018,chandlerSixYearsSustained2019,chandlerCometaryActivityDiscovered2020a}.
\subsection{Data Acquisition}
\label{subsec:datamining}
To locate archival images of (248370)~2005~QN$_{173}${}, we queried our own database of publicly available observation metadata \citep[see][]{chandlerSAFARISearchingAsteroids2018}. This database, which updates daily, includes observing details such as sky coordinates, exposure time, and filter selection. Additionally, we searched Palomar Transient Factory (PTF) and ZTF data through 2021 August 31 through online search tools (listed in Appendix \ref{sec:equipQuickRef}) as well as a ZTF Alert Stream search and retrieval tool we created for this purpose. All instruments and data sources we made use of are listed in Appendix \ref{sec:equipQuickRef}, and we note that some data were found or retrieved via more than one pathway.
\begin{table*}
\centering
\footnotesize
\caption{(248370)~2005~QN$_{173}${} Observations}
\begin{tabular}{cclcrcccrrccl}
Image\footnote{Label in image gallery figures.} & Obs. Date\footnote{UT observing date in year-month-day format.} & Source & N\footnote{Number of images.} & Exp. [s]\footnote{Exposure time for each image.} & Filter & $V$\footnote{Apparent $V$-band magnitude (Horizons).} & r [au]\footnote{Heliocentric distance.} & STO [$\degr$]\footnote{Sun--target--observer angle.} & $\nu$ [$\degr$]\footnote{True anomaly.} & \%$_{Q\rightarrow q}$\footnote{Percentage to perihelion $q$ from aphelion $Q$.} & Act?\footnote{Activity observed.} & Archive\\
\hline
a & 2004-07-08 & MegaPrime & 3 & 180 & \textit{i } & 20.7 & 2.74 & 17.6 & 287.5 & 73\% & N & CADC,* \\
b & 2005-06-08 & SuprimeCam & 3 & 60 & \textit{W-J-VR} & 21.0 & 2.42 & 23.8 & 18.6 & 96\% & N & CADC,SMOKA \\
c & 2010-06-14 & Pan-STARRS1 & 2 & 30 & \textit{z } & 20.2 & 2.42 & 21.1 & 339.2 & 96\% & N & CADC \\
d & 2010-08-02 & Pan-STARRS1 & 1 & 45 & \textit{i } & 19.0 & 2.39 & 3.9 & 353.5 & 99\% & N & CADC \\
e & 2010-08-05 & Pan-STARRS1 & 1 & 40 & \textit{r } & 18.9 & 2.39 & 2.5 & 355.4 & 99\% & N & CADC \\
f & 2010-08-06 & Pan-STARRS1 & 1 & 43 & \textit{g } & 18.8 & 2.39 & 2.0 & 354.7 & 99\% & N & CADC \\
g & 2010-08-28 & PTF & 2 & 60 & \textit{r } & 19.2 & 2.39 & 8.4 & 1.2 & 99\% & N & IRSA/PTF \\
h & 2010-08-31 & Pan-STARRS1 & 2 & 45 & \textit{i } & 19.3 & 2.39 & 9.7 & 2.1 & 99\% & N & CADC \\
i & 2010-09-01 & PTF & 2 & 60 & \textit{r } & 19.3 & 2.39 & 10.1 & 2.3 & 99\% & N & IRSA/PTF \\
j & 2010-09-06 & Pan-STARRS1 & 2 & 43 & \textit{g } & 19.5 & 2.39 & 12.2 & 3.8 & 99\% & N & CADC \\
k & 2010-09-12 & Pan-STARRS1 & 2,2 & 43,40 & \textit{g,r } & 19.6 & 2.40 & 14.5 & 5.6 & 98\% & N & CADC \\
l & 2010-09-15 & PTF & 2 & 60 & \textit{r } & 19.7 & 2.39 & 15.5 & 6.5 & 99\% & N & IRSA/PTF \\
m & 2010-10-30 & Pan-STARRS1 & 2 & 30 & \textit{z } & 20.6 & 2.41 & 23.8 & 19.6 & 97\% & N & CADC \\
n & 2011-07-14 & Pan-STARRS1 & 1 & 40 & \textit{r } & 21.8 & 2.86 & 15.8 & 84.2 & 65\% & N & CADC \\
o & 2011-11-24 & Pan-STARRS1 & 2,2 & 43,40 & \textit{g,r } & 20.6 & 3.15 & 4.0 & 109.3 & 44\% & N & CADC \\
p & 2011-11-30 & Pan-STARRS1 & 2 & 45 & \textit{i } & 20.4 & 3.16 & 1.8 & 110.3 & 43\% & N & CADC \\
q & 2011-12-01 & Pan-STARRS1 & 2 & 43 & \textit{g } & 20.4 & 3.16 & 1.4 & 110.5 & 43\% & N & CADC \\
r & 2014-03-01 & OmegaCam & 5 & 360 & \textit{r } & 21.4 & 3.53 & 7.0 & 218.2 & 17\% & N & CADC,ESO \\
s & 2016-07-22 & DECam & 1 & 89 & \textit{z } & 21.2 & 2.59 & 22.7 & 56.5 & 84\% & Y & CADC,* \\
t & 2019-07-03 & DECam & 9 & 40 & \textit{VR } & 22.6 & 3.55 & 14.6 & 216.9 & 15\% & N & CADC,* \\
u & 2020-02-04 & DECam & 1 & 38 & \textit{r } & 21.6 & 3.16 & 18.1 & 248.9 & 43\% & N & CADC,* \\
v & 2020-02-10 & DECam & 1 & 199 & \textit{z } & 21.8 & 3.15 & 18.2 & 249.9 & 44\% & N & CADC,* \\
w & 2020-04-25 & ZTF & 1,1 & 30,30 & \textit{g,r } & 20.0 & 2.99 & 4.4 & 263.3 & 55\% & N & IRSA/ZTF \\
x & 2020-05-18 & ZTF & 1,1 & 30,30 & \textit{g,r } & 19.9 & 2.93 & 4.7 & 267.7 & 60\% & N & IRSA/ZTF \\
x & 2020-05-27 & ZTF & 1,1 & 30,30 & \textit{g,r } & 20.1 & 2.91 & 8.4 & 269.5 & 61\% & N & IRSA/ZTF \\
x & 2020-06-11 & ZTF & 1,3 & 30,30 & \textit{g,r } & 20.3 & 2.88 & 13.3 & 272.5 & 63\% & N & IRSA/ZTF \\
x & 2020-06-14 & ZTF & 1,1 & 30,30 & \textit{g,r } & 20.4 & 2.87 & 14.1 & 273.1 & 64\% & N & IRSA/ZTF \\
x & 2020-06-17 & ZTF & 2,2 & 30,30 & \textit{g,r } & 20.4 & 2.87 & 14.9 & 273.8 & 64\% & N & IRSA/ZTF \\
x & 2020-06-20 & ZTF & 2,2 & 30,30 & \textit{g,r } & 20.5 & 2.86 & 15.7 & 274.4 & 65\% & N & IRSA/ZTF \\
x & 2020-06-23 & ZTF & 2,1 & 30,30 & \textit{g,r } & 20.5 & 2.85 & 16.5 & 275.0 & 65\% & N & IRSA/ZTF \\
x & 2020-06-26 & ZTF & 4,1 & 30,30 & \textit{g,r } & 20.6 & 2.85 & 17.2 & 275.6 & 65\% & N & IRSA/ZTF\\
\hline
\end{tabular}
%
\footnotesize
\raggedright
Note \textit{W-J-VR} is a single wide-band filter. See Appendix \ref{sec:equipQuickRef} for image source and archive details. * indicates the data were obtained from our local repository. Table entries with multiple comma-separated values contain groups of exposures taken with different filters.
\label{tab:observations}
\end{table*}
We identified candidate images where (248370)~2005~QN$_{173}${} was expected to be within the field of view (FOV) based on observation times, pointing centers, and FOV sizes and orientations, downloaded associated data, and extracted image cutouts. We organized data by instrument and observation date and summarize observation details in Table \ref{tab:observations}.
\begin{figure*}
\centering
\begin{tabular}{ccccc}
\begin{overpic}[width=0.18\linewidth]{2005_QN173_2004-07-08_07.06.29_753395p_chip36_30arcsec_NuEl_3stacked_sum_81pix_overlaid.png}\put (5,7) {\huge\color{green} \textbf{a}}\end{overpic} &
\begin{overpic}[width=0.18\linewidth]{2005_QN173_205-06-08_ds9supa2stack_30arcsec_overlaid.png_1stacked_sum_155pix_overlaid.png}\put (5,7) {\huge\color{green} \textbf{b}}\end{overpic} &
\begin{overpic}[width=0.18\linewidth]{2005_QN173_2010-06-14_14.14.43.007000_rings.v3.skycell.1045.042.wrp.z.55361_59316_chip1_30arcsec_NuElnoArrows_overlaid.png_2stacked_mean_185pix_overlaid.png}\put (5,7) {\huge\color{green} \textbf{c}}\end{overpic} &
\begin{overpic}[width=0.18\linewidth]{2005_QN173_2010-08-02_10.38.34.112000_rings.v3.skycell.1045.028.wrp.i.55410_44306_chip1_30arcsec_NuEl.png}\put (5,7) {\huge\color{green} \textbf{d}}\end{overpic} &
\begin{overpic}[width=0.18\linewidth]{2005_QN173_2010-08-05_11.41.44.038000_rings.v3.skycell.1044.020.wrp.r.55413_48692_chip1_30arcsec_NuEl.png}\put (5,7) {\huge\color{green} \textbf{e}}\end{overpic} \\
\begin{overpic}[width=0.18\linewidth]{2005_QN173_2010-08-06_11.20.25.477000_rings.v3.skycell.1044.010.wrp.g.55414_47212_chip1_30arcsec_NuEl.png}\put (5,7) {\huge\color{green} \textbf{f}}\end{overpic} &
\begin{overpic}[width=0.18\linewidth]{2005_QN173_2010-08-28_03.51.32.938000_PTF_201008281608_i_p_scie_t035132_u011209292_f02_p001955_c05_chip0_30arcsec_NuEl_2stacked_sum_15pix_overlaid.png} \put (5,7) {\huge\color{green} \textbf{g}}\end{overpic} &
\begin{overpic}[width=0.18\linewidth]{2005_QN173_2010-08-31_09.43.32.218000_rings.v3.skycell.0956.089.wrp.i.55439_40484_chip1_30arcsec_NuElnoArrows_overlaid.png_2stacked_mean_185pix_overlaid.png}\put (5,7) {\huge\color{green} \textbf{h}}\end{overpic} &
\begin{overpic}[width=0.18\linewidth]{2005_QN173_2010-09-01_03.49.09.789000_PTF_201009011591_i_p_scie_t034909_u010919697_f02_p001955_c04_chip0_30arcsec_NuEl_2stacked_sum_15pix_overlaid.png} \put (5,7) {\huge\color{green} \textbf{i}}\end{overpic} &
\begin{overpic}[width=0.18\linewidth]{2005_QN173_2010-09-06_07.20.23.785000_rings.v3.skycell.0955.080.wrp.g.55445_30543_chip1_126arcsec_NuElnoArrows_overlaid.png_4stacked_mean_185pix_overlaid.png} \put (5,7) {\huge\color{green} \textbf{j}}\end{overpic}\\
\begin{overpic}[width=0.18\linewidth]{2005_QN173_2010-09-12_06.35.22.823000_rings.v3.skycell.0955.071.wrp.r.55451_27417_chip1_126arcsec_NuElnoArrows_overlaid.png_8stacked_mean_185pix_overlaid.png} \put (5,7) {\huge\color{green} \textbf{k}}\end{overpic} &
\begin{overpic}[width=0.18\linewidth]{2005_QN173_2010-09-15_03.43.16.985000_PTF_201009151550_i_p_scie_t034316_u011020053_f02_p001955_c08_chip0_30arcsec_NuEl_2stacked_sum_15pix_overlaid.png} \put (5,7) {\huge\color{green} \textbf{l}}\end{overpic} &
\begin{overpic}[width=0.18\linewidth]{2005_QN173_2010-10-30_04.51.47.177000_rings.v3.skycell.1044.013.wrp.z.55499_20223_chip1_30arcsec_NuElnoArrows_overlaid.png_2stacked_mean_185pix_overlaid.png} \put (5,7) {\huge\color{green} \textbf{m}}\end{overpic} &
\begin{overpic}[width=0.18\linewidth]{2005_QN173_2011-09-07_14.26.00.251000_rings.v3.skycell.1782.069.wrp.r.55811_60100_chip1_30arcsec_NuEl.png} \put (5,7) {\huge\color{green} \textbf{n}}\end{overpic} &
\begin{overpic}[width=0.18\linewidth]{2005_QN173_2011-11-24_12.38.14.010000_rings.v3.skycell.1781.063.wrp.g.55889_52616_chip1_30arcsec_NuElnoArrows_overlaid.png_4stacked_mean_185pix_overlaid.png}\put (5,7) {\huge\color{green} \textbf{o}}\end{overpic} \\
\begin{overpic}[width=0.18\linewidth]{2005_QN173_2011-11-30_07.30.59.895000_rings.v3.skycell.1781.056.wrp.i.55895_31279_chip1_126arcsec_NuElnoArrows_overlaid.png_4stacked_mean_185pix_overlaid.png}\put (5,7) {\huge\color{green} \textbf{p}}\end{overpic} &
\begin{overpic}[width=0.18\linewidth]{2005_QN173_2011-12-01_11.39.11.507000_rings.v3.skycell.1781.057.wrp.g.55896_48515_chip1_30arcsec_NuElnoArrows_overlaid.png_2stacked_mean_185pix_overlaid.png}\put (5,7) {\huge\color{green} \textbf{q}}\end{overpic} &
\begin{overpic}[width=0.18\linewidth]{2005_QN173_2014-03-01_06.13.46.527000_OMEGA.2014-03-01T06.13.46.527_chip25-ESO_CCD_85_30arcsec_NuEl_19.4dRA_-7.7dDec_5stacked_sum_70pix_overlaid.png}\put (5,7) {\huge\color{green} \textbf{r}}\end{overpic} &
\begin{overpic}[width=0.18\linewidth]{2005_QN173_2016-07-22_10.20.48.181064_c4d_160722_102306_ooi_z_ls9_chip8-S21_30arcsec_NuEl.png}\put (5,7) {\huge\color{green} \textbf{s}}\end{overpic} &
\begin{overpic}[width=0.18\linewidth]{2005_QN173_2019-07-03_22.49.56.808619_c4d_190703_224956_opi_VR_v1_chip58-N27_30arcsec_NuEl_9stacked_mean_57pix_11stacked_sum_57pix_overlaid.png}\put (5,7) {\huge\color{green} \textbf{t}}\end{overpic} \\
\begin{overpic}[width=0.18\linewidth]{2005_QN173_2020-02-04_08.30.31.595447_c4d_200204_083031_opi_r_v1_chip17-S18_30arcsec_NuEl.png}\put (5,7) {\huge\color{green} \textbf{u}}\end{overpic} &
\begin{overpic}[width=0.18\linewidth]{2005_QN173_2020-02-10_08.40.48.950389_c4d_200210_084048_opi_z_v1_chip35-N4_30arcsec_NuEl.png}\put (5,7) {\huge\color{green} \textbf{v}}\end{overpic} &
\begin{overpic}[width=0.18\linewidth]{2005_QN173_2020-04-25_08.07.31.134000_ztf_20200425338542_000326_zg_c07_o_q2_sciimg_chip0-2_30arcsec_NuEl_2stacked_sum_15pix_overlaid.png}\put (5,7) {\huge\color{green} \textbf{w}}\end{overpic} &
\begin{overpic}[width=0.18\linewidth]{2005_QN173_2020-05-18_05.59.05.705000_ztf_20200518249363_000325_zg_c10_o_q2_sciimg_chip0-2_30arcsec_NuEl_21stacked_sum_15pix_overlaid.png}\put (5,7) {\huge\color{green} \textbf{x}}\end{overpic} \\
\end{tabular}
\caption{Archival images of (248370)~2005~QN$_{173}${} with the best activity detection potential (i.e., sufficient depth and observing conditions) for (248370)~2005~QN$_{173}${}. For all images, north is up, east is left, the FOV is 30\arcsec$\times$30\arcsec, the antisolar (yellow -$\odot$) and antimotion (red -$v$) directions are shown with the origin at image center. See Appendix \ref{sec:equipQuickRef} for instrument and archive details. Panel (s) is from the only thumbnail in which we could identify activity unambiguously (Figure \ref{fig:wedgephot}). (a) 2004 July 8 MegaPrime 3$\times$180~s $i$ band. (b) 2005 June 8 SuprimeCam 3$\times$60~s \textit{W-J-VR} band. (c) 2010 June 14 PS1 2$\times$60~s $z$ band. (d) 2010 August 2 PS1 1$\times$45~s $i$ band. (e) 2010 August 5 PS1 1$\times$40~s $r$ band. (f) 2010 August 6 PS1 1$\times$43~s $g$ band. (g) 2010 August 28 PTF 2$\times$60~s $r$ band. (h) 2010 August 31 PS1 $\times$45~s $i$ band. (i) 2010 September 1 PTF 2$\times$60~s $r$ band. (j) 2010 September 6 PS1 2$\times$43~s $g$ band. (k) 2010 September 12 PS1 2$\times$40~s $r$ band + 2$\times$43~s $g$ band. (l) 2010 September 15 PTF 2$\times$60~s $r$ band. (m) 2010 October 30 PS1 2$\times$30~s $z$ band. (n) 2011 July 14 PS1 1$\times$40~s $r$ band. (o) 2011 November 24 PS1 2$\times$40~s $r$ band + 2$\times$43~s $g$ band. (p) 2011 November 30 PS1 2$\times$45~s $i$ band. (q) 2011 December 1 PS1 2$\times$43~s $g$ band. (r) 2014 January 31 OmegaCAM 5$\times$360~s $r$ band. (s) 2016 July 22 1$\times$89~s \textit{z} band. (t) 2019 July 3 DECam 9$\times$40~s \textit{VR} band. (u) 2020 February 4 DECam 1$\times$38~s \textit{r} band. (v) 2020 February 10 DECam 1$\times$199~s $z$ band. (w) 2020 April 25 ZTF 1$\times$30~s \textit{g} band + 1$\times$30~s \textit{r} band. (x) 2020 May 18, 27 + 2020 June 11, 14, 17, 20, 23, \& 26 ZTF 9$\times$30~s \textit{r} band + 12$\times$30~s \textit{g} band.}
\label{fig:gallery}
\end{figure*}
We extracted eighty-one{} $480\times480$ pixel thumbnail images (such as the image shown in Figure \ref{fig:wedgephot}) in which we could confidently identify (248370)~2005~QN$_{173}${} (Table \ref{tab:observations}). We coadded images from the same instrument when observations were close enough in time for computed tail orientation to be in close agreement such that coaddition could enhance activity, if present. The thumbnails with the ``best activity detection potential'' -- meaning the images were judged to have observing conditions (e.g., seeing) and depth (i.e., magnitude limit) amenable to activity detection -- are shown in Figure \ref{fig:gallery}.
To allow for uniform spatial comparisons and to magnify the region of interest around (248370)~2005~QN$_{173}${}, all thumbnail images in Figure \ref{fig:gallery}
are displayed with 30\arcsec$\times$30\arcsec fields of view.
\subsection{Image Assessment}
\label{subsec:imageassessment}
We vetted each thumbnail by visually confirming (248370)~2005~QN$_{173}${} was visible. In cases where the object could not be readily identified, we employed our pipeline to produce comparison thumbnails derived from DECam data that showed the same region of sky, instrument, broadband filter, and exposure time, but from epochs when the object was not in the FOV. We made use of Gaia DR2 \citep{gaiacollaborationGaiaDataRelease2018} and Sloan Digital Sky Survey Release 9 (SDSS~DR-9) catalogs \citep{ahnNinthDataRelease2012} to visually validate World Coordinate System of images within the SAOImageDS9 Vizier \citep{ochsenbeinVizieRDatabaseAstronomical2000} catalog query system.
We next identified vetted thumbnails that were suitable for coaddition by clustering images based on instrument and date. For compatible image sets that included multiple broadband filters, we carried out coaddition among matching filters as well as combining all images. Finally, we visually examined the results and flagged images with potential activity.
We found a single image with clear evidence of activity (Figure \ref{fig:wedgephot}) in an 89~s $z$-band exposure captured 2016 July 22 by Dustin Lang and Alistair Walker as part of the DECam Legacy Survey \citep[DECaLS;][]{deyOverviewDESILegacy2019}. This discovery makes (248370)~2005~QN$_{173}${} the ninth recurrently active main-belt asteroid to be identified to date. The other objects, 238P/Read, 259P/Garradd, 288P, 311P/PANSTARRS, 313P/Gibbs, 324P/La Sagra, (6478)~Gault, and (7968)~Elst-Pizarro, have all demonstrated recurrent activity near perihelion, with the exception of (6478)~Gault \citep{chandlerSixYearsSustained2019}.
We measured the tail length to be about $2\farcm14$ ($2.4\times10^5$~km) in this image but a longer tail may well have been revealed with a longer exposure (see \citet{hsiehPhysicalCharacterizationMainBelt2021} for 2021 apparition tail measurement). Applying our wedge photometry technique (Section \ref{subsec:wedgephotometry}), we produced a diagnostic plot (Figure \ref{fig:wedgephot}) and measured a position angle on the sky of $251\fdg3\pm1\fdg4$ for the tail, in close agreement with the Horizons computed $251\fdg6$ antisolar and $251\fdg7$ antimotion vectors.
\subsection{Wedge Photometry Tail Tool}
\label{subsec:wedgephotometry}
We crafted a new algorithm to (a) identify potentially active objects by detecting likely tails, and (b) quantify alignment between an observed tail and predicted antisolar and antimotion vectors, which are commonly associated with tail direction. Currently, the tool is designed to analyze single tails $<15\degr$ in angular extent, though we plan to address multiple tails and comae in the future. The technique, which we refer to as \textit{wedge photometry}, sums all pixel values in a variable-width wedge bound between an inner and outer radius and identifies wedges containing excess flux relative to other wedges, if present. A similar approach was used in \cite{2011Icar..215..534S} but we have made improvements in angular resolution and algorithmic efficiency. Excess flux within a particular wedge may indicate the presence of a tail, and testing tail alignment with antisolar and antimotion vectors provides additional weight that a detected tail is real. Here we focused on quantifying tail orientation and position angle agreement.
To optimize the process, we convert Cartesian pixel coordinates ($x$,$y$) to polar coordinates ($r$,$\theta$) where the central thumbnail pixel is defined as (0,0). The resulting three-dimensional array has columns $r$, $\theta$, and $c$ (counts).
For a series of wedge sizes $\theta$ ($1\degr$--$10\degr$ in $1\degr$ increments) we summed pixel values in annular segments spanning an angle $\pm\theta/2$ along a radial component $r$ from an inner bound, $r_0=5$ pixels, to a maximum of $r_\mathrm{max}=50$ pixels (13\arcsec for our DECam data), as given by
\begin{equation}
c(\theta, \Delta\theta) = \sum_{\theta=-\Delta\theta/2}^{\theta=+\Delta\theta/2} \sum_{r=r_0}^{r=r_\mathrm{max}} c(r, \theta).
\end{equation}
\noindent We further optimize the procedure by selecting for the target a starting radius $r_0$ outside the FWHM, and choosing a maximum radius $_\mathrm{max}$ that allows for a wedge length long enough to ensure that all bins have sufficient counts to avoid necessitating resampling of any individual pixels. Thus, pixels are assigned to wedges based solely on their precise pixel center coordinate, and any fractional flux from a pixel that spans a wedge boundary is automatically assigned to the wedge containing the pixel center coordinate. We then compute the mean and standard deviation of the resulting counts for each $\theta$ to compare with the predicted antisolar and antimotion vectors.
We produce a polar plot (Figure \ref{fig:wedgephot}) to aid assessing relative radial flux distribution. Most position angles have a $1\sigma$ of $\sim$200 counts. The tail is clearly identified by our algorithm at $> 7 \sigma$ for several $\Delta \theta$ wedge sizes.
\section{Main-belt Comet Classification}
\label{sec:mainBeltComet}
Once we had identified a previous activity epoch we set out to determine if (248370)~2005~QN$_{173}${} could be an MBC.
\subsection{Prerequisites}
\label{subsec:mbc}
For (248370)~2005~QN$_{173}${} to qualify as an MBC it must (1) be an active asteroid, (2) orbit within the Main Asteroid Belt, and (3) exhibit sublimation driven activity.
(1) To qualify as an active asteroid, a body must typically meet three criteria (see \citealt{jewittActiveAsteroids2012} for discussion): (i) A coma or tail must have been observed visually (as is the case in this work) or, potentially, through alternate means such as spectroscopy \citep[e.g.,][]{kuppersLocalizedSourcesWater2014,busarevNewCandidatesActive2018} or detecting magnetic field enhancements \citep[e.g.,][]{russellInterplanetaryMagneticField1984}. (ii) The semi-major axis $a$ must not be exterior to that of Jupiter ($a_\mathrm{J}\approx$ 5.2~au) as is the case for comets and active Centaurs \citep{jewittActiveCentaurs2009}; (248370)~2005~QN$_{173}${} has $a=$3.1~au. And (iii) the Tisserand parameter with respect to Jupiter, $T_\mathrm{J}$, must be greater than 3; this is because objects with $T_\mathrm{J}<3$ are canonically considered cometary and $T_\mathrm{J}>3$ asteroids \citep{vaghiOriginJupiterFamily1973,vaghiOrbitalEvolutionComets1973}.
$T_\mathrm{J}$ describes how an orbit is related to Jupiter by
\begin{equation}
T_\mathrm{J} = \frac{a_\mathrm{J}}{a} + 2 \sqrt{\frac{a\left(1-e^2\right)}{a_\mathrm{J}}}\cos\left(i\right).
\end{equation}
\noindent where $e$ is the eccentricity and $i$ is the orbital inclination. $T_\mathrm{J}$ for (248370)~2005~QN$_{173}${} is 3.192 and thus it qualifies as asteroidal. (248370)~2005~QN$_{173}${} properties are provided in Appendix \ref{sec:ObjectData}, and are established with this criterion.
(2) (248370)~2005~QN$_{173}${} orbits between 2.4~au and 3.76~au and thus does not cross the orbits of either Mars or Jupiter. With a semi-major axis of 3.1~au, (248370)~2005~QN$_{173}${} is an outer main-belt asteroid orbiting between the Kirkwood gaps corresponding to the 7:8 and 2:1 mean motion resonances with Jupiter.
(3) Recurrent activity near perihelion is diagnostic of volatile sublimation as the most likely mechanism responsible for the observed activity \citep[e.g.,][]{hsiehOpticalDynamicalCharacterization2012}. However, other underlying causes of recurrent activity are known, so this point warrants further investigation.
\subsection{Activity Mechanism}
\label{subsec:mechanism}
We demonstrated in Section \ref{sec:secondActivityEpoch} that (248370)~2005~QN$_{173}${} has been active during at least two epochs. This helps rule out activity mechanisms such as \textit{impact events} (e.g., (596)~Scheila; \citealt{bodewitsCollisionalExcavationAsteroid2011,ishiguroObservationalEvidenceImpact2011,moreno596ScheilaOutburst2011}) that are only expected to produce one-time outbursts but which can expel dust and produce comet-like activity. Aside from \textit{temperature-correlated volatile sublimation} (which we examine further in Section~\ref{subsec:tempestimation}) other mechanisms for producing recurrent activity have been proposed.
\textit{Rotational destabilization} causes dust to be flung from a body in a potentially multiepisodic manner, as may be the case for (6478)~Gault \citep{chandlerSixYearsSustained2019,kleynaSporadicActivity64782019}. Taxonomic classification can help diagnose rotational destabilization, as with $S$-type (6478)~Gault, because activity from desiccated asteroid classes is unlikely to be sublimation driven. As discussed in Section \ref{subsec:tempestimation}, the taxonomic class of (248370)~2005~QN$_{173}${} is not yet known but it is likely a C-type asteroid. An accurate rotation period for (248370)~2005~QN$_{173}${} is currently unavailable, and as such, we can neither confirm nor rule out destabilization as a contributing factor to the observed activity at this time.
\textit{Rubbing binaries} is a hypothetical scenario whereby two merging asteroids repeatedly collide and eject material. Proposed as a possible mechanism for the activity of 311P/PANSTARRS \citep{hainautContinuedActivity20132014}, the rubbing binary scenario has yet to be confirmed for that object \citep{jewittNucleusActiveAsteroid2018} or any other. As of this writing, there is no evidence that (248370)~2005~QN$_{173}${} is a binary asteroid, and activity spans two epochs separated by 5 yr, so we would expect merging processes to have either finished or that the binary orbit would have stabilized (see \citealt{jewittNucleusActiveAsteroid2018} for additional discussion concerning dissipation timescales). Therefore we find it unlikely that rubbing causes the activity associated with (248370)~2005~QN$_{173}${}.
Geminid meteor stream parent (3200)~Phaethon undergoes extreme temperature changes ($\sim$600~K) and peaks at 800~K to 1100~K, well above the 573~K serpentine-phyllosilicate decomposition threshold \citep{ohtsukaSolarRadiationHeatingEffects2009}. These temperatures likely induce \textit{thermal fracture} \citep{licandroNatureCometasteroidTransition2007,kasugaObservations1999YC2008} leading to mass shedding \citep{liRecurrentPerihelionActivity2013,huiResurrection3200Phaethon2017}.
Two mechanisms, thermal fracture and temperature-correlated volatile sublimation warrant further inquiry into the thermophysical properties of (248370)~2005~QN$_{173}${}.
\subsection{Temperature Estimation}
\label{subsec:tempestimation}
Estimating temperatures experienced by (248370)~2005~QN$_{173}${} aids us in understanding direct thermal effects (e.g., thermal fracture) as well as assessing long-term volatile survival, especially water. For these reasons, we computed temperatures for an airless body over the course of an orbit similar to that of (248370)~2005~QN$_{173}${}.
Following \citet{hsiehMainbeltCometsPanSTARRS12015}, the energy balance equation for a gray body on which water ice sublimation is occurring is
\begin{equation}
{F_{\odot}\over r_h^2}(1-A) = \chi\left[{\varepsilon\sigma T_{eq}^4 + L f_D\dot m_{w}(T_{eq})}\right]
\label{equation:sublim1}
\end{equation}
where $r_h$ is the object's heliocentric distance, $T_\mathrm{eq}$ is the equilibrium surface temperature, $F_{\odot}=1360$~W~m$^{-2}$ is the solar constant, $r_h$ is in au, $A=0.05$ is the assumed Bond albedo of the body, $\chi$ accounts for the distribution of solar heating over the object's surface, $\sigma$ is the Stefan--Boltzmann constant, and $\varepsilon=0.9$ is the assumed effective infrared emissivity, and $L=2.83$~MJ~kg$^{-1}$ is the latent heat of sublimation of water ice (which we approximate here as being independent of temperature), $f_D$ represents the reduction in sublimation efficiency caused by mantling, where $f_D=1$ in the absence of a mantle, and $\dot m_w$ is the water mass-loss rate due to sublimation of surface ice.
In this equation, $\chi=1$ corresponds to a flat slab facing the Sun, known as the subsolar approximation, and produces the maximum expected temperature for an object,
while $\chi=4$ applies to objects with fast rotation or low thermal inertia, known as the isothermal approximation, and produces the minimum expected temperature for an object.
Next, the sublimation rate of ice into a vacuum can be computed using
\begin{equation}
\dot m_{w} = P_v(T) \sqrt{\mu\over2\pi k T}
\label{equation:sublim2}
\end{equation}
where $\mu=2.991\cdot 10^{-26}$~kg is the mass of one water molecule, and $k$ is the Boltzmann constant, and the equivalent ice recession rate, $\dot \ell_{i}$, corresponding to $\dot m_{w}$ is given by
$\dot \ell_{i} = \dot m_{w}/ \rho$,
where $\rho$ is the bulk density of the object.
Finally, the Clausius--Clapeyron relation,
\begin{equation}
P_v(T) = 611 \times \exp\left[{{\Delta H_\mathrm{subl}\over R_g}\left({{1\over 273.16} - {1\over T}}\right)}\right]
\label{equation:sublim3}
\end{equation}
gives the vapor pressure of water, $P_v(T)$, in Pa, where $\Delta H_\mathrm{subl}=51.06$~MJ~kmol$^{-1}$ is the heat of sublimation for ice to vapor and $R_g=8314$~J~kmol$^{-1}$~K$^{-1}$ is the ideal gas constant.
Solving these three equations iteratively, one can calculate the equilibrium temperature of a gray body at a given heliocentric distance on which water ice sublimation is occurring.
In Figure~\ref{fig:ActivityTimeline}, we plot the object's expected equilibrium temperature over several orbit cycles, as computed by solving the system of equations above. We plot temperatures computed using both $\chi=1$ and $\chi=4$ to show the full range of possible temperatures.
\begin{figure*}[ht]
\centering
\begin{tabular}{c}
\hspace{-4mm}\includegraphics[width=0.85\linewidth]{Activity_dOnly_noYears.pdf}\\
\hspace{9mm}\includegraphics[width=0.92\linewidth]{248370_2005_QN173_at807_noDates.pdf}\\
\hspace{-4mm}\includegraphics[width=0.845\linewidth]{Activity_yearMarkersOnly.pdf}\\
\hspace{-1mm}\includegraphics[width=0.845\linewidth]{newTforChis.pdf}
\end{tabular}
\caption{(248370)~2005~QN$_{173}${} heliocentric distance (top plot), observability timeline (middle plot) and temperature (bottom plot), beginning the year of our first archival data (2004) through 2022.
Top: triangles represent positive (filled red) and negative (unfilled blue) activity detections. Markers indicate when the object was inbound (downward pointing triangles) or outbound (upward pointing triangles). Table \ref{tab:observations} lists observation details.
Middle: apparent $V$-band magnitude (solid green line) and our ``observability'' metric (yellow dashed line) that represents hours during a given UT observing date the object is above $> 15\degr$ elevation while the Sun is below the horizon. Peaks in apparent magnitude coinciding with observability occur during opposition events, and observability troughs indicate solar conjunctions when (248370)~2005~QN$_{173}${} was only above the horizon during daylight. Perihelion (orange $q$) and aphelion (blue $Q$) events are indicated.
Bottom: temperature $T$~(K) by date for different $\chi$ values, where $\chi=1$ (top line) represents a ``flat slab'' and $\chi=4$ (bottom line) an isothermal body.
}
\label{fig:ActivityTimeline}
\end{figure*}
Figure~\ref{fig:ActivityTimeline} (top) shows two parameters key to observing (248370)~2005~QN$_{173}${} between 2015 and 2023: apparent $V$-band magnitude and ``observability,'' which we define as the number of hours an object remains above $15\degr$ elevation during nighttime for a given UT observing date. This plot informs us of potential observational biases or geometric effects that may bias activity detection, such as preferential activity discovery during opposition events, as was the case with (6478)~Gault \citep{chandlerSixYearsSustained2019}.
Figure~\ref{fig:ActivityTimeline} (bottom) illustrates (248370)~2005~QN$_{173}${}'s heliocentric distance and temperature (as computed using Equations \ref{equation:sublim1}-\ref{equation:sublim3}) over time, plus dates of observed activity and images where no visible activity was conclusively identified. Throughout its entire orbit the surface of (248370)~2005~QN$_{173}${} is consistently warmer than 145~K, the temperature above which water ice is not expected to survive over timescales on the order of the age of the solar system \citep{schorghoferLifetimeIceMain2008,snodgrassMainBeltComets2017}.
However, it is possible for water ice to remain preserved over long (Gyr) timescales on small asteroids at depths as shallow as a few centimeters to 30~cm below the surface \citep{schorghoferLifetimeIceMain2008,prialnikCanIceSurvive2009}, where present-day activity may be triggered by meter-scale impactors that excavate subsurface ice. We note that water ice has been detected on the surface of main-belt asteroid (24) Themis \citep{campinsWaterIceOrganics2010,rivkinDetectionIceOrganics2010}, but the mechanism by which that water ice is able to persist on its surface -- likely requiring continual replenishment from subsurface volatile reservoirs -- is not well understood, and furthermore may not have the same effectiveness on kilometer-scale objects like (248370)~2005~QN$_{173}${} as it does on the 200~km diameter (24) Themis.
The surface temperature of (248370)~2005~QN$_{173}${} varies at most between 145~K and 190~K over its orbit (Figure \ref{fig:ActivityTimeline}), far less than the 600~K temperature swings peaking at 800--1000~K described in Section \ref{subsec:mechanism}. We consider it is unlikely that thermal fracture is the primary cause of (248370)~2005~QN$_{173}${}'s activity.
\subsection{Nondetection of Activity}
\label{subsec:nullresults}
The two known epochs of activity for (248370)~2005~QN$_{173}${} both occurred when the object was interior to a heliocentric distance of 2.7~au. However, (248370)~2005~QN$_{173}${} was observed in 2005 and 2010 when the object was interior to 2.7~au but no activity was detected. We believe the circumstances of these epochs preclude a definitive assessment of activity.
The 2010 Pan-STARRS1 data suffer from image artifacts that are significant enough to obscure activity that may have been present. The 2005 SuprimeCam observations should have been well suited to detecting activity as the 8~m Subaru telescope has a large aperture, exposure times (60~s) were sufficiently long, the \textit{W-J-VR} filter covered a broad wavelength range, and the object was well placed in the sky in terms of airmass/elevation during the observations. However, extinction varied significantly over the observing sequence as the summit log for that night\footnote{\url{https://smoka.nao.ac.jp/calendar/slog/2005/slog_20050608.txt}} indicated that conditions were windy with cirrus clouds and the differential image motion monitor measured significant seeing variation (roughly $0\farcs8$ to $>3\arcsec{}$)\footnote{\url{https://smoka.nao.ac.jp/calendar/subaruseeing/20050608.gif}}, potentially contributing to a considerable reduction in our ability to detect activity. All sources in the field could be matched to SDSS~DR-9 stars but the faintest stars we were able match to the SDSS~DR-9 catalog were $r\approx21.3$, very similar to the JPL Horizons computed $V$=21.0 for (248370)~2005~QN$_{173}${}. The Subaru Exposure Time Calculator estimates an equivalent $r$-band exposure would deliver a signal-to-noise ratio of 66 for a source of equivalent magnitude, but we estimate the images are at best $\sim$0.5 mag deeper than necessary to detect the object and thus we find it unlikely that activity would be detectable unless the object was undergoing a significant outburst at the time.
Although we cannot definitively rule out the presence of activity in 2005 or 2010 from these observations, another possibility is that a triggering event (e.g., impact, rotational destabilization) that started (248370)~2005~QN$_{173}${}'s current activity occurred between 2005 June and 2016 July. This would explain the object's apparent inactivity and activity on each of those dates, respectively.
\subsection{Main-belt Comet Membership}
\label{subsec:mbcmembership}
Given the above reasoning, we find it most likely that the activity associated with (248370)~2005~QN$_{173}${} is sublimation driven, in which case the object is an MBC. However, we emphasize that in order to rule out rotational destabilization as the root cause of the observed activity, additional work is needed. Moreover, confirmation of a third activity epoch would lend further evidence favoring sublimation as the primary agency of activity.
\section{Summary and Future Work}
\label{sec:summary}
We harvested eighty-one{} images of (248370)~2005~QN$_{173}${} (also designated 433P) spanning thirty-one{} observing dates. We found clear evidence of a previous activity epoch on 2016 July 22. We provide a catalog of archival observations along with an image gallery. Making use of wedge photometry -- a novel tail detection and characterization tool we introduce in this Letter -- we measure tail orientation to be in close agreement with the antisolar and antimotion vectors computed by Horizons. We showed that (248370)~2005~QN$_{173}${} is a likely member of the MBCs, a group of active asteroids orbiting within the Main Asteroid Belt that are active due to volatile sublimation.
The current observing window for this object ends around 2021 December, and when it returns in late 2022 it will be over 3~au from the Sun and less likely to show activity. We did not find any images showing (248370)~2005~QN$_{173}${} active at beyond 3~au, so we call on observers to make use of the present activity apparition while it is still possible to do so. Continued monitoring to study the evolution of the tail's brightness, including surface brightness measurements, can lead to better characterization of ejected dust grain sizes and total mass loss during this apparition. Once activity subsides, time-series observations to measure a rotation period will be especially useful for diagnosing rotational breakup. Preliminary color measurements suggest (248370)~2005~QN$_{173}${} is a C-type asteroid \citep{hsiehPhysicalCharacterizationMainBelt2021}, but a robust taxonomic classification would help further solidify our assessment of the underlying activity mechanism.
\section{Acknowledgements}
\label{sec:acknowledgements}
The authors thank the anonymous referee whose comments greatly improved the quality of this Letter.
We thank Dr.\ Mark Jesus Mendoza Magbanua (University of California San Francisco) for his frequent and timely feedback on the project.
The authors express their gratitude to
Prof. Mike Gowanlock (NAU),
Dr. Annika Gustafsson (NAU, Lowell Observatory, Southwest Research Institute),
Jay Kueny (Steward Observatory),
and the Trilling Research Group (NAU), all of whom provided invaluable insights which substantially enhanced this work. The unparalleled support provided by Monsoon cluster administrator Christopher Coffey (NAU) and the High Performance Computing Support team facilitated the scientific process.
This material is based upon work supported by the National Science Foundation Graduate Research Fellowship Program under grant No.\ (2018258765). Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the National Science Foundation. C.O.C., H.H.H. and C.A.T. also acknowledge support from the NASA Solar System Observations program (grant 80NSSC19K0869).
Computational analyses were run on Northern Arizona University's Monsoon computing cluster, funded by Arizona's Technology and Research Initiative Fund. This work was made possible in part through the State of Arizona Technology and Research Initiative Program.
World Coordinate System (WCS) corrections facilitated by the \textit{Astrometry.net} software suite \citep{langAstrometryNetBlind2010}.
This research has made use of data and/or services provided by the International Astronomical Union's Minor Planet Center.
This research has made use of NASA's Astrophysics Data System.
This research has made use of The Institut de M\'ecanique C\'eleste et de Calcul des \'Eph\'em\'erides (IMCCE) SkyBoT Virtual Observatory tool \citep{berthierSkyBoTNewVO2006}.
This work made use of the {FTOOLS} software package hosted by the NASA Goddard Flight Center High Energy Astrophysics Science Archive Research Center.
This research has made use of SAOImageDS9, developed by Smithsonian Astrophysical Observatory \citep{joyeNewFeaturesSAOImage2006}.
This work made use of the Lowell Observatory Asteroid Orbit Database \textit{astorbDB} \citep{bowellPublicDomainAsteroid1994,moskovitzAstorbDatabaseLowell2021}.
This work made use of the \textit{astropy} software package \citep{robitailleAstropyCommunityPython2013}.
This project used data obtained with the Dark Energy Camera (DECam), which was constructed by the Dark Energy Survey (DES) collaboration. Funding for the DES Projects has been provided by the U.S. Department of Energy, the U.S. National Science Foundation, the Ministry of Science and Education of Spain, the Science and Technology Facilities Council of the United Kingdom, the Higher Education Funding Council for England, the National Center for Supercomputing Applications at the University of Illinois at Urbana-Champaign, the Kavli Institute of Cosmological Physics at the University of Chicago, Center for Cosmology and Astro-Particle Physics at the Ohio State University, the Mitchell Institute for Fundamental Physics and Astronomy at Texas A\&M University, Financiadora de Estudos e Projetos, Funda\c{c}\~{a}o Carlos Chagas Filho de Amparo, Financiadora de Estudos e Projetos, Funda\c{c}\~ao Carlos Chagas Filho de Amparo \`{a} Pesquisa do Estado do Rio de Janeiro, Conselho Nacional de Desenvolvimento Cient\'{i}fico e Tecnol\'{o}gico and the Minist\'{e}rio da Ci\^{e}ncia, Tecnologia e Inova\c{c}\~{a}o, the Deutsche Forschungsgemeinschaft and the Collaborating Institutions in the Dark Energy Survey. The Collaborating Institutions are Argonne National Laboratory, the University of California at Santa Cruz, the University of Cambridge, Centro de Investigaciones En\'{e}rgeticas, Medioambientales y Tecnol\'{o}gicas–Madrid, the University of Chicago, University College London, the DES-Brazil Consortium, the University of Edinburgh, the Eidgen\"ossische Technische Hochschule (ETH) Z\"urich, Fermi National Accelerator Laboratory, the University of Illinois at Urbana-Champaign, the Institut de Ci\`{e}ncies de l'Espai (IEEC/CSIC), the Institut de Física d'Altes Energies, Lawrence Berkeley National Laboratory, the Ludwig-Maximilians Universit\"{a}t M\"{u}nchen and the associated Excellence Cluster Universe, the University of Michigan, the National Optical Astronomy Observatory, the University of Nottingham, the Ohio State University, the University of Pennsylvania, the University of Portsmouth, SLAC National Accelerator Laboratory, Stanford University, the University of Sussex, and Texas A\&M University.
Based on observations at Cerro Tololo Inter-American Observatory, National Optical Astronomy Observatory (NOAO Prop. ID 2016A-0190, PI: Dey), which is operated by the Association of Universities for Research in Astronomy (AURA) under a cooperative agreement with the National Science Foundation. This research has made use of the NASA/IPAC Infrared Science Archive, which is funded by the National Aeronautics and Space Administration and operated by the California Institute of Technology.
The Legacy Surveys consist of three individual and complementary projects: the Dark Energy Camera Legacy Survey (DECaLS; Proposal ID \#2014B-0404; PIs: David Schlegel and Arjun Dey), the Beijing-Arizona Sky Survey (BASS; NOAO Prop. ID \#2015A-0801; PIs: Zhou Xu and Xiaohui Fan), and the Mayall z-band Legacy Survey (MzLS; Prop. ID \#2016A-0453; PI: Arjun Dey). DECaLS, BASS and MzLS together include data obtained, respectively, at the Blanco telescope, Cerro Tololo Inter-American Observatory, NSF's NOIRLab; the Bok telescope, Steward Observatory, University of Arizona; and the Mayall telescope, Kitt Peak National Observatory, NOIRLab. The Legacy Surveys project is honored to be permitted to conduct astronomical research on Iolkam Du'ag (Kitt Peak), a mountain with particular significance to the Tohono O'odham Nation. BASS is a key project of the Telescope Access Program (TAP), which has been funded by the National Astronomical Observatories of China, the Chinese Academy of Sciences (the Strategic Priority Research Program ``The Emergence of Cosmological Structures'' Grant \# XDB09000000), and the Special Fund for Astronomy from the Ministry of Finance. The BASS is also supported by the External Cooperation Program of Chinese Academy of Sciences (Grant \# 114A11KYSB20160057), and Chinese National Natural Science Foundation (Grant \# 11433005). The Legacy Survey team makes use of data products from the Near-Earth Object Wide-field Infrared Survey Explorer (NEOWISE), which is a project of the Jet Propulsion Laboratory/California Institute of Technology. NEOWISE is funded by the National Aeronautics and Space Administration. The Legacy Surveys imaging of the DESI footprint is supported by the Director, Office of Science, Office of High Energy Physics of the U.S. Department of Energy under Contract No. DE-AC02-05CH1123, by the National Energy Research Scientific Computing Center, a DOE Office of Science User Facility under the same contract; and by the U.S. National Science Foundation, Division of Astronomical Sciences under Contract No. AST-0950945 to NOAO.
Based in part on data collected at Subaru Telescope and obtained from the SMOKA, which is operated by the Astronomy Data Center, National Astronomical Observatory of Japan \citep{2002ASPC..281..298B}.
\vspace{5mm}
\facilities{Astro Data Archive, Blanco (DECam), CFHT (MegaCam), Gaia, IRSA, PO:1.2m (PTF, ZTF), PS1, Sloan, VST (OmegaCAM)}
\software{{\tt astropy} \citep{robitailleAstropyCommunityPython2013},
{\tt astrometry.net} \citep{langAstrometryNetBlind2010},
{\tt FTOOLS},
{\tt JPL Horizons} \citep{Giorgini1996Horizons},
{\tt SAOImageDS} \citep{joyeNewFeaturesSAOImage2006},
{\tt SkyBot} \citep{berthierSkyBoTNewVO2006},
{\tt Vizier} \citep{ochsenbeinVizieRDatabaseAstronomical2000}
}
|
\section{Introduction}
\label{sec:intro}
For over four decades, scaling relations pertaining to the chemical enrichment of galaxies' interstellar medium (ISM) have been used to study galaxy formation. \citet{lequeux1979} investigated the correlation between galaxy luminosity and gas-phase oxygen abundance and found that more luminous galaxies hosted \ion{H}{2} regions with higher O/H. Later, large surveys like the Sloan Digital Sky Survey \citep[SDSS;][]{york2000} enabled studies of large statistical samples of nearby galaxies. \citet{tremonti2004} were among the first to complete a detailed analysis and found a strong positive correlation between stellar mass (M$_{\ast}$) and gas-phase O/H, which was relatively steep at lower masses and started to flatten at M$_{\ast}\gtrsim10^{10.5}$~M$_{\odot}$. The M$_{\ast}$-O/H relation---often referred to as the ``mass-metallicity relation" or simply the MZR---has continued to be studied extensively at $z\sim0$ (see Section~5 of \citealt{maiolino2019} for a review) and has also been studied using resolved observations of nearby galaxies \citep[e.g.,][]{gonzalez-delgado2014,barrera-ballesteros2017,sanchez2019}. As the size and quality of high-redshift spectroscopic surveys has increased, the MZR has been confirmed to exist up to at least $z\sim3.5$ \citep[e.g.,][among many others]{erb2006metal,maiolino2008,steidel2014,zahid2014,onodera2016,sanders2021}.
Still, connecting the results from these studies and constructing a single clear picture of chemical enrichment over cosmic time has remained difficult, in large part because of the many different methods used to infer O/H. Although some studies of individual nearby objects have used ``direct method" abundances based on measurements of the electron temperature\footnote{Oxygen and other heavy elements provide the principal cooling mechanism in photoionized gas, linking temperature to enrichment.} ($T_e$) in \ion{H}{2} regions \citep[e.g.,][]{aller1954,pilyugin2005,berg2012,berg2020,perez-montero2014}, the spectral lines required to determine $T_e$ are exceedingly faint. Therefore, studies of faint, distant, and/or metal-rich galaxies have relied on more indirect tracers of chemical enrichment. The most common method is to use the ratios of strong emission lines in galaxy spectra, which can be empirically linked to gas-phase O/H using direct method abundances measured for $z\sim0$ samples \citep[e.g.,][]{mcgaugh1991,pettini2004}. Even at the same redshift as the calibration samples, however, the MZRs determined using different ``strong-line" methods often do not agree and differ substantially in both shape and normalization \citep{kewley2008,andrews2013}.
There are additional concerns about the application of strong-line methods at high redshift, where galaxies have significantly different stellar populations and conditions in their interstellar gas. The crux is that line ratio diagnostics are frequently sensitive to more than one parameter---e.g., O/H \emph{and} the abundance of other elements like nitrogen, or O/H \emph{and} the ionization conditions in the gas---many of which are observed to be correlated with one another in nearby galaxies. Indeed, the implicit dependence on these underlying correlations is fundamental to the success of strong-line diagnostics. Unfortunately, we do not know \textit{a priori} whether these properties are correlated with one another in the same way in high-$z$ galaxies as in $z\sim0$ galaxies, and so caution must be exercised when using locally-calibrated line ratio diagnostics to study distant galaxies.
Ideally, we would ``re-calibrate" common line ratio diagnostics using a representative sample of high-$z$ galaxies with measurements of $T_e$. However, despite heroic efforts from the ground \citep{yuan2009,erb2010,steidel2014,sanders2016,kojima2017,sanders2020}, this has not yet been achieved, because measurable $T_e$ corresponds to a relatively narrow range of O/H. Over the next several years, we can hope to revisit the task of constructing new, $T_e$-calibrated diagnostics that are more appropriate for measuring O/H at high redshift using observations with the James Webb Space Telescope \citep[JWST;][]{gardner2006}, scheduled to be launched later this year.
Even so, oxygen abundance is only part of the picture. In the nearby universe, it is common to study the chemical enrichment of galaxies using multiple elements. These studies frequently use observations of the stellar continuum to measure, e.g., the iron, magnesium, and carbon content of the integrated stellar population; see Section~2 of \citet{maiolino2019} for a description of the methodology, but also \citet{gallazzi2005}, which is the complementary study of stellar metallicity in the same galaxies studied by \citet{tremonti2004}. More detailed abundance patterns are also central to studies of the circumgalactic medium (CGM) at all redshifts \citep{tumlinson2017,zahedy2019,zahedy2021}. More recently, very deep spectra of massive, quiescent galaxies at high redshift have enabled the measurement of multiple elemental abundances \citep{kriek2016,kriek2019,jafariyazani2020}. Using deep rest-UV spectra, it is also becoming possible to measure iron abundances in high-$z$ star-forming galaxies \citep[e.g.,][]{steidel2016,cullen2019,topping2020,cullen2021}. These studies represent a promising direction for future high-$z$ galaxy research, because understanding the abundance \emph{patterns} of galaxies provides additional insight regarding their assembly histories---particularly the comparison between iron and $\alpha$ elements (including Mg and O), as the $\alpha$/Fe ratio is sensitive to star formation timescales \citep[e.g.,][]{tinsley1979}.
The goal of this paper is to investigate the abundance patterns of a large sample of individual high-$z$ galaxies in a way that avoids the systematic biases introduced by using locally-calibrated strong-line methods. We update the method introduced by \citet[][hereafter Strom18]{strom2018} and use it to self-consistently infer O/H, N/H, and Fe/H for 195 star-forming galaxies at $z\simeq2-2.7$, the largest sample of high-$z$ galaxies for which multiple elemental abundances have been reported. We introduce the galaxy sample and observations in Section~\ref{sec:data}. Section~\ref{sec:galdna} presents the photoionization model method used for measuring chemical abundances. In Section~\ref{sec:omzr}, we report the M$_{\ast}$-O/H relation for our $z\sim2$ sample and compare and contrast the correlation based on our photoionization model method with the relations from more commonly-used strong-line methods. Section~\ref{sec:newmzr} presents the corresponding M$_{\ast}$-N/H and M$_{\ast}$-Fe/H relations. Section~\ref{sec:patterns} discusses the physical insights that result from measuring abundance patterns rather than a single bulk metallicity, including constraints on galaxy star formation histories. We summarize our findings and conclude in Section~\ref{sec:summary}.
Throughout the paper, we refer to specific spectral features using their vacuum wavelengths and adopt the solar metallicity scale from \citet{asplund2009}, with
\begin{gather}
12+\log(\textrm{O/H})_{\odot}=8.69 \nonumber \\ 12+\log(\textrm{N/H})_{\odot}=7.83 \nonumber \\
12+\log(\textrm{Fe/H})_{\odot}=7.50 \nonumber
\end{gather}
Historically, 12 is added when reporting log abundances relative to H to ensure that values are always positive. This is not used when reporting abundance ratios such as $\log(\textrm{N/O})$.
\section{Data}
\label{sec:data}
In this work, we analyze a sub-sample of galaxies drawn from the Keck Baryonic Structure Survey \citep[KBSS;][]{rudie2012,steidel2014}, using the same criteria as \citetalias{strom2018} (summarized in Section~\ref{sec:spectroscopy}) to construct the sample. KBSS is a targeted spectroscopic survey of $1.5\lesssim z\lesssim3.5$ galaxies in 15 separate fields, each centered on a bright quasar. Each survey field is approximately $6\arcmin\times8\arcmin$, resulting in a total survey area of $\approx0.24$~deg$^2$. Extensive multiwavelength imaging and both optical and near-infrared (NIR) spectroscopic observations have been conducted in all fields and are described in detail in other work \citep[e.g.,][]{steidel2003,reddy2012,rudie2012,steidel2014,strom2017}. Below, we review aspects of the survey that are most relevant to the present analysis.
\subsection{Photometry and Parent Sample Selection}
\label{sec:photometry}
The majority of KBSS galaxies are selected on the basis of their rest-UV colors, using observed-optical imaging in the $U_n$, $G$, and $\mathcal{R}$ bands and the color selection criteria introduced by \citet{adelberger2004} and \citet{steidel2004} to identify Lyman Break Galaxy analogues at $z\simeq2-2.7$ (sometimes referred to as ``BX" and ``BM" galaxies). This color selection largely translates to a star formation rate-selected sample and is successful at identifying both young galaxies and older galaxies with significant current star formation or with rising star-formation histories (SFHs). Relying on the shape and brightness of galaxies' rest-UV spectral energy distributions (SEDs) can introduce a bias against very massive galaxies and galaxies with heavily reddened ($E(B-V)_{\rm cont}>0.3$) UV spectra. To mitigate these biases, we incorporate knowledge of the shape of galaxies' rest-optical spectra by relaxing the rest-UV color selection and pairing it with a cut in $\mathcal{R}-K_s$ color, which probes the 4000~\AA\, and Balmer breaks across the redshift range of interest. \citet{strom2017} provides a more thorough description of these extended selection criteria, but in brief: these ``RK" galaxies occupy a region of rest-UV color space where $z\sim2$ galaxies selected using other NIR methods, such as distant red galaxies \citep[DRGs,][]{franx2003} and $BzK$ galaxies \citep{daddi2004}, are found \citep[see also the discussion by][]{reddy2005}.
\subsection{Near-infrared Spectroscopy}
\label{sec:spectroscopy}
Beginning in 2012, NIR spectroscopic observations of KBSS galaxies have been conducted in $J$-, $H$-, and $K$-band using the Multi-Object Spectrometer for InfraRed Exploration \citep[MOSFIRE,][]{mclean2012} installed at the Cassegrain focus of the Keck~I telescope on Mauna Kea. We presented the first results from the MOSFIRE component of KBSS in \citet[][hereafter Steidel14]{steidel2014}, with a comprehensive analysis of an expanded sample reported in \citet{strom2017}. We refer the reader to those two papers for complete details regarding the NIR spectroscopic data acquisition, reduction and analysis.
The spectra were acquired using 0\farcs7 slits, with typical seeing of 0\farcs5$-$0\farcs8. The data were reduced using the publicly-available data reduction pipeline\footnote{\url{https://www2.keck.hawaii.edu/inst/mosfire/drp.html}} and subsequently corrected to vacuum heliocentric velocity. If multiple spectra of the same object were acquired, they were shifted to match in the spatial direction and combined using inverse-variance weighting. Slit losses were determined to first-order by comparing observations of a relatively bright star placed on a mask slit with its photometric magnitude in the same band; individual object slit losses were determined by comparing observations of the same object on multiple masks. Typical slit loss corrections are factors of $\sim2$, consistent with estimates based on comparisons of detected continuum with broadband photometry and SED models.
The reduced 2D spectrograms were analyzed using the custom IDL package \texttt{mospec}\footnote{\url{https://github.com/allisonstrom/mospec}}, developed by A. Strom specifically for interacting with MOSFIRE spectra of emission line galaxies \citep{strom2017}. Galaxy spectra were extracted using boxcar extraction apertures, then fit using the best-fit SED model as the stellar continuum and an Gaussian emission line model with a single redshift and line width per band. The ratio of the [\ion{O}{3}]$\lambda\lambda4959,5008$ and [\ion{N}{2}]$\lambda\lambda6549,6585$ doublets are fixed at 3:1, where the longer-wavelength lines are 3$\times$ the strength of their shorter-wavelength counterparts. This greatly improves emission line measurements when one of the two features is impacted by an OH sky line.
To date, $\sim1500$ KBSS galaxies have been observed with MOSFIRE, $\sim800$ of which fall in the redshift range $1.9\leq z\leq 2.7$, where many of the key rest-optical emission features are accessible from the ground at NIR wavelengths. As in \citetalias{strom2018}, galaxies are selected for analysis when there are observations of the spectral regions near H$\alpha$, H$\beta$, [\ion{O}{3}]$\lambda5008$, and [\ion{N}{2}]$\lambda6585$. Measurements of or limits on [\ion{O}{2}]$\lambda\lambda3727,3729$, [\ion{O}{3}]$\lambda4363$, [\ion{S}{2}]$\lambda\lambda6718,6732$, and [\ion{Ne}{3}]$\lambda3869$ are incorporated when present. Objects are included in the sample regardless of the signal-to-noise (S/N) ratio of any single line measurement, but we require the ratio of the band-to-band slit corrections to be less than a factor of 2, as well as a S/N$~>5$ measurement of the Balmer decrement H$\alpha$/H$\beta$, which is used to account for reddening due to dust.\footnote{The S/N calculation for the Balmer decrement accounts for uncertainties in the relative flux calibration between NIR bands.} Galaxies with evidence of significant AGN contamination (in either their rest-UV or rest-optical spectra) or whose H$\alpha$ or H$\beta$ is compromised by a nearby OH line are excluded. In total, these criteria result in a sample of 196 galaxies with $\langle z \rangle = 2.3$. Of these, 195 galaxies have good abundance measurements, as described below in Section~\ref{sec:galdna}.\footnote{It was not possible to identify a photoionization model solution for Q1623-BX449, although the MCMC chains did converge.}
\subsection{Stellar Masses}
\label{sec:masses}
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{kbss_z2_mstar_hist_mzr_paper-eps-converted-to}
\caption{The distribution of stellar mass (M$_{\ast}$) for the 195 galaxies in our final sample, shown in blue. For comparison, the grey histogram shows the distribution for the complete parent sample of KBSS galaxies with $1.9\leq z\leq2.7$. The paper sample is consistent with the full KBSS sample and has a median M$_{\ast}=10^{10.0}$~M$_{\odot}$.}
\label{fig:masshist}
\end{figure}
Stellar mass (M$_{\ast}$) estimates for the KBSS galaxies are inferred from reddened stellar population synthesis models fit to broad- and intermediate-band photometry spanning the rest-UV to the rest-IR. The rest-optical (observed-NIR) magnitudes are corrected for line emission using the measurements from the MOSFIRE spectra described above. The general SED fitting methodology is explained by \citet{reddy2012}, with a description of its application to the current KBSS sample found in \citet{strom2017}. \citet{theios2019} examined the impact of using other stellar population models (i.e., those with lower $Z_{\ast}$, as expected for stellar populations at high redshift) on the parameters inferred from SED-fitting. Fortunately, rank-ordering in M$_{\ast}$ is generally preserved regardless of the specific choice of model. As a result, we use the stellar masses from \citet{strom2017}, based on \citet{bruzual2003} stellar population synthesis models with a \citet{chabrier2003} initial mass function, to facilitate a more straightforward comparison with other analyses that use similar SED-fitting techniques and models. The M$_{\ast}$ distributions for the full KBSS sample (grey histogram) and for the 195 galaxies with well-measured abundances (blue histogram) are shown in Figure~\ref{fig:masshist}. These two samples are consistent with being drawn from the same parent population, based on a two-sample Kolmogorov-Smirnov (KS) test. The typical statistical uncertainty on log(M$_{\ast}$/M$_{\odot}$) is $\approx0.16$~dex \citep{shapley2005,erb2006mass}.
\section{Chemical abundances with \texttt{GalDNA}}
\label{sec:galdna}
There are a number of ways to determine the level of chemical enrichment or metallicity of distant galaxies. Broadly speaking, all methods rely on measurements of emission lines from the photoionized gas in galaxies' star-forming regions and fall into two categories: (1) empirical diagnostics based on ``direct" $T_e$-based measurements of a calibration sample where both bright nebular lines and faint auroral lines can be detected or (2) photoionization model methods where observed emission lines are compared to predictions made by models assuming an input ionizing source and set of physical conditions in the gas.
Empirical methods implicitly depend on the correlation among various quantities including, e.g., the ionization state of the gas and the relative abundances of elements like oxygen and nitrogen. The existence of these underlying correlations causes galaxies and \ion{H}{2} regions to form tight sequences in line ratio space---for example, the ``BPT" diagram comparing [\ion{O}{3}]/H$\beta$ and [\ion{N}{2}]/H$\alpha$ \citep{baldwin1981,veilleux1987}---and allow many different line ratios to be used to infer O/H. Even if a line ratio is only weakly correlated with O/H, so long as the quantity that it \emph{is} most strongly correlated with also correlates with O/H, the diagnostic can still be used.
In contrast, photoionization model methods allow chosen parameters to vary alongside gas-phase metallicity and thus be explicitly determined for individual objects. Examples include \texttt{IZI} \citep{blanc2015}, which returns ionization parameter $U$ and O/H; \texttt{BOND} \citep{valeasari2016}, which measures both O/H and N/O; and NebulaBayes \citep{thomas2018}, which measures ISM pressure (related to gas density) in addition to $U$ and O/H. Because we are interested in determining multiple chemical abundances for the galaxies in our sample, we use the photoionization model method introduced by \citetalias{strom2018}, which measures O/H, N/H, and Fe/H in addition to $U$. The general method, including recent updates and a revised parameter estimation technique, is described here.
\subsection{Photoionization Model Grid}
\label{sec:model_grid}
We use the same photoionization model predictions for the strong nebular lines in galaxies' rest-optical spectra as \citetalias{strom2018}, which are generated using Cloudy \citep[v13.02;][]{ferland2013}, using stellar population synthesis models from the Binary Population and Spectral Synthesis\footnote{https://bpass.auckland.ac.nz/} code \citep[BPASSv2;][]{eldridge2016,stanway2016} as the input ionizing radiation field. The BPASS models were chosen because of their relative success in reconciling the rest-UV and rest-optical observations of the same $z\sim2$ galaxies, specifically because the implementation of binary evolution physics in BPASS produces harder ionizing radiation fields at fixed $Z_{\ast}$ than, e.g., including stellar rotation or adopting a more top-heavy stellar initial mass function \citep{steidel2016}. This ``boost" of higher energy photons is needed to match the collisional-to-recombination line ratios such as R23\footnote{$\textrm{R23}=\log\left(\frac{[\textrm{\ion{O}{2}}]+[\textrm{\ion{O}{3}}]}{\textrm{H}\beta}\right)$}. In high-$z$ galaxies, these lines ratios tend to be larger than can be explained using stellar models with softer ionizing radiation fields, even at moderate gas-phase O/H corresponding to peak emission from O ions.
We adopt a plane parallel geometry and assume constant star formation histories with an age of 100~Myr and a constant gas density of $n_{\rm H}=300$~cm$^{-3}$. This density was originally chosen to be consistent with the electron density $n_e$\footnote{For ionized gas, $n_{\rm H}\approx n_e$.} determined using the [\ion{O}{2}] ratio measured from a stack of $z\sim2-3$ KBSS galaxies \citep{strom2017}, but is also representative of the typical electron density for \emph{individual} galaxies in the current sample, which ranges from $n_e\simeq150-400$~cm$^{-3}$. Using models with $n_{\rm H}=100$~cm$^{-3}$ or $n_{\rm H}=1000$~cm$^{-3}$ for galaxies with lower or higher densities than this, respectively, does not significantly change their derived abundances or the overall distribution of abundances in the sample. Dust grains are implemented using the ``Orion" mixture, with a dust-to-gas ratio that scales linearly with the metallicity of the gas, $Z_{\rm neb}$. Model grids are calculated for multiple stellar population synthesis models with increasing stellar enrichment ($Z_{\ast}$), allowing $\log(Z_{\rm neb}/Z_{\odot})$ and $\log(U)$, defined as $\log(n_{\gamma}/n_\textrm{H})$, to vary in 0.1~dex steps. Initially, we assume a solar abundance pattern in the gas, but interpolate between the fixed grid points and scale the strength of the nitrogen lines \textit{a posteriori} to accommodate non-solar N/O. These choices allow us to predict emission line intensities (relative to H$\beta$) for the following parameter space:
\begin{eqnarray}
&Z_{\ast}=[0.001,0.014] \textrm{ or } Z_{\ast}/Z_{\odot}\approx[0.07,1.00], \nonumber \\
&Z_{\rm neb}/Z_{\odot}=[0.1,2.0], \nonumber \\
&\log(U)=[-3.5,-1.5], \nonumber \\
&\log({\rm N/O})\geq-2.0. \nonumber
\end{eqnarray}
The lower limit on log(N/O) is $\sim0.5$~dex lower than the primary plateau observed for local \ion{H}{2} regions \citep[e.g.,][]{vanzee1998,izotov1999}.
\begin{figure*}
\centering
\includegraphics[width=\columnwidth]{chain_Q0105_BX147_low-eps-converted-to}
\includegraphics[width=\columnwidth]{chain_Q0105_BX147_high-eps-converted-to}
\caption{The 1D and 2D posterior probability distributions for 12+log(O/H), 12+log(N/H), and [Fe/H] for Q0105-BX147 (M$_{\ast}=10^{9.5}$~M$_{\odot}$), where the low-O/H and high-O/H peaks have been separated into the left and right panels, respectively. These portions of the posterior are considered separately when estimating the abundance pattern in each case (denoted by the purple symbols and solid lines, with the dashed lines indicating the 68\% HDIs). Disallowed regions in N/O and O/Fe are illustrated by the solid grey shading (c.f. Section~\ref{sec:model_grid}), and the solar values of O/H and N/H are identified by the dashed black lines. To decide between potential solutions, we utilize the corresponding ionization parameter estimates to place the galaxy close to the $z\sim2$ locus in mass-ionization space (Figure~\ref{fig:logu_vs_mass}). In this case, the low-O/H solution is preferred.}
\label{fig:two_peaks}
\end{figure*}
These base model parameters can then be converted to abundances:
\begin{eqnarray}
[\rm Fe/H] &=& \log(Z_{\ast}/Z_{\odot}) \nonumber \\
12+\log(\rm O/H) &=& \log(Z_{\rm neb}/Z_{\odot})+8.69 \nonumber \\
12+\log(\rm N/H) &=& \log(Z_{\rm neb}/Z_{\odot})+8.69+\log(\rm N/O), \nonumber
\end{eqnarray}
where the bracket notation for Fe refers to the enrichment relative to solar, $[\rm Fe/H]=\log(\rm Fe/H)-\log(\rm Fe/H)_{\odot}$. These translations are motivated by several considerations: first, the effect of changing $Z_{\ast}$ on the emission line predictions results from differences in the shape of the ionizing radiation field for a given stellar model family, and the shape of the ionizing spectrum of massive stars is most impacted by the amount of Fe present in their atmospheres. Because this method relies on the Fe enrichment in the massive stars, we avoid concerns about depletion onto dust that impact the abundance of Fe in the gas. In contrast to the important role Fe plays in determining the properties of the massive stars and the amount of gas \emph{heating}, it is O that provides the majority of the gas \emph{cooling} via collisionally-excited lines, due to its relatively high abundance with respect to other elements and many low-lying transitions. Thus, we can assume that changes in $Z_{\ast}$ correspond to changes in Fe/H, and changes in $Z_{\rm neb}$ correspond to changes in O/H.
We use a Markov chain Monte Carlo (MCMC) approach to sample the posterior within multivariate parameter space (c.f. Section~3.2 of \citetalias{strom2018}), adopting the additional prior that $0.0\leq \log(Z_{\textrm {neb}}/Z_{\ast}) \leq 0.73$. These limits roughly correspond to a range of $\alpha$-enhancement (i.e., O/Fe) between the solar value and Fe-poor CCSNe yields \citep{nomoto2006}.
The photoionization model grids and IDL routines used to implement this method, which we collectively refer to as \texttt{GalDNA} (referencing the analogy of abundance patterns as ``galaxy DNA") are publicly available\footnote{\url{https://github.com/allisonstrom/galdna}}. We caution, however, that some of the built-in assumptions, including that all of the emission from ionized gas observed in a galaxy's spectrum is produced in \ion{H}{2} regions, may be inappropriate for some samples. This is especially true for galaxies at lower redshifts and those with lower SFR surface densities, where contributions from, e.g., diffuse ionized gas and shocks are likely to be non-negligible \citep{sanders2017}.
\subsection{Parameter Estimation}
\label{sec:param_est}
In \citetalias{strom2018}, we reported the maximum \textit{a posteriori} (MAP) values for the four univariate posteriors and required the associated 68\% highest density intervals (HDIs) to be separated from the boundaries of allowed parameter space for all parameters except N/O. Here, we revisit the manner in which upper limits on both N/H and Fe/H are determined, as well as how we treat galaxies with multiple peaks in their posterior posterior probability distribution function (PDF). These changes allow us to better capture the abundance patterns in the largest possible sample.
\subsubsection{Bimodal posteriors}
The connection between specific emission lines and the model parameters in \texttt{GalDNA} is discussed at length in \citetalias{strom2018}, but there are a few key results that inform our physical intuition about how differences in galaxy spectra correspond to differences in the model parameters and, thus, differences in chemical abundances. Foremost is the fact that ratios of collisionally-excited lines to recombination lines (e.g., [\ion{O}{3}]/H$\beta$ and [\ion{S}{2}]$\lambda\lambda6718,6732$/H$\alpha$) effectively trace the amount of photoionization heating per ionization of an H atom. Thus, all else being equal, these ratios will \emph{increase} with \emph{decreasing} $Z_{\ast}$ (i.e., Fe/H), because Fe-poor stars produce more energetic ionizing photons on average than Fe-rich stars. At the same time, many of these ratios are also double-valued: they increase as gas enrichment ($Z_{\rm neb}$ or, equivalently, O/H) increases, until gas-cooling mediated by the heavy elements in the gas cause the gas temperature to decrease, at which point these line ratios also decrease\footnote{A notable exception is [\ion{N}{2}]$\lambda6585$/H$\alpha$, which continues to increase with increasing O/H due to the rapid increase in N/O at higher O/H.}. This behavior is well-known and is one of the challenges inherent in using line diagnostics like R23 to infer O/H, because a single value of R23 can correspond to two different values of O/H. Further, R23 remains at $\sim0.8-1.0$ (which is where most high-$z$ galaxies fall) for a large range in O/H near the ``turnover" where gas cooling begins to outpace increasing gas enrichment. The combination of these effects mean that maximal values of line ratios like [\ion{O}{3}]$\lambda5008$/H$\beta$, [\ion{O}{2}]/H$\beta$, and [\ion{S}{2}]/H$\beta$ are only achieved at low Fe/H \emph{and} moderate O/H. Lower values of these line ratios will usually correspond to either lower or higher O/H, which manifest as bimodal posteriors in \texttt{GalDNA}. Figure~\ref{fig:two_peaks} the posterior PDFs for one galaxy (Q0105-BX147) where this occurs, with the total posterior divided into low-O/H (left panel) and high-O/H (right panel) ``solutions."
\begin{figure}
\includegraphics[width=\columnwidth]{met_hist_mzr_paper-eps-converted-to}
\caption{The distribution of \texttt{GalDNA}-inferred O/H for the $z\sim2$ galaxies in our sample. The light grey histogram represents the set of 138 objects having either a single local maximum in the posterior PDF or only one valid solution, if the posterior PDF is double-peaked. The open purple histogram shows the distribution of model-inferred O/H for \emph{both} solutions in the 57 remaining galaxies with bimodal posteriors. While the distribution for galaxies with single-peaked posteriors extends to relatively high O/H, the lowest abundances are only probed by galaxies with more than one potential model solution.}
\label{fig:oh_dist}
\end{figure}
Figure~\ref{fig:oh_dist} demonstrates why properly accounting for galaxies with bimodal posteriors is important for characterizing the correlation between O/H and other quantities, such as M$_{\ast}$ or strong-line ratios. Galaxies with a single local maximum in their O/H posterior (the grey histogram) range from $12+\log(\textrm{O/H})\approx8$ (roughly 20\% solar) to just above solar O/H. There is overlap between the upper edge of this distribution and the higher-O/H maxima for galaxies with bimodal posteriors (the open purple histogram). In contrast, lower-O/H maxima only occur in bimodal posteriors and are not seen for galaxies with single-peaked posteriors.
\begin{figure}
\includegraphics[width=\columnwidth]{logu_vs_mass_select_mzr_paper-eps-converted-to}
\caption{Model-inferred $\log(U)$ and M$_{\ast}$ for galaxies with a single peak in the posterior PDF (large grey points) and for galaxies with multi-modal posteriors (colored points). The distributions of both quantities are consistent between the two samples. The best-fit linear relation for the locus of grey points is shown by the solid black line. For galaxies with multiple possible abundance patterns, the darker symbols indicate the solution that is most consistent with this locus, and the lighter symbol shows the location of the second, dispreferred solution. In general, a lower $U$ corresponds to the low-O/H peak in the posterior, and a higher $U$ corresponds to the high-O/H peak in the posterior. The parameter values for Q0105-BX147, which were also shown in Figure~\ref{fig:two_peaks}, are marked by the purple squares; lower $U$ is preferred and corresponds to the lower O/H peak in the posterior (left panel in Figure~\ref{fig:two_peaks}).}
\label{fig:logu_vs_mass}
\end{figure}
To determine which combination of parameters---and thus which O/H---is preferred for individual galaxies, we rely on the anti-correlation between $U$ and M$_{\ast}$. This relation is one of the strongest and most significant correlations among galaxy properties and has been observed in galaxies samples at multiple redshifts using both ionization-sensitive line ratios \citep{sanders2015,sanders2020sulfur,jeong2020} and inferred ionization parameter \citep[e.g.,][]{kaasinen2018}. The tight locus of galaxies in mass-excitation diagram \citep[or MEx; e.g.,][]{juneau2011,juneau2014}, which compares [\ion{O}{3}]/H$\beta$ and M$_{\ast}$, is another manifestation of this relationship.
Using ionization parameter is also practical because it avoids directly imposing a prior on the distribution of abundances and because $U$ is generally the most precisely-determined model parameter when using \texttt{GalDNA}. Further, the distributions of M$_{\ast}$ and $U$ for galaxies with a single maximum in their posterior and for galaxies with bimodal posteriors (including both possible combinations of parameters) are statistically consistent with one another, unlike the distributions of O/H seen in Figure~\ref{fig:oh_dist}. In other words, not only can both sets of galaxies be expected to follow the same locus in the M$_{\ast}$-$U$ plane, but the two $U$ values for galaxies with bimodal posteriors are distinct enough to meaningfully discriminate between the possible combinations of parameters. To accomplish this, we use the single-peaked galaxies to define the best-fit M$_{\ast}$-$U$ relation, shown by the solid black line in Figure~\ref{fig:logu_vs_mass}. For galaxies with bimodal posteriors, we then choose the combination of parameters with $U$ closest to this locus. The preferred ``solutions" determined in this manner are represented by the dark blue squares (low-O/H maxima) and red triangles (high-O/H maxima).
The resulting distributions of $U$ for the low-O/H and high-O/H solutions are significantly different from one another, with the low-O/H galaxies having generally higher $U$ than high-O/H galaxies. For galaxies that are outliers with respect to these $U$ clusters (Q0821-MD5, Q1700-MD103, and Q1623-BX293), we choose the combination of parameters that minimizes this separation. There are 57 galaxies with bimodal posteriors; we select the low-O/H solution for 31 of these and the high-O/H solution for 26. The portion of the multivariate posterior corresponding to the preferred peak is analyzed on its own to determine the MAP values for O/H, N/H, and Fe/H for individual galaxies.
We also investigated other criteria for deciding between unique combinations of parameters for galaxies with bimodal posteriors. These strategies, including assessing the proximity to the N/H-O/H locus of local \ion{H}{2} regions or deciding based on posterior density alone, do not change the qualitative results presented below. We adopt and endorse the method based on the M$_{\ast}$-$U$ relation because it is easily applied in all cases and uses additional knowledge about the individual galaxies themselves to make a more informed choice.
\subsubsection{Nitrogen and Iron Limits}
\label{sec:limits}
Occasionally, the [\ion{N}{2}]$\lambda\lambda6549,6585$ doublet is poorly detected or undetected in the spectra of galaxies in our sample, even when other emission lines (including nearby H$\alpha$) are well-detected. This can occur as a result of low nitrogen abundance and/or high levels of ionization. For our analysis, we do not require the significant detection of any single emission line (just a S/N~$\geq5$ measurement of the Balmer decrement), but these low S/N measurements of [\ion{N}{2}] correspond to posterior PDFs that abut the lower limit of $\log(\textrm{N/O})\geq-2$. Because the imposed prior is on N/O and not N/H, the lowest allowed N/H is a function of O/H. This will tend to produce a well-defined local maximum in N/H space, even when N/O is an upper limit. To account for this, we test whether \emph{both} the N/H and N/O posterior PDFs are bounded (i.e., detached from the limits of the sampled parameter space). If neither the N/H nor the N/O posteriors are bound, or if the N/O posterior alone is not bound, the nitrogen abundance is assumed to be an upper limit given by
\begin{equation}
12+\log(\textrm{N/H}) = 12+\log(\textrm{O/H})+\log(\textrm{N/O})_{68}, \nonumber
\end{equation}
where $\log(\textrm{N/O})_{68}$ is the 68th percentile in log(N/O). Of the 195 galaxies in our sample, 36 have upper limits on N/H determined in this way.
Iron abundance can also be more difficult to estimate because of the more complicated nature of the Fe/H and Fe/H-O/H posteriors. The most common appearance is a horseshoe shape in Fe/H vs. O/H space, opening downward in the lower left panels of the corner plots in Figure~\ref{fig:two_peaks}. This structure in the 2D posterior arises because there is an inherent trade-off between O and Fe for a given galaxy spectrum: as the ionizing radation field softens with increasing Fe/H---moving toward lower O/Fe---it is necessary to have more O to produce the same line emission but eventually gas cooling due to the increasing O/H begins to dominate, and harder ionizing radiation is required counteract this effect---moving back toward higher O/Fe. This is closely related to the double-valued behavior of R23 with O/H, but additionally highlights the crucial interplay between gas heating \emph{and} gas cooling.
Because of the prior requiring $0.0\leq\textrm{[O/Fe]}\leq0.73$, the two arms of the horseshoe are commonly separated and appear as a bimodal O/H posterior, as we have previously discussed. The 1D Fe/H posterior is also frequently bimodal, although the local maxima are less likely to be bounded. An example of this can be seen in the low-O/H panel on the left in Figure~\ref{fig:two_peaks}, where the marginalized PDF in the bottom right histogram has high posterior density even at the lowest allowed Fe/H, and the 2D posteriors involving Fe/H appear as vertical stripes. In these cases, which also occur at higher O/H, we adopt the 68th percentile in [Fe/H] as an upper limit. A smaller number of galaxies have posteriors that abut the low O/Fe boundary at [O/Fe]$=0.0$, but these often have clear local maxima in the univariate Fe/H posterior, and so we treat them as bounded. In total, 24 galaxies have upper limits on Fe/H.
\section{\texorpdfstring{The $\langle z \rangle = 2.3$ Stellar Mass-O/H Relation}{The <z>=2.3 Stellar Mass-O/H Relation}}
\label{sec:omzr}
Most studies of metallicity in star-forming galaxies, which are largely studied via the nebular emission from their \ion{H}{2} regions, focus on oxygen. This is both physically-motivated---oxygen is the most abundant element (by mass) in the universe after hydrogen and helium---and practical, as emission lines of various ions of oxygen are among the brightest features observed in the spectra of galaxies at all redshifts. As a result, many studies use ``metallicity" interchangeably with oxygen abundance, and the relation between galaxies' stellar mass and gas-phase O/H is among the most common scaling relations discussed in the literature.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{omzr_mzr_paper-eps-converted-to}
\caption{The correlation between stellar mass M$_{\ast}$ and \texttt{GalDNA}-based gas-phase O/H for 195 galaxies at $z\sim2$, with the medians in bins of M$_{\ast}$ represented by the yellow stars. The solar value of $12+\log(\textrm{O/H})_{\odot}=8.69$ is represented by the horizontal dashed line. The magenta line shows the center moving average for O/H with M$_{\ast}$. The dark blue line is the best-fit linear relation, with the blue hatching indicating the 1$\sigma$ uncertainties on the slope and normalization. Most $z\sim2$ galaxies have sub-solar O/H, and there is a marginally-significant positive correlation between O/H and M$_{\ast}$ (Spearman $\rho=0.21$, $p=0.003$). However, there is also substantial intrinsic scatter ($\sigma_{\textrm{O/H}}=0.21\pm0.01$~dex), which together with the measurement uncertainties could camouflage a more significant correlation. The O3N2-based O-MZR reported by \citetalias{steidel2014} for a smaller sample of KBSS galaxies is shown by the dashed orange line.}
\label{fig:omzr}
\end{figure}
\begin{deluxetable*}{@{\extracolsep{5pt}}lrcrrrr}
\tablewidth{0pt}
\tablecaption{Abundance scaling relations with M$_{\ast}$ \label{tab:params}}
\tablehead{
\colhead{} & \multicolumn{2}{c}{Spearman test} & \multicolumn{4}{c}{\texttt{LINMIX\_ERR}} \\
\cline{2-3} \cline{4-7}
\colhead{Element} & \colhead{$\rho$} & \colhead{$p$} & \colhead{P(\%)\tablenotemark{a}} & \colhead{$\alpha$\tablenotemark{b}} & \colhead{$\beta$} & \colhead{$\sigma_\textrm{int}$ (dex)}
}
\startdata
12+log(O/H)$_\textrm{GalDNA}$ & 0.21 & $3\times10^{-3}$ & 99.8 &$8.35\pm0.02$ & $0.14\pm0.05$ & $0.21\pm0.01$ \\
12+log(O/H)$_\textrm{O3N2}$ & 0.50 & $9\times10^{-14}$ & \dots & \dots & \dots & \dots \\
12+log(N/H) & 0.28 & $8\times10^{-5}$ & 100.0 & $7.07\pm0.03$ & $0.29\pm0.07$ & $0.33\pm0.02$ \\
$[\textrm{Fe/H}]$ & 0.27 & $1\times10^{-4}$ & 100.0 & $-0.69\pm0.02$ & $0.17\pm0.05$ & $0.22\pm0.01$ \\
\enddata
\tablenotetext{a}{Probability of a positive linear relation, based on the posterior distribution for $\beta$}
\tablenotetext{b}{Determined at M$_{\ast}=10^{10}$~M$_{\odot}$, c.f. Equation~\ref{eq:mzr}}
\end{deluxetable*}
Figure~\ref{fig:omzr} shows the observed correlation between O/H and M$_{\ast}$ (hereafter O-MZR) for our sample of 195 $z\sim2$ galaxies, where O/H has been measured using \texttt{GalDNA} as described in Section~\ref{sec:galdna}. The asymmetric vertical error bars for each point represent the 68\% highest density interval (HDI) on the univariate O/H posterior, and the horizontal error bars represent the typical $\approx0.16$~dex uncertainty on the stellar masses, not taking into account any systematic uncertainties in determining M$_{\ast}$. A Spearman rank correlation test indicates a weak ($\rho=0.21$) but significant ($p=0.003$) positive correlation\footnote{For the purposes of characterizing correlations in this paper, we consider $\rho<0.3$ to be a weak correlation, $0.3\leq\rho\leq0.6$ to be a moderate correlation, and $\rho>0.6$ to be a strong correlation. We use the propability of the null result $p$ to assess the significance of the correlation, with $p\leq0.003$ considered significant.} (Table~\ref{tab:params}). The magenta line shows the center moving average to give a non-parametric sense of this correlation, which increases by $\approx0.12$~dex in $\log(\textrm{O/H})$ (or $\sim30\%$ in O/H) over 1.5 decades in stellar mass. The yellow stars represent the median $12+\log(\textrm{O/H})$ in bins of M$_{\ast}$, which also show an increase in oxygen enrichment with increasing mass.
The large apparent scatter (due to both measurement uncertainties and intrinsic scatter) could camouflage a more significant correlation than indicated by the Spearman test. To account for this, we also use \texttt{LINMIX\_ERR} \citep{kelly2007} to characterize the relation between M$_{\ast}$ and O/H; \texttt{LINMIX\_ERR} is an IDL-based routine that performs linear regression when there are measurement errors in both variables. Using this method, we find a $99.8\%$ probability of a positive correlation, compared to $99.7\%$ from the Spearman test. The solid blue line shows the best-fit linear relation,
\begin{equation}
12+\log(\textrm{O/H}) = \alpha+\beta[\log(\textrm{M}_{\ast}/\textrm{M}_{\odot})-10],
\label{eq:mzr}
\end{equation}
with $\alpha_\textrm{O/H}=8.35\pm0.02$ and $\beta_\textrm{O/H}=0.14\pm0.05$. These fit parameters, as well as the intrinsic scatter of this relation ($\sigma_{\textrm{O/H}}=0.21\pm0.01$), are also reported in Table~\ref{tab:params}. The hatched blue area represents the $1\sigma$ uncertainties in the slope and normalization.
The general sense of this relation is consistent with other studies of the O-MZR at $z\sim2-3$. The majority of galaxies have sub-solar oxygen abundance, with $\approx45\%$ (O/H)$_{\odot}$ at M$_{\ast} = 10^{10}$~M$_{\odot}$ and trending toward solar at high M$_{\ast}$. For comparison, the first O-MZR reported for KBSS galaxies by \citetalias{steidel2014}, which used the O3N2\footnote{$\textrm{O3N2}=\log([\textrm{\ion{O}{3}}]/\textrm{H}\beta)-\log([\textrm{\ion{N}{2}}]/\textrm{H}\alpha)$ } diagnostic from \citet[][hereafter PP04]{pettini2004} to infer O/H, is shown by the dashed orange line in Figure~\ref{fig:omzr} and is defined by $\alpha_\textrm{Steidel14}=8.27\pm0.01$ and $\beta_\textrm{Steidel14}=0.19\pm0.02$. The new \texttt{GalDNA} relation is shallower than the \citetalias{steidel2014} relation, with the slopes differing by a little more than $1\sigma$. The normalization of the \texttt{GalDNA} O-MZR is also higher than the earlier O3N2 relation by $0.08\pm0.02$~dex. These differences can be attributed to changes in the sample and sample size, as well as differences in methodology, which we discuss in more detail below.
In particular, discrepancies in the abundance scale or normalization of the O-MZR determined using different methods are not unexpected. The O3N2 calibration from \citetalias{pettini2004} that is used in \citetalias{steidel2014} is based on a sample of nearby \ion{H}{2} regions with known $T_e$-based abundances. However, the direct method has been shown to result in abundances up to $\simeq0.24$~dex \emph{lower} than those measured using recombination lines \citep{esteban2004,esteban2014,blanc2015}. \citet{steidel2016} found an offset of $0.25$~dex between the $T_e$-based O/H and the photoionization model-inferred O/H for a composite spectrum of 30 KBSS star-forming galaxies; the O/H estimated using the O3N2 index was intermediate between these two values. The offset between the new \texttt{GalDNA} O-MZR and the older O3N2 O-MZR from \citetalias{steidel2014} is less than the offset between the model O/H and O3N2 O/H reported in that work.
\subsection{Slope}
\label{sec:omzr_slope}
The \texttt{GalDNA} O-MZR in Figure~\ref{fig:omzr} has $\beta_{\textrm{O/H}}=0.14\pm0.05$ and is shallower than the previously-reported KBSS O-MZR from \citetalias{steidel2014}, as well as most other high-$z$ O-MZR relations reported in the literature \citep[][among others]{erb2006metal,maiolino2008,yabe2014,zahid2014,sanders2018,sanders2021}, which range from $\beta\sim0.2$ to $\beta\sim0.35$. Our current understanding of galaxy assembly reasonably predicts a positive correlation between M$_{\ast}$ and O/H, given that the more stars that are created, the larger the mass of oxygen is produced. The slope of the O-MZR can also be affected by other astrophysical processes, including inflows and outflows, which determine how much of the oxygen is retained or expelled and how much it is diluted by infalling low-metallicity gas. A comparatively shallow O-MZR may reflect a relationship between gas content and M$_{\ast}$ at $z\sim2$ that keeps O/H higher in low-M$_{\ast}$ galaxies and/or leads to relatively lower O/H in high-M$_{\ast}$ galaxies. For example, a shallower O-MZR at high redshift could result if low-M$_{\ast}$ galaxies are not efficiently ejecting their enriched gas or if high-M$_{\ast}$ galaxies retain more of their original (pristine) gas mass or accrete more low-metallicity gas from their surroundings, both of which would dilute the oxygen. The effects of these processes are explored by multiple cosmological simulations \citep[e.g.,][]{brooks2007,ma2016mzr,derossi2017,torrey2019}, and we briefly return to this issue in Section~\ref{sec:patterns}, where we discuss the O-MZR together with the scaling relations for nitrogen and iron.
However, before considering the astrophysical interpretation of the O-MZR, we must first consider the practical issue of converting observed quantities to O/H and how that may affect the resulting relation. For example, a shallow O-MZR is a natural consequence of a relatively small dynamic range observed in line ratios like R23 in $z\sim2$ samples. However, other line ratios (including O3N2) exhibit a stronger correlation with M$_{\ast}$, even in high-$z$ samples, and these have historically been used to measure the O-MZR instead. As we have discussed, these O-MZR relations can be dramatically different from one another, but these differences should not always be interpreted as having \emph{physical} meaning, because one of the primary drivers of the slope of strong-line relations is the calibration used to infer O/H.
Figure~\ref{fig:omzr_slope} shows how different O3N2 calibrations with different coefficients for the line ratio can result in very different O-MZR slopes. The black points represent the slope of the O3N2 O-MZR measured when different calibrations for O3N2 (identified by the labels) are used to infer O/H for the same sample of galaxies as shown in Figures~\ref{fig:masshist} and \ref{fig:omzr}; these results are also summarized in
Table~\ref{tab:slopes}. Comparing the two most dissimilar calibrations, from \citetalias{strom2018} and \citet{bian2018}, shows that the slope of the O-MZR can differ by up to $\Delta_{\beta}\approx0.1$ \emph{only because a different calibration was used}. One may attempt to circumvent this uncertainty by selecting the calibration deemed most appropriate for the sample being studied, but it is not always clear how that decision should be made. \citet{bian2018} use direct-method abundances from stacks of ``local analogues" to high-$z$ galaxies to construct their calibration. \citet{marino2013} also use the direct method, in combination with a multiple-line-ratio technique, for \ion{H}{2} complexes in local galaxies that were observed as part of the CALIFA survey \citep{sanchez2012}. In \citetalias{strom2018}, we constructed an O3N2 calibration based on the photoionization model oxygen abundances for our $z\sim2$ sample (albeit only the sub-sample with single peaks), which resulted in an equation with a coefficient very similar to one based on \ion{H}{2} regions from \citet{pilyugin2012}. In the absence of a sufficient sample of direct-method abundances in \emph{high-redshift} galaxies, any or all of these calibrations might be good choices, depending on the sample under consideration.
\begin{deluxetable}{lcrl}
\tablewidth{0pt}
\tablecaption{O3N2 O-MZR slopes \label{tab:slopes}}
\tablehead{\colhead{Sample} & \colhead{O3N2 calibration} & \colhead{Coefficient} & \colhead{Slope}}
\startdata
this paper & PP04 & $-0.32$ & $0.18\pm0.02$ \\
this paper & Marino+13 & $-0.214$ & $0.12\pm0.01$ \\
this paper & Bian+18 & $-0.39$ & $0.22\pm0.02$ \\
this paper & Strom+18 & $-0.21$ & $0.11\pm0.01$ \\
\hline
Steidel+14 & PP04 & $-0.32$ & $0.19\pm0.02$ \\
Sanders+18 & PP04 & $-0.32$ & $0.30$ \\
Sanders+21 & Bian+18 & $-0.39$ & $0.34$ \\
\enddata
\end{deluxetable}
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{slope_comp_mzr_paper-eps-converted-to}
\caption{The slope of the O3N2 O-MZR is extremely sensitive to the choice of O3N2 calibration, which is in turn dependent on the correlations among parameters in the original calibration sample. The round points represent the O3N2 O-MZR measured for our $z\sim2$ galaxy sample using four different calibrations. As the slope of the O3N2 calibration becomes more negative, the slope of the O-MZR grows steeper, which can result in a steeper O-MZR than observed in Figure~\ref{fig:omzr} (horizontal grey line and shaded region). Three O3N2 O-MZRs from the literature are shown by the orange square \citep[][based on a smaller sample of KBSS galaxies]{steidel2014}, the red triangle \citep[from MOSDEF;][]{sanders2018}, and the inverted blue triangle \citep[][also from MOSDEF]{sanders2021}. These points illustrate the differences that can arise merely from sample size and sample selection. Notably, the magnitude of the differences that result from choosing different diagnostics is similar to or even larger than sample effects.}
\label{fig:omzr_slope}
\end{figure}
The difference between the \texttt{GalDNA} O-MZR and the \citetalias{pettini2004} O3N2 O-MZR for our current sample is $\Delta_{\beta}=0.04$. This is larger than the difference in slope when the same \citetalias{pettini2004} O3N2 calibration is used for both the current sample and the sample from \citetalias{steidel2014} ($\Delta_{\beta}=0.01$). We therefore conclude that, in this case, the principal explanation for the difference in slope between our current analysis and \citetalias{steidel2014} seen in Figure~\ref{fig:omzr} is the choice of abundance inference technique, with sample effects playing a more minor role.
At the same time, it is worth noting that the slope of the O3N2 O-MZR reported for KBSS by \citetalias{steidel2014} (the orange square in Figure~\ref{fig:omzr_slope}) and the slope reported for MOSDEF by \citet[][the red triangle in Figure~\ref{fig:omzr_slope}]{sanders2018} is still quite large ($\Delta_{\beta}=0.11$), even though both are based on the \citetalias{pettini2004} calibration for O3N2. This indicates that differences in sample \emph{selection} rather than sample \emph{size} can also play a role. Nonetheless, because it is impossible to know \textit{a priori} the specific origin(s) of any reported differences in the O-MZR, it is likely more useful to compare the scaling relations between M$_{\ast}$ and \emph{line ratio}---rather than between M$_{\ast}$ and O/H, if O/H has been inferred using a strong-line diagnostic.
\subsection{Intrinsic Scatter}
\label{sec:omzr_intsc}
\begin{figure*}
\centering
\includegraphics{omzr_w_logu_mzr_paper-eps-converted-to}
\hspace{-1.28in}
\includegraphics{o3n2_omzr_w_logu_mzr_paper-eps-converted-to}
\caption{The O-MZR for the $z\sim2$ galaxy sample presented in this paper, color-coded by ionization parameter (yellow corresponds to low $U$, and blue corresponds to high $U$). The oxygen abundances were determined using \texttt{GalDNA} in the left panel and using the O3N2 diagnostic from \citetalias{strom2018} in the right panel (the distribution of $12+\log(\textrm{O/H})_{\textrm{O3N2}}$ for $z\sim0$ SDSS galaxies is shown in greyscale). The correlation between O/H and M$_{\ast}$ is statistically significant in both cases, but a difference in intrinsic scatter leads to remarkably different appearance in the two panels. A strong correspondence between O3N2 and $U$, which manifests as horizontal isocontours in the right panel, is fundamental to the construction of strong-line diagnostics at $z\sim0$. Consequently, the tighter O3N2 O-MZR is the result of effectively measuring a combination of two parameters---$U$ \emph{and} O/H---and not due to low intrinsic scatter in the O-MZR itself. This trend is not observed in the \texttt{GalDNA} O-MZR in the left panel, as O/H and $U$ are only weakly anti-correlated using the model method.}
\label{fig:omzr_intsc}
\end{figure*}
A notable feature of the \texttt{GalDNA} O-MZR presented in Figure~\ref{fig:omzr} is the substantial scatter, with the intrinsic scatter (after accounting for measurement error) estimated to be $\sigma_{\textrm{O/H}}=0.21\pm0.05$~dex. This is reasonably consistent with the $0.15$~dex intrinsic scatter reported for the $z\sim0$ O-MZR for galaxies with similar M$_{\ast}$ \citep{tremonti2004} but is more visually apparent here because of the comparatively larger O/H measurement uncertainties for the high-$z$ galaxies.
The origin of the scatter in the O-MZR is still a matter of some debate, although there is a general consensus that deviations from the $z\sim0$ relation correlate with SFR, with higher-SFR galaxies exhibiting lower O/H at fixed M$_{\ast}$ \citep[e.g.,][]{ellison2008,lara-lopez2010}. This ``fundamental metallicity relation" \citep[FMR;][]{mannucci2010} is likely the result of differences in gas fraction \citep{bothwell2013}, in the sense that higher-SFR galaxies have larger gas fractions and, thus, have diluted the O formed by their stellar populations. The scatter may also be attributed to the characteristic timescale for oxygen enrichment \citep[][see also Section~\ref{sec:patterns}]{matthee2018}, gas flows \citep{vanloon2021}, or observational effects like inclination angle \citep{tremonti2004}.
Accurately characterizing the intrinsic scatter in chemical scaling relations like the O-MZR is critical, as it may provide insight regarding the importance of these other processes and/or effects. However, as with the slope of such relations, measuring intrinsic scatter is very sensitive to the method used to infer abundances \citep{tremonti2004}. For example, the intrinsic scatter we report for the \texttt{GalDNA} O-MZR is significantly larger than reported for the O-MZR when strong-line diagnostics are used to infer O/H, which is usually $\sigma_{\textrm{O/H}}\lesssim0.08$~dex. However, given the similar (or even larger) scatter in the measurements originally used to establish the diagnostics, such a small amount of scatter is unexpected---and may instead indicate that the O-MZR determined using strong-lines is a projection of a higher-dimensional surface linking M$_{\ast}$, O/H, and other parameters.
Figure~\ref{fig:omzr_intsc} shows that ionization parameter $U$ is a likely candidate for a third parameter that would account for the diminution in observed scatter. The left panel shows the \texttt{GalDNA} O-MZR from Figure~\ref{fig:omzr}, and the right panel shows the O3N2 O-MZR for the same galaxies using the O3N2 calibration from \citetalias{strom2018}. In both panels, points have been color-coded based on the value of $U$ for that galaxy. The trend with $U$ is remarkably different between the panels: the isocontours of $U$ are nearly horizontal in the right panel, while there is considerable scatter in $U$ in the left panel. Using the Pearson correlation coefficients, we can compare the strength of the correlation between model parameters from \texttt{GalDNA} and O3N2: we find $r_{\textrm{O/H-O3N2}}=-0.40$, indicating a moderate anti-correlation between O3N2 and log(O/H), whereas there is a much stronger correlation between O3N2 and log($U$), $r_{U\textrm{-O3N2}}=0.74$. The correlation between O3N2 and \emph{both} $12+\log(\textrm{O/H})$ and $\log(U)$ is $r_{\textrm{O/H+}U}=0.78$. This reflects the fact that O3N2 calibrations depend on a strong underlying (anti-)correlation between O/H and $U$. We did not find strong evidence for such an anti-correlation in our $z\sim2.3$ sample in \citetalias{strom2018}, but our new results for galaxies with bimodal posteriors indicate that the galaxies with higher $U$ tend to have lower O/H and vice versa, although it is not clear whether this is consistent with $z\sim0$ $U$-O/H relation underpinning all locally-calibrated diagnostics. Even if the $U$-O/H relation is the same at $z\sim2$ as at $z\sim0$, the strong correlation between O3N2 and $U$ in most calibration samples means that the scatter in \emph{any} O3N2 O-MZR is not accurately tracing the scatter in the underlying O/H scaling relation. The additional dependence on $U$ will artificially suppress the apparent scatter and thus the inferred intrinsic scatter as well \citep[c.f. Section~7.2,][]{steidel2014}.
Trends with ionization parameter could be the result of underlying differences in ionizing photon flux density, in \ion{H}{2} region gas density and pressure, or both. However, we find no significant correlation between M$_{\ast}$ and $n_e$ in our sample (Spearman $\rho=-0.06$, $p=0.43$), suggesting that differences in $n_{\gamma}$ may be more important. The role of ionization parameter, its relationship to other quantities such as SFR, and how it should be physically interpreted in high-$z$ galaxies will be addressed in future work. Here, we emphasize that methods like \texttt{GalDNA} that explicitly account for other key parameters---including $U$, but also multiple abundances---offer a promising opportunity to investigate scaling relations in greater detail and build more astrophysical intuition from the same observations.
\subsection{Measuring Shallow Relations with Large Scatter}
\label{sec:shallow}
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{recovery_map_mzr_paper-eps-converted-to}
\includegraphics[width=\columnwidth]{recovery_map_b_mzr_paper-eps-converted-to}
\caption{The significance of a monotonic correlation between O/H and M$_{\ast}$, assessed using a Spearman rank correlation test, as a function of the intrinsic scatter and slope of the underlying O-MZR (top) and as a function of the mean measurement error and galaxy sample size (bottom). Darker colors indicate a more significant correlation, and the $95\%$, $99\%$, and $99.9\%$ confidence isocontours are denoted by the dashed and dotted white lines. As seen in the top panel, the intrinsic slope of the O-MZR has a stronger effect on the significance of the observed relation than differences in intrinsic scatter (for a fixed sample size and mean measurement error). Assuming the intrinsic O-MZR has the comparatively shallow slope and moderate intrinsic scatter measured from Figure~\ref{fig:omzr}, improving the precision of abundance measurements could dramatically increase the significance of the observed scaling relation. The mean $1\sigma$ error on $12+\log(\textrm{O/H})_\textrm{O3N2}$ (vertical green line, bottom panel) is small enough to ensure a highly significant observed correlation even for much smaller samples---but this does not include systematic uncertainties stemming from scatter in the original calibration and is likely to be an underestimate of the true uncertainty on O3N2-based O/H abundances.}
\label{fig:recover}
\end{figure}
Above, we highlighted scenarios that might result in a shallow O-MZR with relatively large scatter. However, it is much more challenging to confidently characterize such relations than steeper or tighter correlations.
To understand our results in the context of other O-MZR measured for similar samples, we simulated the correlation between O/H and M$_{\ast}$ for a sample of 195 galaxies with the same distribution of M$_{\ast}$ as shown by the blue histogram in Figure~\ref{fig:masshist}. We consider a range of slopes $\beta_{\textrm{O/H}}=0.05-0.16$ and intrinsic scatter $\sigma_{\textrm{O/H}}=0.10-0.23$~dex. The Spearman rank correlation coefficient and the probability of the null result are then calculated for $n=500$ realizations for each $\beta_{\textrm{O/H}}$ and $\sigma_{\textrm{O/H}}$.
The top panel of Figure~\ref{fig:recover} shows the median correlation significance ($1-p_{\textrm{Spearman}}$) for the O-MZR as a function of slope $\beta_{\textrm{O/H}}$ and intrinsic scatter $\sigma_{\textrm{O/H}}$, with darker shading indicating a more significant correlation. The dashed and dotted white lines denote the $95\%$, $99\%$, and $99.9\%$ confidence isocontours. The location of the \texttt{GalDNA} O-MZR in this parameter space is marked by the white star (with $p=0.997$). This experiment demonstrates why the strong-line-determined O-MZRs from the literature and the O3N2 O-MZR in Figure~\ref{fig:omzr_intsc} appear more obvious visually and tend to be more statistically significant. Such relations are generally steeper and have smaller scatter, which would position them to the upper left in comparison to the corresponding confidence isocontours. As we have noted, however, it is not obvious that the observed scatter in the O3N2 O-MZR reflects the true intrinsic scatter.
The two most straightforward ways to increase the significance of an observed correlation with a given intrinsic $\beta$ and $\sigma$ are to increase the sample size or decrease the statistical measurement uncertainties. The bottom panel shows the results from a similar experiment as the previous one, but now varying the galaxy sample size (where the stellar masses are drawn from a normal distribution with the same mean and variance as the paper sample) and mean measurement error on the abundance. For reference, the mean measurement error on $\log(\textrm{O/H})_{\textrm{O3N2}}$ is shown by the vertical green line, although this does not capture any systematic uncertainties. A modest increase in sample size (to $\sim230$~galaxies) or decrease in measurement error (to $\sim0.11$~dex) would increase our confidence in a positive correlation to $99.9\%+$.
In coming years, we will continue to see progress on both fronts: samples of high-quality rest-optical spectra of $z\gtrsim2$ galaxies are increasing in size, and advances in models of massive star evolution and independent constraints on Fe/H in high-$z$ galaxies will both lead to narrower posterior distributions for O/H. In the meantime, it is important to continue to invest resources in developing methods of abundance inference that can accurately characterize both the slope \emph{and} the scatter in the O-MZR. It is also important to study galaxies' abundance \emph{patterns} rather than a single proxy for their bulk metallicity. \texttt{GalDNA} is an example of this kind of method, and in the next section, we present the corresponding scaling relations for nitrogen (N-MZR) and iron (Fe-MZR) that result from our analysis.
\section{Relations for Other Elements}
\label{sec:newmzr}
Here we present the N-MZR and Fe-MZR for the same galaxies that define the O-MZR in Figure~\ref{fig:omzr}. These results comprise the largest study to date of multiple elemental abundances in individual high-$z$ galaxies.
\subsection{The N-MZR}
\label{sec:nmzr}
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{nmzr_mzr_paper-eps-converted-to}
\caption{The correlation between stellar mass M$_{\ast}$ and gas-phase N/H for our $z\sim2$ galaxy sample, shown in the same manner as Figure~\ref{fig:omzr}. Upper limits on N/H are shown as darker red triangles. The magenta line follows the center moving average, and the solid red line indicates the best-fit linear relation (uncertainties illustrated by the hatched region). Medians in bins of M$_{\ast}$ are shown by the yellow stars. The observed correlation is stronger than the O-MZR, with a rank correlation coefficient of $\rho=0.28$ ($p=8\times10^{-5}$) and a slope of $\beta_{\textrm{N/H}}=0.29\pm0.07$ for the linear relation. The intrinsic scatter in this relation is also larger than that measured for the O-MZR ($\sigma_{\textrm{N/H}}=0.33\pm0.02$~dex).}
\label{fig:nmzr}
\end{figure}
The correlation between M$_{\ast}$ and N/H for our sample of 195 $z\sim2$ galaxies is shown in Figure~\ref{fig:nmzr}. As for the O-MZR in Figure~\ref{fig:omzr}, individual galaxies are represented by the colored points, with the darker triangles indicating upper limits on N/H (c.f. Section~\ref{sec:limits}). The center moving average in magenta provides a non-parametric view of the N-MZR, and the yellow stars represent the median $12+\log(\textrm{N/H})$ in bins of M$_{\ast}$. The Spearman rank correlation coefficient for the N-MZR is $\rho_{\textrm{N/H}}=0.28$ ($p=8\times10^{-5}$), indicating a significant moderate correlation. The correspondence with M$_{\ast}$ is stronger for N/H than for O/H, although it also remains sub-solar across the entire sampled mass range, with $12+\log(\textrm{N/H})=7.07\pm0.03$ at $10^{10}$~M$_{\odot}$, or $\approx17\%$ (N/H)$_{\odot}$.
The best-fit linear relation determined using \texttt{LINMIX\_ERR} is shown by the dark red line, with the hatched region illustrating the 1$\sigma$ uncertainties on the slope and normalization; the fit parameters can be found in Table~\ref{tab:params}. The N-MZR is significantly steeper than the O-MZR with $\beta_{\textrm{N/H}}=0.29\pm0.07$ and has significantly larger intrinsic scatter $\sigma_{\textrm{N/H}}=0.33\pm0.02$. The mean measurement error for $12+\log(\textrm{N/H})$ is also larger, with $s_{\textrm{N/H}}=0.23$, in large part due to the fact that N/H is determined almost entirely by observations of a single feature [\ion{N}{2}]$\lambda\lambda6549,6585$ that tracks the number of nitrogen atoms in the gas. This is in contrast to O/H, which dominates the cooling in \ion{H}{2} regions and thus influences all collisionally-excited transitions. As a result, O/H can be inferred from an ensemble of lines. The larger scatter in the N-MZR may have a physical origin as well, as nitrogen is thought to be produced by intermediate mass stars and returned to the ISM on somewhat longer timescales than oxygen \citep{pettini2002}.
\subsection{The Fe-MZR}
\label{sec:femzr}
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{femzr_mzr_paper-eps-converted-to}
\caption{The correlation between stellar mass M$_{\ast}$ and Fe/H for our $z\sim2$ galaxy sample. As in Figures~\ref{fig:omzr} and \ref{fig:nmzr}, the magenta line follows the center moving average, the solid green line indicates the best-fit linear relation (uncertainties illustrated by the hatched region), and medians in bins of M$_{\ast}$ are shown by the yellow stars. The rank correlation coefficient is $\rho=0.27$ ($p=1\times10^{-4}$), and the linear relation has $\beta_{\textrm{Fe/H}}=0.17\pm0.05$ and $\sigma_{\textrm{Fe/H}}=0.22\pm0.01$~dex. Unlike the O-MZR and N-MZR, the Fe-MZR remains significantly sub-solar even at high M$_{\ast}$.}
\label{fig:femzr}
\end{figure}
Figure~\ref{fig:femzr} shows the correlation between M$_{\ast}$ and inferred [Fe/H] for our $z\sim2$ galaxy sample in the same manner as the O-MZR in Figure~\ref{fig:omzr} and the N-MZR in Figure~\ref{fig:nmzr}. Rather than present the iron abundances at face value, however, we use bracket notation, which indicates the enrichment relative to solar and is commonly used in studies of stellar enrichment. The Spearman rank correlation coefficient for the Fe-MZR is $\rho_{\textrm{Fe/H}}=0.27$ ($p=1\times10^{-4}$), which indicates a stronger and more significant correlation than the O-MZR for the same sample.
The Fe-MZR also has a comparatively low normalization, with $[\textrm{Fe/H}]=-0.69$ ($\approx20\%$ solar) at M$_{\ast}=10^{10}$~M$_{\odot}$. This is consistent with the stellar metallicity relation (read: Fe-MZR) reported by \citet{cullen2019} for a sample of $2.5\leq z\leq5$ galaxies, which had $[\textrm{Fe/H}]\approx-0.7$ based on measurements from rest-UV spectra. This relatively low degree of enrichment is expected, given that the typical ages of star-forming galaxies at $z\sim2$ ($\lesssim300$~Myr) are lower than the characteristic timescale for substantial iron enrichment due to Type~Ia SNe ($\sim1$~Gyr).
As with the O-MZR, the Fe-MZR in Figure~\ref{fig:femzr} is shallower than other recent studies at $z\gtrsim2$ \citep{cullen2019,cullen2021}, which found $\beta_{\textrm{Fe/H}}\sim0.3$ using galaxies' rest-UV spectra to estimate stellar metallicity. \citet{theios2019} also showed a somewhat steeper increase in stellar metallicity as a function of M$_{\ast}$ ($\beta_{\textrm{Fe/H}}\approx0.29$) for KBSS galaxies based on stacks of rest-UV spectra in bins of stellar mass (see Figure~16 in that paper). Using rest-UV photospheric spectra is a more direct method of inferring stellar metallicity than \texttt{GalDNA}, and we showed in \citet{steidel2016} that stellar metallicity inferred this way is relatively insensitive to the choice of stellar model. However, \citet{cullen2021} found that stellar metallicities inferred using BPASS were systematically lower than metallicities inferred using Starburst99 models \citep{leitherer2010}. There is is also substantial diversity in the predicted ionizing spectra of the same massive stars, which is what \texttt{GalDNA} relies on to indirectly estimate Fe/H. Both methods are ultimately model-dependent---albeit in different ways---and it is currently difficult to directly compare them without larger, representative samples of individual galaxies where stellar metallicity can be measured using rest-UV spectra.
\section{Astrophysical Insight from Abundance Patterns}
\label{sec:patterns}
The principal advantage of using photoionization model-based methods like \texttt{GalDNA} is the ability to explicitly account for additional factors that influence the physics of star-forming regions. Photoionization models are also the best way to constrain abundance ratios for the same galaxies using common observables. Routinely acquiring sufficiently deep rest-UV spectra to measure Fe/H for individual high-$z$ galaxies remains prohibitively expensive, although this will hopefully change in the future with the construction of ``extremely large telescopes" (ELTs).
In this section we revisit aspects of the discussion in Sections~\ref{sec:omzr} and \ref{sec:newmzr}, but now consider how the normalization, slope, and scatter of the separate scaling relations can be used \emph{together} to build physical intuition regarding the chemical evolution of high-$z$ galaxies. Figure~\ref{fig:mzrcomp} shows the best-fit O-MZR (blue, dotted white line), N-MZR (red, dashed white line), and Fe-MZR (green, dot-dashed white line) described above on the same log scale relative to solar to facilitate comparison. The shading around the best-fit linear relations indicates the intrinsic scatter in each case.
\begin{figure}
\centering
\includegraphics[width=\columnwidth,trim=0 27 0 55,clip]{mzr_comp_mzr_paper}
\caption{A comparison of the O-MZR (blue), N-MZR (red), and Fe-MZR (green) for our $z\sim2$ galaxy sample, plotted on the same log scale relative to solar. The shading illustrates the intrinsic scatter measured for each relation. The O-MZR and Fe-MZR have very similar slopes and are offset from one another by $0.35\pm0.03$~dex at $M_{\ast}=10^{10}$~M$_{\odot}$, which translates to $\textrm{O/Fe}\approx2.2\times(\textrm{O/Fe})_{\odot}$. This stays relatively constant across the sampled mass range, implying that young stars even in massive galaxies at $z\sim2$ are significantly $\alpha$-enhanced. The N-MZR rises much more quickly than the other relations, starting near the primary plateau in N/O at M$_{\ast}\sim10^{9}$~M$_{\odot}$ and surpassing the Fe-MZR near M$_{\ast}\sim10^{11}$~M$_{\odot}$.}
\label{fig:mzrcomp}
\end{figure}
\subsection{Alpha-enhancement in High-redshift Galaxies}
\label{sec:norm_comp}
The offset between the O-MZR and Fe-MZR yields [O/Fe]~$=(\alpha_{\rm O/H}-8.69)-\alpha_{\rm Fe/H}=0.35\pm0.03$ at $M_{\ast}=10^{10}$~M$_{\odot}$, corresponding to $\textrm{O/Fe}\approx2.2\times(\textrm{O/Fe})_{\odot}$. This is lower than the median ratio reported in \citetalias{strom2018}, [O/Fe]~$=0.42$ or $\textrm{O/Fe}\approx2.6\times(\textrm{O/Fe})_{\odot}$, but still consistent given the typical measurement uncertainties ($\sim0.2$~dex). Figure~\ref{fig:ofe_hist} shows the updated distribution of [O/Fe] for the 63 individual galaxies in the current sample with bounded posterior PDFs for O/Fe. These galaxies have a median [O/Fe]~$=0.39$ or $\textrm{O/Fe}\approx2.5\times(\textrm{O/Fe})_{\odot}$ and an interquartile range of [O/Fe]~$=0.28-0.47$ or $\textrm{O/Fe}\approx1.9-3.0\times(\textrm{O/Fe})_{\odot}$. Both estimates of the typical O/Fe in our $z\sim2-3$ galaxy sample are in line with the $\alpha$-enhancement reported by \citet{cullen2019} and \citet{sanders2020}. These works focused on different spectroscopic features from one another and from our analysis, and the agreement between all of these methods should be seen as strong confirmation of elevated $\alpha/$Fe in high-$z$ galaxies.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{ofe_hist_mzr_paper-eps-converted-to}
\caption{The distribution of [O/Fe] for the 63 individual galaxies with bound posteriors in the abundance ratio. For these galaxies, the median [O/Fe]~$=0.39$, corresponding to $\textrm{O/Fe}\approx2.5\times(\textrm{O/Fe})_{\odot}$. This is slightly higher but still consistent with the [O/Fe] inferred from the offset between the O-MZR and Fe-MZR, which is shown by the dotted blue line, with the shaded region representing the statistical uncertainty in this value.}
\label{fig:ofe_hist}
\end{figure}
Although the slope of Fe-MZR is slightly higher than the slope of the O-MZR, they agree within errors, indicating a relatively constant level of $\alpha$-enhancement across the entire sampled mass range. Likewise, there is no statistically significant anti-correlation between [O/Fe] and M$_{\ast}$ for the subset of galaxies in Figure~\ref{fig:ofe_hist} (Spearman $\rho=-0.29$, $p=0.02$). Perhaps the most direct comparison of this finding to other samples is with the recent results for $z\sim3$ galaxies from \citet{cullen2021}, who compared their rest-UV-measured Fe/H with the strong-line-based O-MZR from \citet{sanders2021}. As we do, the authors found virtually identical---albeit much steeper---slopes for their gas-phase metallicity relation (O-MZR) and stellar metallicity relation (Fe-MZR).
The constancy of super-solar O/Fe with M$_{\ast}$ appears robust to the choice of method and is intriguing, as it may suggest that the mode of production for O and Fe remains the same for galaxies spanning the observed range in M$_{\ast}$. The implication then is that the Fe in even relatively massive galaxies has been produced---along with the O---by CCSNe, rather than by Type Ia SNe. Significant Fe production by Type Ia SNe would tend to decrease O/Fe in older galaxies, and we would expect O/Fe to decrease with increasing M$_{\ast}$ (i.e., for the Fe-MZR to be steeper than the O-MZR).
This is contrary to our understanding of $\alpha$-enhancement in nearby galaxies, which is highest in massive galaxies \citep[e.g.,][]{faber1992,thomas2010,conroy2014,segers2016}. However, our analysis pertains to the abundances in and around young, massive stars and captures the \emph{current} level of enrichment in galaxies, whereas the $\alpha$-enhanced stellar populations in, e.g., elliptical galaxies are composed of lower mass stars that likely formed at earlier times, some at similar redshifts as our galaxy sample. Significant $\alpha$-enhancement in $z\sim2$ galaxies can occur if they are relatively young ($\lesssim300$~Myr) or have rising SFHs. Considering the observed trend in cosmic star formation rate density at these redshifts, it is possible that this is the case and that older galaxies and galaxies with declining SFHs---where we might expect lower O/Fe---are comparatively uncommon. However, it may also suggest that these galaxies are underrepresented in existing surveys of emission-line galaxies.
Setting aside differences in the slopes reported for the O-MZR and Fe-MZR in different studies, the slope of the Fe-MZR must effectively function as an upper limit on the slope of the O-MZR. The Fe-MZR may be \emph{steeper}---if Type~Ia SNe have contributed to the iron in the ISM of older, more massive galaxies---but generally not \emph{shallower} than the O-MZR. This is because the limiting case is an enrichment pattern resulting from pure CCSNe ejecta at all stellar masses, which would tie together the slopes of the two relations. Curiously, the typical $\alpha$-enhancement in our sample ([O/Fe]~$=0.35\pm0.03$) is less than half what is expected from pure Fe-poor CCSNe enrichment ([O/Fe]$\approx0.73$; \citealt{nomoto2006}). Some of this discrepancy could result from our definitions of solar enrichment for O and Fe, but it may also point to other sources of enrichment, for example ``prompt" Type Ia SNe with characteristic delay times $\lesssim400$~Myr \citep{scannapieco2005,matteucci2009,yates2013}.
Nitrogen production is clearly not tied to oxygen and iron production in the same way, as the N-MZR has a significantly steeper slope than both the O-MZR and the Fe-MZR (Figure~\ref{fig:mzrcomp}). At the low-mass end of our sample (M$_{\ast}=10^9$~M$_{\odot}$), comparing the O-MZR and N-MZR yields $\log(\textrm{N/O})\sim-1.45$, close to the primary plateau. Based on the slope and normalization of the two relations, $\log(\textrm{N/O})_{\odot}=-0.86$ would occur at M$_{\ast}=10^{12.8}$~M$_{\odot}$, a much higher mass than one might expect to need to reach solar abundances. This may indicate that the \texttt{GalDNA} N-MZR relation in Figure~\ref{fig:nmzr} is actually too shallow.
Still, a lower N/H coupled with a more rapid increase with M$_{\ast}$ relative to O/H is a robust result that persists regardless of how the final sample is constructed, including the treatment of upper limits on N/H and N/O (c.f. Section~\ref{sec:limits}). These relative abundances suggest that most high-$z$ galaxies with M$_{\ast}\gtrsim10^9$~M$_{\odot}$ are engaging in secondary production of nitrogen. In high-$z$ damped Lyman-$\alpha$ systems (DLAs), where nitrogen and oxygen abundances have also been studied in some detail, N/O is near or below the primary plateau at $\log(\textrm{N/O})\sim-1.5$, but these systems are generally very metal-poor and have lower O/H than the galaxies in our sample \citep[e.g.,][]{pettini1995,pettini2002,pettini2008,petitjean2008}. It seems likely that pushing $z\sim2$ galaxy surveys to lower masses would allow us to begin directly studying galaxies producing primary nitrogen.
\subsection{Timescales of Galaxy Assembly}
\label{sec:intsc_comp}
The behavior of the intrinsic scatter of the MZR for different elements can provide meaningful benchmarks for simulations of galaxy formation and evolution. For example, \citet{matthee2018} argued that the scatter in $z=0.1$ chemical scaling relations should be expected to differ depending on the element, with the O-MZR having the least instrinsic scatter and the N-MZR and Fe-MZR having significantly larger scatter. Their interpretation of the physical origin of these differences relates to the different timescales on which elements are formed and returned to the ISM as a result of their progenitors. For $\alpha$ elements like O, which is produced by CCSNe on relatively short timescales, deviations from equilibrium are short-lived and should therefore result in less observed scatter.
Figure~\ref{fig:intsc} shows the posterior distributions of $\sigma_{\textrm{int}}$ for the O-MZR (blue, dotted line), Fe-MZR (green, dot-dashed line), and N-MZR (red, dashed line). As predicted, the O-MZR exhibits the smallest intrinsic scatter, $\sigma_{\textrm{O/H}}=0.21\pm0.01$. For our sample, however, the intrinsic scatter of the Fe-MZR is $\sigma_{\textrm{Fe/H}}=0.22\pm0.01$~dex, statistically identical to the O-MZR and significantly smaller than that of the N-MZR. The relative scatter suggests that most or all of the iron enrichment in our galaxy sample occurred on comparatively shorter timescales like oxygen (i.e., in CCSNe). Such an interpretation is consistent with the super-solar O/Fe observed for galaxies at all M$_{\ast}$, as discussed previously.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{sig_hist_mzr_paper-eps-converted-to}
\caption{The intrinsic scatter in the correlation between M$_{\ast}$ and abundance varies between elements. The histograms show the posterior distributions of $\sigma_{\textrm{int}}$ from fitting the relations plotted in Figures~\ref{fig:omzr}, \ref{fig:nmzr}, and \ref{fig:femzr}. The intrinsic scatter in the O-MZR is the smallest ($\sigma_\textrm{O/H}=0.21\pm0.01$), followed by the Fe-MZR ($\sigma_\textrm{Fe/H}=0.22\pm0.01$), and the N-MZR ($\sigma_\textrm{N/H}=0.33\pm0.02$). The comparatively low scatter in the Fe-MZR relative to the N-MZR and its relative similarity to the O-MZR may be evidence that most or all of the iron in our galaxy sample was created by CCSNe.}
\label{fig:intsc}
\end{figure}
The large intrinsic scatter of the N-MZR $\sigma_{\textrm{N/H}}=0.33\pm0.02$ could suggest that nitrogen enrichment occurs on longer timescales than oxygen and iron. However, the observed $\sigma_{\textrm{N/H}}$ significantly exceeds the predictions of \cite{matthee2018}, and although the predictions are for scaling relations at $z=0.1$, it is not obvious why one would observe much higher scatter at $z\sim2$. If this increase is real, it could imply differences in the production of nitrogen in lower-metallicity galaxies, such as those at high redshift. Unfortunately, there are not significantly better observational constraints that can be placed on N/H, as [\ion{N}{2}]$\lambda\lambda6549,6585$ is the only nitrogen transition commonly observed in high-$z$ spectra. It is possible that in coming years, it will be possible to measure N/H directly in high-$z$ galaxies with observations of the auroral [\ion{N}{2}]$\lambda5755$ line with JWST; however, samples will be small owing to the extreme faintness of the auroral line relative to the nebular line.
\subsection{Constraints on Gas Flows}
Many studies use the effective yield $y_{\rm eff}\equiv Z/\ln(f_{\rm gas}^{-1})$, to understand the relationship between inflows and outflows and the shape of the O-MZR \citep[e.g.,][]{edmunds1990,garnett2002,dalcanton2007}. In a closed-box model, $y_{\rm eff} = y$, the true yield, which is assumed to be constant; differences in $y_{\rm eff}$ can be attributed to gas flows in and out of the galaxy. In $z\sim0$ galaxies, $y_{\rm eff}$ is seen to increase significantly with increasing baryonic mass \citep{tremonti2004}, which is interpreted as preferential metal loss from lower-mass galaxies due to their shallower potential wells. In contrast, high-$z$ galaxies show the opposite trend \citep{erb2006metal,mannucci2009,troncoso2014}, with $y_{\rm eff}$ \emph{decreasing} slightly with increasing mass. The implication of this result is that the efficiency of outflows does not decline with M$_{\ast}$ in $z\gtrsim2$ galaxies, and high-M$_{\ast}$ galaxies at these redshifts may even be more efficient at ejecting metals than their lower-M$_{\ast}$ counterparts \citep{torrey2019}. Indeed, having mass outflow rates scale proportionally with SFR naturally produces the decline in $y_{\rm eff}$ seen at high masses. Similarly, recent work by \citet{jian2018} suggests that the parallel slopes of the O-MZR and Fe-MZR at high-$z$ and relatively flatter O-MZR at $z\sim0$ are a consequence of larger metal loading factors at early times. The ability to compare the O-MZR and Fe-MZR at multiple redshifts is increasingly used in assessing the success of cosmological simulations \citep{ma2016mzr,derossi2017}.
Although we have not measured $f_{\rm gas}$ and $y_{\rm eff}$ for our sample, the shallower slope of the \texttt{GalDNA} O-MZR should translate to a steeper decline in $y_{\rm eff}$ with M$_{\ast}$ than seen in previous high-$z$ samples. This assumes the correlation between $f_{\rm gas}$ and M$_{\ast}$ for our galaxies is similar to that reported by \citet{erb2006metal}, a reasonable assumption, given that they share the same parent sample. A steep decline in $y_{\rm eff}$ requires massive galaxies to be even more efficient at ejecting metals from their ISM and implies very high mass-loading factors for outflows. This finding is more easily explained by a combination of enriched outflows \emph{and} unenriched inflows that scale with SFR \citep{erb2008}, which also produces a more marked decline in $y_{\rm eff}$ with M$_{\ast}$---and thus a shallower O-MZR slope.
Although disentangling the relative importance of inflows and outflows remains challenging, the CGM provides a ``fossil record" of past outflow activity in the form of heavy elements deposited there by galactic winds. The CGM contains a significant fraction of all the metals created in stars, and the distribution and physical state of this enriched gas can provide constraints on energetic feedback in galaxies \citep{tumlinson2011,peeples2014,oppenheimer2016,rudie2019}. Metal-bearing gaseous structures in the IGM and CGM at $z\sim2-3$ appear to be very small \citep[$\lesssim100$~pc; ][]{rauch2001,rudie2019}, and there are significant variations in the metal content and abundance pattern of the CGM of single galaxies \citep{zahedy2019,zahedy2021,cooper2021}, suggesting that metals are often poorly mixed in the gas around galaxies. Future studies comparing the composition of the CGM with other components of the galaxy, including the ISM and stars, will offer critical insight into the baryonic flows that profoundly shape the evolution of galaxies.
\section{Summary}
\label{sec:summary}
We have inferred the chemical abundances of oxygen, nitrogen, and iron in 195 individual star-forming galaxies at $z\sim2-3$. Self-consistent measurements of O/H, N/H, and Fe/H were made using \texttt{GalDNA} (Section~\ref{sec:galdna}), which uses BPASS stellar population synthesis models together with Cloudy photoionization models to predict the rest-optical nebular spectra of galaxies. These results comprise the largest sample of high-$z$ galaxies with multiple elemental abundances to date and can be used to address questions related to the chemical evolution of galaxies in the early universe. Our key findings are that:
\begin{enumerate}
\item The photoionized gas and massive stars in $z\sim2$ galaxies have sub-solar abundances of O, N, and Fe, which increase with increasing M$_{\ast}$. The fiducial abundances at M$_{\ast}=10^{10}$~M$_{\odot}$ are
\begin{eqnarray}
12+\log(\textrm{O/H}) = 8.35&\pm&0.02\textrm{ or }45\%~ (\textrm{O/H})_{\odot}\nonumber \\
12+\log(\textrm{N/H}) = 7.07&\pm&0.03\textrm{ or }17\%~ (\textrm{N/H})_{\odot}\nonumber \\
\textrm{[Fe/H]} = -0.69&\pm&0.02\textrm{ or }20\%~ (\textrm{Fe/H})_{\odot}\nonumber
\end{eqnarray}
\item The O-MZR based on the \texttt{GalDNA} method is relatively shallow ($\beta_{\textrm{O/H}}=0.14\pm0.05$), with moderate intrinsic scatter ($\sigma_{\textrm{O/H}}=0.21\pm0.01$~dex, Figure~\ref{fig:omzr}). Although qualitatively similar to other O-MZRs in the literature, it is shallower and has larger scatter than relations based on strong-line O/H.
\item The slope of the O3N2 O-MZR (and other strong-line-based relations) depends on the coefficient(s) in the calibration (Figure~\ref{fig:omzr_slope}), which in turn depend on underlying correlations among abundances and other physical conditions like ionization parameter in the calibration sample. In other words, the correlation between $U$ and O/H in the \emph{calibration sample} affects the slope of the O-MZR for \emph{any other sample} for which that calibration is used. This can introduce systematic biases in physical interpretations of the form of such scaling relations.
\item The larger intrinsic scatter observed in the \texttt{GalDNA} O-MZR is the result of explicitly accounting for the ionization conditions in galaxies, which are strongly correlated with many line ratios and with M$_{\ast}$, but not as strongly correlated with O/H. Conversely, the strong correlation between $U$ and many commonly-used line ratios used to infer O/H can artificially suppress the scatter in strong-line-based O-MZRs (Figure~\ref{fig:omzr_intsc}).
\item The O-MZR and Fe-MZR (Figure~\ref{fig:femzr}) have nearly identical slopes across two decades in $M_{\ast}$, indicating a high level of $\alpha$-enhancement equivalent to $\textrm{O/Fe}\approx2.2\times(\textrm{O/Fe})_{\odot}$ even at relatively large M$_{\ast}$ (Figure~\ref{fig:mzrcomp}).
\item The intrinsic scatter in the O-MZR and Fe-MZR is very similar (Figure~\ref{fig:intsc}), which would be the case if the majority of the Fe had been produced along with the O by CCSNe, rather than by Type Ia SNe. The N-MZR has significantly larger scatter, potentially reflecting a longer timescale for nitrogen production.
\end{enumerate}
Our results demonstrate the potential to measure multiple elemental abundances in $z\sim2-3$ galaxies using emission lines that are commonly observed in their rest-optical spectra. Knowledge of galaxies abundance patterns---rather than just a single measure of their enrichment---should prove valuable in comparing galaxies' interstellar medium with their circumgalactic medium to understand gas flows in and out of galaxies. Moreover, abundance patterns can be used to link galaxy populations observed at different redshifts with one another by looking for commonalities in, for example, $\alpha$-enhancement.
Methods like \texttt{GalDNA} capable of accurately characterizing the slope \emph{and} scatter in abundance scaling relations for multiple elements can also be adapted for samples at other redshifts where different emission lines may be accessible (e.g., in the near-UV). Continuing to develop such methods will be key to extracting maximum scientific value from both existing galaxy surveys and observations from future facilities, including JWST, which will provide new opportunities to study some of the earliest galaxies in the universe.
\begin{acknowledgements}
The data presented in this paper were obtained at the W.M. Keck Observatory, which is operated as a scientific partnership among the California Institute of Technology, the University of California, and the National Aeronautics and Space Administration. The Observatory was made possible by the generous financial support of the W.M. Keck Foundation. This work has also been supported in part by the US National Science Foundation (NSF) through grants AST-0908805 and AST-1313472 (ALS and CCS). The authors wish to recognize and acknowledge the significant cultural role and reverence that the summit of Maunakea has within the indigenous Hawaiian community. We are privileged to have the opportunity to conduct observations from this mountain.
\end{acknowledgements}
\facilities{Keck:I (LRIS, MOSFIRE), Magellan:Baade (FourStar), Hale (WIRC), Spitzer (IRAC)}
\software{\texttt{GalDNA}, \texttt{LINMIX\_ERR} \citep{kelly2007}, BPASSv2 \citep{eldridge2016,stanway2016}}
|
\section{Introduction}
Scoring rules provide a principled approach to form and evaluate probabilistic predictions.
The earliest applications go at least back to the evaluation of weather forecasts \cite{brier1950verification}, but have since then developed into a rich theoretical framework that plays a central part in modern statistical inference.
We refer to \cite{gneiting2007strictly} and \cite{dawid2007geometry} for a general background.
The theoretical underpinnings of scoring rules are well-developed, but nearly all of the literature focuses on prediction of vector-valued or scalar-valued quantities.
The aim of this article is develop a scoring rule framework for an important class of non-Euclidean data, namely sequential data -- both in discrete and continuous time.
\paragraph{The Drawbacks of (Na\"ive) Vectorization.} Given a dataset consisting of multi-variate time series (TS), a common approach is to flatten each TS into a long vector and then use a standard pipeline for vector-valued data.
However, this approach has several drawbacks.
Firstly, there's trouble whenever the different TS are irregularly sampled or are of different length since this embeds the different TS in Euclidean spaces of different dimension.
Typically this addressed by ad-hoc approaches such as adding synthetic data by interpolation or dropping data points.
Secondly, often the relevant information is independent of the time-parametrization (``\emph{time-warping invariance}''), at least to a large degree; for instance, the meaning of a spoken word or an object being filmed are both independent of how fast or slow the audio signal or video is presented.
Finally, many models are naturally formulated in continuous time rather than discrete time; for example, stochastic differential equations form a popular class of models in many applications and it is unclear how to evaluate such continuous time models in a scoring rule framework besides above naive vectorization on an arbitrary time grid.
\paragraph{A Non-Euclidean Data Domain.}
Key to our approach is that classic tools from pure mathematics faithfully capture the Non-Euclidean structure of the space of (unparametrized) paths.
While there is no linear structure that allows for addition of paths of different length, any two paths can be concatenated into one path and any path can be run backwards.
Both these operations -- \emph{concatenation and reversal} -- are independent of the choice of parametrization, hence they also apply to equivalence classes of paths under reparametrization.
We refer to an unparametrized path -- that is an equivalence class of paths under reparametrization -- as a \emph{track}, as it is defined uniquely by the track it carves out in the space where it evolves.
A classical result \cite{chen-58} is that there is a \emph{``feature map''} from the set of tracks into a linear space that is \emph{functorial} and \emph{universal}; the former means that operations on tracks (concatenation and reversal) turn into algebraic operations in feature space, the latter means that any function of tracks can be approximated as a linear functional of this map.
Moreover, this map is given as a series of iterated integrals which makes it amenable to computation and we refer to it as the \emph{signature map}.
In fact, the co-domain of this feature map (``\emph{the feature space''}) is not only a linear space but forms a so-called Hopf algebra and it is the Hopf algebra structure that captures operations on tracks as algebraic operations.
The third mathematical ingredient that we use are gradients of functions of
tracks: the usual definition of linear (Fr\'echet) differentiability can sometimes be unsuitable for such functions due to the above lack of linear structure.
However, Pansu generalized classical differentiation to a special class of groups and we leverage this to define \emph{gradients of functions of (unparametrized) paths}.
We show that the usual guarantees of gradient descent algorithms apply which allows us to compute quantities associated with our scoring rule framework (as even in the case of Euclidean data, many quantities are not given in closed form, but can be found by first order methods).
\paragraph{Outline.}
Section~\ref{sec:scoring} recalls the basic definitions and general of scoring rules.
Section~\ref{sec: structure of tracks} contains the theoretical background; it formalizes the structure of the spaces of (unparametrized paths), and introduces the signature feature map and the Hopf algebra structure of its co-domain (the feature space).
Section \ref{sec:scoring tracks} then shows how these quantities lead to natural scoring rules on the non-Euclidean space of tracks; in particular the so-called antipode of the Hopf algebra plays key role to relate the scoring rule framework to structural properties of tracks.
From general results this then immediately leads to definitions of entropy, divergence, and mutual information that -- unlike the (na\"ive) vectorization approach outlined above -- are compatible with the structure of (probability measures) on spaces of (unparametrized) paths.
Section~\ref{sec:gradients} then utilizes that one may identify a track as ``group-like'' element and shows that the concept of Pansu differentiabilty leads to a natural notion of gradient descent on the space of paths resp. tracks.
Finally, Section~\ref{sec:experiments} demonstrates that despite this approach being motivated by pure maths, the resulting quantities lead to efficiently computable quantities that have some advantageous properties compared to other methods with similar invariances.
\subsection{Related Work}
One of the earliest empirical insights for time series data was that time-parametrization (``time warping'') invariance is of key importance \cite{SakoeChiba71, 1163055}.
Arguably the most popular way to address this invariance is via the classical dynamic time warping distance (DTW) and its many variations that introduce a distance between time series by searching over time changes.
For example, \cite{Cuturi2017SoftDTWAD, Blondel2021DifferentiableDB} introduce a regularised version of DTW, so-called soft DTW (s-DTW) which addresses the fact that the DTW distance is not differentiable, making it viable fo use in deep learning pipelines. In the process of doing so it loses the invariance that DTW enjoys and introduces a trade-off of smoothness versus invariance.
A more general point is that DTW and its variations do not aim to provide the full forecasting framework of scoring rules (divergence between measures, entropy, mutual information of TS) and although DTW approaches successfully deal with time-parametrization, they ignore other structural properties such as concatenation and reversal of TS.
Another drawback is that while the focus of DTW is on discrete time it can be formulated in continuous time (so-called Fr\'echet distance) but the computation scales with quadratic complexity in the number of sequence entries which makes it too expensive for many sources of high-frequency data, whereas our distance can be computed in linear time for the price of higher complexity in the state space dimension.
Ultimately the reason for this increase in efficiency is that the time-warpings are never explicitly computed or exhibited.
Another area that is directly related is kernel learning.
Any kernel $k:\mathcal{X} \times \mathcal{X} \to [0,\infty)$ with reproducing kernel Hilbert space $H_k$ induces a scoring rule by setting $s(x,\mu)\coloneqq \| \delta_x - \mu\|_{H_k}^2$ where $\|\mu\|_{H_k}^2 = \int \int k(x,y) \mu(dx)\mu(dx)$, see \cite[Section 5]{Gneiting2007StrictlyPS} and several kernels for sequences have been developed in the literature.
Most relevant to our approach is the ``signature kernel'' introduced in \cite{2016arXiv160108169K}.
However, for any scoring rule given by a kernel, the (generalized) divergence becomes simply the maximum mean discrepancy and the entropy simply the variance in the RKHS $H_k$.
While kernels give rise to a powerful class of scoring rules, the success and popularity of non-kernel based scoring rules on $\mathcal{X}=\mathbb{R}^n$ is motivation enough to look for other interesting, non-linear scoring rules.
The technical key to our approach comes from mathematics where iterated integrals, so-called signatures, and non-commutative algebras are use to represent paths.
This goes at least back to seminal work of Chen~\cite{chen-58} in algebraic topology and subsequent applications in control theory \cite{fliess1981fonctionnelles, RogerW1976167} and more recently rough path theory \cite{MR2314753}.
These results have been influential in stochastic analysis \cite{lyons-qian-02,friz2014course} and only more recently have been started to be explored in a statistical and machine learning context.
We mention pars-pro-toto \cite{papavasiliou2011parameter,ChevyrevOberhauser18, distrReg} for inference about laws of stochastic processes; \cite{2016arXiv160108169K,salvi2021higher,fermanian2021framing} for kernel learning; \cite{toth2020bayesian,dyer2021approximate,lemercier2021siggpde} for Bayesian approaches; \cite{ni2020conditional,buhler2020datadriven,kidger2021neural} for generative modelling; \cite{giusti2021signatures,lee2020path,Chevyrev2018PersistencePA} for applications in topology; \cite{diehl2020generalized,Diehl_2020,toth2021seqtens} for algebraic perspectives.
Finally, we mention that the two topics that are central to us -- invariances and non-Euclidean structure -- have been considered in different contexts in scoring rule frameworks.
For example, \cite{Fissler2017OrderSensitivityAE} studies equi- and in-variances for scoring rules for Euclidean data; non-vector valued data such as sets, contours, intervals, and quantiles have received attention \cite{bolin2015excursion,molchanov2005theory, fissler2021forecast}.
\section{Proper Scoring Rules, Entropies, and Divergences} \label{sec:scoring}
We briefly recall general background on scoring rules following closely the notation in \cite{dawid2007geometry}, see also \cite{gneiting2007strictly,Dawid2014TheoryAA,Grunwaldidid2004,GKMO2018}.
Let $ \mathcal{X} $ be a measurable space (\emph{the outcome space}), $ \mathcal{A} $ be a set (\emph{the action space}), and $ \mathcal{L} : \mathcal{X} \times \mathcal{A} \to \mathbb{R} $ be a function (\emph{the loss function}).
Further, let $X$ be a $\mathcal{X}$-valued random variable.
We consider the following game between a Decision-maker and Nature: the task of the decision maker is to choose an action $a \in \mathcal{A}$, after which Nature
reveals the outcome $x \in \mathcal{X}$ that is given by sampling $X$.
The decision maker then suffers the loss $\mathcal{L}(x,a)$.
Given a set $\mathcal{P}$ of probability measures on $\mathcal{X}$, a principled probabilistic (Bayesian) approach to this decision problem is to proceed as follows:
\begin{enumerate}[(I)]
\item \label{step: bayes act}Associate with every $\mu \in \mathcal{P}$ its \emph{Bayes act} $ a_\mu \in \mathcal{A} $ defined by
\begin{align}
a_\mu \coloneqq \arg\min_{a\in\mathcal{A}} \mathbb{E}_{X\sim \mu}[ \mathcal{L}(X,a)]
\end{align}
(assuming that a minimum exists; if it is not unique, choose $a_\mu$ arbitrary among the minimizers).
\item\label{step: scoring rule} Use the Bayes act $a_\mu$ to define the \emph{scoring rule} $s:\mathcal{X} \times \mathcal{P} \to \mathbb{R}$ on $ \mathcal{X} $ given by
\begin{align}\label{def:scoring rule}
s(x,\mu) \coloneqq \mathcal{L}(x,a_\mu).
\end{align}
\item \label{step: entropy}Use the scoring rule $s$ to define the (generalised) \emph{entropy} $H$, the (generalised) \emph{divergence} $d: \mathcal{P} \times \mathcal{P} \to \mathbb{R}$, and the (generalised) \emph{mutual information} $I: \mathcal{P} \times \mathcal{Q} \to \mathbb{R}$ as
\begin{align}
H: \mathcal{P} &\to \mathbb{R},\quad \mu \mapsto \mathbb{E}_{X\sim \mu}[ s(X,\mu)],\\
d: \mathcal{P} \times \mathcal{P} &\to \mathbb{R},\quad (\nu,\mu) \mapsto \mathbb{E}_{X\sim \nu}[ S(X,\mu)]-H(\nu),\\
I: \mathcal{P} \times \mathcal{Q} &\to \mathbb{R},\quad (\mu,\nu) \mapsto H(\mu) - \mathbb{E}_{U \sim \nu}[ H(\mu|U)].
\end{align}
where $\mathcal{Q}$ denotes a space of probability measures (not necessarily on $\mathcal{X}$) and $\mu|U$ denotes the law of $\mu$ conditioned on the random variable $U$.
\end{enumerate}%
The above definitions and nomenclature is justified as follows:
firstly, it is an instructive exercise to check that for standard choices of state space $\mathcal{X}$, action space $\mathcal{A}$, and loss function $\mathcal{L}$, the above reduce to classical definitions of entropy, divergence and mutual information (e.g.~if $\mathcal{A}$ is the set of densities on $\mathbb{R}^n$ then the \emph{log score} $L(x,a) \coloneqq -\log a(x)$ from \cite{good52} yields the the usual Shannon entropy, Kullback-Leibler divergence, and mutual information); see~\cite{dawid2007geometry} for more examples.
Secondly, Theorem \ref{thm:scoring} below shows that characteristic properties hold in the full generality of the above setup:
\begin{theorem}[\cite{dawid2007geometry}]\label{thm:scoring}
Let $\mathcal{X}$, $\mathcal{A}$, and $L$ be as above.
Further, denote with $H$, $I$, and $d$ the associated (generalized) scoring rule, entropy, and divergence.
Then
\begin{enumerate}
\item\label{itm:proper} the scoring rule \eqref{def:scoring rule} is \emph{proper}, that is
\begin{align}
\mu \mapsto \mathbb{E}_{X \sim \nu}[s(X,\mu)]
\end{align}
is minimized at $\mu=\nu$.
\item $ \mu \mapsto H(\mu) $ is concave,
\item $(\nu, \mu) \mapsto \mathbb{E}_{X \sim \nu}[s(X,\mu)] $ is affine in $ \nu $ for every $\mu$,
\item $ d(\mu,\nu) \geq 0 $ with equality for $ \mu=\nu $,
\item $I(\mu,\nu) \ge 0$ with equality for $\mu \perp \!\!\! \perp \nu$.
\end{enumerate}
\end{theorem}
Different applications areas demand different scoring rules.
Classical choices for the Euclidean case $\mathcal{X}=\mathbb{R}^n$ are besides the already mentioned log score, the Brier score, the Tsallis score, the Bregman score, the Hyv\"arinen score, etc.; see~\cite{dawid2007geometry} for details.
The aim of the remainder of this article is to study the case when $\mathcal{X}$ is a space of paths or a space of equivalence classes of paths (under reparametrisation).
\paragraph{A Toy Example: From Feature Maps to Bayes Actions.}
To motivate our scoring rule for paths let us first revisit the vector-valued case, $\mathcal{X}= \mathbb{R}^n$.
To arrive at a proper scoring rule, the space of Bayes actions $\mathcal{A}$ should be large enough to characterize any (sufficiently nice) probability measures on $\mathcal{X}=\mathbb{R}^n$.
A classic way to characterize a probability measure, is to consider the sequence of moments,
\begin{align}\label{eq:moments}
(1, \mathbb{E}[X], \mathbb{E}[X^{\otimes 2}], \mathbb{E}[X^{\otimes 3}],\ldots)
\end{align}
that is, $\mathbb{E}[X]$ is the mean vector, $\mathbb{E}[X^{\otimes 2}]$ is the covariance matrix, etc.
(We tacitly assume that the sequence of moments is well-defined and decays quickly enough so that it characterizes the measure, see Remark \ref{rem:moments characteristic}).
The sequence \eqref{eq:moments} is an element of the set
\begin{align}\label{eq:define tensor algebra}
\mathcal{H} \coloneqq \prod_{m \ge 0} (\mathbb{R}^n)^{\otimes m}
\end{align}
of sequences of tensor of increasing degree $m$.
In fact, this set $\mathcal{H}$ forms a vector space by element-wise addition of tensors of the same degree.
If we define the ``feature map''
\begin{align}
\varphi: \mathbb{R}^n \to \mathcal{H}, \quad x \mapsto (1, x, x^{\otimes 2}, \ldots).
\end{align}
With the above notation, the moment sequence \eqref{eq:moments} is simply the mean of $\varphi(X)$,
\begin{align}\label{eq:expected feature}
\mathbb{E}[\varphi(X)] = (1, \mathbb{E}[X], \mathbb{E}[X^{\otimes 2}], \mathbb{E}[X^{\otimes 3}],\ldots) \in \mathcal{H}
\end{align}
Using a well-known characterization of the mean as minimizer of a quadratic
we can introduce the loss function
\begin{align}
\mathcal{L}(x,a) \coloneqq \| \varphi(X) - a\|^2
\end{align}
which associates with a probability measure $\mu$ on $\mathbb{R}^n$ the Bayes action
\begin{align}
a_\mu \equiv \mathbb{E}[\varphi(X)] \equiv \underset{m \in \mathcal{H}}{\operatorname{argmin}} \mathbb{E}[ \mathcal{L}(X,a)].
\end{align}
It follows from general principles that the resulting scoring rule on the state $\mathcal{X}=\mathbb{R}^n$ and action space $\mathcal{A} = \mathcal{H}$,
\begin{align}
s(x,\mu) = \mathcal{L}(x,a_\mu)
\end{align}
is proper.
Despite the elementary nature of this example it gives us a simple way to associate with any ``feature map'' a Bayes action and a scoring rule and already simple variations lead to interesting questions: for example, if the quadratic loss function is replaced by the absolute value, one ends with medians of moment as Bayes action and many other choices are possible.
Such questions fall under the framework of ``elicitation'' of properties of probability measures with scoring rules which is an active research area, already in the classical vector-valued (even scalar) case; see \cite{savage1971elicitation} and \cite{Abernethy2012ACO,Steinwart2014ElicitationAI,Frongillo2015VectorValuedPE} for some of the recent advances.
\begin{remark}\label{rem:moments characteristic}
The question which probability measures on $\mathbb{R}^n$ are characterized by the moment sequence \eqref{eq:moments} is classical but quite subtle in general.
But for compactly supported measures this trivially holds.
Our focus will soon shift to probability measures on pathspace but since spaces of paths are generically not even locally compact, compactness is a too strong assumption; in fact, important examples of measures on pathspace such as geometric Brownian motion are not characterized by their ``signature moments'' that we will use in Section~\ref{sec: structure of tracks} and Section~\ref{sec:scoring tracks}.
However, one can replace the moment sequence~\eqref{eq:moments} by a normalized moment sequence that characterizes any probability measure on $\mathbb{R}^n$ and this extends to path space and signature moments, see \cite{ChevyrevOberhauser18} for details.
Hence, for simplicity, we assume throughout that the probability measures are characterized by their expected feature map (since this is possible by a slight modification of the feature map).
\end{remark}
\section{Structure of the Space of (Unparametrized) Paths}\label{sec: structure of tracks}
We review classic mathematical results about spaces of paths going back to seminal work of Chen \cite{chen-58}.
The main result is the existence of a \emph{``feature map''}
\begin{align}\label{eq:signature}
\Phi: \operatorname{Tracks} \to \mathcal{H},\,\quad\mathbf{x} \mapsto \Phi(\mathbf{x})
\end{align}
that has as domain the set $\operatorname{Tracks}$ that consist of equivalence classes of paths that evolve in $\mathbb{R}^n$, and as co-domain the linear space $\mathcal{H}$.
We already encountered $\mathcal{H}$ in Section~\ref{sec:scoring} where it arose as the vector space of sequences of tensors $\mathbf{t}_m \in (\mathbb{R}^n)^{\otimes m}$ of increasing degree $m$, see \eqref{eq:define tensor algebra}.
However, $\mathcal{H}$ is not only a vector space but a so-called Hopf algebra: we can multiply elements of $\mathcal{H}$ and take the ``inverse'' of elements of $\mathcal{H}$.
One of the well-known and attractive properties of the map~\eqref{eq:signature} is that these two algebraic operations (multiplication and inversion) capture the natural operations on $\operatorname{Tracks}$ (conatentation and reversal).
Exploiting this correspondence will be essential for our main results in Section~\ref{sec:scoring tracks}.
\paragraph{The Domain of Paths.}
A bounded variation path\footnote{For simplicity we focus on (equivalence classes of) bounded variation paths but all the results immediately extend to paths with much less regularity such as trajectories of stochastic differential equations or (fractional) Brownian motion by replacing the iterated Riemann--Stieltjes integrals by stochastic integrals or rough path integrals \cite{MR2314753}} $\mathbf{x}$ in $\mathbb{R}^n$ is a continuous map
\begin{align}
\mathbf{x} : [0,T] \to \mathbb{R}^n\, \text{ such that } \|\mathbf{x}\| \coloneqq \sup_{(t_1,\ldots,t_L): 0 \le t_1 < \ldots < t_L < T} \sum \| \mathbf{x}(t_{i+1})- \mathbf{x}(t_i) \| < \infty.
\end{align}
For $a,b \in \mathbb{R}^n$ we denote with $\paths{a}{b}$ the set of all continuous bounded variation paths that start at $a$ and end in $b$,
\begin{align}
\paths{a}{b} \coloneqq \{\mathbf{x} | \mathbf{x}:[0,T] \to \mathbb{R}^n\text{ is of bounded variation, } T>0, \mathbf{x}(0)=a, \mathbf{x}(T)=b\}.
\end{align}
and by
\begin{align}
\operatorname{Paths} \coloneqq \bigcup_{a,b \in \mathbb{R}^n} \paths{a}{b}
\end{align}
the set of all bounded variation paths in $\mathbb{R}^n$.
Although $\operatorname{Paths}$ is not a linear space, it has a rich structure given by concatenation and time reversal.
Informally, this says that if one can go from $a$ to $b$ and from $b$ to $c$ then one can go from $a$ to $c$ and that if one can go from $a$ to $b$ then one can go from $b$ to $a$.
Formally, concatenation and reversal are defined as
\begin{enumerate}
\item For $\mathbf{x} \in \paths{a}{b}$, $\mathbf{y} \in \paths{b}{c}$, their \emph{concatenation} $\mathbf{x} \star \mathbf{y} \in \paths{a}{c}$ is defined by
\begin{align}
(\mathbf{x} \star \mathbf{y})(t) \coloneqq
\begin{cases}
\mathbf{x}(t) \text{, if } t \in [a,b]\\
\mathbf{x}(b)-\mathbf{y}(b)+\mathbf{y}(t) \text{, if } t \in [b,b + c]\\
\end{cases}
\end{align}
\item for any $ \mathbf{x}\in \paths{a}{b}$ there exists an \emph{inverse} path $\overleftarrow \mathbf{x} \in \paths{b}{a}$ defined as
\begin{align}
\overleftarrow\mathbf{x}(t) \coloneqq \mathbf{x}( b - t).
\end{align}
\end{enumerate}
\paragraph{The Domain of Tracks}
As discussed above, often we want to ignore the time parametrization, hence the fundamental object we care about is not the set of paths but equivalence classes of paths.
It turns out that is useful to work with slightly more general equivalence relation, namely that of \emph{tree-like equivalence} $\sim$.
We define
\begin{align}
\tracks{a}{b} \coloneqq \paths{a}{b} / {\sim} \text{, and } \operatorname{Tracks}\coloneqq \operatorname{Paths} / {\sim}
\end{align}
With slight abuse of notation, we use the same notation $\mathbf{x}$ for an element of $\operatorname{Paths}$ and an element of $\operatorname{Tracks}$ but emphasize that an element $\mathbf{x} \in \operatorname{Tracks}$ is a whole equivalence class of paths.
We give the precise definition of the equivalence relation $\sim$ in Appendix \ref{app:TLE} and only note here that if two paths $\mathbf{x}:[0,T] \to \mathbb{R}^n,\mathbf{y}:[0,S] \to \mathbb{R}^n$ differ by time-parametrization, that is $\mathbf{x}(t)=\mathbf{y}(\varphi(t))$ for every $t$ and an increasing function $\varphi:[0,T] \to [0,S]$, then $\mathbf{x} \sim \mathbf{y}$.
However, in addition to time parametrization, tree-like equivalence also identifies paths that backtrack all their excursions, see Appendix \ref{app:TLE}.
We invite readers to think of elements of $\operatorname{Tracks}$ like animal tracks in nature: they provide shape and direction but not the speed at which the track was made.
In particular, we note that the above operations of concatenation and reversal are well-defined for the elements of $\operatorname{Tracks}$; after all, they do not depend on the time-parametrization.
So again, with slight abuse of notation we have concatenation and reversal map,
\begin{align}
\star : \tracks{a}{b} \times \tracks{b}{c} \to \tracks{a}{c} \text{ and } \overleftarrow{\bullet}: \tracks{a}{b} \to \tracks{b}{a}.
\end{align}
\paragraph{The co-domain $\mathcal{H}$.}
Our first encounter of $\mathcal{H}$ was in Section~\ref{sec:scoring} as the state space of the moment map~\eqref{eq:define tensor algebra}.
However, a more abstract way to introduce is by identifying it as the free algebra over $\mathbb{R}^n$.
Informally, this means we want to keep the vector space structure of $\mathbb{R}^n$ but we also would like to have a multiplication and do this in the most general way possible.
Formally, this means $\mathcal{H}$ is the free algebra over $\mathbb{R}^n$.
Despite this abstract characterization as a free object, the space $\mathcal{H}_n$ has a very concrete form which we will take as its definition,
\begin{align}
\mathcal{H} \coloneqq \prod_{m \geq 0} (\mathbb{R}^n)^{\otimes m} \coloneqq \{ \mathbf{t}=(\mathbf{t}_0,\mathbf{t}_1, \mathbf{t}_2,\mathbf{t}_3, \ldots) : \mathbf{t}_m \in (\mathbb{R}^n)^{\otimes m} \}.
\end{align}
(one can then directly verify that this indeed is the free algebra, see~\cite{reutenauer-93}.
That is, an element $\mathbf{t}$ of $\mathcal{H}$ is sequence of tensors $(\mathbf{t}_0,\mathbf{t}_1,\mathbf{t}_2,\ldots)$ of increasing degree $m$ where by convention $(\mathbb{R}^n)^{\otimes 0} = \mathbb{R} $.
The vector space structure of $\mathcal{H}$ is simply given as element-wise addition: addition of $\mathbf{s},\mathbf{t} \in \mathcal{H}$ is defined as
\begin{align}
\mathbf{s} + \mathbf{t} = (\mathbf{s}_0 + \mathbf{t}_0, \mathbf{s}_1 + \mathbf{t}_1, \ldots)
\end{align}
and their multiplication is defined by $ (\mathbf{s} \cdot \mathbf{t})_m = \sum_k \mathbf{s}_k \otimes \mathbf{t}_{m-k} $, i.e
\begin{align} \label{eq:multiplication}
\mathbf{s} \cdot \mathbf{t} \coloneqq (1, \mathbf{s}_1+ \mathbf{t}_1, \mathbf{s}_2 + \mathbf{s}_1 \otimes \mathbf{t}_1 + \mathbf{t}_2, \ldots)
\end{align}
where $\otimes$ denotes the usual tensor (outer) product.
Like matrix multiplication, this multiplication is associative but in general not commutative, $\mathbf{s} \cdot \mathbf{t} \neq \mathbf{t} \cdot \mathbf{s}$ and it has as multiplicative unit $(1,0,\ldots) \in \mathcal{H}$,
\begin{align}
\mathbf{t} \cdot (1,0,0,\ldots) = (1,0,\ldots) \cdot \mathbf{t} = \mathbf{t}.
\end{align}
The existence of a unit for multiplication naturally leads to the question of the existence of inverses, that is for $\mathbf{t} \in \mathcal{H}$ can one find another element in $\mathcal{H}$, denoted by $\mathbf{t}^{-1} \in \mathcal{H}$, such that
\begin{align}
\mathbf{t} \cdot \mathbf{t}^{-1} = \mathbf{t}^{-1} \cdot \mathbf{t} = (1,0,0,\ldots).
\end{align}
This is true whenever $ \mathbf{t}_0 \not= 0 $, and moreover, $ \mathbf{t} \mapsto \mathbf{t}^{-1} $ has the explicit formula
\begin{align}
\mathbf{t}^{-1} = \frac1{\mathbf{t}_0} \Big\{\sum_{m\geq 0} (1-\frac1{\mathbf{t}_0}\mathbf{t})^{\otimes m} \Big\}.
\end{align}
\paragraph{The Feature map $\Phi:\operatorname{Tracks} \to \mathcal{H}$.}
\begin{definition}\label{def:iterated integrals}
For $\mathbf{x} \in \operatorname{Paths}$, $\mathbf{x}:[0,T] \to \mathbb{R}^n$ define
\begin{align}
\int d\mathbf{x}^{\otimes m} \coloneqq \int_{0 \le t_1 <\cdots < t_m \le T} d\mathbf{x}(t_1) \otimes \cdots d \mathbf{x}(t_m) = \int \dot \mathbf{x}(t_1) \otimes \cdots \dot \mathbf{x}(t_m) dt_1 \cdots dt_m.
\end{align}
\end{definition}
It is known that if two paths $\mathbf{x},\mathbf{y} \in \operatorname{Paths}$ are tree-like equivalent, $\mathbf{x} \sim \mathbf{y}$, then $\int d\mathbf{x}^{\otimes m}= \int d\mathbf{y}^{\otimes m}$ for every $m \ge0$, see~\cite{MR2630037}.
In fact, for the case of reparametrization $\mathbf{x}(t)=\mathbf{y}(\tau(t))$ this follows immediately from the change of variables formula.
With slight abuse of notation we now define $\int d \mathbf{x}^{\otimes m}$ for $\mathbf{x}\in \operatorname{Tracks}$.
\begin{definition} For $\mathbf{x} \in \operatorname{Tracks}$, define
\begin{align}
\int d\mathbf{x}^{\otimes m} \coloneqq \int d\mathbf{x}^{\otimes m}_{\text{path}} %
\end{align}
where $\mathbf{x}_{\text{path}} \in \operatorname{Paths}$ is in the equivalence class of $\mathbf{x}$ and $\int d\mathbf{x}^{\otimes m}_{\text{path}}$ is as in Definition~\ref{def:iterated integrals}.
\end{definition}
By~\cite{MR2630037} $\int d\mathbf{x}^{\otimes m}$ for $\mathbf{x} \in \operatorname{Tracks}$ is well-defined in the sense that the choice of $\mathbf{x}_{\text{path}}$ does not matter.
We refer to the resulting map as the signature map (this is also known as the Chen--Fliess series or chronological exponential).
\begin{definition}
We call
\begin{align}
\Phi: \operatorname{Tracks} \to \mathcal{H},\quad \mathbf{x} \mapsto \left( \int d\mathbf{x}^{\otimes m} \right)_{m \ge 0}
\end{align}
the signature map.
\end{definition}
A well-known key property of the map $\Phi$ is that %
concatenation and reversal in $\operatorname{Tracks}$ correspond to multiplication and inversion in $\mathcal{H}$.
Further, the map $\Phi$ is universal (up to fixing the starting point of the track, which is why we fix the starting point $a \in \mathbb{R}^n$ and restrict to the domain $\bigcup_{b\in\mathbb{R}^n} \operatorname{Tracks}(a,b)$) in the sense that it linearizes continuous functions on $\operatorname{Tracks}$.
We summarize all this in Theorem~\ref{thm:signature} below.
\begin{theorem}\label{thm:signature}
For every $a \in \mathbb{R}^n$ the map
\begin{align}
\Phi: \bigcup_{b \in \mathbb{R}^n}\tracks{a}{b}\to \mathcal{H} , \quad \mathbf{x} \mapsto \left(\int d\mathbf{x}^{\otimes m}\right)_{m \ge 0}
\end{align}
is injective and
\begin{enumerate}
\item \label{itm:concat2}
$\Phi(\mathbf{x}\star \mathbf{y}) = \Phi(\mathbf{x}) \cdot \Phi(\mathbf{y})$
\item\label{itm:inverse}
$\Phi(\overleftarrow \mathbf{x}) = \Phi(\mathbf{x})^{-1}$
\item \label{itm:universal}for every $f \in C(\operatorname{Tracks},\mathbb{R})$, $\epsilon>0$ there exists a linear functional $\ell \in \mathcal{H}^\star$ such that
\begin{align}
| f(\mathbf{x}) - \langle \ell, \Phi(\mathbf{x}) \rangle| < \epsilon
\end{align}
uniformly in $\mathbf{x}$ on compacts.
\end{enumerate}
\end{theorem}
\begin{proof}
This is a folk theorem in algebraic topology and control theory; see \cite{chen-58} and \cite{fliess1981fonctionnelles}.
What is less standard is that we use the treelike equivalence from $\sim$ from~\cite{MR2630037}.
\end{proof}
\begin{remark}
\label{rm:compute signature}
The space $\mathcal{H} \equiv \prod_{m \ge 0} (\mathbb{R}^d)^{\otimes m}$ is graded by the tensor degree $m$, and %
$\Phi(\mathbf{x}) \equiv (\int dx^{\otimes m})_{m \ge 0}$ decays exponentially fast in $ m $, that is
\begin{align}
\|{\int dx^{\otimes m}} \|\le \frac{\|\mathbf{x} \|^m }{m!}
\end{align}
(on the right hand side $\|\bullet\|$ denotes the bounded variation (semi-)norm, on the left-hand side it denotes the norm on $(\mathbb{R}^d)^{\otimes m}$).
Hence, in practice one only needs to compute the first $m$ iterated integrals of $\Phi(\mathbf{x})$.
For piecewise linear tracks $\mathbf{x} \in \operatorname{Tracks}$ -- which is how we identify time series -- the first $m$ entries of the map $\Phi(\mathbf{x})$ can be computed in $O(L d^M)$ computational steps: if a track is given by piecewise linear segments $v_1,\ldots,v_L \in \mathbb{R}^n$ then
\begin{align} \label{eq:expsig}
\left(\int d\mathbf{x}^{\otimes m}\right)_{m \in 0,1,\ldots,M} = \exp(v_1) \cdots \exp(v_L),
\end{align}
where $\exp(v) \coloneqq (1, v, \frac{v^{\otimes 2}}{2!}, \ldots) \in \mathcal{H}$.
Hence, for a low dimensional state space, the map $\Phi(\mathbf{x})$ can be approximately in time that scales linearly in the length of the path.
\end{remark}
\paragraph{The Antipode in $\mathcal{H}$.}
The two operations of addition $\mathbf{s} + \mathbf{t}$ and multiplication $\mathbf{s} \cdot \mathbf{t} $ turn $\mathcal{H}$ into a (non-commutative) algebra $(\mathcal{H}, +, \cdot)$.
However, $ \mathcal{H} $ comes with a bit more structure, namely
the so-called \emph{antipode} map
\begin{align}
\alpha : \mathcal{H} \to \mathcal{H}
\end{align}
which is defined as the linear function given by linear extensions of the map
\begin{align}
(\mathbb{R}^n)^{\otimes m} \to (\mathbb{R}^n)^{\otimes m}, \quad v_1 \otimes \cdots v_m \otimes \mapsto (-1)^m v_m \otimes \cdots \otimes v_1.
\end{align}
There is an important subset $ G $ of $ \mathcal{H} $ defined by the property
\begin{align}
G := \{ g \in \mathcal{H} : \alpha(g) = g^{-1} \}.
\end{align}
It turns out that $ G $ in fact forms a group and will play an important role in
our Bayes acts for the simple fact that the feature map $ \Phi $ takes values in
$ G $. We summaries this along with some facts about $ \alpha $ that we use later in the following lemma.
\begin{lemma} \label{lem:antipode}
Let $\alpha$ be the antipode on $\mathcal{H}$.
Then
\begin{enumerate}
\item \label{itm:a1}$ \alpha^2 = 1 $,
\item \label{itm:a2}$ \alpha(\mathbf{s} \cdot \mathbf{t}) = \alpha (\mathbf{t} )\alpha (\mathbf{s}) $,
\item \label{itm:group}
If $ \mathbf{x} \in \operatorname{Tracks} $, then $ \Phi(\mathbf{x}) \in G $,
\item \label{itm:a5}For a power series $ p(\mathbf{t}) = \sum_{m\geq 0} p_m\mathbf{t}^{\otimes m} $, it holds that $ p\circ \alpha (\mathbf{t}) = \alpha \circ p(\mathbf{t}) $ for any $ \mathbf{t} \in \mathcal{H} $,
\item \label{itm:invertible}
If $ \mathbf{t} \in \mathcal{H} $ is invertible, then $\alpha(\mathbf{t}^{-1}) = \alpha( \mathbf{t})^{-1} $,
\item \label{itm:a4}Let $ f : \mathcal{H} \to \mathbb{R} $ have the form
\begin{align}
f(\mathbf{t}) = \sum_m \sum_{i_1 \cdots i_m} f_m(\mathbf{t}^{i_1 \cdots i_m}_m )
\end{align}
where we identify the degree-$m$ tensor $\mathbf{t}_m \in (\mathbb{R}^d)^{\otimes m}$ of $\mathbf{t}=(\mathbf{t}_m)_{m \ge 0} \in \mathcal{H}$ with its coordinates $\mathbf{t}_m \simeq (\mathbf{t}_m^{i_1,\ldots,i_m})_{i_1,\ldots,i_m=1,\ldots,d}$.
If $ f_m(x) = f_m(-x) $ for every $ m\geq 0 $ and $x\in\mathbb{R}$, then
\begin{align}
f(\mathbf{t}) = f(\alpha (\mathbf{t})).
\end{align}
\end{enumerate}
\end{lemma}
\begin{proof}
Items \ref{itm:a1} and \ref{itm:a2} follow from the definition.
Item \ref{itm:group} is well known, see for instance~\cite[Section
5]{chevyrev2016characteristic}.
To see item \ref{itm:a5} note that
\begin{align}
p\circ \alpha (x)
= \sum_{n\geq 0} p_n(\alpha x)^{\otimes n}
= \sum_{n\geq 0} p_n\alpha (x^{\otimes n})
= \alpha(\sum_{n\geq 0} p_n x^{\otimes n})
= \alpha \circ p(x),
\end{align}
by Item \ref{itm:a2} and linearity.
Item \ref{itm:invertible} follows since the inverse map $ \mathbf{t} \mapsto \mathbf{t}^{-1} $ has the power series expansion
\begin{align}
\mathbf{t}^{-1} = \frac1{\mathbf{t}_0} \Big\{\sum_{n\geq 0} (1-\frac1{\mathbf{t}_0}\mathbf{t})^{\otimes n} \Big\},
\end{align}
see \cite[Lemma 7.16]{friz-victoir-book} which together with Item \ref{itm:a5} shows the claim.
For item \ref{itm:a4}, we note that
\begin{align}
f(\alpha x)
= \sum_n \sum_{i_1 \cdots i_n} f_n((-1)^nx^{i_n \cdots i_1})
= \sum_n \sum_{i_1 \cdots i_n} f_n(x^{i_n \cdots i_1})
= f(x).
\end{align}
\end{proof}
A simple example of a function that satisfies the requirements Item \ref{itm:a4} in Lemma~\ref{lem:antipode} is the sum of squares
\begin{align}
L(\mathbf{t}) = \sum_m \sum_{i_1 \cdots i_m} \vert \mathbf{t}^{i_1 \cdots i_m}_m \vert^2
\end{align}
Which will be used to construct a loss function for tracks in Section \ref{sec:experiments}.
\paragraph{From Discrete Time to Continuous Time.}
This section has so far focused on paths [resp.~tracks], that evolves [resp.~equivalence classes of evolutions] in continuous time $\mathbf{x}: [0,T] \to \mathbb{R}^n$.
However, in practice one typically has only access to a discrete time observations $\mathbf{x}(t_1),\ldots,\mathbf{x}(t_L) \in \mathbb{R}^n$ along some grid $0 \le t_1 < \cdots < t_L \le T$, that is a time series.
But any TS can be identified as the piecewise linear path
\begin{align} \label{eq:disctime}
t \mapsto \mathbf{x}(t_i) + \frac{t_i -t}{t_{i+1}-t_i} (\mathbf{x}(t_{i+1})- \mathbf{x}(t_i)) \text{ for } t \in [t_i,t_{i+1})
\end{align}
and hence also as an element of $\operatorname{Tracks}$ after forgetting the parametrisation.
Working in continuous time when the original data is discrete might look cumbersome and unnecessary at first sight but it has several advantages.
Firstly, all TS are embedded into the same space $\operatorname{Paths}$ respectively $\operatorname{Tracks}$, even if the sample grid $t_1< \cdots t_L$ varies from TS to TS, which would not be the case if one identifies TS as vectors.
Secondly, this automatically ensures consistency in terms of high-frequency limits when the grid gets finer, that is $\sup |t_{i+1}^m-t_i^m| \to 0$ as $m \to \infty$ for a sequence $(t_i^m)_{i=1,\ldots,L_m}$.
Finally, many popular models are naturally formulated in continuous time rather than discrete time.
\paragraph{From Tracks to Paths.}
Our guiding philosophy is that the fundamental object is the set $\operatorname{Tracks}$ rather than the set $\operatorname{Paths}$ since the former allows to ignore the time parametrization; note that the set of time parametrisations is infinite-dimensional since every continuous function $\tau: [0,T] \to [0,T']$ can be used to reparametrize a path $t \mapsto \mathbf{x}(t)$ to $t \mapsto \mathbf{x}(\tau(t))$, hence working with $\operatorname{Tracks}$ factors out an infinite-dimensional class of invariances.
Nevertheless, for certain applications the parametrization matters -- at least to a certain degree.
However, this can be easily addressed by adding time as an additional coordinate: to emphasize the dimension $n$ of the state space $\mathbb{R}^n$ in which the paths evolve we write $\operatorname{Paths}_n$ (instead of just $\operatorname{Paths}$ that we used until now); similarly $\operatorname{Tracks}_n$ for the set of equivalence classes of $\operatorname{Paths}_n$.
Given $\mathbf{x} \in \operatorname{Paths}_n$ we embed
\begin{align}
\operatorname{Paths}_n \hookrightarrow \operatorname{Paths}_{n+1}, \quad \mathbf{x} \mapsto \left(t \mapsto (t, \mathbf{x}(t))\right).
\end{align}
That is a path evolving in $\mathbb{R}^n$ is turned into a path in $\mathbb{R}^{n+1}$ by simply adding an additional coordinate that is time itself.
This makes the parametrization part of the ``shape'' of the trajectory which in turn is exactly the information that distinguishes tracks, hence
\begin{align}
\operatorname{Paths}_n \hookrightarrow\operatorname{Tracks}_{n+1} .
\end{align}
This injection shows that any scoring rule for tracks induces a scoring rule for paths.
\section{Scoring Rules For Tracks and Paths}
\label{sec:scoring tracks}
Motivated by the toy example in Section~\ref{sec:scoring} with the moment feature map $\varphi$ for data in $\mathbb{R}^n$, we now follow the analogous reasoning on the non-Euclidean space of tracks by using the feature map $\Phi$ instead of $\varphi$, note that
neither the domain nor the image of $\Phi$ is a linear space as its image is the group $G$ that is embedded into the linear space $\mathcal{H}$. Recall that in Section~\ref{sec: structure of tracks} we have seen that it is exactly the group structure that captures the structure of space of tracks of concatenation and reversal.
This motivates us to
\begin{align}
\text{replace the additive inverse in } \varphi(X)-m \text{ by the group inverse to get } \Phi(\mathbf{X})m^{-1}.
\end{align}
Our first main result Theorem~\ref{thm:scoring} shows that this indeed leads to a proper scoring rule on the space of tracks and operations on $\operatorname{Tracks}$ turn into algebraic operations in decision space.
Consquences of this result are Proposition~\ref{prop:entropy on tracks} and Corollary~\ref{cor:entrev} which show how the associated entropy on the space of tracks is invariant to time-reversal and behaves under conditioning on the past.
\paragraph{A Scoring Rule for Tracks.}
We need to introduce an additional space $ \Hopf^\star $ wedged between $ \mathcal{H} $ and $ G $ defined as the space of all elements $ \mathbf{t} \in \mathcal{H} $ starting with a one, formally
\begin{align}
\Hopf^\star = \{ \mathbf{t} \in \mathcal{H} : \mathbf{t}_0 = 1 \}.
\end{align}
Unlike $ \mathcal{H} $, $ \Hopf^\star $ is not a vector space or a Hopf algebra, but it
is a group like $ G $ while also being convex as a subset of $ \mathcal{H} $ in addition to being topologically closed -- unlike the set of invertible elements of $ \mathcal{H} $.
We have the following sequence of inclusions
\begin{align}
G \subseteq \Hopf^\star \subseteq \text{ invertible elements of } \mathcal{H} \subseteq \mathcal{H}.
\end{align}
\begin{definition} \label{def:BA}
Let $ L : \mathcal{H} \to \mathbb{R} $ be convex with a unique minimum at the unit $(1,0,0,\ldots)$ of $ G $.
Define the left loss function as
\begin{align}
\mathcal{L}^\text{left} : \operatorname{Tracks} \times \mathcal{H}^\star \to [0,\infty),\quad (\mathbf{x}, m) \mapsto L(m^{-1} \Phi(\mathbf{x})).
\end{align}
Applying step~\eqref{step: bayes act} from the scoring rule framework of Section~\ref{sec:scoring}, the left Bayes' act is defined as
\begin{align}
a^\text{left}_\mu \coloneqq \underset{m \in \Hopf^\star_n}{\operatorname{argmin}}\,\mathbb{E}[\mathcal{L}^\text{left}(m,\mathbf{X})]. %
\end{align}
Applying step~\eqref{step: scoring rule} yields the proper scoring rule
\begin{align}
s^\text{left}(\mathbf{x},\mu) \coloneqq \mathbb{E}[\mathcal{L}^\text{left}(a^\text{left}_\mu, \mathbf{x})].
\end{align}
Applying step~\eqref{step: entropy} yields the (generalised) entropy,
divergence, and mutual information
\begin{align}
H^\text{left}(\mu) &\coloneqq \mathbb{E}_{\mathbf{X}\sim \mu} \mathcal{L}^\text{left}(a^\text{left}_\mu, \mathbf{X})\\
d^\text{left}(\nu,\mu) &\coloneqq \mathbb{E}_{\mathbf{X}\sim \nu}\Big[ \mathcal{L}^\text{left}(a^\text{left}_\mu, \mathbf{X})- \mathcal{L}^\text{left}(a^\text{left}_\nu, \mathbf{X}) \Big]\\
I^\text{left}(\mu,\nu) & \coloneqq H(\mu) - \mathbb{E}_{U \sim \nu}[ H(\mu|U)]
\end{align}
on the output space $\mathcal{X}= \operatorname{Tracks}$ and the action space $\mathcal{A}= \mathcal{H}^\star$.
Analogously we define the right loss function $\mathcal{L}^\text{right}(\mathbf{x},m) \coloneqq L(\Phi(\mathbf{x})m^{-1})$, right Bayes act $a^\text{right}_\mu$ and right scoring rule $s^\text{right}(\mathbf{x},\mu)$ as well as right entropy, divergence, and mutual information.
\end{definition}
The scoring rule framework of Definition~\ref{def:BA} turns operations on tracks into algebraic operations in the decision space.
\begin{theorem}\label{thm:scoring tracks}
Let the output space be $\mathcal{X}= \operatorname{Tracks}$, the action space $\mathcal{A}= \mathcal{H}^\star$ and
\begin{align}
\mathcal{L}^\text{left} : \operatorname{Tracks} \times \mathcal{H}^\star \to [0,\infty) \text{ resp. } \mathcal{L}^\text{right} : \operatorname{Tracks} \times \mathcal{H}^\star \to [0,\infty)
\end{align}
the loss functions from Definition \ref{def:BA}.
The following properties hold
\begin{enumerate} %
\item\label{itm:c1} If $ L $ is coercive, that is $ L(\mathbf{t})\to \infty $ whenever $ \lVert \mathbf{t} \rVert \to \infty $, then for any Borel measure $ \mu $ such that $ L $ is $ \mu $-integrable, both $ a^\text{left}_\mu $ and $ a^\text{right}_\mu $ exist. If $ L $ is strictly convex, then they are unique.
\item\label{itm:c2} If $ \mu = \delta_\mathbf{x} $ then $ a^\text{right}_\mu = a^\text{left}_\mu = \Phi(\mathbf{x}) $
\item\label{itm:concat} The Bayes' acts satisfy
\begin{align}
a^\text{left}_{\nu \star \mu \vert \nu} &= \nu(\Phi)a^\text{left}_{\mu\vert\nu} \\
a^\text{right}_{\mu\star\nu \vert \nu} &= a^\text{right}_{\mu\vert\nu} \nu(\Phi)
\end{align}
where $ \nu(\Phi) $ denotes the pushforward measure of $ \nu $ under $ \Phi $ and by $\mu\star \nu| \nu$ we denote the law of $\mathbf{X} \star \mathbf{Y} |\mathbf{Y}$ where $\Law{\mathbf{X}}= \mu$ and $\Law{\mathbf{Y}}=\nu$.
\item \label{itm:reversal}If $ L $ satisfies $ L(\mathbf{t}) = L(\alpha( \mathbf{t})) $, then
\begin{align}
a^\text{right}_{\overleftarrow\mu} = \alpha(a^\text{left}_\mu), \quad
a^\text{left}_{\overleftarrow\mu} = \alpha(a^\text{right}_\mu)
\end{align}
where $ \overleftarrow\mu $ denotes the measure $ \mu $ given by running samples from $\mu$ backwards in time\footnote{Formally $\mathbf{X} \sim \mu$ then $\overleftarrow \mu$ is defined as the law of $\overleftarrow \mathbf{X}$.}
\end{enumerate}
\end{theorem}
\begin{proof} We give the proofs for the right Bayes' act as the proofs for the left Bayes' act is similar.
(\ref{itm:c1})
We equip $ \mathcal{H}_n$ with its $ \ell^2 $ norm,
\begin{align}
\lVert \mathbf{t} \rVert = \sqrt{ \sum_n \vert \mathbf{t}_n \vert^2 }
\end{align}
which makes it into a separable Hilbert Space.
Fix some measure $ \mu $ on $ \mathcal{X} $ and define the map $ \psi : G_n \to \mathbb{R} $ by
\begin{align}
\psi(x) = \mathbb{E}_\mu L(\Phi(\mathbf{X})x).
\end{align}
We want to show that $ \psi $ is convex, coercive and lower semicontinuous on $ (\mathcal{H}_n, \lVert \cdot \rVert) $, as this guarantees the existence of a minimiser. This is because the unit ball of $ (\mathcal{H}_n, \lVert \cdot \rVert) $ is weakly compact, hence we could choose some weakly compact and convex set $ C $ such that $ \psi(x) > M $ outside of $ C $, and since $ \psi $ is lower semicontinuous it is also weakly lower semicontinuous, and therefore since it is convex it achieves a minimum on $ C $ which must be a global minimum. Note that its minimiser must be $ (a^\text{right}_\mu)^{-1} $. It follows that if $ \psi $ is strictly convex, then the minimiser is unique.
Note that if $ L $ is (strictly) convex, then
\begin{align}
\psi(\frac12x+\frac12y)
= \mathbb{E}_\mu L(\frac12\Phi(\mathbf{X})x+\frac12\Phi(\mathbf{X})y)
\leq\!\!(<)\,\,\, \frac12\mathbb{E}_\mu L(\Phi(\mathbf{X})x)+\frac12\mathbb{E}_\mu L(\Phi(\mathbf{X})y)
= \frac12\psi(x) + \frac12\psi(y)
\end{align}
hence $ \psi $ is also (strictly) convex.
Note that $ ( \mathcal{H}_n, \lVert \cdot \rVert ) $ is a Banach algebra, that is $ \lVert \mathbf{t} \cdot \mathbf{s} \rVert \leq \lVert \mathbf{t} \rVert \cdot \lVert \mathbf{s} \rVert $. By taking multiplicative inverses, this implies that
\begin{align}
\lVert \mathbf{t} \cdot \mathbf{s} \rVert \geq \frac1{\lVert \mathbf{s}^{-1} \rVert}\lVert \mathbf{t} \rVert.
\end{align}
for any invertible element $ \mathbf{s} $. As $ \mu $ is Borel, and $ \mathcal{H}_n $ is a Polish space, $ \mu $ is a Radon measure by \cite[Theorem 7.1.7]{Bogachev07} and we may choose a compact set $ C $ such that $ \mathcal P(\mathbf{X} \notin C) \leq \varepsilon $, and define $ M $ to be $ \sup_{X \in C} \lVert \Phi(X)^{-1} \rVert $.
Then on $ C $, $ \lVert \Phi(X)x \rVert \geq \frac1M \lVert x \rVert $, and since
\begin{align}
\psi(x) = \mathbb{E}_\mu L(\Phi(\mathbf{X})x) \geq \mathbb{E}_{\mu\vert_C} L(\Phi(\mathbf{X})x)
\end{align}
and $ L $ is coercive, so is $ \psi $.
To see that $ \psi $ is lower semicontinuous, note that for a sequence $ x_k \to x $
\begin{align}
\lim\inf_k \psi(x_k)
= \lim\inf_k \mathbb{E}_\mu L(\Phi(\mathbf{X})x_k)
\geq \mathbb{E}_\mu L(\Phi(\mathbf{X})x)
= \psi(x)
\end{align}
by Fatous Lemma, the assertion follows.
(\ref{itm:c2})
Since $ L $ is minimised at the unit it is clear that for $ \mu = \delta_\mathbf{x} $, $ a^\text{left}_\mu = a^\text{right}_\mu = \Phi(\mathbf{x}) $ is optimal since $ (a^\text{left}_\mu)^{-1}\Phi(\mathbf{x}) = \Phi(\mathbf{x})(a^\text{right}_\mu)^{-1} = 1 $.
(\ref{itm:concat}) For $ \mathbf{Y} \sim \nu $ we have
\begin{align}
a^\text{right}_{\mu\star\nu \vert \nu} \coloneqq
&\underset{m \in \TAf{\mathbb{R}^n}}{\operatorname{argmin}}\mathbb{E}_{\mathbf{X}\sim\mu}\big[L(\Phi(\mathbf{X}\star \mathbf{Y})m^{-1}) \vert \mathbf{Y} \big] = \\
&\underset{m \in \TAf{\mathbb{R}^n}}{\operatorname{argmin}}\mathbb{E}_{\mathbf{X}\sim\mu}\big[L(\Phi(\mathbf{X})\Phi(\mathbf{Y})m^{-1}) \vert \mathbf{Y} \big] = \\
&\underset{m\Phi(\mathbf{Y})^{-1} \in \TAf{\mathbb{R}^n}}{\operatorname{argmin}}\mathbb{E}_{\mathbf{X}\sim\mu}\big[L(\Phi(\mathbf{X})m^{-1}) \vert \mathbf{Y} \big] \coloneqq
a^\text{right}_{\mu \vert \mathbf{Y}} \Phi(\mathbf{Y}).
\end{align}
(\ref{itm:reversal})
If $ L $ satisfies $ L(\mathbf{t}) = L(\alpha( \mathbf{t})) $, then
\begin{align}
a^\text{right}_{\overleftarrow\mu} \coloneqq
&\underset{m \in \TAf{\mathbb{R}^n}}{\operatorname{argmin}}\mathbb{E}_{\mathbf{X}\sim \overleftarrow\mu}L(\Phi(\mathbf{X})m^{-1}) = \\
&\underset{m \in \TAf{\mathbb{R}^n}}{\operatorname{argmin}}\mathbb{E}_{\mathbf{X}\sim \mu}L(\alpha(\Phi(\mathbf{X}))m^{-1}) = \\
&\underset{m \in \TAf{\mathbb{R}^n}}{\operatorname{argmin}}\mathbb{E}_{\mathbf{X}\sim \mu}L(\alpha(m^{-1})\Phi(\mathbf{X})) \coloneqq
\alpha(a^\text{left}_{\mu}).
\end{align}
\end{proof}
\paragraph{Entropy, Divergence, and Mutual Information on the Space of Tracks.}
We now focus on the (generalized) entropy, divergence and mutual information for probability measures on tracks that results from Definition~\ref{def:BA}.
\begin{proposition}\label{prop:entropy on tracks}
For any two probability measures $\mu$ and $\nu$ on $ \operatorname{Tracks}$ it holds that %
\begin{enumerate}
\item \label{itm:b1}$ H^{\text{left}}(\mu \star \nu \vert \mu) = H^\text{left}(\nu \vert \mu) $ and $ H^{\text{right}}(\mu \star \nu \vert \nu) = H^\text{right}(\mu \vert \nu) $
\item \label{itm:b2}If $ L $ satisfies $ L(\mathbf{t}) = L(\alpha \mathbf{t}) $, then
\begin{align}
H^{\text{right}}(\mu) &= H^{\text{left}}(\overleftarrow\mu), \\
d^{\text{right}}(\nu,\mu) &= d^{\text{left}}(\overleftarrow\nu,\overleftarrow\mu), \\
I^{\text{right}}(\mu,\nu) &= I^{\text{left}}(\overleftarrow\mu,\nu)
\end{align}
\end{enumerate}
\end{proposition}
\begin{proof}
(\ref{itm:b1}) \begin{align}
H^{\text{left}}(\mu \star \nu \vert \mu) &= \mathbb{E}_{X \sim \mu \star \nu \vert \mu} L((a^\text{left}_{\mu \star \nu \vert \mu})^{-1}\Phi(X)) \\
&= \mathbb{E}_{X \sim \nu \vert \mu} L((a^\text{left}_{\nu \vert \mu})^{-1}\nu(\Phi)^{-1}\nu(\Phi)\Phi(X))
= H^{\text{left}}(\nu \vert \mu).
\end{align}
(\ref{itm:b2}) \begin{align}
\mathbb{E}_{X \sim \nu} L(\Phi(X)(a^\text{right}_{\mu})^{-1})
= \mathbb{E}_{X \sim \nu} L(\alpha (a^\text{right}_{\mu})^{-1}\alpha\Phi(X))
= \mathbb{E}_{X \sim \overleftarrow\nu} L((a^\text{left}_{\overleftarrow\mu})^{-1}\alpha\Phi(X)).
\end{align}
The other equalities follow.
\end{proof}
\begin{corollary} \label{cor:entrev}
If $ L $ satisfies $ L(\mathbf{t}) = L(\alpha (\mathbf{t})) $ and a measure $ \mu $ is \emph{reversible}, that is, $ \mu $ and $ \overleftarrow\mu $ are equal up to their starting distribution, then
\begin{align}
H^{\text{right}}(\mu) = H^{\text{left}}(\mu).
\end{align}
\end{corollary}
In the experiments we will simulate sample paths from Brownian motion and since this is a reversible process it will not matter if we use the left- or right entropy.
\begin{remark}\label{rem:flatness}
An alternative to using the group structure in Definition~\ref{def:BA} of the Bayes act is to use that the group $G$ is embedded into the linear space $\mathcal{H}$ and use this linear structure.
That is, we define a Bayes act as $a_\mu \coloneqq \operatorname{argmin}_{m \in \TAf{\mathbb{R}^n}}\mathbb{E}_{\mathbf{X} \sim \mu}[ L(\Phi(\mathbf{X}) - m)^2]$.
It is easy to show that this gives a proper scoring rule and that $a_\mu = \mathbb{E}[\Phi(\mathbf{X})]$.
However, this scoring rule relies on the embedding of the group into its ambient vector space and does not account for or respect the group structure.
Moreover, the resulting divergence and entropy reduce to just the Euclidean distance and usual variance.
The same remark extends to (signature) kernel based scoring, where linear methods are used in an RKHS; see the discussion about non-kernel based scoring in the introduction. %
\end{remark}
\begin{remark}
We identify a stochastic process as a path- or sequence-valued random variable, possibly even ignoring its parametrization.
However, for some applications the filtration of a stochastic process matters and one could ask to extend the scoring rule framework to this.
A kernel that captures the filtration was introduced in~\cite{bonnier2021adapted} and a kernel algorithm and new applications given in \cite{salvi2021higher}.
To get a non-kernel scoring one could try to replace $\Phi$ in Definition \eqref{def:BA} by the higher-rank signature from~\cite{bonnier2021adapted}.
\end{remark}
\section{Gradient Descent on the Space of Tracks}
\label{sec:gradients}
Given a smooth function $f: \mathbb{R}^n \to \mathbb{R}$, the simplest update rule for gradient descent is
\begin{align}\label{eq:classic gradient}
x_{i+1} = x_i -\eta \nabla f(x_i)
\end{align}
and under additional regularity of $f$, the resulting sequence $(x_i)_i \subset \mathbb{R}^n$ converges to a minimizer of $f$.
Our interest lies minimizing functions $f: \operatorname{Tracks} \to \mathbb{R}$.
In accordance with our guiding theme we do not identify these domains as linear spaces where classical gradient descent can be applied.
However, we have seen that $\Phi$ provides an isomorphism between the space of tracks and the free group $G$ (up to forgetting the starting point of the track)
\begin{align}
\operatorname{Tracks} \simeq G
\end{align}
Hence, the minimization problem of a function $f$ of tracks can be re-formulated as a minimization problem of a function $F=F(g) \coloneqq f \circ \Phi^{-1}(g)$ on the free group $G$.
That is, the general problem we try to solve is to find
\begin{align}
\operatorname{argmin}_{g \in G}F(g) \text{ for }F: G \to \mathbb{R}
\end{align}
for any $F$ in class of sufficiently ``smooth'' real-valued functions on $G$.
There have been many attempts to generalize gradient descent to non-linear domains.
Arguably the the case of Riemannian manifolds~\cite{Bonnabel2013} is the most well-developed among these.
However, the group $G$ does not come with a Riemannian structure (to wit, only a Sub-Riemannian structure \cite{montgomery-02}).
We follow here a somewhat different approach inspired by work of Pierre Pansu \cite{pansu1989metriques} that directly uses the group structure to define gradients.
We show that this gradient in turn allows us to give a straightforward generalization of the gradient update rule~\eqref{eq:classic gradient} from $\mathbb{R}^n$ to $G$, so that the resulting sequence $ (g_i)_i \subset G$ converges to the minimizer.
\paragraph{Pansu Derivatives.}
The derivative $ D f (x)(\cdot)$ of a function
\begin{align}
f : \mathbb{R}^n \to \mathbb{R}
\end{align}
at a point $x$ is a linear functional of $\mathbb{R}^n$ that can be defined as the limit
\begin{align}
D f(x)(h) \coloneqq \lim_{h \to 0} \frac{f(x+\lambda h) - f(x)}{\lambda}.
\end{align}
Identifying $\mathbb{R}^n$ as a the additive group $(\mathbb{R}^n, + )$ one can regard the difference quotient that appears in the limit as applying to the group operation to $x$ and $\lambda h$.
Hence, if we have also have a generalization of the multiplication with the scalar $\lambda$, then the above difference quotient makes sense for other groups than the additive group $(\mathbb{R}^n, +)$.
To formalize scalar multiplication, it turns out that the right notion is that of a Carnot group: a Carnot group is a Lie group $C$ that carries a left-invariant geodesic distance $\operatorname{dist}:C \times C \to [0,\infty)$ and for each $\lambda >0$ a bijection
\begin{align}
\delta_\lambda : C \to C \text{ such that } \operatorname{dist}\left(\delta_{\lambda}(g), \delta_\lambda(h)\right) = \lambda \operatorname{dist}\left(g, h\right).
\end{align}
However, our focus is on the free group $G$ and here the scaling $\delta_\lambda$ by a scalar $\lambda>0$ has an explicit form
\begin{proposition}
The group $G \subset \mathcal{H}$ equipped with the geodesic distance and
\begin{align}
\delta: \lambda \times G \to G, \quad \delta_\lambda g = \delta_\lambda(1, g_1, g_2, g_3, \ldots) = (1, \lambda g_1, \lambda^2g_2, \lambda^3g_3, \ldots).
\end{align}
forms a limit of Carnot groups.
\end{proposition}
We now have all we need to define the (Pansu) derivative. We denote by $ G^\star $ the topological dual of $ G $.
\begin{definition}
Let $ f : G \to \mathbb{R} $.
We define $ Df : G \to G^\star $ as
\begin{align}
Df(g)h = \lim_{\lambda\downarrow 0} \frac{f(g\delta_\lambda h) - f(g)}{\lambda}
\end{align}
whenever this limit exists and call $Df(g)h$ the Pansu derivative of $f$ at $g$ in direction $h$.
If $f$ has a Pansu derivative for all $g \in G$ then we say that $f$ is Pansu differentiable. Analogous we define the spaces $C^k(G,\mathbb{R})$ of $k$-times Pansu differentiable functions.
\end{definition}
The Pansu derivative behaves very similar to the classic linear gradient.
For example, for the proof of convergence of gradient descent on $G$ we make use of the following ``Taylor expansion''.
\begin{lemma} \label{lem:taylor}
If $ f \in C^3(G,\mathbb{R}) $, then
\begin{align}
f(gh) = f(g) + Df(g)h_1 + \frac12 D^2f(g)h_2 + O(\lVert h_3 \rVert)
\end{align}
\end{lemma}
\begin{proof}
Consider the function $ A : \mathbb{R} \to \mathbb{R} $
\begin{align}
A(\lambda) = f(g\delta_\lambda h)
\end{align}
then $ A $ is $ C^3 $ and by a (classical linear) Taylor expansion we may write
\begin{align}
A(1) = A(0) + \dot A(0) + \frac12 \ddot A(0) + O(\lVert A^{(3)}(0) \rVert )
\end{align}
which translates into the asserted equation since $ h^{\otimes n} $ is contained in $ h_n $.
\end{proof}
\begin{remark}
A popular approach to differentiating functions of paths is to use a Fr\'echet derivative as in Malliavin calculus, i.e.~one identifies the space of paths as a linear space, see \cite{CASS2011542}.
However, the above Pansu derivative is of a very different nature and -- by construction -- respects the non-Euclidean structure of paths resp.~tracks.
\end{remark}
\paragraph{Gradient Descent on $G$.}
The idea of gradient descent to take a step in a direction that minimises $ f $ in a neighbourhood $ B_h $.
In our (Lie) group $G$, a natural choice is to choose some vector $ v \in \mathbb{R}^n $ and use an exponential neighbourhood $ g e^{\eta v} $ of $ g \in G $ where $\exp$ denotes the exponential from Lie algebra to Lie group.
Hence, the question becomes how choose $ v $ to minimise $ f(ge^{\eta v}) $.
To do so, note that
\begin{align}
f(ge^{\eta v}) = f(g\delta_\eta e^v)
\end{align}
whenever $ v \in \mathbb{R}^n $. Now using Lemma \ref{lem:taylor} shows
\begin{align}
f(g\delta_\eta e^v) = f(g) + Df(g) e^{\eta v}_1 + \eta^2 = f(g) + \eta Df(g)\cdot v + \eta^2.
\end{align}
This suggests that the direction of steepest descent in the exponential neighbourhood is indeed given by the Pansu derivative (henceforth, and with slight abuse of notation, we identify the resulting element of $ G^\star $ as an element of the Hilbert space $ \mathcal{H} $ since $G$ and $G^*$ both embed into $\mathcal{H}$), %
\begin{align}
v = - Df(g).
\end{align}
This leads to the following geometric descent rule
\begin{align} \label{eq:update}
g_{i+1} = g_i e^{ -\eta D f(g_i)}.
\end{align}
As for classic gradient descent, we need a notion of convexity to guarantee convergence to a minimum.
\begin{definition}
We say that a function $ f : G \to \mathbb{R} $ is geometrically convex if
\begin{align}
f(g\delta_\lambda (g^{-1}h)) \leq (1-\lambda)f(g) + \lambda f(h)
\end{align}
for any $ 0 \leq \lambda \leq 1 $.
\end{definition}
As in the linear case, one can show that if $ f : G \to \mathbb{R} $ is geometrically convex and $ C^2 $, then $ D^2f(g) $ is positive definite everywhere.
However, $ D^2 f $ is not symmetric in general unlike in the linear case.
Putting everything together allows us to mimic the convergence proof of gradient descent in linear spaces which ultimately justifies the above informal derivation of the update rule.
\begin{theorem}\label{thm:gradient descent}\,
\begin{enumerate}
\item \label{itm: transitive}
The geometric update rule is transitive, that is, one may go from any point in the group to any other point using updates of the form \eqref{eq:update}.
\item\label{itm: convergence}
If $ f $ is geometrically convex and bounded from below with bounded second Pansu derivatives, then for $ \eta $ sufficiently small the sequence in Equation \eqref{eq:update} converges to a minimum
\end{enumerate}
\end{theorem}
\begin{proof}
Item \ref{itm: transitive} follows from Chow's theorem~\cite{Chow1940} which states that $ G $ is generated by simple exponentials.
For Item~\ref{itm: convergence} let $ L $ be a bound on the derivatives of $ f $.
We may write
\begin{align}
f(gh) \leq f(g) + Df(g)h_1 + \frac12 L\lVert h_2\rVert.
\end{align}
Hence, if $ g_{n+1} = g_n e^{ -\eta D f(g_n)} $, then
\begin{align}
f(g_{n+1}) &\leq f(g_n) - \eta \lVert Df(g_n) \rVert^2 + \frac12 \eta^2 L\lVert Df(g_n) \rVert^2 = f(g_n) + (\frac{L}2\eta^2 - \eta)\lVert Df(g_n) \rVert^2
\end{align}
which is smaller than $ f(g_n) $ for $ \eta $ small enough whenever $ Df(g_n)\not= 0 $. Hence this is a strictly decreasing sequence and converges to a minimum.
\end{proof}
Theorem \ref{thm:gradient descent} gives the analogous convergence guarantees as regular gradient descent and is simple to implement.
\section{Experiments}\label{sec:experiments}
In all our experiments we take as loss function $ L $ the squared norm,
\begin{align}
L(\mathbf{t}) = \lVert \mathbf{t} \rVert^2 = \sum_{m=1}^M \lVert \mathbf{t}_m \rVert^2.
\end{align}
which is symmetric by Lemma \ref{lem:antipode}.
Note that the function $ L $ is smooth as a sum of squares, and since $ \Phi $ is also smooth, since it is polynomial in the increments of its input when computed up to some fixed degree $ M $, we can easily compute $ d, H, $ and $ I $ by automatic differentiation. For the computation of $\Phi$ we use the signatory~\cite{kidger2021signatory} package which allows for fast and easy computations.
\begin{remark}
In the experiments we will simulate sample paths from Brownian motion. As this is a reversible process it does not matter if entropy is computed from the left or the right by Corollary \ref{cor:entrev}, and we will normally compute it from the right.
\end{remark}
\subsection{Comparing a warped time-series to itself}
One natural question to ask is how well the regularised versions of DTW that allow for differentiation are able to incorporate the parametrisation invariance.
One drawback of these regularisation of DTW is that it typically leads to a trade-off between the smoothness and parametrization invariance.
Given two TS $\mathbf{x}$ and $\mathbf{y}$ we then use sDTW $d_{\text{sDTW}}(\mathbf{x},\mathbf{y})$ which is the regularised version of DTW introduced in \cite{Blondel2021DifferentiableDB} which is differentiable and to a certain degree parametrisation invariant quantity.
The scoring rule framework that we presented in the previous sections provides divergence not only between TS but probability measures on TS,
\begin{align}
d(\nu,\mu) &\equiv\mathbb{E}_{X\sim \nu}[ S(X,\mu)]-H(\nu),
\end{align}
but as a special case, we can restrict this to point measures to get a ``divergence'' between two TS akin to sDTW which reduces to the formula
\begin{align}\label{eq:sig divergence}
(\mathbf{x},\mathbf{y}) \mapsto d(\delta_{\mathbf{x}},\delta_{\mathbf{y}}) = L(\Phi(\mathbf{x})\Phi(\mathbf{y})^{-1}).
\end{align}
To study the different behaviour between these two ``divergences'' -- sDTW and \eqref{eq:sig divergence} -- we generated samples paths from a Brownian motion to get $\mathbf{x}$ and warped it a time change $ \varphi : [0,T] \to [0,T], \varphi(t) = T(t/T)^p $ to get $\mathbf{y}$ where $ p \in [1, \infty) $ is a parameter that determines the severity of the warping.
Finally, we sampled on a discrete time grid with resolution $ 10^{-2} $.
The results are shown in Figure \ref{fig:warp_comp}.
The geometric score stays close to $ 0 $ regardless of the value of $ p $ but the sDTW divergence will increase with $ p $ and the rate of increase depends on the $ \gamma $ parameter.
As $ \gamma $ tends to $ 0 $ the sDTW score will tend to the geometric score, but it will lose its smoothness while doing so.
It is worth noting that $ \gamma = 1 $ is considered a default value.
To achieve the same level of invariance enjoyed by the the divergence \eqref{eq:sig divergence} in a DTW paradigm on can use the classical non-smooth DTW algorithms which cannot be updated using gradient descent.
In contrast, the signature divergence \eqref{eq:sig divergence} is always differentiable and parametrization invariant.
Further, the signature divergence is more general in the sense that it is not just a divergence between TS but between probability measures on TS which allows in principle for many other applications such as variational inference.
\begin{figure}[H]
\centering
\includegraphics[width=4.1in]{warpComplog.pdf}
\caption{Geometric scoring compared to sDTW for different values of $ \gamma $ plotted against the logarithm of $ p $ for $ p $ between $ 1 $ and $ 25 $. }
\label{fig:warp_comp}
\end{figure}
The implementation of sDTW is taken from the excellent Python package from \cite{Blondel2021DifferentiableDB}.
\subsection{Mutual information}
Recall that the mutual information between two probability measures $ \mu $ and $ \nu $ is defined as
\begin{align}\label{eq:MI}
I(\mu,\nu) & \coloneqq H(\mu) - \mathbb{E}_{U \sim \nu}[ H(\mu|U)].
\end{align}
and provides a dependency measure between $\mu$ and $\nu$.
Closest related to independence on paths are the \emph{signature cumulants}
\cite{Bonnier2020} which have been proven to be useful in applications
\cite{schell2021nonlinear, friz2021unified}.
However, signature cumulants only compare probability measures on paths [tracks] with other probability measures on paths [tracks].
In contrast, the mutual information \eqref{eq:MI} allows to measures dependency between a probability measure $\mu$ on paths [tracks] and a probability measure $\nu$ on an arbitrary measurable space; in particular, this allows to measure dependence relations between TS and scalar-valued random variables.
\begin{figure}[H]
\centering
\includegraphics[align=t, scale=0.4]{shiftplot.pdf}
\includegraphics[align=t,scale=0.4]{shiftMI.pdf}
\caption{The left hand side shows sample trajectories from $\mathbf{Y}$ for different correlation values $\rho$. The right plot shows the mutual information between $\mathbf{Y}$ and the rotational speed $\omega$ as a function of $\rho$.}
\label{fig:shift_MI}
\end{figure}
To demonstrate this, we consider two examples
\begin{description}
\item[Mutual information between a stochastic process and a scalar.]
We consider the stochastic process $\mathbf{y}$ defined as
\begin{align}
\mathbf{y}_t = \rho t {\cos \omega t \choose \sin \omega t} + \sqrt{1-\rho^2} \mathbf{x}_t
\end{align}
where $\omega$ is sampled from the uniform distribution on $[0, 8\pi) $ and $\mathbf{x}$ is a standard $ 2 $-dimensional Brownian motion independent from $ \omega $. $ \mathbf{x} $ and $ \mathbf{y} $ are generated on the interval $ [0,1] $ on a grid with resolution $ 10^{-2} $.
The right plot in Figure \ref{fig:shift_MI} shows some sample trajectories of the process $\mathbf{y}$.
The left plot in Figure \ref{fig:shift_MI} shows the mutual information between (the law of) $ \omega $ and $ \mathbf{y} $ for various values of $\rho$.
As expected, the mutual information increases monotone with $\rho$ and is bounded by the entropy.
\item[Mutual Information between two unparametrized stochastic processes.]
We consider two independent a Brownian motions $ \mathbf{x} $ and $ \mathbf{y} $ as before, and a random parametrisation $ \phi(t) = T(t/T)^p $ where $ p \in [1,10] $ is uniformly distributed independent of $ \mathbf{x} $. The stochastic process $ \mathbf{z} $ is defined as
\begin{align}
\mathbf{z}_t = \rho \mathbf{x}_{\phi(t)} + \sqrt{1-\rho^2} \mathbf{y}_t.
\end{align}
We then compute the mutual information between $ \mathbf{z} $ and $ \mathbf{x} $ as a function of $ \rho $. The results are shown in Figure \ref{fig:warp_MI}.
\end{description}
\begin{figure}[H]
\centering
\includegraphics[align=t, scale=0.4]{warppaths.pdf}
\includegraphics[align=t,scale=0.4]{warpMI.pdf}
\caption{The left plot shows one-dimensional projections of sample paths of $ \mathbf{x} $ for different warping parameters $ p $. The right plot shows the mutual information between $ \mathbf{x} $ and $ \mathbf{z} $ as a function of $ \rho $.}
\label{fig:warp_MI}
\label{fig:warp_plot}
\end{figure}
\subsection*{Acknowledgements}
PB is supported by the Engineering and Physical Sciences Research Council [EP/R513295/1].
HO is supported by the EPSRC grant ``DATASIG'', the Turing Institute, and the Oxford-Man Institute of Quantitative Finance.
\newpage
|
\section{Introduction}
\label{sec:intro}
Quantum error correction (QEC) is one of the most important and widely studied ideas in quantum information processing~\cite{shor1995scheme,nielsen2002quantum,gottesman2010introduction,lidar2013quantum}. The spirit of QEC is to protect quantum information against noise and errors by suitably encoding logical quantum systems into quantum codes living in a larger physical Hilbert space. Since quantum systems are highly susceptible to noise effects such as decoherence so that errors easily occur, it is clear that QEC is of vital importance to the practical realization of quantum computing and other quantum technologies. Interestingly, besides the enduring efforts on the study of QEC and quantum codes for quantum information processing purposes, in recent years, they are also found to play intriguing roles in many important physical scenarios in e.g. holographic quantum gravity \cite{almheiri2015bulk,pastawski2015holographic} and many-body physics \cite{Kitaev2003,zeng2019quantum,brandao2019quantum,PhysRevLett.120.200503}, and have consequently drawn great interest in physics.
When considering the practical implementation of QEC as well as its connections to physical problems, it is important to take symmetries and conservation laws into account as they are ubiquitous in physical systems. More explicitly, symmetries may constrain the encoders in the way that they must be covariant with respect to the symmetry group, i.e., commute with certain representations of group actions, generating the so-called ``{covariant codes}'' \cite{hayden2017error,faist2019continuous,woods2019continuous,wang2019quasi}.
A principle of fundamental significance in both quantum information and physics is that (finite-dimensional) covariant codes for continuous symmetries (mathematically modeled by Lie groups)\footnote{In what follows, we assume that the associated symmetry group is continuous and that the relevant Hilbert spaces are finite-dimensional when using the term ``covariant codes''.} are in a sense fundamentally
incompatible with exact QEC~\cite{hayden2017error,eastin2009restrictions}.
A well known no-go theorem that unfolds this principle from a quantum computation perspective is the Eastin--Knill theorem \cite{eastin2009restrictions}, which essentially indicates that any QEC code covariant with respect to any continuous symmetry group in the sense that the logical group actions are mapped to ``transversal'' physical actions (that are tensor products on physical subsystems) cannot correct {local} errors perfectly. Crucially, transversal actions are highly desirable for the ``fault tolerance''~\cite{shor1996fault,nielsen2002quantum,gottesman2010introduction,lidar2013quantum} of practical quantum computation schemes because they do not spread errors across physical subsystems within each code block.
An intuitive explanation of this phenomenon is that, due to the conservation laws and transversality, physical subsystems necessarily contain logical charge information that gets leaked into the environment upon errors, so that the perfect recovery of logical information is prohibited.
Due to the Eastin--Knill theorem, unfortunately, it is impossible to find an exact QEC code that implements a universal set of gates transversally, or namely achieves the full power of quantum computation while maintaining transversality. However, it may still be feasible to perform QEC approximately under these constraints, and a natural task is then to characterize the optimal degree of accuracy. Recently, several such bounds on the QEC accuracy achievable by covariant codes (which give rise to ``robust'' or ``approximate'' versions of the Eastin--Knill theorem) as well as explicit constructions of near-optimal covariant codes are found using many different techniques and insights from various areas in quantum information \cite{faist2019continuous,woods2019continuous,wang2019quasi,kubica2020using,zhou2020new,yang2020covariant,fang2020no,tajima2021symmetry,wang2021theory,KongLiu21:random}, showcasing the fundamental nature of the problem.
Remarkably, covariant codes have also found interesting applications to several important scenarios in physics already, in particular, quantum reference frames \cite{hayden2017error,woods2019continuous}, quantum many-body systems \cite{brandao2019quantum,PhysRevLett.120.200503,wang2019quasi}, and AdS/CFT correspondence \cite{harlow2018symmetries,harlow2018constraints,kohler2019toy,faist2019continuous,woods2019continuous}.
These existing studies on covariant codes are mostly concerned with the precision of QEC under exact symmetry conditions.
Indeed, when symmetries arise in theoretical studies, it is commonly assumed by default that they are exactly respected. However, especially for continuous symmetries, it is often important or even necessary to consider cases where the symmetries or conservation laws are only approximate in physical and practical scenarios. First of all, realistic quantum many-body systems are often dirty or defective so that the exact symmetry conditions and conservation laws could generally be violated to a certain extent. Furthermore, there are many important situations in physics where non-exact symmetries need to be considered for fundamental reasons. For example, in particle physics, many important symmetries are known to be only approximate \cite{Witten2018}.
More notably, for quantum gravity, it is commonly believed that exact global symmetries are forbidden \cite{Misner1957,Giddings1988,PhysRevD.52.912,Arkani_Hamed_2007,BanksSeiberg11} (justified in more concrete terms in AdS/CFT \cite{harlow2018constraints,harlow2018symmetries}).
Considering the need for large quantum systems to boost the advantages of quantum technologies and also the broad connections between QEC and physics, it would be important and fruitful to have a theory about QEC codes with approximate symmetries, or ``approximately covariant'' codes.
For example, given that the QEC accuracy of exactly covariant codes is limited, one may wonder whether for codes that achieve exact QEC there are ``dual'' bounds on the degree of symmetry or covariance.
Particularly worth noting, the no-global-symmetry arguments in AdS/CFT indeed have deep connections to covariant codes, and in particular this question \cite{faist2019continuous,harlow2018symmetries}.
However, our understanding of approximate symmetries, especially characterizations and applications on a quantitative level, is very limited to date.
This work aims to establish a quantitative theory of the interplay between the degree of continuous symmetries and QEC accuracy, which in particular allows us to understand symmetry violation in exact QEC codes. (Note that our discussion here mainly proceeds in terms of the most fundamental $U(1)$ symmetry which is sufficient to reveal the key phenomena.)
To this end, we first formally define three different meaningful measures of symmetry violation, respectively in terms of the violation of covariance conditions globally over the entire group or locally at a specific point in the group, and the violation of charge conservation, {which induce corresponding quantitative notions of approximately covariant codes.} Our main results are a series of bounds on the trade-off between QEC accuracy and the above different symmetry measures, each of which may suit certain scenarios the best. (For readers' convenience, we provide in \appref{app:comparison} a table that identifies the key theorems and summarizes their respective strengths and weaknesses.) We introduce two concepts---charge fluctuation and gate implementation error---each providing a framework that allows us to understand and derive the QEC-symmetry trade-off and could be interesting on their own. Furthermore, our derivations feature ideas and techniques from several different fields. More explicitly, various different forms of the trade-off relations are derived by analyzing the ``perturbation'' of the Knill--Laflamme conditions~\cite{knill1997theory,beny2010general}, as well as by leveraging insights and techniques from the fields of quantum metrology~\cite{giovannetti2011advances,zhou2020theory} and quantum resource theory~\cite{chitambar2019quantum,marvian2020coherence,FangLiu19:nogo,fang2020no}.
Our theory provides a comprehensive and formal understanding of the transition between exact QEC and exact symmetry.
On the exact symmetry end, the previous limits on covariant codes (often referred to as ``approximate Eastin--Knill theorems''~\cite{faist2019continuous,woods2019continuous,kubica2020using,zhou2020new}) are recovered, while the exact QEC end provides new lower bounds on various forms of symmetry violation for the commonly studied exact codes. In particular, we demonstrate that our results imply fundamental limitations on the set of transversally implementable logical gates for general QEC codes, which represent a new type of improvement of the Eastin--Knill theorem, and
{apply more broadly than previous results along a similar line about stabilizer codes in Refs.~\cite{zeng2011transversality,bravyi2013classification,pastawski2015fault,anderson2016classification,jochym2018disjointness}.}
Then, to solidify the general theory, we present case studies on two explicit code constructions, which can be seen as interesting examples of approximately covariant codes that exhibit certain key features, as well as upper bounds or achievability results that help understand how strong our fundamental limits are. We first construct a parametrized code family that interpolates between the two ends of exact QEC and exact symmetry and exhibits a full trade-off between QEC and symmetry, by modifying the so-called thermodynamic code~\cite{brandao2019quantum,faist2019continuous}. In the second example, we choose to analyze the quantum Reed--Muller codes in that they exhibit nice structures and features, and in particular, have been widely applied for the transversal implementation of certain non-Clifford gates and magic state distillation~\cite{bravyi2012magic,anderson2014fault,haah2018codes,hastings2018distillation}. We find that both codes can almost saturate the bounds on global covariance and charge conservation (up to constant factors) asymptotically, that is, both the code constructions and bounds are close to being optimal.
This paper can be read as a companion and an extended version of Ref.~\cite{short}, which introduces the most representative results and motivates the study from a more physical standpoint. Here we present the study in a much more comprehensive manner, including all technical details and extended discussions. In particular, this paper contains thorough discussions of all different approaches and many additional results.
This paper is organized as follows. First, in \secref{sec:def}, we overview the formalism of QEC, and formally define the accuracy of approximate QEC codes as well as the different quantitative charaterizations of approximate continuous symmetries associated with QEC codes that will be considered in this work. In \secref{sec:global-1} and \secref{sec:global-2}, we respectively introduce the two frameworks based on the notions of charge fluctuation and gate implementation error, under which we discuss a series of different approaches to deriving the trade-off relations between the QEC inaccuracy and the group-global covariance violation. Then in \secref{sec:transversal}, we specifically discuss the application to fault-tolerant quantum computing, deriving general restrictions on the transversally implementable logical gates in QEC codes from the results above. Afterwards, in \secref{sec:local}, we present our results on the trade-off relations between QEC inaccuracy and group-local symmetry measures including the group-local covariance violation and the charge conservation violation. After the above discussion of fundamental limits, namely lower bounds, in \secref{sec:case-study} we study the modified thermodynamic code and quantum Reed--Muller codes, which gives concrete examples of nearly optimal approximately covariant approximate codes in certain cases. Finally, in \secref{sec:discussion} we summarize our work, and discuss important open problems and future directions.
\section{Approximately covariant approximate QEC codes: Quantitative characterizations of QEC and symmetry}
\label{sec:def}
Here we formally define the quantitative measures of QEC and symmetry that will be used in our study. Specifically, we first overview the notions of QEC and covariant codes and discuss how to quantify the deviation of general quantum codes from them. In particular, we will define the QEC inaccuracy which quantifies the QEC capability of a quantum code under specific noise. We will also define a measure of group-global symmetry which quantifies the approximate covariance of a code over the entire $U(1)$ group and two measures of group-local symmetry which are linked to the covariance of a code at an exact point in $U(1)$. The trade-off relations between QEC and these symmetry measures will be thoroughly studied in later sections.
Note that in this work we extensively use distance metrics defined based on the purified distance \cite{gilchrist2005distance,tomamichel2015quantum}, which are well-behaved and commonly used in the quantum information literature.
Notably, the choice of purified distance directly relates the local covariance violation to the well known quantum Fisher information (QFI)~\cite{helstrom1976quantum,holevo1982probabilistic,hubner1992explicit,sommers2003bures}. The general existence and properties of the Hessian matrix are left for future study.
In principle, one may also consider other metrics. We shall also discuss the situations where one uses the diamond distance~\cite{watrous2018theory}, another standard channel distance measure.
\subsection{Approximate quantum error correction}
\label{sec:def-QEC}
QEC functions by encoding the logical quantum system in some quantum code living in a larger physical system with redundancy, so that a limited number of errors can be corrected to recover the original logical information. A quantum code is defined by an encoding quantum channel ${\mathcal{E}}_{{S\leftarrow L}}$ from a logical system $L$ to a physical system $S$, and it perfectly protects the logical information against a physical noise ${\mathcal{N}}_{S}$ if and only if there exists a recovery channel ${\mathcal{R}}_{{S\leftarrow L}}$ such that
\begin{equation}
{\mathcal{R}}_{{L\leftarrow S}} \circ {\mathcal{N}}_{S} \circ {\mathcal{E}}_{{S\leftarrow L}} = {\mathbbm{1}}_L.
\end{equation}
In particular, when ${\mathcal{E}}_{{S\leftarrow L}}$ is isometric, ${\mathcal{N}}_{S}(\cdot) = \sum_{i=1}^r K_{S,i}(\cdot)K_{S,i}^\dagger$ and $\Pi$ is the projection onto the code subspace in the physical system, such a recovery channel exists if and only if the Knill--Laflamme (KL) conditions, $\forall i,j,\,\Pi K_{S,i}^\dagger K_{S,j} \Pi \propto \Pi$~\cite{knill1997theory}, hold.
In many scenarios, a quantum code is still useful in protecting quantum information when it only achieves approximate QEC, namely, ${\mathcal{R}}_{{L\leftarrow S}} \circ {\mathcal{N}}_{S} \circ {\mathcal{E}}_{{S\leftarrow L}}$ is close to but not exactly equal to ${\mathbbm{1}}_L$. To characterize the inaccuracy of an approximate QEC code, we will use the channel fidelity and the Choi channel fidelity, respectively defined by
\begin{gather}
f(\Phi_1,\Phi_2) = \max_{\rho} f\left((\Phi_1\otimes {\mathbbm{1}})(\rho),(\Phi_2\otimes {\mathbbm{1}})(\rho)\right),\\
\overline{f}(\Phi_1,\Phi_2) = f\left((\Phi_1\otimes {\mathbbm{1}})(\Psi),(\Phi_2\otimes {\mathbbm{1}})(\Psi)\right),
\end{gather}
where $f(\rho,\sigma) = \mathrm{Tr}(\sqrt{\rho^{1/2}\sigma\rho^{1/2}})$ is the fidelity of quantum states, $\ket{\Psi} = \frac{1}{\sqrt{d}}\sum_{i=1}^{d}\ket{i}\ket{i}$ is the maximally entangled state and $\Psi = \ket{\Psi}\bra{\Psi}$. Here the inputs $\rho$ and $\Psi$ lie in a bipartite system consisting of the original system $\Phi_{1,2}$ acting on and a reference system as large as the original. Correspondingly, one can define the purified distance of states $P(\rho,\sigma) = \sqrt{1 - f(\rho,\sigma)^2}$, the purified distance of channels $P(\Phi_1,\Phi_2) = \sqrt{1 - f(\Phi_1,\Phi_2)^2}$ and the Choi purified distance of channels $\overline{P}(\Phi_1,\Phi_2) = \sqrt{1 - \overline{f}(\Phi_1,\Phi_2)^2}$~\cite{gilchrist2005distance,tomamichel2015quantum,LiuWinter19}.
The \emph{(worst-case) QEC inaccuracy} and the \emph{Choi QEC inaccuracy} for approximate QEC codes are then respectively defined as
\begin{gather}
\epsilon({\mathcal{N}}_S,{\mathcal{E}}_{{S\leftarrow L}}) := \min_{{\mathcal{R}}_{{L\leftarrow S}}} P({\mathcal{R}}_{{L\leftarrow S}}\circ{\mathcal{N}}_S \circ {\mathcal{E}}_{{S\leftarrow L}},{\mathbbm{1}}_L),\\
\overline{\epsilon}({\mathcal{N}}_S,{\mathcal{E}}_{{S\leftarrow L}}) := \min_{{\mathcal{R}}_{{L\leftarrow S}}} \overline{P}({\mathcal{R}}_{{L\leftarrow S}}\circ{\mathcal{N}}_S \circ {\mathcal{E}}_{{S\leftarrow L}},{\mathbbm{1}}_L).
\end{gather}
The Choi inaccuracy reflects the average-case behavior as $\overline{\epsilon} = \sqrt{\frac{d_L+1}{d_L}}\epsilon_\mathrm{avg}$~\cite{horodecki1999general,nielsen2002simple} where $\epsilon_\mathrm{avg} := \sqrt{{1 - \max_{{\mathcal{R}}_{{L\leftarrow S}}}\int d\psi f^2(\psi,{\mathcal{R}}_{{L\leftarrow S}}\circ {\mathcal{N}}_S \circ {\mathcal{E}}_{{S\leftarrow L}}(\psi))}}$ and the integral is over the Haar random pure logical states. For simplicity, we will not explicitly write down the arguments of $\epsilon$ or $\overline{\epsilon}$ (and of many other measures defined later) when they are unambiguous.
In the above, we used the channel purified distances as channel distance measures.
As mentioned, we may also consider the the diamond distance $D_\diamond(\Phi_1,\Phi_2)$ induced by the diamond norm of channels~\cite{diamond,watrous2018theory}:
\begin{align}
& \,D_\diamond(\Phi_1,\Phi_2) \nonumber\\
:=\,& \max_\rho \frac{1}{2} \norm{(\Phi_1\otimes {\mathbbm{1}})(\rho)-(\Phi_2\otimes {\mathbbm{1}})(\rho)}_1\\
=\,& \max_{\ket{\psi}} \frac{1}{2} \norm{(\Phi_1\otimes {\mathbbm{1}})(\ket{\psi}\bra{\psi})-(\Phi_2\otimes {\mathbbm{1}})(\ket{\psi}\bra{\psi})}_1.
\end{align}
where $\norm{\cdot}_1$ is the nuclear (trace) norm.
Naturally, the diamond distance version of QEC inaccuracy is defined as
\begin{equation}
\epsilon_\diamond := \min_{{\mathcal{R}}_{{L\leftarrow S}}} D_\diamond({\mathcal{R}}_{{L\leftarrow S}}\circ{\mathcal{N}}_S \circ {\mathcal{E}}_{{S\leftarrow L}},{\mathbbm{1}}_L).
\end{equation}
It is easy to see that lower bounds on $\epsilon$ (that we derive below) directly indicate lower bounds on $\epsilon_\diamond$.
According to the Fuchs--van de Graaf inequality $1 - f(\rho,\sigma) \leq \frac{1}{2}\norm{\rho - \sigma}_1$~\cite{fuchs1999cryptographic}, we have
\begin{align}
\label{eq:fuchs}
D_\diamond(\Phi_1,\Phi_2) \geq 1 - f(\Phi_1,\Phi_2)
\geq \frac{1}{2}P(\Phi_1,\Phi_2)^2.
\end{align}
In the case of our interest where the second channel is the identity, the above inequality can be further improved using $1 - f(\rho,\ket{\psi}\bra{\psi})^2 \leq \frac{1}{2}\norm{\rho - \ket{\psi}\bra{\psi}}_1$:
\begin{align}
D_\diamond(\Phi_1,{\mathbbm{1}}) \geq 1 - f(\Phi_1,{\mathbbm{1}})^2 = P(\Phi_1,{\mathbbm{1}})^2.
\end{align}
Therefore, $\epsilon_\diamond \geq \epsilon^2$.
\subsection{Measuring approximate symmetries of QEC codes}
\label{sec:def-symmetry}
Symmetries of quantum codes manifest themselves in the covariance of the encoder with respect to symmetry transformations.
For the case of current interest, the symmetry transformations on the logical and physical systems are respectively $U_{L,\theta} = e^{-iH_L\theta}$ generated by a logical Hamiltonian (charge observable) $H_L$, and $U_{S,\theta} = e^{-iH_S\theta}$ generated by a physical Hamiltonian (charge observable) $H_S$, both representations of the $U(1)$ Lie group periodic with a common period $\tau$.
The transversality property of symmetry transformations (gate actions) corresponds to the 1-local form of $H_S$, namely, $H_S=\sum_{l=1}^n H_{S_l}$ where each term $H_{S_l}$ acts locally on physical subsystem $S_l$.
We say a quantum code is covariant (with respect to such $U(1)$ representations given by $H_L$ and $H_S$), if
\begin{equation}
\label{eq:covariant}
{\mathcal{U}}_{S,\theta} \circ {\mathcal{E}}_{{S\leftarrow L},\theta} = {\mathcal{E}}_{{S\leftarrow L},\theta} \circ {\mathcal{U}}_{L,\theta},\quad \forall \theta \in {\mathbb{R}}.
\end{equation}
The definitions of covariant codes can be easily extended to general compact Lie groups \cite{faist2019continuous,woods2019continuous}. We also assume $H_L$ and $H_S$ to be both non-trivial, i.e., not a constant operator. Note that applying constant shifts on $H_L$ and $H_S$ do not change the definition of \eqref{eq:covariant} and we will often use this property below.
As mentioned, the covariance of quantum codes is often incompatible with their error-correcting properties and approximate notions of covariance may play important roles in wide-ranging scenarios. For example, here the Eastin--Knill theorem indicates that codes that can perfectly correct single-erasure noise cannot simultaneously be covariant with respect to non-trivial 1-local $H_S$ \cite{eastin2009restrictions}. More generally, exact QEC is known to be incompatible with exact covariance as long as
\begin{equation}
H_S \in {\rm span}\{K_{S,i}^\dagger K_{S,j},\,\forall i,j\},
\end{equation}
which we refer to as the ``Hamiltonian-in-Kraus-span'' (HKS) condition, holds~\cite{zhou2020new,zhou2020theory}.
The HKS condition
holds for many typical scenarios, including the one mentioned above where ${\mathcal{N}}_S$ represents single-erasure noise and $H_S$ is 1-local. When the HKS condition does not hold, examples of exactly covariant QEC codes exist, e.g., when ${\mathcal{N}}_{S} = {\mathbbm{1}}$ (noiseless dynamics), when $H_S$ is a Pauli-X operator and ${\mathcal{N}}_S$ is dephasing noise~\cite{kessler2014quantum,arrad2014increasing}, and when ${\mathcal{N}}_S$ is single-erasure noise but $H_S$ is 2-local~\cite{gottesman2016quantum}. We shall assume that the HKS condition holds for the quantum codes considered in this paper. We also emphasize that there exist examples of exact QEC codes covariant with respect to discrete symmetry groups~\cite{hayden2017error}, so the assumption of continuous groups is important.
Besides quantum computation, approximately symmetries and covariant codes are potentially useful in quantum gravity and condensed matter physics, as discussed in Ref.~\cite{short}. To formally characterize and study approximate covariance, an important first step is to find reasonable ways to quantify it. We now do so.
\subsubsection{Group-global covariance violation}
\label{sec:group-global}
The first, most important type of measure is based on the global covariance violation over the entire symmetry group. Codes that are approximately covariant with respect to $H_L$ and $H_S$ in such a global sense should have small covariance violation for all $\theta$. We define the \emph{group-global\footnote{We shall refer to ``group-global'' and ``group-local'' as just ``global'' and ``local'' for simplicity. However, they should not be confused with the spatially geometrical notions of ``global'' and ``local'' often used in physics.} covariance violation} and the \emph{Choi group-global covariance violation} respectively by
\begin{gather}
\delta_\group := \max_\theta P({\mathcal{U}}_{S,\theta}\circ{\mathcal{E}}_{{S\leftarrow L}},{\mathcal{E}}_{{S\leftarrow L}}\circ{\mathcal{U}}_{L,\theta}),\\
\overline{\delta}_\group := \max_\theta \overline{P}({\mathcal{U}}_{S,\theta}\circ{\mathcal{E}}_{{S\leftarrow L}},{\mathcal{E}}_{{S\leftarrow L}}\circ{\mathcal{U}}_{L,\theta}).
\end{gather}
Intuitively, they measure the maximum deviation of the encoding channel ${\mathcal{E}}_{{S\leftarrow L}}$ from the exact covariance condition \eqref{eq:covariant} in the entire symmetry group.
It is known that $\delta_\group$ and $\epsilon$ cannot be simultaneously zero in non-trivial situations, and previous works~\cite{faist2019continuous,woods2019continuous,wang2019quasi,kubica2020using,zhou2020new,yang2020covariant,tajima2021symmetry} mostly focus on deriving lower bounds on $\epsilon$ for exactly covariant codes ($\delta_\group = 0$). We will present bounds that involve $\delta_\group$ which reveal the trade-off between QEC and global covariance, derived via two notions we introduce called the charge fluctuation and gate implementation error. This extends the scope of previous consideration to general codes including exact QEC codes.
Similar to the case of QEC inaccuracy, we can also consider the diamond distance and define
\begin{equation}
\delta_{\group,\diamond} := \max_\theta D_\diamond({\mathcal{U}}_{S,\theta}\circ{\mathcal{E}}_{{S\leftarrow L}},{\mathcal{E}}_{{S\leftarrow L}}\circ{\mathcal{U}}_{L,\theta}).
\end{equation}
Again, lower bounds on $\delta_{\group,\diamond}$ that we derive below directly indicate lower bounds on $\delta_{\group}$.
Using \eqref{eq:fuchs}, we directly see that $\delta_{\group,\diamond} \geq \delta_{\group}^2/2$. In particular, when ${\mathcal{E}}_{{L\leftarrow S}}$ is isometric, we have $\delta_{\group,\diamond} = \delta_{\group}$, using the fact that $\frac{1}{2}\norm{\ket{\psi_1}\bra{\psi_1} - \ket{\psi_2}\bra{\psi_2}}_1 = P(\ket{\psi_1}\bra{\psi_1},\ket{\psi_2}\bra{\psi_2})$.
\subsubsection{Group-local (point) covariance violation}
One may wonder if the incompatibility between QEC and continuous symmetries can be relieved when we relax the requirement from exact global covariance to exact local covariance, i.e., when we require only the code covariance for $\theta$ inside a small neighborhood of a point $\theta_0$, satisfying ${\mathcal{U}}_{S,\theta_0} \circ {\mathcal{E}}_{{S\leftarrow L}} = {\mathcal{E}}_{{S\leftarrow L}} \circ {\mathcal{U}}_{L,\theta_0}$.
Unfortunately, the no-go results also extend to the local case, meaning that a non-trivial QEC code cannot be exactly covariant even in an arbitrarily small neighborhood of $\theta_0$. Without loss of generality, we assume $\theta_0 = 0$ because we can always redefine ${\mathcal{E}}_{{S\leftarrow L}} \circ {\mathcal{U}}_{L,\theta_0}$ to be the new encoding channel such that the code is covariant at $\theta = 0$. To characterize the local covariance, we introduce the \emph{group-local (point) covariance violation} defined by
\begin{align}
\delta_{\mathrm{P}}
:=&\, \sqrt{{2\partial_\theta^2 P({\mathcal{U}}_{S,\theta}\circ{\mathcal{E}}_{{S\leftarrow L}}\circ{\mathcal{U}}_{L,\theta}^\dagger,{\mathcal{E}}_{{S\leftarrow L}})^2\big|_{\theta = 0}}} \nonumber\\
=&\, \sqrt{F({\mathcal{U}}_{S,\theta}\circ{\mathcal{E}}_{{S\leftarrow L}}\circ{\mathcal{U}}_{L,\theta}^\dagger)\big|_{\theta = 0}}.
\end{align}
Here $F(\Phi_\theta)$ is the quantum Fisher information (QFI) defined using the second order derivative of the purified distance $F(\Phi_\theta) = 2\partial_{\theta'}^2 P(\Phi_\theta,\Phi_{\theta'})^2\big|_{\theta' = \theta}$ which characterizes the amount of information of $\theta$ one can extract from $\Phi_\theta$ around point $\theta$~\cite{fujiwara2008fibre}. Correspondingly, the QFI of quantum states is defined by $F(\rho_\theta) = 2\partial_{\theta'}^2 P(\rho_\theta,\rho_{\theta'})^2\big|_{\theta' = \theta}$~\cite{hubner1992explicit,braunstein1994statistical} which characterizes the amount of information of $\theta$ one can extract from $\rho_\theta$ around point $\theta$ and we have $F(\Phi_\theta) = \max_\rho F((\Phi_\theta\otimes{\mathbbm{1}})(\rho))$. Note that the QFI defined here using the purified distance is usually called the SLD QFI and there are other types of QFIs, e.g., the RLD QFI~\cite{yuen1973multiple,hayashi2011comparison,katariya2020geometric} which we will encounter later in \secref{sec:global-2-res}.
When $\delta_{\mathrm{P}} = 0$, the code is locally covariant up to the lowest order of $d\theta$. We shall see later that for any $\delta_{\mathrm{P}} < \Delta H_L$ (we will use $\Delta(\cdot)$ to denote the difference between the maximum and minimum eigenvalues of $(\cdot)$), there is a non-trivial lower bound on $\epsilon$, leading to a trade-off relation between QEC and local covariance.
\subsubsection{Charge conservation violation}
The correspondence between symmetries and conservation laws is a landmark result of modern physics. Inspired by this correspondence, we can define another intuitive measure of the symmetry violation by the degree of charge deviation.
It can be shown that for an isometric encoding channel ${\mathcal{E}}_{{S\leftarrow L}}$ the covariance condition \eqref{eq:covariant} is equivalent to
\begin{equation}
\label{eq:isometric-charge-relation}
{\mathcal{E}}^\dagger_{{L\leftarrow S}}(H_S) = H_L-\nu{\mathbbm{1}},
\end{equation} for some $\nu\in\mathbb{R}$,
where ${\mathcal{E}}^\dagger$ is the dual channel of ${\mathcal{E}}$ satisfying $\mathrm{Tr}(H{\mathcal{E}}(\rho)) = \mathrm{Tr}({\mathcal{E}}^\dagger(H)\rho)$ for any $H$ and $\rho$. Since $H_S$ and $H_L$ represent the charge observables in the physical and logical systems, \eqref{eq:isometric-charge-relation} implies that the eigenstates of $H_L$ are mapped to the corresponding eigenstates of $H_S$ after the encoding operation~\cite{faist2019continuous}, indicating the charge conservation nature of the encoding map.
The charge conservation law can also be understood through the relation $\mathrm{Tr}(H_S {\mathcal{E}}_{{S\leftarrow L}}(\rho)) = \mathrm{Tr}(H_L\rho)-\nu$ for any $\rho$, where $\nu$ represents a universal constant offset in the charge.
To measure the degree to which the charge conservation law is deviated from, we consider the following quantity which we call the \emph{charge conservation violation} (also defined in Ref.~\cite{faist2019continuous}):
\begin{equation}
\delta_{\mathrm{C}}
:= \Delta\left(H_L - {\mathcal{E}}_{{L\leftarrow S}}^\dagger(H_S)\right).
\end{equation}
Note again that $\Delta(\cdot)$ denotes the difference between the maximum and minimum eigenvalues of $(\cdot)$. {It can be easily verified that $\delta_{\mathrm{C}}/2$ is equal to the difference between physical and logical charges, formally given by $\min_{\nu\in{\mathbb{R}}} \max_\rho| \mathrm{Tr}(H_S{\mathcal{E}}_{{S\leftarrow L}}(\rho))-\mathrm{Tr} ((H_L-\nu {\mathbbm{1}})\rho)|$ (a constant offset on the definitions of charges is allowed).}
For general CPTP encoding maps, $\delta_{\mathrm{C}}$ is not always zero for exactly covariant codes~\cite{cirstoiu2020robustness}, unlike $\delta_\group$ and $\delta_{\mathrm{P}}$. However, for isometric encoding we always have the following relation between $\delta_{\mathrm{P}}$ and $\delta_{\mathrm{C}}$:
\begin{proposition}
\label{prop:point-charge}
When ${\mathcal{E}}_{{L\leftarrow S}}$ is isometric, $\delta_{\mathrm{P}} \geq \delta_{\mathrm{C}}$.
\end{proposition}
\begin{proof}
Suppose ${\mathcal{E}}_{{L\leftarrow S}}(\cdot) = W(\cdot)W^\dagger$ where $W$ is isometric. Then
\begin{align}
\delta_{\mathrm{C}} = \Delta\left({H_L - W^\dagger H_S W}\right),
\end{align}
and
\begin{align}
(\delta_{\mathrm{P}})^2 =&\, F({\mathcal{U}}_{S,\theta}\circ{\mathcal{E}}_{{S\leftarrow L}}\circ{\mathcal{U}}_{L,\theta}^\dagger)\big|_{\theta = 0} \\ =&\, \max_{\ket{\psi}} F(U_{S,\theta} W U_{L,\theta}^\dagger \ket{\psi})\big|_{\theta = 0}.
\end{align}
Let
\begin{align}
\ket{\psi_\theta} = U_{S,\theta} W U_{L,\theta}^\dagger \ket{\psi} = e^{-iH_S\theta}We^{iH_L\theta}\ket{\psi}.
\end{align}
Then
\begin{align}
\ket{\partial_\theta\psi_\theta} \!=\! e^{-iH_S\theta}We^{iH_L\theta}iH_L\ket{\psi}
\!-\! iH_S e^{-iH_S\theta}We^{iH_L\theta}\ket{\psi}\!.
\end{align}
The QFI for pure states is given by
$F(\ket{\psi_\theta}) = 4(\braket{\partial_\theta\psi_\theta|\partial_\theta\psi_\theta} - \abs{\braket{\partial_\theta\psi_\theta|\psi_\theta}}^2 )$~\cite{braunstein1994statistical}. Since
\begin{gather}
\begin{split}
&\braket{\partial_\theta\psi_\theta|\partial_\theta\psi_\theta}\big|_{\theta = 0} \\
=& \bra{\psi}(W H_L - H_S W)^\dagger (W H_L - H_S W)\ket{\psi}\\
\geq& \bra{\psi}(W H_L - H_S W)^\dagger WW^\dagger (W H_L - H_S W)\ket{\psi},
\end{split}\\
\braket{\psi_\theta|\partial_\theta\psi_\theta}\big|_{\theta = 0}
= \bra{\psi}(H_L - W^\dagger H_S W)\ket{\psi},
\end{gather}
we have
\begin{align}
(\delta_{\mathrm{P}})^2 =& \max_{\psi} F(\ket{\psi_\theta})\big|_{\theta = 0} \\
\geq& \max_{\psi} 4\bra{\psi}( H_L - W^\dagger H_S W)^\dagger (H_L - W^\dagger H_S W)\ket{\psi} \nonumber \\&- 4\abs{\bra{\psi}(H_L - W^\dagger H_S W)\ket{\psi}}^2 \\=& (\delta_{\mathrm{C}})^2,
\end{align}
proving the result.
\end{proof}
We shall see later that, similar to the situation of local covariance violation, for any $\delta_{\mathrm{C}} < \Delta H_L$, there is a non-trivial lower bound on $\epsilon$. We refer to both $\delta_{\mathrm{P}}$ and $\delta_{\mathrm{C}}$ as local symmetry measures because their values only depend on the approximate covariance of a code in the neighborhood of the point $\theta = 0$. Note that both $\delta_{\mathrm{P}}$ and $\delta_{\mathrm{C}}$ have the same unit as the charges while $\delta_\group$ is dimensionless, i.e., after replacing $H_S$ and $H_L$ with $c H_S$ and $c H_L$ for some constant $c$, both $\delta_{\mathrm{P}}$ and $\delta_{\mathrm{C}}$ are changed to $c \delta_{\mathrm{P}}$ and $c \delta_{\mathrm{C}}$ while $\delta_\group$ is unchanged.
\subsubsection{Remarks on non-compact groups and infinite-dimensional codes}
In the above discussion, we assumed compact Lie groups and finite-dimensional quantum codes. Here we remark on possible extensions to non-compact Lie groups and infinite-dimensional codes.
First, note that our definitions of $\delta_{\mathrm{P}}$ and $\delta_{\mathrm{C}}$ can be naturally extended to the situations of non-compact groups where $H_S$ and $H_L$ are arbitrary finite-dimensional Hermitian operators but the group transformations are not periodic, because their definitions only depend on the local geometry of the symmetry group. For the global measure $\delta_\group$, our results below need the assumption that the physical and logical group transformations are both periodic with a common period.
Moreover, when the physical system $S$ is infinite-dimensional, one may naturally consider some finite-dimensional truncation $\tilde H_S$ of $H_S$. The trade-off relations we derive below hold for $\tilde H_S$ and $H_L$, so when the truncation is suitably chosen we can still obtain nontrivial results that well indicate the behaviors of $H_S$. For example, when $\|{\mathcal{E}}^\dagger_{{L\leftarrow S}}(H_S) - {\mathcal{E}}^\dagger_{{L\leftarrow S}}(\tilde H_S)\|$ ($\norm{\cdot}$ is the spectral norm) is small, $\tilde H_S$ is a good substitute for $H_S$ in terms of the charge conservation violation~\cite{faist2019continuous}.
\section{Trade-off between QEC and global covariance: charge fluctuation approach}
\label{sec:global-1}
In this section, we derive trade-off relations between the QEC inaccuracy $\epsilon$ and the global covariance violation $\delta_\group$ by connecting them to {a quantity which we call} the charge fluctuation $\chi$ (note that this notion is distinct from the charge conservation violation although they are in some way related as will be discussed). Our approach essentially proceed in two steps. First, we connect $\delta_\group$ and $\chi$ by providing a lower bound of $\delta_\group$ which depends on $\chi$. Then we prove upper bounds on $\abs{\chi}$ in terms of the QEC inaccuracy $\epsilon$ using two different methods. The first one is based on analyzing the deviation of the approximate QEC code from the the KL conditions, which we call the KL-based method, and the second one is based on treating the problem as a channel estimation problem and using quantum metrology techniques. These methods eventually lead to two types of trade-off bounds between QEC and global covariance. Note that we assume quantum codes are isometric throughout this section unless stated otherwise.
\subsection{Bounding global covariance violation by charge fluctuation}
\label{sec:global-1-charge}
Consider a code defined by encoding isometry ${\mathcal{E}}_{{S\leftarrow L}}$. We start by considering the situation where the code achieves exact QEC under the noise channel ${\mathcal{N}}_{S}(\cdot) = \sum_i K_{S,i}(\cdot)K_{S,i}^\dagger$.
According to the KL conditions,
\begin{equation}
\label{eq:KL}
\Pi K_{S,i}^\dagger K_{S,j} \Pi \propto \Pi,
\end{equation}
where $\Pi$ is the projection onto the code subspace.
In particular, let $\ket{0_L}$ and $\ket{1_L}$ be eigenstates corresponding to the largest and the smallest eigenvalues of $H_L$. (We do not specify the exact choices of $\ket{0_L}$ and $\ket{1_L}$ even when $H_L$ is degenerate, as long as they respectively correspond to the largest and smallest eigenvalues.) Using \eqref{eq:KL}, we have
\begin{equation}
\bra{0_L}{\mathcal{E}}^\dagger_{{L\leftarrow S}}(K_{S,i}^\dagger K_{S,j})\ket{0_L} =
\bra{1_L}{\mathcal{E}}^\dagger_{{L\leftarrow S}}(K_{S,i}^\dagger K_{S,j})\ket{1_L}.
\end{equation}
Using the HKS condition $H_S \in {\rm span}\{K_{S,i}^\dagger K_{S,j},\,\forall i,j\}$,
we must also have
\begin{equation}
\label{eq:KL-H}
\bra{0_L}{\mathcal{E}}^\dagger_{{L\leftarrow S}}(H_S)\ket{0_L} =
\bra{1_L}{\mathcal{E}}^\dagger_{{L\leftarrow S}}(H_S)\ket{1_L}.
\end{equation}
The incompatibility between QEC and symmetry could be understood through the incompatibility between \eqref{eq:KL-H} and \eqref{eq:isometric-charge-relation}. \eqref{eq:KL-H} implies that $\braket{0_L|H_L|0_L} - \braket{1_L|H_L|1_L} = 0$ when the code achieves exact QEC, which contradicts with $\braket{0_L|H_L|0_L} - \braket{1_L|H_L|1_L} = \Delta H_L$ for exactly covariant codes from \eqref{eq:isometric-charge-relation}, implying the non-existence of exact QEC codes with exact covariance.
For general codes, we define the \emph{charge fluctuation}:
\begin{equation}
\chi := \bra{0_L}{\mathcal{E}}^\dagger_{{L\leftarrow S}}(H_S)\ket{0_L} -
\bra{1_L}{\mathcal{E}}^\dagger_{{L\leftarrow S}}(H_S)\ket{1_L},
\end{equation}
Based on the discussion above, one can see that $\chi$ embodies the transition between exact QEC and exact symmetry quantitatively---when a code is close to being an exactly covariant code, $\chi$ cannot be too far away from $\Delta H_L$, and when a code is close to being an exact QEC code, $\chi$ cannot be too far away from $0$ (see an illustration in \figref{fig:charge}). Thus the trade-off relation between $\epsilon$ and $\delta_\group$ can be derived by connecting $\chi$ to each of them separately.
\begin{figure}[tb]
\center
\includegraphics[width=0.35\textwidth]{charge-fluctuation.pdf}
\caption{\label{fig:charge}
For exact QEC codes ($\epsilon = 0$) which satisfy the HKS condition, the charge fluctuation $\chi = 0$. For exactly covariant codes ($\delta_\group = 0$), the charge fluctuation $\chi = \Delta H_L$. The trade-off can be derived by investigating the relations between the distances of $\chi$ from $\Delta H_L$ and $0$, and the symmetry and QEC measures.
}
\end{figure}
We now derive the following lower bound on the global covariance violation $\delta_\group$ in terms of the charge fluctuation $\chi$, which directly connects $\delta_\group$ with $\chi$.
\begin{proposition}
\label{prop:global-charge}
Consider an isometric quantum code defined by ${\mathcal{E}}_{{L\leftarrow S}}$. Consider physical Hamiltonian $H_S$, logical Hamiltonian $H_L$, and noise channel ${\mathcal{N}}_S$. Suppose the HKS condition is satisfied. Then when $\abs{\Delta H_L - \chi} \leq \Delta H_S$, it holds that
\begin{multline}
\label{eq:delta-charge}
\delta_\group \geq \\
\min\left\{\frac{\sqrt{\abs{\Delta H_L - \chi}\left(\Delta H_S - \frac{1}{2}\abs{\Delta H_L - \chi}\right)}}{\Delta H_S},\sqrt{\frac{3}{8}}\right\},
\end{multline}
and when $\abs{\Delta H_L - \chi} > \Delta H_S$, $\delta_\group \geq \sqrt{3/8}$.
In particular, when $\abs{\Delta H_L - \chi} \ll \Delta H_S$, we have
\begin{equation}
\delta_\group \gtrsim \sqrt{\frac{\abs{\Delta H_L - \chi}}{\Delta H_S}}.
\end{equation}
\end{proposition}
\noindent {Note that in this paper, ``$\gtrsim$'', ``$\lesssim$'', and ``$\simeq$'' respectively mean ``$\geq$'', ``$\leq$'', and ``$=$'' up to the leading order.}
\begin{proof}
Since $U_{S,\theta}$ and $U_{L,\theta}$ are both periodic with a common period, we assume $H_S$ and $H_L$ both have integer eigenvalues. We also assume the smallest eigenvalue of $H_S$ is zero because constant shifts do not affect the definitions of symmetry measures.
When ${\mathcal{E}}_{{L\leftarrow S}}(\cdot) = W(\cdot)W^\dagger$ is isometric, let $\ket{{\mathfrak{c}}_{0}} = W\ket{0_L}$, $\ket{{\mathfrak{c}}_{1}} = W\ket{1_L}$, and write
\begin{align}
\ket{{\mathfrak{c}}_0} = \sum_{\eta=0}^{\Delta H_S} c_\eta^0 \ket{\eta^0},\quad
\ket{{\mathfrak{c}}_1} = \sum_{\eta=0}^{\Delta H_S} c_\eta^1 \ket{\eta^1},
\end{align}
where $\sum_\eta \abs{c_\eta^0}^2 = \sum_\eta \abs{c_\eta^1}^2 = 1$ and $\ket{\eta^0}$ and $\ket{\eta^1}$ are eigenstates of $H_S$ with eigenvalue $\eta$. $\ket{\eta^0}$ and $\ket{\eta^1}$ may not be the same when $H_S$ is degenerate. Note that when $\eta$ is not an eigenvalue of $H_S$, we simply take $c_\eta^{0} = 0$ (or $c_\eta^{1} = 0$) so that $c_\eta^{0}$ (or $c_\eta^{1}$) is well-defined for any integer $\eta$.
Let $\ket{\psi} = \frac{1}{\sqrt{2}}(\ket{0_L}\ket{0_R}+\ket{1_L}\ket{1_R})$, where $R$ is a reference system.
Then the channel fidelity
\begin{align}
f_\theta :=&\, f({\mathcal{U}}_{S,\theta}\circ {\mathcal{E}}_{{S\leftarrow L}}\circ{\mathcal{U}}_{L,\theta}^\dagger,{\mathcal{E}}_{{S\leftarrow L}}) \nonumber\\
=&\, \min_{\rho} f\big(({\mathcal{U}}_{S,\theta}\!\circ\! {\mathcal{E}}_{{S\leftarrow L}}\!\circ\!{\mathcal{U}}_{L,\theta}^\dagger\!\otimes\!{\mathbbm{1}}_R)(\rho),({\mathcal{E}}_{{S\leftarrow L}}\!\otimes\!{\mathbbm{1}}_R)(\rho)\big)\nonumber\\
\leq&\, f\big(({\mathcal{U}}_{S,\theta}\!\circ\! {\mathcal{E}}_{{S\leftarrow L}}\!\circ\!{\mathcal{U}}_{L,\theta}^\dagger\!\otimes\!{\mathbbm{1}}_R)(\ket{\psi}),({\mathcal{E}}_{{S\leftarrow L}}\!\otimes\!{\mathbbm{1}}_R)(\ket{\psi})\big)\nonumber\\
=&\, \abs{\braket{\psi|W^\dagger U_{S,\theta} W U_{L,\theta}^{\dagger}|\psi}} \nonumber\\
=&\, \abs{\frac{1}{2}\sum_{\eta=0}^{\Delta H_S} \abs{c_\eta^0}^2 e^{-i\eta\theta+i\Delta H_L\theta} + \frac{1}{2}\sum_{\eta=0}^{\Delta H_S} \abs{c_\eta^1}^2 e^{-i\eta\theta}}\nonumber\\
=&\, \abs{\sum_{\eta=-\Delta H_L}^{\Delta H_S} c_\eta e^{-i\eta\theta}} = \abs{c_{\eta_*} \!+\! \sum_{\eta\neq\eta_*} c_\eta e^{-i(\eta-\eta_*)\theta}},
\end{align}
where we define $c_\eta := \frac{1}{2} \abs{c_{\eta+\Delta H_L}^0}^2+ \frac{1}{2}\abs{c_\eta^1}^2$ for $\eta \in [-\Delta H_L,\Delta H_S]$ and choose $\eta_*$ such that $c_{\eta_*} \geq c_{\eta}$ for all $\eta$. Note that there is always a $\theta$ such that $\sum_{\eta\neq\eta_*} c_\eta \cos((\eta-\eta_*)\theta) = 0$ (because the integration of it from $0$ to $2\pi$ is zero) and that $\sum_{\eta\neq\eta_*} c_\eta e^{-i(\eta-\eta_*)\theta}$ is imaginary. Then we must have
\begin{equation}
\min_\theta f_\theta \leq \sqrt{c_{\eta_*}^2 + (1 - c_{\eta_*})^2}.
\end{equation}
To arrive at a non-trivial lower bound on $\delta_\group = \sqrt{1-\min_\theta f_\theta^2}$, we need an upper bound of $\min_\theta f_\theta$ which is smaller than $1$. To this end, we analyze $c_{\eta_*}$ in detail. In particular, we consider two situations:
\begin{enumerate}[(1),wide, labelindent=0pt]
\item $c_{\eta_*} \leq 1/2$ and a constant upper bound on $\min_\theta f_\theta$ exists. We could always find a subset of $\{\eta\}$ denoted by ${\mathfrak{s}}$ such that $ 1/4 \leq \sum_{\eta\in{\mathfrak{s}}} c_\eta \leq 1/2$. To find such a set, we could first include $\eta_*$ in ${\mathfrak{s}}$ and add new elements into ${\mathfrak{s}}$ one by one until their sum is at least $1/4$. Then there is always a $\theta$ such that $(\sum_{\eta\in{\mathfrak{s}}} c_\eta e^{-i\eta\theta})\cdot(\sum_{\eta\notin{\mathfrak{s}}} c_\eta e^{-i\eta\theta})$ is imaginary, in which case $\min_\theta f_\theta \leq \sqrt{(1/4)^2+(3/4)^2} = \sqrt{5/8}$ and we have \begin{equation}\label{eq:delta-1} \delta_\group \geq \sqrt{1-\min_\theta f_\theta^2} \geq \sqrt{3/8}.\end{equation}
\item $c_{\eta_*} > 1/2$. Then $\sqrt{c_{\eta_*}^2 + (1-c_{\eta_*})^2}$ is a monotonically increasing function of $c_{\eta_*}$ and we only need to find an upper bound on $c_{\eta_*}$. To see such an upper bound exists, we first consider the special case where $\epsilon = 0$ and, according to the KL conditions and the HKS condition, $\braket{{\mathfrak{c}}_0|H_S|{\mathfrak{c}}_0} = \braket{{\mathfrak{c}}_1|H_S|{\mathfrak{c}}_1}$ holds true.
On the other hand, if $c_{\eta_*} = \frac{1}{2}\abs{c_{\eta_*+\Delta H_L}^0}^2+ \frac{1}{2}\abs{c_{\eta_*}^1}^2= 1$, we must have $\abs{c_{\eta_*+\Delta H_L}^0}^2 = \abs{c_{\eta_*}^1}^2 = 1$ and $\braket{{\mathfrak{c}}_0|H_S|{\mathfrak{c}}_0} - \braket{{\mathfrak{c}}_1|H_S|{\mathfrak{c}}_1} = \Delta H_L > 0$, leading to a contradition.
In general, to derive a non-trivial upper bound on $c_{\eta_*}$, we first note that $0 \leq \eta^* \leq \Delta H_S - \Delta H_L$ because otherwise either $c_{\eta_*+\Delta H_L}^0 = 0$ or $c_{\eta_*}^1 = 0$ which contradicts with $c_{\eta_*} > 1/2$. We have $\chi = \braket{{\mathfrak{c}}_0|H_S|{\mathfrak{c}}_0} - \braket{{\mathfrak{c}}_1|H_S|{\mathfrak{c}}_1}$ and
\begin{gather*}
\sum_\eta \abs{c_\eta^0}^2 \eta = \sum_\eta \abs{c_\eta^1}^2 \eta + \chi, \\
~\Leftrightarrow~
\Delta H_L - \chi = -\left(1-\abs{c_{\eta_*}^1}^2\right)\left(\eta_* \!-\! \frac{\sum_{\eta\neq\eta_*}{\abs{c_\eta^1}^2\eta}}{\sum_{\eta\neq\eta_*}{\abs{c_\eta^1}^2}}\right) \\
+\left(1-\abs{c_{\eta_*+\Delta H_L}^0}^2\right)\left(\eta_* \!+\! \Delta H_L \!-\! \frac{\sum_{\eta\neq\eta_*+\Delta H_L}{\abs{c_\eta^0}^2\eta}}{\sum_{\eta\neq\eta_*+\Delta H_L}{\abs{c_\eta^0}^2}}\right).
\end{gather*}
Note that both $\bigg|\eta_* - \frac{\sum_{\eta\neq\eta_*}{\abs{c_\eta^1}^2\eta}}{\sum_{\eta\neq\eta_*}{\abs{c_\eta^1}^2}}\bigg|$ and $\bigg|\eta_* + \Delta H_L - \frac{\sum_{\eta\neq\eta_*+\Delta H_L}{\abs{c_\eta^0}^2\eta}}{\sum_{\eta\neq\eta_*+\Delta H_L}{\abs{c_\eta^0}^2}}\bigg|$ are at most $\Delta H_S$. Therefore, $c_{\eta_*} \leq 1 - \abs{\Delta H_L - \chi}/(2\Delta H_S)$ and
\begin{align}
\delta_\group
&\geq \sqrt{1 - \min_\theta f_\theta^2} \geq \sqrt{2 c_{\eta_*}(1-c_{\eta_*})} \\
&\geq \frac{\sqrt{\abs{\Delta H_L - \chi}\left(\Delta H_S - \frac{1}{2}\abs{\Delta H_L - \chi}\right)}}{\Delta H_S}. \label{eq:KL-last-step}
\end{align}
\end{enumerate}
\propref{prop:global-charge} then follows from combining \eqref{eq:delta-1} and \eqref{eq:KL-last-step}.
\end{proof}
\subsection{Bounding charge fluctuation by QEC inaccuracy}
\label{sec:global-1-epsilon}
We now need to establish connections between $\chi$ and the QEC inaccuracy in order to link the global covariance violation to the QEC inaccuracy. We discuss two different methods that achieve this.
\subsubsection{Knill--Laflamme-based method}
\label{sec:global-1-KL}
Intuitively, a non-zero charge fluctuation leads to a violation of the KL conditions (\eqref{eq:KL}), which indicates a non-zero QEC inaccuracy. Therefore, we may bound the QEC inaccuracy through analyzing the deviation from the KL condition. We call this method the KL-based method. Specifically, we have
\begin{proposition}
\label{prop:charge-KL}
Consider an isometric quantum code defined by $\mE_{\StoL}$. Consider physical Hamiltonian $H_S$, logical Hamiltonian $H_L$, and noise channel $\mN_S$. Suppose the HKS condition is satisfied. Then it holds that \begin{equation}
\abs{\chi} \leq 2\epsilon{\mathfrak{J}},
\end{equation}
where ${\mathfrak{J}}$ is a function of $H_S$ and ${\mathcal{N}}_{S}$ defined by
\begin{equation}
\label{eq:def-frakJ}
{\mathfrak{J}}({\mathcal{N}}_S,H_S) := \min_{h: H_S = \sum_{ij} h_{ij} K_{S,i}^\dagger K_{S,j}} \Delta h,
\end{equation}
where $h$ is Hermitian.
\end{proposition}
One can verify that
${\mathfrak{J}}({\mathcal{N}}_S,H_S)$ is efficiently computable using the following semidefinite program~\cite{boyd2004convex}:
\begin{equation}
\begin{split}
\min_{\substack{x,\nu\in{\mathbb{R}}\\h \text{ is Hermitian}}} 2x, \quad\text{s.t. }&\begin{pmatrix}x{\mathbbm{1}} & h-\nu{\mathbbm{1}}\\ h-\nu{\mathbbm{1}} & x{\mathbbm{1}}\end{pmatrix} \geq 0,\\
&H_S = \sum_{ij}h_{ij} K_{S,i}^\dagger K_{S,j}. \\
\end{split}
\end{equation}
The proof of \propref{prop:charge-KL} is partly based on a useful lemma from Ref.~\cite{beny2010general} which connects the QEC inaccuracy $\epsilon$ to the deviation from the KL conditions:
\begin{lemma}[{\cite{beny2010general}}]
\label{lemma:KL-general}
Let $\Pi$ be the projection onto the code subspace of an isometric quantum code ${\mathcal{E}}_{{S\leftarrow L}}$ and the noise channel is ${\mathcal{N}}_{S} = \sum_{i=1}^r K_i(\cdot)K_i^\dagger$.
Then
\begin{equation}
\epsilon({\mathcal{N}}_{S},{\mathcal{E}}_{{S\leftarrow L}}) = \min_{{\Lambda}} \sqrt{1 - f^2({\Lambda},{\Lambda}+{\mathcal{B}})},
\end{equation}
where ${\Lambda}(\rho) = \sum_{ij} \lambda_{ij} \mathrm{Tr}(\rho) \ket{i}\bra{j}$, $({\Lambda}+{\mathcal{B}})(\rho) = {\Lambda}(\rho) + \sum_{ij} \mathrm{Tr}({B }_{ij} {\mathcal{E}}_{{S\leftarrow L}}(\rho) ) \ket{i}\bra{j}$, and $\lambda_{ij}$ and ${B }_{ij}$ are constant numbers and operators satisfying $\Pi K_i^\dagger K_j \Pi = \lambda_{ij} \Pi + \Pi {B }_{ij} \Pi$.
\end{lemma}
\propref{prop:charge-KL} then follows by connecting the deviation from the KL conditions to the charge fluctuation. The proof goes as follows.
\begin{proof}[Proof of {\propref{prop:charge-KL}}]
Let $\Pi$ be the projection onto the code subspace under consideration,
$\ket{{\mathfrak{c}}_{0}} = W\ket{0_L}$, $\ket{{\mathfrak{c}}_{1}} = W\ket{1_L}$ where ${\mathcal{E}}_{{L\leftarrow S}}(\cdot) = W(\cdot)W^\dagger$, and the simplified notations $\braket{K_{S,i}^\dagger K_{S,j}}_{0} := \braket{{\mathfrak{c}}_0|K_{S,i}^\dagger K_{S,j}|{\mathfrak{c}}_0}$ and $\braket{K_{S,i}^\dagger K_{S,j}}_{1} := \braket{{\mathfrak{c}}_1|K_{S,i}^\dagger K_{S,j}|{\mathfrak{c}}_1}$. Assume $\lambda_{ij}$ and ${B }_{ij}$ satisfies $\Pi K_{S,i}^\dagger K_{S,j} \Pi = \lambda_{ij}\Pi + \Pi {B }_{ij} \Pi$. Let $\ket{\psi} = \psi_0\ket{0_L}\ket{0_R} + \psi_1\ket{1_L}\ket{1_R}$ where $\abs{\psi_0}^2 + \abs{\psi_1}^2 = 1$, we have
\begin{align}
&\quad\; f\left({\Lambda}+{\mathcal{B}},{\Lambda}\right) \nonumber\\
&\leq \min_{\psi_{0},\psi_{1}} f\left((({\Lambda}+{\mathcal{B}})\otimes{\mathbbm{1}}_R)(\ket{\psi}),({\Lambda}\otimes{\mathbbm{1}}_R)(\ket{\psi})\right)\nonumber\\
&= \min_{\psi_{0}} f\left(\lambda+{\widetilde{B}},\lambda\right),
\end{align}
where ${\widetilde{B}}_{ij} = \abs{\psi_0}^2 \braket{K_{S,i}^\dagger K_{S,j}}_0 + \abs{\psi_1}^2 \braket{K_{S,i}^\dagger K_{S,j}}_1 - \lambda_{ij}$.
According to the Fuchs--van de Graaf inequality $f(\rho,\sigma) \leq \sqrt{1 - \frac{1}{4}\norm{\rho-\sigma}_1^2}$~\cite{fuchs1999cryptographic},
\begin{align}
\epsilon &= \min_{{\Lambda},{\mathcal{B}}}\sqrt{1-f\left({\Lambda}+{\mathcal{B}},{\Lambda}\right)^2} \nonumber\\
&\geq \min_{\lambda}\max_{\psi_0} \sqrt{1-f\left(\lambda+{\widetilde{B}},\lambda\right)^2}\nonumber\\
&\geq \frac{1}{2}\min_{\lambda}\max_{\psi_0}\norm{{\widetilde{B}}}_1.
\end{align}
According to the HKS condition, $H_S = \sum_{ij} h_{ij}K_{S,i}^\dagger K_{S,j}$ for some Hermitian matrix $h$. Without loss of generality, we assume $h$ is diagonal and $H_S = \sum_{i} h_{ii}K_{S,i}^\dagger K_{S,i}$ because if not, we could always choose another set of Kraus operators that diagonalizes $h$. We can also assume $\max_{i} h_{ii} = - \min_{i} h_{ii} =\frac{\Delta h}{2}$ because we can replace $H_S$ with $H_S -\nu{\mathbbm{1}}$ for any $\nu\in{\mathbb{R}}$. Then we have
\begin{align}
\epsilon
&\geq \frac{1}{2} \min_{\lambda}\max_{\psi_0} \norm{{\widetilde{B}}}_1
\geq \frac{1}{2} \min_{\lambda}\max_{\psi_0} \sum_{i}\abs{{\widetilde{B}}_{ii}} \nonumber\\
&= \frac{1}{2} \min_{\lambda}\max_{\psi_0} \sum_{i}\Big|\abs{\psi_0}^2 \braket{K_{S,i}^\dagger K_{S,i}}_{0} + \nonumber\\ &\quad \qquad \qquad \qquad \quad \abs{\psi_1}^2 \braket{K_{S,i}^\dagger K_{S,i}}_{1} - \lambda_{ii}\Big|\nonumber\\
&\geq \frac{1}{4} \min_{\lambda}\sum_{i} \left|\braket{K_{S,i}^\dagger K_{S,i}}_{0} - \lambda_{ii}\right| + \left|\braket{K_{S,i}^\dagger K_{S,i}}_{1} - \lambda_{ii}\right|\nonumber\\
&\geq \frac{1}{4} \sum_{i} \left|\braket{K_{S,i}^\dagger K_{S,i}}_{0} - \braket{K_{S,i}^\dagger K_{S,i}}_{1}\right|\nonumber\\
&\geq \frac{1}{4} \frac{1}{\max_i h_{ii}} \sum_{i} \abs{h_{ii} \left( \braket{K_{S,i}^\dagger K_{S,j}}_{0} - \braket{K_{S,i}^\dagger K_{S,j}}_{1} \right)} \nonumber\\
&\geq \frac{\abs{\chi}}{2 \Delta h}.
\end{align}
Note that there might be many different choices of $h$ such that $H_S = \sum_{ij} h_{ij}K_{S,i}^\dagger K_{S,j}$ holds true. In order to obtain the tightest lower bound on $\epsilon$, we can minimize $\Delta h$ over all possible $h$ such that $H_S = \sum_{ij} h_{ij}K_{S,i}^\dagger K_{S,j}$, leading to $\epsilon \geq \abs{\chi}/2{\mathfrak{J}}$, where ${\mathfrak{J}} := \min_{h: H_S = \sum_{ij} h_{ij} K_{S,i}^\dagger K_{S,j}} \Delta h$.
\end{proof}
\subsubsection{Quantum metrology method}
\label{sec:global-1-metrology}
Besides the KL-based method, the relationship between the charge fluctuation and the QEC inaccuracy could be understood through the lens of quantum metrology, which results in another inequality concerning $\chi$ and $\epsilon$, as shown in the following. A detailed comparison between the two bounds obtained from the KL-based method and the quantum metrology method (\propref{prop:charge-KL} and \propref{prop:charge-metrology}) will later be given in \secref{sec:global-1-tradeoff} and \secref{sec:global-1-noise}.
\begin{proposition}
\label{prop:charge-metrology}
Consider a quantum code defined by ${\mathcal{E}}_{{L\leftarrow S}}$. Consider physical Hamiltonian $H_S$, logical Hamiltonian $H_L$, and noise channel ${\mathcal{N}}_S$. Suppose the HKS condition is satisfied.
Then it holds that
\begin{equation}
\label{eq:charge-frakB}
\abs{\chi} \leq 2\epsilon\left( \sqrt{(1-\epsilon^2){\mathfrak{F}}} + {\mathfrak{B}}\right).
\end{equation}
Here
\begin{equation}
\label{eq:def-frakB}
{\mathfrak{B}} := \max_{\ket{\psi}}\sqrt{8 {\mathbb{V}}_{H_S}\left({\mathcal{E}}_{{S\leftarrow L}}(\ket{\psi})\right)} \leq \sqrt{2}\Delta H_S,
\end{equation}
where the variance ${\mathbb{V}}_{H}(\rho) := \mathrm{Tr}(H^2\rho) - (\mathrm{Tr}(H\rho))^2$,
and ${\mathfrak{F}}$ is a function of ${\mathcal{N}}_S$ and $H_S$ defined by
\begin{multline}
\label{eq:def-frakF}
{\mathfrak{F}}({\mathcal{N}}_S,H_S) :=
\\ 4\min_{h: H_S = \sum_{ij} h_{ij} K_{S,i}^\dagger K_{S,j}} \bigg\|\sum_{ij}(h^2)_{ij} K_{S,i}^\dagger K_{S,j} - H_S^2\bigg\|,
\end{multline}
where $h$ is Hermitian. In particular, when $\epsilon \ll 1$ and ${\mathfrak{B}} \ll \sqrt{{\mathfrak{F}}}$, we have
\begin{equation}
\abs{\chi} \lesssim 2\epsilon\sqrt{{\mathfrak{F}}}.
\end{equation}
\end{proposition}
Unlike ${\mathfrak{J}}({\mathcal{N}}_S,H_S)$ introduced in \secref{sec:global-1-KL}, ${\mathfrak{F}}({\mathcal{N}}_S,H_S)$ appearing in \propref{prop:charge-metrology} has a clear operational meaning:
\begin{equation}
{\mathfrak{F}}({\mathcal{N}}_S,H_S) \equiv {F^\infty}({\mathcal{N}}_{S,\theta}),\quad \forall \theta \in {\mathbb{R}}.
\end{equation}
Here ${F^\infty}({\mathcal{N}}_{S,\theta})$ is the regularized QFI~\cite{kolodynski2013efficient,zhou2020theory} of the quantum channel ${\mathcal{N}}_{S,\theta} := {\mathcal{N}}_{S}\circ{\mathcal{U}}_{S,\theta}$ where $\theta$ is the unknown parameter to be estimated. (Generally, the regularized QFI of quantum channel $\Phi_\theta$ is defined by ${F^\infty}(\Phi_\theta) := \lim_{N\rightarrow\infty} F(\Phi_\theta^{\otimes N})/N$.) Note that ${F^\infty}({\mathcal{N}}_{S,\theta})$ is independent of $\theta$ and computable using semidefinite programming~\cite{demkowicz2012elusive}. Also, ${F^\infty}({\mathcal{N}}_{S,\theta}) \leq \infty$ if and only if the HKS condition is satisfied. The channel QFI inherits many nice properties from the QFI of quantum states. For example, here ${F^\infty}$ obeys the monotonicity property, i.e.~${F^\infty}({\mathcal{R}}\circ\Phi_\theta\circ{\mathcal{E}}) \leq {F^\infty}(\Phi_\theta)$ for arbitrary parameter-independent channels ${\mathcal{R}}$ and ${\mathcal{E}}$.
{
The operational meaning of the quantity ${\mathfrak{B}}$ is not immediately clear for general encoding channels, but when ${\mathcal{E}}_{{S\leftarrow L}}$ is isometric we have that
\begin{align}
{{\mathfrak{B}}} &= \sqrt{2 F({\mathcal{U}}_{S,\theta}\circ{\mathcal{E}}_{{S\leftarrow L}})}\big|_{\theta = 0},
\end{align}
which satisfies
\begin{equation}
\abs{{\mathfrak{B}} - \sqrt{2}\delta_{\mathrm{P}}}\leq \sqrt{2}\Delta H_L,
\end{equation}
due to the chain rule of the square root of the channel QFI~\cite{katariya2020geometric}:
$\sqrt{F(\Phi_{1,\theta}\circ\Phi_{2,\theta})} \leq \sqrt{F(\Phi_{1,\theta})}+\sqrt{F(\Phi_{2,\theta})}$ for any $\Phi_{1,\theta}$ and $\Phi_{2,\theta}$, and $F({\mathcal{U}}_{L,\theta})= F({\mathcal{U}}_{L,\theta}^\dagger) = (\Delta H_L)^2$.
In general, ${\mathfrak{B}}$ depends on specific encodings and in order to obtain an code-independent bound we should replace ${\mathfrak{B}}$ with its upper bound $\Delta H_S$ so that \eqref{eq:charge-frakB} becomes
\begin{equation}
\label{eq:charge-no-frakB}
\abs{\chi} \leq 2\epsilon\left( \sqrt{(1-\epsilon^2){\mathfrak{F}}} + \sqrt{2}\Delta H_S\right),
\end{equation}
which still leads to useful bounds, e.g., for single-erasure noise as discussed in \secref{sec:global-1-noise}.
However, in many cases ${\mathfrak{B}}$ is negligible, i.e., ${\mathfrak{B}} \ll \sqrt{{\mathfrak{F}}}$ (or $\delta_{\mathrm{P}} \ll \sqrt{{\mathfrak{F}}}$ for isometric codes) in the examples we study later in \secref{sec:case-study}.
}
The monotonicity of the regularized QFI is a key ingredient in the proof of \propref{prop:charge-metrology}. Specifically, we introduce a two-level system $C$ and an ancillary qubit $A$ and consider the channel estimation of the error-corrected noise channel ${\mathcal{N}}_{C,\theta} = {\mathcal{R}}_{{C \leftarrow SA}}\circ({\mathcal{N}}_{S,\theta}\otimes {\mathbbm{1}}_A)\circ{\mathcal{E}}_{{SA\leftarrow C}}$ (see \figref{fig:dephasing}). ${\mathcal{R}}_{{C \leftarrow SA}}$ and ${\mathcal{E}}_{{SA\leftarrow C}}$ is carefully chosen such that ${F^\infty}({\mathcal{N}}_{C,\theta})$ is roughly $\Theta((\chi/\epsilon)^2)$ around $\theta = 0$. Intuitively, one might interpret ${F^\infty}({\mathcal{N}}_{C,\theta})$ handwavily as a quantity proportional to the square of the ``signal-to-noise ratio'' where the QEC inaccuracy $\epsilon$ is roughly the noise rate of ${\mathcal{N}}_{C,\theta}$ and the charge fluctuation $\abs{\chi}$ is roughly the signal strength. \propref{prop:charge-metrology} then follows from the monotonicity of QFI:
\begin{equation}
{F^\infty}({\mathcal{N}}_{C,\theta}) \leq {F^\infty}({\mathcal{N}}_{S,\theta}).
\end{equation}
We now explain the error-corrected metrology protocol in detail. We first introduce an ancilla-assisted two-level encoding. Consider a two-level system $C$ spanned by $\ket{0_C}$ and $\ket{1_C}$ and a Hamiltonian
\begin{equation}
\label{eq:hamt-c}
H_C = \frac{\Delta H_L}{2} \cdot Z_C, \quad {\mathcal{U}}_{C,\theta}(\cdot) = e^{-iH_C\theta} (\cdot) e^{iH_C\theta},
\end{equation}
where $Z_C$ is the Pauli-Z operator.
We define a repetition encoding from $C$ to $LA$,
\begin{equation}
\label{eq:rep-enc}
{\mathcal{E}}^{{\mathrm{rep}}}_{{LA \leftarrow C}}(\ket{0_C}) := \ket{0_L 0_A},\;
{\mathcal{E}}^{{\mathrm{rep}}}_{{LA \leftarrow C}}(\ket{1_C}) := \ket{1_L 1_A},
\end{equation}
where $A$ is the ancillary qubit. The corresponding repetition recovery channel is
\begin{equation}
\label{eq:rep-rec}
{\mathcal{R}}^{{\mathrm{rep}}}_{{C \leftarrow LA}}(\cdot) :=
\sum_{i=0}^{d_L-1} R^{\mathrm{rep}}_i(\cdot)R^{{\mathrm{rep}}\dagger}_i,
\end{equation}
where $R^{\mathrm{rep}}_0 = \ket{0_C}\bra{0_L0_A} + \ket{1_C}\bra{1_L1_A}$ and $R^{\mathrm{rep}}_1 = \ket{0_C}\bra{1_L0_A} + \ket{1_C}\bra{0_L1_A}$ and $R^{\mathrm{rep}}_{i>1} = \ket{0_C}\bra{i_L0_A} + \ket{1_C}\bra{i_L1_A}$. Clearly, ${\mathcal{R}}^{{\mathrm{rep}}}_{{C \leftarrow LA}} \circ {\mathcal{E}}^{{\mathrm{rep}}}_{{LA \leftarrow C}} = {\mathbbm{1}}_C$.
The repetition encoding ${\mathcal{E}}_{{LA \leftarrow C}}^{{\mathrm{rep}}}$ is covariant with respect to $H_C$ and $H_L$, i.e., ${\mathcal{E}}_{{LA \leftarrow C}}^{{\mathrm{rep}}}\circ {\mathcal{U}}_{C,\theta} = ({\mathcal{U}}_{L,\theta}\otimes{\mathbbm{1}}_A) \circ {\mathcal{E}}_{{LA \leftarrow C}}^{{\mathrm{rep}}}$. Moreover, the repetition code corrects all bit-flip noise. When concatenated with ${\mathcal{E}}_{{S\leftarrow L}}$ and ${\mathcal{R}}_{{L\leftarrow S}}$, the error-corrected noisy channel $
{\mathcal{N}}_{C,\theta} = {\mathcal{R}}^{{\mathrm{rep}}}_{{C \leftarrow SA}} \circ ({\mathcal{R}}_{{L\leftarrow S}} \circ{\mathcal{N}}_{S,\theta} \circ {\mathcal{E}}_{{S\leftarrow L}} \otimes {\mathbbm{1}}_A) \circ {\mathcal{E}}_{{LA \leftarrow C}}^{{\mathrm{rep}}}$
becomes a \emph{rotated dephasing channel}, namely, a single-qubit channel which is a composition of dephasing channel $(1-p)(\cdot)+p Z(\cdot)Z$ and a Pauli-$Z$ phase rotation $e^{-i\phi Z}$~\cite{zhou2020new} (see \figref{fig:dephasing}). The regularized QFI of any rotated dephasing channel $\Phi_\theta$ is~\cite{zhou2020theory}
\begin{equation}
{F^\infty}(\Phi_{\theta}) = \frac{\abs{\partial_\theta x_\theta}^2}{1 - \abs{x_{\theta}}^2},
\end{equation}
where the complex number $x_\theta = \braket{0|\Phi_\theta(\ket{0}\bra{1})|1}$. We consider the estimation around $\theta = 0$ for ${\mathcal{N}}_{C,\theta}$. The monotonicity of the regularized QFI guarantees that
\begin{equation}
\label{eq:local-QFI}
{F^\infty}({\mathcal{N}}_{C,\theta})\big|_{\theta = 0} = \frac{\abs{\partial_\theta \xi_\theta}^2\big|_{\theta=0}}{1 - \abs{\xi_{\theta = 0}}^2} \leq {F^\infty}({\mathcal{N}}_{S,\theta}),
\end{equation}
where
\begin{equation}
\xi_\theta := \bra{0_C}{\mathcal{N}}_{C,\theta}(\ket{0_C}\bra{1_C})\ket{1_C}.
\end{equation}
\propref{prop:charge-metrology} can then be proven, connecting $\xi_\theta$ with $\epsilon$ and $\chi$.
\begin{figure}[tb]
\center
\includegraphics[width=0.48\textwidth]{dephasing.pdf}
\caption{\label{fig:dephasing} A two-level encoding scheme for the estimation of $\theta$.
(a) Definition of the encoded channel ${\mathcal{N}}_{C,\theta} $ in the system $C$. ${\mathcal{N}}_{C,\theta} = {\mathcal{R}}_{{C \leftarrow SA}} \circ ({\mathcal{N}}_{S,\theta} \otimes {\mathbbm{1}}_A) \circ {\mathcal{E}}_{{SA\leftarrow C}}$ where ${\mathcal{R}}_{{C \leftarrow SA}} = {\mathcal{R}}_{{LA \leftarrow C}}^{{\mathrm{rep}}} \circ ({\mathcal{R}}^{{\mathrm{opt}}}_{{L\leftarrow S}} \otimes {\mathbbm{1}}_A)$ and ${\mathcal{E}}_{{SA\leftarrow C}} = ({\mathcal{E}}_{{S\leftarrow L}} \otimes {\mathbbm{1}}_A) \circ {\mathcal{E}}_{{LA \leftarrow C}}^{{\mathrm{rep}}}$. It includes a concatenation of the repetition encoding from $C$ to $LA$ and the encoding under consideration ${\mathcal{E}}_{{S\leftarrow L}}$ from $L$ to $S$ with the optimal recovery channels chosen accordingly.
(b) ${\mathcal{N}}_{C,\theta}$ is the composition of ${\mathcal{D}}_{C,\theta}$ and ${\mathcal{U}}_{C,\theta}$ where ${\mathcal{U}}_{C,\theta}$ is the unitary rotation with respect to a Pauli-Z Hamiltonian $H_C$ and ${\mathcal{D}}_{C,\theta}$ is a rotated dephasing channel which is at most $\epsilon$-far from identity at $\theta = 0$.
When $\delta_\group \approx 0$, ${\mathcal{D}}_{C,\theta} \approx {\mathcal{R}}_{{C \leftarrow SA}} \circ ({\mathcal{N}}_{S} \otimes {\mathbbm{1}}_A) \circ {\mathcal{E}}_{{SA\leftarrow C}}$ is almost $\theta$-independent. Note that the gates are applied from right to left.}
\end{figure}
Now we are ready to present the formal proof of \propref{prop:charge-metrology}.
\begin{proof}[Proof of {\propref{prop:charge-metrology}}]
Let ${\mathcal{R}}^{{\mathrm{opt}}}_{{L\leftarrow S}}$ be the recovery channel such that
$\epsilon = P({\mathcal{R}}^{{\mathrm{opt}}}_{{L\leftarrow S}}\circ {\mathcal{N}}_{S}\circ{\mathcal{E}}_{{S\leftarrow L}}, {\mathbbm{1}}_{L})$. Let ${\mathcal{R}}_{{C \leftarrow SA}} = {\mathcal{R}}_{{C \leftarrow SA}}^{\mathrm{rep}} \circ ({\mathcal{R}}_{{L\leftarrow S}}^{{\mathrm{opt}}} \otimes {\mathbbm{1}}_A)$ and ${\mathcal{E}}_{{SA\leftarrow C}} = ({\mathcal{E}}_{{S\leftarrow L}}^{{\mathrm{opt}}} \otimes {\mathbbm{1}}_A) \circ {\mathcal{E}}_{{LA \leftarrow C}}^{\mathrm{rep}} $, we have two rotated dephasing channels (see \figref{fig:dephasing}):
\begin{align}
{\mathcal{N}}_{C,\theta}
&= {\mathcal{R}}_{{C \leftarrow SA}} \circ ({\mathcal{N}}_{S,\theta} \otimes {\mathbbm{1}}_A) \circ {\mathcal{E}}_{{SA\leftarrow C}} \\
&=: {\mathcal{D}}_{C,\theta} \circ {\mathcal{U}}_{C,\theta},
\end{align}
where ${\mathcal{D}}_{C,\theta}$ and ${\mathcal{N}}_{C,\theta}$ are rotated dephasing channels of the following forms:
\begin{gather*}
\begin{split}
{\mathcal{D}}_{C,\theta}(\cdot) =:\;& (1 - p_\theta) e^{-i\frac{\phi_\theta}{2} Z_C} (\cdot) e^{i\frac{\phi_\theta}{2}Z_C} \\
& + p_\theta Z_C e^{-i\frac{\phi_\theta}{2}Z_C} (\cdot) e^{i\frac{\phi_\theta}{2}Z_C} Z_C,
\end{split}\\
\begin{split}
{\mathcal{N}}_{C,\theta}(\cdot) =:\;& (1 - p_\theta) e^{-i\frac{\phi_\theta+\Delta H_L\theta}{2} Z_C} (\cdot) e^{i\frac{\phi_\theta+\Delta H_L\theta}{2} Z_C} \\
& + p_\theta Z_C e^{-i\frac{\phi_\theta+\Delta H_L\theta}{2} Z_C} (\cdot) e^{i\frac{\phi_\theta+\Delta H_L\theta}{2} Z_C} Z_C,
\end{split}
\end{gather*}
and $\xi_\theta = \bra{0_C}{\mathcal{D}}_{C,\theta}\circ{\mathcal{U}}_{C,\theta}(\ket{0_C}\bra{1_C})\ket{1_C} = (1 - 2p_\theta)e^{-i(\phi_\theta+\Delta H_L\theta)}$. ${\mathcal{D}}_{C,\theta}$ is identity when the code is both exactly covariant and exactly error-correcting; it is $\theta$-independent when the code is exactly covariant (see also Ref.~\cite{zhou2020new}). {Note that we will not use the channel ${\mathcal{D}}_{C,\theta}$ in this proof (it will be used later in \secref{sec:global-2}), but we introduce the notation here
to clarify its physical meaning. }
Consider the parameter estimation of $\theta$ in the neighborhood of $\theta = 0$. On one hand, for rotated dephasing channels (see \appref{app:dephasing} for the purified distance of rotated dephasing channels from identity), we have
\begin{equation}
\label{eq:noise-rate-epsilon-1}
\sqrt{p_{\theta=0}} \leq P({\mathcal{D}}_{C,\theta = 0},{\mathbbm{1}}_C) = P({\mathcal{N}}_{C,\theta = 0},{\mathbbm{1}}_C).
\end{equation}
On the other hand,
we have
\begin{align}
&\quad\; P({\mathcal{N}}_{C,\theta=0},{\mathbbm{1}}_C) \nonumber\\
&= P({\mathcal{R}}_{{C \leftarrow SA}} \circ ({\mathcal{N}}_{S} \otimes {\mathbbm{1}}_A) \circ {\mathcal{E}}_{{SA\leftarrow C}} ,{\mathbbm{1}}_C)
\nonumber\\
&\leq P({\mathcal{R}}^{{\mathrm{opt}}}_{{L\leftarrow S}} \circ {\mathcal{N}}_{S} \circ {\mathcal{E}}_{{S\leftarrow L}}, {\mathbbm{1}}_L) \leq \epsilon,
\label{eq:noise-rate-epsilon-2}
\end{align}
where we use the monotonicity of the purified distance~\cite{tomamichel2015quantum} and the definition of $\epsilon$. Combining \eqref{eq:noise-rate-epsilon-1} and \eqref{eq:noise-rate-epsilon-2}, we have
\begin{equation}
\label{eq:xi-epsilon}
\abs{\xi_{\theta = 0}} \geq 1 - 2\epsilon^2.
\end{equation}
As shown in \appref{app:local-proof}, we also have
\begin{equation}
\abs{\partial_\theta\xi_\theta}^2\big|_{\theta = 0} \geq (\abs{\chi}-2\epsilon{\mathfrak{B}})^2,
\end{equation}
when $\abs{\chi} \geq 2\epsilon{\mathfrak{B}}$.
Hence, when $\abs{\chi} \geq 2\epsilon{\mathfrak{B}}$, we must have
\begin{align}
{F^\infty}({\mathcal{N}}_{S,\theta})
&= {F^\infty}({\mathcal{N}}_{S,\theta}) \big|_{\theta = 0} \nonumber\\
&\geq {F^\infty}({\mathcal{N}}_{C,\theta})\big|_{\theta = 0} \geq \frac{(\abs{\chi} -2\epsilon{\mathfrak{B}})^2}{4\epsilon^2(1-\epsilon^2)},
\end{align}
completing the proof.
\end{proof}
\subsection{Consequent bounds on the trade-off between QEC and global covariance}
\label{sec:global-1-tradeoff}
In \secref{sec:global-1-charge} and \secref{sec:global-1-epsilon}, we derived bounds on $\delta_\group$ and $\epsilon$ separately, using the notion of charge fluctuation $\chi$. Combining these results, we immediately obtain the trade-off relations between $\delta_\group$ and $\epsilon$.
\begin{theorem}
\label{thm:global-1}
Consider an isometric quantum code defined by $\mE_{\StoL}$. Consider physical Hamiltonian $H_S$, logical Hamiltonian $H_L$, and noise channel $\mN_S$. Suppose the HKS condition is satisfied. It holds that, when $0 \leq G(\epsilon) \leq \Delta H_S$,
\begin{equation}
\label{eq:tradeoff-1}
\delta_\group \geq
\min\left\{\frac{\sqrt{G(\epsilon)\left(\Delta H_S - \frac{1}{2}G(\epsilon)\right)}}{\Delta H_S},\sqrt{\frac{3}{8}}\right\},
\end{equation}
and when $G(\epsilon) > \Delta H_S$, $\delta_\group \geq \sqrt{3/8}$,
where we could take either
\begin{equation}
G(\epsilon) = \Delta H_L - 2\epsilon {\mathfrak{J}},
\end{equation}
or
\begin{equation}
G(\epsilon) = \Delta H_L - 2 \epsilon (\sqrt{(1-\epsilon^2){\mathfrak{F}}}+{\mathfrak{B}}),
\end{equation}
where ${\mathfrak{J}},{\mathfrak{F}}$ and ${\mathfrak{B}}$ are given by \eqref{eq:def-frakJ}, \eqref{eq:def-frakF} and \eqref{eq:def-frakB} respectively.
\end{theorem}
For the extreme cases of exactly covariant codes and exactly error-correcting codes, we have the following corollaries:
\begin{corollary}
\label{col:global-1-covariant}
Consider an isometric quantum code defined by $\mE_{\StoL}$. Consider physical Hamiltonian $H_S$, logical Hamiltonian $H_L$, and noise channel $\mN_S$. Suppose the HKS condition is satisfied. When $\epsilon = 0$, i.e., the code is exactly error-correcting, it holds that when $\Delta H_L \leq \Delta H_S$,
\begin{equation}
\delta_\group
\geq \frac{\sqrt{\Delta H_L\left(\Delta H_S -\frac{1}{2}\Delta H_L\right)}}{\Delta H_S},
\end{equation}
and when $\Delta H_L > \Delta H_S$, $\delta_\group \geq \sqrt{3/8}$.
\end{corollary}
\begin{corollary}
\label{col:global-1-qec}
Consider an isometric quantum code defined by $\mE_{\StoL}$. Consider physical Hamiltonian $H_S$, logical Hamiltonian $H_L$, and noise channel $\mN_S$. Suppose the HKS condition is satisfied.
When $\delta_\group = 0$, i.e., the code is exactly covariant, it holds that
\begin{equation}
\label{eq:epsilon-covariant-1}
\epsilon \geq \frac{\Delta H_L}{2 {\mathfrak{J}}},
\end{equation}
where ${\mathfrak{J}}$ is given by \eqref{eq:def-frakJ}.
\end{corollary}
\begin{corollary}
\label{col:global-1-qec-2}
Consider a quantum code defined by $\mE_{\StoL}$. Consider physical Hamiltonian $H_S$, logical Hamiltonian $H_L$, and noise channel $\mN_S$. Suppose the HKS condition is satisfied.
When $\delta_\group = 0$, i.e., the code is exactly covariant, we must have either $2\epsilon{\mathfrak{B}} \geq \Delta H_L$ or
\begin{equation}
\label{eq:epsilon-covariant-2}
\epsilon\frac{\sqrt{1-\epsilon^2}}{1-2\epsilon{\mathfrak{B}}/\Delta H_L} \geq \frac{\Delta H_L}{2\sqrt{{\mathfrak{F}}}},
\end{equation}
where ${\mathfrak{F}}$ and ${\mathfrak{B}}$ are given by \eqref{eq:def-frakF} and \eqref{eq:def-frakB} respectively.
\end{corollary}
We make a few remarks on the scope of application of these results. Although \propref{prop:global-charge} and \propref{prop:charge-KL} need the isometric encoding assumption, \propref{prop:charge-metrology} (and thus \corollaryref{col:global-1-qec-2}) holds for arbitrary codes. Also, \propref{prop:global-charge} only holds when $H_L$ and $H_S$ share a common period, but \propref{prop:charge-KL} and \propref{prop:charge-metrology} hold true for arbitrary Hamiltonians without the $U(1)$ assumption. {Finally, a keen reader might have already noticed that the choice of the pair of orthonormal states $\{\ket{0_L},\ket{1_L}\}$ in the proofs of \propref{prop:global-charge}, \propref{prop:charge-KL} and \propref{prop:charge-metrology} is quite arbitrary (chosen only for the purpose of proving \thmref{thm:global-1}) and we could in principle replace it with any other pair and the proofs will still hold, leading to refinements of these propositions. We present these refinements in detail in \appref{app:refine-2}. In particular, \propref{prop:global-charge} leads to an inequality between $\delta_\group$ and $\delta_{\mathrm{C}}$. }
{
To compare the results from the KL-based method and the quantum metrology method, we first consider the limiting situation where $\delta_\group \ll 1$ and $\epsilon \ll 1$.
Then we have
\begin{align}
\label{eq:frakJ-global-asym}
&\delta_\group \gtrsim \sqrt{\frac{\Delta H_L - 2\epsilon {\mathfrak{J}}}{\Delta H_S}} ~\Leftrightarrow~ \epsilon + \delta_\group^2 \frac{\Delta H_S}{2{\mathfrak{J}}} \gtrsim \frac{\Delta H_L}{2{\mathfrak{J}}},\\
&\delta_\group \gtrsim \sqrt{\frac{\Delta H_L - 2\epsilon (\sqrt{{\mathfrak{F}}}+{\mathfrak{B}})}{\Delta H_S}} ~\Leftrightarrow~ \nonumber\\
\label{eq:frakF-global-asym}
&\qquad \qquad \qquad \quad \epsilon + \delta_\group^2 \frac{\Delta H_S}{2(\sqrt{{\mathfrak{F}}}+{\mathfrak{B}})} \gtrsim \frac{\Delta H_L}{2(\sqrt{{\mathfrak{F}}}+{\mathfrak{B}})}.
\end{align}
When ${\mathfrak{B}} \ll \sqrt{{\mathfrak{F}}}$, the metrology bound performs no worse than the KL-based bound because we always have ${\mathfrak{J}}^2 \geq {\mathfrak{F}}$ (proof in \appref{app:inequality}).
For the examples we study later in \secref{sec:case-study}, we find that ${\mathfrak{B}}$ is negligible, but in practice one may need to bound the parameter ${\mathfrak{B}}$ a priori using properties of specific codes to obtain desired trade-off relations. It still open in general under which conditions ${\mathfrak{B}} \ll \sqrt{{\mathfrak{F}}}$ holds, and whether \propref{prop:charge-metrology} might be further improved with ${\mathfrak{B}}$ removed. }
A byproduct of our results are lower bounds on $\epsilon$ (\eqref{eq:epsilon-covariant-1} and \eqref{eq:epsilon-covariant-2}) for exactly covariant codes, a special case which has been extensively studied in previous works~\cite{faist2019continuous,woods2019continuous,kubica2020using,zhou2020new,yang2020covariant}. As discussed below, the bound \eqref{eq:epsilon-covariant-1} for random local erasure noise behaves almost the same as the one in Ref.~\cite{faist2019continuous} and our \propref{prop:charge-KL} provides an alternative method to obtain this result. However, compared to \propref{prop:charge-metrology}, the bound in Ref.~\cite{zhou2020new}
\begin{equation}
\label{eq:epsilon-covariant-3}
\epsilon\frac{\sqrt{1-\epsilon^2}}{1 - 2\epsilon^2} \geq \frac{\Delta H_L}{2\sqrt{{\mathfrak{F}}}}
\end{equation}
does not involve the parameter ${\mathfrak{B}}$, implying that the proof of our \propref{prop:charge-metrology} might be further improved.
\subsection{Noise models and explicit behaviors of the bounds}
\label{sec:global-1-noise}
Now we explicitly discuss how the bounds in \thmref{thm:global-1} behave under difference types of noise in an $n$-partite system.
We consider 1-local Hamiltonians $H_S = \sum_{l=1}^n H_{S_l}$, so $\Delta H_S = O(n)$. In this case we have $\delta_\group = \Omega(1/\sqrt{\Delta H_S}) = \Omega(1/\sqrt{n})$ as long as $G(\epsilon) = \Omega(1)$. On the other hand, when $G(\epsilon) = o(1)$, the scaling of $\epsilon$ must be lower bounded by $\Omega(1/{\mathfrak{J}})$ {(or $\Omega(1/(\sqrt{{\mathfrak{F}}}+{\mathfrak{B}}))$)} so it is important to understand the scalings of ${\mathfrak{J}}$, ${\mathfrak{F}}$ and ${\mathfrak{B}}$. When $n$ is large, the values of ${\mathfrak{J}}$ and ${\mathfrak{F}}$ may be not efficiently computable. However, in \propref{prop:charge-KL}, \propref{prop:charge-metrology} and \thmref{thm:global-1}, we could always replace them with their efficiently computable upper bounds and the trade-off relations still hold then.
We discuss the following two general noise models~\cite{faist2019continuous,woods2019continuous,kubica2020using,zhou2020new,yang2020covariant} (there are still other types of noises that we will not cover, e.g., random long-range phase errors~\cite{woods2019continuous}):
\begin{enumerate}[(1),wide, labelwidth=0pt, labelindent=0pt]
\item Random local noise, where different local noise channels acting on a constant number of subsystems randomly. Specifically, ${\mathcal{N}}_S = \sum_{i} q_i {\mathcal{N}}_{S}^{(i)}$ and $H_S = \sum_i H_{S}^{(i)}$, where ${\mathcal{N}}_{S}^{(i)}$ represent the local noise channels acting on a constant number of subsystems, $q_i$ represent their probabilities ($q_i>0$ and $\sum_i q_i = 1$), and the HKS condition is satisfied for each pair of $({\mathcal{N}}_{S}^{(i)},H_{S}^{(i)})$. Then we have
\begin{align}
\label{eq:frakJ-local}
{\mathfrak{J}}({\mathcal{N}}_S,H_S) &\leq \max_i \frac{1}{q_i}{\mathfrak{J}}({\mathcal{N}}_S^{(i)},H_S^{(i)}),\\
\label{eq:frakF-local}
{\mathfrak{F}}({\mathcal{N}}_S,H_S) &\leq \sum_i \frac{1}{q_i}\widetilde{{\mathfrak{F}}}({\mathcal{N}}_S^{(i)},H_S^{(i)}),
\end{align}
where
\begin{multline}
\widetilde{{\mathfrak{F}}}({\mathcal{N}}_S,H_S) := \\
4\min_{h: H_S = \sum_{ij} h_{ij} K_{S,i}^\dagger K_{S,j}} \bigg| \sum_{ij}(h^2)_{ij} K_{S,i}^\dagger K_{S,j} \bigg|.
\end{multline}
We prove \eqref{eq:frakJ-local} in \appref{app:frakJ} and \eqref{eq:frakF-local} was previously known in Ref.~\cite{zhou2020new}.
For example, consider single erasure noise in an $n$-partite system and let the erasure channel of the $\ell$-th subsystem be ${\mathcal{N}}_{S_l}(\cdot) = \ket{{\emptyset}}\bra{{\emptyset}}_{S_l} \otimes \mathrm{Tr}_{S_l}(\cdot)$ (we use $\ket{{\emptyset}}$ to represent the vacuum state after erasure). When ${\mathcal{N}}_S = \sum_{l=1}^n \frac{1}{n} {\mathcal{N}}_{S_l}$ and the Hamiltonian takes the 1-local form $H_S = \sum_{l=1}^n H_{S_l}$, we have ${\mathfrak{J}}({\mathcal{N}}_{S_l},H_{S_l}) = \Delta H_{S_l}$ and ${\mathfrak{F}}({\mathcal{N}}_{S_l},H_{S_l}) = (\Delta H_{S_l})^2$. Then we have
\begin{align}
\label{eq:erasure-J}
{\mathfrak{J}} &\leq n\max_l \Delta H_{S_l},\\
{\mathfrak{F}} &\leq n\sum_{l=1}^n (\Delta H_{S_l})^2, \\
\label{eq:erasure-frakB}
\sqrt{{\mathfrak{F}}}+{\mathfrak{B}} &\leq n\!\left(\sqrt{\frac{\sum_{l=1}^n (\Delta H_{S_l})^2}{n}} + \sqrt{2}{\frac{\Delta H_S}{n}} \right).
\end{align}
Note that using \eqref{eq:erasure-J} and \eqref{eq:epsilon-covariant-1}, we obtain $\epsilon \geq \Delta H_L/(2n\max_l \Delta H_{S_l})$ which is identical to Theorem~1 in Ref.~\cite{faist2019continuous}.
{In \eqref{eq:erasure-frakB}, we use ${\mathfrak{B}} \leq \sqrt{2}{\Delta H_S}$. Comparing \eqref{eq:erasure-frakB} with \eqref{eq:erasure-J}, we find that when ${\mathfrak{B}}$ is not negligible, the quantum metrology method can still outperform the KL-based method in some cases (e.g., when one of $\Delta H_{S_l}$ is extremely large).
For other types of random local noise acting on each subsystem uniformly randomly, we also have ${\mathfrak{J}}=O(n)$ and $\sqrt{{\mathfrak{F}}}+{\mathfrak{B}}=O(n)$ and the behaviors of the trade-off relations from the KL-based method and the quantum metrology method are similar.} {From~\eqref{eq:frakJ-global-asym} and \eqref{eq:frakF-global-asym}, we have $\epsilon + \Theta(1)\cdot \delta_\group^2 = \Omega(1/n)$, meaning that when both $\epsilon$ and $\delta_\group$ are sufficiently small, their optimal scalings are $\Theta(1/n)$ and $\Theta(1/\sqrt{n})$ respectively. }
\item Independent noise, where noise channels act on each subsystem independently. Note that independent noise is considered a ``stronger'' noise model than local noise because the noise actions are no longer guaranteed to be local. Specifically, ${\mathcal{N}}_S = \bigotimes_{l=1}^n {\mathcal{N}}_{S_l}$ and $H_S = \sum_{l=1}^n H_{S_l}$ where ${\mathcal{N}}_{S_l}$ represent independent noise channels acting on each subsystem and $H_{S_l}$ acts only non-trivially on the subsystem $S_l$, and the HKS condition is satisfied for each pair of $({\mathcal{N}}_{S_l},H_{S_l})$. Here we have
\begin{align}
\label{eq:frakJ-independent}
{\mathfrak{J}}({\mathcal{N}}_S,H_S) &\leq \sum_{l=1}^n {\mathfrak{J}}({\mathcal{N}}_{S_l},H_{S_l}),\\
\label{eq:frakF-independent}
{\mathfrak{F}}({\mathcal{N}}_S,H_S) &= \sum_{l=1}^n {\mathfrak{F}}({\mathcal{N}}_{S_l},H_{S_l}).
\end{align}
The proof of \eqref{eq:frakJ-independent} is provided in \appref{app:frakJ}, and \eqref{eq:frakF-independent} follows directly from the additivity of the regularized QFI ${\mathfrak{F}} = {F^\infty}(\bigotimes_{l=1}^n{\mathcal{N}}_{S_l,\theta}) = \sum_{l=1}^n {F^\infty}({\mathcal{N}}_{S_l,\theta})$~\cite{zhou2020new}. Therefore we have ${\mathfrak{J}} = O(n)$ and $\sqrt{{\mathfrak{F}}} = O(\sqrt{n})$ and there is now a quadratic gap between them.
{If ${\mathfrak{B}}$ can be upper bounded by $O(\sqrt{n})$ (e.g., in \secref{sec:case-study}), from \eqref{eq:frakF-global-asym}, we have $\epsilon + \Theta(\sqrt{n})\cdot \delta_\group^2 = \Omega(1/\sqrt{n})$, meaning that when both $\epsilon$ and $\delta_\group$ are sufficiently small, their optimal scalings are both $\Theta(1/\sqrt{n})$. From \eqref{eq:frakJ-global-asym}, we only have $\epsilon + \Theta(1)\cdot \delta_\group^2 = \Omega(1/n)$ and a worse lower bound $\epsilon = \Omega(1/n)$ for small $\delta_\group$. In general, for independent noise, the trade-off bound from the KL-based method is asymptotically weaker than the one from the quantum metrology method as long as ${\mathfrak{B}} = o(n)$. }
\end{enumerate}
\section{Trade-off between QEC and global covariance: gate implementation error approach}
\label{sec:global-2}
In this section, we introduce another framework that also enables us to derive the trade-off between the QEC inaccuracy $\epsilon$ and the global covariance violation $\delta_\group$ and could be interesting in its own right. Here the idea is to analyze a key notion which we call the gate implementation error $\gamma$ that allow us to treat $\epsilon$ and $\delta_\group$ on the same footing. More specifially, we first formally define $\gamma$ in \secref{sec:global-2-gate} and show that $\epsilon + \delta_\group \geq \gamma$. Then we derive two lower bounds on $\gamma$ using two different methods from quantum metrology and quantum resource theory, which automatically induce two trade-off relations between the QEC inaccuracy and the global covariance violation. We will compare the gate implementation error approach to the charge fluctuation approach at the end of this section.
\subsection{Gate implementation error as a unification of QEC inaccuracy and global covariance violation}
\label{sec:global-2-gate}
Consider a practical quantum computing scenario where we want to implement a set of logical gates $U_{L,\theta}= e^{-iH_L\theta}$ for $\theta \in {\mathbb{R}}$ using physical gates $U_{S,\theta} = e^{-iH_S\theta}$ under noise ${\mathcal{N}}_S$. We would like to design an encoding and a recovery channel such that ${\mathcal{R}}_{{L\leftarrow S}}\circ{\mathcal{N}}_{S,\theta}\circ{\mathcal{E}}_{{S\leftarrow L}}$ simulate ${\mathcal{U}}_{L,\theta}$. We call the error in such simulations the \emph{gate implementation error} and the \emph{Choi gate implementation error}, defined by
\begin{align}
\gamma
&:= \min_{{\mathcal{R}}_{{L\leftarrow S}}} \max_\theta P({\mathcal{R}}_{{L\leftarrow S}}\circ {\mathcal{N}}_{S,\theta}\circ{\mathcal{E}}_{{S\leftarrow L}}, {\mathcal{U}}_{L,\theta}),
\\
\overline{\gamma}
&:= \min_{{\mathcal{R}}_{{L\leftarrow S}}} \max_\theta \overline{P}({\mathcal{R}}_{{L\leftarrow S}}\circ {\mathcal{N}}_{S,\theta}\circ{\mathcal{E}}_{{S\leftarrow L}}, {\mathcal{U}}_{L,\theta}).
\end{align}
Both the QEC inaccuracy and the covariance violation contribute to the gate implementation error (see \figref{fig:gate}). Clearly, $\gamma = 0$ when the quantum code is exactly error-correcting and covariant. In general, $\gamma$ is upper bounded by the sum of $\epsilon$ and $\delta_\group$, as shown in the following
proposition.
\begin{proposition}
\label{prop:gate}
Consider a quantum code defined by ${\mathcal{E}}_{{L\leftarrow S}}$. Consider physical Hamiltonian $H_S$, logical Hamiltonian $H_L$, and noise channel ${\mathcal{N}}_S$. It holds that
\begin{gather}
\epsilon + \delta_\group \geq \gamma,\\
\overline{\epsilon} + \overline{\delta}_\group \geq \overline{\gamma}.
\end{gather}
\end{proposition}
\begin{proof}
Using the triangular inequality of the purified distance~\cite{tomamichel2015quantum}, we have
\begin{align}
&P({\mathcal{R}}_{{L\leftarrow S}}\circ {\mathcal{N}}_{S,\theta}\circ{\mathcal{E}}_{{S\leftarrow L}}, {\mathcal{U}}_{L,\theta}) \nonumber \\ \leq\;& P({\mathcal{R}}_{{L\leftarrow S}}\circ {\mathcal{N}}_{S,\theta}\circ{\mathcal{E}}_{{S\leftarrow L}}, {\mathcal{R}}_{{L\leftarrow S}}\circ {\mathcal{N}}_{S}\circ{\mathcal{E}}_{{S\leftarrow L}}\circ{\mathcal{U}}_{L,\theta}) \nonumber\\&+ P({\mathcal{R}}_{{L\leftarrow S}}\circ {\mathcal{N}}_{S}\circ{\mathcal{E}}_{{S\leftarrow L}}\circ{\mathcal{U}}_{L,\theta}, {\mathcal{U}}_{L,\theta}).
\end{align}
The first term is upper bounded by $P({\mathcal{U}}_{S,\theta}\circ{\mathcal{E}}_{{S\leftarrow L}}, {\mathcal{E}}_{{S\leftarrow L}}\circ{\mathcal{U}}_{L,\theta})$ using the monotonicity of the purified distance and the second term is equal to $P({\mathcal{R}}_{{L\leftarrow S}}\circ {\mathcal{N}}_{S}\circ{\mathcal{E}}_{{S\leftarrow L}}, {\mathbbm{1}}_L)$ by definition. Then $\gamma \leq \epsilon + \delta_\group$ follows by taking the maximization over $\theta$ and the minimization over ${\mathcal{R}}_{{L\leftarrow S}}$ on both sides. The above discussion also holds when we replace the purified distance $P(\cdot,\cdot)$ with the Choi purified distance $\overline{P}(\cdot,\cdot )$, implying that $\overline{\gamma} \leq \overline{\epsilon} + \overline{\delta}_\group$.
\end{proof}
\begin{figure}[tb]
\center
\includegraphics[width=0.4\textwidth]{gate-0.pdf}
\caption{\label{fig:gate}
Both the QEC inaccuracy $\epsilon$ and the covariance violation $\delta_\group$ contribute to the error in approximate gate implementation. Specifically, the (Choi) gate implementation error $\gamma$ ($\overline{\gamma}$) is upper bounded by $\delta_\group + \epsilon$ ($\overline{\delta}_\group + \overline{\epsilon}$). Note that the gates are applied from right to left.}
\end{figure}
\subsection{Bounding gate implementation error}
\subsubsection{Quantum metrology method}
\label{sec:global-2-metrology}
Now we derive a lower bound on the gate implementation error $\gamma$, where we consider the approximate gate implementation of ${\mathcal{U}}_{L,\theta}$ using noisy gates ${\mathcal{N}}_{S,\theta}$ as an error-corrected metrology protocol where $\theta$ is an unknown parameter to be estimated.
Again, we use the ancilla-assisted two-level encoding, as introduced in \secref{sec:global-1-metrology}. We choose the repetition code concatenated with the quantum code under study, so that the error-corrected noise channel $
{\mathcal{N}}_{C,\theta} = {\mathcal{R}}^{{\mathrm{rep}}}_{{C \leftarrow SA}} \circ ({\mathcal{R}}_{{L\leftarrow S}} \circ{\mathcal{N}}_{S,\theta} \circ {\mathcal{E}}_{{S\leftarrow L}} \otimes {\mathbbm{1}}_A) \circ {\mathcal{E}}_{{LA \leftarrow C}}^{{\mathrm{rep}}}$ becomes a rotated dephasing channel.
The main difference between the error-corrected metrology protocol we use here and the one in \secref{sec:global-1-metrology} is that now we choose the recovery channel ${\mathcal{R}}_{{L\leftarrow S}}$ to be the optimal recovery channel which minimizes the gate implementation error (instead of the QEC inaccuracy) and guarantees a lower noise rate over the entire group of $\theta$ (instead of just around $\theta = 0$). In this case, we show that there always exists some $\theta_*$ such that ${F^\infty}({\mathcal{N}}_{C,\theta_*}) = \Theta((\Delta H_L/\gamma)^2)$, which then provide us a lower bound on $\gamma$ using the monotonicity of the regularized QFI. Now we state and prove \thmref{thm:global-2-metrology} which provides a lower bound on $\gamma$ (and thus on $\epsilon + \delta_\group$).
\begin{theorem}
\label{thm:global-2-metrology}
Consider a quantum code defined by ${\mathcal{E}}_{{L\leftarrow S}}$. Consider physical Hamiltonian $H_S$, logical Hamiltonian $H_L$, and noise channel ${\mathcal{N}}_S$. Suppose the HKS condition is satisfied. Then it holds that
\begin{equation}
\epsilon + \delta_\group \geq \gamma \geq \ell_1\left(\frac{\Delta H_L}{2\sqrt{{\mathfrak{F}}}
}\right),
\end{equation}
where ${\mathfrak{F}}$ is given by \eqref{eq:def-frakF}, $\ell_1(x) = x + O(x^2)$ is the inverse function of the monotonic function $x=y\frac{\sqrt{1-y^2}}{1-2y^2}$ on $[0,1/\sqrt{2})$.
\end{theorem}
In particular, for exact QEC codes, we have the following corollary:
\begin{corollary}
\label{col:global-2-covariant-metrology}
Consider a quantum code defined by $\mE_{\StoL}$. Consider physical Hamiltonian $H_S$, logical Hamiltonian $H_L$, and noise channel $\mN_S$. Suppose the HKS condition is satisfied. When $\epsilon = 0$, i.e., when the code is exactly error-correcting, it holds that $\delta_\group \geq \ell_1(\Delta H_L/2\sqrt{{\mathfrak{F}}}
)$, where ${\mathfrak{F}}$ is given by \eqref{eq:def-frakF}.
\end{corollary}
\begin{proof}[Proof of {\thmref{thm:global-2-metrology}}]
By definition, there exists a ${\mathcal{R}}^{{\mathrm{opt'}}}_{{L\leftarrow S}}$ such that
$\gamma = \max_\theta P({\mathcal{R}}^{{\mathrm{opt'}}}_{{L\leftarrow S}}\circ {\mathcal{N}}_{S,\theta}\circ{\mathcal{E}}_{{S\leftarrow L}}, {\mathcal{U}}_{L,\theta})$.
Let ${\mathcal{R}}_{{C \leftarrow SA}} = {\mathcal{R}}_{{C \leftarrow SA}}^{\mathrm{rep}} \circ ({\mathcal{R}}_{{L\leftarrow S}}^{{\mathrm{opt'}}} \otimes {\mathbbm{1}}_A)$ and ${\mathcal{E}}_{{SA\leftarrow C}} = ({\mathcal{E}}_{{S\leftarrow L}}^{{\mathrm{opt'}}} \otimes {\mathbbm{1}}_A) \circ {\mathcal{E}}_{{LA \leftarrow C}}^{\mathrm{rep}} $, we have ${\mathcal{N}}_{C,\theta}
= {\mathcal{R}}_{{C \leftarrow SA}} \circ ({\mathcal{N}}_{S,\theta} \otimes {\mathbbm{1}}_A) \circ {\mathcal{E}}_{{SA\leftarrow C}}
= {\mathcal{D}}_{C,\theta} \circ {\mathcal{U}}_{C,\theta}$,
and ${\mathcal{D}}_{C,\theta}$ and ${\mathcal{N}}_{C,\theta}$ are rotated dephasing channels of the following forms:
\begin{gather*}
\begin{split}
{\mathcal{D}}_{C,\theta}(\cdot) =\;& (1 - p_\theta) e^{-i\frac{\phi_\theta}{2} Z_C} (\cdot) e^{i\frac{\phi_\theta}{2}Z_C} \\
& + p_\theta Z_C e^{-i\frac{\phi_\theta}{2}Z_C} (\cdot) e^{i\frac{\phi_\theta}{2}Z_C} Z_C,
\end{split}\\
\begin{split}
{\mathcal{N}}_{C,\theta}(\cdot) =\;& (1 - p_\theta) e^{-i\frac{\phi_\theta+\Delta H_L\theta}{2} Z_C} (\cdot) e^{i\frac{\phi_\theta+\Delta H_L\theta}{2} Z_C} \\
& + p_\theta Z_C e^{-i\frac{\phi_\theta+\Delta H_L\theta}{2} Z_C} (\cdot) e^{i\frac{\phi_\theta+\Delta H_L\theta}{2} Z_C} Z_C.
\end{split}
\end{gather*}
where $p_\theta \in (0,1)$ and $\phi_\theta \in [0,2\pi)$.
Let $
\xi_\theta
= \bra{0_C}{\mathcal{N}}_{C,\theta}(\ket{0_C}\bra{1_C})\ket{1_C} = (1 - 2p_\theta)e^{-i(\phi_\theta+\Delta H_L\theta)}$.
The regularized channel QFI of rotated dephasing channels is
\begin{align}
\label{eq:dephasing-SLD}
&{F^\infty}({\mathcal{N}}_{C,\theta}) = \frac{\abs{\partial_\theta \xi_\theta}^2}{1 - \abs{\xi_\theta}^2} \\&\quad = \frac{(1-2p_\theta)^2 (\partial_\theta\phi_\theta + \Delta H_L)^2 }{4p_\theta(1-p_\theta)} + \frac{(\partial_\theta p_\theta)^2 }{4p_\theta(1-p_\theta)}.
\end{align}
In order to get a lower bound on ${F^\infty}({\mathcal{N}}_{C,\theta})$ as a function of $\gamma$. We note that the purified distance between ${\mathcal{D}}_{C,\theta}$ and ${\mathbbm{1}}_C$ (see \appref{app:dephasing}) is upper bounded by $\gamma$:
\begin{equation}
\label{eq:purified}
P({\mathcal{D}}_{C,\theta},{\mathbbm{1}}_C) = \sqrt{\frac{1-(1-2p_\theta)\cos\phi_\theta}{2}} \leq \gamma,
\end{equation}
because
\begin{align*}
&\quad\; P({\mathcal{D}}_{C,\theta},{\mathbbm{1}}_C)\nonumber \\
&= P({\mathcal{R}}_{{C \leftarrow SA}} \circ ({\mathcal{N}}_{S,\theta} \otimes {\mathbbm{1}}_A) \circ {\mathcal{E}}_{{SA\leftarrow C}} \circ {\mathcal{U}}_{C,\theta}^\dagger,{\mathbbm{1}}_C)\nonumber
\\
&=
P( {\mathcal{R}}_{{C \leftarrow SA}} \circ ({\mathcal{N}}_{S,\theta}\circ {\mathcal{E}}_{{S\leftarrow L}} \circ {\mathcal{U}}_{L,\theta}^\dagger \otimes {\mathbbm{1}}_A) \circ {\mathcal{E}}_{{LA \leftarrow C}}^{{\mathrm{rep}}} , {\mathbbm{1}}_C )\nonumber\\
&\leq P({\mathcal{R}}^{{\mathrm{opt'}}}_{{L\leftarrow S}} \circ {\mathcal{N}}_{S} \circ {\mathcal{U}}_{S,\theta}\circ {\mathcal{E}}_{{S\leftarrow L}} \circ {\mathcal{U}}_{L,\theta}^\dagger, {\mathbbm{1}}_L) \leq \gamma,
\end{align*}
where we use the monotonicity of the purified distance and the definition of $\gamma$. \eqref{eq:purified} implies $\sqrt{p_\theta} \leq P({\mathcal{D}}_{C,\theta},{\mathbbm{1}}_C) \leq \gamma$ for all $\theta \in {\mathbb{R}}$.
Since $U_{L,\theta}$ and $U_{S,\theta}$ are periodic with a common period $\tau$, we must have $\phi_\theta = \phi_{\theta + \tau}$. Therefore, there must exists a $\theta_*$ such that $\partial_\theta \phi_\theta\big|_{\theta = \theta_*} = 0$. Then using \eqref{eq:dephasing-SLD} and the monotonicity of the regularized QFI, we see that
\begin{align}
&{F^\infty}({\mathcal{N}}_{S,\theta}) = {F^\infty}({\mathcal{N}}_{S,\theta_*}) \geq {F^\infty}({\mathcal{N}}_{C,\theta_*}) \nonumber\\ &\quad\geq
\frac{(1-2p_{\theta_*})^2 (\Delta H_L)^2 }{4p_{\theta_*}(1-p_{\theta_*})} \geq
\frac{(1-2\gamma^2)^2 (\Delta H_L)^2 }{4\gamma^2(1-\gamma^2)}.
\end{align}
\thmref{thm:global-2-metrology} then follows from \propref{prop:gate}.
\end{proof}
Note that \thmref{thm:global-2-metrology} coincides with Theorem 1 in Ref.~\cite{zhou2020new} in the special case where $\delta_\group = 0$.
\subsubsection{Quantum resource theory method}
\label{sec:global-2-res}
Now we present another derivation based on quantum resource theory, which allows us to derive not only a lower bound on the worst-case gate implementation error, but also on the Choi gate implementation error.
We work with a resource theory of coherence~\cite{marvian2020coherence} where the free (incoherent) states are those whose density operators commute with the Hamiltonian and the free (covariant) operations are those that commute with the Hamiltonian evolution, e.g., a covariant operation ${\mathcal{C}}_{L\leftarrow S}$ from $S$ to $L$ satisfies ${\mathcal{C}}_{L\leftarrow S} \circ {\mathcal{U}}_{S,\theta} = {\mathcal{U}}_{L,\theta} \circ {\mathcal{C}}_{L\leftarrow S}$ for all $\theta \in {\mathbb{R}}$. Assuming that the recovery operations ${\mathcal{R}}_{{L\leftarrow S}}$ and the noise channel ${\mathcal{N}}_S$ are covariant, we can formulate the covariant QEC as a resource conversion task from noisy physical states to error-corrected logical states and the noise rate of the latter is upper bounded by $\gamma$, illustrated by the following lemma:
\begin{proposition}
\label{prop:cov}
Consider a quantum code defined by ${\mathcal{E}}_{{L\leftarrow S}}$. Consider physical Hamiltonian $H_S$, logical Hamiltonian $H_L$, and noise channel ${\mathcal{N}}_S$.
Suppose ${\mathcal{N}}_S$ commutes with ${\mathcal{U}}_{S,\theta}$. Then the QEC inaccuracy under the restriction that the recovery channel is covariant satisfies
\begin{align}
\label{eq:cov-ineq-1}
\epsilon_{\mathrm{cov}} = \min_{{\mathcal{R}}^{\mathrm{cov}}_{L\leftarrow S}}
P({\mathcal{R}}^{\mathrm{cov}}_{{L\leftarrow S}}\circ {\mathcal{N}}_{S}\circ{\mathcal{E}}_{{S\leftarrow L}}, {\mathbbm{1}}_L)\leq \gamma ,\\
\label{eq:cov-ineq-2}
\overline{\epsilon}_{\mathrm{cov}} = \min_{{\mathcal{R}}^{\mathrm{cov}}_{L\leftarrow S}}
\overline{P}({\mathcal{R}}^{\mathrm{cov}}_{{L\leftarrow S}}\circ {\mathcal{N}}_{S}\circ{\mathcal{E}}_{{S\leftarrow L}},{\mathbbm{1}}_L) \leq \overline{\gamma},
\end{align}
where ${\mathcal{R}}^{\mathrm{cov}}_{{L\leftarrow S}}$ satisfies ${\mathcal{U}}_{L,\theta}\circ{\mathcal{R}}^{\mathrm{cov}}_{{L\leftarrow S}} = {\mathcal{R}}_{{L\leftarrow S}}^{\mathrm{cov}}\circ{\mathcal{U}}_{S,\theta}$.
\end{proposition}
\begin{proof}
Let ${\mathcal{R}}^{{\mathrm{opt'}}}_{{L\leftarrow S}}$ be the recovery channel such that $\gamma = \max_\theta P({\mathcal{R}}^{{\mathrm{opt'}}}_{{L\leftarrow S}}\circ {\mathcal{N}}_{S,\theta}\circ{\mathcal{E}}_{{S\leftarrow L}}, {\mathcal{U}}_{L,\theta})$. Suppose $U_{S,\theta}$ and $U_{L,\theta}$ share a common period $\tau$.
Consider the following recovery channel:
\begin{equation}
\label{eq:def-cov-rec}
{\mathcal{R}}_{{L\leftarrow S}}^{\mathrm{cov}} = \frac{1}{\tau} \int_0^\tau d\theta\;{\mathcal{U}}^\dagger_{L,\theta} \circ {\mathcal{R}}^{{\mathrm{opt'}}}_{{L\leftarrow S}} \circ {\mathcal{U}}_{S,\theta} .
\end{equation}
Then ${\mathcal{R}}_{{L\leftarrow S}}^{\mathrm{cov}}$ must be covariant and
\begin{equation*}
\begin{split}
&\quad \; P({\mathcal{R}}^{\mathrm{cov}}_{{L\leftarrow S}}\circ {\mathcal{N}}_{S}\circ{\mathcal{E}}_{{S\leftarrow L}}, {\mathbbm{1}}_L)\\
&= P\left(\frac{1}{\tau}\int_0^\tau d\theta\; {\mathcal{U}}_{L,\theta}^\dagger \circ {\mathcal{R}}^{{\mathrm{opt'}}}_{{L\leftarrow S}} \circ {\mathcal{N}}_{S} \circ {\mathcal{U}}_{S,\theta} \circ {\mathcal{E}}_{{S\leftarrow L}} ,{\mathbbm{1}}_L\right)\\
&\leq \max_\theta P\left({\mathcal{U}}_{L,\theta}^\dagger \circ {\mathcal{R}}^{{\mathrm{opt'}}}_{{L\leftarrow S}} \circ {\mathcal{N}}_{S} \circ {\mathcal{U}}_{S,\theta} \circ {\mathcal{E}}_{{S\leftarrow L}} ,{\mathbbm{1}}_L\right) = \gamma,
\end{split}
\end{equation*}
where we used the concavity of $f^2(\Phi,{\mathbbm{1}})$ with respect to $\Phi$~\cite{schumacher1996sending}, leading to \eqref{eq:cov-ineq-1}.
Similarly, let ${\mathcal{R}}^{{\mathrm{opt''}}}_{{L\leftarrow S}}$ be the recovery channel such that $\overline{\gamma} = \max_\theta \overline{P}({\mathcal{R}}^{{\mathrm{opt''}}}_{{L\leftarrow S}}\circ {\mathcal{N}}_{S,\theta}\circ{\mathcal{E}}_{{S\leftarrow L}}, {\mathcal{U}}_{L,\theta})$. We could define ${\mathcal{R}}_{{L\leftarrow S}}^{{\mathrm{cov}}''} = \frac{1}{\tau} \int_0^\tau d\theta\;{\mathcal{U}}^\dagger_{L,\theta} \circ {\mathcal{R}}^{{\mathrm{opt''}}}_{{L\leftarrow S}} \circ {\mathcal{U}}_{S,\theta} $ and verify that $\overline{P}({\mathcal{R}}^{{\mathrm{cov}}''}_{{L\leftarrow S}}\circ {\mathcal{N}}_{S}\circ{\mathcal{E}}_{{S\leftarrow L}}, {\mathbbm{1}}_L) \leq \overline{\gamma}$, leading to \eqref{eq:cov-ineq-2}.
\end{proof}
In order to derive a concrete lower bound on $\gamma$ and $\overline{\gamma}$ using \propref{prop:cov}, we choose a resource monotone based on another type of QFI of quantum states called the RLD QFI~\cite{yuen1973multiple} defined by $F^{\mathscr{R}}(\rho_\theta) = \mathrm{Tr}((\partial_\theta\rho_\theta)^2\rho_\theta^{-1})$ when the support of $\partial_\theta \rho_\theta$ is contained in $\rho_\theta$ and $= +\infty$ otherwise. The resource monotone satisfies
\begin{equation}
\label{eq:monotone-RLD}
F^{\mathscr{R}}({\mathcal{C}}_{{L\leftarrow S}}(\rho_S),H_L) \leq F^{\mathscr{R}}(\rho_S,H_S),
\end{equation}
for all $\rho_S$ and covariant operations ${\mathcal{C}}_{{L\leftarrow S}}$, where
\begin{align}
&F^{\mathscr{R}}(\rho,H) := F^{\mathscr{R}}(e^{-iH\theta}\rho e^{iH\theta}) \nonumber\\
&
=\begin{cases}
\mathrm{Tr}(H\rho^2 H \rho^{-1}) \!-\! \mathrm{Tr}(\rho H^2) & {\mathrm{supp}}(H\rho H)\!\subseteq\! {\mathrm{supp}}(\rho),\\
+\infty & \text{otherwise.}
\end{cases}
\end{align}
Consider an error-corrected logical state ${\mathcal{R}}^{\mathrm{cov}}_{{L\leftarrow S}}\circ{\mathcal{N}}_{S}\circ{\mathcal{E}}_{{S\leftarrow L}}(\ket{\psi_L})$ using covariant recovery operations. On one hand, its RLD QFI is lower bounded by $\Theta(1/\gamma^2)$ when $\rho$ is $\gamma$-close to a coherent pure state in terms of purified distance~\cite{marvian2020coherence,zhou2020new}. On the other hand, its RLD QFI is upper bounded by the RLD QFI of the noisy physical state $F^{\mathscr{R}}({\mathcal{N}}_{S}\circ{\mathcal{E}}_{{S\leftarrow L}}(\ket{\psi_L}),H_S)$ is no less than the channel RLD QFI $F^{\mathscr{R}}({\mathcal{N}}_{S,\theta}) = \max_\rho F^{\mathscr{R}}(({\mathcal{N}}_{S,\theta}\otimes {\mathbbm{1}})(\rho))$~\cite{hayashi2011comparison,katariya2020geometric}.
Specifically,
\begin{equation}
F_{\mathscr{R}}({\mathcal{N}}_{S,\theta}) = \big\|\mathrm{Tr}_S\big(\Psi_{SR}^{{\mathcal{N}}_S,H_S}(\Psi_{SR}^{{\mathcal{N}}_S})^{-1}\Psi_{SR}^{{\mathcal{N}}_S,H_S}\big)\big\|,
\end{equation}
where $\Gamma_{SR}^{{\mathcal{N}}_S} = ({\mathcal{N}}_S \otimes {\mathbbm{1}}_R)(\Gamma_{SR})$, $\Gamma_{SR}^{{\mathcal{N}}_S,H_S} = ({\mathcal{N}}_S \otimes {\mathbbm{1}}_R)\big((H_S\otimes {\mathbbm{1}}_R)\Gamma_{SR}\big) - ({\mathcal{N}}_S \otimes {\mathbbm{1}}_R)\big(\Gamma_{SR}(H_S\otimes {\mathbbm{1}}_R)\big)$, $\Gamma_{SR} = \ket{\Gamma}_{SR}\bra{\Gamma}_{SR}$ and $\ket{\Gamma}_{SR} = \sum_i \ket{i}_S\ket{i}_R$.
We now state and prove \thmref{thm:global-2-res} which provides lower bounds on $\gamma$ and $\overline{\gamma}$ by considering different input logical states $\ket{\psi_L}$.
\begin{theorem}
\label{thm:global-2-res}
Consider a quantum code defined by ${\mathcal{E}}_{{L\leftarrow S}}$. Consider physical Hamiltonian $H_S$, logical Hamiltonian $H_L$, and noise channel ${\mathcal{N}}_S$. Suppose ${\mathcal{N}}_S$ commutes with ${\mathcal{U}}_{S,\theta}$. Then it holds that
\begin{gather}
\label{eq:coh-1}
\epsilon + \delta_\group \geq \gamma \geq \ell_2\left(\frac{\Delta H_L}{\sqrt{4 F^{{\mathscr{R}}}({\mathcal{N}}_{S,\theta})}}\right), \\
\label{eq:coh-2}
\overline{\epsilon} + \overline{\delta}_\group \geq \overline{\gamma} \geq \ell_3\left(\sqrt{\frac{\frac{1}{d_L}\mathrm{Tr}(H_L^2)-\frac{1}{d_L^2}\mathrm{Tr}(H_L)^2}{{ F^{{\mathscr{R}}}({\mathcal{N}}_{S,\theta})}}}\right),
\end{gather}
where $\ell_2(x)= x + O(x^2)$ is the inverse function of the monotonic increasing function $x = \frac{y}{\sqrt{(1 - 3y^2 + y^4)(1 - 6\sqrt{2}y^2)}}$ for $y \in [0,1/(6\sqrt{2}))$ and $\ell_3(x)= x + O(x^2)$ is the inverse function of the monotonic increasing function $x = \frac{y}{\sqrt{(1 - 3y^2 + y^4)\left(1 - ({3(\Delta H_L)^2y})/{\sqrt{2}(\frac{1}{d_L}\mathrm{Tr}(H_L^2)-\frac{1}{d_L^2}\mathrm{Tr}(H_L)^2)}\right)}}$ for $y \in \left[0,\frac{\sqrt{2}(\frac{1}{d_L}\mathrm{Tr}(H_L^2)-\frac{1}{d_L^2}\mathrm{Tr}(H_L)^2)}{3(\Delta H_L)^2}\right)$.
\end{theorem}
In particular, when $\epsilon = 0$, i.e., when the code is exactly error-correcting, we have the following corollary:
\begin{corollary}
\label{col:global-2-covariant-res}
Consider a quantum code defined by ${\mathcal{E}}_{{L\leftarrow S}}$. Consider physical Hamiltonian $H_S$, logical Hamiltonian $H_L$, and noise channel ${\mathcal{N}}_S$.
Suppose ${\mathcal{N}}_S$ commutes with ${\mathcal{U}}_{S,\theta}$.
When $\epsilon = 0$, i.e., when the code is exactly error-correcting, it holds that $\delta_\group \geq \ell_2(\Delta H_L/\sqrt{4 F^{\mathscr{R}}({\mathcal{N}}_{S,\theta})})$ and $\overline{\delta}_\group \geq \ell_3(\Delta H_L/\sqrt{4 F^{\mathscr{R}}({\mathcal{N}}_{S,\theta})})$.
\end{corollary}
\begin{proof}[Proof of {\thmref{thm:global-2-res}}]
Let $\ket{+_L} = \frac{\ket{0_L} + \ket{1_L}}{\sqrt{2}}$. Then according to \propref{prop:cov}, there exists a covariant recovery channel ${\mathcal{R}}_{{L\leftarrow S}}^{\mathrm{cov}}$ such that
\begin{equation}
1 - \braket{+_L|\rho_L|+_L} \leq \epsilon^2_{\mathrm{cov}} \leq \gamma^2,
\end{equation}
where $\rho_L = ( {\mathcal{R}}_{{L\leftarrow S}}^{\mathrm{cov}} \circ {\mathcal{N}}_S \circ {\mathcal{E}}_{{S\leftarrow L}}) (\ket{+_L}\bra{+_L})$.
According to Lemma 3 in Ref.~\cite{zhou2020theory},
\begin{multline}
\label{eq:variance}
F^{\mathscr{R}}(\rho_L,H_L) \\\geq \frac{1 - 3\gamma^2 + \gamma^4}{\gamma^2} \cdot \left({\mathbb{V}}_{H_L}(\ket{+_L} - \frac{3\sqrt{2}\gamma(\Delta H_L)^2}{2}\right) ,
\end{multline}
where the variance ${\mathbb{V}}_{H_L}(\ket{+_L}) = \braket{+_L|H_L^2|+_L} - \braket{+_L|H_L|+_L}^2 = \frac{(\Delta H_L)^2}{4}$. $\gamma < 1/(6\sqrt{2})$ guarantees the right-hand side is positive. On the other hand, using \eqref{eq:monotone-RLD},
\begin{multline}
\label{eq:channel-RLD}
F^{\mathscr{R}}(\rho_L,H_L)
\leq F^{\mathscr{R}}(\rho_S,H_S)
\\=
F^{\mathscr{R}}({\mathcal{N}}_{S,\theta}({\mathcal{E}}_{{S\leftarrow L}}(\ket{+_L}\bra{+_L})) ) \leq F^{\mathscr{R}}({\mathcal{N}}_{S,\theta}),
\end{multline}
where $\rho_S = ({\mathcal{N}}_S \circ {\mathcal{E}}_{{S\leftarrow L}}) (\ket{+_L}\bra{+_L})$.
Using \eqref{eq:variance} and \eqref{eq:channel-RLD}, we have
\begin{equation}
\gamma^2 \cdot \frac{1}{(1 - 3\gamma^2 + \gamma^4)(1 - 6\sqrt{2}\gamma)} \geq \frac{(\Delta H_L)^2}{4 F^{\mathscr{R}}({\mathcal{N}}_{S,\theta})},
\end{equation}
proving \eqref{eq:coh-1}.
Similarly, let $\ket{\Psi_{LR}} = \frac{1}{\sqrt{d_L}}\sum_{i=1}^{d_L} \ket{i}_L \ket{i}_R$. Then according to \propref{prop:cov}, there exists a covariant recovery channel ${\mathcal{R}}_{{L\leftarrow S}}^{{\mathrm{opt''}}}$ such that
\begin{equation}
1 - \braket{\Psi_{LR}|\rho_{LR}|\Psi_{LR}} \leq \overline{\epsilon}_{\mathrm{cov}} \leq \overline{\gamma},
\end{equation}
where $\rho_{LR} = ( {\mathcal{R}}_{{L\leftarrow S}}^{\mathrm{cov}} \circ {\mathcal{N}}_S \circ {\mathcal{E}}_{{S\leftarrow L}}) (\ket{\Psi_{LR}}\bra{\Psi_{LR}})$.
According to Lemma 3 in Ref.~\cite{zhou2020theory},
\begin{align*}
& F^{\mathscr{R}}(\rho_{LR},H_L\otimes{\mathbbm{1}}_R) \nonumber\\
\geq\;& \frac{1 - 3\overline{\gamma}^2 + \overline{\gamma}^4}{\overline{\gamma}^2} \left( {\mathbb{V}}_{H_L\otimes{\mathbbm{1}}_R}(\ket{\Psi_{LR}}) - \frac{3\sqrt{2}\overline{\gamma}(\Delta H_L)^2}{2}\right) \nonumber\\
\geq\;& \frac{1 - 3\overline{\gamma}^2 + \overline{\gamma}^4}{\overline{\gamma}^2} \left( \frac{\mathrm{Tr}(H_L^2)}{d_L}-\frac{\mathrm{Tr}(H_L)^2}{d_L^2} - \frac{3\sqrt{2}\overline{\gamma}(\Delta H_L)^2}{2} \right).
\end{align*}
The rest of the proof is exactly the same as in the proof of the lower bound on the worst-case gate implementation error.
\end{proof}
In fact, the proof of \thmref{thm:global-2-res} follows almost exactly from the proof of Theorem 2 in Ref.~\cite{zhou2020new} and our new contribution here is \propref{prop:cov}.
To compare \thmref{thm:global-2-res} to \thmref{thm:global-2-metrology} , we first note that $F^{\mathscr{R}}(\Phi_\theta) \geq {F^\infty}(\Phi_\theta)$ for any $\Phi_\theta$ because $F^{\mathscr{R}}(\rho_\theta) \geq F(\rho_\theta)$~\cite{petz2011introduction} for any $\rho_\theta$ and $F^{\mathscr{R}}(\Phi_\theta) = \lim_{N\rightarrow \infty} F^{\mathscr{R}}(\Phi_\theta^{\otimes N})/N$~\cite{hayashi2011comparison,katariya2020geometric}.
Moreover, \thmref{thm:global-2-res} requires the commutativity between the noise and the Hamiltonian and $\ell_1(x) \geq \ell_2(x)$, so \thmref{thm:global-2-res} provides a weaker bound on the (worst-case) gate implementation error than \thmref{thm:global-2-metrology}. Note that $F^{\mathscr{R}}({\mathcal{N}}_{S,\theta}) < +\infty$ only when ${\rm span}\{K_{S,i} H_S,\forall i\} \subseteq {\rm span}\{K_{S,i},\forall i\}$ which is also a stronger condition than the HKS condition. The resource theory method leads to a bound on the Choi gate implementation error, however, which is not available using the quantum metrology method.
Also note recent works Refs.~\cite{tajima2018uncertainty,tajima2020coherence} (results implied by Ref.~\cite{tajima2021symmetry}) which considered the coherence cost of implementing unitary gates based on relevant insights.
\subsection{Explicit behaviors of the bounds and comparison with the charge fluctuation approach}
\label{sec:comparison}
We first make a general comparison between the trade-off relations derived using the gate implementation error approach (\thmref{thm:global-2-metrology} and \thmref{thm:global-2-res}) and the charge fluctuation approach (\thmref{thm:global-1}) in \secref{sec:global-1}. Two clear advantages of the gate implementation error approach are that 1) it applies to general quantum codes (e.g., the non-isometric encodings in \cite{woods2019continuous,yang2020covariant}) while the charge fluctuation approach only holds for isometric codes; 2) it leads to a trade-off relation for the Choi measures. Additionally, for the special case of $\delta_\group = 0$, the results based on the gate implementation error approach directly reduce to the previous known result for exactly covariant codes in Ref.~\cite{zhou2020new}, while there is still some discrepancy with previous results using the charge fluctuation approach (see more discussion in \secref{sec:global-1-tradeoff}). For the special case of $\epsilon = 0$ which was not previously studied, we have two lower bounds on $\delta$ from \corollaryref{col:global-2-covariant-metrology} and \corollaryref{col:global-1-covariant} respectively which behave as follows:
\begin{align}
\delta_\group \gtrsim \frac{\Delta H_L}{2\sqrt{{\mathfrak{F}}}}, \qquad
\delta_\group \gtrsim \sqrt{\frac{\Delta H_L}{\Delta H_S}}.
\end{align}
It is interesting to observe that the first bound depends on the noise channel while the second one does not (as long as the HKS condition is satisfied).
We now remark on the explicit scalings of our bounds for different noise models as in \secref{sec:global-1-noise}. Again, consider a $n$-partite system and a local physical Hamiltonian with $\Delta H_S = O(n)$. For random local noise which acts uniformly randomly on each subsystem, ${\mathfrak{F}} = O(n^2)$ and the two bounds give $\delta_\group = \Omega(1/n)$ and $\delta_\group = \Omega(1/\sqrt{n})$ respectively.
That is, the charge fluctuation approach outperforms the gate implementation error approach in this case. For noise acting independently on each subsystems, we have ${\mathfrak{F}} = O(n)$ which gives $\delta_\group = \Omega(1/\sqrt{n})$ using the gate implementation error approach. In this situation, the bounds based on the two approaches are comparable. Note that in situations where ${\mathfrak{F}} = o(n)$, i.e., the noise is even stronger than independent noise so that the regularized QFI is sublinear, the bound based the gate implementation error approach should outperform the bound based on the charge fluctuation approach. In general situations where both the QEC inaccuracy and the global covariance violation are non-vanishing, we expect a similar behavior, i.e., the gate implementation error approach performs better in the extremely strong noise regime, while the charge fluctuation approach performs better in weaker noise regimes.
\section{Limitations on transversal logical gates}
\label{sec:transversal}
Note that a key implication of our results is symmetry constraints on QEC codes that achieve a given accuracy, which extends the scope of previous knowledge on the incompatibility between symmetries and QEC to general codes, especially exact QEC codes which are most commonly studied.
As we shall discuss in this section, such constraints actually allow us to derive restrictions on the transversally implementable gates for general QEC codes, advancing our understanding of fault tolerance. A key intuition is that the precision of gate implementation is associated with the degree of symmetry.
Recall that there are no QEC codes which admit a continuous symmetry acting transversally on physical qubits and thus there are no transversal universal gate sets, according to the Eastin--Knill theorem. For stabilizer codes, the incompatibility between QEC and symmetry are reflected in the classification of transversally logical gates in finite levels of the Clifford hierarchy~\cite{bravyi2013classification,pastawski2015fault,anderson2016classification,jochym2018disjointness}. Here we present new restrictions of transversal gates for arbitrary QEC codes from the perspective of global covariance violations.
The following corollary puts a restriction on the logical transversal gates using \corollaryref{col:global-1-covariant}. Namely, the non-trivial logical gates cannot be too close to the identity operators when implemented by transversal physical gates in the vicinity of identity operators because $\delta_\group$ has a lower bound of $\Theta(1/\sqrt{n})$. Note that here we implicitly consider exact QEC codes under single-erasure noise (so that the HKS condition is satisfied for 1-local Hamiltonians) and in this case \corollaryref{col:global-1-covariant} outperforms \corollaryref{col:global-2-covariant-metrology}, so we will only use \corollaryref{col:global-1-covariant} in this section.
\begin{corollary}
\label{col:gate}
Suppose an $n$-qudit QEC code with distance at least $2$ admits a transversal implementation $V_S = \bigotimes_{l=1}^n e^{-i2\pi T_{S_l}/D}$ of the logical gate $V_L = e^{-i2\pi T_L/D}$ where $D$ is a positive integer and $T_{L,S}$ have integer eigenvalues.
Then it holds that
\begin{multline}
D \leq \max\Bigg\{
4\pi\sqrt{\frac{2}{3}}\bigg(\Delta T_L + \sum_{l=1}^n \Delta T_{S_l}\bigg), \\
{2\sqrt{2}\pi}\sqrt{ \frac{ \sum_{l=1}^n \Delta T_{S_l}}{\Delta T_L}} \left(\Delta T_L + \sum_{l=1}^n \Delta T_{S_l}\right) \Bigg\}.
\end{multline}
In particular, $D = O(\mathrm{poly}(n))$ when $\Delta T_S = O(\mathrm{poly}(n))$, $\Delta T_L = O(\mathrm{poly}(n))$ and $\Delta T_S/\Delta T_L = O(\mathrm{poly}(n))$, where $\Delta T_S = \sum_{\ell=1}^n \Delta T_{S_\ell}$\footnote{
The conditions are satisfied in common settings; see, e.g., the proof of \corollaryref{col:gate-stab}.
}.
\end{corollary}
\begin{proof}
Any codes with distance at least $2$ can correct single-erasure noise. Let $H_{L} = T_L$ and $H_S = \sum_l T_{S_l}$. They have integer eigenvalues implies that ${\mathcal{U}}_{S,\theta}$ and ${\mathcal{U}}_{L,\theta}$ share a common period $2\pi$. According to \corollaryref{col:global-1-covariant}, the code must satisfy
\begin{equation}
\label{eq:app-gate}
\delta_\group
\geq \min\left\{ \frac{\sqrt{\Delta H_L\left(\Delta H_S -\frac{1}{2}\Delta H_L\right)}}{\Delta H_S}, \sqrt{\frac{3}{8}}\right\}.
\end{equation}
We can always write $\theta = \frac{2j\pi}{D} + \theta_1$ for some $j\in{\mathbb{N}}$ and $\theta_1 \in [0,2\pi/D)$. Then we have
\begin{align}
& P({\mathcal{U}}_{S,\theta}\circ {\mathcal{E}}_{{S\leftarrow L}} \circ {\mathcal{U}}_{L,\theta}^\dagger,{\mathcal{E}}_{{S\leftarrow L}})\nonumber\\ =\;& P({\mathcal{U}}_{S,\theta_1}\circ {\mathcal{E}}_{{S\leftarrow L}} \circ {\mathcal{U}}_{L,\theta_1}^\dagger,{\mathcal{E}}_{{S\leftarrow L}}) \nonumber\\
\leq\;& P({\mathcal{U}}_{S,\theta_1}\circ {\mathcal{E}}_{{S\leftarrow L}} \circ {\mathcal{U}}_{L,\theta_1}^\dagger,{\mathcal{U}}_{S,\theta_1}\circ{\mathcal{E}}_{{S\leftarrow L}}) \nonumber\\
& + P({\mathcal{U}}_{S,\theta_1}\circ{\mathcal{E}}_{{S\leftarrow L}},{\mathcal{E}}_{{S\leftarrow L}})\nonumber\\
\leq\;& P({\mathcal{U}}_{L,\theta_1}^\dagger,{\mathbbm{1}}_L) + P({\mathcal{U}}_{S,\theta_1},{\mathbbm{1}}_S),
\end{align}
where we use the monotonicity and the triangular inequality of the purified distance. Without loss of generality, assume $H_L = \sum_{i=1}^{d_L} \lambda_i \ket{i}\bra{i}$ and let $\ket{\psi} = \sum_{i=1}^{d_L} \psi_i \ket{i}\ket{i}$. Consider first the situation where $D > 2\max\{\Delta H_L,\Delta H_S\}$, then we have $\theta_1 \max\{\Delta H_L,\Delta H_S\} < \pi$ and
\begin{equation}
\begin{split}
P({\mathcal{U}}_{L,\theta_1}^\dagger,{\mathbbm{1}}_L) =&\; \max_{\{\psi_i\}} P(({\mathcal{U}}_{L,\theta_1}^\dagger\otimes {\mathbbm{1}})(\ket{\psi}\bra{\psi}),\ket{\psi}\bra{\psi}) \\
=&\; \max_{\{\psi_i\}} \!\sqrt{1 \!- \!\bigg|\!\sum_{i} \abs{\psi_i}^2 e^{i\theta_1 \lambda_i}\!\bigg|^2} \\
=&\;\max_{i,i'} \abs{\sin\left(\frac{1}{2}\theta_1(\lambda_i-\lambda_{i'})\right)} \leq \theta_1 \Delta H_L.
\end{split}
\end{equation}
Similarly, $P({\mathcal{U}}_{S,\theta_1},{\mathbbm{1}}_S) \leq \theta_1 \Delta H_S = \theta_1 \sum_{l=1}^n \Delta H_{S_l}$. Since $\delta_\group = \max_\theta P({\mathcal{U}}_{S,\theta}\circ {\mathcal{E}}_{{S\leftarrow L}} \circ {\mathcal{U}}_{L,\theta}^\dagger,{\mathcal{E}}_{{S\leftarrow L}}) = \max_{\theta_1} P({\mathcal{U}}_{S,\theta_1}\circ {\mathcal{E}}_{{S\leftarrow L}} \circ {\mathcal{U}}_{L,\theta_1}^\dagger,{\mathcal{E}}_{{S\leftarrow L}})$, we obtain
\begin{equation}
\label{eq:app-gate-1}
\delta_\group \leq \frac{2\pi}{D} \left(\Delta H_L + \sum_{l=1}^n \Delta H_{S_l}\right),
\end{equation}
Otherwise,
\begin{equation}
\label{eq:app-gate-2}
D \leq 2\max\{\Delta H_L,\Delta H_S\}.
\end{equation}
The result then follows by combining \eqref{eq:app-gate}, \eqref{eq:app-gate-1} and \eqref{eq:app-gate-2}.
\end{proof}
\corollaryref{col:gate} shows that the precision of transversal logical gates under certain restrictions only increases polynomially in the number of qubits. For the important case of stabilizer codes, this implies that the levels of the Clifford hierarchy that can be reached only increase polynomially in the number of qubits.
Specifically,
consider an $n$-qubit stabilizer code with distance at least $2$.
The following corollary describes the limitation on the transversally implementable logical gates for stabilizer codes:
\begin{corollary}
\label{col:gate-stab}
Let $\tilde{V}_S = V_1 (\bigotimes_{l=1}^n e^{-i 2\pi a_l Z_l/D}) V_2$ be a transversal logical gate for an $n$-qubit stabilizer code with distance at least $2$, where $D$ is a power of two and $a_l$ is an integer and $V_{1,2}$ are transversal Clifford operators. (This describes the most general form of transversal logical gates for stabilizer codes~\cite{zeng2011transversality,anderson2016classification}). When $a_l = O(\mathrm{poly}(n))$, we must have $D = O(\mathrm{poly}(n))$ and $\tilde{V}_S$ implements a logical gate $\tilde V_L$ in the \mbox{$O(\log n)$-th} level of the Clifford hierarchy.
\end{corollary}
\begin{proof}
Let $\Pi$ be the projection onto the stabilizer code under consideration and $Q_S = \bigotimes_{l=1}^n e^{-i 2\pi a_l Z_l/D}$. Then
\begin{equation}
\tilde{V}_S \Pi = \Pi\tilde{V}_S\Pi,~~\Rightarrow~~
Q_S \Pi_2 =
\Pi_1 Q_S \Pi_2 ,
\end{equation}
where $\Pi_1 = V_1^\dagger\Pi V_1$ and $\Pi_2 = V_2\Pi V_2^\dagger$. Both $\Pi_1$ and $\Pi_2$ are stabilizer codes with the same code distance as $\Pi$. Without loss of generality, we assume $\Pi$, $\Pi_1$ and $\Pi_2$ are $2$-dimensional stabilizer codes (by considering subcodes of the original codes).
As proven in Proposition 4 in Ref.~\cite{anderson2016classification}, $V_S = (Q_S)^4$ must be a logical gate on $\Pi_2$, satisfying
\begin{equation}
(Q_S)^4 \Pi_2 = \Pi_2 (Q_S)^4 \Pi_2,
\end{equation}
and the logical gate $V_L$ has the form $V_L = e^{-i2\pi a Z_L/D}$ where $a$ is an integer. By writing down the stabilizer code $\Pi_2$ in its computational basis, it is easy to observe that $a \leq \sum_{l=1}^n 4a_l = O(\mathrm{poly}(n))$.
Let $H_L = a Z_L$ and $H_{S_l} = 4a_l Z_l$. Then we must have
\begin{equation}
\frac{\sqrt{\Delta H_L\left(\Delta H_S -\frac{1}{2}\Delta H_L\right)}}{\Delta H_S} \geq \frac{1}{O(\mathrm{poly}(n))}.
\end{equation}
Using \corollaryref{col:gate}, we have $D = O(\mathrm{poly}(n))$.
Since $D$ is a power of 2, for all $l$, $e^{-i 2\pi a_l Z_l/D}$ (see Proposition 1 in Ref.~\cite{anderson2016classification}) and thus $Q_S = \bigotimes_{l=1}^n e^{-i 2\pi a_l Z_l/D}$ must be in the \mbox{$(\log D)$-th} level of the Clifford hierarchy. \corollaryref{col:gate-stab} then follows from the fact that Clifford operators $V_1$, $V_2$ preserve the level of the Clifford hierarchy and any physical gate in the \mbox{$j$-th} level of the Clifford hierarchy implements a logical gate in the \mbox{$j$-th} level of the Clifford hierarchy (because logical Pauli operators can be implemented by physical Pauli operators for stabilizer codes).
\end{proof}
\corollaryref{col:gate-stab} provides a simple proof on the limitations of transversal logical gates for stabilizer codes from the perspective of continuous symmetries. Note that the relevant results previously known for stabilizer codes \cite{bravyi2013classification,pastawski2015fault,jochym2018disjointness} were obtained using very different techniques.
\section{Trade-off between QEC and local symmetry measures}
\label{sec:local}
In this section, we study relations between QEC and local symmetry measures, that is, the local covariance violation and the charge conservation violation. We will first prove a lemma which links the charge conservation to the charge fluctuation and then derive trade-off relations using \propref{prop:charge-KL} and \propref{prop:charge-metrology}. We will also derive a lower bound on the local covariance violation using the quantum metrology method.
Note that the results in this section (\thmref{thm:local-KL}, \thmref{thm:local-metrology} and \thmref{thm:local-point}) hold true for arbitrary Hermitian operators $H_L$ and $H_S$, which do not necessarily share a common period as generators of $U(1)$ representations.
\subsection{Bounds via charge fluctuation}
We first observe a simple connection between the charge fluctuation $\chi$ and the charge conservation violation $\delta_{\mathrm{C}}$:
\begin{lemma}
\label{lemma:chi-local} Consider a quantum code ${\mathcal{E}}_{{S\leftarrow L}}$, a physical Hamiltonian $H_S$ and a logical Hamiltonian $H_L$. Then
\begin{equation}
\abs{\chi} \geq \Delta H_L - \delta_{\mathrm{C}}.
\end{equation}
\end{lemma}
\begin{proof}
By definition, $\abs{\chi} = \big|\!\braket{0_L|{\mathcal{E}}^\dagger_{{S\leftarrow L}}(H_S)|0_L} - \braket{1_L|{\mathcal{E}}^\dagger_{{S\leftarrow L}}(H_S)|1_L}\!\big|$. Then we must have $\abs{\chi} \geq \Delta H_L - \delta_{\mathrm{C}}$, because $\delta_{\mathrm{C}} \geq \big|\!\braket{0_L|({\mathcal{E}}^\dagger_{{S\leftarrow L}}(H_S)-H_L)|0_L} - \braket{1_L|({\mathcal{E}}^\dagger_{{S\leftarrow L}}(H_S)-H_L)|1_L}\!\big|$, and $\Delta H_L = \big|\!\braket{0_L|H_L|0_L} - \braket{1_L|H_L|1_L}\!\big|$.
\end{proof}
Using the KL-based method (\propref{prop:charge-KL}) and \propref{prop:point-charge}, we immediately have the following trade-off relations:
\begin{theorem}
\label{thm:local-KL}
Consider an isometric quantum code defined by $\mE_{\StoL}$. Consider physical Hamiltonian $H_S$, logical Hamiltonian $H_L$, and noise channel $\mN_S$. Suppose the HKS condition is satisfied. It holds that
\begin{align}
\delta_{{\mathrm{P}}} + 2\epsilon{\mathfrak{J}} &\geq \Delta H_L,\\
\delta_{{\mathrm{C}}} + 2\epsilon{\mathfrak{J}} &\geq \Delta H_L,
\label{eq:tradeoff-3}
\end{align}
where ${\mathfrak{J}}$ is given by \eqref{eq:def-frakJ}.
\end{theorem}
Note that \eqref{eq:tradeoff-3} reduces to Corollary 3 in Ref.~\cite{faist2019continuous} for random local erasure noise.
In particular, for exact QEC codes, we have the following corollary:
\begin{corollary}
\label{col:local-KL}
Consider an isometric quantum code defined by $\mE_{\StoL}$. Consider physical Hamiltonian $H_S$, logical Hamiltonian $H_L$, and noise channel $\mN_S$. Suppose the HKS condition is satisfied. When $\epsilon = 0$, i.e., when the code is exactly error-correcting, we must have $\delta_{\mathrm{P}} \geq \Delta H_L$ and $\delta_{\mathrm{C}} \geq \Delta H_L$.
\end{corollary}
Similarly, using the quantum metrology method (\propref{prop:charge-metrology}) and \propref{prop:point-charge}, we have the following trade-off relations:
\begin{theorem}
\label{thm:local-metrology}
Consider a quantum code defined by $\mE_{\StoL}$. Consider physical Hamiltonian $H_S$, logical Hamiltonian $H_L$, and noise channel $\mN_S$. Suppose the HKS condition is satisfied. It holds that
\begin{align}
\delta_{{\mathrm{C}}} + 2\epsilon\left(\sqrt{(1-\epsilon^2){\mathfrak{F}}
} + {\mathfrak{B}} \right) &\geq \Delta H_L.
\end{align}
In particular, when $\epsilon \ll 1$ and ${\mathfrak{B}} \ll \sqrt{{\mathfrak{F}}}$, we have
\begin{align}
\delta_{{\mathrm{C}}} + 2 \epsilon \sqrt{{\mathfrak{F}}}
&\gtrsim \Delta H_L,
\end{align}
where ${\mathfrak{F}}$ and ${\mathfrak{B}}$ are given by \eqref{eq:def-frakF} and \eqref{eq:def-frakB} respectively,
Furthermore, when the code is isometric, it holds that
\begin{equation}
\delta_{{\mathrm{P}}} + 2\epsilon\left(\sqrt{(1-\epsilon^2){\mathfrak{F}}} + {\mathfrak{B}} \right) \geq \Delta H_L.
\end{equation}
\end{theorem}
\begin{corollary}
\label{col:local-metrology}
Consider a quantum code defined by $\mE_{\StoL}$. Consider physical Hamiltonian $H_S$, logical Hamiltonian $H_L$, and noise channel $\mN_S$. Suppose the HKS condition is satisfied. When $\epsilon = 0$, i.e., when the code is exactly error-correcting, we must have $\delta_{\mathrm{C}} \geq \Delta H_L$.
\end{corollary}
Note that \corollaryref{col:local-metrology} is slightly more general than \corollaryref{col:local-KL} as the former covers the situation where the encoding is non-isometric.
\subsection{Bounding local covariance violation using quantum metrology}
The trade-off relation between $\epsilon$ and $\delta_{\mathrm{P}}$ in \thmref{thm:local-metrology} requires the code to be isometric.
In fact, we can show a cleaner version of the trade-off between the QEC inaccuracy and the local covariance violation using the quantum metrology method which does not contain ${\mathfrak{B}}$ and also covers the non-isometric scenario, as shown below.
\begin{theorem}
\label{thm:local-point}
Consider a quantum code defined by $\mE_{\StoL}$. Consider physical Hamiltonian $H_S$, logical Hamiltonian $H_L$, and noise channel $\mN_S$. Suppose the HKS condition is satisfied. It holds that
\begin{equation}
\label{eq:tradeoff-2}
\delta_{\mathrm{P}} + 2\epsilon\left(\sqrt{(1-\epsilon^2){\mathfrak{F}}} + \epsilon \Delta H_L\right) \geq \Delta H_L.
\end{equation}
When $\epsilon \ll 1$, we have
\begin{equation}
\delta_{\mathrm{P}} + 2\epsilon\sqrt{{\mathfrak{F}}}
\gtrsim \Delta H_L,
\end{equation}
where ${\mathfrak{F}}$ and ${\mathfrak{B}}$ are given by \eqref{eq:def-frakF} and \eqref{eq:def-frakB} respectively.
\end{theorem}
\begin{corollary}
\label{col:local-metrology-2}
Consider a quantum code defined by $\mE_{\StoL}$. Consider physical Hamiltonian $H_S$, logical Hamiltonian $H_L$, and noise channel $\mN_S$. Suppose the HKS condition is satisfied. When $\epsilon = 0$, i.e., when the code is exactly error-correcting, it holds that $\delta_{\mathrm{P}} \geq \Delta H_L$.
\end{corollary}
\begin{proof}
Let ${\mathcal{R}}^{{\mathrm{opt}}}_{{L\leftarrow S}}$ be the recovery channel such that
$\epsilon = P({\mathcal{R}}^{{\mathrm{opt}}}_{{L\leftarrow S}}\circ {\mathcal{N}}_{S}\circ{\mathcal{E}}_{{S\leftarrow L}}, {\mathbbm{1}}_{L})$. Let (see \figref{fig:dephasing})
\begin{align}
{\mathcal{N}}_{C,\theta} =\;&{\mathcal{R}}^{{\mathrm{rep}}}_{{C \leftarrow SA}} \circ ( {\mathcal{R}}_{{L\leftarrow S}}^{{\mathrm{opt}}} \circ{\mathcal{N}}_{S} \otimes {\mathbbm{1}}_A) \circ \nonumber\\& ({\mathcal{U}}_{S,\theta}\circ {\mathcal{E}}_{{S\leftarrow L}} \otimes {\mathbbm{1}}_A) \circ {\mathcal{E}}_{{LA \leftarrow C}}^{{\mathrm{rep}}}.
\end{align}
Consider the parameter estimation of $\theta$ in the neighborhood of $\theta = 0$ and let $\xi_\theta = \bra{0_C}{\mathcal{N}}_{C,\theta}(\ket{0_C}\bra{1_C})\ket{1_C}$. Following the proof of \propref{prop:charge-metrology}, we have from \eqref{eq:xi-epsilon} that
\begin{equation}
\abs{\xi_{\theta = 0}} \geq 1 - 2\epsilon^2.
\end{equation}
As shown in \appref{app:local-proof}, we also have
\begin{equation}
\abs{\partial_\theta\xi_\theta}^2\big|_{\theta = 0} \geq ((1-2\epsilon^2)\Delta H_L - \delta_{\mathrm{P}})^2,
\end{equation}
when $(1-2\epsilon^2)\Delta H_L \geq \delta_{\mathrm{P}}$.
Hence, when $(1-2\epsilon^2)\Delta H_L \geq \delta_{\mathrm{P}}$, we must have
\begin{multline}
{F^\infty}({\mathcal{N}}_{S,\theta})
= {F^\infty}({\mathcal{N}}_{S,\theta}) \big|_{\theta = 0} \\
\geq {F^\infty}({\mathcal{N}}_{C,\theta})\big|_{\theta = 0} \geq \frac{((1-2\epsilon^2)\Delta H_L - \delta_{\mathrm{P}})^2}{4\epsilon^2(1-\epsilon^2)},
\end{multline}
completing the proof.
\end{proof}
\subsection{Remarks on the behaviors of the bounds}
From \thmref{thm:local-KL}, \thmref{thm:local-metrology} and \thmref{thm:local-point}, we observe that in $n$-partite systems, the local covariance violation $\delta_{\mathrm{P}}$ and the charge conservation violation $\delta_{\mathrm{C}}$ are usually lower bounded by constants for small $\epsilon$ which does not vanish
as $n\rightarrow \infty$ like the global covariance violation $\delta_\group$. However, also note that $\delta_{\mathrm{P}}$ and $\delta_{\mathrm{C}}$ may naturally be superconstant (for example, for the trivial encoding ${\mathcal{E}}_{{S\leftarrow L}} = {\mathbbm{1}}$ we usually have $\delta_{\mathrm{P}}=\delta_{\mathrm{C}}=\Delta(H_S-H_L) = \Theta(n)$), indicating that the constant or even sublinear scaling of $\delta_{\mathrm{P}}$ and $\delta_{\mathrm{C}}$ requires non-trivial code structures.
Also note that the bounds on $\delta_{\mathrm{P}}$ in both \thmref{thm:local-KL} and \thmref{thm:local-metrology} rely on the fact that $\delta_{\mathrm{P}} \geq \delta_{\mathrm{C}}$, indicating that these bounds may not be tight when there is a gap between $\delta_{\mathrm{P}}$ and a function of $\delta_{\mathrm{C}}$. Such a gap does exist as we shown later in examples (see \secref{sec:case-study}) and we provide a possible explanation of the existence of the gap in \appref{app:refine}.
\section{Case studies of explicit codes}
\label{sec:case-study}
In the above, we have derived several forms of fundamental limits on the QEC accuracy and degree of symmetry or charge conservation that a quantum code can possibly admit. Then a natural question is to what extent these limits can be attained by certain codes. Furthermore, explicit constructions of approximately covariant codes would be important for our understanding of the QEC-symmetry trade-off and may find broad applications. In this section, we introduce and analyze two code examples with interesting approximate covariance features to address these needs. In the first example, we generalize a covariant code called the thermodynamic code~\cite{brandao2019quantum,faist2019continuous} to a class of general quantum codes which exhibits a full trade-off between symmetry and QEC via a smooth transition from exact covariance to exact QEC. The second one involves a well-known QEC code called the quantum Reed--Muller codes~\cite{steane1999quantum,macwilliams1977theory}, which can be seen as a prominent example of approximately covariant exact QEC codes. In particular, we explicitly compute their QEC and symmetry measures, and compare them to the fundamental limits. Remarkably, the scalings of the global covariance violation and the charge conservation violation for both examples match well with the optimal scalings from our bounds.
\subsection{Modified thermodynamic codes}
\label{sec:thermo}
Thermodynamic codes \cite{brandao2019quantum,faist2019continuous} are $n$-qubit quantum codes given by certain Dicke states with different magnetic charges which become approximately quantum error-correcting for large $n$. Specifically, a two-dimensional thermodynamic code have codewords
\begin{equation}
\ket{{\mathfrak{c}}_0} = \ket{m_n} , \quad
\ket{{\mathfrak{c}}_1} = \ket{(-m)_n} ,
\end{equation}
where $\ket{m_n}$ for $m \in [-n,n]$ is the Dicke state defined by
\begin{equation}
\ket{m_n} = \frac{1}{\sqrt{\binom{n}{\frac{n+m}{2}}}} \sum_{\substack{{\gv{j}}\in\{0,1\}^n:\\\sum_l j_l = \frac{n+m}{2}}}\ket{{\gv{j}}},
\end{equation}
satisfying $\sum_{l=1}^n Z_l\ket{m_n} = -m$. Note that $m+n$ must be an even number and we assume $2\leq m \ll n$. It is easy to verify that the thermodynamic code is exactly covariant with respect to $H_L = \frac{m}{2}Z_L$ and $H_S = - \frac{1}{2} \sum_{l=1}^n Z_l$ and it was proven that for single-erasure noise~\cite{zhou2020theory} $\epsilon = m/2n + O(m^2/n^2)$ which is infinitely small when $m/n\rightarrow 0$.
Here we extend the thermodynamic code in such a way that it smoothly transitions from an exactly covariant code to an exact QEC code as tuned by a continuous parameter $0\leq q \leq 1$. Specifically, our modified thermodynamic code is defined by
\begin{align}
\ket{{\mathfrak{c}}^q_0} &= \sqrt{\frac{n}{n+qm}} \ket{m_n} + \sqrt{\frac{qm}{n+qm}} \ket{(-n)_n}, \\
\ket{{\mathfrak{c}}^q_1} &= \sqrt{\frac{n}{n+qm}} \ket{(-m)_n} + \sqrt{\frac{qm}{n+qm}} \ket{n_n}.
\end{align}
In particular, when $q = 0$, we have the original thermodynamic code, and when $q = 1$, we obtain an modified code which is exactly error-correcting under single-erasure noise.
We shall compute the QEC inaccuracy and the different covariance violation measures, and compare them with our trade-off bounds.
\subsubsection{QEC inaccuracy}
Here we compute the QEC inaccuracy of modified thermodynamic codes $\epsilon({\mathcal{N}}_{S},{\mathcal{E}}_{{L\leftarrow S}})$ where ${\mathcal{E}}_{{L\leftarrow S}}(\cdot) = W(\cdot)W^\dagger$ with $W=\ket{{\mathfrak{c}}^q_0}\bra{0_L} + \ket{{\mathfrak{c}}^q_1}\bra{1_L}$ and ${\mathcal{N}}_S$ is the single-erasure noise channel ${\mathcal{N}}_S = \sum_{l=1}^n \frac{1}{n} {\mathcal{N}}_{S_l}$, where ${\mathcal{N}}_{S_l}(\cdot) = \ket{{\emptyset}}\bra{{\emptyset}}_{S_l} \otimes \mathrm{Tr}_{S_l}(\cdot)$.
We need to use the following lemma which compute the purified distance between error-corrected channels, employing the formalism of complementary channels~\cite{beny2010general}. Let $\widehat \Phi_{B\leftarrow A}(\cdot) = \mathrm{Tr}_A(V_{AB\leftarrow A}(\cdot)V_{AB\leftarrow A}^\dagger)$
be the complementary channel of channel $\Phi_A(\cdot) = \mathrm{Tr}_B(V_{AB\leftarrow A}(\cdot)V_{AB\leftarrow A}^\dagger)$,
where $V_{AB\leftarrow A}$ is a Stinespring dilation of $\Phi_A$. Then we have
\begin{lemma}[\cite{beny2010general}]
\label{lemma:complementary}
\begin{equation}\min_{\mathcal{R}} P({\mathcal{R}}\circ\Phi_1,\Phi_2) = \min_{{\mathcal{R}}'} P(\widehat{\Phi}_1,{\mathcal{R}}' \circ \widehat{\Phi}_2),
\end{equation}
for arbitrary $\Phi_{1,2}$, where the minimizations are taken over all channels with the appropriate input and output spaces.
\end{lemma}
Choosing $\Phi_1 = {\mathcal{N}}_{S} \circ {\mathcal{E}}_{{S\leftarrow L}}$ and $\Phi_2 = {\mathbbm{1}}_L$,
we have
\begin{equation}
\epsilon =
\min_{{\mathcal{R}}'_{{B\leftarrow L}}} P(\widehat{{\mathcal{N}}}_{{B\leftarrow S}}\circ{\mathcal{E}}_{{S\leftarrow L}},{\mathcal{R}}'_{{B\leftarrow L}}\circ \widehat{{\mathbbm{1}}}_{L}),
\end{equation}
As detailed in \appref{app:thermo}, we have that
\begin{equation}
\label{eq:thermo-epsilon}
\epsilon = \frac{(1-q)m}{2n} + O\left(\frac{m^2}{n^2}\right),
\end{equation}
and furthermore, explicitly construct a recovery channel ${\mathcal{R}}^{{\mathrm{opt}}}_{{L\leftarrow S}}$ which achieves the optimal QEC inaccuracy up to the lowest order of $m/n$:
\begin{align}
\tilde \epsilon &= P({\mathcal{R}}^{{\mathrm{opt}}}_{{L\leftarrow S}}\circ{\mathcal{N}}_S\circ{\mathcal{E}}_{{S\leftarrow L}},{\mathbbm{1}}_L) \nonumber \\&= \sqrt{\frac{1}{2}-\sqrt{\frac{(n+m)(n+(2q-1)m)}{4(n+qm)^2}}} \approx \frac{(1-q)m}{2n}.
\end{align}
\subsubsection{Symmetry violation measures}
We now compute all the approximate symmetry measures associated with our modified thermodynamic codes. Note that we let $H_L = \frac{m}{2}Z_L$ and $H_S = -\frac{1}{2}\sum_{l=1}^n Z_l$, which guarantees that the code tends to be covariant as $n\rightarrow\infty$.
We first compute $\delta_\group$ and $\delta_{\mathrm{P}}$. Let $\ket{\psi} = \ket{0_L}\ket{\psi_R^0} + \ket{1_L}\ket{\psi_R^1}$ be an arbitrary pure state on $L \otimes R$. Then
\begin{align*}
&\ket{\psi_\theta} := U_{S,\theta} W U_{L,\theta}^\dagger \ket{\psi} \nonumber\\=
&\left(\!\sqrt{\frac{n}{n\!+\!qm}} \ket{m_n} \!+\! e^{\frac{i(m+n) \theta}{2}} \! \sqrt{\frac{qm}{n\!+\!qm}} \ket{(\!-n\!)_n}\!\right)\!\ket{\psi_R^0} \nonumber
\\&+ \left(\!\sqrt{\frac{n}{n\!+\!qm}} \ket{(\!-m\!)_n}
\!+\! e^{\frac{-i(m+n)\theta}{2}}\! \sqrt{\frac{qm}{n\!+\!qm}} \ket{n_n}\!\right)\!\ket{\psi_R^1},
\end{align*}
and
\begin{multline}
P({\mathcal{U}}_{S,\theta}\circ{\mathcal{E}}_{{S\leftarrow L}},{\mathcal{E}}_{{S\leftarrow L}}\circ{\mathcal{U}}_{L,\theta}) = \max_{\psi} P(\ket{\psi_\theta},\ket{\psi})\\
= \sqrt{1 - \left|\frac{n + qm \cos((m+n)\theta/2)}{n + qm}\right|^2},
\end{multline}
where the maximum of $P(\ket{\psi_\theta},\ket{\psi})$ is attained at $\ket{\psi} = (\ket{0_L}+\ket{1_L})/\sqrt{2}$ (here the reference system can be one-dimensional, namely $\ket{\psi_R^0} = \ket{\psi_R^1} = 1/\sqrt{2}$, because $U_{S,\theta} W U_{L,\theta}^\dagger \ket{0_L}$ does not overlap with $U_{S,\theta} W U_{L,\theta}^\dagger \ket{1_L}$ for all $\theta$).
Therefore, the global covariance violation is given by
\begin{align}
\delta_\group
=\;& \max_\theta P({\mathcal{U}}_{S,\theta}\circ{\mathcal{E}}_{{S\leftarrow L}},{\mathcal{E}}_{{S\leftarrow L}}\circ{\mathcal{U}}_{L,\theta}) \nonumber\\
=\;& \sqrt{1 - \left(\frac{n-qm}{n+qm}\right)^2} = \sqrt{\frac{4qm}{n}} + O\left(\left(\frac{m}{n}\right)^{3/2}\right),
\label{eq:thermo-delta}
\end{align}
and the local covariance violation is given by
\begin{align}
\delta_{\mathrm{P}}
=\;& \sqrt{2 \partial^2_\theta P({\mathcal{U}}_{S,\theta}\circ{\mathcal{E}}_{{S\leftarrow L}},{\mathcal{E}}_{{S\leftarrow L}}\circ{\mathcal{U}}_{L,\theta})^2\big|_{\theta = 0}} \nonumber\\
=\;& \sqrt{\frac{qm(m+n)^2}{n+qm}} = \sqrt{qmn} + O\left(m\sqrt{\frac{m}{n}}\right).
\end{align}
To compute the charge conservation violation $\delta_{\mathrm{C}}$, note that ${\mathcal{E}}_{L\leftarrow S}^\dagger(H_S) = \frac{mn(1-q)}{2(n+qm)} Z_L$ and ${\mathcal{E}}_{L\leftarrow S}^\dagger(H_S^2) = \frac{mn(m+qn)}{4(n+qm)}{\mathbbm{1}}_L$ so we have
\begin{align}
\delta_{\mathrm{C}} =\;& \Delta\left({H_L - {\mathcal{E}}_{{L\leftarrow S}}^\dagger(H_S)}\right)
\nonumber\\
=\;& \frac{qm(n+m)}{(n+qm)} = qm + O\left(m\cdot\frac{m}{n}\right).
\end{align}
Also note that the parameter ${\mathfrak{B}}$ which shows up in \thmref{thm:global-1} and \thmref{thm:local-metrology} is given by
\begin{align}
{\mathfrak{B}} =\;& 2\sqrt{2}\max_{\ket{\psi}}\sqrt{\braket{\psi|{\mathcal{E}}^\dagger(H_S^2)|\psi}-|\!\braket{\psi|{\mathcal{E}}^\dagger(H_S)|\psi}\!|^2} \nonumber\\
=\;& \sqrt{\frac{2mn(m+qn)}{n+qm}}
= \sqrt{2qnm} + O\left(m\sqrt{\frac{m}{n}}\right),
\end{align}
when $qn \gg m$ and ${\mathfrak{B}} = O(m)$ otherwise.
\subsubsection{Trade-off between QEC and symmetry, and explicit comparisons with lower bounds}
\begin{figure}[tb]
\center
\includegraphics[width=0.4\textwidth]{thermo-3.pdf}
\caption{\label{fig:thermo} Various types of symmetry measures and QEC accuracy of thermodynamic codes, where $m = 2$, $n \in [2^6,2^{10}]$ and $q \in [10^{-5},1-10^{-5}]$. Here we plot $\delta_\group$, $\delta_{\mathrm{P}}$, $\delta_C$ and $\tilde{\epsilon}$ (which is approximately equal to $\epsilon$ for large $n$). As $q$ increases from $0$ to $1$, $\tilde{\epsilon}$ decreases while the symmetry measures increase. We can also see that fixing $q$, the slopes of $\delta_\group$, $\delta_{\mathrm{P}}$, $\delta_C$ and $\tilde{\epsilon}$ with respect to $n$ are $-1/2$, $1/2$, $0$ and $-1$ respectively, matching our calculations.
}
\end{figure}
Let us first overview the behavior of modified thermodynamic codes.
Our calculations above indicate that up to the leading order, $\epsilon \simeq (1-q)m/2n$, while $\delta_\group \simeq \sqrt{4qm/n}$, $\delta_{\mathrm{P}} \simeq \sqrt{qmn}$, and $\delta_{\mathrm{C}} \simeq qm$ (see \figref{fig:thermo}). That is, as $q$ varies from $0$ to $1$, the symmetry violation (in terms of different measures) and the QEC inaccuracy exhibit trade-off behaviors---the former increases from $0$ while the latter decreases to $0$.
We now discuss the comparison with our lower bounds, focusing on the large $n$ asymptotics. Note that $H_L = \frac{m}{2}Z_L$ and $H_S = -\frac{1}{2}\sum_{l=1}^n Z_l$, so we have $\Delta H_L = m$ and $\Delta H_{S_l} = 1$ for each $l$. For single-erasure noise channels (as shown in \secref{sec:global-1-noise}), we have ${\mathfrak{J}} \leq n \max_l \Delta H_{S_l} = n$ and ${\mathfrak{F}} \leq n\sum_{l=1}^n (\Delta H_{S_l})^2 = n^2$, and \thmref{thm:global-1} then gives:
\begin{equation}
\label{eq:thermo-tradeoff}
\delta_\group \geq \sqrt{\frac{m-2n\epsilon}{n}} + O\left(\left(\frac{m}{n}\right)^{3/2}\right).
\end{equation}
Plugging in the QEC inaccuracy $\epsilon \simeq \sqrt{(1-q)m/n}$, we have
\begin{equation}
\delta_\group \geq \sqrt{\frac{qm}{n}} + O\left(\left(\frac{m}{n}\right)^{3/2}\right).
\end{equation}
Recall that for the modified thermodynamic code we have $\delta_\group \simeq \sqrt{4qm/n}$, which saturates this lower bound on $\delta_\group$ up to a constant factor of 2 in the leading order of $m/n$.
Similarly, we could also plug the actual value $\delta_\group \simeq \sqrt{4qm/n}$ into \eqref{eq:thermo-tradeoff} and obtain the lower bound
\begin{equation}
\epsilon \geq \frac{(1-4q)m}{2n} + O\left(\frac{m^2}{n^2}\right),
\end{equation}
which shows that the actual value $\epsilon \simeq (1-q)m/2n$ of the modified thermodynamic code saturates this lower bound up to a constant factor in the leading order of $m/n$ for $q < 1/4$.
For the local symmetry measures, we first note that for the modified thermodynamic code with $q>0$ we have $\delta_{\mathrm{P}}/m \simeq \sqrt{qn/m}$ which becomes larger than $1$ as $m/n\rightarrow 0$, thus \thmref{thm:local-point} is not saturated. We provide one possible explanation of this gap between $\delta_{\mathrm{P}}$ and its lower bound in \appref{app:refine}, where we show a refinement of \thmref{thm:local-point} by replacing $\delta_{\mathrm{P}}$ in \thmref{thm:local-point} with $\delta_{\mathrm{P}}^\star$ ($\leq \delta_{\mathrm{P}}$) which is defined using the QFI of the error-corrected channel ${\mathcal{R}}^{\mathrm{opt}}_{{L\leftarrow S}}\circ{\mathcal{N}}_{S,\theta}\circ{\mathcal{E}}_{{S\leftarrow L}}\circ{\mathcal{U}}_{L,\theta}^\dagger$ instead of the QFI of ${\mathcal{U}}_{S,\theta}\circ{\mathcal{E}}_{{S\leftarrow L}}\circ{\mathcal{U}}_{L,\theta}^\dagger$ at $\theta = 0$. We show that the gap between $\delta_{\mathrm{P}}$ and its lower bound could be explained by its gap with $\delta_{\mathrm{P}}^*$, explaining the looseness of \thmref{thm:local-point}.
Recall that for the modified thermodynamic code the charge conservation violation is $\delta_{\mathrm{C}} \simeq qm$. From \thmref{thm:local-KL}, we have
\begin{equation}
\delta_{\mathrm{C}} \geq \Delta H_L - 2\epsilon{\mathfrak{J}} = qm + O\left(\frac{m^2}{n}\right).
\end{equation}
Namely, $\delta_{\mathrm{C}}$ exactly saturates this lower bound in the leading order of $m/n$.
Note that ${\mathfrak{B}} = O(\sqrt{n})$ and ${\mathfrak{F}} = O(n^2)$ satisfies ${\mathfrak{B}} \ll \sqrt{{\mathfrak{F}}}$ in \thmref{thm:global-1} and \thmref{thm:local-metrology} so that ${\mathfrak{B}}$ is negligible in the trade-off relations from \thmref{thm:global-1} and \thmref{thm:local-metrology} for modified thermodynamic codes in the large $n$ asymptotics.
Finally, note that the trade-off relation given by the diamond distance,
\begin{equation}
\delta_{\group,\diamond}^2 + 2\sqrt{\epsilon_\diamond} \geq \delta_\group^2 + 2\epsilon \geq \frac{m}{n},
\end{equation}
which follows from \thmref{thm:global-1} and the discussion in \secref{sec:def}, is also saturated up to a constant factor because $\delta_{\group,\diamond} = \delta_\group$ and $\epsilon_\diamond = \epsilon^2 + O(m^3/n^3)$ (see \appref{app:thermo} for details).
\subsection{Quantum Reed--Muller codes}
\label{sec:RM}
Reed--Muller codes constitute a family of error-correcting codes of great theoretical and technological interest.
The classical Reed--Muller code $R(t',t)$~\cite{macwilliams1977theory} is a $[2^t,\sum_{i=0}^r\binom{t}{i},2^{t-t'}]$ code whose codewords correspond to Boolean functions of $t$ variables of degree at most $r$. Then the shortened Reed--Muller codes $\overline{R(t',t)} = [2^t-1,\sum_{i=1}^r\binom{t}{i},2^{t-t'}]$ are obtained by selecting the codewords of $R(t',t)$ whose first digits are $0$ and deleting their first digits.
The generalization to the quantum regime based on the stabilizer formalism and CSS construction, which leads to the quantum Reed--Muller codes, are also an important type of QEC codes~\cite{steane1999quantum}. Given the nice structures and features of quantum Reed--Muller codes, they provide a natural platform for understanding code properties. {For example, quantum Reed--Muller codes were widely applied in magic state distillation and implementing transversal non-Clifford operations~\cite{bravyi2012magic,anderson2014fault,haah2018codes,hastings2018distillation}. Quantum Reed--Muller codes were also known to reach the highest level of the Clifford hierarchy possible under the disjointness restriction~\cite{jochym2018disjointness}.}
Here we consider the $[[n = 2^t - 1,1,3]]$ quantum Reed--Muller code, which is a CSS code~\cite{nielsen2002quantum} whose $X$ stabilizers correspond to $\overline{R(1,t)}$ and $Z$ stabilizers correspond to $\overline{R(t-2,t)}$. It is exactly error-correcting under single-erasure noise and admits a transversal implementation $\bigotimes_l \big(e^{i\pi Z_l/2^{t-1}}\big)$ of the logical operator $e^{-i\pi Z_L/2^{t-1}}$.
We now compute its symmetry violation measures.
This code has the following form in the computation basis:
\begin{align}
\ket{{\mathfrak{c}}_0} &
= \frac{1}{\sqrt{2^t}} \bigg(\ket{\v{0}} + \sum_{\v{x} \in \overline{R(1,t)}\backslash\{\v{0}\}} \ket{\v{x}}\bigg), \\
\ket{{\mathfrak{c}}_1} &
= \frac{1}{\sqrt{2^t}} \bigg(\ket{\v{1}} + \sum_{\v{x} \in \overline{R(1,t)}\backslash\{\v{0}\}} \ket{\v{1} + \v{x}}\bigg),
\end{align}
where we use $\v{x}$ to denote $n$-bit strings. All strings in $\overline{R(1,t)}\backslash\{\v{0}\}$ have weight $2^{t-1}$.
Let $W$ be the encoding isometry $W = \ket{{\mathfrak{c}}_0}\bra{0_L} + \ket{{\mathfrak{c}}_1}\bra{1_L}$. {Consider $H_L = \frac{1}{2}Z_L$ and $H_S = -\frac{1}{2}\sum_{l=1}^n Z_l$, which again guarantees that the code tends to be covariant as $n\rightarrow\infty$.}
Let $\ket{\psi} = \ket{0_L}\ket{\psi_R^0} + \ket{1_L}\ket{\psi_R^1}$ be an arbitrary pure state on $L \otimes R$. Then
\begin{align*}
&\ket{\psi_\theta}\! = U_{S,\theta} W U_{L,\theta}^\dagger \ket{\psi} \nonumber\\
= &\bigg(\!\frac{1}{\sqrt{2^t}} \bigg(e^{i2^{t-1}{\theta}} \ket{\v{0}} + \sum_{\v{x} \in \overline{R(1,t)}\backslash\{\v{0}\}} \ket{\v{x}}\bigg)\!\bigg)\!\ket{\psi_R^0} \nonumber
\\ &+ \bigg(\!\frac{1}{\sqrt{2^t}} \bigg(e^{-i2^{t-1}{\theta}}\ket{\v{1}} + \sum_{\v{x} \in \overline{R(1,t)}\backslash\{\v{0}\}} \ket{\v{1} + \v{x}}\bigg)\!\bigg)\!\ket{\psi_R^1},
\end{align*}
and thus
\begin{align}\label{eq:rm_p}
P({\mathcal{U}}_{S,\theta}\circ{\mathcal{E}}_{{S\leftarrow L}},{\mathcal{E}}_{{S\leftarrow L}}\circ{\mathcal{U}}_{L,\theta}) = \max_{\psi} P(\ket{\psi_\theta},\ket{\psi})\nonumber\\
= \sqrt{1 - \left|\frac{n + \cos((n+1)\theta/2)}{n + 1}\right|^2}.
\end{align}
Therefore, we have
\begin{align}
\delta_\group & = \max_\theta P({\mathcal{U}}_{S,\theta}\circ{\mathcal{E}}_{{S\leftarrow L}}\circ {\mathcal{U}}_{L,\theta}^\dagger, {\mathcal{E}}_{{S\leftarrow L}})\nonumber\\ & = \sqrt{1 - \left(\frac{n - 1}{n + 1}\right)^2} = \sqrt{\frac{4}{n}} + O\bigg(\bigg(\frac{1}{n}\bigg)^{3/2}\bigg).
\end{align}
The lower bound from \thmref{thm:global-1} gives
\begin{align}
\delta_\group &\geq \sqrt{\frac{\Delta H_L(\Delta H_S - \frac{1}{2}\Delta H_L)}{\Delta H_S}} \nonumber\\ &= \sqrt{\frac{n - \frac{1}{2}}{n^2}} = \sqrt{\frac{1}{n}} + O\bigg(\bigg(\frac{1}{n}\bigg)^{3/2}\bigg).
\end{align}
Similar to the modified thermodynamic code, $\delta_\group$ saturates its lower bound up to a constant factor of 2 in the leading order of $1/n$. Also note that $\delta_{\group,\diamond} = \delta_\group$ in this case according to the discussion in \secref{sec:def}, indicating the saturation of the lower bound when we consider the diamond distance. The local covariance violation can also be easily computed from \eqref{eq:rm_p}:
\begin{equation}
\delta_{\mathrm{P}} \!=\! \sqrt{2 \partial^2_\theta P({\mathcal{U}}_{S,\theta}\circ{\mathcal{E}}_{{S\leftarrow L}},{\mathcal{E}}_{{S\leftarrow L}}\circ{\mathcal{U}}_{L,\theta})^2\big|_{\theta = 0}} \!=\! \sqrt{n+1},
\end{equation}
which has a quadratic gap with its lower bound $\Delta H_L = 1$.
{To compute the charge conservation violation $\delta_{\mathrm{C}}$, we note that ${\mathcal{E}}^\dagger_{{L\leftarrow S}}(H_S) = 0$ and ${\mathcal{E}}^\dagger_{{L\leftarrow S}}(H_S^2) = n/4$, thus
\begin{align}
\delta_{\mathrm{C}} = \Delta\left( H_L - {\mathcal{E}}^\dagger_{{L\leftarrow S}}(H_S)\right) = 1,
\end{align}
matching our lower bound $\Delta H_L$.
Also note that ${\mathfrak{B}} = \sqrt{2n}$, so it is negligible in the trade-off relations from \thmref{thm:global-1} and \thmref{thm:local-metrology}. }
\section{Discussion and outlook}
\label{sec:discussion}
In this work, we devised and explored various approaches that allow us to quantitatively understand the fundamental trade-off between the QEC capability and several different characterizations of the degree of continuous symmetries associated with general QEC codes, including the violation of covariance conditions in both global and local senses as well as the violation of charge conservation (see \appref{app:comparison} for a summary). In particular, we introduced two intuitive and powerful frameworks respectively based on the notions of charge fluctuation and gate implementation error, and
employed several different methods from approximate QEC, quantum metrology, and quantum resource theory, to derive various forms of the trade-off relations in terms of distance metrics that address both worst-case and average-case inputs.
Our results and techniques are expected to have numerous interesting applications to quantum computation and physics. We specifically discussed the consequent restrictions on the transversal logical gates for general QEC codes, which could be of interest for fault tolerance. We also provided detailed analysis of two interesting examples of approximately covariant codes---a parametrized extension of the (covariant) thermodynamic code, which gives a code family that continuously interpolates between exactly covariant and error-correcting, and the quantum Reed--Muller codes. We showed that both codes can saturate the lower bounds asymptotically up to constant factors, indicating that the bounds are quite tight.
We would like to point out a few problems arising from our technical analysis that are not yet satisfactorily understood and could be worth further investigation:
\begin{itemize}
\item For both of the code examples we studied, the global covariance violation and charge conservation violation attain the optimal asymptotic scaling as indicated by the bounds based on the charge fluctuation approach, but the local covariance violation does not (there is a $\Theta(\sqrt{n})$ vs.\ $O(1)$ gap). Note the observation (discussed above and in \appref{app:refine}) that if we additionally consider a recovery step in the definition of local covariance violation then it exhibits a tight scaling. We would hope to close this gap by further understanding both sides of it. This is potentially key to a complete understanding of the behavior and practical meanings of the local symmetry measures.
\item The gate implementation error approach provides bounds that behave worse than the corresponding bounds from the charge fluctuation approach under uniformly random local noise. It would be interesting to further understand whether this gap stems from the looseness of \propref{prop:gate}. On the other hand, the discussion in \secref{sec:comparison} also indicates that for extremely strong noise (so that the regularized QFI is sublinear), the gate implementation error approach outperforms the charge fluctuation approach.
The question remains whether there is a universal bound which exhibits optimal scalings under any noise models.
\end{itemize}
There are also several important directions for future study:
\begin{itemize}
\item Gate implementation error. We introduced the gate implementation error as a notion that nicely unifies QEC inaccuracy and global covariance violation, and in turn serves as a tool for deriving the trade-off between them. We believe that this quantity is interesting in its own right and expect it to find broader applications in the analysis of QEC, distillation etc.
\item General continuous symmetry groups. Here we mainly carried out the discussion in terms of $U(1)$ which corresponds to a single conserved quantity, but obviously the symmetry groups are often more complicated in quantum computation and physical scenarios. It would be useful to extend our study to other important continuous symmetry groups such as $SU(d)$, for which we expect that our analysis for $U(1)$ provides a toolkit and a sub-theory but it is useful to invoke corresponding representation theory machinery (like in Refs.~\cite{faist2019continuous,KongLiu21:random}).
\item Discrete symmetries. Given the incompatibility results for continuous symmetries, it is natural to ask whether discrete symmetries, which are also broadly important, place restrictions on QEC. It is known that for any discrete symmetry group one can in principle construct a covariant code which also achieves exact QEC \cite{hayden2017error}, indicating that the incompatibility is not as fundamental as continuous symmetries. However, we do know interesting cases where exact QEC is forbidden even in the presence of discrete symmetries under simple additional constraints (e.g.,~AdS/CFT codes---see Refs.~\cite{harlow2018constraints,harlow2018symmetries,faist2019continuous}). It would be interesting to further explore both the possible limitations as well as good code constructions for QEC with discrete symmetries in more general terms.
\end{itemize}
Furthermore, we expect the study of how QEC interacts symmetries to be relevant in wide-ranging physical scenarios. In the companion paper \cite{short} we discussed potential applications of our theory and techniques to several topics of great interest in physics, including AdS/CFT, black hole radiation, and many-body physics. It would be interesting to further consolidate these ideas. To this end, an important task is to bridge the language of quantum information used here and those commonly used in high energy and condensed matter physics. To conclude, our study enriches the ``physical'' understanding of QEC using a wide variety of approaches in quantum information. We hope it will stimulate further interest into exploring the interaction between QEC, quantum information, and physics.
\begin{acknowledgments}
We thank Daniel Gottesman, Liang Jiang, Dong-Sheng Wang, Beni Yoshida for relevant discussion and feedback.
Z.-W.L. is supported by Perimeter Institute for Theoretical Physics.
Research at Perimeter Institute is supported in part by the Government of Canada through the Department of Innovation, Science and Economic Development Canada and by the Province of Ontario through the Ministry of Colleges and Universities.
S.Z. acknowledges funding provided by the Institute for Quantum Information and Matter, an NSF Physics Frontiers Center (NSF Grant PHY-1733907). S.Z. also acknowledges support through the University of Chicago from ARO (W911NF-18-1-0020, W911NF-18-1-0212), ARO MURI (W911NF-16-1-0349), AFOSR MURI (FA9550-19-1-0399, FA9550-21-1-0209), DoE Q-NEXT, NSF (EFMA-1640959, OMA-1936118, EEC-1941583), NTT Research, and the Packard Foundation (2013-39273), where part of this work was done.
\end{acknowledgments}
|
\section{Summary}
In this paper, we show that objectiveness could emerge from a computational model by exposing it to unlabeled videos.
We present a novel model which decouples the appearance pathway and the motion pathway, and later binds them into a joint segment flow representation.
As opposed to prior works that rely heavily on accurate dense optical flows for predicting object segmentation, our method learns only from raw pixel observations.
The motion representation in our model is a lot weaker, however the object segmentation is more robust.
The proposed model AMD is the first end-to-end learning approach for zero-shot object segmentation without using any pretrained modules.
Its performance is validated on a number of image and video object segmentation benchmarks.
\section{Experiments}
\label{sec:exp}
We demonstrate that AMD model can be transferred to three downstream applications.
First, the appearance pathway is directly applied on static images for salient object detection in a zero-shot fashion.
Second, both the appearance and the motion pathway are transferred to video object segmentation in novel videos with zero human labels.
Third, we fine-tune the appearance pathway on labeled data for semantic segmentation.
\subsection{Training and Implementation Details}
\label{sec:imp}
AMD is pretrained on the large object-centric video dataset Youtube-VOS~\cite{xu2018youtube}.
The training split for Youtube-VOS contains about 4,000 videos covering 94 categories of objects. The total duration for the dataset is $334$ minutes.
We train the model on the data with a sampling rate 24 frames per second, without using the original segmentation labels.
We train all model parameters from scratch without external pretraining. For the segmentation network, we use ResNet50~\cite{he2016deep} as our backbone followed by a fully convolutional head containing two convolutional blocks. For the motion network, we adopt PWC-Net~\cite{sun2018pwc} architecture because of its effectiveness in estimating optical flows.
We resize the short edge of the input image to $400$ pixels, and random crop a square image of size $384\times 384 $ with random horizontal flipping augmentation.
No other augmentations is engineered.
We adopt the symmetric loss that considers either frame as the target frame and sums the two reconstruction errors.
For training the overall model, we use the Adam optimizer with a learning rate of $\num{1e-4}$ and a weight decay of $\num{1e-6}$.
We train AMD on eight V100 GPUs, with each processing two pairs of sampled adjacent frames.
The network is optimized for 400K iterations.
\subsection{Zero-Shot Saliency Detection}
\label{sec:saliency}
\def\prow#1#2#3#4{
\imw{figs/fig_saliency/#1.png}{0.244}&
\hspace{0.01cm}
\imw{figs/fig_saliency/#2.png}{0.244}&
\hspace{0.01cm}
\imw{figs/fig_saliency/#3.png}{0.244}&
\hspace{0.005cm}
\vrule
\hspace{0.005cm}
\imw{figs/fig_saliency/#4.png}{0.244}\\
}
\begin{figure}[t]
\centering
\subfloat{\small
\tb{@{}cccc@{}}{0.05}{
\prow{0}{1}{22}{12}
\prow{3}{18}{21}{13}
\prow{6}{7}{8}{0003}
\prow{16}{19}{11}{20}
& movable objects & &
stationary objects
}
}
\caption{Qualitative salient object detection results. We directly transfer our pretrained segmentation network to novel images on the DUTS dataset without any finetuning. Surprisingly, we find that the model pretrained on videos to segment moving objects can generalize to detect stationary unmovable objects in a static image, e.g. the statue, the plate, the bench and the tree in the last column.}
\label{fig:saliency}
\vspace{-7pt}
\end{figure}
Once pretrained, AMD's appearance pathway can be directly transferred to object segmentation in novel stationary images without any downstream fine-tuning. To evaluate the quality of the segmentation, we benchmark the results on salient object detection benchmark.
The salient object detection performance is measured on the \textbf{DUTS~\cite{wang2017learning}} benchmark, which contains 5,019 test images with pixel-level ground truth annotations.
We follow two widely used metrics in this area: the $F_\beta$ score and the per-pixel mean squared errors (MAE). $F_\beta$ is defined as the weighted harmonic mean of the precision ($P$) and recall ($R$) scores: $F_\beta = \frac{(1+\beta^2) P \times R}{\beta^2 P + R}$, with $\beta^2 = 0.3$.
MAE is simply the per-pixel averaged error of the soft prediction scores.
\textbf{Experimental results.}
We compare our saliency estimation results against several traditional methods based on low-level cues.
Useful low-level cues and priors include background priors~\cite{zhu2014saliency}, objectness~\cite{jiang2013saliency,jiang2013salient}, and color contrast~\cite{cheng2014global}.
As shown in Table~\ref{tab:saliency}, our method achieves an $F_\beta$ score $60.2$ and an MAE score of $0.13$, outperforming all traditional approaches by a notable margin.
We note that AMD is not designed specifically for this task nor for this particular dataset, and its strong performance demonstrates the generalization ability of the model.
In related work on unsupervised learning of saliency detection~\cite{zhang2018deep,zeng2019multi,nguyen2019deepusps}, the priors of traditional low-level methods are ensembled.
Though they do not use saliency annotations, their models are pretrained on ImageNet classification and even semantic segmentation with pixel-level annotations. These methods are thus not fully unsupervised, so they are omitted from the comparisons.
In Figure~\ref{fig:saliency}, we show some qualitative results on salient object detection.
Surprisingly, we find that our model pretrained on videos to segment moving objects not only detects movable objects in images, but also generalizes to detect stationary unmovable objects, such as statues, benches, trees and plates shown in the last column. This suggests that our model learns a generic objectness prior from the unlabeled videos.
We hypothesize that our model may learn objectness from the camera motion as well. Camera motion may cause the object and the background at various depth to have different observed projective 2D optical flow even though the objects are static.
\subsection{Zero-shot Video Object Segmentation}
\label{sec:vos}
\begin{table}[t]
\begin{minipage}{0.4\linewidth}
\centering
\caption{Salient object detection performance on the DUTS dataset. Our model outperforms traditional low-level methods by notable margins.}
\vspace{3pt}
\setlength{\tabcolsep}{12pt}
\begin{tabular}[t]{c|cc}
\hline
Model & $F_{\beta}$ & MAE \\
\hline
RBD\cite{zhu2014saliency} & 51.0 & 0.20 \\
HS\cite{zou2015harf} & 52.1 & 0.23 \\
MC\cite{jiang2013saliency} & 52.9 & 0.19 \\
DSR\cite{li2013saliency} & 55.8 & 0.14 \\
DRFI\cite{jiang2013salient} & 55.2 & 0.15 \\
\hline
\textbf{AMD} & \textbf{60.2} & \textbf{0.13} \\
\hline
\end{tabular}
\label{tab:saliency}
\end{minipage}
\quad
\begin{minipage}{0.55\linewidth}
\centering
\caption{Transfer performance for semantic segmentation on VOC2012. Our method outperforms TimeCycle and compares favorably with contrastive methods.}
\vspace{3pt}
\setlength{\tabcolsep}{10pt}
\begin{tabular}[t]{c|c|c|c}
\hline
Model & Data & Aug. & mIoU \\
\hline
Scratch & -- & -- & 48.0 \\
TimeCyle\cite{wang2019learning} & VLOG & light & 52.8 \\
MoCo-v2\cite{he2020momentum} & YTB & light & 61.5 \\
\textbf{AMD} & YTB & light & \textbf{62.0} \\
\hline
MoCo-v2\cite{he2020momentum} & YTB & heavy & \textbf{62.8} \\
\textbf{AMD} & YTB & heavy & 62.1 \\
\hline
MoCo-v2\cite{he2020momentum} & IMN & heavy & \textbf{72.4} \\
\hline
\end{tabular}
\label{tab:seg}
\end{minipage}
\vspace{-10pt}
\end{table}
We transfer the pretrained AMD model to object segmentation on novel videos.
Since the segmentation prediction from our model is based on static images, inference on images sequentially in a video essentially estimates objectness.
In order to exploit motion information, we use a test-time adaptation approach.
Concretely, given a novel video, we optimize the training objective in Eq.~\ref{eq:loss} on pairs of frames sampled from the novel testing video. The adaptation takes 100 iterations per video.
We evaluate zero-shot video object segmentation on three testing datasets.
{\bf DAVIS 2016~\cite{Perazzi2016}} contains 20 validation videos with 1,376 annotated frames.
{\bf SegTrackv2~\cite{li2013video}} contains 14 videos with 976 annotated frames. Following prior works, we combine multiple foreground objects in the annotation into a single object for evaluation.
{\bf FBMS59 ~\cite{ochs2013segmentation}} contains 59 videos with 720 annotated frames. The dataset is relatively challenging because the object may be static for a period of time. We pre-process the ground truth labels following prior work~\cite{yang2019unsupervised}.
For evaluation, we report the Jaccard score, which is equivalent to the intersection over union (IoU) between the prediction and the ground truth segmentation.
\textbf{Experimental results.}
We consider baseline methods claiming to be unsupervised for the full pipeline, including traditional non-learning-based approaches~\cite{wang2017saliency,faktor2014video,papazoglou2013fast,keuper2015motion,koh2017primary} and recent self-supervised learning methods~\cite{yang2019unsupervised,yang2021self}.
In Table~\ref{tab:FID}, we summarize the results for all the methods on the three datasets.
Among these methods, NLC~\cite{faktor2014video} actually relies on an edge model trained with human-annotated edge boundaries, and ARP~\cite{koh2017primary} depends on a segmentation model trained on a human-annotated saliency dataset. We thus gray their entries in the table.
For all the traditional methods, since their original papers do not report results on most of these benchmarks, we simply provide the performance values reported in the CIS paper~\cite{yang2019unsupervised}.
We evaluate the performance for AMD with and without test-time adaptation. No adaptation boils down to per-image saliency estimation using only the appearance pathway. Adaptation transfers both appearance and motion pathways.
On DAVIS 2016, our method achieves a Jaccard score of $57.8\%$, surpassing all traditional unsupervised models.
For CIS~\cite{yang2019unsupervised}, their best performing model uses a significant amount of post-processing, including model ensembing, multi-crop, temporal smoothing and spatial smoothing.
We thus refer to their performance obtained from a single model without post-processing.
Our model is slightly worse than CIS on DAVIS, by $1.4\%$.
However, on SegTrackv2 and FBMS59, our method outperforms CIS by large margins of $11.4\%$ and $10.7\%$, respectively.
Motion grouping~\cite{yang2021self} is a work concurrent with ours.
It is essentially a motion segmentation approach which relies on an off-the-shelf pre-computed dense optical flow model.
Motion grouping performs worse than our method on DAVIS2016 and SegTrackv2 when a low-performance unsupervised optical flow model ARFlow is used~\cite{liu2020learning}.
With a state-of-the-art supervised optical flow model~\cite{teed2020raft} that is trained on ground truth flow, their performance improves significantly.
Among all the discussed methods, ours is the first end-to-end self-supervised learning approach which does not require a pretrained optical flow model.
In Figure~\ref{fig:vos}, we show qualitative comparisons to the baseline CIS~\cite{yang2019unsupervised}. We display dense flow from pretrained PWC-Net, the CIS results, our segment flows, and our segmentation results. For most of these examples, our segment flow only coarsely reflects the true pixel-level optical flow. However, our segmentation results are significantly better and less noisy, as our model is relatively insensitive to optical flow quality.
In the first and the third examples, our model produces high-quality object segmentations even though the motion cue for the objects is very subtle.
\def\rowone#1#2#3#4#5#6#7{
\imw{figs/fig_vos/#7/#1.png}{0.152}&
\hspace{0.01cm}
\imw{figs/fig_vos/#7/#2.png}{0.152}&
\hspace{0.01cm}
\imw{figs/fig_vos/#7/#3.png}{0.152}&
\hspace{0.01cm}
\imw{figs/fig_vos/#7/#4.png}{0.152}&
\hspace{0.01cm}
\imw{figs/fig_vos/#7/#5.png}{0.152}&
\hspace{0.01cm}
\imw{figs/fig_vos/#7/#6.png}{0.152}\\
}
\begin{figure}[t]
\centering
\subfloat{\small
\tb{@{}ccccccc@{}}{0.03}{
\rotatebox{90}{\makecell[c]{\hspace{0.15cm}dense\\\hspace{0.15cm}flow}} & \rowone{1}{5}{9}{17}{15}{18}{cis}
\vspace{0.03cm}
\rotatebox{90}{\hspace{0.0cm}CIS~\cite{yang2019unsupervised}} & \rowone{0}{4}{8}{16}{14}{19}{cis}
\rotatebox{90}{\makecell[c]{segment\\flow}} & \rowone{10}{0}{8}{14}{16}{18}{ours}
\rotatebox{90}{\hspace{0.3cm}AMD} & \rowone{11}{1}{9}{15}{17}{19}{ours}
}
}
\caption{Qualitative comparisons to motion segmentation based method CIS~\cite{yang2019unsupervised} with its input dense flow and our segmentation results with segment flow representation. CIS is prone to noise, articulated motion, and camera motion in the dense flow estimations. %
By decomposing appearance from motion, our model AMD suffers less from these vulnerabilities of dense optical flow. This leads to results much better and more robust than the motion segmentation based approach.}
\label{fig:vos}
\vspace{-3pt}
\end{figure}
\begin{table}[t]
\centering
\caption{Performance evaluations for unsupervised video object segmentation on DAVIS 2016, SegTrackv2 and FBMS59 datasets. The numbers are measured in terms of Jaccard score. The table is split into traditional non-learning-based and recent self-supervised learning methods. The model results which rely on other kinds of human supervisions (Sup.) are \textcolor[gray]{0.5}{grayed}. Dependence for pretrained dense flow method is also listed for each model. MG's results on SegTrackv2 and FMBS59 using ARFlow are reproduced by ours and marked with $*$.
We evaluate AMD with appearance pathway only and with both pathways for test time adaptation.
AMD performs favorably to the baseline CIS on DAVIS 2016, while showing large gains on the other two benchmarks.} %
\vspace{5pt}
\setlength{\tabcolsep}{6pt}
\begin{tabular}[t]{ccccc|ccc}
\hline
\hline
& Model & e2e & Sup. & Flow & DAVIS 2016 & SegTrackv2 & FBMS59 \\
\hline
\parbox[t]{0.1mm}{\multirow{5}{*}{\rotatebox[origin=c]{90}{traditional}}} & SAGE\cite{wang2017saliency} & \xmark & \xmark & LDOF\cite{brox2010large} & 42.6 & 57.6 & 61.2 \\
& NLC\cite{faktor2014video} & {\xmark} & {edge} & {SIFTFlow}\cite{liu2009beyond} & \textcolor[gray]{0.5}{55.1} & \textcolor[gray]{0.5}{67.2} & \textcolor[gray]{0.5}{51.5} \\
& CUT\cite{keuper2015motion} & \xmark & \xmark & LDOF\cite{brox2010large} & 55.2 & 54.3 & 57.2 \\
& FTS\cite{papazoglou2013fast} & \xmark & \xmark & LDOF\cite{sundaram2010dense} & 55.8 & 47.8 & 47.7 \\
& {ARP}\cite{koh2017primary} & {\xmark} & {saliency} & {CPMFlow}\cite{hu2016cpm} & \textcolor[gray]{0.5}{76.2} & \textcolor[gray]{0.5}{57.2} & \textcolor[gray]{0.5}{59.8} \\
\hline
\parbox[t]{0.1mm}{\multirow{4}{*}{\rotatebox[origin=c]{90}{learning}}} & CIS\cite{yang2019unsupervised} & \xmark & \xmark & PWC\cite{sun2018pwc} & 59.2 & 45.6 & 36.8 \\
& MG\cite{yang2021self} & \xmark & \xmark & ARFlow\cite{liu2020learning} & 53.2 & 37.8$^*$ & 50.4$^*$ \\
& \textbf{AMD} {\footnotesize (per-img)} & \cmark & \xmark & \xmark & 45.7 & 28.7 & 42.9 \\
& \textbf{AMD} {\footnotesize (per-vid)} & \cmark & \xmark & \xmark & 57.8 & 57.0 & 47.5 \\
\hline
\hline
\end{tabular}
\label{tab:FID}
\vspace{-3pt}
\end{table}
\subsection{Semantic Segmentation}
\label{sec:semantic}
Given that our pretrained segmentation network can produce meaningful generic object segmentations, we further examine its semantic modeling ability on semantic segmentation.
We conduct this experiment on the \textbf{Pascal VOC 2012~\cite{everingham2010pascal}} dataset.
The dataset contains 20 object categories with 10,582 training images and 1,449 validation images.
Given a pretrained model, we finetune the model on the training set and evaluate the performance on the validation set.
The finetuning takes 40,000 iterations with a batch size of 16 and an initial learning rate of 0.01.
The learning rate undergoes polynomial decay with a power parameter of 0.9.
\textbf{Experimental results.}
We compare the pretrained model to a image-based contrastive model, MoCo-v2~\cite{he2020momentum} and a self-supervised video pretraining model, TimeCycle~\cite{wang2019learning}.
TimeCycle~\cite{wang2019learning} is pretrained on the VLOG dataset, which is larger than our Youtube-VOS dataset.
For MoCo-v2, we also pretrain the contrastive model on the Youtube-VOS dataset, to ablate the role of pretraining datasets.
Since the base version of our method does not utilize heavy augmentations as in contrastive models, we also study the effects of data augmentations.
The results are reported in Table~\ref{tab:seg}.
Our method outperforms the video pretraining approach TimeCylce significantly by $9.2\%$.
Compared with MoCo-v2, when light augmentation (resizing, cropping) is used, our model slightly outperforms MoCo-v2 by $0.5\%$. However, when heavy data augmentation (color jitter, grayscale, blurring) is applied, our method underperforms MoCo-v2 by $0.7\%$.
This is possibly because our model is non-contrastive in nature, and thus unable to take advantage of information effectively in augmentations.
MoCo-v2 performs much stronger when pretrained on ImageNet, possibly because the semantic distribution of ImageNet is well aligned with VOC2012.
Overall, our model outperforms a prior self-supervised video model TimeCycle and compares favorably with contrastive model MoCo-v2 under the same data.
\def\rowone#1#2#3#4#5#6{
\imw{figs/fig_segnum/#1}{0.16}&
\hspace{0.003cm}
\imw{figs/fig_segnum/#2}{0.16}&
\hspace{0.003cm}
\imw{figs/fig_segnum/#3}{0.16}&
\hspace{0.0015cm}
\vrule
\hspace{0.0015cm}
\imw{figs/fig_segnum/#4}{0.16}&
\hspace{0.003cm}
\imw{figs/fig_segnum/#5}{0.16}&
\hspace{0.003cm}
\imw{figs/fig_segnum/#6}{0.16}\\
}
\begin{figure}[t]
\begin{minipage}{0.7\linewidth}
\centering
\subfloat{\small
\tb{@{}cccccc@{}}{0.05}{
\rowone{drift-straight_00031_iter0000100_im_msk_l5.png}{drift-straight_00031_iter0000100_im_msk_l6.png}{drift-straight_00031_iter0000300_im_msk_l8.png}{blackswan_00008_iter0000200_im_msk_l5.png}{blackswan_00008_iter0000300_im_msk_l6.png}{blackswan_00008_iter0000300_im_msk_l8.png}
\rowone{drift-straight_00031_iter0000100_f2_l5.png}{drift-straight_00031_iter0000100_f2_l6.png}{drift-straight_00031_iter0000300_f2_l8.png}{blackswan_00008_iter0000200_f1_l5.png}{blackswan_00008_iter0000300_f1_l6.png}{blackswan_00008_iter0000300_f1_l8.png}
$c=5$ & $c=6$ & $c=8$ & $c=5$ & $c=6$ & $c=8$ \\
}
}
\label{fig:ab}
\end{minipage}
\quad
\begin{minipage}{0.2\linewidth}
\centering
\setlength{\tabcolsep}{4pt}
\begin{tabular}[t]{c|c}
\hline
\#segments & DAVIS ($J$) \\
\hline
$c=5$ & \textbf{57.8} \\
$c=6$ & 45.3 \\
$c=8$ & 41.0 \\
\hline
\end{tabular}
\label{tab:ab}
\end{minipage}
\captionof{figure}{Ablation study on different number of segments. Two examples with segmentation masks and segment flows are shown. The object region is split over multiple masks when $c$ becomes large. The over-segmentation decreases the performance of video object segmentation on DAVIS2016.}
\vspace{-5pt}
\label{fig:ab}
\end{figure}
\subsection{Ablation Study}
\label{sec:ablation}
The variable $c$, the number of segmentation channels, is an important hyper-parameter of our model.
We vary the value of $c$ ($5,6,8$) for pretraining the model, and examine its transfer performance for video object segmentation on DAVIS2016.
In Figure~\ref{fig:ab}, we visualize the model predictions under different number of segments. We observe that a large $c$ tends to lead to over-segmentation, and a small $c$ tends to lead to large regions.
The car and the swan is split into multiple regions even if the motion for separated regions are very close.
The model trained with $c=5$ segments a full object, while the model trained with $c=8$ separates the object into parts.
When pretraining the model with $c<4$, the training becomes unstable.
Quantitatively, the video object segmentation performance on DAVIS2016 decreases as we increase the number of segments.
\begin{table}[t]
\end{table}
\section{Introduction}
In recent years, contrastive learning~\cite{wu2018unsupervised,he2020momentum,chen2020simple} has become a powerful model for obtaining high-level representations on generic images~\cite{goyal2021self}.
Despite their encouraging performance, contrastive models have two critical limitations.
First, they heavily rely on hand-crafted image augmentations~\cite{dosovitskiy2015discriminative} to induce invariance, yet fail to account for complex variations such as object deformations and 3D viewpoints.
Second, they still require additional labeled data and a fine-tuning stage for downstream applications, preventing use in a standalone fashion.
In this paper, we seek a zero-shot image model for detecting and segmenting objects by learning from unlabeled videos with minimal
augmentation. As opposed to static images, a dynamic sequence of observations provides information about what is moving in the scene and how it moves. Such patterns not only reveal the boundary segment of an object,
but also indicate hierarchical part organizations and even object semantics. Thus, as an easily accessible source of unlabeled data, videos provide rich natural supervision for learning image representations.
A popular objective for self-supervised learning from videos is view synthesis. Concretely, given a source frame, a function is learned to warp the source frame to the target frame using photometric consistency as supervision. Dense optical flow~\cite{liu2020learning} can be self-supervised in this manner. One could also learn a monocular depth network from videos with additional camera parameters~\cite{godard2019digging}.
The key idea is to find an appropriate representation which not only parameterizes the warping function, but also transfers to the target task. For example, the multi-plane image representation~\cite{zhou2018stereo} is proposed to extrapolate between stereo pairs using view synthesis.
\figTeaser{tp}
Unlike prior works that use view synthesis for low-level vision tasks, our goal is to tackle object segmentation which involves mid-level and high-level visual recognition.
To this end, dense optical flow field which represents the low-level correspondence in a local manner would not suffice (see Figure~\ref{fig:teaser}).
We therefore seek a new representation which could capture a gross characterization of moving regions for the entire scene.
Deriving a representation explicitly for moving regions would allow the model to localize and segment objects.
Our approach decomposes view synthesis into two visual pathways: an appearance pathway to model ``what is moving'' by segmenting a static RGB image into separate regions, and a motion pathway to model ``how it moves'' by extracting motion features on a pair of images.
The motion features are then used to predict flow offsets for individual regions assuming common fate~\cite{wertheimer1923untersuchungen} for all pixels within a region.
The segment masks as well as their flow vectors jointly reconstruct a new representation called {\em segment flow}, which is used for view synthesis.
In this way, object appearance and motion are decoupled, such that the appearance model for predicting segmentation would benefit from rich RGB signals. By conditioning on a region, the motion pathway is also tasked to solve a much simpler problem than dense flow.
The two pathways are jointly trained with a reconstruction loss.
After self-supervised pretraining, we find that generic objectness detection and object segmentation automatically emerge from the model. Our model has the versatility for a variety of applications. First, the appearance pathway can be directly applied to novel images for dominant object segmentation in a zero-shot fashion. Second, it can also be fine-tuned for semantic segmentation on a small labeled dataset. Finally, with unsupervised test-time adaptation, the overall model can be transferred to novel videos for moving object segmentation without labels. Experimentally, we demonstrate strong performance on all of these applications, showing considerable improvements against the baselines.
The contributions of this work can be summarized as follows: 1) the first truly end-to-end zero-shot object segmentation from unlabeled videos; 2) a conceptually novel segment flow representation which goes beyond traditional dense optical flow; 3) a versatile model that can be applied to various image and video segmentation tasks.
Our code is available at \href{https://github.com/rt219/The-Emergence-of-Objectness}{https://github.com/rt219/The-Emergence-of-Objectness}.
\section{Segmentation by Appearance-Motion Decomposition}
\begin{figure}[tp]
\centering
\imw{figs/videoSeg_v2.pdf}{1}
\caption{
We learn a single-image segmentation network and a dual-frame motion network with an unsupervised image reconstruction loss. We sample two frames, $i$ and $j$, from a video. Frame $i$ goes through the \textcolor{red}{segmentation} network and outputs a set of masks, whereas frames $i$ and $j$ go through the \textcolor{cyan}{motion} network and output a feature map. The feature is pooled per mask and a flow is predicted. All the segments and their flows are combined into a segment flow representation from frame $i\to j$, which are used to \textcolor{cyan}{warp} frame $i$ into $j$, and \textcolor{orange}{compared} against frame $j$ to train the two networks.
}
\vspace{-5pt}
\label{fig:model}
\end{figure}
The goal of this work is to learn a zero-shot model to detect and segment objects by merely exposing it to unlabeled videos.
We are only interested in detecting the objectness in this paper instead of further categorizing the objects into specific classes.
We assume that a single moving object appears in one video. When multiple moving objects occur, the model needs to group these objects as one.
We take a learning-based approach to this problem. During training, we are given a collection of unlabeled videos for self-supervised learning.
The pretrained model should be directly applicable for inference on a novel image or a video to produce the object segmentation masks.
The overall pipeline for training our model is illustrated in Figure~\ref{fig:model}.
Our approach appearance-motion decomposition (AMD) takes a pair of RGB frames $X_i$ and $X_j$ sampled from a video for learning.
The model consists of an appearance pathway $f_A(X_i)$ and a motion pathway $f_M(X_i,X_j)$.
The two pathways jointly construct a segment flow representation $F$, which is used to warp frame $X_i$ into $X_j$. The overall model is self-supervised by the reconstruction objective on the frame $X_j$.
In the following, we describe the details for each module in our model.
\subsection{Appearance Pathway for Segmentation}
The appearance pathway is a fully convolutional neural network for segmenting a static RGB image into separate regions.
Formally, given the image $X_i \in \mathbb{R}^{3\times h \times w}$, it is segmented into $c$ regions by
\begin{equation}
S = f_A(X_i) \in \mathbb{R}^{c\times h \times w}.
\end{equation}
In practice, the mask $S$ is a soft probability distribution normalized across $c$ channels.
$c$ is an important hyper-parameter of our approach. A large $c$ may lead to over-segmentation, and a small $c$ may not locate the object. Empirically, we use a default value of $c=5$, and this is examined later in an ablation study.
We note that our segmentation network is designed to operate on static images and thus the network can be transferred to downstream image-based vision tasks.
In Section~\ref{sec:saliency}, we demonstrate that the pretrained segmentation network can be used to detect salient objects in a zero-shot fashion. Fine-tuning the appearance pathway on a labeled dataset is examined in Section~\ref{sec:semantic}.
\subsection{Motion Pathway for Correspondence}
The purpose of the motion pathway is to extract pixel-wise motion features between a pair of images in order to predict the region flow vector detailed in the next subsection.
We follow the network architecture of PWC-Net~\cite{sun2018pwc} for predicting dense optical flow, where the feature for each pixel describes the perceptual similarity to its spatial neighbors in the other frame.
Formally, given input frames $X_i$ and $X_j$, the network extracts features $V$ by
\begin{align}
V = f_M(X_i, X_j) \in \mathbb{R}^{d_v \times h \times w}
\label{eq:fm}
\end{align}
where $d_v$ is the dimension of motion features.
\subsection{Segment Flow Representation}
Given the decoupled appearance pathway and motion pathway, the segment and its motion can be binded for view synthesis.
Concretely, we pool the pixel-wise motion features within each segmentation mask to obtain the mask motion feature as a single vector,
\begin{align}
V_m = \frac{\sum(V \odot S_m)}{\sum S_m} \in \mathbb{R}^{d_v}, \quad m=1,...,c
\label{eq:pool}
\end{align}
where the summation operation is taken across the spatial coordinates, and $m$ is used to index the segmentation masks. The optical flow vector for each segmentation mask is read out from the motion feature by
\begin{align}
F_m = g(V_m) \in \mathbb{R}^2, \quad m=1,...,c
\label{eq:flow}
\end{align}
where the head network $g(\cdot)$ is chosen as a two-layered multilayer perceptron (MLP).
So far, we decompose a pair of images $X_i \to X_j$ into a set of segmentation masks $S_m$ and their associated flow vectors $F_m$.
This decomposition is based on the assumption that pixels within a mask share the same motion, a condition that simplifies optical flow estimation.
This assumption may not hold for articulated objects and inhomogeneous motion.
However, it becomes less problematic when all views in a video are taken for optimization, with
the appearance pathway able to aggregate a smoothly moving region into a meaningful segment.
We reconstruct a novel flow representation for the full image by composing the layers of segments with their motion vectors,
\begin{align}
F = \sum_{m} F_m \odot S_m, \quad m=1,...,c,
\label{eq:G}
\end{align}
where $\odot$ denotes the outer product. Since the flow representation $F$ is segment-based, we refer to it as {\em segment flow}.
This decoupled representation allows each component to cross-supervise each other.
Given an optical flow offset, the segmentation network could be supervised to find pixels that share this offset.
Given a segmentation mask, the correspondence network could be supervised to find the flow offset for this mask.
This approach for supervising object segmentation using motion information is fundamentally different from motion segmentation methods.
Our segmentation mask is predicted from a static appearance model that does not require dense and accurate flow for supervision. It utilizes flow at the region level, which can be approximated from sparse and noisy pixel-level estimates.
\subsection{Reconstruction Objective}
With the segment flow offset map, we are able to warp frame $X_i$ to $X_j$ by
\begin{align}
\hat{X_j}(p)=X_i(p+F(p)),
\end{align}
where $p$ is a spatial location index. The ground-truth frame $X_j$ provides supervision for reconstructed frame $\hat{X_j}$ through the following objective,
\begin{align}
\mathcal{L}=D(X_j, \hat{X_j}),
\label{eq:loss}
\end{align}
where $D$ is a metric defining distance between two images. Among the numerous choices for $D$, such as photometric losses~\cite{wang2004image}, deep-feature-based losses~\cite{johnson2016perceptual}\cite{zhang2018unreasonable}, and contrastive losses~\cite{park2020contrastive}, we adopt the pixel-wise photometric loss of SSIM~\cite{wang2004image} in this work for simplicity.
\subsection{Object Segment Selection}
Since our model outputs $c$ masks for each image, the mask corresponding to the object instead of the background needs to be determined.
We have empirically observed that the primary moving objects all appear in a particular mask channel across the training videos. This channel can be heuristically identified as the one whose segmentation mask has the maximum averaged segment motion.
The object segment from this mask layer is used for evaluating zero-shot downstream tasks.
\section{Related Works}
\textbf{Video object segmentation. }
Segmentation of moving objects requires finding correspondences along the time dimension. A dominant line of work focuses on learning a representation for temporally propagating segmentation masks. Such a representation may be learned with pixel-level object masks in videos with long-term relations~\cite{oh2019video,zhang2020transductive}, or learned through self-supervision such as colorization~\cite{vondrick2018tracking} and cycle-consistency~\cite{jabri2020space}. Given the annotation of object masks in the initial frame, the model tracks the object and propagates the segmentation through the remaining frames.
Fully unsupervised video object segmentation, without initial frame annotations, has received relatively little attention. NLC~\cite{faktor2014video} and ARP~\cite{koh2017primary} take a temporal clustering approach to this problem. Though they do not require segmentation annotations, elements of these algorithms depend on edge and saliency labels, and thus are not completely unsupervised.
FTS~\cite{papazoglou2013fast} calculates the segmentation by obtaining a motion boundary from the optical flow map between frames. SAGE~\cite{wang2015saliency} takes into account multiple cues of edges, motion segmentation, and image saliency for video object segmentation.
Contextual information separation~\cite{yang2019unsupervised} segments moving objects by exploiting the motion independence between the foreground and the background. A concurrent work based on motion grouping~\cite{yang2021self} clusters pixels with similar motion vectors. Both of these works rely on an off-the-shelf optical flow representation, which may be trained with~\cite{sun2018pwc,teed2020raft} or without~\cite{liu2020learning} supervision.
\textbf{Motion Segmentation.}
Classical methods for motion segmentation~\cite{shi1998motion,kumar2008learning,sun2012layered} cluster distinctive motion regions from the background based on two-frame optical flow.
Supervised learning approaches~\cite{tokmakov2017learning,tokmakov2019learning} map the optical flow field to segmentation masks.
The requirement of dense and accurate optical flow may be problematic when the flow vectors are not smooth over time and vulnerable to articulated objects with inhomogeneous motion~\cite{ochs2013segmentation}.
we turn our attention to modeling appearance on RGB representations, which provide rich cues (e.g. texture, color and edges) for perceptual organization, alleviating the need for dense pixel correspondence.
\textbf{Motion Trajectory Segmentation.}
Moving object segmentation has been shown to be effective when motion is considered over a large time interval~\cite{ochs2013segmentation}. An approach based on trajectory clustering~\cite{keuper2015motion} builds point trajectories over hundreds of frames, extracts descriptors for the point trajectories, and clusters them to obtain segmentation results. Though promising, such a global approach is computationally demanding.
\textbf{Layered representations.}
A simple linear model~\cite{wang1993layered,wang1994representing} can factorize a video into layers of foreground objects and background, assuming independence among the objects and background. This layered representation was used to derive better optical flow estimates~\cite{sun2013fully,sun2012layered,sun2010layered} and also for view-interpolation and time retargeting applications~\cite{brostow1999motion,alayrac2019visual,zitnick2004high,lu2020layered}.
Different from prior works, our work demonstrates the emergence of objectness through such layered representations.
\textbf{Unsupervised learning for segmentation.}
Human annotation of pixel-level segmentation is not only time-consuming, but also often inaccurate along object boundaries. Learning segmentation without labels is thus of great interest in practice. Segsort~\cite{hwang2019segsort} predicts segmentation by learning to group super-pixels of similar appearance and context from static images. Later work~\cite{van2021unsupervised} contrasts holistic mask proposals obtained from traditional bottom-up grouping.
A related line of work focuses on learning part segmentation from images and videos of the same object category, such as humans and faces. SCOPS~\cite{hung2019scops} is a representative method learned in a self-supervised fashion. The general idea follows unsupervised landmark detection~\cite{jakab2018unsupervised}, where geometric invariance, representation equivariance and perceptual reconstructions are considered. Co-part segmentation~\cite{xu2019unsupervised} is also explored in videos, where motion provides a strong cue for part organization. A motion-supervised approach~\cite{siarohin2020motion} models part motion between adjacent frames via affine parameters.
Another work~\cite{sabour2020unsupervised} implements a similar idea in capsule networks.
Our work differs significantly in studying learning from generic videos instead of from a single visual category.
\textbf{Image representation learning using motions.}
Motion contains rich cues about object location, shape, and part hierarchy. Motion segmentation has been used as a self-supervision signal for learning image-level object representations~\cite{pathak2017learning}. Motion propagation~\cite{zhan2019self} predicts a dense optical flow field from sparse optical flow vectors, conditioned on an RGB image. Our work also produces an image representation from unlabeled videos. Unlike prior works, our image representation is a by-product of our full framework for video understanding.
\section{Supplementary}
In this supplementary, we provide the network architecture details in Section A.1. In Section A.2, we present more qualitative video object segmentation results on 3 datasets, DAVIS 2016~\cite{Perazzi2016}, FBMS59~\cite{ochs2013segmentation} and SegTrackv2~\cite{li2013video}. We also present the per class quantitative results on these datasets.
\subsection{Network Details}
The network details are shown in Table \ref{tab:detailseg} and Table \ref{tab:detailpwc}. Table \ref{tab:detailseg} shows the detailed network architecture for the segment prediction head of our segmentation network. Our correspondence network adopt the similar framework as PWCNet~\cite{sun2018pwc} which contains a feature extractor, a flow estimator and a context network. The feature extractor is the same as that of PWCNet while we don't use the context network in our correspondence network. Table \ref{tab:detailpwc} shows the detailed layers of the flow estimator.
\subsection{More Results}
More video object segmentation results are shown in Figure~\ref{fig:suppsegt}, Figure~\ref{fig:suppdavis} and Figure~\ref{fig:suppfbms} for SegTrackv2, DAVIS 2016 and FBMS59 correspondingly. We choose those samples from different videos as many as possible.
In Figure~\ref{fig:suppsaliency}, more saliency detection results from DUTS~\cite{wang2017learning} dataset are represented.
\subsection{Broader Impact}
We proposed a self-supervised pretraining method for zero-shot object segmentation. The central idea of decomposing appearance and motion can be implemented with other network architectures, and even training losses.
However, we have not studied the implications of these variations of the approach.
There will also be unpredictable failures, where the generalization of the self-supervised framework still needs deeper understanding.
This method is data-driven thus the data bias problem should be careful during data collection in both pretraining and downstream tasks. As this method can be applied to a wide range of videos without annotation, privacy should be also careful during the data utilization.
\vspace{1cm}
\begin{table}[H]
\centering
\caption{Details about the prediction head in our segmentation network. Our segmentation network consists of a backbone, ResNet50, and a prediction head which predicts the segments through the features from the backbone. Here $c$ is a hyperparameter which represents the segment number. }
\vspace{4pt}
\begin{tabular}{|c|c|}
\hline
Layer & Output size \\
\hline
Input Feature & $2048 \times 48 \times 48$ \\
Conv($3 \times 3, 2048 \rightarrow 256$) + BN + ReLU & $256\times 48 \times 48$ \\
Conv($3 \times 3, 256 \rightarrow 256$) + BN + ReLU & $256 \times 48 \times 48$ \\
Conv($3 \times 3, 256 \rightarrow c$) & $c\times 48 \times 48$ \\
\hline
\end{tabular}
\label{tab:detailseg}
\end{table}
\begin{table}[t]
\centering
\caption{Architecture details about our correspondence network. As it processes the input at different pyramid levels, here $H$ and $W$ represents the size of input in a certain level. And $c$ is a hyperparameter about the segment number. }
\vspace{4pt}
\begin{tabular}{|c|c|c|}
\hline
Index & Layer & Output size \\
\hline
1. & Input Feature & $ 115 \times H \times W $ \\
2. & Conv($3 \times 3, 115 \rightarrow 128$) + ReLU & $ 128 \times H \times W $ \\
3. & Conv($3 \times 3, 128 \rightarrow 128$) + ReLU & $ 128 \times H \times W $ \\
4. & Concatenate 2. and 3. & $ 256 \times H \times W $ \\
5. & Conv($3 \times 3, 256 \rightarrow 96$) + ReLU & $ 96 \times H \times W $ \\
6. & Concatenate 3. and 5. & $ 224 \times H \times W $ \\
7. & Conv($3 \times 3, 224 \rightarrow 64$) + ReLU & $ 64 \times H \times W $ \\
8. & Concatenate 5. and 7. & $ 160 \times H \times W $ \\
9. & Conv($3 \times 3, 160 \rightarrow 32$) + ReLU & $ 32 \times H \times W $ \\
10. & Concatenate 7. and 9. & $ 96 \times H \times W $ \\
11. & Average Pooling & $ 96 \times c $ \\
12. & FC ($96 \rightarrow 2$) & $ 2 \times c $ \\
\hline
\end{tabular}
\label{tab:detailpwc}
\end{table}
\def\rim#1#2#3#4#5#6#7#8{
\imw{figs_supp/#7/#8/#1.jpg}{0.152}&
\hspace{0.01cm}
\imw{figs_supp/#7/#8/#2.jpg}{0.152}&
\hspace{0.01cm}
\imw{figs_supp/#7/#8/#3.jpg}{0.152}&
\hspace{0.01cm}
\imw{figs_supp/#7/#8/#4.jpg}{0.152}&
\hspace{0.01cm}
\imw{figs_supp/#7/#8/#5.jpg}{0.152}&
\hspace{0.01cm}
\imw{figs_supp/#7/#8/#6.jpg}{0.152}\\
}
\begin{figure}[t]
\centering
\subfloat{\small
\tb{@{}ccccccc@{}}{0.03}{
\rotatebox{90}{\hspace{0.5cm}image} & \rim{1}{4}{7}{10}{13}{16}{supp_segt}{im}
\rotatebox{90}{\hspace{0.35cm}segment} & \rim{2}{5}{8}{11}{14}{17}{supp_segt}{mask}
\rotatebox{90}{\makecell[c]{segment\hspace{0.15cm}flow}} & \rim{0}{3}{6}{9}{12}{15}{supp_segt}{f1}
}
}
\caption{Qualitative results of SegTrackv2}
\label{fig:suppsegt}
\end{figure}
\def\rim#1#2#3#4#5#6#7{
\imw{figs_supp/#7/im/#1_iter0000100_im.jpg}{0.152}&
\hspace{0.01cm}
\imw{figs_supp/#7/im/#2_iter0000100_im.jpg}{0.152}&
\hspace{0.01cm}
\imw{figs_supp/#7/im/#3_iter0000100_im.jpg}{0.152}&
\hspace{0.01cm}
\imw{figs_supp/#7/im/#4_iter0000100_im.jpg}{0.152}&
\hspace{0.01cm}
\imw{figs_supp/#7/im/#5_iter0000100_im.jpg}{0.152}&
\hspace{0.01cm}
\imw{figs_supp/#7/im/#6_iter0000100_im.jpg}{0.152}\\
}
\def\rfl#1#2#3#4#5#6#7{
\imw{figs_supp/#7/f1/#1_iter0000100_f1.jpg}{0.152}&
\hspace{0.01cm}
\imw{figs_supp/#7/f1/#2_iter0000100_f1.jpg}{0.152}&
\hspace{0.01cm}
\imw{figs_supp/#7/f1/#3_iter0000100_f1.jpg}{0.152}&
\hspace{0.01cm}
\imw{figs_supp/#7/f1/#4_iter0000100_f1.jpg}{0.152}&
\hspace{0.01cm}
\imw{figs_supp/#7/f1/#5_iter0000100_f1.jpg}{0.152}&
\hspace{0.01cm}
\imw{figs_supp/#7/f1/#6_iter0000100_f1.jpg}{0.152}\\
}
\def\rmask#1#2#3#4#5#6#7{
\imw{figs_supp/#7/mask/#1_iter0000100_im_msk.jpg}{0.152}&
\hspace{0.01cm}
\imw{figs_supp/#7/mask/#2_iter0000100_im_msk.jpg}{0.152}&
\hspace{0.01cm}
\imw{figs_supp/#7/mask/#3_iter0000100_im_msk.jpg}{0.152}&
\hspace{0.01cm}
\imw{figs_supp/#7/mask/#4_iter0000100_im_msk.jpg}{0.152}&
\hspace{0.01cm}
\imw{figs_supp/#7/mask/#5_iter0000100_im_msk.jpg}{0.152}&
\hspace{0.01cm}
\imw{figs_supp/#7/mask/#6_iter0000100_im_msk.jpg}{0.152}\\
}
\begin{figure}[t]
\centering
\subfloat{\small
\tb{@{}ccccccc@{}}{0.03}{
\rotatebox{90}{\hspace{0.2cm}image} & \rim{breakdance_00065}{car-shadow_00003}{dance-twirl_00000}{horsejump-high_00013}{kite-surf_00012}{motocross-jump_00000}{supp_davis}
\rotatebox{90}{\hspace{0.1cm}segment} & \rmask{breakdance_00065}{car-shadow_00003}{dance-twirl_00000}{horsejump-high_00013}{kite-surf_00012}{motocross-jump_00000}{supp_davis}
\rotatebox{90}{\hspace{0.0cm}seg. flow} & \rfl{breakdance_00065}{car-shadow_00003}{dance-twirl_00000}{horsejump-high_00013}{kite-surf_00012}{motocross-jump_00000}{supp_davis}
}
}
\caption{Qualitative results of DAVIS 2016}
\label{fig:suppdavis}
\end{figure}
\def\rim#1#2#3#4#5#6#7#8{
\imw{figs_supp/#7/#8/#1.jpg}{0.152}&
\hspace{0.01cm}
\imw{figs_supp/#7/#8/#2.jpg}{0.152}&
\hspace{0.01cm}
\imw{figs_supp/#7/#8/#3.jpg}{0.152}&
\hspace{0.01cm}
\imw{figs_supp/#7/#8/#4.jpg}{0.152}&
\hspace{0.01cm}
\imw{figs_supp/#7/#8/#5.jpg}{0.152}&
\hspace{0.01cm}
\imw{figs_supp/#7/#8/#6.jpg}{0.152}\\
}
\begin{figure}[t]
\centering
\subfloat{\small
\tb{@{}ccccccc@{}}{0.03}{
\rotatebox{90}{\hspace{0.5cm}image} & \rim{1}{4}{7}{10}{13}{16}{supp_fbms}{im}
\rotatebox{90}{\hspace{0.35cm}segment} & \rim{2}{5}{8}{11}{14}{17}{supp_fbms}{mask}
\vspace{0.3cm}
\rotatebox{90}{\makecell[c]{segment\hspace{0.15cm}flow}} & \rim{0}{3}{6}{9}{12}{15}{supp_fbms}{f1}
\rotatebox{90}{\hspace{0.5cm}image} & \rim{19}{22}{25}{28}{31}{34}{supp_fbms}{im}
\rotatebox{90}{\hspace{0.35cm}segment} & \rim{20}{23}{26}{29}{32}{35}{supp_fbms}{mask}
\rotatebox{90}{\makecell[c]{segment\hspace{0.15cm}flow}} & \rim{18}{21}{24}{27}{30}{33}{supp_fbms}{f1}
}
}
\caption{Qualitative results of FBMS59}
\label{fig:suppfbms}
\end{figure}
\def\prow#1#2#3#4{
\imw{figs_supp/supp_saliency/#1.png}{0.25}&
\hspace{0.01cm}
\imw{figs_supp/supp_saliency/#2.png}{0.25}&
\hspace{0.01cm}
\imw{figs_supp/supp_saliency/#3.png}{0.25}&
\hspace{0.01cm}
\imw{figs_supp/supp_saliency/#4.png}{0.25}\\
}
\begin{figure}[t]
\centering
\subfloat{\small
\tb{@{}cccc@{}}{0.05}{
\prow{0}{1}{22}{16}
\prow{0014}{27}{21}{13}
\prow{6}{7}{8}{2}
\prow{4}{19}{11}{20}
\prow{5}{9}{10}{15}
\prow{17}{26}{24}{25}
}
}
\caption{Qualitative salient object detection results. Our model can detect multiple primary objects and even static object like the chair and the rocks. }
\label{fig:suppsaliency}
\end{figure}
|
\section{Introduction}
From 2019 April 29 to May 4, the \ac{BL} project performed observations to place limits on the prevalence of radio technosignatures (non-human ``objects, substances, and/or patterns whose origins specifically require a [technological] agent'', by analogy with biosignatures \cite{des2008nasa}) in the direction of \ac{ProxCen}. \ac{ProxCen} is an astrobiologically fascinating target due to its proximity: it is the closest star to the Sun at 1.295\,pc \cite{van2007validation}; it is host to Proxima b, the closest-known exoplanet to the Earth \cite{anglada-escude2016terrestrial, mascareno_2020}, which lies in the traditional habitable zone of \ac{ProxCen}; and it has even featured as the target of a proposed in-situ search via Breakthrough Starshot \cite{worden2018philanthropic}.
We employed the CSIRO Parkes ``Murriyang'' telescope with the \ac{UWL} \cite{hobbs2020UWL}, across $0.704$--$4.032$\,GHz, as part of the project ``P1018: Wide-band radio monitoring of space weather on Proxima Centauri'' (Primary Investigator A.Z,). These observations were part of an international multi-wavelength campaign to monitor \ac{ProxCen} for stellar flares \cite{zic2020flare}, collaboratively carried out using the \ac{BL} Parkes Data Recorder backend in shared-risk mode \cite{price2018blpdr, price2021blpdr}. In parallel with the P1018 flare search, the \ac{BL} team used the data to conduct a technosignature search of \ac{ProxCen}. Over these six days, \ac{ProxCen} was observed for a total of 26\,h 9\,min; data are available at \url{seti.berkeley.edu/blc1}.
We searched for narrowband drifting signals in high-resolution dynamic spectra \cite{lebofsky2019} with $\sim 17$\,s subintegrations and $\sim 4$\,Hz frequency resolution. We ran the \texttt{turboSETI} (version 1.2.2) narrowband search algorithm---as described in a companion paper \cite{smith2020inprep}, and the pipeline returned a narrowband event that appeared in two consecutive observations of \ac{ProxCen} and did not appear in any reference ``off-source'' observations toward astronomical calibrator sources between the ``on-source'' \ac{ProxCen} observations, even upon visual inspection. This single event was named ``blc1'' as a shorthand for ``Breakthrough Listen candidate 1''. We find ``signal-of-interest'' to be a more appropriate categorization than ``candidate'' (see, for example, \cite{forgan2019rio} and Supplementary Discussion 1.1) but, as ``blc1'' is already in common usage, we will continue to use it here. blc1 is shown in Figure \ref{fig:1} and its properties are enumerated in Table \ref{tab:general_properties}.
\begin{table}[ht]
\centering
\begin{tabular}{|c|c|}
\hline
\textbf{Parameter} & \textbf{Value} \\
\hline
Detection Date & 29 April 2019 \\
Time at First Detection & 13:17:35.232 UTC \\
Time at Last Detection & 18:19:26.400 UTC \\
Length of Persistence & 5.03 hours \\
Signal Frequency at First Detection & 982.0024 MHz\\
Signal Frequency at Last Detection & 982.0028 MHz\\
Initial Drift Rate & 0.0326 Hz/s\\
Average Signal-to-Noise & 17.956\\
Signal Bandwidth & $<$3.81 Hz\\
\hline
\end{tabular}
\caption{Basic characteristics of blc1. Average signal-to-noise is calculated as the average of the signal-to-noise ratios from the five 30-minute observations in which blc1 appeared.}
\label{tab:general_properties}
\end{table}
\begin{figure}
\centering
\includegraphics[height=0.8\textheight]{blc1_cadence.pdf}
\caption{A waterfall plot (dynamic spectrum) around 982\,MHz depicting alternating observations of \ac{ProxCen} and off-source positions from 2019 April 29. The horizontal axis shows the relative frequency offset from a start frequency, the y-axis shows time progressing from top to bottom (and from left to right, continuing from the end of the first column to the beginning of the second), and the colorbar shows detected power (linearly normalized). blc1 is visible as the diagonal linear feature in ProxCen panels 1, 2, 3, 4, and 7. Data was not taken during purple panels, which are primarily telescope slews.}
\label{fig:1}
\end{figure}
blc1 is intriguing because:
\begin{enumerate}
\item It is a $\sim$Hz-wide narrowband signal, which cannot be created by any known or foreseeable astrophysical system, only by technology \cite{tarter2001review}.
\item It exhibits a non-zero drift rate, as expected for a transmitter that was not on the surface of the Earth.
\item Its drift rate appears approximately linear in each 30 minute ``panel'' (one single-target observation from a cadence of observations that makes up a waterfall plot such as Figure \ref{fig:1}), but the drift rate changes smoothly over time, as expected for a transmitter in a rotational/orbital environment.
\item It is absent in the off-source observations (see Section \ref{sec:initial}), as expected for a signal that is localized on the sky.
\item It persists over several hours, making it unlike other interferers from artificial satellites or aircraft that we have observed before.
\end{enumerate}
For these reasons, especially Point 4 (by which we have ruled out \texttt{turboSETI} outputs in prior searches, e.g., \cite{price2020breakthrough}) blc1 warranted an in-depth analysis beyond any event so far in the course of the \ac{BL} initiative.
\section{Initial investigation and parametrization of blc1}
\label{sec:initial}
We first ensured that the telescope and backend logs showed normal operation, mapped out the telescope pointings against the local site, and checked that data recording was functioning correctly. Upon consultation with the observatory and the Australian Communications and Media Authority (ACMA), there is no catalogued RFI at the Parkes Observatory at 982.002 MHz, nor registered transmitters at that frequency in Australia. The band near 982\,MHz more broadly is reserved for aircraft.
A signal that is truly localized on the sky should only be detected when the telescope is pointed at that location, if the off-source positions are suitably chosen with regard to the sidelobe pattern of the telescope. In order to ensure that the behaviour seen in the off-source observations is consistent with a localized emitter in the on-source, we must consider the cadence, off-source positions, and off-source luminosities, and perform a thorough search for highly-attenuated power on the off-source observations.
The \ac{ProxCen} campaign employed a modified observing strategy that was different than the usual \ac{BL} procedures (\cite{enriquez2017turbo, price2020breakthrough}) in order to fulfill the project's primary goal: to observe and characterize stellar flares from \ac{ProxCen}. The cadence was non-standard, consisting of 30\,min integrations on \ac{ProxCen} that alternated with 5\,min observations of the calibrators. This cadence was chosen to maximize the on-source time. However, the asymmetry in the observing lengths (30\,min vs.\ 5\, min) causes an associated asymmetry in the expected \ac{S/N} from the same signal in a \texttt{turboSETI} analysis by a factor of $\sqrt{\frac{30}{5}}$.
In addition, the off-source positions were two quasars (and a single pulsar, observed once) selected from the ATNF calibrator database. Off-sources for \texttt{turboSETI} are usually chosen to be positions, usually stars, without a detectable radio flux, for consistency when evaluating apparent \ac{S/N}s between on- and off-sources. As these observations were being used for flare detection, the off-sources were chosen as quasars for frequent evaluation of the system temperature. The quasars, given their intrinsic radio emission, have a higher ``noise floor'' that we must contend with when searching for narrowband emission. From \cite{wright1990parkes}, we find a flux of $S(\nu) = 10$\,Jy at a frequency $\nu$ of 982\,MHz for both PKS\,1421-490 or PKS\,1934-638. Given our SEFD of $\sim$38 Jy from above, this leads to a total SEFD of 48\,Jy when observing on a calibrator.
Finally, the off-source positions were significantly further from \ac{ProxCen} than in a normal \ac{BL} observation. PKS\,1421-490 is $12\fdg57$ from \ac{ProxCen} and PKS\,1934-638 is $45\fdg82$. These distances led to the long slew times shown in Figure \ref{fig:1}, which can cause the temporal \ac{RFI} environment to change between the on and off-source observations. There may also be a large spatial difference in directional \ac{RFI}, especially when complicated by the telescope sidelobes. Similarly, the position of \ac{ProxCen} changed in sky coordinates by $~30^{\circ}$, primarily in azimuth, from first detection to final detection.
In this case, the first two effects would downweight the expected \ac{S/N} of blc1 in the off-source observations by a factor of $\sqrt{\frac{30}{5}} \times \frac{48}{38} \approx 3$, which should still be detectable in a de-drifted sum if the signal were present in the off-source. We took all nine off-source panels from Figure \ref{fig:1} and individually examined their waterfall plots, confirming that there was no visible signal in the first eight. The ninth and final off-source shows some power near 982.0028 MHz.
Highly-attenuated signals in an off-source observation may only become apparent after removing the expected drift rate. To this end, we de-drifted each off-source panel to the nearest best-fit drift rates (see Supplementary Methods 2.1). The final off-source \textit{does} show a faint narrowband signal (S/N $<5$), but the best-fit drift rate for this signal is 0 Hz/s, not 0.14 Hz/s. While this could be interpreted as blc1 appearing in an off-source, we deemed it inconclusive given a) the lack of appearance in the previous off-source observations, b) the known presence of a 0 Hz/s interference comb in the spectrum (see Section \ref{ssec:comb}), c) the mismatch of the expected and measured drift rate, and d) the low \ac{S/N}.
Finally, in order to search for constant, highly-attenuated power across the off-sources, the panels can be individually summed across their 5 and 30 minute lengths to create integrated spectra, and those spectra can in turn be summed to produce an incoherent sum across every off-source observation. We saw no resulting signal (\ac{S/N} $ \gtrsim 5$) in the de-drifted spectra of any non-detection, nor any signal in the incoherent sum of the nine off-sources. This indicates that either a) the phenomenon was localized on the sky, either near \ac{ProxCen}'s location or near a source in the sidelobes of the telescope; or b) the phenomenon had a duty-cycle or brightness variation that was matching our chosen cadence particularly well.
\subsection{Frequency Comb}
\label{ssec:comb}
Using an auto-correlation function in frequency, we find evidence for an \ac{RFI} frequency comb---a set of non-drifting signal regularly-spaced in frequency with stable amplitude---in the blc1 observation (Supplementary Figure 16).
The comb is present in the on-sources and the off-sources, has a spacing of $\sim$80.1 Hz and is present throughout the entire 128 MHz sub-band, from 960--1087 MHz. This implies that the comb is not localized to the part of the spectrum around blc1, lowering the likelihood that they are related phenomena. To investigate this further, we searched for reappearances of the frequency comb in other \ac{UWL} observations (using the 7000 observation dataset described in Section \ref{ssec:archival}) and found that the comb appeared occasionally, throughout the year, without any associated signals at 982\,MHz. We do not detect the frequency comb in any of the November 2020 or April 2021 re-observations (Section \ref{ssec:reobs}.
We deem the comb to be independent of blc1 based on a) observing the comb over the week of \ac{ProxCen} observations and in unrelated campaigns throughout the year in both on and off-sources (indicative of \ac{RFI}) without an appearance of blc1; b) the lack of drift in the frequency comb; and c) a lack of correlation between the \ac{S/N} over time of the comb and the \ac{S/N} over time of the signal. However, the \ac{RFI} comb unfortunately complicates the measurement of parameters, especially the \ac{S/N}, in observations of similar signals to blc1 with heavy interference.
\subsection{Re-observations}
\label{ssec:reobs}
Around mid-November 2020, we began to re-observe \ac{ProxCen} during our scheduled observation time at Murriyang, for sessions when it was above the horizon. We re-observed \ac{ProxCen} on November 19, 26, and 30, for 3--4 hours on each occasion. These re-observations were performed with the same \ac{UWL} receiver, \ac{BL} backend, and outputted data product as the original observation of blc1. However, we chose a more standard cadence for the re-observations: 15 minutes on-source on \ac{ProxCen}, 15 minutes off-source, with the off-sources chosen to be nearby stars at different relative angles to the \ac{ProxCen}--ground vector. No signals of any drift rate were detected at $\nu_{\rm blc1} \pm 2$\,kHz in any of these three re-observations.
Two years after the original detection, from April 29--May 3 2021, we observed 12 hours each day to replicate the relevant portion of the initial \ac{ProxCen} session in 2019. These re-observations were performed with the same \ac{UWL} receiver, \ac{BL} backend, outputted data product, local time, and cadence as the original observation of blc1. No signals of any drift rate were detected at $\nu_{\rm blc1} \pm 2$\,kHz in any of these re-observations.
\section{Constraining physical and electronic drift rates}
Signals drifting in frequency are generated in one of two ways: (1) a transmitter accelerating relative to the observer and (2) an electronically varying (deliberately or not) transmitter. Both means may be present at any given time.
Strong ground-based \ac{RFI} entering into a distant sidelobe may remain detectable even when the telescope is pointed at different parts of the sky. The sidelobe may downweight the signal by $10^5$ (50\,dB) or more \cite{hobbs2020UWL}, but it may still be detected with higher \ac{S/N} than an intrinsically weak and/or distant signal coming through the main lobe with as little as one millionth of the flux density. For satellites, the main lobe and near-in sidelobes subtend a very small area so ``direct hits'' from satellites are rare but can saturate the receiver when they do occur.
\subsection{Accelerational Drifts}
For this analysis, we assume that the signal from a transmitter near 982\,MHz is leaking in through either a distant antenna sidelobe or the main beam. We seek to explain both the drift morphology of the signal and the prevalence over a sufficient length of time.
With the frequency fixed, we looked at drift rate characteristics of transmitters undergoing various ``normal'' motions. The drift is proportional to the relative line-of-sight acceleration between the receiver and transmitter, which can be produced via a change of speed or relative direction. The drift due to just telescope motion as it tracks an object is $<10^{-4}\dot{\nu}$, where $\dot{\nu}$ is the signal drift of blc1, and hence does not explain blc1's characteristics.
\subsubsection{Ground-Based and Aerial Transmitters}
A transmitter on the ground may be stationary, or be on a moving platform in the vicinity of the telescope, such as a car, train or other vehicle. In the air, the transmitter could be on a plane, helicopter, drone, balloon or other airborne object. To investigate these hypothetical transmitters, we looked at a range of plausible routes around the observatory (say for someone walking or biking) and for a range of velocities for vehicles along the nearby highway, as shown in Supplementary Figure 1, for a fixed frequency transmitter. Similarly, we explored a range of plausible airplane velocities and routes between various cities.
Supplementary Figure 1 shows the primary issue with all of the local signals described above: that speeds fast enough to exhibit the right order of drift do not persist long enough; conversely, for slow enough speeds to persist, the drifts are too low. None of the drift characteristics from local signals match the measurement and it is extremely difficult to construct a continual motion path that could persist as exhibited by the measured signal, even by varying the speed along the route.
\subsubsection{Satellite Transmitters}
Next we considered artificial satellites orbiting the earth, which are broadly grouped as (1) low-earth orbit (LEO, $\sim$90 minute orbits), (2) medium-earth orbit (MEO, $\sim$12 hour orbits), and (3) geosynchronous orbit (GEO, $\sim$24 hour orbits). The short period of LEOs implies that they cannot be responsible for the signal regardless of their drift characteristics, so we did not include LEO satellites in our analysis. Conversely, we do include GEO satellites; GEO is often referred to as ``geostationary'', implying true zero drift, but even stable GEO satellites generally meander in a figure-8 motion. There are a small number of satellites in other orbits as well, which are encompassed by the following analysis.
To investigate these satellites, we downloaded orbital element parameters for all known, active, non-LEO satellites for the date in question (\url{https://www.space-track.org}). All satellites above the horizon are plotted as the faint lines in Figure \ref{fig:2}. Even for geosynchronous satellites (green lines), the drift characteristics vary too rapidly to explain the signal characteristics.
\begin{figure}
\centering
\includegraphics[width=\textwidth]{space_drift.pdf}
\caption{Drift characteristics of non-LEO satellites and deep space probes above the horizon during the event time. The thin green lines are GEO, the magenta are MEO, and the orange are other non-LEO satellites. The thick lines are deep space probes: Parker Solar Probe (blue), Gaia (orange), Juno (green), New Horizons (red), Ryugu (purple) and Voyager 2 (brown). The black data points show the blc1 data, where the x-axis bars are the time of observation and the y-axis bars are 1$\sigma$ error bars. The deep-space probes are closer to the correct order of blc1’s drift due to their unique status as near-sidereal sources, but none can explain blc1.}
\label{fig:2}
\end{figure}
\subsubsection{Transmitters on Deep-Space Probes}
Deep space probes are exploratory spacecraft that do not orbit the Earth, and are the most celestially-stationary sources produced by humans. Though their radio emission can be strong relative to celestial sources, their distance generally makes them weak enough that they are not readily detectable unless they happen to be in or near the main beam. To investigate space probes, we obtained positional information from NASA Horizons \url{https://ssd.jpl.nasa.gov/horizons.cgi}. Figure \ref{fig:2} shows the drift of space probes above the horizon, though none of them coincide with the main beam of the telescope. Their drifts and positions are inconsistent with blc1.
We also considered---and dismissed---transmitters on asteroids and reflections from Earth-bound radio transmitters off of asteroids in the primary beam as a potential source of blc1 (see Supplementary Methods 2.2).
\subsection{Electronic Drifts}
Finally, we considered electronically-varying signal generators. A modern signal generator can be programmed to produce any drift rate; there are effectively no constraints, and thus we cannot use this to inform our analysis. However, we can constrain other causes of electronic drift related to the temperature, aging and voltage of the oscillator---all oscillators exhibit some drift with these parameters.
This type of \ac{RFI} is prevalent, but typically the drifts are too low (very good regulation) or too high and wandering (as in more typical commodity devices) for them to be mistaken for an object moving sidereally. They are also almost always seen in both the off-source pointings as well as the on-source pointing. However, given their prevalence, occasionally a device could potentially exhibit the expected characteristics of a sidereal source, which could be difficult to ascribe to \ac{RFI}.
\subsection{Astronomically-Expected Drifts}
We can also compare the drift rate magnitude and morphology to the sidereal drift in the direction of \ac{ProxCen}, as well as known accelerations and orbital periods from the planets in the \ac{ProxCen} system. Defiance of expectations for drift rates in a target system does not invalidate a signal, but a match provides additional evidence in favour of an interstellar origin. In an effort to be easily detected and discernable as ``stationary'', a distant transmitter in the direction of the antenna pointing could electronically vary their transmission frequency to compensate for their motion relative to the Sun, the Earth, or the barycenter of the solar system. For the timescales relevant to our observation, the Earth's rotation is the primary contributor to the expected drift rate for \ac{ProxCen}, and is similar for all other sidereal sources in the beam (see Supplementary Discussion 1.2). Supplementary Figure 2 shows the relevant geometry of the Earth-Sun-ProxCen positions and the corresponding residual drift. We find that the drift rate of blc1 is not consistent with the barycentric motion expected from the direction of \ac{ProxCen}, but is consistent with the order-of-magnitude drift that could be produced in the system, based on the orbital and rotational motions of its planets (see Supplementary Discussion 1.3).
\section{Searching for other instances of blc1}
In parallel with the drift rate analysis, we performed a search for reappearances of the signal-of-interest on other days and at other frequencies.
\subsection{Other Murriyang signals near 982 MHz}
\label{ssec:archival}
We searched for signals with the same frequency and same drift as blc1 from both a) the week-long \ac{ProxCen} campaign and b) every archival observation from standard \ac{BL} Murriyang \ac{UWL} observing of other stellar targets.
In order to find all appearances of blc1, even those that were too faint or masked by \ac{RFI} to be flagged by \texttt{turboSETI}, we produced output plots for visual inspection from every \ac{ProxCen} on- and off-source observation from April 29 to May 4. We restricted the plots to $982.002 - 982.004$\,MHz to begin; a few plots were extended up or down by 1\,kHz if there appeared to be interesting behavior near the upper or lower bounds.
We created and searched two kinds of output plots: dynamic time-frequency spectra ``waterfall'' plots (see Figure \ref{fig:1}) and ``butterfly'' plots, which display the power at each drift rate-frequency pair after use of a de-drifting algorithm (see Supplementary Methods 2.1). Examples of these output plots are shown in Supplementary Figure 17.
Through analysis of these diagnostic plots, we identified four occurrences of blc1-like signals during the \ac{ProxCen} observations which, through their low \ac{S/N}s, had failed to reach \texttt{turboSETI}'s detection threshold. An example of one of these similar signals is shown in Figure \ref{fig:3}, while the rest are displayed in Supplementary Figures 3--5.
We performed the same analysis with every archival observation that \ac{BL} had taken with the Murriyang \ac{UWL} receiver. In total, this consisted of about 7000 observations from 2019--2020---primarily nearby stars in the Hipparcos catalog \cite{perryman1997hipparcos,isaacson17}, but also pulsars and quasars used for calibration, and, of course, \ac{ProxCen}. Most of these files had a standard duration of 5\,min, as opposed to the 30\,min observations from the \ac{ProxCen} campaign. In many of these observations, we observed the same zero-drift, frequency comb as detected in the original blc1 observation (see Section \ref{sec:initial}), giving more evidence that this comb is unrelated \ac{RFI}.
We found 15 similar-looking features in these non-\ac{ProxCen} \ac{UWL} observations. Upon visual inspection of the full cadences surrounding these features, 14 of those are different from the blc1-like signals in morphology, length, drift behaviour, and/or signal strength over time. However, one of the 15 features is clearly \ac{RFI} due to its persistence across both on and off-source observations, appeared 4 days \textit{prior} to blc1, and looks to potentially be a member of the set including blc1 and the similar signals from the ProxCen campaign; this signal is shown in Supplementary Figure 6.
\begin{figure}
\centering
\includegraphics[height=0.8\textheight]{SameCamp1.pdf}
\caption{A waterfall plot around 982\,MHz depicting a similar signal to blc1 that occurred on 2019 April 30. The signal is the bright sloping feature, but is clearly overlaid on a non-drifting, comb-like signal (see Section \ref{ssec:comb}). Both the signal and the comb underneath appear to be \ac{RFI} as they appear in the off-source observations. In Supplementary Table 1, we derived the \ac{S/N}, drift rate, and start frequency of each panel of the blc1 observation; we derived the properties of this signal in the same manner (see Supplementary Methods 2.1). The drift rate (median 0.021 Hz/s), \ac{S/N} (median 6.9), and frequency range (982.0021--982.0023 MHz), are consistent with blc1, and this signal is also unresolved.}
\label{fig:3}
\end{figure}
All five of these 982 MHz signals from different days are fainter than blc1; three of them conclusively appear in the off-sources, while two of them are inconclusive.
\subsection{Different Frequencies}
Human-made communication technologies often use multiple simultaneous frequency channels to send information for improved redundancy and bitrate. It is possible that \ac{ETI} would do the same; it is also possible that the appearance of a blc1 twin at another frequency, if clearly \ac{RFI}, would allow us to determine that blc1 is also \ac{RFI}.
To find similar signals to blc1 at different frequencies, we calculated the frequency-normalized drift rate ($\dot{\nu}_{\rm normalized} = \frac{\dot{\nu}}{\nu}$) that would indicate a signal which was drifting proportionally to blc1 in the first observation. This proportional drift is expected for multi-frequency transmitters in the same accelerational environment, but also in multi-frequency transmitters that are electronically-drifting.
We then searched the catalog of narrowband hits created with \texttt{turboSETI} for signals that a) appeared at the same time as blc1 and b) were drifting proportionally in the first panel, plus or minus the drift rate error proxy as given in the first row of Supplementary Table 1. This search returned 112 hits: blc1 itself, and 111 signals that \texttt{turboSETI} had identified as hits but then rejected as \ac{RFI} at an early stage in the pipeline due to their appearance in off-source observations. These hits were plotted in the context of all of the subsequent panels of the \ac{ProxCen} observation on 2019 April 29.
We visually inspected the 111 matches, looking for signals that had the same morphology as blc1 beyond the drift rate in the first panel. To identify a signal with the same morphology, we looked for monotonicity, a shallowing of the slope over time, a vertical length that spanned multiple panels, and the absence of complex additional features (e.g. sinusoidal behaviour). We found that 36 of the 111 \texttt{turboSETI} matches (32\%) were blc1 ``lookalikes'': signals with strikingly similar morphology to blc1. A subset of these lookalikes are shown in Supplementary Figure 7, while some examples of the ``non-lookalikes'' are shown in Supplementary Figure 8.
The lookalikes have a range of variabilities over time, which seem to indicate multiple transmitters producing unusually consistent drifts. We then performed the same search as before but with negative drift rates. We found 310 hits from \ac{RFI}, of which 27, upon visual inspection, were found to be \textit{mirrored} lookalikes: with exactly the same drift structure over time as blc1 but flipped in morphology across the frequency axis. A selection of these signals are shown in Supplementary Figure 9.
We can conclusively state that all lookalike and mirrored lookalike signals are \ac{RFI} due to their appearance in off-source observations.
\subsection{Characterizing the blc1 lookalike population}
The presence of this population of both positively- and negatively-sloped blc1 lookalikes preliminarily suggests that all lookalikes (including blc1) share a common origin. We can further assess this claim by examining the similarity of the parametric distributions of the lookalike signals and blc1. We find that blc1 is consistent with the lookalike population in absolute drift rate, frequency, and \ac{S/N} (Figure \ref{fig:4}).
\begin{figure}
\centering
\includegraphics[width=\textwidth]{lookalike_corner_plot.pdf}
\caption{A corner plot showing the population of lookalikes and mirrored lookalikes compared to blc1 in start frequency, normalized drift rate, and signal-to-noise ratio. For the mirrored lookalikes, which have negative normalized drift rates, we took the absolute value for comparison. The height of the kernel density estimations are not to scale, as the population of blc1 points is so much smaller than the other two populations that it would otherwise not be visible on these axes. The blc1 signal (black x symbols and grey shading) is consistent with the signal-to-noise and normalized drift rate distributions (blue points and shading for lookalikes, red points and shading for mirrored lookalikes) and, although slightly higher in frequency than the peak of the kernel density estimations for the lookalike population, is still consistent with that distribution.}
\label{fig:4}
\end{figure}
\subsection{Determining the origin of the blc1 lookalike population}
We can further strengthen the claim that the lookalike population shares a common origin if we can identify a frequency-shifting instrumental or electronic effect in the data. One potential source of that effect is instrumental harmonic distortion, which can produce replicas of an original frequency $f$ signal at $2f$, $3f$, ..., etc. Another potential source is intermodulation distortion, a superset to harmonic distortion, which can produce a near-arbitrarily complex sequence of replicas which are integer multiples of the sums and differences of two or more original signals.
\subsubsection{Harmonic Analysis}
\label{sec:harmonic}
We investigated whether any of the positive-drift lookalikes could be linked via a harmonic sequence. We generated the first 20 harmonics of a range of fundamental frequencies starting outside the bandpass at 100 MHz and progressing to 1000 MHz in 1 kHz intervals. We defined a potential ``harmonic sequence'' within the data as a set of two or more lookalikes (blc1 included) associated by the same fundamental frequency, within 1 kHz of their theoretical values. blc1 was not consistent with being in a harmonic sequence with any of the observed lookalikes. However, a harmonic sequence does interlinking a set of other lookalikes (Supplementary Table 2). This harmonic sequence contained frequencies in the form $n + 0.1m + 0.099$ MHz, where $m$ and $n$ are integers---because of the constant term, we refer to this set as ``x.y99''.
\subsubsection{Intermodulation Analysis}
Some lookalikes showed additional frequency structure that was not present in blc1. Two sets of lookalikes, which we will refer to as ``Triple Feature'' (TF) and ``Single Feature'' (SF) were distinguishable from their morphology alone. An example from the TF set is shown in Supplementary Figure 10. TF and SF contained both positive lookalikes and mirrored lookalikes. TF had spacings which were integer multiples of 133.33\,MHz. SF had a more complicated relationship of spacings: primarily integer multiples of 15\,MHz, but with an additional appearance of 128\,MHz. In all cases within both sets, these frequencies are consistent with common clock oscillator frequencies used in digital electronics, with matches within 1--1000\,Hz of the expected value.
The three initially-identified sets---TF, SF, and the harmonic ``x.y99'' set (Section \ref{sec:harmonic})---each have a transition region where the morphology of the signal flips, with positive lookalikes on one side and mirrored lookalikes on the other (Supplementary Figure 11). If intermodulation effects were present, we predicted that we should detect a strong, zero-drift interferer at a frequency within the transition region, whose position is dictated by clock oscillator frequencies previously identified in the set.
For TF, we find a strong interferer at the predicted central frequency of 1400\,MHz, consistent to a single channel ($\pm$ 4\,Hz). For SF, one central frequency consistent with the 30\,MHz spacing is 1200\,MHz, where we also see a strong interferer consistent at the Hz level. For x.y99, we predict a central frequency of 1332\,MHz, but see only an extremely faint interferer; we do see strong signals at 1330.0000\,MHz, 1331.2000\,MHz, and 1332.1805\,MHz. This inconsistency could be caused by additional transposition, from an additional oscillator, present within the set, implying that one of the three frequencies listed prior is actually the responsible interferer. We conclude that these intentional-seeming separations seem likely to have been produced by the intermodulation of at least one clock oscillator with a strong interferer.
We found evidence for additional clock oscillator frequencies affecting the lookalike population, including the sets displayed in Figure \ref{fig:5}. In the top panel of Figure \ref{fig:5}, we searched for patterns with integer kHz offsets (as in Section \ref{sec:harmonic}) and uncovered two individual sequences with spacings consistent with a 2.000004\,MHz clock oscillator. Our data have a resolution of $\sim 4$\,Hz, so we expect that the last digit has an error of $\pm 4$, which will propagate in the integer multiplication of spacings. In Figure \ref{fig:5} we display the extra error at the Hz-level to illustrate that these spacings are clearly the product of the same oscillator: not only are the general spacings consistent with powers of 2.000004\,MHz, but the \textit{errors} on those spacings are consistent with propagating errors of order $\pm 4$ Hz, e.g., 16.000038\,MHz. In the bottom panel of Figure \ref{fig:5}, we see these exact spacings relating blc1 (within propagated error, of order 100\,Hz) with three lookalikes at $\sim 712$\,MHz, $\sim 856$\,MHz, and $\sim 1062$\,MHz.
\begin{figure}
\centering
\includegraphics[width=\textwidth]{blc1_aliases.pdf}
\caption{A visualization of 16 signals from the lookalike set and blc1, divided into three sets. Top Panel a): Two sets of positive lookalikes found within the data, originally identified by their position at 985\,kHz above an integer MHz in each appearance. The full spacings, to Hz precision, are shown to demonstrate the set's consistency with mixing from a 2.000004\,MHz clock oscillator. Recall that the spacings have an inherent, multiplicative $\pm 4$Hz error from the discrete frequency resolution. One spacing we see in both sets, apparently resulting from this same oscillator, is 16.000038\,MHz. The two sets are both being affected by the same oscillator, but they are not consistent with \textit{each other} to a multiple of 2.000004\,MHz, illustrating another complexity within the dataset. Bottom Panel b): blc1 (purple) shown in sequence with three additional lookalikes (red) which are consistent with integer multiples of 2.000004\,MHz, including exactly 16.000038\,MHz: a set of clock-oscillator induced spacings that is perfectly consistent with the spacings uncovered in the sets in a).}
\label{fig:5}
\end{figure}
This numerical analysis indicates that blc1 is an intermodulation product being produced by a $\sim 2$\,MHz clock oscillator being mixed with some other zero-drift \ac{RFI} elsewhere in the band. We also find power at the blc1-companion lookalike frequencies from Figure \ref{fig:5} is detected when the four archival signals at 982.002 MHz were detected, just like blc1, and is not detected when the archival signals are not present (see Supplementary Discussion 1.4).
This interpretation provides an explanation for why the signal was appearing in a part of the spectrum reserved for aviation and navigation: the source was not intending to transmit in that frequency region, and the signal was instead generated by the interaction of electronics within the transmitter, the receiver, or both. In this case, the underlying signals that intermodulate to cause the lookalikes are likely from outside Murriyang's receiver system. Signals are digitized using three analog-to-digital cards inside the telescope focus cabin, with further processing done in RF-tight cabinets in the telescope tower \cite{hobbs2020UWL}. As the lookalikes span across all three analog-to-digital cards, they are unlikely to be spurious signals generated within the receiver's digital systems.
blc1 cannot be the original signal because it is two orders-of-magnitude weaker than the strongest signals in the set and it is not seen in the off-panels, which is not replicated across the set. Evidently, blc1’s duty-cycle or variability tracks the observing cadence on ProxCen, leading to the apparent localization on the sky. If blc1 was always ``on'' at its brightest power, it would have been detected in all off-sources. Supplementary Figures 7 and 9 reveal a range of inter-panel brightness behaviours for the lookalikes: some appeared in every panel, while some were as faint as blc1 and missing from all but the first panel. \texttt{turboSETI}, by the thresholding and on-off selection mechanisms, selected the most interesting signal from a set of potentially hundreds of intermodulation products. These large numbers speak to the particular pathology of this case, as this behaviour had never been seen in over a year of \ac{UWL} observations.
\ac{RFI} environment analysis is unfortunately complex, and the \ac{RFI} environment around most astronomical radio facilities is not well-characterized at the frequency resolutions used for \ac{SETI} work. In the case of blc1, the situation is more complicated still, with mixing products that obscure the frequency and character of the original, individual interferers. It is possible that we could untangle the origins of this interference. However, as the goal of this study was to determine whether blc1 had an Earth-based or interstellar origin, we find that this is appropriate for future work.
\section{Creation of a Technosignature Verification Framework}
blc1 is the first signal-of-interest from the Breakthrough Listen program that required extensive signal verification to be undertaken. This case study led to a novel signal verification ``toolkit'' for future \ac{SETI} signals-of-interest. Similar frameworks have been applied to searches for fast radio bursts \cite{Foster2018} and gravitational waves (e.g., \cite{abbott2016}), but prior \ac{SETI} programs relied heavily on re-observation alone, without the application of a thorough checklist (e.g., \cite{horowitz, setihome}) We propose the following verification checks for narrowband technosignature signals-of-interest, once known astrophysical origins have been ruled out:
\begin{enumerate}
\item Verify that all instrumentation was functioning correctly.
\item Verify that the signal-of-interest was not present in the off-source observations at a lower \ac{S/N} threshold.
\item Check for catalogued \ac{RFI} at the same frequency at the observatory where the signal-of-interest was discovered.
\item Compare the drift rate evolution of the signal-of-interest to known accelerational and electronic drifts from human-made technology.
\item Compare the drift rate evolution of the signal-of-interest to the expected drift rates and periods in the target system and the solar system.
\item Search for other potential instances of the signal-of-interest in archival data from the same observatory.
\item Search for similar signals at other frequencies within the observation in which the signal-of-interest was detected. If found, determine if these signals have characteristics consistent with \ac{RFI}.
\item Qualitatively, or, if possible, quantitatively, assess whether similar signals (from Steps 6 and 7 above) are generated by the same phenomenon as the signal-of-interest.
\item If other signals \textit{are} from the same phenomenon as the signal-of-interest, determine whether they are \ac{RFI} using the off-source observations, or, if necessary, Steps 1--8.
\item Re-observe the target with both the same instrument and other instruments to attempt to re-detect the signal-of-interest.
\end{enumerate}
The checklist as written is appropriate for persistent, narrowband technosignature searches with single-beam, single-dish telescopes. The procedure can be applied to multi-beam instruments by using other beams as ``reference pixels'' for the on-source beam, instead of nodding to an off-source position. In addition, re-observation (Step 10) may be performed earlier in the checklist if economical, especially in cases where the signal-of-interest would be expected to be periodic or transient on long timescales (i.e., synchronized with the transit of an exoplanet).
\section{Future Work}
blc1 has underscored that, when practicable, simultaneous observations of potential technosignatures should be conducted at two different observing sites simultaneously. For \ac{ProxCen}, simultaneous observing could be accomplished with MeerKAT (e.g., \cite{jonasMeerKAT2009}) and Murriyang, whose receivers share about 1\,GHz of frequency overlap including the 982\,MHz region of blc1. There is a $\sim$4.5 hour window where the source can be observed simultaneously between the two sites. While re-observations are resource-intensive, they are also scientifically meritorious in their own right. \ac{ProxCen} is still a uniquely fascinating \ac{SETI} target for all of the reasons described in the Introduction, and it will only become better characterized by future, astrobiology-oriented studies.
We are also investigating other ways to further characterize blc1 in both the hardware and software components of the pipeline. Other Murriyang observers are monitoring for interference near 982\,MHz, which could help us identify position-dependent \ac{RFI}. To characterize aliasing behaviour within the receiver, the sky signal could be excluded by covering the feed with cryogenically-cooled RF-absorbing material which can be treated as a thermal blackbody (known as a ``cold load''). This poses engineering and logistical challenges, but may allow us to perhaps identify the different components whose mixing produces the intermodulation product at 982\,MHz. To understand the particular pathology of variability patterns such as blc1's, we could perform noise-injection testing on the lookalike population. Finally, this case study implies that taking data during slews in future single-dish \ac{SETI} observations could help us better understand signal localization and sidelobe behaviour in future programs.
\section{The impact of blc1}
While we attribute blc1 to \ac{RFI}, the benefits from its unique analysis will inform searches for years to come. This paper outlines a ``checklist'' that provides thorough next-steps for this type of signal. We have developed new software which will be incorporated into \ac{BL}'s analysis packages (e.g., \texttt{blimpy} and \texttt{turboSETI}), such that future signals-of-interest can be assessed more quickly and efficiently.
The detection of blc1 shows the success of the \ac{BL} signal detection pipeline. From 26 hours of observations over billions of channels, the \texttt{turboSETI} algorithm was able to retrieve a set of potential signals-of-interest. blc1 was then easily identified from this set upon visual inspection of the software outputs.
Conversely, this signal-of-interest also reveals some novel challenges with radio \ac{SETI} validation. It is well-understood within the community that single-dish, on-off cadence observing could lead to spurious signals-of-interest in the case where the cadence matches the duty cycle of some local \ac{RFI}. blc1 provided the first observational example of that behaviour, albeit in a slightly different manner than expected (variation of signal strength over position and time, which changed for each lookalike within the set). This case study prompts further application of observing arrays, multi-site observing, and multi-beam receivers for radio technosignature searches. For future single-dish observing, we have demonstrated the utility of a deep understanding of the local \ac{RFI} environment. To gain this understanding, future projects could perform omnidirectional \ac{RFI} scans at the observing site, record and process the data with high frequency-resolution instrumentation such as the various \ac{BL} backends, and then use narrowband search software such as \texttt{turboSETI} to obtain a population with which to characterize the statistics (in frequency, drift, power, duty-cycle, etc.) of local \ac{RFI}.
Finally, blc1 encourages us to continue working at logistical challenges that have traditionally vexed large, radio \ac{SETI} efforts. In a project with an incredibly high rate of data inflow, how do we work towards data analysis in real-time? When raw voltage data is exceedingly memory-intensive to store, how do we decide pre-analysis which observations may need additional analysis in the raw voltage products? For example, one solution to the data-storage challenge is to store only ``postage stamps'' of events with limited time and frequency ranges. Here we have learned that neglecting to consider the entire operable bandwidth of a receiver can have serious consequences: for example, losing the context that can be used to show a signal-of-interest is \ac{RFI}.
\section{Data Availability Statement}
All data used in this manuscript are stored as high-resolution filterbank files, which are available through the Breakthrough Listen Open Data Archive at \url{seti.berkeley.edu/opendata}. This includes all observations from the original observing campaign in April 2019, as well as the reobservations in November 2020, January 2021, and April 2021.
\section{Code Availability Statement}
The software tools used to read these files (I/O) and perform the narrowband search are publicly available at \url{https://github.com/UCBerkeleySETI/blimpy} and \url{https://github.com/UCBerkeleySETI/turbo_seti}.
\section{Acknowledgements}
Breakthrough Listen is managed by the Breakthrough
Initiatives, sponsored by the Breakthrough Prize Foundation. The Murriyang radio telescope is part of the Australia Telescope National Facility which is funded by the Australian Government for operation as a National Facility managed by CSIRO. We thank the staff at Murriyang for their observational support. Shane Smith and Steve Croft were supported by the National Science Foundation under the Berkeley SETI Research Center REU Site Grant No. 1950897. We thank Richard Elkins and Luigi Cruz for help with development and debugging of \texttt{turboSETI}.
\section{Author Contributions}
SZS led the data analysis and wrote the manuscript. SS and DCP uncovered blc1 and analyzed data. DD and BL ran simulations and analyzed data. DC, SC, VG, HI, ML, DM, CN, KP, AVPS, and CW assisted with data analysis, scientific interpretation, and manuscript revision. AZ obtained the data via commensal observing. JD and SPW reviewed the manuscript.
\section*{Supplementary Discussion}
\subsection*{Evaluating blc1 on the Rio Scale 2.0}
\label{ssec:rio}
A challenge in discussing potential signals-of-interest in \ac{SETI} is communicating the likelihood and significance of a discovery of \ac{ETI}. The Rio Scale \cite{almar2000discovery} and the revised Rio Scale 2.0 \cite{forgan2019rio} and ranking systems designed to mitigate this challenge. Note that the Rio Scale is built to quantify the novelty and importance of a signal, but it does not provide a framework with which to analyse and determine if a signal is interference. Here we assess blc1 using the Rio Scale 2.0 to provide context for the public for this signal-of-interest.
Using the online Rio Scale interface\footnote{\url{https://dh4gan.github.io/rioscale2/}}, we find that the phenomenon is likely to be instrumental or anthropogenic in origin, largely due to its \ac{RFI}-like form and lack of repeatability, with A (is it amenable to study and repeat observation?), B (are we sure it is not instrumental?), and C (how likely is it to be a technosignature?) values of 6, 8, and 1 respectively. This maps well with conclusions from the main text of this paper. We can then use the A, B, and C values to calculate the warranted journalistic interest $J$ and the likelihood that the signal has an interstellar origin $\delta$ from \cite{forgan2019rio}:
\begin{equation}
J = A + B + C - 20
\end{equation}
\begin{equation}
\delta = \frac{10(J-10)}{2}
\end{equation}
Here, the J value is 0, despite this being an exciting case study into the challenges of a technosignature search. The Q value of the signal, calculated separately to indicate the depth of investigation that we could pursue \textit{if} there was an ETI transmitting from \ac{ProxCen}, is relatively high, at 7. However, $\delta$, the likelihood that this signal has an interstellar origin, is 0.00001, leading to a final Rio Score score of $R = Q \times \delta = 0$.
Unfortunately, the Rio Scale does not actually give us much information about signals like blc1. There is a lack of granularity, noted in the original paper, for events that score between 0 and 1 in Rio Score. We propose that a new scaling system, perhaps logarithmic, should be developed to provide appropriate context and comparison for signals at blc1's level of interest.
\subsection*{Proxima Centauri is not alone in the beam}
In single-dish radio astronomy, the size of the beam on the sky is much larger than the target's angular extent, except for extended/diffuse sources. The low frequencies used in these data translate to approximately a 15\arcmin\ beam, which, according to GAIA DR2 \cite{gaia2018vizier} contains $\sim 22000$ known objects along this sight-line. This number is in fact a lower limit, as it does not account for extragalactic objects in the beam. This underscores the fact that, if we were to find a signal coming from interstellar distances, we would have to perform follow-up observations (assuming the source was continuous or had a relatively high duty-cycle) to determine where in the beam the source was coming from. In this example, while the beam was centered on \ac{ProxCen}, it is more accurate to say that the detection was made ``in the direction of \ac{ProxCen}''.
\subsection*{Expectations for the Proxima Centauri System}
Our analysis is mostly agnostic to the parameters, number, or habitability of planets within the triple-star system of Alpha Centauri; however incorporating this information provides a good case study for the benefits and limitations of including exoplanetary parameters in a drift analysis. Proxima\,b is a 1.3\,M$_{\oplus}$ exoplanet with an 11.2\,d period, and was detected using the ESO HARPS spectrograph \cite{anglada-escude2016terrestrial, mascareno_2020}. It has received the majority of astrobiological attention in the system, as it resides at 0.0485\,AU from its host star, and has an equilibrium temperature of 234\,K \cite{anglada-escude2016terrestrial}. There are additional planets in the \ac{ProxCen} system. In 2019, a second planet, Proxima c, with a minimum of 7 M$_{\oplus}$ and a 5.21 year period was confirmed via radial velocities from HARPS spectra data \cite{Damasso_2020}. A third planet candidate signal, which would be Proxima d, was also detected in \cite{mascareno_2020} with a 5.15 day period and a minimum mass of 0.29 M$_{\oplus}$. Given \ac{ProxCen}'s habitable zone range of $\sim$ 0.03--0.09 AU \cite{Alvarado_2020, Kane_2012}, the Proxima c candidate would be far outside of the habitable zone \cite{Damasso_2020}, while the Proxima d candidate would likely experience extreme weather conditions due to its proximity to \ac{ProxCen} \cite{Alvarado_2020}. Proxima d has yet to be confirmed, and no moons or rings have been confirmed around any planet in the system so far. Still, after finding a bright signal from infrared SPHERE images, \cite{Gratton_2020} speculated the existence of rings or dust around Proxima c.
Given our uncertainty about the number and characteristics of planets in the system, it is difficult to constrain the drift rates that we would expect from a transmitter in the system. Nevertheless, we can apply an order-of-magnitude approach for the currently-known planets. The orbital motion of Proxima b, due to its short orbital period, can produce drift rates up to $10^2$ times higher than that observed in blc1 when it is closest to the Earth. Proxima b's rotational speed is unknown, but if we set the lower and upper limits to the theoretical values when assuming that the planet is tidally-locked (lower) and assuming that the planet is rotating at its break-up speed (upper)\footnote{Using the methodology described in \cite{sheikh2019choosing}}, then the rotation of Proxima b could produce values from $10^{-1}$--$10^3$ of the observed blc1 drift rate. The consensus in the literature is that Proxima b should be tidally-locked (e.g., \cite{ribas2016prox}), and thus it is likely that rotation alone could not cause a large contribution to a drift rate from a transmitter on its surface --- the opposite of the situation of drift rate contributions of the Earth. Proxima c is likely a gas giant, and thus cannot host a transmitter on its surface (no component from rotational motion). In addition, its orbital motion would be an order-of-magnitude too low to produce drifts on the order of blc1.
Drift rate analyses like these can be useful as a heuristic tool to determine consistency with expected drift magnitude and sign, but in practice, they cannot be used to rule out a potential signal-of-interest. Even if we assume that any drifts that remain after a barycentric correction must be accelerational, it is impossible to determine \textit{a priori} exactly which drift rates we would expect at which times. The drift rate vs. time of a signal is dependent on which planet the signal is coming from, whether the transmitter is orbital or on a planetary surface, and the transmitter's latitude/angle of orbit relative to Earth. This doesn't even account for transmitters that are orbiting \ac{ProxCen} itself, orbiting a Lagrange point in the system (such as SOHO in the Solar System, e.g., \cite{domingo1995soho}) or placed on moons. In addition, we cannot neglect drifts from electronic oscillators, either unintentional or intentional. A signal from an \ac{ETI} may be electronically de-drifted for the rotational and/or orbital motion of the host system, de-drifted for the receiver's system (not necessarily Earth), or intentionally sweeping in frequency for some unknown functional reason.
\subsection*{Other appearances of blc1 also show intermodulation lookalikes}
We used the analysis above to provide another line of evidence that the archival 982.002 MHz signals (some of which are clearly RFI) are definitively related to blc1. We searched for lookalikes with the same spacings as in Figure 5. We find lookalikes at 712\,MHz and 856\,MHz in 4 of the 5 archival signals, while the 1062\,MHz lookalike is only present in 1 of them. This is not surprising, given that 1062\,MHz is the faintest lookalike in the blc1 set; the frequency comb interference is stronger at these higher frequencies; and that the entire cadence is washed out by broadband interference for one of the archival signals. To double-check this interpretation, we searched for signals at these frequencies from about a month of \ac{UWL} observations surrounding the \ac{ProxCen} campaign where we do \textit{not} see blc1 and find that, for the majority of observations, we do not see any activity at any of these frequencies either (the exception being three faint appearances of 712\,MHz over the entire month).
\section*{Supplementary Methods}
\subsection*{Determining blc1's Fundamental Parameters}
\label{ssec:blc1params}
While \texttt{turboSETI} provides a first-order best-fit start frequency and drift rate, we needed to determine blc1's fundamental parameters to higher precision in order to perform the more complex analyses detailed in the main text. We determined that the original signal on 2019 April 29 (blc1) appeared at \ac{SNR} $> 10$ in five observations. The properties of the five observations are shown in Supplementary Table S1.
To build Supplementary Table S1, we replicated \texttt{turboSETI}'s search strategy with a de-drifting algorithm. We tested a range of 500 drift rates between 0 and 0.05 Hz/s, informed by the initial \texttt{turboSETI} estimates. At each drift rate, the time rows in the time-frequency-power array were ``rolled'' to de-drift to the test drift rate; an illustration of this method is shown in Supplementary Figure S12. Once the array had been de-drifted, we then summed the spectra in the array over the entire 30\,min observation and searched for a power maximum in frequency space. We assigned the best-fit drift rate to the drift rate that produced the largest power maximum.
It is unfortunately impossible to assign a statistically-valid error value to these measurements without knowing the underlying distribution. In an idealized case, the distribution of power values around the best-fit drift rate will be symmetric and have more data in the tails of the distribution (excess kurtosis) than a Gaussian. However, because the data are discrete, this distribution will be affected by the interplay of frequency resolution, integration time, and drift rate, and these parameters will cause additional, finer structure within the distribution. This is before the noisiness and presence of \ac{RFI} within the dataset are even considered; the problem of drift rate best-fit errors within quantized data is complex, and has not been treated in the literature, to our knowledge. For the purposes of this paper, we determined a proxy for the drift rate error by fitting a Gaussian to the distribution of maxima in drift rate space. Although a Gaussian does not provide a perfect fit for the reasons described above, we are still able to obtain a relative measure of certainty in the drift rate, based on fundamental, interesting parameters such as the signal's linearity, duration and strength.
We determined the start time of the signal by dividing each 30\,min observation into 5\,min slices, and choosing the start time of the first slice where the de-drifted signal becomes significant (\ac{SNR} $\gtrsim 5$) in that observation. We set the start frequencies to the location of the maximum in frequency space for the best-fit drift rate. Finally, we determined the \ac{SNR} by dividing the summed power in the signal region over the entire 30 minute observation by the root-mean-square of power in the non-signal region for the de-drifted (best-fit drift rate) array.
We can use the average \ac{SNR} from Supplementary Table S1 to calculate blc1's apparent brightness. The \ac{UWL} system equivalent flux density (SEFD) of blank sky at 982 MHz is $\sim$38\,Jy. Thus, with an average \ac{SNR} of $\sim 18$, we can get an order-of-magnitude source brightness of
\begin{equation}
S_{\nu} \sim \frac{\textrm{SNR} \times \textrm{SEFD}}{\sqrt{t_{int} \times \Delta \nu}}
= \frac{18 \times 38 \textrm{Jy}}{\sqrt{1800 \textrm{s} \times 4 \textrm{Hz}}}
\sim 8 \textrm{Jy}
\end{equation}
\subsection*{Drift Rates from Reflections of Solar System Objects}
Although blc1 was detected during an observation of \ac{ProxCen}, we wanted to consider whether it could instead be related to a Solar System object that happened to be in or near the beam. This hypothetical object could be an extraterrestrial technosignature itself, e.g. a transmitter \cite{Bracewell1960}, beneficial from the perspective of a hypothetical \ac{ETI} because it would require orders-of-magnitude less power than an interstellar transmitter (e.g., \cite{Freitas1985,Benford2019}). On the other hand, a Solar System object may have caused blc1 via an \ac{NEO} acting as a reflector of an Earthbound radio transmitter.
The main difficulty with either explanation is that \ac{ProxCen} is far off the ecliptic plane ($b = -44\fdg8$), which is inconsistent with an object on (or in orbit around) a major planet. An interplanetary transmitter could easily be bright enough to be detected far into the sidelobes, of course, but this would not explain the null detections when observing the calibrators; without this constraint, a simpler explanation would surely be an anthropogenic satellite. Nonetheless, a Solar System object could appear in the vicinity of \ac{ProxCen} if it either has a large inclination or is a \ac{NEO} making a close approach to Earth.
We checked whether any known minor body was near \ac{ProxCen} using the Jet Propulsion Laboratory's online ``What's observable?'' tool (\url{https://ssd.jpl.nasa.gov/sbwobs.cgi}). We searched for objects in the \ac{RA} range \hms{13}{44}{43} to \hms{15}{14}{43} and \ac{Dec} range \dms{-57}{40}{46} to \dms{-67}{40}{46}, which includes everything within $5^{\circ}$ of \ac{ProxCen}, at UT 2019 April 29 15:30 as observed from Murriyang. 53 objects were returned, but most were several degrees away from \ac{ProxCen} at the time. For comparison, Murriyang's beam width at 982\,MHz is $\sim \lambda / (64{\rm\,m}) \approx 0\fdg27$.
One object stood out, however: 2019\,GG$_3$, which was within $1^{\circ}$ of \ac{ProxCen} at the time (Supplementary Figure S13). During the blc1 observations, 2019\,GG$_3$ was outside the main lobe of Murriyang, although it would pass within its footprint a couple of hours later. Coincidently, further investigation revealed that 2019\,GG$_3$ had passed within 0.07 AU of Earth just four days prior, $< 1$\% of its Minimum Orbit Intersection Distance. JPL's Small-Body Database lists that the only close encounter between 2019\,GG$_3$ and Earth was that April 25 encounter (\url{https://ssd.jpl.nasa.gov/sbdb.cgi?sstr=2019\%20GG3;orb=0;cov=0;log=0;cad=1\#cad}).
We evaluated whether the drift rate was compatible with a transmitter on 2019\,GG$_3$ using its radial velocity as observed from Murriyang, as calculated using JPL's HORIZONS ephemerides system (\url{https://ssd.jpl.nasa.gov/horizons.cgi}). The predicted drift rate (Supplementary Figure S14) is clearly incompatible with the observed drift rate (with $\chi^2 \sim 1100$) -- although the qualitative shape is correct, it has the wrong sign and is about half the magnitude expected. Thus, a constant frequency transmitter on 2019\,GG$_3$ itself was not responsible for blc1. We did not consider rotation on 2019\,GG$_3$ as a possible contributor to the drift rate.
Hypothetically, 2019\,GG$_3$ may have reflected an Earthbound transmitter. The expected returned flux is very small:
\begin{equation}
\label{eqn:Fnu_NEO}
F_{\nu} = 56\ \mathrm{mJy}\ \left(\frac{\mathrm{EIRP}}{1\ \mathrm{TW}}\right)\left(\frac{B}{3\ \mathrm{Hz}}\right)^{-1} \left(\frac{\sqrt{D_R D_T}}{0.069\ \mathrm{AU}}\right)^{-4} \left(\frac{A_r}{3000\ \mathrm{m^2}}\right),
\end{equation}
where $B$ is the transmission bandwidth, $D_R$ and $D_T$ are the reflector's distance from Earth, and $A_r$ is the effective cross section for reflection as viewed from Earth. With an absolute magnitude of $H = 24.2$, 2019\,GG$_3$ has a radius of $9.7/\sqrt{p}$ meters for an albedo $p$. Thus for an $A_r$ comparable to its geometric area, even the brightest radars on Earth would be incapable of generating a detectable return. The maximum possible $A_r$ for an object with geometric area $A_g$ can be estimated with Rayleigh's criterion to be $\sim 4 \pi A_g^2 / \lambda^2$, or $\sim 10^3\,\mathrm{km}^2 (A_g / 3000\,\mathrm{m}^2)^2$ at 982\,MHz. This would allow a return of 0.28 Jy (with a transmitter EIRP of 1 TW, a channel bandwidth of 3 Hz, a distance of 0.069 AU, and an effective radar cross-section of order $\pi/4$ (60 m)$^2$). It would require a rather contrived reflector, essentially a flat sheet oriented precisely to bounce the transmitter's signal into Murriyang.
A further argument against 2019\,GG$_3$ being a reflector comes from the drift rate. The predicted drift rate of a reflection is the drift rate due to the sums of the radial acceleration between transmitter and reflector and that between the reflector and receiver. Given position vectors $\mathbf{r_T}$ for the transmitter, $\mathbf{r_S}$ for the apparent source (reflector), and $\mathbf{r_R}$ for the receiver:
\begin{equation}
\dot{\nu} = \frac{\nu}{c} \left[ \frac{(\mathbf{\ddot{r}_S} - \mathbf{\ddot{r}_T}) \cdot (\mathbf{r_S} - \mathbf{r_T}) + v_{TS}^2 - v_{TS_{\rm rad}}^2}{|\mathbf{r_S} - \mathbf{r_T}|} + \frac{(\mathbf{\ddot{r}_S} - \mathbf{\ddot{r}_R}) \cdot (\mathbf{r_S} - \mathbf{r_R}) + v_{RS}^2 - v_{RS_{\rm rad}}^2}{|\mathbf{r_S} - \mathbf{r_R}|}\right] ,
\end{equation}
where $v_{xy} = |\mathbf{\dot{r}_x} - \mathbf{\dot{r}_y}|$ is the relative speed of $x$ and $y$ and $v_{xy_{\rm rad}} = (\mathbf{\dot{r}_x} - \mathbf{\dot{r}_y}) \cdot (\mathbf{r_x} - \mathbf{r_y}) / |\mathbf{r_x} - \mathbf{r_y}|$ is the radial velocity between $x$ and $y$. When both the transmitter and receiver are Earthbound and the reflector is at a great distance, one can ignore parallax and the drift rate is approximated as:
\begin{equation}
\dot{\nu} \approx \frac{\nu}{c} \left[2 \dot{v}_{S\oplus_{\rm rad}} + \frac{4 \pi^2 R_{\oplus}}{P_{\oplus}^2} \cos \delta_S (\cos \Lambda_R \cos(\alpha_S - \tau_R) + \cos \Lambda_T \cos(\alpha_S - \tau_T))\right] .
\end{equation}
In this equation, $\dot{v}_{S\oplus_{\rm rad}}$ is the radial acceleration between the reflector and Earth's geocenter, $\alpha_S$ and $\delta_S$ are the \ac{RA} and \ac{Dec} of the reflector as viewed from Earth, $R_{\oplus}$ and $P_{\oplus}$ are the radius and sidereal rotation period of the Earth, $\Lambda_T$ and $\Lambda_R$ are the latitudes of the transmitter and receiver, and $\tau_T$ and $\tau_R$ are the sidereal time at the transmitter and receiver expressed in radians.
We modeled transmitters located on a $1^{\circ} \times 1^{\circ}$ equirectangular grid on Earth's surface. The fit between the measured drift rates and the predicted drift rates are evaluated using the $\chi^2$ value. No transmitter location does an adequate job explaining blc1's drift as a reflection off 2019\,GG$_3$ (Supplementary Figure S15). The problem is that the transmitter-asteroid drift must overcome both the negative drift from Murriyang and the negative drift from 2019\,GG$_3$'s recession from Earth. When considering transmitters for which 2019\,GG$_3$ was above the horizon for blc1's duration, the best fit is $\chi^2 = 538$.
2019\,GG$_3$'s encounter with Earth and its near-transit of \ac{ProxCen} during the blc1 time frame is thus almost certainly just an interesting coincidence. These considerations do not rule out that an unknown Solar System object was responsible for blc1, although it would need to be small to have gone undetected.
\begin{table}[ht]
\centering
\begin{tabular}{c|c|c|c|c|c}
\hline
\# & Start Time (UTC) & Start Freq. (MHz) & DR (Hz/s) & DR Error Proxy (\%) & SNR \\
\hline
1 & 13:17:35 & 982.00241 & 0.0312 & 15.1 & 12.64 \\
2 & 13:47:46 & 982.00247 & 0.0303 & 22.8 & 28.09 \\
3 & 14:43:40 & 982.00257 & 0.0305 & 14.8 & 21.17 \\
4 & 15:21:41 & 982.00264 & 0.0273 & 22.3 & 17.70 \\
5 & 17:54:26 & 982.00281 & 0.0112 & 67.0 & 10.18 \\
\hline
\end{tabular}
\caption{Start time and frequency, drift rate DR (with error proxy), and signal-to-noise ratio SNR for the five observations of \ac{ProxCen} which comprise blc1. All five observations have strong (\ac{SNR} $>$ 10) detections, all detections appear to be part of the same phenomenon, and the signal is strongly drifting, with a changing drift rate over time, perhaps indicative of changing radial acceleration or electronically-caused frequency shifting.}
\label{tab:blc1_properties}
\end{table}
\begin{table}[ht]
\centering
\begin{tabular}{|c|c|c|c|}
\hline
Fundamental Frequency (MHz) & Harmonic Frequencies (MHz) & $n$th Harmonic \\
\hline
108.225 & 865.799927, 1298.699886 & 7, 11 \\
133.200 & 799.199932, 1065.599907, 1198.799896 & 5, 7, 8 \\
144.300 & 865.799927, 1298.699886 & 5, 8 \\
149.850 & 899.099922, 1198.799896 & 5, 7 \\
199.800 & 799.199932, 1198.799896 & 3, 5 \\
216.450 & 865.799927, 1298.699886 & 3, 5 \\
266.400 & 799.199932, 1065.599907 & 2, 3 \\
299.700 & 899.099922, 1198.799896 & 2, 3 \\
399.6 & 799.199932, 1198.799896 & 1, 2 \\
432.900 & 865.799927, 1298.699886 & 1, 2 \\
\hline
\end{tabular}
\caption{A set of harmonic sequences that were discovered by the harmonic analysis in Results, defined by their fundamental frequency, the frequencies of the lookalikes that are members of the sequence, and the respective harmonic number $n$. blc1 is not a member of any of these sequences, however, these sequences draw from the same sub-set of lookalikes (with harmonic frequencies at ``x.y99'' MHz), which indicates that at least some of the lookalikes are related in origin.}
\label{tab:harmonics}
\end{table}
\begin{figure}[ht]
\centering
\includegraphics[width=11.5cm]{Drift.pdf}
\caption{Doppler drift characteristics of a fixed-frequency transmitter on a vehicle travelling along a local highway (A39) at 80 km/hr (green), 20 km/hr (orange), and 10 km/hr (blue), as well as on a plane executing a 850 km/hr great circle flight from Melbourne to Sydney (red, only shown while it is above the horizon). The black data points show the blc1 data, where the x-axis bars are the time extent and the y-axis are error bars. All ``trips'' start at the first data point, and none can explain the blc1 drift characteristics.
\label{fig:EarthBoundDrift}}
\end{figure}
\begin{figure}[ht]
\centering
\includegraphics[width=\textwidth]{barygeodrift.pdf}
\caption{The cartoon on the left shows the geometry of the solar system in the plane of the ecliptic and perpendicular to that plane. The sinusoidal pattern shows the net velocity in the pointing direction, as seen from Parkes, over two days. The black portion is during the appearance of blc1. The orange line in the drift rate vs. time plot on the right shows the result of that motion as residual barycentric/geocentric drift (on this scale both corrections are identical), compared to the actual observed datapoints of blc1. The motion from Earth's rotation produced a drift rate in the direction of ProxCen, over the course of the observation, that is not consistent with the drift we see from blc1.
\label{fig:BaryGeoDrift}}
\end{figure}
\begin{figure}[ht]
\centering
\includegraphics[height=0.92\textheight]{SameCamp2.pdf}
\caption{A waterfall plot around 982\,MHz from 2019 May 1, showing a possible blc1 redetection. The \ac{RFI} comb is clearly visible in this observation.}
\label{fig:blc1c_cadence}
\end{figure}
\begin{figure}[ht]
\centering
\includegraphics[height=0.92\textheight]{SameCamp3.pdf}
\caption{A waterfall plot around 982\,MHz on 2019 May 2, showing a possible blc1 redetection. As with Figure S4, the \ac{RFI} comb is clearly visible in the background as well as broadband \ac{RFI} during the final two observations in the sequence.}
\label{fig:blc1d_cadence}
\end{figure}
\begin{figure}[ht]
\centering
\includegraphics[height=0.92\textheight]{SameCamp4.pdf}
\caption{A waterfall plot around 982\,MHz on 2019 May 4, showing a possible blc1 redetection. The possible redetection has a low \ac{SNR} of $\sim$6, comparable with the \ac{RFI} frequency comb.}
\label{fig:blc1e_cadence}
\end{figure}
\begin{figure}[ht]
\centering
\includegraphics[height=0.9\textheight]{DiffCamp.jpg}
\caption{A waterfall plot around 982\,MHz depicting an \ac{RFI} signal detected 4 days prior to blc1, in the direction of HIP 98470 (not part of the \ac{ProxCen} campaign). Note that each panel shows a 5-minute integration, and time still progresses from top to bottom. This signal is also simultaneous with the frequency comb, and appears to have frequency, drift rate, and \ac{SNR} similar to blc1.}
\label{fig:rfi}
\end{figure}
\begin{figure}[ht]
\centering
\includegraphics[height=0.83\textheight]{same_morphology.pdf}
\caption{Waterfall plots for 18 of the 36 \ac{RFI} signals, seen in the same observation as blc1, that were found to have the same drift morphology as blc1. All plots cover the same observing time range on the same targets, but at different frequencies (displayed on the bottom of each plot in MHz). These plots are meant to serve as a visual illustration of the sample and thus have had their axes removed; for a quantitative survey of the parameters, see Figure 5. Note the qualitative differences in signal strength, brightness over time, stability over short timescales, and upon close examination, relative ending frequency. These features, along with a lack of a mathematical relation between every member of this set, indicate that these signals probably do not come from a single source.}
\label{fig:same_morph}
\end{figure}
\begin{figure}[ht]
\centering
\includegraphics[height=0.85\textheight]{non_lookalikes.pdf}
\caption{Two morphologically distinct examples from the set of 76 signals with similar drift rates to blc1 in the first panel that \textit{did not} have the same drift morphology as blc1. These do not appear to be produced by the same mechanism as the lookalikes in Figure \ref{fig:same_morph}.}
\label{fig:non_lookalikes}
\end{figure}
\begin{figure}[ht]
\centering
\includegraphics[height=0.85\textheight]{mirror_morphology.pdf}
\caption{Waterfall plots for 18 of the 27 \ac{RFI} signals, simultaneous to blc1, that were found to have the same drift morphology as blc1 but with negative drift rates. All plots cover the same observing time range on the same targets, but at different frequencies (displayed on the bottom of each plot in MHz). These plots are meant to serve as a visual illustration of the sample and thus have had their axes removed; for a quantitative survey of the parameters, see Figure 5. Again we see significant differences in various signal parameters.}
\label{fig:mirror_morph}
\end{figure}
\begin{figure}[ht]
\centering
\includegraphics[width=\textwidth]{TF.pdf}
\caption{The first panel of a \ac{ProxCen} observation of a blc1 lookalike (top) and blc1 mirrored lookalike (bottom) that show the exact same frequency vs. time morphology, but inverted along the frequency axis. The ``triple feature'' (TF) name was given based on the three wiggles, to the left in the top panel and to the right in the bottom panel, which created a distinctive morphology only displayed by a small sub-set of lookalikes.}
\label{fig:TF}
\end{figure}
\begin{figure}[ht]
\centering
\includegraphics[width=0.97\textwidth]{TF_SF_aliases.jpg}
\caption{Three sets of lookalikes, including both mirrored and non-mirrored members, with spacings that are integer multiples of common clock oscillator frequencies. Red (right-facing arrow) indicates a positive lookalike while black (left-facing arrow) indicates a negative lookalike. The top sequence, labelled ``x.y99 MHz'', is a set of 14 signals separated by integer multiples of 33.3\,MHz, found via similarities in their relative position within each MHz (e.g., 1065.599 MHz and 1098.899 MHz). The middle sequence is labelled TF, for ``Triple Feature'', and consists of 5 signals that have the same frequency structure as the examples shown in Figure \ref{fig:TF} and are consistent with integer multiples of 133.33\,MHz. The bottom sequence is labelled SF, for ``Single Feature'', and consists of 7 signals that show the same frequency structure. All of the SF signals are spaced by multiples of 15\,MHz, except for the first pair, which is separated by 128.003\,MHz.}
\label{fig:TF_SF_aliases.jpg}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\textwidth]{dedrift_roll_figure.pdf}
\caption{A cartoon illustrating the process of de-drifting a waterfall plot by ``rolling'' the rows. The graphic on the left shows a simplified waterfall plot with a drifting, unresolved signal in red. The power from this signal is spread over the majority of the frequency range when integrated over time, as shown in the power spectrum above the grid. However, as shown in the right graphic, when each row is shifted to align the signal in a single frequency bin via discretely shifting each pixel based on its width and the desired drift rate, the power in a single bin increases greatly, indicating a good fit to the actual drift rate of the signal. The frequencies at the edges cannot be used with this method (as indicated in grey in the cartoon); with the \texttt{numpy.roll} method \cite{harris2020array}, each row would actually wrap such that the gray boxes would be consolidated on the right of the figure. Regardless, less than 1\% of each waterfall plot falls in these edge pixels for the actual data used in this analysis.}
\label{fig:dedrift_roll}
\end{figure}
\begin{figure}[ht]
\centering
\includegraphics[width=0.5\textwidth]{ProximaMap_2019GG3_29Apr2019.jpg}
\caption{The track of 2019\,GG$_3$ relative to \ac{ProxCen} (star) on 2019 April 29 as it appeared from Parkes. The blue disk roughly indicates the Murriyang 64-m beam size at 982\,MHz. Red highlighting indicates the duration of the blc1 observations. \label{fig:AsteroidTrack}}
\end{figure}
\begin{figure}[ht]
\centering
\includegraphics[width=0.5\textwidth]{DriftRate_2019GG3.jpg}
\caption{Drift rate for 2019\,GG$_3$ compared to that measured for blc1 (black points). The pale blue dashed line shows the frequency drift due to 2019\,GG$_3$'s orbital motion alone (ignoring Earth's rotation), as observed from the Earth's center. The drift rate of 2019\,GG$_3$ itself as viewed from Parkes (black line), including Earth's rotation and orbital motions, is compared with the best-fit reflector model from an Earthbound transmitter (blue line). Neither a transmitter on GG$_3$ itself nor a reflection off of GG$_3$ from an Earthbound transmitter can explain the observed drift of blc1.
\label{fig:AsteroidDrift}}
\end{figure}
\begin{figure}[ht]
\centering
\includegraphics[width=0.8\textwidth]{chi2_Drift_2019GG3_TSR.jpg}
\caption{A map of the Earth showing the $\chi^2$ when fitting each possible transmitter location with 2019\,GG$_3$ as the reflector (predicted) to the drifts seen in blc1 (measured). Contours are for $\chi^2$ of $600$ and $1000$. No Earthbound transmitter location can explain the drift rate of blc1 in reflection off of GG$_3$. (World map credit: Tom Patterson \textit{et al.}, via Wikimedia, \url{https://commons.wikimedia.org/wiki/File:Blank_Map_Equirectangular_states.svg}) \label{fig:EarthGrid-Asteroid}}
\end{figure}
\begin{figure}[ht]
\centering
\includegraphics[width=\textwidth]{acfs.jpg}
\caption{Auto-correlation functions of an observation of blc1 (top) and an observation of a potential redetection (bottom) in frequency, taken at similar UTCs one day apart. Both observations have signs of a periodic structure at a frequency lag of about 25 channels ($\sim 100$\,Hz, which corresponds to the comb-like structure visible in Figure 3.}
\label{fig:acfs}
\end{figure}
\begin{figure}[ht]
\centering
\includegraphics[width=\textwidth]{waterfall_butterfly.pdf}
\caption{Two example waterfall plots a) and c) and two example butterfly plots b) and d). The first row contains plots from one of the 30-minute ProxCen observations without a detection from 29 April. The waterfall plot a) looks like noise at a glance, but the butterfly plot b) reveals traces of the \ac{RFI} comb mentioned in Section 2.1. Neither plot shows any trace of blc1. The second row contains plots from the second 30-minute detection of blc1. Note the feature offset at a drift rate of about 0.03\,Hz/s in d): the butterfly plot highlights linear features from the waterfall plots in a manner conducive to visual inspection.}
\label{fig:waterfall_butterfly}
\end{figure}
\begin{acronym}
\acro{SETI}{the Search for Extraterrestrial Intelligence}
\acro{SNR}{signal-to-noise ratio}
\acro{RFI}{Radio Frequency Interference}
\acro{BL}{Breakthrough Listen}
\acro{ProxCen}{Proxima Centauri}
\acro{UWL}{Ultra-Wide bandwidth, Low-frequency receiver}
\acro{NEO}{Near-Earth Object}
\acro{ETI}{Extraterrestrial Intelligence}
\acro{RA}{right ascension}
\acro{Dec}{declination}
\end{acronym}
\clearpage
\newpage
|
\section{Introduction}
Let $f,g\in\mathbb{Z}[x]$ be monic polynomials with resultant $r$. Some properties of the greatest common divisor of $f(n)$ and $g(n)$ was examined by P. Frenkel and J. Pelik\'an in their paper \cite{frenkel2017greatest}. They managed to prove the following theorems:
\begin{enumerate}
\item For every integer $n\in \mathbb{Z}$ the $\lnko{f(n),g(n)}$ divides $r$.
\item If $r$ square-free, then every divisor of $r$ will occur in the series $(\lnko{f(n),g(n)})_{n\in\mathbb{Z}}$.
\item If $p^p$ does not divide $r$ for every $p$ prime, then there exists an $n\in\mathbb{Z}$ integer such as $\lnko{f(n),g(n)}=1$.
\end{enumerate}
As the aforementioned claims show, the series $(\lnko{f(n),g(n)})_{n\in\mathbb{Z}}$ and the $r$ resultant are strongly related. This connection motivated further investigation of the topic. Let $v_p$ be the $p$-adic valuation and assume that $v_p(\lnko{f(n),g(n)})=\min{v_p(f(n),v_p(g(n))}\ge s$ for every $n\in\mathbb{Z}$. Furthermore, $S$ denotes the maximum, $S=\max_{n\in\mathbb{Z}}{\min{v_p(f(n)),v_p(g(n))}}$. This notion helps us to rewrite the first result of P. Frenkel and J. Pelikán to the form $v_p(r)\ge S$. In the paper \cite{frenkel2018estimating} the authors, P. Frenkel and G. Z\'abr\'adi, were focusing on this inequality, and their interest was finding the best possible lower bound for $v_p(r)$ in terms of $S,s$ parameters. A short summary of their paper is presented below:
\begin{enumerate}
\item The $v_p(r)\ge S-s+B(2s)-2B(s)$ inequality holds for every non-negative $s$, where $B$ is a function defined in their paper.
\item The corollary of (1) is the following inequality:
$$v_p(r)\ge S-s+(p-1)s^2$$ This lower bound is asymptotically sharp.
\item For every $s\le p$ we have $v_p(r)\ge S-s+ps^2$ and it is sharp.
\item In the case of equality in the last statement, the series $(\min{v_p(f(n)),v_p(g(n))})_{n\in\mathbb{Z}}$ takes every value from the interval $[s,S]$.
\end{enumerate}
In this paper we establish a new lower bound for $v_p(r)$ and also examine more general problems. The main results of this paper are the following theorems.
\begin{Th}
Let $f,g\in\mathbb{Z}[x]$ be monic polynomials with nonzero $r$ resultant and assume that $\min\{v_p(f(n),v_p(g(n))\}\ge s$ for every $n\in \mathbb{Z}$. Then we have the following lower bound for $v_p(r)$.
\begin{align*}
v_p(r)\ge S-s+ps^2\frac{p-1}{p-p^{1-\lfloor \log_p((p-1)s+1)\rfloor}}
\end{align*}
\end{Th}
\bigskip
\begin{Th}
Let $f,g\in\mathbb{Z}[x]$ be monic polynomials with nonzero $r$ resultant and assume that $v_p(f(n))\ge s_1$ and $v_p(g(n))\ge s_1$ for every $n\in \mathbb{Z}$. Then we have the following lower bound for $v_p(r)$.
\begin{align*}
v_p(r)\ge S-\max{s_1,s_2}+ps_1s_2\frac{p-1}{p-p^{-k}}
\end{align*}
Where $k=\lfloor \log_p((p-1)\max\{s_1,s_2\}+1)\rfloor -1$.
\end{Th}
\section{An approach with $p$-adic numbers}
Throughout this chapter $f,g$ will denote monic polynomials with integer coefficients and with nonzero $r$ resultant. Furthermore, assume that $v_p(f(n))\ge s_1$ and $v_p(g(n))\ge s_2$ for every $n\in \mathbb{Z}$. (Note: $s_1,s_n\in \mathbb{N}$) Let $K$ be the splitting field of the polynomials $f,g$ and write $f,g$ as $f(x)=\prod_{i=1}^k(x-\alpha_i)$, $g(x)=\prod_{j=1}^l(x-\beta_j)$. From the definition of $v_p$ it follows that
\begin{align}
v_p(f(n))&=\sum_{i=1}^k v_p(n-\alpha_i)\ge s_1\quad \forall ~n\in\mathbb{Z} \\
v_p(g(n))&=\sum_{j=1}^l v_p(n-\beta_j)\ge s_2 \quad \forall ~n\in\mathbb{Z} \\
v_p(r)&=\sum_{i,j} v_p(\alpha_i-\beta_j)
\end{align}
\begin{Def}
\label{karDef}
For arbitrary $(x-\delta)\in\overline{\mathbb{Q}_p}[x]$ linear polynomial, $m\in\mathbb{Z}_p$ $p$-adic integer and $t\in\mathbb{R}_{\ge0}$ non-negative real we define
$$\kar{x-\delta}=\begin{cases}1&~~\text{if}~ v_p(m-\delta)\ge t\\0&~~\text{otherwise}\end{cases}$$
In general, for higher degree polynomials we define $\kar{h}$ to be $\kar{h}=\sum_i\kar{x-\delta_i}$, where $h(x)=\prod_i(x-\delta_i)$.
\end{Def}
\begin{lemma}
For arbitrary $h\in\mathbb{Z}[x]$ monic polynomial and $m\in\mathbb{Z}_p$ $p$-adic integer
\begin{enumerate}
\item $\kar{h}$ is a monotonically decreasing function in $t$.
\item $v_p(h(m))=\int_0^\infty \kar{h}dt$
\end{enumerate}
\end{lemma}
\begin{proof}
If $h(x)=x-\delta$ is a linear monic polynomial, then
\begin{align*}
\kar{x-\delta}=\begin{cases}
1~~&\text{if}~ t\in [0,v_p(m-\delta)]
\\0~~&\text{otherwise}
\end{cases}
\end{align*}
Therefore $\kar{x-\delta}$ is monotonically decreasing and $\int_0^\infty \kar{h}dt=v_p(m-\delta)$.
\\The statements for higher degree polynomials follow from the linear aspect of $\kar{h}$. Let $h\in\overline{\mathbb{Q}_p}[x]$ be a monic polynomials and $h(x)=\prod_i (x-\delta_i)$ be its factorization. The sum of monotonically decreasing functions is monotonically decreasing and $\kar{h}=\sum_i\kar{x-\delta_i}$, therefore $\kar{h}$ is indeed monotonically decreasing. On the other hand
$$\int_0^\infty \kar{h}dt=\sum_i\int_0^\infty \kar{x-\delta_i}dt=\sum_i v_p(m-\delta_i)=v_p(h(m))$$
which proves (2).
\end{proof}
\begin{lemma}\label{vprchiint}
$$v_p(r)\ge \int_0^\infty \sum_{m=1}^{p^{\lceil t\rceil}}\kar{f}\kar{g} dt$$
\end{lemma}
\begin{proof}
Notice that $v_p(r)=\sum_{i,j}(\alpha_i-\beta_j)$ and $v_p(\alpha_i-\beta_j)\ge \min\{v_p(m-\alpha_i),v_p(m-\beta_j)\}$ for every $m\in\mathbb{Z}_p$ due to the triangular inequality. It follows that
$$v_p(r)=\sum_{i,j}v_p(\alpha_i-\beta_j)\ge\sum_{i,j}\max_{m\in\mathbb{Z}_p}\min\{v_p(m-\alpha_i),v_p(m-\beta_j)\}$$
We need one last observation to transform the last inequality into a much convenient one. Notice that
$$\min\{v_p(m-\alpha_i),v_p(m-\beta_j)\}=\int_0^\infty \kar{x-\alpha_i}\kar{x-\beta_j}dt$$
The $\kar{x-\alpha_i}\kar{x-\beta_j}$ takes $1$ if and only if both terms are $1$ which is equivalent to saying that $t\in[0,v_p(m-\alpha_i)]\cap[0,v_p(m-\beta_j)]=[0,\min\{v_p(m-\alpha_i),v_p(m-\beta_j)\}]$. Therefore the right hand side is equal to the length of the interval within the integrand takes $1$.
\\The next step is finding a way to effortlessly determine $m\in\mathbb{Z}_p$ which maximizes the aformentioned quantity. For given $t$ the $l(m):=\kar{x-\alpha_i}\kar{x-\beta_j}$ takes $1$ for at most one $m$ from the set $\{1,2\dots p^{\lceil t \rceil}\}$, otherwise assume that $l(m_0)=l(m_1)=1$. According the \ref{karDef} definition $v_p(m_0-\alpha_i)\ge t$ and $v_p(m_1-\alpha_i)\ge t$. The triangle inequality implies that $v_p(m_0-m_1)\ge t$ which means that $m_0$ and $m_1$ in the same residue class modulo $p^{\lceil t \rceil}$. Therefore $m_0$ and $m_1$ must be equal if $m_0,m_1\in\{1,2\dots p^{\lceil t \rceil}\}$.
\\There exist $m_0\in \mathbb{Z}_p$ which maximizes $\min\{v_p(m-\alpha_i),v_p(m-\beta_j)\}$ otherwise $v_p(m-\alpha_i)$ and $v_p(m-\beta_j)$ would be arbitrary large for sufficient $m$ and due to the triangle inequality $v_p(\alpha_i-\beta_j)$ would be arbitrary large as well. However, the last one implies $\alpha_i=\beta_j$ which contradicts the assumption that the resultant is not zero.
\\Obviously,
$$\int_0^\infty \chi_t^{(x-\alpha_i)}(m_0)\chi_t^{(x-\beta_j)}(m_0)dt\le
\int_0^\infty \sum_{m=1}^{p^{\lceil t\rceil}}\kar{x-\alpha_i}\kar{x-\beta_j} dt$$
and since there is at most one non-zero term in each sum for a given $t$, we have
$$\int_0^T \chi_t^{(x-\alpha_i)}(m_0)\chi_t^{(x-\beta_j)}(m_0)dt\ge
\int_0^T \sum_{m=1}^{p^{\lceil t\rceil}}\kar{x-\alpha_i}\kar{x-\beta_j} dt$$
where $T=\min\{v_p(m_0-\alpha_i),v_p(m_0-\beta_j)\}$. On the other hand, both integrands are $0$ if $t>T$ otherwise $m_0$ would not satisfy the maximal condition. So, we got that
\begin{align*}
\int_0^\infty \sum_{m=1}^{p^{\lceil t\rceil}}\kar{x-\alpha_i}\kar{x-\beta_j} dt&=\int_0^\infty \chi_t^{(x-\alpha_i)}(m_0)\chi_t^{(x-\beta_j)}(m_0)dt
\\&=\max_{m\in\mathbb{Z}_p}\min\{v_p(m-\alpha_i),v_p(m-\beta_j)\}
\end{align*}
Finally, we need to put the pieces together.
\begin{align*}
v_p(r)&\ge \sum_{i,j}\max_{m\in\mathbb{Z}_p}\min\{v_p(m-\alpha_i),v_p(m-\beta_j)\}\\
&=\sum_{i,j} \int_0^\infty \sum_{m=1}^{p^{\lceil t\rceil}}\kar{x-\alpha_i}\kar{x-\beta_j} dt\\
&=\int_0^\infty \sum_{m=1}^{p^{\lceil t\rceil}}\left(\sum_i\kar{x-\alpha_i}\right)\left(\sum_j\kar{x-\beta_j} \right)dt\\
&=\int_0^\infty \sum_{m=1}^{p^{\lceil t\rceil}}\kar{f}\kar{g} dt
\end{align*}
\end{proof}
\begin{Def}
The integer transformed version of $\kar{h}$ is denoted by $\kkar{h}$ and defined by the equation:
$$\widehat{\chi}_r^{(h)}(m)=\int_{\lceil r\rceil-1}^{\lceil r\rceil}\kar{h} dt$$
\end{Def}
\begin{lemma}
The function, $\kkar{h}$ satisfies the following properties:
\begin{enumerate}
\item It takes only integer values for every $m\in\mathbb{Z}_p$ and $t\in\mathbb{R}_{\ge0}$.
\item It is monotonically decreasing in $t$.
\item $$v_p(h(m))=\int_0^\infty \kkar{h}dt=\sum_{t=1}^\infty \kkar{h}$$
\end{enumerate}
\end{lemma}
\begin{proof}
Let $h(x)=\prod_{i=1}^k (x-\delta_i)$ be a monic irreducible polynomial over $\mathbb{Q}_p$ with $\delta_i$ roots in its splitting field $K$. Since $h$ is irreducible, we know that
$$v_p(m-\delta_i)=\frac{v_p(h(m))}{k}$$
Therefore $\kar{h}=k$ for every $t\in [0,\frac{v_p(h(m))}{k}]$ and $0$ otherwise. Since $v_p(h(m))$ is an integer, the length of the interval $[\lceil r \rceil -1, \lceil r\rceil]\cap [0, \frac{v_p(h(m))}{k}]$ can be written as $\frac{l}{k}$, where $l\in \mathbb{Z}$. So, we have that
\begin{align*}
\widehat{\chi}_r^{(h)}(m)=\int_{\lceil r\rceil-1}^{\lceil r\rceil}\kar{h} dt=k\frac{l}{k}\in \mathbb{Z}
\end{align*}
On the other hand, $\kkar{h_1}+\kkar{h_2}=\kkar{h_1h_2}$ which implies that $\kkar{h}$ takes only integer values for every monic polynomial.
\\Since $\kar{h}$ is monotonically decreasing and we got $\kkar{h}$ by averaging out $\kar{h}$ on every $[\lceil r\rceil-1,\lceil r\rceil]$ interval, $\kkar{h}$ is also monotonically decreasing.
\\Finally, the last equation follows from the facts that $v_p(h(m))=\int_0^\infty \kar{h}dt=\int_0^\infty \kkar{h}dt$ and $\widehat{\chi}_r^{(h)}(m)=\int_{\lceil r\rceil-1}^{\lceil r\rceil}\kkar{h}$.
\end{proof}
\begin{lemma}\label{integralszummabecs}
\begin{align*}
\int_0^\infty \sum_{m=1}^{p^{\lceil t\rceil}}\kar{f}\kar{g} dt&\ge \int_0^\infty \sum_{m=1}^{p^{\lceil t\rceil}}\kkar{f}\kkar{g} dt\\&=\sum_{t=1}^\infty \sum_{m=1}^{p^{t}}\kkar{f}\kkar{g}
\end{align*}
\end{lemma}
\begin{proof}
Recall that $F_r:={\widehat{\chi}_r^{(f)}(m)}=\left(\int_{r-1}^{r}\kar{f}\right)$ and ${\widehat{\chi}_r^{(g)}(m)}=\left(\int_{r-1}^{r}\kar{g}\right)$. Since $\kar{f}$ and $\kar{g}$ are monotonically decreasing non-negative functions, the $\kar{f}-F_r$ is non-negative on the interval $[r-1, r_0]$ and negative on $(r_0, r]$. Furthermore, we have
$$U:=\int_{r-1}^{r_0} (\kar{f}-F_r)dt = - \int_{r_0}^r (\kar{f}-F_r)dt$$
where $U\ge 0$.
The monotonically decreasing property of $\kar{g}$ implies that
\begin{align*}
&\int_{r-1}^{r_0} (\kar{f}-F_r)\kar{g} dt \ge U\chi_{r_0}^{(g)}(m) \ge -\int_{r_0}^r (\kar{f}-F_r)\kar{g}dt
\\\Rightarrow &\int_{r-1}^r (\kar{f}-F_r)\kar{g} dt \ge 0
\\\Rightarrow &\int_{r-1}^r \kar{f}\kar{g} dt \ge F_r\left(\int_{r}^{r+1}\kar{g}\right)={\widehat{\chi}_r^{(f)}(m)}{\widehat{\chi}_r^{(g)}(m)}
\end{align*}
The rest of the proof follows from the linearity aspect of $\kkar{.}$ and $\kar{.}$.
\end{proof}
\begin{lemma}
\label{divisonlemma}
\begin{align*}
\widehat{\chi}_{t-1}^{(f)}(m)\ge \sum_{i=0}^{p-1}\widehat{\chi}_{t}^{(f)}(m+ip^{t-1})
\end{align*}
\end{lemma}
\begin{proof}
Since both sides of the inequality are additive, we only need to prove it for $f(x)=x-\delta$. Assume that $t\in(T-1,T]$ for some $T\in\mathbb{Z}$, then $\chi_t^{(x-\delta)}(m+ip^{T-1})$ is $1$ for at most one $i\in\{0,1\dots p-1\}$. Suppose the opposite, then $v_p(m+ip^{t-1}-\delta)\ge t$ and $v_p(m+jp^{t-1}-\delta)\ge t$, which implies that $v_p((i-j)p^{r-1})\ge t>r-1$, thus $v_p(i-j)> 0$, thus $i=j$. The $\kkar{x-\delta}$ is monotonically decreasing, and it only takes $0,1$ values, thus the sum is zero if the left side is zero, or at most one if the right side is one.
\end{proof}
\section{Combinatorial Approach}
In the previous section, we managed to extract some beneficial equations. Although it has not been clarified yet, those equations let us handle the main problem in a combinatorial manner. In the subsequent chapter, we will show how to derive this form, but first, we want to lay the combinatorial foundation.
\begin{Def}
Let $\mathcal{F}=(V,E)$ be an infinite directed tree with $\delta$ root such as every vertex has $p$ out-degree.
\end{Def}
\begin{Def}
\label{abdef}
An $a:V\to \mathbb{R}_{\ge1}\cup\{0\}$ weight function with $\omega_a$ weight is a function on the vertices of $\mathcal{F}$ satisfying the following properties.
\begin{enumerate}
\item For every $T$ infinite path from $\delta$, $\sum_{v\in T}a(v)\ge \omega_a$.
\item For every $v\in V$ vertex $a(v)\ge \sum_{u\in N_v} a(v)$, where $N_v$ denotes the set of end points of out-edges from $v$.
\end{enumerate}
Furthermore, if the range of $a$ is a subset of $\mathbb{Z}_{\ge 0}$ then we call $a$ an integral weight function and a real weight function otherwise.
\end{Def}
\begin{Def}
The scalar product of two weight is defined by the following equation:
$$\scal{a}{b}=\sum_{v\in V}a(v)b(v)$$
\end{Def}
Our goal is to minimize $\scal{a}{b}$ in terms of the weights of each weight functions, $\omega_a$ and $\omega_b$.
\begin{Def}
\label{gammadef}
We call a $(\gamma_i(\omega))_{i=0}^\infty\subset \mathbb{R}_{\ge1}\cup\{0\}$ series the resolution of $\omega$ if
\begin{align}
\gamma_i(\omega)&\ge p\gamma_{i+1}(\omega)\label{gammacsokken}\\
\omega&=\sum_{i=0}^\infty \gamma_i(\omega)\label{gammaosszeg}
\end{align}
Then the lexicographically minimal is called the minimal resolution of $\omega$ and it is denote by $\widetilde{\gamma}_i$.
\\Furhetmore, a $(\gamma_i(\omega))_{i=0}^\infty\subset \mathbb{Z}_{\ge0}$ series is an integral resolution of $\omega$ and similarly, the lexicographically minimal is called the minimal integral resolution and is denoted by $\widehat{\gamma}_i$.
\end{Def}
Note that $\omega\ge \gamma_i(\omega)$ and there exists and index $N$ such that $\gamma_i(\omega)=0$ for every $i\ge N$. Therefor the set of resolutions is compact and it has a lexicographically minimal element. Similarly, there exists a minimal integral resolution for every $\omega$.
\begin{lemma}
\label{realgammadef}
Let $k=\lfloor\log_p((p-1)\omega+1)\rfloor-1$, then the minimal resolution is
$$\widetilde{\gamma}_i(\omega)=p^{-i}\frac{p-1}{p-p^{-k}}\omega$$
where $i\le i\le k$ and $\widetilde{\gamma}_i=0$ otherwise.
\end{lemma}
\begin{proof}
First of all, we will show that there exist $x\in \mathbb{R}$ and $k'$ such that $\widetilde{\gamma}_i=p^{-i}x$ for every $i\le k'$ and $\widetilde{\gamma}_i=0$ otherwise. Suppose the opposite, then there exists an $i$ index for which $\widetilde{\gamma}_i(\omega)> p\widetilde{\gamma}_{i+1}(\omega)\neq 0$. Let $\varepsilon>0$ be a sufficiently little real number such that $\widetilde{\gamma}_i(\omega)-\varepsilon\ge p(\widetilde{\gamma}_{i+1}(\omega)+\varepsilon)$. By replacing $\widetilde{\gamma}_i$ with $\widetilde{\gamma}_i-\varepsilon$ and $\widetilde{\gamma}_{i+1}$ with $\widetilde{\gamma}_{i+1}+\varepsilon$ we obtain a new real resolution of $\omega$ which differs form the initial one and is also lexicographically less. This contradicts our assumption, hence
\begin{align*}
\widetilde{\gamma}_i(\omega)=\begin{cases}xp^{-i}&\text{if}~i\le k'\\0&\text{if}~i>k'\end{cases}
\end{align*}
According \ref{gammadef}, $\omega=x\sum_{i=0}^{k'} p^{-i}$ which allows us to determine $x$. On the other hand, $\widetilde{\gamma}_i\ge 1$ for every $i\le k'$, because the range of $\widetilde{\gamma}_i$ is $\mathbb{R}_{\ge 1}\cap {0}$, thus we have $x\ge p^{k'}$ which implies that
\begin{align*}
\omega=x\sum_{i=0}^{k'}p^{-i}=x\frac{p-p^{-k'}}{p-1}&\ge \frac{p^{k'+1}-1}{p-1}
\\(p-1)\omega+1&\ge p^{k'+1}
\\ \log_p((p-1)\omega+1)-1&\ge k'
\\\lfloor\log_p((p-1)\omega+1)\rfloor-1&\ge k'
\end{align*}
In the last inequality we used the fact that $k'$ is an integer. We will show that $\widetilde{\gamma}_{k'}<p+\frac{p-1}{p^{k'+1}-1}$. Again, suppose the opposite and let $c=\frac{p-1}{p^{k'+1}-1}$. We define the following real resolution of $\omega$:
\begin{align*}
\widetilde{\gamma}_i'=\begin{cases}
\widetilde{\gamma}_i(\omega)-cp^{k'-i}&\text{for every}~ 0\le i\le k'
\\c\sum_{j=0}^{k'}p^{k'-i}=1&\text{if}~i=k'+1
\\0&\text{if}~ i>k'+1
\end{cases}
\end{align*}
It is easy to check that $\widetilde{\gamma}_i'$ is indeed a resolution of $\omega$ and it is also lexicographically smaller than $\widetilde{\gamma}_i$ which contradicts our assumption. Hence, we have
\begin{align*}
\omega=x\frac{p-p^{-k'}}{p-1}&< \left(p+\frac{p-1}{p^{k'+1}-1}\right)\frac{p^{k'+1}-1}{p-1}=\frac{p^{k'+2}-1}{p-1}
\\(p-1)\omega+1&< p^{k'+2}
\\ \log_p\left((p-1)\omega+1\right)-2&< k'
\end{align*}
From the last two inequality it follows that $k'=\lfloor\log_p((p-1)\omega+1)\rfloor-1=k$ and $k=k'$. Finally, we can calculate each term of the series:
\begin{align*}
\omega=x\frac{p-p^{-k}}{p-1}\implies x=\omega\frac{p-1}{p-p^{-k}}\implies \widetilde{\gamma}_i=p^{-i}\frac{p-1}{p-p^{-k}}\omega
\end{align*}
where $i\le k$. Note that the definition of $k'$ implies $\widetilde{\gamma}_i=0$ if $i>k$.
\end{proof}
\begin{lemma}
\label{gammalemma}
Let $\gamma$ be either a real or integral minimal resolution, then $\gamma$ satisfies the following properties.
\begin{enumerate}
\item $\gamma_1(\omega)=\gamma_0(\omega-\gamma_0(\omega))$.
\item $\gamma_i(\omega)$ monotonically increasing in $\omega$.
\end{enumerate}
\end{lemma}
\begin{proof}
Note that the series $(\gamma_{i+1}(\omega))_{i=0}^\infty$ is a minimal resolution with $\omega-\gamma_0(\omega)$ weight, therefore $\gamma_1(\omega)=\gamma_0(\omega-\gamma_0(\omega))$.
For the second property suppose the opposite. More precisely, let $\omega_1>\omega_2$ be two weight such that $\gamma_0(\omega_1)<\gamma_0(\omega_2)$. By reducing $\gamma(\omega_1)$ to a $\gamma'$ resolution with $\omega_2$ weight, we obtain a lexicographically smaller resolution than $(\gamma_i(\omega_2))_{i=0}^\infty$, which contradicts the fact that $\omega_i$ is a minimal resolution. Let $k$ be the smallest non-negative integer such that $\sum_{i=0}^k \gamma_i(\omega_1) \ge \omega_2$, then we can construct $\gamma'$ as
\begin{align*}
\gamma_i' = \begin{cases}
\gamma_i(\omega_1)~&\text{if}~i<k
\\\omega_2-\sum_{j=0}^{i-1}\gamma_i(\omega_1)~&\text{if}~i=k
\\0~&\text{if}~ i>k
\end{cases}
\end{align*}
Furthermore, assume that $\omega_1-\omega_2<\gamma_0(\omega_1)-\gamma_0(\omega_2)$, then we can define the following resolution
$$\gamma_i'=\begin{cases}\gamma_0(\omega_2)+\omega_1-\omega_2&\text{if}~i=0\\\gamma_i(\omega_2)&\text{if}~i>0\end{cases}$$
Due to our assumption, $\gamma_i'$ is lexicographically smaller than $\gamma_i(\omega_1)$, which is a contradiction. So, we showed that both $\gamma_0(\omega)$ and $\omega-\gamma_0(\omega)$ are monotonically increasing in $\omega$.
Let $\omega^{(k)}$ denote the weight of $(\gamma_i(\omega))_{i=k+1}^\infty$. Note that $\omega^{(0)} = \omega-\gamma_0(\omega)$. We will prove by induction that $\omega^{(k)}$ and $\gamma_k(\omega)$ are monotonically increasing in $\omega$. First, we have already established the base case, when $k=0$. Then, assume that the statement holds for every $k<K$. The $(\gamma_i(\omega))_{i=K}^\infty$ is also a minimal resolution, thus $\gamma_0(\omega^{(K-1)}) =\gamma_K(\omega)$ and $\omega^{(K)} = \omega^{(K-1)}-\gamma_0(\omega^{(K-1)})$. We know that $\gamma_0(\omega')$ and $\omega'-\gamma_0(\omega')$ are monotonically increasing in $\omega'$ and $\omega^{(K-1)}$ is also monotonically increasing in $\omega$, consequently $\gamma_0(\omega^{(K-1)})=\gamma_K(\omega)$ and $\omega^{(K-1)}-\gamma_0(\omega^{(K-1)}) = \omega^{(K)}$ are monotonically increasing. Hence we proved the induction step.
\end{proof}
\begin{lemma}
\label{forind}
Let $\gamma$ be an integral or a real minimal resolution and $\omega<\sum_{i=0}^k p^i$, then we have $\omega-\gamma_0(\omega)<\sum_{i=0}^{k-1} p^i$.
\end{lemma}
\begin{proof}
Let $k$ be the minimal integer that satisfies the condition of the lemma, then $k$ is the same defined in \ref{realgammadef}. First, we prove the lemma for real minimal $\gamma$. Due to \ref{realgammadef}, we have $\widetilde{\gamma}_0(\omega)=\frac{p-1}{p-p^{-k}}\omega$. By using $\omega<\sum_{i=0}^k p^i$, we get
\begin{align*}
\omega-\widetilde{\gamma}_0(\omega)<\left(1-\frac{p-1}{p-p^{-k}}\right)\sum_{i=0}^{k} p^i=\frac{1-p^{-k}}{p-p^{-k}}\frac{p^{k+1}-1}{p-1}=\sum_{i=0}^{k-1}p^i
\end{align*}
We proved the real case. Now, we can move on to the integral case. Since every integral resolution is also a real resolution $\widetilde{\gamma}$ has to be lexicographically smaller than $\widehat{\gamma}$. The latter one implies that $\widehat{\gamma}_0(\omega)\ge \widetilde{\gamma}_0(\omega)\Rightarrow \omega-\widehat{\gamma}_0(\omega)\le \omega-\widetilde{\gamma}_0(\omega)$.
\end{proof}
\begin{lemma}
\label{rootest}
Let $a$ be a real (or integral) weight function with $\omega_a$ weight and $\gamma$ real (or integral) minimal resolution of $\omega_0$. Then $a(\delta)\ge \gamma_0(\omega_a)$.
\end{lemma}
\begin{proof}
We start with the real case. For every $v\in V$ there exists an $u\in N_v$ such that $a(u)\le \frac{a(v)}{p}$, otherwise it follows from the definition that $a(v)\ge \sum_{u\in N_v} a(u)> |N_v|\frac{a(v)}{p}=p\frac{a(v)}{p}$, which is clearly a contradiction. Therefore, we have an infinity path from the root $T=(\delta=u_0, u_1, \dots)$ with the property that for every $u_k\in T$
$$a(u_k)\le \frac{a(u_{k-1})}{p}\le \dots \le \frac{u(\delta)}{p^k}$$
Let $k'$ denote the minimum index such that $a(u_{k'+1})=0$. From the definition of real-valued weight functions, we have
$$\omega_a \le \sum_{u\in T} a(u)\le \sum_{i=0}^{k'} \frac{a(\delta)}{p^i}$$
Let us define the $\gamma_i'=\frac{a(\delta)}{p^i}$ series, which is clearly a resolution. If $\omega'$ is the weight of $\gamma'$, then the last inequality means that $\omega_a\le \omega'$. Since $\widetilde{\gamma}(\omega')$ lexicographically smaller than $\gamma'$ (Because $\widehat{\gamma}(\omega')$ was defined to be minimal.), we also have $a(\delta)=\gamma_0'\ge\widehat{\gamma}_0(\omega')$. Finally, \ref{gammalemma} implies $a(\delta)\ge \widetilde{\gamma}_0(\omega')\ge \widetilde{\gamma}_0(\omega_a)$ because $\omega'\ge \omega_a$.
Now, we move on to the integral case. Due to the fact that every element is an integer, we have a slightly stronger inequality, namely $a(u)\le \left\lfloor\frac{a(\delta)}{p^k}\right\rfloor$. Similarly to the real case, this implies
$$\omega_a\le \sum_{v\in T}a(v)\le \sum_{i=0}^\infty \left\lfloor\frac{a(\delta)}{p^k}\right\rfloor$$
Once again lets take $\gamma_i'=\left\lfloor\frac{a(\delta)}{p^k}\right\rfloor$. It is easy to see that $\gamma_i'$ is an integral resolution and therefore $\omega'\ge \omega_a$ (where $\omega'$ is the weight of $\gamma'$). Since $\widehat{\gamma}_0(\omega')$ is lexicographically minimal and monotonically increasing in $\omega'$, we get that $a(\delta)=\gamma_0\ge \widehat{\gamma}_0(\omega')\ge \widehat{\gamma}_0(\omega_a)$.
\end{proof}
\begin{Th}
\label{mainoptth}
Let $a,b$ be real (or integral) weight functions and $\gamma$ the real (or inegral) minimal resolution, then we have the following inequality
$$\scal{a}{b}\ge \sum_{i=0}^{\infty}p^i\gamma_i(\omega_a)\gamma_i(\omega_b)$$
\end{Th}
\begin{proof}
Assume that $\omega_a<\sum_{i=0}^k p^i$, $\omega_a\le\omega_b$, where $k\in\mathbb{Z}_{>0}$. We apply induction on $k$. First, we show the base case, when $k=1$. It follows from \ref{forind} that $\gamma_0(\omega_a)>\omega_a-1$. We also know that $\gamma_i(\omega_a)=0$ if $i\ge 1$, otherwise $\gamma_j(\omega_a)\ge$ for some $j\ge 1$, and $\omega_a =\sum_{i=0}^\infty \gamma_i(\omega_a) \ge \gamma_0(\omega) +\gamma_j(\omega_a)>\omega_a$, which is clearly contradiction.
Since every term is non-negative, we have $\scal{a}{b}\ge a(\delta)b(\delta)$. By using \ref{rootest} and the fact that $\gamma_i(\omega_a)=0$ if $i\ge 1$, we obtain the following inequality:
$$\scal{a}{b}\ge a(\delta)b(\delta)\ge \gamma_0(\omega_a)\gamma_0(\omega_b)=\sum_{i=0}^{\infty}p^i\gamma_i(\omega_a)\gamma_i(\omega_b) \ .$$
Hence, we proved the theorem for $k=1$.
Let $k>1$ and assume that the statement is true for every integer smaller than $k$. The \ref{forind} lemma states that $\omega_a-\gamma_0(\omega_a)<\sum_{i=0}^{k-1}p^i$. Let $v$ be an arbitrary vertex of the tree ($v\in V$) and let $a^v$ denote the the constraint of the weight function to a sub-tree from the root $v$. (Every ancestor of $v$ is included in the $a^v$ graph.) Moreover, let $\omega_a^v$ denote the weight of $a^v$. This new notion allows us to rewrite the scalar product
$$\scal{a}{b}= a(\delta)b(\delta)+\sum_{v\in N_\delta}\scal{a^v}{b^v}$$
According the induction hypothesis, we already proved the statement for $\scal{a^v}{b^v}$ because $\omega^v_a=\omega_a-\gamma_0(\omega_a)<\sum_{i=0}^{k-1}p^i$. Hence, we obtain the following inequality
\begin{align*}
\scal{a}{b}&\ge a(\delta)b(\delta)+\sum_{v\in N_\delta}\sum_{i=0}^{\infty}p^i\gamma_i(\omega^v_a)\gamma_i(\omega^v_b)\\&=a(\delta)b(\delta)+p\sum_{i=0}^{\infty}p^i\gamma_i(\omega^v_a)\gamma_i(\omega^v_b)
\end{align*}
Recall that $\omega_a^v=\omega_a-a(\delta)$, therefore the right side of the last inequality only depends on $a(\delta)$ and $b(\delta)$. We claim that the right side is minimal if $b(\delta)=\widetilde{\gamma}_0(\omega_b)$. For the sake of contradiction, assume that $b(\delta)>\widetilde{\gamma}_0(\omega_b)$. (Recall that \ref{rootest} states that $b(\delta)\ge\widetilde{\gamma}_0(\omega_b)$.) Let $\varepsilon=b(\delta)-\gamma_0(\omega_b)$. We want to prove that the right side will not increase if we decrease $b(\delta)$ by $\varepsilon$.
\begin{align*}
a(\delta)b(\delta)+p\sum_{i=0}^{\infty}p^i\gamma_i(\omega^v_a)\gamma_i(\omega^v_b) \ge a(\delta)(b(\delta)-\varepsilon)+p\sum_{i=0}^{\infty}p^i\gamma_i(\omega^v_a)\gamma_i(\omega^v_b+\varepsilon)
\end{align*}
Or equivalently,
\begin{align*}
a(\delta)\varepsilon \ge p\sum_{i=0}^\infty p^i\gamma_i(\omega_a^v)\left(\gamma_i(\omega_b^v+\varepsilon)-\gamma_i(\omega_b^v)\right)
\end{align*}
Let $\Delta_i=\gamma_i(\omega_b^v+\varepsilon)-\gamma_i(\omega_b^v)$ denote the difference between each term. (Recall that $\Delta_i\ge 0$ due to \ref{gammalemma} lemma.) Also the \ref{gammadef} definition implies that $\sum \Delta_i=(\omega_b^v+\varepsilon)-\omega_b^v=\varepsilon$. Therefore it is enough to prove that
\begin{align*}
a(\delta)\Delta_i\ge p^{i+1}\gamma_i(\omega_a^v)\Delta_i
\Leftarrow a(\delta)\ge p^{i+1}\gamma_i(\omega_a^v)
\end{align*}
Because we would obtain the desired inequality if we sum them up. Fortunately, these inequalities can be proved with the following chain of inequalities:
\begin{align*}
a(\delta)&\overset{(1)}{\ge} \gamma_0(\omega_a)\overset{(2)}{\ge} p\gamma_1(\omega_a)\overset{(3)}{=}p\gamma_0(\omega_a-\gamma_0(\omega_0))
\\&\overset{(4)}{\ge} p\gamma_0(\omega_a-a(\delta))=p\gamma_0(\omega_a^v)\overset{(2)}{\ge} p^{i+1}\gamma_i(\omega_a^v)
\end{align*}
Where (1) is \ref{rootest}, (2) is \ref{gammadef}.1, (3) is \ref{gammalemma}.1 and (4) follows from \ref{gammalemma}.2 because $a(\delta)\ge \gamma_0(\omega_a)$ due to $\ref{rootest}$. We completed the proof.
\end{proof}
\begin{Cor}
\label{valosegeszsulyfv}
Every real weight function is also integral and therefore we have the following inequality for the $a,b$ integral weight functions
\begin{align*}
\scal{a}{b}\ge \sum_{i=0}^{\infty}p^i\widehat{\gamma}_i(\omega_a)\widehat{\gamma}_i(\omega_b)\ge \sum_{i=0}^{\infty}p^i\widetilde{\gamma}_i(\omega_a)\widetilde{\gamma}_i(\omega_b)
\end{align*}
\end{Cor}
\begin{Cor}
\label{computedrealgamma}
For arbitrary $a$ and $b$ real weight function let $k_c=\lfloor\log_p((p-1)\omega_c+1)\rfloor-1$ for every $c\in\{a,b\}$. Then we have
\begin{align*}
\scal{a}{b}\ge\sum_{i=0}^{\infty}p^i\widetilde{\gamma}_i(\omega_a)\widetilde{\gamma}_i(\omega_b)= \frac{p-1}{p-p^{-\max\{k_a,k_b\}}}\omega_a\omega_b
\end{align*}
\end{Cor}
\begin{proof}
Due to \ref{realgammadef}, we can substitute the values of $\widetilde{\gamma}_i$ into \ref{mainoptth} theorem, which gives us
\begin{align*}
\scal{a}{b}\ge \frac{p-1}{p-p^{-k_a}}\frac{p-1}{p-p^{-k_b}}\omega_a\omega_b\sum_{i=0}^{\min\{k_a,k_b\}}p^{-i}
\end{align*}
where the sum goes only to $\min\{k_a,k_b\}$ because during the proof of \ref{realgammadef} we showed that $\widetilde{\gamma}_i(\omega_a)=0$ for every $k$ index greater than $k_a$. Lastly,
\begin{align*}
\sum_{i=0}^{\min\{k_a,k_b\}}p^{-i} =\frac{p-p^{-\min\{k_a,k_b\}}}{p-1}
\end{align*}
Substituting it back yields the inequality.
\end{proof}
\begin{lemma}
\label{sulyfvbecslese}
Let $\gamma$ be a minimal resolution (real or integral) and introduce $a(v)=\gamma_{d(\delta,v)}(\omega_a)$ weight function with $\omega_a$ weight, where $d(\delta, v)$ denotes the distance of the $v$ vertex from the $\delta$ root. Similarly, we define the $b$ weight function with $\omega_b$ weight. Then the lower bound in \ref{mainoptth} theorem is sharp.
\end{lemma}
\begin{proof}
Since there are $p^i$ vertices $i$ distance from the $\delta$ root. With little calculation we get that
\begin{align*}
\scal{a}{b}=\sum_{v\in V}a(v)b(v)=\sum_{i=0}^\infty p^i\gamma_i(\omega_a)\gamma_i(\omega_b)
\end{align*}
which is exactly the lower bound in \ref{mainoptth} theorem. In conclusion the provided inequality in \ref{mainoptth} is sharp.
\end{proof}
\section{Proof of Main Theorem}
In the last two section we prepared every tool we will need in the proof of the main theorem. In this chapter we will focus on that proof and the corollaries of the main theorem.
\begin{Th}
\label{maintheorem}
Let $f,g$ be monic polynomials with integer coefficients and $r$ resultant such that $v_p(f(n))\ge s_1$ and $v_p(g(n))\ge s_2$ for every integer $n$. Furthermore, let $\gamma$ be the (real or integral) minimal resolution defined in \ref{gammadef}. Then we have the following lower bound for the $p$-adic valuation of the $r$ resultant.
\begin{align*}
v_p(r)\ge p\sum_{i=0}^{\infty}p^i\gamma_i(s_1)\gamma_i(s_2)
\end{align*}
\end{Th}
\begin{proof}
When $r=0$, $v_p(0)=\infty$, thus we can assume that $r\neq 0$.
Recall that the \ref{vprchiint} and \ref{integralszummabecs} lemmas imply the following inequality
\begin{align}
\label{mainsth}
v_p(r)\ge \sum_{t=1}^\infty \sum_{m=1}^{p^t}\kkar{f}\kkar{g}
\end{align}
Take $p$ instances of the graph defined in the 3rd section and let $\mathcal{F}_0,\mathcal{F}_1,\dots,\mathcal{F}_{p-1}$ denote these instances. We identify the vertices of $\mathcal{F}_k$ by the $(l,m)$ pairs ($l\in \mathbb{Z}_{\ge1}$) with the following rule.
Let $\varphi_i:\mathbb{Z}/p^{i+1}\mathbb{Z}\rightarrow\mathbb{Z}/p^i\mathbb{Z}$ be the natural projection and note that $k\xrightarrow{\varphi_1^{-1}}D_1\xrightarrow{\varphi_2^{-1}}D_2\rightarrow\dots$ determines a similar graph structure as $\mathcal{F}_k$, where $D_i=\{x\in \mathbb{Z}/p^{i+1}\mathbb{Z}:x\equiv k~(\text{mod}~p)\}$.
More precisely, if $v\in \mathcal{F}_k$ then we identified $v$ with the pair $(l, m)$, where $m\in D_{d(\delta,v)}$, $l=d(\delta, v)$ and the end points of out-edges are $N_v=\{(l+1, m') \mid m'\in D_{l+1},~ \varphi_{l+1}(m')=m\} $. (Recall that $d(\delta,v)$ is the distance between $\delta$ and $v$. Furthermore, it is easy to see that $|N_v|=p$.)
We claim that $a(t,m)=\kkar{f}$ is an (integral) weight function. First of all, the weight of $a$ is $s_1$ because
$$\sum_{t=1}^\infty \kkar{f}=v_p(f(m))\ge s_1$$
for arbitrary $m\in\mathbb{Z}$. We also need to mention that due to $\ref{divisonlemma}$ lemma, the \ref{abdef}.2 criteria is also satisfied. Similarly, $b(t,m)=\kkar{g}$ is weight function with $s_2$ weight.
Notice that the scalar product of $a$ and $b$ on $\mathcal{F}_k$ is
\begin{align*}
\scal{a}{b}=\sum_{v\in \mathcal{F}_k} a(v)b(v)=\sum_{t=1}^\infty \sum_{\substack{m\equiv k~(p)\\ m\in\mathbb{Z}/p^{t+1}}}\kkar{f}\kkar{g}
\end{align*}
If we add these up for $0\le k<p$, we get the right side of \ref{mainsth}. Therefore, within in the meaning of \ref{mainoptth} theorem we get that
\begin{align*}
v_p(r)\ge p\sum_{i=0}^{\infty}p^i\gamma_i(\omega_a)\gamma_i(\omega_b)\ .
\end{align*}
Hence, we finished the proof.
\end{proof}
Initially, the authors of the \cite{frenkel2018estimating} paper considered a problem setting, where $s_1=s_2$ and $S=\max_{n\in\mathbb{Z}}\{\min{v_p(f(n)), v_p(g(n))}\}$ was also known. This encouraged us to investigate this particular case as well.
\begin{Th}
Let $f,g$ be monic polynomials with integer coefficients and $r$ resultant such that $v_p(f(n))\ge s_1$, $v_p(g(n))\ge s_2$ for every integer $n$ and
$$S=\max_{n\in\mathbb{Z}}\min \{v_p(f(n)),v_p(g(n))\}<\infty$$
Let $\gamma$ be the (real or integral) minimal resolution defined in \ref{gammadef}. Then we have
$$v_p(r)\ge S-\max\{s_1, s_2\}+p\sum_{i=0}p^i\gamma_i(s_1)\gamma_i(s_2)$$
\end{Th}
\begin{proof}
First, we need to understand what the statement on $S$ is really about. Since $S$ is not infinity and $\min \{v_p(f(n)),v_p(g(n))\}$ takes only non-negative integer values, there exists an $m_0\in\mathbb{Z}_p$ such that $S=\min \{v_p(f(m_0)),v_p(g(m_0))\}$. Similarly to the proof of theorem \ref{maintheorem}, let $\mathcal{F}_0,\dots \mathcal{F}_{p-1}$ denote the trees and $a,b$ the corresponding weight functions on them derived from $\kar{f}$ and $\kar{g}$.
Let $k\equiv m_0~(\text{mod}~p)$ and $\mathcal{T}$ denote the path in $\mathcal{F}_k$ determined by $m_0$. We can look up the values of the weight functions, $a,b$, along the $\mathcal{T}$ path. Let these number be $\alpha_0, \alpha_1, \dots $ and $\beta_0, \beta_1, ...$. (In other words, $\alpha_i=a(t_i)$, where $\mathcal{T}=(t_0,t_1,\dots)$.)
We already know that $v_p(f(m_0)), v_p(g(m_0))\ge S$, which implies $\sum_i \alpha_i, \sum_i \beta_i \ge S$. We can form a new $\alpha^*_0, \dots \alpha^*_u$ finite series from $(\alpha_i)_{i\in\mathbb{N}}$ which satisfies $\alpha^*_i=\alpha_j$ if $j<u$ and $\sum_{i=0}^k \alpha^*_i=s_1$. (Where we used the fact that $s_1\le S$.) We can do the same thing for the series $(\beta_i)_{i\in\mathbb{N}}$ and obtain $\beta^*_0, \dots \beta^*_v$.
By changing the values of $a$ along the $\mathcal{T}$ path to $\alpha^*_i$ we get a new weight function $a^*$. We claim that it also has weight $s_1$. The first changed value is $\alpha_u$ and every path containing that node already has $\sum_{i=0}^u \alpha^*_i=s_1$ weight while the other path remained untouched. So, it has at least $s_1$ weight. We should also check the 2nd constraint in Definition \ref{abdef}. Unfortunately, it is not necessary true; however, if we set $a^*(v)$ to $0$ for every $v$ ancestor of $t_u$ (where $\alpha_u$ is located), the weight of $a^*$ is still $s_1$ and the 2nd constraint also holds. Again, we can do the same thing with $b$ and obtain $b^*$.
Let $\mu,\nu$ denote the indexes of the first zeros in $\alpha$ and $\beta$. (Possibly infinity.) If $\mu\le \nu$, then we have the following estimate.
\begin{align*}
\scal{a}{b}&= \scal{a^*}{b} +\sum_{i=0}^\mu (\alpha_i-\alpha^*_i)\beta_i
\\&\ge \scal{a^*}{b^*}+S-s_1
\end{align*}
Where we used that $b^*\le b \Rightarrow \scal{a^*}{b^*}\le \scal{a^*}{b}$ and $\mu\le \nu$ implying that $\beta_i\ge 1$ for every $i\le \mu$, which means that
$\sum_{i=0}^\mu (\alpha_i-\alpha^*_i)\beta_i\ge \sum_{i=0}^\mu (\alpha_i-\alpha^*_i) \ge S-s_1$.
We can do the same thing if $\mu\ge \nu$. Combining the two together, we get that
$$\scal{a}{b}\ge \scal{a^*}{b^*} +\min\{S-s_1,S-s_2\}$$
Since $a^*,b^*$ are weight functions with $s_1, s_2$ weights, we can apply Theorem \ref{mainoptth} and get
$$\scal{a}{b}\ge S-\max\{s_1, s_2\}+\sum_{i=0}p^i\gamma_i(s)$$
In the other trees we don't have any new constraint compared to Theorem \ref{maintheorem}, thus our final lower bound is
$$v_p(r)\ge S-\max\{s_1, s_2\}+p\sum_{i=0}p^i\gamma_i(s)$$
\end{proof}
If we substitute our computations from \ref{computedrealgamma} when $\gamma$ is the real minimal resolution we get the following corollaries
\begin{Cor}
Let $f,g$ be monic polynomials with integer coefficients and $r$ resultant such that $v_p(f(n))\ge s_1$, $v_p(g(n))\ge s_2$ for every integer $n$ and
$$S=\max_{n\in\mathbb{Z}}\min \{v_p(f(n)),v_p(g(n))\}<\infty$$
Then we have the following lower bound for $v_p(r)$.
\begin{align*}
v_p(r)\ge S-\max\{s_1,s_2\}+ ps_1s_2\frac{p-1}{p-p^{-k}}
\end{align*}
Where $k=\lfloor \log_p((p-1)\max\{s_1,s_2\}+1)\rfloor -1$.
\end{Cor}
\begin{Cor}
Let $f,g$ be monic polynomials with integer coefficients and $r$ resultant such that $v_p(f(n))\ge s$, $v_p(g(n))\ge s$ for every integer $n$ and
$$S=\max_{n\in\mathbb{Z}}\min \{v_p(f(n)),v_p(g(n))\}<\infty$$
We have
\begin{align*}
v_p(r)\ge S-s+ ps^2\frac{p-1}{p-p^{1-\lfloor \log_p((p-1)s+1)\rfloor}}
\end{align*}
\end{Cor}
\section{Constructions}
In this section we discuss the sharpness of the provided lower bounds from the previous section. We show that the estimations from the $\gamma$ real minimal resolution give us the best lower bound infinitely many times.
\begin{lemma}
Let $\log_p((p-1)s_1+1)-1, k_2=\log_p((p-1)s_2+1)-1$ integers and $s_1\ge s_2 \Leftrightarrow k_1\ge k_2$. We define $f,g$ polynomials to be
\begin{align*}
f(x)=\prod_{t=0}^{p-1}h(x+t)
\\g(x)=\prod_{t=0}^{p^{k_2+1}-1}(x+t)
\end{align*}
where $h$ is a monic irreducible polynomials over $\mathbb{Q}_p$ with degree $s_1$ such that $v_p(h(0))=s_1$ and $v_p(h(n))=0$ for every other $n$ that is not divisible by $p$. This two polynomial satisfy $v_p(f(n))\ge s_1$ and $v_p(g(n))\ge s_2$ for every $n$ integer and their resultant is
\begin{align*}
v_p(r)=ps_1s_2\frac{p-1}{p-p^{-k_1}}
\end{align*}
(Which is the provided lower bound in Theorem \ref{maintheorem}.)
\end{lemma}
\begin{proof}
First of all, we prove the existence of the polynomial $h$. It is well known that there exists irreducible polynomial over $\mathbb{F}_p$ with degree $s_1$. Let $h_0$ be such polynomial. We claim that $h=p^{s_1}h_0(\frac{x}{p})$ is good choice. It was irreducible over $\mathbb{F}_p$, so it is also irreducible over $\mathbb{Q}_p$. On the other hand, $v_p(h(0))=s_1$ because $h_0(0)\equiv 0~(\text{mod}~p)$ would mean that $h_0$ is not irreducible. Meanwhile, $h(x)\equiv cx^{s_1} ~(\text{mod}~p)$, where $c$ is the non-zero leading coefficient of $h_0$. Therefore, $v_p(h(n))=0$ for every other $n$ that is not divisible by zero.
We can factor $h$ over its splitting field and write $h=\prod_{i=1}^{s_1}(x-\alpha_i)$. Due to the irreducibility of $h$, we have $v_p(n-\alpha_i)=\frac{1}{s_1}v_p(h(n))$ or equivalently \begin{align*}
v_p(n-\alpha_i)=\begin{cases}1&\text{if}~p|n\\ 0&\text{otherwise}\end{cases}
\end{align*}
Hence, the resultant is
\begin{align*}
v_p(r)=\sum_{t=0}^{p^{k_2+1}-1}\sum_{i=1}^{s_1}v_p(t-\alpha_i)=p^{k_2+1}s_1
\end{align*}
but $s_2=\frac{p^{k_2+1}-1}{p-1}$ so
\begin{align*}
v_p(r)=ps_1s_2p^{k_2}\frac{p-1}{p^{k_2+1}-1}=ps_1s_2\frac{p-1}{p-p^{-k_2}}
\end{align*}
Last of all, we need to check that $v_p(f(n))\ge s_1$ and $v_p(g(n))\ge s_2$ holds for every $n\in \mathbb{Z}$.
\\In the case of $g$, we already showed that for an arbitrary $n\in\mathbb{Z}$, we have
$$v_p(g(n))=\sum_{t=0}^{p-1}\sum_{i=1}^{s_1} v_p(n+t-\alpha_i)$$
There exists $t_0\in\{0,1\dots p-1\}$ such that $p| n+t$. As we had showed earlier $v_p(n+t-\alpha_i)=1$, which means $v_p(g(n))=s_1$.
\\In the case of $f$, for a given $n\in \mathbb{Z}$ there are precisely $p^{k_2}$ integers $t\in[0,p^{k_2+1}-1]$ such that $v_p(n+t)\ge 1$. In general, there are $p^{k_2-l}$ integers $t$ for which $v_p(n+t)\ge l+1$. This gives us that
\begin{align*}
v_p(f(n))&=\sum_{l=0}^{k_2}p^{k_2-l}=\frac{p^{k_2+1}-1}{p-1}
\\&=\frac{p^{\log_p((p-1)s_2+1)}-1}{p-1}=s_2
\end{align*}
where we substituted the definition of $k_2=\log_p((p-1)s_2+1)-1$.
We proved the lemma.
\end{proof}
\bibliographystyle{unsrt}
|
\section{Introduction}
Consider the typical univariate regression problem where the intention is to find the relationship between $x$ and $y$. When extended to time series, a number of time specific complexities arise. For the specific case where the input $x$ affects the output $y$, with a random time delay in between, the estimated coefficients (or weights, predictions etc) are significantly attenuated \cite{spearman}. This attenuation occurs both in traditional statistical models and machine learning models, and for certain problems can be a serious limitation. The following document proposes techniques for handling this class of time series regression.
As an example, let us take the management of blood sugar level in diabetes patients. In people with diabetes, the pancreas can't effectively regulate blood sugar levels. Therefore, these levels must be controlled by insulin injections and a special diet. The challenge for many people is that the relationship between the input (insulin) and output (blood sugar) is extremely complex \cite{blood_glucose}. The effect of the insulin injection might be observed after $15$, $20$ or $t$ minutes depending on a number of factors, many of which are unknown. Due to the stochastic nature of the time delays, the actual effect can't be easily determined. It is difficult to differentiate the effect of the insulin injection from other factors and accurately determine how much one should take.
Inference for this type of problem is especially challenging. Typical regression models require a fixed alignment between cause and effect. Using standard methods, we'd need to assume that the effect occurs after some fixed time $t$ which can be inferred from the data. However, if there is any uncertainty in the parameter $t$ ($t$ changes or is noisy) the resulting estimates will be significantly attenuated.
Consider the simple example in Figure 1 where the effect of the input is $1$. The observed input is given by the red line, the blue line is when the effect actually occurs. The first effect happens one time point after the input. The second effect happens at the same time as the input. A fixed time delay isn't valid in this case because the time shifts differ.
\begin{center}
\begin{figure}
\includegraphics[totalheight=6cm]{2_input_delay_example.png}
\caption{Stochastic Time Delay Example}
\label{fig:verticalcell}
\end{figure}
\end{center}
If one were to model the effect using a fixed time delay and OLS, the estimate would only be half the true value because only one of the outputs is aligned. Obviously, this isn't ideal, we want the parameter estimates to be as close to the real values as possible, regardless of any noise in the lag structures. One can mitigate the problem via time aggregations, however in complicated cases with multiple factors, this is just not feasible.
Therefore, we propose a regression model which can handle stochastic time delay structures. We treat the stochastic time delay components as an ‘error’ in the time axis. Next, we find the maximum likelihood estimate, for a given set of parameters, considering the time error ($t$-axis) and the regression error ($y$-axis) simultaneously.\\
An open source Python implementation is available on \href{https://github.com/aaron1rcl/tvs_regression/}{GitHub}.
\section{Related Work}
There is extensive existing literature on time series problems with time delay dependencies. In the statistical disciplines there a number of treatments which focus on
fixed time delay dependencies. For example, Granger Causality \cite{granger} is used to determine whether one time series is useful for forecasting another, across some fixed delay.
Distributed Lag \cite{almon} and Dynamic Regression models (e.g \cite{dynr}) are able to handle linear and non-linear cause and effect relations that occur across multiple time lags.
Nonetheless, these models assume a fixed time lag dependency.
In the machine learning literature, there are a number of models which can handle complex dependencies across time. Sequence models such as the RNN, LSTM and GRU \cite{ml}
are able to generate predictions which incorporate time delays between input and the output variables. More recently, attention based models such as the Transformer have become the state-of-the-art for a
variety of tasks, including time series forecasting \cite{deep_transformers}. Yet, all of these models have a tacit assumption of a fixed time lag dependency. There is also Dynamic Time Warping (DTW) \cite{dtw} which tries to find the optimal alignment between time series sequences by minimizing the
distance between the respective inputs. DTW is able to model varying time lag dependencies but is not purposed for regression, instead being mainly used for pattern matching and sequence alignment.
The field of system identification also has considerable literature on dynamic non-linear systems. For example, the NARX \cite{ml} model can be used to identify
non-linear systems with fixed time delays between input and output. At the same time, there are also treatments on systems with disturbances in the input \cite{eiv_sysid}, known as EIV systems. To our knowledge,
these works do not deal with uncertainty in the time domain.
Finally, recent papers such as Dynamic Time Lag Regression \cite{dtlr} and Variable-Lag Granger Causality \cite{Amornbunchornvej_2019} deal with non-stationary time lag dependencies. In other words, the lag structure is assumed to evolve over time.
In this paper, we deal with the specific case of \emph{stochastic time delays} i.e. time delays which vary randomly. There appears to be little existing research on this topic.
\section{Methodology}
We consider the problem as analogous to the typical error-in-variables (EIV) regression. Ordinary regression analyses (and machine learning models) define the loss function with respect to errors in the $y$ axis only. For EIV, errors are considered in both the $y$-axis and the $x$-axis \cite{deming}. This is useful when there are measurement errors in the independent variable e.g. because the physical measurements have some degree of random error. Similarly, for this problem we assume that we have errors in the $y$-axis and the $t$-axis. That is, there are random prediction errors and random errors in the time domain.
\subsection{Model Specification}
Let $x(t)$ be a discrete time series. We want to determine the functional relationship $f:\mathbb{R} \longrightarrow \mathbb{R}$ to some other (target) variable $y$ on observed data $\{y_i\}_{i=1}^{n}$, that is
$y = f(x(t))$. The series $x(t)$ must be stationary with a known value for the point at which $f(x(t)) = 0$. That is, we know at which point the input series has no effect on the output (when it’s off). The size of the support of $f(x(t))$ should be small relative to number of points in the domain (we we will explore this condition further in 'Limitations'). We utilise the terminology {\em impulse} for an individual element of the support of $f$.
Firstly, let us take the input series $x(t)$ and decompose it into its constituent non-zero impulse components. So, if $x(t)$ is a vector given by
\begin{align*}
x(t) =
\left(
\begin{array}{cccccc}
0 & 0 & 1 & 0 & 1 & 0
\end{array}
\right)
\end{align*}
then we decompose the vector into a matrix
\begin{align*}
X(t) =
\left(
\begin{array}{cccccc}
0 & 0 & 1 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 1 & 0
\end{array}
\right)
\end{align*}
where each impulse is treated separately. \\
Given that each non-zero impulse (row) of $X(t)$ is affected by a random time delay (denoted $\tau$), we then model:
\begin{align}\label{Eqn:ModelSpecification}
y(t)= f(1^TX(t + \mathcal{T})) + \varepsilon
\end{align}
where $X(t + \mathcal{T})$ is the matrix of time shifted impulses, $X(t)$ is the observed input, $\mathcal{T}$ is the set of single impulse time delays ($\tau$'s), $1$ is the vector $(1,1,\cdots, 1)\in
\mathbb{R}^{1 \times k}$, where $k$ is the number of impulses and $\varepsilon$ is a noise term.
The model definition represents the application of time shifts to the matrix rows and a subsequent reduction by summation over the columns.
We also assume that $\mathcal{T}$ is not a constant, but rather a random draw from some discrete distribution (e.g. discrete gaussian, poisson etc). For the discrete gaussian kernel $\mathcal{T}\sim T(\mu, \sigma)$ or for the poisson distribution $\mathcal{T}\sim Pois(\lambda)$. Similarly one can model the errors as $\varepsilon \sim N(\mu, \sigma_{\varepsilon})$.
\subsection{Inference}
In order to find the best estimate of $f$, we would like to find the function $f$ which maximises the joint log-likelihood of the time-domain shifts and the prediction residuals. Specifically, we maximise:
\begin{align}
\mathcal{L} =
\underbrace{\sum_{i=1}^{n} \log(p(y_i | X_{\tau_i} = x_i; \theta_{f}, \tau_i))}_{\mathcal{L}_1}
+
\underbrace{\sum_{i=1}^{n} \log(p(\tau_i| \theta_{\mathcal{T}}))}_{\mathcal{L}_2}
\end{align}
where $\theta_{f}$ and $\theta_{\mathcal{T}}$ represent the parameters of the model $f$ and time shifts $\mathcal{T}$ respectively. In other words, the $\mathcal{L}_2$ term represents likelihood of time shifts and the $\mathcal{L}_1$ term represents the likelihood of the prediction residuals. We maximise these terms simultaneously. For simplicity, we assume that the time shift distribution and error distributions are independent. \\
We will refer to this algorithm as {\em Time Varying Stochastic (TVS) Regression}.
\section{Algorithm}
Before optimisation, the values of each individual time shift $\tau$ are not known. In addition, the prediction errors $\varepsilon$ can only be calculated if each value of $\tau$ is available (because for each time shift there is a different prediction and hence prediction error). Therefore, our algorithm finds the optimum set of time shifts in an inner optimisation ($\mathcal{L}_2$), while iteratively searching for the optimum parameters of $f$ in an outer optimisation loop ($\mathcal{L}_1$).
Firstly, we define a parametric form $f(x;\theta_f)$ and some initial parameters to be estimated for the function $f$.
\begin{example}\label{example:poisson}
Let's take the univariate linear model with gaussian errors, where the $f(x(t))$ is parameterised by $\beta$ and $\sigma_{\varepsilon}$ (error standard deviation). In addition, we choose a poisson distribution for $\mathcal{T}$, where $\theta_{\mathcal{T}}$ is a parameterised by its mean $\lambda_{\mathcal{T}}$. The choice of a poisson distribution for $\mathcal{T}$
ensures discrete, positive time shifts only.
\end{example}
First, let us initialise some starting values for each of these parameters. Now, we want to find the best possible time shift $\tau$ for each input impulse in $X(t)$. It stands to reason that the best possible time shift would be one that is not too distant from the observed impulse (i.e. has a high likelihood given some distribution) and also produces the best possible prediction. From there, the likelihood estimate is derived from the time shift and the prediction error.
Finally, we iterate over a number of values of $\tau$ optimising until we maximise the likelihood for the specific impulse.
\begin{remark}
However, we must also consider that the impulses in $X(t)$ are not independent from each other. After shifting, it's possible that two or more effects occur simultaneously. This is particularly problematic if there are multiple impulses within a short period of time or impulses have a distributed effect over multiple time points. As an example consider the series
\begin{align*}
x(t) =
\left(
\begin{array}{cccccc}
0 & 0 & 1 & 1 & 0 & 0
\end{array}
\right)
\end{align*}
with $\mathcal{T}= (1, 0)$ and $\beta = 1$. For this case,
\begin{align*}
X(t + \mathcal{T}) =
\left(
\begin{array}{cccccc}
0 & 0 & 1 & 0 & 0 & 0 \\
0 & 0 & 1 & 0 & 0 & 0
\end{array}
\right)
\end{align*}
and the effect is therefore
\begin{align*}
y =
\left(
\begin{array}{cccccc}
0 & 0 & 0 & 2 & 0 & 0
\end{array}
\right)
\end{align*}
\end{remark}
To accurately calculate the likelihood, we must optimise the time shifts simultaneously. Therefore, we treat the problem of finding the best set of time shifts as a discrete optimisation problem. For the optimisation step we utilise two assumptions.
\begin{enumerate}
\item Firstly, smaller shifts are more likely than larger shifts (proportionate to the dispersion of the $\mathcal{T}$ distribution). The algorithm should explore the space of smaller shifts more often than larger shifts.
\item Secondly, impulses close to each other are more likely to be dependent than impulses further away.
\end{enumerate}
Accordingly, the optimisation procedure is:
\begin{enumerate}
\item [(1)] Initialise the set of parameters $\theta_{f}$ and $\theta_{\mathcal{T}}$ for the function $f$.
\item [(2)] Find the $\mathcal{T}$ which maximises the likelihood for the given set of parameters:
\begin{enumerate}
\item[(i)] Initialize the discrete optimisation algorithm time shifts ($\mathcal{T}$).
\item[(ii)] Next, randomly select a small number of impulses $m$, with the value of $m$ treated as a hyperparameter. For each impulse, a random time shift is drawn from the $\mathcal{T}$ distribution creating a proposal vector.
\item [(iii)]The likelihood (both time shift and prediction error) for the proposal is calculated.
\item[(iv)] If the proposal likelihood is higher than the current maximum likelihood, our estimate is updated.
\item [(v)] Return to (i) and repeat. The best estimate of the set of $\mathcal{T}$ improves each iteration. The number of iterations ($N$) is also treated as a hyperparameter.
\end{enumerate}
\item [(3)] Optimise the model parameters $\theta_{f}$ and $\theta_{\mathcal{T}}$.
\end{enumerate}
\begin{remark}
In Example \ref{example:poisson} we recommend to initialize the discrete optimisation algorithm with all parameters set to the mean of the $\mathcal{T}$ distribution ($\lambda_{\tau}$).
\end{remark}
For the outer parameter optimisation (Step (3)), typical methods such as gradient descent, genetic algorithms and simulated annealing can be used. In our implementation, we have used the differential evolution algorithm \cite{diff_ev} (scipiy.optimize \cite{scipy}) because of its ability to handle noisy objective functions \cite{diff_ev}. In order to improve convergence, we also standardize all input variables to the range (0,1).
\begin{remark}
We also note that the accuracy of the final parameter estimate is relative to the ratio of the $y$-axis error and the effect size $f(x(t))$.
As the ratio of noise $\sigma_{\varepsilon}$ to effect $f(x(t))$ increases, the time shift distribution $\mathcal{T}$ shrinks.
In the event that the mean of $\theta_{\mathcal{T}}$ become 0, the model becomes a ‘fixed lag’ model, and the parameter estimate $\beta$ tends to the standard linear regression coefficient. Therefore, the method provides no guarantee on recovering the exact time shifts, only that the coefficient estimates are equal to or better than their OLS counterparts.
\end{remark}
\section{Limitations}
\subsection{Scaling}
As the length of $x(t)$ increases, more impulses are introduced and the size of the decomposed matrix $X(t)$ also increases. At some point, handling $X(t)$ becomes impractical. To account for this, we assume that distant impulses do not affect each other. Concretely, if the $i_{th}$ and $j_{th}$ impulse (row) in $X(t)$ are far away then their effect vectors are orthogonal. Hence, we can decompose the matrix $X(t)$ into orthogonal blocks on which the inner optimisation can be run in parallel.
\subsection{Constraints}
Finally, a note on problem constraints. The likelihood estimates are dependent on the inner optimisation procedure. There is no guarantee that the global maximum will be found, particularly for sequences with a high density of impulses.
When the density (in time) of non-zero impulses in $x(t)$ is high, the possible solution space of $\mathcal{T}$ grows exponentially.
In such cases, the estimated likelihood is likely to be close to, but not exactly the same as the real value. Therefore, TVS Regression is most appropriate for sparse time series inputs.
\section{Experiment}
The following section demonstrates a simulated univariate example of TVS Regression. Figure \ref{fig:simulated_example} shows the simulated time series. The full code can be found via \href{https://github.com/aaron1rcl/tvs_regression/blob/master/notebooks/1_univariate_example.ipynb}{TVS Regression on GitHub}.
The input signal has $20$ non-zero impulses which have been drawn from the standard normal distribution. The system is 'off' when $x(t) =0$. In other words, when $x(t) = 0$, $f(x(t)) = 0$. The true shift distribution $\mathcal{T}$ is given by $\mathcal{T} \sim Poisson(\lambda = 2)$. There is one value of $\tau$ for each of the 20 impulses. The green line in Figure \ref{fig:simulated_example} represents the shifted series, corresponding to the time at which the effect occurs. The blue line is the output $y$ which includes a small amount of gaussian noise and an intercept. The values for the parameters were arbitrarily selected. The output $y$ is defined by the following equation:
\begin{align*}
y = 2(1^TX(t + \mathcal{T})) + 6.5
\end{align*}
A histogram of the actual distribution of $\mathcal{T}$ is shown in Figure \ref{fig:real_shifts}.
\begin{center}
\begin{figure}
\includegraphics[scale=0.5]{simulated_example.png}
\caption{Simulated Example Data}
\label{fig:simulated_example}
\end{figure}
\end{center}
After fitting the model, we obtain the following results. Figure \ref{fig:tvs_ols_fit} shows the model fit (denoted TVS) and a comparison with standard OLS. Figures \ref{fig:tvs_errors} and Figure \ref{fig:convergence} show the error distribution of the TVS fit and the convergence of the TVS model respectively.
\begin{center}
\begin{figure}
\includegraphics[scale=0.5]{real_shifts.png}
\caption{Real Shift Distribution $\mathcal{T}$.}
\label{fig:real_shifts}
\end{figure}
\end{center}
\begin{center}
\begin{figure}
\includegraphics[scale=0.6]{tvs_fit.png}
\includegraphics[scale=0.3]{ols_fit.png}
\includegraphics[scale=0.3]{scatterplot.png}
\caption{TVS vs OLS model fit.}
\label{fig:tvs_ols_fit}
\end{figure}
\end{center}
\begin{center}
\begin{figure}
\includegraphics[scale=0.5]{tvs_errors.png}
\caption{Fit Residuals}
\label{fig:tvs_errors}
\end{figure}
\end{center}
\begin{center}
\begin{figure}
\includegraphics[scale=0.6]{convergence.png}
\caption{Convergence}
\label{fig:convergence}
\end{figure}
\end{center}
The true values of $\mathcal{T}$ (true shifts) and the estimated values $\mathcal{T}_{est}$ are:
\begin{align*}
\mathcal{T} &=
\left(
\begin{array}{cccccccccccccccccccc}
2 & 1 & 2 & 0 & 2 & 3 & 1 & 0 & 2 & 4 & 1 & 2 & 3 & 2 & 1 & 0 & 0 & 1 & 1 & 0
\end{array}
\right) \\
\mathcal{T}_{est} &=
\left(
\begin{array}{cccccccccccccccccccc}
2 & 1 & 2 & 0 & 2 & 3 & 1 & 1 & 2 & 4 & 1 & 2 & 3 & 2 & 1 & 0 & 0 & 1 & 1 & 0
\end{array}
\right)
\end{align*}
\begin{center}
\begin{table}
\begin{tabular}{|c|c|c|c|}
\hline
Parameter & TVS Regression & OLS Regression & True Value \\
\hline
$\beta$ & 2.09 & 0.50 & 2.00 \\
\hline
$Intercept$ & 6.53 & 6.62 & 6.50 \\
\hline
$\lambda_{\tau}$ & 1.54 & $N/A$ & 1.40 \\
\hline
$\sigma_{\epsilon}$ & 0.20 & 1.03 & 0.20 \\
\hline
\end{tabular}
\caption{ Parameter comparison between TVS and OLS regression against the true value.}
\label{results_table}
\end{table}
\end{center}
As shown in Table \ref{results_table}, the estimated values for $\beta$ and $\sigma_{\varepsilon}$ are significantly improved by taking into account the stochastic time delay noise. The estimated $\beta$ using OLS regression is $0.5$, compared to the true value $2$. Even a small amount of noise in the time axis causes attenuation limiting the usefulness of OLS for these problems. In comparison, the TVS Regression algorithm estimates $\beta$ to be $2.09$, much closer to the true value.\\
While the example is based on simulated data, we believe that the experiment demonstrates clear potential for improvement in the modelling of real world systems with stochastic time delay noise.
\section{Conclusion}
We have proposed a form of regression analysis suited to the modelling of stochastic time delay problems. In addition, we have shown the feasibility of the approach and its performance on simulated data. Our approach allows for consistently improved estimation and prediction when the input is affected by noise in the time domain. To our knowledge, the method is novel for this class of problem.\\
We propose two extensions as future work. The first is to extend the method to multiple regression. We believe that the extension can be built on the same fundamental ideas presented in this document. Next, the model could be extended to include distributed lag structures. A distributed lag structure is where past values of the impulse influence future values of the output \cite{almon}. Through these extensions we can begin to tackle a number of practical problems that are defined by stochastic relationships between the input and output e.g. the blood glucose problem described in the introduction. In addition, we could also begin to apply the model to forecasting problems where the inputs are affected by stochastic time delay.
\bibliographystyle{acm}
|
\section{Summary about the responses}
As suggested by the reviewers of NeurIPS, we have added new experiments and improved explanations to technical details that help the paper's readability. In this letter, we begin by summarising the critiques raised by the reviewers and provide our response to them. We then make our case as to why the improved, more comprehensive version of our paper makes a good candidate for AAAI.
\begin{itemize}
\item \textbf{In traditional NAS, only one network is optimized. In contrast, LFM requires three networks ($W_1$, $W_2$, $V$). How do authors justify that LFM is indeed superior, given the increase in the number of parameters?} The number of parameters of the network constructed during architecture evaluation are the same for all methods in Tables 1,2,3. On CIFAR-10 and CIFAR-100, the network is a stack of 20 copies of the \textit{searched cell}, which has the same number of parameters for all methods in the comparative Tables 1,2,3. The parameters of $W_1$, $W_2$, $V$ are only used during architecture search; after search is finished, they are discarded. Hence, the better performance of LFM methods is solely due to the better architecture searched by it.
We have updated the paper to make this justification more coherent.
\item \textbf{Could the authors elaborate on the advantages of learning a new network ($W_2$) from scratch, instead of adjusting the the weights of the already learned network ($W_1$)? How does the "learning from mistakes" on $W_2$ (through $a_i$) make $W_1$ learn from the mistakes too?} In our end-to-end tri-level framework, mathematically it is impossible to use a single set of weights to perform stage-1 learning and stage-2 learning. Once the weights are optimized at stage-1, they are not free variables any more and cannot be optimized again at stage-2. So we have to use two different sets of weights to make the tri-level optimization framework mathematically correct. We also add experimental results supporting the benefits provided by this second network in Table 7 of the Appendix due to lack of space.
Further, $W_1$ will learn from its previous mistakes and avoid making similar mistakes in an indirect way. After the weights of $W_2$ are trained by correcting the mistakes made by $W_1$, the architecture will be updated accordingly since the gradient of $A$ depends on $W_2$; an updated $A$ will render $W_1$
to change as well since the gradient of $W_1$
depends on $A$ . Along the chain $W_2 \xrightarrow[]{} A \xrightarrow[]{} W_1$
chain, $W_1$ is indirectly influenced by
$W_2$.
We have updated our paper to include this reasoning and experimental support.
\item \textbf{Ablation studies are provided on limited datasets, and studies on $z$ are missing.} We have added more ablation studies on CIFAR-100 (including the ablation on label similarity) in Figure 3 and Figure 4 of the paper.
The performances of our full LFM methods are significantly better than those of ablation settings, including removing $u$ (No $u$), removing $x$ (No $x$), and removing $z$ (No $z$). The results in Figure 3 of the paper demonstrate the effectiveness and necessity of measuring mistakes on validation examples $u$, calculating visual similarity $x$ and label similarity $z$ between training and validation examples as we would expect intuitively to happen.
Hence, the newly added ablations support our method and make the study complete.
\item \textbf{Why is the visual similarity metric a good one? } For visual similarity, we use dot product attention ~\cite{luong-pham-manning:2015:EMNLP}, which has shown broad effectiveness in many applications and is simple to use. We compared with other metrics such as cosine similarity, L2 distance. The results are shown in Figure 4 of the paper. Dot product attention used in our framework works better than the other two metrics.
Hence, the newly added ablations support our method and make the study complete.
\item \textbf{The memory usage is concerning, since a fixed network architecture would be required to be trained twice. Perhaps there are more clever ways of achieving the same result with smaller cost? Please add discussion} We have added discussion on memory usage and preliminary results of future research. Our method requires the use of two learners that have similar learning capabilities, so that one can learn from the mistakes of other. This increases the memory requirements and makes the learning slow compared to the vanilla approaches.
As future work, we explore reducing memory cost during architecture search by parameter-sharing between the three models $W_1$, $W_2$, and $V$. For $W_1$ and $W_2$, we let them share the same convolutional layers but have different classification heads. For $V$, we replace ResNet-18 with $W_1$. As shown in Table 4, the memory cost of our method via parameter sharing (PS) is similar to that of regular DARTS. However, our method achieves lower test error which is not as good as the full LFM model.
A future work direction is to improve memory usage while keeping similar performance as the full LFM.
We hope this discussion on memory usage is insightful on the workings of a tri-level optimization framework as LFM.
\item \textbf{Missing important baselines such as random sampling or random search} We have added the baselines of random search and random sampling in Tables 1 and 2 of the paper. LFM performs better than random search and random sampling methods, which further speaks in favour of our learning algorithm.
\end{itemize}
In this work, we propose LFM that is an end-to-end tri-level framework for neural architecture search.
In the future works section of our paper, we discuss possible methods to reduce the computational complexity. We also provide some preliminary results of methods that reduce computational complexity, albeit at the expense of a slight drop in effectiveness.
Further, since the architecture parameters are just a set of meta parameters for the learner, LFM could also be easily extended to data re-weighting tasks and other applications. It could also be applied to complex tasks such as semantic segmentation.
We reason that
LFM framework is a novel addition to meta learning and NAS literature. The proposed framework is effective and has potential applicability to a variety of machine learning tasks, making it a strong candidate for publication at a well renowned venue such as AAAI.
\section{Detailed responses}
Below are the detailed rebuttals with reviewers.
\subsection{Reviwer-1}
Final rating: 6
\begin{itemize}
\item
\textbf{Question}: "\textbf{(a)} Trade-off between the soundness of modeling and the ease of optimization. \textbf{(b)} Whether the improvement come from the introduced three-level formulation." \\
\textbf{Answer (a)}: The algorithm is adapted from DARTS and works very well in our empirical experiments. One-step gradient-based approximation is broadly adopted in differentiable NAS, meta learning, and other bi-level optimization problems, and is generatively effective. Though using an approximated optimization algorithm, our framework is still a valid tri-level framework: as shown in Equation 6-9 in the supplements, the gradient of the validation loss to \textbf{$A$} involves the product of three gradients: gradient of the validation loss to \textbf{$W_2'$}, gradient of \textbf{$W_2'$} to \textbf{$W_1'$}, gradient of \textbf{$W_1'$} to \textbf{$A$}. The gradient information is propagated from stage 3 to stage 2, then to stage 1. This shows that the tri-level optimization formulation of our framework is well preserved in the approximated algorithm. \\
\textbf{Answer (b)}: In experimental studies, the major difference between our method and other baselines is the tri-level optimization formulation. This demonstrates that the performance improvement achieved by our method over baselines is due to the tri-level optimization framework.
\item
\textbf{Question}: "Understand the solution through analyzing the learned model parameters". \\
\textbf{Answer}: We selected the top-50 largest values in $r$ and checked the corresponding validation images. These images are visually obscure and it is difficult to judge which classes these images belong to. This indicates that our method pays a lot of attention to hard examples where incorrect predictions tend to happen.
\item
\textbf{Question}: "Give more experimental results about the usage of GPU memory". \\
\textbf{Answer}: To reduce memory cost during architecture search, we can do parameter-sharing between the three models \textbf{$W_1$}, \textbf{$W_2$}, and \textbf{$V$}. For \textbf{$W_1$} and \textbf{$W_2$}, we let them share the same convolutional layers, but have different classification heads. For \textbf{$V$}, we change it from ResNet-18 to \textbf{$W_1$}. The results are shown in Table 4 of the paper. Via parameter sharing (PS), the memory cost of our method is similar to that of regular Darts, but our method achieves lower test error. During architecture evaluation, these methods have similar GPU memory since they follow the same evaluation settings.
\end{itemize}
\subsection{Reviwer-2}
Final rating: 6
\begin{itemize}
\item
\textbf{Question}: "Why does it make sense to calculate all three of validation performance, and visual similarity and label similarity with the training set?" \\
\textbf{Answer}: In Line 104-108 in the NeurIPS's paper, we explained why the three factors are needed during reweighting. The overall rational is: if a training example is similar (both visually and semantically) to many incorrectly-predicted validation examples, then this training example should be paid more attention to in the next round of training. To identify which validation examples are incorrectly predicted, we measure the validation performance. To identify which training examples are visually and semantically similar to the incorrectly-predicted validation examples, we measure the distance between training and validation examples on image features and class labels.
\item
\textbf{Question}: "Why is the visual similarity metric a good one? would there be other that would work well in other settings?" \\
\textbf{Answer}: For visual similarity, we use dot product attention ~\cite{luong-pham-manning:2015:EMNLP}, which has shown broad effectiveness in many applications and is simple to use. We compared with other metrics such as cosine similarity, L2 distance. The results are shown in Figure 4 of the paper. Dot product attention used in our framework works better than the other two metrics.
\item
\textbf{Question}: "What are the implications of the second order approximations explained in section 3.2 (paper of NeurIPS)?" \\
\textbf{Answer}: The second-order approximations imply that our framework is a valid tri-level framework. As shown in Equation 6-9 in the supplements, the gradient of the validation loss to \textbf{$A$} involves the product of three gradients: gradient of the validation loss to \textbf{$W_2'$}, gradient of \textbf{$W_2'$} to \textbf{$W_1'$}, gradient of \textbf{$W_1'$} to \textbf{$A$}. The gradient information is propagated from stage 3 to stage 2, then to stage 1. This shows that the tri-level optimization formulation of our framework is preserved in the approximated algorithm. On the other hand, the second order approximations incur additional computation cost. We reduce this cost by using finite difference approximation as shown in Equation 10-12 in the supplements, following Darts.
\item
\textbf{Question}: Memory usage.\\
\textbf{Answer}: To reduce memory cost during architecture search, we can do parameter-sharing between the three models \textbf{$W_1$}, \textbf{$W_2$}, and \textbf{$V$}. For \textbf{$W_1$} and \textbf{$W_2$}, we let them share the same convolutional layers, but have different classification heads. For \textbf{$V$}, we change it from ResNet-18 to \textbf{$W_1$}. The results are shown in Table 4 of the paper. Via parameter sharing (PS), the memory cost of our method is similar to that of regular Darts, but our method achieves lower test error. During architecture evaluation, these methods have similar GPU memory since they follow the same evaluation settings.
\end{itemize}
\subsection{Reviwer-3}
Final rating: 6
\begin{itemize}
\item
\textbf{Question}: "\textbf{(a)} Why two sets of weights $W_1$ and $W_2$ are required. \textbf{(b)} Do an ablation study which simply uses one set of weights. \textbf{(c)} Model 1 will not learn from its mistake and keeps making the same mistake." \\
\textbf{Answer (a)}: In our end-to-end tri-level framework, mathematically it is impossible to use a single set of weights to perform stage-1 learning and stage-2 learning. Once the weights are optimized at stage-1, mathematically they are not free variables any more, therefore cannot be optimized again at stage-2. So we have to use two different sets of weights to make the tri-level optimization framework mathematically correct. \\
\textbf{Answer (b)}: A simple way of learning a single set of weights is to discard the end-to-end tri-level framework and learn the weights in two separate stages: train the weights, use them to reweight training examples, then retrain the weights on reweighted examples. However, in this case, the method is no longer an end-to-end framework, which leads to inferior performance. We experimented with this simple method and its performance is worse than our tri-level end-to-end framework. The results are shown in Table 7 of the Appendix. \\
\textbf{Answer (c)}: In our end-to-end framework, model 1 will learn from its previous mistakes and avoid making the same mistakes, in an indirect way. After the weights $W_2$
of model 2 are trained by correcting the mistakes made by model 1 ($W_1$), the architecture $A$ will be updated accordingly since the gradient of $A$ depends on $W_2$; an updated $A$ will render $W_1$ to change as well since the gradient of $W_1$ depends on $A$. Along this $W_2 \xrightarrow[]{} A \xrightarrow[]{} W_1$ chain, $W_1$ is indirectly influenced by $W_2$. Since $W_2$ corrects the mistakes, $W_1$ will avoid these mistakes in the next round of training as well.
\item
\textbf{Question}: "Add more ablation studies." \\
\textbf{Answer}: Following the reviewer's suggestion, we provided more ablation studies. The results are shown in Figure 4 of the paper. The performances of our full LFM methods are significantly better than those of ablation settings, including removing $u$ (No u), removing $x$ (No x), and removing $z$ (No z). $V$ is set to ResNet-18. These results demonstrate the effectiveness and necessity of measuring mistakes on validation examples, calculating visual similarity and label similarity between training and validation examples.
\item
\textbf{Question}: "Why the ablation results in Table 4 (PDARTS, CIFAR-10) are not significantly different?" \\
\textbf{Answer}: The reason is: CIFAR-10 is a relatively simple dataset; the performance on CIFAR-10 is highly saturated (close to 100\% accuracy). Therefore, the performance difference between different methods is small on CIFAR-10. In Answer 2 above, we added ablation studies on CIFAR-100. The performances of our full LFM methods are significantly better than those of ablation settings.
\item
\textbf{Question}: "Give evidence for the proposed method learns from its mistake by visualizing some of the learned parameters." \\
\textbf{Answer}: We selected the top-50 largest values in and checked the corresponding validation images. These images are visually obscure and it is difficult to judge which classes these images belong to. This indicates that our method pays a lot of attention to hard examples where incorrect predictions tend to happen.
\end{itemize}
\subsection{Reviwer-4}
Final rating: 5
\begin{itemize}
\item
\textbf{Question}: "Comparison with Boosting." \\
\textbf{Answer}: The difference between our work and boosting is that: boosting sequentially learns multiple models separately: once a model is learned, it won’t be affected by later models. In contrast, our framework is an end-to-end framework, where all models influence each other during training. We will add related works on Boosting.
\item
\textbf{Question}: "In Tables 1/2/3, what does the column 'Param(M)' include?" \\
\textbf{Answer}: The Param(M) in the tables are the parameters of the network constructed during architecture evaluation. On CIFAR100 and CIFAR10, this network is a stack of 20 copies of the searched cell. Param(M) does not involve $W_1$, $W_2$, and $V$. These three sets of weights are only used during architecture search; after search is finished, they are discarded.
\item
\textbf{Question}: "\textbf{(a)} Why a second network $W_2$ is necessary in stage 2? \textbf{(b)} Are there experimental results supporting the benefits provided by this second network?" \\
\textbf{Answer (a)}: In our end-to-end tri-level framework, mathematically it is impossible to use a single set of weights to perform stage-1 learning and stage-2 learning. Once the weights are optimized at stage-1, they are not free variables any more and cannot be optimized again at stage-2. So we have to use two different sets of weights to make the tri-level optimization framework mathematically correct. \\
\textbf{Answer (b)}: A degenerated way of learning a single set of weights is to discard the end-to-end tri-level framework and learn the weights in two separate stages: train the weights, use them to reweight training examples, then retrain the weights on reweighted examples. However, in this case, the method is no longer an end-to-end framework, which leads to inferior performance. We experimented with this degenerated method and its performance is worse than our tri-level end-to-end framework. Below are the results.
\begin{table}[ht]
\centering
\begin{tabular}{lccc}
\toprule
Method & Dataset & Test error(\%)\\
\hline
Single set+DARTS & CIFAR-100 & 19.28$\pm$0.31 \\
Two sets+DARTS & CIFAR-100 & 17.65$\pm$0.45 \\
\hline
Single set+PDARTS & CIFAR-100 & 17.37$\pm$0.23 \\
Two set+PDARTS & CIFAR-100 & 16.44$\pm$0.11 \\
\bottomrule
\end{tabular}
\label{tab:single}
\end{table}
\item
\textbf{Question}: "\textbf{(a)} What do sample weights look like during search? \textbf{(b)} By the end of search, is the model essentially ignoring previously correctly-classified samples (very low weight)? \textbf{(c)} How different from a uniform distribution is this weight distribution?" \\
\textbf{Answer (a)}: At the early epochs during search, the variance of weights of different examples is large and the weight of the same example is changing significantly. At later epochs, the variance of weights gradually decreases, and the weights become more uniform and stable. \\
\textbf{Answer (b)}: The range of the weights is [0.5, 1]. By the end of search, most weights become 0.5. Correctly-classified samples are given relatively smaller weights, but these weights are still large enough (= 0.5) to make sure these examples are not ignored during search. \\
\textbf{Answer (c)}: At the early epochs during search, the weight distribution is very different from a uniform distribution: the weights of different examples are significantly different. At later epochs, the weight distribution becomes closer and closer to a uniform distribution.
\item
\textbf{Question}: "Compare the proposed approach with one where a network searched using standard DARTS (without reweighting) is reweighted at the end only." \\
\textbf{Answer}: We experimented with the baseline suggested by the reviewer. Below are the results. Our method works better than this baseline The reason is: in this baseline, only one round of reweighting and retraining is performed. In our method, multiple rounds of reweighting and retraining are performed iteratively in the following procedure (train $W_1$, reweight training examples, train $W_2$, update architecture, retrain $W_1$, reweight, retrain $W_2$, update architecture, etc.). Iteratively performing retraining and reweighting can continuously improve $W_1$, $W_2$, and the architecture.
\begin{table}[ht]
\centering
\begin{tabular}{lccc}
\toprule
Method & Dataset & Test error(\%)\\
\hline
Baseline+DARTS & CIFAR-100 & 19.05$\pm$0.27 \\
Ours+DARTS & CIFAR-100 & \textbf{17.65$\pm$0.45} \\
\hline
Baseline+PDARTS & CIFAR-100 & 17.44$\pm$0.18 \\
Ours+PDARTS & CIFAR-100 & \textbf{16.44$\pm$0.11} \\
\bottomrule
\end{tabular}
\label{tab:single}
\end{table}
\item
\textbf{Question}: "Comparison with random sampling or random search."\\
\textbf{Answer}: We have added the baselines of random search and random sampling in Tables 1 and 2 of the paper. LFM performs better than random search and random sampling methods, which further speaks in favour of our learning algorithm.
\item
\textbf{Question}: "\textbf{(a)} What increment in performance is due to the architecture? \textbf{(b)} What increment in performance is due to the method itself?"\\
\textbf{Answer (a)}: In Table 1-3 (obtained in architecture evaluation) in the main paper, the improvement of our methods over baselines are purely due to the architecture, since all methods follow the same architecture evaluation protocol. $W_1$, $W_2$, and training example weights learned during architecture search are discarded. The weights of the stacked large network are trained from scratch on unweighted examples. The training settings are the same for our method and baselines. Therefore, the performance difference of different methods is solely due to the quality of architectures.
\textbf{Answer (b)}: As mentioned in (a) above, the performance improvement of our method is directly due to its better searched architectures. The improvement of architectures is due to the end-to-end reweighting mechanism of our method based on tri-level optimization. Other weights ($W_1$, $W_2$, and training example weights) in our method are not directly related to performance improvement in Table 1-3, since these weights are discarded after the search process. But these weights help to improve the architecture during search, and therefore have an indirect influence on the final performance.
\item
\textbf{Question}: "Why the ablation results in Table 4 (PDARTS, CIFAR-10) are not significantly different? Add ablation on label similarity ."
\textbf{Answer}: The reason is: CIFAR-10 is a relatively simple dataset; the performance on CIFAR-10 is highly saturated (close to 100\% accuracy). Therefore, the performance difference between different methods is small on CIFAR-10. In Figure 4 of the paper, we provided ablation studies on CIFAR-100 (including the ablation on label similarity ). The performances of our full LFM methods are significantly better than those of ablation settings, including removing (No u), removing (No x), and removing (No z). is set to ResNet-18. These results demonstrate the effectiveness and necessity of measuring mistakes on validation examples, calculating visual similarity and label similarity between training and validation examples.
\end{itemize}
\section{Introduction}\label{intro}
Over the years, humans have accumulated a lot of practical learning techniques. One such effective learning method is to learn from previous mistakes. Initially, the learner learns a concept and evaluates themselves through a test to measure their level of understanding. The topics in the concept where the learner makes more mistakes can be identified as not having been learned well by the learner. Therefore, the learner will re-study the topic while focusing on the topics where mistakes were made. This will ensure that repetition of similar mistakes in the future is avoided while also strengthening previously well-learned topics. Inspired by this human learning technique, we propose a methodology that can apply the idea of learning from mistakes to machine learning.
With the deep learning era kicking off in machine learning, state-of-the-art neural network performance is achieved mainly by architectures designed manually by experts. The neural architecture search (NAS) and evaluation by human experts require substantial effort and may not give the most optimal performance. Recently, there has been growing interest in automating the manual process of learning architecture design. This paper proposes a general framework that draws inspiration from human learning skills and can be applied to any differentiable architecture search algorithm. We also explore the efficacy of our learning method on NAS.
In our framework, the model consists of two sets of network weights sharing a common learnable architecture, a learnable data encoder, and a coefficient vector - representing different components of the learner. The two sets of network weights correspond to two parts of the main task learning faculties of the learner. The learnable architecture corresponds to skill learning faculty of the learner. The data encoder and coefficient vector correspond to auxiliary faculties of the learner, helping it to recognise and summarize the information respectively. We propose a multi-level optimization framework that uses the above sets of parameters to learn better neural architectures.
We begin by training the first set of network weights on a training dataset. We then see what mistakes our model makes while predicting on the validation set. Then, for each training example, we assign specific weights based on the mistakes made by the model and the similarity of this example to an incorrectly predicted validation example.
The second set of network weights are trained on these weighted examples, essentially making the model learn from its mistakes and correct them. The vanilla approaches in NAS do not factor in the variations in learning difficulty. In contrast, our method re-weights the training examples at each stage based on the current capability of the learner, enabling the learner to deal with more challenging cases in the unseen data. Then finally, the architecture, encoder, and coefficient vectors are learned on the second model's validation performance.
The major contributions of the paper are as follows:
\begin{itemize}
\item Inspired by the human learning technique, we propose a novel approach to apply the Learning from Mistakes (LFM) method in machine learning.
\item We formulate LFM as a multi-level optimization framework that involves three stages of learning: learner learns; learner corrects its mistakes; learner, encoder, and coefficient vector validate themselves.
\item We apply our approach to neural architecture search on CIFAR-100, CIFAR-10, and ImageNet. The results demonstrate the effectiveness of our method.
\end{itemize}
\section{Related Works}
\subsection{Data Re-weighting and Selection}
Several approaches have been proposed for data selection. Matrix column subset selection~\cite{deshpande2010efficient,boutsidis2009improved} aims to select a subset of data examples that can best reconstruct the entire dataset. Similarly, Coreset selection~\cite{bachem2017practical} chooses representative training examples such that models trained on the selected examples have comparable performance with those trained on all training examples. These methods perform data selection and model training separately. As a result, the validation performance of the model cannot be used to guide data selection. \cite{ren2018learning} proposes a meta-learning method to learn the weights of training examples by performing a meta gradient descent step on the weights of the current mini-batch of examples. \cite{shu2019meta} propose a method that can adaptively learn an explicit weighting function
directly from data.
Our work takes inspiration from~\cite{ren2018learning} to use a meta gradient step on the weights, and from~\cite{shu2019meta} to explicitly learn a weighting function and extends them to use not only validation and training performance but also the similarity of data and labels. We make this possible by learning the weights and the learner in different stages of the optimization problem. Moreover, the previous works focus on selecting training (finetuning) examples using a bi-level optimization framework, while our work focuses on selecting pretraining examples using a three-level optimization framework.
\subsection{Neural Architecture Search (NAS)}
Recently, NAS has come to the forefront of deep learning techniques due to its success in discovering neural architectures that can substantially outperform manually designed ones. Early versions of NAS such as \cite{zoph2016neural,pham2018efficient,zoph2018learning} used computationally intensive approaches like reinforcement learning - where the accuracy of the validation set was defined as the reward, and a policy network was trained to generate architectures that can maximize these rewards. Another contemporary approach \cite{liu2018hierarchical,Real_Aggarwal_Huang_Le_2019} was using evolutionary learning techniques - where the set of all architectures represent a population and the fitness score is the validation accuracy of each architecture. Architectures with lower fitness scores would be replaced with higher fitness score architectures. However, even this approach was computationally intensive. To address this problem, differentiable architecture search techniques were explored \cite{cai2018proxylessnas,liu2018darts,xie2018snas} and their results are much more promising because of the use of weight-sharing techniques and the application of gradient descent in a continuous architecture search space.
DARTS~\cite{liu2018darts} made the first breakthrough in the area of Differentiable NAS. Several other DARTS-based techniques \cite{chen_progressive_2019,xu_pc-darts_2019,liang2019darts+,chu2021darts} have explored to further reduce the cost of computation for differentiable NAS. Some of the approaches include -
PDARTS~\cite{chen_progressive_2019} increasing the depth of architectures progressively during the searching,
PC-DARTS~\cite{xu_pc-darts_2019} evaluating only a sub-set of channels and thereby reducing the redundancy in the search space.
The LFM framework proposed in this paper can be applied to any differentiable NAS method for further enhancement.
\section{Method} \label{sec:method}
In this section, we propose a framework that can imitate human learning in the form of Learning from Mistakes (LFM) and present an optimization algorithm for solving the problem of LFM.
\subsection{ The Learning from Mistakes framework} \label{framework}
The framework contains two sets of
network weights $W_1$ and $W_2$ - that are two parts of the same learner and are trying to learn to perform the same target task. They share a learnable architecture $A$. The primary goal here is to help the learner correct the mistakes (made when studying for the first time) during the revision. Further, to help map the topics in the test to topics in the syllabus, there is an encoder with pre-defined neural architecture (by human experts) with learnable network weights $V$; and a coefficient vector $r$. We organized the learning into three stages.
\iffalse
\begin{figure*}[t]
\centering
\includegraphics[width=13cm, height=5cm]{figs/metrics.png}
\caption{The process flow to calculate the weights $a_i$ for training example $i$.
}
\label{fig:1}
\end{figure*}
\fi
\paragraph{Stage I.} In the first stage, we train the first set of network weights $W_1$ by minimizing the loss on the training dataset $D^{(tr)} $.
The optimal weights $W_{1}^{*}(A)$ is a function of architecture A, which at this stage is fixed, and hence:
$$W_1^*(A) = \mathop{\arg\min}_{W_1} L(A, W_1, D^{tr})$$
The architecture $A$ is used to define the training loss but is not updated at this stage. If we were to directly learn $A$ by minimizing this training loss, a trivial solution would be yielded where $A$ is very large and complex, and would perfectly overfit the training data but generalize poorly on unseen data.
\paragraph{Stage II.} In the second stage, the goal is to re-weight the training samples for training the second set of network weights $W_2$ of the learner. We apply $W_{1}^{*}(A)$ to the validation dataset $D^{(val)}$ and check its performance on the validation examples. Without loss of generality, we assume the task is image classification. To make the model re-learn whilst focusing more on mistakes on the validation examples by $W_{1}^{*}(A)$, we re-weight each training example $d^{(tr)}_i$ based on the following metrics:
\begin{itemize}
\item Visual similarity between $d^{(tr)}_i$ and $d^{(val)}_j$, denoted by $x_{ij}$
\item Label similarity of $d^{(tr)}_i$ and $d^{(val)}_j$, denoted by $z_{ij}$
\item Valid performance of $W_{1}^{*}(A)$ on $D^{(val)}$, denoted by $u_j$
\end{itemize}
In human learning, a question that has been incorrectly learned can be corrected during the relearning stage by focusing more on examples that are similar to the wrongly learned question. Here, the metric $x_{ij}$ tries to measure how similar a previously incorrectly predicted $d^{(val)}_j$ is to a training example $d^{(tr)}_i$ and $z_{ij}$ depicts whether they describe the same topic. The term $u_j$ measures by how much $W_{1}^{*}(A)$ is wrong for each validation example $j$.
We use these re-weighted training examples to train $W_2$. This allows $W_2$ to focus on the topics that $W_1$, after training, could not get right.
Visual similarity measures how similar the training example $d^{(tr)}_i$ is to each validation example $d^{(val)}_j$. Let V denote an image encoder. For
each validation example $d^{(val)}_j$, its similarity with $d^{(tr)}_i$ is defined as the dot-product attention~\cite{luong-pham-manning:2015:EMNLP} as:
\begin{equation}
\label{visual_similarity}
x_{ij} = \frac{exp(V(d^{(tr)}_i). V(d^{(val)}_j))}{\sum_{k=1}^{N^{(val)}}exp(V(d^{(tr)}_i) . V(d^{(val)}_k))}
\end{equation}
where $N^{(val)}$ is the number of validation examples. $V (d)$
denotes the $K$-dimensional visual representation of the data example d.
Label similarity measures the similarity between the label of the training example and the label of each validation
example. Let $z_{ij}$ denote the label similarity between a validation example $d^{(val)}_j$
and a training example $d^{(tr)}_i$. We
define $z_{ij}$ as:
\begin{align} \label{z_ij}
z_{ij} = \mathbb{I}\{y^{(tr)}_i=y^{(val)}_j\}
\end{align}
where $y$ is the label of the corresponding data example, and $\mathbb{I}\{\}$ the indicator function on the condition being true or not.
The validation performance $u_j$ of $W_{1}^{*}(A)$ on a validation example $d^{(val)}_j$ is the cross
entropy loss on this example:
\begin{equation} \label{val_perform}
u_j = \text{crossentropy} (f(d^{(val)}_j; W_{1}^{*}(A)), y^{(val)}_j )
\end{equation}
where $f(d^{(val)}_j; W_{1}^{*}(A))$ is the predicted probabilities of $W_{1}^{*}(A)$ on $d^{(val)}_j$ and $y^{(val)}_j$ is the class label of $d^{(val)}_j$.
Let $x_i$, $z_i$, and $u$ be $N^{(val)}$-dimensional vectors where the $j$-th
element is $x_{ij}$ , $z_{ij}$ , and $u_j$ defined before. We calculate the
weight $a_i$ of the training example $d^{(tr)}_i$ as:
\begin{equation} \label{ai}
a_i = \text{sigmoid}((x_i \odot z_i \odot u)^T r)
\end{equation}
where $\odot$ denotes element-wise multiplication, and $r$ is a coefficient vector. Note $a_i$ is a function of $V$, $W_{1}^{*}(A)$, and $r$.
Given the weight $a_i$ of each training example, we train the second set of network weights $W_2$ by minimizing the weighted
training loss, with the architecture $A$, encoder $V$, and $r$ fixed.
\begin{align} \label{w2_eqn}
W_2^*(W_1^*(A), V, r) = \mathop{\arg\min}_{W_2} \sum_{i=1}^{N^{tr}} a_i \ell(A, W_2(A), d_i^{(tr)})
\end{align}
\paragraph{Stage III.} In the third and final stage,
the encoder $V$, coefficient vector $r$, and the architecture $A$ minimise the validation loss of $W_2^*(W_1^*(A), V, r)$.
\begin{align} \label{avr_eqn}
A, V, r = \mathop{\arg\min}_{A,V,r} L(A, W_2^*(W_1^*(A), V, r), D^{(val)})
\end{align}
Putting the above pieces together, we have the following LFM framework, which is a three level optimization problem:
\begin{align}\label{optimisation_eqn}
& A, V, r = \mathop{\arg\min}_{A,V,r} L(A, W_2^*(W_1^*(A), V, r), D^{(val)}) \nonumber \\
& s.t. \quad W_2^*(A) = \mathop{\arg\min}_{W_2} \sum_{i=1}^{N^{tr}} a_i \ell (A, W_2(A), d_i^{(tr)}) \\
& \qquad \ W_1^*(A) = \mathop{\arg\min}_{W_1} L(A, W_1, D^{tr}) \nonumber
\end{align}
We summarise the above equations in the process flow diagram Figure \ref{fig:2}. In our end-to-end framework, $W_1$ will learn from its previous mistakes and avoid making similar mistakes, in an indirect way. After the weights of $W_2$ are trained by correcting the mistakes made by $W_1$, the architecture will be updated accordingly since the gradient of $A$ depends on
$W_2$; an updated $A$ will render $W_1$
to change as well since the gradient of $W_1$
depends on $A$ . Along the chain $W_2 \xrightarrow[]{} A \xrightarrow[]{} W_1$
chain, $W_1$ is indirectly influenced by
$W_2$.
\begin{figure}[ht]
\centering
\includegraphics[width=7cm, height=6cm]{figs/flow.png}
\caption{The overall process flow of our method when applying to NAS. The red arrows indicate stage 1 processes, blue arrows indicate stage 2 processes, and black arrows indicate stage 3 processes.
}
\label{fig:2}
\end{figure}
Similar to~\cite{liu2018darts}, we represent the architecture $A$ of the learner in a differentiable way. The search space of $A$ is composed of many building blocks, where the output of each block is associated with a weight $a$ indicating the importance of the block. After learning, block whose weight $a$ is among the largest are retained to form the final architecture. To this end, architecture search amounts to optimizing the set of architecture weights $A = \{a\}$. Our framework can be applied on top of any differentiable architecture search methods such as DARTS~\cite{liu2018darts}, PDARTS~\cite{chen_progressive_2019}, PC-DARTS~\cite{xu_pc-darts_2019}, DARTS-~\cite{chu2021darts} among others.
\subsection{Optimization Algorithm}\label{optimisation_algo}
In this section, we derive an optimization algorithm to solve the LFM problem defined in Equation \ref{optimisation_eqn}. Inspired by \cite{liu2018darts}, we approximate $W_1^*(A)$ and $W_2^*(W_1^*(A), A, V, r)$ one step gradient descent updates for the inner optimization equations to reduce the computational complexity.
\paragraph{Stage I.} For Stage 1, we approximate $W_1^{*}(A)$ using one step descent for the loss on training data $L(A, W_1, D^{tr})$ keeping $A$ constant as follows:
\begin{align}\label{st1_approx}
W_1'(A) = W_1 - \eta_{W_1} \nabla_{W_1} L(A, W_1, D^{tr})
\end{align}
\paragraph{Stage II.} For Stage 2, we use $W_1'$ from the Equation \ref{st1_approx} to get $u_j$. We compute $x_i$ and $z_i$ for each training sample $d_i^{tr}$ using the equations \ref{visual_similarity} and \ref{z_ij}, to finally compute $a_i$ as:
\begin{align} \label{ai_cal:final}
a_i(d^{val}, d_i^{tr}) &= \text{sigmoid}((x_i \odot z_i \odot u)^T r)
\end{align}
Next, we use one step gradient descent to approximate $W_2^*(A,W_1^*(A),V,r)$ as:
\begin{align} \label{st2_approx}
W_2'(A) = W_2 - \eta_{W_2} \nabla_{W_2} \sum_{i=1}^{N^{tr}} a_i \ell (A, W_2, d_i^{(tr)})
\end{align}
\paragraph{Stage III.} For Stage 3, we plug Equation \ref{st2_approx} to learn architecture $A$ , encoder $V$, and coefficient vector $r$ from the validation loss $L(A, W_2'(W_1'(A), V, r), D^{(val)})$.
The encoder model $V$ can be updated as:
\begin{align} \nonumber
V' &= V - \eta_V \nabla_V L(A, W_2', D^{(val)}) \\ \nonumber
& = V + \eta_V \eta_{W_2} \frac{\partial}{\partial V} \Big(\big( \nabla_{W_2}\sum_{i=1}^{N^{tr}} a_i(V) \ell(A, W_2, d_i^{(tr)})\big) . \nonumber \\
& \qquad \qquad \qquad \qquad \big( \nabla_{W_2'} L(A, W_2',D^{(val)})\big)\Big) \label{v_updt}
\end{align}
Similarly, the coefficient vector $r$ can be updated as:
\begin{align} \nonumber
r' &= r - \eta_r \nabla_r L(A, W_2', D^{(val)}) \\\nonumber
&= r + \eta_r \eta_{W_2} \frac{\partial}{\partial r} \Big(\big( \nabla_{W_2}\sum_{i=1}^{N^{tr}} a_i(V) l(A, W_2, d_i^{(tr)})\big). \nonumber \\
&\qquad \qquad \qquad \qquad \big( \nabla_{W_2'} L(A, W_2',D^{(val)})\big)\Big)
\label{r_updt}
\end{align}
Next, we give the update equation for the architecture $A$.
\begin{align} \nonumber
A' &= A - \eta_A \nabla_A L(A, W_2'(A), D^{(val)}) \\ \nonumber
&= A - \eta_A \big(\frac{\partial L(A, W_2',D^{(val)})}{\partial A} + \nonumber \\
& \qquad \qquad \qquad \frac{\partial L(A, W_2',D^{(val)})} {\partial W_2'} \frac{\partial W_2'(A)}{\partial A}\big) \label{updateA}
\end{align}
To save space, the complete derivation is not given in the paper.
\begin{align} \nonumber
&A' = A - \eta_A \big( \frac{\partial L(A, W_2',D^{(val)})}{\partial A} - \eta_{W_2} \big(\\ \nonumber
& -\eta_{W1} \frac{\nabla_{A}L(A, W_1^+, D^{(tr)}) - \nabla_{A}L(A, W_1^-, D^{(tr)})}{2 \epsilon_1} \nonumber \\
& + \frac{\nabla_A \sum_i a_i \big( L(A, W_2^+, d_i^{(tr)}) - L(A, W_2^-, d_i^{(tr)})\big) }{2\epsilon_2} \big) \big) \label{A_updt}
\end{align}
where \begin{align*}W_1^\pm &= W_1 \pm \epsilon_1 \nabla_{W_1'} \big((\nabla_{W_2} \sum_{i=1}^{N^{(tr)}} \textbf{a}_i(W_1') l(A, W_2, d_i^{(tr)})). \nonumber \\
& \qquad \nabla_{W_2'}L(A, W_2',D^{(val)})\big) \nonumber
\end{align*} and
\begin{align*}
W_2^\pm = W_2 \pm \epsilon_2 \nabla_{W_2'} L(A, W_2',D^{(val)}) \end{align*}
Also, $\epsilon_1$ and $\epsilon_2$ are small scalars.
The overall algorithm of LFM when applying to other related methods can be summarised in Algorithm \ref{algo1}. And LFM can be applied to any differentiable NAS methods. To apply our framework to a differentiable NAS method $M$, we just need to set the architecture variable $A$ in our framework to be the search space of $M$. In other words, $M$ is a special solution in the solution space $S$ of our method.
\iffalse
\paragraph{Update equations for Stage 3} We now show the update equations for Stage 3 to solve the LFM problem defined in Equation \ref{optimisation_eqn}.
We plug Equation \ref{st2_approx} to learn architecture $A$ , encoder $V$, and coefficient vector $r$ from the validation loss $L(A, W_2'(W_1'(A), V, r), D^{(val)})$.
The encoder model $V$ can be updated as:
\begin{align}
&V' = V + \eta_V \eta_{W_2} \frac{\partial}{\partial V} \\ & \left( \left( \nabla_{W_2}\sum_{i=1}^{N^{tr}} a_i(V) \ell(A, W_2, d_i^{(tr)}) \right) . \left( \nabla_{W_2'} L(A, W_2',D^{(val)}) \right) \right) \label{v_updt}
\end{align}
Similarly, the coefficient vector $r$ can be updated as:
\begin{align} \nonumber
& r' = r + \eta_r \eta_{W_2} \frac{\partial}{\partial r} \\ & \left( \left( \nabla_{W_2}\sum_{i=1}^{N^{tr}} a_i(r) \ell (A, W_2, d_i^{(tr)}) \right). \left( \nabla_{W_2'} L(A, W_2',D^{(val)}) \right) \right)\label{r_updt}
\end{align}
Next, we give the update equation for the architecture $A$.
\begin{align}
&A' = A - \eta_A \left(\frac{\partial L(A, W_2',D^{(val)})}{\partial A} + \frac{\partial L(A, W_2',D^{(val)})} {\partial W_2'} \frac{\partial W_2'(A)}{\partial A} \right)\qquad \label{updateA}
\end{align}
To save space, the complete derivation is given in the supplementary material.
\begin{align} \nonumber
&A' = A - \eta_A \big( \frac{\partial L(A, W_2',D^{(val)})}{\partial A} - \eta_{W_2} \big(\\ \nonumber
& -\eta_{W1} \frac{\nabla_{A}L(A, W_1^+, D^{(tr)}) - \nabla_{A}L(A, W_1^-, D^{(tr)})}{2 \epsilon_1} \nonumber \\
& + \frac{\nabla_A \sum_i a_i L(A, W_2^+, d_i^{(tr)}) - \nabla_A \sum_i a_i L(A, W_2^-, d_i^{(tr)}) }{2\epsilon_2} \big) \big) \label{A_updt}
\end{align}
where \begin{align*}& W_1^\pm = W_1 \pm \epsilon_1 \nabla_{W_1'} \\ & \left((\nabla_{W_2} \sum_{i=1}^{N^{(tr)}} \textbf{a}_i(W_1') \ell (A, W_2, d_i^{(tr)})).\nabla_{W_2'}L(A, W_2',D^{(val)})\right)
\end{align*} and
\begin{align*}
W_2^\pm = W_2 \pm \epsilon_2 \nabla_{W_2'} L(A, W_2',D^{(val)}) \end{align*}
Also, $\epsilon_1$ and $\epsilon_2$ are small scalars.
\fi
\begin{algorithm}[H]
\begin{algorithmic}[1]
\WHILE{not converged}
\STATE Update $W_1$
\STATE Update $W_2$
\STATE Update $A, V, r$ respectively
\ENDWHILE
\caption{Algorithm for LFM}
\label{algo1}
\end{algorithmic}
\end{algorithm}
\section{Experiments} \label{sec:experiments}
\subsection{Datasets}
The experiments are performed on three popular NAS datasets, namely CIFAR-10, CIFAR-100~\cite{krizhevsky_learning_nodate} and ImageNet~\cite{deng2009imagenet}. CIFAR-10 contains $10$ classes, and CIFAR-100 contains $100$ classes. Both these datasets contain 60K images each with each class having the same number of images. We split each of these datasets into a training set with 25K images, a validation set with 25K images, and a test set with 10K images. During architecture search in LFM, the training set is used as $D^{tr}$ and the validation set is used as $D^{val}$. During architecture evaluation, the learned network is trained on the combination of $D^{tr}$ and $D^{val}$. Further, ImageNet contains 1.2M training images and 50K test images with 1000 objective classes.
\subsection{Experimental Settings}
Our framework is orthogonal to existing NAS approaches and can be applied to any differentiable NAS method. In our experiments, we applied LFM to DARTS~\cite{liu2018darts} and PDARTS~\cite{chen_progressive_2019}. The search spaces of these methods are composed of (dilated) separable convolutions with sizes of 3 × 3 and 5 × 5, max pooling with the size of 3 × 3, average pooling with the size of 3 × 3, identity, and zero. For the encoder $V$, Res-Nets pre-trained on Imagenet were used. Each LFM experiment was repeated three times with different random seeds. The mean and standard deviation of classification errors obtained from the three runs are reported.
\begin{table}[t]
\small
\centering
\begin{tabular}{l|ccc}
\toprule
Method & Error(\%)\\
\midrule
*ResNet \cite{he2016deep}&22.10\\
*DenseNet \cite{HuangLMW17}&17.18\\
\hline
*PNAS \cite{LiuZNSHLFYHM18}&19.53\\
*ENAS \cite{pham2018efficient}&19.43\\
*AmoebaNet \cite{real2019regularized}&18.93\\
\hline
*GDAS \cite{DongY19}&18.38\\
*R-DARTS \cite{ZelaESMBH20}&18.01$\pm$0.26
\\
*DARTS$^{-}$ \cite{abs-2009-01027}&17.51$\pm$0.25\\
*DARTS$^{-}$ \cite{abs-2009-01027}& 18.97$\pm$0.16\\
*DARTS$^{+}$ \cite{abs-1909-06035}&17.11$\pm$0.43\\
*DropNAS \cite{HongL0TWL020} & 16.39\\
$\;\;$Random search & 21.92$\pm$0.34 \\
$\;\;$Random sampling & 21.37$\pm$0.48 \\
\hline
\hline
*DARTS-2nd \cite{liu2018darts} & 20.58$\pm$0.44\\
$\;\;$LFM-DARTS-2nd-R18 (ours) & \textbf{17.65$\pm$0.45}\\
\hline
*PDARTS \cite{chen_progressive_2019}&17.49\\
$\;\;$LFM-PDARTS-R18 (ours) & \textbf{16.44$\pm$0.11}\\
$\;\;$LFM-PDARTS-R34 (ours) & \textbf{15.69$\pm$0.15}\\
\bottomrule
\end{tabular}
\caption{Test error on CIFAR-100. LFM-DARTS-2nd-R18 represents that LFM is applied onto DARTS-2nd, and ResNet-18 is used as the data encoder. Similar meanings hold for other notations in such a format. Results marked with * are obtained from Skillearn \cite{xie_skillearn_2020}. The information about parameters and search cost are shown in the supplement.}
\label{tab:cifar100}
\end{table}
\paragraph{Architecture Search} During architecture search for CIFAR-10 and CIFAR-100, the architectures of $W_1$ and $W_2$ are a stack of 8 cells. Each cell consists of 7 nodes. We set the initial channel number to 16. For the architecture of the encoder model, we experimented with ResNet-$18$ and ResNet-$34$~\cite{7780459}. The search algorithm was based on SGD, and the hyperparameters of epochs, initial learning rate, and momentum follow the original implementation of the respective DARTS~\cite{liu2018darts} and PDARTS~\cite{chen_progressive_2019}. We use a batch size of 64 for both DARTS and PDARTS. LFM-PDARTS uses first order approximations to be consistent with the original implementation in PDARTS~\cite{chen_progressive_2019}. The LFM experiments in this paper use A100 for DARTS and A40 for PDARTS.
\paragraph{Architecture Evaluation} During architecture evaluation for CIFAR-10 and CIFAR-100, a larger network of each category-specific model is formed by stacking 20 copies of the searched cell. The initial channel number is set to 36. We trained the network with a batch size of 96, an epoch number of 600, on a single Tesla v100 GPU. On ImageNet, we evaluate the architectures searched on CIFAR-10 or CIFAR-100. In either type, 14 copies of optimally searched cells are stacked into a large network, which was trained using two Tesla A100 GPUs on the 1.2M training images, with a batch size of 1024 and an epoch number of 250. The initial channel number is set to 48.
The LFM method is used to learn the architecture $A$, while the weights $W_1$, $W_2$, $V$, and $r$ learnt during the LFM search are discarded during the architecture evaluation. All the architecture evaluations are run using the same standardized setup as described in the above paragraph. This results in a fair comparison between architectures learnt from different methods, as all the models during evaluation have same number of parameters and hyper-parameters such as epochs, learning rate, and batch size.
\subsection{Results}
\begin{table}[t]
\small
\centering
\begin{tabular}{l|ccc}
\toprule
Method& Error(\%)\\
\midrule
*DenseNet
\cite{HuangLMW17}&3.46\\
\hline
*HierEvol \cite{liu2017hierarchical}&3.75$\pm$0.12\\
*PNAS \cite{LiuZNSHLFYHM18} &3.41$\pm$0.09\\
*NASNet-A \cite{zoph2018learning} & 2.65\\
*AmoebaNet-B \cite{real2019regularized} & 2.55$\pm$0.05\\
\hline
*R-DARTS \cite{ZelaESMBH20} &2.95$\pm$0.21\\
*GTN~\cite{abs-1912-07768}& 2.92$\pm$0.06\\
*BayesNAS \cite{ZhouYWP19} &2.81$\pm$0.04\\
*MergeNAS \cite{WangXYYHS20} &2.73$\pm$0.02\\
*NoisyDARTS \cite{abs-2005-03566} &2.70$\pm$0.23\\
*ASAP \cite{NoyNRZDFGZ20} &2.68$\pm$0.11\\
*SDARTS
\cite{abs-2002-05283}&2.61$\pm$0.02\\
*DropNAS \cite{HongL0TWL020} &2.58$\pm$0.14\\
*DrNAS \cite{abs-2006-10355} &2.54$\pm$0.03\\
$\;\;$Random search & 3.07$\pm$0.17 \\
$\;\;$Random sampling & 2.75$\pm$0.09 \\
\hline
\hline
*DARTS-2nd \cite{liu2018darts} &2.76$\pm$0.09\\
$\;\;$LFM-DARTS-2nd-R18 (ours) &\textbf{2.70$\pm$0.06}\\
\hline
*PDARTS \cite{chen_progressive_2019}& 2.50\\
$\;\;$LFM-PDARTS-R18 (ours) &\textbf{2.46$\pm$0.04}\\
\bottomrule
\end{tabular}
\caption{
Test error on CIFAR-10. Results marked with * are obtained from Skillearn~\cite{xie_skillearn_2020}. The other notations are same as described in Table \ref{tab:cifar100}. The information about parameters, search cost and more compared methods are shown in the supplement.
}
\label{tab:cifar10}
\end{table}
\begin{table*}[t]
\centering
\begin{tabular}{l|cccc}
\toprule
\multirow{ 2}{*}{Method} & Top-1 &Top-5 \\
& Error (\%) & Error (\%)\\
\midrule
*Inception-v1 \cite{googlenet}&30.2 &10.1\\
*MobileNet \cite{HowardZCKWWAA17} & 29.4& 10.5\\
*ShuffleNet 2$\times$ (v1) \cite{ZhangZLS18} & 26.4 &10.2 \\
*ShuffleNet 2$\times$ (v2) \cite{MaZZS18} & 25.1 &7.6\\
\hline
*NASNet-A \cite{zoph2018learning} &26.0 &8.4\\
*PNAS \cite{LiuZNSHLFYHM18} &25.8 &8.1\\
*MnasNet-92 \cite{TanCPVSHL19} & 25.2 & 8.0\\
*AmoebaNet-C \cite{real2019regularized} & 24.3 &7.6\\
\hline
*SNAS-CIFAR10 \cite{xie2018snas} & 27.3 &9.2\\
*PARSEC-CIFAR10 \cite{abs-1902-05116} & 26.0 &8.4\\
*DSNAS-ImageNet \cite{HuXZLSLL20} &25.7& 8.1\\
*SDARTS-ADV-CIFAR10 \cite{abs-2002-05283}&25.2& 7.8\\
*FairDARTS-ImageNet \cite{abs-1911-12126} &24.4 &7.4\\
*DrNAS-ImageNet \cite{abs-2006-10355} & 24.2 &7.3\\
*ProxylessNAS-ImageNet \cite{cai2018proxylessnas} & 24.9 &7.5 \\
*GDAS-CIFAR10 \cite{DongY19} & 26.0&8.5\\
\hline
\hline
*DARTS2nd-CIFAR10 \cite{liu2018darts} & 26.7 &8.7\\
$\;\;$LFM-DARTS-2nd-CIFAR10 (ours) & \textbf{25.12} & \textbf{7.65} \\
\hline
*PDARTS (CIFAR10) \cite{chen_progressive_2019}&24.4 &7.4\\
$\;\;$LFM-PDARTS-CIFAR10 (ours) & \textbf{24.14}&\textbf{6.85}\\
\hline
*PDARTS (CIFAR100) \cite{chen_progressive_2019}&24.7& 7.5\\
$\;\;$LFM-PDARTS-CIFAR100 (ours)& \textbf{24.11} & \textbf{6.70}\\
\bottomrule
\end{tabular}
\caption{Top-1 and top-5 classification errors on the test set of ImageNet. Results marked with * are obtained from Skillearn \cite{xie_skillearn_2020}. From top to bottom, on the first three blocks are 1) networks manually designed by humans; 2) non-gradient based NAS methods; and 3) gradient-based NAS methods. Rest of the notations follow Tables \ref{tab:cifar100}, \ref{tab:cifar10}. The information about parameters, search cost and more compared methods are shown in the supplement.
}
\label{tab:imagenet}
\end{table*}
\subsubsection{Result of classification in different datasets}
The results of the classification error(\%) of different NAS methods on CIFAR-100 are showed in Table \ref{tab:cifar100}. We make the following observations from this table:
\begin{itemize}
\item When LFM is applied DARTS-2nd (second-order approximation) and PDARTS, the classification errors of these methods are reduced significantly. For example, when LFM is applied to DARTS-2nd, the error reduces from $20.58\%$ to $17.70\%$. In PDARTS, the error reduces from $17.49\%$ to $16.44\%$ (when using ResNet-18 as encoder $V$). This shows the effectiveness of our method in improving the performance of architecture search. In the baseline NAS approaches, all the training examples have the same weight, which implicitly implies that all examples are equally difficult to learn. The learner, in this case, can give a good performance by performing well on the majority of easy examples and ignoring the minority of difficult examples. In contrast, our method re-weights the training example at each stage based on the current capability of the learner, giving more weight to examples that are difficult to learn, which is a more realistic scenario. The learner gains the ability to deal with more challenging cases in the unseen data.
\item LFM-PDARTS-R34 outperforms LFM-PDARTS-R18 by $0.75\%$, where the former uses ResNet-34 as the image encoder, while the latter uses ResNet-18. ResNet-34 is a deeper and more powerful data encoder than ResNet-18. This shows that mapping the validation examples to similar training examples is a core component contributing to the effectiveness of our proposed LFM method.
\item LFM-PDARTS-R34 achieves the best performance among all methods, which demonstrates the effectiveness of applying LFM to differentiable NAS methods and improving their performance. To the best of our knowledge, LFM-PDARTS-R34 is the new SOTA on CIFAR-100.
\end{itemize}
The results of the classification error(\%) of different NAS methods on CIFAR-10 are showed in Table \ref{tab:cifar10}. As can be seen, LFM applied to DARTS-2nd and PDARTS reduces the errors of these baselines by roughly $0.05\%$. This further demonstrates the effectiveness of our method.
The results of the classification error(\%) - top-1 and top-5 of different NAS methods on ImageNet are showed in Table \ref{tab:imagenet}. In methods LFM-DARTS-2nd-CIFAR10 and LFM-PDARTS-CIFAR10, the architecture searched on CIFAR-10 is evaluated on ImageNet, whereas in LFM-PDARTS-CIFAR100, the architecture searched on CIFAR-100 is evaluated on Imagenet. The LFM-DARTS-2nd-CIFAR10 outperforms the baseline DARTS-2nd-CIFAR10 by $1.6\%$, while LFM-PDARTS-CIFAR100 outperforms its corresponding baseline by $0.6\%$, and LFM-PDARTS-CIFAR10 by $0.3\%$. As shown, the LFM methods outperform their corresponding baselines, and therefore, demonstrate our method's effectiveness.
\subsubsection{Ablation Studies}
\paragraph{Ablation 1} Our method introduces three important components to the re-weighting parameter $a_i$ namely: $x$, $u$, and $z$. In this study, we demonstrate the effect of ablating each component. We perform the experiment of CIFAR-100 and set the encoder to ResNet-18. The other details are are kept same as the base experiments described in earlier sections. The performances of the ablated models are shown in \ref{fig:3}. These results demonstrate the effectiveness and necessity of measuring mistakes $u$ on validation examples, calculating visual similarity $x$ and label similarity $z$ between training and validation examples.
\iffalse
\begin{table}[ht]
\centering
\begin{tabular}{l|c}
\toprule
Model & Test Error(\%) \\
\hline
LFM-DARTS & \textbf{17.65$\pm$0.45} \\
LFM-DARTS (No u) & 19.30$\pm$0.48 \\
LFM-DARTS (No x) & 19.05$\pm$0.12 \\
LFM-DARTS (No z) & 18.77$\pm$0.27 \\
\hline
LFM-PDARTS & \textbf{16.44$\pm$0.11} \\
LFM-PDARTS (No u) & 17.13$\pm$0.09 \\
LFM-PDARTS (No x) & 17.01$\pm$0.12 \\
LFM-PDARTS (No z) & 17.28$\pm$0.17 \\
\bottomrule
\end{tabular}
\caption{Results of ablation experiments on CIFAR-100, including removing validation performance (No u), removing visual similarity (No x), and removing label similarity (No z).}
\end{table}
\fi
\begin{figure}[ht]
\centering
\includegraphics[height=5cm]{figs/xuz_chart.png}
\caption{Ablation on components of $a_i$. The left bar column shows the comparison of ablated models 1) without x, 2) without u, 3) without z , 4) and the full models.
\label{fig:3}
\end{figure}
\paragraph{Ablation 2} For visual similarity, we use dot product attention, which has shown broad effectiveness in many applications and is simple to use. To demonstrate its effectiveness in LFM, we compared with other metrics such as cosine similarity and L2 distance. The other details are are kept same as the base experiments described in earlier sections. Results are shown in Figure \ref{fig:4}. Dot product attention used in our framework works better than the other two metrics.
\begin{figure}[ht]
\centering
\includegraphics[height=5cm]{figs/chart_cosine.png}
\caption{Comparison of different visual similarity metrics.}
\label{fig:4}
\end{figure}
\iffalse
\paragraph{Ablation 3} We applied our method as a data re-weighting strategy for architecture learned by DARTS or P-DARTS to check the effectiveness of our method. We set this as baseline. The results on CIFAR-100 are shown in Table \ref{tab:DR} below. This baseline works better than vanilla DARTS/PDARTS, but worse than our method. Our method is better because it learns architecture and data weights jointly while the baseline learns them separately.
\begin{table}[ht]
\begin{center}
\begin{tabular}{l|c}
\hline
Method & Test error(\%)\\
\hline
Baseline & 19.72$\pm$0.36 \\
Ours+DARTS& 17.65$\pm$0.45 \\
DARTS& 20.58$\pm$0.44\\
\hline
Baseline & 17.03$\pm$0.08 \\
Ours+PDARTS& 16.44$\pm$0.11 \\
PDARTS &17.49\\
\hline
\end{tabular}
\end{center}
\caption{}
\label{tab:DR}
\end{table}
\fi
\paragraph{Ablation 3} Since we use the data encoder that is pre-trained on ImageNet, it may provide an unfair advantage to our method in the CIFAR100 experiment since the encoder is exposed to more data. We set experiments to restrict our method to only using the CIFAR100 dataset and check its performance. We train the encoder solely on CIFAR-100, without using ImageNet pre-training. The results (on Cifar100) are given in Table \ref{tab:pre} below. Removing ImageNet pretraining does not increase test errors significantly, showing that the improvement achieved by our method over DARTS/PDARTS is not due to ImageNet pretraining.
\begin{table}[ht]
\begin{center}
\begin{tabular}{l|c}
\hline
Method & Test error(\%)\\
\hline
LFM+DARTS, no INP & 17.82$\pm$0.39 \\
LFM+DARTS+INP& 17.65$\pm$0.45 \\
DARTS& 20.58$\pm$0.44\\
\hline
LFM+PDARTS, no INP & 16.51$\pm$0.13 \\
LFM+PDARTS+INP& 16.44$\pm$0.11 \\
PDARTS &17.49\\
\hline
\end{tabular}
\end{center}
\caption{Comparisons on different pre-training datasets of the encoder. InP means ImageNet pretrain.}
\label{tab:pre}
\end{table}
\subsection{Limitations and Future Work}\label{sec:limitations}
Our method requires the use of two learners that have similar learning capabilities, so that one can learn from the mistakes of other. This increases the memory requirements and makes the learning slow compared to the vanilla approaches. As future work, we explore to reduce memory cost during architecture search by parameter-sharing between the three models $W_1$, $W_2$, and $V$. For $W_1$ and $W_2$, we let them share the same convolutional layers, but have different classification heads. For $V$, we replace ResNet-18 with $W_1$. As shown in Table \ref{tab:memory} that via parameter sharing (PS), the memory and computation costs of our method are reduced to a level similar to vanilla DARTS and PDARTS, while our method still achieves significantly lower test errors than DARTS and PDARTS. A future work direction is to improve memory usage while keeping the full performance of the LFM method. Another direction is to extend the applicability of LFM to other meta-learning tasks such as data re-weighting, or more complex tasks like semantic segmentation. Further, LFM can be extended to language modeling tasks as well.
\begin{table}[ht]
\centering
\begin{footnotesize}
\begin{tabular}{l|ccc}
\toprule
\multirow{2}{*}{Method} & Test error & Memory & Cost \\
& (\%) & (MiB) & (days) \\
\hline
LFM+DARTS, no PS & 17.65$\pm$0.45 & 23,702 & 5.4 \\
LFM+DARTS+PS & 18.77$\pm$0.31 & 12,138 & 1.6 \\
DARTS & 20.58$\pm$0.44 & 11,053 & 1.5 \\
\hline
LFM+PDARTS, no PS & 16.44$\pm$0.11 & 20,744 & 2.0 \\
LFM+PDARTS+PS & 16.83$\pm$0.08 & 10,582 & 0.3 \\
PDARTS & 17.49 & 9,659 & 0.3 \\
\bottomrule
\end{tabular}
\end{footnotesize}
\caption{Test error(\%), memory cost (MiB) and computation cost (GPU days) of different models on CIFAR-100.}
\label{tab:memory}
\end{table}
\section{Conclusions}
In this paper, we propose a novel strategy to improve the performance of differential Neural Architecture Search approaches - Learning from Mistakes (LFM) inspired by the astounding ability of humans to learn from feedback on their current performance. The learner model continuously focuses more on the mistakes and improves its learning ability on more challenging examples. We propose a multi-level optimization framework to formalize LFM and provide an efficient solution to the problem. We apply LFM to NAS on datasets of CIFAR-10, CIFAR-100, and Imagenet - pushing the frontiers on NAS research.
|
\section{Introduction} \label{sec:intro}
Observations of Galactic Center (GC) over the past several decades have identified stellar SiO masers within a few parsecs of the central galactic black hole Sgr A*. The SiO maser emission originates in the extended envelopes of late type stars and provides high brightness temperature sources for astrometry and narrow spectral profiles for Doppler velocity tracking. Thus, stellar masers are useful as high spatial and spectral resolution tracers of 3D stellar kinematics in the GC.
Circumstellar SiO masers are associated with asymptotic giant branch (AGB) stars, which are late-type red giant stars (see \citealt{Kemball2007} for review). Several lines are observed at frequencies around 43 and 86 GHz, which are rotational transitions in predominantly excited vibrational states, though a smaller number in ground vibrational states have been observed. Stellar SiO masers have been resolved with Very Long Baseline Interferometry (VLBI) observations, showing that the maser emission originates from discrete regions typically within about a few AU around the star (e.g. \citealt{Gonidakis2010}). When the emission is not resolved, SiO masers can be treated as point sources which track the stellar position. The cumulative spectrum typically shows maser emission in a range around $10$ km s$^{-1}$~ relative to the systemic stellar velocity and may have multiple distinct peaks.
Astrometry and velocity tracking of SiO masers in the GC has been ongoing for several decades (e.g. \citealt{Menten1997,Reid2003,Reid2007,Li2010,Borkar2019}). Several masers associated with infrared-bright stars have also been utilized to construct an astrometric reference frame to improve measurements of infrared (IR) stars on short-period orbits around Sgr A* (\citealt{Yelda2010}; \citealt{Plewa2015}; \citealt{Boehle2016}; \citealt{Sakai2019}). The orbits of these IR stars provided some of the first evidence for the existence of black holes and revealed the presence of a $4.15\times10^6$ M$_\odot$ mass black hole \citep{Ghez2008,Genzel2010}. The recent pericentre passage of the star S2 showed a detectable gravitational redshift \citep{GRAVITYCollaboration2018} and Schwarzschild precession \citep{GRAVITYCollaboration2020}. Such precision measurements depend on improvements to the astrometric reference frame derived from SiO masers.
SiO maser stars may also directly probe Sgr A* and the surrounding environment. Stellar kinematics trace the underlying mass distribution, and therefore enable mapping of the total stellar and dark matter mass profiles at various radii in the GC. Additionally, if high velocity stellar masers with short orbital periods are identified, masers may enable measurements of the metric around Sgr A*.
The required accuracy assumed for most general relativistic spectroscopic tests of stars near the central black hole is 10 km s$^{-1}$~ \citep{Angelil2011}, but the spectral resolution for SiO masers is at least two orders of magnitude better, with statistical uncertainties $\sim 10$ - 100 m~s$^{-1}$.
Accurately tracking the systemic stellar velocity using the maser spectra is often a challenge as we demonstrate in this work. However, if the systematics can be quantified, general relativistic tests may be possible with SiO maser orbit measurements, such as tests of the equivalence principle or “no hair” theorem.
We present new observations from the Karl G. Jansky Very Large Array \footnote{The National Radio Astronomy Observatory is a facility of the National Science Foundation operated under cooperative agreement by Associated Universities, Inc.} (VLA) and Atacama Large Millimeter/submillimeter Array (ALMA) of stellar SiO masers within $\sim 2$ pc of Sgr A*.
Observations of the 43 and 86 GHz masers and data reduction methods are detailed in Section \ref{sec:observations}. Results are presented in Section \ref{sec:results}, including stellar proper motions, proper accelerations, radial velocities, and radial accelerations. Analysis of the 3D stellar kinematics is given in Section \ref{sec:kinematics}. Finally, we discuss the implications of the stellar motions in Section \ref{sec:discussion} and give conclusions in Section \ref{sec:conclusions}.
\begin{deluxetable*}{lllccc}
\tablecaption{VLA and ALMA observations \label{tab:obs}}
\tablehead{
\colhead{Telescope} & \colhead{Date} & \colhead{MJD} & \colhead{Lines observed} & \colhead{Beam} & \colhead{Channel width} \\
\colhead{} & \colhead{} & \colhead{} & \colhead{} & \colhead{(" $\times$ ")} & \colhead{(km s$^{-1}$)}
}
\startdata
VLA & 2013 Feb 14 & 2013.13 & $J=1-0$, $v=1$ and $v=2$ & $3.398 \times 1.444$ & 1.0\\
VLA & 2014 Mar 7 & 2014.18 & $J=1-0$, $v=1$ and $v=2$ & $0.1894 \times 0.1032$ & 1.0 \\
ALMA & 2015 Apr 10 & 2015.27 & $J=2-1$, $v=1$ & $4.112 \times 2.512$ & 1.75 \\
ALMA & 2017 Sep 19 & 2017.72 & $J=2-1$, $v=1$ & $0.625 \times 0.3125$ & 0.25 \\
VLA & 2020 Dec 27 & 2020.99 & $J=1-0$, $v=1$ and $v=2$ & $0.1230 \times 0.0483$ & 0.5
\enddata
\end{deluxetable*}
\section{Observations and Data Reduction}\label{sec:observations}
Observations of the inner few parsec of the GC around Sgr A* were carried out over the course of seven years with VLA and ALMA. Three VLA sessions were used to observe the $v=1$ and $v=2$, $J=1-0$ lines at 43.12203 and 42.82048 GHz, and two ALMA sessions were used to observe the $v=1$, $J=2-1$ line at 86.24337 GHz.
The ground state transitions $v=0$, $J=1-0$ and $v=0$, $J=2-1$ were also covered in our observations, but no ground state masers were observed in any VLA or ALMA sessions.
Initial observations with both VLA and ALMA were intended for high velocity maser searches and therefore have lower angular and spectral resolution than followup observations. We detail the observations, calibration, and data reduction in the following subsections.
A summary of the observing sessions is given in Table \ref{tab:obs}.
\subsection{VLA Observations}\label{sec:vla}
We observed the GC with the VLA in the D configuration on 2013 February 14 (13A-071), and in the A configuration on 2014 March 7 (14A-440) and 2020 December 27 (19A-310). The phase center for all observations was Sgr A*. During the 2013 observation, the VLA had 26 antennas in operation with baselines ranging from $\sim$40 to 1486 m, resulting in a $3.398" \times 1.444"$ synthesized beam.
The spectral setup consisted of 32 contiguous spectral windows, each with 1024 channels with 125 kHz channel width in dual polarization. The spectral windows covered frequencies from 42.1121 to 44.1276 GHz.
During calibration, channels between frequencies 43.1171 and 43.1235 GHz had to be discarded due to bandpass edge roll-off, which corresponds to velocities of about $-13$ - $+34$ km s$^{-1}$\ for $v=1$, $J=1-0$ (43.12203 GHz) masers.
The 2014 observation was conducted with 27 antennas in operation and baselines ranging from 793 to 36623 m. The synthesized beam was $0.1894" \times 0.1032"$. The spectral setup was similar to that of 2013, but covered a smaller range of frequencies. The observation had 16 contiguous spectral windows in dual polarization mode, each with 1024 channels with 125 kHz channel width, and covered frequencies from 42.6307 to 43.685 GHz.
The 2020 observation was conducted with 26 antennas in operation and baselines ranging from 793 to 36623 m. The synthesized beam was $0.1230" \times 0.0483"$. The spectral configuration included four spectral windows, each with 62.5 kHz wide channels in dual polarization mode. Three spectral windows contained 2048 channels and were centered on 42.612, 43.124, and 43.380 GHz with 128 MHz bandwidth each. A fourth smaller spectral window contained 1024 channels and was centered on 42.836 GHz with 64 MHz bandwidth.
We performed standard calibration using the Common Astronomy Software Application (CASA; \citealt{McMullin2007}). For all observations, 1733-1304 was used for bandpass and initial phase calibration and 1331+305 (3C286) for the flux density scale. Phases and amplitudes were self-calibrated with the Sgr A* continuum emission. The continuum emission was then subtracted in the $uv$-plane by fitting to the line-free channels. Initial spectral maps were made using the CLEAN algorithm, and these images were used to identify masers by eye.
The measurement sets were regridded to the LSRK frame and channel widths of 1.0 km s$^{-1}$~ for the 2013 and 2014 data, and 0.5 km s$^{-1}$~ for the 2020 data.
We then fit maser positions and spectra to the calibrated, continuum subtracted visibility data directly using the package \emph{uvmultifit} \citep{Marti-Vidal2014}. Positions are measured as offsets from the phase center, which is the Sgr A* continuum. Visibility fitting, rather than image fitting, was done in order to mitigate any systematics introduced between different telescopes, frequencies, and epochs.
\subsection{ALMA Observations}\label{sec:alma}
We observed the 86 GHz masers in the GC with ALMA on 2015 April 10 (2013.1.00834) and 2017 September 19 (2016.1.00940). The 2015 observation was conducted in the C34-1/(2) configuration, with 35 antennas in operation and baselines ranging from 15 to 348 m. The resulting synthesized beam was $4.112" \times 2.512"$. The spectral configuration consisted of four contiguous spectral windows in dual linear polarization, each with 1920 channels with 488.281 kHz channel width. The spectral windows covered frequencies from 84.842 to 88.5688 GHz.
The 2017 observation was conducted in the more extended C40-8/9 configuration, with 44 antennas in operation and baselines ranging from 41 to 12145 m. The resulting synthesized beam was $0.625" \times 0.3125"$. The spectral setup contained four spectral windows in dual linear polarization which were intended for higher spectral resolution than the previous ALMA setup. One window contained 3840 channels with 61.035 kHz width and was centered on 86.355026 GHz. The other three windows each contained 1920 channels with 122.07 kHz width and were centered on 85.752122, 86.958551, and 85.980837 GHz.
Calibrations for both observations were performed initially using the standard ALMA pipeline in CASA. Further self-calibration was done using the Sgr A* continuum emission. The remaining data reduction followed in a similar manner to that described above for the VLA observations. The Sgr A* continuum emission was subtracted from the visibility data by fitting to the line-free channels. Masers were initially identified by eye from spectral line maps created using the CLEAN algorithm to determine initial positions for visibility fitting. The measurement sets were regridded to the LSRK frame and channel widths of 1.75 km s$^{-1}$~ for the 2015 data and 0.25 km s$^{-1}$~ for the 2017 data. We then fit maser offset positions and spectra to the visibility data using the \emph{uvmultifit} package.
\startlongtable
\begin{deluxetable*}{llcccc}
\tabletypesize{\scriptsize}
\tablecaption{Maser positions and velocities \label{tab:masers}}
\tablehead{
\colhead{Star} & \colhead{Epoch} & \colhead{Line} & \colhead{R.A. offset} & \colhead{Dec. offset} & \colhead{Radial velocity} \\
\colhead{} & \colhead{(year)} & \colhead{} & \colhead{(arcsec)} & \colhead{(arcsec)} & \colhead{(km s$^{-1}$)}
}
\startdata
IRS9 & 2013.13 & $J=1-0$, $v=1$ & $5.69 \pm 0.04$ & $-6.2 \pm 0.1$ & $-342.7 \pm 0.2$ \vspace{3pt} \\
& 2014.18 & $J=1-0$, $v=1$ & $5.6986 \pm 0.0002$ & $-6.3157 \pm 0.0006$ & $-339.26 \pm 0.07$ \\
& 2014.18 & $J=1-0$, $v=2$ & $5.6983 \pm 0.0003$ & $-6.3155 \pm 0.0006$ & $-337.82 \pm 0.07$ \vspace{3pt} \\
& 2015.27 & $J=2-1$, $v=1$ & $5.98 \pm 0.05$ & $-6.30 \pm 0.03$ & $-342.0 \pm 0.1$ \vspace{3pt} \\
& 2017.72 & $J=2-1$, $v=1$ & $5.7111 \pm 0.0002$ & $-6.3056 \pm 0.0002$ & $-341.36 \pm 0.01$ \vspace{3pt} \\
& 2020.99 & $J=1-0$, $v=1$ & $5.7188 \pm 0.0003$ & $-6.3009 \pm 0.0006$ & $-341.57 \pm 0.04$ \\
\hline
IRS7 & 2014.18 & $J=1-0$, $v=1$ & $0.0323 \pm 0.0007$ & $5.487 \pm 0.002$ & $-106.8 \pm 0.2$ \\
& 2014.18 & $J=1-0$, $v=2$ & $0.0334 \pm 0.0007$ & $5.487 \pm 0.002$ & $-104.96 \pm 0.08$ \vspace{3pt} \\
& 2015.27 & $J=2-1$, $v=1$ & $0.034 \pm 0.003$ & $5.526 \pm 0.002$ & $-119.50 \pm 0.03$ \vspace{3pt} \\
& 2017.72 & $J=2-1$, $v=1$ & $0.03304 \pm 0.00009$ & $5.47053 \pm 0.00009$ & $-120.28 \pm 0.03$ \vspace{3pt} \\
& 2020.99 & $J=1-0$, $v=1$ & $0.038 \pm 0.001$ & $5.459 \pm 0.003$ & $-108.4 \pm 0.1$ \\
\hline
SiO-14 & 2013.13 & $J=1-0$, $v=1$ & $-7.57 \pm 0.04$ & $-28.2 \pm 0.1$ & $-112.2 \pm 0.2$ \\
& 2013.13 & $J=1-0$, $v=2$ & $-7.56 \pm 0.03$ & $-28.39 \pm 0.08$ & $-109.5 \pm 0.2$ \vspace{3pt} \\
& 2014.18 & $J=1-0$, $v=2$ & $-7.6321 \pm 0.0004$ & $-28.4614 \pm 0.0008$ & $-112.42 \pm 0.07$ \\
& 2014.18 & $J=1-0$, $v=1$ & $-7.6312 \pm 0.0003$ & $-28.4612 \pm 0.0008$ & $-111.96 \pm 0.06$ \vspace{3pt} \\
& 2015.27 & $J=2-1$, $v=1$ & $-7.552 \pm 0.007$ & $-28.413 \pm 0.004$ & $-112.76 \pm 0.02$ \vspace{3pt} \\
& 2017.72 & $J=2-1$, $v=1$ & $-7.6247 \pm 0.0002$ & $-28.4698 \pm 0.0002$ & $-113.06 \pm 0.02$ \vspace{3pt} \\
& 2020.99 & $J=1-0$, $v=2$ & $-7.6185 \pm 0.0005$ & $-28.467 \pm 0.001$ & $-115.92 \pm 0.08$ \\
& 2020.99 & $J=1-0$, $v=1$ & $-7.6185 \pm 0.0005$ & $-28.468 \pm 0.001$ & $-115.55 \pm 0.07$ \\
\hline
SiO-28\tablenotemark{a} & 2014.18 & $J=1-0$, $v=2$ & $-1.113 \pm 0.002$ & $-42.555 \pm 0.005$ & $-103.8 \pm 0.2$ \\
& 2014.18 & $J=1-0$, $v=1$ & $-1.112 \pm 0.002$ & $-42.559 \pm 0.004$ & $-103.11 \pm 0.08$ \vspace{3pt} \\
& 2015.27 & $J=2-1$, $v=1$ & $-0.98 \pm 0.08$ & $-42.54 \pm 0.05$ & $-104.5 \pm 0.1$ \vspace{3pt} \\
& 2017.72 & $J=2-1$, $v=1$ & $-1.114 \pm 0.002$ & $-42.568 \pm 0.002$ & $-105.48 \pm 0.03$ \vspace{3pt} \\
& 2020.99 & $J=1-0$, $v=2$ & $-1.117 \pm 0.002$ & $-42.544 \pm 0.005$ & $-103.67 \pm 0.06$ \\
& 2020.99 & $J=1-0$, $v=1$ & $-1.117 \pm 0.002$ & $-42.545 \pm 0.004$ & $-102.83 \pm 0.06$ \\
\hline
SiO-29\tablenotemark{a} & 2015.27 & $J=2-1$, $v=1$ & $-26.7 \pm 0.1$ & $23.57 \pm 0.07$ & $-93.0 \pm 0.2$ \vspace{3pt} \\
& 2017.72 & $J=2-1$, $v=1$ & $-26.836 \pm 0.002$ & $23.604 \pm 0.002$ & $-90.20 \pm 0.07$ \\
\hline
SiO-31\tablenotemark{a} & 2020.99 & $J=1-0$, $v=1$ & $9.359 \pm 0.002$ & $-18.338 \pm 0.004$ & $-81.1 \pm 0.1$ \\
\hline
SiO-18 & 2014.18 & $J=1-0$, $v=1$ & $-18.684 \pm 0.001$ & $-26.095 \pm 0.003$ & $-78.5 \pm 0.1$ \vspace{3pt} \\
& 2015.27 & $J=2-1$, $v=1$ & $-18.48 \pm 0.03$ & $-25.98 \pm 0.02$ & $-79.02 \pm 0.08$ \vspace{3pt} \\
& 2017.72 & $J=2-1$, $v=1$ & $-18.625 \pm 0.002$ & $-26.077 \pm 0.002$ & $-74.57 \pm 0.07$ \vspace{3pt} \\
& 2020.99 & $J=1-0$, $v=1$ & $-18.715 \pm 0.003$ & $-26.094 \pm 0.007$ & $-76.9 \pm 0.2$ \\
& 2020.99 & $J=1-0$, $v=2$ & $-18.710 \pm 0.002$ & $-26.096 \pm 0.004$ & $-76.49 \pm 0.07$ \\
\hline
SiO-26 & 2015.27 & $J=2-1$, $v=1$ & $22.31 \pm 0.05$ & $23.49 \pm 0.03$ & $-72.5 \pm 0.1$ \vspace{3pt} \\
& 2017.72 & $J=2-1$, $v=1$ & $22.499 \pm 0.001$ & $23.477 \pm 0.001$ & $-72.23 \pm 0.05$ \\
\hline
IRS12N & 2013.13 & $J=1-0$, $v=2$ & $-3.23 \pm 0.03$ & $-6.86 \pm 0.07$ & $-64.6 \pm 0.1$ \\
& 2013.13 & $J=1-0$, $v=1$ & $-3.27 \pm 0.03$ & $-6.91 \pm 0.09$ & $-64.6 \pm 0.1$ \vspace{3pt} \\
& 2014.18 & $J=1-0$, $v=2$ & $-3.2714 \pm 0.0004$ & $-6.931 \pm 0.001$ & $-66.02 \pm 0.08$ \\
& 2014.18 & $J=1-0$, $v=1$ & $-3.2702 \pm 0.0004$ & $-6.9318 \pm 0.0008$ & $-65.41 \pm 0.07$ \vspace{3pt} \\
& 2015.27 & $J=2-1$, $v=1$ & $-3.254 \pm 0.007$ & $-6.898 \pm 0.005$ & $-64.78 \pm 0.03$ \vspace{3pt} \\
& 2017.72 & $J=2-1$, $v=1$ & $-3.2755 \pm 0.0001$ & $-6.9412 \pm 0.0001$ & $-62.64 \pm 0.01$ \vspace{3pt} \\
& 2020.99 & $J=1-0$, $v=1$ & $-3.2796 \pm 0.0002$ & $-6.9494 \pm 0.0005$ & $-63.87 \pm 0.04$ \\
& 2020.99 & $J=1-0$, $v=2$ & $-3.2793 \pm 0.0002$ & $-6.9491 \pm 0.0004$ & $-63.86 \pm 0.04$ \\
\hline
IRS28 & 2013.13 & $J=1-0$, $v=2$ & $10.50 \pm 0.05$ & $-5.9 \pm 0.1$ & $-53.8 \pm 0.1$ \vspace{3pt} \\
& 2014.18 & $J=1-0$, $v=2$ & $10.4919 \pm 0.0002$ & $-5.8675 \pm 0.0004$ & $-53.84 \pm 0.03$ \\
& 2014.18 & $J=1-0$, $v=1$ & $10.4916 \pm 0.0002$ & $-5.8673 \pm 0.0004$ & $-53.41 \pm 0.03$ \vspace{3pt} \\
& 2015.27 & $J=2-1$, $v=1$ & $10.48 \pm 0.01$ & $-5.829 \pm 0.008$ & $-53.89 \pm 0.04$ \vspace{3pt} \\
& 2017.72 & $J=2-1$, $v=1$ & $10.4965 \pm 0.0003$ & $-5.8875 \pm 0.0003$ & $-52.53 \pm 0.02$ \vspace{3pt} \\
& 2020.99 & $J=1-0$, $v=2$ & $10.502 \pm 0.002$ & $-5.911 \pm 0.004$ & $-55.37 \pm 0.08$ \\
& 2020.99 & $J=1-0$, $v=1$ & $10.5029 \pm 0.0009$ & $-5.914 \pm 0.002$ & $-52.91 \pm 0.08$ \\
\hline
SiO-27 & 2014.18 & $J=1-0$, $v=1$ & $-19.930 \pm 0.002$ & $33.672 \pm 0.005$ & $-44.0 \pm 0.2$ \vspace{3pt} \\
& 2015.27 & $J=2-1$, $v=1$ & $-19.82 \pm 0.04$ & $33.68 \pm 0.03$ & $-43.67 \pm 0.07$ \vspace{3pt} \\
& 2017.72 & $J=2-1$, $v=1$ & $-19.9361 \pm 0.0005$ & $33.6805 \pm 0.0005$ & $-43.50 \pm 0.03$ \vspace{3pt} \\
& 2020.99 & $J=1-0$, $v=2$ & $-19.940 \pm 0.002$ & $33.697 \pm 0.005$ & $-44.3 \pm 0.4$ \\
& 2020.99 & $J=1-0$, $v=1$ & $-19.940 \pm 0.002$ & $33.695 \pm 0.004$ & $-42.9 \pm 0.2$ \\
\hline
SiO-30\tablenotemark{a} & 2015.27 & $J=2-1$, $v=1$ & $-23.13 \pm 0.08$ & $21.38 \pm 0.05$ & $-38.6 \pm 0.1$ \vspace{3pt} \\
& 2017.72 & $J=2-1$, $v=1$ & $-23.024 \pm 0.002$ & $21.314 \pm 0.002$ & $-37.34 \pm 0.04$ \\
\hline
SiO-15 & 2014.18 & $J=1-0$, $v=2$ & $-12.4597 \pm 0.0003$ & $-11.0673 \pm 0.0008$ & $-35.35 \pm 0.05$ \\
& 2014.18 & $J=1-0$, $v=1$ & $-12.4593 \pm 0.0007$ & $-11.067 \pm 0.002$ & $-35.2 \pm 0.1$ \vspace{3pt} \\
& 2015.27 & $J=2-1$, $v=1$ & $-12.42 \pm 0.03$ & $-11.01 \pm 0.02$ & $-36.45 \pm 0.06$ \vspace{3pt} \\
& 2017.72 & $J=2-1$, $v=1$ & $-12.4667 \pm 0.0008$ & $-11.0682 \pm 0.0008$ & $-33.91 \pm 0.04$ \vspace{3pt} \\
& 2020.99 & $J=1-0$, $v=2$ & $-12.474 \pm 0.001$ & $-11.070 \pm 0.002$ & $-36.1 \pm 0.1$ \\
& 2020.99 & $J=1-0$, $v=1$ & $-12.475 \pm 0.001$ & $-11.070 \pm 0.003$ & $-35.8 \pm 0.2$ \\
\hline
SiO-19 & 2013.13 & $J=1-0$, $v=2$ & $16.22 \pm 0.06$ & $-21.6 \pm 0.2$ & $-28.8 \pm 0.1$ \vspace{3pt} \\
& 2014.18 & $J=1-0$, $v=2$ & $16.2447 \pm 0.0003$ & $-21.6647 \pm 0.0007$ & $-28.71 \pm 0.07$ \\
& 2014.18 & $J=1-0$, $v=1$ & $16.2442 \pm 0.0004$ & $-21.664 \pm 0.001$ & $-27.80 \pm 0.08$ \vspace{3pt} \\
& 2015.27 & $J=2-1$, $v=1$ & $16.25 \pm 0.01$ & $-21.605 \pm 0.009$ & $-28.58 \pm 0.04$ \vspace{3pt} \\
& 2017.72 & $J=2-1$, $v=1$ & $16.2549 \pm 0.0008$ & $-21.6745 \pm 0.0008$ & $-31.29 \pm 0.04$ \vspace{3pt} \\
& 2020.99 & $J=1-0$, $v=1$ & $16.2626 \pm 0.0008$ & $-21.659 \pm 0.002$ & $-26.18 \pm 0.08$ \\
& 2020.99 & $J=1-0$, $v=2$ & $16.261 \pm 0.001$ & $-21.651 \pm 0.003$ & $-25.80 \pm 0.07$ \\
\hline
IRS10EE & 2013.13 & $J=1-0$, $v=2$ & $7.68 \pm 0.01$ & $4.19 \pm 0.04$ & $-27.26 \pm 0.07$ \vspace{3pt} \\
& 2014.18 & $J=1-0$, $v=2$ & $7.6838 \pm 0.0002$ & $4.1846 \pm 0.0004$ & $-27.48 \pm 0.02$ \\
& 2014.18 & $J=1-0$, $v=1$ & $7.6839 \pm 0.0003$ & $4.1843 \pm 0.0006$ & $-26.98 \pm 0.03$ \vspace{3pt} \\
& 2015.27 & $J=2-1$, $v=1$ & $7.68 \pm 0.02$ & $4.22 \pm 0.01$ & $-28.00 \pm 0.05$ \vspace{3pt} \\
& 2017.72 & $J=2-1$, $v=1$ & $7.6844 \pm 0.0002$ & $4.1767 \pm 0.0002$ & $-28.56 \pm 0.01$ \vspace{3pt} \\
& 2020.99 & $J=1-0$, $v=1$ & $7.68504 \pm 0.00009$ & $4.1701 \pm 0.0002$ & $-27.68 \pm 0.02$ \\
& 2020.99 & $J=1-0$, $v=2$ & $7.68490 \pm 0.00008$ & $4.1702 \pm 0.0002$ & $-27.59 \pm 0.01$ \\
\hline
SiO-20 & 2014.18 & $J=1-0$, $v=1$ & $-13.8629 \pm 0.0004$ & $20.3643 \pm 0.0009$ & $-17.11 \pm 0.07$ \\
& 2014.18 & $J=1-0$, $v=2$ & $-13.8625 \pm 0.0006$ & $20.362 \pm 0.001$ & $-16.9 \pm 0.1$ \vspace{3pt} \\
& 2015.27 & $J=2-1$, $v=1$ & $-13.84 \pm 0.03$ & $20.41 \pm 0.02$ & $-17.13 \pm 0.08$ \vspace{3pt} \\
& 2017.72 & $J=2-1$, $v=1$ & $-13.8604 \pm 0.0006$ & $20.3502 \pm 0.0006$ & $-16.87 \pm 0.02$ \vspace{3pt} \\
& 2020.99 & $J=1-0$, $v=1$ & $-13.858 \pm 0.002$ & $20.343 \pm 0.005$ & $-20.8 \pm 0.1$ \\
& 2020.99 & $J=1-0$, $v=2$ & $-13.857 \pm 0.003$ & $20.344 \pm 0.006$ & $-19.55 \pm 0.05$ \\
\hline
IRS15NE & 2014.18 & $J=1-0$, $v=2$ & $1.1968 \pm 0.0003$ & $11.2295 \pm 0.0007$ & $-11.81 \pm 0.05$ \\
& 2014.18 & $J=1-0$, $v=1$ & $1.1970 \pm 0.0003$ & $11.2293 \pm 0.0006$ & $-11.04 \pm 0.07$ \vspace{3pt} \\
& 2015.27 & $J=2-1$, $v=1$ & $1.24 \pm 0.03$ & $11.25 \pm 0.02$ & $-11.02 \pm 0.05$ \vspace{3pt} \\
& 2017.72 & $J=2-1$, $v=1$ & $1.1903 \pm 0.0004$ & $11.2097 \pm 0.0004$ & $-11.78 \pm 0.03$ \vspace{3pt} \\
& 2020.99 & $J=1-0$, $v=2$ & $1.1847 \pm 0.0005$ & $11.193 \pm 0.001$ & $-14.56 \pm 0.09$ \\
& 2020.99 & $J=1-0$, $v=1$ & $1.1855 \pm 0.0008$ & $11.196 \pm 0.002$ & $-12.4 \pm 0.1$ \\
\hline
IRS14NE & 2014.18 & $J=1-0$, $v=1$ & $0.935 \pm 0.002$ & $-8.159 \pm 0.004$ & $-9.4 \pm 0.1$ \vspace{3pt} \\
& 2015.27 & $J=2-1$, $v=1$ & $0.91 \pm 0.05$ & $-8.31 \pm 0.03$ & $-12.2 \pm 0.1$ \vspace{3pt} \\
& 2017.72 & $J=2-1$, $v=1$ & $0.9468 \pm 0.0009$ & $-8.1701 \pm 0.0009$ & $-11.63 \pm 0.05$ \vspace{3pt} \\
& 2020.99 & $J=1-0$, $v=2$ & $0.965 \pm 0.002$ & $-8.182 \pm 0.005$ & $-12.4 \pm 0.2$ \\
\hline
SiO-16 & 2015.27 & $J=2-1$, $v=1$ & $-26.33 \pm 0.03$ & $-34.39 \pm 0.02$ & $7.90 \pm 0.08$ \vspace{3pt} \\
& 2017.72 & $J=2-1$, $v=1$ & $-26.4178 \pm 0.0009$ & $-34.4788 \pm 0.0009$ & $6.36 \pm 0.03$ \\
\hline
SiO-21 & 2015.27 & $J=2-1$, $v=1$ & $40.88 \pm 0.03$ & $-21.98 \pm 0.02$ & $13.46 \pm 0.06$ \vspace{3pt} \\
& 2017.72 & $J=2-1$, $v=1$ & $40.911 \pm 0.001$ & $-22.043 \pm 0.001$ & $13.44 \pm 0.03$ \\
\hline
SiO-24 & 2015.27 & $J=2-1$, $v=1$ & $17.16 \pm 0.04$ & $-4.71 \pm 0.03$ & $18.10 \pm 0.08$ \vspace{3pt} \\
& 2017.72 & $J=2-1$, $v=1$ & $17.1976 \pm 0.0008$ & $-4.8225 \pm 0.0008$ & $17.95 \pm 0.03$ \vspace{3pt} \\
& 2020.99 & $J=1-0$, $v=1$ & $17.214 \pm 0.001$ & $-4.831 \pm 0.003$ & $19.0 \pm 0.1$ \\
& 2020.99 & $J=1-0$, $v=2$ & $17.213 \pm 0.001$ & $-4.828 \pm 0.003$ & $19.5 \pm 0.1$ \\
\hline
SiO-22 & 2015.27 & $J=2-1$, $v=1$ & $41.44 \pm 0.03$ & $15.21 \pm 0.02$ & $32.79 \pm 0.06$ \vspace{3pt} \\
& 2017.72 & $J=2-1$, $v=1$ & $41.4088 \pm 0.0007$ & $15.1811 \pm 0.0007$ & $35.41 \pm 0.03$ \\
\hline
SiO-6 & 2015.27 & $J=2-1$, $v=1$ & $35.15 \pm 0.02$ & $30.72 \pm 0.01$ & $51.58 \pm 0.09$ \vspace{3pt} \\
& 2017.72 & $J=2-1$, $v=1$ & $35.2719 \pm 0.0005$ & $30.7032 \pm 0.0005$ & $52.17 \pm 0.05$ \\
\hline
SiO-17 & 2013.13 & $J=1-0$, $v=2$ & $8.111 \pm 0.009$ & $-27.47 \pm 0.02$ & $53.17 \pm 0.03$ \\
& 2013.13 & $J=1-0$, $v=1$ & $8.10 \pm 0.02$ & $-27.51 \pm 0.04$ & $53.59 \pm 0.05$ \vspace{3pt} \\
& 2014.18 & $J=1-0$, $v=2$ & $8.0811 \pm 0.0001$ & $-27.6578 \pm 0.0002$ & $53.37 \pm 0.02$ \\
& 2014.18 & $J=1-0$, $v=1$ & $8.0811 \pm 0.0001$ & $-27.6585 \pm 0.0003$ & $53.73 \pm 0.02$ \vspace{3pt} \\
& 2015.27 & $J=2-1$, $v=1$ & $8.07 \pm 0.02$ & $-27.64 \pm 0.02$ & $53.45 \pm 0.03$ \vspace{3pt} \\
& 2017.72 & $J=2-1$, $v=1$ & $8.0900 \pm 0.0005$ & $-27.6538 \pm 0.0004$ & $53.40 \pm 0.02$ \vspace{3pt} \\
& 2020.99 & $J=1-0$, $v=2$ & $8.0978 \pm 0.0005$ & $-27.641 \pm 0.001$ & $54.20 \pm 0.04$ \\
& 2020.99 & $J=1-0$, $v=1$ & $8.0991 \pm 0.0007$ & $-27.641 \pm 0.002$ & $54.48 \pm 0.06$ \\
\hline
SiO-11 & 2015.27 & $J=2-1$, $v=1$ & $1.87 \pm 0.08$ & $40.36 \pm 0.05$ & $70.5 \pm 0.1$ \vspace{3pt} \\
& 2017.72 & $J=2-1$, $v=1$ & $1.764 \pm 0.001$ & $40.316 \pm 0.001$ & $70.94 \pm 0.05$ \vspace{3pt} \\
& 2020.99 & $J=1-0$, $v=2$ & $1.7714 \pm 0.0001$ & $40.3245 \pm 0.0003$ & $70.84 \pm 0.02$ \\
& 2020.99 & $J=1-0$, $v=1$ & $1.7717 \pm 0.0002$ & $40.3241 \pm 0.0004$ & $71.85 \pm 0.02$ \\
\hline
IRS17 & 2014.18 & $J=1-0$, $v=2$ & $13.133 \pm 0.001$ & $5.550 \pm 0.002$ & $74.5 \pm 0.2$ \\
& 2014.18 & $J=1-0$, $v=1$ & $13.132 \pm 0.002$ & $5.546 \pm 0.004$ & $75.6 \pm 0.3$ \vspace{3pt} \\
& 2015.27 & $J=2-1$, $v=1$ & $13.13 \pm 0.02$ & $5.61 \pm 0.01$ & $74.13 \pm 0.05$ \vspace{3pt} \\
& 2017.72 & $J=2-1$, $v=1$ & $13.1310 \pm 0.0003$ & $5.5480 \pm 0.0003$ & $71.81 \pm 0.02$ \vspace{3pt} \\
& 2020.99 & $J=1-0$, $v=2$ & $13.126 \pm 0.001$ & $5.543 \pm 0.003$ & $72.31 \pm 0.06$ \\
& 2020.99 & $J=1-0$, $v=1$ & $13.126 \pm 0.001$ & $5.543 \pm 0.003$ & $72.90 \pm 0.08$ \\
\hline
IRS19NW & 2013.13 & $J=1-0$, $v=1$ & $14.57 \pm 0.05$ & $-18.3 \pm 0.1$ & $84.4 \pm 0.1$ \\
& 2013.13 & $J=1-0$, $v=2$ & $14.46 \pm 0.05$ & $-18.4 \pm 0.1$ & $84.4 \pm 0.1$ \vspace{3pt} \\
& 2014.18 & $J=1-0$, $v=2$ & $14.5695 \pm 0.0003$ & $-18.4681 \pm 0.0007$ & $84.06 \pm 0.05$ \\
& 2014.18 & $J=1-0$, $v=1$ & $14.5694 \pm 0.0003$ & $-18.4683 \pm 0.0006$ & $84.23 \pm 0.05$ \vspace{3pt} \\
& 2015.27 & $J=2-1$, $v=1$ & $14.53 \pm 0.02$ & $-18.44 \pm 0.01$ & $84.08 \pm 0.05$ \vspace{3pt} \\
& 2017.72 & $J=2-1$, $v=1$ & $14.5735 \pm 0.0008$ & $-18.4727 \pm 0.0008$ & $82.85 \pm 0.04$ \vspace{3pt} \\
& 2020.99 & $J=1-0$, $v=2$ & $14.5772 \pm 0.0002$ & $-18.4724 \pm 0.0005$ & $82.60 \pm 0.02$ \\
& 2020.99 & $J=1-0$, $v=1$ & $14.5774 \pm 0.0003$ & $-18.4728 \pm 0.0006$ & $83.68 \pm 0.04$ \\
\hline
SiO-25 & 2013.13 & $J=1-0$, $v=2$ & $-33.07 \pm 0.03$ & $-17.81 \pm 0.08$ & $118.19 \pm 0.07$ \\
& 2013.13 & $J=1-0$, $v=1$ & $-33.05 \pm 0.03$ & $-17.71 \pm 0.08$ & $118.80 \pm 0.05$ \vspace{3pt} \\
& 2014.18 & $J=1-0$, $v=2$ & $-33.056 \pm 0.002$ & $-17.716 \pm 0.005$ & $117.60 \pm 0.06$ \\
& 2014.18 & $J=1-0$, $v=1$ & $-33.061 \pm 0.002$ & $-17.728 \pm 0.005$ & $118.62 \pm 0.06$ \vspace{3pt} \\
& 2015.27 & $J=2-1$, $v=1$ & $-32.96 \pm 0.02$ & $-17.88 \pm 0.02$ & $118.01 \pm 0.06$ \vspace{3pt} \\
& 2017.72 & $J=2-1$, $v=1$ & $-33.105 \pm 0.002$ & $-17.917 \pm 0.002$ & $117.63 \pm 0.03$
\enddata
\tablenotetext{a}{New detection}
\end{deluxetable*}
\begin{figure}[ht!]
\centering
\includegraphics[width=\columnwidth]{ALMA2017_GC_masers_1pc}
\caption{A subset of the 86 GHz SiO masers near Sgr A* detected by ALMA in 2017. The field of view spans $\pm1$ pc. The beam is $0.1617 \times 0.1176$ arcsec ($\sim 6.4 \times 4.7$ mpc) and is shown in the lower left. \label{fig:ALMA masers}}
\end{figure}
\section{Results}\label{sec:results}
We required a $5\sigma$ significance flux measurement reported by \emph{uvmultifit} in at least one channel for detection. A subset of the masers within 1 pc projected distance from Sgr A* detected by ALMA in 2017 are shown in Figure \ref{fig:ALMA masers}. Maser positions where then determined by finding the error-weighted mean position for all channels with $>3\sigma$ measurements. The maser spectra, which are shown in Appendix~\ref{app:spectra}, were used to determine radial velocities by calculating the centroid of the $>3\sigma$ channels.
Offset positions referenced to the Sgr A* continuum centroid and radial velocities of each maser are given in Table~\ref{tab:masers}.
Two new masers were detected in the VLA data: we name these SiO-28 and SiO-31. Three new masers were detected in the ALMA data: SiO-28, SiO-29, and SiO-30.
New masers were named using the convention of \cite{Reid2007}, \cite{Li2010}, and \cite{Borkar2019}.
All new sources were detected in at least two epochs, except for SiO-31 which was only detected at 5$\sigma$-significance in two channels in the 2020 VLA data (see Figure~\ref{fig:spectra and fits}). Further observations will be required to confirm if SiO-31 is a genuine SiO maser source.
Maser positions are measured with typical uncertainties $<1$ mas for the higher resolution epochs (2014.18, 2017.72, and 2020.99) and around 10 mas for the lower resolution epochs (2013.13 and 2015.27). These offset uncertainties are dominated by the maser position fit since the position of Sgr A* is well-measured in all epochs. Extended continuum emission from the GC mini-spiral was detected in both of the lower resolution epochs which was resolved out in the higher resolution data. We note that this emission may impact the resulting astrometry, but these epochs do not strongly affect the proper motion fits described in the next section since the statistical uncertainties of the offset positions are relatively large.
The majority of our spectra show variability in time and between different lines (see Appendix \ref{app:spectra}). The most notable example is IRS 7, where the peak of maser emission from the 43 and 87 GHz spectra are significantly offset by more than 10 km s$^{-1}$, with strong SiO maser emission spread over $\sim 30$ km s$^{-1}$~ in the 87 GHz spectra. Since the VLA and ALMA spectra appear to be self-consistent over time, the variability of the IRS 7 spectra seems to be due to differences in the excitation conditions between the 43 and 87 GHz masers. Generally, however, SiO maser flux is known to be variable on timescales of $\sim 1$ year or longer following the underlying stellar pulsations (e.g. \citealt{Pardo2004}). Since the time interval between VLA and ALMA observations is $>1$ year, it is unclear if the smaller differences between 43 and 86 GHz spectra seen in the other stars in the sample are due to time-dependent variability in the star or differences in the excitation conditions. We also note that the peak and centroid velocities of the $J=1-0$, $v=1$ and $v-2$ spectra among stars with simultaneous detections of both lines often disagree, and we address this further in Section \ref{sec:doppler}.
The spectral variability of IRS 7 also indicates that its astrometry may not be as accurate as that of other stars in the sample. As noted by \cite{Reid2003}, the broad velocity range of the SiO emissions suggests that IRS 7 has a large maser shell that may span 10 mas. The maser position therefore may not match the stellar position to better than $\sim$5 mas. However, we report only the statistical uncertainties on the maser positions in Table \ref{tab:masers}.
\begin{deluxetable*}{lccccccc}
\tablecaption{Offsets, proper motions and proper accelerations of stellar masers \label{tab:pms}}
\tablewidth{700pt}
\tabletypesize{\scriptsize}
\tablehead{
\colhead{Star} & \colhead{$t_0$} & \colhead{$\Delta \alpha_0$} & \colhead{$\mu_\alpha$} & \colhead{$a_\alpha$} & \colhead{$\Delta \delta_0$} & \colhead{$\mu_\delta$} & \colhead{$a_\delta$} \\
\colhead{} & \colhead{(year)} & \colhead{(arcsec)} & \colhead{(mas yr$^{-1}$)} & \colhead{(mas yr$^{-2}$)} & \colhead{(arcsec)} & \colhead{(mas yr$^{-1}$)} & \colhead{(mas yr$^{-2}$)}
}
\startdata
IRS9 & $1998.39$ & $5.65051 \pm 0.00064$ & $3.057 \pm 0.042$ & $-0.0003 \pm 0.0036$ & $-6.3517 \pm 0.0011$ & $2.303 \pm 0.068$ & $-0.0001 \pm 0.0057$ \\
IRS7 & $1995.49$ & $0.0344 \pm 0.0050$ \tablenotemark{a} & $-0.04 \pm 0.11$ & $0.0037 \pm 0.0076$ & $5.5659 \pm 0.0050$\tablenotemark{a} & $-4.25 \pm 0.15$ & $-0.002 \pm 0.011$ \\
SiO-14 & $1998.41$ & $-7.66448 \pm 0.00034$ & $2.054 \pm 0.021$ & $-0.0007 \pm 0.0017$ & $-28.4517 \pm 0.0020$ & $-0.75 \pm 0.12$ & $-0.000 \pm 0.010$ \\
SiO-28 & $2014.18$ & $-1.11207 \pm 0.00037$ & $-0.712 \pm 0.080$ & $-0.006 \pm 0.023$ & $-42.5651 \pm 0.0087$ & $1.8 \pm 2.0$ & $0.31 \pm 0.54$ \\
SiO-29 & \nodata & \nodata & \nodata & \nodata & \nodata & \nodata & \nodata \\
SiO-31 & \nodata & \nodata & \nodata & \nodata & \nodata & \nodata & \nodata \\
SiO-18 & $2014.18$ & $-18.670 \pm 0.031$ & $-2.9 \pm 7.4$ & $-0.7 \pm 2.1$ & $-26.088 \pm 0.010$ & $1.0 \pm 2.7$ & $-0.43 \pm 0.77$ \\
SiO-26 & \nodata & \nodata & \nodata & \nodata & \nodata & \nodata & \nodata \\
IRS12N & $1996.41$ & $-3.25154 \pm 0.00038$ & $-1.120 \pm 0.023$ & $-0.0007 \pm 0.0018$ & $-6.88149 \pm 0.00039$ & $-2.778 \pm 0.023$ & $0.0004 \pm 0.0018$ \\
IRS28 & $1998.41$ & $10.4674 \pm 0.0010$ & $1.541 \pm 0.063$ & $-0.0005 \pm 0.0050$ & $-5.7813 \pm 0.0029$ & $-5.54 \pm 0.18$ & $-0.007 \pm 0.014$ \\
SiO-27 & $2014.18$ & $-19.93043 \pm 0.00061$ & $-1.52 \pm 0.13$ & $0.015 \pm 0.031$ & $33.6670 \pm 0.0028$ & $3.94 \pm 0.70$ & $0.08 \pm 0.15$ \\
SiO-30 & \nodata & \nodata & \nodata & \nodata & \nodata & \nodata & \nodata \\
SiO-15 & $2000.85$ & $-12.42632 \pm 0.00099$ & $-2.426 \pm 0.065$ & $0.0026 \pm 0.0053$ & $-11.0754 \pm 0.0031$ & $0.45 \pm 0.21$ & $-0.009 \pm 0.017$ \\
SiO-19 & $2014.18$ & $16.24461 \pm 0.00047$ & $2.63 \pm 0.12$ & $-0.011 \pm 0.035$ & $-21.6666 \pm 0.0045$ & $0.2 \pm 1.4$ & $0.22 \pm 0.42$ \\
IRS10EE & $1995.49$ & $7.68371 \pm 0.00024$ & $0.036 \pm 0.014$ & $0.0003 \pm 0.0010$ & $4.22082 \pm 0.00044$ & $-1.979 \pm 0.026$ & $-0.0001 \pm 0.0019$ \\
SiO-20 & $2014.18$ & $-13.86276 \pm 0.00018$ & $0.717 \pm 0.062$ & $0.005 \pm 0.020$ & $20.36321 \pm 0.00093$ & $-3.47 \pm 0.35$ & $0.07 \pm 0.13$ \\
IRS15NE & $1995.49$ & $1.23178 \pm 0.00039$ & $-1.852 \pm 0.023$ & $0.0006 \pm 0.0017$ & $11.33613 \pm 0.00079$ & $-5.661 \pm 0.048$ & $0.0013 \pm 0.0037$ \\
IRS14NE & $2014.18$ & $0.9334 \pm 0.0033$ & $4.29 \pm 0.85$ & $0.09 \pm 0.23$ & $-8.15823 \pm 0.00068$ & $-3.38 \pm 0.19$ & $-0.024 \pm 0.049$ \\
SiO-16 & $2008.42$ & $-26.44 \pm 0.13$ & $2 \pm 14$ & $-0.9 \pm 2.4$ & $-34.087 \pm 0.075$ & $-42.1 \pm 8.1$ & $0.2 \pm 1.0$ \\
SiO-21 & $2008.42$ & $40.754 \pm 0.022$ & $17.0 \pm 2.4$ & $-0.11 \pm 0.40$ & $-21.795 \pm 0.018$ & $-26.7 \pm 1.9$ & $-0.02 \pm 0.24$ \\
SiO-24 & $2015.27$ & $17.1853 \pm 0.0011$ & $4.99 \pm 0.26$ & $-0.002 \pm 0.063$ & $-4.8159 \pm 0.0084$ & $-2.5 \pm 2.4$ & $0.04 \pm 0.59$ \\
SiO-22 & $2008.42$ & $41.319 \pm 0.073$ & $9.7 \pm 7.8$ & $-0.5 \pm 1.4$ & $15.38 \pm 0.12$ & $-22 \pm 13$ & $0.4 \pm 1.6$ \\
SiO-6 & $2008.42$ & $34.98 \pm 0.12$ & $32 \pm 13$ & $0.7 \pm 1.9$ & $30.792 \pm 0.092$ & $-9.6 \pm 9.9$ & $0.1 \pm 1.2$ \\
SiO-17 & $1998.41$ & $8.04279 \pm 0.00044$ & $2.450 \pm 0.026$ & $0.0003 \pm 0.0022$ & $-27.7040 \pm 0.0027$ & $2.78 \pm 0.17$ & $-0.001 \pm 0.014$ \\
SiO-11 & $2008.42$ & $1.7424 \pm 0.0083$ & $2.32 \pm 0.69$ & $-0.002 \pm 0.064$ & $40.296 \pm 0.019$ & $2.3 \pm 1.5$ & $0.01 \pm 0.14$ \\
IRS17 & $2000.85$ & $13.1477 \pm 0.0012$ & $-1.052 \pm 0.078$ & $0.0018 \pm 0.0066$ & $5.5657 \pm 0.0015$ & $-1.083 \pm 0.097$ & $-0.0002 \pm 0.0093$ \\
IRS19NW & $1998.41$ & $14.55147 \pm 0.00028$ & $1.143 \pm 0.017$ & $-0.0000 \pm 0.0013$ & $-18.4621 \pm 0.0012$ & $-0.460 \pm 0.067$ & $-0.0011 \pm 0.0052$ \\
SiO-25 & $2013.13$ & $-33.0445 \pm 0.0054$ & $-13.2 \pm 1.8$ & $-0.02 \pm 0.64$ & $-17.669 \pm 0.015$ & $-54.1 \pm 3.9$ & $0.1 \pm 1.4$ \\
\enddata
\tablenotetext{a}{Position uncertainties for IRS 7 have been manually increased to 5 mas to reflect the fact that this star may have an extended maser shell as discussed in Section \ref{sec:results}.}
\end{deluxetable*}
\begin{deluxetable*}{lcccccccc}
\tablecaption{3D velocities and accelerations of stellar masers \label{tab:vels and accels}}
\tablewidth{700pt}
\tabletypesize{\scriptsize}
\tablehead{
\colhead{Star} & \colhead{$v_\alpha$} & \colhead{$a_\alpha$} & \colhead{$v_\delta$} & \colhead{$a_\delta$} & \colhead{$v_{LSR}$} & \colhead{$a_{LSR}$} & \colhead{$|v|$} & \colhead{$|a|$}\\
\colhead{} & \colhead{(km s$^{-1}$)} & \colhead{(km s$^{-1}$ yr$^{-1}$)} & \colhead{(km s$^{-1}$)} & \colhead{(km s$^{-1}$ yr$^{-1}$)} & \colhead{(km s$^{-1}$)} & \colhead{(km s$^{-1}$ yr$^{-1}$) } & \colhead{(km s$^{-1}$)} & \colhead{(km s$^{-1}$ yr$^{-1}$) }
}
\startdata
IRS9 & $115.9 \pm 1.6$ & $-0.01 \pm 0.14$ & $87.3 \pm 2.6$ & $-0.00 \pm 0.22$ & $-341.1 \pm 1.2$ & $-0.05 \pm 0.29$ & $370.7 \pm 1.4$ & $0.05 \pm 0.29$ \\
IRS7\tablenotemark{a} & $-1.4 \pm 4.3$ & $0.14 \pm 0.29$ & $-161.2 \pm 5.5$ & $-0.08 \pm 0.41$ & $-113.6 \pm 7.0$ & $0.1 \pm 1.8$ & $197.2 \pm 6.1$ & $0.19 \pm 0.90$ \\
SiO-14\tablenotemark{b} & $77.91 \pm 0.81$ & $-0.027 \pm 0.065$ & $-28.6 \pm 4.6$ & $-0.01 \pm 0.38$ & $-111.02 \pm 0.43$ & $-0.58 \pm 0.10$ & $138.6 \pm 1.1$ & $0.58 \pm 0.10$ \\
SiO-28 & $-27.0 \pm 3.0$ & $-0.23 \pm 0.87$ & $69 \pm 77$ & $12 \pm 20$ & $-104.16 \pm 0.98$ & $0.02 \pm 0.25$ & $128 \pm 41$ & $12 \pm 20$ \\
SiO-29 & \nodata & \nodata & \nodata & \nodata & $-91.58 \pm 0.98$ & \nodata & \nodata & \nodata \\
SiO-31 & \nodata & \nodata & \nodata & \nodata & $-81.10 \pm 0.13$ & \nodata & \nodata & \nodata \\
SiO-18 & $-110 \pm 280$ & $-28 \pm 80$ & $40 \pm 100$ & $-16 \pm 29$ & $-78.4 \pm 1.5$ & $0.42 \pm 0.38$ & $140 \pm 220$ & $33 \pm 71$ \\
SiO-26 & \nodata & \nodata & \nodata & \nodata & $-72.36 \pm 0.11$ & \nodata & \nodata & \nodata \\
IRS12N & $-42.47 \pm 0.85$ & $-0.026 \pm 0.066$ & $-105.35 \pm 0.88$ & $0.015 \pm 0.069$ & $-65.01 \pm 0.69$ & $0.22 \pm 0.16$ & $130.88 \pm 0.83$ & $0.22 \pm 0.16$ \\
IRS28 & $58.5 \pm 2.4$ & $-0.02 \pm 0.19$ & $-209.9 \pm 6.9$ & $-0.25 \pm 0.52$ & $-53.60 \pm 0.50$ & $-0.01 \pm 0.12$ & $224.4 \pm 6.5$ & $0.25 \pm 0.52$ \\
SiO-27 & $-57.5 \pm 5.1$ & $0.6 \pm 1.2$ & $150 \pm 27$ & $2.9 \pm 5.8$ & $-43.846 \pm 0.080$ & $0.099 \pm 0.025$ & $166 \pm 24$ & $3.0 \pm 5.7$ \\
SiO-30 & \nodata & \nodata & \nodata & \nodata & $-37.96 \pm 0.45$ & \nodata & \nodata & \nodata \\
SiO-15 & $-92.0 \pm 2.5$ & $0.10 \pm 0.20$ & $17.3 \pm 8.0$ & $-0.33 \pm 0.65$ & $-35.4 \pm 1.0$ & $0.01 \pm 0.26$ & $100.1 \pm 2.7$ & $0.34 \pm 0.62$ \\
SiO-19 & $99.7 \pm 4.5$ & $-0.4 \pm 1.3$ & $8 \pm 53$ & $9 \pm 16$ & $-29.3 \pm 1.4$ & $0.21 \pm 0.32$ & $104.2 \pm 6.0$ & $9 \pm 16$ \\
IRS10EE & $1.36 \pm 0.53$ & $0.010 \pm 0.039$ & $-75.07 \pm 0.98$ & $-0.005 \pm 0.072$ & $-27.53 \pm 0.38$ & $-0.070 \pm 0.091$ & $79.97 \pm 0.93$ & $0.071 \pm 0.090$ \\
SiO-20 & $27.2 \pm 2.3$ & $0.20 \pm 0.75$ & $-132 \pm 13$ & $2.7 \pm 5.0$ & $-16.62 \pm 0.67$ & $-0.37 \pm 0.17$ & $136 \pm 13$ & $2.8 \pm 5.0$ \\
IRS15NE\tablenotemark{b} & $-70.25 \pm 0.86$ & $0.021 \pm 0.063$ & $-214.7 \pm 1.8$ & $0.05 \pm 0.14$ & $-10.99 \pm 0.41$ & $-0.34 \pm 0.11$ & $226.2 \pm 1.7$ & $0.35 \pm 0.11$ \\
IRS14NE & $163 \pm 32$ & $3.4 \pm 8.7$ & $-128.1 \pm 7.0$ & $-0.9 \pm 1.8$ & $-10.50 \pm 0.95$ & $-0.32 \pm 0.25$ & $207 \pm 26$ & $3.5 \pm 8.4$ \\
SiO-16 & $90 \pm 520$ & $-33 \pm 91$ & $-1600 \pm 310$ & $8 \pm 39$ & $7.13 \pm 0.55$ & \nodata & $1600 \pm 310$ & \nodata \\
SiO-21 & $643 \pm 91$ & $-4 \pm 15$ & $-1012 \pm 72$ & $-0.8 \pm 9.0$ & $13.447 \pm 0.033$ & \nodata & $1199 \pm 78$ & \nodata \\
SiO-24 & $189.3 \pm 9.8$ & $-0.1 \pm 2.4$ & $-96 \pm 91$ & $2 \pm 22$ & $17.84 \pm 0.47$ & $0.22 \pm 0.13$ & $213 \pm 42$ & $2 \pm 22$ \\
SiO-22 & $370 \pm 300$ & $-18 \pm 53$ & $-820 \pm 480$ & $15 \pm 62$ & $34.10 \pm 0.93$ & \nodata & $900 \pm 450$ & \nodata \\
SiO-6 & $1200 \pm 490$ & $27 \pm 72$ & $-360 \pm 370$ & $5 \pm 47$ & $51.87 \pm 0.21$ & \nodata & $1250 \pm 480$ & \nodata \\
SiO-17 & $92.9 \pm 1.0$ & $0.011 \pm 0.083$ & $105.3 \pm 6.4$ & $-0.06 \pm 0.52$ & $53.27 \pm 0.17$ & $0.105 \pm 0.040$ & $150.2 \pm 4.5$ & $0.12 \pm 0.24$ \\
SiO-11\tablenotemark{b} & $88 \pm 26$ & $-0.1 \pm 2.4$ & $86 \pm 59$ & $0.5 \pm 5.4$ & $70.60 \pm 0.12$ & $0.105 \pm 0.029$ & $142 \pm 39$ & $0.5 \pm 5.2$ \\
IRS17 & $-39.9 \pm 3.0$ & $0.07 \pm 0.25$ & $-41.1 \pm 3.7$ & $-0.01 \pm 0.35$ & $74.36 \pm 0.80$ & $-0.36 \pm 0.21$ & $93.8 \pm 2.1$ & $0.37 \pm 0.21$ \\
IRS19NW & $43.34 \pm 0.63$ & $-0.001 \pm 0.049$ & $-17.4 \pm 2.5$ & $-0.04 \pm 0.20$ & $84.37 \pm 0.22$ & $-0.226 \pm 0.052$ & $96.44 \pm 0.57$ & $0.230 \pm 0.063$ \\
SiO-25\tablenotemark{b} & $-501 \pm 68$ & $-1 \pm 24$ & $-2050 \pm 150$ & $3 \pm 52$ & $118.48 \pm 0.10$ & $-0.196 \pm 0.038$ & $2120 \pm 140$ & $3 \pm 50$ \\
\enddata
\tablenotetext{a}{The velocity centroids of the VLA and ALMA spectra for IRS 7 are offset by $\sim$15 km s$^{-1}$. The radial velocity and acceleration measurements therefore may not accurately track the motion of the star.}
\tablenotetext{b}{Stars with $>3\sigma$ radial acceleration measurements, which may be impacted by maser spectral variability.}
\end{deluxetable*}
\subsection{Proper motions and proper accelerations}\label{sec:pm}
We combined our five epochs with extant maser astrometry of \citet{Reid2007} and \citet{Li2010} to calculate stellar proper motions and accelerations.
For the majority of masers, we included only our three higher resolution epochs (VLA 2014, ALMA 2017, and VLA 2020) and the \cite{Reid2007} astrometry to perform fits. However, for masers fewer than three high resolution epochs to fit, we also included our lower resolution data (VLA 2013 and ALMA 2015) and the \cite{Li2010} astrometry. \cite{Li2010} include observations of the 86 GHz masers from the Australia Telescope Compact Array (ATCA), which has lower resolution than VLA or ALMA astrometry and often showed large offsets from the trends in the masers' positions.
17 masers were fit with only the high resolution data, 7 masers included the lower resolution data, and 4 masers did not have enough epochs of observations for a fit.
All maser positions, including those from literature, are referenced to the in-beam Sgr A* continuum position. The resulting proper motions are therefore referenced to Sgr A* and should be independent of systematic differences in absolute astrometry between epochs. We first find proper motions by fitting a straight line to the maser positions over time, following the equation
\begin{equation}
\Delta x = \Delta x_0 + \mu_x \left( t - t_0 \right), \label{eq:proper motion}
\end{equation}
where $x$ denotes either the R.A. or Dec. coordinate directions, $\alpha$ and $\delta$, respectively. $\Delta x$ is the offset position from Sgr A*, $\mu_x$ is the proper motion, and $\Delta x_0$ and $t_0$ are the first position measurement and date used in the fit, which varies per maser depending on the earliest detection.
Proper accelerations were then measured by fitting the residuals of the proper motion fit, following the equation
\begin{equation}
r_{\Delta x} = r_{\Delta x_0} + \frac{1}{2} a_x \left( t - t_0 \right)^2, \label{eq:proper acceleration}
\end{equation}
where $r_{\Delta x}$ and $r_{\Delta x_0}$ are the residuals of the offset positions and initial positions, and $a_x$ is the proper acceleration. We measure proper motions and accelerations for 24 masers with at least three epochs of observations. Figures showing the proper motion fits are given in Appendix~\ref{app:pm fits} and the results are given in Table~\ref{tab:pms}.
We do not significantly measure proper accelerations for any star in our sample using the residual fitting method. As shown in Table~\ref{tab:pms}, all measurements are consistent with zero transverse acceleration in each coordinate and in magnitude. We also checked for acceleration by simultaneously fitting proper motion and acceleration to the position time series for masers with at least four epochs, however the results gave unrealistically large accelerations for some masers given their projected distance from Sgr A*.
\subsection{Transverse velocities and accelerations}\label{sec:transverse}
We converted the proper motions and proper accelerations to velocities and accelerations using an assumed distance of 8 kpc since the precise distances of the stars in our sample are not known. More precise stellar distances may be constrained with future measurements if 3D acceleration is measured, and therefore the distance from the star to Sgr A* can be calculated. The resulting velocities and accelerations are listed in Table~\ref{tab:vels and accels}.
\begin{figure*}[t!]
\centering
\includegraphics[width=\textwidth]{vel_accel_arrows.pdf}
\caption{Stellar velocities (left) and accelerations (right) assuming a distance of 8 kpc to convert proper motions to physical velocities and accelerations. Masers with proper velocities $>900$ km s$^{-1}$~ or proper accelerations $>0.5$ km s$^{-1}$ yr$^{-1}$~ are not shown. The majority of 2D velocities shown are significant, with the exception of SiO-11, SiO-18, and SiO-28, but all of the radial velocities are significant. Note that none of the accelerations in the plane of the sky are significant, and only SiO-14 has a radial acceleration measured with $>5 \sigma$ significance. Positions are indicated by the arrow tails. The star IRS 9 at ($\Delta x$, $\Delta y$) = (0.22 pc, -0.25 pc) has a radial velocity of -340 km s$^{-1}$, which is below the range of the velocity color bar on the left. \label{fig:arrows}}
\end{figure*}
\subsection{Doppler velocities and accelerations}\label{sec:doppler}
As described previously, we find radial velocities by calculating the centroid of each maser spectrum. However, very long baseline interferometry (VLBI) observations of stellar SiO masers shows that the maser emission is distributed in patches around the star and the velocity inferred from the cumulative spectrum may not match the true stellar velocity (e.g. \citealt{Gonidakis2010}). As can be seen in the spectra shown in Appendix~\ref{app:spectra}, many of the masers in our sample have complicated spectral profiles and the inferred velocities can differ between two line profiles taken during the same observation. We therefore estimated an additional systematic uncertainty on the stellar velocities. First, in cases where two lines were detected in a single epoch, we calculated the weighted mean of the two independent velocity measurements so we have one velocity per epoch.
We then performed a linear fit to the velocities, and calculated a systematic uncertainty from standard deviation of the residuals of the linear fit. The systematic was added in quadrature to the statistical uncertainties.
Radial accelerations were calculated by fitting a slope to the velocity measurements for stars with at least three epochs of observations. Velocity and acceleration fit results are given in Table~\ref{tab:vels and accels}, and figures showing the fits are given in Appendix~\ref{app:spectra}.
For stars with too few measurements to fit an acceleration, we report the error weighted mean velocity in Table~\ref{tab:vels and accels}. Radial velocities and accelerations are expressed in the local standard of rest (LSR) frame. Note that radial velocity and acceleration measurements are only based on the observations detailed in this work and did not include any measurements from literature, unlike the proper motion analysis.
The majority of stars have insignificant radial acceleration limits, with uncertainties ranging from $\sim 0.02$--1 km s$^{-1}$ yr$^{-1}$. However radial acceleration was measured at the $>3\sigma$ level for SiO-14 ($-0.58\pm0.10$ km s$^{-1}$ yr$^{-1}$), SiO-27 ($0.099\pm0.025$ km s$^{-1}$ yr$^{-1}$), IRS15NE ($-0.34\pm0.11$ km s$^{-1}$ yr$^{-1}$), SiO-11 ($0.105\pm0.029$ km s$^{-1}$ yr$^{-1}$), IRS19NW ($-0.226\pm0.052$ km s$^{-1}$ yr$^{-1}$), and SiO-25 ($-0.196\pm0.038$ km s$^{-1}$ yr$^{-1}$). We note that maser variability over time and between lines may introduce a systematic acceleration that does not match the true acceleration of the star.
Velocities from the 2017 ALMA data are often but not always the largest outliers from the fit compared other epochs, with average residuals around $\pm 1.2$ km s$^{-1}$. We verified that there is no systematic velocity shift being introduced by our reduction or spectra fitting methods by comparing the Sgr A* absorption spectrum observed with two ALMA epochs. By cross-correlating the two spectra, we found that the cross-correlation is maximized at $-0.4$ km s$^{-1}$. It is therefore unlikely that the 2017 ALMA velocities are being offset by a significant systematic error.
We also attempted to measure radial accelerations by simultaneously fitting spectral profiles to all spectra for a given star following a method detailed in \cite{Pesce2020} and \cite{Reid2013}, who measured accelerations of extragalactic water megamasers. Per star, we performed 100 fits with varying number of Gaussian components and randomized initial values for the velocities of each component. However, we found that the spectra were poorly fit unless we allowed each Gaussian component to have an independent acceleration measurement, which is not useful for determining the underlying systemic acceleration. This is likely due to the variable nature of stellar SiO masers and the fact that we include spectra from multiple maser transitions, which complicates tracking maser peaks between epochs. This method may have better utility for future works with a higher number of spectra of the same transition.
\section{Stellar kinematics}\label{sec:kinematics}
The 3D stellar velocities and accelerations are shown in Figure~\ref{fig:arrows}.
Masers with proper velocities $>900$ km s$^{-1}$~ or proper accelerations $>0.5$ km s$^{-1}$ yr$^{-1}$~ are not shown in the figure. Note that transverse motions are referenced to the Sgr A* continuum, whereas radial motions are in the LSR frame and do not account for any radial velocity of Sgr A*.
\subsection{Enclosed mass limits}\label{sec:velocities}
Stellar kinematics may be utilized to constrain the stellar and dark matter mass distribution in the GC. We estimate enclosed mass limits using the 3D velocities for each maser with a $>5\sigma$ significance 3D velocity measurement. \citet{Reid2003} derived a strict lower limit on the enclosed mass for a star given its total velocity magnitude, $v$, and projected distance from Sgr A*, $r_{proj}$:
\begin{equation}
M_{encl} \geq \frac{v^2 r_{proj}}{2G}. \label{eq:Mencl}
\end{equation}
The lower limit approaches an equality as $r_{proj}$ approaches the true distance, when the star is near pericenter, and if the eccentricity of the orbit is close to 1. Following \cite{Reid2003}, we subtract $2 \sigma$ from the stellar velocity to obtain a conservatively small velocity allowed by the measurement uncertainties prior to calculating the enclosed mass limit.
\begin{figure}[]
\centering
\includegraphics[width=\columnwidth]{Mencl_lolims.pdf}
\caption{Enclosed mass lower limits based on the stellar 3D velocities as a function of projected distance. The dark matter halo model (green region) is a generalized NFW profile \citep{McMillan2016}, and the maximal dark matter spike model (orange region) includes an additional sharp spike in the dark matter density profile \citep{Lacroix2018}.\label{fig:mass lims}}
\end{figure}
Figure~\ref{fig:mass lims} shows the mass limits calculated for each star compared to two models of the dark matter distribution in the Galactic Center. The first is a dark matter halo based on a generalized Navarro-Frenk-White (NFW) profile, using the main model parameters of \cite{McMillan2016}: a slope index of $\gamma=1$, scale radius of $r_s = 18.8$ kpc, local dark matter density of $\rho_\odot =0.0101$ M$_\odot$~pc$^{-3}$, and distance to the GC of $R_0=8.2$ kpc.
The second model includes a sharp spike in the dark matter density profile scaling as $r^{-7/3}$ within the spike scale radius of $R_{sp}=100$ pc, which is the maximal dark matter spike obtained by \cite{Lacroix2018}.
Both models include a central black hole mass of $4.15\times10^6$ M$_\odot$ \citep{GRAVITYCollaboration2019} and stellar density profiles following $r^{\gamma}$. We display a range of stellar profiles with $\gamma$ values between 1.1 to 1.5 and stellar mass density normalizations between 1.2 to $1.8 \times 10^5$ M$_\odot$~pc$^{-3}$ at 1 pc \citep{Schodel2018}.
All but four of the stars have $M_{encl}$ lower limits consistent with the expected values: IRS 9, SiO-25, SiO-16, and SiO-21, in order of projected distance. We discuss possible reasons that the enclosed mass may be over estimated for these stars in Section~\ref{sec:high vel stars}.
\subsection{Accelerations}\label{sec:accelerations}
\begin{figure}[t!]
\centering
\includegraphics[width=\columnwidth]{accel_lims.pdf}
\caption{3D stellar acceleration magnitude limits as a function of projected distance. Shaded regions show expected accelerations due to models of the interior stellar and dark matter mass. The dark matter halo model (green) is a generalized NFW profile \citep{McMillan2016}, and the maximal dark matter spike model (orange) includes an additional sharp spike in the dark matter density profile \citep{Lacroix2018}.\label{fig:accel lims}}
\end{figure}
Figure~\ref{fig:accel lims} shows the acceleration estimates for our masers as a function of projected distance from Sgr A* compared to expected accelerations for the two mass models described in Section~\ref{sec:velocities}, assuming circular orbits. Approximately 0.04 km s$^{-1}$ yr$^{-1}$~ precision would be required to differentiate between the two dark matter models in the range of radii where our sample of masers are found.
The upper limit accelerations for the majority of masers are at least an order of magnitude higher than the expected acceleration and the difference tends to be larger for masers at larger projected distance. This is due the fact that masers closer to Sgr A* are more likely to have more position measurements and therefore lower acceleration uncertainties than masers farther from Sgr A*.
Three masers, SiO-14, IRS15 and IRS19NW, have at least $3\sigma$-significance acceleration magnitudes, indicated by error bars rather than upper limits in Figure~\ref{fig:accel lims}, which disagree with the model accelerations. These accelerations are dominated by the radial acceleration and may be affected by intrinsic spectral variation as discussed in Section~\ref{sec:doppler} (see also Figure~\ref{fig:spectra and fits}).
\section{Discussion}\label{sec:discussion}
\subsection{High velocity stars}\label{sec:high vel stars}
From Figure \ref{fig:mass lims}, we identified four stars with velocities that exceed the expectation for stars within their projected distance from Sgr A*. The first is IRS 9, which has a 3D velocity of 370 km s$^{-1}$. This result is consistent with that of \citet{Reid2007}, who discussed several possible explanations for the star's high velocity. However, \cite{Trippe2008} found that this velocity is not excessively high compared to the global distribution of 3D stellar velocities in the GC, the tail end of which extends up to about 500 km s$^{-1}$.
SiO-25, SiO-16, and SiO-21 each have significantly higher velocity measurements than predicted for stars in orbit in the GC: approximately 2100, 1600, and 1200 km s$^{-1}$, respectively. In the case of these three masers, however, the total velocity is dominated by the 2D velocity inferred from the proper motion. Observational limitations may be causing spurious large proper motion measurements since these masers have a relative paucity of observations compared to other masers in our sample. In fact, SiO-16 and SiO-21 are outside of the VLA field of view and therefore only detected in the 86 GHz line by ALMA.
However, if the proper motions for these stars are accurate but the stars are closer than the assumed 8 kpc distance to the GC, we may have overestimated the 2D velocity. To estimate the likelihood that three foreground stars would contaminate our sample of masers, we used the model for the distribution of AGB stars in the Galaxy from \citet{Jackson2002}. We integrated their number density model over the volume observed in our ALMA observations towards the GC, resulting in $\sim 0.1$ foreground AGB stars expected in our observations. We therefore conclude it is unlikely that all three masers are actually foreground stars.
\subsection{GC reference frame}\label{reference frame}
Proper motions of SiO masers associated with bright IR stars have been used to derive the GC astrometric reference frame (\citealt{Yelda2010}; \citealt{Plewa2015}; \citealt{Boehle2016}; \citealt{Sakai2019}), which is necessary to mitigate systematic uncertainties in precision measurements of the orbits of IR stars. \cite{Sakai2019} constructed the astrometric reference frame using maser astrometry from 2005 -- 2017 and reported uncertainties on the SiO-based proper motions ranging from about 0.02 -- 0.6 mas yr$^{-1}$~ for seven stars: IRS 9, IRS 7, IRS 12N, IRS 28, IRS 10EE, IRS 15NE, and IRS 17.
Our proper motion analysis incorporates several more epochs of observations and a longer time baseline of about 25 years from 1995 -- 2020. For the seven reference frame stars, we measure proper motions with similar or better precision than those reported in \cite{Sakai2019}, with uncertainties ranging from about 0.01 -- 0.2 mas yr$^{-1}$. The astrometry reported in this work could be used to further improve the GC reference frame since the proper motion fit parameters in Table \ref{tab:pms} may be used with Equation \ref{eq:proper motion} to predict the stellar position at any future epoch.
\section{Conclusions}\label{sec:conclusions}
We presented five epochs of SiO maser astrometry for stars in the GC from 2013 to 2020. Masers were observed at 43 GHz with the VLA in 2013, 2014, and 2020, and at 86 GHz with ALMA in 2015 and 2017. We observed 28 masers total, of which four are new detections.
These masers act as high brightness temperature probes of the underlying stellar kinematics within a few parsec of Sgr A*. By incorporating archival data since 1995 in our maser position time series, we calculated proper motions and proper accelerations with $\sim 0.1$ mas yr$^{-1}$~ and 0.01 mas yr$^{-2}$~ precision, respectively. The proper motions have similar or better precision than those used to derived the GC astrometric reference frame for IR stellar orbits, and may therefore be used to further improve the precision of the reference frame.
The proper motions and proper accelerations correspond to 2D velocities and accelerations with uncertainties of about 4 km s$^{-1}$~ and 0.4 km s$^{-1}$ yr$^{-1}$, though we note that accurate determination of the stellar velocities and accelerations depend on the radial distances to these stars which are currently unknown and assumed to be 8 kpc for all masers. From spectra, we measure radial velocities and accelerations with $\sim$0.5 km s$^{-1}$~ and 0.1 km s$^{-1}$ yr$^{-1}$~ precision, respectively.
We also studied the stellar kinematics in the context of expected stellar velocities and accelerations in the GC based on models of the combined mass of Sgr A*, the stellar population, and the dark matter distribution. Lower limits on the enclosed mass were calculated for each maser based on their 3D velocities and projected distances. Several stars were found to yield enclosed mass limits 1--2 orders of magnitude higher than those predicted due to particularly high transverse velocity measurements. These may caused by inaccurate proper motion measurements due to a paucity of observations or by incorrectly translating proper motions to velocities if the stars are closer than the assumed 8 kpc.
The primary limitation on the analysis presented here is the variability of SiO masers. The majority of our masers show spectral variability in time and between lines, which may degrade the accuracy of measuring the systemic radial velocities and accelerations of the stars themselves. Further high resolution observations will improve the statistical uncertainties in proper motion and radial measurements, as well as potentially help resolve any systematics introduced by maser variability.
Higher precision measurements will enable improved mapping of the stellar and dark matter mass distributions, as well as probes of the metric around Sgr A* and general relativistic tests, which we intend to pursue in future work.
\acknowledgments
We thank the anonymous referee for very helpful comments. The authors acknowledge support from the NSF Graduate Research Fellowship Program under grant DGE-1650115 and the NSF grant AST-1908122. This paper makes use of the following ALMA data: ADS/JAO.ALMA\#2013.1.00834.S, ADS/JAO.ALMA\#2016.1.00940.S. ALMA is a partnership of ESO (representing its member states), NSF (USA) and NINS (Japan), together with NRC (Canada), MOST and ASIAA (Taiwan), and KASI (Republic of Korea), in cooperation with the Republic of Chile. The Joint ALMA Observatory is operated by ESO, AUI/NRAO and NAOJ.
\facilities{ALMA, VLA}
\software{astropy \citep{Astropy2013, Astropy2018}, CASA \citep{McMullin2007}, numpy \citep{Harris2020}, matplotlib \citep{Hunter2007}, scipy \citep{Virtanen2020}, uvmultifit \citep{Marti-Vidal2014}}
\vspace{10cm}
|
\section{Introduction}\label{s:introduction}
Humans perceive the world as a collection of distinct objects. When we interact with an object, we naturally perceive the different parts it consists of.
In visual scene understanding, parts provide intermediate representations that are more invariant to changes in object pose, orientation, camera view or lighting than the objects themselves.
They are useful in analyzing objects for higher level tasks, such as fine-grained recognition, manipulation etc. However, supervised learning of parts requires manual annotations, which are slow, expensive and infeasible to collect for the almost unlimited variety of objects in the real world.
Thus, unsupervised part discovery and segmentation has recently gained the interest of the community.
We thus consider the problem of automatically discovering the parts of visual object classes: given a collection of images of a certain object category (e.g., birds) and corresponding object masks, we want to learn to decompose an object into a collection of repeatable and informative parts.
It is important to define and understand the nature of parts before we begin describing approaches to part discovery.
While there is no universally accepted formal definition for what constitutes a ``part'', the nature of objects and object parts is accepted as different.
For example, for Gibson~\cite{gibson1966senses}, an object is ``detachable'', \textit{i}.\textit{e}.~ something that, at least conceptually, can be picked up and moved to a different place irrespective of the rest of the scene.
Parts, in contrast, are constituent elements of an object, and cannot be removed without breaking the object, \textit{i}.\textit{e}.~ they are essential to the object and occur across most instances of the same object category.
Unsupervised part discovery requires suitable inductive principles and the choice of these principles defines the nature of the parts that will be discovered. Parts could, for example, be defined based on motion following the principle of common fate in Gestalt psychology (\textit{i}.\textit{e}.~ what moves together belongs together)~\cite{spelke1990principles,wagemans2012century}, or they could be defined based on visual appearance or function.
Here, we are interested in \emph{semantic} parts across different instances of an object category (\eg birds, cars, etc.)
and combine three simple learning principles as ``part proxy'': (a) consistency to transformation (equivariance), (b) visual consistency (or self-similarity), and (c) distinctiveness among different parts.
Prior work has suggested that useful cues for part discovery can be obtained from pre-trained neural networks~\cite{bau2017network,gonzalez2018semantic}.
These networks can in fact be used as a dense feature extractors and the feature responses can be clustered or otherwise decomposed to identify parts~\cite{collins2018deep,hung19scops}.
In particular,~\cite{hung19scops} learn part prototypes, and make the latter orthogonal to avoid parts collapsing into a single one.
%
%
%
%
In this paper, we revisit and improve such concepts.
We make the following contributions.
First, we show that contrastive learning can be employed as an effective tool to decompose objects into diverse and yet consistent parts.
In particular, we seek parts whose feature responses are \emph{homogeneous} within the same or different occurrences of the \emph{same part type}, while at the same time being \emph{distinctive} for \emph{different types} of parts.
A second contribution is to discuss whether clustering pre-trained features is indeed sufficient for part discovery.
To this end, we show that simply clustering dense features sometimes captures obviously self-similar structures, such as image edges, rather than meaningful parts (\cref{s:vis}).
This is somewhat intrinsic to using pre-trained feed-forward local features, as these can only analyze a fixed image neighborhood and thus pick up the pattern which is most obvious within their aperture.
As a complementary cue, we thus suggest to look at the visual consistency of parts.
The idea is that most parts are visually homogeneous, sharing a color or texture.
A generative model of the part appearance may thus be able to detect part membership at the level of individual pixels.
We show, in particular, that even very simple models that assume color consistency are complementary and beneficial when added to feature-based grouping.
Finally, we consider the problem of \textit{assessing} automated part discovery.
An issue is the relative scarcity of data labelled with part segmentation.
Another one, technically more challenging, is the fact that parts that are discovered without supervision may not necessarily correspond to the parts that a human annotator would assign to an image.
This makes the use of manual part annotations for evaluation tricky.
Prior work in the area has thus assessed the discovered parts via proxy tasks, such as learning keypoint predictors, using supervision. %
The idea is that, if parts are consistent, they should be good predictors of other geometric primitives.
Unfortunately, as we show empirically, transferring parts to keypoints is unlikely to provide a meaningful metric for the quality of the part segments.
We show, for instance, that knowledge of a \emph{single} keypoint provides a better predictor of other keypoints than \emph{any} of the previous unsupervised models.
To address this issue, we propose a new evaluation protocol.
We still use keypoints as they are readily available, or ground-truth part segmentation when possible; however, instead of learning to regress such ground truth annotations, we simply measure the co-occurrence statistics of the predicted parts and these annotations using Normalized Mutual Information and Adjusted Rand Index.
The latter require the learned parts to be geometrically consistent and distinctive regardless of whether they are in one-to-one correspondence with manually provided labels and results in a more meaningful measure for this task.
Empirically, we demonstrate that these improvements lead to stronger automated part discovery than prior work on standard benchmarks.
\section{Related work}\label{s:related}
There exists a vast amount of literature that studies the problem of decomposing a scene into objects and objects into parts, with or without supervision.
Next, we discuss these lines of work with a focus on unsupervised approaches.
\paragraph{Unsupervised scene decomposition.}
Unsupervised scene decomposition methods aim to spatially decompose a scene into a variable number of components (segments), \eg individual objects and background.
This is typically achieved by encoding scenes into object-centric representations which are then decoded against a reconstruction objective, often in a variational framework~\cite{KingmaW13}.
Representative methods leverage sequential attention~\cite{burgess2019monet, engelcke2019genesis}, iterative inference~\cite{greff2019multiobject, locatello2020object, emami2021Efficient}, spatial attention~\cite{lin2020space, crawford2019spatially} and physical properties~\cite{bear2020learning}, or extend towards temporal sequences~\cite{kosiorek2018sequential,Jiang2020SCALOR,crawford2020exploiting,kabra2021simone,bear2020learning}.
Discriminative approaches~\cite{Kipf2020Contrastive} also exist, using an object-level contrastive objective.
It has been shown that, currently, most of these models perform well on simple, synthetic scenes but struggle with higher visual complexity~\cite{karazija2021clevrtex}.
There are key conceptual differences between decomposing a scene into objects and decomposing an object into its parts.
First, the objects of a scene typically appear in an arbitrary arrangement, whereas the presence and the layout of object parts is generally constrained (\eg the arms of a human connect to the torso).
As such, parts are constituent elements of an object and they occur consistently across most instances of an object category.
Moreover, the segments obtained by the systems described above are orderless and do not have a ``type'' assigned to them, in contrast to parts which usually refer to specific, nameable entities (\eg head vs.~beak).
These differences lead to sufficiently different statistics and technical constraints, which make it difficult to envision methods that can do well both at scene-level and object-level.
\paragraph{Part discovery and segmentation.}
Prior to deep learning, part-based models~\cite{felzenszwalb2009object,fergus2003object,felzenszwalb2010cascade,cootes2001active} played a major role in problems such as object detection and recognition.
In the deep learning era, part discovery remains an integral part of fine-grained recognition, where it acts as an intermediary step with or without part-level supervision~\cite{lam2017fine,yang2018learning,ge2019weakly,zhang2014part,zhang2013deformable,krause2015fine,wang2018learning,chai2013symbiotic,xiao2015application,huang2020interpretable,zhang2016picking,lin2015bilinear,sun2018multi,zheng2017learning,zheng2019looking}.
However, all these methods require joint training with image labels and focus mostly on discovering the most informative (discriminative) regions to ultimately help with the classification task.
Unlike previous methods as well as existing supervised methods that learn from annotated part segments~\cite{wang2015semantic,tsogkas2016deep,jackson2016cnn,liang2016semantic}, our goal is to discover independent and semantically consistent parts without image-level or part-level labels .
\citet{bau2017network} and \citet{gonzalez2018semantic} inspect the hidden units of convolutional neural networks (CNNs) trained with image-level supervision (\eg on ImageNet~\cite{russakovsky2015imagenet}) to understand whether part detectors emerge in them systematically.
This is done by measuring the alignment between each unit and a set of dense part labels, and as such, the availability of manual annotations is required for interpretation.
Most related to our work, however, are approaches for unsupervised part segmentation~\cite{simon2015neural,braun2020parts,hung19scops,collins2018deep,kosiorek2019stacked}.
Based on the observation of \cite{bau2017network,gonzalez2018semantic} that semantic parts do indeed emerge in deep features, \citet{collins2018deep} propose to use non-negative matrix factorization to decompose a pre-trained CNN's activations into parts.
Similar observations had been previously discussed in~\cite{simon2015neural} for constructing part constellations models and in~\cite{wang2015unsupervised} for part detection via feature clustering.
To learn part segmentations in an unsupervised manner, \citet{braun2020parts} propose a probabilistic generative model to disentangle shape and appearance, but focus mostly on human body parts.
Lastly, closest to our work is SCOPS by \citet{hung19scops}; SCOPS is a self-supervised approach for object part segmentation from an image collection of the same coarse category, \eg birds. The authors propose a set of loss functions to train a model to output part segments that adhere to semantic, geometric and equivariance constraints.
Other recent methods~\cite{tritrong2021repurposing,zhang2021datasetgan} use generative adversarial networks for few-shot part segmentation, while \cite{gadre2021act} discover parts without supervision by interacting with articulated objects,
and \cite{siarohin2021motion,xu2019unsupervised,liu2020nothing,sabour2021unsupervised} from motion in videos.
\paragraph{Self-supervised and contrastive learning.}
In self-supervised learning, one typically aims to design pretext tasks~\cite{pathak2016context,zhang2016colorful,noroozi2016unsupervised,doersch15unsupervised,gidaris18unsupervised} for pre-training neural networks; these tasks are constructed such that the model has to capture useful information about the data that leads to learning useful features.
Contrastive learning has recently emerged as a promising paradigm in self-supervised learning in computer vision, with several methods~\cite{oord19representation,chen2020simple,he2020momentum,wu2018unsupervised,henaff2020data,hjelm2019learning,kalantidis2020hard,misra2020self,henaff2021efficient,caron2020unsupervised} learning strong image representations that transfer to downstream tasks.
The key idea in contrastive learning is to encode two similar data points with similar embeddings, while pushing the embeddings of dissimilar data further apart~\cite{hadsell2006dimensionality}.
In absence of labels, most contrastive methods use heavy data augmentations to create different views of the same image to use as a positive pair and are trained to minimize different variants of the InfoNCE loss~\cite{oord19representation}.
We instead follow an approach to contrastive learning that is more tailored to semantic part segmentation, \textit{i}.\textit{e}.~ taking into consideration the dense nature of this problem.
Our method is thus also related to self-supervised learning of dense representations~\cite{pinheiro2020unsupervised,van2021unsupervised,ji2019invariant,zhao2021self}.
As these methods learn an embedding for every pixel, they cannot be directly applied for part segmentation and need either fine-tuning or a clustering step to produce part masks.
\section{Method}\label{s:method}
\begin{figure}[t]
\centering
\includegraphics[width=\textwidth,trim=0cm 3cm 0cm 0cm, clip]{figures/bird_losses.pdf}
\vspace{-1em}
\caption{\textbf{Training objectives.} We train our model with a set of loss functions that enforce several forms of consistency between the discovered parts. The feature loss a) ensures that parts are consistent within themselves. The contrastive loss b) discovers the same part in different images. The equivariance loss c) makes use of the fact that image transformations should not change part segmentations, and the visual consistency d) reconstructs a simplified version of the image from the parts to encourage visual consistency.}
\label{f:objectives}
\end{figure}
Given a collection of images centered around a given type of objects (\eg birds), we wish to automatically learn a part detector, assigning each pixel of the objects to one of $K$ semantic parts.
Formally, we model the part segmentation task as predicting a mask $M \in \{0, 1\}^{K\times H\times W}$ for an image $I \in \mathbb{R}^{3\times H\times W}$, where $\sum_{k=1}^{K} M_u = 1$ for all pixels $u \in \{0,\ldots,H-1\}\times\{0,\ldots,W-1\}$.
The mask thus assigns each pixel $u$ to one of $K$ parts and the part segmenter is a function $f: I \mapsto M$, implemented as a deep neural network, that maps an image $I$ to its part mask $M$.
The mask is relaxed and computed in a differentiable manner, by applying the softmax operator at each pixel.
Since we are tackling this task without supervision, we have to construct a proxy task that will enforce $f$ to learn a meaningful decomposition of the image into its parts without the need for labelled examples.
The rest of the section defines this task.
\subsection{Contrastive feature discovery}
Following prior work~\cite{collins2018deep,hung19scops}, our primary cue for discovering parts is a deep feature extractor $\phi$, obtained as a neural network pre-trained on an off-the-shelf benchmark such as ImageNet, with or without supervision.
In order to obtain repeatable and distinctive parts from these features, we propose to use a \emph{contrastive formulation}~\cite{oord19representation,hjelm2019learning}.
To this end, let $[\phi(I)]_u\in\mathbb{R}^d$ be the feature vector associated by the network to pixel location $u$ in the image.
The idea is that, if pixel $v$ belongs to the same part type as $u$, then their feature vectors should be very similar when \emph{contrasted} to the case in which $v$ belongs to a different part type.
Since parts should be consistent irrespective of the particular object instance, comparisons extend within each image $I$, but also \emph{across different images}.
Thus, a na{\"\i}ve application of this idea would require a number of comparison proportional to the square of the number of pixels in the dataset, which is impractical.
Instead, we approach this issue by noting that contrastive learning would encourage features that belong to the same \textit{part type} to be similar.
This is even more true for features that belong to the same part \emph{occurrence} in a specific image.
We can thus summarize the code for part $k$ in image $I$ via an average part descriptor $z_k \in \mathbb{R}^d$:
\begin{equation}
%
z_k(I) = \frac{1}{|M_{k}|}\sum_{u\in\Omega} M_{ku} \,[\phi(I)]_u,
~~~~
|M_k| = \sum_{u\in\Omega} M_{ku},
\end{equation}
where $\Omega$ represents all foreground pixels in the image.
We can then \emph{directly} enforce that pixels within the same part occurrence respond with similar features by minimizing the variance of descriptors within the part:
\begin{equation}
\mathcal{L}_f(M) = \sum_{k=1}^{K} \sum_{u\in\Omega} M_{ku} \, \| z_k(I) - [\phi(I)]_u \|_2^2.
\end{equation}
By doing so, we gain two advantages.
First, pixels are assigned to the same part occurrence if they have similar feature vectors, as contrastive learning would do.
Second, the part occurrence is now summarized by a single average descriptor vector $z_k(I)$ which has a differentiable dependency on the mask.
Next, we show how we can express the rest of the contrastive learning loss as a function of these differentiable part occurrence summaries.
To this end, we use a random set (\eg the mini-batch) of other images.
Intuitively, we would like to maximize the semantic similarity between all the $k$-th parts \emph{across} images and analogously minimize the semantic similarity between all other parts.
This score is computed over a batch of $N$ images, each with $K$ descriptors $z_k^{(n)}$, where $n$ indexes the image. %
To reduce the number of comparisons, for each part $k$ we randomly choose a \textit{target} $\hat{z}_k^{(n)} \in \{z_k^{(i)} \}_{i \neq n}$ out of the $N-1$ other part $k$ occurrences in the batch.%
\footnote{Note that samples are taken with respect to part occurrences, which are fixed, not with respect to the assignment of pixels to the parts (which are learned as the mask $M$).
As a consequence, we do \emph{not} need to differentiate through this sampler.}
With this, the contrastive loss can be written as usual:
\begin{equation}
\mathcal{L}_c = -\sum_{n=1}^N \sum_{k=1}^K \log \frac{\exp (z_k^{(n)} \cdot \hat{z}_k^{(n)}/\tau)}{ \exp (z_k^{(n)} \cdot \hat{z}_k^{(n)}/\tau) + \sum_{j \neq k}\sum_{i \neq n} \exp (z_k^{(n)} \cdot z_j^{(i)}/\tau)},
\end{equation}
where $\tau$ is a temperature hyper-parameter that controls the ``peakyness'' of the similarity metric.
Note that, while this score function resembles the typical contrastive formulation in current self-supervised approaches, instead of generating the target $\hat{z}_k^{(n)}$ as an augmentation of the original image, here we can actually use a different image, since part $k$ has the same semantic meaning in both images.
This formulation implicitly encourages two properties.
On one hand, it maximizes the similarity of the \textit{same} part type across images, and on the other hand, it maximizes the dissimilarity of \textit{different} part types in the same and other images.
\subsection{Visual consistency}\label{s:vis}
While semantic consistency of part features is an important learning signal, these feature maps are of low spatial resolution and do not accurately align to image boundaries.
We suggest that an effective remedy is to look for the \textit{visual} consistency of the part itself.
We can in fact expect most part occurrences to be characterized by a homogeneous texture.
Generative modelling can then be used to assign individual pixels to different part regions based on how well they fit each part appearance.
\begin{wrapfigure}[10]{r}{20em}
\vspace{-1em}
\includegraphics[width=20em,trim={0 5.4cm 0 0},clip]{figures/feature_clusters.pdf}
\caption{$K$-means clustering ($K=4$) on foreground pixels for features extracted at different layers of a VGG-19~\cite{Simonyan15}.}
\label{fig:feat_cluster}
\end{wrapfigure}
This signal is in part complementary to feature-based grouping.
As shown in \cref{fig:feat_cluster}, clustering features from successive layers of a VGG-19 network~\cite{Simonyan15} (pre-trained on ImageNet), when the receptive field of the features straddles two or more parts, grouping may sometimes highlight self-similar structures such as region boundaries instead of parts.
On the other hand, image pixels can almost always uniquely be attributed to a single part.
In our experiments, we show that even the simplest possible generative model, which assumes that pixels are i.i.d.~samples from identical Gaussians, helps improving the consistency of the discovered parts.
The negative log likelihood of parts under this simple model is given by the loss:
\begin{equation}
\mathcal{L}_v(M) = \sum_{k=1}^{K}
\sum_{u\in\Omega} M_{ku}\,
\Big\| I_u - \frac{1}{|M_k|}\sum_{v\in\Omega} M_{kv}\,I_v \Big\|_2^2.
\end{equation}
This encodes the inductive bias that parts are roughly uniformly colored. Encouraging the model to learn parts that align with image boundaries. While more complex generative models can be used here, in our experiments (\cref{table:ablation}), this simple assumption already improved the results considerably.
\subsection{Transformation equivariance}
Finally, we make use of the fact that an image transformation should not change the assignment of pixels to parts.
We thus sample a random image transformation $T$ and minimize the symmetrized Kullback-Leibler divergence $\mathcal{KL}$ between the original mask and the mask predicted from the transformed image
\begin{equation}
\mathcal{L}_e(I, T(I)) = \sum_{u\in\Omega} \mathcal{KL}\left( T_u(f(I)), f_u(T(I)) \right) + \mathcal{KL}\left( f_u(T(I)), T_u(f(I)) \right)\, .
\end{equation}
Here the $\mathcal{KL}$ divergence is computed per pixel, using the fact that the model predicts, via the softmax operator, a probability distribution over possible parts at each image location.
This objective encourages commutativity of the function $f$ with respect to the transformation as it is minimized if $T(f(I)) = f(T(I))$, in other words, on equivariance under image transformations.
Note that, for equivariance, we need to define the action of $T$ on both the input image $I$ and the output $f(I)$.
We consider simple random geometric warps (affine), which are applicable to any image-like tensor (thus even the pixels-wise predictions $f(I)$).
We also consider photometric augmentations (e.g., color jitter), whose corresponding action in output space is the identity, because we wish the network to learn to be invariant to these effects (they do not change the part identity or location).
\paragraph{Overall objective.}
We learn $f$ by minimizing the weighted sum of the prior losses:
$
\lambda_f \mathcal{L}_f + \lambda_c \mathcal{L}_c + \lambda_v \mathcal{L}_v + \lambda_e \mathcal{L}_e .
$
\section{Experiments}\label{s:experiments}
In the following we validate our approach on three benchmark datasets, the Caltech-UCSD Birds-200 dataset (CUB-200-2011)~\cite{wah2011caltech}, the large-scale fashion database (DeepFashion)~\cite{liuLQWTcvpr16DeepFashion} and PASCAL-Part~\cite{chen2014detect}.
Details regarding the datasets are given in the appendix.
We carry out ablation experiments to study (a) the importance of the proposed objective functions, and (b) the role of supervised vs.~unsupervised pre-training for the different components of our model.
Lastly, we show that our method compares favorably to prior work both quantitatively and qualitatively.
\paragraph{Implementation details.}\label{s:impl}
We model $f$ as a deep neural network, specifically a DeepLab-v2~\cite{chen18deeplab:} with ResNet-50~\cite{he2016deep} as backbone, as it is a standard architecture for semantic image segmentation.
Following SCOPS~\cite{hung19scops} we choose VGG19~\cite{Simonyan15} as the perceptual network $\phi$ and use ground truth foreground masks during training.
Unless otherwise specified the backbone and perceptual network are pre-trained on ImageNet with image-level supervision.
The perceptual network is kept fixed, \textit{i}.\textit{e}.~ its parameters are not further updated during training for part segmentation.
We use the same set of hyper-parameters for both, CUB-200 and Deep-Fashion, whereas some small changes are necessary for PASCAL-Part since the images are in a different resolution which typically impacts the magnitude of feature-based losses.
We provide all implementation details in the appendix.
\subsection{Evaluation Metrics}
\input{tables/cub2}
Prior work on unsupervised part segmentation~\cite{hung19scops} compares against unsupervised landmark regression methods~\cite{zhang2018unsupervised,thewlis2017unsupervised}, due to the similarity between the two tasks and the limited availability of annotations.
To do so, landmarks are obtained from part segmentations by taking the center of each mask, followed by fitting a linear regression model to map the predicted to the ground truth landmarks.
We begin by taking a critical look at this evaluation metric.
We evaluate several baselines on CUB-200-2011\,---\,namely, using the image midpoint, the center of ground truth keypoints and a single selected ground truth keypoint\,---\,and find that the landmark regression error does not correlate well with \emph{segmentation} performance.
For example, if we assume a model can accurately predict one \emph{single} keypoint and nothing else (in this case the ``throat''), the keypoint regression error is already lower than the previous state of the art.
This means that a model that predicts one good part and $K-1$ random parts would already outperform all previous methods.
Thus, the metric does not sufficiently measure the segmentation aspect of the task, which is the main goal of our method, as well as that of \cite{hung19scops,collins2018deep,huang2020interpretable}.
Instead, we propose to measure the information overlap between the predicted labelling and the ground truth with Normalized Mutual Information (NMI) and Adjusted Rand Index (ARI) as we find this does not suffer from this drawback.
Comparing to Intersection-over-Union (IoU)\,---\,which is commonly used to evaluate segmentation and detection performance\,---\,in an unsupervised setting, NMI and ARI have the advantage that they do not require the ground truth annotation to align exactly with the discovered parts and do not impose a constraint in the value of $K$, \textit{i}.\textit{e}.~ it does not need to be the same as the number of annotated categories.
We propose to compute NMI and ARI not only on the full image, but also on foreground pixels only (FG-NMI, FG-ARI).
The latter are stricter metrics that place the focus on \emph{part} quality, dampening the influence of the background, which can be usually predicted with high accuracy using state-of-the-art segmentation or saliency methods.
Importantly, these metrics can be computed even if a subset of the pixels are annotated in the dataset, and in particular even if only keypoint annotations are available, as in the case of CUB-200-2011.
\subsection{Ablation Experiments}
In \Cref{table:ablation} we evaluate the different objectives used to train our model.
We first deactivate each loss and measure the impact it has on performance in two datasets, CUB-200-2011 and DeepFashion.
Interestingly, we find that the different components differ in importance across the two datasets, even though we use the same hyper-parameters for both.
On CUB-200-2011, the most important component is to enforce consistency within parts, whereas on DeepFashion visual consistency appears to have the largest impact.
This likely comes from the different nature of ``parts'' in these two datasets.
For birds, the parts are conceptually defined by shape, function and deformation (which is captured by features), whereas for the fashion dataset, parts such as T-shirts and trousers can be identified by their consistent color and texture (which is better captured by the image).
Nonetheless, to achieve maximum performance both components are necessary in both datasets, as well as the equivariance and contrastive terms.
To better understand the importance of the contrastive formulation, we replace it with a simple $\mathcal{L}_2$ loss, \textit{i}.\textit{e}.~ comparing part feature vectors $z_k$ \emph{across} samples in the batch.
We refer to this variant as ``$\mathcal{L}_2$ instead of contrastive'' and note that it performs significantly worse than the full model with the contrastive loss.
To analyze the effect of using different images, we also train a model where we use parts in differently augmented versions (as is common in representation learning) of the \textit{same} image instead (``$\mathcal{L}_c$ w/ different views''). Exploiting the information in different images leads to better performance.
Finally, we establish a simple baseline by clustering perceptual features of concatenated layers \verb|relu5_2|, \verb|relu5_4| from a VGG19 (same layers as used in~\cite{hung19scops}) with $K$-means ($K=4)$.
This simple clustering baseline performs quite well and almost reaches the performance of previous methods (\cref{table:cub}), but the proposed approach is clearly stronger.
Notably, feature clustering results in weaker performance for DeepFashion, which intuitively also explains why within-part consistency ($\mathcal{L}_f$) is not the most critical component for this dataset.
\input{tables/ablation2}
\subsection{Eliminating Supervision}
\input{tables/unsup2}
The method we have presented is unsupervised with respect to part annotations.
However, similar to previous work~\cite{hung19scops, huang2020interpretable}, we still rely on backbones pre-trained with ImageNet supervision (IN-1lk), and foreground-background segmentation masks.
In~\cref{table:unsup} we remove these remaining, weakly supervised components step by step and replace them with unsupervised models.
We notice, that none of the recent self-supervised methods provides models based on VGG architectures~\cite{Simonyan15}, although VGG is considered a much better architecture for perceptual-type losses than ResNet~\cite{he2016deep}.
We thus use a VGG16 from DeepCluster-v1~\cite{caron2018deep}.
For a fair comparison we directly compare to a supervised VGG16 and not our final model that uses VGG19.
We find that the performance is indeed impacted by changing from supervised to unsupervised visual features ($-6$ NMI) and by replacing the supervised backbone ($-5$ NMI).
But the final performance is still competitive with previous methods such as DFF~\cite{collins2018deep} that use ImageNet supervision and masks.
Using an unsupervised saliency method~\cite{melas2021finding} for segmentation instead of ground truth foreground masks only causes a negligible drop in performance.
We further experiment with a self-supervised vision transformer(ViT)~\cite{caron2021emerging} for $\phi$.
Following~\cite{amir2021deep}, we extract the dense key features of the last transformer block of DINO~\cite{caron2021emerging} network with stride 4.
We observe that self-supervised ViT features perform significantly better than self-supervised CNN features.
When we remove all supervisions, the performance drop is minimal ($-1$ NMI).
\subsection{Comparisons with the State of the Art}\label{s:comp}
\paragraph{CUB-200.}
On CUB-200 (\cref{table:cub} and~\cref{fig:cub}), we evaluate keypoint regression performance to be directly comparable to previous work.
Due to the aforementioned limitations of this metric, we also evaluate NMI and ARI on both the foreground object (denoted with \textbf{FG}) and the whole image.
We use the publicly available checkpoint of SCOPS~\cite{hung19scops} to compute these new metrics for their method.
Additionally, we run DFF~\cite{collins2018deep} using their publicly available code.
Finally, we are even able to improve over \cite{huang2020interpretable} who use class labels for fine-grained recognition during training.
\begin{figure}[t]
\centering
\includegraphics[width=.99\textwidth]{figures/figure_cub.pdf}
\caption{\textbf{CUB-200 dataset.} Qualitative examples for SCOPS~\cite{hung19scops} and our method show that our model is able to find clearer part boundaries even in difficult poses, \eg open wings.}
\label{fig:cub}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=.99\textwidth]{figures/figure_deepfashion.pdf}
%
\caption{\textbf{DeepFashion dataset.} Our model is able to separate the hair from the rest of the head and correctly finds the boundary between upper and lower garments. }
\label{fig:fashion}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=\textwidth]{figures/figure_pascal.pdf}
%
\caption{\textbf{PASCAL-Part dataset.} We train one model per class for both, our model and SCOPS~\cite{hung19scops}. For animals we find are able to separate different body parts. (More examples in the appendix.) }
\label{fig:pascal}
\end{figure}
\paragraph{DeepFashion.}
\begin{wraptable}[6]{r}{19em}
\vspace{-1em}
\input{tables/deepfashion}
\end{wraptable}
Finally, in~\cref{table:fashion} and in~\cref{fig:fashion} we evaluate our method on the DeepFashion dataset, reporting (FG-)NMI and (FG-)ARI scores for $K=4$ predicted parts. Our model is able to identify more meaningful parts (namely: hair, skin, upper-grament, lower-garment) than SCOPS.
\paragraph{PASCAL-Part.}
\input{tables/pascal}
To understand the applicability of the method to a wide variety of objects and animals, we also evaluate on the PASCAL-Part dataset in~\cref{table:pascal} and \cref{fig:pascal}.
We train one model per object class, as in~\cite{hung19scops}.
However, we note that \citet{hung19scops} only evaluate foreground-background segmentation performance in their paper.
We thus train their method on each class and report NMI and ARI for quantitative comparisons.
For DFF we perform non-negative matrix factorization on the set of features of each class separately.
Finally, we consider an interestingly strong baseline, \textit{i}.\textit{e}.~ $K$-means clustering on foreground pixels trained for each class separately.
We find that this baseline even outperforms prior work~\cite{hung19scops,collins2018deep} by a significant margin.
However, our method strikes a balance between feature similarity and visual consistency, achieving superior part segmentation to $K$-means as well as previous methods.
One possible explanation for the disadvantage of SCOPS and DFF to simple $K$-means clustering is that they learn foreground-background separation and discover semantic parts in the foreground simultaneously, which appears to be sub-optimal for either task, \textit{i}.\textit{e}.~ it harms both the foreground and the part segmentation, as also seen in \cref{fig:pascal}.
\section{Discussion}\label{s:discussion}
\paragraph{Limitations.}\label{s:limitations}
One possible drawback of our approach is that there is no underlying reason why parts of an object should have uniform appearance, as enforced by our visual consistency objective (\eg the wheels of a car or a striped garment).
However, our main assumption is that different occurrences of the same part (\eg the mouth for two people) are more similar to each other than two different parts (\eg an eye and a mouth).
While this assumption is of course not universally applicable it is true often enough to be helpful and, complemented by the other objectives, it leads to a considerable performance improvement over prior work that does not use this concept.
Although we experimented with more complex visual modelling (\eg higher level statistics or textures), this did not yield meaningful improvements and is thus left for future investigation.
Another limitation is that parts discovered in a self-supervised manner might not necessarily agree with expected labels or human intuition.
A critical control parameter for this is the number of parts $K$. It controls the granularity of the part segmentation and is left as a hyper-parameter since it is up to the user to decide the level of decomposition.
For example, for humans one could segment arms, legs, torso and head (five parts) or decompose arms into hands, fingers, etc. In the appendix we show results of our method for different $K$.
Finally, the main failure mode of the current model is failing to separate the foreground from the background which leads to messy segmentations and scrambled masks (see the appendix qualitative examples).
\paragraph{Broader Impact.}\label{s:impact}
Supervised learning often requires highly-curated datasets with expensive, time-consuming, manual annotations.
This is especially true for pixel-level tasks (\eg segmentation) or tasks that require expert knowledge (\eg fine-grained recognition).
As a result, increasing attention is being placed on improving image understanding using little or no supervision.
Since part segmentation datasets are limited in number and size, a direct positive impact of our approach is that discovering semantic object parts in a self-supervised manner can significantly increase the amount of data that can be leveraged to train such models.
Finally, as with all methods that learn from data\,---\,and especially in the case of self-supervised learning\,---\,it is likely that underlying biases in the data affect the learning process and consequently predictions made by the model.
\section{Conclusion}\label{s:conclusion}
We have proposed a self-supervised method for discovering and segmenting object parts.
We start from the observation, also discussed in prior work~\cite{bau2017network,collins2018deep}, that deep CNN layers respond to semantic concepts or parts and thus clustering activations across an image collection amounts to discovering dense correspondences among them.
We further expand upon this idea by introducing a contrastive formulation, as well as equivariance and visual consistency constraints.
Our method relies only on the availability of foreground/background masks to separate an object of interest from its background.
However, as we show experimentally, it is possible to leverage unsupervised saliency models to acquire such masks, which allows for a model that has no supervised components at all.
\section*{Acknowledgements and Funding Disclosure}
S. C. is supported by a scholarship sponsored by Facebook.
I. L. is supported by the European Research Council (ERC) grant IDIU-638009 and EPSRC VisualAI EP/T028572/1.
C. R. is supported by Innovate UK (project 71653) on behalf of UK Research and Innovation (UKRI) and ERC grant IDIU-638009.
A. V. is supported by ERC grant IDIU-638009. We thank Luke Melas-Kyriazi for providing precomputed masks for~\cite{melas2021finding}.
{\small\bibliographystyle{plainnat}
|
\section{Introduction}
Given a class of arbitrarily large finite structures, it is a classic
problem to ask,``What does a large random structure from this class look like?" Zero-one laws are one method of
gaining insight into this question for the case of properties expressible in
first-order logic. A class $\mathcal{C}$ of structures is said to admit a zero-one law if, given a sentence $\varphi$ of first-order logic,
the probability that a randomly selected $\mathcal{C}$-structure
of size $n$ satisfies $\varphi$ converges asymptotically to zero or one as $n$ goes
to infinity. For example, in the class of finite graphs (in the language with the edge relation), properties such as containing a specified subgraph or induced
subgraph are expressible in first-order logic, whereas properties such as
connectedness or Hamiltonicity are not. It is a seminal result that this class
admits a zero-one law (\cite{fagin1976probabilities},\cite{glebskii1969range}).
The requirement that such a probability for a class of structures converges to
either zero or one is rather strict, and in general, not many classes admit a
zero-one law. This is particularly evident when dealing with ordered structures.
For example, consider the class of finite linear orders with two colors:
red and blue. The probability that a randomly selected structure's first point is
red (or blue) is $1/2$. A class of structures is said to admit a {\em logical
limit law} when the probability that a randomly selected structure of size $n$
satisfies any given first order property converges as $n \to \infty$, so a zero-one law is the special case where the limiting probability is always zero or one.
We prove logical limit laws for various classes of ordered
structures. In particular, we show the following result.
\begin{theorem}\label{theorem:intro}
Convex linear orders and layered permutations admit both unlabeled and labeled
limit laws. Compositions admit an unlabeled limit law.
\end{theorem}
We first prove a logical limit law for convex linear orders using an adaptation of
Ehrenfeucht's proof for colored linear orders, which itself uses stationary states of
Markov chains. We then use some basic definability results to transfer the limit law to layered permutations and compositions.
We now discuss some connections to previous work. First-order properties of finite permutations (when viewed as pairs of linear
orders) were studied in \cite{Albert}. There the existence of a zero-one law was
disproven, and it was asked whether or not permutations could admit a logical
limit law; the answer to this turns out to be negative as well (as shown in \cite{Foy}). However, there is a rich study of various permutation classes other than the
class of all permutations, and we can also ask whether these admit logical limit
laws, as has been done for various restricted graph classes (e.g
\cite{kopczynski2019logical}). Layered permutations are a
simple and well-studied example of such a permutation class.
In \cite{Cameron}, the limiting probability distributions of several
properties are computed for random compositions (what we call ``compositions,'' \cite{Cameron} refers to as ``preorders''). Our result complements this by showing that the limiting probability of any first-order
property converges, although we do not describe any sort of limiting distribution.
\section{Convex Linear Orders}
\begin{definition}
The language of convex linear orders. $\mathcal{L}$, consists of two
binary relations: $<$ (a total order on points) and $E$ (an equivalence relation).
A {\em convex linear order} is an $\mathcal{L}$-structure $\mathfrak{C}$ such that $E$-classes are $<$-intervals (i.e., for $x, y, z \in \mathfrak{C}$ with $x\mathrel{E} y$ and $x < z<y$, it holds that $z \mathrel{E} x,y$).
\end{definition}
In this section, we prove a logical limit law for the class of all finite convex
linear orders, closely following Ehrenfeucht's argument for colored linear orders as presented in Chapter 10 of \cite{RG}.
We denote the convex linear order with one point by $\bullet$.
\subsection{Preliminaries}
\begin{definition}\label{def:convlinorder}
Let $\mathfrak{C}$ be a convex linear order. Define
$\widehat{\mathfrak{C}}$ to be the convex linear order obtained by
adding one additional point to the last class of $\mathfrak{C}$.
\end{definition}
\begin{definition}\label{def:oplus}
For convex linear orders $\mathfrak{C}, \mathfrak{D}$, define $\mathfrak{C}
\oplus \mathfrak{D}$ as the convex linear order placing
$\mathfrak{D}$ $<$-after $\mathfrak{C}$.
\end{definition}
Clearly, the last class of any convex linear order either contains exactly one
element or more than one element. This leads naturally to:
\begin{lemma} \label{lemma:con}
Every finite convex linear order containing $n$ points can be uniquely
constructed by applying $\widehat{(-)}$ and $- \oplus \bullet$ to $\bullet$
repeatedly. This construction is done in $n-1$ steps.
\end{lemma}
\begin{proof}
We proceed by induction. Let $\mathfrak{C}$ be a convex linear order of size $n$. If $n=1$, $\mathfrak{C} \simeq \bullet$.
Now assume that for arbitrary $n$, any convex linear order of size $n-1$ can be
uniquely constructed from the operations above, in $n-2$ steps. Let $\mathfrak{B}$ be $\mathfrak{C}$ with the $<$-last point removed. Either the last class of
$\mathfrak{C}$ contains one point, or it contains more than one point.
If the former is true, then $\mathfrak{C} \simeq \mathfrak{B} \oplus \bullet$.
In this case,
$\mathfrak{C}$ cannot be obtained from $\widehat{(-)}$ from a structure of size $n-1$, as the last class of $\mathfrak{C}$ needs
to contain exactly one point. If the latter is true, then $\mathfrak{C} \simeq
\widehat{\mathfrak{B}}$, and cannot be obtained from applying $- \oplus \bullet$ to a structure of size $n-1$.
\end{proof}
From Lemma \ref{lemma:con}, we see that starting with $\bullet$ and randomly applying $- \oplus \bullet$
and $\widehat{(-)}$ (each with probability $1/2$) $n-1$ times will uniformly randomly sample all
possible convex linear orders of size $n$.
\begin{definition}\label{def:ef}
Suppose $\mathfrak{A}$ and $\mathfrak{B}$ are structures in a
language consisting of binary relations $R_1, \ldots, R_n$. An {\em
Ehrenfeucht–Fraïssé game} (sometimes referred to as a {\em ``back and forth game''}) of length $k$ between $\mathfrak{A}$ and $\mathfrak{B}$
is a game between two players (referred to as Duplicator and
Spoiler). In each round, Spoiler plays by selecting and marking a point on
either structure. Duplicator responds by marking a corresponding point on the
structure which Spoiler did not choose from. After $k$ rounds, the points
$x_1, \ldots, x_k$ have been marked on $\mathfrak{A}$, and $y_1, \ldots, y_k$ on
$\mathfrak{B}$. Duplicator has won if the map sending each $x_i$ to the
corresponding $y_i$ is an isomorphism (i.e., $x_1, \ldots, x_k, y_1, \dots, y_k$ satisfy $x_i
\mathrel{R}_\ell x_j \iff y_i \mathrel{R}_\ell y_j$ for all $i,j \in [k]$ and $\ell \in [n]$). Otherwise, Spoiler has won.
\end{definition}
The quantifier depth of a first order sentence counts the maximum depth of
nested quantifiers. A formal definition via induction is found on page 16
of \cite{RG}.
\begin{theorem}
In the Ehrenfeucht–Fraïssé game of length $k$ between $\mathfrak{A}$ and $\mathfrak{B}$, Duplicator has a winning strategy iff $\mathfrak{A}$ and $\mathfrak{B}$ agree
on all sentences of quantifier depth at most $k$.
\end{theorem}
\begin{proof}
A proof can be found in Section 2.3.1 of \cite{RG} for the case of graphs; the
general case follows an essentially identical argument.
\end{proof}
We write $\mathfrak{A} \equiv_k \mathfrak{B}$ when $\mathfrak{A}$ and $\mathfrak{B}$ agree on all sentences of quantifier depth at most $k$.
\begin{lemma} \label{lemma:add}
Let $\mathfrak{M}, \mathfrak{N}, \mathfrak{M}', \mathfrak{N}'$ be convex linear orders such that $\mathfrak{M} \equiv_k \mathfrak{N}$ and
$\mathfrak{M}' \equiv_k \mathfrak{N}'$. Then, $\mathfrak{M} \oplus
\mathfrak{M}' \equiv_k \mathfrak{N} \oplus \mathfrak{N}'$.
\end{lemma}
\begin{proof}
We will show that in any Ehrenfeucht–Fraïssé game of length $k$, Duplicator has a
winning strategy. Consider such a game between $\mathfrak{M} \oplus \mathfrak{M}'$
and $\mathfrak{N} \oplus \mathfrak{N}'$.
If Spoiler chooses any element in $\mathfrak{M}$ (respectively, $\mathfrak{M}'$),
then Duplicator responds as they would in a length $k$ game between
$\mathfrak{M}$ and $\mathfrak{N}$ (respectively, between $\mathfrak{M}'$ and
$\mathfrak{N}'$), and vice-versa if Spoiler chooses a point in $\mathfrak{N}
\oplus \mathfrak{N}'$. We will show this gives Duplicator a winning strategy in
the Ehrenfeucht–Fraïssé game between $\mathfrak{M} \oplus \mathfrak{M}'$ and
$\mathfrak{N} \oplus \mathfrak{N}'$.
Let $A \subset \mathfrak{M}, A' \subset \mathfrak{M}', B \subset \mathfrak{N}, B' \subset \mathfrak{N}'$ be the elements chosen in the game. Then $A \simeq B$ and $A' \simeq B'$. But it is easy to see that $\oplus$ preserves isomorphism, so $A \oplus A' \simeq B \oplus B'$.
\end{proof}
A very similar argument gives the following:
\begin{lemma} \label{lemma:hat}
Suppose $\mathfrak{M} \equiv_k \mathfrak{N}$, then, $\widehat{\mathfrak{M}} \equiv_k
\widehat{\mathfrak{N}}$.
\end{lemma}
\begin{proof}
We again show that in an Ehrenfeucht–Fraïssé game of length $k$, Duplicator has a
winning strategy. For any move by Spoiler in $\mathfrak{M}$ (or $\mathfrak{N}$),
Duplicator responds by playing as they would normally would in an
Ehrenfeucht–Fraïssé game between $\mathfrak{M}$ and $\mathfrak{N}$; because
$\mathfrak{M} \equiv_k \mathfrak{N}$, Duplicator always has a winning move in
response to Spoiler for any such play. If Spoiler plays the last point in the
last class of $\widehat{\mathfrak{M}}$ or $\widehat{\mathfrak{N}}$ (that is, the
point added by $\widehat{(-)}$), Duplicator can always respond with the
corresponding point at the end of $\widehat{\mathfrak{N}}$ or $\widehat{\mathfrak{M}}$. Hence, Duplicator has a response for any of Spoiler's moves in a length $k$
Ehrenfeucht–Fraïssé game between $\widehat{\mathfrak{M}}$ and $\widehat{\mathfrak{N}
}$, so $\widehat{\mathfrak{M}} \equiv_k \widehat{\mathfrak{N}}$.
\end{proof}
\begin{lemma} \label{lemma:2k}
For two finite linear orders $N, M$ having $n$ and $m$ points respectively,
$N \equiv_k M$ iff $n = m$ or $n, m \geq 2^k-1$.
\end{lemma}
\begin{proof}
This is Lemma $2.6.3$ in \cite{RG}.
\end{proof}
\begin{lemma} \label{lemma:1024}
For a convex linear order $\mathfrak{M}$ and $k \in \mathbb{N}$, there exists $\ell \in
\mathbb{N}$ such that for all $s, t > \ell$, \[ \bigoplus_s \mathfrak{M} \equiv_k
\bigoplus_t \mathfrak{M}\]
\end{lemma}
\begin{proof}
We reduce this to a case of the previous lemma. Let $\ell = 2^{k-1}$. Let $O_s$ be a linear order
with $s$ points, each corresponding to a copy of $\mathfrak{M}$ in
$\bigoplus_s \mathfrak{M}$, and define $O_t$ likewise for $t$ and
$\bigoplus_t \mathfrak{M}$. In an Ehrenfeucht–Fraïssé game of length $k$ between
$\bigoplus_s \mathfrak{M}$ and $\bigoplus_t \mathfrak{M}$, we will show that
Duplicator has a winning strategy. If Spoiler picks a point in the $i$th copy
of $\mathfrak{M}$ in $\bigoplus_s \mathfrak{M}$, we view this as
Spoiler picking the $i$th point in $O_s$ if it were playing a length-$k$
Ehrenfeucht–Fraïssé game between $O_s$ and $O_t$. By Lemma \ref{lemma:2k}, Duplicator has a
response in $O_t$; suppose this response is the $j$th point. To have a winning
strategy in the Ehrenfeucht–Fraïssé game between $\bigoplus_s \mathfrak{M}$ and
$\bigoplus_t \mathfrak{M}$, Duplicator can select the same point in $\mathfrak{M}$
which Spoiler selected, but in the $j$th copy of $\mathfrak{M}$ in $\bigoplus_t
\mathfrak{M}$.
\end{proof}
\subsection{The limit law}
It is in general important to note the distinction between labeled
and unlabeled limit laws, which count structures of size $n$ differently.
Labeled limit laws count all possible structures
over the universe $\{1, \ldots, n\}$ as $n \to \infty$, whereas unlabeled limit laws count all
structures up to isomorphism over $\{1, \ldots, n\}$. In general, a labeled
limit law for a given class does not imply an unlabeled limit law for
the class, and vice-versa. However, as finite linearly ordered structures have
no nontrivial automorphisms, labeled and unlabeled limit laws are equivalent. Thus we will not distinguish between them in the section or the next.
Given a first-order sentence $\varphi$ having quantifier rank $k$, we compute
the limiting probability of $\varphi$ by associating to it a Markov chain
$M_\varphi$. For a $\equiv_k$-class of convex linear orders $C$, and for some/any $\mathfrak{M} \in C$, we make the following definitions.\[ C \oplus \bullet := \left [ \mathfrak{M}
\oplus \bullet \right ]_{\equiv_k} \]
and \[\widehat{C} := \left [ \widehat{\mathfrak{M}} \right ]_{\equiv_k} \]
These operations are well-defined, as any choice of $\mathfrak{M}$ yields a $\equiv_k$-equivalent result by Lemmas \ref{lemma:add} and \ref{lemma:hat}.
The states of $M_\varphi$ are $\equiv_k$-classes
of $\mathcal{L}$-structures (where $k$ is the quantifier depth of $\varphi$); there are finitely many of such classes, by
Theorem 2.2.1 of \cite{RG}. For a $\equiv_k$-class $C$, there are two possible transitions out of $C$: one to
$C \oplus \bullet$, and one to $\widehat{C}$, each having probability $1/2$.
The starting state of $M_\varphi$ is $\bullet$ (we slightly abuse notation by writing $\bullet$ to also
mean $\left [ \bullet \right ]_{\equiv_k}$).
\begin{definition}\label{def:periodic}
A Markov chain $M$ is {\em fully aperiodic} if there do not exist
disjoint sets of $M$-states $P_0, P_1, \ldots, P_{d-1}$ for some $d > 1$ such
that for every state in $P_i$, $M_\varphi$ transitions to a state in $P_{i+1}$
with probability $1$ (with $P_{d-1}$ transitioning to $P_0$), i.e every
state is aperiodic.
\end{definition}
We next state a variant of the fundamental theorem of Markov chains which does not
assume irreducibility.
\begin{proposition} \label{prop:markov} Let $M$ be a finite, fully aperiodic
Markov chain with initial state $S$, and let $Pr^{n-1}(S, Q)$ denote the
probability that $M$ is in state $Q$ after $n-1$ steps. Then, for any
$Q$, $\lim_{n \to \infty} Pr^{n-1}(S, Q)$ converges.
\end{proposition}
\begin{proof}
This is contained in the discussion following Theorem 0.3.1 of \cite{Markov}.
\end{proof}
\begin{lemma} \label{lemma:fully-aperiodic}
For any first-order sentence $\varphi$, $M_\varphi$ is fully aperiodic.
\end{lemma}
\begin{proof}
Suppose $M_\varphi$ were not fully aperiodic. Then, there would exist disjoint sets of
$M$-states ($\equiv_k$-classes) $P_0, P_1, \ldots, P_{d-1}$ for some $d > 1$ such that for every
state in $P_i$, $M_\varphi$ transitions to a state in $P_{i+1}$ with probability
$1$ (with $P_{d-1}$ transitioning to $P_0$). Write $i \bullet$ to mean
$\bigoplus_i \bullet$. Thus, for any $Q \in P_0$, $Q \oplus
i \bullet$ is in $P_0$ iff $d \mid i$. But by
Lemma \ref{lemma:add} and Lemma \ref{lemma:1024}, $Q \oplus i \bullet
\equiv_k Q \oplus (i+1) \bullet$ for sufficiently large $i$, contradicting
this claim.
\end{proof}
\begin{theorem} \label{theorem:convlinorder}
Convex linear orders admit a logical limit law.
\end{theorem}
\begin{proof}
Fix a first-order sentence $\varphi$, and consider the Markov chain $M_\varphi$.
For each state $S$ in $M_\varphi$, either
each structure in $S$ satisfies
$\varphi$ or no structures in $S$ satisfy $\varphi$. Let $S_\varphi$ denote the set of
states in $M_\varphi$ for which all structures in that
state satisfy $\varphi$. By the fact that $\widehat{(-)}$ and $- \oplus
\bullet$ are well-defined on $\equiv_k$-classes (it does not matter which
structure in the class is chosen), we can view moving $n-1$ steps in $M_\varphi$ as
starting with any structure in the current state, applying $\widehat{(-)}$ or
$- \oplus \bullet $ to it $n-1$ times, and taking the resulting $\equiv_k$-class at the end. Thus by the comment after Lemma \ref{lemma:con}, moving $n-1$ steps from the starting state $\bullet$ is the same as uniformly randomly picking a size $n$
structure and then taking its $\equiv_k$-class.
Therefore, the probability that after $n$ steps, the chain is in a state of
$S_\varphi$ is same as probability that uniformly randomly selected structure of
size $n$ satisfies $\varphi$.
\if{false}
By the comment after Lemma \ref{lemma:con}, the probability that a uniformly,
randomly selected structure of size $n$ satisfies $\varphi$ is precisely the
probability that, after $n-1$ steps, $M_\varphi$ is in a state in $S_\varphi$.
This probability is given by
\[ \sum_{Q \in S_\varphi}^{}Pr^{n-1}(\bullet, Q) \]
where $Pr^{n-1}(\bullet, Q)$ denotes the transition probability from
$\bullet$ to $P$ after $n$ steps .
\fi
So it suffices to show that $\lim_{n \to \infty} \sum_{Q \in
S_\varphi}^{}Pr^{n-1}(\bullet, Q)$ converges. Because $M_\varphi$ has
finitely many states, \[\lim_{n \to \infty} \sum_{Q \in
S_\varphi}^{}Pr^{n-1}(\bullet, Q) = \sum_{Q \in
S_\varphi}^{}\lim_{n \to \infty}Pr^{n-1}(\bullet, Q) \]
It now suffices to show that $\lim_{n \to \infty}
Pr^n(\bullet, Q)$ exists for every state $Q$ of $M_\varphi$. But this follows from Proposition \ref{prop:markov}.
\end{proof}
\section{Layered Permutations}
Permutations can be viewed as structures in a language with two linear orders,
$<_1$, $<_2$. The order $<_1$ gives the unpermuted order of the points (before
applying the permutation) and $<_2$ describes the points in permuted order. An embedding of one such structure into another then corresponds to the usual notion of pattern containment.
The properties of permutations expressible by a first-order sentence in this language are explored in Section 3 of \cite{Albert}. These include the containment and avoidance of (generalized) patterns, concepts related to the substitution decomposition, and sortability properties such as $k$-stack sortability.
\begin{definition}\label{def:lperm}
Given a
permutation $P$, a {\em block} is a maximal subset $B \subset P$ that is an interval with respect to $<_1$ and $<_2$, and is monotone.
A {\em layered permutation} consists of increasing blocks of decreasing permutations.
\end{definition}
\begin{figure}[H] \label{figure:lperm}
\centering
\caption{Illustration of a layered permutation. From left to right, $<_1$
is increasing; from bottom to top, $<_2$ is increasing.}
\label{fig:label1}
\begin{tikzpicture}[x=0.75pt,y=0.75pt,yscale=-1,xscale=1]
\draw (64,145) -- (114,145) -- (114,195) -- (64,195) -- cycle ;
\draw (114,95) -- (164,95) -- (164,145) -- (114,145) -- cycle ;
\draw (164,45) -- (214,45) -- (214,95) -- (164,95) -- cycle ;
\draw (36,220) -- (249,220) ;
\draw [shift={(251,220)}, rotate = 180] [color={rgb, 255:red, 0; green, 0; blue, 0 } ][line width=0.75] (10.93,-3.29) .. controls (6.95,-1.4) and (3.31,-0.3) .. (0,0) .. controls (3.31,0.3) and (6.95,1.4) .. (10.93,3.29) ;
\draw (36,220) -- (36,19) ;
\draw [shift={(36,19)}, rotate = 450] [color={rgb, 255:red, 0; green, 0; blue, 0 } ][line width=0.75] (10.93,-3.29) .. controls (6.95,-1.4) and (3.31,-0.3) .. (0,0) .. controls (3.31,0.3) and (6.95,1.4) .. (10.93,3.29) ;
\draw (66,150.4) node [anchor=north west][inner sep=0.75pt] {$\bullet $};
\draw (97,175.4) node [anchor=north west][inner sep=0.75pt] {$\bullet $};
\draw (116,100.4) node [anchor=north west][inner sep=0.75pt] {$\bullet $};
\draw (150,128.4) node [anchor=north west][inner sep=0.75pt] {$\bullet $};
\draw (132,113.4) node [anchor=north west][inner sep=0.75pt] {$\bullet $};
\draw (183,60.4) node [anchor=north west][inner sep=0.75pt] {$\bullet $};
\draw (128,225.4) node [anchor=north west][inner sep=0.75pt] {$< _{1}$};
\draw (4,102.4) node [anchor=north west][inner sep=0.75pt] {$< _{2}$};
\end{tikzpicture}
\end{figure}
Note that in a layered permutation, two points are in the same block iff $<_1$ and $<_2$ disagree between them.
In many cases, it is useful to transfer a logical limit law on one class of
structures to another class of similar structures.
This is possible when the following interdefinability condition is satisfied. (This is surely well-known, but we could not find an explicit statement.)
\begin{definition}(Uniform Interdefinability) \label{id}\\
Let $\mathcal{L}_0$, $\mathcal{L}_1$ be languages, and $\mathcal{C}_0$,
$\mathcal{C}_1$ be classes of finite $\mathcal{L}_0$, $\mathcal{L}_1$ structures
respectively, with a common domain of $\left [ n \right ] = \left \{ 1,\cdots,n
\right \}$. $\mathcal{C}_0$ and $\mathcal{C}_1$ are said to be
{\em uniformly interdefinable} if there exists a map $f_I: \mathcal{C}_0 \to
\mathcal{C}_1$ (which is a bijection on structures of size $n$, for all $n$), along
with formulae $\varphi_{R_{0,i}}, \varphi_{R_{1,i}}$ for each relation $R_{0,
i}$ in $\mathcal{L}_0$
and $R_{1, i}$ in $\mathcal{L}_1$ respectively such that,
for each $\mathfrak{M}_0$ in $\mathcal{C}_0$ and $\mathfrak{M}_1$ in $\mathcal{C}_1$:
\begin{itemize}
\item $\mathfrak{M}_0 \models R_{0,i}(\bar{x}) \iff
f_I(\mathfrak{M}_0) \models \varphi_{R_{0,i}}(\bar{x})$
\item $\mathfrak{M}_1 \models R_{1,i}(\bar{x}) \iff f_I^{-1}(\mathfrak{M}_1)
\models \varphi_{R_{1,i}}(\bar{x})$
\end{itemize}
\end{definition}
Although uniform interdefinability may appear to be asymmetric, it is in fact a
symmetric relation, using that $f_I$ is bijective.
\begin{lemma} \label{lemma:translation}
Let $\mathcal{L}_0$, $\mathcal{L}_1$ be languages, $\mathcal{C}_0$,
$\mathcal{C}_1$ classes of $\mathcal{L}_0$, $\mathcal{L}_1$-structures
respectively, $f$ a map from the set of $\mathcal{L}_0$-structures to the set
of $\mathcal{L}_1$-structures, and $g$ a map from the set of
$\mathcal{L}_0$-sentences to the set of $\mathcal{L}_1$-sentences such that, for
any $\mathcal{C}_0$-structure $\mathfrak{M}$ and $\mathcal{L}_0$-sentence
$\varphi$:
\begin{enumerate}
\item $\mathfrak{M} \models \varphi \iff
f(\mathfrak{M}) \models g(\varphi)$ \item $f$ is a bijection between
$\mathcal{C}_0$ and $\mathcal{C}_1$ structures of size $n$
\item The class $\mathcal{C}_1$ admits a logical limit law
\end{enumerate}
Then, $\mathcal{C}_0$ admits a logical limit law as well.
\end{lemma}
\begin{proof}
Let $\varphi$ be a sentence in $\mathcal{L}_0$ and $a_0$ the number of $\mathcal{C}_0$-structures of size $n$ satisfying $\varphi$.
Likewise, let $a_1$ be the number of $\mathcal{L}_1$-structures of size $n$ which satisfy $g(\varphi)$.
For a randomly selected $\mathcal{C}_0$-structure $\mathfrak{M}$ (of size $n$),
the probability that $\mathfrak{M} \models \varphi$ is $\frac{a_0}{\left |
\mathcal{C}_0 \right |}$, and the probability that $f(\mathfrak{M}) \models g(\varphi)$ in $\mathcal{C}_1$ is $\frac{a_1}{\left | \mathcal{C}_1 \right |}$. Since $f$ is a
bijection on structures of size $n$, we have that $\left | \mathcal{C}_1 \right | = \left | \mathcal{C}_0 \right |$ and together with (1), that $a_1 = a_0$. Hence, the probabilities are equal for any
$\varphi$; because $\mathcal{C}_1$ admits a limit law,
$\mathcal{C}_0$ admits a limit law as well.
\end{proof}
\begin{lemma}\label{lemma:inter-translation}
Let $\mathcal{C}_0$, $\mathcal{C}_1$ be uniformly interdefinable classes of
$\mathcal{L}_0$, $\mathcal{L}_1$ structures. If $\mathcal{C}_1$ admits a
logical limit law, $\mathcal{C}_0$ admits one as well.
\end{lemma}
\begin{proof}
We show that the maps $f$, $g$ exist as in Lemma \ref{lemma:translation}. Take $f=f_I$ and $g$ as the map
which sends an $\mathcal{L}_0$ sentence to the $\mathcal{L}_1$ sentence where each
occurrence of $R_{0,i}$ is replaced by $\varphi_{R_{0,i}}$. Given a $\mathcal{C}_0$-structure $\mathfrak{M}_0$
and an atomic $\mathcal{L}_0$-
formula $\varphi(\bar{x})$, and a tuple $\bar{m} \subset \mathfrak{M}_0$, we have $\mathfrak{M}_0 \models \varphi(\bar{m}) \iff f_I(\mathfrak{M}_0) \models g(\varphi)(\bar{m})$ by the definition
of uniform interdefinability. When $\varphi$ is
nonatomic, the same statement follows from a standard induction on the complexity of
$\varphi$.
\if{false}
$$\mathfrak{M}_0 \models R_0(\bar{x}) \iff f_I(\mathfrak{M}_0) \models \varphi_{R_0}(\bar{x}) =
g(R_0(\bar{x}))$$
and identically for $R_1$, $\varphi_{R_1}$.
\fi
Furthermore, by Definition \ref{id}, $f_I$ is a bijection on structures of size
$n$, and therefore, $f_I, g$ are as desired.
\end{proof}
Let $\mathcal{L}_0 = \{ <_1, <_2 \}$ and $ \mathcal{L}_1 = \{ E, < \}$ denote the
languages of layered permutations and convex linear orders respectively, and
let $\mathcal{C}_0, \mathcal{C}_1$ classes of isomorphism types of
$\mathcal{L}_0, \mathcal{L}_1$-structures respectively. We define a map
from layered permutations to convex linear orders which sends blocks of a layered permutation to convex equivalence classes, and
points in each block of a layered permutation to points in the same equivalence
class such that $<_1$ agrees with $<$ (see Figure 2). Formally, this is a map
$f_I\colon \mathcal{C}_0 \to \mathcal{C}_1$ is defined such that for
$\mathfrak{M}_0$ in $\mathcal{C}_0$ and $\mathfrak{M}_1$ in $\mathcal{C}_1$:
\begin{itemize}
\item $f_I(\mathfrak{M}_0) \models a < b \iff \mathfrak{M}_0 \models a <_1 b $
\item $f_I(\mathfrak{M}_0) \models a \mathrel{E} b \iff \mathfrak{M}_0 \models
(a <_1 b \wedge a >_2 b) \vee (b <_1 a \wedge b >_2 a)$
\end{itemize}
The relations $<_1$, $<_2$ in the language of layered permutations are rewritten
in the language of convex equivalence relations using the following
rules on atomic formulas:
\begin{itemize}
\item $\varphi_{<_1}:a <_1 b \rightsquigarrow a < b $
\item $\varphi_{<_2}:a <_2 b \rightsquigarrow (a \mathrel{E} b \wedge b < a)
\vee (\neg (a \mathrel{E} b) \wedge a < b)$
\end{itemize}
\begin{figure}[H] \label{fig:2}
\tikzset{every picture/.style={line width=0.75pt}}
\center
\begin{tikzpicture}[x=0.75pt,y=0.75pt,yscale=-1,xscale=1]
\draw (299,115) -- (331,115) -- (331,145) -- (299,145) -- cycle ;
\draw (332,84) -- (364,84) -- (364,114) -- (332,114) -- cycle ;
\draw (301,28) .. controls (301,26.34) and (302.34,25) .. (304,25) -- (325,25) .. controls (326.66,25) and (328,26.34) .. (328,28) -- (328,37) .. controls (328,38.66) and (326.66,40) .. (325,40) -- (304,40) .. controls (302.34,40) and (301,38.66) .. (301,37) -- cycle ;
\draw (337,28) .. controls (337,26.34) and (338.34,25) .. (340,25) -- (359,25) .. controls (360.66,25) and (362,26.34) .. (362,28) -- (362,37) .. controls (362,38.66) and (360.66,40) .. (359,40) -- (340,40) .. controls (338.34,40) and (337,38.66) .. (337,37) -- cycle ;
\draw (306,122.5) -- (306,48.5) ;
\draw [shift={(306,46.5)}, rotate = 450] [color={rgb, 255:red, 0; green, 0; blue, 0 } ][line width=0.75] (10.93,-3.29) .. controls (6.95,-1.4) and (3.31,-0.3) .. (0,0) .. controls (3.31,0.3) and (6.95,1.4) .. (10.93,3.29) ;
\draw (321,132.5) -- (321,48.5) ;
\draw [shift={(321,46.5)}, rotate = 450] [color={rgb, 255:red, 0; green, 0; blue, 0 } ][line width=0.75] (10.93,-3.29) .. controls (6.95,-1.4) and (3.31,-0.3) .. (0,0) .. controls (3.31,0.3) and (6.95,1.4) .. (10.93,3.29) ;
\draw (348,97.5) -- (348,46.5) ;
\draw [shift={(348,44.5)}, rotate = 450] [color={rgb, 255:red, 0; green, 0; blue, 0 } ][line width=0.75] (10.93,-3.29) .. controls (6.95,-1.4) and (3.31,-0.3) .. (0,0) .. controls (3.31,0.3) and (6.95,1.4) .. (10.93,3.29) ;
\draw (301,118.4) node [anchor=north west][inner sep=0.75pt] {$\bullet $};
\draw (316,129.4) node [anchor=north west][inner sep=0.75pt] {$\bullet $};
\draw (343,93.4) node [anchor=north west][inner sep=0.75pt] {$\bullet $};
\draw (302,27.4) node [anchor=north west][inner sep=0.75pt] {$\bullet $};
\draw (317,27.4) node [anchor=north west][inner sep=0.75pt] {$\bullet $};
\draw (344,27.4) node [anchor=north west][inner sep=0.75pt] {$\bullet $};
\draw (243,122.4) node [anchor=north west][inner sep=0.75pt] {$ \begin{array}{l}
\mathfrak{M}_{0}\\
\end{array}$};
\draw (228,20.4) node [anchor=north west][inner sep=0.75pt] {$f_I(\mathfrak{M}_{0})$};
\draw (289,72.4) node [anchor=north west][inner sep=0.75pt] {$f_I$};
\end{tikzpicture}
\caption{Illustration of the map $f_I$. Blocks of the layered permutation
$\mathfrak{M}_0$ are mapped to equivalence classes of $f_I(\mathfrak{M}_0)$, and
points are mapped in an order-preserving manner.}
\end{figure}
It is perhaps clear from Figure 2 that convex linear orders and layered
permutations are uniformly interdefinable, but we now verify the details.
\begin{lemma} \label{lemma:id2}
Layered permutations and convex linear orders are uniformly interdefinable.
\end{lemma}
\begin{proof}
A finite layered permutation is determined, up to isomorphism, by the number of
points in each of its blocks; likewise, a finite convex linear order is determined
in the same manner by the number of points in each equivalence class. Because
the map $f$ sends blocks of $\mathfrak{M}_0$ to equivalence classes of
$f(\mathfrak{M}_0)$ of the same size, $f$ is injective. The number of possible
size $n$ convex linear orders is equal to the number of size $n$ layered
permutations (because every block structure is allowed), therefore, $f$ is
bijective as well.\\
Let $\mathfrak{M}_0$, $\mathfrak{M}_0'$ be two $\mathcal{C}_0$ -structures (layered
permutations) with $f(\mathfrak{M}_0) = f(\mathfrak{M}_0')$. From the definition of $<_1$, it is clear
that $$\mathfrak{M}_0 \models a <_1 b \iff f(\mathfrak{M}_0) \models a < b$$
We verify $$\mathfrak{M}_0 \models a <_2 b \iff f(\mathfrak{M}_0) \models (a \mathrel{E} b \wedge b < a) \vee (\neg a \mathrel{E} b \wedge a < b)$$ \label{id:permutation}
\noindent $(\Rightarrow)$
Suppose $a$ and $b$ are in the same block, then, because $\mathfrak{M}_0 \models
a <_2 b$, we have that $\mathfrak{M}_0
\models b <_1 a$, so $f(\mathfrak{M}_0) \models b < a$ (because the orders $<_1$
and $<$ agree). When $a$ and $b$ are in
different blocks, $f(\mathfrak{M}_0) \models \neg a \mathrel{E} b$; furthermore,
because $\mathfrak{M}_0 \models a <_2 b$, and $a$, $b$ are in different blocks,
$\mathfrak{M}_0 \models a <_1 b$, so $f(\mathfrak{M}_0) \models a < b$.\\
\noindent $(\Leftarrow)$
First suppose $a$ and $b$ are in the same equivalence class.
Then, $f(\mathfrak{M}_0) \models b < a$ and $\mathfrak{M}_0 \models b <_1 a$.
Because $a$, $b$ are in the same class, they are in the same block of
$\mathfrak{M}_0$. Since $\mathfrak{M}_0 \models b <_1 a$ and $a$, $b$ are
in the same block, we have $\mathfrak{M}_0 \models a <_2 b$.
When $a$ and $b$ are in
different equivalence classes, $a < b$ in $\mathfrak{M}_0$, so $a <_1 b$ in $\mathfrak{M}_1$. Since
$a$ and $b$ are in classes and thus different blocks, the orders $<_1$ and $<_2$ agree, giving $a <_2 b$.
\end{proof}
\begin{theorem}\label{theorem:permlimlaw}
Layered permutations admit a logical limit law.
\end{theorem}
\begin{proof}
By Lemma \ref{lemma:id2}, layered permutations are uniformly interdefinable with
convex linear orders. Because convex linear orders admit a logical limit law,
layered permutations admit one as well by Lemma \ref{lemma:inter-translation}.
\end{proof}
Zero-one laws have been extensively studied in the context of {\em homogeneous} structures, in the sense of \cite{cameron2002homogeneous}. The homogeneous permutations are classified in \cite{cameron2002homogeneous}, and Theorem \ref{theorem:permlimlaw} completes the picture of their logical limit laws. The increasing and decreasing permutations are uniformly interdefinable with linear orders, and so admit a zero-one law (e.g., see Section 2.6.2 of \cite{RG}), while the class of all permutations doesn't admit a logical limit law by \cite{Foy}. The remaining classes are the layered and ``skew layered'' permutations (consisting of decreasing blocks of increasing permutations), which are uniformly interdefinable by replacing $<_2$ with its reverse.
\section{Compositions}
Informally, a composition consists of an equivalence relation
$\mathrel{E}$ along with a linear order $\prec_1$ on $\mathrel{E}$-classes (but
not on points of the classes themselves). This description naturally
corresponds to a composition of $n$ in the usual sense, i.e. an ordered tuple of
positive integers summing to $n$, and the embeddability order agrees with the order
on compositions in the usual sense given in \cite{bergeron1995standard}.
Compositions can be obtained from
convex linear orders by forgetting the order between points within the same
$E$-class. We formalize this by passing through the following notion of fractured
orders.
Let $\mathcal{L}_0 = \{ \mathrel{E}, < \}$ be the language of convex linear orders
as before. Define a language $\mathcal{L}_1 = \{\mathrel{E}, \prec_1, \prec_2 \}$
consisting of three relation symbols (an equivalence relation and two partial
orders), and let the reduct $\mathcal{L}_2 \subset \mathcal{L}_1$ be given by
$\mathcal{L}_2 = \{ \mathrel{E}, \prec_1 \}$. In a fractured order, we start with a
convex linear order $<$ and break it into two parts: $\prec_1$ between
$\mathrel{E}$-classes, and $\prec_2$ within $\mathrel{E}$-classes.
Formally, we define the class of finite fractured orders $\mathcal{F}$
to be the class of $\mathcal{L}_1$-structures satisfying:
\begin{enumerate}
\item $\prec_1$, $\prec_2$ are partial orders
\item $\mathrel{E}$ is an equivalence relation
\item Distinct points $a$, $b$ are $\prec_1$-comparable iff they are not
$E$-related
\item Distinct points $a$, $b$ are $\prec_2$-comparable iff they are
$E$-related
\item $a \mathrel{E} a'$, $a \prec_1 b \Rightarrow a' \prec_1 b$ (convexity)
\end{enumerate}
Although $\prec_1$ is a partial order on points, Axioms 3 and 5 say it is
essentially equivalent to a linear order on E-classes. So, compositions may be
defined formally as $\mathcal{L}_2$-reducts of
fractured orders. To show that compositions admit a logical limit law, we show
that fractured orders are uniformly interdefinable with convex linear orders,
and that any composition admits a unique expansion to a fractured
order.
\begin{lemma} \label{lemma:c-oer}
Convex linear orders and finite fractured orders are uniformly interdefinable.
\end{lemma}
\begin{proof}
Define a map $f_I:\mathcal{F} \to \mathcal{C}_0$ such that:
\begin{itemize}
\item $\mathfrak{M}_1 \models a \mathrel{E} b \iff f_I(\mathfrak{M}_1)
\models a \mathrel{E} b$
\item $\mathfrak{M}_1 \models a \prec_1 b \iff f_I(\mathfrak{M}_1) \models
\neg a \mathrel{E} b \wedge a < b$
\item $\mathfrak{M}_1 \models a \prec_2 b \iff f_I(\mathfrak{M}_1) \models a
\mathrel{E} b \wedge a < b$
\end{itemize}
The order $<$ is total and convex by Axioms $3$ and $4$ in the definition of
$\mathcal{F}$. Both convex linear orders and compositions are determined
(up to isomorphism) by the number of points in each class, therefore the number of fractured orders of
size $n$ is equal to the number of convex linear orders of
size $n$. Because $f_I$ preserves $E$-classes, it is injective, and therefore a
bijection on structures of size $n$.
\end{proof}
\begin{lemma} \label{lemma:reduct}
Let $\mathcal{L}$ be a language and $\mathcal{L}'$ a reduct of $\mathcal{L}$.
Given a class $\mathcal{C}$ of $\mathcal{L}$-structures which admits a
logical limit law, any class $\mathcal{C}' $ of $\mathcal{L}'$-structures
which expand uniquely to $\mathcal{C}$-structures also admits a logical limit law.
\end{lemma}
\begin{proof}
The map $f$ is taken to be the map sending a structure in $\mathcal{C}'$ to
its unique expansion in $\mathcal{C}$. Because this expansion is unique, $f$ is
bijective on structures of size $n$ for all $n$. We take $g$ to be the identity
map on formulas (as $\mathcal{L}'$ is a reduct of $\mathcal{L}$). Then these maps satisfy the requirements of Lemma \ref{lemma:translation}.
\end{proof}
\begin{lemma}\label{lemma:expand}
Every composition expands uniquely to a fractured order, up to
isomorphism.
\end{lemma}
\begin{proof}
There is a unique way (up to isomorphism) to linearly order
each $\mathrel{E}$-class individually. Because ordering these classes determines
$\prec_2$, there is a unique (up to isomorphism) way to define $\prec_2$ on any
composition, expanding it to a fractured order.
\end{proof}
\begin{theorem}\label{theorem:oeqlimlaw}
Compositions admit an unlabeled logical limit law.
\end{theorem}
\begin{proof}
By Lemma \ref{lemma:c-oer}, fractured orders are uniformly interdefinable
with convex linear orders (which admit an unlabeled logical limit law).
Because every unlabeled composition expands uniquely to a
fractured order, by Lemma \ref{lemma:reduct} we have an unlabeled limit law for
compositions.
\end{proof}
\section{Questions}
Our methods prove only an unlabeled limit law for compositions, hence, it is natural to ask the following:
\begin{question}
Do compositions admit a labeled logical limit law?
\end{question}
The original motivation for this work was to consider limit laws for permutation
classes.
Compton \cite{Bell} devised a method for proving limit laws on
classes of structures based on analyzing growth rate, i.e., function
counting the number of unlabeled structures of size $n$ in the class, for each $n$, assuming the classes are
closed under disjoint union. Although linearly ordered structures are not
closed under disjoint union, $\oplus$ provides a non-symmetric analogue.
\begin{question} \label{question:compton}
Can the method of Compton outlined in \cite{Bell} be extended to classes of
ordered structures using the $\oplus$ operation?
\end{question}
In particular, Compton's method shows that slow growth rate gives rise to
limit laws. Although the class of all permutations has
growth rate $n!$, any other class has at most exponential growth rate \cite{marcus2004excluded}, which
is comparatively slow.
A positive answer to Question \ref{question:compton} would suggest the following,
although it might be approached by other means.
\begin{question}
Do all sum-closed permutation classes (besides the class of all permutations)
admit a logical limit law?
\end{question}
The gap in growth rates between proper permutation classes and the class of all permutations is one manifestation of the fact that proper permutation classes are comparatively tame. This is also witnessed by the fact that they have bounded twin-width, in the sense of \cite{bonnet2021twin}, which may help answer the following.
\begin{question}
Do all permutation classes (besides the class of all permutations) admit a
logical limit law?
\end{question}
\subsection{Acknowledgements}
We thank the referees for their comments improving the exposition of the paper.
|
\section{Deterministic graphical finite $n$-person games}
\label{s1}
Nash-solvability (NS) of these games is discussed in \cite{BGMOV18}.
Here we use the concepts, definitions, and notation from this paper.
NS holds for $n=2$ and may fail for $n > 2$; see respectively Sections 5 and 2 of \cite{BGMOV18}.
The first NE-free example was constructed for $n=4$ in the preprint \cite{GO14A}
and published in \cite{Gur15}
\footnote{Vladimir Oudalov refused to co-author paper~\cite{Gur15} published in Russia.}
Then, a much more compact example for $n=3$ appeared in \cite[Section 3]{BGMOV18}.
In this example $c$ is the worst outcome for player 2
but for players 1 and 3 $c$ is better than (at least) two terminal outcomes.
Somewhat surprisingly, computations show that such two players exist
in each NE-free $n$-person DG game (for all $n > 2$).
In other words, condition (C22) given in Abstract implies NS.
We call this conjecture "Catch 22".
It is stronger than conjecture "(C) implies NS",
since condition (C22) is weaker than (C).
\smallskip
The latter means that
every player strictly prefers each terminal play to each infinite one,
which looks natural and can be easily interpreted \cite[Section 3.2]{GN21A} for examples.
If (C) does not imply NS then there exists a DG game that has a cyclic NE
(with outcome $c$) but has no terminal NE
(with an outcome $a \in V_T$) although
$V_T$ can be reached from the initial position \cite[Proposition 3]{BGMOV18}.
This property looks strange but not impossible.
\medskip
To formulate (C22) more accurately
denote by $k_c(i)$ the number of terminals that are
worse than $c$ for player $i \in I = \{1, \ldots, n\}$.
Then, (C22) does not hold if and only if
$k_c(i) \geq 2$ for (at least) two players.
\bigskip
Digraph $G$ is called {\em bidirected} if
each non-terminal move in it is reversible, that is,
$v' v'' \in E(G)$ if and only if $v'' v' \in E(G)$
unless $v'$ or $v''$ is a terminal position.
We conjecture that every $n$-person DG game on a finite bidirected digraph is NS.
\section{Deterministic graphical multi-stage (DGMS) finite $n$-person games}
\label{s2}
Recently, DG games were generalized and DGMS games were introduced
in \cite[Section 1]{Gur18}; see also \cite[Section 4.1]{GN21B}.
Here we use to the concepts, definitions, and notation from these papers.
It is well-known that every digraph $G$ is uniquely partitioned
into the strongly connected components (SCC).
An SCC will be called;
\begin{itemize}
\item [] {\em terminal} if it has no outgoing edges (exits);
\item [] {\em interior} if it is not terminal and contains a directed cycle;
\item [] {\em transient} if it is neither terminal not interior.
\end{itemize}
Obviously, each transient component consists of a single non-terminal vertex.
Contracting every SCC to a single vertex, one gets an acyclic digraph.
Conversely, given an acyclic digraph $G$,
let us substitute some of its non-terminal vertices
by strongly connected digraphs that contain directed cycles.
They will form the interior SCCs of the obtained digraph $G'$,
while each of the remaining non-terminal vertices of $G$ will form
a transient SCC in $G'$.
\medskip
It seems logical to assign a separate outcome $a$
(resp., $c$) to each terminal (resp., interior) SCC, thus,
getting the set of outcomes
\newline
$O = \{a_1, \ldots, a_p\} \cup \{c_1 \ldots, c_q\}$,
which define a DGMS game.
Merging the latter $q$ outcomes in it we obtain a DG game.
Obviously, merging outcomes of a game form respects NS.
Hence, an $n$-person DGMS game may have no NE when $n > 2$.
The main examples of \cite{GO14A,BGMOV18} show this.
\smallskip
Yet, any two-person DGMS game has a NE, which
can be determined by a slightly modified backward induction (BI) procedure
\cite[Section 1.5]{Gur18}; see also \cite[Section 4.1]{MN21B}.
It was proven that finite $n$-person DG and DGMS game forms are tight, for all $n$.
(Note that merging outcomes respects tightness.)
Yet, tightness is equivalent with NS only when $n=2$;
for $n > 2$ tightness is neither necessary nor sufficient for NS.
Note also that the modified BI works only for $n=2$ and
even in this case the obtained NE may be not subgame perfect,
unlike the NE obtained by the standard BI.
\medskip
Obviously, contracting each terminal SCC of a DGSM game
to a single vertex one does not change the game.
Finally, note that a DGSM game with a unique interior SCC is DG game.
\bigskip
Conditions (C) and (C22) can be adapted to DGMS games as follows.
\medskip
$(C')$ For any player $i \in I$
every interior SCC $c$ is better than each terminal SCC $a$.
\medskip
To formulate $(C'22)$
denote by $k(i, c_j)$ the number of terminal SCC that are
worse than $c_j$ for player $i \in I$.
Then, (C'22) does not hold if and only if
there are (at least) two distinct players $i', i'' \in I$
and two, not necessarily distinct,
interior SCC $c_{j'}$ and $c_{j'}$ such that
$k(i', c_{j'}) \geq 2$ and $k(i'', c_{j''}) \geq 2$.
Obviously $(C')$ implies $(C'22)$.
We conjecture that $(C'22)$ implies NS of finite $n$-person DGMS games
(for $n > 2$). .
\section{Subgame perfect NE-free DG games}
\label{s3}
NE $s = (s_1, \ldots, s_n)$ in a finite $n$-person DG game
is called {\em uniform} if it is a NE with respect to
every initial position $v \in V \setminus V_T$.
In the literature uniform NE (UNE) are frequently
referred to as {\em subgame perfect NE}.
By definition, any UNE is a NE, but not vice versa.
A large family of $n$-person UNE-free DG games satisfying (C)
can be found in \cite[Section 3.3]{GN21A} for $n > 2$,
and even for $n=2$ in \cite[the last examples in Figures 1 and 3]{BEGM12}.
Every NE-free game contains a UNE-free subgame \cite[Remark 3]{BGMOV18}.
Indeed, consider an arbitrary NE-free DG game $\Gamma$
and eliminate the initial position $v_0$ from its graph $G$.
The obtained subgame $\Gamma'$ is UNE-free.
Indeeed, assume for contradiction that $\Gamma'$ has a UNE $s = (s_1, \ldots, s_n)$.
Then, $\Gamma$ has a NE, which can be obtained by backward induction.
The player beginning in $v_0$ chooses a move that maximizes his/her reward,
assuming that $s$ is played in $\Gamma'$ by all players.
Clearly $s$ extended by this move forms a NE in $\Gamma$, which is a contradiction.
Thus, searching for a NE-free DG game
(satisfying (C) or (C22))
one should begin with a UNE-free DG game
trying to extend it with an acyclic prefix.
This was successfully reallized in \cite{GO14A,BGMOV18},
where condition (C22) (not to mention (C)) was waved.
However, under these assumptions all tries failed.
\subsection*{Acknowledgement}
The authors was partially supported by the RSF grant 20-11-20203.
|
\section{Random permutations}\label{sec:rp}
\subsection{General introduction}
The number of cycles $\mathscr{C}(\sigma)$ of a random permutation $\sigma$ is one the most elementary aspects one might want to describe; its distribution is well known and particularly elegant when $\sigma$ is distributed uniformly in $\mathfrak{S}_M$. This famous result will serve here as a paradigm.
\begin{theorem}[Cycles of a uniform permutation]\label{thm:uniform_permutation} For $M \geq 1$, if $\sigma$ is a uniform element of $\mathfrak{S}_M$, then
\begin{equation}\label{uniform_gen_fun_1}
\mathbb{E} \left( t^{\mathscr{C}(\sigma)} \right)
= \frac{1}{M!} t (t+1) \cdots (t+M-1),
\end{equation}
which is equivalent to the equality in distribution
\begin{equation}
\mathscr{C}(\sigma) \stackrel{d}{=} B_1 + B_{2} + \cdots + B_{M}
\end{equation}
where $(B_k)_{k \geq 1}$ is a sequence of independent Bernoulli variables with $B_k \stackrel{d}{=} \mathrm{Ber}\left( \frac{1}{k}\right)$.
\end{theorem}
Polynomials such as \eqref{uniform_gen_fun_1} appear repeatedly in this context. This motivates the following notation: for any integer $n$, we define the rising and falling factorials as the polynomials
\begin{align}
\label{rising_fact} F_n^+ (X) & := X (X+1) \dots (X+n-1) \\
\label{falling_fact} F_n^- (X) & := X (X-1) \dots (X-n+1) = (-1)^{n} F_n^+(-X).
\end{align}
By convention, $F_0^+ = F_0^- = 1$.
Theorem \ref{thm:uniform_permutation} can then be rephrased as
\begin{equation}\label{uniform_gen_fun_2}
\mathbb{E} \left( t^{\mathscr{C}(\sigma)} \right)
= \frac{1}{M!} F_{M}^+ (t),
\end{equation}
a simple identity which describes entirely the distribution of $\mathscr{C}(\sigma)$, the number of cycles of a uniform permutation. The object of interest here is the number of cycles in the commutator
\begin{equation}
[\sigma, \tau]=\sigma \tau \sigma^{-1} \tau^{-1}
\end{equation}
between a uniform permutation $\sigma$ and a \textit{fixed} permutation $\tau$, so that the distribution at stake depends on $\tau$; more precisely, it only depends on the cycle structure of $\tau$. Diaconis \& al. \cite{Diaconis2014} studied the number of fixed points in such a commutator. Note that there are also several studies of the cycles of commutators between two uniform permutations, and even general words involving i.i.d. uniform permutations and their inverses: see for example \cites{Nic94, LP10}. Fixing $\tau$ can be thought of as a `quenched' version of these questions. \medskip
For some specific choices of $\tau$, we are able to describe exactly the distribution at stake; a related question of great interest is the asymptotic behavior of $\mathscr{C}([\sigma,\tau])$. For this, the number of fixed points of $\tau$ seems to be the main key parameter; one could hope to establish that, under the condition that $\tau$ has `few' fixed points, the number of cycles of $[\sigma,\tau]$ is asymptotically well approximated by that of a uniform even permutation. This, however, would have to be done by entirely different methods. The present work is solely concerned with a few explicit descriptions for fixed $M$, that can be derived \textit{via} random matrix theory. \\
In the rest of this section, we state our main results and establish a few preliminary facts. All proofs are provided in Section \ref{sec:rmt}.
\subsection{Results}
We prove the following results, where $F_M^{+}$ and $F_M^{-}$ denote the rising and falling factorial respectively, as defined in \eqref{rising_fact} and \eqref{falling_fact}.
\begin{theorem}\label{thm:one_cycle}
For any $M \geq 1$, if $\tau=(1, \dots, M) \in \mathfrak{S}_M$ is an $M$-cycle and $\sigma$ is uniformly distributed in $\mathfrak{S}_M$, then
\begin{equation}\label{one_cycle_gen_fun}
\mathbb{E} \left( t^{\mathscr{C}([\sigma,\tau])} \right)
= \frac{1}{(M+1)!} \left( F_{M+1}^+(t) - F_{M+1}^-(t) \right),
\end{equation}
which is equivalent to
\begin{equation}\label{odd_interpretation}
\mathscr{C}([\sigma,\tau])
\stackrel{d}{=}
\mathscr{C}(\nu),
\end{equation}
where $\nu$ is distributed according to the uniform distribution on $\mathscr{A}_{M+1}^c$, the complement of the alternating group.
\end{theorem}
Theorem \ref{thm:one_cycle} is equivalent to a closed formula for Hultman numbers, which was already known from different techniques \cite{Stanley}; our input in that case is simply to provide a new conceptual proof of this formula, relying on random matrix theory. Another case that this proof technique allows to solve explicitly is when $\tau$ the product of two cycles of equal size $M$ in $\mathfrak{S}_{2M}$.
\begin{theorem}\label{thm:two_cycles}
For any $M \geq 1$, if $\tau=(1, \dots, M)(M+1, \dots, 2M) \in \mathfrak{S}_{2M}$ is a product of two disjoint $M$-cycles and $\sigma$ is uniformly distributed in $\mathfrak{S}_{2M}$, then
\begin{multline}
\mathbb{E} \left( t^{\mathscr{C}([\sigma,\tau])} \right)
= \frac{F_{2M+1}^+ (t) - F_{2M+1}^- (t)}{(2M+1)!}
+ \frac{2}{(2M)!} \left( \frac{ F_{M+1}^+ (t) - F_{M+1}^- (t)}{M+1} \right)^2 \\
- \frac{2}{(2M)!} \sum_{k=0}^{M} (-1)^k
\frac{k!}{2M-k+1}
\binom{M}{k}^2 \left( F_{2M-k+1}^+ (t) + F_{2M-k+1}^+ (-t) \right)
\end{multline}
\end{theorem}
The last solvable case we present is in a sense the opposite extreme, when $\tau$ is a product of many transpositions.
\begin{theorem}\label{thm:many_trans} For any $M\geq 1$, if $
\tau = (1,2) \cdots (2M-1, 2M) \in \mathfrak{S}_{2M}
$
is a product of $M$ disjoint transpositions, and $\sigma$ be uniformly distributed in $\mathfrak{S}_{2M}$, then
\begin{equation}
\mathbb{E} \left( t^{\mathscr{C}([\sigma, \tau])} \right)
=
\frac{ 2^M M!}{(2M) !} \ F_{M}^+ \left( t^2/2 \right),
\end{equation}
which is equivalent to
\begin{equation}
\mathscr{C}([\sigma, \tau]) \stackrel{d}{=} 2 \sum_{k=1}^M B_k
\end{equation}
where $(B_k)_{k \geq 1}$ is a sequence of independent Bernoulli variables with $B_k \stackrel{d}{=} \mathrm{Ber}\left( \frac{1}{2k-1}\right)$.
\end{theorem}
The present methods and results suggest possible generalization to other particular cases of $\tau$. For instance, it is in principle possible to derive the distribution of $\mathscr{C}([\sigma,\tau])$ where $\tau$ is a product of disjoints $m$-cycles, thanks to the identities of \cite{DubachPowers}.
\subsection{First elementary facts}
Note that the commutator $[\sigma, \tau]$ can be interpreted as the product $\tau_1 \tau_2$, where
\begin{equation}
\tau_1 = \sigma \tau \sigma^{-1} , \qquad
\tau_2 = \tau^{-1},
\end{equation}
which implies that $\tau_1$ is a random permutation chosen uniformly among those who have the same cycle structure as $\tau$ (but independent of any other aspect of $\tau$). This is the reason why the distribution of $\mathscr{C}([\sigma, \tau])$ only depends on the cycle structure of $\tau$, and also explains the link with Hultman numbers and the result of Stanley \cite{Stanley}. \medskip
Another important remark is that the number of cycles in a commutator $[\sigma, \tau]$ always is of the same parity as $M$, the size of the permutations. This is for the simple reason that a commutator is an element of the alternating group $\mathfrak{A}_M$, which is equivalent to this property:
\begin{fact}\label{odd_fact} In the symmetric group $\mathfrak{S}_M$,
$\sigma \in \mathfrak{A}_M$ if and only if $\mathscr{C}(\sigma) \equiv M [2]$.
\end{fact}
\begin{proof}
If $\sigma \in \mathfrak{S}_M$ has cycles of sizes $C_1, \dots, C_{\mathscr{C}(\sigma)}$, then
\begin{equation}
\epsilon (\sigma) = \prod_{i=1}^{\mathscr{C}(\sigma)} (-1)^{C_i+1} = (-1)^{M + \mathscr{C}(\sigma)}
\end{equation}
and the result follows directly.
\end{proof}
This remark allows us to compute the generating function of the number of cycles in a uniform element of $\mathfrak{A}_M$.
\begin{proposition}\label{odd_distr}
If $\nu$ is uniform in $\mathfrak{A}_M$, then
\begin{equation}
\mathbb{E}_{\mathfrak{A}_M} \left( t^{\mathscr{C}(\nu)} \right)
= \frac{1}{M!} \left( F_{M}^+(t) + F_{M}^-(t) \right),
\end{equation}
and if $\nu$ is uniform in $ \mathfrak{A}_M^c = \mathfrak{S}_M \backslash \mathfrak{A}_M$, then
\begin{equation}
\mathbb{E}_{\mathfrak{A}_M^c} \left( t^{\mathscr{C}(\nu)} \right)
= \frac{1}{M!} \left( F_{M}^+(t) - F_{M}^-(t) \right).
\end{equation}
\end{proposition}
\begin{proof}
These formulas follow from Theorem \ref{thm:uniform_permutation} and a conditioning argument coherent with Fact \ref{odd_fact}: a uniform even (resp. odd) permutation is a uniform permutation, conditionally on the number of cycle having the right parity.
\end{proof}
\begin{remark} The family $(F_n^+)_{n \geq 0}$ is an eigenbasis for the discrete differential operator
\begin{equation}\label{def:discrete_differential}
\mathscr{D}: P(X) \mapsto P(X) - P(X-1).
\end{equation}
Indeed, $\mathscr{D} F_n^+ = n F_{n-1}^+$. In particular, the integer values of $F_{n}^+$ can be written in the following form, for $n,k \geq 1$:
\begin{equation}\label{fact_identity}
F_{n}^+(k) = \sum_{j=1}^k \mathscr{D} F_{n}^+(j)
= \sum_{j=1}^k n F_{n-1}^+ (j) = n \sum_{j=1}^k \frac{\Gamma(n+j-1)}{\Gamma(j)}.
\end{equation}
\end{remark}
\section{Random matrices}\label{sec:rmt}
This section presents a few key facts about the relevant ensembles of non-Hermitian random matrices; we then apply these to provide the proofs of the results announced in Section \ref{sec:rp}. We consider the following two types of Gaussian random matrices, related to the complex and the real Ginibre ensembles; the difference being that, for the sake of clarity, we work with \textit{unscaled} matrices, meaning that individual entries have variance $1$ instead of the usual $N^{-1/2}$. Throughout this section, then,
\begin{enumerate}
\item $G$ is an $N \times N$ complex Gaussian matrix, by which we mean that the entries $G_{ij}$ are i.i.d. standard complex Gaussian random variable.
\item $R$ is an $N \times N$ real Gaussian matrix, by which we mean that the entries $R_{ij}$ are i.i.d. standard real Gaussian random variable.
\end{enumerate}
The fact that all entries of these matrices are Gaussian makes it possible to access some relevant statistics using Wick's law, leading to a topological understanding of the contributions -- an approach known as \textit{genus expansion}. Some consequences of genus expansion for non-Hermitian matrices have been exposed in \cite{DubachPeled}, where a new conceptual proof of Theorem \ref{thm:uniform_permutation} was mentioned as a byproduct (p.9). Although this proof may seem a bit far-fetched (considering that the same propety follows from Feller coupling, which is a more `natural' explanation), we recall it here as it serves as a first example of the use of random matrices to answer questions about random permutations.
\begin{proof}[Proof of Theorem \ref{thm:uniform_permutation} via random matrix theory.]
For $G$ a complex Gaussian matrix, we have, using the independence of the entries and the additivity property of independent gamma variables,
\begin{equation}
\mathrm{tr} \left( G^* G \right)
= \sum_{i,j=1}^N |G_{ij}|^2
\stackrel{d}{=} \sum_{i,j=1}^N \gamma_1
\stackrel{d}{=} \gamma_{N^2},
\end{equation}
so that the moments are given explicitly by
$$
\mathbb{E} \left( \mathrm{tr} \left( G^* G \right)^M \right)
= \mathbb{E} \left( \gamma_{N^2}^M \right)
= \frac{\Gamma(N^2 + M)}{\Gamma(N^2)}
= F_{M}^+ (N^2).
$$
On the other hand, we have by genus expansion,
$$
\mathbb{E} \left( \mathrm{tr} \left( G^* G \right)^M \right)
=
\sum_{\phi \in \mathscr{P}( G^*G, \dots, G^*G )} N^{2c(S_{\phi}) - 2g (S\phi)},
$$
using the notations of \cite{DubachPeled}, namely: $\phi$ is an admissible pairing of edges, $S_{\phi}$ the resulting surface, $c(S_{\phi})$ its number of connected components, and $g(S_{\phi})$ the sum of their genera. In that case, as illustrated on Figure \ref{fig:Example1}, each surface $S_\phi$ obtained by the gluing of a certain numbers of $2$-gons is simply a union of spheres, so that $g (S_{\phi}) =0$. Moreover, there is a straightforward one-to-one correspondence between admissible pairings $\phi$ and elements $\sigma \in \mathfrak{S}_M$: the $G$-edge of face $j$ is matched with the $G^*$-edge of face $\sigma(j)$. In particular, $c (S_{\phi})= \mathscr{C}(\sigma)$. This allows us to write
$$
\mathbb{E}_{\sigma} \left( N^{2 \mathscr{C}(\sigma)} \right)
= \frac{1}{M!} \sum_{\sigma \in \mathfrak{S}_M} N^{2 \mathscr{C}(\sigma)}
= \mathbb{E}_G \left( \mathrm{tr} \left( G^* G \right)^M \right)
= \frac{1}{M!} F_{M}^+ (N^2).
$$
The above equality holds for any $N, M \geq 1$, so that it characterizes all relevant generating functions and proves the claim.
\end{proof}
\begin{figure}[ht]
\begin{center}
\begin{tikzpicture}
\draw (-5.3,1.7) node {$\sigma=\rm{Id} = (1)(2)(3)$};
\draw (-5.3,0) node {
\begin{tikzpicture}[scale=1.1]
\begin{scope}[thick,decoration={
markings,
mark=at position 0.51 with {\arrow{>}}}
]
\draw[thick,color=blue, postaction={decorate}] (-1,0.2) to[bend right=50] (-.8,0.7) to[bend right=50] (-1.2,0.7) to[bend right=50] (-1,0.2) ;
\draw[thick,color=blue, postaction={decorate}] (1,0.2) to[bend right=50] (1.2,0.7) to[bend right=50] (.8,0.7) to[bend right=50] (1,0.2) ;
\draw[thick,color=blue, postaction={decorate}] (0,-.8) to[bend left=50] (-.2,-.3) to[bend left=50] (.2,-.3) to[bend left=50] (0,-.8) ;
\filldraw[fill=black!5!white,draw=white] (-1,-0.5) to[bend right=50] (-1,0.5) to[bend right=50] (-1,-0.5);
\draw[postaction={decorate}] (-1,-0.5) to[bend right=50] (-1,0.5);
\draw[postaction={decorate}] (-1,-0.5) to[bend left=50] (-1,.5);
\filldraw[fill=black!5!white,draw=white] (1,-0.5) to[bend right=50] (1,0.5) to[bend right=50] (1,-0.5);
\draw[postaction={decorate}] (1,-0.5) to[bend right=50] (1,0.5);
\draw[postaction={decorate}] (1,-0.5) to[bend left=50] (1,.5);
\filldraw[fill=black!5!white,draw=white] (0,-1.5) to[bend right=50] (0,-.5) to[bend right=50] (0,-1.5);
\draw[postaction={decorate}] (0,-.5) to[bend left=50] (0,-1.5);
\draw[postaction={decorate}] (0,-.5) to[bend right=50] (0,-1.5);
\draw (-.5,0) node {$G$};
\draw (-1.5,0) node {$G^*$};
\draw (1.5,0) node {$G$};
\draw (.5,0) node {$G^*$};
\draw (.5,-1) node {$G^*$};
\draw (-.5,-1) node {$G$};
\draw (-1,0) node {$1$};
\draw (1,0) node {$2$};
\draw (0,-1) node {$3$};
\end{scope}
\end{tikzpicture}
};
\draw (0,1.7) node {$\sigma=(1,2)(3)$};
\draw (0,-.2) node {
\begin{tikzpicture}[scale=1.1]
\begin{scope}[thick,decoration={
markings,
mark=at position 0.51 with {\arrow{>}}}
]
\draw[thick,color=blue,postaction={decorate}] (-1,0) to[bend right=40] (1,0);
\draw[thick,color=blue,postaction={decorate}] (1,.2) to[bend right=90] (1.1,.7) to[bend right=20] (-1.1,.7) to[bend right=90] (-1,.2);
\draw[thick,color=blue,postaction={decorate}] (0,-1.2) to[bend right=50] (-.2,-1.7) to[bend right=50] (.2,-1.7) to[bend right=50] (0,-1.2) ;
\filldraw[fill=black!5!white,draw=white] (-1,-0.5) to[bend left=45] (-1,0.5) to[bend left=45] (-1,-0.5);
\draw[postaction={decorate}] (-1,-0.5) to[bend right=50] (-1,0.5);
\draw[postaction={decorate}] (-1,-0.5) to[bend left=50] (-1,.5);
\filldraw[fill=black!5!white,draw=white] (1,-0.5) to[bend right=50] (1,0.5) to[bend right=50] (1,-0.5);
\draw[postaction={decorate}] (1,-0.5) to[bend right=50] (1,0.5);
\draw[postaction={decorate}] (1,-0.5) to[bend left=50] (1,.5);
\filldraw[fill=black!5!white,draw=white] (0,-1.5) to[bend right=50] (0,-.5) to[bend right=50] (0,-1.5);
\draw[postaction={decorate}] (0,-.5) to[bend left=50] (0,-1.5);
\draw[postaction={decorate}] (0,-.5) to[bend right=50] (0,-1.5);
\draw (-.5,0) node {$G$};
\draw (-1.5,0) node {$G^*$};
\draw (1.5,0) node {$G$};
\draw (.5,0) node {$G^*$};
\draw (.5,-1) node {$G^*$};
\draw (-.5,-1) node {$G$};
\draw (-1,0) node {$1$};
\draw (1,0) node {$2$};
\draw (0,-1) node {$3$};
\end{scope}
\end{tikzpicture}
};
\draw (5.3,1.7) node {$\sigma=(1,2,3)$};
\draw (5.3,-.1) node {
\begin{tikzpicture}[scale=1.1]
\begin{scope}[thick,decoration={
markings,
mark=at position 0.51 with {\arrow{>}}}
]
\draw[thick,color=blue,postaction={decorate}] (-1,.1) to[bend left=20] (1,.1);
\draw[thick,color=blue,postaction={decorate}] (1,0) to[bend left=60] (1.5,-.5) to[bend left=40] (0,-1.3);
\draw[thick,color=blue,postaction={decorate}] (0,-1.3) to[bend left=40] (-1.5,-.5) to[bend left=60] (-1,0);
\filldraw[fill=black!5!white,draw=white] (-1,-0.5) to[bend right=50] (-1,0.5) to[bend right=50] (-1,-0.5);
\draw[postaction={decorate}] (-1,-0.5) to[bend right=50] (-1,0.5);
\draw[postaction={decorate}] (-1,-0.5) to[bend left=50] (-1,.5);
\filldraw[fill=black!5!white,draw=white] (1,-0.5) to[bend right=45] (1,0.5) to[bend right=50] (1,-0.5);
\draw[postaction={decorate}] (1,-0.5) to[bend right=50] (1,0.5);
\draw[postaction={decorate}] (1,-0.5) to[bend left=50] (1,.5);
\filldraw[fill=black!5!white,draw=white] (0,-1.5) to[bend right=50] (0,-.5) to[bend right=50] (0,-1.5);
\draw[postaction={decorate}] (0,-.5) to[bend left=50] (0,-1.5);
\draw[postaction={decorate}] (0,-.5) to[bend right=50] (0,-1.5);
\draw (-.5,0) node {$G$};
\draw (-1.5,0) node {$G^*$};
\draw (1.5,0) node {$G$};
\draw (.5,0) node {$G^*$};
\draw (.5,-1) node {$G^*$};
\draw (-.5,-1) node {$G$};
\draw (-1,0) node {$1$};
\draw (1,0) node {$2$};
\draw (0,-1) node {$3$};
\end{scope}
\end{tikzpicture}
};
\end{tikzpicture}
\end{center}
\vspace{-.2in}
\caption{Illustration of the random matrix proof of Theorem \ref{thm:uniform_permutation} for $M=3$. We represent three admissible pairings (out of a total of $6$) corresponding to the identity, a transposition and a $3$-cycle; the resulting surfaces are unions of spheres: three, two and one respectively.}
\label{fig:Example1}
\end{figure}
The proof of Theorem \ref{thm:many_trans} relies on the same phenomenon, with a real Gaussian matrix $R$ (see Section \ref{sec:last}). The proof of Theorems \ref{thm:one_cycle} and \ref{thm:two_cycles} on the other hand rely on another feature, specific to the complex Gaussian matrix $G$, namely: the decorrelation of high powers of its eigenvalues. This fact is known to hold more generally for determinantal point processes with radial symmetry in the complex plane.
\begin{theorem}[Hough, Krishnapur, Peres, Vir\'ag]\label{thm:high_powers}
Let $\{ \lambda_1, \cdots, \lambda_N \}$ be the set of eigenvalues of $G$ defined as above. For any $M \geq N$, the following identity in distribution holds
\begin{equation}
\{ \lambda_1^M, \cdots, \lambda_N^M \} \stackrel{d}{=} \{ \gamma_1^M e^{i \theta_1}, \dots, \gamma_N^M e^{i \theta_N} \}
\end{equation}
where the $\theta_i$'s are i.i.d. \! real variables uniform in $[0,2\pi]$, and $\gamma_1, \dots, \gamma_N$ are independent gamma variables with corresponding parameters $1, \dots, N$.
\end{theorem}
See \cite{HKPV} for a general presentation of this rather counter-intuitive property, and \cite{DubachPowers} for a generalization to small powers $M < N$. \medskip
The following lemma is the essential link between complex Gaussian random matrices and random permutations that will be used repeatedly in this note.
\begin{lemma}\label{lem:corresp}
If $C_1 + \cdots+ C_K =M$, and $G$ is a complex Gaussian matrix as defined above, then
\begin{equation}
\mathbb{E}_{G} \left( \prod_{k=1}^K \left| \mathrm{tr} G^{C_k} \right|^2 \right)
=
M! \
\mathbb{E}_{\sigma} \left( N^{\mathscr{C}([\sigma, \tau])} \right),
\end{equation}
where $\tau$ has $K$ cycles of respective sizes $C_1, \dots C_K$, and $\sigma \in \mathfrak{S}_M$ is a uniform permutation.
\end{lemma}
\begin{proof}
This follows from the diagramatic approach for non-Hermitian matrices presented in \cite{DubachPeled}, and the following remarks. First, we choose an arbitrary numbering of the $G$-edges from $1$ to $M$, and also number the $G^*$-edges symmetrically. To each admissible pairing (i.e. a bipartite pairing between the $G$-edges and the $G^*$-edges), one associates the permutation $\sigma \in \mathfrak{S}_M$ such that $\sigma(i)=j$ if and only if the $G$-edge number $i$ is paired with the $G^*$-edge number $j$. We call $\tau$ the fixed permutation such that $\tau(i)=j$ if and only if the $G$-edge $i$ follows directly the $G$-edge $j$ in the diagram, which implies that $\tau$ has the corresponding cycle structure. The key remark is then that
\begin{equation}
V(S_{\phi}) = \mathscr{C}([\sigma, \tau]),
\end{equation}
which can be justified in the following way: there are initially $M$ vertices of the $G$-edges, which eventually will be associated to the symmetric $M$ vertices of the $G^*$-edges; $V(S_{\phi})$ being the number of vertices remaining after the associations. It is enough to count the cycles of associations created in the $G$-vertices. If one defines `vertex number $i$' as the end-vertex of the oriented edge number $i$, the pairing associated to $\sigma$ forces the assimilation of the $G$-vertex $i$ with the $G^*$-vertex number $\tau (\sigma(i))$, and consequently the assimilation of the $G$-vertices encoded in permutation $\tau^{-1} \sigma^{-1} \tau \sigma$. Therefore, the number of free vertices in the end is the number of cycles in the commutator $[\sigma, \tau]$.
\end{proof}
\begin{remark}
We can also write, using Euler's identity:
\begin{equation}
\mathscr{C}([\sigma,\tau])
= V(S_{\phi})
= \chi(S_{\phi}) + E(S_{\phi}) -F(S_{\phi}) = M - 2(g(S_{\phi}) + K - c(S_{\phi}))
\end{equation}
It follows, in particular, that $M- \mathscr{C}([\sigma,\tau])$ is even, coherently with Fact \ref{odd_fact}.
\end{remark}
A closely related random matrix interpretation has been previously suggested by Alexeev \& Zograf \cites{Alexeev2011, Alexeev2014}, who related Hultman numbers to the quantities $\mathbb{E} \mathrm{tr} ( G^M G^{*M} )$, and therefore to the singular values of $G^M$. Lemma \ref{lem:corresp}, on the other hand, relates Hultman numbers to $\mathbb{E} |\mathrm{tr} G^M|^2$, and more generally, statistics of cycles of commutators to eigenvalues of powers of $G$. It is this choice of statistics that enables explicit computations, for instance \textit{via} Theorem \ref{thm:high_powers}.
\subsection{Proof of Theorem \ref{thm:one_cycle}: one large cycle}
\begin{proof}[Proof of Theorem \ref{thm:one_cycle}]
Let $M \geq 1$ be a fixed integer. We denote
$
P(t) := \mathbb{E}_{\sigma} \left( t^{\mathscr{C}([\sigma,\tau])} \right)
$
the generating function of the number of cycles of a commutator between a uniform permutation $\sigma$ and a large cycle $\tau$ in $\mathfrak{S}_M$. Clearly, $P$ is a polynomial of degree $M$, and $P(0)=0$. For any $N\geq 1$, let $G$ be a complex Gaussian matrix of size $N \times N$. Lemma \ref{lem:corresp} gives
\begin{equation}
P(N) = \frac{1}{M!} \mathbb{E}_G | \mathrm{tr} G^M |^2.
\end{equation}
We compute this quantity using Theorem \ref{thm:high_powers}, for any $1 \leq N \leq M$:
\begin{equation}
\mathbb{E} | \mathrm{tr} G^M |^2 = \sum_{i=1}^N \mathbb{E} |\lambda_i|^{2M} = \sum_{i=1}^N \frac{{\Gamma}(M+i)}{{\Gamma}(i)},
\end{equation}
so that $P(0) = 0$ and for every $1 \leq N \leq M$, using (\ref{fact_identity}),
\begin{equation}
P (N) = \frac{1}{M!} \sum_{i=1}^N \frac{{\Gamma}(M+i)}{{\Gamma} (i)} = \frac{1}{(M+1)!} F_{M+1}^+(N).
\end{equation}
Thus, the polynomial
$$
F_{M+1}^+(t) - (M+1)! P(t)
$$
is of degree $M+1$ and cancels at $t=0, 1, \dots, M$. Therefore it is a multiple of $F_{M+1}^-$. By inspection of the leading coefficient, it is exactly $F_{M+1}^-$. The main result follows; consequence \eqref{odd_interpretation} holds by Proposition \ref{odd_distr}.
\end{proof}
One remarquable consequence of this identity is the fact that $\mathscr{C}([\sigma,\tau])$ is distributed like a sum of independent Bernoulli variables, although their parameters are not as straightforward as those appearing in Theorems \ref{thm:uniform_permutation} and \ref{thm:many_trans}.
\begin{corollary}
If $\tau = (1, \dots, M) \in \mathfrak{S}_M$, there are parameters $(p_j)_{j=1}^{\lfloor M/2 \rfloor}$ such that
$$
M - \mathscr{C}([\sigma,\tau]) \stackrel{d}{=} \sum_{j=1}^{\lfloor M-1/2 \rfloor } 2 B_{p_j}
$$
where the $B_p$'s are independent Bernoulli variables.
\end{corollary}
\begin{proof}
The generating function (\ref{one_cycle_gen_fun}) satisfies the Lee-Yang property, i.e. it has only purely imaginary zeros. Indeed, if $\Re z > 0$, then for every $j=1, \dots, M$, $|z-j|>|z+j|$. But if $z \neq 0$ is a zero of $P_M = F_{M+1}^+ - F_{M+1}^-$, then obviously $z \neq 1, \dots, M$ and
$$
1 = \left| \frac{F_M^+ (z)}{F_M^- (z)} \right| = \prod_{j=1}^M \frac{|z+j|}{|z-j|} < 1
$$
which is a contradiction. The conclusion follows by using the parity of $P_M$ (same as $M$). It follows that the decomposition of $P_M$ as a product of irreducible polynomials in $\mathbb{R}[X]$ is
$$
P_M(X) = X^{1 + \delta_{2 | M}} \prod_{j=1}^{\lfloor M/2 \rfloor} (X^2 + |z_j|^2)
$$
where $z_j$ are the (purely imaginary) roots. Each factor can then be interpreted as the generating function of a variable $2 B_{p_j}$, with $p_j = \frac{1}{1+|z_j|^2}$.
\end{proof}
\subsection{Proof of Theorem \ref{thm:two_cycles}: two large cycles}
For any $M$, we define the polynomial of degree $2M+1$
\begin{equation} \label{def:Q}
Q_{2M+1} (X) := \sum_{k=0}^{M} (-1)^k
\frac{k!}{2M-k+1}
\binom{M}{k}^2 F_{2M-k+1}^+ (X)
\end{equation}
The proof of Theorem \ref{thm:two_cycles} requires to identify a particular values of this polynomial, which is done in Lemma \ref{lem:identify_Q}.
\begin{lemma}\label{lem:identify_Q}
The following identity holds:
\begin{equation}\label{ideux}
\sum_{k=1}^N \left( \frac{\Gamma(k+M)}{\Gamma(k)} \right)^2
= Q_{2M+1}(N).
\end{equation}
\end{lemma}
\begin{proof}
We will make use of the following classical identity, for any $m \leq n$,
\begin{equation}\label{connection_coefs}
F_m^- (X) F_n^- (X)
= \sum_{k=0}^{m} \binom{m}{k} \binom{n}{k} k! \ F_{m+n-k}^- (X),
\end{equation}
which can be combinatorially understood, for integer values of $X$, as counting in two different ways the number of possible pairs of lists of length $m$ and $n$ respectively, chosen among $X$ possible items. It follows from identity \eqref{connection_coefs} with $m=n=M+1$ that
\begin{equation}
F_{M+1}^+ (X)^2
= \sum_{k=0}^{M+1} (-1)^k \binom{M+1}{k}^2 k! F_{2M+2-k}^+ (X)
\end{equation}
So that $Q_{2M+1}$ is the unique polynomial of degree $2M+1$ such that
\begin{equation}
Q_{2M+1}(0) = 0 \quad \& \quad
\mathscr{D} Q_{2M+1} = (F_{M}^+)^2.
\end{equation}
In particular, for every $N \geq 1$,
\begin{equation}
Q_{2M+1}(N)
= \sum_{k=1}^N \mathscr{D} Q_{2M+1}(k)
= \sum_{k=1}^N F_M^+(k)^2
\end{equation}
which is the claim.
\end{proof}
We can now establish Theorem \ref{thm:two_cycles}, following the same method as for Theorem \ref{thm:one_cycle}.
\begin{proof}[Proof of Theorem \ref{thm:two_cycles}]
Let $M\geq 1$ be a fixed integer and denote
$P(t):= \mathbb{E}_{\sigma} (t^{\mathscr{C}([\sigma, \tau])})$, where $\tau = (1, \dots, M)(M+1, \dots, 2M) \in \mathfrak{S}_{2M}$ and $\sigma \in \mathfrak{S}_{2M}$ is uniform. $P$ is an polynomial of degree $2M$ such that $P(0)=0$. By Fact \ref{odd_fact}, $P$ is also even. For any $N\geq 1$, Lemma \ref{lem:corresp} gives
\begin{equation}
P(N)
= \frac{1}{(2M)!} \mathbb{E} |\mathrm{tr} G^M |^4.
\end{equation}
This quantity can be computed exactly for $1 \leq N \leq M$, using Theorem \ref{thm:high_powers} and then Lemma \ref{lem:identify_Q}:
\begin{align*}
\mathbb{E} |\mathrm{tr} G^M |^4 & = \mathbb{E} \left( \sum_{i,j,k,l=1}^N \lambda_i^M \lambda_j^M \overline{\lambda_k}^M \overline{\lambda_l}^M \right) \\
& = \sum_{i=1}^N \mathbb{E} |\lambda_i|^{4M} + 2\left(\sum_{i=1}^N \mathbb{E} |\lambda_i|^{2M}\right)^2 - 2\sum_{i=1}^N \left(\mathbb{E} |\lambda_i|^{2M}\right)^2 \\
& = \sum_{i=1}^N \frac{\Gamma(i+2M)}{\Gamma(i)} + 2 \left( \sum_{i=1}^N \frac{\Gamma(i+M)}{\Gamma(i)} \right)^2 - 2 \sum_{i=1}^N \left( \frac{\Gamma(i+M)}{\Gamma(i)} \right)^2 \\
& = \frac{1}{2M+1} F_{2M+1}^+(N) + \frac{2}{(M+1)^2} F_{M+1}(N)^2 - 2 Q_{2M+1}(N).
\end{align*}
so that we have, for $ 0 \leq N \leq M$,
\begin{align*}
P(N) & = \frac{1}{(2M+1)!} F_{2M+1}^+ (N) + \frac{2}{(2M)!(M+1)^2} F_{M+1}^+ (N)^2 - \frac{2}{(2M)!} Q_{2M+1} (N).
\end{align*}
Note that all terms $F_{2M+1}^+, F_{M+1}^+, Q_{2M}$ cancel at $-M, \dots, 0$. Therefore, the following polynomials have degree $2M$ and, by parity, coincide with $P$ on $2M+1$ values:
\begin{multline}
\frac{F_{2M+1}^+ (t) - F_{2M+1}^- (t)}{(2M+1)!}
+ \frac{2}{(2M)!} \left( \frac{ F_{M+1}^+ (t) - F_{M+1}^- (t)}{M+1} \right)^2 \\
- \frac{2}{(2M)!} \left( Q_{2M+1} (t) + Q_{2M+1} (-t) \right)
\end{multline}
therefore it is equal to $P$, which is the claim.
\end{proof}
\subsection{Proof of theorem \ref{thm:many_trans}: product of many transpositions}\label{sec:last}
We conclude this note with a proof of Theorem \ref{thm:many_trans}, as well as a byproduct of this method: namely, identities in distribution for $\mathrm{tr} (G^2)$ and $\mathrm{tr} (G_1 G_2)$ where $G_1,G_2$ are i.i.d. complex Gaussian matrices.
\begin{proof}[Proof of Theorem \ref{thm:many_trans}]
Let $R$ be a real Gaussian matrix as defined in the beginning of Section \ref{sec:rmt}. In particular, its coefficients are i.i.d. variables such that $R_{ij}^2 \stackrel{d}{=} 2 \gamma_{1/2}$, so that we have, on the one hand,
\begin{equation}
\mathrm{tr} ( R R^t ) = \sum_{i,j=1}^N R_{ij}^2
\stackrel{d}{=}
\sum_{i,j=1}^N 2 {\gamma}_{1/2}
\stackrel{d}{=} 2 {\gamma}_{N^2/2},
\end{equation}
which moments are given by
\begin{equation}
\mathbb{E} (\mathrm{tr} R R^t )^M = 2^M \frac{\Gamma(N^2/2 + M)}{\Gamma(N^2/2)}
= 2^M F_{M}^+ (N^2/2).
\end{equation}
On the other hand, in the genus expansion for real Gaussian matrices presented in \cite[Section 4.4]{DubachPeled}, the admissible pairings include all associations of $R$-edges and $R^t$-edges, with the \textit{caveat} that the corresponding edge-gluing is either direct (for mixed pairs $R$--$R^t$) or indirect (for pairs $R$--$R$ or $R^t$--$R^t$). In general, the resulting surface need not be oriented, and connected components may include, for instance, real projective planes, Klein bottles, etc. However, in this case, as in the proof of Theorem \ref{thm:uniform_permutation}, one can see that the resulting surface is always a union of spheres: indeed, only $2$-gons are being used, so that the resulting connected components are either spheres or projective planes; moreover, no connected component is a projective plane, as indirect pairings always happens an even number of times due to the specific structure of the faces (all have one $R$ edge and one $R^t$ edge). \medskip
The next important remark is that an admissible pairing between these $2M$ edges can be interpreted as a permutation $\nu \in \mathfrak{S}_{2M}$ which is a product of $M$ disjoint transpositions: one numbers the edges from $1$ to $2M$ and pairs edges $i$ and $j$ whenever $\nu(i)=j$, $\nu(j)=i$. Note that the main difference with the proof of Theorem \ref{thm:uniform_permutation} is that the pairing is not bipartite here, as all associations are admissible. Moreover, in the resulting surface,
\begin{equation}
V(S_{\phi}) = 2 c(S_{\phi}) - 2 g(S_{\phi}) = 2 c(S_{\phi}) = \mathscr{C}(\nu \tau)
\end{equation}
where $\tau$ is another product of $2M$ disjoint transpositions, which corresponds to the edges themselves (so that $\tau = (1,2) \cdots (2M-1,2M)$ if the edges have been numbered from left to right). If we denote by $\mathfrak{D}_{2M}$ the set of such products of transpositions, which has cardinal $(2M-1)!!$, we find that
$$
\mathbb{E} (\mathrm{tr} R R^t )^M
=
\sum_{\phi \in \mathscr{P}(RR^t, \dots, RR^t)} N^{ V(\Gamma_{\phi})}
=
\sum_{\nu \in \mathfrak{D}_{2M}} N^{ \mathscr{C}(\nu \tau)}
=
(2M-1)!! \ \mathbb{E}_{\sigma} \left( N^{ \mathscr{C}( [ \sigma, \tau ] )} \right)
$$
where $\nu \stackrel{d}{=} \sigma \tau \sigma^{-1}$ with $\sigma$ uniform. We conclude that:
$$
\mathbb{E}_{\sigma} \left( N^{\mathscr{C}([ \sigma, \tau ] )} \right) =
\frac{2^M}{(2M-1)!!}
F_{M}^+(N^2/2)
=
\prod_{k=1}^{M} \frac{N^2 + 2 k -2 }{2k-1}.
$$
As this is true for any $N,M \geq 1$, we deduce that for any $M$, the relevant generating function is given by
$$
\mathbb{E}_{\sigma} \left( t^{ \mathscr{C}(\sigma, \tau)} \right)
=
\prod_{k=1}^{M} \frac{t^2 + 2 k-2 }{2k-1}
=
\prod_{k=1}^{M} \mathbb{E} \left( t^{2 \mathrm{Ber}\left(\frac{1}{2k-1}\right))} \right),
$$
which is the claim.
\end{proof}
\begin{figure}[ht]
\begin{center}
\begin{tikzpicture}
\draw (-5.3,1.2) node {$\nu=(1,2)(3,4)$};
\draw (-5.3,0) node {
\begin{tikzpicture}
\begin{scope}[thick,decoration={
markings,
mark=at position 0.52 with {\arrow{>}}}
]
\draw[thick,color=blue] (-1,0.2) to[bend right=50] (-.8,0.7) to[bend right=50] (-1.2,0.7) to[bend right=50] (-1,0.2) ;
\draw[thick,color=blue] (1,0.2) to[bend left=50] (.8,0.7) to[bend left=50] (1.2,0.7) to[bend left=50] (1,0.2) ;
\filldraw[fill=black!5!white,draw=white] (-1,-0.5) to[bend right=45] (-1,0.5) to[bend right=45] (-1,-0.5);
\draw[postaction={decorate}] (-1,-0.5) to[bend right=45] (-1,0.5);
\draw[postaction={decorate}] (-1,-0.5) to[bend left=45] (-1,.5);
\filldraw[fill=black!5!white,draw=white] (1,-0.5) to[bend right=45] (1,0.5) to[bend right=45] (1,-0.5);
\draw[postaction={decorate}] (1,-0.5) to[bend right=45] (1,0.5);
\draw[postaction={decorate}] (1,-0.5) to[bend left=45] (1,.5);
\draw (-.5,0) node {$R$};
\draw (-1.5,0) node {$R^t$};
\draw (1.5,0) node {$R$};
\draw (.5,0) node {$R^t$};
\end{scope}
\end{tikzpicture}
};
\draw (0,1.2) node {$\nu=(1,4)(2,3)$};
\draw (0,-.2) node {
\begin{tikzpicture}
\begin{scope}[thick,decoration={
markings,
mark=at position 0.52 with {\arrow{>}}}
]
\draw[thick,color=blue] (-1,0.1) to[bend left=50] (1,0.1);
\draw[thick,color=blue] (1,-.2) to[bend left=90] (1.1,-.7) to[bend left=20] (-1.1,-.7) to[bend left=90] (-1,-.2);
\filldraw[fill=black!5!white,draw=white] (-1,-0.5) to[bend right=45] (-1,0.5) to[bend right=45] (-1,-0.5);
\draw[postaction={decorate}] (-1,-0.5) to[bend right=45] (-1,0.5);
\draw[postaction={decorate}] (-1,-0.5) to[bend left=45] (-1,.5);
\filldraw[fill=black!5!white,draw=white] (1,-0.5) to[bend right=45] (1,0.5) to[bend right=45] (1,-0.5);
\draw[postaction={decorate}] (1,-0.5) to[bend right=45] (1,0.5);
\draw[postaction={decorate}] (1,-0.5) to[bend left=45] (1,.5);
\draw (-.5,0) node {$R$};
\draw (-1.5,0) node {$R^t$};
\draw (1.5,0) node {$R$};
\draw (.5,0) node {$R^t$};
\end{scope}
\end{tikzpicture}
};
\draw (5.3,1.2) node {$\nu=(1,3)(2,4)$};
\draw (5.3,-.1) node {
\begin{tikzpicture}
\begin{scope}[thick,decoration={
markings,
mark=at position 0.52 with {\arrow{>}}}
]
\draw[thick,color=blue] (1,0.2) to[bend right=80] (1.2,.65) to[bend right=40] (-1,.15);
\draw[thick,color=blue] (-1,-0.2) to[bend right=80] (-1.2,-.65) to[bend right=40] (1,-0.15);
\filldraw[fill=black!5!white,draw=white] (-1,-0.5) to[bend right=45] (-1,0.5) to[bend right=45] (-1,-0.5);
\draw[postaction={decorate}] (-1,-0.5) to[bend right=45] (-1,0.5);
\draw[postaction={decorate}] (-1,-0.5) to[bend left=45] (-1,.5);
\filldraw[fill=black!5!white,draw=white] (1,-0.5) to[bend right=45] (1,0.5) to[bend right=45] (1,-0.5);
\draw[postaction={decorate}] (1,-0.5) to[bend right=45] (1,0.5);
\draw[postaction={decorate}] (1,-0.5) to[bend left=45] (1,.5);
\draw (-.5,0) node {$R$};
\draw (-1.5,0) node {$R^t$};
\draw (1.5,0) node {$R$};
\draw (.5,0) node {$R^t$};
\end{scope}
\end{tikzpicture}
};
\end{tikzpicture}
\end{center}
\vspace{-.1in}
\caption{Illustration of the proof of Theorem \ref{thm:many_trans} for $M=2$. The three possible pairings correspond to the three double-transpositions of $\mathfrak{S}_4$.}
\label{Example3}
\end{figure}
According to Lemma \ref{lem:corresp}, the same statistics can be accessed from the quantities $\mathbb{E} |\mathrm{tr} G^2 |^M$. This gives us the following description of the random variable $\mathrm{tr} (G^2)$.
\begin{proposition}\label{distrib_of_tr_G2}
Let $G$ be an complex Gaussian matrix, then
$$
\mathrm{tr} ( G^2 )
\stackrel{d}{=}
e^{i \theta} \sqrt{ 4 {\gamma}_1 {\gamma}_{\frac{N^2}{2}}}
$$
where $\theta$ is uniform in $[0,2 \pi]$ and all variables are independent.
\end{proposition}
\begin{proof}
We proceed according to the following steps. For every $M \geq 1$, with $\tau$ a product of $M$ disjoint transpositions,
\begin{align*}
\mathbb{E}_{G} \left( |\mathrm{tr}(G^{2})|^{2M} \right)
& =
(2M)! \
\mathbb{E}_{\sigma} \left( N^{\mathscr{C}([\sigma, \tau])} \right) & \text{(Lemma \ref{lem:corresp})} \\
& =
\frac{(2M)!}{(2M-1)!!}
\mathbb{E}_{R} \left(
\mathrm{tr} \left( R R^t \right)^M \right) & \text{(Genus expansion for $R$)} \\
& = 2^M M! \ \mathbb{E} \left( \left( 2 {\gamma}_{N^2/2}\right)^M \right) & \text{(Distribution of $\mathrm{tr} ( R R^t)$)} \\
& = \mathbb{E} \left( \left( 4 {\gamma}_1 {\gamma}_{N^2/2} \right)^M \right) &
\end{align*}
The claim follows by equality of all relevant moments, as
$$
\mathbb{E} \left( \mathrm{tr} G^{M_1} \mathrm{tr} G^{*M_2} \right) = 0
$$
when $M_1 \neq M_2$.
\end{proof}
Note that Proposition \ref{distrib_of_tr_G2} with $N=1$ is a particular case of a general identity for product of gamma distributions when parameters are half of consecutive integers. \\
It makes sense to compare Proposition \ref{distrib_of_tr_G2} with the corresponding identity with two independent complex Gaussian matrices:
\begin{proposition}\label{distrib_of_tr_G1G2}
Let $G_1, G_2$ be i.i.d. complex Gaussian matrices, then
$$
\mathrm{tr} ( G_1 G_2 )
\stackrel{d}{=}
e^{i \theta} \sqrt{ {\gamma}_1 {\gamma}_{N^2}}
$$
where $\theta$ is uniform in $[0,2 \pi]$ and all variables are independent.
\end{proposition}
\begin{proof}
Genus expansion yields cycles of $\sigma=\sigma_1 \sigma_2$, which is uniform:
$$
\mathbb{E} |\mathrm{tr} G_1 G_2|^{2M}
= (M!)^2 \mathbb{E}(N^{\mathscr{C}(\sigma_1 \sigma_2)})
= (M!)^2
\mathbb{E}_{\sigma}(N^{2\mathscr{C}(\sigma)})
= M! (M+N^2)!,
$$
and the result follows.
\end{proof}
\section*{Acknowledgments}
The author acknowledges funding from the European Union's Horizon 2020 research and innovation programme, under the Marie Sk{\l}odowska-Curie Grant Agreement No. 754411. \medskip
I would like to thank Percy Deift, Paul Bourgade, Yuval Peled and Igor Kortchemski for many discussions and advice on that topic.
\begin{bibdiv}
\begin{biblist}
\bib{Alexeev2011}{article}{
title={Hultman numbers, polygon gluings and matrix integrals},
author={Alexeev, N.},
author={Zograf, P.},
eprint={arXiv:1111.3061},
year={2011}
}
\bib{Alexeev2014}{article}{
title={Random matrix approach to the distribution of genomic distance},
author={Alexeev, N.},
author={Zograf, P.},
journal={Journal of Computational Biology},
volume={21},
number={8},
pages={622--631},
year={2014},
publisher={Mary Ann Liebert, Inc. 140 Huguenot Street, 3rd Floor New Rochelle, NY 10801 USA}
}
\bib{Diaconis2014}{article}{
title={Unseparated pairs and fixed points in random permutations},
author={Diaconis, P.},
author={Evans, S.N.},
author={Graham, R.},
journal={Advances in Applied Mathematics},
volume={61},
pages={102--124},
year={2014},
publisher={Elsevier}
}
\bib{DubachPowers}{article}{
author={Dubach, G.},
title={Powers of Ginibre eigenvalues},
journal={Electronic Journal of Probability},
volume={23},
pages={1--31},
year={2018}
}
\bib{DubachPeled}{article}{
author={Dubach, G.},
author={Peled, Y.},
title={On words of non-Hermitian random matrices},
journal={The Annals of Probability},
volume={49},
number={4},
year={2021},
pages={1886--1916}
}
\bib{Ginibre1965}{article}{
title={Statistical ensembles of complex, quaternion, and real matrices},
author={Ginibre, J.},
journal={Journal of Mathematical Physics},
volume={6},
number={3},
pages={440--449},
year={1965},
publisher={American Institute of Physics}
}
\bib{HKPV}{article}{
title={Determinantal processes and independence},
author={Hough, J. B.},
author={Krishnapur, M.},
author={Peres, Y.},
author={Vir{\'a}g, B.},
journal={Probability surveys},
volume={3},
pages={206--229},
year={2006},
publisher={The Institute of Mathematical Statistics and the Bernoulli Society}
}
\bib{LP10}{article}{
AUTHOR = {Linial, N.},
author= {Puder, D.},
TITLE = {Word maps and spectra of random graph lifts},
JOURNAL = {Random Structures Algorithms},
VOLUME = {37},
YEAR = {2010},
NUMBER = {1},
PAGES = {100--135},
ISSN = {1042-9832},
}
\bib{Nic94}{article}{
AUTHOR = {Nica, Alexandru},
TITLE = {On the number of cycles of given length of a free word in
several random permutations},
JOURNAL = {Random Structures Algorithms},
VOLUME = {5},
YEAR = {1994},
NUMBER = {5},
PAGES = {703--730},
ISSN = {1042-9832},
DOI = {10.1002/rsa.3240050506}
}
\bib{Stanley}{article}{
title={Two enumerative results on cycles of permutations},
author={Stanley, R. P.},
journal={European Journal of Combinatorics},
volume={32},
number={6},
pages={937--943},
year={2011},
publisher={Elsevier}
}
\end{biblist}
\end{bibdiv}
\end{document} |
\section{Continuum model for ATMG}
In this work we consider a class of alternating twisted multilayer graphene (ATMG), which consist of three sets of graphene multilayers with the number of layers denoted by $M$, $L$, $N$ respectively. The stacking sequence within each set of multilayers can be Bernal ($ABA$), rhombohedral ($ABC$), or a mixture of the two. These multilayers are stacked from bottom to up in the $M$-$L$-$N$ sequence, where the $N$ ($L$) layers and the $L$ ($M$) layers are twisted by an angle $\theta$ (-$\theta$) as schematically shown in Fig.~\ref{fig1}(a). Such a system forms a moir\'e pattern in real space with the moir\'e superlattice constant $L_s\!=\!a/(2\sin{\theta/2})$, where $a\!=\!2.46\,\mbox{\normalfont\AA}$ is the graphene lattice constant. The corresponding moir\'e Brillouin zone is shown in Fig.~\ref{fig1}(b). Similar to twisted bilayer graphene (TBG), the low energy states of the ATMG system are contributed by those from the atomic $K$ and $K'$ valleys, which are approximately decoupled from each other at the non-interacting level for small twist angles. Thus, it is generally assumed that the system preserves valley charge conservation at small twist angles \cite{BM}.
Therefore, we generalize the Bistritzer-MacDonald continuum model \cite{BM} to describe the low-energy states of the ATMG system for each valley and each spin, assuming the states from the $K$ and $K'$ valleys are completely decoupled. The continuum model for valley $\mu$ ($\mu=\mp$ for $K$ and $K'$ valleys) is expressed as
\begin{align}
&H^{\mu}_{ATMG}
= \left(\begin{array}{ccc}
H^{\mu}_{N} & \mathbb{U}_{\mu}^{\dagger}e^{-i\mu\Delta\mathbf{K}\cdot\mathbf{r}} & 0\\
\mathbb{U}_{\mu}e^{i\mu\Delta\mathbf{K}\cdot\mathbf{r}} & H^{\mu}_{L} & \mathbb{U}_{\mu}e^{i\mu\Delta\mathbf{K}\cdot\mathbf{r}}\\
0 & \mathbb{U}_{\mu}^{\dagger}e^{-i\mu\Delta\mathbf{K}\cdot\mathbf{r}} & H^{\mu}_{M}
\end{array}\right)
\label{eq:H-atmg}
\end{align}
where $H^{\mu}_{N}$, $H^{\mu}_{L}$ and $H^{\mu}_{M}$ denote the $\mathbf{k}\!\cdot\!\mathbf{p}$ Hamiltonians of the untwisted graphene multilayers, which consist of the Dirac fermions of each monolayer graphene and the interlayer hopping terms. $\mathbb{U}_{\mu}e^{i\mu\Delta\mathbf{K}\cdot\mathbf{r}}$ stands for the moir\'e potential term for valley $\mu$, which arises from the mutual twist between two sets of adjacent multilayers. $\Delta\mathbf{K}=(0,4\pi/(3L_s))$ is a vector characterizing the shift of Dirac points due to the twist.
The details of the continuum Hamiltonian Eq.~(\ref{eq:H-atmg}) are presented in Supplementary Information.
\section{Generic partition rules}
One can obtain various types of low-energy band structures from the Hamiltonian given by Eq.~(\ref{eq:H-atmg}). A careful study reveals that the ATMG systems can be roughly divided into three types based on their low-energy band dispersion: for type (\lowercase\expandafter{\romannumeral 1}) there is only one pair of flat bands for each valley and spin, which is similar to TBG; for type (\lowercase\expandafter{\romannumeral 2}) there is one pair of flat bands co-existing with some low-energy bands characterized by the dispersion $E(\mathbf{k})\sim k^{J}$($J$ is positive integer); and for type ({\lowercase\expandafter{\romannumeral 3}}) there are two pairs of flat bands.
In Fig.~\ref{fig1}(c)-(d) we show the band structures of two typical ATMG systems with $A-A-A$ and $A-ABA-A$ stacking, where the solid and dashed lines denote energy bands from the $K$ and $K'$ valleys respectively in chiral limit, i.e the intrasublattice coupling between twisted layers are zero. For the $A-A-A$ system, there is one pair of flat bands co-existing with a Dirac cone (per valley per spin), which can be categorized as type (\lowercase\expandafter{\romannumeral 2}) ATMG; while there are two pairs of flat bands for each valley and spin for the $A-ABA-A$ system, which is the simplest example of type ({\lowercase\expandafter{\romannumeral 3}}) ATMG. In what follows we will explain the origin of such intriguing low-energy dispersion and derive partition rules for generic ATMG systems.
To better illustrate the origin of these intriguing low-energy dispersion, we first consider the chiral limit in which all the intrasublattice couplings are turned off. Within the chiral limit, we first analyze the low-energy states of the untwisted multilayers, then discuss the effects of the moir\'e potentials at the twisted interfaces. It has been proposed that an untwisted multilayer graphene with arbitrary stacking sequence and with the total number of layers $N$ can be decomposed into $S_N$ chiral segments, within each of which the stacking chirality is unchanged. Then the low-energy states contributed by the $i$th chiral segment with the number of layers $J_i$ consists of a chiral doublet described by the following effective Hamiltonian
\begin{equation}
H_{J_{i}}^{\mu}(\textbf{k})\propto k^{J_{i}}[\cos(J_i \phi_{\textbf{k}})\sigma_{x}\pm\sin(J_i \phi_{\textbf{k}})\sigma_{y}]
\label{eq:doublet}
\end{equation}
where $\tan{\phi_{\mathbf{k}}}=k_y/k_x$, and $\sigma_{x,y}$ denote Pauli matrices in the sublattice space. Then the low-energy Hamiltonian of the untwisted $N$ layers from valley $\mu$ can be written as a direct sum of those of the $N_D$ chiral segments:
\begin{equation}
H^{\mu}_{N}\approx H^{\mu}_{J_1}\oplus H^{\mu}_{J_2}... \oplus H^{\mu}_{J_{S_N}}\;.
\label{eq:segment}
\end{equation}
Each segment contributes to a chiral doublet with the dispersion $E\sim k^{J_{i}}$ around $\textbf{K}^{\mu}$ point.
Then we consider the chirally decomposed $N$ layers are stacked with the other $M$ layers and are twisted by angle $\theta$. The moir\'e potential at the interface would couple the topmost chiral segment of the $N$ with the bottom-most segment of the $M$ layers, giving rise to a pair of flat bands for each spin and each valley. These flat bands would co-exist with the dispersive chiral doublets contributed by the remaining chiral segments (if any) of the two sets of multilayers\cite{tmgdecomposition}. The ATMG system introduces additional complexity due to the additional multilayers ($L$ layers) and the additional twist. It turns out that the situations with the number of middle layers $L\!=\!1$ and $L\!>\!1$ need to be treated separately.
To evaluate the difference between ATMG systems with $L\!=\!1$ and $L\!>\!1$, we first consider alternating twisted trilayer graphene (TTG), i.e $M\!=\!L\!=\!N\!=\!1$. We will show that for TTG, in which three alternating twisted layers are coupled together, can be decoupled into a TBG-like Hamiltonian and a free Dirac-fermion Hamiltonian.
For the sake of convenience, we first apply a gauge transformation to the basis functions of the ATMG system:
\begin{equation}
\widetilde{\psi}_{ls,\mathbf{k}}^{\mu}(\textbf{r})=\psi_{ls,\mathbf{k}}^{\mu}(\textbf{r})e^{i\textbf{K}_{l}^{\mu}\cdot\textbf{r}}
\label{eq:gauge}
\end{equation}
where $\mathbf{K}^{\mu}_{l}$ denotes the Dirac point of valley $\mu$ and layer $l$, and $s$ is the sublattice index. Such a gauge transformation would remove the phase factor $e^{i\pm\mu\Delta\mathbf{K}\cdot\mathbf{r}}$ in the moir\'e potential term, and would move the Dirac points of the different twisted layers to the same origin.
Then we apply a unitary transformation to the alternating TTG Hamiltonian, $\tilde{H}^{\mu}_{TTG}=W^{\dagger}H^{\mu}_{TTG}W$. The Hamiltonian after the unitary transformation $\tilde{H}^{K}_{TTG}$ is expressed as
\begin{align}
\begin{split}
\tilde{H}^{\mu}_{TTG}
= \left(\begin{array}{ccc}
h^{\mu}(\mathbf{k}) &0 & \sqrt{2}\,\mathbb{U}_{\mu}\\
0 & h^{\mu}(\mathbf{k}) & 0\\
\sqrt{2}\,\mathbb{U}_{\mu}^{\dagger} & 0 & h^{\mu}(\mathbf{k})
\end{array}\right)\;,
\end{split}
\label{eq:H-ttg}
\end{align}
where $h^{\mu}(\mathbf{k})=-\hbar v_{F}\textbf{k}\cdot\bm{\sigma}_{\mu}$, with the Pauli matrices $\bm{\sigma}_{\mu}=(\mu\sigma_x,\sigma_y)$ defined in the sublattice space, and the unitary transformation matrix $W$ is expressed as
\begin{align}
\begin{split}
W=\left(\begin{array}{ccc}
0 & -\frac{1}{\sqrt{2}} & \frac{1}{\sqrt{2}}\\
1 &0 & 0\\
0& \frac{1}{\sqrt{2}} & \frac{1}{\sqrt{2}}
\end{array}\right)\;.
\end{split}
\label{eq:transform}
\end{align}
From Eq.~(\ref{eq:H-ttg}) it is immediately seen that the total Hamiltonian of alternating twisted trilayer graphene consists of a TBG-like part with the moir\'e potential rescaled by $\sqrt{2}$ and a free Dirac fermion part. The TBG part and Dirac fermion part are completely decoupled from each other, which contribute to one pair of flat bands co-existing with a Dirac cone as shown in Fig.~\ref{fig1}(c). Moreover, since the magic angle is determined by the ratio between the intersublattice component of the moir\'e potential and the Fermi velocity, the re-scaled moir\'e potential in Eq.~(\ref{eq:H-ttg}) implies that the magic angle for the TTG system is rescaled by the same factor, i.e., the new magic angle should be $\sqrt{2}\times 1.05^\circ\approx 1.5^{\circ}$.
For $L\!=\!1$ but $M, N\!>\! 1$, one can apply the chiral decomposition rule as discussed in Eqs.~(\ref{eq:doublet})-(\ref{eq:segment}) to the $M$ layers and $N$ layers. The topmost chiral segment from the $M$ layer and the bottom-most segment from the $N$ layers are coupled with the $L\!=\!1$ middle layer through the moir\'e potentials, contributing to one pair of flat bands co-existing with either a Dirac cone or a pair of quadratic bands. The remaining chiral segments (if any) in the $N$ layers and $M$ layers would contribute to additional $E(\mathbf{k})\sim k^{J}$ dispersive bands. On the other hand, when $L\!>\! 1$, one need to apply the chiral decomposition rule to the $L$ multilayers as well, and carefully study how the chiral doublets contributed by the $M$, $L$, and $N$ layers are coupled to each other through the moir\'e potentials at the two twisted interfaces.
After a comprehensive theoretical analysis based on a simplified $\mathbf{k}\!\cdot\!\mathbf{p}$ model approach (to be explained in the following section), we have derived a set of generic partition rules describing the low-energy band structures of ATMG systems in the chiral limit. First, the $M$, $L$, and $N$ multilayers are divided into $S_{M}$, $S_{L}$, and $S_N$ chiral segments (see Eqs.~(\ref{eq:doublet})-(\ref{eq:segment})), and the number of layers of the $i$th segment, say, in $N$ multilayer is denoted as $J_{N, i}$ ($i=1,..., S_N$). We also need to keep the chiral segments that are closest to the twisted interfaces to be as long as possible, i.e., we need to make a choice of chiral decomposition to make $J_{M,S_{M}}$, $J_{L,1}$, $J_{L,S_{L}}$ and $J_{N,1}$ as large as possible. Based on the above choice of chiral segments, we reach the following partition rules for the low-energy dispersion in the chiral limit:
(a) For $L=1$: the $J_{M,S_{M}}$, $J_{L,1}$ and $J_{N,1}$ chiral segments are coupled through the moir\'e potential generated by the alternating twisted structure. When the stacking chirality of $J_{M,S_{M}}$ and $J_{N,1}$ are the same, there are one pair of flat bands and one Dirac cone co-existing near $\textbf{K}_{\mu}$ point (per spin per valley); while if $J_{M,S_{M}}$ and $J_{N,1}$ have opposite stacking chiralities, there are one pair of flat bands and one pair of quadratic bands co-existing near $\textbf{K}_{\mu}$ point. The remaining chiral segments in the $M$ ($N$) multilayers would contribute to additional chiral doublets with the dispersion $E(\mathbf{k})\!\sim\! k^{J_{M,i}}$ ($E(\mathbf{k})\!\sim\! k^{J_{N,i}}$) near $\mathbf{K}_{\mu}$ ($\mathbf{K}_{\mu}'$) point.
(b) For $L>$1: if the $L$ multilayer can be divided into more than one chiral segments ($S_L>1$), there are two pairs of flat bands (double flat bands) around CNP; while if the $L$ multilayer is in the chiral (or rhombohedral) stacking sequence ($S_L=1$), there is only one pair of flat bands.
When $S_L\!>\!1$, the remaining chiral segment $\{J_{L,i},\; 2\leq i\leq S_L-1\}$ that are not coupled with the $M$ and $N$ multilayers would contribute to dispersive bands $E(\mathbf{k})\sim k^{J_{L,i}}$ around CNP. Similarly, the remaining chiral segments $\{J_{M,i}, \;1\leq i\leq S_M-1\}$ ($\{J_{N,i},\; 2\leq i\leq S_N$\}) from the $M$ ($N$) multilayer that are not coupled with the middle $L$ multilayer would contribute to the low-energy dispersive bands with $E(\mathbf{k})\sim k^{J_{M,i}}$ ($E(\mathbf{k})\sim k^{J_{N,i}}$).
In Table~\ref{tab:partitioning}, we illustrate some ATMG systems as typical cases and apply the partition rules described above to these systems to characterize their low energy band structures, where the notation $(m,n)$ means that there are $m$ pairs of bands with dispersion $E(\textbf{k})\sim\textbf{k}^{n}$ around $K_s$ or $K_s'$ points. For example, for $A$-$A$-$ABAC$ system, it can be divided into two parts including the alternating twisted layers $A$-$A$-$AB$ and untwisted layers $AC$. The twisted layers would contribute one pair of flat bands co-existing with a Dirac cone around $K_{\mu}$ point, while the untwisted layers give rise to a pair of quadratic bands($E(\textbf{k})\sim\textbf{k}^{2}$) centered at $K_{\mu}$ point.
\begin{table}[]\scriptsize
\centering
\caption{Typical cases for generic partition rules. The band structures are measured with full continuum model considering nearest neighbor for untwisted bands. We label $m$ bands with dispersion $E(\textbf{k})\sim\textbf{k}^{n}$ at $\textbf{K}_{\mu}(\textbf{K}^{\prime}_{\mu})$ as (m,n). }
\begin{tabular}{l|c|c|c}
\hline\hline
partitioning & number of & bands at $\textbf{K}$ & bands at $\textbf{K}^{\prime}$ \\
&falt bands & & \\
\hline
A-A-A & 2 & (1,1) & 0 \\
A-A-AB+AC & 2 & (1,1),(1,2) &0 \\
AB-A-BA & 2 & (1,1) & 0 \\
AB-A-AB & 2 & (1,2) & 0 \\
A-AB+A-A & 4 & 0 & 0 \\
A-ABC-A & 2 & / & / \\
A-AB+ABC-A & 4 & 0 & 0 \\
A+BA-AB+A+BA-AB+A &4 & (1,1) & (2,1) \\
\hline\hline
\end{tabular}
\label{tab:partitioning}
\end{table}
\section{Simplified $\textbf{k}\cdot\textbf{p}$ model}
The partition rules presented above can be derived using a simplified $\mathbf{k}\cdot\mathbf{p}$ model approach. In this approach, we write a $\mathbf{k}\cdot\mathbf{p}$ model within the moir\'e Brillouin zone by expanding the flat bands and the Dirac cones around the moir\'e $K_s$ or $K_s'$ points including the coupling terms between them. In the chiral limit both the flat bands and the Dirac cone can be solved exactly \cite{ashvin-analytical-tbg}, then we can analytically construct a greatly simplified $\mathbf{k}\cdot\mathbf{p}$ model in the basis of the zero modes (flat-band wavefunctions) and the Dirac fermions, and solve it exactly. From the analytic solutions of the simplified $\mathbf{k}\cdot\mathbf{p}$ model, we derive the partition rules for generic ATMG systems presented above. Such an approach can capture the essential low-energy physics, while neglecting the irrelevant high-energy bands obtained from a direct numerical diagonalization of the original continuum Hamiltonian.
To construct such a simplified $\textbf{k}\cdot\textbf{p}$ model for a generic ATMG system (in the chiral limit), we should first find proper unitary transformations to the original continuum Hamiltonian to decompose it into a form consisting of a TBG-like continuum Hamiltonian and free Dirac fermions, e.g., as illustrated in Eq.~(\ref{eq:H-ttg}).
To be specific, for any ATMG system with $L\!=\!1$, we can apply the unitary transformation given by Eq.~(\ref{eq:transform}) to the three alternating twisted layers, and keep other layers unchanged. The corresponding layer-mixed basis functions are:
\begin{align}
&\bar{\psi}_{\alpha s,\mathbf{k}}^{\mu}(\textbf{r})=\widetilde{\psi}_{Ls,\mathbf{k}}^{\mu}(\textbf{r})\;\nonumber\\
&\bar{\psi}_{\beta s,\mathbf{k}}^{\mu}(\textbf{r})=\frac{-1}{\sqrt{2}}\widetilde{\psi}_{Ms,\mathbf{k}}^{\mu}(\textbf{r})+\frac{1}{\sqrt{2}}\widetilde{\psi}_{Ns,\mathbf{k}}^{\mu}(\textbf{r})\;\nonumber\\
&\bar{\psi}_{\gamma s,\mathbf{k}}^{\mu}(\textbf{r})=\frac{1}{\sqrt{2}}\widetilde{\psi}_{Ms,\mathbf{k}}^{\mu}(\textbf{r})+\frac{1}{\sqrt{2}}\widetilde{\psi}_{Ns,\mathbf{k}}^{\mu}(\textbf{r})
\label{eq:unitary-transform}
\end{align}
where $\alpha$, $\beta$, $\gamma$ are the three mix-layer indices marking the basis functions after the unitary transformation, $\widetilde{\psi}_{l s,\mathbf{k}}^{\mu}(\textbf{r})$ with $l=M, L, N$ denote the basis wave functions (after the gauge transformation of Eq.~(\ref{eq:gauge})) of the three alternating twisted layers in the $M$, $L$, and $N$ sequence respectively, and $s=A, B$ is the sublattice index.
After such a transformation, the continuum Hamiltonian consists of a TBG-like part and a free Dirac fermion (as shown in Eq.~(\ref{eq:H-ttg}) contributed by the three alternating twisted layers, with the terms from the other layers being unchanged. One can solve for the zero-mode solutions for the TBG-like part at the renormalized magic angle $\sqrt{2}\times 1.05^{\circ}$, and expand all the Dirac cones around the Dirac points within the moir\'e Brillouin zone, then eventually obtain a $\mathbf{k}\cdot\mathbf{p}$ model in a greatly simplified form.
For L$>$1, above transformation is unnecessary, since there are two pairs of twisted layers giving rise to two TBG-like terms. For each of the TBG-like terms, we can obtain the zero-mode solution of magic-angle TBG in the chiral limit. The analytical wave functions for the zero modes in magic-angle TBG in the chiral limit are expressed as \cite{ashvin-analytical-tbg}:
\begin{align}
\Psi_{lA,\textbf{k}}(\textbf{r})=f_{\textbf{k}}(z)\bar{\psi}_{l A,\textbf{K}}(\textbf{r}),\;\nonumber\\
\Psi_{lB,\textbf{k}}(\textbf{r})=f^{*}_{\textbf{k}}(-z)\bar{\psi}_{l B,\textbf{K}}(\textbf{r})
\end{align}
where $l=\alpha,\gamma$ refers to the two mix-layer indices as defined in Eq.~(\ref{eq:unitary-transform})
, $\Psi_{l s,\textbf{K}}(\textbf{r})$ refers to the $s$ ($s=A, B$) sublattice component of the zero-mode solution at the Dirac point $\mathbf{K}$ \cite{ashvin-analytical-tbg}.
$f_{\textbf{k}}(z)=\vartheta_{a_{1},b_{1}}(z)/\vartheta_{a_{2},b_{2}}(z)$, where $\vartheta_{a,b}(z)$ is the theta function defined in the previous work\cite{ashvin-analytical-tbg}.
For the untwisted layers, we take the $\textbf{k}\cdot\textbf{p}$ Hamiltonians of the Dirac fermions and expand them around the Dirac points within the moir\'e Brillouin zone. The coupling between the zero modes from the twisted layers and the Dirac fermions from the untwisted layers can be evaluated by re-expressing the original interlayer hopping matrix in the basis of the zero-mode wavefunctions and the free Dirac-fermion states.
We take the $AB$-$A$-$AB$ and $AB$-$A$-$BA$ stacked ATMG systems as two typical examples to illustrate how the partition rule (a) works.
Following the procedures described above, we obtain the simplified $\textbf{k}\cdot\textbf{p}$ Hamiltonian for $AB$-$A$-$AB$ ($AB$-$A$-$BA$) stacked ATMG system:
%
\begin {align}
\begin{split}
H^{\mu}_{\textbf{k}\cdot\textbf{p}}&=\left(\begin{array}{ccccc}
h^{\mu}(\textbf{k}) & 0 &-h_{+}& \tilde{h}_{+}&0\\
0 &0 &0& 0&0\\
-h_{-}& 0& h^{\mu}(\textbf{k}) & 0 &h_{+(-)} \\
\tilde{h}_{-}&0&0& 0 &\tilde{h}_{+(-)}\\
0&0&h_{-(+)}&\tilde{h}_{-(+)}&h^{\mu}(\textbf{k})
\end{array}\right)\;,
\end{split}
\label{eq:AB-A-ABkp}
\end{align}
where
\begin{align}
h_{+}&=\left(\begin{array}{cc}
0 & 0 \\
t_{\perp} &0
\end{array}\right)\;,\tilde{h}_{+}=\left(\begin{array}{cc}
0 & 0 \\
\tilde{t}_{\perp} &0
\end{array}\right)\;
\label{eq:tildeh}
\end{align}
$h^{\mu}(\textbf{k})=-\hbar v_{F}\textbf{k}\cdot\bm{\sigma}^{\mu}$, and $h_{-}=h_{-}^{\dagger}$ for different stacking chirality. The ``$+(-)$" sign in $\tilde{h}_{+(-)}$ in Eq.~(\ref{eq:AB-A-ABkp}) applies to the $AB$-$A$-$AB$ ($AB$-$A$-$BA$) stacking. $t_{\perp}$ is the interlayer intersublattice hopping parameter. In the chiral limit we only consider nearest neighbor interlayer hopping term, and $\tilde{t}_{\perp}$ is the renormalized hopping parameter. We refer the readers to Supplementary Information(Appendix II) for more details about the simplified $\textbf{k}\cdot\textbf{p}$ model. %
We present the band structures of the above two ATMG systems in Fig.~(\ref{fig2}). In particular, in Fig.~\ref{fig2}(a)-(b) we show the band structures for $AB$-$A$-$AB$ and $AB$-$A$-$BA$ systems obtained from direct numerical diagonalizations of the full continuum Hamiltonian in chiral limit, while in Fig.~\ref{fig2}(c)-(d) the band structures are calculated by the $10\times 10$ simplified $\mathbf{k}\cdot\mathbf{p}$ model expanded around either $K_s$ or $K_s'$ point as given by Eq.~(\ref{eq:AB-A-ABkp}). For the $AB$-$A$-$AB$ system there are two flat bands co-existing with a Dirac cone (Fig.~\ref{fig2}(a) and (c)), while for the $AB$-$A$-$BA$ system there are two flat bands and one pair of quadratic band touching around $K_s$ ($K_s$') for $K$ ($K'$) valley. We see that the $10\times 10$ simplified $\mathbf{k}\cdot\mathbf{p}$ model can fully capture the essential low-energy physics, which fulfills the partition rules introduced above.
\begin{figure}[b]
\begin{center}
\includegraphics[width=8.5cm]{fig2f.png}
\caption{
The band structures of $AB-A-AB$ for $K$ valley(solid line) and $K^{\prime}$ valley(dashed line) obtained from (a) continuum model and (c) simplified $\textbf{k}\cdot\textbf{p}$; $AB-A-BA$ for K valley obtained from (b) continuum model and (d) simplified $\textbf{k}\cdot\textbf{p}$ model. The continuum models are constructed including nearest neighbor interlayer hopping and intersublattice moir\'e potential.}
\label{fig2}
\end{center}
\end{figure}
\section{Origin of double flat bands} Some of the ATMG systems may have two pairs of flat bands (double flat bands) for each valley and spin.
According to the partition rules, the double flat bands can only occur when $S_{L}>$1.
To illustrate the origin of the double flat bands, we first study two ATMG systems for $L\!=\!3$ with $A$-$ABA$-$A$ (with double flat bands) and $A$-$ABC$-$A$ (with one pair of flat bands) stacking based on the simplified $\mathbf{k}\cdot\mathbf{p}$ model.
To be specific, the simplified $\textbf{k}\cdot\textbf{p}$ Hamiltonian of the $A$-$ABA$-$A$ system can be written as a direct sum of a $6\times6$ matrix and two zero modes
\begin {align}
\begin{split}
H^{\textbf{k}\cdot\textbf{p}}_{A-ABA-A}=\left(\begin{array}{ccc}
0& \tilde{h}_{+}& 0 \\
\tilde{h}_{-}&h(\textbf{k})&\tilde{h}_{-}\\
0&\tilde{h}_{+}&0
\end{array}\right) \oplus \left(\begin{array}{cc}
0 &0 \\
0&0
\end{array}\right)\;,
\end{split}
\label{eq:A-ABA-A}
\end{align}
where the $6\times 6$ matrix consists of two zero modes coupled with a free Dirac fermion. Then it is straightforward to check that the eigenvalues of this $6\times 6$ matrix has two zero modes, so that the entire system has four zero modes, leading to double flat bands.
For $A$-$ABC$-$A$, the simplified $\textbf{k}\cdot\textbf{p}$ model can be written in a similar form
\begin {align}
\begin{split}
H_{A-ABC-A}=\left(\begin{array}{ccc}
0& \tilde{h}_{+}& 0 \\
\tilde{h}_{-}&h(\textbf{k})&\tilde{h}_{+}\\
0&\tilde{h}_{-}&0
\end{array}\right) \oplus \left(\begin{array}{cc}
0 &0 \\
0&0
\end{array}\right)
\end{split}
\label{eq:A-ABC-A}
\end{align}
Due to the change of stacking chirality, the $6\times 6$ Hamiltonian consisting of two zero modes coupled with a free Dirac fermion no longer has zero-mode solution. Thus, there are only two zero modes for the $A$-$ABC$-$A$ system, contributing to one pair of flat bands per spin per valley.
We can generalize the above argument for $L>3$ and $M,N>$1. According to the partition rules, the number of chiral segments in the $L$ multilayers have a dramatic influence on the number of flat bands, while the segments in the $M$ and $N$ multilayers do not. First, if the middle $L$ multilayers has a chiral stacking, i.e., $S_L\!=\!1$, then the system is equivalent to the $A$-$ABC$-$A$ system with a renormalized interlayer hopping parameter between untwisted layers $t^{\prime}_{\perp}$.
As a result, there is only one pair of flat bands (in the chiral limit) for the middle $L$ layers with chiral stacking. Second, if the $L$ multilayers can be decomposed into two segments($S_{L}\!=\!2$), the corresponding simplified $\mathbf{k}\cdot\mathbf{p}$ model is equivalent to that of the $A-ABA-A$ system with renormalized hopping parameters, which has double flat bands. Finally, for $S_{L}\!>\!2$:
we can argue that for alternating twist mutilayers, two pairs of moir\'e potentials are coupled together through several chiral segments within the $L$ multilayers, so that the more the segments are, the weaker the couple is. As a result, we can treat them as nearly decoupled and there are four flat bands for the $S_{L}\!>\!2$ situation. Therefore, if the middle $L$ multilayers has $S_L\!>\!1$, there would be double flat bands for each spin and valley degrees of freedom.
It is worthwhile to note that the simplified $\mathbf{k}\cdot\mathbf{p}$ model is constructed in the chiral limit neglecting all the intrasublattice couplings. Consequently, the partition rules derived from the simplified $\mathbf{k}\cdot\mathbf{p}$ model and the above discussions about double flat bands are rigorous only in the chiral limit. In a more realistic situation, one needs to include the intrasublattice component of the moir\'e potential and the further neighbor interlayer hoppings within the untwisted layers as in Eq.~(\ref{eq:fur}), which break chiral symmetry.
With these additional coupling terms, the otherwise exactly flat bands at the magic angle may acquire nonzero bandwidths, and the $E(\mathbf{k})\!\sim\!k^J$ bands may have slightly modified dispersion. However, despite these perturbative changes, the main conclusions sketched by the partition rules are unchanged.
It follows from the previous arguments that a mirror-symmetric ATMG system with $L\!>\!1$ must satisfy the condition of $S_{L}\!>\!1$, thus there must be two pairs of flat bands.
The double flat bands can be classified by the opposite mirror eigenvalues $\pm 1$ for ATMG with $m_z$ symmetry\cite{li-arxiv19}. In Fig.~\ref{fig:131}(a) we present the band structures of the $A$-$ABA$-$A$ system including the intrasublattice moire potential and the further neighbor interlayer hopping, where the color coding indicates the weight projected onto the middle layer. First, we note that compared with the band structures in the chiral limit shown Fig.~\ref{fig1}(d), in the realistic situation two of the four flat bands become more dispersive with the bandwidth $\sim\!25\,$meV, while the other pair of flat bands remain flat with very small bandwidth $\sim 10\,$meV. Second, we note that the weight of the middle layer for the pair of flat bands lower in energy with small bandwidth is vanishing, while the upper pair of flat bands with relatively large bandwidth have significant contributions from the middle-layer states. This is because the two flat bands lower/upper in energy have mirror eigenvalues $\mp1$, and the Bloch states with $-1$ mirror eigenvalue must have zero contribution from the middle layer. In the presence of Coulomb interactions, the $m_z$ symmetry could be broken spontaneously at certain filling factors.
\section{Correlation effects in mirror-symmetric ATMG system}
The ubiquitous flat bands in ATMG make these systems strongly susceptible to Coulomb interactions. Moreover, unlike magic-angle TBG, in magic-angle ATMG typically there are flat bands co-existing with other dispersive bands (such as Dirac cone) or double flat bands in these systems. The extra low-energy dispersive bands (e.g., Dirac cone) may be coupled with the flat bands under weak displacement fields and display different correlated states from those in magic-angle TBG at certain filling factors \cite{lian-trilayer2-prb21,bernevig-trilayer1-prb21}. On the other hand, in mirror-symmetric ATMG with double flat bands, e.g., in $A$-$ABA$-$A$ system, the extra pair of flat bands marked by opposite mirror eigenvalues introduce additional degrees of freedom. What are the correlated ground states in such double-flat-band systems at different filling factors of the flat bands, how the extra degrees of freedom (mirror eigenvalues) would play a role, and how the correlated states would differ from those of TBG, are all open questions. We try to answer these questions by studying the correlated states at different integer fillings of ATMG with $A$-$ABA$-$A$ stacking, the simplest mirror-symmetric ATMG system with double flat bands.
\subsection{Symmetry-breaking ground states}
We have performed unrestricted self-consistent Hartree-Fock calculations for the $A$-$ABA$-$A$ system within the subspace of the double flat bands. In particular, the intersite Coulomb interactions in graphene-based systems can be written as
\begin{equation}
\begin{split}
H_{C}=\frac{1}{2N_{s}}&\sum_{\alpha\alpha^{\prime}}\sum_{\textbf{k}_{a}\textbf{k}^{\prime}_{a}\textbf{q}_{a}}\sum_{\sigma\sigma^{\prime}}V(\textbf{q}_{a}) \\
&\times\hat{c}^{\dagger}_{\textbf{k}_{a}+\textbf{q}_{a},\alpha\sigma}\hat{c}^{\dagger}_{\textbf{k}_{a}^{\prime}-\textbf{q}_{a},\alpha^{\prime}\sigma^{\prime}}\hat{c}_{\textbf{k}_{a}^{\prime},\alpha^{\prime}\sigma^{\prime}}\hat{c}_{\textbf{k}_{a},\alpha\sigma}
\end{split}
\end{equation}
where $N_s$ is the total number of moir\'e cells in the system, $\textbf{k}_{a}, \textbf{k}^{\prime}_{a}, \textbf{q}_{a}$ are the atomic wavevectors of graphene, $\alpha$, $\alpha'$ denote the layer and sublattice indices, and $\sigma$ is the spin index. If one expands the wavevector around the Dirac points $\mathbf{K}$ and $\mathbf{K}'$, the Coulomb interaction can be further decomposed into the intravalley part and the intervalley part, with the former interaction strength being two orders of magnitudes greater than the latter for small twist angles $\theta\!\sim\!1^{\circ}$. Therefore, we only consider the intravalley Coulomb interactions in this work, which is expressed as
\begin{align}
H_{C}^{\textrm{intra}}=\frac{1}{2N_{s}}\sum_{\textbf{k}_{a}\textbf{k}_{a}^{\prime}\textbf{q}_{a}}\sum_{\mu\mu^{\prime},\sigma\sigma^{\prime},\alpha\alpha^{\prime}}\,&V(\textbf{q}_{a})\hat{c}^{\dagger}_{\textbf{k}_{a}+\textbf{q}_{a},\mu\alpha\sigma}\hat{c}^{\dagger}_{\textbf{k}_{a}^{\prime}-\textbf{q}_{a},\mu^{\prime}\alpha^{\prime}\sigma^{\prime}}\;\nonumber\\
&\times\hat{c}_{\textbf{k}_{a}^{\prime},\mu^{\prime}\alpha^{\prime}\sigma^{\prime}}\hat{c}_{\textbf{k}_{a},\mu\alpha\sigma}
\end{align}
where $\mu,\mu^{\prime}=\pm$ are the valley indices, and now the wavevectors $\textbf{k}$, $\textbf{k}^{\prime}$, and $\textbf{q}$ are expanded around the Dirac point of valley $\mu$ ($\mathbf{K}^{\mu}$), which can decomposed as $\textbf{k}_{a}=\textbf{k}+\textbf{Q}$, where $\textbf{k}$ is the moir\'e wavevector in moir\'e Brillouin zone, and $\textbf{Q}$ is the moir\'e lattice vector.
A single-gate screened Coulomb interaction $V(\textbf{q}_{a})=e^{2}(1-e^{-2\lvert \textbf{q}_{a}\rvert d_{s}})/(2\Omega_{M}\epsilon_{\rm{BN}}\lvert \textbf{q}_{a}\rvert)$ is adopted in this work, where $\Omega_{M}$ is the area of moir\'e primitive cell, $d_{s}\!=\!40\,$nm is the distance between ATMG layers and the metallic gate, $\epsilon_{\rm{BN}}\!\approx\!4$ is the dielectric constant of the BN substrate. The Coulomb interactions are further projected onto the double flat bands, and we perform self consistent Hartree-Fock calculations within the subspace of the double flat bands. Besides, the Coulomb interactions between electrons in flat bands can be further screened by remote-band particle-hole excitation, and such screening effects in our calculation are treated with the constrained random phase approximation(cRPA) \cite{zhang-tbghf-arxiv20}, where the cRPA dielectric constant $\epsilon(\textbf{q}+\textbf{Q})=\epsilon_{\rm{BN}}(\mathbbm{1}+\hat{\chi}^{0}(\textbf{q})\hat{V}(\textbf{q}))_{\textbf{Q},\textbf{Q}}$, where $\hat{\chi}^{0}(\textbf{q})$ is the zero-frequency bare susceptibility at moir\'e wavevector $\textbf{q}$, and $\hat{V}(\textbf{q})$ is the Coulomb interaction matrix defined in the space of reciprocal moir\'e lattice vector $\mathbf{Q}$, with $\hat{V}(\textbf{q})_{\mathbf{Q},\mathbf{Q}}=V(\textbf{q}+\mathbf{Q})$. We refer the readers to Supplementary Information (Appendix III-IV) for more details about the Hartree-Fock and cRPA formalism.
To start with, we calculate the ground states at different integer fillings of the double flat bands using the Hartree-Fock and cRPA methods described above. The filling factor is counted with respect to the CNP, i.e., the filling factor is defined as $\nu\!=\!n-8$ when $n$ out of the 16 flat bands (including valley and spin degeneracy) are filled. Then we calculate the expectation values of the order parameters of the Hartree-Fock ground states at each integer filling, and figure out the dominant ones which are presented in Table~\ref{tab:filling}, where $\mathbf{\tau}$, $\mathbf{s}$, and $\mathbf{\sigma}$ denote Pauli matrices defined in valley, spin, and sublattice space respectively. For example, the ground state at filling -3 is a gaped spin-valley polarized state. In order to depict the spontaneous $m_z$ symmetry breaking, we also calculate the vertical electric polarization at different filling factors. The vertical electric polarization per moir\'e supercell $p_z$ is defined as: $p_{z}\!=\!\sum^{5}_{l=1}\,(l-3)\,q_{l}d_{0}$, where $d_{0}\!=\!3.35\,\mbox{\normalfont\AA}$ is the interlayer distance of Bernal bilayer graphene, $q_{l}=e\left\langle \tau_{0}\textrm{s}_{0}\mathbb{L}_{l}\sigma_{0}\right\rangle$ is the layer resolved charge density, where $\mathbb{L}_l$ is the projection operator onto layer $l$, a $5\times 5$ matrix with the $l$th diagonal element identity and all other elements being zeros.
The unit of the electric polarization is $e\cdot\mbox{\normalfont\AA}$ per moir\'e supercell. We also evaluate the orbital magnetization and valley polarization. The valley polarization $\xi_z$ is defined as: $\xi_z\!=\!\sum_{l=1}^{5}\xi_z(l)\!=\!\sum_{l=1}^{5}\left\langle\tau_{z}s_{0}\mathbb{L}_l\sigma_{0}\right\rangle$, where $\langle\tau_z(l)\rangle$ is defined as the valley polarization projected onto layer $l$. A finite valley polarization would split the two valleys and would give rise to nonzero net orbital magnetization. In Table~\ref{tab:filling}, we present the calculated vertical electric polarization and valley polarization of the spontaneous symmetry-breaking states at different filling factors. We find that $m_z$ symmetry is spontaneously broken by Coulomb interactions at all integer fillings, which generate small but nonzero electric polarization.
\begin{table}[]\scriptsize
\centering
\caption{Gap, main order parameter, polarization and valley polarization of the Hartree-Fock ground states at different filling factor. The unit of polarization is $e\cdot\mbox{\normalfont\AA}$ per moir\'e supercell. The unit of valley polarization is charge per unit cell.}
\begin{tabular}{c|c|c|c|c}
\hline\hline
filling factor & gap(eV) & main order & polarization &valley \\
& & parameter & &polarization\\
\hline
-7 & 0.0119 & $s_{z}\tau_{0}\sigma_{0}$ & -0.0103& 0.398 \\
-6 & 0.0208& $s_{z}\tau_{0}\sigma_{0}$ & -0.0009 & 0.048\\
-5 & 0.0132& $s_{z}\tau_{0}\sigma_{0}$, $s_{0}\tau_{z}\sigma_{0}$ & -0.0110 & 0.624 \\
-4& / & $s_{z}\tau_{0}\sigma_{0}$ & 0.0197 & -1.179\\
-3 & 0.0148& $s_{z}\tau_{0,z}\sigma_{0}$ & -0.0112 & -0.406 \\
-2 &/& $s_{0}\tau_{z}\sigma_{0}$ & 0.0732 & -1.364 \\
-1 & 0.0073 & $s_{0}\tau_{z}\sigma_{0}$ & 0.0585 & -3.046 \\
0 & /& $s_{0}\tau_{x}\sigma_{y}$ & 0.1004 & 0.014 \\
1 & / & $s_{0}\tau_{z}\sigma_{0}$ & 0.0708 & -2.705 \\
2 &0.0095 & $s_{z}\tau_{0}\sigma_{0}$ & 0.0834 & -0.030 \\
3 & 0.0063 & $s_{0,z}\tau_{0,z}\sigma_{0}$ & 0.0713 & 1.000 \\
4 &0.0131 & $s_{0}\tau_{0}\sigma_{0}$ & 0.0834 & 0.000 \\
5 & / & $s_{0,z}\tau_{0,z}\sigma_{0}$ & 0.0771 & -0.988 \\
6 & / & $s_{0}\tau_{z}\sigma_{0}$ & 0.0741 & -1.963 \\
7 & / & $s_{0,z}\tau_{0,z}\sigma_{0}$ & 0.0264 & -0.967 \\
\hline\hline
\end{tabular}
\label{tab:filling}
\end{table}
\begin{figure}[b]
\begin{center}
\includegraphics[width=8.5cm]{fig3f.png}
\caption{
The band structure of $A-ABA-A$ for two valleys including intrasublattice moir\'e potential and the further neighbor interlayer hopping. (a)The non-interacting band structures with weight projected to the middlemost layer. (b) The Hartree-Fock band structure for two valleys. (c) The vertical polarization, and (d) the valley polarization and the corresponding orbital magnetization of the symmetry-breaking ground state at -3 filling as a function of vertical electrostatic potential drop $U_d$.
\label{fig:131}
\end{center}
\end{figure}
\subsection{Orbital magnetoelectric effect through intertwined orders}
To characterize the effect of vertical electric field, we calculate both the vertical electric polarization and valley polarization of the Hartree-Fock ground states for the $A-ABA-A$ system at filling factor -3 with increasing displacement fields. The displacement field ($D$) is introduced by applying a homogeneous vertical electrostatic potential difference $U_d$ between the topmost and bottommost laye
, i.e., $U_d=eD d_{0}\times 4/\epsilon_{\rm{BN}}$, where $\epsilon_{\rm{BN}}\approx 4$ is the dielectric constant of the BN substrate.
Our calculations indicate that the dominant order parameters of the ground states at filling -3 are unchanged under different $U_d$, i.e., the system always stays in the spin-valley polarized state with broken $m_z$ symmetry, suggesting that the no phase transition occurs at least for $0\!\leq\!U_d\!\leq\!0.02\,$eV. However, by virtue of the $m_z$ symmetry and the additional layer degrees of freedom, the valley polarization acquires nontrivial layer distributions as shown by the $\xi_z(l)$ ($l=1,...,5$) values in Table~\ref{tab:vz}. We see that $\xi_{z}(1)$ and $\xi_z(2)$ are approximately the same, while $\xi_{z}(4)$ and $\xi_z(5)$ are approximately the same, which are different from $\xi_z(3)$. As a result, the layer-resolved valley polarization can be approximately decomposed into three terms:
\begin{align}
\hat{\xi_z} \approx \xi_z^{s} \,\mathbbm{1}_{5\times 5}+
\xi_{z}^{a} \hat{P}_z + \xi_{z}^{a} \hat{Q}_z
\end{align}
where
\begin{align}
\hat{P}_{z}=\left(\begin{array}{ccccc}
-2& 0& 0 &0&0 \\
0&-1&0&0&0\\
0&0&0&0&0\\
0&0&0&1&0\\
0&0&0&0&2 \end{array}\right)\;,\;\; \hat{Q}_{z}=\left(\begin{array}{ccccc}
1& 0& 0 &0&0 \\
0&0&0&0&0\\
0&0&0&0&0\\
0&0&0&0&0\\
0&0&0&0&-1 \end{array}\right)
\end{align}
and $\hat{\xi}_z$ is a $5\times 5$ matrix with its $l$th diagonal element denoting the valley polarization contributed by layer $l$, i.e., $\hat{\xi}_{z,ll}=\xi_{z}(l)$.
$\xi_z^{s}$ and $\xi_z^{a}$ denote the layer-symmetric and layer anti-symmetric components of the valley polarization, with $\xi_z^{s}=\sum_{l=1}^{5}\xi_z(l)/5$, and $\xi_z^{a}=\sum_{l\neq 3} \rm{sgn}[(l-3)](\xi_z(l)-\xi_z^{s})/4$. In other words, $\xi_z^{s}=\langle \tau_z\otimes \mathbbm{1}_{5\times 5}\rangle/5$ is the layer average of valley polarization ($\mathbbm{1}_{5\otimes 5}$ denotes identity matrix in layer space), and $\xi_z^{a}=\langle \tau_z\otimes (\hat{P}_z+\hat{Q}_z) \rangle /4$, where $\langle \hat{O} \rangle$ denotes the expectation value of operator $\hat{O}$ evaluated with respect to the symmetry-breaking ground state. The layer-symmetric and layer anti-symmetric valley polarization for the ground states at filling $-3$ with different $U_d$ are presented in the last two rows of Table~\ref{tab:vz}.
We note that $\hat{P}_z$ is exactly the vertical electric polarization operator $\hat{p}_z\!=\!e d_0\hat{P}_z$, which couples linearly to external electric field; while the valley polarization operator is proportional to the orbital magnetization operator $\hat{M}_z\!=\!g_z\mu_B\tau_z$ which couples linearly to external magnetic field, where $g_z$ is introduced as an effective $g$ factor and $\mu_B$ is the Bohr magneton.
Given the above discussions, we introduce an effective mean-field Hamiltonian to describe how the symmetry-breaking state at filling -3 would respond to external electric and magnetic fields
\begin{equation}
\begin{split}
H(\textbf{k})=&H_{0}(\textbf{k})+ \Delta_{z0}\,\tau_{z}\otimes\mathbbm{1}_{5\times5}+\Delta_{zz}\,\tau_{z}\otimes{(\hat{P}_z + \hat{Q}_z)}\\
& -g_z\mu_B B_z\,\tau_{z}\otimes\mathbbm{1}_{5\times5} + \frac{1}{4}U_d \,\tau_{0}\otimes \hat{P}_{z}\;,
\end{split}
\label{eq:order}
\end{equation}
where $\Delta_{z0}$ and $\Delta_{zz}$ are the ``mean fields" that are self consistently generated by Coulomb interactions which are coupled with the $\tau_{z}\otimes\mathbbm{1}_{5\times5}$ operator and $\tau_{z}\otimes{(\hat{P}_z + \hat{Q}_z)}$ operator respectively, while $B_z$ is the vertical magnetic field and $U_d$ is the vertical electrostatic energy drop.
As the electric polarization operator and the valley polarization operator are intertwined together, Eq.~(\ref{eq:order}) implies a tunable electric polarization by magnetic field and conversely a tunable valley polarization (orbital magnetization) by electric field.
To be specific, a vertical electric field is coupled to the $\hat{P}_z$ operator, which is in turn intertwined with the layer anti-symmetric component of the valley polarization operator, thus would change the valley polarization and orbital magnetization of the system. Conversely, a vertical magnetic $B_z$ is coupled to orbital magnetization (valley polarization), and the valley polarization operator is intertwined with the electric polarization operator, which would change electric polarization of the system.
It is clearly seen from Table~\ref{tab:vz} that the layer symmetric valley polarization is larger than the layer anti-symmetric one, implying that the orbital magnetization still has the strongest coupling to magnetic field, but can tuned by electric field. In Fig.~\ref{fig:131}(c)and (d), we present the calculated electric polarization and valley polarization of the symmetry-breaking states at filling -3 under different $U_d$. As $U_d$ increases, clearly the electric polarization is linearly enhanced as shown in Fig.~\ref{fig:131}(c). On the other hand, the valley polarization and the corresponding orbital magnetization are also dramatically enhanced with the increase of $U_d$ as shown by the blue and red dots in Fig.~\ref{fig:131}(d).
This indicates a novel type of orbital magnetoelectric effect driven by Coulomb interactions in mirror-symmetric ATMG system with double flat bands.
\begin{table}[
\centering
\caption{Layer resolved valley polarization for the ground states at filling -3 under different $U_d$. $\xi_z^{s}=\sum_{l=1}^{5}\xi_z(l)/5$, and $\xi_z^{a}=\sum_{l\neq 3} \rm{sgn}[(l-3)](\xi_z(l)-\xi_z^{s})/4$.}
\begin{tabular}{lccccc}
\hline\hline
$U_d$ (eV) & 0 & 0.005 & 0.010 & 0.015 & 0.020\\
\hline
$\xi_z(1)$ & -0.0994 & -0.1690 &-0.2556 &-0.2724&-0.2792\\
\hline
$\xi_z(2)$ & -0.0996 & -0.1676 & -0.2547 &-0.2720&-0.2780\\
\hline
$\xi_z(3)$ & -0.0065 & 0.0023 & -0.0685 &-0.1317&-0.1469\\
\hline
$\xi_z(4)$ & -0.1003 & -0.1309 & -0.1683 &-0.1626&-0.1482\\
\hline
$\xi_z(5)$ & -0.1000 & -0.1309 & -0.1674 &-0.1617&-0.1477\\
\hline
$\xi^{s}_{z}$ & -0.0998 & -0.1496 &-0.2115 &-0.2172&-0.2133\\
\hline
$\xi^{a}_{z}$ & 0.0003 & 0.0187 & 0.0437 &0.0550&0.0653\\
\hline\hline
\end{tabular}
\label{tab:vz}
\end{table}
To summarize, in this work we have theoretically studied the electronic structures and interaction effects of alternating twisted multilayer graphene (ATMG) systems. We find that these ATMG systems exhibit various intriguing non-interacting band dispersion including one pair of flat bands, one pair of flat bands co-existing with Dirac cones or more generally $E(\mathbf{k})\sim k^{J}$($J$ is positive integer) dispersion, as well as two pairs of flat bands which may also co-exist with $E(\mathbf{k})\sim k^{J}$($J$ is positive integer)dispersion. Based on an analytic analysis from a simplified $\textbf{k}\cdot\textbf{p}$ model approach, we find that the low energy band structure can be described by a set of generic partition rules.
We have also considered Coulomb interaction effects in ATMG with $A-ABA-A$ stacking, the simplest mirror symmetric ATMG system having two pairs of flat bands. We have studied the symmetry-breaking ground states at different integer filling factors under zero external fields based on unrestricted Hartree-Fock calculations. We find that at certain fillings both time-reversal symmetry and the mirror symmetry can be broken spontaneously by Coulomb interactions, leading to insulator states with intertwined electric polarization and orbital magnetization. As a result of such intertwined ordering, the system can exhibit a novel type of orbital magnetoelectric effect with the orbital magnetization (electric polarization) being highly tunable by external electric (magnetic) field. Our work is a significant step forward in understanding the electronic structures and correlation effects of alternating twisted graphene systems, and will provide useful guidelines for future experimental and theoretical studies.
\acknowledgements
This work is supported by the National Key R \& D program of China (grant no. 2020YFA0309601), the National Science Foundation of China (grant no. 12174257), and the start-up grant of ShanghaiTech University.
|
\section{Introduction}
We have witnessed a large number of crisis situations in recent years, from natural disasters to man-made disasters and also to deadly animal and human health crises, culminating with the ongoing COVID-19 public health crisis.
Affected individuals often turn to social media (e.g., Twitter or Facebook) to report useful information,
or ask for help
\cite{sakaki2010earthquake,vieweg2010microblogging,king2018social}. Information contributed on social media by people on the ground can be invaluable to emergency response organizations in terms of gaining situational awareness, prioritizing resources to best assist the affected population, addressing concerns, and even saving lives \cite{king2018social}.
\begin{table*}[ht]
\scriptsize
\centering
\begin{tabular}{|c|l|}
\hline
\multicolumn{1}{|l|}{{\bf No.}} & {\bf Tweet text} \\ \hline
\multirow{2}{*}{1} &
Roads in {\colorbox{red!25!white}{Calhoun County}} are underwater, access to the {\colorbox{red!25!white}{Port Lavaca Causeway}} is flooded, the bridge is closed. \\ \cline{2-2}
&
{\color{white}{oo}}O {\color{white}o}
O{\color{white}} {\colorbox{blue!25!white}{B-ctc{\color{blue!25!white}OO} I-ctc {\color{blue!25!white}oi}}} {\color{white}o}O{\color{white}ooooo}O{\color{white}oooooo} O{\color{white}o}O{\color{white}o}O{\color{white}oi}{\colorbox{blue!25!white}{B-lan {\color{blue!25!white}i}I-lan {\color{blue!25!white}ooo} I-lan{\color{blue!25!white}o}}}{\color{white}{o}}O{\color{white}{oo}}O{\color{white}{oooo}}O{\color{white}{ooo}}O{\color{white}{oo}}O{\color{white}{oo}}O
\\
\hline
\multirow{2}{*}{2} &
Very extensive damage sustained throughout {\colorbox{red!25!white}{Wilmington}} {\colorbox{red!25!white}{NC {\color{red!25!white}o}}} from Hurricane Florence \\ \cline{2-2}
&
{\color{white}{oo}}O {\color{white}ooo}
O{\color{white}ooooo} O{\color{white}ooooo} O{\color{white}ooooooo}
O{\color{white}oooo}
{\colorbox{blue!25!white}{\hspace{0.3cm} B-ctc {\color{blue!25!white}oo}}} {\color{white}}{\colorbox{blue!25!white}{{\color{blue!25!white}}B-sta{\color{blue!25!white}}}}{\color{white}I} {\color{white}o}
O{\color{white}oooo}
O{\color{white}ooooo}
O{\color{white}ooooo}
\\ \hline
\multirow{2}{*}{3} &
Big tree fell on power lines and blocking {\colorbox{red!25!white}{Brown Ave}} near {\colorbox{red!25!white}{Washington St}} in {\colorbox{red!25!white}{Orlando}} s {\colorbox{red!25!white}{Thornton Park}} \\ \cline{2-2}
&
{\color{white}{o}}O {\color{white}o}
O{\color{white}o} O{\color{white}o} O{\color{white}oo}
O{\color{white}ooo}
O{\color{white}oo}
O{\color{white}ooo}
O{\color{white}ooo}
{\colorbox{blue!25!white}{B-oth I-oth{\color{blue!25!white}}}}{\color{white}o} O{\color{white}ii} {\color{white}}{\colorbox{blue!25!white}{{\color{blue!25!white}O}B-oth I-oth{\color{blue!25!white}O}}} O{\color{white}}{\color{white}i}{\colorbox{blue!25!white}{B-sta\hspace{0.3cm}}}{\color{white}i}O{\color{white}} {\colorbox{blue!25!white}{B-lan I-lan{\color{blue!25!white}ooi}}}
\\ \hline
\multirow{2}{*}{4} &There are now more confirmed cases of coronavirus in {\colorbox{red!25!white}{New {\color{red!25!white}i} York {\color{red!25!white}i} City}} than there are in all of {\colorbox{red!25!white}{South Korea}}
\\ \cline{2-2}
&
{\color{white}{oi}}O{\color{white}oo}
O{\color{white}ii} O{\color{white}ooo} O {\color{white}{oooo}}O {\color{white}{ooooo}}O {\color{white}{oi}}O {\color{white}{oooo}}O {\color{white}{oooii}}O {\colorbox{blue!25!white}{B-ctc I-ctc \hspace{.1cm} I-ctc }}{\color{white}oo}O{\color{white}ooo}O{\color{white}o} O{\color{white}o}O{\color{white}o}O {\color{white}i}O{\color{white}i}{\colorbox{blue!25!white}{B-con I-con}}
\\ \hline
\multirow{2}{*}{5} & {\colorbox{red!25!white}{South Asia \hspace{.1cm}}} is quickly marching towards being the new epicenter of covid 19 \\ \cline{2-2}
&
{\colorbox{blue!25!white}{B-reg {\color{blue!25!white}i}I-reg }}
{\color{white}{}}O {\color{white}oo}
O{\color{white}ooooo} O{\color{white}ooooo} O{\color{white}oooo} O{\color{white}oo}
O{\color{white}oo}
O{\color{white}oooo}
O{\color{white}ooo}
O{\color{white}o}
O{\color{white}oo}
O{\color{white}}
\\ \hline
\multirow{2}{*}{6} & The difference in COVID 19 cases and deaths between {\colorbox{red!25!white}{New York}} and {\colorbox{red!25!white}{California}} continues to be astounding
\\ \cline{2-2}
&
{\color{white}{o}}O{\color{white}oooo}
O{\color{white}ooo} O{\color{white}ooi} O {\color{white}{ooo}}O {\color{white}{oo}}O {\color{white}{oi}}O {\color{white}{ooi}}O {\color{white}{ooooi}}O
{\color{white}{oo}}
{\colorbox{blue!25!white}{B-sta I-sta}}{\color{white}o}O{\color{white}ii} {\colorbox{blue!25!white}{\hspace{.2cm}B-sta \hspace{.2cm}}}{\color{white}oooo}O{\color{white}ooo} O{\color{white}i}O{\color{white}iooo}O {\color{white}}
\\ \hline
\end{tabular}
\caption{Examples of crisis tweets tagged with fine-grained location types. The subsequences representing location mentions are highlighted with pink, and their corresponding tags (in BIO format) are highlighted with blue.}
\label{examples}
\end{table*}
Many recent studies have focused on identifying informative tweets posted by individuals affected by a crisis, and classifying those tweets according to situational awareness categories useful for crisis response and management \cite{imran2015processing}. However, for situational awareness information extracted from social media to be actionable, knowing the corresponding geographic location is of key importance. For example, location information enables responders to perform fast assessment of the damage produced by a natural disaster \cite{villegas2018lessons}, or to respond to requests for help coming from affected individuals or institutions (e.g., hospitals or schools). In the case of COVID-19 health crisis, location information can also be used to identify trends by locations (e.g., stance of a community towards various health recommendations) \cite{mutlu2020stance, miao-etal-2020-twitter}, and subsequently employ that information to prevent dissemination of misinformation and rumors, and resurgence of the novel coronavirus.
Unfortunately, only a very small percentage of tweets are geotagged \citep{ICWSM124605}. Furthermore, even when geolocation information is available, that location may not be the location mentioned in the tweet text \citep{ikawa2013location}. According to \citet{vieweg2010microblogging}, the location in the tweet text is usually the location needed for monitoring and/or responding to an emergency. Table \ref{examples} shows several examples of tweets posted during recent hurricanes (first three tweets) and during the COVID-19 crisis (last three tweets). As can be seen, locations are mentioned at different levels of granularity, from region and landmark to city, state and country. Furthermore, the same location name, in our COVID-19 examples - {\it New York}, can be associated with different location types, such as {\it city} (tweet 4) and {\it state} (tweet 6). Information about the tags of the ambiguous entities can be used to disambiguate the corresponding locations and link them to physical locations. Therefore, tools for identifying fine-grained locations directly from the texts of crisis tweets are greatly needed.
Location identification has been frequently addressed as part of the broader named entity recognition (NER) task \citep{goyal2018recent,li2020survey}. Some studies have focused specifically on the task of identifying generic location mentions (without considering the type of location) in tweet text \citep{hoang2017predicting}, and even disaster tweet text \citep{kumar2019location}.
Other studies have focused on identifying fine-grained points-of-interest (POI), useful for location-based services \citep{li2014fine,malmasi2015location,ji2016joint,xu2019dlocrl}.
To the best of our knowledge, there are no publicly available, manually annotated datasets that can facilitate progress on the task of identifying fine-grained locations (including, city, state, country, region, landmark) in crisis tweets, despite the benefits provided by the use of social media data in monitoring and responding to a crisis.
To address this need, we have assembled two datasets for identifying fine-grained locations in crisis tweets. The first dataset, called {\it MIXED}, consists of tweets crawled during five crisis events, specifically, Nepal Earthquake, Queensland Floods, Srilanka Bombing, Hurricane Michael and Hurricane Florence. The second dataset, called {\it COVID}, consists of a set of coronavirus-related tweets crawled between February 27 and April 7, 2020. We used Amazon Mechanical Turk (AMT)\footnote{\url{https://www.mturk.com/}} to annotate the datasets using six location types (Country, State, Region, City, Landmark and Others).
Given the success of deep learning approaches for NER tasks \citep{li2020survey}, we use different state-of-the-art models to establish baseline results on the dataset.
In summary, the contributions of this work are as follows:
\begin{comment}
we propose to use a sequence labeling deep learning approach to identify fine-grained locations in crisis tweets. More specifically, we use an architecture considered state-of-the-art for NER \citep{li2020survey}, which has three main components:
a {\it distributed representation} model for input, a {\it context encoder} and a {\it tag decoder}. We use a BERT model (short for Bidirectional Encoder Representations from Transformers) \citep{DBLP:journals/corr/abs-1810-04805} as a distributed representation of the input, a BiLSTM model (short for Bidirectional Long Short-Term Memory) \citep{hochreiter1997long} as a context encoder, and a CRF model (short for Conditional Random Fields) \citep{lafferty2001conditional} as a tag decoder.
We study the combined BERT-BiLSTM-CRF model in both in-domain and cross-domain settings to gain insights into its effectiveness, as well as its transferability to a new emerging crisis for which labeled data may not yet be available.
\end{comment}
\begin{comment}
\begin{figure}[tpb]
\centering
\includegraphics[width=20em]{diagram-10.png}
\caption{Overview of the the deep learning architecture for fine-grained location identification. }
\label{model}
\end{figure}
\end{comment}
\begin{itemize}
\item We create two datasets of tweets from a mixed set of crisis events and from COVID-19, respectively. The tweets are manually annotated with fine-grained location types, including city, state, country, region, landmark.
\item {We use state-of-the-art models including a contextual encoder coupled with a tag decoder in a multi-task learning setting, and a model based on contextualized word and entity representations, combined with entity-aware self-attention to establish baseline results for our datasets.}
\item We perform extensive experiments on the {\it MIXED} and {\it COVID} datasets, respectively, in both {\it in-domain} and {\it cross-domain settings} to understand the usefulness of the data from the domain of interest, as well as the transferability of the models from one domain to another.
\end{itemize}
Given this introduction, we proceed with a discussion of related work in the next section, followed by the description of the datasets constructed, and then background and approaches, experimental setup, results and error analysis, and finally, conclusions and an ethics statement.
\section{Related Work}
\label{related}
We organize the related work based on several categories relevant to the research in this paper. Specifically, we first briefly discuss the location mention identification as a specific task in the area of NER. Subsequently, we review works on fine-grained location types, followed by approaches used for identifying locations, and finally, other existing and relevant location datasets.
\subsection{NER and location mention identification} NER is a well-researched problem in natural language processing (NLP) \citep{goyal2018recent,li2020survey}.
Text-based location identification has been traditionally addressed as part of the broader NER task, although some works focus specifically on location identification \citep{lingad-location,han2014identifying,kumar2019location,magge2019bi}.
Most of the works that identify locations simply tag location mentions, as opposed to identifying fine-grained location types \citep{li2020survey}. For example,
\citet{lingad-location} aim to identify mentions of locations (including geographic locations and points of interest) in disaster tweets, by using standard NER taggers (pre-trained or retrained), and report best performance using retrained Stanford NER \citep{Finkel:2005:INI:1219840.1219885}.
Also in the context of emergencies, \citet{kumar2019location} use a convolutional neural network (CNN) approach to identify location references
in crisis tweets, regardless of their specific types.
\subsection{Fine-grained location types} Some recent works have considered fine-grained location types, such as city, state, country \citep{10.1007/978-3-319-18117-2_24,anand2017fine,lal2019sane,qazi2020geocov19}. While focused on COVID-19 tweets, \citet{qazi2020geocov19} use a gazetteer approach to infer the geolocation of tweets, based on user and tweet information. Closest to our goal of identifying fine-grained locations in disaster tweet texts, \citet{10.1007/978-3-319-18117-2_24} propose a CRF-based approach to identify countries, states/provinces and cities using a Twitter dataset annotated according to guidelines provided in \citep{10.2307/40666361}. They make use of hand-crafted features, including gazetteer features, to train a CRF model. As opposed to \citep{10.1007/978-3-319-18117-2_24}, we use a larger set of location types and approaches that preclude the need for manually crafted features and gazetteers.
Other works on fine-grained location focus on identifying point of interests locations, such as restaurants, hotels, parks, etc. and linking them to pre-defined location profiles \citep{li2014fine,ji2016joint,xu2019dlocrl}.
\citet{li2014fine} build a POI inventory (which can be seen as a noisy version of a gazetteer), and a time-aware POI tagger. The time-aware POI tagger is a CRF trained to extract and disambiguate fine-grained POIs.
\citet{ji2016joint} extend the POI tagger in \citet{li2014fine} by proposing a joint framework that achieves POI recognition and linking to pre-defined POI profiles simultaneously. \citet{xu2019dlocrl} address the same problem of identifying fine-grained POIs and linking them to location profiles. However, they use a deep learning model (specifically, BiLSTM-CRF) to avoid the need for manually designed features, and subsequently use a collection of location profiles to perform the linking. The definition of fine-grained POI tagging is different from our definition of fine-grained location tagging - we aim to assign specific types/tags to location entities, as opposed to identifying generic (yes/no) POI tags, and then linking the tags to pre-defined profiles, as in prior works \citep{li2014fine,ji2016joint,xu2019dlocrl}. Moreover, we want to avoid the use of gazetteers to ensure that the models are resilient to the informal nature of the language used in tweets. Similar to \citep{xu2019dlocrl}, we also want to avoid the need for manually designed features, and thus focus on deep learning approaches.
\subsection{State-of-the-art approaches for NER} State-of-the-art approaches for NER, in general, and location identification, in particular, are sequence labeling type approaches based on deep learning language models \citep{li2020survey}. More specifically, competitive architectures consist of three components: distributed representations of the input, a context encoder model, and a tag decoder mode
. Both character-level and word-level embeddings (or their combination) have been used to represent the NER input in recent works \cite{goyal2018recent}, with BERT \citep{DBLP:journals/corr/abs-1810-04805} contextual embeddings being among the most successful \cite{li2020survey}. In terms of context encoders and tag decoders, recurrent neural networks, most often, BiLSTM networks (short for Bidirectional Long Short-Term Memory) \citep{hochreiter1997long}, and CRF (short for Conditional Random Fields) \citep{lafferty2001conditional}, respectively, contribute to some of the best results on benchmark NER datasets \citep{luo2019hierarchical,baevski2019cloze,liu2019gcdt,jiang-etal-2019-improved}. Given these successful architectures for the NER task, one of our baseline models consists of three components: BERT, BiLSTM and CRF, for the input representation, context encoder and tag decoder, respectively. As another strong baseline, we investigate a recent state-of-the-art architecture, called LUKE, \citep{yamada-etal-2020-luke}, based on a bidirectional transformer architecture pre-trained to output both word and entity contextualized representations. LUKE uses an entity-aware self-attention to identify entities.
\subsection{Existing location datasets}
Most previous works on location identification in tweet texts are focused on general tweets \cite{liu2014automatic,10.1007/978-3-319-18117-2_24} with a few notable exceptions of works focused on crisis tweets \citep{lingad-location,kumar2019location,qazi2020geocov19}. However, the datasets used in these works are not all available \citep{lingad-location,kumar2019location}. Even when available, the datasets focus on identifying location mentions without specifically identifying the fine-grained type of the location mentions \citep{liu2014automatic}. \citet{qazi2020geocov19} used a gazetteer-only approach to annotate tweets with geolocations, and the resulting annotations are not very accurate. While not specifically focused on crisis tweets, the dataset published by \citet{10.1007/978-3-319-18117-2_24} is the closest to our dataset in terms of fine-grained location types used (which include city, country, state or province, etc.). However, most locations in their dataset are not mentioned in the tweet, but are inferred from auxiliary information. Specifically, only about 3\% of the tweet texts in their dataset have location entities, for a total of only 220 different location entities. Furthermore, they also used a gazetteer approach to annotate most of the tweets, and performed manual annotations just for a small subset of their dataset. Given the above-mentioned differences between existing datasets and our datasets, it is not possible to directly use the existing datasets to transfer information to our tasks in a cross-domain setting.
\section{Datasets}
\label{datasets}
One main contribution of our work is to construct two benchmark datasets for identifying fine-grained locations (see Table \ref{tab:category_distribution}) useful for crisis monitoring and response. The datasets cover events that are different in nature, to enable studies in both in-domain and cross-domain settings.
\subsection{Data collection} The first dataset, called {\it MIXED}, contains tweets posted during four natural disasters and one man-made disaster that happened in specific geographical regions. The second dataset, called {\it COVID}, contains tweets posted during the COVID-19 pandemic, and thus has worldwide coverage. More specifically, the tweets in the {\it MIXED} dataset were crawled during the following events: Nepal Earthquake, Queensland Floods, Srilanka Bombing, Hurricane Michael and Hurricane Florence. The tweets from Nepal Earthquake and Queensland Floods were obtained from \citep{alam-etal-2018-domain}. Tweets from Srilanka Bombing, Hurricane Michael and Hurricane Florence were crawled locally using the Twitter streaming API. A random sample of unique English tweets was included in the {\it MIXED} dataset that was annotated using AM. More than 133 million tweets from COVID-19 pandemic were also crawled locally between February 27th and April 7th, 2020. A random sample of unique English tweets was included in the {\it COVID} dataset for AMT annotation. The keywords used to crawl the tweets and the final number of tweets included in the dataset for each event are provided in the appendi
{\it Unlabaled tweets.} In addition to the {\it MIXED} and {\it COVID} datasets that are annotated as part of this work, we also used a large number of unlabeled mixed crisis and COVID-19 tweets to pre-trained BERT \cite{DBLP:journals/corr/abs-1810-04805} models and obtain crisis-specific embeddings. In particular, to pre-train the BERT model for the \textit{MIXED} dataset, we collected a larger set of tweets pertaining to various crisis events from prior works \citep{imran2016lrec,nguyen2017robust,crisismmd2018icwsm,alam-etal-2018-domain,olteanu2014crisislex,Olteanu:2015:EUH:2675133.2675242} in addition to the locally crawled tweets. For the \textit{COVID} dataset, however, we only used the locally crawled tweets to pre-train the BERT model.
\begin{table}[t]
\scriptsize
\centering
\begin{tabular}{|l|l|rr|rr|}
\hline
\multicolumn{1}{|l|}{\multirow{2}{*}{\textbf{Type}}} & \multirow{2}{*}{\textbf{Descr.}} & \multicolumn{2}{c|}{\textbf{MIXED Distr.}} & \multicolumn{2}{c|}{\textbf{COVID Distr.}} \\
\multicolumn{1}{|l|}{} & & \multicolumn{1}{r}{\textbf{\#}} & \multicolumn{1}{c|}{\textbf{\%}} & \multicolumn{1}{r}{\textbf{\#}} & \multicolumn{1}{c|}{\textbf{\%}} \\ \hline
con & Country & 1,763 & 28.31 & 1,819 & 53.06 \\ \hline
sta & State & 1,242 & 19.95 & 396 & 11.56 \\ \hline
reg & \begin{tabular}[c]{@{}l@{}}Region, \\ Continent\end{tabular} & 764 & 12.27 & 158 & 4.61 \\ \hline
ctc & \begin{tabular}[c]{@{}l@{}}City, Town,\\ County\end{tabular} & 797 & 12.80 & 518 & 15.11 \\ \hline
lan & \begin{tabular}[c]{@{}l@{}}Building,\\Landmark \end{tabular} & 1,190 & 19.11 & 391 & 11.42 \\ \hline
oth & \begin{tabular}[c]{@{}l@{}}Other \\ \end{tabular} & 471 & 7.56 & 146 & 4.24 \\
\hline
\multicolumn{1}{|l|}{All} & Entities & \multicolumn{1}{r}{6,227} & \multicolumn{1}{r|}{100.00} & \multicolumn{1}{r}{3,428} & 100.00 \\
\hline
\end{tabular}
\caption{Location types and their descriptions, together with type distribution (as raw numbers \# and percentages \%) in the {\it MIXED} and {\it COVID} datasets, respectively.}
\label{tab:category_distribution}
\end{table}
\subsection{Data annotation and quality assessment}
To prepare the tweets for annotation, the following pre-processing was performed. User mentions were anonymized by replacing them with a generic {\tt user} keyword, and links were removed from the tweet text. Special characters, including -=~!\#\$\% $\hat{ }$ \&*()+[]\{\};\textbackslash':"$\mid$$<$$>$?, and non-printable ASCII characters
were also removed. The tweet text was tokenized
to enforce annotation at the token level and avoid accidental annotation of token fragments. Tweet tokens were annotated with six location types using the BIO scheme (where B stands for Beginning, I stands for Inside and O stands for Outside of a location entity). The location types together with their brief descriptions are shown in Table \ref{tab:category_distribution}. Examples of annotated tweets are shown in Table \ref{examples}, where the first three tweets are representative of the {\it MIXED} dataset, and the last three are representative {\it COVID}.
We used feedback from a local annotator to iteratively develop and improve a custom annotation tool for our task. The tool was subsequently deployed to AMT.
Annotators were provided with definitions of the location types included in our study, together with precise instructions for annotation, and examples of annotated tweets, such as those in Table \ref{examples}.
Each tweet was annotated by at least 3 annotators. Only entities where two or more annotators agreed were included in the final datasets.
The Cohen's Kappa scores that we obtained for inter-annotator agreement were 0.63 and 0.62, and the average pairwise F1-scores for inter-annotator agreement were 68.87 and 65.86 for the {\it MIXED} and {\it COVID} datasets, respectively. According to \citet{doi:10.1177/001316446002000104}, these scores represent substantial agreement.
The distributions of the location entities over the six location types included in our study are shown in Table \ref{tab:category_distribution},
As can be seen, the annotated entities are more evenly distributed over the types considered in the {\it MIXED} dataset, while more than half of the entities are of type {\it country} in the {\it COVID} dataset. The datasets also show differences in terms of the number of entities per tweet, with the {\it MIXED} dataset containing a majority of tweets with one or two entities (and a small number of tweets with more than two entities), and {\it COVID} containing mostly tweets with one entity (and a small number of tweets with two or more entities). Such differences emphasize specific characteristics and challenges in the two domains, and are useful in studying the transferability of the models from one domain to another.
\begin{table}[t]
\scriptsize
\centering
\begin{tabular}{|l|l|r|r|r|r|}
\hline
\textbf{Dataset} & \textbf{No.} & \multicolumn{1}{l}{\textbf{Train}} & \multicolumn{1}{|l|}{\textbf{Test}} & \multicolumn{1}{l}{\textbf{Dev}} & \multicolumn{1}{|l|}{\textbf{Total}} \\ \hline
\multicolumn{1}{|c|}{} & Tweets & 2,620 & 820 & 656 & 4,096 \\
\multicolumn{1}{|c|}{} & Tokens & 73,622 & 23,253 & 18,511 & 115,386 \\
\multicolumn{1}{|c|}{} & Entities & 4,001 & 1,237 & 989 & 6,227\\
\cline{2-6}
\multicolumn{1}{|l|}{} & \multicolumn{5}{c|}{Entity type distribution} \\
\cline{2-6}
\multicolumn{1}{|c|}{} & con & 1,135 & 360 & 268 & 1,763\\
MIXED & sta & 752 & 267 & 223 & 1,242\\
\multicolumn{1}{|c|}{} & reg & 514 & 144 & 106 & 764\\
\multicolumn{1}{|c|}{} & ctc & 522 & 141 & 134 & 797\\
\multicolumn{1}{|c|}{} & lan & 768 & 238 & 184 & 1,190\\
\multicolumn{1}{|c|}{} & oth & 310 & 87 & 74 & 471\\
\hline
\hline
& Tweets & 3,355 & 1,049 & 839 & 5,243 \\
\multicolumn{1}{|c|}{} & Tokens & 103,646 & 32,674 & 25,798 & 162,118 \\
& Entities & 2,206 & 656 & 566 & 3,428 \\
\cline{2-6}
\multicolumn{1}{|l|}{} & \multicolumn{5}{c|}{Entity type distribution} \\
\cline{2-6}
\multicolumn{1}{|c|}{} & con & 1,162 & 347 & 310 & 1,819\\
COVID & sta & 264 & 75 & 57 & 396\\
\multicolumn{1}{|c|}{} & reg & 101 & 34 & 23 & 158\\
\multicolumn{1}{|c|}{} & ctc & 328 & 103 & 87 & 518\\
\multicolumn{1}{|c|}{} & lan & 265 & 62 & 64 & 391\\
\multicolumn{1}{|c|}{} & oth & 86 & 35 & 27 & 146\\
\hline
\end{tabular}
\caption{Statistics for the number of tweets, tokens and the number of location entities in the {\tt train/test/dev} subsets of the {\it MIXED} and {\it COVID} datasets, respectively. The entity type distribution in the {\tt train/test/dev} subsets is also shown for each dataset.}
\label{train-test-dev}
\end{table}
\subsection{Benchmark Datasets} To enable progress on fine-grained location identification in crisis tweets, and facilitate comparisons between models developed for this task (in-domain and cross-domain), we created benchmark datasets by randomly splitting our {\it MIXED} and {\it COVID} datasets into training ({\tt train}), development ({\tt dev}) and test ({\tt test}) subsets, respectively. We use the training subset to train our models, the development subset to select hyperparameters and the test subset to evaluate the final performance of the models. Statistics for the {\it MIXED} and {\it COVID} datasets in terms of number of tweets, tokens, entities in the {\tt train}, {\tt test} and {\tt dev} subsets, respectively, are shown in Table \ref{train-test-dev}.
The benchmark datasets, together with the pre-processing script, will be made publicly available upon publication of this work. More specifically, to comply with Twitter's Developer Agreement and Policy\footnote{https://developer.twitter.com/en/developer-terms/agreement-and-policy}, the datasets will be made available as pairs of tweet ID and corresponding locations. The locations will be specified as a list of location-type tags corresponding to the tokens in the tweet as shown in Table \ref{examples} (i.e., a list of tags such as B-ctc, I-ctc, B-sta, O, etc. - one tag for each tweet token). Given that the pre-processing script will also be made available, the index of the location tags should precisely match the index of the tweet tokens. In addition to the location annotated datasets of tweets, the IDs of the unlabeld tweets that are used to pre-trained BERT will also be made available for both mixes crises and COVID-19 health crisis.
\section{Background and Approaches}
\label{approaches}
The task of identifying fine-grained locations in tweet text can be formulated as follows: Given a set of $(X,Y)$ pairs, where $X=\{x_1,\cdots,x_n\}$ is a text sequence/tweet with $n$ tokens, and $Y=\{y_1,\cdots,y_n\}$ is a tag sequence with $n$ location tags/types (in BIO format) corresponding to the tokens in the sequence $X$; our sequence tagging task is to find a mapping $f_{\theta}:{\bf X}\rightarrow {\bf Y}$ (with parameters $\theta$) from input sequences to output sequences of fine-grained location types.
\begin{comment}
\subsection{Word and Character Embeddings.}
As our datasets consist of tweet texts, and Twitter data exhibits specific characteristics due to the informal nature of the platform,
we use GloVe embeddings \citep{pennington2014glove} with $100$ dimensions, pre-trained on a Twitter dataset with 2 billion tweets.
\citet{ma2016end} use a CNN based encoder to learn character-level embeddings, and subsequently concatenate them with word-level embeddings to generate the final distributed representation of the input for a sequence labeling task. Given the effectiveness of this approach, we also combine character-level embeddings with pre-trained GloVe embeddings to generate a distributed input representation.
\subsection{BERT Embeddings}
A generic BERT model, pre-trained on a large corpus, can be further {\it pre-trained} and/or {\it fine-tuned} for specific NLP tasks \cite{DBLP:journals/corr/abs-1810-04805}.
To customize BERT to crisis tweets, we start with a pre-trained model, {\tt bert-base-uncased}, and further pre-train it on a large collection of mixed tweets and COVID-19 tweets, respectively, using the masked language modeling (MLM) task. We use the Huggingface library \citep{Wolf2019HuggingFacesTS} to pre-train BERT.
The pre-trained BERT embeddings are provided as input to a BiLSTM, whose output is further provided as input to a CRF model, which identifies locations.
BERT uses WordPiece \citep{DBLP:journals/corr/WuSCLNMKCGMKSJL16} tokenization, which can break a single word into multiple pieces. This presents a challenge in our input representation, where the tokenized input has a different token collections than the tagged output.
To account for this, when an input token is broken down into pieces by the WordPiece tokenizer, we assign the original tag to the first piece of the token and label the remaining pieces \textcolor{blue}{with a special token \textit{X} masked for the final prediction.
This approach preserves the token-label mapping of the input sequence and aligns with the tokenization requirement of BERT at the same time.
\noindent
{\bf BiLSTM.} LSTM networks \cite{hochreiter1997long} are a type of recurrent neural networks (RNN) that can be used to capture dependencies in sequence data, including long term dependencies. At the core of an RNN network, including LSTM networks, there is a recurrent cell, which enables the network to pass information from one time step to the next one. A standard LSTM has one layer corresponding to one cell, and processes the sequence in the forward direction. A bidirectional LSTM (BiLSTM) includes a second layer/cell, which processes the sequence in reverse direction. We use a BiLSTM network as context encoder to capture sequence dependencies in both forward and reverse directions.
\noindent
{\bf Conditional Random Fields.} Conditional Random Fields (CRF) \citep{lafferty2001conditional} are commonly used as decoders in sequence labeling tasks. A CRF model takes as input a distributed representation ${\bf X}=\{{\bf x_1},\cdots,{\bf x_n}\}$ of the tokens in a sequence $X=\{x_1,\cdots,x_n\}$, and outputs a decoding ${\bf y}$ of the input representation ${\bf X}$. Context-sensitive representations produced by BiLSTM models are frequently used as input representations for CRF models. A CRF model is trained to maximize the log-probability of the correct sequence of tags as defined below (or equivalently to minimize the loss given by the negative log-probability):
\begin{equation}
\log (p({\bf y}|{\bf X})) = s({\bf X},{\bf y}) - \log ( \sum_{{\tilde y}\in {\bf Y_x}} e^{s({\bf X,{\tilde y}})})
\end{equation}
where $s({\bf X},{\bf y})$ is a tagging score function, and ${\bf Y_x}$ represents the set of all possible sequences of tags for the input sequence ${\bf X}$. During inference, a sequence ${\bf X}$ is decoded into the tag sequence ${\bf y}^*$, which maximizes the tagging score below:
\begin{equation}
{\bf y}^* = \arg\max_{{\tilde y}\in {\bf Y_x}} {s({\bf X,{\tilde y}})}
\end{equation}
We use a CRF as our tag decoder in the generic model architecture shown in Figure \ref{model}.
\end{comment}
\begin{comment}
\subsection{ Multi-Task Learning} Multi-task learning (MTL) \citep{caruana1997multitask} is focused on training multiple related tasks in parallel to leverage shared patterns and improve the performance of each individual task.
Generally, in the MTL framework, the related tasks share some lower layers that extract common features, while they also have task-specific layers. In this work, we use the task of tagging generic location mentions, as an auxiliary task for the main fine-grained location tagging task.
\end{comment}
\subsection{Baseline Models}
\subsubsection{Feature-Engineered Baseline.}
\textbf{Stanford NER} \citep{Finkel:2005:INI:1219840.1219885}
uses an arbitrary order linear chain CRF model over a set of predefined word and character level features extracted from the input. The model has been used as a strong baseline for many NER models. We retrain the model with both \textit{MIXED} and \textit{COVID} datasets, respectively, to learn fine-grained location types.
\subsubsection{Character and Word Embedding Baselines.}
One model architecture in this category consists of a distributed representation layer learning the embeddings at character and word level followed by an LSTM-based context-encoder layer and a CRF tag-decoder. The model is referred as \textbf{CNN-GloVe-BiLSTM-CRF} in what follows.
Considering the recent success of transformer-based models, we also experiment with a similar model where BERT is used as the embedding layer instead of CNN+GloVe. We call this model \textbf{BERT-BiLSTM-CRF}.
For both CNN-GloVe-BiLSTM-CRF and BERT-BiLSTM-CRF models, we employ a multitask learning approach \cite{caruana1997multitask}, in which the main task of fine-grained location tagging is learned simultaneously with the auxiliary task of a generic yes/no location tagging (see Appendix for more details). We refer this model using the \textit{-MTL} suffix in what follows.
\subsubsection{Word and Entity Embedding Baseline.}
{In addition to using contextualized word embeddings learned from a transformer-based language model, \textbf{LUKE} \citep{yamada-etal-2020-luke} also learns contextualized entity embeddings and subsequently uses an entity-aware self-attention mechanism to perform tasks such as entity typing, relation classification, NER, etc. The LUKE approach has achieved state-of-the-art results on standard NER datasets (among others). We fine-tune the pre-trained \textit{LUKE-base} model with the \textit{COVID} and \textit{MIXED} datasets, respectively. The LUKE model selects candidate entity spans before making the entity type category predictions, a task that is comparable to the auxiliary task in the MTL models discussed earlier. Hence, we do not use the multitask learning setting for LUKE.
\section{Experimental Setup}
\label{setup}
In this section, we discuss the metrics used in the evaluation, implementation details and experiments performed.
\subsection{Metrics} We use standard metrics, including precision (Pr), recall (Re) and F1-measure (F1), to evaluate the performance of the models trained.
\subsection{Implementation details}
We performed a grid-search with $5$ trials and used the development subsets to identify best-overall hyperparameter values (see the Appendix for details on the values included in the grid and best-overall values)
We used the best-overall values in the experiments. We used the Glorot uniform initializer \citep{pmlr-v9-glorot10a} to initialize the model weights. The optimization was performed using the AdamW optimizer \citep{loshchilov2017decoupled}, with a learning rate of $1e-3$, weight decay of $1e-2$, and gradient clipping with max norm of $5$. We used a dropout of $0.5$ and mini-batch size of $32$ in all the experiments.
We set a patience of $5$ epochs on the development F1-measure, as early stopping of training. All experiments are run on NVIDIA Tesla V100 GPU.\color{black}
\subsection{Experiments} We conducted experiments in two settings, {\it in-domain} and {\it cross-domain}. In the {\bf in-domain setting}, models were trained and tested on the same dataset (e.g., models were trained on MIXED-train, tuned on MIXED-dev, and tested on MIXED-test). The goal was to study: 1) the performance of the deep learning models by comparison with the traditional Stanford NER model;
2) the effect of the auxiliary task in the MTL framework; 3) the effect of different types of embeddings.
In the {\bf cross-domain setting}, we used the best in-domain model to investigate several ways to perform transfer of information between domains: 1) a {\it zero-shot} transfer setting, where models trained on one dataset were tested on the other dataset (e.g., models trained on MIXED-train, tuned on MIXED-dev and tested on COVID-test); 2) an {\it embedding-level} transfer, where the transformer block fine-tuned on one dataset (e.g., MIXED) was used as a starting point for the transformer block of the model trained/tuned/tested on the other dataset (e.g., COVID); 3) a {\it model-level} transfer, where the model trained/tuned on a dataset (e.g., MIXED-train, MIXED-dev) is used as the starting point of the model for the other dataset (e.g., COVID-train, COVID-dev, COVID-test, respectively).
\section{Results and Discussion}
\label{results}
We first present and discuss the in-domain results, followed by the cross-domain results. In addition, we also perform error analysis and discuss the robustness of the models.
\begin{table}[t]
\scriptsize
\centering
\begin{tabular}{|l|lll|}
\hline
\rowcolor{_blue!25!white}
\textbf{Dataset} & \multicolumn{3}{c|}{\textbf{MIXED}} \\ \hline
\textbf{Model} & \multicolumn{1}{l}{\textbf{Pr}} & \multicolumn{1}{l}{\textbf{Re}} & \multicolumn{1}{l|}{\textbf{F1}} \\ \hline
Stanford NLP (retrained) & \multicolumn{1}{l}{\textbf{82.52}} & \multicolumn{1}{l}{65.64} & \multicolumn{1}{l|}{73.12} \\
CNN-GloVe-BiLSTM-CRF & \multicolumn{1}{l}{80.26} & \multicolumn{1}{l}{65.20} & \multicolumn{1}{l|}{71.95} \\
CNN-GloVe-BiLSTM-CRF-MTL & \multicolumn{1}{l}{76.92} & \multicolumn{1}{l}{59.30} & \multicolumn{1}{l|}{66.97} \\
BERT-BiLSTM-CRF & \multicolumn{1}{l}{74.98} & \multicolumn{1}{l}{70.07} & \multicolumn{1}{l|}{74.52} \\
BERT-BiLSTM-CRF-MTL & \multicolumn{1}{l}{74.58} & \multicolumn{1}{l}{\textbf{74.75}} & \multicolumn{1}{l|}{74.67} \\
LUKE & 80.71 & 73.08 & \textbf{76.71} \\\hline
\rowcolor{_blue!25!white}
\textbf{Dataset} & \multicolumn{3}{c|}{\textbf{COVID}} \\ \hline
{\bf Model} & \multicolumn{1}{l}{\textbf{Pr}} & \multicolumn{1}{l}{\textbf{Re}} & \multicolumn{1}{l|}{\textbf{F1}} \\ \hline
Stanford NLP (retrained) & \textbf{85.71} & 57.62 & 68.92 \\
CNN-GloVe-BiLSTM-CRF & 77.27 & 68.81 & 71.66 \\
CNN-GloVe-BiLSTM-CRF-MTL & 78.64 & 52.74 & 63.14 \\
BERT-BiLSTM-CRF & 73.41 & 0.74 & 72.05 \\
BERT-BiLSTM-CRF-MTL & \multicolumn{1}{l}{77.06} & \multicolumn{1}{l}{69.43} & \multicolumn{1}{l|}{73.04} \\
LUKE & 78.49 & \textbf{71.12 }& \textbf{74.66} \\\hline
\end{tabular}
\caption{In-domain results. Comparison of the following models: Stanford NER, CNN-GloVe-BiLSTM-CRF/BERT-BiLSTM-CRF and their MTL variants, and LUKE.
\label{in-domain}
\end{table}
\subsection{In-domain Setting} Table \ref{in-domain} shows the in-domain results of the models. As can be seen in Table \ref{in-domain}, the entity-embedding based LUKE model is the best overall in terms of F1-measure for both {\it MIXED} and {\it COVID} datasets, with a relatively high recall compared to most of the other models. Specifically, the F1-measure is 76.71\% for the {\it MIXED} dataset and 74.66\% for the {\it COVID} dataset. While the Stanford NLP has the highest precision overall, we argue that in the context of disaster monitoring and response, recall is more important than precision, as the final results will be reviewed by humans before any action is taken. Comparing the results for the {\it MIXED} and {\it COVID} datasets, we can see that the models have slightly better performance on the {\it MIXED} dataset. While this dataset contains a variety of crisis events, the events are relatively localized to specific geographical regions, which may make it easier for the models to identify the locations. As opposed to that, the {\it COVID} dataset has a big variety of locations as it covers a global pandemic. Nevertheless, the F1 score of the LUKE model on {\it COVID} is 8.3\% higher than the score of the Stanford NLP model, which uses manually designed features for training. We can also observe that the contextualized word and/or entity embeddings obtained from transformer architectures are better than both the engineered features in Stanford NLP and the character/word-embeddings in the CNN-GloVE-BiLSTM-CRF models. Finally, when comparing the BERT-BiLSTM-CRF-MTL model (with auxiliary task) to its BERT-BiLSTM-CRF variant (without the auxiliary task), the results show that the auxiliary task can help improve the F1-measure, especially in the case of {\it COVID}. However, for CNN-GloVe-BiLSTM-CRF, the addition of the auxiliary task decreases the F1-measure. This suggests that the transformer allows for a richer transfer of knowledge between similar tasks as compared to the CNN/GloVe architectures
\begin{table}[t]
\scriptsize
\centering
\begin{tabular}{|l|l|lll|}
\hline
\rowcolor{_blue!25!white}
\multicolumn{2}{|c|}{\textbf{Datasets}} & \multicolumn{3}{l|}{\textbf{COVID$\rightarrow$MIXED}} \\
\hline
\multicolumn{1}{|l|}{\textbf{Model}} &
\multicolumn{1}{l|}{\textbf{Transfer style}} & \textbf{Pr} & \textbf{Re} & \textbf{F1} \\
\hline
\multirow{3}{*}{BERT-BiLSTM-CRF-MTL} &
\multicolumn{1}{l|}{zero-shot} & 77.05 & 54.80 & 64.05 \\
&\multicolumn{1}{l|}{embedding-level} & 76.40 & 71.20 & 73.71 \\
&\multicolumn{1}{l|}{model-level} & 79.86 & 71.05 & 75.20 \\ \hline
\multirow{3}{*}{LUKE} &
zero-shot & 50.65 & 47.41 & 48.98 \\
&embedding-level & 78.13 & \textbf{74.21} & 76.12 \\
&\multicolumn{1}{l|}{model-level} & \textbf{81.32} & 73.57 & \textbf{77.25} \\ \hline
\rowcolor{_blue!25!white}
\multicolumn{2}{|c|}{\textbf{Datasets}} & \multicolumn{3}{l|}{\textbf{MIXED$\rightarrow$COVID}} \\ \hline
\multicolumn{1}{|l|}{\textbf{Model}} &
\multicolumn{1}{l|}{\textbf{Transfer style}} & \textbf{Pr} & \textbf{Re} & \textbf{F1} \\ \hline
\multirow{3}{*}{BERT-BiLSTM-CRF-MTL} &
zero-shot & 36.23 & 45.85 & 40.41 \\
&embedding-level & 67.19 & 68.85 & 68.01 \\
&\multicolumn{1}{l|}{model-level} & 66.44 & 71.47 & 68.86 \\ \hline
\multirow{3}{*}{LUKE} &
zero-shot & \textbf{79.94} & 43.17 & 56.06 \\
&embedding-level & 76.95 & \textbf{73.78} & 75.33 \\
&\multicolumn{1}{l|}{model-level} & 78.08 & 73.32 & \textbf{75.63} \\ \hline
\end{tabular}
\caption{Cross-domain results. Comparison between three transfer styles: zero-shot, embedding-level and model-level.
\label{cross-domain}
\end{table}
\subsection{Cross-domain setting} Table \ref{cross-domain} shows the results of the BERT-BiLSTM-CRF-MTL and LUKE models (which give the best overall results in the in-domain setting) in the cross-domain setting. Specifically, we compare three transfer styles, {\it zero shot}, {\it embedding-level}, and {\it model-level}, when {\it COVID} is used as source and {\it MIXED} as target, and the other way around. As expected, the model-level transfer style gives the best results overall, while the zero-shot style gives the worst results overall.
Notably, in the case of the {\it COVID} to {\it MIXED} transfer, the model-level transfer improves the results of the in-domain LUKE model, from 76.71\% to 77.25\%. This is probably due to the diversity in the {\it COVID} dataset, which enables more accurate locations to be identified in the {\it MIXED} dataset. As opposed to that, the transfer from {\it MIXED} to {\it COVID} causes more specific locations to be identified, which improves the recall but negatively affects the precision (and the overall F1-measure).
\subsection{Error analysis}
We performed error analysis of the model-level transfer from Table \ref{cross-domain} for both BERT-BiLSTM-CRF-MTL and LUKE (specifically, model-level transfer from {\it COVID} to {\it MIXED} and from {\it MIXED} to {\it COVID}). The analysis is based on the framework proposed by \citet{ribeiro-etal-2020-beyond}, where a model is tested for a capability using three tests: minimum functionality test (MFT), invariance test (INV) and directional expectation test (DIR). We performed the tests on the model's capability to generalize the concept of a location entity. In our case, MFT is the model's performance on the original {\it MIXED} or {\it COVID} test set, respectively. For INV, the location entities in the original test set were replaced with other randomly selected location entities of the same type from the test set. Finally, for DIR, the original location entities were replaced with randomly selected location entities of different types from the test set.
The results of the analysis are shown in Table \ref{error_analysis}. The MFT score serves as a baseline for the other two tests. As can be seen, in both cases, the performance degrades when the locations are mixed up - tests INV and DIR as compared with the test MFT - suggesting that the model captures correlations between locations and their context.
However, the F1 score for INV is better than the F1 score for DIR, which shows that the model expects a particular type of location in a given context.
Table \ref{error_analysis_examples} shows sample predictions for different tests (MFT, INV, DIR). In the first example, for the MFT test, the model makes a correct prediction for a tweet where a location entity of type \textit{ctc} is followed by a location entity of type \textit{sta}, which is the general convention for specifying a {\it city, state} location. However, for the DIR test, when the entities are replaced with others in reverse order of the type as compared to the original tweet (i.e., {\it sta, ctc} instead of {\it ctc, sta}),
the model incorrectly, but not surprisingly, predicts \textit{sta} as \textit{ctc} and vice versa. In the second example, for the MFT test, the model correctly predicts Sri Lanka as a country (i.e., {\it con}). However, when \textit{Sri Lanka} is replaced with \textit{South Africa} in the case of the INV test, the model predicts it as \textit{reg}. This is probably because Africa as a continent is a location of type \textit{reg}, and also because cardinal directions are commonly associated with \textit{reg} locations. Hence, without any external knowledge about \textit{South Africa} as a country, \textit{reg} is the next best prediction.
\begin{table}[t]
\centering
\scriptsize
\begin{tabular}{|c|l|l|l|l|}
\hline
\rowcolor{_blue!25!white}
\multicolumn{5}{|c|}{\textbf{Model level COVID$\rightarrow$MIXED transfer}} \\
\hline
\textbf{Model}&\textbf{Test} & \textbf{Pr} & \textbf{Re} & \textbf{F1} \\ \hline
\multirow{3}{*}{BERT-BiLSTM-CRF-MTL} & MFT & \textbf{79.86} & \textbf{71.05} & \textbf{75.20}\\\cline{2-5}
& INV & 67.78 & 52.29 & 59.03 \\ \cline{2-5}
& DIR & 47.48 & 31.34 & 37.76 \\ \hline
\multirow{3}{*}{LUKE} & MFT & \textbf{81.32} & \textbf{73.57} & \textbf{77.25}\\\cline{2-5}
& INV & 70.52 & 50.71 & 58.99 \\ \cline{2-5}
& DIR & 56.87 & 34.41 & 42.88 \\ \hline
\rowcolor{_blue!25!white}
\multicolumn{5}{|c|}{\textbf{Model level MIXED$\rightarrow$COVID transfer}} \\
\hline
\multirow{3}{*}{BERT-BiLSTM-CRF-MTL} & MFT & \textbf{66.44} & \textbf{71.47} & \textbf{68.86}\\\cline{2-5}
& INV & 57.64 & 59.35 & 58.48 \\ \cline{2-5}
& DIR & 40.26 & 33.80 & 36.75 \\ \hline
\multirow{3}{*}{LUKE} & MFT & \textbf{78.08} & \textbf{73.32} & \textbf{75.63}\\\cline{2-5}
& INV & 69.86 & 53.43 & 60.55 \\ \cline{2-5}
& DIR & 49.47 & 29.28 & 36.79 \\ \hline
\end{tabular}
\caption{Error analysis tests (MFT, INV and DIR) for the capability of the model-level transfer approach to generalize the concept of a location entity.}
\label{error_analysis}
\end{table}
\section{Conclusions and Future Work}
\label{conclusions}
In this paper, we introduced two new crisis tweet datasets manually tagged with specific fine-grained location types. These are the first manually annotated datasets for fine-grained location identification in crisis tweet texts, and can foster research in this area of great importance for crisis monitoring and response. The two datasets are different in nature, with one of them focused on mixed natural and man-made crisis events, which are generally localized to specific regions, and the second one focused on the worldwide COVID-19 pandemic. The different nature of the two datasets enables studies on location identification for localized and global events, as well as studies on the transferability of information between localized and global events.
In addition to introducing these datasets, we reported baseline results for the fine-grain location identification task using state-of-the-art models based on different embedding styles. Our results suggest that the entity-embedding style of the LUKE model gives the best results. We also used MTL to incorporate an auxiliary task in one of the models and showed its effectiveness in transferring information between datasets. As part of future work, we plan to improve the results of the models by including other crisis-related tagging and classification tasks in the LUKE/MTL settings.
\begin{table}[t]
\scriptsize
\centering
\begin{tabular}{|c|l|l|}
\hline
\multicolumn{1}{|l|}{{\bf No.}} &{\bf Test} &{\bf Tweet text} \\ \hline
\multirow{5}{*}{1} & \multirow{3}{*}{MFT} & NEW Hurricane Florence has made landfall near
\\
& & \locationtag[ctc$\rightarrow$ctc]{Wrightsville Beach,} \locationtag[sta$\rightarrow$sta]{North Carolina}\\
& & at 7 15 AM EDT\\% 1115 UTC with estimated maximum winds of 90 mph\\
\cline{2-3}
&\multirow{2}{*}{DIR} & NEW Hurricane Florence has made landfall near \\
& & \locationtag[sta$\rightarrow$ctc]{WI,} \locationtag[ctc$\rightarrow$sta]{Panama} at 7 15 AM EDT\\%
\hline
\multirow{6}{*}{2} & \multirow{3}{*}{MFT} & On this Easter Sunday my thoughts are with \\
& & \locationtag[con$\rightarrow$con]{Sri Lanka} following the horrific attacks
\\
& & on worshippers there.\\
\cline{2-3}
& \multirow{3}{*}{INV} & On this Easter Sunday my thoughts are with \\
& & \locationtag[con$\rightarrow$reg]{South Africa} following the horrific attacks on
\\
& & worshippers there.\\
\hline
\end{tabular}
\caption{Examples of location predictions for different error analysis test settings (MFT, INV and DIR). The examples are from the {\it MIXED} datasets and the predictions are made using model-level transfer from {\it COVID} to {\it MIXED}. The locations are highlighted with pink. The labels and predictions for entities are shown as a subscript to the corresponding locations using the convention \textit{gold label$\rightarrow$model prediction}.}
\label{error_analysis_examples}
\end{table}
\section{Ethics and Impact Statement}
\label{ethics}
The dataset that we plan to share will not provide any personally identifiable information, as only the tweet IDs and human annotated location tags (i.e., tags such as B-ctc, I-sta, O, etc., but not specific locations) will be shared. Thus, our dataset complies with
Twitter's Developer Agreement and Policy\footnote{https://developer.twitter.com/en/developer-terms/agreement-and-policy} in terms of privacy.
Furthermore, in compliance with the Twitter's Developer Agreement and Policy, Section III.E, the location information is used only in conjunction with the tweet content, and, as allowed by Twitter, we {\tt "only use such location data and geographic information to identify the location tagged by the Twitter Content."} In terms of impact, the research enabled by this dataset has the potential to help officials and health organizations identify actionable information useful for fast response during a crisis situation, or facilitate the health organizations to aggregate information relevant to COVID-19 by locations (which in turn can be useful in preventing a serious resurgence of the novel coronavirus in a particular region). However, we want to emphasize that we do not use any of the information in Twitter content, in particular the location information, to infer any sensitive information about the user, and most importantly our models do not infer any information about users' health\footnote{https://developer.twitter.com/en/developer-terms/more-on-restricted-use-cases}. The models are simply trained to identify location tags in tweets (as explicitly allowed by Twitter) and nothing more. Also important, our pre-processing script removes any user mentions from the tweet content before feeding the tweets to the models for training.
|
\section{Introduction}
\label{sec:intro}
There is a growing heterogeneous group of transients with peak luminosities in the general range from luminosities of classical novae to typical luminosities of supernovae that last from days to years (e.g., \citealt{Mouldetal1990, Rau2007, Ofek2008, Masonetal2010, Kasliwal2011, Tylendaetal2011, Tylendaetal2013, Kasliwal2013, Tylendaetal2015, Blagorodnovaetal2017, Kaminskietal2018, Pastorelloetal2018, BoianGroh2019, Caietal2019, Jencsonetal2019, PastorelloMasonetal2019, Banerjeeetal2020, Stritzingeretal2020, Blagorodnovaetal2021, Kaminskietal2021CKVul, Pastorelloetal2021}). Some of these transients are powered by thermonuclear reactions while other are powered by gravitational energy. Gravitational energy includes either a mass transfer in a binary system, possibly with the launching of jets by the mass-accreting star (e.g., \citealt{Kashietal2010, SokerKashi2016TwoI, Kashi2018Galax, Soker2020ILOTjets}), and/or mass ejection in the equatorial plane (e.g., \citealt{Pejchaetal2017, HubovaPejcha2019}), or a merger process of two objects. The merger itself can be a process where one star destroys the other star to form an accretion disk, where the merger of the two stars forms a more massive star, or the merger can form a common envelope evolution where one star spirals-in inside the envelope of the larger star (e.g., \citealt{RetterMarom2003,Tylendaetal2011, Ivanovaetal2013a, Nandezetal2014, Kaminskietal2015, Pejchaetal2016a, Pejchaetal2016b, Soker2016GEEI, MacLeodetal2017, MacLeodetal2018, Michaelisetal2018, Segevetal2019, Howittetal2020, MacLeodLoeb2020, Schrderetal2020}). In some cases both CEE and mass transfer can take place, as in the common envelope jets supernova impostor scenario where a neutron star or a black hole spirals-in inside the envelope of a red supergiant star accretes mass and launches jets \citep{SokerGilkis2018, Gilkisetal2019, YalinewichMatzner2019, Schreieretal2021}.
There is no nomenclature in consensus for this heterogeneous transient group. We refer to all transients that are powered by gravitational energy as Intermediate Luminosity Optical Transients (ILOTs; \citealt{Berger2009, KashiSoker2016, MuthukrishnaetalM2019}), but note that some do not use this term, e.g., \cite{Jencsonetal2019}. \cite{KashiSoker2016}\footnote{See \url{http://physics.technion.ac.il/~ILOT/} for an updated list.} use one set of sub-classes while \cite{PastorelloMasonetal2019} and \cite{PastorelloFraser2019} use another set.
In this paper we concentrate on the formation of a CEE where a planet enters the envelope of a low mass star that is just evolving off the main sequence, i.e., mainly during the Hertzsprung gap.
There are tens of papers on evolved stars, in particular on their giant branches, that swallow planets (e.g., \citealt{Soker1996, SiessLivio1999AGB, Massarotti2008, Carlbergetal2009, VillaverLivio2009, MustillVillaver2012, Nordhausetal2010, NordhausSpiegel2013, GarciaSeguraetal2014, Staffetal2016, AguileraGomezetal2016, Geieretal2016, Guoetal2016, Priviteraetal2016, Veras2016, Raoetal2018, SabachSoker2018a, SabachSoker2018b, Guidarellietal2019, Schaffenrothetal2019, Hegazietal2020, Jimenezetal2020, Krameretal2020, Chamandyetal2021, Guidarellietal2021, Lagosetal2021, Merlovetal2021}). In studying specific observed exoplanetary systems we follow earlier such studies (e.g., \citealt{NordhausSpiegel2013, SabachSoker2018a, Hegazietal2020}), although these concentrated on the giant branches. We focus on engulfment events at earlier evolutionary phases that lead to ILOTs.
There are several earlier studies of ILOTs powered by star-planet interaction (for a recent list see \citealt{Kashietal2019Galax}). \cite{RetterMarom2003} and \cite{Retteretal2006} suggested that planets that entered the envelope of a young star powered the ILOT V838~Mon. They suggested this star-planets interaction scenario as an alternative to the stellar merger scenario of V838~Mon \citep{SokerTylenda2003, Tylendaetal2005, Kaminskietal2021V838Mon}.
\cite{Bearetal2011} discuss ILOTs that result from the destruction of a planet
by a brown dwarf. These ILOTS will be of only several days duration and their luminosity might be even below those of classical novae. We here also deal with such faint ILOTs. \cite{KashiSoker2017planet} proposed that
the unusual outburst of the young stellar object ASASSN-15qi was a result of a young stellar object that tidally destroyed a sub-Jupiter young
planet. The gas from the destroyed planet formed an accretion disc around the young stellar object, and the accretion process powered this outburst.
\cite{Kashietal2019Galax} suggested that the $\approx 800$ days long eruption of the young stellar object ASASSN-13db was powered by a young stellar object that accreted the remains of a planet that it shredded by tidal forces.
We here use the evolutionary code \textsc{mesa}-binary (section \ref{sec:Numeical}) to examine the engulfment of observed exoplanets by their parent stars as the stars just leave the main sequence and have radii of several solar radii (section \ref{sec:Evolution}), so that the outbursts might be large enough to be observed (section \ref{sec:ILOTs}).
We summarise our results in section \ref{sec:Summary}.
\section{Numerical scheme}
\label{sec:Numeical}
We follow the evolution of five exoplanetary systems using the evolutionary stellar code Modules for Experiments in Stellar Astrophysics (\textsc{mesa}; \citealt{Paxtonetal2011, Paxtonetal2013, Paxtonetal2015, Paxtonetal2018, Paxtonetal2019}), version 10398 in its binary mode.
We follow the example of \textsc{mesa}-binary plus point until a CEE is formed, i.e., until $R_1>a(1-e)$, where $R_1$ is the stellar radius, $a$ is the semi-major axis of the orbit and $e$ is the eccentricity.
We set the parameters of the binary (in inlist~project) \textit{do\_tidal\_circ} and \textit{do\_tidal\_sync} to be true using the default prescriptions for both. Similar to our earlier study in \cite{Rapoportetal2021}, we take the stellar mass-loss rate on the red giant branch from Reimers \citep{Bascheketal1975} with the mass loss parameter $\eta=0.12$. We set the initial stellar equatorial velocity to be similar to that of the sun, between $ 1-2 {~\rm km} {~\rm s}^{-1}$. To chose a number, we simply set the initial rotation period to be equal to the planet's orbital period (however, this is just an initial condition, and does not result from tidal synchronisation at this early time). We also simulated one case with a zero equatorial rotation velocity and basically we found the same results.
\section{Evolution to CEE}
\label{sec:Evolution}
We examine the evolution of five exoplanets that we found in the Extrasolar Planets Encyclopaedia; (exoplanet.eu; \citealt{Schneideretal2011}). We list these systems in Table \ref{tab:Table1}. We study in details only the four systems where the star has a mass of $M_{\rm i} > 0.85 M_\odot$, as only such similar exoplanet systems might lead to observe ILOTs.
\begin{table*}[]
\centering
\begin{tabular}{|c|c|c|c|c|c|l|c|c|c|c|c|c|c|}
\hline
System & $M_*$ & $M_{\rm p}$ & $a_{\rm i}$ & $e_{\rm i}$ & $R_{\rm i}$ & Ref. & $t_{\rm f}$ & $e_{\rm f}$ & $R_{\rm f}$ & $L_{\rm f}$ & $A$ & $B$ & $t_0$ \\ \hline
\textbf{} & $M_\odot$ & $M_{\rm J}$ & $R_\odot$ & & $R_\odot$ & & $10^9{~\rm yr}$ & & $R_\odot$ & $L_\odot$ & day & days & day\\ \hline
V 1298 Tau b & 1.1 & 2.2 & 36.98 & 0.29 & 0.98 & D2019 & 8.43 & 0.23 & 6.78 & 19.65 & 4.23 & 13.38 & 12,170 \\ \hline
TOI-216 c & 0.77 & 0.56 & 34.55 & 0 & 0.7 & D21/N21 & 31.76 & 0 & 5.53 & 12.05 & & & \\ \hline
NGTS-11 b & 0.86 & 0.37 & 35.46 & 0.11 & 0.77 & G2020 & 21.15 & 0.11 & 4.1 & 7.23 & 1.736 & 10.82 & 12,340 \\ \hline
Kepler-145 c & 1.28 & 0.25 & 54.32 & 0.11 & 1.25 &C16/V15& 4.71 & 0.11 & 5.42 & 14.3 & 2.144 & 8.767 & 12,350 \\ \hline
Kepler-89 e & 1.28 & 0.11 & 42.88 & 0.02 & 1.25 & W2013 & 4.71 & 0.02 & 5.26 & 13.78 & 1.788 & 19.88 & 14,820 \\ \hline
\end{tabular}
\caption{The five systems that we examine in this paper, from which we study in details four systems. From left to right we list the exoplanet name, the mass of the star, the mass of the planet, the semi-major axis, and the eccentricity. These are our initial conditions for the simulations. In the sixth column we list the observed stellar radius and in the seventh column we give the reference to these observational data. We then list the final time of our simulation, i.e., the onset time of the CEE relative to the zero age main sequence, the final eccentricity of the orbit, the final radius of the star, and the final luminosity of the star. In the last three columns we list the parameters that we fit to equation (\ref{eq:Orbit}).
References from top to bottom are as follow.
D2019: \cite{Davidetal2019};
D21/N21: \cite{Dawsonetal2021} and \cite{Nesvornyetal2021};
G2020: \cite{Gilletal2020};
W2013: \cite{Weissetal2013};
C16/V15: \cite{Campanteetal2016} and \cite{VanEylenetal2015}.
}
\label{tab:Table1}
\end{table*}
The final evolution time of our simulation $t_f$ is at the onset of the CEE, namely, when the periastron distance just becomes smaller than stellar radius, i.e., $a_{\rm f} (1-e_{\rm f}) \le R_{\rm f}$. In Fig. \ref{fig:periastron_radius_e_age} we present the periastron distances and the stellar radii as function of time from the main sequence to the CEE for the four systems we study here.
The evolution is typical for stars that engulf exoplanet (e.g., \citealt{Hegazietal2020}), i.e., a slow change with a final rapid drop towards engulfment and the formation of a CEE.
\begin{figure*
\hskip -1.00 cm
\includegraphics[trim= 0cm 4.7cm 0.0cm 4.9cm,clip=true,width=0.9\textwidth]{periastron_radius_e_age.pdf}
\vskip +0.00 cm
\caption{The periastron distance $(1-e)a$ (blue line), the eccentricity $e$ (black line scale on the right) and the stellar radius $R$ (red line), as function of time for the four exoplanetary systems that we study here. }
\label{fig:periastron_radius_e_age}
\end{figure*}
To facilitate comparison with future observations of faint ILOTs we present in Fig. \ref{fig:P(t)} the evolution of the orbital period in the last $10,000 {~\rm days}$ before CEE onset.
\begin{figure
\includegraphics[trim= 7.0cm 4.5cm -0cm 5.0cm,clip=true,width=0.85\textwidth]{T_t_new.pdf}
\caption{The orbital period as function of time in the last $10,000 {~\rm days}$ before CEE onset for four systems as indicated. Open circles are the results from the \textsc{mesa}-binary simulations while the solid lines are the fittings according to equation (\ref{eq:Orbit}) with the parameters in the last three columns of Table \ref{tab:Table1}. }
\label{fig:P(t)}
\end{figure}
\cite{Tylendaetal2011} fit a formula for the decrease of the orbital period with time in the last thousands of days before merger of V1309~Sco.
We fit the same form of the formula but with different parameters. The formula for the orbital period reads
\begin{equation}
P=A \exp \left( \frac{B}{t-t_0} \right).
\label{eq:Orbit}
\end{equation}
We list the value of the parameters in the last three column of Table \ref{tab:Table1}. We take $t=0$ in this fitting to be at $10,000 {~\rm days}$ before the onset of the CEE (merger), i.e., merger takes place at $t=10,000 {~\rm days}$ in this formula.
We note that the value of $t_0$ is larger than the time of merger in this fitting ($10,000 {~\rm days}$). This is as the fitting of \cite{Tylendaetal2011} to V1309~Sco where $t_0$ is about 700 days after the start of the main peak in the light curve.
The parameters \cite{Tylendaetal2011} fit to equation (\ref{eq:Orbit}) for V1309~Sco are $(A,B,t_0)=(1.4456, 15.29, 10,700){~\rm days}$.
We find that the final orbital shrinkage towards CEE of the cases we study here is qualitatively similar to the observations of V1309~Sco.
The exoplanets we study here have too low masses to eject much of the envelope of their parent stars. Therefore, after the planet enters a CEE, the star is likely to destroy the planet before the planet can reach the core. In principle, the planet destruction can be by tidal interaction with the center of the star or by evaporation. In Fig. \ref{fig:logRho_profile} we present the temperature and density profiles of the stars at the moment they engulf the planet. As the planet average density is $\simeq 1 {~\rm g} {~\rm cm}^{-3}$, it will not be tidally destroyed before it is deep inside the star, $r \ll 1 R_\odot$, where the temperature is $\ga 3 \times 10^6 {~\rm K}$. This is more than an order of magnitude above the virial temperature of the planet. The condition for evaporation, which is for the stellar envelope sound speed to be larger than the escape velocity from the planet, is fulfilled at larger radii than the radii where the tidal destruction condition holds (e.g., \citealt{NelemansTauris1998, Soker1998, Krameretal2020}). We therefore expect the planet to be evaporated inside the stellar envelope.
The planet will not influence much the later evolution of the star, e.g., the star will continue very close to its original evolutionary track.
\begin{figure
\includegraphics[trim= 3.0cm 7.5cm 3.5cm 8.0cm,clip=true,width=0.5\textwidth]{logRho-profile.pdf} \\
\includegraphics[trim= 3.0cm 7.5cm 3.5cm 8.0cm,clip=true,width=0.5\textwidth]{logT-profile.pdf}
\caption{Upper panel: Stellar density (in ${~\rm g} {~\rm cm}^{-3}$) versus radius at the onset of the CEE for four systems as we list in the inset. Lower panel: The temperature (in K) profiles at the same time. }
\label{fig:logRho_profile}
\end{figure}
The remnant of the merger in the case of V1309~Sco is a “blue straggler” star (e.g., \citealt{Ferreiraetal2019}). The planets we study here will have a much smaller effect years after merger, and the star will return to about its pre-merger state.
\section{ILOT events}
\label{sec:ILOTs}
We here crudely estimate the expected ILOT luminosity and time scale. For that we compare the systems we study to V1309~Sco.
The basic process in our systems is of a planet that enters a CEE and by that ejects mass from the stellar envelope and energise an ILOT. In V1309~Mon it was a low-mass main sequence star that entered the envelope of a more massive star.
From Fig. \ref{fig:P(t)} we see that the systems we study have final orbital periods of $P_{\rm f} \simeq 1.7-4.2 {~\rm days}$ compared with $P=1.42 {~\rm days}$ in the case of V1309~Sco \citep{Tylendaetal2011}. \cite{Tylendaetal2011} further estimate that the V1309~Sco binary luminosity was $3-8.6 L_\odot$ and that orbital separation at the merger was of $5.4 - 7.7 R_\odot$.
Overall, the orbital and stellar properties of the systems we study here are similar to those of V1309~Sco.
The main difference between the systems we study and V1309~Sco is that in the systems we study the companions are planets rather than a main sequence star.
\cite{Nandezetal2014} take the stellar masses of the binary progenitor system of V1309~Sco to be $1.52 M_\odot$ and $0.16 M_\odot$, but these are highly uncertain.
From their simulations \cite{Nandezetal2014} estimate the ejecta mass as $M_{\rm ej} \simeq 0.04-0.05 M_\odot$ (observations give only a lower limit of about $0.001 M_\odot$, e.g., \citealt{Kaminskietal2018}), its energy as $E_{\rm ej} \simeq 1.5 \times 10^{46} {~\rm erg}$, and that the ejecta takes up to one third of the initial orbital energy. They also find from their simulations that the typical ejecta velocity at large distances is about half the escape velocity from the merging system. These velocities are $\simeq 200 {~\rm km} {~\rm s}^{-1}$ as observed \citep{Kaminskietal2015}.
The total radiated energy, accounting for bolometric luminosity and not only visible (e.g.,\citealt{TylendaKaminski2016}), is $\approx 10^{45} {~\rm erg}$. Namely, only $\approx 5-10\%$ of the outburst energy of V1309~Sco was carried in radiation.
In the case of planets this fraction might in some cases be larger because of much lower ejecta mass, which it turn implies that the ejecta is more transparent.
Consider then a simple scenario by which a planet of mass $M_{\rm p}$ ejects a mass of
\begin{equation}
M_{\rm ej} \approx 0.05 \left( \frac{M_{\rm p}}{0.16 M_\odot} \right) M_\odot =
3 \times 10^{-4} \left( \frac{M_{\rm p}}{M_{\rm J}} \right) M_\odot,
\label{eq:Mej}
\end{equation}
where we scale according to the estimate of \cite{Nandezetal2014} for the ejected mass in V1309~Sco. The ejecta velocity is about the same (or might be somewhat lower), namely, about a half of the escape velocity. Although the energy of the ejecta is much smaller for planets, because of this lower mass the ejecta is less opaque and radiation diffuses out through the ejecta early on.
Below we assume a spherical mass ejecta. This is a strong assumption since the onset of the CEE ejects mass in a highly-non-spherical geometry (e.g., \citealt{Nandezetal2014}). However, when we use this crude estimate together with the scaling to the observations and simulations of V1309~Sco, we suggest that the results do give an order of magnitude correct estimate of the luminosity and timescale of the event.
Consider an ejecta expanding with a velocity of $v_{\rm s}$ in a shell of width $\Delta r_{\rm s}$ at radius $r_{\rm s}$ and with a mass of $M_{\rm ej}$. The ejecta starts hot and cools via radiation and adiabatic cooling. The fraction of thermal energy that the radiation carries out of the initial thermal energy is $f \approx t^{-1}_{\rm diff}/(t^{-1}_{\rm diff} + t^{-1}_{\rm exp})$, where $t_{\rm exp} = r_{\rm s} / v_{\rm s}$ is the expansion time of the shell and where the radiation diffusion time from the shell is
\begin{eqnarray}
\begin{aligned}
t_{\rm diff} & \simeq \frac{3 \tau \Delta r_{\rm s}}{c}
\simeq 55
\left( \frac{M_{\rm ej}}{0.05 M_\odot} \right)
\left( \frac{\kappa}{1 {~\rm cm}^2 {~\rm g}^{-1}} \right)
\\ & \times
\left( \frac{r_{\rm s}}{5 \times 10^{13} {~\rm cm}} \right)^{-1}
\left( \frac{\Delta r_{\rm s}}{0.3 r_{\rm s}} \right) {~\rm days},
\end{aligned}
\label{eq:tdiff}
\end{eqnarray}
and where $c$ is the light speed, $\kappa$ is the opacity, and $\tau=\rho_{\rm s} \kappa \Delta r_{\rm s}$ is the shell optical depth.
During this time the gas reach a distance of $r_{\rm s}(55 {~\rm days})= 9.5 \times 10^{13} (v_{\rm s}/200 {~\rm km} {~\rm s}^{-1})^{-1} {~\rm cm}$. Namely, the average radius during this time was $\simeq 5 \times 10^{13} {~\rm cm}$, as the scaling in equation (\ref{eq:tdiff}).
This might crudely account for the decline time after the peak luminosity of V1309~Sco.
The actual situation is much more complicated as collision of different ejecta parts can transfer kinetic energy to more thermal energy hence more radiation (e.g., \citealt{Pejchaetal2016a, SokerKashi2016}) and a stellar companion might accrete mass and launch jets to power ILOTs (e.g., \citealt{SokerKashi2016, Soker2020ILOTjets}). However, for the purpose of the present crude estimate we proceed with our assumptions.
In the case of of a planet, the mass is two to three orders of magnitude lower (equation \ref{eq:Mej}), and the equality of photon diffusion time and expansion time is at a radius about ten times smaller, namely, $r_{\rm s} \approx 5 \times 10^{12} {~\rm cm}$.
The expanding shell suffers less of adiabatic cooling and a larger fraction of the thermal energy is radiated away. As well, the time scale is ten times shorter for the same expansion velocity. It might be that for planet companions the ejecta velocity and thermal energy are somewhat lower than the simple scaling. However, in that case the shell will suffer less adiabatic cooling as well.
We take then the ILOT timescale to be the diffusion time when they are both about equal to the expansion time
$\tau_{\rm ILOT} = t_{\rm diff} = t_{\rm exp}$.
Using equations (\ref{eq:Mej}) and (\ref{eq:tdiff}) and $t_{\rm exp} = r_{\rm s} / v_{\rm s}$ we find
\begin{eqnarray}
\begin{aligned}
& \tau_{\rm ILOT} \approx 3
\left( \frac{M_{\rm p}}{M_{\rm J}} \right)^{1/2}
\left( \frac{\kappa}{1 {~\rm cm}^2 {~\rm g}^{-1}} \right)^{1/2}
\\ & \times
\left( \frac{\Delta r_{\rm s}}{0.3 r_{\rm s}} \right)^{1/2}
\left( \frac{v_{\rm s}}{200 {~\rm km} {~\rm s}^{-1}} \right)^{-1/2}
{~\rm days}.
\end{aligned}
\label{eq:ILOTtau}
\end{eqnarray}
We assume that the total radiated energy is proportional to the companion mass and we scale by the above mentioned quantities for V1309~Sco. Namely, the total radiated energy is $E_{\rm rad} \approx 10^{45} (M_{\rm p}/0.16 M_\odot) {~\rm erg} =6 \times 10^{42} (M_{\rm p}/M_{\rm J}) {~\rm erg} $. Dividing by the ILOT timescale in equation (\ref{eq:ILOTtau}) we find the typical luminosity of the ILOT to be
\begin{equation}
L_{\rm ILOT}
\approx 6000 \left( \frac{M_{\rm p}}{M_{\rm J}} \right)^{1/2} L_\odot .
\label{eq:ILOTL}
\end{equation}
The total radiated energy is a small fraction of the orbital energy at merger $E_{\rm orb} \simeq 3 \times 10^{44} ({M_{\rm p}}/{M_{\rm J}}) {~\rm erg}$.
These parameters place the ILOT below the luminosity and energy of classical novae.
If the unusual outburst of the young stellar object ASASSN-15qi was indeed an ILOT, as \cite{KashiSoker2017planet} suggest, then this zone below novae is already populated.
The stellar binary interaction most likely forms a bipolar nebula from the ejecta, as observed for example in the cases of the ILOTs V4332~Sgr and V1309~Sco (e.g., \citealt{Kaminskietal2018}) and in CK~Vul (Nova~1670; \citealt{Kaminskietal2021CKVul}). This might be due to jets that the companion launches as it accretes mass (e.g., \citealt{SokerKashi2016, Soker2020ILOTjets}). We do not expect the planets in the systems we study to form such bipolar structures.
\section{Summary}
\label{sec:Summary}
We followed the evolution of four observed exoplanets to the time that their parent stars engulf them (Fig.\ref{fig:periastron_radius_e_age}). This takes place just after the stars have left the main sequence, i.e., on or near the Hertzsprung gap. Our simulations with \textsc{mesa}-binary show that the final spiralling-in towards forming a CEE is similar to what was observed in the ILOT V1309~Sco (Fig. \ref{fig:P(t)}).
The main difference between the observed ILOT V1309~Sco and the four systems we study here is that the typical planet mass in the systems we study is about two to three orders of magnitude below the mass of the lower-mass companion in the progenitor of V1309~Sco (section \ref{sec:ILOTs}). Scaling from V1309~Sco to the systems we study here, we crudely estimate the typical duration and luminosity of the ILOTs to be according to equations (\ref{eq:ILOTtau}) and (\ref{eq:ILOTL}) , respectively.
On a broader scope, our study adds to the variety of ILOTs that planets can power as they interact with a more massive companion. We listed some earlier studies of planet-powered ILOTs in section \ref{sec:intro} (\citealt{RetterMarom2003, Retteretal2006, Bearetal2011, KashiSoker2017planet, Kashietal2019Galax}). Many of these ILOTs have peak luminosity and total energy below those of classical nova, and, therefore, literally speaking they are not `intermediate luminosity'.
Because of their low luminosity and short duration, planet-driven ILOTs on the Hertzsprung-gap are more difficult to identify. However, because they are simpler than ILOTs driven by stellar companions they can teach us on more energetic ILOTs that are driven by stellar binary systems. They are simpler because they are expected to eject less dust, hence be more transparent, and because, unlike a stellar companion, the planet does not release much energy by accreting mass.
\section*{Acknowledgements}
We thank Amit Kashi and an anonymous referee for helpful comments.
This research was supported by a grant from the Israel Science Foundation (769/20).
\section*{Data availability}
The data underlying this article will be shared on reasonable request to the corresponding author.
|
\section{Introduction}
\label{sec:introduction}
\IEEEPARstart{I}{n} 21\textsuperscript{th} century, the production lines of a large proportion of the manufacturing enterprises worldwide still rely on processing and assembly activities that are performed manually by human operators \cite{frohm2008levels}.
As a result, an alarming worsening of the workers' health conditions in the last decade has been reported \cite{bortolini2018automatic}.
Indeed, roughly three out of every five workers in the European Union (EU) report on a daily basis tiring or even painful conditions while performing their regular tasks, which frequently result in critical chronic work-related musculoskeletal disorders (WMSDs) \cite{EUOSHA2019}.
WMSDs are not just conditions of older age but also affect workers across the life-course, in all sectors and occupations, with an extremely heavy charge to companies and healthcare systems \cite{james2018global,punnett2004work}.
Hence, the careful monitoring of workers' exposure to the factors that may contribute to their development is a key requirement in the manufacturing sector, aiming to lay the foundation of risk prevention and reduction programs.
\begin{figure}[t]
\centering
\vspace{-3 mm}
\includegraphics[trim={14.5cm 6cm 5cm 2cm},clip,width=0.45\textwidth]{Figure1.pdf}
\label{fig:scenario001}
\caption{An online multi-index approach is proposed to monitor both kinematic and dynamic quantities on the human whole body. Setting proper thresholds for such indexes, the associated ergonomic risk can be categorised and more convenient body configurations and loading conditions can be identified.}
\label{fig:concept}
\vspace{-5 mm}
\end{figure}
To this end, a considerable number of methods and approaches have been proposed by researchers over the years \cite{li1999current,david2005ergonomic}.
Concerning the industrial scenario, the great majority of the tools currently adopted to evaluate workers' ergonomics relies on the so-called ``pen-and-paper'' observational techniques. Most of these analyse a particular aspects or a specific activity
(NIOSH: carrying and lifting, Snook and Ciriello: pushing and pulling, OCRA: low loads at high frequency, OWAS, RULA \& REBA: posture and movements \cite{david2005ergonomic}), however such a high specificity severely limits their application field. Conversely, the more recently developed Ergonomic Assessment Worksheet (EAWS) \cite{schaub2013european} provides a unique ergonomic score considering workers' activities in a more comprehensive view.
However, all these methods must be performed as an off-line procedure and mainly focus on the kinematics of the activities (i.e. body postures and movements) while the dynamic aspects are deemed to a limited extent.
On the other hand, to consider also the moments and forces developed within the body due to external interactions, many studies were conducted wherein direct measurements collected on the subjects are integrated with models of the human body. Most of the proposed approaches adopt detailed biomechanical models of the human musculoskeletal structure to estimate dynamic states such as joint reactions (forces and torques) by using inverse dynamics and then optimisation techniques to compute the muscle tensions \cite{delp2007opensim,damsgaard2006analysis}. Alternatively, using electromyography (EMG) signals the muscle activity can be measured directly, and then empirical models can be exploited to convert such activity into muscle tensions \cite{murai2010musculoskeletal,sartori2012emg}. Nevertheless, all the models underlying these techniques require the estimation of a large number of parameters, or otherwise, they can be obtained by means of anthropometric standards \cite{winter1995human} thus the achieved estimated quantities are not subject-specific. Moreover, the sensor system employed (e.g. EMG, marker-based optical motion-capture) are often cumbersome and the procedures involved are computationally expensive, affecting the online capabilities of most of them.
In the brand-new industrial background, the accurate assessment of a workers' ergonomics should be performed via noninvasive sensing technologies that require short preparation time as well as advanced yet rapid probing techniques that can provide instant data on human physical load \cite{ajoudani2020smart}. In light of this, many researchers have recently focused their effort on developing online and practical strategies to account for human physical ergonomic risk factors by means of robotics-inspired indexes \cite{maurice2017human,rapettipartner} but also low-dimensional space models \cite{marin2018optimizing} or exploiting vision-based algorithms \cite{parsa2019toward,liu2016tracking}.
In the same line, the foremost objective of this paper is to introduce an online human ergonomics monitoring framework that can provide on the spot, just as the workers are carrying out their activities, an estimate of the required physical load. An exhaustive set of indexes is defined that account for multiple ergonomic risk factors (e.g. awkward postures, mechanical overloading of the body, repetition frequency, etc.) and thus allow the evaluation of different manual activities typically conducted in the manufacturing sector.
Besides considering kinematic variables such as joint angles, velocities, and accelerations, a great emphasis is placed on the dynamic aspects of the tasks (i.e. forces or torques acting on and inside the body). Indeed, some indicators that were investigated and developed in previous studies \cite{kim2017real,lorenzini2019new,fortini2020compressive} such as overloading torque, fatigue, and power, the human centre of mass (CoM) potential energy, and compressive forces are integrated into the framework to establish a comprehensive method for the assessment of humans' ergonomics in the workplace (see Figure \ref{fig:concept}). The proposed framework is conceived to work with wearable and easy-to-use sensor systems, which allow the users to move freely and can be used for long periods, meeting the requirements of real factories. Furthermore, to address workers' individual demands, the estimation of such ergonomic indexes is based on a subject-specific and fast re-identifiable human model.
To evaluate the ergonomics monitoring framework introduced in this study, an experimental analysis was conducted on twelve human subjects. Participants were required to perform three different tasks that represent typical working activities in the manufacturing sector and, additionally, are associated with different potential risk factors to the development of WMSDs.
The proposed set of indexes was computed and then, their outcome was investigated to establish their link with the most common ergonomic risk factors. It should be underlined that, as previously said, the proposed indexes are conceived to perform an online evaluation, as demonstrated in our previous works \cite{kim2017real,lorenzini2019new,fortini2020compressive}. However, in this study, they are computed a posteriori for the sake of a thorough analysis. A statistical method was also employed to verify, for each ergonomic index, the significance of the differences among the different experimental conditions for each task. Indeed, in this work, we were mainly interested in the analysis of the indexes by varying the physical load (i.e. experimental conditions) to identify their capability to address the workload associated with a specific activity. Thus, we focus more on the comparison among different conditions than on the indexes' absolute values.
In addition, to provide a benchmark of the effective physical effort required for the performed activities, muscle activity was monitored through a sEMG system over the entire duration of the experiments. Since the proposed indexes are expressed on the joint level and cannot be directly compared to sEMG signals that are on the muscle level, the overall trend of the muscle activity was compared with the trend of the developed metrics. Finally, a comparison was made with the ergonomic risk scores resulting from the EAWS \cite{schaub2013european} inasmuch as they are a well-recognised method to evaluate workers' physical exposure to WMSDs.
Due to the growing interest in occupational ergonomics on one hand, and the effort needed to create and organise a database, on the other, numerous human motion datasets on industry-oriented activities were recently produced and made available to researchers and ergonomics practitioners.
To cite a few, in \cite{mandery2015kit,maurice2019human} comprehensive whole-body motion capture datasets that include trials of typical working tasks (e.g. manipulating a screw, driver or hammer, lifting loads) were proposed. In view of this, another contribution of this study is the creation of a comprehensive database that contains measurements collected with multiple sensors systems on healthy human subjects performing the aforementioned working activities. Both kinematic, dynamic and sEMG quantities are included, as well as videos recorded to facilitate the management of the data by future users.
The rest of the paper is organized as follows. In Section \ref{sec:overview}, an overview of the proposed human ergonomics monitoring framework is provided.
Section \ref{sec:experiments} illustrates the experimental analysis conducted to evaluate the framework. Section \ref{sec:database} describes the developed human measurements database on factory-like activities. In Section \ref{sec:conclusion}, the contribution and future improvements of this study are highlighted.
\section{Method Overview}
\label{sec:overview}
In this Section, the human body model adopted in this study is first presented. Then, a set of kinodynamic indexes to monitor online the humans' physical load during daily working activities is introduced.
\subsection{Human Modeling Bases}
The framework introduced in this paper has been conceived in view of its application and usage in a real industrial environment. Hence, it aims at a simplified approach that focuses on fast identifiability of the models and online capabilities of the methods, respectively.
Based on that, the human model adopted in this study is a reduced-complexity representation of the human musculoskeletal structure. Specifically, it is a floating-based sequence of rigid links interconnected by revolute joints and limited to the sagittal plane.
The pelvis frame is set as the human base frame $\Sigma_{0}$ and it is attached to the inertial frame $\Sigma_{W}$ through six virtual degrees of freedom (DoFs).
$\boldsymbol{q}=[\boldsymbol{x}_0^T \enspace \boldsymbol{\theta}_0^T \enspace \boldsymbol{q}_h]$ refers to the generalized coordinate of the system. $\boldsymbol{x_0} \in \mathbb{R}^3$ and $\boldsymbol{\theta}_0 \in \mathbb{R}^3$ are the position and the orientation of the human base frame $\Sigma_{0}$. The rigid links are articulated through $n_{j}$ revolute joints whose angular position is denoted by $\boldsymbol{q}_h=[q_1 \enspace \ldots \enspace q_{n_{j}}]^T\in \mathbb{R}^{n_{j}}$.
To obtain the human body segment inertial parameters (BSIPs), a reduced-complexity approach called statically equivalent serial chain (SESC) technique \cite{cotton2009estimation} is employed in this study.
The SESC parameters can be employed to achieve a subject-specific estimation of the whole-body CoM.
As a result, the whole-body centre of pressure (CoP) estimation, which is required in the computation of some of the ergonomic indexes that will be addressed in the next paragraph, can be accomplished.
In static conditions, the CoP can be computed by simply projecting the whole-body CoM onto the $x$-$y$ plane. Conversely, when considering dynamic conditions, the position of CoP w.r.t. the CoM can be obtained by using the differences between the angular momentum variations and the acceleration about the CoM
\cite{popovic2005ground}.
A key strength of this method is that, due to its simplified approach, it can be applied online \cite{kim2017anticipatory}.
\subsection{Set of Ergonomic Indexes}
By integrating the data about human motion and interaction forces, which can be collected through suitable sensor systems, within a model of the human body, it is possible to estimate several kinematic and dynamic quantities (e.g. joint position/velocity/acceleration, joint overloading torque/fatigue) that can be linked to ergonomic targets such as mechanical overloading of the musculoskeletal structure and body posture. Hence, the proposed human ergonomic monitoring framework is defined as a set of indexes, which are expressed as $\bm{\omega}_{h}(\mathbf{q}) \in \mathbb{R}^{n_{j}}$, that account for the physical exposure of the workers to different risk factors to the development of WMSDs. Such indexes, with the corresponding equation, scope, and the ergonomic target that each one of them seeks to address are listed in Table \ref{tab:indexes}. In general, $\mathbf{X}^{\text{max}}$ and $\mathbf{X}^{\text{min}}$ denote the vectors including the maximum and minimum values of the variable $X$, respectively. $|\mathbf{X}|$ denotes instead the vector including the absolute values of the variable $X$.
\begin{table}[h]
\centering
\caption{The proposed set of ergonomic indexes to assess human kinodynamic quantities accounting for multiple ergonomic risk factors. The corresponding scope (kinematics or dynamics), equation and ergonomic target are included for each one of them.}
\label{tab:indexes}
\begin{adjustbox}{max width=0.49\textwidth}
\begin{tabular}{cccc}
\toprule \\
\Large{SCOPE} & \Large{INDEX} & \Large{EQUATION} & \Large{RISK FACTOR}\\ \\
\midrule \\
\multirow{3}{*}{\rotatebox[origin=c]{90}{\Large{Kinematics}}} & \textbf{\large{Joint Displacement}} & {\Large{ ${\bm{\omega}}_{1}(\mathbf{q}) = \frac{|\mathbf{q}_h| }{\mathbf{q}_h^{\text{max}} - \mathbf{q}_h^{\text{min}}}$} } & \normalsize{Awkward body postures} \\ \\
& \textbf{\large{Joint Velocity}} & {\Large ${\bm{\omega}}_{2}(\mathbf{q}) = \frac{\dot{\mathbf{q}_h}}{\dot{\mathbf{q}_h}^{\text{max}}}$ } & \normalsize{Abrupt and sudden efforts} \\ \\
& \textbf{\large{Joint Acceleration}} & {\Large${\bm{\omega}}_{3}(\mathbf{q}) = \frac{\ddot{\mathbf{q}_h}}{\ddot{\mathbf{q}_h}^{\text{max}}}$ } & \normalsize{\begin{tabular}[c]{@{}c@{}}High-intensity forces \\ due to inertia\end{tabular}} \\ \\ \midrule
& \begin{tabular}[c]{@{}c@{}}\textbf{\normalsize{Overloading}} \\ \textbf{\large{Joint Torque}}\end{tabular} & {\Large${\bm{\omega}}_{4}(\mathbf{q}) = \frac{\Delta{\bm{\tau}}}{\Delta{\bm{\tau}}^{\text{max}}} $} & \normalsize{\begin{tabular}[c]{@{}c@{}}Mechanical overburden \\ of the musculoskeletal stucture\end{tabular}} \\ \\ \cmidrule{2-4}
& \begin{tabular}[c]{@{}c@{}}\textbf{\large{Overloading}} \\ \textbf{\large{Joint Fatigue}}\end{tabular} & {\Large${\bm{\omega}}_{5}(\mathbf{q}) = \frac{\bm{\tau}^{F}(t)}{\bm{\tau}^{F,\text{max}}(t)}$} & \normalsize{\begin{tabular}[c]{@{}c@{}}Repetitive and monotonous \\ movements \end{tabular}}\\ \\ \cmidrule{2-4}
\multirow{-3}{*}{\rotatebox[origin=c]{90}{\Large{Dynamics}}}& \begin{tabular}[c]{@{}c@{}}\textbf{\large{Overloading}} \\ \textbf{\large{Joint Power}}\end{tabular} & {\Large${\bm{\omega}}_{6}(\mathbf{q}) = \frac{\mathbf{P}}{\mathbf{P}^{\text{max}}} = \frac{\dot{\mathbf{q}_h}\Delta{\bm{\tau}}}{\dot{\mathbf{q}_h}^{\text{max}}\Delta{\bm{\tau}}^{\text{max}}}$} & \normalsize{\begin{tabular}[c]{@{}c@{}}Mechanical overburden \\ of the musculoskeletal stucture\end{tabular}} \\ \\
& \begin{tabular}[c]{@{}c@{}}\textbf{\large{CoM}} \\ \textbf{\large{Potential Energy}}\end{tabular} & {\Large${\bm{\omega}}_{7}= \frac{\Delta {E}_{P}}{ \Delta {E}_{P}^{\text{max}}} = \frac{\Delta C_{M}|z}{ \Delta C_{M}|z^{\text{max}}}$} & \normalsize{Awkward body postures} \\ \\ \cmidrule{2-4}
& \textbf{\large{Compressive Forces}} & {\Large${\bm{\omega}}_{8}(\mathbf{q}) = \frac{\mathbf{f}_{C}}{\mathbf{f}_{C}^{\text{max}}} $} & \normalsize{\begin{tabular}[c]{@{}c@{}}Mechanical overburden \\ of the musculoskeletal stucture\end{tabular}} \\ \\ \bottomrule
\end{tabular}
\end{adjustbox}
\end{table}
It should be noted that the indexes listed in Table \ref{tab:indexes} are all normalised, i.e. divided by their corresponding maximum values, to concurrently analyse and compare their outcome. In addition, as indicated in the leftmost column, they can be divided into two categories: kinematics and dynamics, respectively. The explanation (with reference to the equations reported in Table \ref{tab:indexes}), as well as the corresponding rationale of each ergonomic index, will be addressed as follow.
\subsubsection{Kinematic Indexes}
The first category includes the variables related to the kinematics (i.e. motions without regard to forces that cause it) of the workers' activities.
The first index $\bm{\omega}_{1}(\mathbf{q})$ is the joint displacement, which is considered here as the distance from the mechanical joint limits. Given the joint angles vector $\mathbf{q}_h$ as defined in the previous paragraph, $\bm{\omega}_{1}(\mathbf{q})$ can be computed considering its absolute values $|\mathbf{q}_h|$ and the joint angles upper and lower boundaries, $\mathbf{q}_h^{\text{max}}$ and $\mathbf{q}_h^{\text{min}}$, respectively, which can be found in literature \cite{whitmore2012nasa}.
The aim of monitoring the joint displacement is to detect wherever the body configurations that workers adopt to perform their tasks are not ergonomic. In fact, certain human postures have proven to be a potential cause of musculoskeletal diseases \cite{lutmann2003protecting}.
The latter are mostly related to specific sections of the human range of motion (RoM) that should be avoided (e.g. in proximity to the maximum limits).
Next, the second index $\bm{\omega}_{2}(\mathbf{q})$ is the joint normalised angular velocity, where $\dot{\mathbf{q}}_h \in \mathbb{R}^{n_{j}}$ is the joint velocities vector and $\dot{\mathbf{q}_h}^{\text{max}}$ is the vector including the joints maximum velocities. The values of $\dot{\mathbf{q}_h}^{\text{max}}$ can be found in literature \cite{jessop2016maximum}.
Finally, the third index $\bm{\omega}_{3}(\mathbf{q})$ is the joint normalised angular acceleration, where $\ddot{\mathbf{q}}_h \in \mathbb{R}^{n_{j}}$ is the joint accelerations vector and $\ddot{\mathbf{q}}_h^{\text{max}}$ is the vector including the joints maximum accelerations. The values of $\ddot{\mathbf{q}}_h^{\text{max}}$ can be obtained experimentally, by asking the subjects to move dynamically to the maximum extent possible, exciting all the joints, and recording the maximum accelerations achieved.
As reported by some researchers, when assessing dynamically varying activities, the direct measurement of human motion - in terms, for example, of joint velocity and acceleration - can provide more relevant information than posture analysis for the ergonomic risk of the low back \cite{marras1995biomechanical} and upper limbs \cite{malchaire1997relationship}.
Joint velocities, in particular, are suggestive of eventual abrupt movements that may be frequently associated with sudden and acute efforts potentially leading to severe injuries. On the other hand, the forces induced in the human joints due to the inertia of the body masses are directly proportional to the accelerations (e.g. Newton's second law) and may lead to the overload of musculoskeletal tissues. This can happen, for example, when the workers perform actions that require high accelerations to be accomplished efficiently (e.g. hammering).
\subsubsection{Dynamic Indexes}
The second category includes the variables related to the dynamics (i.e. motions that result from forces) of the workers' activities.
The fourth index arises from the ``overloading joint torque'' method that was originally proposed in \cite{kim2017real} to account online for the torque variations induced on the human main joints by an external heavy load. Its computation is based on the displacement of the CoP, computed from the difference between an estimated one and a measured one. The estimated CoP vector $\hat{\bm{C}}_P$ can be obtained as explained in the previous paragraph, by taking advantage of the SESC technique. On the other hand, the measured CoP vector ${\bm{C}}_{P}$ can be collected using an external sensor system.
Details of the method can be found in \cite{kim2017real}.
Specifically, ${w}_{4}(\mathbf{q})$ is the joint normalised overloading torque, where $\Delta{\bm\tau} \in \mathbb{R}^{n_{j}}$ is the overloading joint torque vector and $\Delta{\bm\tau}^{\text{max}}$ is the vector including the maximum values of the human overloading joint torque, which can be estimated directly through experiments on the subjects.
Increasing torque profiles are applied on the selected body joints, one at a time, until the subjects start to feel discomfort. In such a specific instant, the resulting torque values are estimated (based on the applied force and the lever arm) and compared to the ones extracted from literature
\cite{anderson2007maximum}.
If these values are comparable, the experimental ones are used as the maximum torque values. If the differences are significant, the ``safest'' choice, i.e. the smallest value for the maximum torque, is chosen.
The overloading joint torque index $\bm{\omega}_{4}(\mathbf{q})$ can be employed to account for the mechanical overburden of musculoskeletal structures induced by the weight of a tool or object. It should be pointed out that this method takes into account only the vertical component of external forces thus it can be deployed in a certain class of industrial tasks, which nevertheless, are quite numerous (lifting/lowering, carrying, pick and place, etc.).
The fifth index is the ``overloading joint fatigue'' that was originally introduced in \cite{lorenzini2019new} as an extension of the previous index.
Indeed, when dealing with a light payload (e.g. a lightweight tool), the overloading torques induced on the joints are low/moderate and the associated risk of injuries is not significant. On the other hand,
the building up of their effect over a protracted period of time could become hazardous thus a whole-body fatigue model was developed to evaluate this aspect.
Details of the method can be found in \cite{lorenzini2019new}.
Specifically, ${\bm\omega}_{5}(\mathbf{q})$ is the joint normalised overloading fatigue, where $\bm{\tau}^{F}\in \mathbb{R}^{n_{j}}$ is the overloading joint fatigue vector and $\bm\tau^{F,\text{max}}$ is the vector including the maximum values of the human overloading joint fatigue. Since fatigue is strictly related to the subject’s physical capacity and feelings, the maximum endurance time (MET) concept presented in \cite{imbeau2006percentile} was employed to find experimentally the subject-specific parameters of the model (see \cite{lorenzini2019new}).
The accumulation of local joint fatigue due to external load, which can be monitored by means of ${\bm\omega}_{5}(\mathbf{q})$, can result from repetitive and monotonous actions, which are among the most frequently cited contributors to WMSDs \cite{bernard1997musculoskeletal}.
The sixth index ${\bm\omega}_{6}(\mathbf{q})$ is the ``overloading joint power'' that basically combines the second and the fourth indexes previously explained.
Specifically, ${\bm\omega}_{6}(\mathbf{q})$ is the joint normalised overloading power where $\mathbf{P}\in \mathbb{R}^{n_{j}}$ results from the overloading joint torque vector $\Delta{\bm{\tau}}$ multiplied by the joint velocity vector $\dot{\mathbf{q}}_h$, whereas $\mathbf{P}^{\text{max}}$ results from $\Delta{\bm\tau}^{\text{max}}$ multiplied by $\dot{\mathbf{q}_h}^{\text{max}}$ and includes the corresponding maximum values of $\mathbf{P}$.
Based on human joint power, several procedures can be found in the literature with the aim to estimate the physical effort needed during gait \cite{ren2007predictive} but also to detect potential risk in manual material handling \cite{gagnon1991muscular}. Overloading joint power is different from joint power in his classical conception since the overloading joint torques (which takes into account only the effect of an external load) and not the net joint torques are employed in the computation.
Nevertheless, it is considered worthwhile to investigate also ${\bm\omega}_{6}(\mathbf{q})$ as a possible indicator of the physical expenditure during occupational activities.
The seventh index ${\bm\omega}_{7}$ is the ``CoM potential energy'' that is developed based on the study presented in \cite{yang2004multi}, wherein the change in potential energy of human body masses was employed as a measure of performance within an optimisation-based approach to model human motion. The potential energy for the CoM of the human body can be obtained by multiplying the height of the CoM, thus the $z$-coordinate, by the force of gravity and the human body mass. However, such a quantity is not evaluated directly but rather its variation between different body configurations is considered. Accordingly, two potential energies are defined: ${E}_{P}^{0} = mg^{T}C^{0}_{M}|z$ that is associated with a neutral and natural body configuration and ${E}_{P} = mg^{T}C_{M}|z$ that is associated with the current one. Hence, delta CoM potential energy can be defined as
\begin{equation*}
\Delta{E}_{P} = {E}_{P} - {E}_{P}^{0} = Mg\Delta C_{M}|z = Mg(C_{M}|z - C^{0}_{M}|z),
\end{equation*}
\noindent where $M$ is the mass of the subject, $g$ is the gravity acceleration, and $ \Delta C_{M}|z $ is the variation of the $z$-coordinate of the CoM, with $C^{0}_{M}|z$ the CoM in the neutral posture and $C_{M}|z$ the CoM in the current one.
Such a quantity provides the degree of deviation of the posture of the subjects from a neutral and convenient body configuration. Hence, to some extent, it is capable to monitor the risk associated with awkward and unfavourable postures.
Accordingly, the seventh index ${\bm\omega}_{7}$ is the normalised delta CoM potential energy, where $\Delta C_{M}|z$ is the displacement of the CoM height ($z$-coordinate) and $\Delta CoM_{z}^{max}$ is its maximum value, which can be estimated experimentally, by asking the subjects to move along the $z$-coordinate to the maximum extent possible and recording the maximum CoM displacements achieved. ${\bm\omega}_{7}$ is the only joint-independent index.
Finally, the eighth index ${\bm{\omega}}_{8}(\mathbf{q})$ derives from the ``compressive force'' concept that was originally conceived in \cite{fortini2020compressive} to account for the balanced inward (``pushing'') forces that result from the interaction of the human with the environment or objects. The aim of this study was to overcome the limitations of the overloading joint torque index. In fact, there are certain body configurations or task-dependent force profiles (magnitude and direction) that are characterised by negligible joint torques but considerable compressive forces, which may be harmful to the musculoskeletal system.
Additionally, the compressive force index considers all the components of the external force. Details of the method can be found in \cite{fortini2020compressive}.
Based on the above, the eighth index ${\bm{\omega}}_{8}(\mathbf{q})$ is the joint normalised compressive force where $\mathbf{f}_{C}\in \mathbb{R}^{n_{j}}$ is the compressive force vector and $\mathbf{f}_{C}^{\text{max}}$ is the vector including the corresponding maximum values, which can be found experimentally, by asking the subjects to exert the maximum force possible and recording the maximum values achieved.
\section{Experimental analysis}
\label{sec:experiments}
To investigate the proposed monitoring framework as a tool to assess human ergonomics in the workplace, an experimental analysis was conducted on human subjects. Three different tasks were performed by each participant with the aim to simulate, in the laboratory settings, activities that are commonly carried out by workers in the current industrial scenario. Such tasks were selected to encompass the most significant risk factors in the workplace: mechanical overloading of the body joints, variable and high-intensity interaction forces, and repetitive and monotonous movements. Accordingly, lifting/lowering of a heavy object (task $1$), drilling (task $2$), and painting with a lightweight tool (task $3$), were considered, respectively, in this study.
While the subjects were carrying out such activities, the data regarding the whole-body motion and the forces exchanged with the environment (both GRF and interaction forces at the end-effector) were collected and the full set of ergonomic indexes (see Table \ref{tab:indexes}) was estimated. In addition, the ergonomic risk scores resulting from the EAWS \cite{schaub2013european} was computed in an off-line phase for the sake of comparison. The corresponding analysis and computations are carried out in collaboration with Fondazione Ergo (headquarters: Varese, Italia).
The whole experimental procedure was carried out at Human-Robot Interfaces and Physical Interaction (HRII) Lab, Istituto Italiano di Tecnologia (IIT) in accordance with the Declaration of Helsinki, and the protocol was approved by the ethics committee Azienda Sanitaria Locale (ASL) Genovese N.3 (Protocol IIT\_HRII\_ERGOLEAN 156/2020).
\subsection{Experimental setup}
\begin{figure*}[h]
\centering
\begin{subfigure}[c]{0.37\textwidth}
\centering
\includegraphics[width=0.9\textwidth]{Figure2.jpg}
\caption{Task $1$}
\label{fig:scenario001}
\end{subfigure}
\begin{subfigure}[c]{0.28\textwidth}
\centering
\includegraphics[trim = 0 10 0 30 mm,clip,width=0.9\textwidth]{Figure3.jpg}
\caption{Task $2$}
\label{fig:scenario002}
\end{subfigure}
\begin{subfigure}[c]{0.31\textwidth}
\centering
\includegraphics[trim = 0 20 0 0 mm,clip,width=0.9\textwidth]{Figure4.jpg}
\caption{Task $3$}
\label{fig:scenario003}
\end{subfigure}
\caption{Experimental setups to investigate the proposed monitoring framework as a tool to assess human ergonomics while performing three typical working activities in the current industrial scenario: \textbf{(a)} lifting/lowering of a heavy object (task $1$), \textbf{(b)} drilling (task $2$), \textbf{(c)} and painting with a spray gun (task $3$). The sensor systems employed are also highlighted.}
\label{fig:scenarios}
\vspace{-5 mm}
\end{figure*}
Twelve healthy volunteers, eight males and four females, (age: $28.9 \pm 3.5$ years; mass: $69.3 \pm 13.5$ kg; height: $173.9 \pm 6.8$ cm)\footnote{Subject data is reported as: mean $\pm$ standard deviation.} were recruited in the experimental analysis. Participants were students and researchers with no or limited experience of industrial work. Written informed consent was obtained after explaining the experimental procedure and a numerical ID was assigned to anonymise the data.
The experimental setups for the three selected tasks are illustrated in Figure \ref{fig:scenarios}. The sensor systems employed are also highlighted.
The subjects were required to wear the Xsens MVN Biomech suit, an inertial-based motion-capture suit, commercialised by Xsens Technologies B.V. (headquarters: Enschede, Netherlands) to track the whole-body motion and to stand on the Kistler force plate, commercialised by Kistler Holding AG (headquarters: Winterthur, Switzerland) to measure the whole-body CoP and GRF.
In addition, a sEMG system, the Delsys Trigno Wireless platform, commercialised by Delsys Inc. (headquarters: Natick, MA, United States), was employed to measure muscle activity as a reference to the effective physical effort required for the tasks. Specifically, ten muscles were analysed: the Anterior Deltoid (AD), the Posterior Deltoid (PD), the Biceps Brachii (BC), the Triceps Brachii (TC), the Trapezius Descendens (TR), the Erector Spinae (ES), the Gluteus Maximus (GM), the Rectus Femoris (RF), the Biceps Femoris (BF) and the Tibialis Anterior (TA). The subjects were required to bound their movements to the sagittal plane since a human sagittal model was adopted in this study. Accordingly, the selection of the muscles was made considering the ones more involved in flexion/extension movements. Prior to the experiment, the maximum voluntary contraction (MVC) exerted by each muscle was recorded for all the participants with the aim to normalise the muscle activity. The acquisition and synchronization of all the sensor data are managed using Robot Operating System (ROS) environment.
All the subjects were required to perform all three tasks. No specific instructions were provided to the subjects about the way to perform the task and no preliminary learning phase was conducted. For each task, they were asked to perform three trials and between one trial and the following, there was a resting period of $2$ minutes.
Task $1$ consisted in lifting/lowering a box on a shelf and put it/take it at/from three different height levels w.r.t. floor that were located under the human knees (V1), at the level of the pelvis (V2) and over the shoulder (V3), respectively (see Figure \ref{fig:scenario001}). Indeed, such an activity induces high-intensity forces and torques within the human body that may lead to the mechanical overload of biological structures such as muscles and joints.
Since the detrimental effect of mechanical overload primarily depends on the magnitude of the force required to accomplish it, three boxes of different weights (thus demanding different effort) were considered. Specifically, the weights of the boxes were $2.5$, $5.0$, and $10.0$ kg, respectively. During a trial, all the possible combinations of box weight/height level were considered in a randomized order to counter-balance the effect of fatigue on the participants' muscles.
Even though the load affecting the musculoskeletal system during a job depends mainly on the magnitude of the force exerted/experienced by workers, other factors can contribute to its intensification. Namely: the direction of the force, its variability over time, and the postural demands required to develop it. Accordingly, task $2$ consisted in drilling three wood panels that are located in three different positions (D1, D2, D3) on a shelf, highlighted in green in Figure \ref{fig:scenario002}. The holes, made with the driller, had to be performed approximately in the points where the red crosses are depicted.
During a trial, each point had to be drilled for $10.0$ s, in the sequence \footnote{It should be noted that the randomized order was considered only for task $1$ that was longer and more demanding. For tasks $2$ and $3$, the order was the same for all the trials since the effect of fatigue was assumed negligible.} D1, D2, D3, with a break of approximately $5.0$ s in between. Throughout the whole duration of the task, the interaction forces were measured via the developed hand/tool interface. For this purpose, the subjects had to keep the driller through a dedicated handle, which was connected to a Mini45 F/T sensor, commercialised by ATI Industrial Automation (headquarters: Apex, NC, United States).
The weight of the tool (driller and handle) was $2.5$ kg.
As mentioned above, repetitive actions with lightweight objects may result in the accumulation of local muscle fatigue, which
can cause severe injuries \cite{shin2007measurement}.
Hence, the last task selected was spray painting with a lightweight tool. Task $3$ consisted in painting two helmets along their whole surface
with a spray gun. The helmets were hanging from a hook at two different height levels (V2 and V3, respectively). During a trial, the subjects had to spray each helmet (from V3 to V2 position) moving smoothly and slowly the spray gun in proximity to the helmet: from top to bottom for $10$ s, then from to bottom to top for $10$ s and finally again from top to bottom for $10$ s. In total, each helmet has to be painted for $30$ s, with a break of $5$ s approximately in between.
After the full set of ergonomic indexes was estimated, a statistical method, i.e. the analysis of variance (ANOVA), was employed to verify the significance of the results, considering a level of statistical significance equal to $0.05$. Separate one-way repeated measures ANOVAs were applied to determine the capability of each ergonomic index to discriminate between different load levels (i.e. experimental conditions). Pairwise comparisons were also conducted by using post-hoc paired t-tests. The analysis was performed separately for each considered joint.
Finally, the sEMG signals, which were collected as a benchmark for the human physical effort, were first filtered (passband: 2-500 Hz) to remove movement artifacts for all the ten muscles \cite{stegeman2007standards}. Then, they were rectified and normalised using the MVC values previously recorded for each subject.
\subsection{Results}
\label{sec:results}
In this section, the results of the simulation experiments of typical working activities are presented for all the three selected tasks performed by twelve subjects. In addition, the findings of the analysis carried out using the EAWS method are provided.
\subsubsection{Working activities simulation experiments}
\begin{figure}[!h]
\centering
\begin{subfigure}[c]{0.13\textwidth}
\centering
\caption{2.5 kg, V1}
\includegraphics[trim= 0 0 0 0 mm, clip,width=\textwidth]{Figure5.pdf}
\end{subfigure}\hspace{5 mm}
\begin{subfigure}[c]{0.13\textwidth}
\centering
\caption{2.5 kg, V2}
\includegraphics[trim= 0 0 0 0 mm, clip,width=\textwidth]{Figure6.pdf}
\end{subfigure}\hspace{5 mm}
\begin{subfigure}[c]{0.13\textwidth}
\centering
\caption{2.5 kg, V3}
\includegraphics[trim= 0 0 0 0 mm, clip,width=\textwidth]{Figure7.pdf}
\end{subfigure}
\begin{subfigure}[c]{0.13\textwidth}
\centering
\vspace{5 mm}
\caption{5.0 kg, V1}
\includegraphics[trim= 0 0 0 0 mm, clip,width=\textwidth]{Figure8.pdf}
\end{subfigure}\hspace{5 mm}
\begin{subfigure}[c]{0.13\textwidth}
\centering
\vspace{5 mm}
\caption{5.0 kg, V2}
\includegraphics[trim= 0 0 0 0 mm, clip,width=\textwidth]{Figure9.pdf}
\end{subfigure}\hspace{5 mm}
\begin{subfigure}[c]{0.13\textwidth}
\centering
\vspace{5 mm}
\caption{5.0 kg, V3}
\includegraphics[trim= 0 0 0 0 mm, clip,width=\textwidth]{Figure10.pdf}
\end{subfigure}
\begin{subfigure}[c]{0.13\textwidth}
\centering
\vspace{5 mm}
\caption{10.0 kg, V1}
\includegraphics[trim= 0 0 0 0 mm, clip,width=\textwidth]{Figure11.pdf}
\end{subfigure}\hspace{5 mm}
\begin{subfigure}[c]{0.13\textwidth}
\centering
\vspace{5 mm}
\caption{10.0 kg, V2}
\includegraphics[trim= 0 0 0 0 mm, clip,width=\textwidth]{Figure12.pdf}
\end{subfigure}\hspace{5 mm}
\begin{subfigure}[c]{0.13\textwidth}
\centering
\vspace{5 mm}
\caption{10.0 kg, V3}
\includegraphics[trim= 0 0 0 0 mm, clip,width=\textwidth]{Figure13.pdf}
\end{subfigure}
\begin{subfigure}[c]{0.4\textwidth}
\centering
\vspace{5 mm}
\includegraphics[trim= 0 0 0 0 mm, clip,width=\textwidth]{Figure14.png}
\end{subfigure}
\vspace{2 mm}
\caption{Polar plots to represent the overall outcome of the proposed ergonomic indexes for task $1$. For each experimental condition, i.e. box weight ($2.5$, $5.0$, $10.0$ kg) and height level (V1, V2, V3), the maximum values of each index over the trial are shown in the human main joints: hip (H), knee (K), ankle (A), back (B), shoulder (S), elbow (E) and wrist (W).}
\label{fig:Task001_polar}
\vspace{-8mm}
\end{figure}
The outcome of the proposed ergonomic indexes has been resumed in the aggregated polar plots illustrated in Figures \ref{fig:Task001_polar}, \ref{fig:Task002_polar} and \ref{fig:Task003_polar} for task $1$, task $2$ and task $3$, respectively. For task $1$, a plot is provided for each one of the nine actions corresponding to a specific experimental condition (box weight, height level). For task $2$, a plot is provided for each one of the three actions corresponding to the three different positions (D1, D2, D3) of the panels to be drilled. For task $3$, two plots are considered to represent the two phases, i.e. the two helmets to be processed, of the painting activity.
Each plot includes eight axes to correspond to the eight ergonomic indexes listed in Table \ref{tab:indexes}. For each index, the maximum value of each action/phase is exhibited for all the considered joints, specifically: hip (H), knee (K), ankle (A), back (B), shoulder (S), elbow (E) and wrist (W). It should be noted that the choice of the maximum value is due to the short duration of the task, however, also other values such as the root mean square could have been considered. To highlight the scope of the ergonomic indexes, two separate areas are enhanced on the polar plots: the pink area encompasses the indexes accounting for kinematic aspects while the light blue area encompasses the indexes accounting for dynamic aspects.
\begin{figure}[h]
\centering
\vspace{-3 mm}
\includegraphics[width=0.5\textwidth]{Figure15.pdf}
\caption{Boxplots for the overloading joint torque index $\bm\omega_{4}$ averaged among all the trials and subjects for the experimental conditions of task $1$, i.e. box weight ($2.5$, $5.0$, $10.0$ kg) and height level (V1, V2, V3), in the human main joints. The $F$-ratio ($F$) and $p$-values ($p$) resulting from the ANOVA are also reported.}
\label{fig:boxplot_1}
\end{figure}
Let us consider the plots in Figure \ref{fig:Task001_polar}, relative to task $1$. Focusing on the pink area, which comprises the indexes accounting for the kinematic aspects, it can be noted that the highest values were reported by $\bm\omega_{1}$, namely, the joint displacement index. At the hip, knee, and elbow level the joint angles approached the boundaries of the RoM (i.e. mechanical joint limits) to a slightly different extent among the experimental conditions, suggesting that the subjects were required to adopt potentially unfavourable postures to execute this specific task. As regards $\bm\omega_{2}$, the values are gathered approximately in the middle of the range and no meaningful differences in joint velocities can be found among the different experimental conditions. Similarly, as shown by $\bm\omega_{3}$, comparable joint accelerations were exhibited by the subjects, even varying the conditions, with values in the lower half of the index range for all the joints. Interestingly, the only exception is the acceleration at the wrist level that presented rather high values. Since the variations of wrist acceleration have proven to correlate with ergonomic risk factors in the workplace \cite{estill2000use}, this data acquire certain importance for the purposes of the current analysis. On the other hand, focusing on the light blue area of the polar plots, which covers the indexes accounting for the dynamic aspects, it can be noted that the highest values were reported by $\bm\omega_{4}$, the overloading joint torque index. $\bm\omega_{5}$ and $\bm\omega_{6}$ remained generally low and steady among the different experimental conditions, $\bm\omega_{7}$ (which is actually joint-independent) was slightly higher but likewise without considerable variations. $\bm\omega_{8}$ could not just be estimated due to the absence of full knowledge about interaction forces for this task. Conversely, the overloading joint torque index exhibited a clear growth with increasing box weight and to a small extent with increasing height level.
A statistical analysis was also performed to compare the variance between experimental conditions on the dynamic ergonomic indexes.
In Figure \ref{fig:boxplot_1}, the boxplots representing the values of $\bm\omega_{4}$ averaged among all the trials and subjects for all the experimental conditions of task $1$,
are illustrated in the human main joints. The $F$-ratio ($F$) and $p$-values ($p$) resulting from the ANOVA are also reported. For all the joints, significant differences can be observed among the different experimental conditions ($p$ $<0.05$ for the hip, $p$ $<0.001$ for the other joints). On the other hand, the differences between subgroups, examined by using the post-hoc paired t-tests for $\bm\omega_{4}$ among all the experimental conditions, are reported in Table S-I and Table S-II (supplementary file).
The comparison among different height levels within the same box weight condition shows significant difference only for the upper body. Instead, the comparison among different box weights, especially when considering $10.0$ kg, shows significant differences ($p$ $<0.05$) in most of the cases/joints.
Due to the large datasets, results of the full statistical analysis are not reported for all the indexes but only for the overloading joint torque index $\bm\omega_{4}$, which was able to provide more relevant information (i.e. the dynamic aspects) for this task.
However, concerning $\bm\omega_{5}$ and $\bm\omega_{6}$, no significant differences were found for the upper body while in the lower body the results were comparable to $\bm\omega_{4}$, in line with the fact that the former are a function of the latter. Conversely, for $\bm\omega_{7}$ no significant differences were present.
Considering the outcome of the sEMG investigation, the trend of the muscle activity was comparable to the one of $\bm\omega_{4}$ in most of the considered muscles, indicating its capability to address human physical effort. Specifically, the percentage increase of the muscle activity between the $2.5$ kg box and the $5.0$/$10.0$ kg boxes (averaged among height levels which showed less significant variations) was $51.3/112.9$ \% in the AD, $33.4/115.9$ \% in the PD, $67.9/175.1$ \% in the BC, $9.1/69.3$ \% in the TC, $49.8/143.5$ \% in the TR and $27.1/90.1$ \% in the ES, $13.2/33.5$ \% in the GM, $23.8/57.6$ \% in the RF, respectively \footnote{The reported values of sEMG are the ones exhibited by the muscles in the same instant in which the maximum value of the compared index $\bm\omega$, which is shown in the polar plots, was found.}. The muscle activity in BF and TA instead was approximately constant among different experimental conditions. The percentage differences of all the dynamic indexes and the associated muscle activity for task $1$ can be found in Table S-III (supplementary file).
\begin{figure}[!h]
\centering
\begin{subfigure}[c]{0.13\textwidth}
\centering
\caption{D1}
\includegraphics[trim= 0 0 0 0 mm, clip,width=\textwidth]{Figure16.pdf}
\end{subfigure}\hspace{5 mm}
\begin{subfigure}[c]{0.13\textwidth}
\centering
\caption{D2}
\includegraphics[trim= 0 0 0 0 mm, clip,width=\textwidth]{Figure17.pdf}
\end{subfigure}\hspace{5 mm}
\begin{subfigure}[c]{0.13\textwidth}
\centering
\caption{D3}
\includegraphics[trim= 0 0 0 0 mm, clip,width=\textwidth]{Figure18.pdf}
\end{subfigure}
\begin{subfigure}[c]{0.4\textwidth}
\centering
\vspace{5 mm}
\includegraphics[trim= 0 0 0 0 mm, clip,width=\textwidth]{Figure14.png}
\end{subfigure}
\vspace{2 mm}
\caption{Polar plots to represent the overall outcome of the proposed ergonomic indexes for task $2$. For each experimental condition, i.e. the positions (D1, D2, D3) of the panels to be drilled, the maximum values of each index over the trial are illustrated in the human main joints: hip (H), knee (K), ankle (A), back (B), shoulder (S), elbow (E) and wrist (W).}
\label{fig:Task002_polar}
\end{figure}
\begin{figure}[h]
\centering
\includegraphics[width=0.5\textwidth]{Figure19.pdf}
\caption{Boxplots for the overloading joint torque index $\bm\omega_{4}$ and the compressive force index $\bm\omega_{8}$ averaged among all the trials and subjects for the experimental conditions of task $2$, i.e. the positions (D1, D2, D3) of the panels to be drilled, in the human main joints. The $F$-ratio ($F$) and $p$-values ($p$) resulting from the ANOVA are also reported.}
\label{fig:boxplot_2}
\end{figure}
Let us then observe the plots relative to task $2$, illustrated in Figure \ref{fig:Task002_polar}. Considering the indexes accounting for the kinematic aspects (pink area), it can be noted that either $\bm\omega_{1}$, $\bm\omega_{2}$ and $\bm\omega_{3}$ showed low values in all the experimental conditions, except the joint displacement at the elbow level (in line with the task requirements), meaning that kinematic variables are not significantly predictive of the ergonomic risk associated with this specific activity. This is consistent with the fact that the subjects were basically steady throughout its whole duration. On the other hand, considering the indexes accounting for the dynamic aspects, the predominance of $\bm\omega_{8}$ over the other indexes is clearly noticeable. The compressive forces index was overall far higher than all the other indexes and presented a certain increment in D1 with respect to D2 and D3, most likely due to the fact that in D1 the gravitational component of the external force was greater because of the tool weight. In particular, in Figure \ref{fig:boxplot_2}, the boxplots representing the values of the overloading joint torque index $\bm\omega_{4}$ and the compressive force index $\bm\omega_{8}$ averaged among all the trials and subjects for all the experimental conditions of task $2$,
are illustrated in the human main joints.
These two indexes, among the dynamics ones, are expected to provide more relevant information with respect to the type of task. In fact, similarly as for joint kinematics, neither overloading power nor CoM potential energy showed no meaningful variations, again in line with the fact that subjects were steady throughout the task.
In Figure \ref{fig:boxplot_2}, the $F$-ratio ($F$) and $p$-values ($p$) resulting from the ANOVA are also reported while in Table S-IV (supplementary file), the $p$-values resulting from the post-hoc paired t-tests among all the experimental conditions are presented. As it can be observed in the boxplots, $\bm\omega_{4}$ was quite low in the lower body and even less elevated in the upper body. The differences among drilling locations were significant for the back and the arm joints (p $<.001$), but not for the leg joints. On the other hand, $\bm\omega_{8}$ showed substantially higher values and more variability between joints and experimental conditions. The differences among drilling locations were significant for all the considered joints except the knee. Similarly, the results of the paired t-tests indicate significant differences in more cases/joints for $\bm\omega_{8}$ than for $\bm\omega_{4}$.
As a matter of fact, the overloading joint torque, as it is defined, considered only the gravitational component of the force, i.e. the weight of the tool. Since the tool was partially supported by the drilled interface while drilling, such a component was not significant during this specific task. On the other hand, the compressive force index took into account all the components of the external force induced on the human end-effector since this information was provided by the F/T sensor mounted purposely on the driller.
Considering the muscles activity analysis, similarly as for $\bm\omega_{8}$, sEMG signals displayed more significant values in D1 than in the other two conditions in several muscles, meaning that $\bm\omega_{8}$ correlated to muscle activity to some extent. Specifically, the percentage increase of the muscle activity between D2/D3 and D1 was $130.3/1153.1$ \% in the AD, $32.2/93.0$ \% in the BC $6.3/267.7$ \% in the TR, $112.7/24.1$ \% in the ES, $3.5/18.9$ \% in the GM, $358.7/106.1$ \% in the RF and $47.6/85.2$ \% in the TA, respectively. The muscle activity in PD was $86.7$ \% lower in D2 but $14.9$ \% higher in D3 than in D1, probably due to the task requirements (i.e. the position of the arm while holding the driller). The same explanation can be given for the muscle activity in TC, which was highest in D3, lower in D2, and even less significant in D1. The percentage differences of all the dynamic indexes and the associated muscle activity for task $2$ can be found in Table S-V (supplementary file).
\begin{figure}[!h]
\centering
\begin{subfigure}[c]{0.22\textwidth}
\centering
\caption{First helmet}
\includegraphics[trim= 0 0 0 0 mm, clip,width=\textwidth]{Figure20.pdf}
\end{subfigure}\hspace{5 mm}
\begin{subfigure}[c]{0.22\textwidth}
\centering
\caption{Second helmet}
\includegraphics[trim= 0 0 0 0 mm, clip,width=\textwidth]{Figure21.pdf}
\end{subfigure}
\begin{subfigure}[c]{0.4\textwidth}
\centering
\vspace{5 mm}
\includegraphics[trim= 0 0 0 0 mm, clip,width=\textwidth]{Figure14.png}
\end{subfigure}
\vspace{2 mm}
\caption{Polar plots to represent the overall outcome of the proposed ergonomic indexes for task $3$. For each phase of the task, i.e. one of the two helmets to be painted, the maximum values of each index over the phase are illustrated in the human main joints: hip (H), knee (K), ankle (A), back (B), shoulder (S), elbow (E) and wrist (W).}
\label{fig:Task003_polar}
\vspace{-2 mm}
\end{figure}
\begin{figure}[!h]
\centering
\includegraphics[width=0.5\textwidth]{Figure22.pdf}
\caption{Boxplots for the overloading joint torque index $\bm\omega_{4}$ and the overloading joint fatigue index $\bm\omega_{5}$ averaged among all the trials and subjects for the initial (i) and final (f) instants of the phases (first and second) of task $3$, i.e. one of the two helmets to be painted, in the human main joints. The $F$-ratio ($F$) and $p$-values ($p$) resulting from the ANOVA are also reported.}
\label{fig:boxplot_3}
\end{figure}
Lastly, let us focus on the plots depicted in Figure \ref{fig:Task003_polar}, relative to task $3$. Considering the kinematics-related variables (pink area), it can be observed that rather meaningful values were shown by $\bm\omega_{1}$ at the elbow and knee level, to a lesser degree at the hip level, meaning that potentially risky body configurations were adopted by the subjects to accomplish the task. On the other hand, $\bm\omega_{2}$ and $\bm\omega_{3}$ were altogether very low thus bearing no information for the purposes of the ergonomics assessment. Similarly as in task $1$, the only exception was the acceleration of the wrist that, as already mentioned, may be suggestive of potential hazards. Considering then the dynamics-related variables, it is evident that the most relevant indexes to explain the physical exposure of the task were $\bm\omega_{4}$ and $\bm\omega_{5}$, in comparison with all the other indexes that are basically negligible.
In Figure \ref{fig:boxplot_3}, the boxplots representing the values of the overloading joint torque index $\bm\omega_{4}$ and the overloading joint fatigue index $\bm\omega_{5}$ averaged among all the trials and subjects for the initial (i) and final (f) instants of the phases of task $3$, i.e. one of the two helmets to be painted, are illustrated in the human main joints. Concerning the lower body, the overloading torque showed low/moderate values while the overloading fatigue was less considerable. Concerning the upper body, the overloading torque presented higher values in the first phase for all joints except the wrist but its level remained moderate. Similarly, the overloading fatigue exhibited higher values in the first phase but, on the contrary, it reaches a considerable level, mostly in the shoulder and in the back, as it can be noted from the considerable standard deviation values. In fact, as already explained, $\bm\omega_{4}$ considered the instantaneous effect of the external load while $\bm\omega_{5}$ was capable to address its cumulative effect thus accounting for the accumulation of fatigue. Moreover, $\bm\omega_{5}$ presented more variability between different trials and subjects since fatigue occurred differently among individuals as explained by the subject-specific parameters that the model is based on. Nevertheless, what is not evident observing these results but it should be underlined, is that the overloading fatigue model was extremely sensitive to the threshold which has been set to distinguish among fatigue and recovery phase. In Figure \ref{fig:boxplot_3}, the $F$-ratio ($F$) and $p$-values ($p$) resulting from the ANOVA are also reported. The differences among the considered instants were significant for all the considered joints for both $\bm\omega_{4}$ and $\bm\omega_{5}$. On the other hand, in Table S-VI (supplementary file), the $p$-values resulting from the post-hoc paired t-tests among all the experimental conditions are reported. Significant differences can be found mostly in the back and in the shoulder for both the considered indexes, in line with the fact that these are more involved joints in task $3$.
In addition, the trend of the muscle activity was comparable to the one displayed by both the overloading torque and fatigue in the lower body and in the arm, suggesting the capability of $\bm\omega_{4}$ and also $\bm\omega_{5}$ to account for human effort. Specifically, the raising of the overloading torque and fatigue in the legs between the first and the second phase was accompanied by a percentage increase of muscle activity of $348.5$ \% in the RF, $177.5$ \% in the BF, and $624.5$ \% in the TA. Conversely, the overloading torque and fatigue in the arms were reduced between the first and the second phase but still matching the activity in most of the muscles with a percentage decrement of $78.2$ \% in the AD, $15.2$ \% in the BC, $48.8$ \% in the TC and $42.9$ \% in the TR. The percentage differences of all the dynamic indexes and the associated muscle activity for task $3$ can be found in Table S-VII (supplementary file).
\subsubsection{Analysis with EAWS method}
As mentioned above, to assess the potential of the proposed set of ergonomic indexes, its outcome was compared with the ergonomic risk scores provided by a well-recognised and widely used tool to evaluate workers' physical load, namely, the EAWS method \cite{schaub2013european}.
It is critical to note that EAWS were conceived and usually applied in real industrial scenarios. On the contrary, for the sake of this experimental analysis, they were adopted with simplified activities that were designed to simulate occupational tasks and were conducted during experiments in laboratory settings. Accordingly, several assumptions were made to contextualise such simplified tasks in the workplace that was intended to be replicated.
Since the level of performance of the recruited subjects was not realistic compared to an actual worker, the task timing was defined according to standard base time using the Methods-Time Measurement - Universal Analyzing System (MTM-UAS) method \cite{maynard1948methods}. In fact, in the industrial environment, it is necessary to quantify work with respect to a regulatory level of performance \cite{caragnano2018ergo}.
The EAWS method enables a comprehensive and unique ergonomic risk evaluation. In fact, four different sections that focus on each specific aspect of manual material activities are included and their results can be integrated into a combined score presented in an intuitive traffic light scheme (green, yellow, red) according to the Machinery Directive 2006/42/EC (EN 614). Accordingly, the evaluation of the three occupational tasks considered in this experimental analysis by means of the EAWS method resulted in a total score for each of them, indicative of the associated ergonomic risk level.
The EAWS score was computed for each trial of every task and then the mean value among the resulting score was computed. This procedure was repeated for all twelve subjects. As a result, each subject presented an EAWS final score for each task. It should be noted that for task $1$ three different ratings were computed, considering the box weight conditions separately. The mean and the standard deviation values which were finally computed among all the subjects and the corresponding risk level are illustrated in Table \ref{tab:EAWSscore}.
\begin{table}[h!]
\centering
\caption{Mean and standard deviation\footnoteref{footnote_eaws} computed among twelve subjects of the EAWS final scores for all the task considered in the experimental analysis.}
\label{tab:EAWSscore}
\begin{adjustbox}{max width=0.48\textwidth}
\begin{tabular}{ccccc}
\toprule
\textbf{Task} & \textbf{Activity} & \textbf{Condition} & \textbf{EAWS Final Score} & \textbf{Risk level\footnotemark} \\
\midrule
1 & lifting/lowering & $2.5$ kg box & $38.8$ $\pm$ $1.2^{\dag\dag}$ & Moderate \\
& & $5.0$ kg box & $41.7$ $\pm$ $1.7^{\dag\dag}$ & Moderate\\
& & $10.0$ kg box & $48.7$ $\pm$ $6.9^{\dag\dag}$ & Moderate \\
\midrule
2 & drilling & & $23.1$ $\pm$ $5.0^{\dag}$ & Low\\
\midrule
3 & painting & & $14$ $\pm$ $0.0^{\dag}$ & Low \\
\bottomrule
\end{tabular}
\end{adjustbox}
\end{table}
\footnotelabeled{footnote_eaws}{The risk level associated with the EAWS score was determined as follows: the range $0-25$ points corresponded to a low risk ($^{\dag}$), the range $25-50$ points corresponded to a moderate risk ($^{\dag\dag}$) and a value $>50$ points corresponded to a high risk, respectively.}
For task $1$, the inter-subject variability of the EAWS score only depended on the worker gender, i.e. the male and the female subjects, respectively, had the exact same rating among them. This was due to the level of risk associated with the box weight which was indeed gender-specific. It can be noted that, as expected, the EAWS score of the lifting/lowering task grown with increasing box weight. Looking at the mean among the subjects, the risk level associated was moderate ($^{\dag\dag}$) for all the experimental conditions. Nevertheless, considering the $10.0$ kg box one, the EAWS score was quite close to the higher threshold ($>50$) and the standard deviation equal to $6.9$ suggests that for some subjects the risk level turned into high. It should be observed that the growing trend of the risk level with the box weight was likewise identified by the proposed human ergonomics monitoring framework.
For task $2$, each trial of each subject exhibited a rather different EAWS score since the maximum value (the worst condition) of the force exerted while drilling was taken into account. Similarly, as for the last experimental condition in task $1$, the rating was quite close to the mid threshold ($>25$) and presented a significant standard deviation ($5.0$). Thus, the risk level considering the mean value was low ($^{\dag}$) but for some subjects it turned into moderate.
On the other hand, for task $3$, the EAWS score was exactly the same for each subject, leading to a standard deviation equal to $0.0$. In fact, neither gender nor other individual factors influenced the evaluation of this specific activity but only the task timing. The risk level associated was then low for all the subjects, abundantly within the low part of the range.
\subsection{Discussion}
In view of the results presented in the previous paragraph, some remarks can be drawn for the human ergonomics monitoring framework proposed in this study.
As regards kinematics-related variables, the most beneficial index in assessing physical load for the considered task was represented by $\bm\omega_{1}$. The joint displacement index allows to detect wherever the human current body configuration lays within specific sections of the human RoM which should be avoided (e.g. in proximity to the maximum limits). Hence, by monitoring how often and how long these body configurations are maintained, it is possible to determine if a potential ergonomic risk exists. As a matter of fact, joint kinematics, in general, has extensively been used in the analysis of human movement and joint angles, in particular, have widely been employed to identify human awkward or unfavourable postures while performing working activities. Concerning joint velocities and accelerations, no significant information was offered by $\bm\omega_{2}$ and $\bm\omega_{3}$ in the context of this experimental analysis, excluding the rather significant accelerations exhibited by the wrist, which should be taken into account since they were demonstrated to correlate with occupational risks. Nevertheless, the three tasks considered were quite steady and regular and did not require the development of significant speeds and accelerations. Hence, the joint velocity and acceleration indexes should be investigated more thoroughly.
Concerning dynamics-related variables, for each considered task, the ergonomic indexes that better explain the required physical load were identified, as evaluated through statistical analysis and supported by the outcome of a sEMG investigation. Specifically, the overloading joint torque index $\bm\omega_{4}$ was the more promising mean to account for the mechanical overburden of the body structures in tasks involving the handling of heavy objects (task $1$). It showed far higher values than the other indexes with significant differences among all the experimental conditions and it correlated with the muscle activity. Instead, considering a task in which all the components of the interaction forces at the hand/tool interface are relevant - not only the gravitational one - (task $2$), the joint compressive forces index $\bm\omega_{8}$ proved to be more predictive of the risk associated with dynamically varying interaction forces. In fact, in these cases, the overloading joint torque may underestimate the hazard. Finally, concerning repetitive and monotonous prolonged activities (task $3$) both the overloading joint torque and fatigue indexes, $\bm\omega_{4}$ and $\bm\omega_{5}$, respectively, exhibited significant results. The overloading fatigue index happened to be more valuable since it better accounts for the variability among the subjects. Moreover, it took into consideration the cumulative effect of the risk and not the instantaneous one thus it was able to account for the accumulation of fatigue. However, the sensitivity of the overloading fatigue model to thresholds must be further investigated, also considering longer-lasting activities. Similarly, concerning the overloading joint power index $\bm\omega_{6}$ and the CoM potential energy index $\bm\omega_{7}$ a deeper examination should be made. Not remarkable or not easily interpretable information appeared to be provided by such variables but this may be due to the fact that, as already noticed, no considerable velocities or unexpected movements were required to accomplish the analysed tasks. Conversely, regular, smooth, and constant-speed actions were repeated. Hence, the potential of both the indexes should be investigated more deeply while conducting more dynamically varying jobs, likewise for $\bm\omega_{2}$ and $\bm\omega_{3}$.
Ultimately, a comparison should be made between the proposed set of ergonomic indexes and the EAWS. Since the EAWS only provides a single score for each task rather than a trend, the similarities of the differences among the experimental conditions were simply analysed.
It should be observed that EAWS happened to be subject-specific only to a certain degree. In fact, the inter-subject variations were significant only for task $2$. For task $1$, only gender was crucial to the computation of the final score
while for task 3, the exact same score was assigned to all the subjects.
Consequently, it is possible to state that the set of ergonomic indexes proposed in this study pays much more attention to the individual's behaviours and demands due to a subject-specific model of the human body.
In addition, the EAWS analysis consists in a rather complex and articulated procedure that can be conducted only in an off-line phase and, according to the author, to some extent is affected by the subjective opinions of the expert who carries out the analysis.
Finally, this method seems still to lack a full evaluation of all the dynamic aspects underlying the execution of occupational tasks.
\section{Human Measurements Database}
\label{sec:database}
The created database including the data from all the sensors system employed is available in Zenodo\footnote{Digital Object Identifier: 10.5281/zenodo.5575139} and is organised as follows. There is a folder for each task (task $1$: lifting/lowering of a heavy object, task $2$: drilling, task $3$: painting with a lightweight tool). Inside, there is a folder for each one of the twelve participants (named after the subject’s ID) that includes a file containing the ID, age, gender, handedness, body mass, body dimensions, MVC values of the subject, and then, four folders containing the data in the available formats, i.e. \textit{.mat}, \textit{.bag}, \textit{.mp4} and \textit{.mvnx} (plus \textit{.c3d}). In all these folders, there is a file for each one of the trials performed by the subjects.
Each \textit{.mat} is a binary MATLAB\textregistered~ (MathWorks, Inc.; headquarters: Natick, MA, United States) file that stores workspace variables and contains the information that is listed in Table \ref{tab:database}, divided according to the sensor system employed to collect it. Similarly, each \textit{.bag} file contains the same information but in the form of ``ROS bags'' that are a file format developed in ROS for storing ROS message data and can be conveniently played back (details can be found in \cite{ROS}). Each \textit{.mp4} file contains the video recording of each trial to facilitate the data usage. Finally, each \textit{.mvnx} is the Xsens custom open XML format that contains all the information collected by the Xsens system. Also the \textit{.c3d} exported from the latter is available. Besides the participants' folders, in the database home folder, some additional files are also provided to easily manage the data. The latter include: a short guide with general information about the database, a file describing the experimental setup, and a file describing in detail the human model employed and the reference frame adopted.
\begin{table}[]
\centering
\caption{Information included in the human measurements database on factory-like activities divided according to the sensor system employed to collect it.}
\label{tab:database}
\begin{adjustbox}{max width=0.5\textwidth}
\begin{tabular}{cl} \toprule
\textbf{Sensor system} & \textbf{Information} \\ \midrule
\multirow{2}{*}{Motion-capture suit} & \begin{tabular}[c]{@{}l@{}} 3D position, orientation (quaternion), linear and angular \\ velocity, and linear and angular acceleration of the origin \\ of the 21 body segments of the human model.\end{tabular} \\
& Angles of the 20 joints (3 DoFs each) of the human model.\\ \midrule
\multirow{3}{*}{Force plate} & 2D position of the CoP. \\
& GRF. \\
& Ground reaction moment (GRM). \\ \midrule
sEMG system & Muscle activity of the ten recorded muscles \\ \midrule
\multirow{2}{*}{F/T sensor} & 3D force \\
& 3D torque \\ \bottomrule
\end{tabular}
\end{adjustbox}
\vspace{-4 mm}
\end{table}
\section{Conclusions}
\label{sec:conclusion}
In this paper, a novel method to assess human ergonomics in the workplace was introduced. Several physical variables are estimated that can lay the foundations for the ergonomic risk categorisation.
Since the proposed set of indexes was conceived to address multiple physical risk factors, it was capable to evaluate different working activities. Indeed, for each considered task, a subset of ergonomic indexes was found to be predictive of the required physical load. Significant differences were found among different experimental conditions, as evaluated by the statistical analysis and further evidence is provided by the trend of the sEMG signals. The potential of the remaining indexes must be investigated more deeply in future studies considering more dynamically varying and activities.
The comparison between the EAWS and the proposed set of ergonomic indexes highlighted the pros and cons of the two approaches. Hence, upcoming studies may focus on their integration, allowing a more comprehensive and thorough analysis of the occupational risk factors, taking equally into account the relative kinematics and dynamic aspects and addressing workers' specific requirements.
The proposed framework was developed envisaging its use in an industrial environment thus it can operate with fit-for-workplaces (i.e. wearable, long-life, and easy-to-use) sensor systems. In this work, the computations were made off-line to ensure higher accuracy but in future studies its online capabilities will be demonstrated.
A force plate was employed for the sake of a thorough validation, however, it can be replaced with sensor insoles. On the other hand, a wearable inertial-based system was employed to track human motion but vision-based systems can be considered as an alternative.
A sagittal model was employed to represent the human body since only activities performed along the sagittal plane were analysed. Nevertheless, its extension to a 3D model to consider more complex tasks will be investigated in the future.
Lastly, the database created in this study would serve a twofold purpose. First, to advance the research in the monitoring and assessment of human ergonomics in the context of the ERC project Ergo-Lean, that aims to propose collaborative robotics technology to improve human physical factors in the workplace. Next, to share experimental data with the scientific community for comparisons in ergonomics studies as well as model identification or learning for assistive and collaborative robotics. In future works, the extension and enrichment of this database will be covered as well.
\ifCLASSOPTIONcaptionsoff
\newpage
\fi
\bibliographystyle{IEEEtran}
|
\section{Parameterizing Spaces of Signals with Neural Fields}
\label{sect:function}
An arbitrary signal $\mathcal{I}$ in $\mathbb{R}^{D_1 \times \cdots \times D_k}$ can be represented efficiently as a function $\Phi$ which maps each coordinate ${\bm{x}} \in \mathbb{R}^k$ to the value ${\bm{v}}$ of the feature coordinate at that dimension.
\begin{equation}
\Phi : \mathbb{R}^{k} \rightarrow \mathbb{R}, \quad {\bm{x}} \rightarrow \Phi({\bm{x}}) = {\bm{v}}
\end{equation}
Such a formulation enables us to represent any signal, such as an image, shape, or waveform -- or a set of signals, like images with waveforms -- as a continuous function $\Phi$.
We use a three-layer multilayer perceptron (MLP) with hidden dimension 512 as our $\Phi$. Only the input dimension of $\Phi$ is varied, depending on the dimensionality of the underlying signal -- all other architectural choices are the same across signal types. We refer to $\Phi$ as a neural field.
\myparagraph{Parameterizing Spaces over Signals.}
A set of signals may be parameterized as a subspace $\mathcal{F}$ of functions $\Phi_i$. To efficiently parameterize functions $\Phi$, we represent each $\Phi$ using a low-dimensional latent space in $\mathbb{R}^n$ and map latents to $\mathcal{F}$ via a hypernetwork $\Psi$:
\begin{equation}
\Psi : \mathbb{R}^{n} \rightarrow \mathcal{F}, \quad {\bm{z}} \rightarrow \Psi({\bm{z}}) = \Phi,
\end{equation}
To regress $\Phi$ with our hypernetwork, we predict individual weights $\bm{W}^{\mathcal{L}}$ and biases $\bm{b}^{\mathcal{L}}$ for each layer $\mathcal{L}$ in $\Phi$. Across all experiments, our hypernetwork is parameterized as a three-layer MLP (similar to our $\Phi$), where all weights and biases of the signal representation $\Phi$ are predicted in the final layer.
Directly predicting weights $\bm{W}^{\mathcal{L}}$ is a high-dimensional regression problem, so following \citep{inr_gan}, we parameterize $\bm{W}$ via a low-rank decomposition as $\bm{W}^{\mathcal{L}} = \bm{W}^{\mathcal{L}}_s \odot \sigma(\bm{W}^{\mathcal{L}}_h)$, where $\sigma$ is the sigmoid function.
$\bm{W}^{\mathcal{L}}_s$ is shared across all predicted $\Phi$, while $\bm{W}^{\mathcal{L}}_h = A^{\mathcal{L}} \times B^{\mathcal{L}}$ is represented as two low-rank matrices $A^{\mathcal{L}}$ and $B^{\mathcal{L}}$ -- and is regressed separately for each individual latent ${\bm{z}}$.
\section{Learning Structured Manifold of Signals}
\input{figText/method_figure}
Given a training set $\mathcal{C} = \{\mathcal{I}_i \}_{i=1}^N$, of $N$ distinct arbitrary signals $\mathcal{I}_i \in \mathbb{R}^{D_1 \times \cdots \times D_k}$, our goal is to learn -- in a \textit{signal-agnostic} manner -- the underlying low-dimensional \textit{manifold} $\mathcal{M} \subset \mathbb{R}^{D_1 \times \cdots \times D_k}$.
In particular, a manifold $\mathcal{M}$ \citep{Roweis2000Nonlinear, Tenenbaum2000global}, over signals $\mathcal{I}$, is a locally low-dimensional subspace consisting and describing all possible variations of $\mathcal{I}$. A manifold has three crucial properties. First, it must embed all possible instances of $\mathcal{I}$. Second, $\mathcal{M}$ should be locally metric around each signal $\mathcal{I}$, enabling us to effectively navigate and interpolate between nearby signals. Finally, it must be locally perceptually consistent; nearby points on $\mathcal{M}$ should be similar in identity. To learn a manifold $\mathcal{M}$ with these properties, we introduce a loss for each property: $\mathcal{L}_{\text{Rec}}$, $\mathcal{L}_{\text{LLE}}$, and $\mathcal{L}_{\text{Iso}}$, respectively. $\mathcal{L}_{\text{Rec}}$ ensures that all signals $\mathcal{I}$ are embedded inside $\mathcal{M}$ (\sect{sect:autodecoder}), $\mathcal{L}_{\text{LLE}}$ forces the space near each $\mathcal{I}$ to be locally metric (\sect{sect:lle}), and $\mathcal{L}_{\text{Iso}}$ encourages the underlying manifold to be perceptually consistent (\sect{sect:iso}). As such, our training loss takes the form:
\begin{equation}
\mathcal{L}_{\text{Total}} = \mathcal{L}_{\text{Rec}} + \mathcal{L}_{\text{LLE}} + \mathcal{L}_{\text{Iso}}
\end{equation}
Please see Fig.~\ref{fig:method} for an overview of the proposed manifold learning approach. We utilize equal weighting across each introduced loss term. We discuss each component of the loss in more detail below.
\subsection{Data Coverage via Auto-Decoding}
\label{sect:autodecoder}
Given a parameterization of a subspace $\mathcal{F}$ (as introduced in \sect{sect:function}) which represents our manifold $\mathcal{M}$, we must ensure that $\Psi$, our hypernetwork, can effectively cover the entire subspace in which our signal $\mathcal{I}$ lies. A common approach to learn subspaces over signals is to employ a GAN \citep{Goodfellow2014Generative}, whereby a discriminator trains a hypernetwork $\Psi({\bm{z}})$ to generate signals that are indistinguishable from real examples. However, an issue with such an approach is that many $\mathcal{I}$ may be missing in the resultant mapping.
Here, to ensure that our manifold covers all instances, we utilize the auto-decoder framework~\cite{park2019deepsdf,sitzmann2019srns}. We learn a separate latent ${\bm{z}}_i$ for each training signal $\mathcal{I}_i$ and train models utilizing the loss
\begin{equation}
\mathcal{L}_{\text{Rec}} = \|\Psi({\bm{z}}_i)({\bm{x}}) - \mathcal{I}_i({\bm{x}}) \|^2
\end{equation}
over each possible input coordinate ${\bm{x}}$. By explicitly learning to reconstruct each training point in our latent space, we enforce that our latent space, and thus our manifold $\mathcal{M}$, covers all training signals $\mathcal{I}$.
\subsection{Local Metric Consistency via Linear Decomposition}
\label{sect:lle}
A manifold $\mathcal{M}$ is locally metric around a signal $\mathcal{I}$ if there exist local linear directions of variation on the manifold along which the underlying signals vary in a perceptually coherent way. Inspired by \citep{Roweis2000Nonlinear}, we enforce this constraint by encouraging our manifold to consist of a set of local convex linear regions $\mathcal{R}_i$ in our learned latent space.
We construct these convex linear regions $\mathcal{R}_i$ using autodecoded latents ${\bm{z}}_i$ (\sect{sect:autodecoder}). Each latent ${\bm{z}}_i$ defines a convex region, $\mathcal{R}_i$ -- obtained by combining nearest latent neighbors ${\bm{z}}_i^1, \ldots, {\bm{z}}_i^j$, as illustrated in \fig{fig:method}.
To ensure that our manifold $\mathcal{M}$ consists of these convex local regions, during training, we enforce that each latent ${\bm{z}}_i$ can be represented in a local convex region, $\mathcal{R}_i$, e.g. that it can be expressed as a set of weights ${\bm{w}}_i$, such that ${\bm{z}}_i = \sum_j w_i^j {\bm{z}}_i^j$ and $\sum_j w_i^j = 1$. Given a ${\bm{z}}_i$ we may solve for a given ${\bm{w}}_i$ by minimizing the objective $\|{\bm{z}}_i -\sum_j w_i^j {\bm{z}}_i^j\|^2 = {\bm{w}}_i^T \bm{G}_i {\bm{w}}_i$ with respect to ${\bm{w}}_i$. In the above expression, $\bm{G}_i$ is the Gram matrix, and is defined as $\bm{G}_i = {\bm{k}} {\bm{k}}^T$, where ${\bm{k}}$ is a block matrix, with row $j$ corresponding to ${\bm{k}}_j = {\bm{z}}_i - {\bm{z}}_i^j$. Adding the constraint that all weights add up to one via a Legrange multiplier, leads to the following optimization objective:
\begin{equation}
\mathcal{L}({\bm{w}}_i, \lambda) = {\bm{w}}_i^T \bm{G}_i {\bm{w}}_i + \lambda (\bm{1}^T{\bm{w}}_i - 1)
\end{equation}
where $\bm{1}$ corresponds to a vector of all ones. Finding the optimal ${\bm{w}}_i$ in the above expression then corresponds to the best projection of ${\bm{z}}_i$ on the linear region $\mathcal{R}_i$ it defines. Taking gradients of the above expression, we find that
${\bm{w}}_i = \frac{\lambda}{2}\bm{G}_i^{-1} \bm{1}$, where $\lambda$ is set to ensure that elements of ${\bm{w}}_i$ sum up to 1. The mapped latent ${\bm{z}}_i' = \sum_j {\bm{w}}_i^j {\bm{z}}_i^j$ corresponds to the best projection ${\bm{z}}_i$ on $\mathcal{R}_i$.
To enforce that ${\bm{z}}_i$ lies in $\mathcal{R}_i$, we enforce that this projection
\begin{equation}
\mathcal{L}_{\text{LLE}} = \|\Psi({\bm{z}}_i')({\bm{x}}) - \mathcal{I}_i({\bm{x}}) \|^2 ,
\end{equation}
also decodes to the training signal $\mathcal{I}_i$, where we differentiate through intermediate matrix operations. To encourage ${\bm{w}}_i$ to be positive, we incorporate an additional L1 penalty to negative weights.
\subsection{Perceptual Consistency through Local Isometry}
\label{sect:iso}
Finally, a manifold $\mathcal{M}$ should be perceptually consistent. In other words, two signals that are perceptually similar should also be closer in the underlying manifold, according to a distance metric on the respective latents, than signals that are perceptually distinct.
While in general the MSE distance between two signals is not informative, when such signals are relatively similar to one another, their MSE distance \textit{does} contain useful information.
Therefore, we enforce that the distance between latents ${\bm{z}}_i$ and ${\bm{z}}_j$ in our manifold $\mathcal{M}$ are locally isometric to the MSE distance of the corresponding samples $\mathcal{I}_i$ and $\mathcal{I}_j$, when relative pairwise distance of samples is small. On the lowest 25\% of pairwise distances, we utilize the following loss:
\begin{equation}
\mathcal{L}_{\text{Iso}} = \| \alpha * \|({\bm{z}}_i - {\bm{z}}_j)\| - \| \mathcal{I}_i - \mathcal{I}_j \|\|.
\end{equation}
Such local isometry may alternatively be enforced by regularizing the underlying Jacobian map.
\subsection{Applications}
By learning a modality-independent manifold $\mathcal{M}$, GEM\xspace permits the following \textit{signal-agnostic} downstream applications, each of which we explore in this work.
\myparagraph{Interpolation.} By directly interpolating between the learned latents ${\bm{z}}_i$ and ${\bm{z}}_j$ representing each signal, we can interpolate in a perceptually consistent manner between neighboring signals in $\mathcal{M}$. Further, latents for novel signals may be obtained by optimizing $\mathcal{L}_{\text{Rec}}$ on the signal.
\myparagraph{Conditional Completion.} Given a partial signal $\hat{\mathcal{I}}$, we can obtain a point on the manifold $\mathcal{M}$ by optimizing a latent ${\bm{z}}$ using $\mathcal{L}_{\text{Rec}} + \mathcal{L}_{\text{LLE}}$. Decoding the corresponding latent ${\bm{z}}$ then allows us to complete the missing portions of the image.
\myparagraph{Sample Generation.} Our manifold $\mathcal{M}$ consists of a set of locally linear regions $\mathcal{R}_i$ in a underlying latent space ${\bm{z}}$. As such, new samples in $\mathcal{M}$ can be generated by walking along linear regions $\mathcal{R}_i$, and sampling points each such region. To sample a random point within $\mathcal{R}_i$, we generate $\alpha$ between 0 and 1, and compute a latent $\hat{{\bm{z}}_i} = \alpha {\bm{z}}_i + (1 - \alpha) * {\bm{z}}_i^j + \beta * \mathcal{N}(0, 1)$ (corresponding to a random sample in the neighborhood of ${\bm{z}}_i$) and project it onto the underlying region $\mathcal{R}_i$.
\section{Conclusion}
\label{sect:conclusion}
We have presented GEM\xspace, an approach to learn a modality-independent manifold. We demonstrate how our model enables us to interpolate between signals, complete partial signals, and further generate new signals. A limitation of our approach is that while underlying manifolds are recovered, they are not captured with high-fidelity. We believe that a promising direction of future work involves the pursuit of further engineered inductive biases, and general structure, which enable the recovery of high-fidelity manifolds of the plethora of signals in the world around us. We emphasize, however, that as our manifold is constructed over a dataset; GEM\xspace may learn to incorporate and propogate existing prejudices and biases present in such data, posing risks for mass deployment of our model. Additionally, while our work offers exciting avenues for cross-modal modeling and generation, we note that GEM\xspace has the potential to be used to create enhanced "deep fakes" and other forms of synthetic media.
\section{Related Work}
\myparagraph{Manifold Learning.}
Manifold learning is a large and well-studied topic \citep{journals/neco/ScholkopfSM98, Tenenbaum2000global, Roweis2000Nonlinear, 10.1162/089976603321780317, diffusion_map, Donoho5591, vanDerMaaten2008} which seeks to obtain the underlying non-linear, low-dimensional subspace in which naturally high-dimensional input signals lie.
Many early works in manifold learning utilize a nearest neighbor graph to obtain such a low-dimensional subspace. For instance, Tenenbaum~\textit{et al}.~\cite{Tenenbaum2000global} employs the geodesic distance between nearby points, while Roweis~\textit{et al}.~\cite{Roweis2000Nonlinear} use locally linear subregions around each subspace.
Subsequent work has explored additional objectives \citep{10.1162/089976603321780317, diffusion_map, Donoho5591, vanDerMaaten2008} to obtain underlying non-linear, low-dimensional subspaces.
Recently, \citep{JIA2015250, MISHNE2019259, pai2018dimal} propose to combine traditional manifold learning algorithms with deep networks by training autoencoders to explicitly encourage latents to match embeddings obtained from classical manifold learning methods.
Further, \citep{zhou2020learning} propose to utilize a heat kernel diffusion process to uncover the underlying manifold of data.
In this work, we show how we may combine insights from earlier works in manifold learning with modern deep learning techniques to learn continuous manifolds of data in high-dimensional signal spaces.
Drawing on these two traditions enables our model to smoothly interpolate between samples, restore partial input signals, and further generate new data samples in high-dimensional signal space.
\myparagraph{Learning Distributions of Neural Fields.}
Recent work has demonstrated the potential of treating fully connected networks as continuous, memory-efficient implicit (or coordinate-based) representations for shape parts \cite{genova2019learning,genova2019deep}, objects~\cite{park2019deepsdf,michalkiewicz2019implicit,atzmon2019sal,gropp2020implicit}, or scenes~\cite{sitzmann2019srns,jiang2020local,peng2020convolutional,chabra2020deep}.
Sitzmann et al.~\cite{sitzmann2020implicit} showed that these coordinate-based representations may be leveraged for modeling a wide array of signals, such as audio, video, images, and solutions to partial differential equations.
These representations may be conditioned on auxiliary input via conditioning by concatenation~\cite{park2019deepsdf,mescheder2019occupancy}, hypernetworks~\cite{sitzmann2019srns}, gradient-based meta-learning~\cite{sitzmann2020metasdf}, or activation modulation~\cite{chan2020pi}.
Subsequently, they may be embedded in a generative adversarial~\cite{schwarz2020graf,chan2020pi} or auto-encoder-based framework~\cite{sitzmann2019srns}.
However, both of these approaches require modality-dependent architecture choices. Specifically, modification is necessary in the encoder of auto-encoding frameworks, or the discriminator in adversarial frameworks.
In contrast, we propose to represent the problem of learning a distribution as one of learning the underlying \textit{manifold} in which the signals lie, and show that this enables signal-agnostic modeling.
Concurrent to our work, Dupont et al.~\cite{dupont2021generative} propose a signal-agnostic discriminator based on pointwise convolutions, which enables discrimination of signals that do not lie on regular grids.
In contrast, we do not take an adversarial approach, thereby avoiding the need for a convolution-based discriminator. We further show that our learned manifold better captures the underlying data manifold and demonstrate that our approach is able to capture manifolds across cross-modal distributions, which we find destabilizes training of \citep{dupont2021generative}.
\myparagraph{Generative Modeling.}
Our work is also related to existing work in generative modeling. GANs \citep{Goodfellow2014Generative} are a popular framework used to model modalities such as images \citep{Karras_2020_CVPR} or audio \citep{donahue2019adversarial}, but often suffer from training instability and mode collapse. VAEs \citep{Kingma2014Semi}, another class of generative models, utilize an encoder and decoder to learn a shared latent space for data samples.
Recently autoregressive models have also been used for image \citep{oord2016pixelrec}, audio \citep{oord2016wavenet}, and multi-modal text and image generation \citep{ramesh2021zeroshot}. In contrast to previous generative approaches, which require custom architectures for encoding and decoding, our approach is a \textit{modality-independent} method for generating samples distribution and presents a new view of generative modeling: one of learning an underlying manifold of data.
\section{Introduction}
\label{introduction}
Every moment, we receive and perceive high-dimensional signals from the world around us. These signals are in constant flux; yet remarkably, our perception system is largely invariant to these changes, allowing us to efficiently infer the presence of coherent objects and entities across time. One hypothesis for how we achieve this invariance is that we infer the underlying manifold in which perceptual inputs lie \citep{seung:science00}, naturally enabling us to link high-dimensional perceptual changes with local movements along such a manifold. In this paper, we study how we may learn and discover a low-dimensional manifold in a \textit{signal-agnostic} manner, over arbitrary perceptual inputs.
Manifolds are characterized by three core properties \citep{Tenenbaum2000global, Roweis2000Nonlinear}. First, a manifold should exhibit data coverage, i.e., all instances and variations of a signal are explained in the underlying low-dimensional space. Second, a manifold should be locally metric, enabling perceptual manipulation of a signal by moving around the surrounding low-dimensional space. Finally, the underlying structure of a manifold should be globally-consistent; e.g. similar signals should be embedded close to one another.
Existing approaches to learning generative models, such as GANs \citep{Goodfellow2014Generative}, can be viewed as instances of manifold learning. However, such approaches have two key limitations. First, low-dimensional latent codes learned by generative models do not satisfy all desired properties for a manifold; while the underlying latent space of a GAN enables us to perceptually manipulate a signal, GANs suffer from mode collapse, and the underlying latent space does not cover the entire data distribution. Second, existing generative architectures are biased towards particular signal modalities, requiring custom architectures and losses depending on the domain upon which they are applied -- thereby preventing us from discovering a manifold across arbitrary perceptual inputs in a signal-agnostic manner.
As an example, while existing generative models of images are regularly based on convolutional neural networks, the same architecture in 1D does not readily afford high-quality modeling of audio signals. Rather, generative models for different domains require significant architecture adaptations and tuning.
Existing generative models are further constrained by the common assumption that training data lie on a regular grid, such as grids of pixels for images, or grids of amplitudes for audio signals.
As a result, they require uniformly sampled data, precluding them from adequately modeling irregularly sampled data, like point clouds.
Such a challenge is especially prominent in cross-modal generative modeling, where a system must jointly learn a generative model over multiple signal sources (e.g., over images and associated audio snippets). While this task is trivial for humans -- we can readily recall not only an image of an instrument, but also a notion of its timbre and volume -- existing machine learning models struggle to jointly fit such signals without customization.
\input{figText/teaser_gan_manifold}
To obtain \textit{signal-agnostic} learning of manifolds over signals, we propose to model distributions of signals in the function space of neural fields, which are capable of parameterizing image, audio, shape, and audiovisual signals in a modality-independent manner.
We then utilize hypernetworks \citep{ha2016hypernetworks}, to regress individual neural fields from an underlying latent space to represent a signal distribution.
To further ensure that our distribution over signals corresponds to a manifold, we formulate our learning objective with explicit losses to encourage the three desired properties for a manifold: data coverage, local linearity, and global consistency.
The resulting model, which we dub GEM\xspace{}\footnote{short for GEnerative Manifold learning}, enables us to capture the manifold of a variety of signals -- ranging from audio to images and 3D shapes -- with almost no architectural modification, as illustrated in \fig{fig:teaser}.
We further demonstrate that our approach reliably recovers distributions over cross-modal signals, such as images with a correlated audio snippet, while also eliciting sample diversity.
We contribute the following: first, we present GEM\xspace, which we show can learn manifolds over images, 3D shapes, audio, and cross-modal audiovisual signals in a signal-agnostic manner. Second, we demonstrate that our model recovers the global structure of each signal domain, permitting easy interpolation between nearby signals, as well as completion of partial inputs. Finally, we show that walking along our learned manifold enables us to generate new samples for each modality.
\section{Learning Manifolds of Data}
\input{tbl/psnr_fit}
We validate the generality of GEM\xspace{} by first showing that our model is capable of fitting diverse signal modalities.
Next, we demonstrate that our approach captures the underlying structure across these signals; we are not only able to cluster and perceptually interpolate between signals, but inpaint to complete partial ones.
Finally, we show that we can draw samples from the learned manifold of each signal type, illustrating the power of GEM\xspace{} to be used a signal-agnostic generative model.
We re-iterate that \textit{nearly identical} architectures and training losses are used across separate modalities.
\myparagraph{Datasets.} We evaluate GEM\xspace on four signal modalities: image, audio, 3D shape, and cross-modal image and audio signals, respectively. For the image modality, we investigate performance on the CelebA-HQ dataset \citep{Karras2017Progressive} fit on 29000 $64\times64$ training celebrity images, and test on 1000 $64\times64$ test images. To study GEM\xspace behavior on audio signals, we use the NSynth dataset \citep{nsynth2017}, and fit on a training set of 10000 one-second 16kHz sounds clips of different instruments playing, and test of 5000 one-second 16kHz sound clips. We process sound clips into spectrograms following \citep{donahue2019adversarial}. For the 3D shape domain, we work with the ShapeNet dataset from \citep{chen2019learning}. We train on 35019 training shapes at $64\times64\times64$ resolution and test on 8762 shapes at $64\times64\times64$ resolution. Finally, for the cross-modal image and audio modality, we utilize the cello image and audio recordings from the Sub-URMP dataset \citep{chen2017deep}. We train on 9800 images at $128\times128$ resolution and 0.5 second 16kHz audio clips, which are also processed into spectrograms following \citep{donahue2019adversarial} and test on 1080 images and associated audio clips. We provide additional cross-modal results in the supplement.
\myparagraph{Setup and Baselines.} We benchmark GEM\xspace against existing manifold learning approaches. Namely, we compare our approach with that of StyleGAN2 \citep{Karras_2020_CVPR}, VAE \citep{Kingma2014Auto}, and concurrent work FDN of \citep{dupont2021generative}, as well as heat kernel manifold learning \citep{zhou2020learning}. We utilize the authors' provided codebase for StyleGAN2, the PytorchVAE library \footnote{https://github.com/AntixK/PyTorch-VAE} for the VAE, and the original codebase of FDN \citep{dupont2021generative}, which the authors graciously provided. We train all approaches with a latent dimension of 1024, and re-scale the size of the VAE to ensure parameter counts are similar. We report model architecture details in the appendix. We need to change the architecture of both the VAE and StyleGAN2 baselines, respectively, in order to preserve proper output dimensions, depending on signal modality. However, we note that identical architectures readily can and are used for GEM\xspace. For fitting cross-modal audiovisual signals, all models utilize two separate decoders, with each decoder in GEM\xspace identical to each other. Architectural details are provided in the appendix.
\input{figText/nsynth}
\subsection{Covering the Data Manifold}
\label{sect:data_coverage}
We first address whether our model is capable of covering the underlying data distribution of each modality, compared against our baselines, and study the impact of our losses on this ability.
\myparagraph{Modality Fitting.} We quantitatively measure reconstruction performance on test signals in \fig{fig:reconstruct_multimodal} and find the GEM\xspace outperforms each of our baselines. In the cross-modal domain in particular, we find that FDN training destabilizes quickly and struggles to adequately fit examples. We illustrate test signal reconstructions in \fig{fig:reconstruct_multimodal} and \fig{fig:nsynth}. Across each modality, we find that GEM\xspace{} obtains the sharpest samples. In \fig{fig:reconstruct_multimodal} we find that while StyleGAN2 can reconstruct certain CelebA-HQ images well, it fails badly on others, indicating a lack of data coverage. Our results indicate that our approach best captures the underlying data distribution.
\input{figText/interp}
\myparagraph{Ablations.} In \tbl{tbl:ablations}, we assess the impact of two of our proposed losses -- $\mathcal{L}_{\text{LLE}}$ and $\mathcal{L}_{\text{Iso}}$ -- at enabling test-signal reconstruction, and find to both help improve the resultant test construction. We posit that both losses serve to regularize the underlying learned data manifold, enabling GEM\xspace to more effectively cover the overall signal manifold.
\myparagraph{Applications to Other Models.} While we apply $\mathcal{L}_{\text{Rec}}$ , $\mathcal{L}_{\text{LLE}}$, $\mathcal{L}_{\text{Iso}}$ to GEM\xspace, our overall losses may be applied more generally across different models, provided the model maps an underlying latent space to an output domain. We further assess whether these losses improve the manifold captured by StyleGAN2 via measuring reconstruction performance on test CelebA-HQ images. We find that while a base StyleGAN2 model obtains a reconstruction error of MSE 0.0044 and PSNR 24.03, the addition of $\mathcal{L}_{\text{Rec}}$ improves test reconstruction to MSE 0.0041 and PSNR 24.29 with $\mathcal{L}_{\text{LLE}}$ losses further improving test reconstruction to MSE 0.0038 and PSNR 24.61. In contrast, we found that $\mathcal{L}_{\text{Iso}}$ leads to a slight regression in performance. We posit that both $\mathcal{L}_{\text{Rec}}$ and $\mathcal{L}_{\text{LLE}}$ serve to structure the underlying StyleGAN2 latent space, while the underlying Gaussian nature of the StyleGAN2 latent space precludes the need for $\mathcal{L}_{\text{Iso}}$. Our results indicate the generality of our proposed losses towards improving the recovery of data manifolds.
\subsection{Learning the Structure of Signals}
\label{sect:eval_structure}
Next, we explore the extent to which the manifold learned by GEM\xspace{} captures the underlying global \textit{and} local structure inherent to each signal modality.
Additionally, we probe our model's understanding of individual signals by studying GEM\xspace's ability to reconstruct partial signals.
\myparagraph{Global Structure.} We address whether GEM\xspace captures the global structure of a signal distribution by visualizing the latents learned by GEM\xspace on the NSynth dataset with t-SNE \citep{VanderMaaten2008Visualizing}. We find in \fig{fig:tsne_pitch} that the underlying pitch of a signal maps onto the resulting t-SNE. We further visualize the inferred connectivity structure of our manifold in \fig{fig:celebahq_completion} (left) and visualize the nearest neighbors in latent space of the autodecoded latents (right), finding that nearby latents correspond to semantically similar faces -- further supporting that GEM\xspace learns global structure. Note, we find that without either $\mathcal{L}_{\text{LLE}}$ or $\mathcal{L}_{\text{Iso}}$, the underlying connectivity of our manifold is significantly poorer.
\input{figText/celebahq_completions}
\input{figText/audiovis_completion}
\myparagraph{Local Structure.} We next probe whether GEM\xspace learns a densely connected manifold, e.g. one which allows us to interpolate between separate signals, suggestive of capturing the local manifold structure. In \fig{fig:interp-nearneighbor}, we visualize nearest neighbor interpolations in our manifold and observe that the combination of $\mathcal{L}_{\text{LLE}}$ and $\mathcal{L}_{\text{Iso}}$ enables us to obtain perceptually smooth interpolations over both image and shape samples. We provide additional interpolations and failure cases in the supplement.
\myparagraph{Signal Completion.} To investigate the ability of GEM\xspace to understand individual signals, we assess the recovery of a full signal when only a subset of such a signal is given. In \fig{fig:celebahq_completion}, we consider the case of inpainting CelebA-HQ faces, where a subset of the face is missing. We observe that GEM\xspace is able to obtain several different perceptually consistent inpaintings of a face, with individual completions exhibiting consistent skin and lip colors, and restoring structural features such as eyeglasses. We provide additional examples and failure cases in the supplement.
Additionally, we consider signal completion in the cross-modal audiovisual domain. Here, we provide only an input image and ask GEM\xspace to generate possible audio spectograms. As seen in \fig{fig:audiovis_completion}, we observe that GEM\xspace is able to generate empty spectograms when the bow is not on the cello, as well as different spectrograms dependent on the position of the bow. Alternatively, in \fig{fig:audiovis_completion}, we provide only an input audio spectrogram and ask GEM\xspace to generate possible images. In this setting, we find that generated images have the bow off the cello when the input spectrogram is empty (with the bow missing due to GEM\xspace blurring the uncertainty of all possible locations of the bow), and at different positions depending on the input audio spectrogram. In contrast, we found that all baselines were unable to fit the audiovisual setting well, as discussed in \sect{sect:data_coverage} and shown in \fig{fig:uncond-gen-multimodal}.
\subsection{Generating Data Samples}
\input{tbl/image_generative}
Finally, we investigate the ability of GEM\xspace{} to generate new data samples. %
\input{figText/unconditional-gen-multimodal}
\myparagraph{Qualitative Generations.} We show random samples drawn from GEM\xspace and compare these to our baselines. We consider images in \fig{fig:celebahq_gen}, shapes in \fig{fig:uncond-gen-multimodal}, audio snippets in \fig{fig:nsynth_gen} and audiovisual inputs in \fig{fig:uncond-gen-multimodal}. While we find that GEM\xspace performs comparably to FDN in the image regime, our model significantly outperforms all baselines on domains of audio, shape, and audiovisual modalities. We further display nearest latent space neighbors of generations on CelebA-HQ in \fig{fig:celebahq_gen} and find that our generations are distinct from those in the training dataset.
\myparagraph{Quantitative Comparisons.} Next, we provide quantitiative evaluations of generations in \tbl{tbl:main_quant} on image and shape modalities. We report the FID \citep{heusel2017gans}, precision, and recall \citep{Kynkaanniemi2019} metrics on CelebA-HQ $64\times64$. We find that GEM\xspace performs better than StyleGAN2 and FDN on precision and recall, but worse in terms of FID. We note that our qualitative generations in \fig{fig:celebahq_gen} are comparable to those of FDN; however, we find that our approach obtains high FID scores due to the inherent sensitivity of the FID metric to blur. Such sensitivity to bluriness has also been noted in \citep{razavi2019generating}, and we find that even our autodecoded training distribution obtains an FID of 23.25, despite images appearing near perceptually perfect (\fig{fig:reconstruct_multimodal}). On ShapeNet, we report coverage and MMD metrics from \citep{achlioptas2017latent_pc} using Chamfer Distance. To evaluate generations, we sample 8762 shapes (the size of test Shapenet dataset) and generate 2048 points following \citep{chen2019learning}. We compare generations from GEM\xspace with those of FDN and latent GAN trained on IMNET (using the provided code in \citep{chen2019learning}). We find that GEM\xspace outperforms both approaches on the task of 3D shape generation.
\section{Appendix for Learning Signal-Agnostic Manifolds of Neural Fields}
Please visit our project website at \url{https://yilundu.github.io/gem/} for additional qualitative visualizations of test-time reconstruction of audio and audiovisual samples, traversals along the underlying manifold of GEM\xspace on CelebA-HQ as well as interpolations between audio samples. We further illustrate additional image in-painting results, as well as audio completion results. Finally, we visualize several audio and audiovisual generations.
In \sect{sect:experimental} below, we provide details on training settings, as well as the underlying baseline model architectures utilized for each modality. We conclude with details on reproducing our work in \sect{sect:reproducibility}.
\subsection{Experimental Details}
\label{sect:experimental}
\myparagraph{Training Details} For each separate training modality, all models and baselines are trained for one day, using one 32GB Volta machine. GEM\xspace is trained with the Adam optimizer \citep{Kingma2015Adam}, using a training batch size of 128 and a learning rate of 1e-4. Each individual datapoint is fit by fitting the value of 1024 sampled points in the sample (1024 for each modality in the multi-modal setting). We normalize the values of a signals to be between -1 and 1. When computing $\mathcal{L}_{\text{Iso}}$, a scalar constant of $\alpha=100$ is employed to scale distances in the underlying manifold to that of distances of signals in sample space. When enforcing $\mathcal{L}_{\text{LLE}}$, a total of 10 neighbors are considered to compute the loss across modalities. We utilize equal loss weight across $\mathcal{L}_{\text{Rec}}$, $\mathcal{L}_{\text{Iso}}$, $\mathcal{L}_{\text{LLE}}$, and found that the relative magnitudes of each loss had little impact on the overall performance.
\myparagraph{Model Details} We provide the architectures of the hypernetwork $\psi$ and implicit function $\phi$ utilized by GEM\xspace across separate modalities in \tbl{tbl:hyper} and \tbl{tbl:implicit}, respectively. Additionally, we provide the architectures used in each domain for our baselines: StyleGAN2 in \tbl{tbl:stylegan} each domain in \tbl{tbl:stylegan} and VAE in \tbl{tbl:vae}. Note that for the VAE, the ReLU nonlinearity is used, with each separate convolution having stride 2.
We obtained the hyperparameters for implicit functions and hypernetworks based off of \citep{inr_gan}. Early in the implementation of the project, we explored a variety of additional architectural choices; however, we ultimately found that neither changing the number of layers in the hypernetworks, nor changing the number of underlying hidden units in networks, significantly impacted the performance of GEM. We will add these details to the appendix of the paper.
\subsection{Reproducibility}
\label{sect:reproducibility}
We next describe details necessary to reproduce each of other underlying empirical results.
\myparagraph{Hyperparameter Settings for Baselines} We employ the default hyperparameters, as used in the original papers for StyleGAN2 \citep{Karras_2020_CVPR} and FDN \citep{dupont2021generative}, to obtain state-of-the-art performance on their respective tasks. Due to computational constraints, we were unfortunately unable to do a complete hyperparameter search for each method over all tasks considered. Despite this, we were able to run the models on toy datasets and found that these default hyperparameters performed the best. We utilized the author’s original codebases for experiments.
\myparagraph{Variance Across Seeds} Results in the main tables of the paper are run across a single evaluated seed. Below in \tbl{tbl:main_confidence}, we rerun test reconstruction results on CelebA-HQ across different models utilizing a total of 3 separate seeds. We find minimal variance across separate runs, and still find the GEM\xspace performs significantly outperforms baselines.
\begin{table}[h]
\centering
\scalebox{0.8}{
\begin{tabular}{l|l|cc}
\toprule
Modality & Model & MSE $\downarrow$ & PSNR $\uparrow$ \\
\midrule
\multirow{4}{*}{Images} & VAE & 0.0327 $\pm$ 0.0035 & 15.16 $\pm$ 0.06 \\
& FDN & 0.0062 $\pm$ 0.0003 & 22.57 $\pm$ 0.02 \\
& StyleGANv2 & 0.0044 $\pm$ 0.0001 & 24.03 $\pm$ 0.01\\
& GEM\xspace & \textbf{0.0025} $\pm$ 0.0001 & \textbf{26.53} $\pm$ 0.01\\
\bottomrule
\end{tabular}
}
\vspace{5pt}
\caption{Test CelebA-HQ reconstruction results of different methods evaluated across 3 different seeds. We further report standard deviation between different runs.}
\label{tbl:main_confidence}
\end{table}
\myparagraph{Datasets} We provide source locations to download each of the datasets we used in the paper. The CelebA-HQ dataset can be downloaded at \url{https://github.com/tkarras/progressive_growing_of_gans/blob/master/dataset_tool.py} and is released under the Creative Commons license. The NSynth dataset may be downloaded at \url{https://magenta.tensorflow.org/datasets/nsynth} and is released under the Creative Commons license. The ShapeNet dataset can be downloaded at \url{https://github.com/czq142857/IM-NET} and is released under the MIT License, and finally the Sub-URMP dataset we used may be downloaded at \url{https://www.cs.rochester.edu/~cxu22/d/vagan/}.
\begin{figure}[H]
\centering
\begin{minipage}{0.45\linewidth}
\centering
\scalebox{0.8}{
\begin{tabular}{c}
\toprule
\toprule
Dense $\rightarrow$ 512\\
\midrule
Dense $\rightarrow$ 512\\
\midrule
Dense $\rightarrow$ 512\\
\midrule
Dense $\rightarrow$ $\phi$ Parameters \\
\bottomrule
\end{tabular}
}
\captionof{table}{The architecture of the hypernetwork utilized by GEM\xspace.}
\label{tbl:hyper}
\end{minipage}%
\hfill
\begin{minipage}{0.45\linewidth}
\centering
\scalebox{0.8}{
\begin{tabular}{c}
\toprule
\toprule
Pos Embed (512)\\
\midrule
Dense $\rightarrow$ 512\\
\midrule
Dense $\rightarrow$ 512\\
\midrule
Dense $\rightarrow$ 512 \\
\midrule
Dense $\rightarrow$ Output Dim \\
\bottomrule
\end{tabular}
}
\captionof{table}{The architecture of the implicit function $\phi$ used to agnostically encode each modality. We utilize the Fourier embedding from \citep{mildenhall2020nerf} to embed coordinates.}
\label{tbl:implicit}
\end{minipage}
\end{figure}
\begin{figure}[H]
\centering
\begin{minipage}{0.22\linewidth}
\centering
\resizebox{\columnwidth}{!}{%
\begin{tabular}{c}
\toprule
\toprule
3x3 Conv2d, 64\\
\midrule
3x3 Conv2d, 128\\
\midrule
3x3 Conv2d, 256\\
\midrule
3x3 Conv2d, 512 \\
\midrule
3x3 Conv2d, 512 \\
\midrule
z $\leftarrow$ Encode \\
\midrule
Reshape(2, 2) \\
\midrule
3x3 Conv2d Transpose, 512\\
\midrule
3x3 Conv2d Transpose, 512\\
\midrule
3x3 Conv2d Transpose, 256\\
\midrule
3x3 Conv2d Transpose, 128\\
\midrule
3x3 Conv2d Transpose, 3\\
\bottomrule
\end{tabular}
}
\captionof{table}{The encoder and decoder of the VAE utilized for CelebA-HQ.}
\label{tbl:vae_celebahq}
\end{minipage}%
\hfill
\begin{minipage}{0.22\linewidth}
\centering
\resizebox{\columnwidth}{!}{%
\begin{tabular}{c}
\toprule
\toprule
3x3 Conv2d, 32\\
\midrule
3x3 Conv2d, 64\\
\midrule
3x3 Conv2d, 128\\
\midrule
3x3 Conv2d, 256 \\
\midrule
3x3 Conv2d, 512 \\
\midrule
z $\leftarrow$ Encode \\
\midrule
Reshape(4, 2) \\
\midrule
3x3 Conv2d Transpose, 512\\
\midrule
3x3 Conv2d Transpose, 256\\
\midrule
3x3 Conv2d Transpose, 128\\
\midrule
3x3 Conv2d Transpose, 64\\
\midrule
3x3 Conv2d Transpose, 32\\
\midrule
3x3 Conv2d Transpose, 1\\
\midrule
Crop \\
\bottomrule
\end{tabular}
}
\captionof{table}{The encoder and decoder of the VAE utilized for NSynth}
\label{tbl:vae_nsynth}
\end{minipage}%
\hfill
\begin{minipage}{0.22\linewidth}
\centering
\resizebox{\columnwidth}{!}{%
\begin{tabular}{c}
\toprule
\toprule
3x3 Conv2d, 32\\
\midrule
3x3 Conv2d, 64\\
\midrule
3x3 Conv2d, 128\\
\midrule
3x3 Conv2d, 256 \\
\midrule
3x3 Conv2d, 512 \\
\midrule
z $\leftarrow$ Encode \\
\midrule
Reshape(2, 2) \\
\midrule
3x3 Conv2d Transpose, 512\\
\midrule
3x3 Conv2d Transpose, 256\\
\midrule
3x3 Conv2d Transpose, 128\\
\midrule
3x3 Conv2d Transpose, 64\\
\midrule
3x3 Conv2d Transpose, 32\\
\midrule
3x3 Conv2d Transpose, 3\\
\bottomrule
\end{tabular}
}
\captionof{table}{The architecture of encoder and decoder of the VAE utilized for audiovisual dataset on images. Latent encodings from image and audio modalities are added together.}
\label{tbl:vae_audiovis_audio}
\end{minipage}%
\hfill
\begin{minipage}{0.22\linewidth}
\centering
\resizebox{\columnwidth}{!}{%
\begin{tabular}{c}
\toprule
\toprule
3x3 Conv2d, 32\\
\midrule
3x3 Conv2d, 64\\
\midrule
3x3 Conv2d, 128\\
\midrule
3x3 Conv2d, 256 \\
\midrule
3x3 Conv2d, 512 \\
\midrule
z $\leftarrow$ Encode \\
\midrule
Reshape(4, 1) \\
\midrule
3x3 Conv2d Transpose, 512\\
\midrule
3x3 Conv2d Transpose, 256\\
\midrule
3x3 Conv2d Transpose, 128\\
\midrule
3x3 Conv2d Transpose, 64\\
\midrule
3x3 Conv2d Transpose, 32\\
\midrule
3x3 Conv2d Transpose, 1\\
\midrule
Crop \\
\bottomrule
\end{tabular}
}
\captionof{table}{The architecture of encoder and decoder of the VAE utilized for audiovisual dataset on audio. Latent encodings from image and audio modalities are added together.}
\label{tbl:vae_audiovis_vision}
\end{minipage}
\captionof{table}{The architecture of the VAE utilized across datasets.}
\label{tbl:vae}
\end{figure}
\begin{figure}[H]
\centering
\begin{minipage}{0.22\linewidth}
\centering
\resizebox{\columnwidth}{!}{%
\begin{tabular}{c}
\toprule
\toprule
Constant Input (512, 4, 4)\\
\midrule
StyleConv 512\\
\midrule
StyleConv 512\\
\midrule
StyleConv 512 \\
\midrule
StyleConv 512 \\
\midrule
StyleConv 256 \\
\midrule
3x3 Conv2d, 3 \\
\bottomrule
\end{tabular}
}
\captionof{table}{The generator architecture of StyleGAN2 for CelebA-HQ.}
\label{tbl:stylegan_celebahq}
\end{minipage}%
\hfill
\begin{minipage}{0.22\linewidth}
\centering
\resizebox{\columnwidth}{!}{%
\begin{tabular}{c}
\toprule
\toprule
Constant Input (512, 8, 4)\\
\midrule
StyleConv 512\\
\midrule
StyleConv 512\\
\midrule
StyleConv 512 \\
\midrule
StyleConv 512 \\
\midrule
StyleConv 256 \\
\midrule
3x3 Conv2d, 1 \\
\midrule
Crop \\
\bottomrule
\end{tabular}
}
\captionof{table}{The generator architecture of StyleGAN2 for NSynth}
\label{tbl:stylegan_nsynth}
\end{minipage}%
\hfill
\begin{minipage}{0.22\linewidth}
\centering
\resizebox{\columnwidth}{!}{%
\begin{tabular}{c}
\toprule
\toprule
Constant Input (512, 4, 4)\\
\midrule
StyleConv 512\\
\midrule
StyleConv 512\\
\midrule
StyleConv 512 \\
\midrule
StyleConv 512 \\
\midrule
StyleConv 256 \\
\midrule
3x3 Conv2d, 3 \\
\bottomrule
\end{tabular}
}
\captionof{table}{The generator architecture of StyleGAN2 for audiovisual domain for images.}
\label{tbl:stylegan_audiovis_audio}
\end{minipage}%
\hfill
\begin{minipage}{0.22\linewidth}
\centering
\resizebox{\columnwidth}{!}{%
\begin{tabular}{c}
\toprule
\toprule
Constant Input (512, 8, 2)\\
\midrule
StyleConv 512\\
\midrule
StyleConv 512\\
\midrule
StyleConv 512 \\
\midrule
StyleConv 512 \\
\midrule
StyleConv 256 \\
\midrule
3x3 Conv2d, 1 \\
\midrule
Crop \\
\bottomrule
\end{tabular}
}
\captionof{table}{The generator architecture of StyleGAN2 for audiovisual domain for audio.}
\label{tbl:stylegan_audiovis_vision}
\end{minipage}
\captionof{table}{The architecture of the StyleGAN generator utilized across datasets.}
\label{tbl:stylegan}
\end{figure}
|
\section{Introduction}\label{sec:intro}
Stochastic approximation (SA) algorithms are the major work-horses for solving large-scale optimization and machine learning problems. Theoretically, to achieve asymptotic convergence, we should use diminishing stepsizes (learning rates) with proper decay rate \citep{nemirovski2009robust}. However, constant stepsize SA algorithms are preferred in practice due to their faster convergence \citep{goodfellow2016deep}. In that case, instead of converging asymptotically to the desired solution, the iterates of constant stepsize SA have a stationary distribution. Although in many cases such weak convergence to a stationary distribution was established in the literature, it is not possible to fully characterize the limiting distribution. The reason is that, when constant stepsize is used, the distribution of the noise sequence within the SA algorithm plays an important role in the stationary distribution of the iterates. Since the distribution of the noise is in general unknown, the stationary distribution cannot be analytically characterized. In this work, building upon the works on stationary distribution of constant stepsize SA algorithms, we aim at understanding the limiting behavior of the properly scaled stationary distribution as the constant stepsize goes to zero.
More formally, with initialization $X_0^{(\alpha)}\in\mathbb{R}^d$, consider the SA algorithm
\begin{align}\label{algo:sa}
X_{k+1}^{(\alpha)}=X_k^{(\alpha)}+\alpha\left(F(X_k^{(\alpha)})+w_k\right),
\end{align}
where $F:\mathbb{R}^d\mapsto\mathbb{R}^d$ is a general nonlinear operator, $\alpha$ is the constant stepsize, and $\{w_k\}$ is the noise sequence. Observe that SA algorithm (\ref{algo:sa}) can be viewed as an iterative algorithm for solving the equation $F(x)=0$ in the presence of noise \citep{robbins1951stochastic}. A typical example is when $F(x)=-c \nabla f(x)$ (where $c>0$ is a constant) for some objective function $f(\cdot)$, in this case Algorithm (\ref{algo:sa}) becomes the popular stochastic gradient descent (SGD) algorithm for minimizing $f(\cdot)$ \citep{lan2020first,bottou2018optimization}. Another example lies in the context of reinforcement learning, where $F(x)=\mathcal{T}(x)-x$, and $\mathcal{T}(\cdot)$ is the Bellman operator \citep{sutton2018reinforcement}. In this case, Algorithm (\ref{algo:sa}) captures popular reinforcement learning algorithms such as TD-learning \citep{sutton1988learning} and $Q$-learning \citep{watkins1992q}.
Under some mild conditions on the operator $F(\cdot)$, it was shown in the literature that the sequence $\{X_k^{(\alpha)}\}$ converges weakly to some random variable $X^{(\alpha)}$ \citep{dieuleveut2017bridging, bianchi2020convergence,yu2020analysis, durmus2021riemannian}. However, for a fixed $\alpha$, it is not possible to fully characterize the distribution of $X^{(\alpha)}$ because it depends on the distribution of the noise sequence $\{w_k\}$, which is usually unknown. In this work, we further consider letting $\alpha$ go to zero, and study the distribution of a properly centered and scaled iterate. Specifically, let $Y_k^{(\alpha)}:=(X_k^{(\alpha)}-x^*)/g(\alpha)$, where $x^*$ is the solution of $F(x)=0$, and $g:\mathbb{R}\mapsto\mathbb{R}$ is a properly chosen scaling function\footnote{The scaling function is unique up to a numerical factor}. When $k$ goes to infinity, we expect that $Y_k^{(\alpha)}$ converges weakly to some random variable $Y^{(\alpha)}$. Then we let $\alpha$ go to zero, and our goal is to further characterize the weak limit of $Y^{(\alpha)}$. Notice that proper scaling of the iterates is essential for raveling its fine grade behavior because otherwise the limiting distribution of the un-scaled iterates will converge to a singleton as the stepsize $\alpha$
goes to zero, which is analogous to the almost sure convergence result for using diminishing stepsizes in SA literature.
To summarize, we want to find a suitable scaling function $g(\cdot)$ and characterize the following two-step weak convergence of the centered scaled iterate $Y^{(\alpha)}_k=(X^{(\alpha)}_k-x^*)/g(\alpha)$:
\begin{align}\label{eq:two_limit}
Y_k^{(\alpha)}\overset{k\rightarrow\infty}{\Longrightarrow} Y^{(\alpha)}\overset{\alpha\rightarrow 0}{\Longrightarrow}Y,
\end{align}
where we use the notation $\Rightarrow$ for weak convergence (or convergence in distribution).
\subsection{Main Contributions}\label{subsec:contribution}
Our main contributions are twofold.
\textit{Characterizing the Distribution of $Y$.}
We propose a general framework for characterizing the distribution of $Y$ in the following 3 cases:
(1) SGD with a smooth and strongly convex objective, (2) linear SA with a Hurwitz matrix, and (3) SA involving a contractive operator. In particular, we show that in all three cases above the correct scaling function is $g(\alpha)=\sqrt{\alpha}$, and the distribution of $Y$ is Gaussian with mean zero and covariance matrix being the unique solution of an appropriate Lyapunov equation. Our proof is to use the characteristic function as a test function to obtain an integral equation of the distribution of $Y$, and then show that the desired Gaussian distribution solves the integral equation.
\textit{Determining the Suitable Scaling Function.} For more general SA algorithms, we show empirically that the scaling function need not be $g(\alpha)=\sqrt{\alpha}$ and the distribution of $Y$ need not be Gaussian. Inspired by this observation, we propose a method to find the the correct scaling function for general SA algorithms. In particular, our results indicate that the scaling function $g(\alpha)$ should be chosen such that (1) $\lim_{\alpha\rightarrow 0}\frac{\alpha}{g(\alpha)}=0$, and (2) the function $\Tilde{F}(\cdot)$ defined by $\Tilde{F}(y)=\lim_{\alpha\rightarrow 0}\frac{g(\alpha)F(yg(\alpha)+x^*)}{\alpha}$ is non-trivial in the sense that it is not identically zero or infinity. Our proposed condition is verified in numerical experiments. Moreover, we make an insightful connection between the choice of $g(\alpha)$ and the Euler-Maruyama discretization scheme for approximating a stochastic differential equation (SDE) -- Langevin diffusion \citep{sauer2012numerical}.
\subsection{An Illustrative Example}\label{subsec:illustrative}
In this section, we provide an example to further illustrate the problem we are going to study.
Consider SA algorithm (\ref{algo:sa}). Suppose that $F(x)=-x$ is a scalar valued function, and $\{w_k\}$ is a sequence of i.i.d. standard normal random variables. We make such noise assumption here only for ease of exposition, and it will be relaxed in later sections. In this case, the SA algorithm (\ref{algo:sa}) becomes
\begin{align}\label{eq:sa1}
X_{k+1}^{(\alpha)}=(1-\alpha)X_k^{(\alpha)}+\alpha w_k.
\end{align}
This algorithm has the following two simple interpretations: (1) it can be viewed as the SGD algorithm for minimizing a quadratic objective function $f(x)=x^2/2$, which has a unique minimizer $x^*=0$, (2) it can also be viewed as an SA algorithm for solving the fixed-point equation $\mathcal{T}(x)=x$ with $\mathcal{T}(x)$ being identically equal to zero, therefore $x^*=0$ is the unique fixed-point.
Since $x^*=0$, let $Y_k^{(\alpha)}=X_k^{(\alpha)}/\sqrt{\alpha}$ be the centered scaled iterate. To obtain an update equation for $Y_k^{(\alpha)}$, dividing both sides of Eq. (\ref{eq:sa1}) by $\sqrt{\alpha}$ and we have for all $k\geq 0$:
\begin{align*}
Y_k^{(\alpha)}&=(1-\alpha)Y_{k-1}^{(\alpha)}+\sqrt{\alpha} w_{k-1}\\
&=(1-\alpha)^2Y_{k-2}^{(\alpha)}+(1-\alpha)\sqrt{\alpha} w_{k-2}+\sqrt{\alpha} w_{k-1}\\
&=\cdots\\
&=(1-\alpha)^kY_0^{(\alpha)}+\sum_{i=0}^{k-1}(1-\alpha)^{k-i-1}\sqrt{\alpha} w_i.
\end{align*}
Since $Y_k^{(\alpha)}$ is a linear combination of independent Gaussian random variables, $Y_k^{(\alpha)}$ itself is also Gaussian. Since
\begin{align*}
\mathbb{E}[Y_k^{(\alpha)}]&=(1-\alpha)^kY_0^{(\alpha)}+\sum_{i=0}^{k-1}(1-\alpha)^{k-i-1}\sqrt{\alpha} \mathbb{E}[w_i]=(1-\alpha)^kY_0^{(\alpha)},
\end{align*}
and
\begin{align*}
\mathbb{V}[Y_k^{(\alpha)}]&=\mathbb{V}\left[(1-\alpha)^kY_0^{(\alpha)}+\sum_{i=0}^{k-1}(1-\alpha)^{k-i-1}\sqrt{\alpha} w_i\right]=\frac{1}{2-\alpha}\left(1-(1-\alpha)^{2k}\right),
\end{align*}
where $\mathbb{E}[\;\cdot\;]$ and $\mathbb{V}[\;\cdot\;]$ denote the mean and variance, respectively,
we have $\lim_{k\rightarrow \infty}\mathbb{E}[Y_k^{(\alpha)}]=0$ and $\lim_{k\rightarrow \infty}\mathbb{V}[Y_k^{(\alpha)}]=\frac{1}{2-\alpha}$. Therefore, the sequence $Y_k^{(\alpha)}$ converges weakly to a random variable $Y^{(\alpha)}$, whose distribution is $ \mathcal{N}(0,\frac{1}{2-\alpha})$. In this case, we are able to analytically characterize the distribution of $Y^{(\alpha)}$ for a fixed $\alpha$ because of the simplicity of the algorithm (\ref{eq:sa1}) and the noise $\{w_k\}$ being i.i.d. Gaussian. For the general SA algorithm (\ref{algo:sa}) with limited information on the noise sequence $\{w_k\}$, it is in general not possible to fully characterize the distribution of $Y^{(\alpha)}$.
Now consider the second weak convergence in Eq. (\ref{eq:two_limit}). Since we have already shown that $Y^{(\alpha)}\sim \mathcal{N}(0,\frac{1}{2-\alpha})$. As $\alpha$ goes to zero, we see that $Y^{(\alpha)}$ converges weakly to a random variable $Y$, whose distribution is $\mathcal{N}(0,\frac{1}{2})$. As opposed to the first weak convergence in Eq. (\ref{eq:two_limit}), where the distribution of $Y^{(\alpha)}$ in general cannot be fully characterized, we are able to characterize (in later sections) the distribution of $Y$ for the more general algorithm (\ref{algo:sa}), and for more general noise assumptions. Intuitively, the reason is that as the constant stepsize decreases, the
effect of the entire distribution of the noise $\{w_k\}$ on the distribution of $Y^\alpha$ is weakened. In the limit, only the mean and the variance of $w_k$ play roles in determining the distribution of $Y$. This is analogous to central limit theorem type of results.
To summarize, we have shown in the special case of Algorithm (\ref{eq:sa1}) that the correct scaling function is $g(\alpha)=\sqrt{\alpha}$, and the distribution of the limiting random variable $Y$ is a Gaussian distribution with mean zero and variance $1/\sqrt{2}$. In Section \ref{sec:distribution}, we extend this result to the more general algorithm (\ref{algo:sa}) with weaker noise assumptions.
\subsection{Related Literature}\label{subsec:literature}
Since proposed in \citep{robbins1951stochastic}, SA has been popular for solving large scale optimization in modern machine learning applications. Although require using diminishing stepsizes to achieve convergence \citep{hu2017diffusion, xie2020linear, mertikopoulos2020almost,shamir2013stochastic, li2019convergence,fehrman2020convergence, gower2021sgd}, constant stepsize is used in practice \citep{goodfellow2016deep}. In contrast to the success in machine learning practice, there is little discussion about the stationary distribution of constant step size SGD \citep{dieuleveut2017bridging, bianchi2020convergence,yu2020analysis, durmus2021riemannian}. Among them, \cite{dieuleveut2017bridging} bridges Markov chain theory and the constant step size SGD algorithm. They provided an explicit asymptotic expansion of the moments of the averaged SGD iterates. \cite{bianchi2020convergence} studied the asymptotic behavior of constant step size SGD for a nonconvex nonsmooth, locally Lipchitz objective function. It was shown that in a small step size regime, the interpolated trajectory of the algorithm converges in probability towards the solutions of the differential inclusion $\dot{x}=\partial F(x)$ and the invariant distribution of the corresponding Markov chain converges weakly to the set of invariant distribution of the differential inclusion. \cite{yu2020analysis} established an asymptotic normality result for the constant step size SGD algorithm for a non-convex and non-smooth objective function satisfying a dissipativity property. \cite{durmus2021riemannian} first established non-asymptotic performance bounds under Lyapunov conditions and then proved that for any step size, the corresponding Markov chain admits a unique stationary distribution.
The work mentioned before established the stationary distribution for almost strongly convex and smooth functions for a fixed constant stepsize. Since the SGD iterates will converge to a singleton as the constant step size goes to zero, none of the previously mentioned literates can be applied to study the limiting behavior of SGD in this regime. To understand such behavior, we propose to study the properly centered and scaled iterates. Although not directly related, it shares a similar flavor when studying the limiting behavior of the stationary distribution of the stochastic gradient Langevin dynamics (SGLD) iterates as step size goes to zero.
Another set of related literature is on the diffusion approximation of SGD \citep{li2017stochastic,feng2017semi,yang2021fast,sirignano2020stochastic, latz2021analysis}. Authors aim to approximate the trajectory of SGD by a diffusion process which solves an SDE. Notice that they also study the scaled version of the diffusion limit of SGD. However, different from our approach, their scale is in temporal domain and cannot be applied in our research.
The Markov chain perspective of studying SGD iterates when step size goes to zero \citep{dieuleveut2017bridging} is related to the heavy traffic analysis in queuing theory \citep{eryilmaz2012asymptotically}. It has been studied in the literature using fluid and diffusion limits
\citep{gamarnik2006validity,harrison1988brownian,har_state_space,harlop_state_space,stolyar2004maxweight,Williams_state_space} where the interchange of limit is usually problematic \citep{eryilmaz2012asymptotically}. An alternative approach in stochastic networks is based on drift arguments introduced by \citep{eryilmaz2012asymptotically} and further generalized by \citep{maguluri2016heavy,maguluri2018optimal,hurtado2020transform,hurtado2020logarithmic, mou2020heavy}. We adopt similar techniques in quantifying the limiting distribution of the scaled SGD iterates. Notice that in stochastic networks, people mainly focus on finite state space Markov chains. However, when it comes to SGD iterates, the state space is continuous and thus more challenging.
The rest of this paper is organized as follows. In Section \ref{sec:distribution}, we characterize the distribution of $Y$ (cf. Eq (\ref{eq:two_limit})) in the following cases: (1) $F(\cdot)$ is the negative gradient for some smooth and strongly convex function $f(\cdot)$, (2) $F(x)=Ax+b$, where $A$ is a Hurwitz matrix, and (3) $F(x)=\mathcal{T}(x)-x$, where $\mathcal{T}(\cdot)$ is a contraction operator. In all three cases above, the scaling function is chosen to be $g(\alpha)=\sqrt{\alpha}$. Then in Section \ref{sec:scaling}, we first empirically show that for more general SA algorithms beyond these cases, the scaling function need not be $g(\alpha)=\sqrt{\alpha}$ and the distribution of $Y$ need not be Gaussian. Then we present a method to determine the scaling function for more general SA algorithms and make connection to the Euler-Maruyama discretization scheme for approximating SDE. Finally, we conclude this paper in Section \ref{sec:conclusion}.
\section{Characterizing the Asymptotic Stationary Distribution}\label{sec:distribution}
Through out this section, we make the following assumption regarding the noise sequence $\{w_k\}$.
\begin{assumption}\label{as:noise}
The sequence $\{w_k\}$ is independent and identically distributed with mean zero and a positive definite covariance matrix $\Sigma\in\mathbb{R}^{d\times d}$.
\end{assumption}
Note that Assumption \ref{as:noise} is much weaker than the assumption used in Section \ref{subsec:illustrative}, where the noise is assumed to obey the Gaussian distribution. Nevertheless, extending our results to the more general noise setting (e.g. martingale difference noise, Markovian noise, etc) is one of our future directions.
\subsection{SGD for Minimizing a Smooth and Strongly Convex Objective}\label{subsec:SGD}
Suppose that $F(x)=-\nabla f(x)$, where $f(\cdot)$ is an objective function. Then the SA algorithm becomes
\begin{align}\label{algo:SGD}
X_{k+1}^{(\alpha)}=X_k^{(\alpha)}+\alpha\left(-\nabla f(X_k^{(\alpha)})+w_k\right),
\end{align}
which is the well-known SGD algorithm for minimizing $f(\cdot)$. To proceed, we require the following definition.
\begin{definition}
A convex differentiable function $h:\mathbb{R}^d\mapsto\mathbb{R}$ is said to be $L$-smooth and $\sigma$-strongly convex with respect to the Euclidean norm $\|\cdot\|_2$ if and only if
\begin{align*}
h(y)&\leq h(x)+\langle \nabla h(x),y-x\rangle+\frac{L}{2}\|x-y\|_2^2\tag{$L$-smooth},\\
h(y)&\geq h(x)+\langle \nabla h(x),y-x\rangle+\frac{\sigma}{2}\|x-y\|_2^2\tag{$\sigma$-convex}
\end{align*}
for all $x,y\in\mathbb{R}^d$.
\end{definition}
To characterize the asymptotic behavior of Algorithm (\ref{algo:SGD}), we make the following assumption.
\begin{assumption}\label{as:SGD}
The function $f:\mathbb{R}^d\mapsto\mathbb{R}$ is twice differentiable, and is $L$-smooth and $\sigma$-strongly convex.
\end{assumption}
Under Assumption \ref{as:SGD}, the function $f(x)$ has a unique minimizer (or $F(x)=0$ has a unique solution), which we have denoted by $x^*$. To proceed, let $Y_k^{(\alpha)}=(X_k^{(\alpha)}-x^*)/\sqrt{\alpha}$ be the centered scaled iterate. We first write down the corresponding update equation of $Y_k^{(\alpha)}$ in following:
\begin{align}\label{algo:Y:SGD}
Y_{k+1}^{(\alpha)}=Y_k^{(\alpha)}-\sqrt{\alpha}\nabla f\left(\sqrt{\alpha} Y_k^{(\alpha)}+x^*\right)+\sqrt{\alpha} w_k,
\end{align}
which is obtained by first subtracting both sides of Eq. (\ref{algo:SGD}) by $x^*$, and then dividing by $\sqrt{\alpha}$.
We next characterize the two-step weak convergence (cf. Eq. (\ref{eq:two_limit})) in the following theorem. Let $H_f\in\mathbb{R}^{d\times d}$ be the Hessian matrix of the objective function $f(\cdot)$ evaluated at $x^*$, which is well-defined because $f(\cdot)$ is twice
differentiable (cf. Assumption \ref{as:SGD}).
\begin{theorem} \label{thm:SGDLimDist}
Consider the iterates $\{Y_k^{(\alpha)}\}$ generated by Algorithm (\ref{algo:Y:SGD}). Suppose that Assumptions \ref{as:noise} and \ref{as:SGD} are satisfied, then the following statements hold.
\begin{enumerate}[(1)]
\item There exists a threshold $\Bar{\alpha}>0$ such that for all $\alpha\in (0,\Bar{\alpha})$, the sequence of random variables $\{Y_k^{(\alpha)}\}$ converges weakly to some random variable $Y^{(\alpha)}$, which satisfies $\mathbb{E}[\|Y^{(\alpha)}\|_2^2]<\infty$.
\item For any positive sequence $\{\alpha_i\}$ satisfying $\alpha_i\in (0,\Bar{\alpha})$ for all $i$ and $\lim_{i\rightarrow\infty}\alpha_i=0$, the sequence $\{Y^{(\alpha_i)}\}$ converges weakly to a random variable $Y$, which satisfies the following integral equation
\begin{align}\label{eq:SGD}
\mathbb{E}\left[\left(t^\top \Sigma t+2it^\top H_f Y\right)e^{i t^\top Y}\right]=0
\end{align}
for all $t\in\mathbb{R}^d$.
In addition, suppose that Eq. (\ref{eq:SGD}) has a unique solution (in terms of the distribution of $Y$), then the distribution of $Y$ is a Multivariate normal distribution with mean zero and covariance matrix $\Sigma_Y$ being the unique solution of the Lyapunov equation $H_f\Sigma_Y+\Sigma_Y H_f^\top=\Sigma$.
\end{enumerate}
\end{theorem}
\begin{remark}
To establish Theorem \ref{thm:SGDLimDist} (2), we require Eq. (\ref{eq:SGD}) to have a unique solution in terms of the distribution of $Y$. Such assumption will be relaxed to some extent in Section \ref{subsec:uniqueness}.
\end{remark}
Since $\Sigma$ is positive definite (cf. Assumption \ref{as:noise}), and $H_f$ is also positive definite under Assumption \ref{as:SGD}, it is well established in the literature that the Lyapunov equation $H_f\Sigma_Y+\Sigma_Y H_f^\top=\Sigma$ has a unique solution, which is explicitly given by
\begin{align*}
\Sigma_Y=\int_{0}^\infty e^{-H_f u}\Sigma e^{-H_f^\top u}du.
\end{align*}
Consider the special case where $f(x)=x^2/2$. In this case we have $H_f=1$, and hence $\Sigma_Y=\frac{1}{2}\Sigma$ by the Lyapunov equation. As a result, the distribution of the limiting random variable $Y$ is Gaussian with mean zero, and covariance matrix being $\frac{1}{2}\Sigma$. This agrees with the illustrative example (which is for scalar valued iterates) presented in Section \ref{subsec:illustrative}.
From Theorem \ref{thm:SGDLimDist}, we see that the distribution of $Y$ only depends on the Hessian of $f(\cdot)$ at $x^*$. This makes intuitive sense because we are studying the asymptotic behavior of SA algorithm (\ref{algo:SGD}), and only the properties of $f(\cdot)$ around $x^*$ should play a role in determining the stationary distribution.
\subsection{Stochastic Approximation for Solving Linear Systems of Equations}\label{subsec:linear_SA}
Suppose that $F(x)=Ax+b$, where $A\in\mathbb{R}^{d\times d}$ and $b\in\mathbb{R}^d$. Then the SA algorithm (\ref{algo:sa}) becomes
\begin{align}\label{algo:linear_SA}
X_{k+1}^{(\alpha)}=X_k^{(\alpha)}+\alpha\left(AX_k^{(\alpha)}+b+w_k\right),
\end{align}
which aims at iteratively solving the linear equation $Ax+b=0$.
Note that since the matrix $A$ is not necessarily symmetric, $F(x)$ need not be a gradient of any objective function. Such linear SA algorithm arises in many realistic applications. One typical example is the TD-learning algorithm for solving the policy evaluation problem in reinforcement learning, where the goal is to solve a linear Bellman equation. See \cite{bertsekas1996neuro,srikant2019finite} for more details.
To study the asymptotic behavior of Algorithm (\ref{algo:linear_SA}), we make the following assumption regarding the matrix $A$.
\begin{assumption}\label{as:linear_sa}
The matrix $A$ is Hurwitz, i.e., all the eigenvalues of the matrix $A$ have strictly negative real part.
\end{assumption}
\begin{remark}
Since $A$ being Hurwitz implies $A$ being non-singular, Assumption \ref{as:linear_sa} implies that the target equation $Ax+b=0$ has a unique solution, which we denote by $x^*$.
\end{remark}
Assumption \ref{as:linear_sa} is standard in studying linear SA algorithms. In particular, it was shown in the literature that under Assumption \ref{as:linear_sa} and some mild conditions on the noise $\{w_k\}$, Algorithm (\ref{algo:linear_SA}) converges in the mean square sense to a neighborhood around $x^*$ \citep{bertsekas1996neuro}.
To study the asymptotic distribution, for a fixed stepsize $\alpha$, we define the centered scaled iterate $Y_k^{(\alpha)}$ by $Y_k^{(\alpha)}=(X_k^{(\alpha)}-x^*)/\sqrt{\alpha}$ for all $k\geq 0$. To find the corresponding update equation for $Y_k^{(\alpha)}$, using the update equation for $X_k^{(\alpha)}$ and the fact that $Ax^*+b=0$, we obtain
\begin{align}\label{algo:linear_SA_scaled}
Y_{k+1}^{(\alpha)}=(I+\alpha A) Y_k^{(\alpha)}+\sqrt{\alpha} w_k.
\end{align}
The full characterization of the two-step weak convergence (cf. Eq. (\ref{eq:two_limit})) of $\{Y_k^{(\alpha)}\}$ is presented in the following.
\begin{theorem}\label{thm:linear_SA}
Consider the iterates $\{Y_k^{(\alpha)}\}$ generated by Algorithm (\ref{algo:linear_SA_scaled}). Suppose that Assumptions \ref{as:noise} and \ref{as:linear_sa} are satisfied, then the following statements hold.
\begin{enumerate}[(1)]
\item There exists a threshold $\Bar{\alpha}'>0$ such that for all $\alpha\in (0,\Bar{\alpha}')$, the sequence of random variables $\{Y_k^{(\alpha)}\}$ converges weakly to some random variable $Y^{(\alpha)}$, which satisfies $\mathbb{E}[\|Y^{(\alpha)}\|_2^2]<\infty$.
\item For any positive sequence $\{\alpha_i\}$ satisfying $\alpha_i\in (0,\Bar{\alpha}')$ for all $i$ and $\lim_{i\rightarrow\infty}\alpha_i=0$, the sequence of random variables $\{Y^{(\alpha_i)}\}$ converges weakly to a random variable $Y$, which satisfies the following equation
\begin{align}\label{eq:linearSA}
\mathbb{E}\left[\left(t^\top \Sigma t-2it^\top AY\right)e^{it^\top Y}\right]=0,\quad \forall\;t\in\mathbb{R}^d.
\end{align}
In addition, suppose that Eq. (\ref{eq:linearSA}) has a unique solution, then the distribution of $Y$ is a Multivariate normal distribution with mean zero and covariance matrix being the unique solution of the Lyapunov equation $A \Sigma_Y+\Sigma_Y A^\top+\Sigma=0$.
\end{enumerate}
\end{theorem}
Since the matrix $A$ is Hurwitz, and $\Sigma$ is positive definite, the existence and uniqueness of a positive definition solution to the Lyapunov equation $A \Sigma_Y+\Sigma_Y A^\top+\Sigma=0$ are guaranteed \citep{haddad2011nonlinear}. Lyapunov equations were used extensively in studying the stability of linear ordinary differential equations (ODE). Interestingly, it also shows up in determining the limit distribution of centered scaled iterates of discrete linear SA algorithms.
\subsection{Stochastic Approximation under Contraction Assumption}\label{subsec:contractive_SA}
Suppose that $F(x)=\mathcal{T}(x)-x$, where $\mathcal{T}:\mathbb{R}^d\times\mathbb{R}^d$ is a general nonlinear operator. In this case, Algorithm (\ref{algo:sa}) becomes
\begin{align}\label{algo:contractive_SA}
X_{k+1}^{(\alpha)}=X_k^{(\alpha)}+\alpha\left(\mathcal{T}\left(X_k^{(\alpha)}\right)-X_k^{(\alpha)}+w_k\right),
\end{align}
which can be interpreted as an SA algorithm for finding the fixed-point of the operator $\mathcal{T}(\cdot)$. These type of algorithms arise in the context of reinforcement learning, where $\mathcal{T}(\cdot)$ is the Bellman operator. To proceed, we need the following definition.
\begin{definition}
Let $\nu_i$, $1\leq i\leq d$ be positive real numbers. Then the weighted $\ell_2$ norm $\|\cdot\|_{\nu}$ with weights $\{\nu_i\}_{1\leq i\leq d}$ is defined by $\|x\|_{\nu}=(\sum_{i=1}^d\nu_ix_i^2)^{1/2}$ for all $x\in\mathbb{R}^d$.
\end{definition}
Next, we state our assumption regarding the operator $\mathcal{T}(\cdot)$.
\begin{assumption}\label{as:contraction}
The operator $\mathcal{T}(\cdot)$ is differentiable, and
there exists $\gamma\in (0,1)$ such that $\|\mathcal{T}(x_1)-\mathcal{T}(x_2)\|_\mu\leq \gamma\|x_1-x_2\|_\mu$ for any $x_1,x_2\in\mathbb{R}^d$, where $\|\cdot\|_\mu$ is some weighted $\ell_2$-norm with weights $\{\mu_i\}_{1\leq i\leq d}$.
\end{assumption}
Assumption \ref{as:contraction} essentially states that the operator $\mathcal{T}(\cdot)$ is a contraction mapping with respect to the weighted $\ell_2$-norm $\|\cdot\|_\mu$. By Banach fixed-point theorem \citep{banach1922operations}, the operator $\mathcal{T}(\cdot)$ has a unique fixed-point, which we denote by $x^*$.
We next write down the update equation of the centered scaled iterate $Y_k^{(\alpha)}=(X_k^{(\alpha)}-x^*)/\sqrt{\alpha}$ in the following:
\begin{align}\label{algo:Y_contractive_SA}
Y_{k+1}^{(\alpha)}=Y_k^{(\alpha)}+\sqrt{\alpha}\left(\mathcal{T}\left(\sqrt{\alpha} Y_k^{(\alpha)}+x^*\right)-\left(\sqrt{\alpha} Y_k^{(\alpha)}+x^*\right)\right)+\sqrt{\alpha} w_k.
\end{align}
In the next theorem, we characterize the distribution of the limiting random vector $Y$ (cf. Eq. (\ref{eq:two_limit})). Let $J\in\mathbb{R}^{d\times d}$ be the Jacobian matrix of $\mathcal{T}(\cdot)$ evaluated at $x^*$.
\begin{theorem}\label{thm:contraction}
Consider the iterates $\{Y_k^{(\alpha)}\}$ generated by Algorithm (\ref{algo:Y_contractive_SA}). Suppose that Assumptions \ref{as:noise} and \ref{as:contraction} are satisfied, then the following statements hold.
\begin{enumerate}[(1)]
\item There exists a threshold $\Bar{\alpha}''>0$ such that for all $\alpha\in (0,\Bar{\alpha}'')$, the sequence of random variables $\{Y_k^{(\alpha)}\}$ converges weakly to some random variable $Y^{(\alpha)}$, which satisfies $\mathbb{E}[\|Y^{(\alpha)}\|_2^2]<\infty$.
\item For any positive sequence $\{\alpha_i\}$ satisfying $\alpha_i\in (0,\Bar{\alpha}'')$ for all $i$ and $\lim_{i\rightarrow\infty}\alpha_i=0$, the sequence of random variables $\{Y^{(\alpha_i)}\}$ converges weakly to a random variable $Y$, which satisfies the following equation
\begin{align}\label{eq:contractive}
\mathbb{E}\left[\left(t^\top \Sigma t-2it^\top (J-I)Y\right)e^{it^\top Y}\right]=0,\quad \forall\;t\in\mathbb{R}^d.
\end{align}
In addition, suppose that Eq. (\ref{eq:contractive}) has a unique solution, then the distribution of $Y$ is a Multivariate normal distribution with mean zero and covariance matrix being the unique solution of the Lyapunov equation $(J-I) \Sigma_Y+\Sigma_Y (J-I)^\top+\Sigma=0$.
\end{enumerate}
\end{theorem}
Under the contraction assumption, each eigenvalue of the matrix $J$ is contained in the open unit ball on the complex plane. Therefore, the matrix $J-I$ is Hurwitz and hence the Lyapunov equation $(J-I) \Sigma_Y+\Sigma_Y (J-I)^\top+\Sigma=0$ has a unique positive definite solution $\Sigma_Y$ \citep{khalil2002nonlinear}.
\subsection{Regarding the Uniqueness Assumption}\label{subsec:uniqueness}
In Theorems \ref{thm:SGDLimDist}, \ref{thm:linear_SA}, and \ref{thm:contraction}, after obtaining the corresponding integral equation (e.g., Eqs. (\ref{eq:SGD}), (\ref{eq:linearSA}), and (\ref{eq:contractive})), to conclude that the distribution of $Y$ is Gaussian, we need to assume that the equation has a unique solution. In this section, we show that such uniqueness assumption can be relaxed to some extend.
\subsubsection{Uni-Dimensional Setting}
Suppose that we are in the uni-dimensional setting, i.e., $d=1$. Then Eqs. (\ref{eq:SGD}), (\ref{eq:linearSA}), and (\ref{eq:contractive}) all reduce to an equation of the following form: $\mathbb{E}[(a t+2biY)e^{itY}]=0$ for all $t\in\mathbb{R}$, where $a$ and $b$ are positive constants. Let $\phi_Y(t)=\mathbb{E}[e^{itY}]$ be the characteristic function of $Y$. Then we can rewrite the previous equation as
\begin{align}\label{eq:81}
a t \phi_Y(t)+2b\frac{d\phi_Y(t)}{dt}=0,
\end{align}
where the interchange of integral and differentiation is justified \citep{flanders1973differentiation}. Now Eq. (\ref{eq:81}) is an ODE, which has solutions of the form
\begin{align*}
\phi_Y(t)=C\exp\left(-\frac{a}{4b}t^2\right),
\end{align*}
where $C$ is a constant. Since $\phi_Y(t)$ as a characteristic function satisfies $\phi_Y(0)=1$, we have $C=1$ and hence $\phi_Y(t)=\exp(-\frac{a}{4b}t^2)$, which is characteristic function for a Gaussian random variable with mean zero and covariance $\sqrt{a/(2b)}$. Therefore, the uniqueness assumption can be removed in the uni-dimensional setting.
\subsubsection{Multi-Dimensional Setting}
Moving to the multi-dimensional setting, consider Eq. (\ref{eq:SGD}) of Theorem \ref{thm:SGDLimDist} as a representative example. To show the same result of Theorem \ref{thm:SGDLimDist} (2) without imposing the uniqueness assumption, we consider the case where (1) the Hessian matrix $H_f$ of the objective function $f(\cdot)$ evaluated at $x^*$ is the identity matrix, and (2) the covariance matrix of the noise $w_k$ is also an identity matrix. Extending the result to the more general setting where $H_f$ and $\Sigma$ can be any positive definite matrices is a future research direction.
Similarly let $\phi_Y(t)=\mathbb{E}[e^{it^\top Y}]$ be the characteristic function of the random vector $Y$. Then in this case Eq. (\ref{eq:SGD}) becomes $t^\top t\phi_Y(t)+2t^\top \nabla \phi_Y(t)=0$, which is equivalent to
\begin{align}\label{eq:120}
0&=t^\top t+2t^\top \frac{\nabla \phi_Y(t)}{\phi_Y(t)}\nonumber\\
&=t^\top t+2t^\top \nabla \psi_Y(t),
\end{align}
where $\psi_Y(t):=\log(\phi_Y(t))$. To solve the partial differential equation (PDE) (\ref{eq:120}), we will first convert the PDE from Cartesian coordinates to spherical coordinates, which then is directly solvable.
The $d$-dimensional spherical coordinate system consists of a radial coordinate $\rho$, and $d-1$ angular coordinates $\{\theta_i\}_{1\leq i\leq d-1}$. The relation between the Cartesian coordinates $(t_1,\cdots,t_d)$ and the spherical coordinates $(\rho,\theta_1,\cdots,\theta_{d-1})$ is given by
\begin{align*}
t_1&=\rho cos(\theta_1)\\
t_2&=\rho sin(\theta_1)cos(\theta_2)\\
t_3&=\rho sin(\theta_1)sin(\theta_2)cos(\theta_3)\\
&\vdots\\
t_{d-1}&=\rho sin(\theta_1)sin(\theta_2)\cdots sin(\theta_{d-2})cos(\theta_{d-1})\\
t_d&=\rho sin(\theta_1)sin(\theta_2)\cdots sin(\theta_{d-2})sin(\theta_{d-1}).
\end{align*}
For simplicity of notation, denote $S\in\mathbb{R}^d$ as the spherical coordinate representation of $(t_1,\cdots,t_d)$ given above, i.e., $S_1=\rho cos(\theta_1)$, $\cdots$, $S_d=\rho sin(\theta_1)sin(\theta_2)\cdots sin(\theta_{d-2})sin(\theta_{d-1})$.
To proceed, we first compute the Jacobian matrix $J_d$ of the transformation in the following:
{\small \begin{align*}
J_d=\begin{bmatrix}
cos(\theta_1)& -\rho sin(\theta_1) & 0 &0 &\cdots &0\\
sin(\theta_1)cos(\theta_2)& \rho cos(\theta_1)cos(\theta_2)&-\rho sin(\theta_1)sin(\theta_2)&0&\cdots &0\\
\vdots &\vdots &\vdots&&\ddots&\vdots\\
\prod_{i=1}^{d-2}sin(\theta_i)cos(\theta_{d-1})&\cdots&\cdots& & &-\rho \prod_{i=1}^{d-1}sin(\theta_i)\\
\prod_{i=1}^{d-1}sin(\theta_i)&\rho\prod_{i=1}^{d-2}cos(\theta_i)sin(\theta_{d-1})&\cdots& & &\rho \prod_{i=1}^{d-2}sin(\theta_i)cos(\theta_{d-1})\\
\end{bmatrix}.
\end{align*}}
Using spherical coordinate system, Eq. (\ref{eq:120}) is equivalent to
\begin{align}\label{eq:121}
\rho^2+2S^\top J_d \nabla \psi_Y(\rho,\theta_1,\cdots,\theta_{d-1}).
\end{align}
By direct computation (where we use $sin^2(\theta)+cos^2(\theta)=1$ for any $\theta$), we have $S^\top J_d=(\rho,0,\cdots,0)$. As a result, Eq. (\ref{eq:121}) simplifies to
\begin{align}\label{eq:122}
\rho+2\frac{\partial \psi_Y(\rho,\theta_1,\cdots,\theta_{d-1})}{\partial \rho}=0,
\end{align}
which implies that $\psi_Y(\rho,\theta_1,\cdots,\theta_{d-1})=-\frac{\rho^2}{4}+C(\theta_1,\cdots,\theta_{d-1})$. Using the initial condition that $\psi_Y(0,\theta_1,\cdots,\theta_{d-1})=\log(\phi_Y(0))=\log(1)=0$ for any $\theta_1,\cdots,\theta_{d-1}$, we see that $C(\theta_1,\cdots,\theta_{d-1})=0$ and hence
$\phi_Y(\rho,\theta_1,\cdots,\theta_{d-1})=\frac{\rho^2}{4}$. Therefore, we have that $\psi_Y(t)=-\frac{t^\top t}{4}$, which implies $\phi_Y(t)=\exp(-\frac{t^\top t}{4})$. Therefore, the distribution of $Y$ is a multinormal distribution with mean zero and covariance matrix being $I_d/\sqrt{2}$. This agrees with Theorem \ref{thm:SGDLimDist} (2) when $H_f=\Sigma=I$, but the uniqueness assumption is not required to establish the result.
\subsection{Proof of Theorem \ref{thm:SGDLimDist}}
In this section, we present our proof for Theorem \ref{thm:SGDLimDist}. The proofs for Theorems \ref{thm:linear_SA} and \ref{thm:contraction} are similar and hence are omitted.
Before going into details, we first highlight the main ideas for the proof. For Theorem \ref{thm:SGDLimDist} (1), to show that $\{Y^{(\alpha_i)}\}$ converges weakly to a some random variable $Y$, we show that any sub-sequence of $\{Y^{(\alpha_i)}\}$ further contains a weakly convergent sub-sequence, with a common weak limit. We do it by first showing that the sequence of random variables $\{Y^{(\alpha_i)}\}$ is tight. In particular, we show that there exists a constant $B$ independent of $\alpha$ such that $\mathbb{E}[\|Y^{(\alpha)}\|_2^2]\leq B$ for all small enough $\alpha$. This result in conjunction with the Markov inequality implies the tightness of $\{Y^{(\alpha_i)}\}$. As a result of tightness, $\{Y^{(\alpha_i)}\}$ contains a weakly convergent sub-sequence.
Now consider Theorem \ref{thm:SGDLimDist} (2). For any positive sequence $\{\alpha_i\}$ such that $\lim_{i\rightarrow\infty}\alpha_i=0$, since the family of random variables $\{Y^{(\alpha_i)}\}$ is tight, there is a weakly convergent subsequence $\{Y^{\alpha_{i_k}}\}$. We further show that the weak limit $Y$ of the subsequence $\{Y^{(\alpha_{i_k})}\}$ solves Eq. (\ref{eq:SGD}). In this case, under the assumption that Eq. (\ref{eq:SGD}) has a unique solution, the random variable $Y$ is a Gaussian random variable with mean zero, and covariance matrix $\Sigma_Y$ being the unique solution of the Lyapunov equation $H_f \Sigma_Y+\Sigma_Y H_f^\top =\Sigma$. Since for every sequence $\{Y^{(\alpha_i)}\}$, there is a weakly convergent subsequence $\{Y^{(\alpha_{i_k})}\}$ with a common weak limit, the sequence of random variables $\{Y^{(\alpha_i)}\}$ also converge weakly to the same random variable $Y$.
\subsubsection{Proof of Theorem \ref{thm:SGDLimDist} (1)}
To prove the result, we will apply Proposition 2.1 in \cite{yu2020analysis}. For completeness, we first state this proposition using our notation in the following.
\begin{proposition}\label{prop:SGD}
Consider $\{X_k^{(\alpha)}\}$ generated by Algorithm (\ref{algo:SGD}). Suppose that
\begin{enumerate}[(a)]
\item There exists $L'>0$ such that $\|\nabla f(x)\|_2\leq L'(1+\|x\|_2)$ for any $x\in\mathbb{R}^d$.
\item There exist $\ell_1,\ell_2>0$ such that $\langle x, \nabla f(x)\rangle\geq \ell_1\|x\|_2^2-\ell_2$ for all $x\in\mathbb{R}^d$.
\item The noise sequence $\{w_k\}$ is an i.i.d. sequence satisfying $\mathbb{E}[w_k]=0$ and $\mathbb{E}^{1/2}[\|w_k\|_2^2]\leq L''(1+\|x\|_2)$ for all $k\geq 0$, where $L''>0$ is a constant.
\end{enumerate}
Then, when the constant stepsize $\alpha<\frac{\ell_1-\sqrt{\max(\ell_1^2-(3L'^2+L''),0)}}{3L'^2+L''^2}$, the following statements hold.
\begin{enumerate}[(1)]
\item The iterates $\{X_k^{(\alpha)}\}$ admit a unique stationary distribution $\pi_\alpha$, which depends on the choice of $\alpha$. In addition, let $X^{(\alpha)}\sim \pi_\alpha$, then we have $\mathbb{E}[\|X^{(\alpha)}\|_2^2]<\infty$.
\item For a test function $\phi:\mathbb{R}^d\mapsto\mathbb{R}$ satisfying $|\phi(x)|\leq L_\phi(1+\|x\|_2)$ for all $x\in\mathbb{R}^d$ and some $L_\phi>0$, and for any initialization $X_0^{(\alpha)}\in\mathbb{R}^d$ of the SGD algorithm (\ref{algo:SGD}), there exists $\rho\in (0,1)$ and $\kappa$
(both depending on $\alpha$) such that we have $|\mathbb{E}[\phi(X_k^{(\alpha)})]-\pi_{\alpha}(\phi)|\leq \kappa \rho^k(1+\|X_0^{(\alpha)}\|_2^2)$, where $\pi_\alpha(\phi)=\mathbb{E}[\phi(X^{(\alpha)})]$.
\end{enumerate}
\end{proposition}
Note that Proposition \ref{prop:SGD} (2) implies that $\{X_k^{(\alpha)}\}$ converges weakly to $X^{(\alpha)}$. To apply Proposition \ref{prop:SGD}, we next verify the assumptions.
\begin{enumerate}[(a)]
\item Since the objective function $f(\cdot)$ is assumed to be $L$-smooth, we have for any $x\in\mathbb{R}^d$ that $\|\nabla f(x)-\nabla f(0)\|_2\leq L\|x\|_2$, which implies
\begin{align*}
\|\nabla f(x)\|_2\leq \|\nabla f(0)\|_2+L\|x\|_2\leq \underbrace{\max(\|\nabla f(0)\|_2,L)}_{L'}(\|x\|_2+1).
\end{align*}
\item Since the objective function is assumed to be $\sigma$-strongly convex, we have for any $x\in\mathbb{R}^d$:
\begin{align*}
f(0)-f(x)\geq \langle \nabla f(x),-x\rangle+\frac{\sigma}{2}\|x\|_2^2,
\end{align*}
which implies that
\begin{align*}
\langle \nabla f(x),x\rangle\geq \frac{\sigma}{2}\|x\|_2^2+f(x)-f(0)\geq \underbrace{\frac{\sigma}{2}}_{\ell_1}\|x\|_2^2+\underbrace{f(x^*)-f(0)-1}_{\ell_2}.
\end{align*}
\item This is immediately implied by Assumption \ref{as:noise}, with $L''=\text{Trace}(\Sigma)^{1/2}$.
\end{enumerate}
Now apply Proposition \ref{prop:SGD}, when the stepsize $\alpha$ satisfies $\alpha<\frac{\sigma}{2(3L'^2+\text{Trace}(\Sigma))}$,
the SGD iterates $\{X^{(\alpha)}_k\}$ converge weakly to some random variable $X^{(\alpha)}$, which is distributed according to the unique stationary distribution $\pi_\alpha$. In addition, we have $\mathbb{E}[\|X^{(\alpha)}\|_2^2]<\infty$. Since $Y_k^{(\alpha)}$ is the centered scaled variant of $X_k^{(\alpha)}$, the sequence $\{Y_k^{(\alpha)}\}$ converges weakly to some random variable $Y^{(\alpha)}$ and $\mathbb{E}[\|Y^{(\alpha)}\|_2^2]<\infty$.
\subsubsection{Proof of Theorem \ref{thm:SGDLimDist} (2)}
Following the road map described in the beginning of this section, we present and prove a sequence of lemmas in the following. Together they imply the desired result.
\begin{lemma}\label{le:tightness}
Let $\alpha_0=\sigma/L^2$. the family of random variables $\{Y^{(\alpha)}\}_{0<\alpha\leq \alpha_0}$ is tight.
\end{lemma}
\begin{proof}[Proof of Lemma \ref{le:tightness}]
We first show that that there exists an absolute constant $C>0$ such that $\mathbb{E}[\|Y^{(\alpha)}\|^2]\leq C$ for any $\alpha\in (0,\alpha_0]$. Using the update equation (\ref{algo:sa}), we have
\begin{align*}
Y_{k+1}^{(\alpha)}&=Y_k^{(\alpha)}+\frac{\alpha}{g(\alpha)} \left(-\nabla f(Y_k^{(\alpha)}g(\alpha)+x^*)+w_k\right)\\
&=Y_k^{(\alpha)}-\sqrt{\alpha}\nabla f(\sqrt{\alpha} Y_k^{(\alpha)}+x^*)+\sqrt{\alpha} w_k
\end{align*}
The existence and uniqueness of a stationary distribution $Y^{(\alpha)}$ is proved in Part (1) of this theorem. We next show that the family of random variables $\{Y^{(\alpha)}\}_{0\leq \alpha\leq \alpha_0}$ is tight. Using the equation
\begin{align*}
Y^{(\alpha)} \stackrel{D}{=}Y^{(\alpha)}-\sqrt{\alpha}\nabla f(\sqrt{\alpha} Y^{(\alpha)}+x^*)+\sqrt{\alpha} w,
\end{align*}
and we have
\begin{align*}
\mathbb{E}[\|Y^{(\alpha)}\|_2^2]=\;&\mathbb{E}[\|Y^{(\alpha)}\|_2^2]+\alpha\mathbb{E}\left[\left\|\nabla f(\sqrt{\alpha} Y^{(\alpha)}+x^*)\right\|_2^2\right]+\alpha\text{Trace}(\Sigma)\\
&-2\sqrt{\alpha}\mathbb{E}\left[{Y^{(\alpha)}}^\top\nabla f(\sqrt{\alpha} Y^{(\alpha)}+x^*) \right].
\end{align*}
By smoothness, we have
\begin{align*}
\left\|\nabla f(\sqrt{\alpha} Y^{(\alpha)}+x^*)\right\|_2^2\leq L^2\alpha\|Y^{(\alpha)}\|_2^2.
\end{align*}
By strong convexity, we have
\begin{align*}
{Y^{(\alpha)}}^\top\nabla f(\sqrt{\alpha} Y^{(\alpha)}+x^*)&=\frac{1}{\sqrt{\alpha}}(\sqrt{\alpha}Y^{(\alpha)}+x^*-x^*)^\top\left(\nabla f(\sqrt{\alpha} Y^{(\alpha)}+x^*)-\nabla f(x^*)\right)\\
&\geq \sigma\sqrt{\alpha}\|Y^{(\alpha)}\|_2^2.
\end{align*}
Therefore, we obtain
\begin{align*}
0\leq\;& L^2\alpha^2\mathbb{E}[\|Y^{(\alpha)}\|_2^2]+\alpha \text{Trace}(\Sigma)-2\sigma\alpha\mathbb{E}[\|Y^{(\alpha)}\|_2^2].
\end{align*}
When $\alpha\in (0,\alpha_0]$, we have from the previous inequality that
\begin{align*}
\mathbb{E}[\|Y^{(\alpha)}\|_2^2]\leq \frac{\text{Trace}(\Sigma)}{2\sigma-L^2\alpha}\leq \frac{\text{Trace}(\Sigma)}{\sigma}.
\end{align*}
Hence, for any $\alpha>0$, let $M=\sqrt{\text{Trace}(\Sigma)/\sigma\alpha}$, then we have
\begin{align*}
\mathbb{P}(\|Y^{(\alpha)}\|>M)\leq \frac{\mathbb{E}[\|Y^{(\alpha)}\|^2]}{M^2}\leq \frac{\text{Trace}(\Sigma)}{\sigma M^2}=\alpha
\end{align*}
for any $\alpha\in (0,\alpha_0]$. It follows that the family of random variables $\{Y^{(\alpha)}\}_{0<\alpha\leq \alpha_0}$ is tight.
\end{proof}
\begin{lemma}\label{le:equation}
Let $\{\alpha_i\}$ be a positive sequence of real numbers such that $\lim_{i\rightarrow\infty}\alpha_i=0$. Suppose that $\{Y^{\alpha_i}\}$ converges weakly to some random variable $Y$. Then $Y$ satisfies the following equation
\begin{align}\label{eq:40}
\mathbb{E}\left[\frac{t^\top \Sigma t}{2} e^{i t^\top Y}\right]=-\mathbb{E}\left[\exp(it^\top Y)it^\top H_f Y\right].
\end{align}
\end{lemma}
\begin{proof}[Proof of Lemma \ref{le:equation}]
For any $i\geq 0$, we have
\begin{align*}
Y^{(\alpha_i)} \overset{D}{=} Y^{(\alpha_i)}-\sqrt{\alpha_i}\nabla f(\sqrt{\alpha_i} Y^{(\alpha_i)}+x^*)+\sqrt{\alpha_i}w,
\end{align*}
which implies for any $t\in\mathbb{R}^d$:
\begin{align}\label{eq:1}
\mathbb{E}\left[e^{i t^\top Y^{(\alpha_i)}}\right]=\mathbb{E}\left[\exp\left(it^\top Y^{(\alpha_i)}\right)\exp\left(-\sqrt{\alpha_i}it^\top \nabla f(\sqrt{\alpha_i }Y^{(\alpha_i)}+x^*)\right)\right]\mathbb{E}\left[e^{\sqrt{\alpha_i} it^\top w}\right]
\end{align}
Using the Taylor's theorem and we have
\begin{align*}
&\exp\left(-\sqrt{\alpha_i}it^\top \nabla f(\sqrt{\alpha_i }Y^{(\alpha_i)}+x^*)\right)\\
=\;&1-\sqrt{\alpha_i} it^\top \nabla f(\sqrt{\alpha_i}Y^{(\alpha_i)}+x^*)+\mathcal{O}\left(\alpha_i \|t\|^2\|\nabla f(\sqrt{\alpha_i}Y^{(\alpha_i)}+x^*\|^2\right).
\end{align*}
Using Theorem 3.3.20 from \cite{durrett2019probability} and we have
\begin{align*}
\mathbb{E}\left[e^{\sqrt{\alpha_i}it^\top Y^{(\alpha_i)}}\right]=1-\frac{\alpha_i t^\top \Sigma t}{2}+o(\alpha_i \|t\|^2).
\end{align*}
Using the previous two inequalities in Eq. (\ref{eq:1}) and we have
\begin{align*}
&\mathbb{E}\left[e^{i t^\top Y^{(\alpha_i)}}\right]\\
=\;&\mathbb{E}\left[\exp(it^\top Y^{(\alpha_i)})\exp(-\sqrt{\alpha_i}it^\top \nabla f(\sqrt{\alpha_i }Y^{(\alpha_i)}+x^*))\right]\mathbb{E}[e^{\sqrt{\alpha_i} it^\top w}]\\
=\;&\mathbb{E}\left[\exp(it^\top Y^{(\alpha_i)})\left(1-\sqrt{\alpha_i} it^\top \nabla f(\sqrt{\alpha_i}Y^{(\alpha_i)}+x^*)+\mathcal{O}\left(\alpha_i \|t\|^2\|\nabla f(\sqrt{\alpha_i}Y^{(\alpha_i)}+x^*\|^2\right)\right)\right]\times\\
&\left(1-\frac{\alpha_i t^\top \Sigma t}{2}\right)
+\mathbb{E}\left[\exp(it^\top Y^{(\alpha_i)})\exp(-\sqrt{\alpha_i}it^\top \nabla f(\sqrt{\alpha_i }Y^{(\alpha_i)}+x^*))\right]o(\alpha_i \|t\|^2)\\
=\;&\mathbb{E}\left[e^{i t^\top Y^{(\alpha_i)}}\right]-\mathbb{E}\left[\frac{\alpha_i t^\top \Sigma t}{2} e^{i t^\top Y^{(\alpha_i)}}\right]-\mathbb{E}\left[\exp(it^\top Y^{(\alpha_i)})\sqrt{\alpha_i} it^\top \nabla f(\sqrt{\alpha_i}Y^{(\alpha_i)}+x^*)\right]\\
&+\mathbb{E}\left[\frac{\alpha_i t^\top \Sigma t}{2}\exp(it^\top Y^{(\alpha_i)})\sqrt{\alpha_i} it^\top \nabla f(\sqrt{\alpha_i}Y^{(\alpha_i)}+x^*)\right]\\
&+\mathbb{E}\left[e^{i t^\top Y^{(\alpha_i)}}\mathcal{O}\left(\alpha_i \|t\|^2\|\nabla f(\sqrt{\alpha_i}Y^{(\alpha_i)}+x^*\|^2\right)\right]\\
&-\mathbb{E}\left[\frac{\alpha_i t^\top \Sigma t}{2}e^{i t^\top Y^{(\alpha_i)}}\mathcal{O}\left(\alpha_i \|t\|^2\|\nabla f(\sqrt{\alpha_i}Y^{(\alpha_i)}+x^*\|^2\right)\right]\\
&+\mathbb{E}\left[\exp(it^\top Y^{(\alpha_i)})\exp(-\sqrt{\alpha_i}it^\top \nabla f(\sqrt{\alpha_i }Y^{(\alpha_i)}+x^*))\right]o(\alpha_i \|t\|^2).
\end{align*}
Simplify the above equality and we obtain
\begin{align*}
\underbrace{\mathbb{E}\left[\frac{t^\top \Sigma t}{2} e^{i t^\top Y^{(\alpha_i)}}\right]}_{T_1}=\;&-\underbrace{\mathbb{E}\left[\exp(it^\top Y^{(\alpha_i)})\frac{it^\top \nabla f(\sqrt{\alpha_i}Y^{(\alpha_i)}+x^*)}{\sqrt{\alpha_i}} \right]}_{T_2}\\
&+\underbrace{\mathbb{E}\left[\frac{t^\top \Sigma t}{2}\exp(it^\top Y^{(\alpha_i)})\sqrt{\alpha_i} it^\top \nabla f(\sqrt{\alpha_i}Y^{(\alpha_i)}+x^*)\right]}_{T_3}\\
&+\underbrace{\mathbb{E}\left[e^{i t^\top Y^{(\alpha_i)}}\mathcal{O}\left( \|t\|^2\|\nabla f(\sqrt{\alpha_i}Y^{(\alpha_i)}+x^*\|^2\right)\right]}_{T_4}\\
&-\underbrace{\mathbb{E}\left[\frac{t^\top \Sigma t}{2}e^{i t^\top Y^{(\alpha_i)}}\mathcal{O}\left(\alpha_i \|t\|^2\|\nabla f(\sqrt{\alpha_i}Y^{(\alpha_i)}+x^*\|^2\right)\right]}_{T_5}\\
&+\underbrace{\mathbb{E}\left[\exp(it^\top Y^{(\alpha_i)})\exp(-\sqrt{\alpha_i}it^\top \nabla f(\sqrt{\alpha_i }Y^{(\alpha_i)}+x^*))\right]\frac{o( \alpha_i \|t\|^2)}{\alpha_i}}_{T_6}.
\end{align*}
We next let $i$ go to infinity on both sides of the previous inequality and evaluate the limit of the terms $\{T_i\}_{1\leq i\leq 6}$.
Since $\{Y^{(\alpha_i)}\}$ converges weakly to some random variable $Y$, we have by continuity theorem (Theorem 3.3.17 in \cite{durrett2019probability}) that
\begin{align*}
\lim_{i\rightarrow \infty} \mathbb{E}\left[\frac{t^\top \Sigma t}{2} e^{i t^\top Y^{(\alpha_{i})}}\right]=\frac{t^\top \Sigma t}{2}\mathbb{E}\left[ e^{i t^\top Y}\right].
\end{align*}
For the term $T_6$, we have by bounded convergence theorem that $\lim_{\alpha_i\rightarrow 0}T_6=0$. To evaluate the terms $T_2$ to $T_5$, the following definition and result from \cite{van2000asymptotic} is needed.
\begin{definition}
A sequence of random variables $\{X_n\}$ is called asymptotically uniformly integrable if $\lim_{M\rightarrow\infty}\limsup_{n\rightarrow\infty}\mathbb{E}[|X_n|\mathbb{I}\{|X_n|>M\}]=0$.
\end{definition}
\begin{theorem}[Theorem 2.20 in \cite{van2000asymptotic}]\label{thm:weak_convergence}
Let $f:\mathbb{R}^d\mapsto\mathbb{R}$ be measurable and continuous at every point in a set $C$. Let $X_n\Rightarrow X$, where $X$ takes its values in $C$. Then $\mathbb{E}[f (X_n)]\rightarrow\mathbb{E}[f(X)]$ if and only if the
sequence of random variables $f(X_n)$ is asymptotically uniformly integrable.
\end{theorem}
Now consider the term $T_2$. Since
\begin{align*}
&\mathbb{E}\bigg[\;\left|\exp(it^\top Y^{(\alpha_i)})\frac{it^\top \nabla f(\sqrt{\alpha_i}Y^{(\alpha_i)}+x^*)}{\sqrt{\alpha_i}}\right|\times\\
&\mathbb{I}\left\{\left|\exp(it^\top Y^{(\alpha_i)})\frac{it^\top \nabla f(\sqrt{\alpha_i}Y^{(\alpha_i)}+x^*)}{\sqrt{\alpha_i}}\right|>M\right\}\bigg]\\
\leq \;&\frac{1}{M}\mathbb{E}\bigg[\;\left|\exp(it^\top Y^{(\alpha_i)})\frac{it^\top \nabla f(\sqrt{\alpha_i}Y^{(\alpha_i)}+x^*)}{\sqrt{\alpha_i}}\right|^2\times\\
&\mathbb{I}\left\{\left|\exp(it^\top Y^{(\alpha_i)})\frac{it^\top \nabla f(\sqrt{\alpha_i}Y^{(\alpha_i)}+x^*)}{\sqrt{\alpha_i}}\right|>M\right\}\bigg]\\
\leq \;&\frac{1}{\alpha_i M }\mathbb{E}\left[\;\left|t^\top \nabla f(\sqrt{\alpha_i}Y^{(\alpha_i)}+x^*)\right|^2\mathbb{I}\left\{\left|\exp(it^\top Y^{(\alpha_i)})\frac{it^\top \nabla f(\sqrt{\alpha_i}Y^{(\alpha_i)}+x^*)}{\sqrt{\alpha_i}}\right|>M\right\}\right]\\
\leq \;&\frac{\|t\|^2}{\alpha_i M }\mathbb{E}\left[\;\|\nabla f(\sqrt{\alpha_i}Y^{(\alpha_i)}+x^*)\|^2\right]\\
\leq \;&\frac{L\|t\|^2}{M }\mathbb{E}\left[\;\|Y^{(\alpha_i)}\|^2\right]\\
\leq \;&\frac{L\|t\|^2}{M }\mathbb{E}\left[\;\|Y^{(\alpha_i)}\|^2\right]\\
\leq \;&\frac{L\text{Trace}(\Sigma)\|t\|^2}{\sigma M },
\end{align*}
which goes to zero as $M\rightarrow\infty$, we have by Theorem \ref{thm:weak_convergence} that
\begin{align*}
\lim_{i\rightarrow \infty}T_2=\mathbb{E}\left[\exp(it^\top Y)it^\top H_f Y\right].
\end{align*}
Using the same line of analysis, we have $\lim_{i\rightarrow \infty}T_3=\lim_{i\rightarrow \infty}T_4=\lim_{i\rightarrow \infty}T_5=0$. It follows that
\begin{align*}
\mathbb{E}\left[\frac{t^\top \Sigma t}{2} e^{i t^\top Y}\right]=-\mathbb{E}\left[\exp(it^\top Y)it^\top H_f Y\right].
\end{align*}
Rearranging terms and we obtain the resulting equation.
\end{proof}
\begin{lemma}\label{le: unique solution}
Suppose Eq. (\ref{eq:40}) admits a unique solution. Then the random variable $Y$ given in Lemma \ref{le:equation} follows a Gaussian distribution with mean zero, and covariance matrix $\Sigma_Y$ being the unique solution of the Lyapunov equation $H_f \Sigma_Y+\Sigma_Y H_f^\top =\Sigma$.
\end{lemma}
\begin{proof}[Proof of Lemma \ref{le: unique solution}]
Suppose that Eq. (\ref{eq:SGD}) has a unique solution, we only need to verify that the multinormal distribution with mean zero and covariance matrix being the solution to the Lyapunov equation $H_f^\top \Sigma_Y+\Sigma_Y H_f=\Sigma$ solves equation \eqref{eq:SGD}.
\begin{align*}
&\mathbb{E}\left[(2it^\top H_f Y + t^\top \Sigma t)e^{it^\top Y}\right]\\
=\;&C\int_{\mathbb{R}^d}(2it^\top H_f y + t^\top \Sigma t)e^{it^\top y}e^{-\frac{1}{2}y^\top \Sigma_Y^{-1} y}dy\tag{$C=\frac{1}{\sqrt{\left( 2\pi \right)^d \textit{det}(\Sigma_Y)}}$}\\
=\;&Ce^{-\frac{1}{2}t^\top \Sigma_Y t}\int_{\mathbb{R}^d}(2it^\top H_f y + t^\top \Sigma t)e^{-\frac{1}{2}(y-i\Sigma_Y t)^\top \Sigma_Y^{-1}(y-i\Sigma_Y t)}dy\\
=\;&Ce^{-\frac{1}{2}t^\top \Sigma_Y t}\int_{\mathbb{R}^d}(2it^\top H_f(z+i\Sigma_Y t)+t^\top \Sigma t)e^{-\frac{1}{2}z^\top \Sigma_Y^{-1} z} dz\tag{change of variable}\\
=\;&Ce^{-\frac{1}{2}t^\top \Sigma_Y t}\int_{\mathbb{R}^d}(-2t^\top H_f\Sigma_Y t+t^\top \Sigma t)e^{-\frac{1}{2}z^\top \Sigma_Y^{-1} z} dz\\
=\;&Ce^{-\frac{1}{2}t^\top \Sigma_Y t}\int_{\mathbb{R}^d}(-t^\top (H_f\Sigma_Y+\Sigma_Y H_f^\top) t-t^\top t)e^{-\frac{1}{2}z^\top \Sigma_Y^{-1} z} dz\\
=\;&Ce^{-\frac{1}{2}t^\top \Sigma_Y t}\int_{\mathbb{R}^d}(-t^\top \Sigma t+t^\top \Sigma t)e^{-\frac{1}{2}z^\top \Sigma_\alpha z} dz\tag{The Lyapunov equation}\\
=\;&0.
\end{align*}
\end{proof}
\section{Identifying the Suitable Scaling Function for More General SA Algorithms}\label{sec:scaling}
In the previous section, we have shown that for several particular SA algorithms (e.g. SGD, linear SA, and contractive SA), the scaling function is $g(\alpha)=\sqrt{\alpha}$ and distribution of the limiting random variable $Y$ is Gaussian. In this section, we consider more general SA algorithms. We first show impirically in the following section that in general the scaling function need not be $g(\alpha)=\sqrt{\alpha}$, and the distribution of $Y$ need not be Gaussian.
\subsection{Numerical Experiments}\label{subsec:experiment}
Suppose that Algorithm (\ref{algo:sa}) is the SGD algorithm for minimizing the scalar objective $f(x)=x^4/4$. That is:
\begin{align}\label{algo:sa:x^4}
X_{k+1}^{(\alpha)}=X_k^{(\alpha)}+\alpha\left(-(X_k^{(\alpha)})^3+w_k\right).
\end{align}
Note that $f(\cdot)$ in this case is neither smooth nor strongly convex. It is clear that the unique minimizer of $f(\cdot)$ is zero. Let the centered scaled iterate $Y_k^{(\alpha)}$ be defined by $Y_k^{(\alpha)}=X_k^{(\alpha)}/g(\alpha)$. We next use numerical simulation to show that the correct scaling function in this case is $g(\alpha)=\alpha^{1/4}$ instead of $g(\alpha)=\sqrt{\alpha}$.
\begin{figure}[h]
\begin{minipage}{0.45\textwidth}
\centering
\includegraphics[width=\linewidth]{x4_1_2.png}
\caption{Estimated Density Functions when choosing $g(\alpha)=\alpha^{1/2}$}
\label{fig:1}
\end{minipage}
\hspace{0.02\textwidth}
\begin{minipage}{0.45\textwidth}
\centering
\includegraphics[width=\linewidth]{x4_1_4.png}
\caption{Estimated Density Functions when choosing $g(\alpha)=\alpha^{1/4}$}
\label{fig:2}
\end{minipage}
\vfill
\begin{minipage}{0.45\textwidth}
\centering
\includegraphics[width=\linewidth]{x4_1_4_line.png}
\caption{$\log(p_Y(y))$ as a function of $y^4$}\label{fig:3}
\end{minipage}
\end{figure}
In Figures \ref{fig:1} and \ref{fig:2}, we plot the empirical density function of $Y^{(\alpha)}$ for different $\alpha$. For the right scaling function, we expect the density function to converge as $\alpha$ decreases, while for the wrong scaling function, we expect the density function to change drastically for order-wise different $\alpha$. As we see, it is clear that $g(\alpha)=\sqrt{\alpha}$ is not suitable in this case, and $g(\alpha)=\alpha^{1/4}$ seems to be the right scaling.
To further verify this result, we plot the logarithmic empirical density function as a function of $y^4$ in Figure \ref{fig:3}. We observe linear growth in Figure \ref{fig:3}. This indicates that the density function $p_Y(y)$ is proportional to $e^{\beta y^4}$, where $\beta$ is some numerical constant. Therefore, numerical experiments suggest that the distribution of $Y$ is not Gaussian but super Gaussian in this problem.
\subsection{A Method to Determine the Suitable Scaling Function}
Inspired by the numerical simulations provided in the previous section, we here provide a method to determine the correct scaling function for general SA algorithms.
To gain intuition, we consider the centered scaled iterates $Y_k^{(\alpha)}=X_k^{(\alpha)}/\alpha^{1/4}$ for SA algorithm (\ref{algo:sa:x^4}). The update equation of $Y_k^{(\alpha)}$ is given by
\begin{align*}
Y_{k+1}^{(\alpha)}=Y_k^{(\alpha)}-\alpha^{3/2}(Y_k^{(\alpha)})^3+\alpha^{3/4}w_k.
\end{align*}
Notably, the factor in terms of the stepsize $\alpha$ in front of the term $(Y_k^{(\alpha)})^3$ is $\alpha^{3/2}$, which is equal to the square of the factor $\alpha^{3/4}$ in front of the noise term $w_k$.
Now for general SA algorithm (\ref{algo:sa}), by rewriting the update equation (\ref{algo:sa}) in terms of the centered scaled iterate $Y_k^{(\alpha)}=(X_k^{(\alpha)}-x^*)/g(\alpha)$, we have
\begin{align}\label{eq:50}
Y_{k+1}^{(\alpha)}
=Y_k^{(\alpha)}+\left(\frac{\alpha}{g(\alpha)}\right)^2\frac{g(\alpha)F(Y_k^{(\alpha)}g(\alpha)+x^*)}{\alpha}+\frac{\alpha}{g(\alpha)}w_k.
\end{align}
In view of the previous equation and the empirical observations in the previous section, we see that we need to choose a scaling function $g(\alpha)$ such that the following condition is satisfied.
\begin{condition}\label{con:1}
The scaling function $g(\cdot)$ should be chosen such that
\begin{enumerate}[(1)]
\item $\lim_{\alpha\rightarrow 0}\frac{\alpha}{g(\alpha)}=0$ and $\lim_{\alpha\rightarrow 0}g(\alpha)=0$
\item The function $\Tilde{F}:\mathbb{R}^d\mapsto\mathbb{R}^d$ defined by $\Tilde{F}(y)=\lim_{\alpha\rightarrow 0}\frac{g(\alpha)F(yg(\alpha)+x^*)}{\alpha}$
is a nontrivial function in the sense that $\Tilde{F}(\cdot)$ is not identically equal to zero or infinity.
\end{enumerate}
\end{condition}
We next verify the choice of scaling functions in Section \ref{sec:distribution} using our proposed Condition \ref{con:1}. For SGD with a smooth and strong convex objective, since
\begin{align*}
\sigma\|x-x^*\|_2\leq \|\nabla f(x)-\nabla f(x^*)\|_2=\|\nabla f(x)\|_2\leq L\|x-x^*\|_2,\quad \forall\;x\in\mathbb{R}^d,
\end{align*}
we have
\begin{align*}
\sigma\frac{g(\alpha)^2}{\alpha}\|y\|_2\leq \left\|\frac{g(\alpha) \nabla f(g(\alpha) y+x^*)}{\alpha}\right\|_2\leq L\frac{g(\alpha)^2}{\alpha}\|y\|_2.
\end{align*}
In view of the previous inequality and Condition \ref{con:1}, it is clear that the only possible choice of $g(\alpha)$ is $g(\alpha)=\sqrt{\alpha}$.
For linear SA algorithm studied in Section \ref{subsec:linear_SA}, one can also easily show using Condition \ref{con:1} that $g(\alpha)=\sqrt{\alpha}$. As for contractive SA studied in Section \ref{subsec:contractive_SA}, using the contraction property and we have
\begin{align*}
(1-\gamma)\|x-x^*\|_\mu\leq \|\mathcal{T}(x)-x\|_\mu=\|\mathcal{T}(x)-\mathcal{T}(x^*)-(x-x^*)\|_\mu\leq (1+\gamma)\|x-x^*\|_\mu.
\end{align*}
It follows that
\begin{align*}
\frac{g(\alpha)^2}{\alpha}(1-\gamma)\|y\|_\mu\leq \left\|\frac{g(\alpha) [\mathcal{T}(g(\alpha) y+x^*)-(g(\alpha) y+x^*)]}{\alpha}\right\|_\mu\leq \frac{g(\alpha)^2}{\alpha}(1+\gamma)\|y\|_\mu.
\end{align*}
Since all norms are ``equivalent'' in finite dimensional spaces, the previous inequality implies that we must choose $g(\alpha)=\sqrt{\alpha}$.
Now to further verify the correctness of the scaling function suggested by Condition \ref{con:1}, consider the SGD algorithm
\begin{align*}
X_{k+1}^{(\alpha)}=X_k^{(\alpha)}+\alpha(-\nabla f(X_k^{(\alpha)})+w_k)
\end{align*}
with the following two choices of objective functions: (1) $f(x)=e^{x^2}$, and (2) $f(x)=\frac{x^4}{4}+\frac{\sin^2(x)}{2}$. Note that in these two cases the function $f(\cdot)$ is not smooth and strongly convex.
\textbf{Case 1.} In the first case where $f(x)=e^{x^2}$, since
\begin{align*}
\left\|\frac{g(\alpha) F(yg(\alpha))}{\alpha}\right\|_2=\frac{g(\alpha)^2}{\alpha}2|y|e^{(yg(\alpha))^2},
\end{align*}
when choosing $g(\alpha)=\sqrt{\alpha}$, we have $\Tilde{F}(y)=\lim_{\alpha\rightarrow 0}\frac{g(\alpha)^2}{\alpha}2ye^{(yg(\alpha))^2}=2y$.
One interesting implication of this example is the following. In the three SA algorithms considered in Section \ref{sec:distribution}, it seems that it is the function $F(\cdot)$ that appears in the algorithm determines the distribution of $Y$. However, the above example suggests that it is the function $\Tilde{F}(\cdot)$ of Condition \ref{con:1} instead of $F(\cdot)$ that directly impacts the distribution $Y$. In SGD with a smooth and strongly convex objective, linear SA, and contractive SA, $F(\cdot)$ and $\Tilde{F}(\cdot)$ happen to coincides, but this is in general not the case. In fact, since we have $\frac{de^{x^2}}{dx}=\sum_{i=1}^\infty (2i)x^{2i-1}$ by Taylor series, $\Tilde{F}(\cdot)$ in this example is exactly the dominant terms appears in the Taylor series. In addition, this suggests that the distribution of the limiting random variable $Y$ has a density function proportional to $e^{\beta' x^2}$, where $\beta'$ is a numerical constant.
We next verify this choice of $g(\alpha)$ and the distribution of $Y^{(\alpha)}$ for small enough $\alpha$ using numerical simulation in the following.
\begin{figure}[h]
\begin{minipage}{0.45\textwidth}
\centering
\includegraphics[width=\linewidth]{ex2_1_2.png}
\caption{Estimated Density Functions when choosing $g(\alpha)=\alpha^{1/2}$}
\label{fig:4}
\end{minipage}
\hfill
\begin{minipage}{0.45\textwidth}
\centering
\includegraphics[width=\linewidth]{ex2_1_2_line.png}
\caption{$\log(p_Y(y))$ as a function of $y^2$}\label{fig:5}
\end{minipage}
\end{figure}
We see from Figure \ref{fig:4} that with the scaling function $g(\alpha)=\sqrt{\alpha}$, the empirical density function of the random variable $Y^{(\alpha)}$ seems to converge. and Figure \ref{fig:5} further justifies this result by showing that the density function $p_Y(y)$ of the distribution of $Y$ in this case is proportional to $e^{\beta' x^2}$, where $\beta'$ is a numerical constant.
\textbf{Case 2.} Now consider case where $f(x)=\frac{x^4}{4}+\frac{\sin^2(x)}{2}$. Observe that
\begin{align*}
\left\|\frac{g(\alpha)F(yg(\alpha))}{\alpha}\right\|_2=\frac{g(\alpha)}{\alpha}|y^3g(\alpha)^3+sin(yg(\alpha))cos(yg(\alpha))|.
\end{align*}
Since $\lim_{x\rightarrow 0}\frac{sin(x)}{x}=1$, the only possible choice of the scaling function $g(\alpha)$ to satisfy Condition \ref{con:1} (2) is $g(\alpha)=\sqrt{\alpha}$. In this case, we have $\Tilde{F}(y)=\lim_{\alpha\rightarrow 0} \frac{1}{\sqrt{\alpha}}y^3\alpha^{3/2}+sin(y\sqrt{\alpha})cos(y\sqrt{\alpha})=y$ by L'Hôpital's rule. This is another example where $F(\cdot)\neq \Tilde{F}(\cdot)$. In fact, since $x^4$ is dominated by $sin^2(x)$ as $x$ approaches $x^*$ (which is $0$), the scaling function and the function $\Tilde{F}(\cdot)$ are determined only by the dominant term.
.
Similarly, we verify this choice of scaling function via numerical experiments. In Figures \ref{fig:10} and \ref{fig:11}, we plot the empirical density function of the random variable $Y^{(\alpha)}$ for different stepsize $\alpha$, and see if the density function converges as $\alpha$ goes to zero. The results suggest that $g(\alpha)=\alpha^{1/2}$ seems to be the correct scaling. To further verify this result, we plot the logarithmic function of the empirical density of $Y^{(\alpha)}$ as a function of $y^2$ and observe straight lines. Therefore, the distribution of $Y^{(\alpha)}$ is proportional to $e^{\beta'' x^2}$, where $\beta''$ is a numerical constants.
\begin{figure}[h]
\begin{minipage}{0.45\textwidth}
\centering
\includegraphics[width=\linewidth]{xcomposite_func_1_2.png}
\caption{Estimated Density Functions when choosing $g(\alpha)=\alpha^{1/2}$}
\label{fig:10}
\end{minipage}
\hspace{0.02\textwidth}
\begin{minipage}{0.45\textwidth}
\centering
\includegraphics[width=\linewidth]{xcomposite_func_1_4.png}
\caption{Estimated Density Functions when choosing $g(\alpha)=\alpha^{1/4}$}
\label{fig:11}
\end{minipage}
\vfill
\begin{minipage}{0.45\textwidth}
\centering
\includegraphics[width=\linewidth]{x2_1_4_line_composite_func.png}
\caption{$\log(p_Y(y))$ as a function of $y^2$}
\label{fig:12}
\end{minipage}
\end{figure}
\subsection{Connection to Euler-Maruyama Discretization Scheme for Approximating SDE}
The choice of the scaling function suggested by Condition \ref{con:1} has an insightful connection to the Euler-Maruyama discretization scheme for approximate the solution of an SDE, as elaborated below. Let $(B_t)_{t\geq 0}$ be a Brownian motion. Consider the following SDE:
\begin{align}\label{eq:SDE}
dX_t=F(X_t)dt+dB_t
\end{align}
with initial condition $X_0$. The Euler-Maruyama discretization $\{\hat{X}_k\}$ to the solution $(X_t)$ of SDE (\ref{eq:SDE}) is defined as follows. Let $\Delta t$ be the discretization accuracy. Set $\hat{X}_0=X_0$, and recursively define $\hat{X}_k$ according to
\begin{align*}
\hat{X}_{k+1}=\hat{X}_k+\Delta t F(\hat{X}_k)+ (B_{(k+1)\Delta t}-B_{k \Delta t}).
\end{align*}
Since $(B_t)_{t\geq 0}$ is a Brownian motion, we have $(B_{(k+1)\Delta t}-B_{k \Delta t})\sim \mathcal{N}(0,\Delta t)$. Therefore, by letting $\{Z_k\}$ be an i.i.d. sequence of standard normal random variables, we can rewrite the previous equation as
\begin{align}\label{eq:Euler-discretization}
\hat{X}_{k+1}=\hat{X}_k+\Delta t F(\hat{X}_k)+ \sqrt{\Delta t} Z_k.
\end{align}
The approximation property of the Euler-Maruyama discretization (\ref{eq:Euler-discretization}) to its corresponding SDE (\ref{eq:SDE}) has been studied in the literature, see \cite{mou2019improved}. Specifically, it was shown that under some mild conditions on $F(\cdot)$, the Euler-Maruyama scheme is known to have the first-order accuracy of the SDE (\ref{eq:SDE}). As a consequence, intuitively, when $(X_t)_{t\geq 0}$ has a stationary distribution $\mu$, the limiting distribution $\mu_{\Delta t}$ of $\{\hat{X}_k\}$ as a function of the discretization accuracy $\Delta t$ should converge weakly to $\mu$ as $\Delta t$ tends to zero. If we view the discretization accuracy $\Delta t$ as the stepsize in Eq. (\ref{eq:Euler-discretization}). In order for $\mu_{\Delta t}$ to converge to some nontrivial distribution $\mu$ as $\Delta t$ tends to zero, it is important to notice that the scaling factor of the noise $Z_k$ in terms of $\Delta t$ must be order-wise equal to the square root of the scaling factor of $F(\hat{X}_k)$. This observation coincides with Eq. (\ref{eq:50}) in the previous section, which eventually leads to our Condition \ref{con:1}.
\section{Conclusion}\label{sec:conclusion}
In this paper, we characterize the asymptotic stationary distribution of properly centered scaled iterate of SA algorithms. In particular, we show that for (1) SGD with smooth and strongly convex objective, (2) linear SA, and (3) contractive SA, the scaling function is $g(\alpha)=\sqrt{\alpha}$ and the corresponding stationary distribution is Gaussian. For SA beyond these cases, we empirical show that the stationary distribution need not be Gaussian, and provide a method for determine the suitable scaling function. Since our paper is the first study for this problem, it might open a door for research in this direction.
\bibliographystyle{imsart-nameyear}
|
\section{Introduction}
\label{sec:intro}
The arrival directions of Galactic cosmic rays (CRs) are highly isotropic due to their interactions with interstellar turbulent magnetic fields. This random scattering process effectively isotropizes the arrival directions of CRs and leads to their diffusive transport~\citep{1964ocr..book.....G}. However, current high-statistics observatories like IceCube~\citep{2019ApJ...871...96A} and HAWC~\citep{2018ApJ...865...57A} have observed significant deviations from isotropy down to angular scales of $10^\circ$ at energies between a few TeV and a few PeV. These deviations from isotropy can be quantified by the angular power spectrum, which is defined as
\begin{equation}
C_\ell(t) = \frac{1}{4\pi}\int\mathrm{d}\uvct{p}_A\int\mathrm{d}\uvct{p}_B P_\ell(\uvct{p}_A\cdot\uvct{p}_B)f_A(t)f_B(t) \, . \label{eqn:def_APS}
\end{equation}
Here, $\uvct{p}_i=\vct{p}_i/|\vct{p}_i|$ are unit momentum vectors, $P_\ell$ are the Legendre polynomials of degree $\ell$, and \mbox{$f_i(t) \equiv f(\vct{r}_\oplus,\vct{p}_i,t)$} is the phase-space density measured by an observer at position $\vct{r}_\oplus$ and time $t$ with momentum $\vct{p}_i$.
In fact, small-scale anisotropies have been the subject of intensive research not only observationally but also theoretically. Many different scenarios have been proposed to explain the presence of anisotropies at small angular scales which includes, for example, effects of the heliosphere~\citep{Drury:2013uka,2014ApJ...790....5Z,2014Sci...343..988S}, non-uniform pitch-angle diffusion~\citep{2010ApJ...721..750M}, non-diffusive propagation of Galactic CRs~\citep{battaner2011,harding2016}, and also more exotic explanations~\citep{kotera2013}.
Interestingly, it has been suggested that the turbulent Galactic magnetic field could be the source of small-scale anisotropies \citep{giacinti2012,2014PhRvL.112b1101A,2015ApJ...815L...2A,battaner2015,lopez-barquero2016,2019JCAP...11..048M}. This idea has been investigated using test particle simulations of CRs in synthetic turbulence where sky maps of the CR arrival directions can be obtained from the phase-space density back-tracked along the CR trajectories to an earlier time (see \citealt{2017PrPNP..94..184A} for a review). Most of the previous studies, however, limit themselves to particles with a ratio of gyroradius to correlation length $\rho \equiv r_g/L_c \sim 10^{-2}-10^{1}$. This corresponds to energies much larger than what is relevant for observational data, e.g.\ $\rho \sim 10^{-4}$, for a $10\,\text{TeV}$ particle in a $4\,\mu \text{G}$ RMS magnetic field of correlation length $L_c \simeq 30\,\text{pc}$. Comparison with observed anisotropies is thus based on some extrapolations in energy.
More importantly, the standard picture of CR transport known as Quasi-Linear Theory (QLT)~\citep{1966ApJ...146..480J,1966PhFl....9.2377K,1967PhFl...10.2620H,1970ApJ...162.1049H} fails to explain the observed anisotropies at scales smaller than the dipole. This can be traced back to the fact that QLT allows computing only the phase-space density averaged over an ensemble of turbulent magnetic fields, $\langle f\rangle$. Computing the ensemble-averaged angular power spectrum via Eq.~\eqref{eqn:def_APS}, however, requires the two-point function of phase-space densities, $\langle f_A f_B\rangle$. Use of QLT therefore relies on assuming statistical independence, $\langle f_Af_B\rangle = \langle f_A\rangle\langle f_B\rangle$.
Correlations are expected to be present, however: Imagine two particles arriving under a small angle. For about a scattering time before observation, they have experienced the same magnetic field configuration. Fluxes from nearby directions will therefore be similar and $\langle f_A f_B \rangle \geq \langle f_A\rangle \langle f_B \rangle$. \citet{2019JCAP...11..048M} take into account these correlations and put forward a model to predict the angular power spectrum based on a perturbative expansion of the time-evolution operator. However, a rather unrealistic white-noise power spectrum of turbulence was adopted to allow for some explicit analytical results.
The aim of this work is, thus, to further improve our understanding of turbulence-induced small-scale anisotropies. A better understanding of these small-scale anisotropies will allow deriving independent constraints on the nature of turbulence in our Galactic neighborhood, e.g.\ on the CR scattering time or the power spectrum of the local turbulent magnetic field. These quantities are crucial in achieving a better precision of predictions of CR fluxes, but currently, they are only poorly constrained.
The outline of this paper is as follows. In Sec.~\ref{sec:method}, we will outline the methodological approach which is both numerical and analytical. Sec.~\ref{sec:numerical} introduces the concept of backtracking and gives some details on the numerical test particle simulations. Sec.~\ref{sec:analytical} contains the analytical framework introduced by \citet{2019JCAP...11..048M} for computing the angular power spectrum, extended to the case of slab turbulence. In Sec.~\ref{sec:results}, we will present the numerical and analytical results and compare the computed angular power spectra in the energy range relevant to observations. We will summarise our findings and conclude in Sec.~\ref{sec:summary}.
\section{Method}
\label{sec:method}
\subsection{Numerical simulations}
\label{sec:numerical}
\subsubsection{The concept of backtracking}
The origin of small-scale anisotropies can be understood by exploiting Liouville's theorem that states that the phase-space density is conserved along particle trajectories through phase-space~\citep{Goldstein:2001}. We can therefore evaluate the phase-space density for an arbitrary point $\{\vct{r}(t), \vct{p}(t)\}$ at time $t'=t$, $f(\vct{r}(t), \vct{p}(t), t)$, if we know the phase-space density at another time $t' = t_0$ along the trajectory $\{\vct{r}(t'), \vct{p}(t')\}$ from that initial point,
\begin{equation}
f(\vct{r}(t), \vct{p}(t), t) = f(\vct{r}(t_0), \vct{p}(t_0), t_0) \, .
\label{eq:Liouville}
\end{equation}
Applied to CRs, out of all possible trajectories we pick those that converge at the observer's position $\vct{r}(t) = \vct{r}_{\oplus}$ with $\vct{p}(t) = p \uvct{p}_{\oplus}$ at time $t$. (For simplicity, we have assumed that the absolute value of the momentum $p = |\vct{p}|$ does not change which is a very good approximation for CR nuclei of rigidity above a few GV.) The angular dependence in $f(\vct{r}_{\oplus}, p \uvct{p}_{\oplus}, t)$ then constitutes the CR sky map. This so-called Liouville mapping or back-tracking is also the basis for the computation of the angular power spectrum from test particle simulations of CRs in synthetic turbulence, to be presented in Sec.~\ref{sec:simulations}. We assume the phase-space density to be the quasi-steady solution of a diffusion equation at the earlier time $t_0$ (see e.g. \citealt{2017PrPNP..94..184A}),
\begin{equation}
\avg{ f(\vct{r}(t_0),\uvct{p}(t_0),t_0) } = \bar{f} + \mathbf{r}(t_0)\cdot \nabla\bar{f} - 3\hat{\mathbf{p}}(t_0)\cdot \mathbf{K}\cdot \nabla \bar{f} \, ,
\label{eqn:quasi_steady}
\end{equation}
parametrized by a time-independent isotropic part $\bar{f}$ and its gradient, $\nabla\bar{f}$. Note that in general the phase-space density $f$ in a particular realisation of the turbulent magnetic field differs from the ensemble average, $\avg{ f }$, that is $\delta f \equiv f - \avg{ f } \neq 0$. In backtracking, however, this is negligible since $\delta f(t_0)$ is uncorrelated with $f(t)$ as long as the backtracking time $T \equiv (t - t_0)$ is large enough.
\begin{figure*}[tbh]
\includegraphics[scale=1]{figures/skymap_row1.pdf}\\
\includegraphics[scale=1]{figures/skymap_row35.pdf}\\
\includegraphics[scale=1]{figures/skymap_row100.pdf}\\
\includegraphics[scale=1]{figures/skymap_row700.pdf}
\caption{Trajectories (left column), sky maps (middle column) and angular power spectra (right column) for different times $T$ (top through bottom row). The times $T_1$ through $T_4$ can be considered as backtracking times $(t - t_0)$ when following particles back in time along their trajectories or as the times passed since the preparation of the initial phase-space density at $t_0$. See the text for further explanations.}
\label{fig:illustration}
\end{figure*}
For illustration, we follow a set of trajectories, their sky map and its angular power spectrum as we increase the backtracking time $T$ from $T_1$ to $T_2$, $T_3$ and $T_4$ as shown in Fig.~\ref{fig:illustration}. Backtracking for a small time $T_1$ only (top row), CRs have travelled mostly ballistically, their trajectories being straight lines (top left panel). The phase-space density at the other end of the trajectory $i$ at time $t_0 = t-T_1$ and therefore the flux in a direction $\vhat{p}_i$ at time $t$ is proportional to the projection of this direction onto the gradient $\vct{\nabla} \bar{f}$, $f(\vct{r}_{\oplus}, \vct{p}_i, t) \propto \uvct{p}_i \cdot \vct{\nabla} \bar{f}$. The sky map (top center panel) is then approximately dipolar and the angular power spectrum $C_{\ell}$ (top right panel) vanishes for $\ell > 1$.
If we backtrack for a longer time $T_2 > T_1$ (second row from top), we can see (left panel) that trajectories that arrive at the observer from nearby directions at time $t$ were mostly already closely aligned a time $T_2$ earlier: For instance, a number of trajectories towards the top end of the panel have just started gyrating around the large-scale magnetic field. Another set of trajectories, towards the right side of the panel are instead moving towards the right without much gyration since their momentum vectors are aligned with the background magnetic field. Therefore, the phase-space densities at the ends of the trajectories are correlated for trajectories from nearby observed directions. The sky map (centre panel), while not perfect dipolar anymore, is still rather smooth. There is some angular power (right panel) on scales different from the dipole, but the smaller scales are still largely suppressed.
For even larger backtracking times $T_3 > T_2$ (second row from top), the trajectories (left panel) have further differentiated out, but nearby trajectories are still correlated. On the sky map (centre panel), this is visible as structure on smaller scales. Correspondingly, we now observe power on almost all scales in the angular power spectrum (right panel).
Finally, for very late times $T_4 > T_3$ (bottom row), the particle directions and positions are finally uncorrelated. In the sky map (centre panel) this is reflected in noise on small angular scales, also visible as a flat angular power spectrum for $\ell \gtrsim 10$. We are thus led to conclude that the specific, non-dipolar pattern is imprinted onto the map of CR arrival directions at intermediate times, between the epochs of purely ballistic and of purely diffusive transport. This transition has been the topic of recent interest in the study of CR transport~\citep[e.g.][]{2014ApJ...783...15E,2015PhRvD..92h3003P,2017PhRvD..95b3007M}.
So far, we have read the succession of the trajectories, sky maps and angular power spectra in Fig.~\ref{fig:illustration} from top to bottom as going successively further in backtracking time. It is however perfectly equivalent to read them from top to bottom as going forward in the time that has passed since preparing the phase-space density of Eq.~\eqref{eqn:quasi_steady}. Imagine we \emph{globally} prepare the phase-space density at time $t_0$ to Eq.~\eqref{eqn:quasi_steady} and then observe its evolution at the observer's position. The particles arriving at time $(t_0 + T_1)$ will have been at the positions indicated by the markers in the top left panel of Fig.~\ref{fig:illustration} at time $t_0$. Given Liouville's theorem, the phase-space densities at these points have been conserved along the individual trajectories and are observed at time $(t_0 + T_1)$ as the sky map (middle panel). Correspondingly, the right panel shows the angular power spectrum a time $T_1$ after setting up the initial state. Similarly, the particles arriving at time $(t_0 + T_2)$ (second row) were at the end of the trajectories shown in the left panel at time $t_0$. The sky map observed at time $(t_0 + T_2)$ and its angular power spectrum are shown in the middle and right panels. The same applies to the other rows for times $(t_0 + T_3)$ and $(t_0 + T_4)$.
The succession of sky maps and angular power spectra from top (time $t_0 + T_1$) to bottom (time $t_0 + T_4$) therefore shows the evolution of angular structures as seen by the observer as time progresses. Initially, all the angular power is on large scales, but at later times power starts appearing on smaller scales. While the processes of this generation has a \emph{non-local} character (due to the Liouville mapping), in the ensemble average we can describe it \emph{locally} as an evolution of angular power from large scales to smaller scales. Given the discussion above, it is clear that the angular power on small scales is produced by mixing of power from larger into smaller scales: Structure coherent on a certain angular scale starts developing substructure on smaller scales as the set of trajectories corresponding to the initial scale start diverging.
\subsubsection{Synthetic turbulence}
\label{sec:simulations}
We start by numerically simulating test particles in synthetic turbulence~\citep[e.g.][]{2020Ap&SS.365..135M}. The monoenergetic test particles are initialized
with isotropic directions on a HEALPix~\citep{2005ApJ...622..759G} grid with $N_\text{side} = 256~\mathrm{or}~512$ leading to a total number of test particles $N_\text{particles} \sim 8\times 10^5~\text{and}~3\times 10^6$. These particles are then tracked back in time through the magnetic field by solving the Newton-Lorentz equations using the rigidity conserving Boris method~\citep{boris}. As the test particles do not interact with each other or backreact on the magnetic field this can be parallelised very efficiently. We therefore run these simulations on GPUs which allow for very cost-effective parallelisation \citep{2016NewA...45....1T,GPU-lecturenotes}.
In the literature two different methods have been used to generate synthetic magnetic field turbulence. In the method proposed by \citet{1994ApJ...430L.137G}, the magnetic field is calculated as a superposition of waves. Only the phases and amplitudes for the waves are stored. In the other method the magnetic field is set up on a grid in Fourier space, transformed to and saved in real space. This has the advantage that no large sums have to be evaluated at every particle position and time step. The magnetic field evaluation is reduced to a simple interpolation between grid points \citep{Schlegel:2019uww}. The disadvantage is the large amount of memory required to store the entire field grid.
The rigidity of particles required to compare to observational data from IceCube and HAWC is of the order of $10\,\text{TV}$ assuming only protons. Therefore the resulting gyroradii that need to be resolved in our simulations are \mbox{$r_g/L_c\sim 8.3\times 10^{-5}$} given $B_{\rm RMS}\simeq 4\,\mu$G and $L_c\simeq 30 \, \text{pc}$. Even though there could be artefacts due to grid periodicity, it has proven sufficient to make the grid a factor of $8$ larger than the largest wavelength. To get the correct normalization of the angular power spectrum the smallest wavelength has to be much smaller than the gyroradius. In practice we choose the smallest wavelength to be a factor $\sim 150$ smaller than the gyroradius at the smallest rigidity and resolve it by at least 10 grid points. Spanning this large dynamical range with a single grid would require at least $n=675,000,000$ grid points. To reduce the memory requirement on our GPUs we therefore use 3 nested grids with different grid spacings as proposed by~\citet{2012JCAP...07..031G}. Each individual grid covers a part of the magnetic field power spectrum. The 3 grids are then superimposed \citep[see][for an illustration]{2020Ap&SS.365..135M}.
Equation \eqref{eq:Liouville} is applied for the simulated CR trajectories to provide a sky map for a particular realization of turbulence. We then calculate and average the angular power spectra for all realizations to obtain the ensemble-averaged numerical angular power spectrum $\langle C_\ell\rangle^{\rm num}$. At this point, we redefine the angular power spectrum from Eq.~\eqref{eqn:def_APS} as follows
\begin{equation}
\langle C_\ell\rangle = \frac{1}{4\pi}\int\mathrm{d}\uvct{p}_A\int\mathrm{d}\uvct{p}_B P_\ell(\uvct{p}_A\cdot\uvct{p}_B)\frac{\langle f_Af_B\rangle}{\bar{f}^2} \, . \label{eqn:redef_APS}
\end{equation}
which essentially replaces the phase-space density by the relative intensity of CRs as commonly adopted for the observed angular power spectrum.
We note also that even for a large number of particles the higher multipoles of the angular power spectrum are strongly effected by noise. For large backtracking times the particle transport is dominated by diffusion. The sky map therefore gets dominated by Gaussian noise with variance
\begin{equation}
\sigma^2 =\langle r_i r_j\rangle \frac{\partial_i \bar{f}\partial_j\bar{f}}{\bar{f}^2}= 2T K_{ij}\frac{\partial_i \bar{f}\partial_j\bar{f}}{\bar{f}^2}.
\label{eqn:sky_variance}
\end{equation}
The angular power spectrum coming from this noise contribution can be calculated by substituting \mbox{$\langle f_A f_B \rangle = \sigma^2 \delta^{(3)} (\uvct{p}_A - \uvct{p}_B)$} into the definition of the angular power spectrum given in Eq.~\eqref{eqn:redef_APS}. In the case of a discrete pixelized sky map this gives
\begin{equation}
\mathcal{N}_\ell = \frac{1}{4\pi}\sum_i \Delta\Omega \sum_j \Delta \Omega P_\ell(\uvct{p}_i\cdot\uvct{p}_j) \sigma^2 \delta_{ij},
\end{equation}
with the pixel size $\Delta \Omega = 4\pi/N_\text{pix}$, where $N_\text{pix}$ is the total number of pixels in the arrival direction sky map.
Evaluating the angular power spectrum of the noise contribution, using that $P_\ell(1) = 1$, results in a constant contribution for all $\ell$ that is given by~\citep{2015ApJ...815L...2A}
\begin{equation}
\mathcal{N}_\ell = \frac{4\pi}{N_\text{pix}} 2 T K_{ij} \frac{\partial_i \bar{f}\partial_j \bar{f}}{\bar{f}^2} .\label{eqn:noise}
\end{equation}
It has been argued by \citet{2015ApJ...815L...2A} that the best estimator for the angular power spectrum in this case is $\langle C_\ell\rangle^{\rm sub}=\langle C_\ell\rangle^{\rm num} -\mathcal{N}_\ell$ which we refer to throughout this work as the ensemble-averaged, noise-subtracted angular power spectrum. The variance of $\langle C_\ell\rangle^{\rm sub}$ can be estimated as $2\mathcal{N}_\ell^2/(2\ell+1)$~\citep{2015MNRAS.448.2854C}.
\subsubsection{Parameter values}
\label{sec:parameters}
For all simlations we choose an outer scale of turbulence $L_\text{max} = 150\,\text{pc}$, that was used also in previous studies~\citep[e.g.][]{2018JCAP...07..051G}. For a kolmogorov turbulence like power spectrum this corresponds to a correlation length $L_c \simeq L_\text{max}/5 = 30\,\text{pc}$~\citep{2002JHEP...03..045H}. Current observational constraints on the outer scale differ significantly~\citep{2013A&A...558A..72I} depending on galactic latitude~\citep{2010ApJ...710..853C} and the position in the galactic plane~\citep{2008ApJ...680..362H}. Since the simulations are only sensitive to the normalized rigidity $\rho=r_g/L_c$, simulations at one $\rho$ can correspond to a range of outer scales and rigidities. The total root mean square magnetic field strength is set to $\sqrt{B_0^2+\langle\delta B^2\rangle} = 4\,\mu\text{G}$~\citep{2013pss5.book..641B}, motivated by energy equipartition calculations. The turbulence level indicated by modelling efforts of the galactic magnetic field, based on the observed synchrotron intensity and polarization, $\eta = \langle\delta B^2\rangle/(B_0^2+\langle\delta B^2\rangle) \sim 0.7-0.9$~\citep{2013pss5.book..641B,2017A&A...603A..15O} seems to be incompatible with the CR diffusion coefficient obtained from measurements of the Boron to Carbon ratio~\citep{Evoli:2019wwu,Weinrich:2020cmw} at GeV energies under standard assumptions~\citep{2018JCAP...07..051G}, pointing to much smaller turbulence levels. We will therefore focus on an intermediate turbulence level of $\eta=0.5$. The results of the numerical simulations will be presented in Sec.~\ref{sec:results}.
\subsection{Analytical computation}
\label{sec:analytical}
\subsubsection{The mixing equation}
\label{sec:mixing-equation}
We will start by reviewing the diagrammatic formalism for solving the Liouville equation that describes the time evolution of the phase-space density, as outlined by \citet{2019JCAP...11..048M}. We shall work in natural units ($c=1$) and consider only relativistic particles for simplicity. The Liouville equation for relativistic charged particles propagating through the magnetic field with both static regular and turbulent components, denoted as $\vct{B}_0$ and $\delta\vct{B}$, could be written as
\begin{equation}
\label{eq:vlasov}
\partial_t f + \uvct{p}\cdot\vct{\nabla} f + \mathcal{L} f = -\delta \mathcal{L} f \, ,
\end{equation}
where $\mathcal{L} = -\imath \vct{\Omega} \cdot \vct{L}$ and $\delta\mathcal{L} = - \imath \vct{\omega} \cdot \vct{L}$ represent the deterministic and stochastic Liouville operators. Here, $\vct{\Omega} = q \vct{B}_0/p$ and $\vct{\omega} = q\delta\vct{B}/p$ are the gyrovectors in the regular and turbulent magnetic fields, respectively. The $L_m=\imath\varepsilon_{mnk}p_n\partial/\partial p_{k}$ are angular momentum operators that obey the usual commutation relations $[ L_m, L_n ] = \imath \epsilon_{mnk} L_k$.
In the absence of any spatial dependence, $\vct\nabla f \equiv 0$, this equation can be solved formally,
\begin{equation}
f(\vct{p}, t) = U_{t,t_0} f(\vct{p}, t_0) \, ,
\end{equation}
by introducing a time evolution operator or propagator
\begin{equation}
U_{t,t_0} = \mathcal{T}\exp{\left[-\int_{t_0}^{t}\mathrm{d}t'(\mathcal{L}+\delta\mathcal{L}(t'))\right]} \, , \label{eqn:time_evolution_operator}
\end{equation}
with the usual time-ordered ("latest-to-left") exponential. The time evolution operator depends, through $\delta\mathcal{L}$, on the specific realisation of the turbulent magnetic field, $\vct{\omega}$ to which we do not have access in any practical sense. What we can predict, however, are statistical averages. The ensemble averaged phases space density $\langle f \rangle$, for instance satisfies,
\begin{equation}
\langle f(\vct{p}, t) \rangle = \langle U_{t,t_0} f(\vct{p}, t_0) \rangle \simeq \langle U_{t,t_0} \rangle \langle f(\vct{p}, t_0) \rangle \, .
\end{equation}
In the last step, we have assumed that $(t - t_0)$ is so large, that correlations between the initial state $f(\vct{p}, t_0)$ and the time evolution operator $U_{t,t_0}$ can be ignored.
Relaxing the requirement of homogeneity, we can expand the phase-space density around the position of the observer $\vct{r}_\oplus$ as $f(\vct{r},\vct{p},t) \simeq f_\oplus + (\vct{r} - \vct{r}_\oplus) \cdot \bm\nabla\bar{f}$, where $\bm\nabla\bar{f}$ denotes a CR gradient at $\vct{r}_\oplus$, and insert this into Eq.~\eqref{eq:vlasov} to obtain an equation for the time evolution of the phase-space density at the position of the observer
\begin{equation}
\label{eq:vlasov2}
\partial_t f_\oplus + \mathcal{L} f_\oplus + \delta\mathcal{L}f_\oplus \simeq - \uvct{p}\cdot \nabla\bar{f}.
\end{equation}
As we will show later, the right hand side sources a dipole anisotropy coming from the assumption of a CR gradient. The formal solution to Eq.~\eqref{eq:vlasov2} can then be written as
\begin{equation}
\label{eq:solution}
\begin{aligned}
f_\oplus(\vct{p},t)
&\simeq U_{t,t_0}f_{\oplus}(\vct{p},t_0)-\int^t_{t_0}\mathrm{d} t' U_{t,t'} \uvct{p}(t) \cdot \vct\nabla \bar{f} \\
&=U_{t,t_0}f_\oplus(\vct{p},t_0) + \Delta\vct{r}(t_0) \cdot \vct\nabla \bar{f}.
\end{aligned}
\end{equation}
where we have denoted $\Delta\vct{r}(t_0)=\vct{r}(t_0) - \vct{r}_\oplus$. Note also that we have employed $U_{t,t'} \uvct{p}(t)=\uvct{p}(t')$ for the last identity.
As briefly discussed in Sec.~\ref{sec:intro}, the trajectories of a pair of CRs are expected to stay correlated for a finite amount of time prior to observation which results in the presence of small-scale anisotropies. This means that in order to predict the angular power spectrum, we have to examine the ensemble average for the products of phase-space densities $\left\langle f_\oplus(\vct{p}_A,t) f^*_\oplus(\vct{p}_B,t)\right\rangle$ which from Eq.~\eqref{eq:solution} could be written as
\begin{equation}
\label{eq:formal_solution_product}
\begin{aligned}
\avg{f_A(t) f^*_B(t)} & \simeq \avg{\Up{A}{t}{t_0} \Up{B*}{t}{t_0}} \avg{f_A(t_0) f_B^*(t_0)} \\
& +\avg{ \big( \Delta \vct{r}_A(t_0) \cdot \vct\nabla \bar{f} \big) \Up{B*}{t}{t_0} f^*_B(t_0)} \\
& + \avg{ \big( \Delta \vct{r}_B(t_0) \cdot \vct\nabla \bar{f}^* \big) \Up{A}{t}{t_0}f_A(t_0) } \\
& \mkern-9mu + \avg{ \big( \Delta \vct{r}_A(t_0) \cdot \vct\nabla \bar{f} \big) \big( \Delta \vct{r}_B(t_0) \cdot \vct\nabla \bar{f}^* \big) }
\end{aligned}
\end{equation}
where we have again adopted the shorthand notation $f_A(t)=f_\oplus(\vct{p}_A,t)$ and $f_B(t)=f_\oplus(\vct{p}_B,t)$. In addition, we have also assumed that the correlations between the propagators and the initial state $f_\oplus(\vct{p},t_0)$ are negligible. Note that we have used the asterisk symbol to mark the complex conjugates of both functions and operators (including the angular momentum operators).
If we could evaluate the pair propagator $\avg{\Up{A}{t}{t_0} \Up{B*}{t}{t_0}}$, the angular power spectrum could be computed at anytime for a given initial state $f(\vct{p},t_0)$ using \mbox{Eqs. \eqref{eqn:redef_APS} and \eqref{eq:formal_solution_product}}. However, it might be difficult to estimate directly the ensemble average of all the terms on the RHS of Eq.~\eqref{eq:formal_solution_product} for an arbitrary period of time $(t-t_0)$ due to the stochastic nature of particles' trajectories in the turbulent magnetic field. Thus, we shall first limit ourselves to the evolution of the angular power spectrum over a short period of time from $t_0$ to $(t_0+\Delta T)$ with small enough $\Delta T$ such that we could approximate $U_{t,t_0}\simeq 1 + \mathcal{O}(\Delta T)$ and also $\lim_{\Delta T\rightarrow 0}\Delta\vct{r}(t_0)\simeq -\mu\Delta T\uvct{z}$. (Note that we have adopted a coordinate system with the $z$-axis directed along $\vct{B}_0$ for definiteness.) Equation~\eqref{eq:formal_solution_product} in this case becomes
\begin{equation}
\begin{aligned}
& \mkern-18mu \avg{f_A(t_0+\Delta T) f^*_B(t_0+\Delta T)} \\
& \simeq \avg{\Up{A}{t_0+\Delta T}{t_0} \Up{B*}{t_0+\Delta T}{t_0}} \avg{f_A(t_0) f_B^*(t_0)} \\
& + (\mu_A\Delta T\partial_z\bar{f})\avg{f^*_B(t_0)} + (\mu_B\Delta T\partial_z\bar{f})\avg{f_A(t_0)} \\
& + \mathcal{O}(\Delta T^2) \, .
\label{eqn:product-small-time}
\end{aligned}
\end{equation}
We can now integrate both sides of Eq.~\eqref{eqn:product-small-time} and adopt the definition of the angular power spectrum in Eq.~\eqref{eqn:redef_APS} to obtain an equation for the time evolution of the angular power spectrum over a small time step $\Delta T$,
\begin{equation}
\begin{aligned}
&\langle C_\ell(t_0+\Delta T)\rangle \simeq \, \! \frac{1}{4\pi} \! \int\mathrm{d}\uvct{p}_A \! \int\mathrm{d}\uvct{p}_B P_\ell(\uvct{p}_A\cdot\uvct{p}_B) \\
& \times \Bigg[ \avg{ \Up{A}{t_0+\Delta T}{t_0} \Up{B*}{t_0+\Delta T}{t_0} } \frac{\avg{f_A(t_0) f_B^*(t_0)}}{\bar{f}^2} \\
&+ \Delta T\frac{\partial_z\bar{f}}{\bar{f}^2}\Big( \mu_A\avg{f^*_B(t_0)} + \mu_B\avg{f_A(t_0)} \vphantom{\frac{a}{b}} \Big) \Bigg] \, .
\label{eq:product-small-time}
\end{aligned}
\end{equation}
The expression above could be re-written in a more compact form by introducing the so-called mixing matrix as presented in \citet{2019JCAP...11..048M},
\begin{widetext}
\begin{equation}
M_{\ell\ell_0}(\Delta T) = \frac{1}{4\pi}\int\mathrm{d}\uvct{p}_A\int\mathrm{d}\uvct{p}_B P_\ell(\uvct{p}_A\cdot\uvct{p}_B)\langle U^A_{t_0+\Delta T,t_0}U^{B*}_{t_0+\Delta T,t_0}\rangle \frac{2\ell_0+1}{4\pi}P_{\ell_0}(\uvct{p}_A\cdot\uvct{p}_B).
\label{eq:mixing-matrix}
\end{equation}
\end{widetext}
Note that by construction the mixing matrix depends only on the small time step $\Delta T$ in the case of static turbulence. We shall further assume that the quasi-stationary solution of the ensemble averaged phase-space density applies, meaning $\avg{f(\uvct{p},t_0)}\simeq\bar{f}-3\uvct{p}\cdot\vct{K}\cdot\nabla\bar{f}$, where $\vct{K}$ is the diffusion tensor. This allows us to rewrite the time evolution of the angular power spectrum over a small time step $\Delta T$ as follows
\begin{equation}
\begin{aligned}
\langle C_{\ell}(t_0+\Delta T) \rangle &= \! \sum_{\ell_0 = 0}^{\infty}M_{\ell \ell_0}(\Delta T)C_{\ell_0}(t_0) \\
&+ \delta_{\ell 1}\frac{8\pi}{3}K_{zz}\Delta T\left(\frac{\partial_z\bar{f}}{\bar{f}}\right)^2
\end{aligned}
\label{eq:APS-small-time}
\end{equation}
It is clear from Eq. \eqref{eq:APS-small-time} that the angular power spectrum at some later time $(t_0+\Delta T)$ gets contributions both from the mixing induced by the turbulent magnetic field of the angular power spectrum at earlier time $t_0$ (the first term on the RHS) and the dipole source term (the second term on the RHS). In the following we will suppress the summation symbol and use the Einstein sum convention for the matrix product.
We can generalize this time evolution equation to an arbitrary time $t = t_0 + n\Delta T$ with $n\geq 1$ by applying Eq. \eqref{eq:APS-small-time} consecutively to evolve the angular power spectrum from $C_\ell(t_0)$ to $C_\ell(t)$, yielding
\begin{equation}
\begin{aligned}
&\langle C_{\ell}(t)\rangle
= \left( M_{\ell\ell_0}(\Delta T) \right)^n C_{\ell_0}(t_0) \\
& + \sum_{n_0=1}^{n} \left( M_{\ell\ell_0}(\Delta T) \right)^{n_0-1} \delta_{\ell_0 1} \frac{8\pi}{3} K_{zz} \Delta T \left(\frac{\partial_z\bar{f}}{\bar{f}}\right)^2 \, .
\label{eq:APS-full-time}
\end{aligned}
\end{equation}
It might be instructive to gain some insight into the picture of mixing angular power by considering for example an initial state which has only the dipole meaning $C_\ell(t_0)\neq 0$ only for $\ell=1$ similar to the one illustrated in the right panel of the first row in Fig.~\ref{fig:illustration}.
As we evolve the system in time, the source term (the last term on the RHS of Eq.~\eqref{eq:APS-full-time}) will continuously provide the angular power to the dipole and the mixing matrix starts to transfer angular power from the dipole to smaller scales corresponding to $\ell>1$.
This would result in the angular power spectrum as in the right panels in the other rows of Fig. \ref{fig:illustration}. This mixing will persist until the system reaches a steady-state and the angular power spectrum becomes time-independent with angular power on both large and small scales. Note that we are making the analogy between the angular power spectra in Eq.~\eqref{eq:APS-full-time} and in Fig.~\ref{fig:illustration} only for illustration and they are not exactly the same. This is because Eq.~\eqref{eq:APS-full-time} actually provides the ensemble-averaged angular power spectrum while Fig. \ref{fig:illustration} only represents a specific realization of the angular power spectrum.
Another comment on the nature of the dipole source term is in order. The fact that there is only a dipole source term on the RHS of Eq.~\eqref{eq:APS-small-time} can be traced back to the fact that in the derivation of the ensemble-averaged phase-space density $\avg{ f }$ only the monopole and dipole terms in $\uvct{p}$ are retained~\citep{jones1990}. Source terms with $\ell > 1$ would also appear if the presence of higher order terms in $\uvct{p}$ was considered (see e.g.~\citet{giacinti2017}).
We could now seek the time-independent angular power spectrum by considering the steady-state limit of Eq.~\eqref{eq:APS-small-time} or \eqref{eq:APS-full-time}, that is demanding $C_\ell( t ) = C_\ell (t_0)$ which results in the steady-state equation for the angular power spectrum or \textit{the mixing equation},
\begin{equation}
\label{eq:mm}
\frac{\delta_{\ell\ell_0}-M_{\ell\ell_0}(\Delta T)}{\Delta T} \langle C_{\ell_0}\rangle^{\rm mat} = \frac{8\pi}{3}K_{zz}\left(\frac{\partial_z \bar{f}}{\bar{f}}\right)^2 \delta_{\ell 1}.
\end{equation}
where $\langle C_\ell\rangle^{\rm mat}$ denotes the steady-state ensemble-averaged angular power spectrum from the mixing matrix approach. The right hand side of Eq.~\eqref{eq:mm} is again a dipole term sourced by a CR gradient according to Fick's law.
The left hand side describes how power from the dipole source is mixed into higher multipoles by the mixing matrix.
At this point, it might be appropriate to start a brief discussion on the physical interpretation of $\Delta T$ which is, in fact, the only parameter required to predict the angular power spectrum for a given turbulence model. Since we first specify $\Delta T$ as a small time step in the derivation of Eq. \eqref{eq:APS-small-time}, one might expect that we should take the limit $\Delta T\rightarrow 0$ in the mixing matrix $M_{\ell \ell_0}$. However, we shall see below that the interactions between particles mediated by the turbulent field are essential for the generation of small-scale anisotropies as they allow the mixing of angular power from large to small scales. It might be, therefore, more suitable to choose $\Delta T$ as a timescale on which these interactions occur. In practice, this means that the time integration of the pair propagators in Eq. \eqref{eq:mixing-matrix} for the mixing should be carried out for a finite amount of time. Determining the correct value of $\Delta T$ requires further examination of the time evolution operators and will be done later in Sec.~\ref{sec:OmegaDT}. Nonetheless it is important to note at this point that there are two constraints on this parameter: i) $\Omega\Delta T \gtrsim 2\pi$, to allow for gyroresonant interactions and ii) $\Delta T < \tau_s$ where $\tau_s=K_{zz}/3$ is the scattering time in QLT (setting again $c=1$ in natural unit). The latter constraint is due to the fact that we have assumed unperturbed trajectories in the derivation of the mixing equation.
\subsubsection{The mixing matrix}
\label{sec:mixing_matrix}
In order to compute the mixing matrix $M_{\ell\ell_0}$, we need to expand the time-ordered exponential of Eq.~\eqref{eqn:time_evolution_operator} into an infinite series and compute the expectation values of the individual terms. Note that under the assumption of Gaussianity, the $n$-point functions in the turbulent magnetic field $\vct{\omega}$ only contribute for even $n$. Further, through a cumulant expansion~\citep{1962JPSJ...17.1100K} they can be expressed as sums of permutations of products of $2$-point functions, e.g.,
\begin{equation*}
\begin{aligned}
\langle \vct{\omega}(t_1) \vct{\omega}(t_2) \vct{\omega}(t_3) \vct{\omega}(t_4) \rangle
&\!=\! \langle \vct{\omega}(t_1) \vct{\omega}(t_2) \rangle \langle \vct{\omega}(t_3) \vct{\omega}(t_4) \rangle \\
&\!+\! \langle \vct{\omega}(t_1) \vct{\omega}(t_3) \rangle \langle \vct{\omega}(t_2) \vct{\omega}(t_4) \rangle \\
&\!+\! \langle \vct{\omega}(t_1) \vct{\omega}(t_4) \rangle \langle \vct{\omega}(t_2) \vct{\omega}(t_3) \rangle \, .
\end{aligned}
\end{equation*}
In fact, the Bethe-Salpeter equation \citep{salpeter1951} allows us to expand the pair propagator $\avg{\Up{A}{t}{t_0} \Up{B*}{t}{t_0}}$ in Eq.~\eqref{eq:formal_solution_product} into a perturbative series in powers of the turbulent magnetic field which could be conveniently described with a diagrammatic representation similar to Feynman diagrams in Quantum Field Theory,
\begin{equation}
\begin{matrix}\includegraphics[scale=1.0,trim={1cm 0cm 1cm 0cm}]{figures/double1.pdf}
\end{matrix}
\label{eq:diagrams}
\end{equation}
Here, the solid lines denote the ``free'' propagator \mbox{$U_{t,t_0}^{(0)} = \exp\left[-(t-t_0)\mathcal{L}\right]$} that describes particles following helical trajectories through the regular background magnetic field. The dashed lines then denote correlations between insertions of interactions with the turbulent magnetic field via the stochastic Liouville operator $\delta\mathcal{L}(t')$ at an intermediate time $t'$. All intermediate times are integrated over.
So far, we have discussed the time evolution of the angular power spectrum without specifying the exact form of $M_{\ell \ell_0}$. In fact, the mixing of angular power would happen only if $M_{\ell \ell_0}$ has non-zero off-diagonal elements. As we shall show later, one does not have mixing when considering only self-interacting diagrams in Eq.~\eqref{eq:diagrams} (diagrams with dashed lines connecting the dots on the same solid lines) since they only lead to diagonal matrices. The off-diagonal elements of $M_{\ell \ell_0}$ would actually appear when we take into account ``interaction'' diagrams (diagrams with dashed lines connecting the dots on two different solid lines). It is actually the interaction between two particles mediated by the correlation of magnetic turbulence that induces the correlation of two phase-space densities and leads to mixing of angular power from large to small scales (see Section \ref{sec:calculation-diagrams} for more details).
We could now start to evaluate the steady-state angular power spectrum for a given turbulence model if the mixing matrix for each diagrams in Eq.~\eqref{eq:diagrams} could be calculated and resummed. However, this is quite challenging even for the simplified setup considered by \citet{2019JCAP...11..048M}. Thus, we will limit ourselves to a first order calculation and only compute the diagrams up to and including the first parenthesis of Eq.~\eqref{eq:diagrams}. The difficulty in resumming the diagrams of eq.~\eqref{eq:diagrams} lies in the appearance of interaction diagrams. For the unconnected diagrams alone, the resummation of a subseries, the so-called Bourret series, is possible. This is shown in detail in appendix~\ref{sec:Bourret}. It is then possible to show that the repeated action of the time evolution operator or the mixing matrices for the unconnected diagrams is perfectly equivalent to the action of the Bourret propagator and physically represents nothing but pitch-angle scattering. We show this in appendix~\ref{sec:pitch-angle-sacttering}.
We label the contributions from the different diagrams to the correlation function as
\begin{align}
\centering
&\vcenter{\hbox{\includegraphics[scale=1,trim={0 0 0 0.4cm},clip=True]{figures/term0}}} \\
&\vcenter{\hbox{\includegraphics[scale=1,trim={0 0 0 0.4cm},clip=True]{figures/term1a}}} \label{eqn:def_U1a} \\
&\vcenter{\hbox{\includegraphics[scale=1,trim={0 0 0 0.3cm},clip=True]{figures/term1b}}} \\
&\vcenter{\hbox{\includegraphics[scale=1,trim={0 0 0 0.3cm},clip=True]{figures/term1c}}}
\end{align}
We note again that the zeroth order diagram plus the corrections from the $(1a)$ and $(1b)$ diagrams together correspond to the usual pitch-angle scattering from QLT and are therefore not expected to lead to the mixing of the angular power spectrum. The contribution from the correlation of two phase-space densities in this formalism comes from the interaction diagram $(1c)$.
\subsubsection{The various contributions to the propagator and mixing matrix}
\label{sec:calculation-diagrams}
We note again that the time evolution operator should depend only on the time difference between the initial and final state in the case of static turbulence (see Section \ref{sec:mixing-equation}). Since we are only interested in the time evolution operator over the time step $\Delta T$, we shall from now on consider $U_{t,t_0}$ with $t=t_0+\Delta T$.
\paragraph{The (0) diagram}
Given the chosen coordinate system has the $z$-axis directed along $\vct{B}_0$, the free propagator evaluates to $U^{(0)}_{t,t_0}=\exp\left[-\imath\Omega \Delta T L_z\right]$. The double propagator then reads
\begin{equation}
\avg{\Up{A}{t}{t_0} U^{B*}_{t,t_0}}^{(0)} = \exp \left[-\imath \Omega \Delta T L^A_z + \imath \Omega \Delta T L^{B*}_z \right]
\end{equation}
which corresponds to a simple form of the leading order term for the mixing matrix element according to Eq.~\eqref{eq:mixing-matrix},
\begin{equation}
M_{\ell \ell_0}^{(0)}=\delta_{\ell\ell_0}.
\label{eq:diagram-0}
\end{equation}
\paragraph{The (1a) and (1b) diagrams}
The next-to-leading order terms (1a), (1b), and (1c) are more complicated and require a detailed discussion. In the following, we will provide the analytic formulae for these terms and their mixing matrix elements in the case of a general slab model with the turbulence tensor of the form \citep[see e.g.][and references therein]{Shalchi_book}
\begin{equation}
\avg{ \tilde{\omega}_a(\vct{k}) \tilde{\omega}^*_b(\vct{k}') } = \delta(\vct{k}-\vct{k}') g(k_\parallel) \frac{\delta(k_\perp)}{k_\parallel} \delta_{ab} \, ,
\end{equation}
if $a,b \in \{ x, y \}$ and zero otherwise. Here, $k_\parallel$ and $k_\perp$ are components of the wave vector $\vct{k}$ parallel and perpendicular to the ordered field $\vct{B}_0$. In the following we evaluate the turbulent magnetic field along unperturbed helical CR trajectories. For a particle with momentum $\vct{p} = p \left( \sqrt{1 - \mu^2} \cos{\phi}, \sqrt{1 - \mu^2} \sin{\phi}, \mu \right)$ and position $\vct{r}_\oplus$ at time t, the position at time $t'$ reads
\begin{equation}
\begin{aligned}
\vct{r}(t') &= \vct{r}_\oplus - \uvct{x} \frac{\sqrt{1-\mu^2}}{\Omega} \sin\left[ \phi - \Omega (t-t') \right] \\
& + \uvct{y} \frac{\sqrt{1-\mu^2}}{\Omega} \cos\left[ \phi - \Omega(t-t') \right] + \uvct{z} \mu (t-t') \, .
\end{aligned}
\end{equation}
As the two terms (1a) and (1b) give the same contribution to the mixing matrix, we will illustrate the evaluation of the matrix element only for the (1a) term. The (1a) contribution to the time evolution operator is
\begin{equation}
\begin{aligned}
& \avg{U_{t,t_0}^AU_{t,t_0}^{B*}}^{(1a)} \\
&= \!\!\int_{t_0}^t\!\!\!\mathrm{d}t_2\!\int_{t_0}^{t_2}\!\!\!\!\mathrm{d}t_1 U_{t,t_2}^{A(0)} \!\left\langle\delta\mathcal{L}_{t_2}^AU_{t_2,t_1}^{A(0)}\delta\mathcal{L}_{t_1}^A\right\rangle\! U_{t_1,t_0}^{A(0)}U_{t,t_0}^{B*(0)}\\
&= -\int_{0}^{\Delta T}\mathrm{d}T\int_{0}^{T}\mathrm{d}\tau \int\mathrm{d}^3\vct{k}\,g(k_\parallel)\frac{\delta(k_\perp)}{k_\perp}e^{\imath k_\parallel \mu_A \tau}\\
& \times\left[\cos{\Omega\tau}({L}^{A})^{2} + \imath \sin{\Omega\tau}{L}^A_z-\cos{\Omega\tau}({L}_z^{A})^{2}\right] \\
& \times e^{-\imath \Omega \Delta T {L}_z^A + \imath \Omega \Delta T L_z^{B*}},
\end{aligned}
\label{eqn:U1a}
\end{equation}
where we have made the change of variables $\tau = t_2-t_1$ and $T = t_2-t_0$ for the two time integrals and also rearranged the angular momentum operators using Hadamard's lemma \citep{miller1972}.
Next, we substitute Eq.~\eqref{eqn:U1a} into Eq.~\eqref{eq:mixing-matrix} and expand the Legendre polynomials into products of spherical harmonics $Y_\ell^m(\uvct{p})$,
\begin{equation}
P_{\ell}(\uvct{p}_A\cdot\uvct{p}_B) = \frac{4\pi}{2\ell+1}\sum_{m=-\ell}^{\ell} Y_\ell^m(\uvct{p}_A)Y_\ell^{m*}(\uvct{p}_B) \, .
\end{equation}
The integral over $\uvct{p}_B$ can be evaluated trivially and acting on the spherical harmonics with the angular momentum operators gives
\begin{equation}
\begin{aligned}
M^{(1a)}_{\ell\ell_0} &= -\int_0^{\Delta T} \mathrm{d}T\int_0^T\mathrm{d}\tau\int\mathrm{d}\vct{P}_A\int\mathrm{d}^3k \\
& \times g(k_\parallel) \frac{\delta(k_\perp)}{k_\perp} \sum_{m=-\ell}^\ell |Y_\ell^m(\uvct{p}_A)|^2\frac{e^{\imath \vct{k} \cdot \uvct{p}_A\tau}}{2\ell+1} \delta_{\ell \ell_0} \\
& \mkern-18mu \times \left[\ell(\ell+1)\cos{\Omega\tau} - \imath m \sin{\Omega\tau} - m^2 \cos{\Omega\tau} \right] \, .
\end{aligned}
\end{equation}
The sum over $m$ from the expansion of the Legendre polynomials can be evaluated using Uns\"{o}ld's theorem~ \citep{1927AnP...387..355U},
\begin{equation}
\sum_{m=-\ell}^{\ell} |Y_{\ell}^{m}(\mathbf{x})|^2 = \frac{2\ell+1}{4\pi}
\end{equation}
and its extensions,
\begin{align}
\sum_{m=-\ell}^{\ell} m |Y_{\ell}^{m}(\mathbf{x})|^2 &= 0 \, , \\
\sum_{m=-\ell}^{\ell} m^2 |Y_{\ell}^{m}(\mathbf{x})|^2 &= \frac{\ell(\ell+1)(2\ell+1)}{8\pi}\sin^2{\theta} \, . \label{step:unsoeld3}
\end{align}
In order to perform the remaining $\uvct{p}_A$ integration the exponential function is expanded by virtue of the plane wave expansion
\begin{equation}
e^{i\mathbf{k}\cdot\mathbf{r}} = 4\pi\sum_{\ell=0}^\infty\sum_{m=-\ell}^{\ell}i^\ell j_\ell(kr)Y_{\ell}^{m}(\mathbf{\hat{k}})Y_{\ell}^{m*}(\mathbf{\hat{r}}),
\end{equation}
where $j_\ell(\cdot)$ denotes the spherical Bessel function of the first kind. The mixing matrix element then reads
\begin{equation}
\begin{aligned}
& M_{\ell\ell_0}^{(1a)} = -\int_0^{\Delta T}\!\! \mathrm{d}T\! \int_0^T\!\! \mathrm{d}\tau\! \int \!\mathrm{d}\uvct{p}_A\! \int\mathrm{d}^3k \, g(k_\parallel)\\
& \times\! \frac{\delta(k_\perp)}{k_\perp}\sum_{\ell_A=0}^\infty\sum_{m_A=-\ell_A}^{\ell_A}\!\! \ell(\ell+1)\cos{\Omega\tau} \imath^{\ell_A} \\
& \times j_{\ell_A}(k_\parallel\tau) Y_{\ell_A}^{m_A}(\uvct{k}) Y_{\ell_A}^{m_A*}(\uvct{p}_A) \\
&\times
\left(\frac{4}{3}\sqrt{\pi}Y_0^0(\uvct{p}_A)+\frac{2}{3}\sqrt{\frac{\pi}{5}}Y_2^0(\uvct{p}_A)\right) \delta_{\ell\ell_0},
\end{aligned}
\end{equation}
where the $\uvct{p}_A$ integral can be evaluated by using the orthonormality condition of the spherical harmonics. Next, the azimuthal and perpendicular directions of the $k$ integration can be solved trivially. The $k_\parallel$ integral is rewritten as an integral over the positive part only to avoid an absolute value in the power spectrum. The $(1a)$ contribution to the mixing matrix is then given by
\begin{equation}
\begin{aligned}
M_{\ell\ell_0}^{(1a)}& = -\int_0^{\Delta T} \! \mathrm{d}T\int_0^T\mathrm{d}\tau \! \int_0^{\infty} \!\! \mathrm{d}k_\parallel g(k_\parallel)4\pi\ell(\ell+1) \\
& \times \cos{\Omega\tau}\left[\frac{2}{3}j_0(k_\parallel\tau)-\frac{1}{3}j_2(k_\parallel\tau)\right]\delta_{\ell \ell_0}.
\end{aligned}
\end{equation}
Introducing the function $\Lambda_\ell(\Delta T)$ for arbitrary integer $\ell$,
\begin{equation}
\Lambda_{\ell}(\Delta T) \! = \!\! \int_0^{\Delta T}\!\!\!\!\!\!\! \mathrm{d}T\! \int_0^T \mkern-11mu \mathrm{d}\tau \!\!\int_0^{\infty} \!\!\!\!\mathrm{d} k_\parallel g(k_\parallel) \cos{(\Omega\tau)} j_{\ell}(k_\parallel\tau) ,
\label{eq:diagram-1a}
\end{equation}
that will later be evaluated numerically, the mixing matrix contribution can be rewritten in the compact form
\begin{equation}
M_{\ell\ell_0}^{(1a)}
\! = \! - 8\pi\ell(\ell+1) \! \left[\frac{2}{3}\Lambda_0(\Delta T) - \frac{1}{3} \Lambda_2(\Delta T)\right] \! \delta_{\ell\ell_0}.
\end{equation}
\paragraph{The (1c) diagram}
The time evolution operator corresponding to the interacting diagram reads,
\begin{equation}
\label{eq:1ctimeevolution}
\begin{aligned}
&\mkern-8mu \left\langle U_{t,t_0}^A U_{t,t_0}^{B*}\right\rangle^{(1c)} \\
&=\!\!\! \int_{t_0}^t\!\!\!\mathrm{d}t_2\!\!\int_{t_0}^{t}\!\!\!\mathrm{d}t_1 U_{t,t_1}^{A(0)}
U_{t,t_2}^{B*(0)}\!\!\left\langle\delta\mathcal{L}_{t_1}^A\delta\mathcal{L}_{t_2}^{B,*}\!\right\rangle\!
U_{t_1,t_0}^{A(0)}U_{t_2,t_0}^{B*(0)}
\end{aligned}
\end{equation}
Using again Hadamard's lemma, we can simplify the operator algebra,
\begin{align}
e^{i\lambda L_z} L_x e^{-i\lambda L_z} &= \cos{\lambda} L_x - \sin{\lambda} L_y \, , \\
e^{i\lambda L_z} L_y e^{-i\lambda L_z} &= \sin{\lambda} L_x + \cos{\lambda} L_y \, .
\end{align}
Products of angular momentum operators can be expressed in terms of the ladder operators
\begin{align}
L_x^A L_x^B + L_y^A L_y^B &= \frac{1}{2} \left( L_{+}^A L_{-}^B + L_-^A L_{+}^B \right) \, , \\
L_x^A L_y^B - L_y^A L_x^B &= \frac{1}{2 \imath} \left( L_{-}^A L_{+}^B - L_{+}^A L_{-}^B \right) \, .
\end{align}
The action of the ladder operators on the spherical harmonics is $L_\pm Y_{\ell_0}^{m_0} = \sigma_\pm Y_{\ell_0}^{m_0\pm 1}$ and $L_\pm Y_\ell^{m_0*} = \sigma_\mp Y_{\ell_0}^{m_0\mp 1*}$ where \mbox{$\sigma_\pm = \sqrt{\ell_0(\ell_0+1)-m_0(m_0\pm1)}$}.
Using these relations the contribution from the (1c) diagram can be written as
\begin{equation}
\begin{aligned}
& \avg{ U_{t,t_0}^A U_{t,t_0}^{B*} }^{(1c)} \!\!=\!\! \int_{t_0}^{t}\!\!\mathrm{d}t_1\!\! \int_{t_0}^{t}\!\!\mathrm{d}t_2 \!\!\int\mathrm{d}^3\vct{k} \, g(k_\parallel)\frac{\delta(k_\perp)}{k_\perp} \\
& \times e^{\imath k_\parallel \mu_A(t-t_1) - \imath k_\parallel\mu_B(t-t_2)} e^{- \imath \Omega\Delta T L_z^A - \imath \Omega\Delta T L_z^B} \\
& \times \!\bigg\{(L_x^AL_x^B\! +\! L_y^AL_y^B)\sin(\Omega(t-t_1))\sin(\Omega(t-t_2))\vphantom{\frac{}{}}\\
& +(L_x^AL_x^B \!+\! L_y^AL_y^B)\cos(\Omega(t-t_1))\cos(\Omega(t-t_2))\vphantom{\frac{}{}}\\
& +(L_x^AL_y^B\! -\! L_y^AL_x^B)\cos(\Omega(t-t_1))\sin(\Omega(t-t_2))\vphantom{\frac{}{}}\\
& -(L_x^AL_y^B \!-\! L_y^AL_x^B)\sin(\Omega(t-t_1)\cos(\Omega(t-t_2))\vphantom{\frac{}{}}\bigg\} \, .
\end{aligned}
\end{equation}
When computing the mixing matrix, integrals over three spherical harmonics can be solved by introducing the Wigner-3j symbol
\begin{equation*}
\begin{aligned}
& \int\mathrm{d}\uvct{p} \, Y_{\ell_1}^{m_1}(\uvct{p}) Y_{\ell_2}^{m_2}(\uvct{p}) Y_{\ell_3}^{m_3}(\uvct{p}) \\
=& \sqrt{\frac{(2\ell_1\!+\!1)(2\ell_2\!+\!1)(2\ell_3\!+\!1)}{4\pi}} \wignerj{\ell_1}{\ell_2}{\ell_3}{0}{0}{0} \!\!\! \wignerj{\ell_1}{\ell_2}{\ell_3}{m_1}{m_2}{m_3}
\end{aligned}
\end{equation*}
Combining the above equation with Eq.~\eqref{eq:mixing-matrix} the mixing matrix receives the following contribution
\begin{widetext}
\begin{equation}
\begin{aligned}
\label{eq:diagram-1c}
M_{\ell\ell_0}^{(1c)} =& \pi\sum_{\ell_A,\ell_B} \imath^{\ell_B-\ell_A} (2\ell_0+1) (2\ell_A+1) (2\ell_B+1) \left[ 1 + (-1)^{\ell_A+\ell_B} \right] \wignerj{\ell_A}{\ell}{\ell_0}{0}{0}{0} \wignerj{\ell_B}{\ell}{\ell_0}{0}{0}{0} \\
& \times \sum_{m_0,m} \bigg[ (2 \ell_0 (\ell_0 + 1) - 2 m_0^2) \wignerj{\ell_A}{\ell}{\ell_0}{0}{m}{m_0} \wignerj{\ell_B}{\ell}{\ell_0}{0}{m}{m_0} \bigg] \kappa_{\ell_A,\ell_B}(\Delta T) \, ,
\end{aligned}
\end{equation}
where
\begin{equation}
\kappa_{\ell_A,\ell_B}(\Delta T) = \int_{0}^{\Delta T}\mathrm{d}t_A\int_{0}^{\Delta T}\mathrm{d}t_B \int_0^{\infty} \mathrm{d}k_\parallel~
g(k_{\parallel})j_{\ell_A}(k_{\parallel} t_A)j_{\ell_B}(k_{\parallel} t_B)\cos(\Omega (t_A-t_B)).
\end{equation}
\end{widetext}
Note that we have used relations presented in Appendix~\ref{appendix:1crelations} in order to shift some summation indices to achieve further simplifications. The $\kappa_{\ell_A,\ell_B}(\Delta T)$ integrals can be evaluated numerically using the Levin integration method~\citep{Levin1996FastIO}.
Due to the symmetry of exchanging particles the contribution from the $(1b)$ diagram is equal to the contribution from the $(1a)$ diagram. Since the $(0)$, $(1a)$ and $(1b)$ contributions to the mixing matrix are diagonal they do not lead to mixing between different multipoles, as expected. If it was for these two contributions only, the steady-state angular power spectrum would be only dipolar. The (1c) matrix element in Eq.~\eqref{eq:diagram-1c} is however non-diagonal and will thus lead to mixing of the dipole into higher multipoles. Putting these contributions to the mixing matrix together, Eq.~\eqref{eq:mm} can be solved numerically to get the steady state angular power spectrum.
In theory, solving Eq.~\eqref{eq:mm} requires inverting the mixing matrix for all $\ell,\ell_0$ up to infinity. In practice, however, this is not possible, but the contributions from very large $\ell,\ell_0$ on small multipoles are expected to be small. We therefore truncate the mixing matrix at $\ell_\text{max}=100$ and extrapolate the angular power spectrum as a function of $\ell_\text{max}$ to infinity assuming a power law scaling.
We have tested the validity of this extrapolation by comparing the extrapolation to $\ell=100$ for $\ell_\text{max} = 50$ to the full calculation for $\ell_\text{max} = 100$ and found very good agreement.
\subsubsection{The free parameter $\Omega \Delta T$}
\label{sec:OmegaDT}
To determine the correct value of $\Omega\Delta T$ the resonance structure of the different contributions has to be examined. Na\"ively one would expect the relevant timescale for the durations of the wave-particle interactions to correspond to the quasi-linear scattering time $\tau_s$. This however holds only for pure pitchangle scattering.
Starting from Eq.~\eqref{eq:1ctimeevolution} and performing the same steps as described previously for the $(1a)$ time evolution operator it can be shown that the $(1c)$ contribution in the quasi-linear theory limit becomes
\begin{equation}
\begin{aligned}
& \avg{ U_{t,t_0}^A U_{t,t_0}^{B*} }^{(1c)} \\
&= \left(\frac{\Delta\mu}{\bar\mu}\right)^{-1}\frac{2\pi^2\Omega}{\bar\mu}g\left(\frac{\Omega}{\bar\mu}\right)\\
&\times \!\bigg[ L_+^A L_-^B \left(e^{-i\Omega\frac{\Delta\mu}{\bar\mu}\Delta T}\!\!-\!\!1\right) \!\!-\!L_-^AL_+^B \left(e^{i\Omega\frac{\Delta\mu}{\bar\mu}\Delta T}\!\!-\!\!1\right) \!\bigg]\\
&\times e^{-i\Omega L_z^A\Delta T}e^{-i\Omega L_z^{B*}\Delta T} \, ,
\end{aligned}
\end{equation}
with $\Delta \mu = \mu_A-\mu_B$ the difference of the two particles pitch angles and $\bar\mu = (\mu_A+\mu_B)/2$ the mean pitch angle. The term
\begin{equation}
\left(\frac{\Delta \mu}{\bar\mu}\right)^{-1}\left(e^{\Omega\Delta T\frac{\Delta\mu}{\bar\mu}}-1\right)
\end{equation}
has the form of a resonance with finite support for $|\Omega \Delta T \Delta\mu/\bar\mu| < 1$. Further assuming that the relative pitchangle of particles is diffusing $\langle \left(\Delta\mu/\bar\mu\right)^2\rangle \propto D_{\mu\mu}\Delta T \propto \Delta T/\tau_s$ with the pitchangle diffusion coefficient $D_{\mu\mu}$ and the quasi-linear scattering time $\tau_s$ gives
\begin{equation}
\Omega \Delta T \propto (\Omega\tau_s)^{1/3} \, .
\label{eqn:relation}
\end{equation}
The constant of proportionality depends on the magnetic turbulence geometry and the power spectrum and will be determined from numerical simulations in Sec.~\ref{sec:results_APS}.
This parameter can be interpreted as the number of gyrotimes over which correlations in the (1c) diagram decay. In QLT the particles trajectories are approximated as unperturbed trajectories. The particles can thus interact with the waves for an infinitely long time leading to a sharp resonance. In reality however particles trajectories are perturbed by the interactions with the turbulent field leading to a decay of correlations on a timescale related to the scattering time $\tau_s$.
In summary, we expect $\Omega\Delta T$ to be related to the scattering time as $\Omega\Delta T\propto (\Omega\tau_s)^{1/3}$ due to the resonance structure of the interacting diagram.
\section{Results}
\label{sec:results}
\subsection{Sky maps}
\label{sec:sky_maps}
\begin{figure*}
\centering
\includegraphics[scale=1]{figures/skymaps0t40r_b.pdf}\includegraphics[scale=1]{figures/skymaps0t40r_b_wodipole.pdf}\\
\includegraphics[scale=1]{figures/skymaps1t40r_b.pdf}\includegraphics[scale=1]{figures/skymaps1t40r_b_wodipole.pdf}\\
\includegraphics[scale=1]{figures/skymaps2t40r_b.pdf}\includegraphics[scale=1]{figures/skymaps2t40r_b_wodipole.pdf}\\
\includegraphics[scale=1]{figures/skymaps4t40r_b.pdf}\includegraphics[scale=1]{figures/skymaps4t40r_b_wodipole.pdf}
\caption{CR anisotropy maps in various random field realizations, before (left column) and after (right column) subtraction of monopole and dipole, smoothed on a scale of $0.5^\circ$. We also compare the directions of the dipole anisotropy, the background magnetic field and the mean magnetic field averaged over a length of $L_\text{avg} = (1/30) L_c$. The simulations are done at turbulence level $\eta = \delta B^2/(B_0^2+\delta B^2) = 0.5$ and at $\rho=r_g/L_c = 8.9\times 10^{-4}$. We adopt a backtracking time $\Omega T = 2000$.}
\label{fig:dipoledir}
\end{figure*}
We present sky maps of the CR arrival directions in four different random realizations of the turbulent magnetic field in the left column of Fig.~\ref{fig:dipoledir}. These sky maps are chosen at $T \gtrsim \tau_s$ since we do not expect significant variations of the large-scale features in these sky maps (except for the contamination by the noise, see Sec.~\ref{sec:results_APS}) once particles are backtracked into the diffusive regime. This might be better visualized in Fig.~\ref{fig:illustration} where the shape of the CR excess region (the red patch in the sky map) remains roughly unchanged as we proceed to larger backtracking time. Roughly speaking, this means that the angular power spectrum on large scales converges for $T\gtrsim \tau_s$ as we will discuss in depth in Sec.~\ref{sec:results_APS}. We also see that a dipole appears in all realizations as the dominant anisotropy. Interestingly, the direction of the dipole vector in each realization $i$ (diamonds),
\begin{eqnarray}
\vct{\Phi}_i=\frac{1}{4\pi}\int\textrm{d}\uvct{p}\,\uvct{p}\frac{f(\vct{r}_\odot,\uvct{p},t)}{\bar{f}}
\end{eqnarray}
seems to point in the direction of the local mean magnetic field (empty circles) and not in the direction of the regular background magnetic field (crosses).
The right column of Fig.~\ref{fig:dipoledir} shows the sky maps for the same magnetic field realizations as in the left column of Fig.~\ref{fig:dipoledir} with the monopole and dipole fitted and subtracted. The sky maps still clearly show signs of anisotropy.
The results of QLT can be approximated by averaging the phase-space density over the ensemble of magnetic field realizations.
The ensemble averaged phase-space density calculated from the test particle simulations by averaging over $N=35$ different field realizations is shown in Fig.~\ref{fig:skymaps_qlt}. As expected from QLT, the dipole component of the averaged map,
\begin{eqnarray}
\langle\vct{\Phi}\rangle = \frac{1}{N}\sum_{i=1}^{N}\vct{\Phi}_i,
\end{eqnarray}
is pointing in the direction of the regular magnetic field. The misalignment between $\langle \vct{\Phi} \rangle$ and $\vct{\Phi}_i$ in all realizations is due to a relatively large value of $\eta$ adopted for these simulations and, as we shall see later in Sec.~\ref{sec:dipole}, this effect has to be taken into account in order to reconcile the normalization of the numerical and analytic angular power spectra for strong turbulence. We show also the ensemble averaged phase-space density with the monopole and dipole removed in the bottom panel of Fig.~\ref{fig:skymaps_qlt}. It has significantly less structure than the corresponding sky maps in the right column of Fig.~\ref{fig:dipoledir}, but a small octupole is visible.
\begin{figure}
\centering
\includegraphics[scale=1]{figures/skymaps34t40r_b_qlt.pdf} \includegraphics[scale=1]{figures/skymaps34t40r_b_wodipole_qlt.pdf}\\
\caption{Ensemble averaged sky map of the CR anisotropy from test particle simulations in slab turbulence with turbulence level $\eta = 0.5$ at normalized rigidity $\rho = r_g/L_c \simeq 8.9\times 10^{-4}$ averaged over 35 magnetic field realizations. The bottom panel shows the same ensemble averaged sky map with the monopole and dipole subtracted.}
\label{fig:skymaps_qlt}
\end{figure}
\subsection{Angular power spectrum}
\label{sec:results_APS}
Fig.~\ref{fig:time-dependent} shows the ensemble-averaged angular power spectrum from simulations $\langle C_\ell\rangle^{\rm num}$ as a function of backtracking time $T$ for the normalised rigidity \mbox{$\rho = r_{\text{g}} / L_{\text{c}} \simeq 8.9\times 10^{-3} $}, which corresponds to $R = 1 \, \text{PV}$ for the parameters detailed in Sec.~\ref{sec:parameters}. All the multipoles initially grow with $T$. For $T \gtrsim \tau_s$, the angular power spectrum on large scales starts converging towards a constant value (for the adopted parameters of turbulence $\tau_s(R=1\textrm{ PV})\simeq 100\,\Omega^{-1}$). We can also see that all the multipoles, especially the ones on small scales, are distorted by the shot noise at sufficiently large values of $T$ since $\mathcal{N}_\ell\sim T$ for $T\gtrsim \tau_s$, see Eq.~\eqref{eqn:noise}. In fact, the ensemble-averaged noise-subtracted angular power spectrum $\langle C_\ell\rangle^{\rm sub}=\langle C_\ell\rangle^{\rm num}-\mathcal{N}_\ell$ is expected to become constant for $T\gtrsim \tau_s$ which is essentially due to the angular power spectrum being sensitive only to the particular realization of the \emph{local} magnetic field. Clusters of particles being tracked back through time away from the observer initially travel ballistically and get deflected coherently as they experience similar magnetic fields. When they reach the regime of diffusive transport the trajectories decorrelate and hence the sky maps becomes noisier, but the ensemble-averaged, noise-subtracted angular power spectrum stays constant. The angular power spectrum is thus the imprint of the turbulent magnetic field in the transitory regime between ballistic and diffusive particle transport.
In addition, we present the noise-subtracted total angular power, which is $\sum_{\ell}(2\ell+1)\langle C_\ell\rangle^{\rm sub}$, as the green dotted line in Fig.~\ref{fig:time-dependent}. Note that in the summation we have used the empirical relation \mbox{$\langle C_\ell\rangle^{\rm sub}\sim \ell^{-2.7}$}, to be shown below, for $\ell\geq 5$. The sky-map total angular power, that is the ensemble-averaged variance of the sky maps $4\pi\sigma^2=\sum_{\ell}(2\ell+1)\langle C_\ell\rangle^{\rm num}$, is also shown as the solid black line which grows linearly with $T$ for $T\gg \tau_s$, see Eq.~\eqref{eqn:sky_variance}. More importantly, we have found that $4\pi\sigma^2\gg \sum_{\ell}(2\ell+1)\langle C_\ell\rangle^{\rm sub}$ for $T\gg\tau_s$ and, thus, the dominant contribution to the sky-map total angular power at large backtracking time is the noise total angular power $\sum_{\ell}(2\ell+1)\mathcal{N}_\ell$ (orange dot-dashed line in Fig.~\ref{fig:time-dependent}).
We note also that this result is valid regardless of the number of pixels $N_{\rm pix}$ of the simulations since
the sum over $\ell$ with an $N_\text{pix}$ dependent summation range exactly cancels the $N_\text{pix}$ dependence of $\mathcal{N}_\ell$, making the noise total angular power independent of $N_\text{pix}$.
\begin{figure}
\centering
\includegraphics[scale=1]{figures/cltime1PeVslabneweta0.5.pdf}
\caption{Ensemble-averaged angular power spectrum as a function of backtracking time $T$ from the test particle simulations for a normalized rigidity of $\rho \simeq 8.9\times 10^{-3}$. The solid lines show the different multipoles, the dashed line shows the estimated noise level growing linearly with time. The other lines show the total noise subtracted angular power (green dotted line), the noise total angular power (orange dot-dashed line) and the sky map total angular power (solid black line).}
\label{fig:time-dependent}
\end{figure}
\begin{figure*}[tbh]
\centering
\includegraphics[width=\textwidth]{figures/cellkolmogorov_unnormalized_paper.pdf}
\caption{Angular power spectra of the arrival directions of CRs. The points show the angular power spectra computed from the results of test particle simulations in slab turbulence with a turbulence level of $\eta = 0.5$. Also shown is the best fit line from the analytical results for each individual rigidity. The numerical and analytical results agree well down to the smallest angular scales.}
\label{fig:fit}
\end{figure*}
The noise-subtracted angular power spectra from the numerical simulations, normalised to the dipole, for the normalised rigidities $\rho = r_{\text{g}} / L_{\text{c}} \simeq 8.9 \times 10^{-5}$, $8.9\times 10^{-4}$, $8.9\times 10^{-3}$ and $8.9\times 10^{-2}$ are shown in Fig.~\ref{fig:fit}. With the parameters as adopted in Sec.~\ref{sec:parameters}, that is $B_\text{RMS} = 4\,\mu\text{G}$ and $L_c = 30\,\text{pc}$, this corresponds to the rigidities $10 \, \text{TV}, 100 \, \text{TV}, 1 \, \text{PV}$ and $10 \, \text{PV}$. The angular power spectra are falling power laws in $\ell$ for $\ell \gtrsim 2$ and the magnitude of the spectral index is larger for smaller rigidities.
\begin{figure}[tbh]
\centering
\includegraphics[scale=1]{figures/scatteringtimefit_code_wokrai_paper.pdf}
\caption{Best fit $\Omega \Delta T$ parameter as a function of the quasi-linear scattering time $\Omega \Delta \tau_s$ for simulations at different rigidities and different turbulence levels. The solid line shows the fitted relation $\Omega \Delta T \simeq 4 (\Omega\tau_s)^{1/3}$.}
\label{fig:taus}
\end{figure}
Also shown in Fig.~\ref{fig:fit} are the steady-state angular power spectra from the analytical computation, showing excellent agreement with the numerical results. We have also normalised to the dipole here, but we discuss this normalisation in Sec.~\ref{sec:dipole} below. In computing the mixing matrices, we have adopted the relation between the free parameter $\Omega \Delta T$ and the quasi-linear scattering time $\Omega \tau_{\text{s}}$ predicted in Eq.~\eqref{eqn:relation}. This is a prediction up to the constant of proportionality and we have found this constant to be $\simeq 4$ by fitting to the results from the numerical simulations, that is $\Omega \Delta T \simeq 4 (\Omega\tau_s)^{1/3}$. This fit is illustrated in Fig.~\ref{fig:taus}. The steady state angular power spectra initially fall off very quickly depending on rigidity via the $\Omega\Delta T$ parameter and then, for $\ell\gtrsim 5$ exhibit power-law like falling behaviour with a spectral index close to $-2.7$ for all rigidities. This results in a slightly flatter angular power spectrum than the $\ell^{-3}$ scaling that was found in~\cite{2014PhRvL.112b1101A} and is more similar to the scaling found in~\cite{2015ApJ...815L...2A} for their model parameter $p=0.6$.
\subsection{Dipole amplitude}
\label{sec:dipole}
We have already shown that the dipole typically aligns with the average magnetic field direction instead of the background field direction, see Sec.~\ref{sec:sky_maps}. It turns out that the normalisations of the angular power spectra from the numerical simulations and from the analytical calculation differ, while the shapes agree. We will discuss the origin of this disagreement and explain how to correct the overall normalisation of the angular power spectra from the analytical calculation.
Both of these results could be better understood by looking at the dipole direction within each realization which can be related to the dipole strength as follows
\begin{eqnarray}
C^{(i)}_1=4\pi\vct{\Phi}_i^2.
\end{eqnarray}
This quantity is technically the ``observed'' dipole strength for a specific realization of turbulence and, correspondingly, the numerical or noise subtracted ensemble-averaged dipole strength (before the noise also starts dominating the dipole) is
\begin{eqnarray}
\langle C_1 \rangle^{\rm sub}\simeq \langle C_1 \rangle^{\rm num}=\frac{4\pi}{N}\sum_{i=1}^{N}\vct{\Phi}_i^2=\frac{1}{N}\sum_{i=1}^{N}C_1^{(i)}.\label{eq:C1-observe}
\end{eqnarray}
where $N$ is the total number of realizations considered. Note again that this statement is only true if the dipole is not dominated by the shot noise which should be the case for $T$ roughly a few times $\tau_s$. In the standard QLT approach, we could also predict the ensemble average of the dipole vector following Fick's law \citep{2017PrPNP..94..184A}
\begin{eqnarray}
\langle\vct{\Phi}\rangle \simeq \frac{1}{N}\sum_{i=1}^{N}\vct{\Phi}_i=-\frac{\vct{K}\cdot\nabla\bar{f}}{\bar{f}},
\label{eq:dipole}
\end{eqnarray}
from which we could also write down the QLT ensemble average dipole strength $\langle C_1 \rangle^{\rm QLT}$ as
\begin{eqnarray}
\langle C_1 \rangle^{\rm QLT} = 4\pi \langle\vct{\Phi}\rangle^2 = \frac{4\pi}{N^2} \sum_{i,j=1}^{N} \vct{\Phi}_i \vct{\Phi}_j.\label{eq:C1-QLT}
\end{eqnarray}
Having presented the various definitions of the dipole, we can now compare them with each other. In the following, we shall fix the direction of the CR gradient along $\vct{B}_0$ for simplicity. This means that the ensemble average dipole direction should be aligned with $\vct{B}_0$ as could be seen from Eq. \eqref{eq:dipole}. This might not be too problematic for a small turbulence level as we expect the dipoles in all realizations to point more or less in the direction of $\vct{B}_{0}$ which, roughly speaking, means $\vct{\Phi}_i\simeq\langle \vct{\Phi} \rangle$ or equivalently $\langle C_1\rangle^{\rm sub} \simeq \langle C_1\rangle^{\rm QLT}$. If the turbulence level is large enough, e.g. $\langle\delta\vct{B}^2\rangle\simeq\vct{B}_0^2$, the prediction for the dipole might not be accessible in a straightforward way within the framework of QLT \citep{mertsch2015}. This is because CRs propagate principally along the large-scale magnetic field $\vct{B}_i$ within the $i$th realization meaning the flux of particles determining the dipole direction should also point along $\vct{B}_i$ which has a large contribution from the turbulent field and do not exactly follow $\vct{B}_0$. More importantly, Eq. \eqref{eq:C1-observe} and \eqref{eq:C1-QLT} also tell us that $\langle C_1\rangle^{\rm sub} > \langle C_1\rangle^{\rm QLT}$ for strong turbulence (see also \cite{2015ApJ...815L...2A} for the case of isotropic turbulence). We shall now explore this scenario with test particle simulations.
The sky maps of CR arrival directions within a few different realizations of the turbulence with $\eta=0.5$ were already shown in Fig. \ref{fig:dipoledir}. Note that the direction of the $\vct{B}_0$-field is in the middle of the map. The direction of the local large-scale magnetic field $\vct{B}_i$ within each realizations (circle) could also be obtained by performing the spatial average of the turbulent field over $L_{\rm avg}$. We note that $L_{\rm avg}$ has to be sufficiently smaller than the correlation length. The results are shown for $L_{\rm avg}=L_c/30$, but they are insensitive to the exact choice of this parameter.
We also present the dipole strength for different rigidities in Fig.~\ref{fig:dipole-amplitude}. The green and orange points represent respectively the ensemble-averaged QLT and noise-subtracted dipole strength obtained from the numerical simulations. We also include the dipole strengths at different rigidities from the mixing matrix approach $\langle C_\ell\rangle^{\rm mat}$ (solid blue line) and they seem to match very well with $\langle C_1\rangle^{\rm QLT}$. This is understandable in the sense that the mixing matrix formalism only considers unperturbed trajectories and particles propagating along $\vct{B}_0$ similar to the standard QLT approach as indicated in Eq. \eqref{eq:product-small-time} where we have assumed $\lim_{\Delta T\rightarrow 0}\langle \Delta \vct{r}\rangle =\mu\Delta T\uvct{z}$.
It is also clear that $\langle C_1\rangle^{\rm sub} > \langle C_1\rangle^{\rm mat} \simeq \langle C_1\rangle^{\rm QLT}$ since we are in the case of $\eta=0.5$. Note that we have assumed $\nabla\bar{f}$ to follow the same scaling in rigidity as $\bar{f}$ and, thus, $\langle C_\ell\rangle^{\rm QLT}\sim K_{zz}^2 \sim R^{2/3}$. Interestingly, $\langle C_\ell\rangle^{\rm sub}$ seems to have the same rigidity scaling as $\langle C_\ell\rangle^{\rm QLT}$ with the normalization shifted by a factor of about two. Näively, we might expect $\vct{B}_i$ and $\vct{B}_0$ to make an angle of around $\alpha\simeq\pi/4$ in the case of $\eta=0.5$ which means the projection of the dipole vectors $\vct{\Phi}_i$ along $\vct{B}_0$ would make $\langle C_1\rangle^{\rm sub}/\langle C_1\rangle^{\rm mat} \simeq 1/\eta \simeq 1/\cos^2\alpha \simeq 2$.
In principle, we should be able to predict $\langle C_\ell\rangle^{\rm mat}$ more accurately and better match $\langle C_\ell\rangle^{\rm sub}$ even for strong turbulence once the particle's motion along $\vct{B}_i$ is taken into account. This requires a description of the CR transport perpendicular to $\vct{B}_0$. However, we expect this effect to modify only the dipole source term in Eq. \eqref{eq:mm} by introducing an additional factor which depends only on the turbulence level $\eta$ and, thus, the relation between $\tau_s$ and $\Delta T$ obtained from fitting the shape of the angular power spectrum spectra should remain roughly the same. For now, we shall adopt a more straightforward phenomenological approach which is to obtain the cosine of the misalignment angle $\alpha$ between $\vct{B}_i$ and $\vct{B}_0$ for all realizations and then rescale $\langle C_\ell\rangle^{\rm mat}$ by a factor $1/\langle\cos\alpha\rangle^2\simeq 2.2$.
In addition, due to limited dynamical range in the numerical testparticle simulations the diffusion coefficient at the smallest rigidities deviates from the expected $R^{-1/3}$ scaling and thus the normalization of the angular power spectrum is sligthly too large. Therefore the numerical angular power spectrum at $10\,\text{PV}$ has to be rescaled to give the correct rigidity dependence of the dipole anisotropy.
This provides good fits of $\langle C_\ell\rangle^{\rm mat}$ to $\langle C_\ell\rangle^{\rm sub}$ as shown in Fig.~\ref{fig:fit}.
\begin{figure}
\centering
\includegraphics[scale=1]{figures/dipole_scaling.pdf}
\caption{Noise subtracted and QLT dipole strengths at different rigidities for test particle simulations with a turbulence level $\eta=0.5$ in comparison to the mixing matrix prediction.}
\label{fig:dipole-amplitude}
\end{figure}
\section{Summary and conclusion}
\label{sec:summary}
The angular power spectrum of CR arrival directions is an important observable. As it is sourced by correlations of particles experiencing the same turbulent magnetic field it can be used to infer the local field configuration. Understanding the origin of the angular power spectrum is therefore important as an independent probe of the outer scale of turbulence and the local turbulence geometry.
Here we have used a perturbative calculation to predict the angular power spectrum of CR arrival directions taking into account the correlations of phase-space densities implied by the correlations in the turbulent magnetic field. We have assumed a homogeneous background magnetic field such that the unperturbed orbits are helical. The perturbative expansion up to first order in the turbulence strength then includes resonance effects between particles and the turbulent magnetic field similarly to QLT. The difference in the angular power spectrum compared to QLT arises because also correlations between phase-space densities that are induced by particles travelling through the same field realization are treated explicitly. This leads to a finite angular power spectrum even at larger multipoles $\ell$.
To validate and test the assumptions that were made in this calculation we have compared to test particle simulations done in the same turbulence model at the rigidities relevant for observations by IceCube and HAWC. This comparison shows very good agreement between the analytical model and the numerical test particle simulations.
If the turbulence in the local interstellar medium indeed has the slab geometry, we can use this model to constrain the parameters of the turbulent magnetic field. In particular, fitting the shape of the angular power spectrum to the available observational data will allow us to probe the scattering time which might, in turn, provide more information on the correlation length and the turbulence level in our Galactic neighborhood. This will however require more information on the uncertainties from observations. It is clear that the analytic model always predicts the angular power spectrum to be rigidity-dependent and yet sky maps of the arrival directions are observed for particles of different energies. In fact, the conversion between rigidity and energy might not be straightforward as the CR compositions assumed could vary between observatories. More importantly, even if the CR composition is known with confidence, we have to note that observational data always come with a finite energy resolution and, thus, the observed anisotropies might contain features different from the ansiotropies expected for individual energies. Finally, we should note also that the limited sky coverage of observatories like IceCube and HAWC could also hinder the reconstructions of the angular power spectrum from observed sky maps. Futher investigations are needed in order to better address these problems and gain more insight into the nature of local turbulence hidden in the small-scale anisotropies.
\begin{acknowledgments}
This project was funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) -- project number 426614101.
\end{acknowledgments}
\bibliographystyle{aasjournal}
|
\section{Introduction}
The solar wind provides a natural laboratory to study space plasma turbulence which is believed to be responsible for its acceleration and heating. Despite being weakly collisional in nature, turbulence in the solar wind plasma can be studied in the framework of magnetohydrodynamics (MHD) for the length scales superior to the ion gyroscale and ion inertial scale ($\sim 100$ km). Similar to neutral fluids, homogeneous MHD turbulence also entails an inertial range energy cascade with a scale invariant flux rate $\varepsilon$. This holds good only for relatively short intervals of solar wind where the assumption of homogeneity can be justified. From spacecraft, the turbulent fluctuations of the plasma fluid and the electromagnetic field are directly obtained in the form of a time series. Since both the spacecraft speed ($\sim$ 1 km/s) and the Alfv\'en speed ($\sim$ 50 km/s) of the solar wind at 1 AU are much lower than the solar wind speed ($v \sim$ 600 km/s), one can use Taylor's hypothesis to map length scales ($\ell$) into corresponding time scales ($\tau$) as $ \ell = v \tau$ or different frequencies ($f$) to equivalent wave numbers ($k$) as $ \omega = v k$ \citep{Taylor, banerjeethese}. The power spectra for the magnetic field, velocity and density fluctuations in the solar wind have been extensively studied for many decades. In the MHD frequency range of $10^{-4}-10^{-1} Hz$, where both the magnetic energy and density power spectrum are found to scale as $f^{-5/3}$ (or $\sim k^{-5/3}$), velocity power spectrum is found to follow a shallower $f^{-3/2}$ spectrum \citep{Podesta2006}. Turbulence power spectrum for higher frequencies are also investigated systematically \citep{Alexandrova2009, Sahraoui2009}. It is only recently that the first measured power spectrum of electric fluctuations has been reported \citep{Bale}. The spectrum is shown to have a power law index of $-5/3$ and follows the magnetic field fluctuation spectrum until around 0.45 Hz (breakpoint) at the spacecraft measured frequencies. Beyond this breakpoint, the magnetic spectrum becomes steeper whereas the electric spectra becomes enhanced.
All these power spectra have been obtained under the assumption of isotropy. In MHD, unlike velocity, Galilean transformation cannot eliminate the effect of mean magnetic field ($B_0$) which leads to unequal deformation of the Alfv\'enic wave packets along and perpendicular to $B_0$. Due to a non-negligible $B_0$, anisotropy in solar wind turbulence (SWT) is normally expected in the turbulence power level as well as in the spectral indices. However, in studies conducted using a global mean magnetic field, only power anisotropy and no spectral index anisotropy were reported \citep{Tessein, Sari, Chen2011}. This puzzle can be addressed using the results of several numerical studies of MHD turbulence showing that the deformation of wave packets of a specific length scale will be regulated by the mean magnetic field of comparable length scale rather than that of the largest scale \citep{Cho2000, Milano2001}. For solar wind, the minimum variance direction of the inertial range fluctuations is found to closely follow the scale dependent local mean field (LMF) \citep{Horbury95}. It is therefore reasonable to investigate the power and spectral index anisotropies of inertial range solar wind turbulence with respect to the scale dependent mean magnetic field rather than the global mean field. For single spacecraft data (in the form of a one dimensional time series), an estimate of the LMF and the analysis of anisotropy with respect to the LMF are performed using the method of wavelets. Unlike the previous studies with global mean field, here, prominent power and spectral index anisotropies of magnetic power spectra are reported both for polar and ecliptic solar winds \citep{Horbury, Podesta2009}. Within the frequency range $10^{-2}-10^{-1}$ Hz, the parallel power is found to be at least 5 times less than the perpendicular power whereas the spectral index varies from around -2 to -5/3 as the angle between the LMF and the flow direction changes from $0^\circ$ to $90^\circ$, being consistent with the critically balanced MHD turbulence theory by \citet{Goldreich}. Clear signatures of anisotropy for (i) magnetic power spectra at MHD and sub-ion scales and (ii) inertial range velocity spectra are also found for near ecliptic solar wind using both single and multi spacecraft data \citep{Chen2010, Chen2011, Wicks11}. Recently, it is realized that the scale dependent spectral anisotropy may indeed emerge from the mixing of the scaling property of the LMF and that of the magnetic field fluctuations \citep{Oughton}. This effectively makes the magnetic power spectra to be derived from moments of order higher than 2 thereby leading to a sensitivity towards the phase randomization. The signature of critical balance cannot therefore be ascertained just by the spectral anisotropy with respect to an LMF. In fact using Hilbert spectral analysis of solar wind data, both the parallel and the perpendicular magnetic power spectra are found to give a $k^{-5/3}$ behaviour \citep{Telloni}. Despite this limitation, wavelet based analysis of scale-dependent anisotropy can be interesting due to its simplicity and statistical robustness both for the velocity and the magnetic field.\\
Unlike velocity and magnetic field, the anisotropy of electric field fluctuations is not studied extensively. The first (and probably the only till date) study of power and spectral index anisotropy in the electric field power spectra was carried out by projecting the electric field data along and perpendicular to the global mean magnetic field \citep{Mozer}. After following stringent selection criteria, they could find only three suitable intervals from the \textit{THEMIS} and Cluster data. From their analysis, it was found that the parallel power magnitudes were comparable to or greater than the perpendicular powers. Within the frequency range $10^{-2}-1$ Hz, both the spectra had similar shape and inertial range power law slopes of about -5/3. But the results obtained could not be validated over more intervals due to lack of data. In this paper, we investigate the power and spectral index anisotropies of electric field fluctuations with respect to local mean magnetic field. Following the wavelet transform technique as implemented in \citet{Horbury} for magnetic field data, here we decompose a time series of electric field data into wavelet coefficients which are localized in both time and frequency (or wavelet scale). In order to capture the local mean magnetic field, we use Gaussian windows of different standard deviations thereby giving a measure of the corresponding length scale.
\begin{table*}
\centering
\caption{Parameter values for the streams analyzed near solar maximum and minimum}
\label{tab:parameter_table}
\begin{tabular}{c c c c c c c}
\hline
Year & From & To & Solar wind velocity & Proton Density & Proton Parallel Temperature & Proton Perpendicular Temperature \\
~ & ~ & ~ & $km\ s^{-1}$ & $(n_p)\ cm^{-3}$ & $(T_{\parallel})\ eV$ & $(T_{\perp})\ eV$ \\
\hline \hline
2002 & 02 Feb 14:00 & 03 Feb 00:06 & (-393.9, 8.7, -41.7) & 5.7 & 38.9 & 32.2 \\
2007 & 17 Jan 17:33 & 17 Jan 20:44 & (-623.7, -6.7, -38.8) & 1.04 & 132.8 & 80.6 \\
\hline
\end{tabular}
\end{table*}
\section{Data Intervals}
We use the data from Cluster spacecraft (ESA), when it is in the pristine fast solar wind (speed $\sim 550-700 km/s)$ or moderate solar wind (speed $\sim 400-550 km/s)$, outside the earth's magnetic environment. Electric field measurements are made by the electric field and wave (EFW) experiment \citep{Gustafsson} and magnetic field is measured by flux gate magnetometer (FGM) instrument \citep{Balogh} on board the spacecraft. We use the magnetic field and electric field data of cadences 22 and 25 Hz respectively, measured in geocentric solar ecliptic (GSE) coordinate system. Ion moments (velocity, density and temperature) of the solar wind are obtained from cluster ion spectrometry (CIS) experiment \citep{Reme}. All the above data are taken from Cluster 4 (C4) spacecraft. To choose the data intervals for our analysis, first we searched for the streams with longest duration possible (to get the maximum number of data points) with no discontinuities and then checked for the stationarity, which means that the average value of the parameter should change the least during the entire interval. This selection criteria enabled us to choose total 60 intervals for this study with 30 data intervals from the solar activity period (year 2001-2003) and another 30 intervals from the solar calm period (2007) of the solar cycle 23. The selected intervals are from the periods when spacecraft is at geocentric distances of $15 R_E$ to $20 R_E$ with interval lengths lasting from 1 hr to 10 hrs. In addition, during the high solar activity period the data is chosen so as to avoid any transient phenomenon like CMEs (note that the only CME that took place in 2002 was in January). \\
In table~\ref{tab:parameter_table}, we mention parameters for two of the longest duration streams that we could find, one each from the period of sun's maximum (year 2002) and minimum activity (year 2007). The spacecraft C4 was at a distance of $18.25 R_e$ and $19 R_e$ from the earth (using ephemeris data), respectively. The proton temperatures $(T_{\parallel})$ and $(T_{\perp})$, proton density $(n_p)$ and proton bulk velocity $(V_x, V_y, V_z)$ values mentioned in the table have been averaged over the respective entire intervals. Fig.~\ref{fig:para} shows the time variation of various plasma properties over the interval from the year 2007. The presented data consists of negligible amount (less than 0.5\%) of data gaps which we have tackled using linear interpolation.
\begin{figure}
\includegraphics[width=1.0\linewidth]{param_2007.png}
\caption{Plasma parameters (proton temperatures, proton density and proton bulk velocity), electric and magnetic field components for the interval of 17 Jan 2007 from 17h33 to 20h44.}
\label{fig:para}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\linewidth]{angle_hist.png}
\caption{Histogram of number of counts for each angle bin computed at the frequency 0.04 Hz (wavelet scale value of 22.59 sec) for $(a)$ the data interval from February 2, 2002 during sun's maximum activity (in blue) and $(b)$ the data interval from January 17, 2007 during solar minimum (in orange).}
\label{fig:hist}
\end{figure}
\section{Analysis Techniques}
In this study, we measure the reduced spectrum which is defined as \citep{Fredricks, Horbury, Papen2015}:
\begin{equation}
\mathcal{P}(f) = \displaystyle \int_{}^{} d^3\textbf{k}\,P(\textbf{k})\,\delta(2\pi f - \textbf{k\,.\,V})
\end{equation}
Scale dependent anisotropy of electric field fluctuation spectrum can be measured by studying how its spacecraft-frame power spectrum $\mathcal{P}(f)$ varies with angle between the average flow direction and the LMF. In order to compute the power spectra, 25 Hz electric field data is sub-sampled onto the time tags of 22 Hz magnetic field data by linear interpolation. Wavelet transformation of the electric field data is carried out to decompose the time series into time-frequency space thereby allowing us to calculate the power localised in both time and frequency \citep{Torrence}. For component i of electric field, $E_i(t_k)$, where $t_k = t_0 + k\delta t$, with equal time spacing $\delta t = 1/22 \ s$ and $k=0,1,2...N-1$, the continuous wavelet transform is defined as the convolution of $E_i(t_k)$ with a scaled and translated version of $\psi(\eta)$:
\begin{equation}
w_i(t_j, f_l) = \displaystyle \sum\limits_{k=0}^{N-1} \ E_i(t_k) \ \psi\left(\frac{t_k - t_j}{s_l}\right)
\label{eq:wavelet}
\end{equation}
where, $\psi(\eta)$ is called the Morlet wavelet and is given as:
\begin{equation}
\psi(\eta) = \pi^{-1/4} \ e^{i\omega_0\eta} \ e^{-\eta^{2}/2} \ .
\end{equation}
Here $\eta$ is a non-dimensional time parameter and $\omega_0$ is the non-dimensional frequency taken as 6 for it to construct a nearly orthonormal set of wavelets \citep{Farge}.
To obtain the set of frequencies for Morlet wavelet, it is convenient to have:
\begin{equation}
s_m = s_0 \ 2^{m \delta m}, \ m = 0,1,...,M
\end{equation}
where smallest resolvable scale, $s_0 = 2 \ \delta t/(1.03)$. For adequate sampling of scales, we choose $\delta m$ as 0.5 and conduct the analysis for a total of 25 scales. For $\omega_0=6$, the wavelet scale $s_l$ is related to the frequency $f_l$ as $f_l=0.97/s_l$ from which we get a set of 25 frequencies. Instead of computing the wavelet coefficients from equation~(\ref{eq:wavelet}), we calculate them in Fourier space using fast Fourier transform (FFT) and convolution theorem, which is considered to be faster. We calculate the wavelet coefficients at 25 frequencies ranging from 2.7 mHz to 11 Hz (Nyquist frequency for the data used). Power in component i at time $t_j$ and frequency $f_l$ is defined as:
\begin{equation}
\mathcal{P}_{ii}(t_j, f_l) \displaystyle \propto \ \frac{2\delta t}{N} \ \mid w_i(t_j, f_l) \mid^2 \ .
\end{equation}
For the vector field \textbf{E}(t), the trace power is the sum of the power of the three orthogonal components, such that:\\
Trace power, $\mathcal{P} = \mathcal{P}_{xx} + \mathcal{P}_{yy} + \mathcal{P}_{zz}$ . We now calculate the scale dependent mean magnetic field or the LMF at time $t_j$ and wavelet scale $s_l$. This is obtained by modulating $B_i$, the ith component of the magnetic field time series, with a Gaussian curve centered at time $t_j$ and scaled with $s_l$, and is given by:
\begin{equation}
\bar{b}_i(t_j, s_l) = \displaystyle \sum\limits_{k=0}^{N-1} \ B_i(t_k) \ \exp\left[-\frac{(t_k-t_j)^2}{2s_l^2}\right] \ .
\end{equation}
This results in a time series of vectors $\bar{\textbf{b}}(t_j, s_l)$ for each scale $s_l$ pointing in the direction of LMF. The inclination ($\theta_{vb}$, again a scale-dependent quantity) of this mean field with respect to the average flow direction $\textbf{V}_{sw}$ of solar wind (sw) can be given by:
\begin{equation}
\theta_{vb} (t_j, s_l) = \displaystyle \cos^{-1} \left(\frac{\textbf{V}_{sw}\cdot \ \bar{\textbf{b}}}{\mid\textbf{V}_{sw}\mid \ \mid\bar{\textbf{b}}\mid} \right).
\label{eq:angle}
\end{equation}
\color{black}
The inertial range turbulence is found to be axisymmetric about the magnetic field direction \textit{i.e.} the power spectra are independent of the azimuthal angle $\phi$ and hence we consider the power values averaged over all $\phi$ \citep{Horbury}. Upon obtaining the directions of the local mean magnetic field, we now bin the powers of each scale as a function of $\theta_{vb}$. We construct bins of equal spacing of $10^\circ$. As it is evident from the Fig.~\ref{fig:hist}, for the 2002 interval, we get 16 bins ($20^\circ$ to $180^\circ$) and for that of 2007, we have only 14 bins ($0^\circ$ to $140^\circ$). The distribution of bin counts are found to be almost similar at all frequencies from 2.7 mHz to 11 Hz. For each bin and for a given frequency $f_l$, we average the trace power values corresponding to all the angles that lie in the bin. By doing this, we obtain the average trace power value $\mathcal{P}(f_l,\,\theta _{vb})$ corresponding to the frequency $f_l$ for that particular angle bin. For the sake of reasonable statistical sampling, we reject those angle bins which have less than 100 contributing power levels.
\begin{figure*}
\includegraphics[width=\linewidth]{ladder_plot.png}
\caption{Electric power spectra at two different angle ranges of the local mean magnetic field to the flow direction: $80^\circ$-$100^\circ$ (blue) and $170^\circ$-$180^\circ$ (orange) for the data interval 02 Feb, 2002 14:00 to 03 Feb, 2002 00:06 (solar activity) whereas $80^\circ$-$100^\circ$ (blue) and $0^\circ$-$10^\circ$ (orange) for the data interval 17 Jan, 2007 17:33 to 20:44 (solar calm) }
\label{fig:ladder_plot}
\end{figure*}
Of the 60 intervals used in this study, only for 4 intervals, we found contribution from both the parallel (or antiparallel) and perpendicular direction with respect to the LMF. Finally, in this paper, we have chosen to present the study of two intervals, one from each of the years 2002 and 2007.
\section{Results}
Fig.~\ref{fig:ladder_plot} shows, for two aforementioned intervals, the turbulent power spectra of electric field fluctuations along and perpendicular to the LMF. From the figures it is evident that unlike the previous studies \citep{Mozer}, here we obtain anisotropies in both power levels and spectral indices. In particular, the nature of anisotropy is found to depend strongly on the frequency. In both intervals, clear discrepancy between the parallel (or the antiparallel) and the perpendicular spectra is observed in the upper MHD frequency range \textit{i.e.} from $10^{-2}-10^{-1}$ Hz. Most of the time, it is found that the parallel power is larger and steeper than the perpendicular power. For 2002, where the wind is moderately slow, the antiparallel electric power is found to be roughly 4 times the perpendicular power. The antiparallel and the perpendicular spectral slopes are obtained as $-1.98$ and $-1.64$, which looks very similar to that of magnetic power anisotropy which is consistent with critical balance theory. However here we have further subdivided the parallel spectra into a very flat ($-1.46$) and a very steep ($-2.52$) part which give together a slope of nearly $-1.98$. During the interval of 2007, the MHD range anisotropy also represents an initial dominance of parallel electric power over the perpendicular one (by a factor of around 2.5). The parallel and perpendicular spectral indices are $-2.4$ and $-1.48$ respectively. Note that, the electric field amplitudes were much stronger in 2007 and hence a common factor of around 10 is observed between the power amplitudes of 2002 and 2007. After this range of prominent anisotropy, the parallel and perpendicular spectra behave randomly for different intervals of our study, therefore making it difficult to interpret. However, beyond 1 Hz, both the spectra get flat possibly due to the instrumental noise floor (which lies around $\sim 10^{-10} (mV/m)^2/Hz$) \citep{Salem}.
\begin{figure}
\centering
\includegraphics[width=\linewidth]{E_compare_B.png}
\caption{Perpendicular trace electric power spectrum (flow direction to local magnetic field angle ranging from $80^\circ$-\ $100^\circ$) in blue, Parallel electric fluctuations spectrum (angle ranging from $0^\circ$-\ $10^\circ$) in orange and trace magnetic field spectrum (in green) plotted with respect to frequency for the data interval 17 Jan, 2007.}
\label{fig:comparison_plot}
\end{figure}
\\
To observe how the magnetic power spectrum scales with respect to that of parallel and perpendicular electric field fluctuations, we plot the three spectra together for the 2007 interval, as shown in Fig.~\ref{fig:comparison_plot}. It can be seen that the trace power spectrum of magnetic field fluctuations follows the perpendicular electric power spectrum upto the frequency of approximately $0.3$ Hz. Quantitatively, the spectral indices are measured to be $-1.54$ for the former and $-1.43$ for the latter in the range of frequencies $0.01 - 0.3$ Hz. After around $\sim 0.3$ Hz, magnetic field spectrum becomes steeper via smooth transition unlike both the electric field spectra. Furthermore, it can be seen that both the magnetic and perpendicular electric fluctuations spectrum have a breakpoint near $\sim 0.6$ Hz after which the magnetic spectrum becomes more steeper than the electric one. Due to the unavailability of high cadence data, it is difficult to conclude reasonably for the frequency range above 1 Hz \citep{Mozer}. In the discussion section of \citep{Mozer}, they state that the parallel electric field spectrum coincides with trace magnetic field spectrum which is not expected for Alfv\'enic turbulence in the inertial range. They observe identical slopes for both the magnetic and parallel electric field spectra possibly because of using the global mean magnetic field and not the LMF. Interestingly, with the use of local mean magnetic field in our study, we find that parallel electric spectrum does not coincide with that of magnetic field but the perpendicular spectrum, which is normally expected for Alfv\'enic turbulence in the solar wind \citep{Bianetal}.
\begin{figure*}
\includegraphics[width=\linewidth]{si_tr_theta.png}
\caption{(a), (b): Variation of average trace powers of electric field (in blue) with the angle between average flow and local mean magnetic field for both the data intervals from 2002 (at frequency 21 mHz) and 2007 (at frequency 11 mHz) respectively. (c), (d): Power law indices (in orange) plotted against the flow-to-field angles, fitted for frequency range $0.043 - 0.086$ Hz for 2002 data and frequency range $0.011 - 0.043$ Hz for 2007 data interval.}
\label{fig:si_tr}
\end{figure*}
Next, we present the average electric powers (normalized by parallel power) and spectral indices as a function of angles between the average flow direction and the local mean magnetic field. Following \citet{Wicks11}, we calculate the errors in average powers from the respective standard deviation and those in spectral indices from the errors introduced by least square fitting. For the interval from 02 February 2002 (Fig.~\ref{fig:si_tr} (a)), it is clearly evident that the power in antiparallel electric fluctuations is approximately 4 times more than that in the perpendicular direction. Similarly, for the data interval from 17 January, 2007 (Fig.~\ref{fig:si_tr} (b)) the parallel power is found to be greater than the perpendicular one. Interestingly, we can see nearly similar behaviour for the angle variation of trace powers during the high solar activity and low solar activity but the shape of these variations is different from that obtained for magnetic field fluctuations as reported in \citet{Horbury}.However the variation of spectral indices with angles is found to follow roughly the same nature as reported by \citet{Horbury}. \citet{Mozer} reported in their paper that the parallel electric powers can be greater than or comparable to the perpendicular powers. But from our scale dependent analysis of both the intervals, we can firmly say that the parallel electric powers are greater than the perpendicular powers. As for the spectral index anisotropy, shown in Fig.~\ref{fig:si_tr} (c), (d) the parallel (or antiparallel) electric field spectrum is found to be steeper than the perpendicular spectrum for both the intervals from two different periods of solar cycle. Qualitatively, this feature of power laws of electric fluctuations spectrum in two different angle ranges is similar to that found for magnetic fluctuations spectrum. Out of two types of anisotropies studied in this paper for electric field fluctuations, it is interesting to note that the power anisotropy is opposite and spectral index anisotropy is similar to what has been found for magnetic field fluctuations.
To conclude, our study clearly shows that (i) electric fluctuations spectrum undergoes prominent power and spectral index anisotropies when the analysis is done with respect to the local mean field, (ii) unlike \citet{Mozer}, the parallel electric power spectrum does not share the same slope as that of the magnetic power spectrum thereby indicating the clear signature of Alfv\'enic turbulence \citep{Bianetal}. \citet{Bianetal} (and with in references) also mention that the presence of the parallel electric field produced by Alfv\'enic turbulence and its further study through theory and observations might be helpful for understanding the turbulent acceleration of particles through wave-particle interactions. \citet{Kellogg} mentioned that diffusion due to the electric fields is the dominant process resulting in the isotropy of the proton distributions. Hence, the study of spectral structure of electric field in direction parallel and perpendicular to mean magnetic field could also shed some light on the particle distribution in space. With the launch of future missions and the availability of higher resolution data, it will be possible to probe into the anisotropy in the smaller scales. Another possible future work can consist of the use and analysis of the multi-spacecraft data of Cluster to study the anisotropy of electric power using the structure functions method following \citet{Chen2011}.
\section*{Acknowledgements}
This research has been supported by the DST INSPIRE research grant (DST/PHY/2017514).
\section*{Data Availability}
The data used for this study is available at the Cluster Science Archive (\url{https://csa.esac.esa.int/csa/aio/html/home_main.shtml}).
\bibliographystyle{mnras}
|
\section{Introduction}
\input{pages/intro}
\section{Design Considerations}
\input{pages/design}
\section{Implementation}
\input{pages/implementation}
\section{Conclusion}
\input{pages/conclusion}
\bibliographystyle{ACM-Reference-Format}
\subsection{Data Flow}
Raw data is rarely consumed in its original form - it often needs to go through a series of transformation, aggregation, and enrichment steps before it can be acted upon. In ODF, we make a clear distinction between the following two types of data:
\begin{itemize}
\item \textbf{Source data}, represented in ODF by \textbf{Root Datasets}, comes directly from external systems. The organization that publishes such data has complete authority over it and is fully accountable for its veracity.
\item \textbf{Derivative data}, represented in ODF by \textbf{Derivative Data-sets} is produced by transforming and combining other data. It is secondary to the source data but is equally important since derivative data is what's being presented to the decision-makers, used for training models, or fed into various automation.
\end{itemize}
The multi-stage process through which source data is transformed into actionable (derivative) data has many forms. It can manifest itself as reporting chains where people perform recurrent analysis and summarization tasks (e.g. payroll, company management reports), or as complex automated workflows (e.g. enterprise data pipelines). In ODF, we find it more intuitive and effective to reason about these processes as computational graphs (see Figure \ref{fig:dataset_kinds}), where data continuously flows from its sources to consumers via multiple transformation stages.
\subsection{The Foundation of Collaboration}
As shown previously, verifying the validity of derivative data can be cost-prohibitive compared to re-doing the work from scratch. This is the reason why complex multi-stage data processing chains mainly exist within the boundaries of organizations, where employees can implicitly trust one-another. Collaboration on data between the organizations today often leads to centralization - creation of data spaces where access control, audit and user privileges can be enforced. But centralization is time-consuming, expensive, and often impossible, as different parties want to maintain full control and ownership of their data.
Collaboration on data in a decentralized setting, such as research, remains unsolved. The validity of source data used in publications is often left unchecked during the review process, and the issue of trust hinges mostly on the reputation of the authors or the institution they represent. This may unfairly penalize young researchers and lesser known universities when it comes to publishing their work.
It is therefore ODF's goal to reduce the time it takes consumers to establish validity of derivative data down to minutes and enable collaboration on data in decentralized environments where complete trust between participants is not possible.
A modern epitome of effective collaboration between independent parties is the Open-Source Software ecosystem. What started as an exchange of ideas within a close-knit community quickly grew into a worldwide movement. The emergence of distributed version control systems (DVCS) \cite{otte2009version} ensured that the collaboration could function even at a rapidly growing scale. Many recently developed data management systems try to apply some of the ideas behind DVCS, like diff-based history and branching, directly to data \cite{dolt,deltalake}. Instead of borrowing specific technical decisions, however, in ODF we tried to understand what properties of DVCS made them so good when it comes to collaboration. We think the answer is that they help to \textit{build trust}. They can build trust even between people who have never met before, who may not even share a common language. They do so via properties like tracking the complete history of changes, allowing to attribute any change to its author, to safely roll back any previous change, and by making any malicious behavior easy to expose and remedy.
Applying these ideas to data, we have identified three pillars that would build trust and are essential to collaboration around data:
\begin{itemize}
\item \textbf{Reproducibility} - the ability to reproduce the results is the cornerstone of the scientific method without which the process and findings of one party cannot be followed by others.
\item \textbf{Verifiability} - when considering to use some data for a project it is essential to be able to understand whether it comes from a reliable publisher and whether it truthfully declares which transformations it underwent.
\item \textbf{Provenance} - regardless of how many transformation stages the data went through, it's important to be able to trace any individual value back to its source and understand what data contributed to its existence and its value.
\end{itemize}
\subsection{Reproducibility and Verifiability in Source Data}
The path towards better reproducibility and verifiability has to start at the source. The requirements here are very simple: two different parties at different times should be able to access the exact same data and validate that this data comes unaltered from the trusted source. Surprisingly, a majority of data sources today fail to satisfy these basic requirements.
Consider for example GeoNames \footnote{GeoNames. https://www.geonames.org/} and NaturalEarth \footnote{NaturalEarth. http://naturalearthdata.com/} - a few of the major publishers of open GIS data. On a periodic basis they provide datasets containing the latest known state of their domain. These state "snapshots" are published destructively by overwriting all the previous data. Naturally, everyone who uses the same URL to download data from these sources at different times is likely to get somewhat different data.
Another example is the NYC Open Data Catalog \footnote{NYC Open Data. https://opendata.cityofnewyork.us/}, which includes some datasets with a full history of certain events. Such "ledger" only grows over time as new events get appended. This approach is much better than snapshotting as it never loses data, however, it is still quite hard for two parties to obtain the same data. Whoever downloads data first is likely to have a subset of the ledger obtained by one who downloads data later. Consumers have to rely on other means to coordinate which subset of the ledger they will be using (e.g. counting rows, using timestamps, record identifiers), which is highly unreliable (see Section 3.4).
The problems of reproducibility and verifiability of modern data start at the source data, as the bad practices like state snapshots and destructive updates are considered to be the norm for many major data publishers today. For our purposes however we have identified two crucial properties we'd like to achieve:
\begin{itemize}
\item It must be possible to obtain a stable reference to data that can be shared between parties and used to obtain the same data at any future point in time.
\item Data source has to provide a mechanism to ensure that data obtained this way was not maliciously or accidentally altered.
\end{itemize}
One of the most prevalent strategy for achieving these properties today is to copy the entire dataset from the source onto a durable storage and assign it a version or a unique identifier. This approach is very common in the enterprise data science and popularized by data management tools like Quilt \footnote{Quilt. https://quiltdata.com/} and DVC \footnote{DVC. https://dvc.org/}. While it may work well in the closed, trusted environments such as enterprises, this approach is not suitable for a distributed setting, when working with open data, and with fast-moving data sources. Once copied, versioned snapshots essentially become fully independent datasets since no mechanism exists to reliably link them to the trusted source. Such copies contribute to the overall noise, only exacerbate the problem of provenance, and should be avoided.
Similar issues arise in many modern attempts to create various data hubs and data portals, such as Dataverse \footnote{Dataverse. https://dataverse.org/}, DataWorld \footnote{DataWorld. https://data.world/}, Knoema \footnote{Knoema. https://knoema.com/}, and increasingly popular data sharing on platforms like Kaggle \footnote{Kaggle. https://www.kaggle.com/datasets} and GitHub \footnote{GitHub. https://github.com/}. While the goal of simplifying discovery and federating data is noble, without a mechanism to link that data back to the trusted source all they do in fact is create more disjoint and non-trustworthy datasets.
To satisfy these properties in ODF, we changed our perspective on what "data" means to us, and make its definition more strict.
\subsection{Data Model}
\paragraph{Events}
ODF was primarily designed for mission-critical data. When data is used to gain insight and drive decision-making, discarding or modifying data is akin to rewriting history. The history of all data observed by the system must be preserved.
In the ODF, data is treated as a ledger of historical records. History only grows, it is never deleted or altered - thus all data that gets into the system is immutable. ODF treats all records in data as events or relational propositions that were believed to be true \textit{at a specific time}.
This view entirely rejects the idea of storing state snapshot data since it lacks the necessary properties. Building on the ideas of Event Sourcing \cite{EventSourcing} and Stream-Table Duality \cite{StreamTableDuality} we treat state data as a byproduct of history, which can always be reconstructed by \textit{projecting the events onto the time axis}. State data is therefore considered as a simple optimization for queries that operate with \textit{current time projections}.
\paragraph{Bitemporality}
Storing historical events alone is not enough to implement stable data references. As we mentioned previously, data consumers could agree to use a subset of the data history by utilizing some internal properties of the dataset (e.g. event timestamps, identifiers, or record counting), but it's highly error-prone.
Imagine a scenario where two parties agree to use the event timestamps to delimit the data and use all events until current time T. The reproducibility in this case can be compromised by multiple factors, such as the delay it takes data to appear in the dataset, publisher performing back-fills of the events for a certain period prior to T, or corrective events with time less than T being emitted in future upon discovering errors in some old data
To prevent such situations and issue stable references to data, ODF applies the ideas of Bitemporal Data Modelling \cite{date2002temporal,JohnstonTom2014BDTa} to the event records. We augment the schema with an extra "System Time" column, which contains the time when an event first entered the system. System Time is guaranteed to be monotonically increasing, so a stable reference to data can be as simple as having and ID of the dataset and a system time timestamp. Note that this requirement does not apply for Event Time.
\textbf{Event time}, therefore, tells us when something happened from the outside world's perspective, and is usually the most useful one for querying and joining data. \textbf{System time}, on the other hand, gives us a reference point for when something has occurred from the perspective of the system and allows us to establish before-after relationships for data within one dataset and datasets that are part of the same computation graph.
\subsection{Reproducibility and Verifiability in Derivative Data}
For derivative data, which is obtained by transforming and combining data from other datasets, reproducibility can be thought of as having an ability to repeat all transformation steps and obtain the same results as the original. Conversely, verifiability is an ability to ensure that the data presented to you as the result of some transformation was produced without being accidentally or maliciously altered. Verifiability allows us to assess trustworthiness in two simple steps: ensuring all source data comes from reliable publishers, and auditing all applied transformations.
From these definitions, we can extract the following requirements: \textbf{determinism} - all transformations should be guaranteed to result in the same output given the same input, and \textbf{transparency} - all transformations should be known.
These requirements are simple but extremely hard to meet in modern data science. A typical project can consist of hundreds of moving parts such as frameworks and libraries, operating systems, and hardware. Most of these components aren't purposely built with determinism in mind, so the burden of achieving reproducibility lies entirely on the person who implements the project. Achieving determinism is therefore a non-trivial problem that requires deep understanding of the execution environment and eliminating all sources of randomness. It is not surprising that such a significant undertaking is often left out completely to meet the project timelines. As a result, the modern data science is currently in a state of \textit{reproducibility crisis} \cite{MiyakawaTsuyoshi2020Nrdn, baker2016reproducibility}.
We believe that there is no way of getting around this problem - data processing systems have to be built with reproducibility in mind. Until determinism becomes an intrinsic property of such systems, we employ a series of techniques to remove as much burden as possible to ensure reproducibility (see Section 4).
\paragraph{Derivative Data Transience} Considering that source data is immutable and all derivative transformations are deterministic, derivative data of any transformation graph in ODF can be fully reconstructed by starting from the source data and re-applying all transformations. The derived data thus can be considered as a \textit{form of caching}. As we will discuss later, this approach can potentially reduce overall data storage costs globally, since such data doesn't need to be stored durably or be heavily replicated.
\subsection{Applying Stream Processing to Historical Data}
Data science today is dominated by the batch processing workflows, whose key characteristic is treating data as a finite set of records. But a significant portion of data is not static - new data points are constantly being produced and datasets are continuously updated. The more data-driven we are, the more we will continue to the push data processing towards real-time speeds. Applying the same batch processing techniques to recent data, however, is a harmful oversimplification that exposes us to the multitude of problems associated with temporal data: data arriving late, arriving out of order, accounting for corrections that could be issued for data that has been already processed, misalignment of data arrival cadences between datasets that are being joined etc. The reliance of batch processing on data completeness often results in incorrect results, with errors concentrated in recent data - data everyone cares about the most. What's worse, as new data arrives batch workflows may produce different results for the same time periods, effectively overwriting the history - special care has to be taken to make such corrections explicit.
It would be practically impossible to write a conventional batch processing logic that correctly handles all the temporal edge cases on a mass scale. Even if we would write such logic, its complexity would negate the benefits of verifiability - what's the use of being able to audit the transformation code that is too complex to fully understand?
In the past few years, there has been some major advancements in the field of streaming data processing. Systems like Google Data Flow \cite{akidau2015dataflow}, Apache Spark \cite{zaharia2016apache}, Apache Flink \cite{carbone2015apache}, and Naiad and its modern implementation Timely Dataflow \cite{murray2013naiad} have developed a great apparatus for dealing with many of these problems in a very intuitive way. Stream processing paradigm acknowledges a simple fact - that data processing is a balancing act between latency and correctness. It gives user the power to control this trade-off in a fully automated way.
\begin{figure*}[htb]
\centering
\includegraphics[width=0.9\linewidth]{images/watermarks_vs_time.png}
\caption{Bitemporality and watermarks in event streams}
\label{fig:watermarks_vs_time}
\end{figure*}
One of the key mechanisms of stream processing is called the \textbf{watermark}. Watermark is a type of metadata that flows along regular data in the stream and tells the processing system that at certain system time $Ts$ with a high probability $P$ the system has observed all events prior to event time $Te$. The Figure \ref{fig:watermarks_vs_time} represents the watermark as a curve on a bitemporal event diagram that separates data that is late by an expected amount and data that is exceptionally late. Watermarks can be predictive (a system can constantly adjust it based on the observed difference between event and system time), or it can be set manually (e.g. an owner of the dataset can "hint" all data consumers that the data for the time period prior $Te$ has been fully entered). Using watermarks a stream processing system can delay the processing by the exact amount needed to handle out-of-order and late data and let users perform computations in the event time space, which is a lot more natural and easier to reason about than the arrival time. The exceptional cases of late data and backfills can also be dealt with automatically and explicitly by issuing correction or retraction events or recording that certain data points were ignored. All this makes stream processing a lot more autonomous, reliable, and composable than batch.
Conventionally, stream processing is employed to build highly responsive systems that process near real-time data. In ODF, however, we saw many benefits in applying the semantics of stream processing to historical data, even data that is updated very infrequently. Every dataset in ODF is treated as a potentially infinite stream of events. Expanding on the idea that batch processing is a special case of stream processing \cite{batchisstreaming}, we use stream processing as our primary data transformation method.
ODF does not prescribe any specific language or framework and aims to support multiple implementations. Our first two prototype transformation engines, which we will cover in Section 4.3, use streaming dialects of SQL. An example streaming SQL query that is using stream-to-stream anti-join to detect late shipments can be seen below:
\begin{minipage}{0.95\linewidth}
\begin{lstlisting}[
language=SQL]
SELECT o.order_time, o.order_id
FROM orders as o
LEFT JOIN shipments as s
ON o.order_id = s.order_id
AND s.shipment_time BETWEEN
o.order_time AND o.order_time + INTERVAL '1' WEEK
WHERE s.shipment_id = NULL
\end{lstlisting}
\end{minipage}
\vfill
The benefits of this approach include:
\begin{itemize}
\item Users can define a query once and potentially run it forever. This allows us to minimize the latency with which data propagates through the system.
\item Streaming queries are expressive and are closer to "which question is being asked" as opposed to "how to compute the result". They are usually much more concise than equivalent batch queries.
\item Queries can be expressed in a way that is agnostic of how and how often the new data arrives. Whether the data is ingested once a month in Gigabyte batches, in micro-batches every hour, or as a true near real-time stream - processing logic can stay the same, produce the same results, and guarantee the best propagation times possible.
\item Streaming queries are declarative, while batch processing is usually imperative. Declarative nature lets us perform static analysis of queries and automatically derive provenance in many cases without tracking any extra data.
\item High-level abstractions like windowing, watermarks, and triggers allow users to produce maximally correct results within the configurable latency window, preventing cascading error effects typically seen in similar batch workflows.
\end{itemize}
\begin{figure*}[htb]
\centering
\includegraphics[width=0.8\linewidth]{images/metadata_slices}
\caption{Metadata captures all dataset life-cycle events}
\label{fig:metadata_slices}
\end{figure*}
The property of low latency that emerges from the described combination of our data model and stream processing is worth highlighting. Currently, we often see situations where critical data (e.g. employment situation reports, or COVID-19 cases reports in the early days of pandemic) is released so infrequently that it always has a dramatic effect on the stock markets, or prompts over-corrective actions from leaders and governments. We believe that the prevalence of batch workflows is a major contributing factor that adds significant delays to every data transformation stage, no matter how well-automated it is. By using stream processing and allowing people to define transformations in a way agnostic to how often data arrives the end-to-end propagation time of data can be reduced from weeks/months to mere seconds. As new data arrives, it is immediately made available to the consumer in its most usable form. This satisfies one the main guiding design principles of ODF that the frequency with which data is presented to consumers should be optimized for their experience and usability, not dictated by limitations of the data pipeline.
\subsection{Metadata Tracking}
In ODF, data never appears in the system alone, as we would not be able to tell whether it can be trusted. Metadata, therefore, is an essential part of a dataset. It contains every aspect of where the data came from, how it was transformed, and everything that ever influenced how data looks like throughout its entire lifetime (see Figure \ref{fig:metadata_slices}). We think of metadata as a digital passport of data that is instrumental to reproducibility, verifiability, and data provenance.
So far, when talking about datasets, we have been assuming that the inputs, the transformations, the schema of the result and more were constant. While new events can be ingested and would propagate through the system, the processing graph itself was frozen in time. This was a deliberate oversimplification.
As the nature of businesses change, new requirements arrive, defects are detected and fixed - it's not a matter of if but when the time comes to make changes. Calling data a potentially infinite stream would not make sense without providing a way to improve and evolve it over time. Having to create a new dataset every time you need to change the schema or update the transformation would mean that the entire downstream processing graph would have to be rebuilt from scratch.
In order to support backward-compatible schema changes, evolution of transformations over time, and to provide a way to correct past mistakes in data and queries we have developed a ledger-based mechanism for recording the dataset metadata over time. We will look at it in more detail when discussing the Metadata Chain (see Section \ref{sec:metadata_chain}).
\subsection{Provenance}
An ability to easily understand how a specific data point came to be is crucial for building trust and showing that data can be relied upon. While verifiability can tell us which data sources were used to produce the results and which transformations were performed, this is often too coarse-grained. Provenance, on the other hand, is the ability to trace a specific piece of data back to its ultimate source, understand which events have directly contributed to its value, and what data was considered to determine its existence in the output.
Provenance \cite{cheney2009provenance} is an is extensively studied problem in modern data science \cite{simmhan2005survey}, but in practice it still didn't fully manifest in any widely applied system or a state of the art enterprise data pipeline. Most solutions implement it only on the dataset level, as a simplified form called \textit{lineage}. Granular provenance is much challenging to achieve as it needs to span through virtually every component of the data pipeline, and potentially across many independent systems. Even if fully implemented, its value would be limited due to mutable nature of data in many modern data systems.
We believe that not being able to answer provenance questions fast can undermine the credibility of even fully trustworthy data, so ODF was designed with complete provenance in mind and supports it on multiple levels. Firstly, the metadata tracking creates a link between output and input data blocks for every iteration of a transformation. This lets ODF improve provenance granularity by limiting the search space from entire datasets to small blocks of data within them. Since both data and metadata are immutable, this link is never lost. Secondly, the declarative nature of the streaming transformations allows us to easily analyze the structure of queries. For simple map-style queries, provenance can be derived automatically without tracking any additional information. Thirdly, for more complex queries we require provenance to be supported by an underlying data processing engine \cite{provenanceSQL}. ODF defines the provenance query API that specific engine implementations need to implement (see Section \ref{sec:engine}).
\begin{figure*}[ht]
\centering
\includegraphics[width=0.9\linewidth]{images/metadata_chain}
\caption{Metadata as a chain of immutable blocks}
\label{fig:metadata_chain}
\end{figure*}
\subsection{Data Sharing}
As a distributed protocol, ODF was designed with data sharing efficiency in mind and the previously covered features directly contribute towards this goal.
The source data is irreducible by definition. ODF makes no assumptions that this data can be retrieved from anywhere else in case its lost, thus every peer that publishes root datasets is responsible for storing them durably, in a replicated and highly-available way.
The derivative data, as we covered, is considered transient; therefore all parties that publish derivative datasets can use the cheapest data hosting available (or no hosting at all) without the need for durability or heavy replication.
The immutability property of both data and metadata ensures they can be easily and safely replicated without complex synchronization mechanisms. Metadata being cryptographically linked to the raw data means that there is no need to encrypt the data itself or use a secure channel for distributing it unless we want data to remain private. It is usually sufficient to securely distribute the metadata and use it to establish the authenticity of the downloaded data. Metadata is several orders of magnitude smaller than associated data, so it can be easily hosted and widely shared.
With all these properties combined, ODF has the potential to significantly reduce data storage and distribution costs globally, compared to the widespread copy-and-version approach that often results in circulation of thousands of similar copies of a dataset taken at different points in time.
\subsection{Data Ingestion}
Root datasets are the points of entry for external data into the system. Our vision is that eventually all root datasets will be owned by organizations with full authority over certain data (i.e. trusted publishers) and will be provided directly in ODF-compliant format.
As a fallback mechanism, we also provide a way to link a root dataset to some external source (e.g. using the URL) and periodically ingest its data into the system. Such configuration duplicates the data, but it is necessary to guarantee that the properties of data are preserved and insulate the rest of the system from a wide range of current bad practices in data publishing.
For data sources that always preserve the entire history ODF has to do little but copy and de-duplicate the data with records that were ingested previously. For data sources that publish data in a destructive way (e.g. periodic state snapshots) ODF takes on the task of "historization" - transforming state data into event form using the Change Data Capture \cite{PetrieKevin2018SCDC,ankorion2005change} (CDC) techniques.
\subsection{Metadata Chain}
\label{sec:metadata_chain}
The Metadata Chain's purpose is to capture all information and events that influenced the way data looks right now. This includes: where the data comes from, how it was processed, its schema, and current watermark (see Figure \ref{fig:metadata_chain}).
Its design borrows heavily from the existing ledger-based systems such as blockchain \cite{zheng2017overview} and version control systems \cite{otte2009version}. Just like all data in ODF, the metadata chain is append-only and immutable. It consists of individual blocks that are cryptographically linked together and also contain hashes of the associated data slices, so the overall data structure is similar to a Merkle Tree \cite{10.10073-540-48184232}.
Metadata chain is designed for extensibility and can carry other kinds of information, such as:
\begin{itemize}
\item Extra meaning and structure of knowledge (semantics, ontology)
\item Relevant policies, terms, rules, compliance, and regulations (governance)
\item License, privacy and security concerns (stewardship)
\item Information that aids discovery
\item Interoperability data to connect ODF to other ledger-based systems
\end{itemize}
We see it as a crucial building block that will allow us to collectively push the quality of data further by standardizing and automating best data science and engineering practices.
\begin{figure*}[htb]
\centering
\includegraphics[width=0.85\linewidth]{images/transform.png}
\caption{Engine execution environment}
\label{fig:transform}
\end{figure*}
\subsection{Engine}
\label{sec:engine}
Data processing technologies are evolving rapidly. As a data exchange protocol, we would like ODF to outlive most of them and be able to adapt to new technologies as they emerge. We also want ODF to be inclusive of any data processing languages and dialects used in different branches of data science. Therefore we designed ODF to be unopinionated as to which language is used to define transformations or which framework performs them, as long as they satisfy all the criteria necessary for the protocol to function (e.g. determinism).
A specific implementation of ODF's data processing contract is called an \textbf{Engine}. Engines are responsible for applying queries defined in the datasets to input data and returning the result. For example, the two of our prototype engines based on Apache Spark \cite{zaharia2016apache} and Apache Flink \cite{carbone2015apache} allow us to transform data using a series of Streaming SQL \cite{calcite} statements. Since engines are in full control of all data transformations, they are also responsible for answering the provenance queries.
ODF takes a few extra steps to guarantee the deterministic and reproducible properties of transformations done by the engines.
\subsubsection{Execution Environment}
Engines run in a fully isolated environment called the "sandbox", implemented using the OCI Containers \cite{oci} technology (see Figure \ref{fig:transform}). Sandbox is designed to prevent engines from accessing any external resources except for the inputs and outputs of a current transformation (e.g. on the Internet or user's file system) as potential sources of undesired non-determinism.
This may sound very restrictive, and it is. After all many common data processing tasks like geolocation rely on the use of external APIs which would be inaccessible under the sandbox model. However, we strongly believe that this is a necessary step in managing data correctly. External resources like APIs are run by companies that can disappear over night, they also often evolve without following strict versioning policies - it is impossible to achieve reproducible results in such environment. Running any "black box" operations like API calls would require us to re-classify derivative datasets as source data and admit that such data is non-reproducible. Instead we envision that the software algorithms and ML models used by such transformations will be incorporated into the ODF as the engine extensions or pure data, and this transition will be one of the focus points of our future research.
\subsubsection{Engine Versioning}
To further strengthen the reproducibility guarantees of the system we associate every transformation with an exact version of an engine that was used to perform it. This excludes the possibility of any code changes in the engine producing different results than what was originally observed. For this purpose we use the full SHA digest of the engine's OCI image.
As dataset evolves over time it may start depending on too many different versions of a certain engine, unsustainably increasing the amount of images user needs to download to fully validate the dataset. We use a special \textit{engine upgrade} procedure to remedy this problem.
\subsubsection{Checkpoints \& Watermarks}
Some computations over the input data like windowed aggregations or temporal joins may require engine to maintain some state. Since engine is required to fully consume input data during transformation (as it will never see it again) some of this state may need to be preserved in between the invocations of a query. For this purpose ODF allows engines to maintain \textit{checkpoints} - a piece of opaque and fully engine-specific data used to store intermediate state. Along with data and metadata, checkpoints are an integral part of a dataset. If checkpoint is lost the entire computation will have to be restarted from scratch.
\begin{figure*}[htb!]
\centering
\includegraphics[width=0.9\linewidth]{images/watermarks_in_stream.png}
\caption{Example use of watermarks in the event stream with an allowed lateness of t=4.}
\label{fig:watermarks_in_stream}
\end{figure*}
Every dataset in ODF also has a watermark \cite{akidau2015dataflow} - a metadata tuple $(Ts, Te)$ we described earlier. For example, if a root dataset receives new data on a monthly basis, its watermark can lag by over a month behind the wall clock time. It effectively prevents all derivative processing from proceeding past that time point until the data arrives, ensuring the correctness of computations. Figures \ref{fig:watermarks_vs_time} and \ref{fig:watermarks_in_stream} shows examples of how watermarks can prevent late processing errors in event streams. Watermarks can be set on root datasets both via fixed offset or manually, and they then fully automatically propagate through derivative datasets based on the nature of transformations. Watermarks are elevated from the checkpoints into the metadata as they are an important consumer-facing property.
\subsection{Coordinator}
The coordinator is an application responsible for maintaining the invariants and transactional semantics of the system. It handles all operations related to the metadata chain and guarantees its integrity and validity. The coordinator implements data ingestion and data sharing logic, but delegates all data processing to the engines.
Complexity of the metadata management is fully contained in the coordinator. From an engine's perspective the transformations look just like conventional stream processing, so the underlying data processing frameworks don't require any customizations or being made ODF-aware. Turning an existing data processing library into an engine is a matter of creating a thin wrapper around it that conforms to the ODF engine interface.
Besides the query execution logic, the coordinator also expects engines to implement a few extra operations related to the dataset life-cycle, such as input schema change, query change, and an engine upgrade handlers.
\section{Introduction}
\input{pages/intro}
\section{Design Considerations}
\input{pages/design}
\section{Implementation}
\input{pages/implementation}
\section{Conclusion}
\input{pages/conclusion}
\bibliographystyle{ACM-Reference-Format}
\subsection{Data Flow}
Raw data is rarely consumed in its original form - it often needs to go through a series of transformation, aggregation, and enrichment steps before it can be acted upon. In ODF, we make a clear distinction between the following two types of data:
\begin{itemize}
\item \textbf{Source data}, represented in ODF by \textbf{Root Datasets}, comes directly from external systems. The organization that publishes such data has complete authority over it and is fully accountable for its veracity.
\item \textbf{Derivative data}, represented in ODF by \textbf{Derivative Data-sets} is produced by transforming and combining other data. It is secondary to the source data but is equally important since derivative data is what's being presented to the decision-makers, used for training models, or fed into various automation.
\end{itemize}
The multi-stage process through which source data is transformed into actionable (derivative) data has many forms. It can manifest itself as reporting chains where people perform recurrent analysis and summarization tasks (e.g. payroll, company management reports), or as complex automated workflows (e.g. enterprise data pipelines). In ODF, we find it more intuitive and effective to reason about these processes as computational graphs (see Figure \ref{fig:dataset_kinds}), where data continuously flows from its sources to consumers via multiple transformation stages.
\subsection{The Foundation of Collaboration}
As shown previously, verifying the validity of derivative data can be cost-prohibitive compared to re-doing the work from scratch. This is the reason why complex multi-stage data processing chains mainly exist within the boundaries of organizations, where employees can implicitly trust one-another. Collaboration on data between the organizations today often leads to centralization - creation of data spaces where access control, audit and user privileges can be enforced. But centralization is time-consuming, expensive, and often impossible, as different parties want to maintain full control and ownership of their data.
Collaboration on data in a decentralized setting, such as research, remains unsolved. The validity of source data used in publications is often left unchecked during the review process, and the issue of trust hinges mostly on the reputation of the authors or the institution they represent. This may unfairly penalize young researchers and lesser known universities when it comes to publishing their work.
It is therefore ODF's goal to reduce the time it takes consumers to establish validity of derivative data down to minutes and enable collaboration on data in decentralized environments where complete trust between participants is not possible.
A modern epitome of effective collaboration between independent parties is the Open-Source Software ecosystem. What started as an exchange of ideas within a close-knit community quickly grew into a worldwide movement. The emergence of distributed version control systems (DVCS) \cite{otte2009version} ensured that the collaboration could function even at a rapidly growing scale. Many recently developed data management systems try to apply some of the ideas behind DVCS, like diff-based history and branching, directly to data \cite{dolt,deltalake}. Instead of borrowing specific technical decisions, however, in ODF we tried to understand what properties of DVCS made them so good when it comes to collaboration. We think the answer is that they help to \textit{build trust}. They can build trust even between people who have never met before, who may not even share a common language. They do so via properties like tracking the complete history of changes, allowing to attribute any change to its author, to safely roll back any previous change, and by making any malicious behavior easy to expose and remedy.
Applying these ideas to data, we have identified three pillars that would build trust and are essential to collaboration around data:
\begin{itemize}
\item \textbf{Reproducibility} - the ability to reproduce the results is the cornerstone of the scientific method without which the process and findings of one party cannot be followed by others.
\item \textbf{Verifiability} - when considering to use some data for a project it is essential to be able to understand whether it comes from a reliable publisher and whether it truthfully declares which transformations it underwent.
\item \textbf{Provenance} - regardless of how many transformation stages the data went through, it's important to be able to trace any individual value back to its source and understand what data contributed to its existence and its value.
\end{itemize}
\subsection{Reproducibility and Verifiability in Source Data}
The path towards better reproducibility and verifiability has to start at the source. The requirements here are very simple: two different parties at different times should be able to access the exact same data and validate that this data comes unaltered from the trusted source. Surprisingly, a majority of data sources today fail to satisfy these basic requirements.
Consider for example GeoNames \footnote{GeoNames. https://www.geonames.org/} and NaturalEarth \footnote{NaturalEarth. http://naturalearthdata.com/} - a few of the major publishers of open GIS data. On a periodic basis they provide datasets containing the latest known state of their domain. These state "snapshots" are published destructively by overwriting all the previous data. Naturally, everyone who uses the same URL to download data from these sources at different times is likely to get somewhat different data.
Another example is the NYC Open Data Catalog \footnote{NYC Open Data. https://opendata.cityofnewyork.us/}, which includes some datasets with a full history of certain events. Such "ledger" only grows over time as new events get appended. This approach is much better than snapshotting as it never loses data, however, it is still quite hard for two parties to obtain the same data. Whoever downloads data first is likely to have a subset of the ledger obtained by one who downloads data later. Consumers have to rely on other means to coordinate which subset of the ledger they will be using (e.g. counting rows, using timestamps, record identifiers), which is highly unreliable (see Section 3.4).
The problems of reproducibility and verifiability of modern data start at the source data, as the bad practices like state snapshots and destructive updates are considered to be the norm for many major data publishers today. For our purposes however we have identified two crucial properties we'd like to achieve:
\begin{itemize}
\item It must be possible to obtain a stable reference to data that can be shared between parties and used to obtain the same data at any future point in time.
\item Data source has to provide a mechanism to ensure that data obtained this way was not maliciously or accidentally altered.
\end{itemize}
One of the most prevalent strategy for achieving these properties today is to copy the entire dataset from the source onto a durable storage and assign it a version or a unique identifier. This approach is very common in the enterprise data science and popularized by data management tools like Quilt \footnote{Quilt. https://quiltdata.com/} and DVC \footnote{DVC. https://dvc.org/}. While it may work well in the closed, trusted environments such as enterprises, this approach is not suitable for a distributed setting, when working with open data, and with fast-moving data sources. Once copied, versioned snapshots essentially become fully independent datasets since no mechanism exists to reliably link them to the trusted source. Such copies contribute to the overall noise, only exacerbate the problem of provenance, and should be avoided.
Similar issues arise in many modern attempts to create various data hubs and data portals, such as Dataverse \footnote{Dataverse. https://dataverse.org/}, DataWorld \footnote{DataWorld. https://data.world/}, Knoema \footnote{Knoema. https://knoema.com/}, and increasingly popular data sharing on platforms like Kaggle \footnote{Kaggle. https://www.kaggle.com/datasets} and GitHub \footnote{GitHub. https://github.com/}. While the goal of simplifying discovery and federating data is noble, without a mechanism to link that data back to the trusted source all they do in fact is create more disjoint and non-trustworthy datasets.
To satisfy these properties in ODF, we changed our perspective on what "data" means to us, and make its definition more strict.
\subsection{Data Model}
\paragraph{Events}
ODF was primarily designed for mission-critical data. When data is used to gain insight and drive decision-making, discarding or modifying data is akin to rewriting history. The history of all data observed by the system must be preserved.
In the ODF, data is treated as a ledger of historical records. History only grows, it is never deleted or altered - thus all data that gets into the system is immutable. ODF treats all records in data as events or relational propositions that were believed to be true \textit{at a specific time}.
This view entirely rejects the idea of storing state snapshot data since it lacks the necessary properties. Building on the ideas of Event Sourcing \cite{EventSourcing} and Stream-Table Duality \cite{StreamTableDuality} we treat state data as a byproduct of history, which can always be reconstructed by \textit{projecting the events onto the time axis}. State data is therefore considered as a simple optimization for queries that operate with \textit{current time projections}.
\paragraph{Bitemporality}
Storing historical events alone is not enough to implement stable data references. As we mentioned previously, data consumers could agree to use a subset of the data history by utilizing some internal properties of the dataset (e.g. event timestamps, identifiers, or record counting), but it's highly error-prone.
Imagine a scenario where two parties agree to use the event timestamps to delimit the data and use all events until current time T. The reproducibility in this case can be compromised by multiple factors, such as the delay it takes data to appear in the dataset, publisher performing back-fills of the events for a certain period prior to T, or corrective events with time less than T being emitted in future upon discovering errors in some old data
To prevent such situations and issue stable references to data, ODF applies the ideas of Bitemporal Data Modelling \cite{date2002temporal,JohnstonTom2014BDTa} to the event records. We augment the schema with an extra "System Time" column, which contains the time when an event first entered the system. System Time is guaranteed to be monotonically increasing, so a stable reference to data can be as simple as having and ID of the dataset and a system time timestamp. Note that this requirement does not apply for Event Time.
\textbf{Event time}, therefore, tells us when something happened from the outside world's perspective, and is usually the most useful one for querying and joining data. \textbf{System time}, on the other hand, gives us a reference point for when something has occurred from the perspective of the system and allows us to establish before-after relationships for data within one dataset and datasets that are part of the same computation graph.
\subsection{Reproducibility and Verifiability in Derivative Data}
For derivative data, which is obtained by transforming and combining data from other datasets, reproducibility can be thought of as having an ability to repeat all transformation steps and obtain the same results as the original. Conversely, verifiability is an ability to ensure that the data presented to you as the result of some transformation was produced without being accidentally or maliciously altered. Verifiability allows us to assess trustworthiness in two simple steps: ensuring all source data comes from reliable publishers, and auditing all applied transformations.
From these definitions, we can extract the following requirements: \textbf{determinism} - all transformations should be guaranteed to result in the same output given the same input, and \textbf{transparency} - all transformations should be known.
These requirements are simple but extremely hard to meet in modern data science. A typical project can consist of hundreds of moving parts such as frameworks and libraries, operating systems, and hardware. Most of these components aren't purposely built with determinism in mind, so the burden of achieving reproducibility lies entirely on the person who implements the project. Achieving determinism is therefore a non-trivial problem that requires deep understanding of the execution environment and eliminating all sources of randomness. It is not surprising that such a significant undertaking is often left out completely to meet the project timelines. As a result, the modern data science is currently in a state of \textit{reproducibility crisis} \cite{MiyakawaTsuyoshi2020Nrdn, baker2016reproducibility}.
We believe that there is no way of getting around this problem - data processing systems have to be built with reproducibility in mind. Until determinism becomes an intrinsic property of such systems, we employ a series of techniques to remove as much burden as possible to ensure reproducibility (see Section 4).
\paragraph{Derivative Data Transience} Considering that source data is immutable and all derivative transformations are deterministic, derivative data of any transformation graph in ODF can be fully reconstructed by starting from the source data and re-applying all transformations. The derived data thus can be considered as a \textit{form of caching}. As we will discuss later, this approach can potentially reduce overall data storage costs globally, since such data doesn't need to be stored durably or be heavily replicated.
\subsection{Applying Stream Processing to Historical Data}
Data science today is dominated by the batch processing workflows, whose key characteristic is treating data as a finite set of records. But a significant portion of data is not static - new data points are constantly being produced and datasets are continuously updated. The more data-driven we are, the more we will continue to the push data processing towards real-time speeds. Applying the same batch processing techniques to recent data, however, is a harmful oversimplification that exposes us to the multitude of problems associated with temporal data: data arriving late, arriving out of order, accounting for corrections that could be issued for data that has been already processed, misalignment of data arrival cadences between datasets that are being joined etc. The reliance of batch processing on data completeness often results in incorrect results, with errors concentrated in recent data - data everyone cares about the most. What's worse, as new data arrives batch workflows may produce different results for the same time periods, effectively overwriting the history - special care has to be taken to make such corrections explicit.
It would be practically impossible to write a conventional batch processing logic that correctly handles all the temporal edge cases on a mass scale. Even if we would write such logic, its complexity would negate the benefits of verifiability - what's the use of being able to audit the transformation code that is too complex to fully understand?
In the past few years, there has been some major advancements in the field of streaming data processing. Systems like Google Data Flow \cite{akidau2015dataflow}, Apache Spark \cite{zaharia2016apache}, Apache Flink \cite{carbone2015apache}, and Naiad and its modern implementation Timely Dataflow \cite{murray2013naiad} have developed a great apparatus for dealing with many of these problems in a very intuitive way. Stream processing paradigm acknowledges a simple fact - that data processing is a balancing act between latency and correctness. It gives user the power to control this trade-off in a fully automated way.
\begin{figure*}[htb]
\centering
\includegraphics[width=0.9\linewidth]{images/watermarks_vs_time.png}
\caption{Bitemporality and watermarks in event streams}
\label{fig:watermarks_vs_time}
\end{figure*}
One of the key mechanisms of stream processing is called the \textbf{watermark}. Watermark is a type of metadata that flows along regular data in the stream and tells the processing system that at certain system time $Ts$ with a high probability $P$ the system has observed all events prior to event time $Te$. The Figure \ref{fig:watermarks_vs_time} represents the watermark as a curve on a bitemporal event diagram that separates data that is late by an expected amount and data that is exceptionally late. Watermarks can be predictive (a system can constantly adjust it based on the observed difference between event and system time), or it can be set manually (e.g. an owner of the dataset can "hint" all data consumers that the data for the time period prior $Te$ has been fully entered). Using watermarks a stream processing system can delay the processing by the exact amount needed to handle out-of-order and late data and let users perform computations in the event time space, which is a lot more natural and easier to reason about than the arrival time. The exceptional cases of late data and backfills can also be dealt with automatically and explicitly by issuing correction or retraction events or recording that certain data points were ignored. All this makes stream processing a lot more autonomous, reliable, and composable than batch.
Conventionally, stream processing is employed to build highly responsive systems that process near real-time data. In ODF, however, we saw many benefits in applying the semantics of stream processing to historical data, even data that is updated very infrequently. Every dataset in ODF is treated as a potentially infinite stream of events. Expanding on the idea that batch processing is a special case of stream processing \cite{batchisstreaming}, we use stream processing as our primary data transformation method.
ODF does not prescribe any specific language or framework and aims to support multiple implementations. Our first two prototype transformation engines, which we will cover in Section 4.3, use streaming dialects of SQL. An example streaming SQL query that is using stream-to-stream anti-join to detect late shipments can be seen below:
\begin{minipage}{0.95\linewidth}
\begin{lstlisting}[
language=SQL]
SELECT o.order_time, o.order_id
FROM orders as o
LEFT JOIN shipments as s
ON o.order_id = s.order_id
AND s.shipment_time BETWEEN
o.order_time AND o.order_time + INTERVAL '1' WEEK
WHERE s.shipment_id = NULL
\end{lstlisting}
\end{minipage}
\vfill
The benefits of this approach include:
\begin{itemize}
\item Users can define a query once and potentially run it forever. This allows us to minimize the latency with which data propagates through the system.
\item Streaming queries are expressive and are closer to "which question is being asked" as opposed to "how to compute the result". They are usually much more concise than equivalent batch queries.
\item Queries can be expressed in a way that is agnostic of how and how often the new data arrives. Whether the data is ingested once a month in Gigabyte batches, in micro-batches every hour, or as a true near real-time stream - processing logic can stay the same, produce the same results, and guarantee the best propagation times possible.
\item Streaming queries are declarative, while batch processing is usually imperative. Declarative nature lets us perform static analysis of queries and automatically derive provenance in many cases without tracking any extra data.
\item High-level abstractions like windowing, watermarks, and triggers allow users to produce maximally correct results within the configurable latency window, preventing cascading error effects typically seen in similar batch workflows.
\end{itemize}
\begin{figure*}[htb]
\centering
\includegraphics[width=0.8\linewidth]{images/metadata_slices}
\caption{Metadata captures all dataset life-cycle events}
\label{fig:metadata_slices}
\end{figure*}
The property of low latency that emerges from the described combination of our data model and stream processing is worth highlighting. Currently, we often see situations where critical data (e.g. employment situation reports, or COVID-19 cases reports in the early days of pandemic) is released so infrequently that it always has a dramatic effect on the stock markets, or prompts over-corrective actions from leaders and governments. We believe that the prevalence of batch workflows is a major contributing factor that adds significant delays to every data transformation stage, no matter how well-automated it is. By using stream processing and allowing people to define transformations in a way agnostic to how often data arrives the end-to-end propagation time of data can be reduced from weeks/months to mere seconds. As new data arrives, it is immediately made available to the consumer in its most usable form. This satisfies one the main guiding design principles of ODF that the frequency with which data is presented to consumers should be optimized for their experience and usability, not dictated by limitations of the data pipeline.
\subsection{Metadata Tracking}
In ODF, data never appears in the system alone, as we would not be able to tell whether it can be trusted. Metadata, therefore, is an essential part of a dataset. It contains every aspect of where the data came from, how it was transformed, and everything that ever influenced how data looks like throughout its entire lifetime (see Figure \ref{fig:metadata_slices}). We think of metadata as a digital passport of data that is instrumental to reproducibility, verifiability, and data provenance.
So far, when talking about datasets, we have been assuming that the inputs, the transformations, the schema of the result and more were constant. While new events can be ingested and would propagate through the system, the processing graph itself was frozen in time. This was a deliberate oversimplification.
As the nature of businesses change, new requirements arrive, defects are detected and fixed - it's not a matter of if but when the time comes to make changes. Calling data a potentially infinite stream would not make sense without providing a way to improve and evolve it over time. Having to create a new dataset every time you need to change the schema or update the transformation would mean that the entire downstream processing graph would have to be rebuilt from scratch.
In order to support backward-compatible schema changes, evolution of transformations over time, and to provide a way to correct past mistakes in data and queries we have developed a ledger-based mechanism for recording the dataset metadata over time. We will look at it in more detail when discussing the Metadata Chain (see Section \ref{sec:metadata_chain}).
\subsection{Provenance}
An ability to easily understand how a specific data point came to be is crucial for building trust and showing that data can be relied upon. While verifiability can tell us which data sources were used to produce the results and which transformations were performed, this is often too coarse-grained. Provenance, on the other hand, is the ability to trace a specific piece of data back to its ultimate source, understand which events have directly contributed to its value, and what data was considered to determine its existence in the output.
Provenance \cite{cheney2009provenance} is an is extensively studied problem in modern data science \cite{simmhan2005survey}, but in practice it still didn't fully manifest in any widely applied system or a state of the art enterprise data pipeline. Most solutions implement it only on the dataset level, as a simplified form called \textit{lineage}. Granular provenance is much challenging to achieve as it needs to span through virtually every component of the data pipeline, and potentially across many independent systems. Even if fully implemented, its value would be limited due to mutable nature of data in many modern data systems.
We believe that not being able to answer provenance questions fast can undermine the credibility of even fully trustworthy data, so ODF was designed with complete provenance in mind and supports it on multiple levels. Firstly, the metadata tracking creates a link between output and input data blocks for every iteration of a transformation. This lets ODF improve provenance granularity by limiting the search space from entire datasets to small blocks of data within them. Since both data and metadata are immutable, this link is never lost. Secondly, the declarative nature of the streaming transformations allows us to easily analyze the structure of queries. For simple map-style queries, provenance can be derived automatically without tracking any additional information. Thirdly, for more complex queries we require provenance to be supported by an underlying data processing engine \cite{provenanceSQL}. ODF defines the provenance query API that specific engine implementations need to implement (see Section \ref{sec:engine}).
\begin{figure*}[ht]
\centering
\includegraphics[width=0.9\linewidth]{images/metadata_chain}
\caption{Metadata as a chain of immutable blocks}
\label{fig:metadata_chain}
\end{figure*}
\subsection{Data Sharing}
As a distributed protocol, ODF was designed with data sharing efficiency in mind and the previously covered features directly contribute towards this goal.
The source data is irreducible by definition. ODF makes no assumptions that this data can be retrieved from anywhere else in case its lost, thus every peer that publishes root datasets is responsible for storing them durably, in a replicated and highly-available way.
The derivative data, as we covered, is considered transient; therefore all parties that publish derivative datasets can use the cheapest data hosting available (or no hosting at all) without the need for durability or heavy replication.
The immutability property of both data and metadata ensures they can be easily and safely replicated without complex synchronization mechanisms. Metadata being cryptographically linked to the raw data means that there is no need to encrypt the data itself or use a secure channel for distributing it unless we want data to remain private. It is usually sufficient to securely distribute the metadata and use it to establish the authenticity of the downloaded data. Metadata is several orders of magnitude smaller than associated data, so it can be easily hosted and widely shared.
With all these properties combined, ODF has the potential to significantly reduce data storage and distribution costs globally, compared to the widespread copy-and-version approach that often results in circulation of thousands of similar copies of a dataset taken at different points in time.
\subsection{Data Ingestion}
Root datasets are the points of entry for external data into the system. Our vision is that eventually all root datasets will be owned by organizations with full authority over certain data (i.e. trusted publishers) and will be provided directly in ODF-compliant format.
As a fallback mechanism, we also provide a way to link a root dataset to some external source (e.g. using the URL) and periodically ingest its data into the system. Such configuration duplicates the data, but it is necessary to guarantee that the properties of data are preserved and insulate the rest of the system from a wide range of current bad practices in data publishing.
For data sources that always preserve the entire history ODF has to do little but copy and de-duplicate the data with records that were ingested previously. For data sources that publish data in a destructive way (e.g. periodic state snapshots) ODF takes on the task of "historization" - transforming state data into event form using the Change Data Capture \cite{PetrieKevin2018SCDC,ankorion2005change} (CDC) techniques.
\subsection{Metadata Chain}
\label{sec:metadata_chain}
The Metadata Chain's purpose is to capture all information and events that influenced the way data looks right now. This includes: where the data comes from, how it was processed, its schema, and current watermark (see Figure \ref{fig:metadata_chain}).
Its design borrows heavily from the existing ledger-based systems such as blockchain \cite{zheng2017overview} and version control systems \cite{otte2009version}. Just like all data in ODF, the metadata chain is append-only and immutable. It consists of individual blocks that are cryptographically linked together and also contain hashes of the associated data slices, so the overall data structure is similar to a Merkle Tree \cite{10.10073-540-48184232}.
Metadata chain is designed for extensibility and can carry other kinds of information, such as:
\begin{itemize}
\item Extra meaning and structure of knowledge (semantics, ontology)
\item Relevant policies, terms, rules, compliance, and regulations (governance)
\item License, privacy and security concerns (stewardship)
\item Information that aids discovery
\item Interoperability data to connect ODF to other ledger-based systems
\end{itemize}
We see it as a crucial building block that will allow us to collectively push the quality of data further by standardizing and automating best data science and engineering practices.
\begin{figure*}[htb]
\centering
\includegraphics[width=0.85\linewidth]{images/transform.png}
\caption{Engine execution environment}
\label{fig:transform}
\end{figure*}
\subsection{Engine}
\label{sec:engine}
Data processing technologies are evolving rapidly. As a data exchange protocol, we would like ODF to outlive most of them and be able to adapt to new technologies as they emerge. We also want ODF to be inclusive of any data processing languages and dialects used in different branches of data science. Therefore we designed ODF to be unopinionated as to which language is used to define transformations or which framework performs them, as long as they satisfy all the criteria necessary for the protocol to function (e.g. determinism).
A specific implementation of ODF's data processing contract is called an \textbf{Engine}. Engines are responsible for applying queries defined in the datasets to input data and returning the result. For example, the two of our prototype engines based on Apache Spark \cite{zaharia2016apache} and Apache Flink \cite{carbone2015apache} allow us to transform data using a series of Streaming SQL \cite{calcite} statements. Since engines are in full control of all data transformations, they are also responsible for answering the provenance queries.
ODF takes a few extra steps to guarantee the deterministic and reproducible properties of transformations done by the engines.
\subsubsection{Execution Environment}
Engines run in a fully isolated environment called the "sandbox", implemented using the OCI Containers \cite{oci} technology (see Figure \ref{fig:transform}). Sandbox is designed to prevent engines from accessing any external resources except for the inputs and outputs of a current transformation (e.g. on the Internet or user's file system) as potential sources of undesired non-determinism.
This may sound very restrictive, and it is. After all many common data processing tasks like geolocation rely on the use of external APIs which would be inaccessible under the sandbox model. However, we strongly believe that this is a necessary step in managing data correctly. External resources like APIs are run by companies that can disappear over night, they also often evolve without following strict versioning policies - it is impossible to achieve reproducible results in such environment. Running any "black box" operations like API calls would require us to re-classify derivative datasets as source data and admit that such data is non-reproducible. Instead we envision that the software algorithms and ML models used by such transformations will be incorporated into the ODF as the engine extensions or pure data, and this transition will be one of the focus points of our future research.
\subsubsection{Engine Versioning}
To further strengthen the reproducibility guarantees of the system we associate every transformation with an exact version of an engine that was used to perform it. This excludes the possibility of any code changes in the engine producing different results than what was originally observed. For this purpose we use the full SHA digest of the engine's OCI image.
As dataset evolves over time it may start depending on too many different versions of a certain engine, unsustainably increasing the amount of images user needs to download to fully validate the dataset. We use a special \textit{engine upgrade} procedure to remedy this problem.
\subsubsection{Checkpoints \& Watermarks}
Some computations over the input data like windowed aggregations or temporal joins may require engine to maintain some state. Since engine is required to fully consume input data during transformation (as it will never see it again) some of this state may need to be preserved in between the invocations of a query. For this purpose ODF allows engines to maintain \textit{checkpoints} - a piece of opaque and fully engine-specific data used to store intermediate state. Along with data and metadata, checkpoints are an integral part of a dataset. If checkpoint is lost the entire computation will have to be restarted from scratch.
\begin{figure*}[htb!]
\centering
\includegraphics[width=0.9\linewidth]{images/watermarks_in_stream.png}
\caption{Example use of watermarks in the event stream with an allowed lateness of t=4.}
\label{fig:watermarks_in_stream}
\end{figure*}
Every dataset in ODF also has a watermark \cite{akidau2015dataflow} - a metadata tuple $(Ts, Te)$ we described earlier. For example, if a root dataset receives new data on a monthly basis, its watermark can lag by over a month behind the wall clock time. It effectively prevents all derivative processing from proceeding past that time point until the data arrives, ensuring the correctness of computations. Figures \ref{fig:watermarks_vs_time} and \ref{fig:watermarks_in_stream} shows examples of how watermarks can prevent late processing errors in event streams. Watermarks can be set on root datasets both via fixed offset or manually, and they then fully automatically propagate through derivative datasets based on the nature of transformations. Watermarks are elevated from the checkpoints into the metadata as they are an important consumer-facing property.
\subsection{Coordinator}
The coordinator is an application responsible for maintaining the invariants and transactional semantics of the system. It handles all operations related to the metadata chain and guarantees its integrity and validity. The coordinator implements data ingestion and data sharing logic, but delegates all data processing to the engines.
Complexity of the metadata management is fully contained in the coordinator. From an engine's perspective the transformations look just like conventional stream processing, so the underlying data processing frameworks don't require any customizations or being made ODF-aware. Turning an existing data processing library into an engine is a matter of creating a thin wrapper around it that conforms to the ODF engine interface.
Besides the query execution logic, the coordinator also expects engines to implement a few extra operations related to the dataset life-cycle, such as input schema change, query change, and an engine upgrade handlers.
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.